6B parameters.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhu et al., 2021)\", \"Explanation\": \"The cited work by Zhu et al. (2021) provides a method for fine-tuning model parameters using a norm constraint on weight changes, which the citing paper adopts in its implementation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wei et al., 2022)\", \"Explanation\": \"The cited work by Wei et al. (2022) provides a method called chain-of-thought (CoT) prompting, which the citing paper adopts to maximize model performance in their research.\"}]"}}},{"rowIdx":70,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b4","b5","b6","b13","b14","b15"],"table_ref":[],"text":"In contemporary times, the Language Model (LM) [1; 2] has emerged as a pivotal player in the field of Natural Language Processing (NLP). It accomplishes this by unifying a range of diverse NLP tasks into a shared prompt paradigm [3; 4]. To elaborate, a LM, unsupervised trained via conditional language modeling on a substantial volume of non-annotated data collected from the web, can reformulate various downstream tasks into fitting textual prompts. These prompts contain slots, the word probabilities of which are calculated using the pre-trained LM. This process obviates the necessity for gradient updates to the parameters of LM, thus mitigating the challenges associated with additional data collection and fine-tuning. To further enhance the effectiveness of this paradigm, the few-shot prompt (or in-context learning) [5] is proposed where a few examples are provided as additional contexts to guide the LM to generate the desired result.\nWitnessing the success of the prompt paradigm in the NLP field, there has been a concerted effort by researchers to replicate its function in the Vision-Language Model (VLM). To facilitate this, Flamingo [6] is proposed to align well-trained large-scale vision and language models through some trainable cross-modal adapters. Consequently, the resultant model is capable of addressing Vision-Language (VL) tasks by processing a prompt sequence, which includes several interleaved image and text examples for in-context learning. Since the primary objective of Flamingo is to build a VLM for the few-shot prompt, they only apply a straightforward strategy to configure the in-context A women riding a bike on the bike path.\n[Good Caption]\nPeople riding on a street on a street.\n[Bad Caption]\nPeople walking down a street and a street.\n[Bad Caption]\nThe man is riding a pink bike threw traffic.\n[Good Caption]\nA women riding a bike on the bike path.\n[Good Caption]\nPeople riding on a street on a street.\n[Bad Caption]\nPeople walking down a street and a street.\n[Bad Caption]\nThe man is riding a pink bike threw traffic. A kitchen with a stove and a stove.\n[Bad Caption]\nA kitchen scene with a fridge and an oven.\n[Good Caption] People ride on elephants through a field.\n[Good Caption] Some elephants are walking and walking.\n[Bad Caption]\nA kitchen with a stove and a stove.\n[Bad Caption]\nA kitchen scene with a fridge and an oven.\n[Good Caption]\nA group of people riding on a city street.\n[CIDEr 72.9]\nA bicycler riding in the bike lane next to traffic.\n[CIDEr 126 .8] (e) (f)\nA group of people riding on a city street.\n[CIDEr 72.9]\nA bicycler riding in the bike lane next to traffic.\n[CIDEr 126 .8] (e) (f)\nFigure 1: The distinction between LM and VLMs as few-shot learners. LM generally excel with examples akin to the test case (blue blocks in (a)). In contrast, for VLMs, the performance is not strictly correlated with image similarity but heavily relies on the caption quality. For instance, when low-quality captions are used, similar images (d) lead to worse performance than dissimilar ones (f) since VLMs may build a short-cut by reusing in-context captions without seeing the given images.\nsequence by randomly sampling a few image-text pairs. Nevertheless, a plethora of studies within the NLP field [7] have demonstrated that diverse in-context configurations lead to dramatic effects on few-shot performance, e.g., the selection or ordering of in-context samples [8; 9; 10], while only a limited number of studies systematically explore such effects in the VL case.\nTo narrow this gap, we explore the effects of various in-context configurations on the performance of few-shot VL tasks. Among various VL tasks, Image Captioning (IC) aims at generating a text conditioned on the source image, and thus can be considered as the visually-conditioned LM. Just as a multitude of NLP tasks can be recast as LM tasks, IC performs a similar function [11; 12; 13], which motivates our decision to select IC as the subject of our case study. However, unlike in NLP, where only single-modal texts are considered in the in-context configuration, in IC, the synergy between multiple modalities significantly influences the performance. For instance, our experiments revealed that selecting images similar to the test image for the in-context sequence does not always lead to good performance, a result closely tied to the quality of the associated captions. Figure 1 shows the comparison between LM and VLM as the few-shot learners.\nConsequently, we design diverse ways to select images as the in-context images. After selection, the captions of different qualities are assigned to these images for constructing the multi-modal in-context sequence. By combining diverse image selection and caption assignment techniques, we undertake a comprehensive exploration of the effects of multi-modal mutual synergy on VL in-context captioning. We implement all experiments using the prevalent captioning dataset, MSCOCO [14], employing its training set as the database for image selection. Through extensive evaluation of various image selection and caption assignment strategies, we uncover two counter-intuitive yet valuable insights. (1) Caption quality is determined by descriptiveness and language patterns, but their influence on in-context captioning performance is unequal. When captions adequately describe salient image objects, simpler language patterns may yield better results.\n(2) The efficacy of similar images depends on the quality of the paired captions. Excessive similarity might cause VLM to create a short-cut inference [15] from in-context captions, potentially misleading the model with low-quality captions. Beyond these findings, we introduce a practical in-context captioning strategy, Iterative Prompting, for cases with limited or no Ground-Truth Captions. Furthermore, when Ground-Truth Captions are available, we recommend using Model-Generated Captions as anchors to identify which Ground-Truth Caption is a more suitable in-context caption. Experimental results indicate that even when utilizing low-quality model-generated captions, there is an average CIDEr improvement of 7.3. Moreover, in optimal conditions, the average enhancement reaches up to 20.9 points compared to the random sampling baseline.\nSince the Flamingo code is not publicly available, our experiments mainly utilize the unofficial Open-Flamingo [16] 2 . It's worth noting that the performance of Open-Flamingo is not on par with the official Flamingo due to its training on less data."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b2","b16","b0","b17","b1","b33","b40","b5","b15","b10","b41","b35","b42","b43","b2","b49","b50","b49","b5","b53"],"table_ref":[],"text":"Prompting Language Model (LM). The research paradigms of NLP have encountered two sea changes in the past few years [3]. The first one is the LM that is pre-trained by predicting the next word conditioned on observed words and can be fine-tuned for solving various downstream tasks, including GPT [17], BERT [1], and BART [18]. The second sea change is the emergence of the prompt paradigm, which was introduced with GPT-3 [2]. Within this paradigm, a pre-trained LM does not require fine-tuning to solve downstream tasks; instead, tasks are reformulated into appropriate prompts with empty slots to be filled. Subsequently, more advanced prompt-based techniques have been proposed, including prompt-tuning [19; 20; 21] and Chain-of-Thought [22; 23; 24; 25].\nPrompting Vision-Language Model (VLM). In contrast to NLP, the Vision-Language (VL) domain has made significant strides in the first sea change, as evident by the development of various VL-BERT models. These models leverage large volumes of web-collected image-caption pairs to learn VL-generalizable embeddings [26; 27; 28; 29; 30; 31; 32; 33]. However, the prompt paradigm, despite revolutionizing NLP studies, only appears when a certain scale of the model is reached [34]. This scale prerequisite poses further challenges to the development of a VLM with prompt and in-context learning ability.\nTo mitigate training burdens, instead of updating all the parameters of a VLM [35; 36], some VLMs [35; 36; 37] freeze well-trained Language Models and only train a smaller network, referred to as adapters [38; 39; 40], to align the pre-trained vision and language models. Inspired by these models, both Frozen [41] and Flamingo [6] evolve into multi-modal few-shot learners by training vision and cross-modal adapters, respectively. Given its superior in-context learning ability, we use Flamingo to explore the effects of various in-context configurations [16].\nThere are also models that address VL tasks through in-context learning. For instance, PICa [11] utilizes captions as mediators to construct an in-context text for solving Visual Question Answering (VQA) tasks, while this may lose the mutual synergy in the representation space of different modalities. The models proposed in [42] and [36] both fine-tune VLMs for specific VQA tasks, but lack the generalized few-shot prompt ability for other VL tasks. UNIFIED-IO [43]demonstrates a unified approach to a myriad of tasks, from classical computer vision to natural language processing, without task-specific fine-tuning. And ProGrad [44] introduces a technique to prevent prompt tuning from forgetting pre-trained vision-language models' general knowledge by only updating aligned prompts, outperforming other methods in various few-shot learning scenarios.\nExploring In-Context Configurations in NLP. Upon observing that pre-trained LMs are good fewshot learners, researchers also discover that diverse in-context configurations have dramatic effects on performance [3]. This observation sparks numerous studies aimed at determining the optimal in-context configurations, such as the format of the in-context examples [45; 21; 5], the selection of these examples [8; 9; 46; 47], and even the order in which these examples are presented [48; 49; 10]. However, these studies are predominantly conducted within the NLP field and fail to consider the unique characteristics and complexity of multi-modal data. In order to address this gap, we propose a series of strategies to explore the effects of various multi-modal in-context sequence configurations.\nImage Caption. Image Captioning (IC) [50] aims at correctly verbalizing one image using descriptive languages, which can be solved through both retrieve [51] and generation [50], the former retrieves complete sentences from a corpus, while the latter generates words sequentially. Researchers have recently combined these approaches by first retrieving image-caption pairs and inputting them into generation models [52; 53]. This process resembles in-context captioning, which also retrieves image-caption pairs to help captioning. However, in contrast to them [52; 53], our work introduces novel image and caption selection strategies to study in-context captioning and these methods can also enhance traditional retrieval-generation methods. \nwhere the probability P (•) is calculated by a pre-trained Vision-Language Model (VLM) (e.g., Flamingo [6] or Otter [54]).\nVarious studies in NLP [8; 45; 48; 46] field have shown that the performance of in-context learning varies significantly with different in-context configurations. We explore these effects in the case of Image Captioning (IC). Unlike pure NLP tasks, IC is a dual-modal task, which makes it more complex than NLP tasks. Specifically, the mutual synergy of image-caption examples must be considered in in-context learning, as opposed to considering the images or the captions independently. We next respectively introduce the image selection (cf. 3.1) and caption assignment (cf. 3.2) strategies used to configure in-context image-caption pairs."},{"figure_ref":[],"heading":"Selecting Images","publication_ref":[],"table_ref":[],"text":"Random Selection (RS). Given a set D = {(I 1 , C 1 ), ..., (I M , C M )} with M image-caption pairs, we randomly sample n images as {I 1 , ..., I N } in S."},{"figure_ref":["fig_3","fig_3","fig_3","fig_3"],"heading":"Similarity-based Image-Image Retrieval (SIIR).","publication_ref":["b34","b54","b55","b34","b56"],"table_ref":[],"text":"Certain NLP studies suggest that performance can be enhanced by retrieving examples similar to the test case [8; 46; 47]. Following this approach, we retrieve n images with the highest similarity scores to the test image Î from D. We employ two methods to compute these similarities: 1) SIIR-CLIP (Figure 2 (a)): Using the vision encoder of CLIP [35], we extract image embeddings to determine similarities. 2) SIIR-TAG (Figure 2 (b)): We utilize some scene graph extractors to derive tags for each image. Specifically, we employ Vinvl [55] to extract objects and their attributes, and IETrans [56] to determine the relations present within the image. Following this extraction, we conduct an AND operation to compute the similarity between tags.\nSimilarity-based Image-Caption Retrieval (SICR-CLIP) (Figure 2 (d)). Taking advantage of the cross-modal retrieval capability of CLIP [35], we use its vision and language encoders to embed images and captions into a shared space for computing similarities. Given Î, we calculate its cross-modal embedding similarities with {C 1 , ..., C M } ∈ D, and select the images whose captions have top-n similarities with Î. Note that we use different kinds of captions as mediators; the methods to generate these captions will be detailed in Section 3.2. These captions are also used as\nC 1 , ..., C N ∈ S.\nDiversity-based Image-Image Retrieval (DIIR). For a single test sample, beside using similar incontext examples, some NLP studies find that diversity is also crucial [57]. Consequently, we retrieve a diverse set of images from D to configure S. Specifically, we employ the following strategies: 1) DIIR-TR: We extract discrete tags from the aforementioned VinVL and IETrans. Then we randomly divide the tags into N clusters and apply SIIR-TAG to retrieve the most similar image from each cluster. 2) DIIR-TT (Figure 2 (c)): To conveniently control the number of shots, we incorporate class tags generated by the IETrans model, extending the basis of SIIR-Tags to four categories: object, class, attribute, and relation. We then employ SIIR-TAG to identify the top-k similar images within each category, allowing us to create 4k-shot images in S. Note that both DIIR methods take into account a certain level of similarity during retrieval, rather than selecting entirely distinct images."},{"figure_ref":[],"heading":"Assigning Captions","publication_ref":[],"table_ref":[],"text":"After selecting images, we need to assign one caption to each image to construct in-context imagecaption pairs. To explore the effects of multi-modal mutual synergy, we use captions with diverse qualities as introduced in the following."},{"figure_ref":[],"heading":"Ground-Truth Captions (GTC).","publication_ref":["b13"],"table_ref":[],"text":"In the MSCOCO dataset [14], each image has 5 GTCs and we simply use the first one in S."},{"figure_ref":[],"heading":"Model-Generated Captions (MGC).","publication_ref":["b57","b13","b58","b15","b5"],"table_ref":[],"text":"Compared with GTC, MGC has lower quality due to two disadvantages. Firstly, MGC uses poorer language, e.g., simple words or rigid sentence patterns.\nSecondly, MGC has less descriptiveness that it may mis-verbalize or miss certain salient vision patterns of an image. However, we will see these two disadvantages do not equally cause worse performance compared with GTC in in-context captioning and surprisingly, we find that sometimes simple words or rigid sentence patterns even help generate good captions.\nHere we apply two different models to generate the captions with diverse qualities. MGC-TF@X: a Transformer is trained from scratch by using VinVL features and X denotes the CIDEr score [58] on the test set. To get captions of different qualities, we use the checkpoints from different training epochs and totally generate three kinds of MGCs. 1) MGC-TF@66 contains grammar mistakes while can describe the most salient objects. 2) MGC-TF@88 can use relatively correct grammar to describe the salient objects. 3) MGC-TF@135 is generated by a well-trained Transformer, i.e. , the loss converges. MGC-VLM(N )@X: Another way to get MGCs is to use VLM in a N -shot manner.\nTo achieve this, for each image I ∈ D, we treat I as the test image and then use Eq. ( 1) to generate a corresponding caption C conditioned on S constructed by only N image-caption pairs. As a result, two kinds of captions are got which are MGC-VLM(0)@63 and MGC-VLM(32)@81.\nBoth two ways can construct the set D with M image-caption pairs. Then for a novel test image Î, we can select some image-caption pairs from D by some above-mentioned approaches, e.g., using RS or SIIR-CLIP to get images and assigning the MGCs to the image, to configure S for generating a new caption. Compared with MGC-TF, MGC-VLM is more practical as it addresses scenarios where only a handful or even no human-labelled captions are available, which means we do not have enough data to train a Transformer from scratch.\nIteratively Prompting (IP). For MGC-VLM introduced before, one natural extension is IP. To achieve this, in the first iteration, we generate a caption for each image I ∈ D by MGC-VLM. In the subsequent iteration, these generated captions are paired with the selected images to prompt VLM to generate enhanced captions. This process can be repeated across multiple iterations, thereby iteratively prompting the VLM.\nModel-Generated Captions as Anchors (MGCA). A MGC can serve not only as an in-context caption but also as an anchor for selecting a suitable caption from the five GTCs. As MGCs typically verbalize salient visual patterns in an image but may miss finer details, using them as anchors can lead to the selection of GTCs that highlight these salient patterns. Furthermore, the selected GTC may supplement interesting details about these patterns, potentially assisting VLM to generate superior captions during in-context learning. In the MGCA implementation, for each selected image, we measure the similarity between the MGC and five GTCs using CIDEr and select the GTC with the highest CIDEr score. We evaluate the proposed strategies on MSCOCO dataset [14], which is the most widely used benchmark in image captioning. We used the Karpathy split [59] in the experiments, which contains 113,287/5000/5000 training/validation/test images and each image is associated with 5 human-annotated captions.\nImplementation Details We employ the Open-Flamingo model [16] to test our strategies, setting the length penalty to -2.0 and a maximum generation length of 20. We follow Flamingo [6] to use 4, 8, 16, and 32 shots. We respectively use ViT-L/14 and as the vision and language encoders to extract image and sentence embeddings that used in SIIR-CLIP and SICR-CLIP. For MGC-TF, we train the standard Transformer encoder-decoder architecture on the MSCOCO dataset and use the checkpoints underwent 1000, 3000, and 170,000 iterations respectively. These checkpoints generate the captions with CIDEr scores of 66, 88, and 135 on the test set. We implement all experiments on a single RTX 3090 using FP16."},{"figure_ref":["fig_4","fig_5","fig_6","fig_7"],"heading":"Results and Analyses","publication_ref":[],"table_ref":[],"text":"Given our varied strategies for image selection and caption assignment, displaying results for each configuration in table format, especially at 4, 8, 16, and 32-shot levels, could become overwhelming.\nFor clarity, we've chosen to present results using line charts and histograms in the main paper, while detailed numerical outcomes are in the supplementary material. The line charts in Figures 3 and4 illustrate trends as the shot number grows. Each subplot within these figures corresponds to a unique strategy for image selection or caption assignment. Furthermore, Figures 5 and6 show histograms of average CIDEr scores for the various shot results. To facilitate comprehension, we first analyze the effects of caption qualities 4.2.1 and then of image selection strategies 4.2.2."},{"figure_ref":["fig_5","fig_5","fig_6","fig_6","fig_4","fig_4","fig_5","fig_5","fig_6","fig_6"],"heading":"Effects of Caption Qualities","publication_ref":[],"table_ref":[],"text":"Figure 4(a)-(e) reveals that performance typically improves with an increase in shot numbers. However, the rate of improvement varies, or even declines, depending on the quality of the captions used. For instance, in Figure 4(a), using Ground-Truth Captions (GTC), the increase rates of the Ground-Truth Captions (GTC) vs. Model-Generated Captions (MGC). As discussed in Sec 3.2, MGC exhibits two primary shortcomings when compared to GTC: poorer language and less descriptiveness. However, we will see that these two shortcomings do not always make MGC achieve poorer performance when compared to GTC. Specifically, we find that up to a certain level of descriptiveness, simpler sentence patterns are more easily recognized by the VLM, thereby improving caption generation.\nEvidence for this can be observed in Figure 5 by comparing GTC with MGC-TF@135. Compared to the caption generated by MGC-TF@135, ground-truth caption has better language patterns, e.g., rich vocabulary and complex sentence pattern, and better descriptiveness. Then when the selected images cannot provide enough vision cues, i.e. , when Random Selection (RS) is applied in Figure 5 (a), GTC outperforms MGC-TF@135. However, once the similarity-based retrieval methods like Similaritybased Image-Image Retrieval (SIIR) or Similarity-based Image-Caption Retrieval (SICR-CLIP) is used, the selected similar images offer useful visual patterns that help address the descriptiveness issue. Consequently, the VLM is more likely to recognize the consistent, simple patterns in MGC than the rich, diverse patterns in GTC and then generate better captions. This effect is more pronounced in 4-shot cases, where VLM lacks sufficient in-context captions to discern sentence patterns, making simpler patterns advantageous. As long as the descriptiveness issue is addressed, MGC often outperforms GTC, e.g., MGC-TF@88>GTC in Figure 3(b-d). Especially, when SICR-CLIP is employed to select captions with high cross-modal similarity to the test image, it significantly mitigates the descriptiveness problem. Then as demonstrated in Figure 3(d), even MGC-TF@66 surpasses GTC.\nMGC-TF vs. MGC-VLM. Before we see that using more low-quality captions like MGC-TF@66 will misguide VLM to generate worse captions. Yet, Figure 4 (f) indicates that using MGC-VLM(0)@63 improves performance with increased shot numbers, contrasting MGC-TF@66 in Figure 4 (c). This raises the question: why do weaker captions (MGC-VLM(0)@63) surpass those ... ... Several motor scooters are jammed into a small market street.\nA row of parked bicycles sitting in front of a store.\nRows of motor scooters are parked in front of a store.\nThis slice of cake looks like half cheesecake and half vanilla.\nA bite is taken out of a piece of cake. This slice of cake looks like half cheesecake and half vanilla cake.\nA piece of cake on a plate with a fork.\nA piece of cake on a plate with a fork.\nA piece of cake on a plate with a fork and a spoon.\nA row of motorcycles parked in front of a street.\nA group of motorcycles parked in front of a street.\nA group of motorcycles parked in front of a street.\nA piece of cake on a plate with a fork.\nA piece of cake on a plate with a fork.\nA piece of cake on a plate with a fork and a spoon.\nA row of motorcycles parked in front of a street.\nA group of motorcycles parked in front of a street.\nA group of motorcycles parked in front of a street.\n(a) MGC-TF@135(blue) v.s. GTC(red) ... ...\nA stop sign that is on a street.\nA stop sign on a street with a street.\nA stop sign on the side of a street with a street light.\nA group of boats in the water and a boat.\nA boat on a boat in the water.\nA boat with a boat in the water and a boat in the water .\nTwo sailing boats are moored in a harbour.\nA view of a lake with a boat and a dock in the foreground.\nA view of a lake with a boat and a dock in the foreground.\nA close up view of a traffic light that's red.\nA close up picture of a red traffic light.\nA picture of a traffic light that's red.\nTwo sailing boats are moored in a harbour.\nA view of a lake with a boat and a dock in the foreground.\nA view of a lake with a boat and a dock in the foreground.\nA close up view of a traffic light that's red.\nA close up picture of a red traffic light.\nA picture of a traffic light that's red.\n(b) MGC-VLM(0)@63(blue) v.s. MGC-TF@66(red) Figure 7: (a) Two examples show that GTC uses more diverse words than MGC-TF@135, making VLM hard to recognize the major pattern, e.g., it mis-generates \"a store\" in left or neglects \"a fork\" in right. (b) Two examples demonstrating how VLM is misguided by syntactic errors in MGC-VLM@66, where certain phrases are repeated such as \"a street\" or \"a boat in the water\". with higher CIDEr (MGC-TF@66)? This discrepancy aligns with our assumption: descriptiveness and language pattern influence in-context captioning differently. MGC-TF@66 excels in object detection but struggles with language decoding, causing salient objects to be identified correctly but with syntactical errors in captions. As such examples increase, VLM produces worse captions due to these errors. Conversely, MGC-VLM(0)@63, though limited in object recognition, maintains better grammar. When more vision cues are provided, VLM leverages these along with the better-formed captions from MGC-VLM(0)@63, resulting in improved captions. Model-Generated Captions as Anchors (MGCA). From Figure 5, we see that using MGC as the in-context captions usually underperforms GTC. However, by the MGCA strategy, we observe consistent improvements over GTCs, as demonstrated by the higher blue histograms of different MGCs compared to the grey dashed line. For example, despite MGC-TF@66 only achieves 64.24 CIDEr score in Figure 5(a), we still observe a 3.2 CIDEr improvement when using MGC-TF@66 as the anchor, compared to simply selecting the first GTC. Specifically, when using MGC-TF@66/MGC-TF@88/MGC-TF@135/MGC-VLM(0)@63/MGC-VLM(32)@81 as the anchors, the average improvements over six image selection strategies compared to GTC are 7.3/8.0/8.8/3.6/4.8 respectively. In contrast to solely leveraging the RS+GTC method, the combination of SIIR-CLIP and using MGC-TF@135 as anchor chalked up an average boost of 20.9.\nThe primary reason for such improvement is likely that MGC, to some extent, verbalize the major patterns, such as the salient objects of an image. This helps identify which GTC provides more detailed information about these patterns. Such detailed information of the salient objects provide help VLM generate better captions. This assumption is further supported by comparisons between MGC-TF and MGC-VLM. Given that MGC-TF prioritizes verbalizing the salient objects of an image, using MGC-TF@66/MGC-TF@88 as anchors tends to select superior GTC than MGC-VLM(0)@63/MGC-VLM(32)@81, thus yielding higher improvements. ..."},{"figure_ref":[],"heading":"A long table with a plant on top of it surrounded with wooden chairs.","publication_ref":[],"table_ref":[],"text":"A table is adorned with wooden chairs with blue accents.\nA table is adorned with wooden chairs with red accents.\n..."},{"figure_ref":[],"heading":"(a) Experiment (1)","publication_ref":[],"table_ref":[],"text":"A long table with a plant on top of it surrounded with wooden chairs.\nA table is adorned with wooden chairs with blue accents.\nBananas on a wooden table.\n..."},{"figure_ref":[],"heading":"A long table with a plant on top of it surrounded with wooden chairs.","publication_ref":[],"table_ref":[],"text":"A table is adorned with wooden chairs with blue accents.\nA car with a sign that says stop.\n..."},{"figure_ref":[],"heading":"(c) Experiment (3)","publication_ref":[],"table_ref":[],"text":"A long table with a plant on top of it surrounded with wooden chairs.\nA table is adorned with wooden chairs with blue accents.\nA long table with bananas on it.\n..."},{"figure_ref":["fig_6"],"heading":"A long table with a plant on top of it surrounded with wooden chairs.","publication_ref":[],"table_ref":[],"text":"A table is adorned with wooden chairs with blue accents.\nA stop sign on a road.\n... the second, indicating that extended VLM iterations might be redundant. Remarkably, even when confined to just 32-shot GTCs, merely two iterations of IP -for instance, where MGC-VLM(32) attains an average CIDEr of 80.5 -can rival performances seen when all GTCs are utilized, as exemplified by RS-GTC's average CIDEr of 80.04 in Figure 5(a)."},{"figure_ref":["fig_7","fig_7"],"heading":"Effects of Image Qualities","publication_ref":["b56"],"table_ref":["tab_3"],"text":"We evaluated the outcomes of several image selection techniques. SIIR-CLIP, leveraging vision embeddings to compute retrieval similarities, generally identifies images that are more analogous than those found by SIIR-TAG. This is likely attributed to the intrinsic noise present in semantic tags. SICR-CLIP, emphasizing captions that spotlight prominent objects, naturally gravitates towards images showcasing similar objects. In contrast, both DIIR-TR and DIIR-TT produce more varied selections. Nevertheless, when benchmarked against RS, every retrieval-based model consistently fetches images that exhibit greater similarity.\nAt first glance, one could reasonably infer that using more similar images would invariably lead to superior performance. Yet, as demonstrated in Figure 6, this assumption doesn't universally hold. When engaging high-quality captions, namely GTC in (a) and MGC-TF@135 in (d), the correlation stands, with more analogous images indeed translating to enhanced results. For instance, all retrieval-based techniques surpass RS in this context. However, with medium-quality captions, as in MGC-TF@88 in (b) and VLM(32)@81 in (e), only the similarity-based retrieval methods like SIIR or SICR-CLIP manage to outdo RS, and this is specifically observed in (e). Most intriguingly, when low-quality captions like MGC-TF@66 in (c) and VLM(32)@63 in (f) are in play, analogous images inversely impact performance, as illustrated by RS outperforming SIIR-CLIP in (f). This critical insight underscores a pivotal revelation: the efficacy of utilizing similar images is intricately tied to the caliber of the corresponding captions.\nSimilar Images Lead to Short-Cut Inference. A relatively bold hypothesis to elucidate this phenomenon is: when in-context images are similar to the test image, VLM may take a short-cut by leveraging in-context captions to generate a new one, rather than learning how to caption from the in-context pairs. Consequently, the greater the similarity between in-context and test images, the more the VLM is influenced by the in-context captions. To robustly test our underlying assumption, we designed three distinct experimental setups. In each setup, the incontext captions remain consistent, comprising 5 groundtruth captions sourced randomly from an image unrelated to the test image. However, the in-context images are chosen differently in each experiment: (1) they are identical to the test image; (2) they are picked using SIIR-CLIP; (3) they are chosen via RS. This progression results in decreasing similarity between the in-context and test images from experiments (1) through (3). Table 2 showcases two sets of CIDEr scores. One set compares the generated captions to the five ground-truth captions (GTC) of the test image, and the other set contrasts them with the five in-context captions (ICC). Our findings suggest a clear pattern: the closer the in-context images are to the test image, the more the VLM tends to mirror the ICC in its generated caption. For illustration, method (1) registers the highest CIDEr score when compared to the ICC. However, the caption it produces doesn't accurately depict the image, as reflected by its notably lower CIDEr score in relation to the GTC. These observations solidify our hypothesis that images with high similarity can inadvertently prompt short-cut inference. Additionally, Figure 8 provides visual representations to further elucidate the phenomenon of short-cut inference.\nDIIR. As depicted in Figure 6, the performance of the two DIIR methods noticeably lags behind SIIR-TAG. This observation underscores the contention that the strength of diversity may not always translate to superior in every context. One plausible explanation we propose is the inherent nature of captioning as a task. Contrary to certain complex NLP challenges, where diversity can be instrumental in offering a multifaceted understanding of a problem [57], captioning is relatively straightforward and may not benefit as much from diverse in-context examples. Delving deeper into the DIIR methods, DIIR-TT consistently outshines DIIR-TR. This leads us to infer that the clustering based on semantic tag types might be a more optimal strategy. Such a strategy, we believe, not only ensures diversity but also completeness in the selection of images. For instance, given a caption like \"a brown dog is running\", DIIR-TT could potentially source images that highlight elements like \"brown objects\", \"dogs\", and the \"action of running\"."},{"figure_ref":[],"heading":"Conclusion and Limitations","publication_ref":["b15","b5","b53"],"table_ref":[],"text":"In this study, we utilize image captioning as a case study to examine the impacts of varied multimodal in-context configurations on few-shot vision-language prompting. Specifically, we design 4 different ways to select images and 4 different strategies to assign captions to the selected images for constructing multi-modal in-context configurations. Exhaustive experiments reveal that, contrary to single-modal NLP cases, multi-modal mutual synergy significantly influences performance. Notably, we observe that the descriptiveness and language patterns of the captions differently affect performance: better performance may be achieved with simpler and consistent sentence patterns when selected images compensate for descriptiveness issues. And we also discover that when the in-context images are similar to the test one, VLM may build a short-cut by directly using the in-context captions instead of really learning to captioning. Moreover, our optimal strategy lead to a 20.9 average increase in CIDEr scores compared to a random sampling baseline. This study's primary limitation is that at the inception of our exploration, the only open-source multi-modal few-shot learner available is Open-Flamingo [16]. However, Open-Flamingo, when compared with the official Flamingo [6], underperforms due to its training on significantly less data. Consequently, some findings in this paper might shift if a more robust multi-modal few-shot learner is employed. Nevertheless, even in such a scenario, the diverse configuration strategies proposed in this paper maintain their utility, aiding researchers in swiftly exploring the characteristics of the employed VLMs. Additionally, we have provided experimental results on Otter [54] and smaller version of Open-Flamingo, with detailed findings available in Appendix."},{"figure_ref":[],"heading":"Supplementary Material","publication_ref":[],"table_ref":["tab_5","tab_6"],"text":"A Experimental results on Open-Flamingo v1 9B\nHere we present all the numerical results from the experiment, divided into four tables (Table 3, Table 4, Table 5, and "},{"figure_ref":[],"heading":"B Experimental results on Open-Flamingo v2 3B","publication_ref":[],"table_ref":["tab_7"],"text":"Here we present the numerical results on Open-Flamingo v2 3B model3 in Table 7 based on two different Image Selection strategies. Additionally, we have calculated the average values for each method in the \"mean\" column.\nFrom the values in the table, it can be seen that the trend is basically consistent with the v1 model. It can be considered that our strategy and analysis can be transferred to different models."},{"figure_ref":[],"heading":"C Experimental results on Otter","publication_ref":[],"table_ref":[],"text":"Here we present the numerical results on Otter model in "},{"figure_ref":["fig_14","fig_17"],"heading":"D More Results of MGC-TF@135 vs. GTC","publication_ref":[],"table_ref":[],"text":"We further elucidate the performance of MGC-TF@135(blue) and GTC(red), by offering additional examples shown in Figure 9. It can be easily observed that GTC has a more diverse range of sentence structures in comparison to MGC-TF@135. In the initial two examples, the words \"cat\" and \"dachshunds\" were inaccurately recognized by GTC, demonstrating its limitation in some specific instances. A shift can be noticed in the subsequent three examples, where the GTC generates captions with complex sentence structures, with a notable proportion of incorrect information. This finding serves to reinforce the notion that simpler sentence patterns, up to a certain degree of descriptiveness, are more readily deciphered by the VLM, thereby enhancing the quality of generated captions.\nE More Results of MGC-TF@66 vs. MGC-VLM(0)@63 Similarly, we supplement with more examples, as depicted in Figure 10, to facilitate the comparison between MGC-TF@66(red) and MGC-VLM(0)@63(blue). A striking observation from this comparative study reveals that MGC-TF@66 demonstrates a significant number of syntactical errors. This flaw becomes problematic as it misdirects VLM, resulting in a substantial volume of syntactical mistakes in the produced sentences.This correlation implies that the syntax errors in the initial input by MGC-TF@66 tend to propagate into the VLM's output. Therefore, it becomes clear that the accuracy of grammar in the prompt is crucial for achieving better results. Five examples demonstrate that more diverse words will be used in GTC than in MGC-TF@135 which making VLM hard to catch the major pattern. e.g., the first line and the second line incorrectly state \"cat\" and \"dachshunds.\" Five examples show that VLM will be misguided by the syntactic errors in MGC-VLM@66. Lots of outputs have repeated phrases, such as \"a table\", \"a clock\", \"a desk\", \"a motorcycle\" and \"a vase\"."},{"figure_ref":["fig_18"],"heading":"F More Results of short-cut inference","publication_ref":[],"table_ref":[],"text":"In this section, we provide additional examples to demonstrate how similar images can lead our VLM to exhibit a tendency for short-cut inference, where it might relies heavily on in-context captions to generate the captions, disregarding the image information in the prompt and the inherent relationships within the in-context pairs. In Figure 11, we present the results for the same test image under two different scenarios of choosing in-context images: \"identical to the test image\" (top row) and \"via random sampling\" (bottom row).\nIn the first scenario, for the majority of cases, our results are largely unrelated to the image content but similar to the in-context captions, with only a few instances capturing some elements from the image, such as \"pink\" in (a) and \"on the beach\" in (b). However, in the second scenario, our results are heavily influenced by the in-context captions, as evident in (c) and (d) with the mention of \"wooden chairs\". "},{"figure_ref":[],"heading":"(b)","publication_ref":[],"table_ref":[],"text":"A long restaurant table with rattan rounded back chairs.\nA long table with a plant on top of it surrounded with wooden chairs.\nA restaurant has modern wooden tables and chairs.\nA long table with a flower arrangement in the middle for meetings.\nA table is adorned with wooden chairs with blue accents.\nA man is running on the beach.\nA long restaurant table with rattan rounded back chairs.\nA long table with a plant on top of it surrounded with wooden chairs.\nA restaurant has modern wooden tables and chairs. "},{"figure_ref":[],"heading":"Acknowledgments","publication_ref":[],"table_ref":[],"text":"This work is supported by National Science Foundation of China (62206048), Natural Science Foundation of Jiangsu Province (BK20220819), Young Elite Scientists Sponsorship Program of Jiangsu Association for Science and Technology Tj-2022-027 and the Big Data Computing Center of Southeast University."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b4\",\n \"b5\",\n \"b6\",\n \"b13\",\n \"b14\",\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"In contemporary times, the Language Model (LM) [1; 2] has emerged as a pivotal player in the field of Natural Language Processing (NLP). It accomplishes this by unifying a range of diverse NLP tasks into a shared prompt paradigm [3; 4]. To elaborate, a LM, unsupervised trained via conditional language modeling on a substantial volume of non-annotated data collected from the web, can reformulate various downstream tasks into fitting textual prompts. These prompts contain slots, the word probabilities of which are calculated using the pre-trained LM. This process obviates the necessity for gradient updates to the parameters of LM, thus mitigating the challenges associated with additional data collection and fine-tuning. To further enhance the effectiveness of this paradigm, the few-shot prompt (or in-context learning) [5] is proposed where a few examples are provided as additional contexts to guide the LM to generate the desired result.\\nWitnessing the success of the prompt paradigm in the NLP field, there has been a concerted effort by researchers to replicate its function in the Vision-Language Model (VLM). To facilitate this, Flamingo [6] is proposed to align well-trained large-scale vision and language models through some trainable cross-modal adapters. Consequently, the resultant model is capable of addressing Vision-Language (VL) tasks by processing a prompt sequence, which includes several interleaved image and text examples for in-context learning. Since the primary objective of Flamingo is to build a VLM for the few-shot prompt, they only apply a straightforward strategy to configure the in-context A women riding a bike on the bike path.\\n[Good Caption]\\nPeople riding on a street on a street.\\n[Bad Caption]\\nPeople walking down a street and a street.\\n[Bad Caption]\\nThe man is riding a pink bike threw traffic.\\n[Good Caption]\\nA women riding a bike on the bike path.\\n[Good Caption]\\nPeople riding on a street on a street.\\n[Bad Caption]\\nPeople walking down a street and a street.\\n[Bad Caption]\\nThe man is riding a pink bike threw traffic. A kitchen with a stove and a stove.\\n[Bad Caption]\\nA kitchen scene with a fridge and an oven.\\n[Good Caption] People ride on elephants through a field.\\n[Good Caption] Some elephants are walking and walking.\\n[Bad Caption]\\nA kitchen with a stove and a stove.\\n[Bad Caption]\\nA kitchen scene with a fridge and an oven.\\n[Good Caption]\\nA group of people riding on a city street.\\n[CIDEr 72.9]\\nA bicycler riding in the bike lane next to traffic.\\n[CIDEr 126 .8] (e) (f)\\nA group of people riding on a city street.\\n[CIDEr 72.9]\\nA bicycler riding in the bike lane next to traffic.\\n[CIDEr 126 .8] (e) (f)\\nFigure 1: The distinction between LM and VLMs as few-shot learners. LM generally excel with examples akin to the test case (blue blocks in (a)). In contrast, for VLMs, the performance is not strictly correlated with image similarity but heavily relies on the caption quality. For instance, when low-quality captions are used, similar images (d) lead to worse performance than dissimilar ones (f) since VLMs may build a short-cut by reusing in-context captions without seeing the given images.\\nsequence by randomly sampling a few image-text pairs. Nevertheless, a plethora of studies within the NLP field [7] have demonstrated that diverse in-context configurations lead to dramatic effects on few-shot performance, e.g., the selection or ordering of in-context samples [8; 9; 10], while only a limited number of studies systematically explore such effects in the VL case.\\nTo narrow this gap, we explore the effects of various in-context configurations on the performance of few-shot VL tasks. Among various VL tasks, Image Captioning (IC) aims at generating a text conditioned on the source image, and thus can be considered as the visually-conditioned LM. Just as a multitude of NLP tasks can be recast as LM tasks, IC performs a similar function [11; 12; 13], which motivates our decision to select IC as the subject of our case study. However, unlike in NLP, where only single-modal texts are considered in the in-context configuration, in IC, the synergy between multiple modalities significantly influences the performance. For instance, our experiments revealed that selecting images similar to the test image for the in-context sequence does not always lead to good performance, a result closely tied to the quality of the associated captions. Figure 1 shows the comparison between LM and VLM as the few-shot learners.\\nConsequently, we design diverse ways to select images as the in-context images. After selection, the captions of different qualities are assigned to these images for constructing the multi-modal in-context sequence. By combining diverse image selection and caption assignment techniques, we undertake a comprehensive exploration of the effects of multi-modal mutual synergy on VL in-context captioning. We implement all experiments using the prevalent captioning dataset, MSCOCO [14], employing its training set as the database for image selection. Through extensive evaluation of various image selection and caption assignment strategies, we uncover two counter-intuitive yet valuable insights. (1) Caption quality is determined by descriptiveness and language patterns, but their influence on in-context captioning performance is unequal. When captions adequately describe salient image objects, simpler language patterns may yield better results.\\n(2) The efficacy of similar images depends on the quality of the paired captions. Excessive similarity might cause VLM to create a short-cut inference [15] from in-context captions, potentially misleading the model with low-quality captions. Beyond these findings, we introduce a practical in-context captioning strategy, Iterative Prompting, for cases with limited or no Ground-Truth Captions. Furthermore, when Ground-Truth Captions are available, we recommend using Model-Generated Captions as anchors to identify which Ground-Truth Caption is a more suitable in-context caption. Experimental results indicate that even when utilizing low-quality model-generated captions, there is an average CIDEr improvement of 7.3. Moreover, in optimal conditions, the average enhancement reaches up to 20.9 points compared to the random sampling baseline.\\nSince the Flamingo code is not publicly available, our experiments mainly utilize the unofficial Open-Flamingo [16] 2 . It's worth noting that the performance of Open-Flamingo is not on par with the official Flamingo due to its training on less data.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b2\",\n \"b16\",\n \"b0\",\n \"b17\",\n \"b1\",\n \"b33\",\n \"b40\",\n \"b5\",\n \"b15\",\n \"b10\",\n \"b41\",\n \"b35\",\n \"b42\",\n \"b43\",\n \"b2\",\n \"b49\",\n \"b50\",\n \"b49\",\n \"b5\",\n \"b53\"\n ],\n \"table_ref\": [],\n \"text\": \"Prompting Language Model (LM). The research paradigms of NLP have encountered two sea changes in the past few years [3]. The first one is the LM that is pre-trained by predicting the next word conditioned on observed words and can be fine-tuned for solving various downstream tasks, including GPT [17], BERT [1], and BART [18]. The second sea change is the emergence of the prompt paradigm, which was introduced with GPT-3 [2]. Within this paradigm, a pre-trained LM does not require fine-tuning to solve downstream tasks; instead, tasks are reformulated into appropriate prompts with empty slots to be filled. Subsequently, more advanced prompt-based techniques have been proposed, including prompt-tuning [19; 20; 21] and Chain-of-Thought [22; 23; 24; 25].\\nPrompting Vision-Language Model (VLM). In contrast to NLP, the Vision-Language (VL) domain has made significant strides in the first sea change, as evident by the development of various VL-BERT models. These models leverage large volumes of web-collected image-caption pairs to learn VL-generalizable embeddings [26; 27; 28; 29; 30; 31; 32; 33]. However, the prompt paradigm, despite revolutionizing NLP studies, only appears when a certain scale of the model is reached [34]. This scale prerequisite poses further challenges to the development of a VLM with prompt and in-context learning ability.\\nTo mitigate training burdens, instead of updating all the parameters of a VLM [35; 36], some VLMs [35; 36; 37] freeze well-trained Language Models and only train a smaller network, referred to as adapters [38; 39; 40], to align the pre-trained vision and language models. Inspired by these models, both Frozen [41] and Flamingo [6] evolve into multi-modal few-shot learners by training vision and cross-modal adapters, respectively. Given its superior in-context learning ability, we use Flamingo to explore the effects of various in-context configurations [16].\\nThere are also models that address VL tasks through in-context learning. For instance, PICa [11] utilizes captions as mediators to construct an in-context text for solving Visual Question Answering (VQA) tasks, while this may lose the mutual synergy in the representation space of different modalities. The models proposed in [42] and [36] both fine-tune VLMs for specific VQA tasks, but lack the generalized few-shot prompt ability for other VL tasks. UNIFIED-IO [43]demonstrates a unified approach to a myriad of tasks, from classical computer vision to natural language processing, without task-specific fine-tuning. And ProGrad [44] introduces a technique to prevent prompt tuning from forgetting pre-trained vision-language models' general knowledge by only updating aligned prompts, outperforming other methods in various few-shot learning scenarios.\\nExploring In-Context Configurations in NLP. Upon observing that pre-trained LMs are good fewshot learners, researchers also discover that diverse in-context configurations have dramatic effects on performance [3]. This observation sparks numerous studies aimed at determining the optimal in-context configurations, such as the format of the in-context examples [45; 21; 5], the selection of these examples [8; 9; 46; 47], and even the order in which these examples are presented [48; 49; 10]. However, these studies are predominantly conducted within the NLP field and fail to consider the unique characteristics and complexity of multi-modal data. In order to address this gap, we propose a series of strategies to explore the effects of various multi-modal in-context sequence configurations.\\nImage Caption. Image Captioning (IC) [50] aims at correctly verbalizing one image using descriptive languages, which can be solved through both retrieve [51] and generation [50], the former retrieves complete sentences from a corpus, while the latter generates words sequentially. Researchers have recently combined these approaches by first retrieving image-caption pairs and inputting them into generation models [52; 53]. This process resembles in-context captioning, which also retrieves image-caption pairs to help captioning. However, in contrast to them [52; 53], our work introduces novel image and caption selection strategies to study in-context captioning and these methods can also enhance traditional retrieval-generation methods. \\nwhere the probability P (•) is calculated by a pre-trained Vision-Language Model (VLM) (e.g., Flamingo [6] or Otter [54]).\\nVarious studies in NLP [8; 45; 48; 46] field have shown that the performance of in-context learning varies significantly with different in-context configurations. We explore these effects in the case of Image Captioning (IC). Unlike pure NLP tasks, IC is a dual-modal task, which makes it more complex than NLP tasks. Specifically, the mutual synergy of image-caption examples must be considered in in-context learning, as opposed to considering the images or the captions independently. We next respectively introduce the image selection (cf. 3.1) and caption assignment (cf. 3.2) strategies used to configure in-context image-caption pairs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Selecting Images\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Random Selection (RS). Given a set D = {(I 1 , C 1 ), ..., (I M , C M )} with M image-caption pairs, we randomly sample n images as {I 1 , ..., I N } in S.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\",\n \"fig_3\",\n \"fig_3\",\n \"fig_3\"\n ],\n \"heading\": \"Similarity-based Image-Image Retrieval (SIIR).\",\n \"publication_ref\": [\n \"b34\",\n \"b54\",\n \"b55\",\n \"b34\",\n \"b56\"\n ],\n \"table_ref\": [],\n \"text\": \"Certain NLP studies suggest that performance can be enhanced by retrieving examples similar to the test case [8; 46; 47]. Following this approach, we retrieve n images with the highest similarity scores to the test image Î from D. We employ two methods to compute these similarities: 1) SIIR-CLIP (Figure 2 (a)): Using the vision encoder of CLIP [35], we extract image embeddings to determine similarities. 2) SIIR-TAG (Figure 2 (b)): We utilize some scene graph extractors to derive tags for each image. Specifically, we employ Vinvl [55] to extract objects and their attributes, and IETrans [56] to determine the relations present within the image. Following this extraction, we conduct an AND operation to compute the similarity between tags.\\nSimilarity-based Image-Caption Retrieval (SICR-CLIP) (Figure 2 (d)). Taking advantage of the cross-modal retrieval capability of CLIP [35], we use its vision and language encoders to embed images and captions into a shared space for computing similarities. Given Î, we calculate its cross-modal embedding similarities with {C 1 , ..., C M } ∈ D, and select the images whose captions have top-n similarities with Î. Note that we use different kinds of captions as mediators; the methods to generate these captions will be detailed in Section 3.2. These captions are also used as\\nC 1 , ..., C N ∈ S.\\nDiversity-based Image-Image Retrieval (DIIR). For a single test sample, beside using similar incontext examples, some NLP studies find that diversity is also crucial [57]. Consequently, we retrieve a diverse set of images from D to configure S. Specifically, we employ the following strategies: 1) DIIR-TR: We extract discrete tags from the aforementioned VinVL and IETrans. Then we randomly divide the tags into N clusters and apply SIIR-TAG to retrieve the most similar image from each cluster. 2) DIIR-TT (Figure 2 (c)): To conveniently control the number of shots, we incorporate class tags generated by the IETrans model, extending the basis of SIIR-Tags to four categories: object, class, attribute, and relation. We then employ SIIR-TAG to identify the top-k similar images within each category, allowing us to create 4k-shot images in S. Note that both DIIR methods take into account a certain level of similarity during retrieval, rather than selecting entirely distinct images.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Assigning Captions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"After selecting images, we need to assign one caption to each image to construct in-context imagecaption pairs. To explore the effects of multi-modal mutual synergy, we use captions with diverse qualities as introduced in the following.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ground-Truth Captions (GTC).\",\n \"publication_ref\": [\n \"b13\"\n ],\n \"table_ref\": [],\n \"text\": \"In the MSCOCO dataset [14], each image has 5 GTCs and we simply use the first one in S.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Model-Generated Captions (MGC).\",\n \"publication_ref\": [\n \"b57\",\n \"b13\",\n \"b58\",\n \"b15\",\n \"b5\"\n ],\n \"table_ref\": [],\n \"text\": \"Compared with GTC, MGC has lower quality due to two disadvantages. Firstly, MGC uses poorer language, e.g., simple words or rigid sentence patterns.\\nSecondly, MGC has less descriptiveness that it may mis-verbalize or miss certain salient vision patterns of an image. However, we will see these two disadvantages do not equally cause worse performance compared with GTC in in-context captioning and surprisingly, we find that sometimes simple words or rigid sentence patterns even help generate good captions.\\nHere we apply two different models to generate the captions with diverse qualities. MGC-TF@X: a Transformer is trained from scratch by using VinVL features and X denotes the CIDEr score [58] on the test set. To get captions of different qualities, we use the checkpoints from different training epochs and totally generate three kinds of MGCs. 1) MGC-TF@66 contains grammar mistakes while can describe the most salient objects. 2) MGC-TF@88 can use relatively correct grammar to describe the salient objects. 3) MGC-TF@135 is generated by a well-trained Transformer, i.e. , the loss converges. MGC-VLM(N )@X: Another way to get MGCs is to use VLM in a N -shot manner.\\nTo achieve this, for each image I ∈ D, we treat I as the test image and then use Eq. ( 1) to generate a corresponding caption C conditioned on S constructed by only N image-caption pairs. As a result, two kinds of captions are got which are MGC-VLM(0)@63 and MGC-VLM(32)@81.\\nBoth two ways can construct the set D with M image-caption pairs. Then for a novel test image Î, we can select some image-caption pairs from D by some above-mentioned approaches, e.g., using RS or SIIR-CLIP to get images and assigning the MGCs to the image, to configure S for generating a new caption. Compared with MGC-TF, MGC-VLM is more practical as it addresses scenarios where only a handful or even no human-labelled captions are available, which means we do not have enough data to train a Transformer from scratch.\\nIteratively Prompting (IP). For MGC-VLM introduced before, one natural extension is IP. To achieve this, in the first iteration, we generate a caption for each image I ∈ D by MGC-VLM. In the subsequent iteration, these generated captions are paired with the selected images to prompt VLM to generate enhanced captions. This process can be repeated across multiple iterations, thereby iteratively prompting the VLM.\\nModel-Generated Captions as Anchors (MGCA). A MGC can serve not only as an in-context caption but also as an anchor for selecting a suitable caption from the five GTCs. As MGCs typically verbalize salient visual patterns in an image but may miss finer details, using them as anchors can lead to the selection of GTCs that highlight these salient patterns. Furthermore, the selected GTC may supplement interesting details about these patterns, potentially assisting VLM to generate superior captions during in-context learning. In the MGCA implementation, for each selected image, we measure the similarity between the MGC and five GTCs using CIDEr and select the GTC with the highest CIDEr score. We evaluate the proposed strategies on MSCOCO dataset [14], which is the most widely used benchmark in image captioning. We used the Karpathy split [59] in the experiments, which contains 113,287/5000/5000 training/validation/test images and each image is associated with 5 human-annotated captions.\\nImplementation Details We employ the Open-Flamingo model [16] to test our strategies, setting the length penalty to -2.0 and a maximum generation length of 20. We follow Flamingo [6] to use 4, 8, 16, and 32 shots. We respectively use ViT-L/14 and as the vision and language encoders to extract image and sentence embeddings that used in SIIR-CLIP and SICR-CLIP. For MGC-TF, we train the standard Transformer encoder-decoder architecture on the MSCOCO dataset and use the checkpoints underwent 1000, 3000, and 170,000 iterations respectively. These checkpoints generate the captions with CIDEr scores of 66, 88, and 135 on the test set. We implement all experiments on a single RTX 3090 using FP16.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\",\n \"fig_5\",\n \"fig_6\",\n \"fig_7\"\n ],\n \"heading\": \"Results and Analyses\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Given our varied strategies for image selection and caption assignment, displaying results for each configuration in table format, especially at 4, 8, 16, and 32-shot levels, could become overwhelming.\\nFor clarity, we've chosen to present results using line charts and histograms in the main paper, while detailed numerical outcomes are in the supplementary material. The line charts in Figures 3 and4 illustrate trends as the shot number grows. Each subplot within these figures corresponds to a unique strategy for image selection or caption assignment. Furthermore, Figures 5 and6 show histograms of average CIDEr scores for the various shot results. To facilitate comprehension, we first analyze the effects of caption qualities 4.2.1 and then of image selection strategies 4.2.2.\"\n },\n {\n \"figure_ref\": [\n \"fig_5\",\n \"fig_5\",\n \"fig_6\",\n \"fig_6\",\n \"fig_4\",\n \"fig_4\",\n \"fig_5\",\n \"fig_5\",\n \"fig_6\",\n \"fig_6\"\n ],\n \"heading\": \"Effects of Caption Qualities\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Figure 4(a)-(e) reveals that performance typically improves with an increase in shot numbers. However, the rate of improvement varies, or even declines, depending on the quality of the captions used. For instance, in Figure 4(a), using Ground-Truth Captions (GTC), the increase rates of the Ground-Truth Captions (GTC) vs. Model-Generated Captions (MGC). As discussed in Sec 3.2, MGC exhibits two primary shortcomings when compared to GTC: poorer language and less descriptiveness. However, we will see that these two shortcomings do not always make MGC achieve poorer performance when compared to GTC. Specifically, we find that up to a certain level of descriptiveness, simpler sentence patterns are more easily recognized by the VLM, thereby improving caption generation.\\nEvidence for this can be observed in Figure 5 by comparing GTC with MGC-TF@135. Compared to the caption generated by MGC-TF@135, ground-truth caption has better language patterns, e.g., rich vocabulary and complex sentence pattern, and better descriptiveness. Then when the selected images cannot provide enough vision cues, i.e. , when Random Selection (RS) is applied in Figure 5 (a), GTC outperforms MGC-TF@135. However, once the similarity-based retrieval methods like Similaritybased Image-Image Retrieval (SIIR) or Similarity-based Image-Caption Retrieval (SICR-CLIP) is used, the selected similar images offer useful visual patterns that help address the descriptiveness issue. Consequently, the VLM is more likely to recognize the consistent, simple patterns in MGC than the rich, diverse patterns in GTC and then generate better captions. This effect is more pronounced in 4-shot cases, where VLM lacks sufficient in-context captions to discern sentence patterns, making simpler patterns advantageous. As long as the descriptiveness issue is addressed, MGC often outperforms GTC, e.g., MGC-TF@88>GTC in Figure 3(b-d). Especially, when SICR-CLIP is employed to select captions with high cross-modal similarity to the test image, it significantly mitigates the descriptiveness problem. Then as demonstrated in Figure 3(d), even MGC-TF@66 surpasses GTC.\\nMGC-TF vs. MGC-VLM. Before we see that using more low-quality captions like MGC-TF@66 will misguide VLM to generate worse captions. Yet, Figure 4 (f) indicates that using MGC-VLM(0)@63 improves performance with increased shot numbers, contrasting MGC-TF@66 in Figure 4 (c). This raises the question: why do weaker captions (MGC-VLM(0)@63) surpass those ... ... Several motor scooters are jammed into a small market street.\\nA row of parked bicycles sitting in front of a store.\\nRows of motor scooters are parked in front of a store.\\nThis slice of cake looks like half cheesecake and half vanilla.\\nA bite is taken out of a piece of cake. This slice of cake looks like half cheesecake and half vanilla cake.\\nA piece of cake on a plate with a fork.\\nA piece of cake on a plate with a fork.\\nA piece of cake on a plate with a fork and a spoon.\\nA row of motorcycles parked in front of a street.\\nA group of motorcycles parked in front of a street.\\nA group of motorcycles parked in front of a street.\\nA piece of cake on a plate with a fork.\\nA piece of cake on a plate with a fork.\\nA piece of cake on a plate with a fork and a spoon.\\nA row of motorcycles parked in front of a street.\\nA group of motorcycles parked in front of a street.\\nA group of motorcycles parked in front of a street.\\n(a) MGC-TF@135(blue) v.s. GTC(red) ... ...\\nA stop sign that is on a street.\\nA stop sign on a street with a street.\\nA stop sign on the side of a street with a street light.\\nA group of boats in the water and a boat.\\nA boat on a boat in the water.\\nA boat with a boat in the water and a boat in the water .\\nTwo sailing boats are moored in a harbour.\\nA view of a lake with a boat and a dock in the foreground.\\nA view of a lake with a boat and a dock in the foreground.\\nA close up view of a traffic light that's red.\\nA close up picture of a red traffic light.\\nA picture of a traffic light that's red.\\nTwo sailing boats are moored in a harbour.\\nA view of a lake with a boat and a dock in the foreground.\\nA view of a lake with a boat and a dock in the foreground.\\nA close up view of a traffic light that's red.\\nA close up picture of a red traffic light.\\nA picture of a traffic light that's red.\\n(b) MGC-VLM(0)@63(blue) v.s. MGC-TF@66(red) Figure 7: (a) Two examples show that GTC uses more diverse words than MGC-TF@135, making VLM hard to recognize the major pattern, e.g., it mis-generates \\\"a store\\\" in left or neglects \\\"a fork\\\" in right. (b) Two examples demonstrating how VLM is misguided by syntactic errors in MGC-VLM@66, where certain phrases are repeated such as \\\"a street\\\" or \\\"a boat in the water\\\". with higher CIDEr (MGC-TF@66)? This discrepancy aligns with our assumption: descriptiveness and language pattern influence in-context captioning differently. MGC-TF@66 excels in object detection but struggles with language decoding, causing salient objects to be identified correctly but with syntactical errors in captions. As such examples increase, VLM produces worse captions due to these errors. Conversely, MGC-VLM(0)@63, though limited in object recognition, maintains better grammar. When more vision cues are provided, VLM leverages these along with the better-formed captions from MGC-VLM(0)@63, resulting in improved captions. Model-Generated Captions as Anchors (MGCA). From Figure 5, we see that using MGC as the in-context captions usually underperforms GTC. However, by the MGCA strategy, we observe consistent improvements over GTCs, as demonstrated by the higher blue histograms of different MGCs compared to the grey dashed line. For example, despite MGC-TF@66 only achieves 64.24 CIDEr score in Figure 5(a), we still observe a 3.2 CIDEr improvement when using MGC-TF@66 as the anchor, compared to simply selecting the first GTC. Specifically, when using MGC-TF@66/MGC-TF@88/MGC-TF@135/MGC-VLM(0)@63/MGC-VLM(32)@81 as the anchors, the average improvements over six image selection strategies compared to GTC are 7.3/8.0/8.8/3.6/4.8 respectively. In contrast to solely leveraging the RS+GTC method, the combination of SIIR-CLIP and using MGC-TF@135 as anchor chalked up an average boost of 20.9.\\nThe primary reason for such improvement is likely that MGC, to some extent, verbalize the major patterns, such as the salient objects of an image. This helps identify which GTC provides more detailed information about these patterns. Such detailed information of the salient objects provide help VLM generate better captions. This assumption is further supported by comparisons between MGC-TF and MGC-VLM. Given that MGC-TF prioritizes verbalizing the salient objects of an image, using MGC-TF@66/MGC-TF@88 as anchors tends to select superior GTC than MGC-VLM(0)@63/MGC-VLM(32)@81, thus yielding higher improvements. ...\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A long table with a plant on top of it surrounded with wooden chairs.\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"A table is adorned with wooden chairs with blue accents.\\nA table is adorned with wooden chairs with red accents.\\n...\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"(a) Experiment (1)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"A long table with a plant on top of it surrounded with wooden chairs.\\nA table is adorned with wooden chairs with blue accents.\\nBananas on a wooden table.\\n...\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A long table with a plant on top of it surrounded with wooden chairs.\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"A table is adorned with wooden chairs with blue accents.\\nA car with a sign that says stop.\\n...\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"(c) Experiment (3)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"A long table with a plant on top of it surrounded with wooden chairs.\\nA table is adorned with wooden chairs with blue accents.\\nA long table with bananas on it.\\n...\"\n },\n {\n \"figure_ref\": [\n \"fig_6\"\n ],\n \"heading\": \"A long table with a plant on top of it surrounded with wooden chairs.\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"A table is adorned with wooden chairs with blue accents.\\nA stop sign on a road.\\n... the second, indicating that extended VLM iterations might be redundant. Remarkably, even when confined to just 32-shot GTCs, merely two iterations of IP -for instance, where MGC-VLM(32) attains an average CIDEr of 80.5 -can rival performances seen when all GTCs are utilized, as exemplified by RS-GTC's average CIDEr of 80.04 in Figure 5(a).\"\n },\n {\n \"figure_ref\": [\n \"fig_7\",\n \"fig_7\"\n ],\n \"heading\": \"Effects of Image Qualities\",\n \"publication_ref\": [\n \"b56\"\n ],\n \"table_ref\": [\n \"tab_3\"\n ],\n \"text\": \"We evaluated the outcomes of several image selection techniques. SIIR-CLIP, leveraging vision embeddings to compute retrieval similarities, generally identifies images that are more analogous than those found by SIIR-TAG. This is likely attributed to the intrinsic noise present in semantic tags. SICR-CLIP, emphasizing captions that spotlight prominent objects, naturally gravitates towards images showcasing similar objects. In contrast, both DIIR-TR and DIIR-TT produce more varied selections. Nevertheless, when benchmarked against RS, every retrieval-based model consistently fetches images that exhibit greater similarity.\\nAt first glance, one could reasonably infer that using more similar images would invariably lead to superior performance. Yet, as demonstrated in Figure 6, this assumption doesn't universally hold. When engaging high-quality captions, namely GTC in (a) and MGC-TF@135 in (d), the correlation stands, with more analogous images indeed translating to enhanced results. For instance, all retrieval-based techniques surpass RS in this context. However, with medium-quality captions, as in MGC-TF@88 in (b) and VLM(32)@81 in (e), only the similarity-based retrieval methods like SIIR or SICR-CLIP manage to outdo RS, and this is specifically observed in (e). Most intriguingly, when low-quality captions like MGC-TF@66 in (c) and VLM(32)@63 in (f) are in play, analogous images inversely impact performance, as illustrated by RS outperforming SIIR-CLIP in (f). This critical insight underscores a pivotal revelation: the efficacy of utilizing similar images is intricately tied to the caliber of the corresponding captions.\\nSimilar Images Lead to Short-Cut Inference. A relatively bold hypothesis to elucidate this phenomenon is: when in-context images are similar to the test image, VLM may take a short-cut by leveraging in-context captions to generate a new one, rather than learning how to caption from the in-context pairs. Consequently, the greater the similarity between in-context and test images, the more the VLM is influenced by the in-context captions. To robustly test our underlying assumption, we designed three distinct experimental setups. In each setup, the incontext captions remain consistent, comprising 5 groundtruth captions sourced randomly from an image unrelated to the test image. However, the in-context images are chosen differently in each experiment: (1) they are identical to the test image; (2) they are picked using SIIR-CLIP; (3) they are chosen via RS. This progression results in decreasing similarity between the in-context and test images from experiments (1) through (3). Table 2 showcases two sets of CIDEr scores. One set compares the generated captions to the five ground-truth captions (GTC) of the test image, and the other set contrasts them with the five in-context captions (ICC). Our findings suggest a clear pattern: the closer the in-context images are to the test image, the more the VLM tends to mirror the ICC in its generated caption. For illustration, method (1) registers the highest CIDEr score when compared to the ICC. However, the caption it produces doesn't accurately depict the image, as reflected by its notably lower CIDEr score in relation to the GTC. These observations solidify our hypothesis that images with high similarity can inadvertently prompt short-cut inference. Additionally, Figure 8 provides visual representations to further elucidate the phenomenon of short-cut inference.\\nDIIR. As depicted in Figure 6, the performance of the two DIIR methods noticeably lags behind SIIR-TAG. This observation underscores the contention that the strength of diversity may not always translate to superior in every context. One plausible explanation we propose is the inherent nature of captioning as a task. Contrary to certain complex NLP challenges, where diversity can be instrumental in offering a multifaceted understanding of a problem [57], captioning is relatively straightforward and may not benefit as much from diverse in-context examples. Delving deeper into the DIIR methods, DIIR-TT consistently outshines DIIR-TR. This leads us to infer that the clustering based on semantic tag types might be a more optimal strategy. Such a strategy, we believe, not only ensures diversity but also completeness in the selection of images. For instance, given a caption like \\\"a brown dog is running\\\", DIIR-TT could potentially source images that highlight elements like \\\"brown objects\\\", \\\"dogs\\\", and the \\\"action of running\\\".\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion and Limitations\",\n \"publication_ref\": [\n \"b15\",\n \"b5\",\n \"b53\"\n ],\n \"table_ref\": [],\n \"text\": \"In this study, we utilize image captioning as a case study to examine the impacts of varied multimodal in-context configurations on few-shot vision-language prompting. Specifically, we design 4 different ways to select images and 4 different strategies to assign captions to the selected images for constructing multi-modal in-context configurations. Exhaustive experiments reveal that, contrary to single-modal NLP cases, multi-modal mutual synergy significantly influences performance. Notably, we observe that the descriptiveness and language patterns of the captions differently affect performance: better performance may be achieved with simpler and consistent sentence patterns when selected images compensate for descriptiveness issues. And we also discover that when the in-context images are similar to the test one, VLM may build a short-cut by directly using the in-context captions instead of really learning to captioning. Moreover, our optimal strategy lead to a 20.9 average increase in CIDEr scores compared to a random sampling baseline. This study's primary limitation is that at the inception of our exploration, the only open-source multi-modal few-shot learner available is Open-Flamingo [16]. However, Open-Flamingo, when compared with the official Flamingo [6], underperforms due to its training on significantly less data. Consequently, some findings in this paper might shift if a more robust multi-modal few-shot learner is employed. Nevertheless, even in such a scenario, the diverse configuration strategies proposed in this paper maintain their utility, aiding researchers in swiftly exploring the characteristics of the employed VLMs. Additionally, we have provided experimental results on Otter [54] and smaller version of Open-Flamingo, with detailed findings available in Appendix.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Supplementary Material\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_5\",\n \"tab_6\"\n ],\n \"text\": \"A Experimental results on Open-Flamingo v1 9B\\nHere we present all the numerical results from the experiment, divided into four tables (Table 3, Table 4, Table 5, and \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B Experimental results on Open-Flamingo v2 3B\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_7\"\n ],\n \"text\": \"Here we present the numerical results on Open-Flamingo v2 3B model3 in Table 7 based on two different Image Selection strategies. Additionally, we have calculated the average values for each method in the \\\"mean\\\" column.\\nFrom the values in the table, it can be seen that the trend is basically consistent with the v1 model. It can be considered that our strategy and analysis can be transferred to different models.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C Experimental results on Otter\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Here we present the numerical results on Otter model in \"\n },\n {\n \"figure_ref\": [\n \"fig_14\",\n \"fig_17\"\n ],\n \"heading\": \"D More Results of MGC-TF@135 vs. GTC\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We further elucidate the performance of MGC-TF@135(blue) and GTC(red), by offering additional examples shown in Figure 9. It can be easily observed that GTC has a more diverse range of sentence structures in comparison to MGC-TF@135. In the initial two examples, the words \\\"cat\\\" and \\\"dachshunds\\\" were inaccurately recognized by GTC, demonstrating its limitation in some specific instances. A shift can be noticed in the subsequent three examples, where the GTC generates captions with complex sentence structures, with a notable proportion of incorrect information. This finding serves to reinforce the notion that simpler sentence patterns, up to a certain degree of descriptiveness, are more readily deciphered by the VLM, thereby enhancing the quality of generated captions.\\nE More Results of MGC-TF@66 vs. MGC-VLM(0)@63 Similarly, we supplement with more examples, as depicted in Figure 10, to facilitate the comparison between MGC-TF@66(red) and MGC-VLM(0)@63(blue). A striking observation from this comparative study reveals that MGC-TF@66 demonstrates a significant number of syntactical errors. This flaw becomes problematic as it misdirects VLM, resulting in a substantial volume of syntactical mistakes in the produced sentences.This correlation implies that the syntax errors in the initial input by MGC-TF@66 tend to propagate into the VLM's output. Therefore, it becomes clear that the accuracy of grammar in the prompt is crucial for achieving better results. Five examples demonstrate that more diverse words will be used in GTC than in MGC-TF@135 which making VLM hard to catch the major pattern. e.g., the first line and the second line incorrectly state \\\"cat\\\" and \\\"dachshunds.\\\" Five examples show that VLM will be misguided by the syntactic errors in MGC-VLM@66. Lots of outputs have repeated phrases, such as \\\"a table\\\", \\\"a clock\\\", \\\"a desk\\\", \\\"a motorcycle\\\" and \\\"a vase\\\".\"\n },\n {\n \"figure_ref\": [\n \"fig_18\"\n ],\n \"heading\": \"F More Results of short-cut inference\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we provide additional examples to demonstrate how similar images can lead our VLM to exhibit a tendency for short-cut inference, where it might relies heavily on in-context captions to generate the captions, disregarding the image information in the prompt and the inherent relationships within the in-context pairs. In Figure 11, we present the results for the same test image under two different scenarios of choosing in-context images: \\\"identical to the test image\\\" (top row) and \\\"via random sampling\\\" (bottom row).\\nIn the first scenario, for the majority of cases, our results are largely unrelated to the image content but similar to the in-context captions, with only a few instances capturing some elements from the image, such as \\\"pink\\\" in (a) and \\\"on the beach\\\" in (b). However, in the second scenario, our results are heavily influenced by the in-context captions, as evident in (c) and (d) with the mention of \\\"wooden chairs\\\". \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"(b)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"A long restaurant table with rattan rounded back chairs.\\nA long table with a plant on top of it surrounded with wooden chairs.\\nA restaurant has modern wooden tables and chairs.\\nA long table with a flower arrangement in the middle for meetings.\\nA table is adorned with wooden chairs with blue accents.\\nA man is running on the beach.\\nA long restaurant table with rattan rounded back chairs.\\nA long table with a plant on top of it surrounded with wooden chairs.\\nA restaurant has modern wooden tables and chairs. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This work is supported by National Science Foundation of China (62206048), Natural Science Foundation of Jiangsu Province (BK20220819), Young Elite Scientists Sponsorship Program of Jiangsu Association for Science and Technology Tj-2022-027 and the Big Data Computing Center of Southeast University.\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2024-01-23"},"doi":{"kind":"string","value":""},"references":{"kind":"list like","value":[{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton ; Lee Kristina; Toutanova ","journal":"","ref_id":"b0","title":"Bert: Pre-training of deep bidirectional transformers for language understanding","year":"2019"},{"authors":"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell","journal":"Advances in neural information processing systems","ref_id":"b1","title":"Language models are few-shot learners","year":"2020"},{"authors":"Pengfei Liu; Weizhe Yuan; Jinlan Fu; Zhengbao Jiang; Hiroaki Hayashi; Graham Neubig","journal":"ACM Computing Surveys","ref_id":"b2","title":"Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing","year":"2023"},{"authors":"Grégoire Mialon; Roberto Dessì; Maria Lomeli; Christoforos Nalmpantis; Ram Pasunuru; Roberta Raileanu; Timo Baptiste Rozière; Jane Schick; Asli Dwivedi-Yu; Celikyilmaz","journal":"","ref_id":"b3","title":"Augmented language models: a survey","year":"2023"},{"authors":"Sewon Min; Xinxi Lyu; Ari Holtzman; Mikel Artetxe; Mike Lewis; Hannaneh Hajishirzi; Luke Zettlemoyer","journal":"","ref_id":"b4","title":"Rethinking the role of demonstrations: What makes in-context learning work?","year":"2022"},{"authors":"Jean-Baptiste Alayrac; Jeff Donahue; Pauline Luc; Antoine Miech; Iain Barr; Yana Hasson; Karel Lenc; Arthur Mensch; Katherine Millican; Malcolm Reynolds","journal":"Advances in Neural Information Processing Systems","ref_id":"b5","title":"Flamingo: a visual language model for few-shot learning","year":"2022"},{"authors":"Eric Zelikman; Jesse Mu; Yuhuai Tony Noah D Goodman; Wu","journal":"","ref_id":"b6","title":"Star: Self-taught reasoner bootstrapping reasoning with reasoning","year":"2022"},{"authors":"Jiachang Liu; Dinghan Shen; Yizhe Zhang; Bill Dolan; Lawrence Carin; Weizhu Chen","journal":"","ref_id":"b7","title":"What makes good in-context examples for gpt-3? DeeLIO","year":"2022"},{"authors":"Tianyu Gao; Adam Fisch; Danqi Chen","journal":"","ref_id":"b8","title":"Making pre-trained language models better few-shot learners","year":"2021"},{"authors":"Yao Lu; Max Bartolo; Alastair Moore; Sebastian Riedel; Pontus Stenetorp","journal":"","ref_id":"b9","title":"Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity","year":"2022"},{"authors":"Zhengyuan Yang; Zhe Gan; Jianfeng Wang; Xiaowei Hu; Yumao Lu; Zicheng Liu; Lijuan Wang","journal":"","ref_id":"b10","title":"An empirical study of gpt-3 for few-shot knowledge-based vqa","year":"2022"},{"authors":"Jaemin Cho; Jie Lei; Hao Tan; Mohit Bansal","journal":"PMLR","ref_id":"b11","title":"Unifying vision-and-language tasks via text generation","year":"2021"},{"authors":"Jialin Wu; Jiasen Lu; Ashish Sabharwal; Roozbeh Mottaghi","journal":"","ref_id":"b12","title":"Multi-modal answer validation for knowledge-based vqa","year":"2022"},{"authors":"Tsung-Yi Lin; Michael Maire; Serge Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence; Zitnick ","journal":"Springer","ref_id":"b13","title":"Microsoft coco: Common objects in context","year":"2014"},{"authors":"Robert Geirhos; Jörn-Henrik Jacobsen; Claudio Michaelis; Richard Zemel; Wieland Brendel; Matthias Bethge; Felix A Wichmann","journal":"Nature Machine Intelligence","ref_id":"b14","title":"Shortcut learning in deep neural networks","year":"2020"},{"authors":"Anas Awadalla; Irena Gao; Josh Gardner; Jack Hessel; Yusuf Hanafy; Wanrong Zhu; Yonatan Kalyani Marathe; Samir Bitton; Shiori Gadre; Jenia Sagawa; Simon Jitsev; Pang Kornblith; Gabriel Wei Koh; Mitchell Ilharco; Ludwig Wortsman; Schmidt","journal":"","ref_id":"b15","title":"Openflamingo: An opensource framework for training large autoregressive vision-language models","year":"2023"},{"authors":"Alec Radford; Karthik Narasimhan","journal":"","ref_id":"b16","title":"Improving language understanding by generative pre-training","year":"2018"},{"authors":"Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Ves Stoyanov; Luke Zettlemoyer","journal":"","ref_id":"b17","title":"Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension","year":"2019"},{"authors":"Karen Hambardzumyan; Hrant Khachatrian; Jonathan May","journal":"","ref_id":"b18","title":"Warp: Word-level adversarial reprogramming","year":"2021"},{"authors":"Lisa Xiang; Percy Li; Liang","journal":"","ref_id":"b19","title":"Prefix-tuning: Optimizing continuous prompts for generation","year":"2021"},{"authors":"Brian Lester; Rami Al-Rfou; Noah Constant","journal":"","ref_id":"b20","title":"The power of scale for parameter-efficient prompt tuning","year":"2021"},{"authors":"Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Fei Xia; Ed H Chi; V Quoc; Denny Le; Zhou","journal":"","ref_id":"b21","title":"Chain-of-thought prompting elicits reasoning in large language models","year":""},{"authors":"Xuezhi Wang; Jason Wei; Dale Schuurmans; Quoc Le; Ed Chi; Denny Zhou","journal":"","ref_id":"b22","title":"Selfconsistency improves chain of thought reasoning in language models","year":"2022"},{"authors":"Hangfeng He; Hongming Zhang; Dan Roth","journal":"","ref_id":"b23","title":"Rethinking with retrieval: Faithful large language model inference","year":"2022"},{"authors":"Harsh Trivedi; Niranjan Balasubramanian; Tushar Khot; Ashish Sabharwal","journal":"","ref_id":"b24","title":"Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions","year":"2022"},{"authors":"Jiasen Lu; Dhruv Batra; Devi Parikh; Stefan Lee","journal":"Advances in neural information processing systems","ref_id":"b25","title":"Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks","year":"2019"},{"authors":"Hao Tan; Mohit Bansal","journal":"","ref_id":"b26","title":"Lxmert: Learning cross-modality encoder representations from transformers","year":"2019"},{"authors":"Yen-Chun Chen; Linjie Li; Licheng Yu; Ahmed El Kholy; Faisal Ahmed; Zhe Gan; Yu Cheng; Jingjing Liu","journal":"Springer","ref_id":"b27","title":"Uniter: Universal image-text representation learning","year":"2020"},{"authors":"Xiujun Li; Xi Yin; Chunyuan Li; Pengchuan Zhang; Xiaowei Hu; Lei Zhang; Lijuan Wang; Houdong Hu; Li Dong; Furu Wei","journal":"Springer","ref_id":"b28","title":"Oscar: Object-semantics aligned pre-training for vision-language tasks","year":"2020"},{"authors":"Jiasen Lu; Vedanuj Goswami; Marcus Rohrbach; Devi Parikh; Stefan Lee","journal":"","ref_id":"b29","title":"12-in-1: Multitask vision and language representation learning","year":"2020"},{"authors":"Ziwang Fu; Feng Liu; Hanyang Wang; Siyuan Shen; Jiahao Zhang; Jiayin Qi; Xiangling Fu; Aimin Zhou","journal":"","ref_id":"b30","title":"Lmr-cbt: Learning modality-fused representations with cb-transformer for multimodal emotion recognition from unaligned multimodal sequences","year":"2021"},{"authors":"Zhong Ji; Jingwei Ni; Xiyao Liu; Yanwei Pang","journal":"Frontiers of Computer Science","ref_id":"b31","title":"Teachers cooperation: team-knowledge distillation for multiple cross-domain few-shot learning","year":"2023"},{"authors":"Shukang Yin; Chaoyou Fu; Sirui Zhao; Ke Li; Xing Sun; Tong Xu; Enhong Chen","journal":"","ref_id":"b32","title":"A survey on multimodal large language models","year":"2023"},{"authors":"Jason Wei; Yi Tay; Rishi Bommasani; Colin Raffel; Barret Zoph; Sebastian Borgeaud; Dani Yogatama; Maarten Bosma; Denny Zhou; Donald Metzler; Ed H Chi; Tatsunori Hashimoto; Oriol Vinyals; Percy Liang; Jeff Dean; William Fedus","journal":"Transactions on Machine Learning Research","ref_id":"b33","title":"Emergent abilities of large language models","year":"2022"},{"authors":"Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark","journal":"PMLR","ref_id":"b34","title":"Learning transferable visual models from natural language supervision","year":"2021"},{"authors":"Zhuosheng Zhang; Aston Zhang; Mu Li; Hai Zhao; George Karypis; Alex Smola","journal":"","ref_id":"b35","title":"Multimodal chain-of-thought reasoning in language models","year":"2023"},{"authors":"Jing Yu Koh; Ruslan Salakhutdinov; Daniel Fried","journal":"","ref_id":"b36","title":"Grounding language models to images for multimodal inputs and outputs","year":"2023"},{"authors":"Junnan Li; Dongxu Li; Caiming Xiong; Steven Hoi","journal":"PMLR","ref_id":"b37","title":"Blip: Bootstrapping languageimage pre-training for unified vision-language understanding and generation","year":"2022"},{"authors":"Junnan Li; Dongxu Li; Silvio Savarese; Steven Hoi","journal":"","ref_id":"b38","title":"Blip-2: Bootstrapping languageimage pre-training with frozen image encoders and large language models","year":"2023"},{"authors":"Jiahui Yu; Zirui Wang; Vijay Vasudevan; Legg Yeung; Mojtaba Seyedhosseini; Yonghui Wu","journal":"","ref_id":"b39","title":"Coca: Contrastive captioners are image-text foundation models","year":"2022"},{"authors":"Maria Tsimpoukelli; Jacob L Menick; Serkan Cabi; Oriol Eslami; Felix Vinyals; Hill","journal":"Advances in Neural Information Processing Systems","ref_id":"b40","title":"Multimodal few-shot learning with frozen language models","year":"2021"},{"authors":"Pan Lu; Swaroop Mishra; Tanglin Xia; Liang Qiu; Kai-Wei Chang; Song-Chun Zhu; Oyvind Tafjord; Peter Clark; Ashwin Kalyan","journal":"Advances in Neural Information Processing Systems","ref_id":"b41","title":"Learn to explain: Multimodal reasoning via thought chains for science question answering","year":"2022"},{"authors":"Jiasen Lu; Christopher Clark; Rowan Zellers; Roozbeh Mottaghi; Aniruddha Kembhavi","journal":"","ref_id":"b42","title":"Unified-io: A unified model for vision, language, and multi-modal tasks","year":"2022"},{"authors":"Beier Zhu; Yulei Niu; Yucheng Han; Yue Wu; Hanwang Zhang","journal":"","ref_id":"b43","title":"Prompt-aligned gradient for prompt tuning","year":"2023"},{"authors":"Zhengbao Jiang; Frank F Xu; Jun Araki; Graham Neubig","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b44","title":"How can we know what language models know?","year":"2020"},{"authors":"Ohad Rubin; Jonathan Herzig; Jonathan Berant","journal":"","ref_id":"b45","title":"Learning to retrieve prompts for incontext learning","year":"2022"},{"authors":"Jungo Su Hongjin; Chen Henry Kasai; Weijia Wu; Tianlu Shi; Jiayi Wang; Rui Xin; Mari Zhang; Luke Ostendorf; Noah A Zettlemoyer; Smith","journal":"","ref_id":"b46","title":"Selective annotation makes language models better few-shot learners","year":""},{"authors":"Sawan Kumar; Partha Talukdar","journal":"","ref_id":"b47","title":"Reordering examples helps during priming-based few-shot learning","year":"2021"},{"authors":"Denny Zhou; Nathanael Schärli; Le Hou; Jason Wei; Nathan Scales; Xuezhi Wang; Dale Schuurmans; Olivier Bousquet; Quoc Le; Ed Chi","journal":"","ref_id":"b48","title":"Least-to-most prompting enables complex reasoning in large language models","year":"2022"},{"authors":"Oriol Vinyals; Alexander Toshev; Samy Bengio; Dumitru Erhan","journal":"","ref_id":"b49","title":"Show and tell: A neural image caption generator","year":"2015"},{"authors":"Sanghyuk Chun; Joon Seong; Rafael Oh; Yannis Sampaio De Rezende; Diane Kalantidis; Larlus","journal":"","ref_id":"b50","title":"Probabilistic embeddings for cross-modal retrieval","year":"2021"},{"authors":"Rita Ramos; Bruno Martins; Desmond Elliott; Yova Kementchedjhieva","journal":"","ref_id":"b51","title":"Smallcap: lightweight image captioning prompted with retrieval augmentation","year":"2023"},{"authors":"Sara Sarto; Marcella Cornia; Lorenzo Baraldi; Rita Cucchiara","journal":"","ref_id":"b52","title":"Retrieval-augmented transformer for image captioning","year":"2022"},{"authors":"Bo Li; Yuanhan Zhang; Liangyu Chen; Jinghao Wang; Jingkang Yang; Ziwei Liu","journal":"","ref_id":"b53","title":"Otter: A multi-modal model with in-context instruction tuning","year":"2023"},{"authors":"Pengchuan Zhang; Xiujun Li; Xiaowei Hu; Jianwei Yang; Lei Zhang; Lijuan Wang; Yejin Choi; Jianfeng Gao","journal":"","ref_id":"b54","title":"Vinvl: Revisiting visual representations in vision-language models","year":"2021"},{"authors":"Ao Zhang; Yuan Yao; Qianyu Chen; Wei Ji; Zhiyuan Liu; Maosong Sun; Tat-Seng Chua","journal":"","ref_id":"b55","title":"Fine-grained scene graph generation with data transfer","year":"2022"},{"authors":"Zhuosheng Zhang; Aston Zhang; Mu Li; Alex Smola","journal":"","ref_id":"b56","title":"Automatic chain of thought prompting in large language models","year":"2022"},{"authors":"Ramakrishna Vedantam; Lawrence Zitnick; Devi Parikh","journal":"","ref_id":"b57","title":"Cider: Consensus-based image description evaluation","year":"2015"},{"authors":"Andrej Karpathy; Li Fei-Fei","journal":"CLIP MGCA-VLM","ref_id":"b58","title":"Deep visual-semantic alignments for generating image descriptions","year":"0287"},{"authors":"","journal":"TR MGC-VLM","ref_id":"b59","title":"15 Table 5: Using Similarity-based Image-Caption Retrieval (SICR) image selection strategy results","year":""},{"authors":"","journal":"","ref_id":"b60","title":"37 Table 6: Using Diversity-based Image-Image Retrieval (DIIR) image selection strategy results","year":""}],"string":"[\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton ; Lee Kristina; Toutanova \",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Bert: Pre-training of deep bidirectional transformers for language understanding\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b1\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Pengfei Liu; Weizhe Yuan; Jinlan Fu; Zhengbao Jiang; Hiroaki Hayashi; Graham Neubig\",\n \"journal\": \"ACM Computing Surveys\",\n \"ref_id\": \"b2\",\n \"title\": \"Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Grégoire Mialon; Roberto Dessì; Maria Lomeli; Christoforos Nalmpantis; Ram Pasunuru; Roberta Raileanu; Timo Baptiste Rozière; Jane Schick; Asli Dwivedi-Yu; Celikyilmaz\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Augmented language models: a survey\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Sewon Min; Xinxi Lyu; Ari Holtzman; Mikel Artetxe; Mike Lewis; Hannaneh Hajishirzi; Luke Zettlemoyer\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Rethinking the role of demonstrations: What makes in-context learning work?\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jean-Baptiste Alayrac; Jeff Donahue; Pauline Luc; Antoine Miech; Iain Barr; Yana Hasson; Karel Lenc; Arthur Mensch; Katherine Millican; Malcolm Reynolds\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b5\",\n \"title\": \"Flamingo: a visual language model for few-shot learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Eric Zelikman; Jesse Mu; Yuhuai Tony Noah D Goodman; Wu\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Star: Self-taught reasoner bootstrapping reasoning with reasoning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jiachang Liu; Dinghan Shen; Yizhe Zhang; Bill Dolan; Lawrence Carin; Weizhu Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"What makes good in-context examples for gpt-3? DeeLIO\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tianyu Gao; Adam Fisch; Danqi Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Making pre-trained language models better few-shot learners\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yao Lu; Max Bartolo; Alastair Moore; Sebastian Riedel; Pontus Stenetorp\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Zhengyuan Yang; Zhe Gan; Jianfeng Wang; Xiaowei Hu; Yumao Lu; Zicheng Liu; Lijuan Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"An empirical study of gpt-3 for few-shot knowledge-based vqa\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jaemin Cho; Jie Lei; Hao Tan; Mohit Bansal\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b11\",\n \"title\": \"Unifying vision-and-language tasks via text generation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jialin Wu; Jiasen Lu; Ashish Sabharwal; Roozbeh Mottaghi\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Multi-modal answer validation for knowledge-based vqa\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tsung-Yi Lin; Michael Maire; Serge Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence; Zitnick \",\n \"journal\": \"Springer\",\n \"ref_id\": \"b13\",\n \"title\": \"Microsoft coco: Common objects in context\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Robert Geirhos; Jörn-Henrik Jacobsen; Claudio Michaelis; Richard Zemel; Wieland Brendel; Matthias Bethge; Felix A Wichmann\",\n \"journal\": \"Nature Machine Intelligence\",\n \"ref_id\": \"b14\",\n \"title\": \"Shortcut learning in deep neural networks\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Anas Awadalla; Irena Gao; Josh Gardner; Jack Hessel; Yusuf Hanafy; Wanrong Zhu; Yonatan Kalyani Marathe; Samir Bitton; Shiori Gadre; Jenia Sagawa; Simon Jitsev; Pang Kornblith; Gabriel Wei Koh; Mitchell Ilharco; Ludwig Wortsman; Schmidt\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Openflamingo: An opensource framework for training large autoregressive vision-language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Alec Radford; Karthik Narasimhan\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Improving language understanding by generative pre-training\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Ves Stoyanov; Luke Zettlemoyer\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Karen Hambardzumyan; Hrant Khachatrian; Jonathan May\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Warp: Word-level adversarial reprogramming\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Lisa Xiang; Percy Li; Liang\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Prefix-tuning: Optimizing continuous prompts for generation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Brian Lester; Rami Al-Rfou; Noah Constant\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"The power of scale for parameter-efficient prompt tuning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Fei Xia; Ed H Chi; V Quoc; Denny Le; Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Chain-of-thought prompting elicits reasoning in large language models\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Xuezhi Wang; Jason Wei; Dale Schuurmans; Quoc Le; Ed Chi; Denny Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Selfconsistency improves chain of thought reasoning in language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hangfeng He; Hongming Zhang; Dan Roth\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Rethinking with retrieval: Faithful large language model inference\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Harsh Trivedi; Niranjan Balasubramanian; Tushar Khot; Ashish Sabharwal\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jiasen Lu; Dhruv Batra; Devi Parikh; Stefan Lee\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b25\",\n \"title\": \"Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Hao Tan; Mohit Bansal\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Lxmert: Learning cross-modality encoder representations from transformers\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yen-Chun Chen; Linjie Li; Licheng Yu; Ahmed El Kholy; Faisal Ahmed; Zhe Gan; Yu Cheng; Jingjing Liu\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b27\",\n \"title\": \"Uniter: Universal image-text representation learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Xiujun Li; Xi Yin; Chunyuan Li; Pengchuan Zhang; Xiaowei Hu; Lei Zhang; Lijuan Wang; Houdong Hu; Li Dong; Furu Wei\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b28\",\n \"title\": \"Oscar: Object-semantics aligned pre-training for vision-language tasks\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jiasen Lu; Vedanuj Goswami; Marcus Rohrbach; Devi Parikh; Stefan Lee\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"12-in-1: Multitask vision and language representation learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ziwang Fu; Feng Liu; Hanyang Wang; Siyuan Shen; Jiahao Zhang; Jiayin Qi; Xiangling Fu; Aimin Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Lmr-cbt: Learning modality-fused representations with cb-transformer for multimodal emotion recognition from unaligned multimodal sequences\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zhong Ji; Jingwei Ni; Xiyao Liu; Yanwei Pang\",\n \"journal\": \"Frontiers of Computer Science\",\n \"ref_id\": \"b31\",\n \"title\": \"Teachers cooperation: team-knowledge distillation for multiple cross-domain few-shot learning\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Shukang Yin; Chaoyou Fu; Sirui Zhao; Ke Li; Xing Sun; Tong Xu; Enhong Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"A survey on multimodal large language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Jason Wei; Yi Tay; Rishi Bommasani; Colin Raffel; Barret Zoph; Sebastian Borgeaud; Dani Yogatama; Maarten Bosma; Denny Zhou; Donald Metzler; Ed H Chi; Tatsunori Hashimoto; Oriol Vinyals; Percy Liang; Jeff Dean; William Fedus\",\n \"journal\": \"Transactions on Machine Learning Research\",\n \"ref_id\": \"b33\",\n \"title\": \"Emergent abilities of large language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b34\",\n \"title\": \"Learning transferable visual models from natural language supervision\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zhuosheng Zhang; Aston Zhang; Mu Li; Hai Zhao; George Karypis; Alex Smola\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Multimodal chain-of-thought reasoning in language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Jing Yu Koh; Ruslan Salakhutdinov; Daniel Fried\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Grounding language models to images for multimodal inputs and outputs\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Junnan Li; Dongxu Li; Caiming Xiong; Steven Hoi\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b37\",\n \"title\": \"Blip: Bootstrapping languageimage pre-training for unified vision-language understanding and generation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Junnan Li; Dongxu Li; Silvio Savarese; Steven Hoi\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Blip-2: Bootstrapping languageimage pre-training with frozen image encoders and large language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Jiahui Yu; Zirui Wang; Vijay Vasudevan; Legg Yeung; Mojtaba Seyedhosseini; Yonghui Wu\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Coca: Contrastive captioners are image-text foundation models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Maria Tsimpoukelli; Jacob L Menick; Serkan Cabi; Oriol Eslami; Felix Vinyals; Hill\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b40\",\n \"title\": \"Multimodal few-shot learning with frozen language models\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Pan Lu; Swaroop Mishra; Tanglin Xia; Liang Qiu; Kai-Wei Chang; Song-Chun Zhu; Oyvind Tafjord; Peter Clark; Ashwin Kalyan\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b41\",\n \"title\": \"Learn to explain: Multimodal reasoning via thought chains for science question answering\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jiasen Lu; Christopher Clark; Rowan Zellers; Roozbeh Mottaghi; Aniruddha Kembhavi\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Unified-io: A unified model for vision, language, and multi-modal tasks\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Beier Zhu; Yulei Niu; Yucheng Han; Yue Wu; Hanwang Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"Prompt-aligned gradient for prompt tuning\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Zhengbao Jiang; Frank F Xu; Jun Araki; Graham Neubig\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b44\",\n \"title\": \"How can we know what language models know?\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ohad Rubin; Jonathan Herzig; Jonathan Berant\",\n \"journal\": \"\",\n \"ref_id\": \"b45\",\n \"title\": \"Learning to retrieve prompts for incontext learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jungo Su Hongjin; Chen Henry Kasai; Weijia Wu; Tianlu Shi; Jiayi Wang; Rui Xin; Mari Zhang; Luke Ostendorf; Noah A Zettlemoyer; Smith\",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"Selective annotation makes language models better few-shot learners\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Sawan Kumar; Partha Talukdar\",\n \"journal\": \"\",\n \"ref_id\": \"b47\",\n \"title\": \"Reordering examples helps during priming-based few-shot learning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Denny Zhou; Nathanael Schärli; Le Hou; Jason Wei; Nathan Scales; Xuezhi Wang; Dale Schuurmans; Olivier Bousquet; Quoc Le; Ed Chi\",\n \"journal\": \"\",\n \"ref_id\": \"b48\",\n \"title\": \"Least-to-most prompting enables complex reasoning in large language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Oriol Vinyals; Alexander Toshev; Samy Bengio; Dumitru Erhan\",\n \"journal\": \"\",\n \"ref_id\": \"b49\",\n \"title\": \"Show and tell: A neural image caption generator\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Sanghyuk Chun; Joon Seong; Rafael Oh; Yannis Sampaio De Rezende; Diane Kalantidis; Larlus\",\n \"journal\": \"\",\n \"ref_id\": \"b50\",\n \"title\": \"Probabilistic embeddings for cross-modal retrieval\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Rita Ramos; Bruno Martins; Desmond Elliott; Yova Kementchedjhieva\",\n \"journal\": \"\",\n \"ref_id\": \"b51\",\n \"title\": \"Smallcap: lightweight image captioning prompted with retrieval augmentation\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Sara Sarto; Marcella Cornia; Lorenzo Baraldi; Rita Cucchiara\",\n \"journal\": \"\",\n \"ref_id\": \"b52\",\n \"title\": \"Retrieval-augmented transformer for image captioning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Bo Li; Yuanhan Zhang; Liangyu Chen; Jinghao Wang; Jingkang Yang; Ziwei Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b53\",\n \"title\": \"Otter: A multi-modal model with in-context instruction tuning\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Pengchuan Zhang; Xiujun Li; Xiaowei Hu; Jianwei Yang; Lei Zhang; Lijuan Wang; Yejin Choi; Jianfeng Gao\",\n \"journal\": \"\",\n \"ref_id\": \"b54\",\n \"title\": \"Vinvl: Revisiting visual representations in vision-language models\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ao Zhang; Yuan Yao; Qianyu Chen; Wei Ji; Zhiyuan Liu; Maosong Sun; Tat-Seng Chua\",\n \"journal\": \"\",\n \"ref_id\": \"b55\",\n \"title\": \"Fine-grained scene graph generation with data transfer\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Zhuosheng Zhang; Aston Zhang; Mu Li; Alex Smola\",\n \"journal\": \"\",\n \"ref_id\": \"b56\",\n \"title\": \"Automatic chain of thought prompting in large language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ramakrishna Vedantam; Lawrence Zitnick; Devi Parikh\",\n \"journal\": \"\",\n \"ref_id\": \"b57\",\n \"title\": \"Cider: Consensus-based image description evaluation\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Andrej Karpathy; Li Fei-Fei\",\n \"journal\": \"CLIP MGCA-VLM\",\n \"ref_id\": \"b58\",\n \"title\": \"Deep visual-semantic alignments for generating image descriptions\",\n \"year\": \"0287\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"TR MGC-VLM\",\n \"ref_id\": \"b59\",\n \"title\": \"15 Table 5: Using Similarity-based Image-Caption Retrieval (SICR) image selection strategy results\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b60\",\n \"title\": \"37 Table 6: Using Diversity-based Image-Image Retrieval (DIIR) image selection strategy results\",\n \"year\": \"\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[4,108,713.17,66.91,9.68],"formula_id":"formula_1","formula_text":"C 1 , ..., C N ∈ S."}],"string":"[\n {\n \"formula_coordinates\": [\n 4,\n 108,\n 713.17,\n 66.91,\n 9.68\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"C 1 , ..., C N ∈ S.\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"Exploring Diverse In-Context Configurations for Image Captioning"},"abstract":{"kind":"string","value":"After discovering that Language Models (LMs) can be good in-context few-shot learners, numerous strategies have been proposed to optimize in-context sequence configurations. Recently, researchers in Vision-Language (VL) domains also develop their few-shot learners, while they only use the simplest way, i.e. , randomly sampling, to configure in-context image-text pairs. In order to explore the effects of varying configurations on VL in-context learning, we devised four strategies for image selection and four for caption assignment to configure in-context image-text pairs for image captioning. Here Image Captioning is used as the case study since it can be seen as the visually-conditioned LM. Our comprehensive experiments yield two counter-intuitive but valuable insights, highlighting the distinct characteristics of VL in-context learning due to multi-modal synergy, as compared to the NLP case. Furthermore, in our exploration of optimal combination strategies, we observed an average performance enhancement of 20.9 in CIDEr scores compared to the baseline. The code is given in https://github.com/yongliang-wu/ExploreCfg."},"authors":{"kind":"string","value":"Xu Yang; Yongliang Wu; Mingzhuo Yang; Haokun Chen; Xin Geng"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 2 :2Figure 2: Image selection strategies: (a) SIIR-CLIP, (b) SIIR-TAG, (c) DIIR-TT, (d) SICR-CLIP. 3 Configuring In-Context Sequences The in-context captioning can be treated as a vision-language conditioned text generation task. Given the multi-modal in-context sequence S = {(I 1 , C 1 ); (I 2 , C 2 ); ...; (I n , C n ); Î} that contains n-shot image-caption pairs (I, C) and one test image Î, we hope to generate the corresponding image caption Ĉ = { ŵ1 , ..., ŵT } in an auto-regressive manner. Here the t-th word ŵt is sampled from the following word distribution: P ( ŵt |S, ŵ1:t-1 ),(1)","figure_data":"","figure_id":"fig_3","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: The line charts of various in-context captions with diverse image-selection strategies.","figure_data":"","figure_id":"fig_4","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: The line charts of various in-context images with diverse caption-assignment strategies.4 Experiments 4.1 Dataset and Implementation Details MSCOCO.We evaluate the proposed strategies on MSCOCO dataset[14], which is the most widely used benchmark in image captioning. We used the Karpathy split[59] in the experiments, which contains 113,287/5000/5000 training/validation/test images and each image is associated with 5 human-annotated captions.","figure_data":"","figure_id":"fig_5","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: The histograms of various in-context captions with diverse image-selection strategies.","figure_data":"","figure_id":"fig_6","figure_label":"5","figure_type":"figure"},{"figure_caption":"Figure 6 :6Figure 6: The histograms of various in-context images with diverse caption-assignment strategies. six image selection strategies surpass those in Figure 4(b) where MGC-TF@88 is used. Further, low-quality captions, such as MGC-TF@66, may become \"toxic examples\" which can misguide the Vision-Language Model (VLM) as the shot number increases. Next we compare different kinds of captions to figure out what characteristics of the captions affect the performance of in-context captioning.","figure_data":"","figure_id":"fig_7","figure_label":"6","figure_type":"figure"},{"figure_caption":"Figure 7 (7a) visualizes 2 examples about the above comparisons.","figure_data":"","figure_id":"fig_8","figure_label":"7","figure_type":"figure"},{"figure_caption":"Figure 7 (7b) offers two comparison examples.","figure_data":"","figure_id":"fig_9","figure_label":"7","figure_type":"figure"},{"figure_caption":"(b) Experiment ( 2 )Figure 8 :28Figure 8: Four examples illustrate the phenomenon of short-cut inference. (a) When the in-context images are identical to the test image, the generated caption mirrors the in-context captions. (b) When using SIIR-CLIP to select similar examples, the generated caption tends to amalgamate features from both the in-context and test images, sometimes leading to ambiguous or partially accurate descriptions. (c) In contrast, when the in-context images are distinct from the test image, the generated caption more aptly describes the image, including specific words such as \"car\" or \"bananas\".","figure_data":"","figure_id":"fig_10","figure_label":"28","figure_type":"figure"},{"figure_caption":"Dog and cat sleeping on a couch. two dogs laying on top of a brown couch. Dog and cat lying on floor side by side sleeping. a dog and a cat laying on the floor. Cats sleeping on top of a bed with blue and white striped blankets. two cats laying on top of a bed. two white black and brown dogs are lying on a red couch. two dogs laying on top of a red couch. A couple of dogs sleeping on top of a couch. three dogs laying on top of a couch. two dachshunds and a cat sleep on a couch. two dogs laying on top of a couch. two dachshunds and a cat sleep on a couch. two dogs laying on top of a couch. two dachshunds and a cat sleep on a bed. two dogs laying on top of a bed. TWO DOGS LYING ON A BED ON TOP OF THE COVERS. dogs laying on top of a bed. A couple of dogs sleeping on top of a couch. three dogs laying on top of a couch. Cats sleeping on top of a bed with blue and white striped blankets. two cats laying on top of a bed. A bed in a hotel room with a window looking into the bathroom. a hotel room with two beds and a window. A bed in a hotel room with a window looking into the bathroom. a hotel room with two beds and a window. A very clean bedroom with two beds, lights and a TV. a hotel room with two beds and a television. Two beds in a hotel room with a window looking into the bathroom. a hotel room with two beds and a mirror. A bed in a bedroom next to a slide glass door. a hotel room with two beds and a table.","figure_data":"","figure_id":"fig_11","figure_label":"","figure_type":"figure"},{"figure_caption":"Apink bedspread is featured in this bedroom. a bedroom with a bed with two lamps. A pizza sitting on top of a wooden cutting board with different toppings. a person with a pizza sitting in a box on a","figure_data":"","figure_id":"fig_12","figure_label":"","figure_type":"figure"},{"figure_caption":"Akid is trying to catch a frisbee while standing in a park. a young girl flying a kite in the sky. A kid is trying to catch a frisbee while standing in a park. a young girl flying a kite in the sky. A person in a park holding a kite. a person holding a kite in a field. A man that is chasing a frisbee in the grass. a man playing with a frisbee in a park. A kid is trying to catch something while standing in a park. a woman playing with a frisbee in a park. A kid is trying to catch a frisbee while standing in a park. a young girl flying a kite in the sky. A person in a park holding a kite. a person holding a kite in a field. A man that is chasing a frisbee in the grass. a man playing with a frisbee in a park. A kid is trying to catch something while standing in a park. a woman playing with a frisbee in a park. A young boy flying a kite in a blue sky with clouds. a young boy flying a kite in the sky.","figure_data":"","figure_id":"fig_13","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 9 :9Figure 9: MGC-TF@135(blue) vs. GTC(red).Five examples demonstrate that more diverse words will be used in GTC than in MGC-TF@135 which making VLM hard to catch the major pattern. e.g., the first line and the second line incorrectly state \"cat\" and \"dachshunds.\"","figure_data":"","figure_id":"fig_14","figure_label":"9","figure_type":"figure"},{"figure_caption":"arome with a cup of coffee and a kettle. a clock tower with a clock on it. a close up of a clock on the side of a building reads 12:00. a clock tower with a clock on it. a close up of a clock on the side of a building reads 12:00. a clock tower with a clock on it. A clock on the side of a building. a clock tower with a clock on it. A clock on the side of a building reads 12:00. a clock tower with a clock on it. a close up of a clock on the side of a building. a clock tower with a clock on it. a clock on the side of a building. a laptop sitting on top of a desk with a desk and a desk.A woman sitting at a desk with a laptop in front of her. a laptop sitting on top of a desk with a desk and a desk.A woman sitting at a desk with a laptop in front of her. a laptop sitting on a desk with a desk.A laptop computer on a table with a cup of coffee. a laptop sitting on top of a desk.A laptop on a desk in a hotel room. a laptop sitting on top of a desk.A laptop on a table.","figure_data":"","figure_id":"fig_15","figure_label":"","figure_type":"figure"},{"figure_caption":"adesk with a desk and a desk.A man sitting at a desk with a laptop in front of him. a man riding a motorcycle on a motorcycle. A man riding a motorbike with a police radio. a man riding a motorcycle on a motorcycle. A man riding a motorbike with a police radio. a man riding a motorcycle on a motorcycle. a man riding a motorbike with a woman and a child on the back. a man riding a motorcycle on a motorcycle. A man riding a bicycle in front of a row of motorcycles. a group of people walking down a street. A man riding a motorbike with his daughter. a man riding a motorcycle on a motorcycle. A man rides a motorbike while talking on a cell phone. a vase with a vase in a vase and a vase. A vase filled with orange flowers. a vase with a vase in a vase and a vase. A vase filled with orange flowers. a table with a vase and flowers on it. a glass vase with orange flowers. a vase with a vase and a vase. A blue vase with a purple flower in it. a vase with a vase and a vase. A vase filled with purple flowers. a vase with a vase in a vase and a vase. A vase filled with orange flowers. a table with a vase and flowers on it. a glass vase with orange flowers. a vase with a vase and a vase. A blue vase with a purple flower in it. a vase with a vase and a vase. A vase filled with purple flowers. a vase with a vase and a vase. A vase made out of a newspaper.","figure_data":"","figure_id":"fig_16","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 10 :10Figure10: MGC-VLM(0)@63(blue) vs. MGC-TF@66(red). Five examples show that VLM will be misguided by the syntactic errors in MGC-VLM@66. Lots of outputs have repeated phrases, such as \"a table\", \"a clock\", \"a desk\", \"a motorcycle\" and \"a vase\".","figure_data":"","figure_id":"fig_17","figure_label":"10","figure_type":"figure"},{"figure_caption":"Figure 11 :11Figure 11: Five examples of verifying short-cut inference. we provide two different ways of choosing in-context images: \"identical to the test image\" (top row) and \"via random sampling\" (bottom row).","figure_data":"","figure_id":"fig_18","figure_label":"11","figure_type":"figure"},{"figure_caption":"The CIDEr scores of IP in different iterations.","figure_data":"Iter12345MGC-VLM(0)63.074.179.979.377.3MGC-VLM(32)85.380.579.478.977.1","figure_id":"tab_1","figure_label":"1","figure_type":"table"},{"figure_caption":"The results for verifying shortcut inference.","figure_data":"","figure_id":"tab_3","figure_label":"2","figure_type":"table"},{"figure_caption":"based on different Image Selection strategies. Additionally, we have calculated the average values for each method in the \"mean\" column.","figure_data":"Image Selection Caption Assignment4-shot8-shot16-shot 32 -shotmeanRSGTC75.8078.9381.9783.4780.04RSMGC-TF@6669.1864.5962.7360.4664.24RSMGCA-TF@6678.5282.4485.4886.5383.24RSMGC-TF@8874.6074.4777.8078.7176.39RSMGCA-TF@8878.4081.2884.9387.6283.06RSMGC-TF@13572.3570.1072.7377.7673.23RSMGCA-TF@13578.8180.6483.8686.7482.51RSMGC-VLM(0)@6370.4573.9274.8377.0074.05RSMGCA-VLM(0)@6376.1379.6182.1483.6380.38RSMGC-VLM(32)@8578.3380.9482.6282.7581.16RSMGCA-VLM(32)@8577.2280.1682.9785.0181.34","figure_id":"tab_4","figure_label":"6","figure_type":"table"},{"figure_caption":"Using Random Selection (RS) image selection strategy results.","figure_data":"Image Selection Caption Assignment4-shot8-shot16-shot 32 -shotmeanSIIR-CLIPGTC80.3288.7695.1898.2490.62SIIR-CLIPMGC-TF@6669.0866.7865.7965.5166.79SIIR-CLIPMGCA-TF@6689.6997.53102.31104.6398.54SIIR-CLIPMGC-TF@8881.8682.4384.7784.8983.49SIIR-CLIPMGCA-TF@8891.1198.23102.74106.2499.58SIIR-CLIPMGC-TF@13595.6496.6297.6698.3297.06SIIR-CLIPMGCA-TF@13592.7399.47104.09107.52100.95SIIR-CLIPMGC-VLM(0)@6365.9869.5271.8873.4970.22SIIR-CLIPMGCA-VLM(0)@6386.7293.3098.42100.8594.82SIIR-CLIPMGC-VLM(32)@8579.4681.3484.1484.8682.45SIIR-CLIPMGCA-VLM(32)@8588.0995.7098.98102.1296.22SIIR-TAGSGTC78.3786.4092.0894.9487.95SIIR-TAGSMGC-TF@6671.2368.2766.7765.1267.85SIIR-TAGSMGCA-TF@6687.4093.3197.85102.1695.18SIIR-TAGSMGC-TF@8882.7081.8082.5183.3382.58SIIR-TAGSMGCA-TF@8887.8694.2698.69102.6495.86SIIR-TAGSMGC-TF@13589.3591.1492.7394.8692.02SIIR-TAGSMGCA-TF@13588.8994.45100.34104.4497.03SIIR-TAGSMGC-VLM(0)@6367.1170.2172.2174.4571.00SIIR-TAGSMGCA-VLM(0)@6383.2389.3594.5198.7991.47SIIR-TAGSMGC-VLM(32)@8579.6981.9682.8284.2582.18SIIR-TAGSMGCA-VLM(32)@8585.0491.2396.10100.1093.12","figure_id":"tab_5","figure_label":"3","figure_type":"table"},{"figure_caption":"Using Similarity-based Image-Caption Retrieval (SIIR-CLIP) image selection strategy results.","figure_data":"","figure_id":"tab_6","figure_label":"4","figure_type":"table"},{"figure_caption":"Table 8 based on two different Image Selection strategies. Additionally, we have calculated the average values for each method in the \"mean\" column. Open-Flamingo v2 3B results with various strategies.From the values in the table, it can be seen that the trend is basically consistent with the v1 model. It can be considered that our strategy and analysis can be transferred to different models.","figure_data":"Image Selection Caption Assignment 4-shot8-shot16-shot 32 -shotmeanRSGT77.9086.1490.8093.1788.47RSMGC-TF@6682.1686.4988.1686.6186.55RSMGC-TF@8882.4788.4391.8193.8390.12RSMGC-TF@13584.6489.9390.7092.0390.31RSMGCA-TF@6677.9587.6492.5495.1490.09RSMGCA-TF@8878.9588.2192.1795.2890.19RSMGCA-TF@13579.6187.9991.9295.2289.95SIIRGT84.5392.3696.3698.6694.36SIIRMGC-TF@6684.7277.6975.2976.4977.09SIIRMGC-TF@8893.3891.2990.5089.2590.90SIIRMGC-TF@135104.10 103.80103.31102.39103.56SIIRMGCA-TF@6690.0097.51102.04103.5099.78SIIRMGCA-TF@8891.1598.88102.17104.55100.53SIIRMGCA-TF@13591.9199.01103.26104.93101.14Image Selection Caption Assignment 4-shot8-shot16-shot 32 -shotmeanRSGT83.4388.3691.8693.0990.11RSMGC-TF@13575.4980.7785.4689.5383.115RSMGCA-TF@13584.6789.692.4593.9791.025SIIRGT87.490.7294.4295.9492.57SIIRMGC-TF@13595.0297.3798.899.8898.085SIIRMGCA-TF@13590.9396.5997.08101.396.835","figure_id":"tab_7","figure_label":"7","figure_type":"table"},{"figure_caption":"Otter results with various strategies.","figure_data":"","figure_id":"tab_8","figure_label":"8","figure_type":"table"},{"figure_caption":"","figure_data":"a person with a pizza in a box.a pizza in a box on a table.a pizza sitting in a box on a table.a person with a pizza sitting in a box on a table.A box with pizza in it that has different toppings.A pizza sitting on top of a cardboard box on a table.There is an open box with a pizza inside the box.A pizza sitting on top of a wooden cutting board.A pizza sitting on top of a wooden cutting board with different toppings.","figure_id":"tab_9","figure_label":".","figure_type":"table"},{"figure_caption":"A long table with a flower arrangement in the middle for meetings.A table is adorned with wooden chairs with blue accents.A table is adorned with wooden chairs with pink accents.","figure_data":"A restaurant has modernA long restaurant tableA long table with a plantwooden tables andwith rattan roundedon top of it surroundedchairs.back chairs.with wooden chairs.(a)A restaurant has modernA long restaurant tableA long table with a plantA long table with aA table is adorned withA mother and daughterwooden tables andwith rattan roundedon top of it surroundedflower arrangement inwooden chairs with blueare carrying theirchairs.back chairs.with wooden chairs.the middle for meetings.accents.luggage.A long table with aA table is adorned withflower arrangement inwooden chairs with bluethe middle for meetings.accents.A restaurant has modernA long restaurant tableA long table with a plantA long table with aA table is adorned withA table is adorned withwooden tables andwith rattan roundedon top of it surroundedflower arrangement inwooden chairs with bluewooden chairs with redchairs.back chairs.with wooden chairs.the middle for meetings.accents.accents.(c)A restaurant has modernA long restaurant tableA long table with a plantA long table with aA table is adorned withA table is adorned withwooden tables andwith rattan roundedon top of it surroundedflower arrangement inwooden chairs with bluewooden chairs and achairs.back chairs.with wooden chairs.the middle for meetings.accents.bowl of dog food.A restaurant has modernA long restaurant tableA long table with a plantA long table with aA table is adorned withA table is adorned withwooden tables andwith rattan roundedon top of it surroundedflower arrangement inwooden chairs with bluewooden chairs with redchairs.back chairs.with wooden chairs.the middle for meetings.accents.accents.(d)A restaurant has modern wooden tables and chairs.A long restaurant table with rattan rounded back chairs.A long table with a plant on top of it surrounded with wooden chairs.A long table with a flower arrangement in the middle for meetings.A table is adorned with wooden chairs with blue accents.A small boy sits on a wooden chair.","figure_id":"tab_11","figure_label":"","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Image selection strategies: (a) SIIR-CLIP, (b) SIIR-TAG, (c) DIIR-TT, (d) SICR-CLIP. 3 Configuring In-Context Sequences The in-context captioning can be treated as a vision-language conditioned text generation task. Given the multi-modal in-context sequence S = {(I 1 , C 1 ); (I 2 , C 2 ); ...; (I n , C n ); Î} that contains n-shot image-caption pairs (I, C) and one test image Î, we hope to generate the corresponding image caption Ĉ = { ŵ1 , ..., ŵT } in an auto-regressive manner. Here the t-th word ŵt is sampled from the following word distribution: P ( ŵt |S, ŵ1:t-1 ),(1)\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: The line charts of various in-context captions with diverse image-selection strategies.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: The line charts of various in-context images with diverse caption-assignment strategies.4 Experiments 4.1 Dataset and Implementation Details MSCOCO.We evaluate the proposed strategies on MSCOCO dataset[14], which is the most widely used benchmark in image captioning. We used the Karpathy split[59] in the experiments, which contains 113,287/5000/5000 training/validation/test images and each image is associated with 5 human-annotated captions.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: The histograms of various in-context captions with diverse image-selection strategies.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 :6Figure 6: The histograms of various in-context images with diverse caption-assignment strategies. six image selection strategies surpass those in Figure 4(b) where MGC-TF@88 is used. Further, low-quality captions, such as MGC-TF@66, may become \\\"toxic examples\\\" which can misguide the Vision-Language Model (VLM) as the shot number increases. Next we compare different kinds of captions to figure out what characteristics of the captions affect the performance of in-context captioning.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 7 (7a) visualizes 2 examples about the above comparisons.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_8\",\n \"figure_label\": \"7\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 7 (7b) offers two comparison examples.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_9\",\n \"figure_label\": \"7\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"(b) Experiment ( 2 )Figure 8 :28Figure 8: Four examples illustrate the phenomenon of short-cut inference. (a) When the in-context images are identical to the test image, the generated caption mirrors the in-context captions. (b) When using SIIR-CLIP to select similar examples, the generated caption tends to amalgamate features from both the in-context and test images, sometimes leading to ambiguous or partially accurate descriptions. (c) In contrast, when the in-context images are distinct from the test image, the generated caption more aptly describes the image, including specific words such as \\\"car\\\" or \\\"bananas\\\".\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_10\",\n \"figure_label\": \"28\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Dog and cat sleeping on a couch. two dogs laying on top of a brown couch. Dog and cat lying on floor side by side sleeping. a dog and a cat laying on the floor. Cats sleeping on top of a bed with blue and white striped blankets. two cats laying on top of a bed. two white black and brown dogs are lying on a red couch. two dogs laying on top of a red couch. A couple of dogs sleeping on top of a couch. three dogs laying on top of a couch. two dachshunds and a cat sleep on a couch. two dogs laying on top of a couch. two dachshunds and a cat sleep on a couch. two dogs laying on top of a couch. two dachshunds and a cat sleep on a bed. two dogs laying on top of a bed. TWO DOGS LYING ON A BED ON TOP OF THE COVERS. dogs laying on top of a bed. A couple of dogs sleeping on top of a couch. three dogs laying on top of a couch. Cats sleeping on top of a bed with blue and white striped blankets. two cats laying on top of a bed. A bed in a hotel room with a window looking into the bathroom. a hotel room with two beds and a window. A bed in a hotel room with a window looking into the bathroom. a hotel room with two beds and a window. A very clean bedroom with two beds, lights and a TV. a hotel room with two beds and a television. Two beds in a hotel room with a window looking into the bathroom. a hotel room with two beds and a mirror. A bed in a bedroom next to a slide glass door. a hotel room with two beds and a table.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_11\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Apink bedspread is featured in this bedroom. a bedroom with a bed with two lamps. A pizza sitting on top of a wooden cutting board with different toppings. a person with a pizza sitting in a box on a\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_12\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Akid is trying to catch a frisbee while standing in a park. a young girl flying a kite in the sky. A kid is trying to catch a frisbee while standing in a park. a young girl flying a kite in the sky. A person in a park holding a kite. a person holding a kite in a field. A man that is chasing a frisbee in the grass. a man playing with a frisbee in a park. A kid is trying to catch something while standing in a park. a woman playing with a frisbee in a park. A kid is trying to catch a frisbee while standing in a park. a young girl flying a kite in the sky. A person in a park holding a kite. a person holding a kite in a field. A man that is chasing a frisbee in the grass. a man playing with a frisbee in a park. A kid is trying to catch something while standing in a park. a woman playing with a frisbee in a park. A young boy flying a kite in a blue sky with clouds. a young boy flying a kite in the sky.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_13\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 9 :9Figure 9: MGC-TF@135(blue) vs. GTC(red).Five examples demonstrate that more diverse words will be used in GTC than in MGC-TF@135 which making VLM hard to catch the major pattern. e.g., the first line and the second line incorrectly state \\\"cat\\\" and \\\"dachshunds.\\\"\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_14\",\n \"figure_label\": \"9\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"arome with a cup of coffee and a kettle. a clock tower with a clock on it. a close up of a clock on the side of a building reads 12:00. a clock tower with a clock on it. a close up of a clock on the side of a building reads 12:00. a clock tower with a clock on it. A clock on the side of a building. a clock tower with a clock on it. A clock on the side of a building reads 12:00. a clock tower with a clock on it. a close up of a clock on the side of a building. a clock tower with a clock on it. a clock on the side of a building. a laptop sitting on top of a desk with a desk and a desk.A woman sitting at a desk with a laptop in front of her. a laptop sitting on top of a desk with a desk and a desk.A woman sitting at a desk with a laptop in front of her. a laptop sitting on a desk with a desk.A laptop computer on a table with a cup of coffee. a laptop sitting on top of a desk.A laptop on a desk in a hotel room. a laptop sitting on top of a desk.A laptop on a table.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_15\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"adesk with a desk and a desk.A man sitting at a desk with a laptop in front of him. a man riding a motorcycle on a motorcycle. A man riding a motorbike with a police radio. a man riding a motorcycle on a motorcycle. A man riding a motorbike with a police radio. a man riding a motorcycle on a motorcycle. a man riding a motorbike with a woman and a child on the back. a man riding a motorcycle on a motorcycle. A man riding a bicycle in front of a row of motorcycles. a group of people walking down a street. A man riding a motorbike with his daughter. a man riding a motorcycle on a motorcycle. A man rides a motorbike while talking on a cell phone. a vase with a vase in a vase and a vase. A vase filled with orange flowers. a vase with a vase in a vase and a vase. A vase filled with orange flowers. a table with a vase and flowers on it. a glass vase with orange flowers. a vase with a vase and a vase. A blue vase with a purple flower in it. a vase with a vase and a vase. A vase filled with purple flowers. a vase with a vase in a vase and a vase. A vase filled with orange flowers. a table with a vase and flowers on it. a glass vase with orange flowers. a vase with a vase and a vase. A blue vase with a purple flower in it. a vase with a vase and a vase. A vase filled with purple flowers. a vase with a vase and a vase. A vase made out of a newspaper.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_16\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 10 :10Figure10: MGC-VLM(0)@63(blue) vs. MGC-TF@66(red). Five examples show that VLM will be misguided by the syntactic errors in MGC-VLM@66. Lots of outputs have repeated phrases, such as \\\"a table\\\", \\\"a clock\\\", \\\"a desk\\\", \\\"a motorcycle\\\" and \\\"a vase\\\".\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_17\",\n \"figure_label\": \"10\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 11 :11Figure 11: Five examples of verifying short-cut inference. we provide two different ways of choosing in-context images: \\\"identical to the test image\\\" (top row) and \\\"via random sampling\\\" (bottom row).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_18\",\n \"figure_label\": \"11\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"The CIDEr scores of IP in different iterations.\",\n \"figure_data\": \"Iter12345MGC-VLM(0)63.074.179.979.377.3MGC-VLM(32)85.380.579.478.977.1\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"The results for verifying shortcut inference.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"based on different Image Selection strategies. Additionally, we have calculated the average values for each method in the \\\"mean\\\" column.\",\n \"figure_data\": \"Image Selection Caption Assignment4-shot8-shot16-shot 32 -shotmeanRSGTC75.8078.9381.9783.4780.04RSMGC-TF@6669.1864.5962.7360.4664.24RSMGCA-TF@6678.5282.4485.4886.5383.24RSMGC-TF@8874.6074.4777.8078.7176.39RSMGCA-TF@8878.4081.2884.9387.6283.06RSMGC-TF@13572.3570.1072.7377.7673.23RSMGCA-TF@13578.8180.6483.8686.7482.51RSMGC-VLM(0)@6370.4573.9274.8377.0074.05RSMGCA-VLM(0)@6376.1379.6182.1483.6380.38RSMGC-VLM(32)@8578.3380.9482.6282.7581.16RSMGCA-VLM(32)@8577.2280.1682.9785.0181.34\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Using Random Selection (RS) image selection strategy results.\",\n \"figure_data\": \"Image Selection Caption Assignment4-shot8-shot16-shot 32 -shotmeanSIIR-CLIPGTC80.3288.7695.1898.2490.62SIIR-CLIPMGC-TF@6669.0866.7865.7965.5166.79SIIR-CLIPMGCA-TF@6689.6997.53102.31104.6398.54SIIR-CLIPMGC-TF@8881.8682.4384.7784.8983.49SIIR-CLIPMGCA-TF@8891.1198.23102.74106.2499.58SIIR-CLIPMGC-TF@13595.6496.6297.6698.3297.06SIIR-CLIPMGCA-TF@13592.7399.47104.09107.52100.95SIIR-CLIPMGC-VLM(0)@6365.9869.5271.8873.4970.22SIIR-CLIPMGCA-VLM(0)@6386.7293.3098.42100.8594.82SIIR-CLIPMGC-VLM(32)@8579.4681.3484.1484.8682.45SIIR-CLIPMGCA-VLM(32)@8588.0995.7098.98102.1296.22SIIR-TAGSGTC78.3786.4092.0894.9487.95SIIR-TAGSMGC-TF@6671.2368.2766.7765.1267.85SIIR-TAGSMGCA-TF@6687.4093.3197.85102.1695.18SIIR-TAGSMGC-TF@8882.7081.8082.5183.3382.58SIIR-TAGSMGCA-TF@8887.8694.2698.69102.6495.86SIIR-TAGSMGC-TF@13589.3591.1492.7394.8692.02SIIR-TAGSMGCA-TF@13588.8994.45100.34104.4497.03SIIR-TAGSMGC-VLM(0)@6367.1170.2172.2174.4571.00SIIR-TAGSMGCA-VLM(0)@6383.2389.3594.5198.7991.47SIIR-TAGSMGC-VLM(32)@8579.6981.9682.8284.2582.18SIIR-TAGSMGCA-VLM(32)@8585.0491.2396.10100.1093.12\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Using Similarity-based Image-Caption Retrieval (SIIR-CLIP) image selection strategy results.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Table 8 based on two different Image Selection strategies. Additionally, we have calculated the average values for each method in the \\\"mean\\\" column. Open-Flamingo v2 3B results with various strategies.From the values in the table, it can be seen that the trend is basically consistent with the v1 model. It can be considered that our strategy and analysis can be transferred to different models.\",\n \"figure_data\": \"Image Selection Caption Assignment 4-shot8-shot16-shot 32 -shotmeanRSGT77.9086.1490.8093.1788.47RSMGC-TF@6682.1686.4988.1686.6186.55RSMGC-TF@8882.4788.4391.8193.8390.12RSMGC-TF@13584.6489.9390.7092.0390.31RSMGCA-TF@6677.9587.6492.5495.1490.09RSMGCA-TF@8878.9588.2192.1795.2890.19RSMGCA-TF@13579.6187.9991.9295.2289.95SIIRGT84.5392.3696.3698.6694.36SIIRMGC-TF@6684.7277.6975.2976.4977.09SIIRMGC-TF@8893.3891.2990.5089.2590.90SIIRMGC-TF@135104.10 103.80103.31102.39103.56SIIRMGCA-TF@6690.0097.51102.04103.5099.78SIIRMGCA-TF@8891.1598.88102.17104.55100.53SIIRMGCA-TF@13591.9199.01103.26104.93101.14Image Selection Caption Assignment 4-shot8-shot16-shot 32 -shotmeanRSGT83.4388.3691.8693.0990.11RSMGC-TF@13575.4980.7785.4689.5383.115RSMGCA-TF@13584.6789.692.4593.9791.025SIIRGT87.490.7294.4295.9492.57SIIRMGC-TF@13595.0297.3798.899.8898.085SIIRMGCA-TF@13590.9396.5997.08101.396.835\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"7\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Otter results with various strategies.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"8\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"a person with a pizza in a box.a pizza in a box on a table.a pizza sitting in a box on a table.a person with a pizza sitting in a box on a table.A box with pizza in it that has different toppings.A pizza sitting on top of a cardboard box on a table.There is an open box with a pizza inside the box.A pizza sitting on top of a wooden cutting board.A pizza sitting on top of a wooden cutting board with different toppings.\",\n \"figure_id\": \"tab_9\",\n \"figure_label\": \".\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"A long table with a flower arrangement in the middle for meetings.A table is adorned with wooden chairs with blue accents.A table is adorned with wooden chairs with pink accents.\",\n \"figure_data\": \"A restaurant has modernA long restaurant tableA long table with a plantwooden tables andwith rattan roundedon top of it surroundedchairs.back chairs.with wooden chairs.(a)A restaurant has modernA long restaurant tableA long table with a plantA long table with aA table is adorned withA mother and daughterwooden tables andwith rattan roundedon top of it surroundedflower arrangement inwooden chairs with blueare carrying theirchairs.back chairs.with wooden chairs.the middle for meetings.accents.luggage.A long table with aA table is adorned withflower arrangement inwooden chairs with bluethe middle for meetings.accents.A restaurant has modernA long restaurant tableA long table with a plantA long table with aA table is adorned withA table is adorned withwooden tables andwith rattan roundedon top of it surroundedflower arrangement inwooden chairs with bluewooden chairs with redchairs.back chairs.with wooden chairs.the middle for meetings.accents.accents.(c)A restaurant has modernA long restaurant tableA long table with a plantA long table with aA table is adorned withA table is adorned withwooden tables andwith rattan roundedon top of it surroundedflower arrangement inwooden chairs with bluewooden chairs and achairs.back chairs.with wooden chairs.the middle for meetings.accents.bowl of dog food.A restaurant has modernA long restaurant tableA long table with a plantA long table with aA table is adorned withA table is adorned withwooden tables andwith rattan roundedon top of it surroundedflower arrangement inwooden chairs with bluewooden chairs with redchairs.back chairs.with wooden chairs.the middle for meetings.accents.accents.(d)A restaurant has modern wooden tables and chairs.A long restaurant table with rattan rounded back chairs.A long table with a plant on top of it surrounded with wooden chairs.A long table with a flower arrangement in the middle for meetings.A table is adorned with wooden chairs with blue accents.A small boy sits on a wooden chair.\",\n \"figure_id\": \"tab_11\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"[1; 2]\", \"Explanation\": \"The cited works are the Language Model (LM) that have emerged as a pivotal player in the field of Natural Language Processing (NLP), providing a unified approach to a range of diverse tasks by using a shared prompt paradigm.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[3; 4]\", \"Explanation\": \"The cited works are the use of the prompt paradigm in NLP tasks, which has been extended to a shared prompt paradigm in the field of Natural Language Processing (NLP).\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[5]\", \"Explanation\": \"The cited work is the few-shot prompt (or in-context learning) that has been proposed to enhance the effectiveness of the prompt paradigm in NLP tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work, Flamingo, is the basis for the in-context learning approach used in the citing paper to align well-trained large-scale vision and language models through cross-modal adapters.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[7]\", \"Explanation\": \"The cited work provides a body of research within the NLP field that demonstrates the effects of in-context configurations on few-shot performance, which the citing paper builds upon to explore the same effects in the context of visual language tasks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[8; 9; 10]\", \"Explanation\": \"The cited works explore the effects of in-context configurations in the context of few-shot performance, which the citing paper extends by focusing on the effects in the context of visual language tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work, MSCOCO, is used as the database for image selection in the citing paper, providing the basis for the image selection and captioning process.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work, Open-Flamingo, is used as a training model for the unofficial version utilized in the citing paper, indicating a methodological basis for the research conducted.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[17]\", \"Explanation\": \"The cited work, GPT, is a pre-trained language model that the citing paper adopts for predicting the next word in a sequence.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[1]\", \"Explanation\": \"The cited work, BERT, is a pre-trained language model that the citing paper uses for solving various downstream tasks in NLP.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[18]\", \"Explanation\": \"The cited work, BART, is a pre-trained language model that the citing paper adopts for solving downstream tasks in NLP.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[2]\", \"Explanation\": \"The cited work, GPT-3, is the first to introduce the prompt paradigm in NLP, which the citing paper further builds upon in the context of pre-trained LMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[19; 20; 21]\", \"Explanation\": \"The cited works propose prompt-tuning techniques for fine-tuning pre-trained LMs in NLP.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[22; 23; 24; 25]\", \"Explanation\": \"The cited works introduce the Chain-of-Thought technique for solving downstream tasks in NLP using pre-trained LMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[17]\", \"Explanation\": \"The cited work, GPT, is a pre-trained language model that the citing paper adopts for predicting the next word in a sequence.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[1]\", \"Explanation\": \"The cited work, BERT, is a pre-trained language model that the citing paper uses for solving various downstream tasks in NLP.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[18]\", \"Explanation\": \"The cited work, BART, is a pre-trained language model that the citing paper adopts for solving downstream tasks in NLP.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[2]\", \"Explanation\": \"The cited work, GPT-3, is the first to introduce the prompt paradigm in NLP, which the citing paper further builds upon in the context of pre-trained LMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[19; 20; 21]\", \"Explanation\": \"The cited works propose prompt-tuning techniques for fine-tuning pre-trained LMs in NLP.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[22; 23; 24; 25]\", \"Explanation\": \"The cited works introduce the Chain-of-Thought technique for solving downstream tasks in NLP using pre-trained LMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[17]\", \"Explanation\": \"The cited work, GPT, is a pre-trained language model that the citing paper adopts for predicting the next word in a sequence.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[1]\", \"Explanation\": \"The cited work, BERT, is a pre-trained language model that the citing paper uses for solving various downstream tasks in NLP.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[18]\", \"Explanation\": \"The cited work, BART, is a pre-trained language model that the citing paper adopts for solving downstream tasks in NLP.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[2]\", \"Explanation\": \"The cited work, GPT-3, is the first to introduce the prompt paradigm in NLP, which the citing paper further builds upon in the context of pre-trained LMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[19; 20; 21]\", \"Explanation\": \"The cited works propose prompt-tuning techniques for fine-tuning pre-trained LMs in NLP.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[22; 23; 24; 25]\", \"Explanation\": \"The cited works introduce the Chain-of-Thought technique for solving downstream tasks in NLP using pre-trained LMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[26; 27; 28; 29; 30; 31; 32; 33]\", \"Explanation\": \"The cited works provide a foundation for the development of VLM models that leverage large volumes of web-collected image-caption pairs to learn VL-generalizable embeddings, which the citing paper builds upon in their research on VLM models.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[34]\", \"Explanation\": \"The cited work on the prompt paradigm in NLP studies serves as a basis for the extension of the research in the citing paper to explore the use of prompt in VLM models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[35; 36]\", \"Explanation\": \"The cited works on updating all the parameters of VLM models are referenced in the citing paper to address the training burdens in the development of VLM models with prompt and in-context learning ability.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[35; 36; 37]\", \"Explanation\": \"The cited works on training vision and language models with adapters are used as a basis for the development of VLM models with the ability to align pre-trained vision and language models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[38; 39; 40]\", \"Explanation\": \"The cited works on training smaller networks with adapters to align pre-trained vision and language models are referenced in the citing paper to develop VLM models with in-context learning ability.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[41]\", \"Explanation\": \"The cited work on Frozen and Flamingo models is used as a basis for the development of VLM models with the ability to train vision and cross-modal adapters for in-context learning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work on in-context learning configurations in VLM models is referenced in the citing paper to explore the effects of different in-context configurations in the development of VLM models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[50]\", \"Explanation\": \"The cited work on image captioning provides a method for generating descriptive languages for images, which the citing paper adopts in their research on in-context captioning.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[51]\", \"Explanation\": \"The cited work on image retrieval provides a method for retrieving image-caption pairs, which the citing paper uses in their research on in-context captioning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[52; 53]\", \"Explanation\": \"The cited works on image captioning and retrieval are extended in the citing paper to explore the effects of in-context sequence configurations in multi-modal data.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[52]\", \"Explanation\": \"The cited work provides a method for in-context captioning that the citing paper adopts to study the same research topic.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[53]\", \"Explanation\": \"The cited work introduces a method for in-context captioning that the citing paper uses to enhance traditional retrieval-generation methods.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work, Flamingo, is a pre-trained Vision-Language Model (VLM) that the citing paper uses to calculate the probability of in-context learning in image captioning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[54]\", \"Explanation\": \"The cited work, Otter, is another pre-trained VLM that the citing paper uses to calculate the probability of in-context learning in image captioning.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[8]\", \"Explanation\": \"The cited work in the NLP field has shown that the performance of in-context learning varies significantly with different in-context configurations, which the citing paper uses as a data source to study the same effect in the case of image captioning.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[45]\", \"Explanation\": \"The cited work in the NLP field has shown the effect of in-context learning on the performance of image captioning, which the citing paper uses as a data source to study the same effect in the case of image captioning.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[48]\", \"Explanation\": \"The cited work in the NLP field has shown the effect of in-context learning on the performance of image captioning, which the citing paper uses as a data source to study the same effect in the case of image captioning.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[46]\", \"Explanation\": \"The cited work in the NLP field has shown the effect of in-context learning on the performance of image captioning, which the citing paper uses as a data source to study the same effect in the case of image captioning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[35]\", \"Explanation\": \"The cited work, CLIP, is used as a vision encoder in the SIIR-CLIP method to extract image embeddings for computing similarities in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[55]\", \"Explanation\": \"The cited work, Vinvl, is used to extract scene graph tags for images in the SIIR-TAG method, which is employed in the citing paper to compute similarities between images.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[56]\", \"Explanation\": \"The cited work, IETrans, is used to extract relations from images in the SIIR-TAG method, which is employed in the citing paper to compute similarities between images.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[35]\", \"Explanation\": \"The cited work, CLIP, is also used in the SICR-CLIP method to compute similarities between images and captions in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[57]\", \"Explanation\": \"The cited work provides a method for controlling the number of shots in the DIIR-TT strategy by incorporating class tags generated by the IETrans model, which the citing paper adopts to extend the basis of SIIR-Tags to four categories.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work, the MSCOCO dataset, is the source of the GTCs used in the citing paper. The citing paper uses the GTCs from the first image in the dataset to conduct their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(VinVL features)\", \"Explanation\": \"The cited work provides the features used in the MGC model, which serves as a methodological basis for the generation of captions in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(CIDEr score [58])\", \"Explanation\": \"The cited work provides the CIDEr score metric used to evaluate the quality of the generated captions in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work serves as a benchmark for image captioning, providing a standard for evaluation and comparison in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[59]\", \"Explanation\": \"The cited work provides the Karpathy split dataset used in the experiments of the citing paper, which is a crucial data source for the research conducted.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The Open-Flamingo model cited in the text is used as a method to test the strategies discussed in the citing paper, providing a specific approach to evaluate the research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[6]\", \"Explanation\": \"The use of 4, 8, 16, and 32 shots in the experiments is an extension of the research conducted in the cited work, exploring new dimensions and variables in the study of image captioning.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[57]\", \"Explanation\": \"The cited work provides evidence that diversity can be crucial in certain complex NLP challenges, which supports the contention that the strength of diversity may not always translate to superior performance in all contexts.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work, Open-Flamingo, is the only open-source multi-modal few-shot learner available to the citing paper. The citing paper uses this work as a basis for their research on image captioning and the impacts of multimodal in-context configurations on few-shot vision-language prompting.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work serves as the data source for the training of Open-Flamingo, which the citing paper uses in its research to compare the performance of the model with the official Flamingo.\"}]"}}},{"rowIdx":71,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b0"],"table_ref":[],"text":"In-context learning (ICL) with large language models (LLMs) has shown great potential in performing a wide range of language tasks (Brown et al., 2020). ICL has the unique advantages of being data-efficient (i.e., only a few labeled training examples are needed) and accessible (i.e., expertise in training models is no longer required). With these advantages, a non-expert user can create a system to perform a new task within minutes by writing a few examples. This gives rise to the popularity of ating ICL with a labeled test set is a direct solution to know whether ICL will be effective, it greatly reduces the appeal of ICL, as one of ICL's key selling points is that it does not require a large labeled dataset. In addition, many tasks do not come with a labeled test set due to high annotation costs (e.g., medical/law-related questions that require professional knowledge to answer). In such cases, it is highly desirable to estimate the ICL performance without a labeled test set. This would help system developers determine whether ICL is likely to be useful for their problems of interest.\nGuided by this motivation, we formalize the problem of few-shot ICL accuracy estimation: given a handful of labeled in-context examples and a set of unlabeled test examples, our goal is to estimate the overall accuracy of ICL on these test examples. Our contributions are twofold:\n• We propose to address the accuracy estimation problem by training a \"meta-model,\" which takes in LLM confidence features as input and outputs the task accuracy. The meta-model is trained with observed ICL accuracies on seen datasets, and then used to estimate ICL accuracy on unseen datasets (see Figure 1).\n• We obtain 42,360 observations of LLM ICL performance, by conducting extensive ICL experiments spanning two tasks (multiplechoice QA and closed-book QA), 91 datasets, and 4 LLMs. We then benchmark the metamodel method and multiple baselines on a total of 12 evaluation settings derived from these observations.\nOur meta-model can estimate ICL accuracies without the need for labeled test examples. In 10 out of 12 settings, the meta-model estimates are at least as accurate as directly evaluating on 16 labeled examples. In 2 out of 12 settings, they match with evaluating on 128 labeled examples. On average, we are able to save the annotation cost of 40 test labels per task by using the meta-model. Further, the meta-model outperforms all baseline methods in 8 out of 12 settings, improving the relative estimation error by 23.6% However, we also find that there exists substantial room for improvement across all settings. We envision estimating ICL accuracy without labeled test data as an open challenge and encourage the community to develop new techniques that can more accurately predict when ICL will be effective."},{"figure_ref":[],"heading":"Related Work","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Model Confidence and Calibration","publication_ref":["b5","b13","b14","b9","b5","b15"],"table_ref":[],"text":"Calibration of LLMs has been studied on a diverse range of tasks such as classification (Desai and Durrett, 2020) and question answering (QA) (Jiang et al., 2021;Kadavath et al., 2022). It aims to study whether LLMs assign meaningful correctness likelihood-also known as model confidence-to the outputs (Guo et al., 2017). Most prior work evaluates calibration at the example level (Desai and Durrett, 2020;Kamath et al., 2020); in this paper, we focus on using overall model confidence distributions to estimate dataset-level accuracies. We propose a method to learn model calibration patterns based on observations of LLMs' performance at the dataset level."},{"figure_ref":[],"heading":"In-context Learning","publication_ref":["b0","b4","b19","b31","b21"],"table_ref":[],"text":"LLMs pre-trained with auto-regressive language modeling objectives have been shown to be capable of \"learning\" in context when given a prompt composed of a prompt template and a few labeled demonstrations (Brown et al., 2020;Chowdhery et al., 2022). While LLMs can learn a new task only through model inference, the accuracy is sensitive to the choices of prompt templates and incontext examples (Lu et al., 2021;Zhao et al., 2021;Perez et al., 2021). Therefore, we aim to develop a method to accurately estimate ICL performance for a dataset prompted with any prompt template and combination of in-context examples."},{"figure_ref":[],"heading":"Out-of-distribution (OOD) Prediction","publication_ref":["b8","b6","b28","b24","b18","b6","b24"],"table_ref":[],"text":"Machine learning models in the real world commonly encounter distribution shifts between training and test time. Prior work (Guillory et al., 2021;Garg et al., 2022;Yu et al., 2022;Singhal et al., 2022;Li et al., 2022) aims to predict models' OOD performance under different setups. Garg et al. (2022) predict target domain accuracy for image classification tasks with distribution by fitting a threshold on model confidence using only labeled source data and unlabeled target data. Singhal et al. (2022) use a few additional target-domain examples to predict the accuracy, focusing on known source-target dataset pairs on which models often have low OOD accuracy due to overfitting to spurious correlations (e.g., MNLI-HANS and QQP-PAWS). They find that accuracy on the given small set of target examples is a strong baseline to approximate accuracy on the full-test set. We include the accuracy for a small set of labeled test examples as an oracle baseline (see Section 3.3). These papers all try to predict the OOD accuracy of a model trained on in-distribution training data; in contrast, in our setting we have access to some labeled datasets but the language models we study were never finetuned on those datasets. In order to avoid confusion, we instead use the terms \"seen/unseen tasks\" to describe the datasets available to us, rather than \"in-distribution/out-of-distribution.\"\n3 Accuracy Prediction"},{"figure_ref":[],"heading":"Problem Definition","publication_ref":[],"table_ref":[],"text":"We formalize the task of ICL accuracy estimation for unseen datasets given observations of the same model's performance on other datasets. A method for the ICL accuracy estimation task takes in four inputs: a language model M ; a set of labeled seen datasets {D i } r i=1 , where each D i consists of a set of labeled examples {(x\n(1) i , y(1)\ni ), . . . , (x\n(n i ) i , y (n i ) i\n)} and n i = |D i |; a prompt c for the test task; and an unlabeled test dataset\nD test = {x (1) test , . . . , x(m)\ntest } of size m. In a typical setting, each seen task should consist of a sufficient amount of labeled examples, i.e., n i ≥ 100. The method should output the estimated accuracy of M on D test when prompted with prompt c; we denote the actual accuracy of the model as acc M,c test and acc M,c test as the predicted accuracy. Note that with the labeled datasets D i and a corresponding prompt c, we can compute the corresponding dataset-level ICL accuracy acc M,c i for i = 1, . . . , r."},{"figure_ref":[],"heading":"Prompt Formulation and Data Splits","publication_ref":[],"table_ref":[],"text":"We construct prompts by sampling k in-context examples uniformly at random from available labeled data and formatting them with prompt templates to form a prompt (see Section B and "},{"figure_ref":[],"heading":"Comparing with Labeled Test Data","publication_ref":["b24"],"table_ref":[],"text":"To put our results in context, we compare all methods to the Oracle approach of sampling l labeled examples from the test dataset D test and measuring accuracy on those l examples, which we call oracle l . This approach is used by Singhal et al. (2022) and it represents how well we can evaluate ICL performance for D test by collecting labeled examples. With a large value of l, we get a better evaluation of the test dataset at the cost of collecting expensive annotations. In proposing the task of accuracy prediction, we hope to develop methods that outperform the l-labeled oracle for values of l that represent non-trivial annotation costs."},{"figure_ref":[],"heading":"Confidence Profile Meta-Model","publication_ref":["b0","b30"],"table_ref":[],"text":"We propose a new method that trains a meta-model based on the confidence profiles of seen datasets {D i } r i=1 to estimate ICL performance. We use the term confidence profile to denote the distribution of model confidence scores on each example in the dataset. We extract the confidence profiles (see Figure 1) from each seen dataset and convert them to a feature vector. We then train a meta-model to map confidence feature vectors to the datasetlevel ICL accuracies. The benefits of using the confidence feature vector are twofold. First, we do not need any labeled test data, which saves annotation costs. Second, this approach is applicable to any pre-trained language model like GPT3 (Brown et al., 2020) and OPT (Zhang et al., 2022)."},{"figure_ref":[],"heading":"Confidence Profile","publication_ref":[],"table_ref":[],"text":"In general, given a (not-necessarily labeled) dataset D, LM M , and a prompt c, we obtain the confidence profile by first computing the confidence score s M,c (x) for each x ∈ D. The score for each input x can be computed by one forward pass of M ; the exact value of the score differs based on the task, as described below. Next, we sort the scores to obtain a list [s 1 , . . . , s |D| ] where each s i ≤ s i+1 . Then we create a d conf -dimensional feature vector conf M,c D , whose i-th component is a linear interpolation between s ⌊|D|×i/d⌋ and s ⌈|D|×i/d⌉ . Intuitively, the i-th feature represents the i/|D|-th percentile confidence score. We refer to the feature vectors derived from confidence profiles as confidence vectors."},{"figure_ref":[],"heading":"Confidence scores","publication_ref":["b14"],"table_ref":[],"text":"The confidence score s M,c (x) is calculated differently for closed-set generation and open-ended generation.\nClosed-set generation. Closed-set generation tasks have a pre-defined label space Y. We take outputs from LLMs and identify the answers only by labels (Kadavath et al., 2022). For each example, we take model confidence as the normalized probability across the label space:\ns M,c (x) = p ŷ ỹ∈Y p ỹ (1)\nwhere p ỹ is the model-assigned probability for label ỹ on input x, p ŷ is the probability for the output label ŷ from model M , and ŷ = arg max ỹ∈Y p ỹ.\nOpen-ended generation. We refer to tasks that require sequence generation (e.g., closed-book QA, summarization, machine reading comprehension, etc.) as open-ended generation tasks. We use negative log-likelihood (NLL) 2 to obtain confidence scores from each generated sequence. Let ŷ be the model-generated sequence. We compute the confidence score as:\ns M,c (x) = - |ŷ| t=1 log p t (ŷ t ). (2\n)\np t is the model-assigned probability distribution at output token t and ŷt is the t-th output token"},{"figure_ref":[],"heading":"Meta-Model Training Data","publication_ref":[],"table_ref":[],"text":"For each seen dataset {D i } r i=1 , we sample K prompts {c ij } K j=1 . Then for each prompt sampled we compute the confidence vector conf 2 We also tried perplexity but found NLL to yield better results.\nM,c ij i and accuracy acc M,c ij i to create one meta-training ex- ample (conf M,c ij i , acc M,c ij i ). This creates a total of r × K meta-"},{"figure_ref":[],"heading":"Meta-Model Architectures","publication_ref":[],"table_ref":[],"text":"We choose meta-models that are easy to train and contain far fewer parameters than LLMs for computational efficiency. In this paper, we consider three meta-model architectures. First, we use k Nearest Neighbors regression (k-NN), which measures feature similarity. In the context of this paper, k-NN retrieves the most similar confidence profile from the seen datasets to the test dataset confidence profile and predicts based on the observed ICL accuracy on the retrieved in-distribution datasets. We use the implementation in scikit-learn library. 3Second, we use a two-layer Multilayer Perceptron (MLP) that takes confidence feature vectors as input. Third, we use the tree-based method XG-Boost (Chen and Guestrin, 2016) with the same confidence features. We use XGBoostRegressor implemented in the XGBoost library4 and tune the hyperparameters as described in Appendix C.2.5 "},{"figure_ref":[],"heading":"Evaluation","publication_ref":[],"table_ref":[],"text":"For task performance evaluation, we use Exact Match (EM) accuracy to measure accuracy for closed-set QA, and F1-score to measure accuracy for open-ended QA.\nWe evaluate accuracy prediction models based on absolute error, defined as |acc M,ctest test acc M,ctest test |, where both are computed using the test dataset D test with a test prompt c test . We then average the absolute error over all prompts C test and compute the dataset-specific mean average error:\nerr Dtest = 1 |C test | c∈Ctest |acc M,c test -acc M,c test |.\nFinally, to evaluate the overall success of accuracy prediction across a collection of test datasets T , we measure mean absolute error (MAE), defined as:\nerr T = 1 |T | Dtest∈T err Dtest\n(3)"},{"figure_ref":[],"heading":"Baselines","publication_ref":["b20","b9","b16","b5","b12","b9","b23","b6"],"table_ref":[],"text":"We consider four baselines for accuracy estimation.\nAverage training accuracy (AVGTRAIN). We simply take the average dataset-level accuracy of the seen datasets as our accuracy estimation:\nacc AVGTRAIN = 1 r × K r i=1 K j=1 acc M,c ij i .\nAverage Calibration Error (AVGCONF). We take the average confidence across the test dataset as the accuracy estimation:\nacc AVGCONF = 1 |C test | × m c∈Ctest x∈Dtest s M,c (x).\nNote that this baseline is only applicable to closedset generation and not open-ended generation tasks since the accuracy metric for open-ended generation (F1 score) and confidence metric (NLL) do not share the same range ([0, 1] vs. (-∞, 0]). The intuition behind AVGCONF is that if the model confidence scores are well-calibrated (at the example level), then the expected value of the model's confidence scores should be equal to the accuracy. In fact, we note that the MAE of AVGCONF is similar to Expected Calibration Error (ECE), which measures the example-level calibration error (Naeini et al., 2015;Guo et al., 2017;Kumar et al., 2019;Desai and Durrett, 2020). 6Temperature Scaling (TS) Temperature scaling is a widely used calibration method (Hinton et al., 2015;Guo et al., 2017;Si et al., 2022). By fitting a single scaler parameter called temperature τ , it produces softer model-assigned probabilities\np ŷ = exp (z ŷ/τ ) ỹ exp (z ỹ∈Y /τ )\n.\nWe then obtain scaled confidence scores with Equation 1, and evaluate AVGCONF on the test dataset. Note that we optimize temperature τ based on the AVGCONF of the training datasets instead of the common approach of using NLL as an objective function.\nAverage Threshold Confidence (ATC) We use ATC (Garg et al., 2022) as one of our OOD accuracy estimation baselines. ATC takes accuracy estimation by fitting a confidence threshold on a single source dataset and generalizes to the target dataset. We take the estimated accuracy for the test dataset to be the average of the ATC estimates from each seen dataset:\nacc AT C = 1 r r i=1 atc M,c i,test ,\nwhere atc M,c i,test is the D i to D test ATC estimate."},{"figure_ref":[],"heading":"Alternative Featurizations","publication_ref":[],"table_ref":[],"text":"In addition to the confidence profiles, we experiment with another featurization method that uses model embeddings from the LM M . Given a dataset D, LM M , and prompt c, we obtain the model embedding by first taking the last-layer, lasttoken embedding e M,c (x) for each x ∈ D, and then averaging across the dataset:\nembed M,c D = 1 |D| x∈D e M,c (x).\nSince embed\nM,c D\nis very high-dimensional (e.g., 5120 dimensional for 13B models), we use Principle Component Analysis (PCA) to reduce its dimensionality. We fit the PCA model on all dataset embedding vectors embed M,c D and transform them into d e -dimensional vectors embed M,c D , which we can use as a feature vector. As an additional experiment, we can concatenate the confidence vector and embedding vector to form a combined feature vector:\nce M,c D = conf M,c D + embed M,c D .\nReducing the dimensionality makes the comparison with confidence features more fair, and does not dilute the influence of confidence features when concatenating them with embedding features."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Accuracy Estimation Benchmark","publication_ref":["b11"],"table_ref":[],"text":"We benchmark both our meta-model method for ICL accuracy estimation and the baseline methods mentioned in Section 4.6 on a total of 12 LLMdataset collection pairs (3 dataset collections × 4 LLMs). For each evaluation setting, we evaluate 3 different featurization methods mentioned in Section 4.7. This adds up to 36 experiment settings.\nDatasets. We use three different collections of datasets in total: multiple-choice QA (MCQA) from MMLU (Hendrycks et al., 2020) and both"},{"figure_ref":[],"heading":"LLaMA-7B","publication_ref":["b17","b30","b26","b26","b22"],"table_ref":[],"text":"LLaMA-13B OPT-6.7B OPT-13B 5 for the full list of tasks). We henceforth use MCQA and CBQA to refer to the CrossFit dataset collections respectively. We use the implementations and training/test partitions from HuggingFace Datasets (Lhoest et al., 2021). We split each collection of datasets into meta-training/test splits using 5-fold cross-validation-we partition each dataset collection into 5 equal-sized subsets and run five versions of each experiment, one with each subset being used as the meta-test set and the remaining subsets used as meta-training data. We take the average of the meta-test results as our final result.\nLLMs. We run our experiments on four LLMs: OPT-6.7B, OPT-13B (Zhang et al., 2022), LLaMA-7B, and LLaMA-13B (Touvron et al., 2023). We use the OPT models from HuggingFace Models7 and the LLaMA model from Meta AI.8 More details are included in Appendix C.1.\nExperimental Details. We generate prompts for each dataset using the method noted in Section 3.2. For each dataset in MMLU,9 we combine the \"validation\" set and \"dev\"10 set to be the training set 11 We choose up to 5-shot setting because it is studied in previous studies (Touvron et al., 2023;Rae et al., 2021). prompts for MCQA/CBQA datasets) for MMLU because it contains a very large number of datasets."},{"figure_ref":["fig_1"],"heading":"Main Results","publication_ref":[],"table_ref":["tab_2","tab_4"],"text":"The meta-model outperforms all baselines under certain evaluation settings. Table 1 shows the meta-model estimation error for each evaluation setting. For 8 out of 12 settings (all CBQA settings, LLaMA-7B on MCQA, LLaMA-13B on MMLU, both OPT models on MCQA), the best meta-model architecture has 23.67% lower relative MAE than the best baseline method on average. In the best case (OPT-6.7B on MCQA), the meta-model can achieve 43.5% lower relative MAE than all baselines. However, for the other 4 settings (both OPT models on MMLU, LLaMA-7B on MMLU, and LLaMA-13B on MCQA), baseline methods provide more accurate estimates of ICL accuracy. Fig- ure 2 shows the evaluation results graphically. On average across all 12 settings, the best estimation errors from the meta-models are 32.5% less than the actual accuracy standard deviations. In 11 out of 12 settings, the estimation errors are within one standard deviation of the actual accuracy.\nOracle baselines indicate useful accuracy estimations. In comparison to the Oracle baselines, the meta-model outperforms the oracle 32 baseline in all MMLU and MCQA settings except for LLaMA-13B on MCQA (achieves oracle 8 ) and outperforms the oracle 16 baseline in all CBQA settings except for LLaMA-13B (achieves oracle 8 ).\nIn the two best-case settings (using XGBoost as the meta-model on MCQA with either OPT model), the meta-model achieves the oracle 128 baseline, i.e., is equivalent to estimating the accuracy using 128 annotations.\nBaseline methods are effective in some settings.\nWhile ATC is a weak baseline for ICL accuracy estimation, AVGTRAIN, AVGCONF, and TS are strong baselines for MMLU and MCQA. AVG-TRAIN is able to achieve oracle 32 for 3 out of 12 settings (LLaMA-7B, OPT-6. for 5 settings (LLaMA-13B on MMLU, both OPT models on MMLU, and OPT-6.7B on MCQA).12 \nAblation on Model Architecture Across three meta-model structures, the XGBoost meta-model overall provides the most accurate estimation as it has the lowest MAE for 7 out of 12 evaluation settings. The average MAE is 5.88 for XG-Boost meta-models, 5.94 for 3-NN meta-models, and 7.18 for MLP meta-models. Surprisingly, 3-NN meta-models have a lower average MAE than MLP meta-models despite having a simpler model structure. In Figure 3, we show that the XGBoost meta-model provides well-correlated accuracy estimation across 4 different evaluation settings.\nAblation on Featurization Methods We consider three featurization methods as described in Section 4.7. Table 2 in the appendix shows that the best overall accuracy estimation for all settings is attained by using the confidence vectors as metafeatures (achieves the lowest MAE for 26 out of 36 evaluation settings). The average MAE is 6.27 for conf , 8.34 for embed, and 7.36 for ce. Further, using conf as features demonstrates a more dominant advantage for all CBQA tasks, achieving the lowest MAE for 11 out of 12 evaluation settings."},{"figure_ref":["fig_2","fig_2"],"heading":"Effect of Unlabeled Data and Confidence Vector Dimensions","publication_ref":[],"table_ref":[],"text":"We now study confidence feature vector ablations by varying the number of unlabeled test examples m in each unseen dataset and the dimension of the confidence vector d conf . We test with OPT-13B on MCQA datasets using the XGBoost meta-model since we achieve the lowest MAE in this setting.\nFigure 4 shows that increasing m enables better accuracy estimation, reducing the average MAE (across all d conf ) from 3.92 for m = 200 to 2.56 for m = 1000. Note that increasing m requires performing additional LLM inferences on unlabeled examples, so leveraging unlabeled test data is constrained by computational cost considerations. The quality of our accuracy estimates does not vary much as we change the confidence vector dimension d conf , as shown in Figure 4."},{"figure_ref":[],"heading":"Effect of Number of Shots","publication_ref":[],"table_ref":[],"text":"We compare ICL accuracy estimation performance given different k-shot ICL accuracy observations for LLaMA-13B on MMLU datasets. Table 4 in the Appendix shows that the meta-model produces a slightly better ICL accuracy estimation for the 3-shot setting. Overall, the meta-model gives consistent accuracy estimates across different k-shot settings as they all achieve oracle 32 ."},{"figure_ref":["fig_3"],"heading":"Prompt Selection","publication_ref":["b31","b21","b23"],"table_ref":[],"text":"Previous works demonstrated that ICL performance is highly sensitive to the prompt templates as well as in-context examples (Zhao et al., 2021;Perez et al., 2021;Chen et al., 2022); we are thus interested in whether our ICL accuracy estimation method can be applied to select the best ICL prompt c ∈ C test for the test dataset. For each dataset, we use the XGBoost meta-model to select the best prompt c * , as opposed to the actual best prompt c * . We then compute the corresponding ICL accuracies and compare them to the average accuracy across all test prompts. Figure 5 shows that there is a significant difference in ICL accuracy given different prompts for all 12 settings, and the selected prompts lead to better ICL accuracies than the average accuracy for 7 out of 12 settings. On average, the selected prompt is 15.6% as effective as the actual best prompt. The limited improvement from the random baseline indicates there's a large room for improvement and we encourage future work to derive a better prompt selection standard."},{"figure_ref":[],"heading":"Discussion and Conclusion","publication_ref":[],"table_ref":[],"text":"In this paper, we study the problem of few-shot ICL accuracy estimation. We propose training a meta-model based on LLM confidence features and observed accuracies on seen datasets. We show that without using any labeled test data, the meta-model is often able to attain accurate estimates of ICL accuracy, which is practically useful for predicting LLMs' accuracy on datasets that have high annotation costs. We also construct a large-scale benchmark for dataset-level ICL accuracy estimation by evaluating the meta-model and multiple baseline methods across 12 evaluation settings and 3 metafeature options. We observe that while some baseline methods can provide good accuracy estimates, our meta-model demonstrates non-trivial improvement in estimation abilities over baseline methods in 8 out of 12 evalutation settings. We encourage future work to develop better meta-model architectures as well as better metafeatures and study potential implications for the meta-model, such as acting as a prompt template/ICL example selection method. We believe that our benchmark can serve as an open challenge for improving dataset-level ICL accuracy estimations, leading to an improved understanding of when ICL is likely to be effective."},{"figure_ref":[],"heading":"Limitations","publication_ref":["b7"],"table_ref":[],"text":"While we conducted extensive experiments to study ICL accuracy estimations, there are many more LLMs that have exhibited impressive capabilities on a variety of tasks. Due to computational constraints, we do not benchmark accuracy estimations based on LLMs with limited access (e.g., GPT-4 (OpenAI, 2023)) as it is difficult to extract model embedding features, or those larger than 13B. We also don't consider instruction-tuned models to avoid possible overlaps between their training datasets and our evaluation datasets. Meanwhile, instruction tuning sometimes hurts model performance on canonical datasets such as MMLU, as shown in Gudibande et al. (2023). It might also significantly hurt calibration as reported in OpenAI (2023). For the same reasons, we include only a limited number of prompt templates and in-context example variations for ICL prompting. While we choose only 3 few-shot settings for MMLU and 2 for MCQA and CBQA, it is possible to achieve better accuracy estimations with more observations in the training data.\nIn terms of dataset selection, we use 13 closedbook QA tasks for the open-ended generation setting. Our findings might not generalize to other open-ended generation tasks such as summarization or long-form question answering. Overall, the meta-model provides effective accuracy estimations, but there's still substantial room for improvement. "},{"figure_ref":[],"heading":"Acknowledgements","publication_ref":[],"table_ref":[],"text":"We would like to thank Ting-Yun Chang for her valuable contributions. We thank Ameya Godbole, Johnny Wei, and Wang Zhu for their valuable discussions. We also thank all members of the Allegro lab at USC for their support and valuable feedback. RJ was supported by an Open Philanthropy research grant, a Cisco Research Award, and HF was supported by the USC Provost Fellowship Award."},{"figure_ref":[],"heading":"A Dataset Implementation Details","publication_ref":[],"table_ref":[],"text":"Following Section 5.1's discussion of specific dataset implementations for MCQA and CBQA, our approach for each setting is: for tasks that have a defined train/test split in HuggingFace Datasets "},{"figure_ref":[],"heading":"B Prompt Templates","publication_ref":["b22","b11"],"table_ref":[],"text":"We collect 5 general prompt templates for MMLU datasets: 1 Null template, 1 self-constructed prompt template, 2 from previous work (Rae et al., 2021;Hendrycks et al., 2020), and 1 generated by ChatGPT. For MCQA and CBQA, we only use the Null template due to resource considerations. "},{"figure_ref":[],"heading":"C Implementation Details","publication_ref":[],"table_ref":[],"text":"We will release code to reproduce our results upon publication."},{"figure_ref":[],"heading":"C.1 LLMs Implementations","publication_ref":["b26","b30"],"table_ref":[],"text":"We use LLaMA-7B and LLaMA-13B (Touvron et al., 2023) from Meta AI, and OPT-6.7B and OPT-13B (Zhang et al., 2022) from HuggingFace transformers. For LLaMA-7B, OPT-6.7B, and OPT-13B, we run evaluations using a single RTXA6000 GPU (48GB). For LLaMA-13B, we run evaluations by parallel inference on two RTXA6000 GPUs. We use half-precision for both OPT models. Note that some LLMs (except for OPT-13B) can be run on GPUs with a smaller memory. We evaluate 42,360 ICL observations in total, where each observation is a dataset with 100 to 1000 examples. The total inference process takes around 2000 GPU hours."},{"figure_ref":[],"heading":"C.2 Meta-model Implementations","publication_ref":["b2"],"table_ref":[],"text":"All meta-model architectures can be trained on an i7-10700 CPU. The total training time of the metamodel on one experiment setting varies from 1.5 hours to 24 hours depending on training data dimensions. We include the implementation details for each of the meta-model architectures. We use the random seed 1 for all processes involving randomness. For K Nearest Neighbors regression, we use the implementation of KNeighborsRegressor from sklearn library. We use euclidean distance as the weight metric, and fit the model on the metatraining data.\nFor MLP, we implement with Pytorch. We use a 2-layer MLP with the size of the hidden_state = 1536, learning_rate = 1e -5, and dropout_rate = 0.2. We use Adam Optimizer and MSELoss, and we perform early stopping with the validation data. The validation data is a 20% random partition of the meta-training data. For early stopping, the max epoch is 50 and the patience is 7.\nFor XGBoost, we implement with the XGBRegressor from the XGBoost library (Chen and Guestrin, 2016). We use a 5-fold random search cross-validation for 300 iterations to choose the hyperparameter. The candidates are: 1 { 2 \" lr \" : uniform ( 0 . 0 1 , 0 . 5 ) , 3\n\" max_depth \" : randint ( 3 , 1 0 ) , 4\n\" n_estimators \" : randint ( 1 0 0 , 1 0 0 0 ) , 5\n\" colsample_bytree \" : [ 0 . 5 , 0 . 6 , 0 . 7 , 0 . 8 , 0 . 9 , 0. 9 5 , 1 ] , 6\n\" subsample \" : [ 0 . 5 , 0 . 6 , 0 . 7 , 0 . 8 , 0 . 9 , 0 . 9 5 , 1], 7\n\" gamma \" : uniform ( 0 , 1 ) , 8\n\" reg_alpha \" : uniform ( 0 , 1 ) , 9\n\" reg_lambda \" : uniform ( 0 , 1 ) 10 }"},{"figure_ref":[],"heading":"C.3 Other Implementation Details","publication_ref":[],"table_ref":[],"text":"Output Collection For multiple-choice QA tasks (MMLU and MCQA), we collect generated choice labels (e.g., \"(A)\") from the first 5 tokens generated. For closed-book QA tasks (CBQA), we collect the first 16 newly generated tokens as the model output and truncate the outputs by the newline token.\nTemperature Scaling We search for the optimal temperature τ based on the meta-training set. The search grid is: np.linspace(1.0, 3.0, 100) D Few-shot setting ablation results We take the same prompt template for MCQA as MMLU. "}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b0\"\n ],\n \"table_ref\": [],\n \"text\": \"In-context learning (ICL) with large language models (LLMs) has shown great potential in performing a wide range of language tasks (Brown et al., 2020). ICL has the unique advantages of being data-efficient (i.e., only a few labeled training examples are needed) and accessible (i.e., expertise in training models is no longer required). With these advantages, a non-expert user can create a system to perform a new task within minutes by writing a few examples. This gives rise to the popularity of ating ICL with a labeled test set is a direct solution to know whether ICL will be effective, it greatly reduces the appeal of ICL, as one of ICL's key selling points is that it does not require a large labeled dataset. In addition, many tasks do not come with a labeled test set due to high annotation costs (e.g., medical/law-related questions that require professional knowledge to answer). In such cases, it is highly desirable to estimate the ICL performance without a labeled test set. This would help system developers determine whether ICL is likely to be useful for their problems of interest.\\nGuided by this motivation, we formalize the problem of few-shot ICL accuracy estimation: given a handful of labeled in-context examples and a set of unlabeled test examples, our goal is to estimate the overall accuracy of ICL on these test examples. Our contributions are twofold:\\n• We propose to address the accuracy estimation problem by training a \\\"meta-model,\\\" which takes in LLM confidence features as input and outputs the task accuracy. The meta-model is trained with observed ICL accuracies on seen datasets, and then used to estimate ICL accuracy on unseen datasets (see Figure 1).\\n• We obtain 42,360 observations of LLM ICL performance, by conducting extensive ICL experiments spanning two tasks (multiplechoice QA and closed-book QA), 91 datasets, and 4 LLMs. We then benchmark the metamodel method and multiple baselines on a total of 12 evaluation settings derived from these observations.\\nOur meta-model can estimate ICL accuracies without the need for labeled test examples. In 10 out of 12 settings, the meta-model estimates are at least as accurate as directly evaluating on 16 labeled examples. In 2 out of 12 settings, they match with evaluating on 128 labeled examples. On average, we are able to save the annotation cost of 40 test labels per task by using the meta-model. Further, the meta-model outperforms all baseline methods in 8 out of 12 settings, improving the relative estimation error by 23.6% However, we also find that there exists substantial room for improvement across all settings. We envision estimating ICL accuracy without labeled test data as an open challenge and encourage the community to develop new techniques that can more accurately predict when ICL will be effective.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Model Confidence and Calibration\",\n \"publication_ref\": [\n \"b5\",\n \"b13\",\n \"b14\",\n \"b9\",\n \"b5\",\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"Calibration of LLMs has been studied on a diverse range of tasks such as classification (Desai and Durrett, 2020) and question answering (QA) (Jiang et al., 2021;Kadavath et al., 2022). It aims to study whether LLMs assign meaningful correctness likelihood-also known as model confidence-to the outputs (Guo et al., 2017). Most prior work evaluates calibration at the example level (Desai and Durrett, 2020;Kamath et al., 2020); in this paper, we focus on using overall model confidence distributions to estimate dataset-level accuracies. We propose a method to learn model calibration patterns based on observations of LLMs' performance at the dataset level.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"In-context Learning\",\n \"publication_ref\": [\n \"b0\",\n \"b4\",\n \"b19\",\n \"b31\",\n \"b21\"\n ],\n \"table_ref\": [],\n \"text\": \"LLMs pre-trained with auto-regressive language modeling objectives have been shown to be capable of \\\"learning\\\" in context when given a prompt composed of a prompt template and a few labeled demonstrations (Brown et al., 2020;Chowdhery et al., 2022). While LLMs can learn a new task only through model inference, the accuracy is sensitive to the choices of prompt templates and incontext examples (Lu et al., 2021;Zhao et al., 2021;Perez et al., 2021). Therefore, we aim to develop a method to accurately estimate ICL performance for a dataset prompted with any prompt template and combination of in-context examples.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Out-of-distribution (OOD) Prediction\",\n \"publication_ref\": [\n \"b8\",\n \"b6\",\n \"b28\",\n \"b24\",\n \"b18\",\n \"b6\",\n \"b24\"\n ],\n \"table_ref\": [],\n \"text\": \"Machine learning models in the real world commonly encounter distribution shifts between training and test time. Prior work (Guillory et al., 2021;Garg et al., 2022;Yu et al., 2022;Singhal et al., 2022;Li et al., 2022) aims to predict models' OOD performance under different setups. Garg et al. (2022) predict target domain accuracy for image classification tasks with distribution by fitting a threshold on model confidence using only labeled source data and unlabeled target data. Singhal et al. (2022) use a few additional target-domain examples to predict the accuracy, focusing on known source-target dataset pairs on which models often have low OOD accuracy due to overfitting to spurious correlations (e.g., MNLI-HANS and QQP-PAWS). They find that accuracy on the given small set of target examples is a strong baseline to approximate accuracy on the full-test set. We include the accuracy for a small set of labeled test examples as an oracle baseline (see Section 3.3). These papers all try to predict the OOD accuracy of a model trained on in-distribution training data; in contrast, in our setting we have access to some labeled datasets but the language models we study were never finetuned on those datasets. In order to avoid confusion, we instead use the terms \\\"seen/unseen tasks\\\" to describe the datasets available to us, rather than \\\"in-distribution/out-of-distribution.\\\"\\n3 Accuracy Prediction\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Problem Definition\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We formalize the task of ICL accuracy estimation for unseen datasets given observations of the same model's performance on other datasets. A method for the ICL accuracy estimation task takes in four inputs: a language model M ; a set of labeled seen datasets {D i } r i=1 , where each D i consists of a set of labeled examples {(x\\n(1) i , y(1)\\ni ), . . . , (x\\n(n i ) i , y (n i ) i\\n)} and n i = |D i |; a prompt c for the test task; and an unlabeled test dataset\\nD test = {x (1) test , . . . , x(m)\\ntest } of size m. In a typical setting, each seen task should consist of a sufficient amount of labeled examples, i.e., n i ≥ 100. The method should output the estimated accuracy of M on D test when prompted with prompt c; we denote the actual accuracy of the model as acc M,c test and acc M,c test as the predicted accuracy. Note that with the labeled datasets D i and a corresponding prompt c, we can compute the corresponding dataset-level ICL accuracy acc M,c i for i = 1, . . . , r.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Prompt Formulation and Data Splits\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We construct prompts by sampling k in-context examples uniformly at random from available labeled data and formatting them with prompt templates to form a prompt (see Section B and \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Comparing with Labeled Test Data\",\n \"publication_ref\": [\n \"b24\"\n ],\n \"table_ref\": [],\n \"text\": \"To put our results in context, we compare all methods to the Oracle approach of sampling l labeled examples from the test dataset D test and measuring accuracy on those l examples, which we call oracle l . This approach is used by Singhal et al. (2022) and it represents how well we can evaluate ICL performance for D test by collecting labeled examples. With a large value of l, we get a better evaluation of the test dataset at the cost of collecting expensive annotations. In proposing the task of accuracy prediction, we hope to develop methods that outperform the l-labeled oracle for values of l that represent non-trivial annotation costs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Confidence Profile Meta-Model\",\n \"publication_ref\": [\n \"b0\",\n \"b30\"\n ],\n \"table_ref\": [],\n \"text\": \"We propose a new method that trains a meta-model based on the confidence profiles of seen datasets {D i } r i=1 to estimate ICL performance. We use the term confidence profile to denote the distribution of model confidence scores on each example in the dataset. We extract the confidence profiles (see Figure 1) from each seen dataset and convert them to a feature vector. We then train a meta-model to map confidence feature vectors to the datasetlevel ICL accuracies. The benefits of using the confidence feature vector are twofold. First, we do not need any labeled test data, which saves annotation costs. Second, this approach is applicable to any pre-trained language model like GPT3 (Brown et al., 2020) and OPT (Zhang et al., 2022).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Confidence Profile\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In general, given a (not-necessarily labeled) dataset D, LM M , and a prompt c, we obtain the confidence profile by first computing the confidence score s M,c (x) for each x ∈ D. The score for each input x can be computed by one forward pass of M ; the exact value of the score differs based on the task, as described below. Next, we sort the scores to obtain a list [s 1 , . . . , s |D| ] where each s i ≤ s i+1 . Then we create a d conf -dimensional feature vector conf M,c D , whose i-th component is a linear interpolation between s ⌊|D|×i/d⌋ and s ⌈|D|×i/d⌉ . Intuitively, the i-th feature represents the i/|D|-th percentile confidence score. We refer to the feature vectors derived from confidence profiles as confidence vectors.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Confidence scores\",\n \"publication_ref\": [\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"The confidence score s M,c (x) is calculated differently for closed-set generation and open-ended generation.\\nClosed-set generation. Closed-set generation tasks have a pre-defined label space Y. We take outputs from LLMs and identify the answers only by labels (Kadavath et al., 2022). For each example, we take model confidence as the normalized probability across the label space:\\ns M,c (x) = p ŷ ỹ∈Y p ỹ (1)\\nwhere p ỹ is the model-assigned probability for label ỹ on input x, p ŷ is the probability for the output label ŷ from model M , and ŷ = arg max ỹ∈Y p ỹ.\\nOpen-ended generation. We refer to tasks that require sequence generation (e.g., closed-book QA, summarization, machine reading comprehension, etc.) as open-ended generation tasks. We use negative log-likelihood (NLL) 2 to obtain confidence scores from each generated sequence. Let ŷ be the model-generated sequence. We compute the confidence score as:\\ns M,c (x) = - |ŷ| t=1 log p t (ŷ t ). (2\\n)\\np t is the model-assigned probability distribution at output token t and ŷt is the t-th output token\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Meta-Model Training Data\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"For each seen dataset {D i } r i=1 , we sample K prompts {c ij } K j=1 . Then for each prompt sampled we compute the confidence vector conf 2 We also tried perplexity but found NLL to yield better results.\\nM,c ij i and accuracy acc M,c ij i to create one meta-training ex- ample (conf M,c ij i , acc M,c ij i ). This creates a total of r × K meta-\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Meta-Model Architectures\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We choose meta-models that are easy to train and contain far fewer parameters than LLMs for computational efficiency. In this paper, we consider three meta-model architectures. First, we use k Nearest Neighbors regression (k-NN), which measures feature similarity. In the context of this paper, k-NN retrieves the most similar confidence profile from the seen datasets to the test dataset confidence profile and predicts based on the observed ICL accuracy on the retrieved in-distribution datasets. We use the implementation in scikit-learn library. 3Second, we use a two-layer Multilayer Perceptron (MLP) that takes confidence feature vectors as input. Third, we use the tree-based method XG-Boost (Chen and Guestrin, 2016) with the same confidence features. We use XGBoostRegressor implemented in the XGBoost library4 and tune the hyperparameters as described in Appendix C.2.5 \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Evaluation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"For task performance evaluation, we use Exact Match (EM) accuracy to measure accuracy for closed-set QA, and F1-score to measure accuracy for open-ended QA.\\nWe evaluate accuracy prediction models based on absolute error, defined as |acc M,ctest test acc M,ctest test |, where both are computed using the test dataset D test with a test prompt c test . We then average the absolute error over all prompts C test and compute the dataset-specific mean average error:\\nerr Dtest = 1 |C test | c∈Ctest |acc M,c test -acc M,c test |.\\nFinally, to evaluate the overall success of accuracy prediction across a collection of test datasets T , we measure mean absolute error (MAE), defined as:\\nerr T = 1 |T | Dtest∈T err Dtest\\n(3)\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Baselines\",\n \"publication_ref\": [\n \"b20\",\n \"b9\",\n \"b16\",\n \"b5\",\n \"b12\",\n \"b9\",\n \"b23\",\n \"b6\"\n ],\n \"table_ref\": [],\n \"text\": \"We consider four baselines for accuracy estimation.\\nAverage training accuracy (AVGTRAIN). We simply take the average dataset-level accuracy of the seen datasets as our accuracy estimation:\\nacc AVGTRAIN = 1 r × K r i=1 K j=1 acc M,c ij i .\\nAverage Calibration Error (AVGCONF). We take the average confidence across the test dataset as the accuracy estimation:\\nacc AVGCONF = 1 |C test | × m c∈Ctest x∈Dtest s M,c (x).\\nNote that this baseline is only applicable to closedset generation and not open-ended generation tasks since the accuracy metric for open-ended generation (F1 score) and confidence metric (NLL) do not share the same range ([0, 1] vs. (-∞, 0]). The intuition behind AVGCONF is that if the model confidence scores are well-calibrated (at the example level), then the expected value of the model's confidence scores should be equal to the accuracy. In fact, we note that the MAE of AVGCONF is similar to Expected Calibration Error (ECE), which measures the example-level calibration error (Naeini et al., 2015;Guo et al., 2017;Kumar et al., 2019;Desai and Durrett, 2020). 6Temperature Scaling (TS) Temperature scaling is a widely used calibration method (Hinton et al., 2015;Guo et al., 2017;Si et al., 2022). By fitting a single scaler parameter called temperature τ , it produces softer model-assigned probabilities\\np ŷ = exp (z ŷ/τ ) ỹ exp (z ỹ∈Y /τ )\\n.\\nWe then obtain scaled confidence scores with Equation 1, and evaluate AVGCONF on the test dataset. Note that we optimize temperature τ based on the AVGCONF of the training datasets instead of the common approach of using NLL as an objective function.\\nAverage Threshold Confidence (ATC) We use ATC (Garg et al., 2022) as one of our OOD accuracy estimation baselines. ATC takes accuracy estimation by fitting a confidence threshold on a single source dataset and generalizes to the target dataset. We take the estimated accuracy for the test dataset to be the average of the ATC estimates from each seen dataset:\\nacc AT C = 1 r r i=1 atc M,c i,test ,\\nwhere atc M,c i,test is the D i to D test ATC estimate.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Alternative Featurizations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In addition to the confidence profiles, we experiment with another featurization method that uses model embeddings from the LM M . Given a dataset D, LM M , and prompt c, we obtain the model embedding by first taking the last-layer, lasttoken embedding e M,c (x) for each x ∈ D, and then averaging across the dataset:\\nembed M,c D = 1 |D| x∈D e M,c (x).\\nSince embed\\nM,c D\\nis very high-dimensional (e.g., 5120 dimensional for 13B models), we use Principle Component Analysis (PCA) to reduce its dimensionality. We fit the PCA model on all dataset embedding vectors embed M,c D and transform them into d e -dimensional vectors embed M,c D , which we can use as a feature vector. As an additional experiment, we can concatenate the confidence vector and embedding vector to form a combined feature vector:\\nce M,c D = conf M,c D + embed M,c D .\\nReducing the dimensionality makes the comparison with confidence features more fair, and does not dilute the influence of confidence features when concatenating them with embedding features.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Accuracy Estimation Benchmark\",\n \"publication_ref\": [\n \"b11\"\n ],\n \"table_ref\": [],\n \"text\": \"We benchmark both our meta-model method for ICL accuracy estimation and the baseline methods mentioned in Section 4.6 on a total of 12 LLMdataset collection pairs (3 dataset collections × 4 LLMs). For each evaluation setting, we evaluate 3 different featurization methods mentioned in Section 4.7. This adds up to 36 experiment settings.\\nDatasets. We use three different collections of datasets in total: multiple-choice QA (MCQA) from MMLU (Hendrycks et al., 2020) and both\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"LLaMA-7B\",\n \"publication_ref\": [\n \"b17\",\n \"b30\",\n \"b26\",\n \"b26\",\n \"b22\"\n ],\n \"table_ref\": [],\n \"text\": \"LLaMA-13B OPT-6.7B OPT-13B 5 for the full list of tasks). We henceforth use MCQA and CBQA to refer to the CrossFit dataset collections respectively. We use the implementations and training/test partitions from HuggingFace Datasets (Lhoest et al., 2021). We split each collection of datasets into meta-training/test splits using 5-fold cross-validation-we partition each dataset collection into 5 equal-sized subsets and run five versions of each experiment, one with each subset being used as the meta-test set and the remaining subsets used as meta-training data. We take the average of the meta-test results as our final result.\\nLLMs. We run our experiments on four LLMs: OPT-6.7B, OPT-13B (Zhang et al., 2022), LLaMA-7B, and LLaMA-13B (Touvron et al., 2023). We use the OPT models from HuggingFace Models7 and the LLaMA model from Meta AI.8 More details are included in Appendix C.1.\\nExperimental Details. We generate prompts for each dataset using the method noted in Section 3.2. For each dataset in MMLU,9 we combine the \\\"validation\\\" set and \\\"dev\\\"10 set to be the training set 11 We choose up to 5-shot setting because it is studied in previous studies (Touvron et al., 2023;Rae et al., 2021). prompts for MCQA/CBQA datasets) for MMLU because it contains a very large number of datasets.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Main Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\",\n \"tab_4\"\n ],\n \"text\": \"The meta-model outperforms all baselines under certain evaluation settings. Table 1 shows the meta-model estimation error for each evaluation setting. For 8 out of 12 settings (all CBQA settings, LLaMA-7B on MCQA, LLaMA-13B on MMLU, both OPT models on MCQA), the best meta-model architecture has 23.67% lower relative MAE than the best baseline method on average. In the best case (OPT-6.7B on MCQA), the meta-model can achieve 43.5% lower relative MAE than all baselines. However, for the other 4 settings (both OPT models on MMLU, LLaMA-7B on MMLU, and LLaMA-13B on MCQA), baseline methods provide more accurate estimates of ICL accuracy. Fig- ure 2 shows the evaluation results graphically. On average across all 12 settings, the best estimation errors from the meta-models are 32.5% less than the actual accuracy standard deviations. In 11 out of 12 settings, the estimation errors are within one standard deviation of the actual accuracy.\\nOracle baselines indicate useful accuracy estimations. In comparison to the Oracle baselines, the meta-model outperforms the oracle 32 baseline in all MMLU and MCQA settings except for LLaMA-13B on MCQA (achieves oracle 8 ) and outperforms the oracle 16 baseline in all CBQA settings except for LLaMA-13B (achieves oracle 8 ).\\nIn the two best-case settings (using XGBoost as the meta-model on MCQA with either OPT model), the meta-model achieves the oracle 128 baseline, i.e., is equivalent to estimating the accuracy using 128 annotations.\\nBaseline methods are effective in some settings.\\nWhile ATC is a weak baseline for ICL accuracy estimation, AVGTRAIN, AVGCONF, and TS are strong baselines for MMLU and MCQA. AVG-TRAIN is able to achieve oracle 32 for 3 out of 12 settings (LLaMA-7B, OPT-6. for 5 settings (LLaMA-13B on MMLU, both OPT models on MMLU, and OPT-6.7B on MCQA).12 \\nAblation on Model Architecture Across three meta-model structures, the XGBoost meta-model overall provides the most accurate estimation as it has the lowest MAE for 7 out of 12 evaluation settings. The average MAE is 5.88 for XG-Boost meta-models, 5.94 for 3-NN meta-models, and 7.18 for MLP meta-models. Surprisingly, 3-NN meta-models have a lower average MAE than MLP meta-models despite having a simpler model structure. In Figure 3, we show that the XGBoost meta-model provides well-correlated accuracy estimation across 4 different evaluation settings.\\nAblation on Featurization Methods We consider three featurization methods as described in Section 4.7. Table 2 in the appendix shows that the best overall accuracy estimation for all settings is attained by using the confidence vectors as metafeatures (achieves the lowest MAE for 26 out of 36 evaluation settings). The average MAE is 6.27 for conf , 8.34 for embed, and 7.36 for ce. Further, using conf as features demonstrates a more dominant advantage for all CBQA tasks, achieving the lowest MAE for 11 out of 12 evaluation settings.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\",\n \"fig_2\"\n ],\n \"heading\": \"Effect of Unlabeled Data and Confidence Vector Dimensions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We now study confidence feature vector ablations by varying the number of unlabeled test examples m in each unseen dataset and the dimension of the confidence vector d conf . We test with OPT-13B on MCQA datasets using the XGBoost meta-model since we achieve the lowest MAE in this setting.\\nFigure 4 shows that increasing m enables better accuracy estimation, reducing the average MAE (across all d conf ) from 3.92 for m = 200 to 2.56 for m = 1000. Note that increasing m requires performing additional LLM inferences on unlabeled examples, so leveraging unlabeled test data is constrained by computational cost considerations. The quality of our accuracy estimates does not vary much as we change the confidence vector dimension d conf , as shown in Figure 4.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Effect of Number of Shots\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We compare ICL accuracy estimation performance given different k-shot ICL accuracy observations for LLaMA-13B on MMLU datasets. Table 4 in the Appendix shows that the meta-model produces a slightly better ICL accuracy estimation for the 3-shot setting. Overall, the meta-model gives consistent accuracy estimates across different k-shot settings as they all achieve oracle 32 .\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Prompt Selection\",\n \"publication_ref\": [\n \"b31\",\n \"b21\",\n \"b23\"\n ],\n \"table_ref\": [],\n \"text\": \"Previous works demonstrated that ICL performance is highly sensitive to the prompt templates as well as in-context examples (Zhao et al., 2021;Perez et al., 2021;Chen et al., 2022); we are thus interested in whether our ICL accuracy estimation method can be applied to select the best ICL prompt c ∈ C test for the test dataset. For each dataset, we use the XGBoost meta-model to select the best prompt c * , as opposed to the actual best prompt c * . We then compute the corresponding ICL accuracies and compare them to the average accuracy across all test prompts. Figure 5 shows that there is a significant difference in ICL accuracy given different prompts for all 12 settings, and the selected prompts lead to better ICL accuracies than the average accuracy for 7 out of 12 settings. On average, the selected prompt is 15.6% as effective as the actual best prompt. The limited improvement from the random baseline indicates there's a large room for improvement and we encourage future work to derive a better prompt selection standard.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Discussion and Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we study the problem of few-shot ICL accuracy estimation. We propose training a meta-model based on LLM confidence features and observed accuracies on seen datasets. We show that without using any labeled test data, the meta-model is often able to attain accurate estimates of ICL accuracy, which is practically useful for predicting LLMs' accuracy on datasets that have high annotation costs. We also construct a large-scale benchmark for dataset-level ICL accuracy estimation by evaluating the meta-model and multiple baseline methods across 12 evaluation settings and 3 metafeature options. We observe that while some baseline methods can provide good accuracy estimates, our meta-model demonstrates non-trivial improvement in estimation abilities over baseline methods in 8 out of 12 evalutation settings. We encourage future work to develop better meta-model architectures as well as better metafeatures and study potential implications for the meta-model, such as acting as a prompt template/ICL example selection method. We believe that our benchmark can serve as an open challenge for improving dataset-level ICL accuracy estimations, leading to an improved understanding of when ICL is likely to be effective.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [\n \"b7\"\n ],\n \"table_ref\": [],\n \"text\": \"While we conducted extensive experiments to study ICL accuracy estimations, there are many more LLMs that have exhibited impressive capabilities on a variety of tasks. Due to computational constraints, we do not benchmark accuracy estimations based on LLMs with limited access (e.g., GPT-4 (OpenAI, 2023)) as it is difficult to extract model embedding features, or those larger than 13B. We also don't consider instruction-tuned models to avoid possible overlaps between their training datasets and our evaluation datasets. Meanwhile, instruction tuning sometimes hurts model performance on canonical datasets such as MMLU, as shown in Gudibande et al. (2023). It might also significantly hurt calibration as reported in OpenAI (2023). For the same reasons, we include only a limited number of prompt templates and in-context example variations for ICL prompting. While we choose only 3 few-shot settings for MMLU and 2 for MCQA and CBQA, it is possible to achieve better accuracy estimations with more observations in the training data.\\nIn terms of dataset selection, we use 13 closedbook QA tasks for the open-ended generation setting. Our findings might not generalize to other open-ended generation tasks such as summarization or long-form question answering. Overall, the meta-model provides effective accuracy estimations, but there's still substantial room for improvement. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgements\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We would like to thank Ting-Yun Chang for her valuable contributions. We thank Ameya Godbole, Johnny Wei, and Wang Zhu for their valuable discussions. We also thank all members of the Allegro lab at USC for their support and valuable feedback. RJ was supported by an Open Philanthropy research grant, a Cisco Research Award, and HF was supported by the USC Provost Fellowship Award.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Dataset Implementation Details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Following Section 5.1's discussion of specific dataset implementations for MCQA and CBQA, our approach for each setting is: for tasks that have a defined train/test split in HuggingFace Datasets \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B Prompt Templates\",\n \"publication_ref\": [\n \"b22\",\n \"b11\"\n ],\n \"table_ref\": [],\n \"text\": \"We collect 5 general prompt templates for MMLU datasets: 1 Null template, 1 self-constructed prompt template, 2 from previous work (Rae et al., 2021;Hendrycks et al., 2020), and 1 generated by ChatGPT. For MCQA and CBQA, we only use the Null template due to resource considerations. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C Implementation Details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We will release code to reproduce our results upon publication.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.1 LLMs Implementations\",\n \"publication_ref\": [\n \"b26\",\n \"b30\"\n ],\n \"table_ref\": [],\n \"text\": \"We use LLaMA-7B and LLaMA-13B (Touvron et al., 2023) from Meta AI, and OPT-6.7B and OPT-13B (Zhang et al., 2022) from HuggingFace transformers. For LLaMA-7B, OPT-6.7B, and OPT-13B, we run evaluations using a single RTXA6000 GPU (48GB). For LLaMA-13B, we run evaluations by parallel inference on two RTXA6000 GPUs. We use half-precision for both OPT models. Note that some LLMs (except for OPT-13B) can be run on GPUs with a smaller memory. We evaluate 42,360 ICL observations in total, where each observation is a dataset with 100 to 1000 examples. The total inference process takes around 2000 GPU hours.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.2 Meta-model Implementations\",\n \"publication_ref\": [\n \"b2\"\n ],\n \"table_ref\": [],\n \"text\": \"All meta-model architectures can be trained on an i7-10700 CPU. The total training time of the metamodel on one experiment setting varies from 1.5 hours to 24 hours depending on training data dimensions. We include the implementation details for each of the meta-model architectures. We use the random seed 1 for all processes involving randomness. For K Nearest Neighbors regression, we use the implementation of KNeighborsRegressor from sklearn library. We use euclidean distance as the weight metric, and fit the model on the metatraining data.\\nFor MLP, we implement with Pytorch. We use a 2-layer MLP with the size of the hidden_state = 1536, learning_rate = 1e -5, and dropout_rate = 0.2. We use Adam Optimizer and MSELoss, and we perform early stopping with the validation data. The validation data is a 20% random partition of the meta-training data. For early stopping, the max epoch is 50 and the patience is 7.\\nFor XGBoost, we implement with the XGBRegressor from the XGBoost library (Chen and Guestrin, 2016). We use a 5-fold random search cross-validation for 300 iterations to choose the hyperparameter. The candidates are: 1 { 2 \\\" lr \\\" : uniform ( 0 . 0 1 , 0 . 5 ) , 3\\n\\\" max_depth \\\" : randint ( 3 , 1 0 ) , 4\\n\\\" n_estimators \\\" : randint ( 1 0 0 , 1 0 0 0 ) , 5\\n\\\" colsample_bytree \\\" : [ 0 . 5 , 0 . 6 , 0 . 7 , 0 . 8 , 0 . 9 , 0. 9 5 , 1 ] , 6\\n\\\" subsample \\\" : [ 0 . 5 , 0 . 6 , 0 . 7 , 0 . 8 , 0 . 9 , 0 . 9 5 , 1], 7\\n\\\" gamma \\\" : uniform ( 0 , 1 ) , 8\\n\\\" reg_alpha \\\" : uniform ( 0 , 1 ) , 9\\n\\\" reg_lambda \\\" : uniform ( 0 , 1 ) 10 }\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.3 Other Implementation Details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Output Collection For multiple-choice QA tasks (MMLU and MCQA), we collect generated choice labels (e.g., \\\"(A)\\\") from the first 5 tokens generated. For closed-book QA tasks (CBQA), we collect the first 16 newly generated tokens as the model output and truncate the outputs by the newline token.\\nTemperature Scaling We search for the optimal temperature τ based on the meta-training set. The search grid is: np.linspace(1.0, 3.0, 100) D Few-shot setting ablation results We take the same prompt template for MCQA as MMLU. \"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-10-26"},"doi":{"kind":"string","value":""},"references":{"kind":"list like","value":[{"authors":"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel Ziegler; Jeffrey Wu; Clemens Winter; Chris Hesse; Mark Chen; Eric Sigler; Mateusz Litwin; Scott Gray; Benjamin Chess; Jack Clark; Christopher Berner; Sam Mccandlish; Alec Radford; Ilya Sutskever; Dario Amodei","journal":"","ref_id":"b0","title":"Language models are few-shot learners","year":"2020"},{"authors":"Curran Associates; Inc ","journal":"","ref_id":"b1","title":"","year":""},{"authors":"Tianqi Chen; Carlos Guestrin","journal":"","ref_id":"b2","title":"Xgboost: A scalable tree boosting system","year":"2016"},{"authors":"Yanda Chen; Chen Zhao; Zhou Yu; Kathleen Mckeown; He He","journal":"","ref_id":"b3","title":"On the relation between sensitivity and accuracy in in-context learning","year":"2022"},{"authors":"Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann","journal":"","ref_id":"b4","title":"Palm: Scaling language modeling with pathways","year":"2022"},{"authors":"Shrey Desai; Greg Durrett","journal":"","ref_id":"b5","title":"Calibration of pre-trained transformers","year":"2020"},{"authors":"Saurabh Garg; Sivaraman Balakrishnan; Zachary C Lipton; Behnam Neyshabur; Hanie Sedghi","journal":"","ref_id":"b6","title":"Leveraging unlabeled data to predict out-of-distribution performance","year":"2022"},{"authors":"Arnav Gudibande; Eric Wallace; Charlie Snell; Xinyang Geng; Hao Liu; Pieter Abbeel; Sergey Levine; Dawn Song","journal":"","ref_id":"b7","title":"The false promise of imitating proprietary llms","year":"2023"},{"authors":"Devin Guillory; Vaishaal Shankar; Sayna Ebrahimi; Trevor Darrell; Ludwig Schmidt","journal":"","ref_id":"b8","title":"Predicting with confidence on unseen distributions","year":"2021"},{"authors":"Chuan Guo; Geoff Pleiss; Yu Sun; Kilian Q Weinberger","journal":"","ref_id":"b9","title":"On calibration of modern neural networks","year":"2017"},{"authors":" Pmlr","journal":"","ref_id":"b10","title":"","year":""},{"authors":"Dan Hendrycks; Collin Burns; Steven Basart; Andy Zou; Mantas Mazeika; Dawn Song; Jacob Steinhardt","journal":"","ref_id":"b11","title":"Measuring massive multitask language understanding","year":"2020"},{"authors":"Geoffrey Hinton; Oriol Vinyals; Jeff Dean","journal":"","ref_id":"b12","title":"Distilling the knowledge in a neural network","year":"2015"},{"authors":"Zhengbao Jiang; Jun Araki; Haibo Ding; Graham Neubig","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b13","title":"How can we know when language models know? on the calibration of language models for question answering","year":"2021"},{"authors":"Saurav Kadavath; Tom Conerly; Amanda Askell; Tom Henighan; Dawn Drain; Ethan Perez; Nicholas Schiefer; Zac Hatfield Dodds; Nova Dassarma; Eli Tran-Johnson","journal":"","ref_id":"b14","title":"Language models (mostly) know what they know","year":"2022"},{"authors":"Amita Kamath; Robin Jia; Percy Liang","journal":"","ref_id":"b15","title":"Selective question answering under domain shift","year":"2020"},{"authors":"Ananya Kumar; Percy S Liang; Tengyu Ma","journal":"Advances in Neural Information Processing Systems","ref_id":"b16","title":"Verified uncertainty calibration","year":"2019"},{"authors":"Quentin Lhoest; Albert Villanova Del Moral; Yacine Jernite; Abhishek Thakur; Suraj Patrick Von Platen; Julien Patil; Mariama Chaumond; Julien Drame; Lewis Plu; Tunstall","journal":"","ref_id":"b17","title":"Datasets: A community library for natural language processing","year":"2021"},{"authors":"Zeju Li; Konstantinos Kamnitsas; Mobarakol Islam; Chen Chen; Ben Glocker","journal":"Springer","ref_id":"b18","title":"Estimating model performance under domain shifts with classspecific confidence scores","year":"2022-09-18"},{"authors":"Yao Lu; Max Bartolo; Alastair Moore; Sebastian Riedel; Pontus Stenetorp","journal":"","ref_id":"b19","title":"Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity","year":"2021"},{"authors":"Gregory F Mahdi Pakdaman Naeini; Hauskrecht Cooper","journal":"OpenAI","ref_id":"b20","title":"Obtaining well calibrated probabilities using bayesian binning","year":"2015"},{"authors":"Ethan Perez; Douwe Kiela; Kyunghyun Cho","journal":"Advances in neural information processing systems","ref_id":"b21","title":"True few-shot learning with language models","year":"2021"},{"authors":"Sebastian Jack W Rae; Trevor Borgeaud; Katie Cai; Jordan Millican; Francis Hoffmann; John Song; Sarah Aslanides; Roman Henderson; Susannah Ring; Young","journal":"","ref_id":"b22","title":"Scaling language models: Methods, analysis & insights from training gopher","year":"2021"},{"authors":"Chenglei Si; Chen Zhao; Sewon Min; Jordan Boyd-Graber","journal":"","ref_id":"b23","title":"Re-examining calibration: The case of question answering","year":"2022"},{"authors":"Prasann Singhal; Jarad Forristal; Xi Ye; Greg Durrett","journal":"","ref_id":"b24","title":"Assessing out-of-domain language model performance from few examples","year":"2022"},{"authors":"Aarohi Srivastava; Abhinav Rastogi; Abhishek Rao; Abu Awal; Md Shoeb; Abubakar Abid; Adam Fisch; Adam Adam R Brown; Aditya Santoro; Adrià Gupta; Garriga-Alonso","journal":"","ref_id":"b25","title":"Beyond the imitation game: Quantifying and extrapolating the capabilities of language models","year":"2022"},{"authors":"Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar","journal":"","ref_id":"b26","title":"Llama: Open and efficient foundation language models","year":"2023"},{"authors":"Qinyuan Ye; Bill Yuchen Lin; Xiang Ren","journal":"","ref_id":"b27","title":"Crossfit: A few-shot learning challenge for cross-task generalization in nlp","year":"2021"},{"authors":"Yaodong Yu; Zitong Yang; Alexander Wei; Yi Ma; Jacob Steinhardt","journal":"","ref_id":"b28","title":"Predicting out-ofdistribution error with the projection norm","year":"2022"},{"authors":" Pmlr","journal":"","ref_id":"b29","title":"","year":""},{"authors":"Susan Zhang; Stephen Roller; Naman Goyal; Mikel Artetxe; Moya Chen; Shuohui Chen; Christopher Dewan; Mona Diab; Xian Li; Xi Victoria Lin","journal":"","ref_id":"b30","title":"Opt: Open pre-trained transformer language models","year":"2022"},{"authors":"Zihao Zhao; Eric Wallace; Shi Feng; Dan Klein; Sameer Singh; ; Mmlu; Mcqa Cbqa","journal":"","ref_id":"b31","title":"Calibrate before use: Improving few-shot performance of language models","year":"2021"}],"string":"[\n {\n \"authors\": \"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel Ziegler; Jeffrey Wu; Clemens Winter; Chris Hesse; Mark Chen; Eric Sigler; Mateusz Litwin; Scott Gray; Benjamin Chess; Jack Clark; Christopher Berner; Sam Mccandlish; Alec Radford; Ilya Sutskever; Dario Amodei\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Curran Associates; Inc \",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Tianqi Chen; Carlos Guestrin\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Xgboost: A scalable tree boosting system\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Yanda Chen; Chen Zhao; Zhou Yu; Kathleen Mckeown; He He\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"On the relation between sensitivity and accuracy in in-context learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Palm: Scaling language modeling with pathways\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Shrey Desai; Greg Durrett\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Calibration of pre-trained transformers\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Saurabh Garg; Sivaraman Balakrishnan; Zachary C Lipton; Behnam Neyshabur; Hanie Sedghi\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Leveraging unlabeled data to predict out-of-distribution performance\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Arnav Gudibande; Eric Wallace; Charlie Snell; Xinyang Geng; Hao Liu; Pieter Abbeel; Sergey Levine; Dawn Song\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"The false promise of imitating proprietary llms\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Devin Guillory; Vaishaal Shankar; Sayna Ebrahimi; Trevor Darrell; Ludwig Schmidt\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Predicting with confidence on unseen distributions\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Chuan Guo; Geoff Pleiss; Yu Sun; Kilian Q Weinberger\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"On calibration of modern neural networks\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \" Pmlr\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Dan Hendrycks; Collin Burns; Steven Basart; Andy Zou; Mantas Mazeika; Dawn Song; Jacob Steinhardt\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Measuring massive multitask language understanding\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Geoffrey Hinton; Oriol Vinyals; Jeff Dean\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Distilling the knowledge in a neural network\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Zhengbao Jiang; Jun Araki; Haibo Ding; Graham Neubig\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b13\",\n \"title\": \"How can we know when language models know? on the calibration of language models for question answering\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Saurav Kadavath; Tom Conerly; Amanda Askell; Tom Henighan; Dawn Drain; Ethan Perez; Nicholas Schiefer; Zac Hatfield Dodds; Nova Dassarma; Eli Tran-Johnson\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Language models (mostly) know what they know\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Amita Kamath; Robin Jia; Percy Liang\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Selective question answering under domain shift\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ananya Kumar; Percy S Liang; Tengyu Ma\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b16\",\n \"title\": \"Verified uncertainty calibration\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Quentin Lhoest; Albert Villanova Del Moral; Yacine Jernite; Abhishek Thakur; Suraj Patrick Von Platen; Julien Patil; Mariama Chaumond; Julien Drame; Lewis Plu; Tunstall\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Datasets: A community library for natural language processing\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zeju Li; Konstantinos Kamnitsas; Mobarakol Islam; Chen Chen; Ben Glocker\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b18\",\n \"title\": \"Estimating model performance under domain shifts with classspecific confidence scores\",\n \"year\": \"2022-09-18\"\n },\n {\n \"authors\": \"Yao Lu; Max Bartolo; Alastair Moore; Sebastian Riedel; Pontus Stenetorp\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Gregory F Mahdi Pakdaman Naeini; Hauskrecht Cooper\",\n \"journal\": \"OpenAI\",\n \"ref_id\": \"b20\",\n \"title\": \"Obtaining well calibrated probabilities using bayesian binning\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Ethan Perez; Douwe Kiela; Kyunghyun Cho\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b21\",\n \"title\": \"True few-shot learning with language models\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Sebastian Jack W Rae; Trevor Borgeaud; Katie Cai; Jordan Millican; Francis Hoffmann; John Song; Sarah Aslanides; Roman Henderson; Susannah Ring; Young\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Scaling language models: Methods, analysis & insights from training gopher\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Chenglei Si; Chen Zhao; Sewon Min; Jordan Boyd-Graber\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Re-examining calibration: The case of question answering\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Prasann Singhal; Jarad Forristal; Xi Ye; Greg Durrett\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Assessing out-of-domain language model performance from few examples\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Aarohi Srivastava; Abhinav Rastogi; Abhishek Rao; Abu Awal; Md Shoeb; Abubakar Abid; Adam Fisch; Adam Adam R Brown; Aditya Santoro; Adrià Gupta; Garriga-Alonso\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Beyond the imitation game: Quantifying and extrapolating the capabilities of language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Llama: Open and efficient foundation language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Qinyuan Ye; Bill Yuchen Lin; Xiang Ren\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Crossfit: A few-shot learning challenge for cross-task generalization in nlp\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yaodong Yu; Zitong Yang; Alexander Wei; Yi Ma; Jacob Steinhardt\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Predicting out-ofdistribution error with the projection norm\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \" Pmlr\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Susan Zhang; Stephen Roller; Naman Goyal; Mikel Artetxe; Moya Chen; Shuohui Chen; Christopher Dewan; Mona Diab; Xian Li; Xi Victoria Lin\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Opt: Open pre-trained transformer language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Zihao Zhao; Eric Wallace; Shi Feng; Dan Klein; Sameer Singh; ; Mmlu; Mcqa Cbqa\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Calibrate before use: Improving few-shot performance of language models\",\n \"year\": \"2021\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,86.8,368.05,32.73,16],"formula_id":"formula_0","formula_text":"(1) i , y(1)"},{"formula_coordinates":[3,158.99,368.05,40.86,16],"formula_id":"formula_1","formula_text":"(n i ) i , y (n i ) i"},{"formula_coordinates":[3,105.45,396.69,112.26,15.63],"formula_id":"formula_2","formula_text":"D test = {x (1) test , . . . , x(m)"},{"formula_coordinates":[4,133.75,232.74,156.11,27.29],"formula_id":"formula_3","formula_text":"s M,c (x) = p ŷ ỹ∈Y p ỹ (1)"},{"formula_coordinates":[4,118.67,442.77,166.95,34.6],"formula_id":"formula_4","formula_text":"s M,c (x) = - |ŷ| t=1 log p t (ŷ t ). (2"},{"formula_coordinates":[4,285.63,455.77,4.24,9.46],"formula_id":"formula_5","formula_text":")"},{"formula_coordinates":[4,70.87,574.95,220.08,59.15],"formula_id":"formula_6","formula_text":"M,c ij i and accuracy acc M,c ij i to create one meta-training ex- ample (conf M,c ij i , acc M,c ij i ). This creates a total of r × K meta-"},{"formula_coordinates":[4,317.21,510.9,196.12,29.64],"formula_id":"formula_7","formula_text":"err Dtest = 1 |C test | c∈Ctest |acc M,c test -acc M,c test |."},{"formula_coordinates":[4,352.15,603.36,125.25,29.64],"formula_id":"formula_8","formula_text":"err T = 1 |T | Dtest∈T err Dtest"},{"formula_coordinates":[5,92.27,97.29,175.46,33.71],"formula_id":"formula_9","formula_text":"acc AVGTRAIN = 1 r × K r i=1 K j=1 acc M,c ij i ."},{"formula_coordinates":[5,70.87,192.73,221.1,29.64],"formula_id":"formula_10","formula_text":"acc AVGCONF = 1 |C test | × m c∈Ctest x∈Dtest s M,c (x)."},{"formula_coordinates":[5,125.43,508.74,104.91,27.29],"formula_id":"formula_11","formula_text":"p ŷ = exp (z ŷ/τ ) ỹ exp (z ỹ∈Y /τ )"},{"formula_coordinates":[5,358.43,107.33,113.69,33.71],"formula_id":"formula_12","formula_text":"acc AT C = 1 r r i=1 atc M,c i,test ,"},{"formula_coordinates":[5,346.6,301.69,137.36,29.64],"formula_id":"formula_13","formula_text":"embed M,c D = 1 |D| x∈D e M,c (x)."},{"formula_coordinates":[5,366.14,344.15,14.54,18.88],"formula_id":"formula_14","formula_text":"M,c D"},{"formula_coordinates":[5,346.7,499.74,137.15,15.73],"formula_id":"formula_15","formula_text":"ce M,c D = conf M,c D + embed M,c D ."}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 86.8,\n 368.05,\n 32.73,\n 16\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"(1) i , y(1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 158.99,\n 368.05,\n 40.86,\n 16\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"(n i ) i , y (n i ) i\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 105.45,\n 396.69,\n 112.26,\n 15.63\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"D test = {x (1) test , . . . , x(m)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 133.75,\n 232.74,\n 156.11,\n 27.29\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"s M,c (x) = p ŷ ỹ∈Y p ỹ (1)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 118.67,\n 442.77,\n 166.95,\n 34.6\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"s M,c (x) = - |ŷ| t=1 log p t (ŷ t ). (2\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 285.63,\n 455.77,\n 4.24,\n 9.46\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 70.87,\n 574.95,\n 220.08,\n 59.15\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"M,c ij i and accuracy acc M,c ij i to create one meta-training ex- ample (conf M,c ij i , acc M,c ij i ). This creates a total of r × K meta-\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 317.21,\n 510.9,\n 196.12,\n 29.64\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"err Dtest = 1 |C test | c∈Ctest |acc M,c test -acc M,c test |.\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 352.15,\n 603.36,\n 125.25,\n 29.64\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"err T = 1 |T | Dtest∈T err Dtest\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 92.27,\n 97.29,\n 175.46,\n 33.71\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"acc AVGTRAIN = 1 r × K r i=1 K j=1 acc M,c ij i .\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 70.87,\n 192.73,\n 221.1,\n 29.64\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"acc AVGCONF = 1 |C test | × m c∈Ctest x∈Dtest s M,c (x).\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 125.43,\n 508.74,\n 104.91,\n 27.29\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"p ŷ = exp (z ŷ/τ ) ỹ exp (z ỹ∈Y /τ )\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 358.43,\n 107.33,\n 113.69,\n 33.71\n ],\n \"formula_id\": \"formula_12\",\n \"formula_text\": \"acc AT C = 1 r r i=1 atc M,c i,test ,\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 346.6,\n 301.69,\n 137.36,\n 29.64\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"embed M,c D = 1 |D| x∈D e M,c (x).\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 366.14,\n 344.15,\n 14.54,\n 18.88\n ],\n \"formula_id\": \"formula_14\",\n \"formula_text\": \"M,c D\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 346.7,\n 499.74,\n 137.15,\n 15.73\n ],\n \"formula_id\": \"formula_15\",\n \"formula_text\": \"ce M,c D = conf M,c D + embed M,c D .\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"Estimating Large Language Model Capabilities without Labeled Test Data"},"abstract":{"kind":"string","value":"Large Language Models (LLMs) have the impressive ability to perform in-context learning (ICL) from only a few examples, but the success of ICL varies widely from task to task. Thus, it is important to quickly determine whether ICL is applicable to a new task, but directly evaluating ICL accuracy can be expensive in situations where test data is expensive to annotate-the exact situations where ICL is most appealing. In this paper, we propose the task of ICL accuracy estimation, in which we predict the accuracy of an LLM when doing in-context learning on a new task given only unlabeled test data for that task. To perform ICL accuracy estimation, we propose a method that trains a meta-model using LLM confidence scores as features. We compare our method to several strong accuracy estimation baselines on a new benchmark that covers 4 LLMs and 3 task collections. The meta-model improves over all baselines across 8 out of 12 settings and achieves the same estimation performance as directly evaluating on 40 collected labeled test examples per task. At the same time, no existing approach provides an accurate and reliable ICL accuracy estimation in every setting, highlighting the need for better ways to measure the uncertainty of LLM predictions."},"authors":{"kind":"string","value":"Harvey Yiyun Fu; Qinyuan Ye; Albert Xu; Xiang Ren; Robin Jia"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 2 :2Figure2: Bar graph of evaluation results (MAE) for all meta-models, baseline methods, and Oracle baselines of all 3 dataset collections with all 4 LLMs. We use the confidence vector as the meta-feature. Red/blue bars represent the meta-model/baseline evaluation results and the horizontal lines show the Oracle baselines.","figure_data":"","figure_id":"fig_0","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: We plot the meta-model predicted accuracy versus the actual accuracy across 4 settings. We use the XGBoost meta-model and the confidence vector meta-feature. Each entity represents an observation for one dataset. Red/blue represents higher/lower absolute error.","figure_data":"","figure_id":"fig_1","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Estimation results for ablating the number of unlabeled examples m (x-axis) and confidence vector dimensions d c (y-axis), evaluated on the OPT-13B on MCQA using the XGBoost meta-model.","figure_data":"","figure_id":"fig_2","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: Prompt selection results for all evaluation settings, measured by the absolute difference between ICL accuracy when prompted with c and the average accuracy. Blue bars show the actual best prompt (i.e., c = c * ), and red bars show the selected best prompt (i.e., c = c * )","figure_data":"","figure_id":"fig_3","figure_label":"5","figure_type":"figure"},{"figure_caption":"","figure_data":"","figure_id":"","figure_label":"","figure_type":"figure"},{"figure_caption":"use c to denote a prompt in general, C i to denote the set of training prompts for dataset D i , and C test to denote test prompts for dataset D test .","figure_data":"in","figure_id":"tab_0","figure_label":"6","figure_type":"table"},{"figure_caption":"training examples. The meta-model is trained on the meta-training examples and predicts the estimated accuracy acc M,ctest test based on the test dataset feature vector conf M,ctest Dtest for each test prompt c test ∈ C test . Note that since closedset/open-ended generations have different confidence scores and accuracy evaluation metrics, the meta-model does not train on datasets that have a different task formulation than the test datasets.","figure_data":"","figure_id":"tab_1","figure_label":"","figure_type":"table"},{"figure_caption":"","figure_data":"MethodsLLaMA-7BLLaMA-13BMMLUMCQACBQAMMLUMCQACBQAMeta ModelsMLP7.06 ± 1.038.82 ± 2.646.32 ± 2.135.80 ± 0.6311.04 ± 3.308.08 ± 2.993-NN5.98 ± 0.975.62 ± 2.357.10 ± 2.155.50 ± 0.6611.26 ± 4.568.56 ± 2.85XGBoost5.42 ± 4.525.22 ± 2.177.00 ± 2.445.00 ± 0.6211.52 ± 5.209.00 ± 4.85BaselinesAVGTRAIN5.26 ± 1.4010.26 ± 2.5011.62 ± 5.889.90 ± 1.7011.40 ± 3.3613.14 ± 7.25AVGCONF5.54 ± 0.546.88 ± 3.14n/a5.10 ± 0.868.58 ± 1.75n/aTS5.60 ± 1.636.32 ± 3.58n/a14.06 ± 1.5314.00 ± 6.95n/aATC20.34 ± 4.10 34.66 ± 9.36 31.80 ± 13.44 20.50 ± 4.9224.14 ± 7.72 31.80 ± 12.49Oracle5.82 (32)6.44 (32)7.06 (16)6.18 (32)13.44 (8)10.82 (8)ACC31.08 ± 6.32 39.00 ± 10.52 23.7 ± 10.44 45.50 ± 11.74 50.34 ± 12.84 29.30 ± 12.38MethodsOPT-6.7BOPT-13BMMLUMCQACBQAMMLUMCQACBQAMeta ModelsMLP5.70 ± 0.776.66 ± 1.257.28 ± 2.756.30 ± 1.147.18 ± 1.695.90 ± 1.183-NN4.06 ± 0.332.98 ± 0.576.46 ± 2.055.16 ± 0.242.78 ± 1.065.84 ± 2.19XGBoost3.76 ± 0.322.60 ± 0.738.06 ± 3.074.54 ± 0.352.54 ± 1.166.00 ± 2.43BaselinesAVGTRAIN3.48 ± 0.3710.66 ± 1.758.38 ± 3.864.28 ± 0.3511.46 ± 2.1518.00 ± 3.77AVGCONF8.42 ± 0.6013.42 ± 1.70n/a6.20 ± 0.777.14 ± 1.77n/aTS3.92 ± 0.174.60 ± 0.80n/a4.36 ± 0.312.72 ± 1.03n/aATC24.54 ± 2.12 37.64 ± 7.93 32.16 ± 14.60 24.72 ± 3.32 29.40 ± 10.28 30.34 ± 12.48Oracle5.58 (32)2.76 (128)6.50 (16)5.58 (32)2.76 (128)6.70 (16)ACC26.68 ± 4.42 33.16 ± 9.8017.54 ± 5.9226.82 ± 5.28 33.68 ±12.06 19.34 ± 6.3013),compared to OPT-6.7B (5.28), LLaMA-7B (6.50), LLaMA-13B(8.41). We report overall accuracy (ACC) by ExactMatch for MMLU and MCQA, and F1-score for CBQAthat we sample in-context examples from. We sam-ple 10 3-shot, 10 4-shot, and 10 5-shot prompts 11and decorate each of them with 5 prompt tem-plates chosen for MMLU (see Table 3). We choosed conf = 20 here since many of the datasets containonly 100 text examples. For MCQA and CBQA,we sample in-context examples from a pool of 100examples as the training set, and obtain a test setof 1000 examples (see Section A in the appendixfor implementation details). For each dataset, wesample 30 3-shot and 30 4-shot prompts and deco-rate them with only the null template. We choosed conf = 100 for both MCQA and CBQA settings.","figure_id":"tab_2","figure_label":"1","figure_type":"table"},{"figure_caption":"Full estimation results for all 4 LLMs, 3 dataset collections, and 3 meta-feature choices. XGBoost is the best overall meta-model structure with an average MAE of 6.60. The confidence vector is the best overall feature with an average MAE of 6.38 across all evaluation settings. The best meta-model outperforms all baseline methods in 9 out of 12 evaluation settings. We report overall accuracy (ACC) by Exact Match for MMLU and MCQA, and F1-score for CBQA","figure_data":"MMLU","figure_id":"tab_4","figure_label":"2","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 2 :2Figure2: Bar graph of evaluation results (MAE) for all meta-models, baseline methods, and Oracle baselines of all 3 dataset collections with all 4 LLMs. We use the confidence vector as the meta-feature. Red/blue bars represent the meta-model/baseline evaluation results and the horizontal lines show the Oracle baselines.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: We plot the meta-model predicted accuracy versus the actual accuracy across 4 settings. We use the XGBoost meta-model and the confidence vector meta-feature. Each entity represents an observation for one dataset. Red/blue represents higher/lower absolute error.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Estimation results for ablating the number of unlabeled examples m (x-axis) and confidence vector dimensions d c (y-axis), evaluated on the OPT-13B on MCQA using the XGBoost meta-model.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: Prompt selection results for all evaluation settings, measured by the absolute difference between ICL accuracy when prompted with c and the average accuracy. Blue bars show the actual best prompt (i.e., c = c * ), and red bars show the selected best prompt (i.e., c = c * )\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"\",\n \"figure_id\": \"\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"use c to denote a prompt in general, C i to denote the set of training prompts for dataset D i , and C test to denote test prompts for dataset D test .\",\n \"figure_data\": \"in\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"training examples. The meta-model is trained on the meta-training examples and predicts the estimated accuracy acc M,ctest test based on the test dataset feature vector conf M,ctest Dtest for each test prompt c test ∈ C test . Note that since closedset/open-ended generations have different confidence scores and accuracy evaluation metrics, the meta-model does not train on datasets that have a different task formulation than the test datasets.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"MethodsLLaMA-7BLLaMA-13BMMLUMCQACBQAMMLUMCQACBQAMeta ModelsMLP7.06 ± 1.038.82 ± 2.646.32 ± 2.135.80 ± 0.6311.04 ± 3.308.08 ± 2.993-NN5.98 ± 0.975.62 ± 2.357.10 ± 2.155.50 ± 0.6611.26 ± 4.568.56 ± 2.85XGBoost5.42 ± 4.525.22 ± 2.177.00 ± 2.445.00 ± 0.6211.52 ± 5.209.00 ± 4.85BaselinesAVGTRAIN5.26 ± 1.4010.26 ± 2.5011.62 ± 5.889.90 ± 1.7011.40 ± 3.3613.14 ± 7.25AVGCONF5.54 ± 0.546.88 ± 3.14n/a5.10 ± 0.868.58 ± 1.75n/aTS5.60 ± 1.636.32 ± 3.58n/a14.06 ± 1.5314.00 ± 6.95n/aATC20.34 ± 4.10 34.66 ± 9.36 31.80 ± 13.44 20.50 ± 4.9224.14 ± 7.72 31.80 ± 12.49Oracle5.82 (32)6.44 (32)7.06 (16)6.18 (32)13.44 (8)10.82 (8)ACC31.08 ± 6.32 39.00 ± 10.52 23.7 ± 10.44 45.50 ± 11.74 50.34 ± 12.84 29.30 ± 12.38MethodsOPT-6.7BOPT-13BMMLUMCQACBQAMMLUMCQACBQAMeta ModelsMLP5.70 ± 0.776.66 ± 1.257.28 ± 2.756.30 ± 1.147.18 ± 1.695.90 ± 1.183-NN4.06 ± 0.332.98 ± 0.576.46 ± 2.055.16 ± 0.242.78 ± 1.065.84 ± 2.19XGBoost3.76 ± 0.322.60 ± 0.738.06 ± 3.074.54 ± 0.352.54 ± 1.166.00 ± 2.43BaselinesAVGTRAIN3.48 ± 0.3710.66 ± 1.758.38 ± 3.864.28 ± 0.3511.46 ± 2.1518.00 ± 3.77AVGCONF8.42 ± 0.6013.42 ± 1.70n/a6.20 ± 0.777.14 ± 1.77n/aTS3.92 ± 0.174.60 ± 0.80n/a4.36 ± 0.312.72 ± 1.03n/aATC24.54 ± 2.12 37.64 ± 7.93 32.16 ± 14.60 24.72 ± 3.32 29.40 ± 10.28 30.34 ± 12.48Oracle5.58 (32)2.76 (128)6.50 (16)5.58 (32)2.76 (128)6.70 (16)ACC26.68 ± 4.42 33.16 ± 9.8017.54 ± 5.9226.82 ± 5.28 33.68 ±12.06 19.34 ± 6.3013),compared to OPT-6.7B (5.28), LLaMA-7B (6.50), LLaMA-13B(8.41). We report overall accuracy (ACC) by ExactMatch for MMLU and MCQA, and F1-score for CBQAthat we sample in-context examples from. We sam-ple 10 3-shot, 10 4-shot, and 10 5-shot prompts 11and decorate each of them with 5 prompt tem-plates chosen for MMLU (see Table 3). We choosed conf = 20 here since many of the datasets containonly 100 text examples. For MCQA and CBQA,we sample in-context examples from a pool of 100examples as the training set, and obtain a test setof 1000 examples (see Section A in the appendixfor implementation details). For each dataset, wesample 30 3-shot and 30 4-shot prompts and deco-rate them with only the null template. We choosed conf = 100 for both MCQA and CBQA settings.\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Full estimation results for all 4 LLMs, 3 dataset collections, and 3 meta-feature choices. XGBoost is the best overall meta-model structure with an average MAE of 6.60. The confidence vector is the best overall feature with an average MAE of 6.38 across all evaluation settings. The best meta-model outperforms all baseline methods in 9 out of 12 evaluation settings. We report overall accuracy (ACC) by Exact Match for MMLU and MCQA, and F1-score for CBQA\",\n \"figure_data\": \"MMLU\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Supporting Evidence\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. (2020) provides foundational evidence of the potential of in-context learning (ICL) with large language models (LLMs) in performing a wide range of language tasks, which supports the claims and hypotheses of the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Desai and Durrett, 2020)\", \"Explanation\": \"The cited work by Desai and Durrett (2020) studies calibration of LLMs on classification tasks, providing a methodological basis for the citing paper to build upon in their research on model confidence in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Jiang et al., 2021)\", \"Explanation\": \"The cited work by Jiang et al. (2021) studies calibration of LLMs on question answering (QA) tasks, providing a methodological basis for the citing paper to build upon in their research on model confidence in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kadavath et al., 2022)\", \"Explanation\": \"The cited work by Kadavath et al. (2022) also studies calibration of LLMs on question answering (QA) tasks, providing a methodological basis for the citing paper to build upon in their research on model confidence in LLMs.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Guo et al., 2017)\", \"Explanation\": \"The cited work by Guo et al. (2017) studies the concept of model confidence in LLMs, providing supporting evidence for the citing paper to build upon in their research on model confidence in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides the foundational concept of using auto-regressive language modeling objectives to train LLMs for in-context learning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chowdhery et al., 2022)\", \"Explanation\": \"The cited work by Chowdhery et al. further builds upon the concept of in-context learning by demonstrating the capabilities of LLMs in learning new tasks from prompt templates and in-context examples.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Lu et al., 2021)\", \"Explanation\": \"The cited work by Lu et al. provides evidence that the accuracy of in-context learning is sensitive to the choices of prompt templates and in-context examples, which supports the need for accurate estimation methods in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Zhao et al., 2021)\", \"Explanation\": \"The cited work by Zhao et al. further highlights the sensitivity of in-context learning accuracy to the choices of prompt templates and in-context examples, providing additional support for the need for accurate estimation methods in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Perez et al., 2021)\", \"Explanation\": \"The cited work by Perez et al. also contributes to the discussion on the sensitivity of in-context learning accuracy to the choices of prompt templates and in-context examples, providing additional support for the need for accurate estimation methods in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Singhal et al., 2022)\", \"Explanation\": \"The cited work by Singhal et al. (2022) provides a benchmark for evaluating the performance of ICL methods in the test dataset D test, which is used as a reference point in the citing paper to compare the results of the proposed methods.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work, GPT3, is used as a pre-trained language model in the citing paper to train a meta-model for estimating ICL performance based on confidence profiles of seen datasets.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2022)\", \"Explanation\": \"The cited work, OPT, is also used as a pre-trained language model in the citing paper to train a meta-model for estimating ICL performance based on confidence profiles of seen datasets.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kadavath et al., 2022)\", \"Explanation\": \"The cited work by Kadavath et al. provides a method for identifying answers in closed-set generation tasks by using model confidence as the normalized probability across the label space, which the citing paper adopts in their research on closed-set generation.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(2)\", \"Explanation\": \"The cited work by the author uses negative log-likelihood to obtain confidence scores from generated sequences in open-ended generation tasks, which the citing paper utilizes in their research on open-ended generation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Naeini et al., 2015)\", \"Explanation\": \"The cited work by Naeini et al. provides the foundational concept of example-level calibration error, which the citing paper builds upon to measure the calibration error in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Guo et al., 2017)\", \"Explanation\": \"The cited work by Guo et al. contributes to the development of the example-level calibration error measurement method, which the citing paper adopts in their research to evaluate the calibration of their model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kumar et al., 2019)\", \"Explanation\": \"The cited work by Kumar et al. provides a method for measuring example-level calibration error, which the citing paper uses to evaluate the calibration of their model in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Desai and Durrett, 2020)\", \"Explanation\": \"The cited work by Desai and Durrett contributes to the development of a method for measuring example-level calibration error, which the citing paper adopts in their research to evaluate the calibration of their model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Hinton et al., 2015)\", \"Explanation\": \"The cited work by Hinton et al. introduces the concept of temperature scaling as a calibration method, which the citing paper uses in their research to scale model-assigned probabilities and obtain confidence scores.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Si et al., 2022)\", \"Explanation\": \"The cited work by Si et al. provides a method for using temperature scaling to calibrate model-assigned probabilities, which the citing paper adopts in their research to scale the probabilities and obtain confidence scores.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Garg et al., 2022)\", \"Explanation\": \"The cited work by Garg et al. introduces the concept of average threshold confidence (ATC) as a method for OOD accuracy estimation, which the citing paper uses as a baseline in their research to evaluate the accuracy of their model in OOD settings.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Hendrycks et al., 2020)\", \"Explanation\": \"The cited work provides the multiple-choice QA datasets used in the evaluation of the ICL accuracy estimation method in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Lhoest et al., 2021)\", \"Explanation\": \"The cited work provides the implementations and training/test partitions used in the experiments conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2022)\", \"Explanation\": \"The cited work provides the OPT models used in the experiments, which the citing paper adopts in its research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Touvron et al., 2023)\", \"Explanation\": \"The cited work introduces the LLaMA model, which the citing paper extends the research on by using the LLaMA model in its experiments.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Touvron et al., 2023)\", \"Explanation\": \"The cited work provides the method used to generate prompts for the MMLU dataset, which the citing paper adopts in its research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Rae et al., 2021)\", \"Explanation\": \"The cited work is used to choose the up to 5-shot setting for the MMLU dataset, as it is studied in the research of Rae et al.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhao et al., 2021)\", \"Explanation\": \"The cited work by Zhao et al. provides a basis for understanding the sensitivity of ICL performance to prompt templates and in-context examples, which the citing paper uses to guide its own research on prompt selection for ICL.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Perez et al., 2021)\", \"Explanation\": \"The cited work by Perez et al. contributes to the understanding of prompt templates and in-context examples in ICL, which the citing paper uses to develop its own method for prompt selection.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chen et al., 2022)\", \"Explanation\": \"The cited work by Chen et al. further builds upon the research on prompt templates and in-context examples in ICL, providing insights that the citing paper uses in its own study on prompt selection.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhao et al., 2021)\", \"Explanation\": \"The cited work by Zhao et al. serves as a starting point for the citing paper to explore the potential of prompt selection in improving ICL performance, leading to a continuation of research in this area.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Perez et al., 2021)\", \"Explanation\": \"The cited work by Perez et al. serves as a foundational study for the citing paper to build upon in its own research on prompt selection for ICL, extending the research in this area.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Chen et al., 2022)\", \"Explanation\": \"The cited work by Chen et al. further extends the research on prompt selection in ICL, providing insights that the citing paper uses in its own study to improve ICL performance.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Gudibande et al., 2023)\", \"Explanation\": \"The cited work by Gudibande et al. (2023) is mentioned as a source of information that shows how instruction tuning can impact model performance on canonical datasets such as MMLU, which the citing paper uses to inform its own research on instruction calibration.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(OpenAI, 2023)\", \"Explanation\": \"The cited work by OpenAI (2023) is mentioned as a source of information on the impact of instruction tuning on calibration in LLMs, which the citing paper uses to support its own research on instruction calibration in LLMs.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Rae et al., 2021)\", \"Explanation\": \"The cited work by Rae et al. provides a data source for the MMLU datasets used in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Hendrycks et al., 2020)\", \"Explanation\": \"The cited work by Hendrycks et al. also serves as a data source for the MMLU datasets in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(ChatGPT)\", \"Explanation\": \"The use of ChatGPT to generate a prompt template for the MMLU datasets indicates a reliance on external data or pre-existing models for the study conducted in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Touvron et al., 2023)\", \"Explanation\": \"The cited work by Touvron et al. (2023) provides the LLaMA-7B and LLaMA-13B models that the citing paper uses in its research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2022)\", \"Explanation\": \"The cited work by Zhang et al. (2022) provides the OPT-6.7B and OPT-13B models that the citing paper uses in its research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Touvron et al., 2023)\", \"Explanation\": \"The cited work by Touvron et al. (2023) may have provided the methods or techniques used in the inference process of the LLaMA-7B and LLaMA-13B models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2022)\", \"Explanation\": \"The cited work by Zhang et al. (2022) may have provided the methods or techniques used in the inference process of the OPT-6.7B and OPT-13B models.\"}]"}}},{"rowIdx":72,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b16","b27","b5","b10","b7","b0","b25","b23","b36","b24","b10","b40","b3","b5","b13","b23","b18","b38","b8","b38","b14","b24","b21","b10","b32","b0","b36","b24","b30","b12","b29","b10","b1","b25","b19","b39","b22","b21","b7","b23","b37","b33"],"table_ref":[],"text":"Large pre-trained transformer models have demonstrated impressive performance across popular abstractive summarization benchmarks (Lewis et al., 2020;Raffel et al., 2020). Yet, transformer's quadratic memory complexity presents challenges for summarizing long documents with more than hundreds of words, such as scientific papers and investigation reports (Cohan et al., 2018;Huang et al., 2021), making it infeasible for researchers and developers with limited hardware resources (e.g., GPUs with insufficient memories) to contribute to this important research field.\nThe NLP community has made several innovations to address the long document challenge. Prior work divides a document into smaller chunks and summarizes each separately (Gidiotis and Tsoumakas, 2020), reduces the complexity of attention calculations (Beltagy et al., 2020), and removes unimportant content before running an abstractor (Pilault et al., 2020). In terms of memory efficiency, divide-and-conquer methods obtain the most significant advantage (Moro and Ragazzi, 2022). However, information outside of a document segment and their corresponding summaries become inaccessible, leading to uninformative and incoherent summaries. Unsurprisingly, state-of-the-art performance is obtained by models that can maintain global context, e.g., by combining global attentions with local attentions in transformer-based summarization models (Zaheer et al., 2021;Phang et al., 2022). Yet, they still require a large GPU memory footprint in practice. 1Though large language models like GPT-4 (Ope-nAI, 2023) are trained to handle up to 32K tokens, the privacy and security of data transmitted and shared through the API remain concerning, particularly in sectors dealing with sensitive information, e.g., clinical notes. Local model development can bolster privacy and security; however, limited computational resources in these scenarios necessitate the exploration of efficient modeling techniques.\nTherefore, this work aims to address the problem of long document summarization using constrained resources, specifically focusing on constrained GPU memory. We propose AWESOME 2 , which is built on the memory-efficient divide-and-conquer approach, and Augmented With Estimated Salient cOntent and MEmory mechanism. In essence, AWESOME maintains global context of both the source document and the summary generated so far with a limited memory usage, to enhance summary informativeness, faithfulness, and coherence.\nFirst, external memory mechanism is used on the encoder side of AWESOME to store information as it reads in document segments in sequence. This maintains relevant context for improved document understanding and salient content detection, thus promoting summary informativeness and faithfulness. Another memory is applied on the decoder side to improve generation coherence by tracking the partial summaries generated for prior document segments. Importantly, to ensure the GPU memory efficiency of AWESOME, we curb gradients from propagating to other document and summary segments and only allow a limited number of layers to maintain the external memory.\nSecond, AWESOME incorporates global salient content selected by an efficiently trained extractor through (1) direct text concatenation, or (2) inserting their key-value matrices into attention calculation. This lets the summarizer be aware of important topics at a global level, to enhance salience estimation and summary informativeness.\nWe experiment with five popular long-input benchmarks of different genres: investigation reports in GovReport (Huang et al., 2021), meeting transcripts in QMSum (Zhong et al., 2021), TV screenplays in SummScreen (Chen et al., 2022), scientific papers in arXiv (Cohan et al., 2018), and fictions in BookSum (Kryscinski et al., 2022). First, on all the five datasets, all AWE-SOME variants uniformly outperform Se3 (Moro and Ragazzi, 2022), the divide-and-conquer baseline, on summary informativeness as evaluated by ROUGE (Lin, 2004) and on coherence as measured by DiscoScore (Zhao et al., 2022) and a metric based on entity graphs (Guinaudeau and Strube, 2013)-both metrics are highly correlated with human judgment, according to Zhao et al. (2022). Second, AWESOME with memory mechanisms also improves summary faithfulness over Se3 on Gov-Report, according to SummaC (Laban et al., 2022), an entailment-based faithfulness metric. Lastly, compared with more memory-intensive models that also maintain global context, such as Phang et al. (2022) and Liu et al. (2022), AWESOME achieves higher automatic scores for informativeness, coherence, and faithfulness on GovReport (Huang et al., 2021). On BookSum which comprises the lengthiest documents and summaries among the Efficient attentions are designed to reduce the quadratic complexity of the original transformer architecture (Vaswani et al., 2017) and maintain full encoding context by combining global attentions with local attentions built on sliding windows (Beltagy et al., 2020;Zaheer et al., 2021), text blocks (Phang et al., 2022;Tay et al., 2020), or clusters of similar tokens (Kitaev et al., 2020;Roy et al., 2021). Besides the aforementioned attention variants designed for self-attentions, recent work has reduced the memory usage of decoder cross attentions by distributing encoder outputs to different attention heads (Huang et al., 2021) or selecting attendable encoder outputs via KNN search (Bertsch et al., 2023). Despite the reduced complexity, efficient attention-based systems effectively require reading the full document x to generate a summary y during model training and thus still need huge GPU memory that scales with the input length.\nApproach In→Out Enc Enc←Dec Dec Efficient Attention x → y ■ ■ Extract-Abstract xe → y □ ⋆ Dynamic Weight x → y □ ■ + ⋆ Divide-Conquer xi → yi □ □\nExtract-then-abstract systems circumvent the long sequence challenge by first identifying the salient segments, x e (e.g., sentences), using an extractor, and then running an abstractor over x e to produce the final summary (Pilault et al., 2020;Liu and Lapata, 2019;Zhao et al., 2020). However, the extracted segments may contain incomplete and out-of-context information that leads to incomprehensible and unfaithful summaries.\nTo mitigate the error propagation issue of a twostage approach, recent studies bridge the extractor and abstractor via dynamic weights over document segments. Rather than feeding the extracted segments directly to the abstractor, at each summary decoding step, DYLE (Mao et al., 2022) first predicts an output token distribution for each segment separately, and then aggregates over all the extracted segments as weighted by their extraction salience. PageSum (Liu et al., 2022) further alleviates context loss by averaging decoder output representations conditioned on all document segments. Though their abstractor processes each document segment x i separately, jointly training the extractor and the abstractor still requires loading the full document x into the GPU memory.\nDivide-and-conquer systems split a document into multiple non-overlapping segments and summarize each segment separately, as done in Gidiotis and Tsoumakas (2020) and Se3 (Moro and Ragazzi, 2022). Summ N (Zhang et al., 2022) uses an additional summarization stage to further condense the segmented summaries. As each document segment x i is summarized separately, the divide-andconquer approach's fixed GPU memory footprint is independent from the document length. This fits well with our goal of long document summarization with limited memory. However, without access to other parts of the document and their summaries, the summarizer struggles for content salience estimation in each isolated segment, and generates incoherent outputs when piecing together summaries. Though Wu et al. (2021) concatenate previously generated summaries as part of the input, a complicated strategy is required for training sample construction.\nAWESOME is built on the memory-efficient divide-and-conquer approach, and improves summary informativeness, coherence, and faithfulness by using newly designed external memories for accumulating salient information from other document segments and their generated summaries. We further augment AWESOME with global salient content to provide important topics at the document level, when summarizing each segment."},{"figure_ref":[],"heading":"Memory and Content Augmentation","publication_ref":["b6","b26","b2"],"table_ref":[],"text":"Different memory mechanisms have been studied for long-range text understanding tasks. For instance, Transformer-XL (Dai et al., 2019) caches intermediate representations produced in the last document segment and attends over these representations. Compressive Transformer (Rae et al., 2020) further increases the context range by compressing the oldest cached representations. To simulate memory reading and writing, Recurrent Memory Transformer (Bulatov et al., 2022) includes extra memory vectors in each text segment and passes their corresponding output vectors to the next segment. Instead of using a memory with a fixed size, Memorizing Transformer (Wu et al., 2022a) stores all prior representations as key-value pairs, and performs an approximate kNN lookup to retrieve representations to augment the current segment. However, existing work on memory mechanisms focuses on language modeling, while incorporating memory mechanisms into the decoding process for generation tasks is nontrivial as it requires updating both decoding states (e.g., beams) and memory states. Our work is the first to leverage memory mechanisms and content augmentation to incorporate global context for the purpose of memory-efficient long document summarization."},{"figure_ref":[],"heading":"External Memory and Global Salient Content Augmentation","publication_ref":["b23","b0","b16"],"table_ref":[],"text":"The architecture of AWESOME (Figure 1) is based on Se3 (Moro and Ragazzi, 2022), where a document is summarized segment by segment, with the final summary obtained by concatenating the resultant summaries. Document sentences are split into segments with up to 768 tokens each, while reference summary sentences are assigned to their most overlapping segment to create the oracle summary, as detailed in Appendix A. Following Longformer (Beltagy et al., 2020), we initialize the encoder and decoder parameters from BART (Lewis et al., 2020). AWESOME preserves the global context and builds communications across segments with minimal GPU memory increase, by (1) employing external memories in both the encoder and the decoder to gather relevant information ( §3.1), and (2) augmenting the encoder with salient content from other segments ( §3.2)."},{"figure_ref":[],"heading":"External Memory Mechanisms","publication_ref":["b26","b15","b15"],"table_ref":[],"text":"We design two external memory mechanisms to efficiently enable the information flow from prior segments to the current segment. Specifically, each memory module maintains a matrix M ∈ R m×d , where m = 1024 is the memory size and d = 1024 is the hidden state dimension of BART. M is updated after encoding each document segment and then passed to the next segment. We denote the memory matrix after the t-th segment as M t . Each layer of the encoder and decoder can be equipped with one such external memory. Below we describe two mechanisms to update M t and incorporate it in both the encoding and decoding processes. The layer index in the formulas is omitted for simplicity.\nCompressive Memory. For each document seg-ment, compression-based memory caches its input vectors to be fed into self-attention calculation.\nSince storing the input vectors as-is requires the memory usage m to scale linearly with the context length, we dedicate half of M t to store the compressed memory, with a compression ratio of r. With H t inp denoting the matrix that contains input vectors to the transformer self-attention, the memory compression and update processes are:\nM t-1 c , M t-1 u = M t-1 [: m 2 ], M t-1 [ m 2 :](1)\nM ′ u = concat(M t-1 u , SG(H t inp )) (2) M ′ c = compress(M ′ u [: - m 2 ])(3)\nM t u = M ′ u [- m 2 :](4)\nM t c = concat(M t-1 c , M ′ c )[- m 2 :] (5) M t = concat(M t c , M t u )(6)\nwhere SG(•) denotes stopping the gradient backpropagation to lower GPU usage, and compress(•) performs convolutions with their stride and kernel size set to the compression ratio r. r is set to 5 after tuning on the development sets.\nNext, to leverage the memory from the previous segment in summarizing the current segment, M t-1 is concatenated with the inputs to the selfattentions to obtain the key-value matrices:\nH t mem = concat(M t-1 , H t inp )(7)\nH t self = Attn(H t inp query , H t mem key , H t mem value )(8)\nwhere H t self is the output of the self-attention. Our compression-based memory is adopted from Compressive Transformer (Rae et al., 2020), a decoder-only model for language modeling. We are the first to apply it to both the encoder and the decoder of a Transformer model and on long document summarization tasks.\nCompressive memory favors recency, particularly the previous segment and its summary, potentially causing older relevant history to be lost during compression. Attentive Memory. To mitigate the recency bias by compressive memory, we further investigate an attention-based memory updating mechanism, to selectively include content in M t . First, the memory is additionally accompanied by an extra crossattention in each of the encoder and decoder layers, specialized in retrieving relevant information from M t . Following prior study (Lei et al., 2020) that uses memories in video captioning, we update M t with a gate matrix G t to control the amount of content to be updated:\nM t = G t ⊙ U t + (1 -G t ) ⊙ M t-1 (9)\nwhere ⊙ denotes the element-wise product and U t is the matrix containing vectors to update the memory. U t and G t are obtained as follows:\nU t = tanh(W u1 M t-1 + W u2 S t )(10)\nG t = σ(W g1 M t-1 + W g2 S t )(11)\nS t = Attn(M t-1 query , SG(H t self ) key , SG(H t self ) value )(12)\nwhere W * are learnable matrices, S t synthesizes the current segment via an attention calculation, and SG(•) indicates stopping the gradient backpropagation. In each encoder and decoder layer, an extra cross-attention is inserted after the selfattention, where M t-1 is attended and incorporated into the current segment's summarization process. Unlike our approach, the memory in Lei et al. (2020) does not employ gradient stopping. This omission eliminates the memory efficiency gained from the divide-and-conquer strategy, leading to comparable high memory usage as the efficient attention strategy. 3 While their memory is suitable for generating short image captions, our design with gradient stopping is crucial for efficient long document summarization.\nSelective Addition of External Memory. External memory incurs overhead in GPU memory usage. To mitigate this overhead, we consider selectively adding external memory to specific layers, as the importance of external memory varies according to the different functions of layers in the model. Our pilot study suggests that the last layers of the Transformer model more effectively utilize external memory compared to the first layers. To avoid an exhaustive search for the optimal layer or combination of layers for each dataset, we choose to uniformly equip the last three layers with external memory across all datasets unless otherwise specified. 4"},{"figure_ref":[],"heading":"Global Salient Content Augmentation","publication_ref":[],"table_ref":[],"text":"The memory mechanisms only grant access to prior content in the documents, yet subsequent context can also help with salience estimation, e.g., elaborating the pros and cons of a proposed solution makes it necessary to introduce the problem and the solution. Moreover, memories store content implicitly, so it is unclear whether relevant information can be stored and retrieved effectively. Therefore, we inform the system of a document's important sentences, which are pre-identified by a separatelytrained extractor. The details of extractor training can be found in Appendix D. After extracting important sentences in a document, we study two methods of injecting them into the summarizer.\nText Concatenation. For each segment, we include the extracted sentences in the following way to prioritize long-term context. We start with the \"outermost\" extracted sentences, i.e., the earliest sentence in the past segments and the last sentence in the future segments, and repeat this process until the input has reached the maximum length accepted by the positional encoding of the model (1024 for BART). 5 To differentiate the content in the current segment from the added sentences, we prefix the current segment and the added sentences from before/after the current segment with \"Current chunk:\", \"Previous important sentences:\", and \"Next important sentences:\", respectively. Text concatenation is easy to implement and most compatible with the source modality, but the memory usage increase is quadratic to the length of the augmented content.\nKey-value Vectors. To circumvent the quadratic memory increase, we join the key-value representations of tokens in important sentences in the encoder self-attentions, and directly inject them into the summarizer encoder. The memory increase is only linear to the augmented content's length.\nConcretely, the summarizer encoder first encodes all document segments and obtains the representations (i.e., encoder outputs) of tokens belonging to the extracted important sentences. During training, the token representations of these sentences are concatenated with the key-value matrices in the encoder self-attentions while the query matrix remains in its original form. Up to 1024 tokens are concatenated via the same inclusion method for text concatenation, to prioritize the outermost sentences. A similar idea has been used by Memorizing Transformer (Wu et al., 2022a) to include retrieved text representations from past segments for long-form language modeling. Our method differs in two aspects. First, we extract representations from future segments, which are crucial for accurately identifying salient content. Second, we apply a learnable projection to the augmented representations prior to key-value concatenation. This process is crucial in improving compatibility with the original key-value matrices.\nModel R-1 ↑ R-2 ↑ R-L ↑ Ent Prec ↑ SummaC ↑ Disco ↓ Ent Graph ↑ GPU"},{"figure_ref":[],"heading":"Experimental Setups","publication_ref":["b10","b40","b3","b1","b21","b18","b38","b11","b14"],"table_ref":[],"text":"Datasets. We conduct experiments on GovReport (Huang et al., 2021), QMSum (Zhong et al., 2021), SummScreen (Chen et al., 2022) (Bertsch et al., 2023). We also include an extract-then-abstract model (Extract-Abstract) and PageSum (Liu et al., 2022) that leverages dynamic weights, as discussed in §2. All models are initialized from BART-large, except for LongT5 that is pre-trained on long-form data. Details of baseline models are reported in Appendix D.\nEvaluation Metrics. We evaluate summary informativeness using ROUGE (Lin, 2004). To measure coherence, we use DiscoScore (Zhao et al., 2022) (Disco), a reference-based metric that evaluates discourse coherence by comparing focus (e.g., nouns) frequency and semantics between the system summary and the reference. We also report a graph-based reference-free coherence metric (Guinaudeau and Strube, 2013) (Ent Graph), which measures the connectivity of summary sentences linked by entities, reflecting the coherence of topic transitions. For summary faithfulness, we follow prior work on text generation (Iv et al., 2022) and show the precision of the entities (Ent Prec) in the summary with respect to the document. Additionally, a recent model-based faithfulness metric, SummaC (Laban et al., 2022), is used. Finally, we show the maximum size of allocated GPU memory by each model during training.\nSe3: VA is required to publish information on appointment wait times at each VA medical facility for primary care, specialty care, and hospital care and medical services, which it does through two public websites. VA has taken a number of actions to address deficiencies GAO found in wait-time measurement and implementation of its scheduling policy. For wait-time measurement, these actions included changes to the wait-time measurement definitions, provision and documentation of scheduler training, and improved oversight through audits, all of which have been in a state of flux for the past 6 years. On July 12, 2019, VA provided GAO additional updates on efforts to implement GAO's related recommendations. AWESOME: GAO recommended that VA either clarify its scheduling policy to better define the desired date, or identify clearer wait-time measures that are not subject to interpretation and prone to scheduler error. VA concurred with the recommendation, which GAO has identified as among those recommendations that warrant priority attention. VA has taken a number of actions to address GAO's recommendations regarding deficiencies GAO found in wait-time measurement and implementation of its scheduling policy. For wait-time measurement, these actions included changes to the wait-time measurement definitions, provision and documentation of scheduler training, and improved oversight through audits, all of which have been in a state of flux for the past 6 years. On July 12, 2019, VA provided GAO additional updates on efforts to implement GAO's related recommendations.\nTable 3: Summary snippets generated by Se3 and AWE-SOME. AWESOME's summary is more coherent, with natural transitions surrounding \"GAO's recommendation\", while Se3 abruptly introduces the topic."},{"figure_ref":[],"heading":"Results","publication_ref":["b10"],"table_ref":["tab_2","tab_4","tab_5","tab_6","tab_7"],"text":"We report results by all AWESOME variants and comparison models on GovReport in Table 2. Compared with Se3, AWESOME variants consistently achieve better performance on both ROUGE and coherence scores, indicating the importance of maintaining global context for accurate salience estimation of local content and enforcing coherent transitions across segment-level summaries. This can also be demonstrated by the sample outputs in Table 3. Summaries generated by Se3 tend to be shorter, as Se3 fails to plan at a global level. On faithfulness, AWESOME with attentive memory has the best entity precision among all models and also improves SummaC over Se3, while only augmenting AWESOME with global salient content hurts faithfulness. Inspecting the model outputs, we find that using attentive memory improves understanding concepts of long-term dependencies, e.g., connecting a strategy with its related information that appears earlier in the report.\nOf the two types of external memory mechanisms, attentive memory outperforms compression- based memory on all metrics, which highlights the advantage of adaptively updating the stored context. Meanwhile, directly concatenating salient content with the input yields higher ROUGE scores than injecting key-value vectors into the attention calculation, though the latter is less memory-intensive.\nModel R-1 ↑ R-2 ↑ R-L ↑ Disco ↓ GPU ↓Se3\nWe believe natural language-based augmentation better interleaves with the document segment, echoing the findings by prior work on using retrieval for question answering (Wu et al., 2022b). Importantly, under a strict GPU memory constraint, AWESOME with external memory mechanisms and global salient content augmentation achieves the best ROUGE scores among all models, while obtaining competitive results on other measures. Though efficient attention models and Page-Sum can perform remarkably when given highercapacity GPUs as in the original work, they generate less informative summaries when truncation is required to comply with the memory constraint, emphasizing the importance of studying memoryefficient long document summarization models. Furthermore, with selective addition of external memory, AWESOME adds only about 4GB of GPU memory usage, enhancing the model performance efficiently.\nOn QMSum (Table 4), AWESOME with attention-based memory outperforms all comparisons on ROUGE scores. While our models' summaries are more coherent than the summaries by Se3, as measured by DiscoScore, the differences among all models are less pronounced compared to the ones on GovReport. This is because QM-Sum contains shorter summaries than GovReport (69 vs. 553), thus involving fewer topic transitions.\nModel R-1 ↑ R-2 ↑ R-L ↑ Ent G ↑ GPU ↓ Se3 38\nWe also find that the extractor performs poorly on QMSum, leading to degraded results after augmenting our model with the extracted salient content. Specifically, the F1 score of the extractor on the test set is only 1.29, as opposed to 27.85 on GovReport. Compared to our model, the extract-then-abstract model is more prone to its errors and produce summaries of the lowest quality on QMSum. This trend is similarly observed on Summ-Screen (Table 5) and arXiv (Table 6), where the extract-then-abstract method performs poorly and adding extracted content leads to performance drop of AWESOME due to the low performance of the extractor. Meanwhile, AWESOME with the attentive memory is able to obtain the best ROUGE-1 and ROUGE-L scores. On arXiv, models that use efficient attentions obtain the higher ROUGE scores, because truncating arXiv documents has little effect on summary generation-arXiv articles have the most uneven distributions of salient content, where only about 10% of new salient bigrams are located in the second halves of the documents (Huang et al., 2021).\nFinally, experiments on BookSum show that the divide-and-conquer method produces better summaries for long novels, while our method can further boost its performance (Table 7). However, we find it necessary to incorporate external memory into all layers, suggesting a more complex interaction of external memory with the summarization process for novel plots. Unlike other document types tested, novel plots are typically sequential with less redundancy, which reduces the necessity of the memory mechanism. \nR-1 ↑ R-2 ↑ R-L ↑ Disco ↓ GPU ↓Se3"},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"We present AWESOME for summarizing long documents in a memory-constrained setting. Based on the divide-and-conquer strategy, AWESOME uses two mechanisms to gather global context and improve summary quality. First, external memories on the encoder and decoder are employed to track previously read document content and the corresponding summaries. Second, the encoder is informed of global salient content predicted by an extractor via text or representation concatenation. On five summarization datasets, AWESOME generates summaries with better informativeness, faithfulness, and coherence than a baseline divideand-conquer system. Under the same memory constraint, AWESOME outperforms competitive models that leverage efficient attentions or dynamic extraction to preserve global context, highlighting its effectiveness in supplying global context. AWESOME's external memory mechanism is restricted to operating solely from past segments to the current segment. This means that the model does not leverage the information contained in future segments, which can be relevant for a comprehensive understanding of the current segment.\nTo address this limitation, we have designed the global salient content augmentation mechanism to cover context from the future segments, yet more advanced solutions can be explored in future work. For example, on the encoder, making the external memory bidirectional is a potential approach.\nWhile being memory-efficient, the external memory mechanism of AWESOME necessitates a longer running time due to its recurrent nature. The need for recurrent computations may lead to increased processing requirements, which could impact real-time applications or scenarios where rapid responses are crucial. The running times of different models are provided in Appendix B.1 for reference. Although our model is slower than that of LongT5 and Se3, it still outperforms several other competitive models in terms of speed, and we will investigate methods for reducing the running time in future work."},{"figure_ref":[],"heading":"Ethical Consideration","publication_ref":[],"table_ref":[],"text":"We anticipate that one of the major use cases of AWESOME is to allow ordinary users who have computing devices with limited memory to quickly understand government policies and other types of long documents. However, we recognize that the system generated summaries might not comprehensively cover the salient content that is essential for correctly understanding the policies, causing risks ranging from capital loss to legal liability. Moreover, system summaries might contain statements that cannot be verified through the document, which further adds to the risks of real-world deployment. We suggest developers who intend to use our model for real-world application carefully study the outputs by our model before the actual deployment."},{"figure_ref":[],"heading":"A Divide-and-Conquer Architecture","publication_ref":["b23","b28"],"table_ref":[],"text":"We choose Se3 (Moro and Ragazzi, 2022) as our base divide-and-conquer architecture because it can be applied to any document-summary pair. In order to create divide-and-conquer training data for summarization, for each document-summary pair, the document is first divided into segments ( §A.1) and each summary sentence is then assigned to a document segment as part of the generation target ( §A.2). The length of each document segment is between 512 and 768 tokens. During segmentation, the algorithm loops through all document sentences, as shown in Algorithm 1. A document sentence will be added to the current segment if the segment contains less than 512 tokens. The current segment will be finalized if the current segment contains more than 768 tokens or the current sentence is more semantically similar to the next pseudo segment than the current segment, where the next pseudo segment is created by including future sentences until reaching 512 tokens. To measure the similarity between the current sentence and a segment, we use the average cosine similarity between the representation of the current sentence and representations of the sentences in the segment. Sentence representations are obtained using Sentence Transformer (Reimers and Gurevych, 2019) with the all-roberta-large-v1 model."},{"figure_ref":[],"heading":"A.1 Document","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"A.2 Target Assignment","publication_ref":[],"table_ref":[],"text":"For each sentence in the reference summary, we calculate its ROUGE scores with the document segments. The sentence will then be assigned to the document segment with which yields the highest ROUGE-1 and ROUGE-2 scores."},{"figure_ref":[],"heading":"B Additional Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_1"],"heading":"B.1 Running Time","publication_ref":[],"table_ref":[],"text":"We compare the model running time on GovReport (Figure 2). The input document is truncated to 16384 tokens and each model is separately train for 1000 steps with a batch size of 1. No other computation-heavy program is running at the same time. While AWESOME take longer time to complete training than Se3, it is still the third fastest model."},{"figure_ref":[],"heading":"C Dataset Details C.1 Statistics","publication_ref":["b10","b40","b5","b13","b3","b1"],"table_ref":["tab_10"],"text":"We conduct experiments on five long document summarization datasets with diverse genres. Gov-Report (Huang et al., 2021) contains long reports and their summaries written by government research agencies. QMSum (Zhong et al., 2021) is a query-focused long meeting transcript summarization dataset, with summary-worthy content spread over the documents. We prepend the query to all segments. We further use a screenplay summarization dataset, SummScreen (Chen et al., 2022), which contains the transcripts of TV series. The TMS subset, with more samples and longer summaries, is selected. Moreover, we experiment with the scientific papers and their abstracts from arXiv (Cohan et al., 2018). Finally, we test our models on summarizing full novels in Book-Sum (Kryscinski et al., 2022). For all datasets, we use the official train/dev/test splits if their original data files are released. Statistics of datasets are reported in Table 8. For GovReport6 , QMSum7 , and SummScreen (Chen et al., 2022), we use the data released by the original papers. For arXiv, we use the version provided by Huggingface Datasets. 8 As the original data files for BookSum are not released due to summary copyright, we use the version reproduced by Unlimiformer (Bertsch et al., 2023)."},{"figure_ref":[],"heading":"C.2 Input Truncation","publication_ref":[],"table_ref":[],"text":"In our main experiments, we employ a GPU memory constraint of 27GB. As some baseline models require the input length to be a multiplier of 1024, setting a constraint of 24GB, a more common num-"},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"ber, would lead to further truncation and significant performance drop.\nTo fit models into our memory constraint, we truncate the model inputs. The truncation thresholds used by each model on different datasets are shown in Table 9. Although Se3 and AWESOME theoretically maintain a consistent GPU memory consumption during training regardless of the number of input tokens processed, we have chosen to restrict the maximum number of input tokens in a training sample to 51200 for reasonable training time."},{"figure_ref":[],"heading":"D Implementation Details","publication_ref":["b24","b0","b9","b1","b21","b22","b18","b38","b8","b14"],"table_ref":[],"text":"Baselines. BlockAttn and Longformer use blockwise attentions (Phang et al., 2022) and slidingwindow attentions (Beltagy et al., 2020), where a global token can attend to and be attended by all tokens, while other tokens can only attend to tokens in the same block or window. LongT5 (Guo et al., 2022) is a sliding-window attention model pre-trained on long sequences, and Unlimiformer (Bertsch et al., 2023) extends BART by selecting input tokens to be attended to via KNN searching. For the extract-then-abstract approach, we use the same extractor as in the global salient content augmentation of our model, and the abstractor takes as input oracle extracted sentences during training. Lastly, PageSum (Liu et al., 2022) synthesizes the output representations given by different document segments with dynamic weights.\nExtractor. The extractor first uses a RoBERTa (Liu et al., 2019) to encode each sentence and takes the average of the final layer's outputs as the sentence representation. It then applies a self-attention on top of all sentence representations. The resulting representations are converted to extraction scores after applying a multi-layer perception with one hidden layer. The extractor is trained with oracle extractive labels that are constructed by greedily searching for document sentences that maximize the sum of ROUGE-1 and ROUGE-2 scores, compared against the reference summary. We do not compute ROUGE-L as in DYLE (Mao et al., 2022), because finding the longest common subsequence is computationally expensive and does not yield performance gain.\nTraining Parameters. We train all models with a maximum learning rate of 5 × 10 -5 , except that LongT5 is trained with a maximum learning rate of 1 × 10 -4 . We use a running batch size of 1 and apply gradient accumulation to achieve an effective batch size of 8. The numbers of training epochs are 3, 9, 6, 2, 10 on GovReport, QMSum, SummScreen, arXiv, and BookSum, with warmup steps of 300, 100, 300, 1000, and 40. Due to the computational cost of training long document summarization, each model is trained for a single run.\nModel Size. AWESOME is based on BART-large 9 and has 708 millions of parameters.\nComputing Infrastructure. All experiments are conducted on RTX A6000 GPUs.\nEvaluation Metrics. For ROUGE (Lin, 2004), we use the Python implementation by Google. 10 The official code for DiscoScore (Zhao et al., 2022) is used 11 , which also provides an implementation of the Ent Graph metric (Guinaudeau and Strube, 2013). We implement the entity precision measure ourselves and run the official code for Sum-maC (Laban et al., 2022). 12 All metrics used are open-source and can be distributed for research purposes."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b16\",\n \"b27\",\n \"b5\",\n \"b10\",\n \"b7\",\n \"b0\",\n \"b25\",\n \"b23\",\n \"b36\",\n \"b24\",\n \"b10\",\n \"b40\",\n \"b3\",\n \"b5\",\n \"b13\",\n \"b23\",\n \"b18\",\n \"b38\",\n \"b8\",\n \"b38\",\n \"b14\",\n \"b24\",\n \"b21\",\n \"b10\",\n \"b32\",\n \"b0\",\n \"b36\",\n \"b24\",\n \"b30\",\n \"b12\",\n \"b29\",\n \"b10\",\n \"b1\",\n \"b25\",\n \"b19\",\n \"b39\",\n \"b22\",\n \"b21\",\n \"b7\",\n \"b23\",\n \"b37\",\n \"b33\"\n ],\n \"table_ref\": [],\n \"text\": \"Large pre-trained transformer models have demonstrated impressive performance across popular abstractive summarization benchmarks (Lewis et al., 2020;Raffel et al., 2020). Yet, transformer's quadratic memory complexity presents challenges for summarizing long documents with more than hundreds of words, such as scientific papers and investigation reports (Cohan et al., 2018;Huang et al., 2021), making it infeasible for researchers and developers with limited hardware resources (e.g., GPUs with insufficient memories) to contribute to this important research field.\\nThe NLP community has made several innovations to address the long document challenge. Prior work divides a document into smaller chunks and summarizes each separately (Gidiotis and Tsoumakas, 2020), reduces the complexity of attention calculations (Beltagy et al., 2020), and removes unimportant content before running an abstractor (Pilault et al., 2020). In terms of memory efficiency, divide-and-conquer methods obtain the most significant advantage (Moro and Ragazzi, 2022). However, information outside of a document segment and their corresponding summaries become inaccessible, leading to uninformative and incoherent summaries. Unsurprisingly, state-of-the-art performance is obtained by models that can maintain global context, e.g., by combining global attentions with local attentions in transformer-based summarization models (Zaheer et al., 2021;Phang et al., 2022). Yet, they still require a large GPU memory footprint in practice. 1Though large language models like GPT-4 (Ope-nAI, 2023) are trained to handle up to 32K tokens, the privacy and security of data transmitted and shared through the API remain concerning, particularly in sectors dealing with sensitive information, e.g., clinical notes. Local model development can bolster privacy and security; however, limited computational resources in these scenarios necessitate the exploration of efficient modeling techniques.\\nTherefore, this work aims to address the problem of long document summarization using constrained resources, specifically focusing on constrained GPU memory. We propose AWESOME 2 , which is built on the memory-efficient divide-and-conquer approach, and Augmented With Estimated Salient cOntent and MEmory mechanism. In essence, AWESOME maintains global context of both the source document and the summary generated so far with a limited memory usage, to enhance summary informativeness, faithfulness, and coherence.\\nFirst, external memory mechanism is used on the encoder side of AWESOME to store information as it reads in document segments in sequence. This maintains relevant context for improved document understanding and salient content detection, thus promoting summary informativeness and faithfulness. Another memory is applied on the decoder side to improve generation coherence by tracking the partial summaries generated for prior document segments. Importantly, to ensure the GPU memory efficiency of AWESOME, we curb gradients from propagating to other document and summary segments and only allow a limited number of layers to maintain the external memory.\\nSecond, AWESOME incorporates global salient content selected by an efficiently trained extractor through (1) direct text concatenation, or (2) inserting their key-value matrices into attention calculation. This lets the summarizer be aware of important topics at a global level, to enhance salience estimation and summary informativeness.\\nWe experiment with five popular long-input benchmarks of different genres: investigation reports in GovReport (Huang et al., 2021), meeting transcripts in QMSum (Zhong et al., 2021), TV screenplays in SummScreen (Chen et al., 2022), scientific papers in arXiv (Cohan et al., 2018), and fictions in BookSum (Kryscinski et al., 2022). First, on all the five datasets, all AWE-SOME variants uniformly outperform Se3 (Moro and Ragazzi, 2022), the divide-and-conquer baseline, on summary informativeness as evaluated by ROUGE (Lin, 2004) and on coherence as measured by DiscoScore (Zhao et al., 2022) and a metric based on entity graphs (Guinaudeau and Strube, 2013)-both metrics are highly correlated with human judgment, according to Zhao et al. (2022). Second, AWESOME with memory mechanisms also improves summary faithfulness over Se3 on Gov-Report, according to SummaC (Laban et al., 2022), an entailment-based faithfulness metric. Lastly, compared with more memory-intensive models that also maintain global context, such as Phang et al. (2022) and Liu et al. (2022), AWESOME achieves higher automatic scores for informativeness, coherence, and faithfulness on GovReport (Huang et al., 2021). On BookSum which comprises the lengthiest documents and summaries among the Efficient attentions are designed to reduce the quadratic complexity of the original transformer architecture (Vaswani et al., 2017) and maintain full encoding context by combining global attentions with local attentions built on sliding windows (Beltagy et al., 2020;Zaheer et al., 2021), text blocks (Phang et al., 2022;Tay et al., 2020), or clusters of similar tokens (Kitaev et al., 2020;Roy et al., 2021). Besides the aforementioned attention variants designed for self-attentions, recent work has reduced the memory usage of decoder cross attentions by distributing encoder outputs to different attention heads (Huang et al., 2021) or selecting attendable encoder outputs via KNN search (Bertsch et al., 2023). Despite the reduced complexity, efficient attention-based systems effectively require reading the full document x to generate a summary y during model training and thus still need huge GPU memory that scales with the input length.\\nApproach In→Out Enc Enc←Dec Dec Efficient Attention x → y ■ ■ Extract-Abstract xe → y □ ⋆ Dynamic Weight x → y □ ■ + ⋆ Divide-Conquer xi → yi □ □\\nExtract-then-abstract systems circumvent the long sequence challenge by first identifying the salient segments, x e (e.g., sentences), using an extractor, and then running an abstractor over x e to produce the final summary (Pilault et al., 2020;Liu and Lapata, 2019;Zhao et al., 2020). However, the extracted segments may contain incomplete and out-of-context information that leads to incomprehensible and unfaithful summaries.\\nTo mitigate the error propagation issue of a twostage approach, recent studies bridge the extractor and abstractor via dynamic weights over document segments. Rather than feeding the extracted segments directly to the abstractor, at each summary decoding step, DYLE (Mao et al., 2022) first predicts an output token distribution for each segment separately, and then aggregates over all the extracted segments as weighted by their extraction salience. PageSum (Liu et al., 2022) further alleviates context loss by averaging decoder output representations conditioned on all document segments. Though their abstractor processes each document segment x i separately, jointly training the extractor and the abstractor still requires loading the full document x into the GPU memory.\\nDivide-and-conquer systems split a document into multiple non-overlapping segments and summarize each segment separately, as done in Gidiotis and Tsoumakas (2020) and Se3 (Moro and Ragazzi, 2022). Summ N (Zhang et al., 2022) uses an additional summarization stage to further condense the segmented summaries. As each document segment x i is summarized separately, the divide-andconquer approach's fixed GPU memory footprint is independent from the document length. This fits well with our goal of long document summarization with limited memory. However, without access to other parts of the document and their summaries, the summarizer struggles for content salience estimation in each isolated segment, and generates incoherent outputs when piecing together summaries. Though Wu et al. (2021) concatenate previously generated summaries as part of the input, a complicated strategy is required for training sample construction.\\nAWESOME is built on the memory-efficient divide-and-conquer approach, and improves summary informativeness, coherence, and faithfulness by using newly designed external memories for accumulating salient information from other document segments and their generated summaries. We further augment AWESOME with global salient content to provide important topics at the document level, when summarizing each segment.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Memory and Content Augmentation\",\n \"publication_ref\": [\n \"b6\",\n \"b26\",\n \"b2\"\n ],\n \"table_ref\": [],\n \"text\": \"Different memory mechanisms have been studied for long-range text understanding tasks. For instance, Transformer-XL (Dai et al., 2019) caches intermediate representations produced in the last document segment and attends over these representations. Compressive Transformer (Rae et al., 2020) further increases the context range by compressing the oldest cached representations. To simulate memory reading and writing, Recurrent Memory Transformer (Bulatov et al., 2022) includes extra memory vectors in each text segment and passes their corresponding output vectors to the next segment. Instead of using a memory with a fixed size, Memorizing Transformer (Wu et al., 2022a) stores all prior representations as key-value pairs, and performs an approximate kNN lookup to retrieve representations to augment the current segment. However, existing work on memory mechanisms focuses on language modeling, while incorporating memory mechanisms into the decoding process for generation tasks is nontrivial as it requires updating both decoding states (e.g., beams) and memory states. Our work is the first to leverage memory mechanisms and content augmentation to incorporate global context for the purpose of memory-efficient long document summarization.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"External Memory and Global Salient Content Augmentation\",\n \"publication_ref\": [\n \"b23\",\n \"b0\",\n \"b16\"\n ],\n \"table_ref\": [],\n \"text\": \"The architecture of AWESOME (Figure 1) is based on Se3 (Moro and Ragazzi, 2022), where a document is summarized segment by segment, with the final summary obtained by concatenating the resultant summaries. Document sentences are split into segments with up to 768 tokens each, while reference summary sentences are assigned to their most overlapping segment to create the oracle summary, as detailed in Appendix A. Following Longformer (Beltagy et al., 2020), we initialize the encoder and decoder parameters from BART (Lewis et al., 2020). AWESOME preserves the global context and builds communications across segments with minimal GPU memory increase, by (1) employing external memories in both the encoder and the decoder to gather relevant information ( §3.1), and (2) augmenting the encoder with salient content from other segments ( §3.2).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"External Memory Mechanisms\",\n \"publication_ref\": [\n \"b26\",\n \"b15\",\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"We design two external memory mechanisms to efficiently enable the information flow from prior segments to the current segment. Specifically, each memory module maintains a matrix M ∈ R m×d , where m = 1024 is the memory size and d = 1024 is the hidden state dimension of BART. M is updated after encoding each document segment and then passed to the next segment. We denote the memory matrix after the t-th segment as M t . Each layer of the encoder and decoder can be equipped with one such external memory. Below we describe two mechanisms to update M t and incorporate it in both the encoding and decoding processes. The layer index in the formulas is omitted for simplicity.\\nCompressive Memory. For each document seg-ment, compression-based memory caches its input vectors to be fed into self-attention calculation.\\nSince storing the input vectors as-is requires the memory usage m to scale linearly with the context length, we dedicate half of M t to store the compressed memory, with a compression ratio of r. With H t inp denoting the matrix that contains input vectors to the transformer self-attention, the memory compression and update processes are:\\nM t-1 c , M t-1 u = M t-1 [: m 2 ], M t-1 [ m 2 :](1)\\nM ′ u = concat(M t-1 u , SG(H t inp )) (2) M ′ c = compress(M ′ u [: - m 2 ])(3)\\nM t u = M ′ u [- m 2 :](4)\\nM t c = concat(M t-1 c , M ′ c )[- m 2 :] (5) M t = concat(M t c , M t u )(6)\\nwhere SG(•) denotes stopping the gradient backpropagation to lower GPU usage, and compress(•) performs convolutions with their stride and kernel size set to the compression ratio r. r is set to 5 after tuning on the development sets.\\nNext, to leverage the memory from the previous segment in summarizing the current segment, M t-1 is concatenated with the inputs to the selfattentions to obtain the key-value matrices:\\nH t mem = concat(M t-1 , H t inp )(7)\\nH t self = Attn(H t inp query , H t mem key , H t mem value )(8)\\nwhere H t self is the output of the self-attention. Our compression-based memory is adopted from Compressive Transformer (Rae et al., 2020), a decoder-only model for language modeling. We are the first to apply it to both the encoder and the decoder of a Transformer model and on long document summarization tasks.\\nCompressive memory favors recency, particularly the previous segment and its summary, potentially causing older relevant history to be lost during compression. Attentive Memory. To mitigate the recency bias by compressive memory, we further investigate an attention-based memory updating mechanism, to selectively include content in M t . First, the memory is additionally accompanied by an extra crossattention in each of the encoder and decoder layers, specialized in retrieving relevant information from M t . Following prior study (Lei et al., 2020) that uses memories in video captioning, we update M t with a gate matrix G t to control the amount of content to be updated:\\nM t = G t ⊙ U t + (1 -G t ) ⊙ M t-1 (9)\\nwhere ⊙ denotes the element-wise product and U t is the matrix containing vectors to update the memory. U t and G t are obtained as follows:\\nU t = tanh(W u1 M t-1 + W u2 S t )(10)\\nG t = σ(W g1 M t-1 + W g2 S t )(11)\\nS t = Attn(M t-1 query , SG(H t self ) key , SG(H t self ) value )(12)\\nwhere W * are learnable matrices, S t synthesizes the current segment via an attention calculation, and SG(•) indicates stopping the gradient backpropagation. In each encoder and decoder layer, an extra cross-attention is inserted after the selfattention, where M t-1 is attended and incorporated into the current segment's summarization process. Unlike our approach, the memory in Lei et al. (2020) does not employ gradient stopping. This omission eliminates the memory efficiency gained from the divide-and-conquer strategy, leading to comparable high memory usage as the efficient attention strategy. 3 While their memory is suitable for generating short image captions, our design with gradient stopping is crucial for efficient long document summarization.\\nSelective Addition of External Memory. External memory incurs overhead in GPU memory usage. To mitigate this overhead, we consider selectively adding external memory to specific layers, as the importance of external memory varies according to the different functions of layers in the model. Our pilot study suggests that the last layers of the Transformer model more effectively utilize external memory compared to the first layers. To avoid an exhaustive search for the optimal layer or combination of layers for each dataset, we choose to uniformly equip the last three layers with external memory across all datasets unless otherwise specified. 4\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Global Salient Content Augmentation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The memory mechanisms only grant access to prior content in the documents, yet subsequent context can also help with salience estimation, e.g., elaborating the pros and cons of a proposed solution makes it necessary to introduce the problem and the solution. Moreover, memories store content implicitly, so it is unclear whether relevant information can be stored and retrieved effectively. Therefore, we inform the system of a document's important sentences, which are pre-identified by a separatelytrained extractor. The details of extractor training can be found in Appendix D. After extracting important sentences in a document, we study two methods of injecting them into the summarizer.\\nText Concatenation. For each segment, we include the extracted sentences in the following way to prioritize long-term context. We start with the \\\"outermost\\\" extracted sentences, i.e., the earliest sentence in the past segments and the last sentence in the future segments, and repeat this process until the input has reached the maximum length accepted by the positional encoding of the model (1024 for BART). 5 To differentiate the content in the current segment from the added sentences, we prefix the current segment and the added sentences from before/after the current segment with \\\"Current chunk:\\\", \\\"Previous important sentences:\\\", and \\\"Next important sentences:\\\", respectively. Text concatenation is easy to implement and most compatible with the source modality, but the memory usage increase is quadratic to the length of the augmented content.\\nKey-value Vectors. To circumvent the quadratic memory increase, we join the key-value representations of tokens in important sentences in the encoder self-attentions, and directly inject them into the summarizer encoder. The memory increase is only linear to the augmented content's length.\\nConcretely, the summarizer encoder first encodes all document segments and obtains the representations (i.e., encoder outputs) of tokens belonging to the extracted important sentences. During training, the token representations of these sentences are concatenated with the key-value matrices in the encoder self-attentions while the query matrix remains in its original form. Up to 1024 tokens are concatenated via the same inclusion method for text concatenation, to prioritize the outermost sentences. A similar idea has been used by Memorizing Transformer (Wu et al., 2022a) to include retrieved text representations from past segments for long-form language modeling. Our method differs in two aspects. First, we extract representations from future segments, which are crucial for accurately identifying salient content. Second, we apply a learnable projection to the augmented representations prior to key-value concatenation. This process is crucial in improving compatibility with the original key-value matrices.\\nModel R-1 ↑ R-2 ↑ R-L ↑ Ent Prec ↑ SummaC ↑ Disco ↓ Ent Graph ↑ GPU\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experimental Setups\",\n \"publication_ref\": [\n \"b10\",\n \"b40\",\n \"b3\",\n \"b1\",\n \"b21\",\n \"b18\",\n \"b38\",\n \"b11\",\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"Datasets. We conduct experiments on GovReport (Huang et al., 2021), QMSum (Zhong et al., 2021), SummScreen (Chen et al., 2022) (Bertsch et al., 2023). We also include an extract-then-abstract model (Extract-Abstract) and PageSum (Liu et al., 2022) that leverages dynamic weights, as discussed in §2. All models are initialized from BART-large, except for LongT5 that is pre-trained on long-form data. Details of baseline models are reported in Appendix D.\\nEvaluation Metrics. We evaluate summary informativeness using ROUGE (Lin, 2004). To measure coherence, we use DiscoScore (Zhao et al., 2022) (Disco), a reference-based metric that evaluates discourse coherence by comparing focus (e.g., nouns) frequency and semantics between the system summary and the reference. We also report a graph-based reference-free coherence metric (Guinaudeau and Strube, 2013) (Ent Graph), which measures the connectivity of summary sentences linked by entities, reflecting the coherence of topic transitions. For summary faithfulness, we follow prior work on text generation (Iv et al., 2022) and show the precision of the entities (Ent Prec) in the summary with respect to the document. Additionally, a recent model-based faithfulness metric, SummaC (Laban et al., 2022), is used. Finally, we show the maximum size of allocated GPU memory by each model during training.\\nSe3: VA is required to publish information on appointment wait times at each VA medical facility for primary care, specialty care, and hospital care and medical services, which it does through two public websites. VA has taken a number of actions to address deficiencies GAO found in wait-time measurement and implementation of its scheduling policy. For wait-time measurement, these actions included changes to the wait-time measurement definitions, provision and documentation of scheduler training, and improved oversight through audits, all of which have been in a state of flux for the past 6 years. On July 12, 2019, VA provided GAO additional updates on efforts to implement GAO's related recommendations. AWESOME: GAO recommended that VA either clarify its scheduling policy to better define the desired date, or identify clearer wait-time measures that are not subject to interpretation and prone to scheduler error. VA concurred with the recommendation, which GAO has identified as among those recommendations that warrant priority attention. VA has taken a number of actions to address GAO's recommendations regarding deficiencies GAO found in wait-time measurement and implementation of its scheduling policy. For wait-time measurement, these actions included changes to the wait-time measurement definitions, provision and documentation of scheduler training, and improved oversight through audits, all of which have been in a state of flux for the past 6 years. On July 12, 2019, VA provided GAO additional updates on efforts to implement GAO's related recommendations.\\nTable 3: Summary snippets generated by Se3 and AWE-SOME. AWESOME's summary is more coherent, with natural transitions surrounding \\\"GAO's recommendation\\\", while Se3 abruptly introduces the topic.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results\",\n \"publication_ref\": [\n \"b10\"\n ],\n \"table_ref\": [\n \"tab_2\",\n \"tab_4\",\n \"tab_5\",\n \"tab_6\",\n \"tab_7\"\n ],\n \"text\": \"We report results by all AWESOME variants and comparison models on GovReport in Table 2. Compared with Se3, AWESOME variants consistently achieve better performance on both ROUGE and coherence scores, indicating the importance of maintaining global context for accurate salience estimation of local content and enforcing coherent transitions across segment-level summaries. This can also be demonstrated by the sample outputs in Table 3. Summaries generated by Se3 tend to be shorter, as Se3 fails to plan at a global level. On faithfulness, AWESOME with attentive memory has the best entity precision among all models and also improves SummaC over Se3, while only augmenting AWESOME with global salient content hurts faithfulness. Inspecting the model outputs, we find that using attentive memory improves understanding concepts of long-term dependencies, e.g., connecting a strategy with its related information that appears earlier in the report.\\nOf the two types of external memory mechanisms, attentive memory outperforms compression- based memory on all metrics, which highlights the advantage of adaptively updating the stored context. Meanwhile, directly concatenating salient content with the input yields higher ROUGE scores than injecting key-value vectors into the attention calculation, though the latter is less memory-intensive.\\nModel R-1 ↑ R-2 ↑ R-L ↑ Disco ↓ GPU ↓Se3\\nWe believe natural language-based augmentation better interleaves with the document segment, echoing the findings by prior work on using retrieval for question answering (Wu et al., 2022b). Importantly, under a strict GPU memory constraint, AWESOME with external memory mechanisms and global salient content augmentation achieves the best ROUGE scores among all models, while obtaining competitive results on other measures. Though efficient attention models and Page-Sum can perform remarkably when given highercapacity GPUs as in the original work, they generate less informative summaries when truncation is required to comply with the memory constraint, emphasizing the importance of studying memoryefficient long document summarization models. Furthermore, with selective addition of external memory, AWESOME adds only about 4GB of GPU memory usage, enhancing the model performance efficiently.\\nOn QMSum (Table 4), AWESOME with attention-based memory outperforms all comparisons on ROUGE scores. While our models' summaries are more coherent than the summaries by Se3, as measured by DiscoScore, the differences among all models are less pronounced compared to the ones on GovReport. This is because QM-Sum contains shorter summaries than GovReport (69 vs. 553), thus involving fewer topic transitions.\\nModel R-1 ↑ R-2 ↑ R-L ↑ Ent G ↑ GPU ↓ Se3 38\\nWe also find that the extractor performs poorly on QMSum, leading to degraded results after augmenting our model with the extracted salient content. Specifically, the F1 score of the extractor on the test set is only 1.29, as opposed to 27.85 on GovReport. Compared to our model, the extract-then-abstract model is more prone to its errors and produce summaries of the lowest quality on QMSum. This trend is similarly observed on Summ-Screen (Table 5) and arXiv (Table 6), where the extract-then-abstract method performs poorly and adding extracted content leads to performance drop of AWESOME due to the low performance of the extractor. Meanwhile, AWESOME with the attentive memory is able to obtain the best ROUGE-1 and ROUGE-L scores. On arXiv, models that use efficient attentions obtain the higher ROUGE scores, because truncating arXiv documents has little effect on summary generation-arXiv articles have the most uneven distributions of salient content, where only about 10% of new salient bigrams are located in the second halves of the documents (Huang et al., 2021).\\nFinally, experiments on BookSum show that the divide-and-conquer method produces better summaries for long novels, while our method can further boost its performance (Table 7). However, we find it necessary to incorporate external memory into all layers, suggesting a more complex interaction of external memory with the summarization process for novel plots. Unlike other document types tested, novel plots are typically sequential with less redundancy, which reduces the necessity of the memory mechanism. \\nR-1 ↑ R-2 ↑ R-L ↑ Disco ↓ GPU ↓Se3\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We present AWESOME for summarizing long documents in a memory-constrained setting. Based on the divide-and-conquer strategy, AWESOME uses two mechanisms to gather global context and improve summary quality. First, external memories on the encoder and decoder are employed to track previously read document content and the corresponding summaries. Second, the encoder is informed of global salient content predicted by an extractor via text or representation concatenation. On five summarization datasets, AWESOME generates summaries with better informativeness, faithfulness, and coherence than a baseline divideand-conquer system. Under the same memory constraint, AWESOME outperforms competitive models that leverage efficient attentions or dynamic extraction to preserve global context, highlighting its effectiveness in supplying global context. AWESOME's external memory mechanism is restricted to operating solely from past segments to the current segment. This means that the model does not leverage the information contained in future segments, which can be relevant for a comprehensive understanding of the current segment.\\nTo address this limitation, we have designed the global salient content augmentation mechanism to cover context from the future segments, yet more advanced solutions can be explored in future work. For example, on the encoder, making the external memory bidirectional is a potential approach.\\nWhile being memory-efficient, the external memory mechanism of AWESOME necessitates a longer running time due to its recurrent nature. The need for recurrent computations may lead to increased processing requirements, which could impact real-time applications or scenarios where rapid responses are crucial. The running times of different models are provided in Appendix B.1 for reference. Although our model is slower than that of LongT5 and Se3, it still outperforms several other competitive models in terms of speed, and we will investigate methods for reducing the running time in future work.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ethical Consideration\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We anticipate that one of the major use cases of AWESOME is to allow ordinary users who have computing devices with limited memory to quickly understand government policies and other types of long documents. However, we recognize that the system generated summaries might not comprehensively cover the salient content that is essential for correctly understanding the policies, causing risks ranging from capital loss to legal liability. Moreover, system summaries might contain statements that cannot be verified through the document, which further adds to the risks of real-world deployment. We suggest developers who intend to use our model for real-world application carefully study the outputs by our model before the actual deployment.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Divide-and-Conquer Architecture\",\n \"publication_ref\": [\n \"b23\",\n \"b28\"\n ],\n \"table_ref\": [],\n \"text\": \"We choose Se3 (Moro and Ragazzi, 2022) as our base divide-and-conquer architecture because it can be applied to any document-summary pair. In order to create divide-and-conquer training data for summarization, for each document-summary pair, the document is first divided into segments ( §A.1) and each summary sentence is then assigned to a document segment as part of the generation target ( §A.2). The length of each document segment is between 512 and 768 tokens. During segmentation, the algorithm loops through all document sentences, as shown in Algorithm 1. A document sentence will be added to the current segment if the segment contains less than 512 tokens. The current segment will be finalized if the current segment contains more than 768 tokens or the current sentence is more semantically similar to the next pseudo segment than the current segment, where the next pseudo segment is created by including future sentences until reaching 512 tokens. To measure the similarity between the current sentence and a segment, we use the average cosine similarity between the representation of the current sentence and representations of the sentences in the segment. Sentence representations are obtained using Sentence Transformer (Reimers and Gurevych, 2019) with the all-roberta-large-v1 model.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.1 Document\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.2 Target Assignment\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"For each sentence in the reference summary, we calculate its ROUGE scores with the document segments. The sentence will then be assigned to the document segment with which yields the highest ROUGE-1 and ROUGE-2 scores.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B Additional Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"B.1 Running Time\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We compare the model running time on GovReport (Figure 2). The input document is truncated to 16384 tokens and each model is separately train for 1000 steps with a batch size of 1. No other computation-heavy program is running at the same time. While AWESOME take longer time to complete training than Se3, it is still the third fastest model.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C Dataset Details C.1 Statistics\",\n \"publication_ref\": [\n \"b10\",\n \"b40\",\n \"b5\",\n \"b13\",\n \"b3\",\n \"b1\"\n ],\n \"table_ref\": [\n \"tab_10\"\n ],\n \"text\": \"We conduct experiments on five long document summarization datasets with diverse genres. Gov-Report (Huang et al., 2021) contains long reports and their summaries written by government research agencies. QMSum (Zhong et al., 2021) is a query-focused long meeting transcript summarization dataset, with summary-worthy content spread over the documents. We prepend the query to all segments. We further use a screenplay summarization dataset, SummScreen (Chen et al., 2022), which contains the transcripts of TV series. The TMS subset, with more samples and longer summaries, is selected. Moreover, we experiment with the scientific papers and their abstracts from arXiv (Cohan et al., 2018). Finally, we test our models on summarizing full novels in Book-Sum (Kryscinski et al., 2022). For all datasets, we use the official train/dev/test splits if their original data files are released. Statistics of datasets are reported in Table 8. For GovReport6 , QMSum7 , and SummScreen (Chen et al., 2022), we use the data released by the original papers. For arXiv, we use the version provided by Huggingface Datasets. 8 As the original data files for BookSum are not released due to summary copyright, we use the version reproduced by Unlimiformer (Bertsch et al., 2023).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.2 Input Truncation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In our main experiments, we employ a GPU memory constraint of 27GB. As some baseline models require the input length to be a multiplier of 1024, setting a constraint of 24GB, a more common num-\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"ber, would lead to further truncation and significant performance drop.\\nTo fit models into our memory constraint, we truncate the model inputs. The truncation thresholds used by each model on different datasets are shown in Table 9. Although Se3 and AWESOME theoretically maintain a consistent GPU memory consumption during training regardless of the number of input tokens processed, we have chosen to restrict the maximum number of input tokens in a training sample to 51200 for reasonable training time.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D Implementation Details\",\n \"publication_ref\": [\n \"b24\",\n \"b0\",\n \"b9\",\n \"b1\",\n \"b21\",\n \"b22\",\n \"b18\",\n \"b38\",\n \"b8\",\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"Baselines. BlockAttn and Longformer use blockwise attentions (Phang et al., 2022) and slidingwindow attentions (Beltagy et al., 2020), where a global token can attend to and be attended by all tokens, while other tokens can only attend to tokens in the same block or window. LongT5 (Guo et al., 2022) is a sliding-window attention model pre-trained on long sequences, and Unlimiformer (Bertsch et al., 2023) extends BART by selecting input tokens to be attended to via KNN searching. For the extract-then-abstract approach, we use the same extractor as in the global salient content augmentation of our model, and the abstractor takes as input oracle extracted sentences during training. Lastly, PageSum (Liu et al., 2022) synthesizes the output representations given by different document segments with dynamic weights.\\nExtractor. The extractor first uses a RoBERTa (Liu et al., 2019) to encode each sentence and takes the average of the final layer's outputs as the sentence representation. It then applies a self-attention on top of all sentence representations. The resulting representations are converted to extraction scores after applying a multi-layer perception with one hidden layer. The extractor is trained with oracle extractive labels that are constructed by greedily searching for document sentences that maximize the sum of ROUGE-1 and ROUGE-2 scores, compared against the reference summary. We do not compute ROUGE-L as in DYLE (Mao et al., 2022), because finding the longest common subsequence is computationally expensive and does not yield performance gain.\\nTraining Parameters. We train all models with a maximum learning rate of 5 × 10 -5 , except that LongT5 is trained with a maximum learning rate of 1 × 10 -4 . We use a running batch size of 1 and apply gradient accumulation to achieve an effective batch size of 8. The numbers of training epochs are 3, 9, 6, 2, 10 on GovReport, QMSum, SummScreen, arXiv, and BookSum, with warmup steps of 300, 100, 300, 1000, and 40. Due to the computational cost of training long document summarization, each model is trained for a single run.\\nModel Size. AWESOME is based on BART-large 9 and has 708 millions of parameters.\\nComputing Infrastructure. All experiments are conducted on RTX A6000 GPUs.\\nEvaluation Metrics. For ROUGE (Lin, 2004), we use the Python implementation by Google. 10 The official code for DiscoScore (Zhao et al., 2022) is used 11 , which also provides an implementation of the Ent Graph metric (Guinaudeau and Strube, 2013). We implement the entity precision measure ourselves and run the official code for Sum-maC (Laban et al., 2022). 12 All metrics used are open-source and can be distributed for research purposes.\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":""},"doi":{"kind":"string","value":"10.18653/v1/2022.acl-long.589"},"references":{"kind":"list like","value":[{"authors":"Iz Beltagy; Matthew E Peters; Arman Cohan","journal":"","ref_id":"b0","title":"Longformer: The long-document transformer","year":"2020"},{"authors":"Amanda Bertsch; Uri Alon; Graham Neubig; Matthew R Gormley","journal":"","ref_id":"b1","title":"Unlimiformer: Longrange transformers with unlimited length input","year":"2023"},{"authors":"Aydar Bulatov; Yuri Kuratov; Mikhail Burtsev","journal":"","ref_id":"b2","title":"Recurrent memory transformer","year":"2022"},{"authors":"Mingda Chen; Zewei Chu; Sam Wiseman; Kevin Gimpel","journal":"Association for Computational Linguistics","ref_id":"b3","title":"SummScreen: A dataset for abstractive screenplay summarization","year":"2022"},{"authors":"Tianqi Chen; Bing Xu; Chiyuan Zhang; Carlos Guestrin","journal":"","ref_id":"b4","title":"Training deep nets with sublinear memory cost","year":"2016"},{"authors":"Arman Cohan; Franck Dernoncourt; Soon Doo; Trung Kim; Seokhwan Bui; Walter Kim; Nazli Chang; Goharian","journal":"Association for Computational Linguistics","ref_id":"b5","title":"A discourse-aware attention model for abstractive summarization of long documents","year":"2018"},{"authors":"Zihang Dai; Zhilin Yang; Yiming Yang; Jaime Carbonell; Quoc Le; Ruslan Salakhutdinov","journal":"Association for Computational Linguistics","ref_id":"b6","title":"Transformer-XL: Attentive language models beyond a fixed-length context","year":"2019"},{"authors":"Alexios Gidiotis; Grigorios Tsoumakas","journal":"IEEE/ACM Transactions on Audio, Speech, and Language Processing","ref_id":"b7","title":"A divide-and-conquer approach to the summarization of long documents","year":"2020"},{"authors":"Camille Guinaudeau; Michael Strube","journal":"Association for Computational Linguistics","ref_id":"b8","title":"Graphbased local coherence modeling","year":"2013"},{"authors":"Mandy Guo; Joshua Ainslie; David Uthus; Santiago Ontanon; Jianmo Ni; Yun-Hsuan Sung; Yinfei Yang","journal":"Association for Computational Linguistics","ref_id":"b9","title":"LongT5: Efficient text-to-text transformer for long sequences","year":"2022"},{"authors":"Luyang Huang; Shuyang Cao; Nikolaus Parulian; Ji Heng; Lu Wang","journal":"Association for Computational Linguistics","ref_id":"b10","title":"Efficient attentions for long document summarization","year":"2021"},{"authors":"Robert Iv; Alexandre Passos; Sameer Singh; Ming-Wei Chang","journal":"Association for Computational Linguistics","ref_id":"b11","title":"FRUIT: Faithfully reflecting updated information in text","year":"2022"},{"authors":"Nikita Kitaev; Lukasz Kaiser; Anselm Levskaya","journal":"","ref_id":"b12","title":"Reformer: The efficient transformer","year":"2020"},{"authors":"Wojciech Kryscinski; Nazneen Rajani; Divyansh Agarwal; Caiming Xiong; Dragomir Radev","journal":"Association for Computational Linguistics","ref_id":"b13","title":"BOOKSUM: A collection of datasets for long-form narrative summarization","year":"2022"},{"authors":"Philippe Laban; Tobias Schnabel; Paul N Bennett; Marti A Hearst","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b14","title":"SummaC: Re-visiting NLIbased models for inconsistency detection in summarization","year":"2022"},{"authors":"Jie Lei; Liwei Wang; Yelong Shen; Dong Yu; Tamara Berg; Mohit Bansal","journal":"Association for Computational Linguistics","ref_id":"b15","title":"MART: Memoryaugmented recurrent transformer for coherent video paragraph captioning","year":"2020"},{"authors":"Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer","journal":"Association for Computational Linguistics","ref_id":"b16","title":"BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension","year":"2020"},{"authors":"Chuan Li","journal":"","ref_id":"b17","title":"Best gpu for deep learning in 2022","year":"2022"},{"authors":"Chin-Yew Lin","journal":"Association for Computational Linguistics","ref_id":"b18","title":"ROUGE: A package for automatic evaluation of summaries","year":"2004"},{"authors":"Yang Liu; Mirella Lapata","journal":"Association for Computational Linguistics","ref_id":"b19","title":"Hierarchical transformers for multi-document summarization","year":"2019"},{"authors":"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b20","title":"Roberta: A robustly optimized bert pretraining approach","year":"2019"},{"authors":"Yixin Liu; Ansong Ni; Linyong Nan; Budhaditya Deb; Chenguang Zhu; Ahmed H Awadallah; Dragomir Radev","journal":"","ref_id":"b21","title":"Leveraging locality in abstractive text summarization","year":"2022"},{"authors":"Ziming Mao; Chen Henry Wu; Ansong Ni; Yusen Zhang; Rui Zhang; Tao Yu; Budhaditya Deb; Chenguang Zhu; Ahmed Awadallah; Dragomir Radev","journal":"Association for Computational Linguistics","ref_id":"b22","title":"DYLE: Dynamic latent extraction for abstractive long-input summarization","year":"2022"},{"authors":"Gianluca Moro; Luca Ragazzi","journal":"OpenAI","ref_id":"b23","title":"Semantic selfsegmentation for abstractive summarization of long documents in low-resource regimes","year":"2022"},{"authors":"Jason Phang; Yao Zhao; Peter J Liu","journal":"","ref_id":"b24","title":"Investigating efficiently extending transformers for long input summarization","year":"2022"},{"authors":"Jonathan Pilault; Raymond Li; Sandeep Subramanian; Chris Pal","journal":"","ref_id":"b25","title":"On extractive and abstractive neural document summarization with transformer language models","year":"2020"},{"authors":"Jack W Rae; Anna Potapenko; M Siddhant; Chloe Jayakumar; Timothy P Hillier; Lillicrap","journal":"","ref_id":"b26","title":"Compressive transformers for long-range sequence modelling","year":"2020"},{"authors":"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu","journal":"Journal of Machine Learning Research","ref_id":"b27","title":"Exploring the limits of transfer learning with a unified text-to-text transformer","year":"2020"},{"authors":"Nils Reimers; Iryna Gurevych","journal":"","ref_id":"b28","title":"Sentence-bert: Sentence embeddings using siamese bert-networks","year":"2019"},{"authors":"Aurko Roy; Mohammad Saffar; Ashish Vaswani; David Grangier","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b29","title":"Efficient content-based sparse attention with routing transformers","year":"2021"},{"authors":"Yi Tay; Dara Bahri; Liu Yang; Donald Metzler; Da-Cheng Juan","journal":"","ref_id":"b30","title":"Sparse sinkhorn attention","year":"2020"},{"authors":" Pmlr","journal":"","ref_id":"b31","title":"","year":""},{"authors":"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin","journal":"Advances in neural information processing systems","ref_id":"b32","title":"Attention is all you need","year":"2017"},{"authors":"Jeff Wu; Long Ouyang; M Daniel; Nisan Ziegler; Ryan Stiennon; Jan Lowe; Paul Leike; Christiano","journal":"","ref_id":"b33","title":"Recursively summarizing books with human feedback","year":"2021"},{"authors":"Yuhuai Wu; Markus Norman Rabe; Delesley Hutchins; Christian Szegedy","journal":"","ref_id":"b34","title":"Memorizing transformers","year":"2022"},{"authors":"Yuxiang Wu; Yu Zhao; Baotian Hu; Pasquale Minervini; Pontus Stenetorp; Sebastian Riedel","journal":"","ref_id":"b35","title":"An efficient memory-augmented transformer for knowledge-intensive nlp tasks","year":"2022"},{"authors":"Manzil Zaheer; Guru Guruganesh; Avinava Dubey; Joshua Ainslie; Chris Alberti; Santiago Ontanon; Philip Pham; Anirudh Ravula; Qifan Wang; Li Yang; Amr Ahmed","journal":"","ref_id":"b36","title":"Big bird: Transformers for longer sequences","year":"2021"},{"authors":"Yusen Zhang; Ansong Ni; Ziming Mao; Chen Henry Wu; Chenguang Zhu; Budhaditya Deb; Ahmed Awadallah; Dragomir Radev; Rui Zhang","journal":"Association for Computational Linguistics","ref_id":"b37","title":"Summ n : A multi-stage summarization framework for long input dialogues and documents","year":"2022"},{"authors":"Wei Zhao; Michael Strube; Steffen Eger","journal":"","ref_id":"b38","title":"Discoscore: Evaluating text generation with bert and discourse coherence","year":"2022"},{"authors":"Yao Zhao; Mohammad Saleh; Peter J Liu","journal":"","ref_id":"b39","title":"Seal: Segment-wise extractive-abstractive long-form text summarization","year":"2020"},{"authors":"Ming Zhong; Da Yin; Tao Yu; Ahmad Zaidi; Mutethia Mutuma; Rahul Jha; Ahmed Hassan Awadallah; Asli Celikyilmaz; Yang Liu; Xipeng Qiu; Dragomir Radev","journal":"Association for Computational Linguistics","ref_id":"b40","title":"QMSum: A new benchmark for querybased multi-domain meeting summarization","year":"2021"}],"string":"[\n {\n \"authors\": \"Iz Beltagy; Matthew E Peters; Arman Cohan\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Longformer: The long-document transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Amanda Bertsch; Uri Alon; Graham Neubig; Matthew R Gormley\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Unlimiformer: Longrange transformers with unlimited length input\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Aydar Bulatov; Yuri Kuratov; Mikhail Burtsev\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Recurrent memory transformer\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Mingda Chen; Zewei Chu; Sam Wiseman; Kevin Gimpel\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b3\",\n \"title\": \"SummScreen: A dataset for abstractive screenplay summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tianqi Chen; Bing Xu; Chiyuan Zhang; Carlos Guestrin\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Training deep nets with sublinear memory cost\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Arman Cohan; Franck Dernoncourt; Soon Doo; Trung Kim; Seokhwan Bui; Walter Kim; Nazli Chang; Goharian\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b5\",\n \"title\": \"A discourse-aware attention model for abstractive summarization of long documents\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Zihang Dai; Zhilin Yang; Yiming Yang; Jaime Carbonell; Quoc Le; Ruslan Salakhutdinov\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b6\",\n \"title\": \"Transformer-XL: Attentive language models beyond a fixed-length context\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Alexios Gidiotis; Grigorios Tsoumakas\",\n \"journal\": \"IEEE/ACM Transactions on Audio, Speech, and Language Processing\",\n \"ref_id\": \"b7\",\n \"title\": \"A divide-and-conquer approach to the summarization of long documents\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Camille Guinaudeau; Michael Strube\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b8\",\n \"title\": \"Graphbased local coherence modeling\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Mandy Guo; Joshua Ainslie; David Uthus; Santiago Ontanon; Jianmo Ni; Yun-Hsuan Sung; Yinfei Yang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b9\",\n \"title\": \"LongT5: Efficient text-to-text transformer for long sequences\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Luyang Huang; Shuyang Cao; Nikolaus Parulian; Ji Heng; Lu Wang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b10\",\n \"title\": \"Efficient attentions for long document summarization\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Robert Iv; Alexandre Passos; Sameer Singh; Ming-Wei Chang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b11\",\n \"title\": \"FRUIT: Faithfully reflecting updated information in text\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Nikita Kitaev; Lukasz Kaiser; Anselm Levskaya\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Reformer: The efficient transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Wojciech Kryscinski; Nazneen Rajani; Divyansh Agarwal; Caiming Xiong; Dragomir Radev\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b13\",\n \"title\": \"BOOKSUM: A collection of datasets for long-form narrative summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Philippe Laban; Tobias Schnabel; Paul N Bennett; Marti A Hearst\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b14\",\n \"title\": \"SummaC: Re-visiting NLIbased models for inconsistency detection in summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jie Lei; Liwei Wang; Yelong Shen; Dong Yu; Tamara Berg; Mohit Bansal\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b15\",\n \"title\": \"MART: Memoryaugmented recurrent transformer for coherent video paragraph captioning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Veselin Stoyanov; Luke Zettlemoyer\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b16\",\n \"title\": \"BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Chuan Li\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Best gpu for deep learning in 2022\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Chin-Yew Lin\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b18\",\n \"title\": \"ROUGE: A package for automatic evaluation of summaries\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Yang Liu; Mirella Lapata\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b19\",\n \"title\": \"Hierarchical transformers for multi-document summarization\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Roberta: A robustly optimized bert pretraining approach\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yixin Liu; Ansong Ni; Linyong Nan; Budhaditya Deb; Chenguang Zhu; Ahmed H Awadallah; Dragomir Radev\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Leveraging locality in abstractive text summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ziming Mao; Chen Henry Wu; Ansong Ni; Yusen Zhang; Rui Zhang; Tao Yu; Budhaditya Deb; Chenguang Zhu; Ahmed Awadallah; Dragomir Radev\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b22\",\n \"title\": \"DYLE: Dynamic latent extraction for abstractive long-input summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Gianluca Moro; Luca Ragazzi\",\n \"journal\": \"OpenAI\",\n \"ref_id\": \"b23\",\n \"title\": \"Semantic selfsegmentation for abstractive summarization of long documents in low-resource regimes\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jason Phang; Yao Zhao; Peter J Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Investigating efficiently extending transformers for long input summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jonathan Pilault; Raymond Li; Sandeep Subramanian; Chris Pal\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"On extractive and abstractive neural document summarization with transformer language models\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jack W Rae; Anna Potapenko; M Siddhant; Chloe Jayakumar; Timothy P Hillier; Lillicrap\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Compressive transformers for long-range sequence modelling\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu\",\n \"journal\": \"Journal of Machine Learning Research\",\n \"ref_id\": \"b27\",\n \"title\": \"Exploring the limits of transfer learning with a unified text-to-text transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Nils Reimers; Iryna Gurevych\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Sentence-bert: Sentence embeddings using siamese bert-networks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Aurko Roy; Mohammad Saffar; Ashish Vaswani; David Grangier\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b29\",\n \"title\": \"Efficient content-based sparse attention with routing transformers\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yi Tay; Dara Bahri; Liu Yang; Donald Metzler; Da-Cheng Juan\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Sparse sinkhorn attention\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \" Pmlr\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b32\",\n \"title\": \"Attention is all you need\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Jeff Wu; Long Ouyang; M Daniel; Nisan Ziegler; Ryan Stiennon; Jan Lowe; Paul Leike; Christiano\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Recursively summarizing books with human feedback\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yuhuai Wu; Markus Norman Rabe; Delesley Hutchins; Christian Szegedy\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Memorizing transformers\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yuxiang Wu; Yu Zhao; Baotian Hu; Pasquale Minervini; Pontus Stenetorp; Sebastian Riedel\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"An efficient memory-augmented transformer for knowledge-intensive nlp tasks\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Manzil Zaheer; Guru Guruganesh; Avinava Dubey; Joshua Ainslie; Chris Alberti; Santiago Ontanon; Philip Pham; Anirudh Ravula; Qifan Wang; Li Yang; Amr Ahmed\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Big bird: Transformers for longer sequences\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yusen Zhang; Ansong Ni; Ziming Mao; Chen Henry Wu; Chenguang Zhu; Budhaditya Deb; Ahmed Awadallah; Dragomir Radev; Rui Zhang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b37\",\n \"title\": \"Summ n : A multi-stage summarization framework for long input dialogues and documents\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Wei Zhao; Michael Strube; Steffen Eger\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Discoscore: Evaluating text generation with bert and discourse coherence\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yao Zhao; Mohammad Saleh; Peter J Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Seal: Segment-wise extractive-abstractive long-form text summarization\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ming Zhong; Da Yin; Tao Yu; Ahmad Zaidi; Mutethia Mutuma; Rahul Jha; Ahmed Hassan Awadallah; Asli Celikyilmaz; Yang Liu; Xipeng Qiu; Dragomir Radev\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b40\",\n \"title\": \"QMSum: A new benchmark for querybased multi-domain meeting summarization\",\n \"year\": \"2021\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[2,315.64,75.72,199.27,71.08],"formula_id":"formula_0","formula_text":"Approach In→Out Enc Enc←Dec Dec Efficient Attention x → y ■ ■ Extract-Abstract xe → y □ ⋆ Dynamic Weight x → y □ ■ + ⋆ Divide-Conquer xi → yi □ □"},{"formula_coordinates":[4,327.69,211.73,197.39,22.31],"formula_id":"formula_1","formula_text":"M t-1 c , M t-1 u = M t-1 [: m 2 ], M t-1 [ m 2 :](1)"},{"formula_coordinates":[4,366.16,235.82,158.92,36.48],"formula_id":"formula_2","formula_text":"M ′ u = concat(M t-1 u , SG(H t inp )) (2) M ′ c = compress(M ′ u [: - m 2 ])(3)"},{"formula_coordinates":[4,366.16,271.83,158.92,22.31],"formula_id":"formula_3","formula_text":"M t u = M ′ u [- m 2 :](4)"},{"formula_coordinates":[4,366.78,293.68,158.3,36.58],"formula_id":"formula_4","formula_text":"M t c = concat(M t-1 c , M ′ c )[- m 2 :] (5) M t = concat(M t c , M t u )(6)"},{"formula_coordinates":[4,339.73,485.71,185.35,12.69],"formula_id":"formula_5","formula_text":"H t mem = concat(M t-1 , H t inp )(7)"},{"formula_coordinates":[4,342.88,501.91,182.2,27.2],"formula_id":"formula_6","formula_text":"H t self = Attn(H t inp query , H t mem key , H t mem value )(8)"},{"formula_coordinates":[5,106.24,152.95,183.56,11.03],"formula_id":"formula_7","formula_text":"M t = G t ⊙ U t + (1 -G t ) ⊙ M t-1 (9)"},{"formula_coordinates":[5,84.44,238.5,205.36,11.72],"formula_id":"formula_8","formula_text":"U t = tanh(W u1 M t-1 + W u2 S t )(10)"},{"formula_coordinates":[5,84.5,254.68,205.3,11.72],"formula_id":"formula_9","formula_text":"G t = σ(W g1 M t-1 + W g2 S t )(11)"},{"formula_coordinates":[5,85.65,270.86,204.15,27.73],"formula_id":"formula_10","formula_text":"S t = Attn(M t-1 query , SG(H t self ) key , SG(H t self ) value )(12)"},{"formula_coordinates":[6,78.53,75.72,409.21,8.06],"formula_id":"formula_11","formula_text":"Model R-1 ↑ R-2 ↑ R-L ↑ Ent Prec ↑ SummaC ↑ Disco ↓ Ent Graph ↑ GPU"},{"formula_coordinates":[7,309.24,75.72,212.08,23.79],"formula_id":"formula_12","formula_text":"Model R-1 ↑ R-2 ↑ R-L ↑ Disco ↓ GPU ↓Se3"},{"formula_coordinates":[8,73.86,75.72,217.25,23.79],"formula_id":"formula_13","formula_text":"Model R-1 ↑ R-2 ↑ R-L ↑ Ent G ↑ GPU ↓ Se3 38"},{"formula_coordinates":[8,309.13,75.72,215,23.79],"formula_id":"formula_14","formula_text":"R-1 ↑ R-2 ↑ R-L ↑ Disco ↓ GPU ↓Se3"}],"string":"[\n {\n \"formula_coordinates\": [\n 2,\n 315.64,\n 75.72,\n 199.27,\n 71.08\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"Approach In→Out Enc Enc←Dec Dec Efficient Attention x → y ■ ■ Extract-Abstract xe → y □ ⋆ Dynamic Weight x → y □ ■ + ⋆ Divide-Conquer xi → yi □ □\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 327.69,\n 211.73,\n 197.39,\n 22.31\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"M t-1 c , M t-1 u = M t-1 [: m 2 ], M t-1 [ m 2 :](1)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 366.16,\n 235.82,\n 158.92,\n 36.48\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"M ′ u = concat(M t-1 u , SG(H t inp )) (2) M ′ c = compress(M ′ u [: - m 2 ])(3)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 366.16,\n 271.83,\n 158.92,\n 22.31\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"M t u = M ′ u [- m 2 :](4)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 366.78,\n 293.68,\n 158.3,\n 36.58\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"M t c = concat(M t-1 c , M ′ c )[- m 2 :] (5) M t = concat(M t c , M t u )(6)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 339.73,\n 485.71,\n 185.35,\n 12.69\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"H t mem = concat(M t-1 , H t inp )(7)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 342.88,\n 501.91,\n 182.2,\n 27.2\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"H t self = Attn(H t inp query , H t mem key , H t mem value )(8)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 106.24,\n 152.95,\n 183.56,\n 11.03\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"M t = G t ⊙ U t + (1 -G t ) ⊙ M t-1 (9)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 84.44,\n 238.5,\n 205.36,\n 11.72\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"U t = tanh(W u1 M t-1 + W u2 S t )(10)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 84.5,\n 254.68,\n 205.3,\n 11.72\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"G t = σ(W g1 M t-1 + W g2 S t )(11)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 85.65,\n 270.86,\n 204.15,\n 27.73\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"S t = Attn(M t-1 query , SG(H t self ) key , SG(H t self ) value )(12)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 78.53,\n 75.72,\n 409.21,\n 8.06\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"Model R-1 ↑ R-2 ↑ R-L ↑ Ent Prec ↑ SummaC ↑ Disco ↓ Ent Graph ↑ GPU\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 309.24,\n 75.72,\n 212.08,\n 23.79\n ],\n \"formula_id\": \"formula_12\",\n \"formula_text\": \"Model R-1 ↑ R-2 ↑ R-L ↑ Disco ↓ GPU ↓Se3\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 73.86,\n 75.72,\n 217.25,\n 23.79\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"Model R-1 ↑ R-2 ↑ R-L ↑ Ent G ↑ GPU ↓ Se3 38\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 309.13,\n 75.72,\n 215,\n 23.79\n ],\n \"formula_id\": \"formula_14\",\n \"formula_text\": \"R-1 ↑ R-2 ↑ R-L ↑ Disco ↓ GPU ↓Se3\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"AWESOME: GPU Memory-constrained Long Document Summarization using Memory Mechanism and Global Salient Content"},"abstract":{"kind":"string","value":"Long document summarization systems are critical for domains with lengthy and jargonladen text, yet they present significant challenges to researchers and developers with limited computing resources. Existing solutions mainly focus on efficient attentions or divideand-conquer strategies. The former reduces theoretical time complexity, but is still memoryheavy. The latter methods sacrifice global context, leading to uninformative and incoherent summaries. This work aims to leverage the memory-efficient nature of divide-and-conquer methods while preserving global context. Concretely, our framework AWESOME uses two novel mechanisms: (1) External memory mechanisms track previously encoded document segments and their corresponding summaries, to enhance global document understanding and summary coherence. (2) Global salient content is further identified beforehand to augment each document segment to support its summarization. Extensive experiments on diverse genres of text, including government reports, meeting transcripts, screenplays, scientific papers, and novels, show that AWESOME produces summaries with improved informativeness, faithfulness, and coherence than competitive baselines on longer documents, while having a smaller GPU memory footprint."},"authors":{"kind":"string","value":"Shuyang Cao; Lu Wang"},"figures":{"kind":"list like","value":[{"figure_caption":"Model","figure_data":"","figure_id":"fig_0","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Running time (batch per second) of each model. A higher number of batches processed per second indicates a faster running speed. All models use a batch size of 1 and the input is truncated to 16384 tokens.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Existing approaches to long document summarization ( §2.1). In→Out: Longer inputs (|x| > |x e | > |x i |) or outputs (|y| > |y i |) produce more nodes in the","figure_data":"computation graph, thus the higher memory consump-tion. Enc: Encoder accessing partial documents (□)hurts document understanding, compared to reading thefull text (■). Enc←Dec: Decoder reading the full docu-ment (■) or pre-identified salient content (⋆) enhancessummary informativeness, compared to a segment (□).Dec: Decoder accessing previously generated summarycontent ( ) is crucial for generation coherence thanreading a current summary segment only ( ).five datasets, AWESOME produces more informa-tive and coherence outputs than recent models.2 Related Worksummary context used by each approach whengenerating summaries. Specifically, we check (1)full vs. partial documents that are consumed toobtain the encoder representations (Enc); (2) fullvs. partial encoder representations that are attendedby the decoder (Enc←Dec); and (3) full vs. partialoutput that is accessed by the decoder (Dec).","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"Mem ↓ Results on GovReport. The best and second best results per metric are bolded and underlined. Results by AWESOME variants that are better than all comparisons and Se3 are shaded with green and blue, respectively. AWESOME with attentive memory only and its full version that additionally uses salient content through text concatenation obtain the highest ROUGE scores (in green) and are comparable or better on faithfulness (Ent Prec & SummaC) and coherence (Disco & Ent Graph) than base model Se3.","figure_data":"Se346.56 23.22 44.36 98.2414.717.371.4111.1BlockAttn57.46 26.78 54.82 97.4520.435.912.0525.6Longformer57.40 26.92 54.70 97.5220.395.682.0525.3LongT554.21 24.87 51.06 96.4113.344.811.5625.4Unlimiformer56.35 25.94 53.83 92.196.055.361.9627.0Extract-Abstract56.89 24.76 54.26 92.8222.074.032.0913.2PageSum56.80 23.26 54.11 89.566.823.041.8824.9AWESOME using External Memory OnlyCompressive50.71 † 23.91 48.45 † 89.1715.345.16 †1.94 †12.5Attentive (Attn)58.44 * 27.71 * 55.98 * 98.3318.98 †3.62 †1.98 †14.0AWESOME using Global Salient Content OnlyText-concat (Txt)56.65 † 27.68 * 54.11 † 97.9312.235.05 †2.09 †12.0Key-value Vectors55.02 † 26.39 † 52.41 † 98.2211.524.75 †1.75 †14.3AWESOME (Attn + Txt) 58.76 * 28.18 * 56.05 * 98.3119.22 †3.86 †2.03 †14.8","figure_id":"tab_2","figure_label":"2","figure_type":"table"},{"figure_caption":"Results on meeting transcripts in QMSum. Equipped with attentive memory only, AWESOME achieves the best ROUGE scores. Though better than some baselines, adding extracted salient content does not further boost the performance, due to the low performance of the extractor on dialog data.","figure_data":"29.28 10.51 25.93 0.778.1BlockAttn30.768.26 26.49 0.5022.8Longformer29.187.82 24.94 3.0726.5LongT531.88 10.07 27.82 0.4425.4Unlimiformer30.578.82 26.89 0.4926.9Extract-Abstract 17.635.65 16.02 4.0210.3PageSum29.557.38 26.11 0.3121.5AWESOMEAttn Only34.86 † 12.69 31.09 * 0.6812.9Attn + Txt31.16 † 10.11 27.66 0.6913.3","figure_id":"tab_4","figure_label":"4","figure_type":"table"},{"figure_caption":"Results on TV transcripts in SummScreen. We report Ent Graph instead of DiscoScore, as DiscoScore encounters errors when identifying focus. AWESOME with the attentive memory obtains the best R1 and RL scores, while the low accuracy of the extracted salient content leads to performance drop of the summarizer.","figure_data":".09 11.30 36.56 0.5011.3BlockAttn32.018.99 30.90 1.6125.7Longformer42.78 13.21 41.34 0.9725.3LongT542.03 12.67 40.76 1.0325.4Unlimiformer35.17 11.98 34.28 1.3327.0Extract-Abstract 19.955.58 19.70 0.0613.1AWESOMEAttn Only46.05 † 13.09 † 44.21 † 0.81 †13.2Attn + Txt45.30 † 12.63 † 43.51 † 0.90 †14.2","figure_id":"tab_5","figure_label":"5","figure_type":"table"},{"figure_caption":"Results on arXiv papers. AWESOME variants again outperform Se3. For 80% of the arXiv documents, efficient attention models and PageSum can fully train on their first halves, covering 90% of the salient content that appear in the references(Huang et al., 2021), thus the better ROUGE scores than models encoding smaller segments.","figure_data":"40.74 17.96 36.87 1.3312.8BlockAttn49.12 21.69 44.40 1.7725.7Longformer48.59 21.45 43.99 2.1725.2LongT548.25 20.74 43.41 0.9725.5Unlimiformer47.78 20.58 43.22 1.2226.8Extract-Abstract 42.37 16.43 38.62 1.0315.3PageSum46.01 18.77 41.55 0.8826.2AWESOMEAttn Only42.51 † 18.96 † 38.56 † 1.3016.0Attn + Txt44.20 † 18.89 † 40.07 † 1.3216.5ModelR-1 ↑ R-2 ↑ R-L ↑ Disco ↓ GPU ↓Se340.78 10.16 39.77 10.46 11.5BlockAttn23.45 3.09 22.09 190.27 25.7Longformer20.20 2.45 18.55 204.48 25.3LongT533.15 6.74 32.62 24.24 25.5Unlimiformer38.09 9.55 37.41 47.72 27.0AWESOME (Attn) 41.11 10.63 40.20 10.36 24.0","figure_id":"tab_6","figure_label":"6","figure_type":"table"},{"figure_caption":"Results on novels in BookSum. AWESOME with attentive memory in all layers achieves the best performance on all metrics. Methods requiring external extractors are not included due to the computational cost of building extractive oracles for long novels.","figure_data":"","figure_id":"tab_7","figure_label":"7","figure_type":"table"},{"figure_caption":"Statistics of datasets used in our experiments.","figure_data":"DatasetModelGov arXiv QMSum SumScrn BookSe350x50x50x50x50xExt-Abs †1x (∞) 1x (∞) 1x (∞) 1x (∞)-BlockAttn6x6x8x6x6xLongformer8x8x8x8x8xLongT56x6x6x6x6xUnlimiformer 2x2x2x2x2xPageSum3x5x2x--AWESOME 50x50x50x50x50x","figure_id":"tab_10","figure_label":"8","figure_type":"table"},{"figure_caption":"Truncation thresholds (multiply by 1024) used by each model on different datasets to comply with the memory constraint during training. †: For the extractthen-abstract model, the abstractor has a maximum input length of 1024, while the extractor can consume all sentences in the document.","figure_data":"","figure_id":"tab_11","figure_label":"9","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Model\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Running time (batch per second) of each model. A higher number of batches processed per second indicates a faster running speed. All models use a batch size of 1 and the input is truncated to 16384 tokens.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Existing approaches to long document summarization ( §2.1). In→Out: Longer inputs (|x| > |x e | > |x i |) or outputs (|y| > |y i |) produce more nodes in the\",\n \"figure_data\": \"computation graph, thus the higher memory consump-tion. Enc: Encoder accessing partial documents (□)hurts document understanding, compared to reading thefull text (■). Enc←Dec: Decoder reading the full docu-ment (■) or pre-identified salient content (⋆) enhancessummary informativeness, compared to a segment (□).Dec: Decoder accessing previously generated summarycontent ( ) is crucial for generation coherence thanreading a current summary segment only ( ).five datasets, AWESOME produces more informa-tive and coherence outputs than recent models.2 Related Worksummary context used by each approach whengenerating summaries. Specifically, we check (1)full vs. partial documents that are consumed toobtain the encoder representations (Enc); (2) fullvs. partial encoder representations that are attendedby the decoder (Enc←Dec); and (3) full vs. partialoutput that is accessed by the decoder (Dec).\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Mem ↓ Results on GovReport. The best and second best results per metric are bolded and underlined. Results by AWESOME variants that are better than all comparisons and Se3 are shaded with green and blue, respectively. AWESOME with attentive memory only and its full version that additionally uses salient content through text concatenation obtain the highest ROUGE scores (in green) and are comparable or better on faithfulness (Ent Prec & SummaC) and coherence (Disco & Ent Graph) than base model Se3.\",\n \"figure_data\": \"Se346.56 23.22 44.36 98.2414.717.371.4111.1BlockAttn57.46 26.78 54.82 97.4520.435.912.0525.6Longformer57.40 26.92 54.70 97.5220.395.682.0525.3LongT554.21 24.87 51.06 96.4113.344.811.5625.4Unlimiformer56.35 25.94 53.83 92.196.055.361.9627.0Extract-Abstract56.89 24.76 54.26 92.8222.074.032.0913.2PageSum56.80 23.26 54.11 89.566.823.041.8824.9AWESOME using External Memory OnlyCompressive50.71 † 23.91 48.45 † 89.1715.345.16 †1.94 †12.5Attentive (Attn)58.44 * 27.71 * 55.98 * 98.3318.98 †3.62 †1.98 †14.0AWESOME using Global Salient Content OnlyText-concat (Txt)56.65 † 27.68 * 54.11 † 97.9312.235.05 †2.09 †12.0Key-value Vectors55.02 † 26.39 † 52.41 † 98.2211.524.75 †1.75 †14.3AWESOME (Attn + Txt) 58.76 * 28.18 * 56.05 * 98.3119.22 †3.86 †2.03 †14.8\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Results on meeting transcripts in QMSum. Equipped with attentive memory only, AWESOME achieves the best ROUGE scores. Though better than some baselines, adding extracted salient content does not further boost the performance, due to the low performance of the extractor on dialog data.\",\n \"figure_data\": \"29.28 10.51 25.93 0.778.1BlockAttn30.768.26 26.49 0.5022.8Longformer29.187.82 24.94 3.0726.5LongT531.88 10.07 27.82 0.4425.4Unlimiformer30.578.82 26.89 0.4926.9Extract-Abstract 17.635.65 16.02 4.0210.3PageSum29.557.38 26.11 0.3121.5AWESOMEAttn Only34.86 † 12.69 31.09 * 0.6812.9Attn + Txt31.16 † 10.11 27.66 0.6913.3\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Results on TV transcripts in SummScreen. We report Ent Graph instead of DiscoScore, as DiscoScore encounters errors when identifying focus. AWESOME with the attentive memory obtains the best R1 and RL scores, while the low accuracy of the extracted salient content leads to performance drop of the summarizer.\",\n \"figure_data\": \".09 11.30 36.56 0.5011.3BlockAttn32.018.99 30.90 1.6125.7Longformer42.78 13.21 41.34 0.9725.3LongT542.03 12.67 40.76 1.0325.4Unlimiformer35.17 11.98 34.28 1.3327.0Extract-Abstract 19.955.58 19.70 0.0613.1AWESOMEAttn Only46.05 † 13.09 † 44.21 † 0.81 †13.2Attn + Txt45.30 † 12.63 † 43.51 † 0.90 †14.2\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Results on arXiv papers. AWESOME variants again outperform Se3. For 80% of the arXiv documents, efficient attention models and PageSum can fully train on their first halves, covering 90% of the salient content that appear in the references(Huang et al., 2021), thus the better ROUGE scores than models encoding smaller segments.\",\n \"figure_data\": \"40.74 17.96 36.87 1.3312.8BlockAttn49.12 21.69 44.40 1.7725.7Longformer48.59 21.45 43.99 2.1725.2LongT548.25 20.74 43.41 0.9725.5Unlimiformer47.78 20.58 43.22 1.2226.8Extract-Abstract 42.37 16.43 38.62 1.0315.3PageSum46.01 18.77 41.55 0.8826.2AWESOMEAttn Only42.51 † 18.96 † 38.56 † 1.3016.0Attn + Txt44.20 † 18.89 † 40.07 † 1.3216.5ModelR-1 ↑ R-2 ↑ R-L ↑ Disco ↓ GPU ↓Se340.78 10.16 39.77 10.46 11.5BlockAttn23.45 3.09 22.09 190.27 25.7Longformer20.20 2.45 18.55 204.48 25.3LongT533.15 6.74 32.62 24.24 25.5Unlimiformer38.09 9.55 37.41 47.72 27.0AWESOME (Attn) 41.11 10.63 40.20 10.36 24.0\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Results on novels in BookSum. AWESOME with attentive memory in all layers achieves the best performance on all metrics. Methods requiring external extractors are not included due to the computational cost of building extractive oracles for long novels.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"7\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Statistics of datasets used in our experiments.\",\n \"figure_data\": \"DatasetModelGov arXiv QMSum SumScrn BookSe350x50x50x50x50xExt-Abs †1x (∞) 1x (∞) 1x (∞) 1x (∞)-BlockAttn6x6x8x6x6xLongformer8x8x8x8x8xLongT56x6x6x6x6xUnlimiformer 2x2x2x2x2xPageSum3x5x2x--AWESOME 50x50x50x50x50x\",\n \"figure_id\": \"tab_10\",\n \"figure_label\": \"8\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Truncation thresholds (multiply by 1024) used by each model on different datasets to comply with the memory constraint during training. †: For the extractthen-abstract model, the abstractor has a maximum input length of 1024, while the extractor can consume all sentences in the document.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_11\",\n \"figure_label\": \"9\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(Lewis et al., 2020)\", \"Explanation\": \"The cited work by Lewis et al. provides the performance benchmarks for large pre-trained transformer models in abstractive summarization, which the citing paper uses to evaluate the performance of their own model.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Raffel et al., 2020)\", \"Explanation\": \"The cited work by Raffel et al. provides a dataset of long documents for the citing paper to use in their research on summarizing long documents.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Cohan et al., 2018)\", \"Explanation\": \"The cited work by Cohan et al. discusses the challenges of summarizing long documents, which the citing paper builds upon to develop solutions for addressing the long document challenge.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Huang et al., 2021)\", \"Explanation\": \"The cited work by Huang et al. provides a method for summarizing long documents, which the citing paper adopts to address the long document challenge.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Gidiotis and Tsoumakas, 2020)\", \"Explanation\": \"The cited work by Gidiotis and Tsoumakas presents a method for dividing a document into smaller chunks for summarization, which the citing paper uses to address the long document challenge.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Beltagy et al., 2020)\", \"Explanation\": \"The cited work by Beltagy et al. discusses methods for reducing the complexity of attention calculations in summarization, which the citing paper adopts to address the long document challenge.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Pilault et al., 2020)\", \"Explanation\": \"The cited work by Pilault et al. presents a method for removing unimportant content in a document before running an abstractor, which the citing paper uses to address the long document challenge.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Moro and Ragazzi, 2022)\", \"Explanation\": \"The cited work by Moro and Ragazzi discusses the divide-and-conquer method for addressing the long document challenge, which the citing paper adopts to address the long document challenge.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zaheer et al., 2021)\", \"Explanation\": \"The cited work by Zaheer et al. introduces a method of combining global and local attention in transformer-based summarization models, which the citing paper adopts to improve performance in long document summarization.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Phang et al., 2022)\", \"Explanation\": \"The cited work by Phang et al. also contributes to the method of combining global and local attention in transformer-based summarization models, which the citing paper builds upon to achieve state-of-the-art performance in long document summarization.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(OpenAI, 2023)\", \"Explanation\": \"The cited work by OpenAI introduces the large language model GPT-4, which the citing paper uses as a reference for training to handle up to 32K tokens in long document summarization.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lin, 2004)\", \"Explanation\": \"The cited work by Lin (2004) provides the ROUGE metric, which is used in the citing paper to evaluate summary informativeness.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhao et al., 2022)\", \"Explanation\": \"The cited work by Zhao et al. (2022) introduces the DiscoScore metric, which is used in the citing paper to measure coherence in summaries.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Guinaudeau and Strube, 2013)\", \"Explanation\": \"The cited work by Guinaudeau and Strube (2013) presents a metric based on entity graphs that is used in the citing paper to measure coherence in summaries.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Laban et al., 2022)\", \"Explanation\": \"The cited work by Laban et al. (2022) introduces the SummaC metric, which is used in the citing paper to assess summary faithfulness in a divide-and-conquer approach.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Phang et al., 2022)\", \"Explanation\": \"The cited work introduces a method of combining global and local attention to reduce the quadratic complexity of the original transformer architecture, which the citing paper adopts in their research to improve the efficiency of attention mechanisms.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The cited work is an extension of the research on attention mechanisms, as it also aims to improve the efficiency of attention in language processing by maintaining full encoding context through a combination of global and local attention.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Huang et al., 2021)\", \"Explanation\": \"The cited work provides evidence that the method of distributing encoder outputs to different attention heads can reduce the memory usage of decoder cross attentions, which the citing paper uses to support their research on attention efficiency.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Bertsch et al., 2023)\", \"Explanation\": \"The cited work is a data source for the KNN search method used in the citing paper to select attendable encoder outputs for attention in language processing.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Pilault et al., 2020)\", \"Explanation\": \"The cited work introduces the concept of extract-then-abstract systems, which the citing paper adopts to address the long sequence challenge in attention-based systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu and Lapata, 2019)\", \"Explanation\": \"The cited work provides a method of using an abstractor to produce the final summary in extract-then-abstract systems, which the citing paper builds upon to address the long sequence challenge in attention-based systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhao et al., 2020)\", \"Explanation\": \"The cited work offers a method of using an abstractor to produce the final summary in extract-then-abstract systems, which the citing paper builds upon to address the long sequence challenge in attention-based systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Mao et al., 2022)\", \"Explanation\": \"DYLE predicts output token distributions for each segment separately, which is adopted in the citing paper to improve the accuracy of summary decoding.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"PageSum alleviates context loss by averaging decoder output representations conditioned on all document segments, which the citing paper leverages to improve the summarization process.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhang et al., 2022)\", \"Explanation\": \"Summ N uses an additional summarization stage to further condense the segmented summaries, which the citing paper builds upon to improve the accuracy of long document summarization.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu et al., 2021)\", \"Explanation\": \"The cited work by Wu et al. (2021) is used as a basis for the design of the memory-efficient divide-and-conquer approach in the citing paper, which is used to improve summary informativeness, coherence, and faithfulness.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu et al., 2022a)\", \"Explanation\": \"The cited work on Memorizing Transformer provides a new method of storing all prior representations as key-value pairs and performing an approximate kNN lookup to retrieve representations for content augmentation, which the citing paper adopts in their research for long document summarization.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Moro and Ragazzi, 2022)\", \"Explanation\": \"The cited work provides the basis for the document summarization architecture in AWESOME, which is based on Se3 and involves segmenting documents and reference summaries into segments and assigning them to the most overlapping segment in the oracle summary.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Rae et al., 2020)\", \"Explanation\": \"The cited work, Compressive Transformer, is adopted by the citing paper to develop a compression-based memory mechanism for language modeling in the encoder and decoder of a Transformer model. The cited work provides the methodology and techniques for the memory compression process.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lei et al., 2020)\", \"Explanation\": \"The cited work by Lei et al. provides a method of using memories in video captioning, which the citing paper updates with a gate matrix to control the content to be updated in the memory.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Huang et al., 2021)\", \"Explanation\": \"The dataset GovReport is used in the experiments conducted in the citing paper to evaluate the performance of the models.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhong et al., 2021)\", \"Explanation\": \"The dataset QMSum is also used in the experiments to evaluate the performance of the models.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Chen et al., 2022)\", \"Explanation\": \"The dataset SummScreen is used in the experiments to evaluate the performance of the models.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Bertsch et al., 2023)\", \"Explanation\": \"The dataset SummScreen is used in the experiments to evaluate the performance of the models.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The model PageSum is used in the experiments to evaluate the performance of the models.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Lin, 2004)\", \"Explanation\": \"The evaluation metric ROUGE is used to measure summary informativeness, providing a measure of the quality of the system summary.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Zhao et al., 2022)\", \"Explanation\": \"The evaluation metric DiscoScore is used to measure discourse coherence, providing a measure of the quality of the system summary.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Guinaudeau and Strube, 2013)\", \"Explanation\": \"The graph-based reference-free coherence metric is used to measure the connectivity of summary sentences linked by entities, providing a measure of the quality of the system summary.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Iv et al., 2022)\", \"Explanation\": \"The cited work provides a method for measuring the precision of entities in a summary with respect to a document, which the citing paper adopts for its own research on text generation.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Laban et al., 2022)\", \"Explanation\": \"The cited work introduces a model-based faithfulness metric for text generation, which the citing paper builds upon to measure the quality of summaries in their research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(VA, 2019)\", \"Explanation\": \"The cited work by VA provides updates on efforts to implement recommendations related to wait-time measurement and scheduling policy, which the citing paper uses to support its research on the topic.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu et al., 2022b)\", \"Explanation\": \"The cited work on using retrieval for question answering provides a methodological basis for the citing paper in terms of interleaving natural language-based augmentation with document segments.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Huang et al., 2021)\", \"Explanation\": \"The cited work provides evidence that the distribution of salient content in arXiv articles is uneven, which has a significant impact on summary generation and the use of efficient attention mechanisms.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Table 7)\", \"Explanation\": \"The citing paper further extends the research on BookSum by exploring the use of external memory in all layers to improve the performance of long novel plots in summary generation.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Huang et al., 2021)\", \"Explanation\": \"The cited work provides the Gov-Report dataset, which the citing paper uses in their experiments on long document summarization.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhong et al., 2021)\", \"Explanation\": \"The QMSum dataset is cited for the long meeting transcript summarization task, and the citing paper uses the data in their experiments.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Chen et al., 2022)\", \"Explanation\": \"The SummScreen dataset is cited for the TV series summarization task, and the citing paper uses the data in their experiments.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Cohan et al., 2018)\", \"Explanation\": \"The citing paper uses the data from the arXiv scientific papers and abstracts dataset in their experiments on summarizing full novels in the Book-Sum dataset.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Kryscinski et al., 2022)\", \"Explanation\": \"The citing paper uses the data from the Book-Sum dataset in their experiments on summarizing full novels.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Bertsch et al., 2023)\", \"Explanation\": \"The cited work by Bertsch et al. (2023) is the source of the data files for BookSum that the citing paper uses in its research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Phang et al., 2022)\", \"Explanation\": \"The cited work introduces the concept of blockwise attention, which the citing paper adopts in the design of the BlockAttn and Longformer models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Beltagy et al., 2020)\", \"Explanation\": \"The cited work presents the sliding-window attention technique, which the citing paper incorporates in the design of the Longformer model.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Guo et al., 2022)\", \"Explanation\": \"The cited work introduces the LongT5 model pre-trained on long sequences, which the citing paper uses as a data source for the LongT5 model.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Bertsch et al., 2023)\", \"Explanation\": \"The cited work presents the Unlimiformer model that extends BART by selecting input tokens to be attended to via KNN searching, which the citing paper utilizes as a data source for the Unlimiformer model.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The cited work introduces the PageSum model that synthesizes the output representations given by different document segments with dynamic weights, which the citing paper builds upon in the design of the PageSum model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Mao et al., 2022)\", \"Explanation\": \"The cited work by Mao et al. (2022) provides a method for computing ROUGE-1 and ROUGE-2 scores, which the citing paper uses in the process of training the extractor to obtain oracle extractive labels.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Lin, 2004)\", \"Explanation\": \"The cited work by Lin (2004) provides the evaluation metric used in the citing paper for ROUGE, which is a standard measure in text generation tasks.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhao et al., 2022)\", \"Explanation\": \"The cited work by Zhao et al. (2022) provides the official code for DiscoScore, a popular evaluation metric in text generation tasks that the citing paper utilizes.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Guinaudeau and Strube, 2013)\", \"Explanation\": \"The cited work by Guinaudeau and Strube (2013) provides the implementation of the Ent Graph metric, which the citing paper uses in their evaluation.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Laban et al., 2022)\", \"Explanation\": \"The cited work by Laban et al. (2022) provides the official code for Sum-maC, a metric used in the citing paper to evaluate the performance of their model.\"}]"}}},{"rowIdx":73,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"INTRODUCTION","publication_ref":["b16","b17","b2","b10"],"table_ref":[],"text":"In search engines, the queries issued by users are mostly broad and vague [15,16]. This problem is extremely crucial for product search scenarios, since users may only issue the query such as \"Microsoft\" (the brand), \"Personal Computer\"(the category), or \"Surface\"(part of the product name), instead of the exact name of the product \"Microsoft Surface Pro 9\" to find and purchase it. If online shopping platforms only use ad-hoc ranking models to provide results, the products that users intend to purchase may be displayed at the bottom positions, which makes users unsatisfied and the sales of platforms low. To solve this problem, similar to the personalized web search area, many personalized product search approaches [1-6, 9, 11] have been proposed. These methods utilize user histories to rank products considering both the relevance to the current query and user interests. Applying this strategy, online platforms can show products that users want to buy at top positions.\nUnfortunately, the lack of large-scale datasets based on real user behaviors blocks the studies of personalized product search. The most widely-used personalized product search datasets are simulated Amazon sub-category datasets 1 . These datasets are collected from the interactions between products and users in the amazon online shopping websites from 1996 to 2014. However, these datasets are originally developed for evaluating recommendation models and only contain the reviews of users and products, without the genuine queries issued by users. To apply these datasets in personalized product search experiments, researchers usually concatenate the terms in the category lists of products to build pseudo queries [3,10]. These generated queries can somehow resemble the ambiguity of actual queries, but in reality, users may issue diverse queries based on several fields of the products they want to buy, including brands, categories, names, and detailed descriptions. This may make the experimental performances based on pseudo queries in these datasets differ from the real product search scenarios. Furthermore, these datasets also put some restrictions on the products and users. For example, they only include products and users whose total amounts of reviews are larger than five (5-core products and users), which is an obstacle for few-shot model studies. Another drawback of this dataset is that each dataset solely contains products belonging to the same category. However, user histories usually contain products belonging to a variety of categories in reality, and modeling user interests across diverse categories may be beneficial for personalization. As a result, the simulated user behaviors in these datasets can vary from the patterns in practice, which makes the results on them unreliable. Thus, the models designed using these datasets may be not applicable in real situations. There also exists a few open-resource datasets such as DIGINETICA2 . However, these datasets are in a small scale and few previous approaches utilize them. To conclude, research progress in the personalized product search area requires a largescale dataset containing real user behaviors, but existing datasets are unsatisfactory.\nIn this paper, we construct and release a new dataset, JDsearch, based on a popular Chinese online shopping platform JD.com. JDsearch dataset is a ready-to-use and well-documented anonymized dataset. It is licensed under CC BY-NC-SA 4.0 3 . The dataset contains about 170,000 users, 12,000,000 products, 9,400,000 real searches, and 26,000,000 user-product interactions between them. During the construction of this dataset, we record real user queries and reserve all products belonging to all categories, regardless of their popularity. Users with various history lengths are also included in this dataset. To protect the privacy of users, we anonymize all the sensitive information including ids and texts in JDsearch dataset. We analyze this dataset from several perspectives including users, products, queries, and personalization potential to show its advantages. Finally, we test a wide range of existing personalized product search models on this dataset to show the feasibility of using it to conduct personalization studies. Overall, JDsearch dataset has the following advantages and can inspire some new research topics :\n1) Different from the pseudo queries in the Amazon dataset, the queries in JDsearch dataset are real. This can make the experimental results on this dataset closer to the online serving scenarios.\n2) We reserve all the products belonging to diverse categories in JDsearch dataset, which can support future studies considering multiple user preferences and few-shot products.\n3) We include both users with extremely long and short histories, which may lead the future approaches to design different strategies for different kinds of users."},{"figure_ref":[],"heading":"RELATED WORK","publication_ref":["b7","b13","b7"],"table_ref":[],"text":"Previous personalized product search works usually utilize the Amazon sub-category datasets to conduct experiments. These datasets are originally recommendation datasets, which contain product reviews and metadata on the Amazon websites from May 1996 to July 2014. As the scale of the overall dataset is too huge for models to personalize results, researchers usually select some sub-category datasets to set up experiments and these datasets only contain products belonging to one single category. Besides, these datasets also filter out products and users whose reviews are less than five to obtain the denser 5-core datasets. Since these datasets only contain the review information of products, researchers need to heuristically build pseudo queries from the metadata of products. Previous approaches commonly utilize the categories of products to build queries: they concatenate the words in the category lists and remove the duplicated words and stopwords to construct the queries for corresponding products. In this way, the generated queries can mimic the ambiguity of real queries. However, users don't issue queries solely based on the category information of the products they want to purchase in reality.\nExcept for the widely-used simulated Amazon sub-category products, there only exists a few open-resource datasets. DIGINETICA or CIKMCUP2016 dataset is only utilized in two previous works [7,12]. This dataset is collected from a Russian online shopping platform. However, over half of the purchase behaviors in this dataset are done by guests [7]. In other words, these behaviors don't have user ids and cannot be utilized for personalization. After filtering these anonymous behaviors, the remaining dataset is quite small and difficult for conducting personalization experiments. Besides the DIGINETICA dataset, there also exist some online competitions containing users' online shopping histories. However, to the best of our knowledge, no previous studies apply these competition datasets in their experiments."},{"figure_ref":[],"heading":"JDSEARCH: A NEW PERSONALIZED PRODUCT SEARCH DATASET","publication_ref":["b10"],"table_ref":[],"text":"As we mentioned in Section 1, queries in Amazon datasets are heuristically constructed by the categories of products [10]. In contrast to this hypothetical situation, in reality, users can issue a variety of queries based on different attributes such as titles, brands, and categories to search products. Thus, the simple pseudo query generation process in Amazon datasets cannot mimic the real search behaviors well. Further, Amazon sub-category datasets also put several limitations on the products and users. First, it only contains products and users that have at least 5 reviews or purchases (5-core products and users) and leaves others out. Second, in each sub-category dataset, products in the corpus all fall under the same category. However, in real situations, what personalization models face is that tail users and products are common while products in user histories may belong to diverse categories. In conclusion, the above restrictions may cause the simulated Amazon datasets to differ from the real situation. This may lead the methods developed, tested, and assessed using Amazon datasets to become inapplicable in reality. As a result, it is essential to construct and release a new dataset based on real user behaviors."},{"figure_ref":[],"heading":"Dataset Construction","publication_ref":[],"table_ref":["tab_0"],"text":"To solve the problems in Amazon datasets and support research in the personalized product search area, we construct a new personalized product search dataset based on genuine user behaviors from a Chinese online shopping platform JD.com. The preprocessing pipeline for constructing JDsearch dataset is as follows:\n1) User Behavior Collection. First, we randomly sample about 170,000 users who have issued queries on the platform on a specific date, 2022-10-17. The histories of users are formed by their issued queries and interacted products from 2021-10-18 to 2022-10-17. Unlike the Amazon datasets, we place no restrictions on the categories or populations of the history products or the history lengths of users. Therefore, all the products in the dataset corpus belong to a variety of categories, which makes the user histories more diverse. In particular, we include all types of user behaviors including click, add to cart, follow and purchase, and record these interaction type labels in JDsearch dataset, which can provide signals for future works considering multiple feedback. The timestamps of these actions are also recorded in our dataset which can offer temporal information to models. Specifically, we don't require the interactions with products must be under issued queries, which means users may interact with products from diverse channels including search, recommendation, and casual browsing. For example, if a user purchases an iPhone by searching \"smart cellphone\" and then clicks the AirPods product through the recommendation systems, these behaviors will all be recorded in JDsearch but Amazon datasets may only record the first one. In summary, we record the historical queries, interacted products, interaction type labels, and their corresponding timestamps of users in our JDsearch dataset. : 2) Product Metadata Collection. Then, for the product meta information, we record the names, categories, brands, and shops of products in our dataset. There exist four-level categories in the JD online shopping platform and we reserve all of them. However, the JDsearch dataset doesn't contain the related item relationships (such as \"bought together\" in the metadata part of Amazon datasets) because the platform doesn't release these data.\n3) Anonymization. Next, since this dataset is collected from a commercial shopping platform, we need to anonymize JDsearch to remove personal private information. For the ids of products, brands, categories, and shops, we randomly hash them to numbers in a wide range. For the textual information, we first conduct word segmentations for texts including queries, product titles, category names, etc. Then, we randomly hash these term ids to integers, too.\n4) Dataset Partition. Finally, similar to the popular leave-oneout evaluation methods in recommendation systems, we use the last queries of the users issued on 2022-10-17 as the test queries. The behaviors before the last queries can be used to train models. Different from the Amazon datasets that don't have the displayed results of queries, we obtain the exposed product lists and their labels under the test queries issued by users in this commercial platform. We further remove the duplicated products in these displayed results and reserve at most 200 products as the candidate product lists for these test queries. Besides, users who don't interact with any products under their last queries (about 2,000 users) are removed from the test part of JDsearch dataset. Because of including the exposed products, personalization models can conduct fine-grained ranking in JDsearch dataset instead of the coarse-grained retrieving in Amazon datasets. In a nutshell, we record the test queries, their candidate product lists, and the labels for candidates for each user in JDsearch dataset.\nThe fields and explanation in user behavior data and product meta data in JDsearch dataset are shown in Table 1. The detailed content and format description of our JDsearch dataset can be found in our repository. 4 .\nOverall, compared with the Amazon sub-category datasets, The main advantages of our JDsearch dataset are: 1) it includes real user queries, 2) it reserves all products with different categories and comprises both cold and popular products, 3) it contains various types of users whose history lengths are diverse. Furthermore, we also record all interactions with various types and prepare the candidate product lists for test queries. We "},{"figure_ref":[],"heading":"DATASET ANALYSIS","publication_ref":[],"table_ref":["tab_3"],"text":"First, we provide basic statistics in our dataset and some Amazon sub-category datasets in Table 3. Compared with widely used Amazon sub-category datasets, we don't filter out the products which have only been interacted with a few times and keep track of all the interactions in user history. Thus, we can find that JDsearch dataset is much sparser and the average user history length is much longer, making it more challenging for models to capture user interests and conduct personalized ranking. Besides, our dataset contains more test queries than Amazon datasets, which can make the evaluation more stable and convincing. "},{"figure_ref":["fig_2"],"heading":"Product and User Analysis","publication_ref":[],"table_ref":[],"text":"In this part, we investigate the characteristics of products and users in JDsearch dataset.\n4.1.1 Product's Interaction Frequency Analysis. First, we demonstrate the distribution of the product's interaction frequency in the JDsearch dataset in Figure 1. We can find the distribution of product's interaction frequency aligns with power law distribution, which means that many products have only interacted with users a few times (cold products) while the amounts of hot products (frequently interacted with users) are relatively limited. This phenomenon suggests that the preprocess manipulation of only including 5-core products in Amazon datasets may destruct the continuous user behaviors in reality and may result in different performances compared with retaining all products. Besides, this configuration may help models filter out noisy histories and makes personalization easier. However, in reality, personalization models need to discriminate which part of user histories is more important for personalization and more related to current queries."},{"figure_ref":["fig_4","fig_5"],"heading":"User History","publication_ref":[],"table_ref":[],"text":"Analysis. Then, we analyze the pattern of user histories. The frequencies of user history length in JDsearch dataset are shown in Figure 2. We can discover that there exist users with very short histories (only have one or two interacted products). Meanwhile, some users that have extremely rich interaction histories are also included in our dataset. This phenomenon indicates that users may have different personalities: some users rarely do online shopping while some users frequently browse products online. Thus, it may stimulate fresh research on designing different personalization strategies for users with different characters. For example, models may incorporate more universal preferences from all users into the user modeling while the users have limited histories and pay more attention to personal interests while they have rich interactions. Besides, we also show the numbers of products' first-level categories in user histories in Figure 3. For a fair comparison, we only investigate users whose history length is larger than five in our dataset, since Amazon datasets only include 5-core users. We can find that most users' histories are formed with products belonging to different categories. Different from this characteristic, in the Amazon 5-core sub-category datasets, user histories only contain products belonging to one certain top-level category. However, customers' interests in one category may extend to other categories, hence only containing products belonging to one category can be harmful to personalization. Therefore, it may be better to include products with all categories in user histories as we did in the JDsearch dataset."},{"figure_ref":["fig_6","fig_7"],"heading":"Query Analysis","publication_ref":["b8"],"table_ref":["tab_4"],"text":"As we mentioned in Section 1, the most crucial problem in Amazon datasets is that the queries in them are pseudo ones and are generated by the categories of products. These queries can somehow reflect the ambiguity of the real issued queries by users. However, based on our JDsearch dataset gathered from real user behaviors, we observe that the source of query terms can be the brands, names, or even detailed descriptions of products, not only the categories of the products they want to purchase. Specifically, depending on the sources of the terms, we categorize queries in JDsearch dataset into the following types:\n• Category: All the query terms belong to the interacted product's category (e.g., cellphone). • Brand: All the query terms belong to the interacted product's brand (e.g., Apple). • Name: All the query terms belong to the interacted product's name (e.g., Surface Pro 8). • Category&Brand: All the query terms belong to the interacted product's category and brand (e.g., Apple cellphone). • Category&Name: All the query terms belong to the interacted product's category and name (e.g., Laptop Surface). • Brand&Name: All the query terms belong to the interacted product's brand and name (e.g., Microsoft Surface).\nWe show the type distribution of queries in Figure 4. We can observe that the terms of most queries issued by users only derive from the names of products. After that, users also usually enter queries based on the categories or the combination of categories and names. Besides, there also exist queries based on the brands of products. Noticing that in this part, we only investigate queries that can be categorized into the above types. There still exist many queries whose terms originate from other fields of products. These findings show that the naive strategy of constructing queries from categories of products is inadequate, a better way may be to utilize diverse fields of products to assemble pseudo queries. Then, we analyze all the history queries and show their ambiguity. To quantify the ambiguity of queries, we propose a metric interaction entropy(IE) for queries. The interaction entropy is a natural extension of click entropy [8] and can be calculated from the corresponding interactions under queries. Generally speaking, when the interaction entropy of a query is larger than or equal to one, we can infer that this query is an informational query and may be ambiguous. The interaction entropy is calculated as follows:\nIE(𝑞) = ∑︁ 𝑝 ∈ I (𝑞) -𝑃 (𝑝 |𝑞) log 2 𝑃 (𝑝 |𝑞)(1)\n𝑃 (𝑝 |𝑞) = |Interaction(𝑞, 𝑝)| 𝑝 ′ ∈ I (𝑞) |Interaction(𝑞, 𝑝 ′ )| , (2\n)\nwhere I (𝑞) is the collection of products interacted (including clicked, added to cart, followed, and purchased) with users under query 𝑞, 𝑃 (𝑝 |𝑞) is the percentage of interactions on product 𝑝 among all interactions under query 𝑞. we calculate the interaction entropy for all repeated queries in JDsearch dataset and show the numbers of queries whose IE is less than / equal to / larger than one in Table 4.\nFrom the statistics, we can find that most queries in our corpus are ambiguous and different products have been interacted under them. This can also verify the personalization is necessary in our dataset. (the latest ten interactions). The support distribution of 𝑢 is calculated based on her own early histories (interactions before the latest ten ones). The overall distribution is calculated based on all users' early histories. The personal divergence of 𝑢 is calculated between her valid distribution and her support distribution. The overall divergence is calculated between her valid distribution and the overall distribution. All these divergences are calculated by JS divergence. We calculate the personal divergence and overall divergence of each user 𝑢 based on the first-level category and brand respectively. We argue that if user interests are continuous and personalization is rational, the personal divergence should be smaller than the overall divergence. We show the results in Figure 5.\nIn the boxplot figure, since there must exist some users' late purchase behaviors are totally different from the early ones and the overall preferences, the upper bound of the personal divergences and overall divergences are all close to log 2, which is the upper bounds of JS divergence. However, we can infer that most users' personal divergence is smaller than the overall divergence, which suggests that most users' interests in JDsearch dataset are in consistency and personalization can be effective in our dataset."},{"figure_ref":[],"heading":"EXPERIMENTS AND ANALYSIS","publication_ref":[],"table_ref":[],"text":"The JDsearch dataset can be used for personalized product search studies to conduct experiments. In this part, we evaluate representative personalized product search models in our JDsearch dataset to verify the feasibility of performing personalization in this dataset. Further, we also conduct experiments on several dataset variations to show the different characteristics of our dataset compared with Amazon datasets."},{"figure_ref":[],"heading":"Settings and Evaluation Metrics","publication_ref":[],"table_ref":[],"text":"In our experiments, we train the models using the behaviors with queries in user histories. For inference, as we mentioned in Section 4, we use the last queries of users as the test queries. So we apply the trained models to rank the candidate product lists to evaluate their performances. For evaluation metrics, we use MRR@200, Precision@1, and NDCG@10 to evaluate the ranking results."},{"figure_ref":[],"heading":"Benchmark models","publication_ref":["b15","b4","b0","b5","b1"],"table_ref":[],"text":"We experiment with the following ad-hoc and personalized models: BM25: BM25 [14] is a classical sparse ad-hoc retrieval model. QEM: QEM only considers the matching scores between products and queries and can be regarded as a neural ad-hoc model. HEM: HEM [3] is a latent vector based personalized model. It builds the representations of users and items by generative language models based on reviews.\nDREM: DREM [4] is a KG based personalized model. It utilizes the metadata of items to establish a knowledge graph.\nAEM, ZAM: AEM [1] is an attention-based personalized model. It aggregates the user historical interacted items with the current query to construct a query-specific user profile. ZAM improves AEM by concatenating a zero vector to the item list to adjust the extent of conducting personalization.\nTEM: TEM [5] is a transformer-based personalized model. It upgrades the attention layer in AEM with transformer encoder.\nHGN: HGN [2] integrates DREM and ZAM models. It uses the relations in KG to boost the representations of products and users."},{"figure_ref":[],"heading":"Implementation Details","publication_ref":["b14","b2","b4"],"table_ref":[],"text":"As we anonymize the textual information in the dataset, which is much sparser than Amazon datasets, it is hard to optimize the word embeddings from scratch. So we adopt word2vec [13] to initialize the word embedding in models. Besides, different from the settings in Amazon datasets, products in the test set may have not been seen in the training process (they may be cold products). In addition to the large scale of product corpus, simply optimizing the product embedding table as previous models may cause extreme overfitting. Therefore, we obtain products' representations in models by calculating dynamic vectors based on their texts including title, brand, and category. We use the average term vectors of products' texts and apply a simple non-linear function to obtain product representations. Due to the privacy protection policy, we don't have the review information of users and products. Thus, we remove the generative language modeling loss for items and users from all models. Thus, in HEM [3] and DREM [4], again, we apply the non-linear function to build user embeddings from all the terms in their history interacted items. For the KG-enhanced models, as JD commercial platform doesn't provide the related product relationships (such as \"bought together\" and \"also bought') to us, so we only utilize the category and brand information to build the item-attribute graph. For the hyper-parameter settings, we set the maximum user history length in all models as 50. We set the embedding dimension as 128 and train them for 30 epochs. The number of transformer layers in transformer-based models is chosen from {1, 2}. The number of attention heads in attention-based models is set as 8. For all negative sampling in models, we randomly choose five negative samples in uniform distribution over the corpus."},{"figure_ref":[],"heading":"Overall Results","publication_ref":[],"table_ref":["tab_6"],"text":"The overall results are shown in Table 5. We can find that all the neural product search models overperform the sparse retrieval method BM25, which shows the effectiveness of neural ranking models. Except for the KG-enhanced DREM and HGN models, all the personalized product search models achieve improvements over the ad-hoc QEM models, demonstrating the feasibility of conducting personalization in our dataset. The poor performances of KG-based models may result from that we don't have efficient meta relationships among products and attributes. A promising way of solving this issue is to select some user histories to build a denser knowledge graph. This experiment proves that it is possible to train and evaluate personalized product search models in JDsearch dataset. However, the improvements over the ad-hoc search models are relatively incremental, which indicates that there still exists potential research topics for more accurate personalization using our JDsearch dataset."},{"figure_ref":[],"heading":"Dataset Ablation Study","publication_ref":[],"table_ref":["tab_7"],"text":"As we introduced in Section 1, some manipulations in previous Amazon datasets are obstacles for personalized product search research. For example, the pseudo queries generated artificially may be different from the real queries issued by users. To further show the characteristics and advantages of our JDsearch dataset, we resemble some operations in processing Amazon datasets and adjust the construction procedure of JDsearch dataset to produce several dataset variants. These variants include : JDsearch fakequery : We replace all the user historical queries with the artificial queries generated by concatenating the corresponding interacted products' category terms but keep the real test queries.\nJDsearch samecate : We only reserve the products and their corresponding queries (if they have any) in user histories whose firstlevel category is the same as the products that the users finally interacted with under the test queries.\nJDsearch w/o cold : We only reserve the products and their corresponding queries (if they have any) in user histories that have at least five interactions with users.\nJDsearch short : We don't change the training process of models but only keep users whose history length is not larger than two in the test part.\nJDsearch long : We don't change the training process of models but only keep users whose history length is larger than three hundred in the test part.\nThen we train and evaluate the TEM model in these dataset variants and the results are shown in Table 6.\nWe can find that all these performances on different dataset variants are significantly different. These results show that the manipulations in Amazon sub-category datasets have effects on the model's performances and may skew the evaluation of approaches. We can find by replacing the queries in the training set with the pseudo queries in the JDsearch fakequery dataset, the improve performance as it heuristically removes some noises from user histories. However, a more reasonable and rational way may be to reserve all products and leave the process of cold products to the model designer and researchers. From the performances in JDsearch short and JDsearch long dataset, we can find that personalized models usually can achieve high results on users who frequently do online shopping but perform badly while facing cold users. This finding can inspire future research applying different personalization strategies for users with different personalities."},{"figure_ref":[],"heading":"DISCUSSION OF APPLICATION SCENARIOS","publication_ref":[],"table_ref":[],"text":"Besides utilizing it to run experiments and evaluate performances for personalized product search models, the JDsearch dataset can also support research in other areas. First, researchers can simply ignore the query information or remove the interactions under queries in the dataset to use it to assess product recommendation models. In addition, since we record all user behaviors from diverse channels including both search and recommendation, the JDsearch dataset can also support studies in the unified recommendation and search model, which has no publicly available datasets to the best of our knowledge."},{"figure_ref":[],"heading":"CONCLUSION","publication_ref":[],"table_ref":[],"text":"In this work, we introduce a new personalized product search JDsearch dataset collected from real user behaviors. Different from the simulated Amazon sub-category dataset, our dataset includes real user queries. Besides, there also exist products belonging to various categories. Cold products and users are also included. These lead the JDsearch dataset closer to the real product search situations. We also investigate the characteristics of the dataset from several perspectives and test existing personalized product search models. These analyses and experiments verify the feasibility of the proposed dataset. This dataset can also support some potential personalization directions including few-shot scenarios, multi-interest modeling, and separate strategies for different types of users."},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"ACKNOWLEDGMENTS Zhicheng Dou is the corresponding author. This work is done during Jiongnan Liu's internship at JD. This work was supported by the National Key R&D Program of China (2022ZD0120103), National Natural Science Foundation of China (62272467), Beijing Outstanding Young Scientist Program NO. BJJWZYJH012019100020098, Public Computing Cloud, Renmin University of China, and Intelligent Social Governance Platform, Major Innovation & Planning Interdisciplinary Platform for the \"Double-First Class\" Initiative, Renmin University of China. The work was partially done at Beijing Key Laboratory of Big Data Management and Analysis Methods."},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"https://github.com/rucliujn/JDsearch."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"INTRODUCTION\",\n \"publication_ref\": [\n \"b16\",\n \"b17\",\n \"b2\",\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"In search engines, the queries issued by users are mostly broad and vague [15,16]. This problem is extremely crucial for product search scenarios, since users may only issue the query such as \\\"Microsoft\\\" (the brand), \\\"Personal Computer\\\"(the category), or \\\"Surface\\\"(part of the product name), instead of the exact name of the product \\\"Microsoft Surface Pro 9\\\" to find and purchase it. If online shopping platforms only use ad-hoc ranking models to provide results, the products that users intend to purchase may be displayed at the bottom positions, which makes users unsatisfied and the sales of platforms low. To solve this problem, similar to the personalized web search area, many personalized product search approaches [1-6, 9, 11] have been proposed. These methods utilize user histories to rank products considering both the relevance to the current query and user interests. Applying this strategy, online platforms can show products that users want to buy at top positions.\\nUnfortunately, the lack of large-scale datasets based on real user behaviors blocks the studies of personalized product search. The most widely-used personalized product search datasets are simulated Amazon sub-category datasets 1 . These datasets are collected from the interactions between products and users in the amazon online shopping websites from 1996 to 2014. However, these datasets are originally developed for evaluating recommendation models and only contain the reviews of users and products, without the genuine queries issued by users. To apply these datasets in personalized product search experiments, researchers usually concatenate the terms in the category lists of products to build pseudo queries [3,10]. These generated queries can somehow resemble the ambiguity of actual queries, but in reality, users may issue diverse queries based on several fields of the products they want to buy, including brands, categories, names, and detailed descriptions. This may make the experimental performances based on pseudo queries in these datasets differ from the real product search scenarios. Furthermore, these datasets also put some restrictions on the products and users. For example, they only include products and users whose total amounts of reviews are larger than five (5-core products and users), which is an obstacle for few-shot model studies. Another drawback of this dataset is that each dataset solely contains products belonging to the same category. However, user histories usually contain products belonging to a variety of categories in reality, and modeling user interests across diverse categories may be beneficial for personalization. As a result, the simulated user behaviors in these datasets can vary from the patterns in practice, which makes the results on them unreliable. Thus, the models designed using these datasets may be not applicable in real situations. There also exists a few open-resource datasets such as DIGINETICA2 . However, these datasets are in a small scale and few previous approaches utilize them. To conclude, research progress in the personalized product search area requires a largescale dataset containing real user behaviors, but existing datasets are unsatisfactory.\\nIn this paper, we construct and release a new dataset, JDsearch, based on a popular Chinese online shopping platform JD.com. JDsearch dataset is a ready-to-use and well-documented anonymized dataset. It is licensed under CC BY-NC-SA 4.0 3 . The dataset contains about 170,000 users, 12,000,000 products, 9,400,000 real searches, and 26,000,000 user-product interactions between them. During the construction of this dataset, we record real user queries and reserve all products belonging to all categories, regardless of their popularity. Users with various history lengths are also included in this dataset. To protect the privacy of users, we anonymize all the sensitive information including ids and texts in JDsearch dataset. We analyze this dataset from several perspectives including users, products, queries, and personalization potential to show its advantages. Finally, we test a wide range of existing personalized product search models on this dataset to show the feasibility of using it to conduct personalization studies. Overall, JDsearch dataset has the following advantages and can inspire some new research topics :\\n1) Different from the pseudo queries in the Amazon dataset, the queries in JDsearch dataset are real. This can make the experimental results on this dataset closer to the online serving scenarios.\\n2) We reserve all the products belonging to diverse categories in JDsearch dataset, which can support future studies considering multiple user preferences and few-shot products.\\n3) We include both users with extremely long and short histories, which may lead the future approaches to design different strategies for different kinds of users.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"RELATED WORK\",\n \"publication_ref\": [\n \"b7\",\n \"b13\",\n \"b7\"\n ],\n \"table_ref\": [],\n \"text\": \"Previous personalized product search works usually utilize the Amazon sub-category datasets to conduct experiments. These datasets are originally recommendation datasets, which contain product reviews and metadata on the Amazon websites from May 1996 to July 2014. As the scale of the overall dataset is too huge for models to personalize results, researchers usually select some sub-category datasets to set up experiments and these datasets only contain products belonging to one single category. Besides, these datasets also filter out products and users whose reviews are less than five to obtain the denser 5-core datasets. Since these datasets only contain the review information of products, researchers need to heuristically build pseudo queries from the metadata of products. Previous approaches commonly utilize the categories of products to build queries: they concatenate the words in the category lists and remove the duplicated words and stopwords to construct the queries for corresponding products. In this way, the generated queries can mimic the ambiguity of real queries. However, users don't issue queries solely based on the category information of the products they want to purchase in reality.\\nExcept for the widely-used simulated Amazon sub-category products, there only exists a few open-resource datasets. DIGINETICA or CIKMCUP2016 dataset is only utilized in two previous works [7,12]. This dataset is collected from a Russian online shopping platform. However, over half of the purchase behaviors in this dataset are done by guests [7]. In other words, these behaviors don't have user ids and cannot be utilized for personalization. After filtering these anonymous behaviors, the remaining dataset is quite small and difficult for conducting personalization experiments. Besides the DIGINETICA dataset, there also exist some online competitions containing users' online shopping histories. However, to the best of our knowledge, no previous studies apply these competition datasets in their experiments.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"JDSEARCH: A NEW PERSONALIZED PRODUCT SEARCH DATASET\",\n \"publication_ref\": [\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"As we mentioned in Section 1, queries in Amazon datasets are heuristically constructed by the categories of products [10]. In contrast to this hypothetical situation, in reality, users can issue a variety of queries based on different attributes such as titles, brands, and categories to search products. Thus, the simple pseudo query generation process in Amazon datasets cannot mimic the real search behaviors well. Further, Amazon sub-category datasets also put several limitations on the products and users. First, it only contains products and users that have at least 5 reviews or purchases (5-core products and users) and leaves others out. Second, in each sub-category dataset, products in the corpus all fall under the same category. However, in real situations, what personalization models face is that tail users and products are common while products in user histories may belong to diverse categories. In conclusion, the above restrictions may cause the simulated Amazon datasets to differ from the real situation. This may lead the methods developed, tested, and assessed using Amazon datasets to become inapplicable in reality. As a result, it is essential to construct and release a new dataset based on real user behaviors.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Dataset Construction\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_0\"\n ],\n \"text\": \"To solve the problems in Amazon datasets and support research in the personalized product search area, we construct a new personalized product search dataset based on genuine user behaviors from a Chinese online shopping platform JD.com. The preprocessing pipeline for constructing JDsearch dataset is as follows:\\n1) User Behavior Collection. First, we randomly sample about 170,000 users who have issued queries on the platform on a specific date, 2022-10-17. The histories of users are formed by their issued queries and interacted products from 2021-10-18 to 2022-10-17. Unlike the Amazon datasets, we place no restrictions on the categories or populations of the history products or the history lengths of users. Therefore, all the products in the dataset corpus belong to a variety of categories, which makes the user histories more diverse. In particular, we include all types of user behaviors including click, add to cart, follow and purchase, and record these interaction type labels in JDsearch dataset, which can provide signals for future works considering multiple feedback. The timestamps of these actions are also recorded in our dataset which can offer temporal information to models. Specifically, we don't require the interactions with products must be under issued queries, which means users may interact with products from diverse channels including search, recommendation, and casual browsing. For example, if a user purchases an iPhone by searching \\\"smart cellphone\\\" and then clicks the AirPods product through the recommendation systems, these behaviors will all be recorded in JDsearch but Amazon datasets may only record the first one. In summary, we record the historical queries, interacted products, interaction type labels, and their corresponding timestamps of users in our JDsearch dataset. : 2) Product Metadata Collection. Then, for the product meta information, we record the names, categories, brands, and shops of products in our dataset. There exist four-level categories in the JD online shopping platform and we reserve all of them. However, the JDsearch dataset doesn't contain the related item relationships (such as \\\"bought together\\\" in the metadata part of Amazon datasets) because the platform doesn't release these data.\\n3) Anonymization. Next, since this dataset is collected from a commercial shopping platform, we need to anonymize JDsearch to remove personal private information. For the ids of products, brands, categories, and shops, we randomly hash them to numbers in a wide range. For the textual information, we first conduct word segmentations for texts including queries, product titles, category names, etc. Then, we randomly hash these term ids to integers, too.\\n4) Dataset Partition. Finally, similar to the popular leave-oneout evaluation methods in recommendation systems, we use the last queries of the users issued on 2022-10-17 as the test queries. The behaviors before the last queries can be used to train models. Different from the Amazon datasets that don't have the displayed results of queries, we obtain the exposed product lists and their labels under the test queries issued by users in this commercial platform. We further remove the duplicated products in these displayed results and reserve at most 200 products as the candidate product lists for these test queries. Besides, users who don't interact with any products under their last queries (about 2,000 users) are removed from the test part of JDsearch dataset. Because of including the exposed products, personalization models can conduct fine-grained ranking in JDsearch dataset instead of the coarse-grained retrieving in Amazon datasets. In a nutshell, we record the test queries, their candidate product lists, and the labels for candidates for each user in JDsearch dataset.\\nThe fields and explanation in user behavior data and product meta data in JDsearch dataset are shown in Table 1. The detailed content and format description of our JDsearch dataset can be found in our repository. 4 .\\nOverall, compared with the Amazon sub-category datasets, The main advantages of our JDsearch dataset are: 1) it includes real user queries, 2) it reserves all products with different categories and comprises both cold and popular products, 3) it contains various types of users whose history lengths are diverse. Furthermore, we also record all interactions with various types and prepare the candidate product lists for test queries. We \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"DATASET ANALYSIS\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_3\"\n ],\n \"text\": \"First, we provide basic statistics in our dataset and some Amazon sub-category datasets in Table 3. Compared with widely used Amazon sub-category datasets, we don't filter out the products which have only been interacted with a few times and keep track of all the interactions in user history. Thus, we can find that JDsearch dataset is much sparser and the average user history length is much longer, making it more challenging for models to capture user interests and conduct personalized ranking. Besides, our dataset contains more test queries than Amazon datasets, which can make the evaluation more stable and convincing. \"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Product and User Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this part, we investigate the characteristics of products and users in JDsearch dataset.\\n4.1.1 Product's Interaction Frequency Analysis. First, we demonstrate the distribution of the product's interaction frequency in the JDsearch dataset in Figure 1. We can find the distribution of product's interaction frequency aligns with power law distribution, which means that many products have only interacted with users a few times (cold products) while the amounts of hot products (frequently interacted with users) are relatively limited. This phenomenon suggests that the preprocess manipulation of only including 5-core products in Amazon datasets may destruct the continuous user behaviors in reality and may result in different performances compared with retaining all products. Besides, this configuration may help models filter out noisy histories and makes personalization easier. However, in reality, personalization models need to discriminate which part of user histories is more important for personalization and more related to current queries.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\",\n \"fig_5\"\n ],\n \"heading\": \"User History\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Analysis. Then, we analyze the pattern of user histories. The frequencies of user history length in JDsearch dataset are shown in Figure 2. We can discover that there exist users with very short histories (only have one or two interacted products). Meanwhile, some users that have extremely rich interaction histories are also included in our dataset. This phenomenon indicates that users may have different personalities: some users rarely do online shopping while some users frequently browse products online. Thus, it may stimulate fresh research on designing different personalization strategies for users with different characters. For example, models may incorporate more universal preferences from all users into the user modeling while the users have limited histories and pay more attention to personal interests while they have rich interactions. Besides, we also show the numbers of products' first-level categories in user histories in Figure 3. For a fair comparison, we only investigate users whose history length is larger than five in our dataset, since Amazon datasets only include 5-core users. We can find that most users' histories are formed with products belonging to different categories. Different from this characteristic, in the Amazon 5-core sub-category datasets, user histories only contain products belonging to one certain top-level category. However, customers' interests in one category may extend to other categories, hence only containing products belonging to one category can be harmful to personalization. Therefore, it may be better to include products with all categories in user histories as we did in the JDsearch dataset.\"\n },\n {\n \"figure_ref\": [\n \"fig_6\",\n \"fig_7\"\n ],\n \"heading\": \"Query Analysis\",\n \"publication_ref\": [\n \"b8\"\n ],\n \"table_ref\": [\n \"tab_4\"\n ],\n \"text\": \"As we mentioned in Section 1, the most crucial problem in Amazon datasets is that the queries in them are pseudo ones and are generated by the categories of products. These queries can somehow reflect the ambiguity of the real issued queries by users. However, based on our JDsearch dataset gathered from real user behaviors, we observe that the source of query terms can be the brands, names, or even detailed descriptions of products, not only the categories of the products they want to purchase. Specifically, depending on the sources of the terms, we categorize queries in JDsearch dataset into the following types:\\n• Category: All the query terms belong to the interacted product's category (e.g., cellphone). • Brand: All the query terms belong to the interacted product's brand (e.g., Apple). • Name: All the query terms belong to the interacted product's name (e.g., Surface Pro 8). • Category&Brand: All the query terms belong to the interacted product's category and brand (e.g., Apple cellphone). • Category&Name: All the query terms belong to the interacted product's category and name (e.g., Laptop Surface). • Brand&Name: All the query terms belong to the interacted product's brand and name (e.g., Microsoft Surface).\\nWe show the type distribution of queries in Figure 4. We can observe that the terms of most queries issued by users only derive from the names of products. After that, users also usually enter queries based on the categories or the combination of categories and names. Besides, there also exist queries based on the brands of products. Noticing that in this part, we only investigate queries that can be categorized into the above types. There still exist many queries whose terms originate from other fields of products. These findings show that the naive strategy of constructing queries from categories of products is inadequate, a better way may be to utilize diverse fields of products to assemble pseudo queries. Then, we analyze all the history queries and show their ambiguity. To quantify the ambiguity of queries, we propose a metric interaction entropy(IE) for queries. The interaction entropy is a natural extension of click entropy [8] and can be calculated from the corresponding interactions under queries. Generally speaking, when the interaction entropy of a query is larger than or equal to one, we can infer that this query is an informational query and may be ambiguous. The interaction entropy is calculated as follows:\\nIE(𝑞) = ∑︁ 𝑝 ∈ I (𝑞) -𝑃 (𝑝 |𝑞) log 2 𝑃 (𝑝 |𝑞)(1)\\n𝑃 (𝑝 |𝑞) = |Interaction(𝑞, 𝑝)| 𝑝 ′ ∈ I (𝑞) |Interaction(𝑞, 𝑝 ′ )| , (2\\n)\\nwhere I (𝑞) is the collection of products interacted (including clicked, added to cart, followed, and purchased) with users under query 𝑞, 𝑃 (𝑝 |𝑞) is the percentage of interactions on product 𝑝 among all interactions under query 𝑞. we calculate the interaction entropy for all repeated queries in JDsearch dataset and show the numbers of queries whose IE is less than / equal to / larger than one in Table 4.\\nFrom the statistics, we can find that most queries in our corpus are ambiguous and different products have been interacted under them. This can also verify the personalization is necessary in our dataset. (the latest ten interactions). The support distribution of 𝑢 is calculated based on her own early histories (interactions before the latest ten ones). The overall distribution is calculated based on all users' early histories. The personal divergence of 𝑢 is calculated between her valid distribution and her support distribution. The overall divergence is calculated between her valid distribution and the overall distribution. All these divergences are calculated by JS divergence. We calculate the personal divergence and overall divergence of each user 𝑢 based on the first-level category and brand respectively. We argue that if user interests are continuous and personalization is rational, the personal divergence should be smaller than the overall divergence. We show the results in Figure 5.\\nIn the boxplot figure, since there must exist some users' late purchase behaviors are totally different from the early ones and the overall preferences, the upper bound of the personal divergences and overall divergences are all close to log 2, which is the upper bounds of JS divergence. However, we can infer that most users' personal divergence is smaller than the overall divergence, which suggests that most users' interests in JDsearch dataset are in consistency and personalization can be effective in our dataset.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"EXPERIMENTS AND ANALYSIS\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The JDsearch dataset can be used for personalized product search studies to conduct experiments. In this part, we evaluate representative personalized product search models in our JDsearch dataset to verify the feasibility of performing personalization in this dataset. Further, we also conduct experiments on several dataset variations to show the different characteristics of our dataset compared with Amazon datasets.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Settings and Evaluation Metrics\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In our experiments, we train the models using the behaviors with queries in user histories. For inference, as we mentioned in Section 4, we use the last queries of users as the test queries. So we apply the trained models to rank the candidate product lists to evaluate their performances. For evaluation metrics, we use MRR@200, Precision@1, and NDCG@10 to evaluate the ranking results.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Benchmark models\",\n \"publication_ref\": [\n \"b15\",\n \"b4\",\n \"b0\",\n \"b5\",\n \"b1\"\n ],\n \"table_ref\": [],\n \"text\": \"We experiment with the following ad-hoc and personalized models: BM25: BM25 [14] is a classical sparse ad-hoc retrieval model. QEM: QEM only considers the matching scores between products and queries and can be regarded as a neural ad-hoc model. HEM: HEM [3] is a latent vector based personalized model. It builds the representations of users and items by generative language models based on reviews.\\nDREM: DREM [4] is a KG based personalized model. It utilizes the metadata of items to establish a knowledge graph.\\nAEM, ZAM: AEM [1] is an attention-based personalized model. It aggregates the user historical interacted items with the current query to construct a query-specific user profile. ZAM improves AEM by concatenating a zero vector to the item list to adjust the extent of conducting personalization.\\nTEM: TEM [5] is a transformer-based personalized model. It upgrades the attention layer in AEM with transformer encoder.\\nHGN: HGN [2] integrates DREM and ZAM models. It uses the relations in KG to boost the representations of products and users.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Implementation Details\",\n \"publication_ref\": [\n \"b14\",\n \"b2\",\n \"b4\"\n ],\n \"table_ref\": [],\n \"text\": \"As we anonymize the textual information in the dataset, which is much sparser than Amazon datasets, it is hard to optimize the word embeddings from scratch. So we adopt word2vec [13] to initialize the word embedding in models. Besides, different from the settings in Amazon datasets, products in the test set may have not been seen in the training process (they may be cold products). In addition to the large scale of product corpus, simply optimizing the product embedding table as previous models may cause extreme overfitting. Therefore, we obtain products' representations in models by calculating dynamic vectors based on their texts including title, brand, and category. We use the average term vectors of products' texts and apply a simple non-linear function to obtain product representations. Due to the privacy protection policy, we don't have the review information of users and products. Thus, we remove the generative language modeling loss for items and users from all models. Thus, in HEM [3] and DREM [4], again, we apply the non-linear function to build user embeddings from all the terms in their history interacted items. For the KG-enhanced models, as JD commercial platform doesn't provide the related product relationships (such as \\\"bought together\\\" and \\\"also bought') to us, so we only utilize the category and brand information to build the item-attribute graph. For the hyper-parameter settings, we set the maximum user history length in all models as 50. We set the embedding dimension as 128 and train them for 30 epochs. The number of transformer layers in transformer-based models is chosen from {1, 2}. The number of attention heads in attention-based models is set as 8. For all negative sampling in models, we randomly choose five negative samples in uniform distribution over the corpus.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Overall Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_6\"\n ],\n \"text\": \"The overall results are shown in Table 5. We can find that all the neural product search models overperform the sparse retrieval method BM25, which shows the effectiveness of neural ranking models. Except for the KG-enhanced DREM and HGN models, all the personalized product search models achieve improvements over the ad-hoc QEM models, demonstrating the feasibility of conducting personalization in our dataset. The poor performances of KG-based models may result from that we don't have efficient meta relationships among products and attributes. A promising way of solving this issue is to select some user histories to build a denser knowledge graph. This experiment proves that it is possible to train and evaluate personalized product search models in JDsearch dataset. However, the improvements over the ad-hoc search models are relatively incremental, which indicates that there still exists potential research topics for more accurate personalization using our JDsearch dataset.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Dataset Ablation Study\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_7\"\n ],\n \"text\": \"As we introduced in Section 1, some manipulations in previous Amazon datasets are obstacles for personalized product search research. For example, the pseudo queries generated artificially may be different from the real queries issued by users. To further show the characteristics and advantages of our JDsearch dataset, we resemble some operations in processing Amazon datasets and adjust the construction procedure of JDsearch dataset to produce several dataset variants. These variants include : JDsearch fakequery : We replace all the user historical queries with the artificial queries generated by concatenating the corresponding interacted products' category terms but keep the real test queries.\\nJDsearch samecate : We only reserve the products and their corresponding queries (if they have any) in user histories whose firstlevel category is the same as the products that the users finally interacted with under the test queries.\\nJDsearch w/o cold : We only reserve the products and their corresponding queries (if they have any) in user histories that have at least five interactions with users.\\nJDsearch short : We don't change the training process of models but only keep users whose history length is not larger than two in the test part.\\nJDsearch long : We don't change the training process of models but only keep users whose history length is larger than three hundred in the test part.\\nThen we train and evaluate the TEM model in these dataset variants and the results are shown in Table 6.\\nWe can find that all these performances on different dataset variants are significantly different. These results show that the manipulations in Amazon sub-category datasets have effects on the model's performances and may skew the evaluation of approaches. We can find by replacing the queries in the training set with the pseudo queries in the JDsearch fakequery dataset, the improve performance as it heuristically removes some noises from user histories. However, a more reasonable and rational way may be to reserve all products and leave the process of cold products to the model designer and researchers. From the performances in JDsearch short and JDsearch long dataset, we can find that personalized models usually can achieve high results on users who frequently do online shopping but perform badly while facing cold users. This finding can inspire future research applying different personalization strategies for users with different personalities.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"DISCUSSION OF APPLICATION SCENARIOS\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Besides utilizing it to run experiments and evaluate performances for personalized product search models, the JDsearch dataset can also support research in other areas. First, researchers can simply ignore the query information or remove the interactions under queries in the dataset to use it to assess product recommendation models. In addition, since we record all user behaviors from diverse channels including both search and recommendation, the JDsearch dataset can also support studies in the unified recommendation and search model, which has no publicly available datasets to the best of our knowledge.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"CONCLUSION\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this work, we introduce a new personalized product search JDsearch dataset collected from real user behaviors. Different from the simulated Amazon sub-category dataset, our dataset includes real user queries. Besides, there also exist products belonging to various categories. Cold products and users are also included. These lead the JDsearch dataset closer to the real product search situations. We also investigate the characteristics of the dataset from several perspectives and test existing personalized product search models. These analyses and experiments verify the feasibility of the proposed dataset. This dataset can also support some potential personalization directions including few-shot scenarios, multi-interest modeling, and separate strategies for different types of users.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"ACKNOWLEDGMENTS Zhicheng Dou is the corresponding author. This work is done during Jiongnan Liu's internship at JD. This work was supported by the National Key R&D Program of China (2022ZD0120103), National Natural Science Foundation of China (62272467), Beijing Outstanding Young Scientist Program NO. BJJWZYJH012019100020098, Public Computing Cloud, Renmin University of China, and Intelligent Social Governance Platform, Major Innovation & Planning Interdisciplinary Platform for the \\\"Double-First Class\\\" Initiative, Renmin University of China. The work was partially done at Beijing Key Laboratory of Big Data Management and Analysis Methods.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"https://github.com/rucliujn/JDsearch.\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-05-24"},"doi":{"kind":"string","value":"10.1145/3539618.3591900"},"references":{"kind":"list like","value":[{"authors":"Qingyao Ai; Daniel N Hill; S V N Vishwanathan; W Bruce Croft","journal":"ACM","ref_id":"b0","title":"A Zero Attention Model for Personalized Product Search","year":"2019-11-03"},{"authors":"Qingyao Ai; Lakshmi Narayanan; Ramasamy ","journal":"ACM","ref_id":"b1","title":"Model-agnostic vs. Modelintrinsic Interpretability for Explainable Product Search","year":"2021-11-01"},{"authors":"Qingyao Ai; Yongfeng Zhang; Keping Bi; Xu Chen; W Bruce Croft","journal":"","ref_id":"b2","title":"Learning a Hierarchical Embedding Model for Personalized Product Search","year":"2017-08-07"},{"authors":"","journal":"ACM","ref_id":"b3","title":"","year":""},{"authors":"Qingyao Ai; Yongfeng Zhang; Keping Bi; W Bruce Croft","journal":"ACM Trans. Inf. Syst","ref_id":"b4","title":"Explainable Product Search with a Dynamic Relation Embedding Model","year":"2020"},{"authors":"Keping Bi; Qingyao Ai; W Bruce Croft","journal":"ACM","ref_id":"b5","title":"A Transformer-based Embedding Model for Personalized Product Search","year":"2020-07-25"},{"authors":"Keping Bi; Qingyao Ai; W Bruce Croft","journal":"ACM","ref_id":"b6","title":"Learning a Fine-Grained Review-based Transformer Model for Personalized Product Search","year":"2021-07-11"},{"authors":"Dian Cheng; Jiawei Chen; Wenjun Peng; Wenqin Ye; Fuyu Lv; Tao Zhuang; Xiaoyi Zeng; Xiangnan He","journal":"ACM","ref_id":"b7","title":"IHGNN: Interactive Hypergraph Neural Network for Personalized Product Search","year":"2022-04-25"},{"authors":"Zhicheng Dou; Ruihua Song; Ji-Rong Wen","journal":"Association for Computing Machinery","ref_id":"b8","title":"A Large-Scale Evaluation and Analysis of Personalized Search Strategies (WWW '07)","year":"2007"},{"authors":"Yangyang Guo; Zhiyong Cheng; Liqiang Nie; Yinglong Wang; Jun Ma; Mohan S Kankanhalli","journal":"ACM Trans. Inf. Syst","ref_id":"b9","title":"Attentive Long Short-Term Preference Modeling for Personalized Product Search","year":"2019"},{"authors":"Christophe Van Gysel; Maarten De Rijke; Evangelos Kanoulas","journal":"ACM","ref_id":"b10","title":"Learning Latent Vector Spaces for Product Search","year":"2016-10-24"},{"authors":"Jiongnan Liu; Zhicheng Dou; Qiannan Zhu; Ji-Rong Wen","journal":"","ref_id":"b11","title":"A Categoryaware Multi-interest Model for Personalized Product Search","year":"2022-04-25"},{"authors":"","journal":"ACM","ref_id":"b12","title":"","year":""},{"authors":"Shang Liu; Wanli Gu; Gao Cong; Fuzheng Zhang","journal":"ACM","ref_id":"b13","title":"Structural Relationship Representation Learning with Graph Embedding for Personalized Product Search","year":"2020-10-19"},{"authors":"Tomas Mikolov; Kai Chen; Greg Corrado; Jeffrey Dean","journal":"","ref_id":"b14","title":"Efficient Estimation of Word Representations in Vector Space","year":"2013"},{"authors":"Stephen Robertson; Hugo Zaragoza","journal":"Found. Trends Inf. Retr","ref_id":"b15","title":"The Probabilistic Relevance Framework: BM25 and Beyond","year":"2009-04"},{"authors":"Craig Silverstein; Monika Rauch Henzinger; Hannes Marais; Michael Moricz","journal":"SIGIR Forum","ref_id":"b16","title":"Analysis of a Very Large Web Search Engine Query Log","year":"1999"},{"authors":"Yuki Yano; Yukihiro Tagami; Akira Tajima","journal":"","ref_id":"b17","title":"Quantifying Query Ambiguity with Topic Distributions","year":"2016"}],"string":"[\n {\n \"authors\": \"Qingyao Ai; Daniel N Hill; S V N Vishwanathan; W Bruce Croft\",\n \"journal\": \"ACM\",\n \"ref_id\": \"b0\",\n \"title\": \"A Zero Attention Model for Personalized Product Search\",\n \"year\": \"2019-11-03\"\n },\n {\n \"authors\": \"Qingyao Ai; Lakshmi Narayanan; Ramasamy \",\n \"journal\": \"ACM\",\n \"ref_id\": \"b1\",\n \"title\": \"Model-agnostic vs. Modelintrinsic Interpretability for Explainable Product Search\",\n \"year\": \"2021-11-01\"\n },\n {\n \"authors\": \"Qingyao Ai; Yongfeng Zhang; Keping Bi; Xu Chen; W Bruce Croft\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Learning a Hierarchical Embedding Model for Personalized Product Search\",\n \"year\": \"2017-08-07\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"ACM\",\n \"ref_id\": \"b3\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Qingyao Ai; Yongfeng Zhang; Keping Bi; W Bruce Croft\",\n \"journal\": \"ACM Trans. Inf. Syst\",\n \"ref_id\": \"b4\",\n \"title\": \"Explainable Product Search with a Dynamic Relation Embedding Model\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Keping Bi; Qingyao Ai; W Bruce Croft\",\n \"journal\": \"ACM\",\n \"ref_id\": \"b5\",\n \"title\": \"A Transformer-based Embedding Model for Personalized Product Search\",\n \"year\": \"2020-07-25\"\n },\n {\n \"authors\": \"Keping Bi; Qingyao Ai; W Bruce Croft\",\n \"journal\": \"ACM\",\n \"ref_id\": \"b6\",\n \"title\": \"Learning a Fine-Grained Review-based Transformer Model for Personalized Product Search\",\n \"year\": \"2021-07-11\"\n },\n {\n \"authors\": \"Dian Cheng; Jiawei Chen; Wenjun Peng; Wenqin Ye; Fuyu Lv; Tao Zhuang; Xiaoyi Zeng; Xiangnan He\",\n \"journal\": \"ACM\",\n \"ref_id\": \"b7\",\n \"title\": \"IHGNN: Interactive Hypergraph Neural Network for Personalized Product Search\",\n \"year\": \"2022-04-25\"\n },\n {\n \"authors\": \"Zhicheng Dou; Ruihua Song; Ji-Rong Wen\",\n \"journal\": \"Association for Computing Machinery\",\n \"ref_id\": \"b8\",\n \"title\": \"A Large-Scale Evaluation and Analysis of Personalized Search Strategies (WWW '07)\",\n \"year\": \"2007\"\n },\n {\n \"authors\": \"Yangyang Guo; Zhiyong Cheng; Liqiang Nie; Yinglong Wang; Jun Ma; Mohan S Kankanhalli\",\n \"journal\": \"ACM Trans. Inf. Syst\",\n \"ref_id\": \"b9\",\n \"title\": \"Attentive Long Short-Term Preference Modeling for Personalized Product Search\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Christophe Van Gysel; Maarten De Rijke; Evangelos Kanoulas\",\n \"journal\": \"ACM\",\n \"ref_id\": \"b10\",\n \"title\": \"Learning Latent Vector Spaces for Product Search\",\n \"year\": \"2016-10-24\"\n },\n {\n \"authors\": \"Jiongnan Liu; Zhicheng Dou; Qiannan Zhu; Ji-Rong Wen\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"A Categoryaware Multi-interest Model for Personalized Product Search\",\n \"year\": \"2022-04-25\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"ACM\",\n \"ref_id\": \"b12\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Shang Liu; Wanli Gu; Gao Cong; Fuzheng Zhang\",\n \"journal\": \"ACM\",\n \"ref_id\": \"b13\",\n \"title\": \"Structural Relationship Representation Learning with Graph Embedding for Personalized Product Search\",\n \"year\": \"2020-10-19\"\n },\n {\n \"authors\": \"Tomas Mikolov; Kai Chen; Greg Corrado; Jeffrey Dean\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Efficient Estimation of Word Representations in Vector Space\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Stephen Robertson; Hugo Zaragoza\",\n \"journal\": \"Found. Trends Inf. Retr\",\n \"ref_id\": \"b15\",\n \"title\": \"The Probabilistic Relevance Framework: BM25 and Beyond\",\n \"year\": \"2009-04\"\n },\n {\n \"authors\": \"Craig Silverstein; Monika Rauch Henzinger; Hannes Marais; Michael Moricz\",\n \"journal\": \"SIGIR Forum\",\n \"ref_id\": \"b16\",\n \"title\": \"Analysis of a Very Large Web Search Engine Query Log\",\n \"year\": \"1999\"\n },\n {\n \"authors\": \"Yuki Yano; Yukihiro Tagami; Akira Tajima\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Quantifying Query Ambiguity with Topic Distributions\",\n \"year\": \"2016\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[5,371.15,416,187.59,21.99],"formula_id":"formula_0","formula_text":"IE(𝑞) = ∑︁ 𝑝 ∈ I (𝑞) -𝑃 (𝑝 |𝑞) log 2 𝑃 (𝑝 |𝑞)(1)"},{"formula_coordinates":[5,366.14,443.22,189.43,21.78],"formula_id":"formula_1","formula_text":"𝑃 (𝑝 |𝑞) = |Interaction(𝑞, 𝑝)| 𝑝 ′ ∈ I (𝑞) |Interaction(𝑞, 𝑝 ′ )| , (2"},{"formula_coordinates":[5,555.57,449.86,3.17,7.94],"formula_id":"formula_2","formula_text":")"}],"string":"[\n {\n \"formula_coordinates\": [\n 5,\n 371.15,\n 416,\n 187.59,\n 21.99\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"IE(𝑞) = ∑︁ 𝑝 ∈ I (𝑞) -𝑃 (𝑝 |𝑞) log 2 𝑃 (𝑝 |𝑞)(1)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 366.14,\n 443.22,\n 189.43,\n 21.78\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"𝑃 (𝑝 |𝑞) = |Interaction(𝑞, 𝑝)| 𝑝 ′ ∈ I (𝑞) |Interaction(𝑞, 𝑝 ′ )| , (2\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 555.57,\n 449.86,\n 3.17,\n 7.94\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \")\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"JDsearch: A Personalized Product Search Dataset with Real Queries and Full Interactions"},"abstract":{"kind":"string","value":"Recently, personalized product search attracts great attention and many models have been proposed. To evaluate the effectiveness of these models, previous studies mainly utilize the simulated Amazon recommendation dataset, which contains automatically generated queries and excludes cold users and tail products. We argue that evaluating with such a dataset may yield unreliable results and conclusions, and deviate from real user satisfaction. To overcome these problems, in this paper, we release a personalized product search dataset comprised of real user queries and diverse userproduct interaction types (clicking, adding to cart, following, and purchasing) collected from JD.com, a popular Chinese online shopping platform. More specifically, we sample about 170,000 active users on a specific date, then record all their interacted products and issued queries in one year, without removing any tail users and products. This finally results in roughly 12,000,000 products, 9,400,000 real searches, and 26,000,000 user-product interactions. We study the characteristics of this dataset from various perspectives and evaluate representative personalization models to verify its feasibility. The dataset can be publicly accessed at Github:"},"authors":{"kind":"string","value":"Jiongnan Liu; Zhicheng Dou; Guoyu Tang; Sulong Xu"},"figures":{"kind":"list like","value":[{"figure_caption":"4 https://github.com/rucliujn/JDsearch","figure_data":"","figure_id":"fig_0","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 1 :1Figure 1: The log-log distribution of product's interaction frequency","figure_data":"","figure_id":"fig_2","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: The log-log distribution of user history length","figure_data":"","figure_id":"fig_4","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: The category sizes of products in user histories","figure_data":"","figure_id":"fig_5","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Distribution of query types","figure_data":"","figure_id":"fig_6","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: The boxplot of the personal and overall interest divergence of each user","figure_data":"","figure_id":"fig_7","figure_label":"5","figure_type":"figure"},{"figure_caption":"Field and explanation in the datasets","figure_data":"FiledExplanationUser Behavior Dataquerythe anonymized term ids of the test query.candidate_wid_listthe anonymized id list of candidate prod-ucts displayed under the test query.candidate_label_listthe corresponding label for the candidateproducts.history_qry_listthe sequence of anonymized term ids ofissued queries in user histories.history_wid_listthe sequence of anonymized ids of prod-ucts in user histories.history_type_listthe sequence of interaction levels in userhistories.history_time_listthe sequence of timestamps of interactionsin user histories.Product Meta Datawidthe anonymized id of the product.namethe anonymized term ids of the product'sname.brand_idthe anonymized id of the product's brand.brand_namethe anonymized term ids of the product'sbrand name.category_id_{1,2,3,4}the anonymized ids of the the product'sfour level categories.category_name_{1,2,3,4} the anonymized term ids of the product'sfour level categories' names.shop_idthe anonymized id of the product's shop.","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"characteristics of the datasets","figure_data":"CharacteristicAmazon datasetsJDsearch datasetQueryArtificialRealItem popularity5-core itemsAll itemsItem categorySame categoryDiverse categoriesUser popularity5-core usersAll usersInteraction typePurchaseClick, add to cart, follow, purchasesummarize the characteristics of JDsearch dataset and the previ-ous sub-category dataset in","figure_id":"tab_1","figure_label":"2","figure_type":"table"},{"figure_caption":"In the following part, we will analyze our dataset to demonstrate these advantages.","figure_data":"","figure_id":"tab_2","figure_label":"2","figure_type":"table"},{"figure_caption":"Statistics of the datasetsThe numbers of test queries in Amazon datasets are calculated based on the sequential-based dataset division in previous works[5].","figure_data":"DatasetCell Phones & Accessories Clothing, Shoes & Jewelry Sports & Outdoors Electronics JDsearch#Users27,87939,38735,598192,403173,831#Items10,42923,03318,35763,00112,872,636#Interactions194,43963,001278,6771,689,188 26,667,260#Test Queries *4266,9394723,221171,728","figure_id":"tab_3","figure_label":"3","figure_type":"table"},{"figure_caption":"Interaction entropy distribution of repeated queries in the JDsearch dataset Query Ambiguity Analysis. First, we analyze the test queries and their candidates in our dataset. We provide how many test queries have candidate products that come from various first-level categories (category ambiguous) and various brands (brand ambiguous). Among all 171,728 test queries, 114,955 are category ambiguous, 109,874 are brand ambiguous, and 68,478 are both category and brand ambiguous. Through these statistics, we can find that most test queries in our JDsearch dataset are vague. In this case, it is hard for simple ad-hoc ranking methods to provide satisfying results for users, so personalization is required.","figure_data":"EntropyIE < 1.0IE = 1.0IE > 1.0#Queries 147,095 (18.69%) 231,638 (29.44%) 408,195(51.87%)4.3 Potential of PersonalizationIn this part, we analyze the personalization potential in our JDsearchdataset.4.3.1","figure_id":"tab_4","figure_label":"4","figure_type":"table"},{"figure_caption":"Overall performances of models. The best and the second results are denoted in bold and underlined fonts respectively.","figure_data":"ModelMRRPrecNDCGAd-hocBM25 QEM0.1114 0.17740.0402 0.07280.0940 0.1705HEM0.19550.08470.1905DREM0.16470.06320.1578PersonalizedHGN AEM0.1662 0.19710.0634 0.08510.1591 0.1920ZAM0.19690.08490.1920TEM0.22290.10490.2192","figure_id":"tab_6","figure_label":"5","figure_type":"table"},{"figure_caption":"The performances of TEM models in different dataset variants. performance while facing real queries issued by users can be poor. Only including products belonging to the same category also harms the user interest construction as evidenced by the results on JDsearch samecate . The operations of removing the cold products from user histories in JDsearch w/o cold can help models","figure_data":"DatasetMRRPrecNDCGJDsearch0.22290.10490.2192JDsearch fakequery0.16440.06610.1539JDsearch samecate0.19250.08400.1869JDsearch w/o cold0.22960.10880.2272JDsearch short0.20800.10120.1979JDsearch long0.22940.10960.2215model's","figure_id":"tab_7","figure_label":"6","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"4 https://github.com/rucliujn/JDsearch\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: The log-log distribution of product's interaction frequency\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: The log-log distribution of user history length\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: The category sizes of products in user histories\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Distribution of query types\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: The boxplot of the personal and overall interest divergence of each user\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Field and explanation in the datasets\",\n \"figure_data\": \"FiledExplanationUser Behavior Dataquerythe anonymized term ids of the test query.candidate_wid_listthe anonymized id list of candidate prod-ucts displayed under the test query.candidate_label_listthe corresponding label for the candidateproducts.history_qry_listthe sequence of anonymized term ids ofissued queries in user histories.history_wid_listthe sequence of anonymized ids of prod-ucts in user histories.history_type_listthe sequence of interaction levels in userhistories.history_time_listthe sequence of timestamps of interactionsin user histories.Product Meta Datawidthe anonymized id of the product.namethe anonymized term ids of the product'sname.brand_idthe anonymized id of the product's brand.brand_namethe anonymized term ids of the product'sbrand name.category_id_{1,2,3,4}the anonymized ids of the the product'sfour level categories.category_name_{1,2,3,4} the anonymized term ids of the product'sfour level categories' names.shop_idthe anonymized id of the product's shop.\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"characteristics of the datasets\",\n \"figure_data\": \"CharacteristicAmazon datasetsJDsearch datasetQueryArtificialRealItem popularity5-core itemsAll itemsItem categorySame categoryDiverse categoriesUser popularity5-core usersAll usersInteraction typePurchaseClick, add to cart, follow, purchasesummarize the characteristics of JDsearch dataset and the previ-ous sub-category dataset in\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"In the following part, we will analyze our dataset to demonstrate these advantages.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Statistics of the datasetsThe numbers of test queries in Amazon datasets are calculated based on the sequential-based dataset division in previous works[5].\",\n \"figure_data\": \"DatasetCell Phones & Accessories Clothing, Shoes & Jewelry Sports & Outdoors Electronics JDsearch#Users27,87939,38735,598192,403173,831#Items10,42923,03318,35763,00112,872,636#Interactions194,43963,001278,6771,689,188 26,667,260#Test Queries *4266,9394723,221171,728\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Interaction entropy distribution of repeated queries in the JDsearch dataset Query Ambiguity Analysis. First, we analyze the test queries and their candidates in our dataset. We provide how many test queries have candidate products that come from various first-level categories (category ambiguous) and various brands (brand ambiguous). Among all 171,728 test queries, 114,955 are category ambiguous, 109,874 are brand ambiguous, and 68,478 are both category and brand ambiguous. Through these statistics, we can find that most test queries in our JDsearch dataset are vague. In this case, it is hard for simple ad-hoc ranking methods to provide satisfying results for users, so personalization is required.\",\n \"figure_data\": \"EntropyIE < 1.0IE = 1.0IE > 1.0#Queries 147,095 (18.69%) 231,638 (29.44%) 408,195(51.87%)4.3 Potential of PersonalizationIn this part, we analyze the personalization potential in our JDsearchdataset.4.3.1\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Overall performances of models. The best and the second results are denoted in bold and underlined fonts respectively.\",\n \"figure_data\": \"ModelMRRPrecNDCGAd-hocBM25 QEM0.1114 0.17740.0402 0.07280.0940 0.1705HEM0.19550.08470.1905DREM0.16470.06320.1578PersonalizedHGN AEM0.1662 0.19710.0634 0.08510.1591 0.1920ZAM0.19690.08490.1920TEM0.22290.10490.2192\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"The performances of TEM models in different dataset variants. performance while facing real queries issued by users can be poor. Only including products belonging to the same category also harms the user interest construction as evidenced by the results on JDsearch samecate . The operations of removing the cold products from user histories in JDsearch w/o cold can help models\",\n \"figure_data\": \"DatasetMRRPrecNDCGJDsearch0.22290.10490.2192JDsearch fakequery0.16440.06610.1539JDsearch samecate0.19250.08400.1869JDsearch w/o cold0.22960.10880.2272JDsearch short0.20800.10120.1979JDsearch long0.22940.10960.2215model's\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Data Source\", \"Citation\": \"[1-6, 9, 11]\", \"Explanation\": \"The cited works are the most widely-used personalized product search datasets based on real user behaviors, which serve as the foundational data for the research conducted in the citing paper on personalized product search.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[3,10]\", \"Explanation\": \"The cited works are used to build pseudo queries in the amazon online shopping websites datasets, but the generated queries may not fully reflect the ambiguity of actual queries in real product search scenarios.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[7]\", \"Explanation\": \"The cited work is the source of the DIGINETICA dataset, which is used in the citing paper for personalization experiments but with a limitation of only including behaviors with user ids.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[12]\", \"Explanation\": \"The cited work is the source of the CIKMCUP2016 dataset, which is also utilized in the citing paper for research purposes but with a small and limited dataset after filtering anonymous behaviors.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[10]\", \"Explanation\": \"The cited work is the source of the heuristic query generation process used in Amazon datasets, which is discussed in Section 1 of the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[8]\", \"Explanation\": \"The cited work introduces the concept of click entropy, which the citing paper extends to develop a new metric for quantifying query ambiguity, called interaction entropy.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work introduces the BM25 model, which the citing paper adopts as a classical ad-hoc retrieval model in their experiments.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[3]\", \"Explanation\": \"The cited work introduces the HEM model, which the citing paper uses as a latent vector based personalized model in their experiments.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[4]\", \"Explanation\": \"The cited work introduces the DREM model, which the citing paper utilizes as a knowledge graph based personalized model in their experiments.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[1]\", \"Explanation\": \"The cited work introduces the AEM model, which the citing paper adopts as an attention-based personalized model in their experiments.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[2]\", \"Explanation\": \"The cited work introduces the HGN model, which the citing paper integrates with the DREM and ZAM models to create a hybrid model in their experiments.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work, word2vec, is used to initialize the word embedding in the models in the citing paper. The cited work provides the method and technique for obtaining the word embedding.\"}]"}}},{"rowIdx":74,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_0","fig_0","fig_0","fig_0"],"heading":"Introduction","publication_ref":["b40","b50","b38","b66","b5","b57","b42","b90","b62","b63","b0","b73","b81","b2","b78","b58","b52","b39","b15","b69","b42","b15","b42","b15","b69","b42","b50","b4","b59","b72","b76","b86","b70","b91","b91"],"table_ref":["tab_9"],"text":"Though object detection has been significantly advanced in the supervised learning domain by neural network-based detectors [41,51,39,67,6]), there is still a large room for improvement in semi-supervised object detection (SSOD). In practice, SSOD is desirable because annotating bounding boxes and their object classes are both costly and time-consuming. Most existing semi-supervised object detectors [58,43,91,63,64,1,74,82]) are learned by estimated pseudo-labels, which are assigned to bounding box proposals and filtered by a single fixed confidence threshold. Such a combination of pseudo-labeling and confidence thresholdsbased filtering has been largely inspired by research on semi-supervised image classification [3,79,59,53]).\nMost existing studies are conducted on the COCO dataset [40] that has curated categories and highly balanced data distributions. However, real-world problems are much more challenging than what the COCO dataset represents in that data distributions are often long-tailed, i.e., a majority of classes have only a few labeled images, which could easily result in an extremely biased detector. In recent years, the research community has paid increasing attention to long-tailed object detection, with several relevant datasets released, such as LVIS [16] and COCO-LT [70]. However, to our knowledge, none of the existing studies has been devoted to long-tailed object detection in the semi-supervised setting, a more challenging yet practical problem.\nImplementing semi-supervised object detection algorithms on long-tailed datasets is not trivial. By train- ing a state-of-the-art semi-supervised detector, i.e., Unbiased Teacher [43], using a long-tailed LVIS [16] dataset, we identify the following three major problems. First, a fixed confidence threshold often fails to provide a good trade-off between precision and recall. The shortcoming is evidenced in Figure 1(a), which shows none of the commonly used thresholds gives the best performance in both the AP and AR metrics, e.g., a fixed threshold of 0.6 returns the highest recall but has the lowest precision. Second, by digging deeper into the distribution of prediction scores, we observe that the model's predictions are biased toward the frequent classes (see Figure 1(b)). Finally, we identify the reason why using a fixed threshold leads to low confidenceand hence low prediction accuracy-on the common and rare classes: the model's exposure to these classes during training is substantially reduced compared to that to the frequent classes (see Figure 1(c)).\nTo overcome these problems, we propose Cascade-Match, a novel pseudo-labeling-based approach to addressing long-tailed and semi-supervised object detection. Specifically, CascadeMatch features a cascade pseudo-labeling (CPL) design, which contains multistage detection heads. To control the precision-recall trade-off, we set progressive confidence thresholds for detection heads to focus on different parts. The early detection head is assigned a small confidence threshold to improve recall, while the subsequent heads are assigned larger confidence thresholds to ensure precision. The use of multiple heads also allows the unique chance for us to deal with confirmation bias -a phenomenon where a model is iteratively reinforced by incorrect pseudo labels produced by itself. In particular, we show the possibility of using ensemble predictions from all detection heads as the teacher's supervision signal to obtain more reliable pseudo labels for training each individual detection head. To deal with the issue of biased prediction score distributions to frequent classes, we propose an adaptive pseudo-label mining mechanism (APM) that automatically identifies suitable class-wise threshold values from data with minimal human intervention. As shown in Figure 1(c), with the APM module, our approach can retain more pseudolabels for common and rare classes than the previous SOTA approach [43], boosting the performance for classes with small sample sizes.\nWe present comprehensive experiments on two challenging long-tailed object detection datasets, namely LVIS v1.0 [16] and COCO-LT [70], under the SSOD setting. Overall, CascadeMatch achieves the best performance on both datasets in all metrics. Notably, on LVIS, CascadeMatch improves upon the most competitive method, i.e., Unbiased Teacher [43], by 2.3% and 1.8% AP Fix in the rare and common classes, which confirm the effectiveness of our design for long-tailed data. Importantly, CascadeMatch is general and obtains consistent improvements across a variety of detection architectures, covering both anchor-based R-CNN detectors [51,5] and the recent Sparse R-CNN detector [60] with the Pyramid Vision Transformer encoder (PVT) [73] (Table 7). We also conduct various ablation studies to confirm the effectiveness of each of our proposed modules.\nWe also apply CascadeMatch to another challenging sparsely-annotated object detection (SAOD) setting [77,87,71,92] where training data are only partially annotated and contain missing annotated instances. Again, CascadeMatch yields considerable improvements over the supervised-only baseline and a state-of-the-art method [92] (Table 10). Finally, we provide several qualitative results and analyses to show that our proposed CascadeMatch method generates high-quality pseudo labels on both SSOD and SAOD settings."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b53","b45","b64","b68","b54","b25","b12","b35","b57","b62","b26","b90","b81","b79","b85","b41","b8","b7","b31","b44","b14","b33","b41","b25","b26","b57","b42","b63","b81","b79","b85","b8","b7","b63","b90","b63","b42","b38","b73","b90","b33","b7","b8","b41","b39","b15","b25","b57","b42","b79","b50","b38","b4","b66","b5","b59","b88","b15","b56","b21","b75","b60","b49","b71","b61","b87","b67","b11","b6","b91","b32","b19","b83","b34","b13","b15","b36","b69","b60","b61","b67","b84","b55","b2","b3","b58","b89","b80","b2","b78","b58","b78","b1","b29","b24","b77","b46","b58","b48","b28","b65","b42","b0","b58","b0","b52","b16","b47","b23","b27","b82","b74","b30","b10","b46","b18","b20"],"table_ref":[],"text":"Semi-Supervised Object Detection has been a topical research area due to its importance to practical applications [54,46,65,69,55,26,13,36,58,63,27,91,82,80,86,42,9,8,32,45,15,34,42]. Various semi-supervised object detectors have been proposed in the literature, and many of them borrow ideas from the semi-supervised learning (SSL) community.\nIn CSD [26] and ISD [27], consistency regularization is applied to the mined bounding boxes for unlabeled images. STAC [58] uses strong data augmentation for self-training.\nRecently, pseudo-labeling-based methods have shown promising results on several benchmark datasets, which are attributed to a stronger teacher model trained by, e.g., a weighted EMA ensemble [43,64,82,80,86,9,8], a data ensemble [64], or advanced data augmentation [91,64]. To overcome the confirmation bias, Unbiased Teacher [43] employs focal loss [39] to reduce the weights on overconfident pseudo labels, while others use uncertainty modeling [74] or co-training [91] as the countermeasure. Li, et al. [34] propose dynamic thresholding for each class based on both localization and classification confidence. LabelMatch [8] introduces a re-distribution mean teacher based on the KL divergence distribution between teacher and student models. DSL [9] assigns pixel-wise pseudo-labels for anchor-free detectors. Unbiased Teacherv2 [42] introduces a new pseudolabeling mechanism based on the relative uncertainties of teacher and student models.\nIt is worth noting that most existing methods are designed for class-balanced datasets like MS COCO [40], while their capabilities to handle longtailed datasets like LVIS [16] have been largely understudied-to our knowledge, none of existing research has specifically investigated long-tailed object detection in the SSL setting. Instead, the majority of existing SSL algorithms are evaluated on class-balanced datasets [26,58,43,80]. Our work takes the first step toward a unified approach to solving unlabeled data and the longtailed object detection problem, which we hope to inspire more work to tackle this challenging setting.\nLong-tailed Object Detection Though object detection has witnessed significant progress in recent years [51,39,5,67,6,60], how to deal with the long-tailed problem remains an open question [89]. Most existing methods fall into two groups: data resampling [16,57,22,76] and loss re-weighting [61,50,72,62,88,68,12,7,92,33,20]. Some recent works [84,35,14] suggest that data augmentation is useful for long-tailed recognition. In terms of data re-sampling, Repeated Factor Sampling (RFS) [16] assigns high sampling rates to images of rare classes. A couple of studies [37,70] have suggested using different sampling schemes in decoupled training stages. When it comes to data re-weighting, a representative method is equalization loss [61,62], which raises the weights for rare classes based on inverse class frequency. Seesaw Loss [68] automatically adjusts class-specific loss weights based on a statistical ratio between the positive and negative gradients computed for each class. MosaicOS [85] is one of the early studies that uses weakly-supervised learning to help long-tailed detection. Their study assumes the availability of weaklyannotated class labels. In contrast, we take a pure semisupervised setting without assuming any annotations in the unlabeled set. In our work, we first investigate how to exploit unlabeled data to improve the performance of detectors trained on long-tailed datasets.\nSemi-Supervised Learning (SSL) Numerous SSL methods are based on consistency learning [56,3,4,59,90,81], which forces a model's predictions on two different views of the same instance to be similar. Recent state-of-the-art consistency learning methods like MixMatch [3], UDA [79] and FixMatch [59] introduce strong data augmentations [79] to the learning paradigm-they use predictions on weakly augmented images as the target to train the model to produce similar outputs given the strongly augmented views of the same images.\nAnother research direction related to our work is pseudo-labeling [2,30,25,78,47], which is typically based on a teacher-student architecture: a teacher model's predictions are used as the target to train a student model. The teacher model can be either a pretrained model [59] or an exponential moving average of the student model [49,29,66,43]. Some studies [1] have also demonstrated that using the student model being trained to produce the target can reach decent performance-the trick is to inject strong noise to the student model, such as applying strong data augmentations to the input [59].\nA common issue encountered in pseudo-labeling methods is confirmation bias [1], which is caused by a constant feed of incorrect pseudo labels with high confidence to the model. And such a vicious cycle would reinforce since the model will become increasingly inaccurate and subsequently provide more erroneous pseudo labels. To mitigate the issue of confirmation bias, existing methods have tried using an uncertainty-based metric [53] to modulate the confidence threshold or using the co-training framework [17,48] that simultaneously trains two neural networks each giving pseudo labels to the other. In this work, to prevent each detection head from overfitting its own prediction errors, the pseudo labels to train each detection head are formed by the ensemble predictions of multiple detection heads. This strategy is new in the literature.\nIt is worth noting that most aforementioned algorithms are evaluated on class-balanced datasets while only very few recent works apply SSL for long-tailed image classification [24,28,83,75,31,11,47] or semantic segmentation [19,21]. The detection task requires predicting both the class labels and object locations, which is much harder than the classification-only task. The pseudo-labeling-based semi-supervised methods are unable to predict high-quality pseudo labels for detection task as accurately as for classification task, in the presence of class imbalance. This motivates us to improve the pseudo-labeling quality for semi-supervised and long-tailed detection using a cascade mechanism."},{"figure_ref":[],"heading":"Our Approach: CascadeMatch","publication_ref":["b15"],"table_ref":[],"text":"Problem Definition Given a labeled dataset D l = {(x, y * , b * )} with x, y * and b * denoting image, label and bounding box, respectively,1 and an unlabeled dataset D u = {x}, the goal is to learn a robust object detector using both D l and D u . We further consider the issue of long-tailed distribution [16], which is common in real-world data but have been largely unexplored in existing semi-supervised object detection methods. More specifically, let n i and n j denote the number of images for class i and j respectively, and assume i is a frequent class while j is a rare class. In a long-tailed scenario, we might have n i ≫ n j ."},{"figure_ref":["fig_1"],"heading":"An Overview","publication_ref":["b4","b59","b17","b43"],"table_ref":[],"text":"A brief overview of the main paradigm of our proposed CascadeMatch is illustrated in Figure 2. CascadeMatch features a cascade pseudolabeling (CPL) design and an adaptive pseudo-label mining (APM) mechanism. The former aims to generate pseudo-labels and filter out low-quality labels in a cascade fashion to improve the trade-off between precision and recall, while the latter aims to automate threshold tuning. CascadeMatch only modifies a detector's head structure and thus can be seen as a plug-andplay module that fits into most existing object detectors including the popular anchor-based R-CNN series like Cascade R-CNN [5] or more recent end-to-end detectors like Sparse R-CNN [60]. CascadeMatch can also take either CNNs [18] or Transformers [44] as the backbone. Discussion A cascade structure benefits from the \"divide and conquer\" concept, where each stage is ded-icated to a specific sub-task. This notion of cascading has been found practical and useful in many computer vision systems. For the detection task, finding an accurate IoU threshold to separate the positive and negative region proposals is impossible. To allow a better precision-recall trade-off, Cascade R-CNN uses the cascade structure to progressively increase the IoU threshold for different stages. Recall that pseudo labeling faces a similar dilemma in pinpointing a single confidence threshold to separate the valid pseudo-labels and noisy background region proposals. It is thus natural for Cas-cadeMatch to use the cascade structure with a set of progressive confidence thresholds. Note that the confidence threshold of CascadeMatch is class-specific and self-adaptive. We will provide the details in Section 3.2.\nBelow we provide the technical details of the two key components in CascadeMatch, namely cascade pseudolabeling (Section 3.1) and adaptive pseudo-label mining (Section 3.2). For clarity, in Section 3.1 we first present CascadeMatch in an anchor-based framework and later explain the modifications needed for an end-to-end detector."},{"figure_ref":["fig_1"],"heading":"Cascade Pseudo-Labeling","publication_ref":["b50","b4","b17","b50","b4","b4","b59"],"table_ref":[],"text":"Model Architecture For an anchor-based framework [51,5], the CascadeMatch-based detector starts with a CNN as the backbone for feature extraction, e.g., ResNet50 [18], which is then followed by a region proposal network (RPN) [51] for generating object proposals. See Figure 2(a) for the architecture.\nThe detector has K heads following the Cascade R-CNN [5] pipeline. The parameter K controls the tradeoff between performance and efficiency, which can be adjusted by practitioners based on their needs. Increasing the number of heads will improve the performance at the cost of speed. In the paper, we followed previous cascade methods [5,60] to use K = 3 heads. We will provide the ablation studies of varying the value of K in Table 4 of Section 4.1. Formally, given an image x, the first-stage detection head predicts for an object proposal b 0 (generated by the RPN) a class probability distribution p 1 (y|x, b 0 ) and the bounding box offsets b 1 . Then, the second-stage detection head predicts another probability p 2 (y|x, b 1 ) using the refined bounding box from the first stage;2 and so on and so forth."},{"figure_ref":[],"heading":"Labeled Losses","publication_ref":["b50","b57","b42","b90","b62","b63","b73","b0"],"table_ref":["tab_3"],"text":"With labeled data D l = {(x, y * , b * )}, we train each detection head using the classification loss Cls(•, •) (for proposal classification) and the bounding box regression loss Reg(•, •) [51]. Formally, we have\nℓ labeled cls = (x,y * )∼D l K k=1 Cls(y * , p k (y|x, b k-1 )),(1)\nℓ labeled reg = (x,b * )∼D l K k=1 Reg(b * , b k ). (2\n)\nUnlabeled Losses To cope with unlabeled images, we adopt a pseudo-labeling approach with a teacherstudent architecture where the teacher's estimations on unlabeled data are given to the student as supervision. Such a paradigm has been widely used in previous semisupervised methods [58,43,91,63,64,74]. Different from previous methods, we focus on tackling the confirmation bias issue [1] when designing our architecture. We observe that the ensemble predictions are more accurate than using each individual prediction (please refer to Table 5 of Section 4.1 for more details), so we use the ensemble predictions from all detection heads as the teacher supervision signal (teacher module in Figure . 2 (a)). Formally, given an unlabeled image x ∼ D u , the ensemble prediction p t is computed as\np t = 1 K K k=1 p k (y|x, b k-1 ) and b t = 1 K K k=1 b k ,(3)\nwhere K is the number of heads. Let q t = max(p t ) be the confidence and qt = arg max(p t ) the pseudo label, we compute the classification loss and the bounding box regression loss for unlabeled data using\nℓ unlabeled cls = x∼Du K k=1 1(q t ≥ τ qt k ) Cls(q t , p k (y|x, b k-1 )),(4)\nℓ unlabeled reg = x∼Du K k=1 1(q t ≥ τ qt k ) Reg(b t , b k ),(5)\nwhere τ qt k is a self-adaptive confidence threshold specific to class qt . We detail the design of class-specific selfadaptive thresholds in Section 3.2."},{"figure_ref":[],"heading":"Training","publication_ref":["b50","b59","b5","b38","b51"],"table_ref":[],"text":"Similar to most region-based object detectors, our CascadeMatch model is learned using four losses: a region-of-interest (ROI) classification loss\nℓ roi cls = ℓ labeled cls + λ u • ℓ unlabeled cls , an ROI regression loss ℓ roi reg = ℓ labeled reg + λ u • ℓ unlabeled reg\n, and two other losses for the RPN, i.e., the objectness classification loss ℓ rpn cls and the proposal regression loss ℓ rpn reg , as defined in [51]. The loss parameter λ u controls the weight between the supervised term ℓ l cls and the unsupervised term ℓ u cls . By default, we set the unsupervised loss weight λ u = 1.0.\nTransfer to End-to-End Object Detector Cas-cadeMatch is readily applicable to an end-to-end detector. We use Sparse R-CNN [60] as an example. Two main modifications are required: 1) Since region proposals are learned from a set of embedding queries as in DETR [6], we do not need an RPN and the RPN loss ℓ rpn ; 2) The classification loss is replaced by the focal loss [39] while the regression loss is replaced by L1 and GIoU loss [52]. We show the universality of Cascade-Match on anchor-based detector (i.e., Cascade R-CNN) and an end-to-end detector (i.e., Sparse R-CNN) in the experiments, see Table 7."},{"figure_ref":[],"heading":"Adaptive Pseudo-label Mining","publication_ref":[],"table_ref":[],"text":"Determining a confidence threshold for pseudo labels is a non-trivial task, not to mention that each class requires a specific threshold to overcome the classimbalance issue-many-shot classes may need a higher threshold while few-shot classes may favor a lower threshold. Moreover, predictive confidence typically increases as the model observes more data (see Figure 3) (a), and therefore, dynamic thresholds are more desirable.\nTo solve the aforementioned problems, we propose an Adaptive Pseudo-label Mining (APM) module, which is an automatic selection mechanism for predicted pseudo-labels. Specifically, at each iteration, we first aggregate the ensemble predictions made on each ground-truth class using the labeled proposals (see The formulation above is simple but meaningful. In particular, since the predictive confidence values for each class are updated every iteration, the mean µ c will increase gradually, which naturally makes τ c k selfadaptive to the learning process without extra designs. By increasing ϵ k moderately in different stages, we maintain the progressive pattern of confidence threshold for different stages (e.g., τ 1 < τ 2 < • • • < τ K ) for any class. In this work, we choose ϵ k ∈ {1, 1.5, 2} for the three stages. The ablation study is provided in Table 3 of Section 4.1. In the experiments, we show that the progressive design is useful to control the precision and recall trade-off."},{"figure_ref":[],"heading":"Experiments","publication_ref":["b15","b69","b60","b61","b36","b69","b9","b9","b0","b19","b19","b4","b37","b17","b59","b72","b42","b57","b42","b63"],"table_ref":["tab_0"],"text":"Datasets We evaluate our approach on two longtailed object detection datasets: LVIS v1.0 [16] and COCO-LT [70]. LVIS v1.0 widely serves as a testbed for the long-tailed object detection task [61,62,37,70 Metrics We adopt the recently proposed Fixed AP (denoted by AP Fix ) metric [10], which does not restrict the number of predictions per image and can better characterize the long-tailed object detection performance. Following Dave et al. [10], we adopt the following notations for the metrics of different class groups: AP Fix r for rare classes, AP Fix c for common classes, and AP Fix f for frequent classes. For COCO-LT dataset, the symbols AP 1 , AP 2 , AP 3 and AP 4 correspond to the bins of [1,20), [20,400), [400, 8000) and [8000, -) (i.e., number of training instances).\nImplementation Details For the anchor-based detector, we employ the two-stage detector, Cascade R-CNN [5] with the FPN [38] neck. ResNet50 [18] pretrained from ImageNet is used as the CNN backbone. For the end-to-end detector, we adopt Sparse R-CNN [60] with the Pyramid Vision Transformer (PvT) [73] encoder. All settings for the parameters, such as learning rate, are kept the same as previous work [43]. We list the value of our used hyperparameters in Table 1. All models are trained with the standard SGD optimizer on 8 GPUs. Similar to previous methods [58,43,64], we also have a \"burn-in\" stage to stabilize training. Specifically, we pre-train the detector using labeled data first for several iterations, and then include unlabeled data in the training process. With burn-in Without burn-in Fig. 3: (a) Visualization of predictive confidence scores throughout training. We find that the predicted scores have the increasing tendency, which motivates us to propose the Adaptive Pseudo-label Mining (APM) module that using dynamic thresholds. (b) Impact of the burn-in stage. Clearly, the burn-in stage improves the performance."},{"figure_ref":[],"heading":"Ablation Studies","publication_ref":["b4","b15"],"table_ref":["tab_1","tab_1","tab_3","tab_4"],"text":"Before discussing the main results of long-tailed and semi-supervised object detection, we investigate the effects of the two key components of CascadeMatch, i.e., the cascade pseudo-labeling (CPL) and adaptive pseudo-label mining (APM), as well as some hyperparameters. The experiments are conducted on the LVIS v1.0 validation dataset.\nCascade Pseudo-Labeling The results are detailed in Table 2. We first examine the effect of the cascade pseudo-labeling module. The top row contains the results of the supervised baseline, while the second row corresponds to the combination of the baseline and CPL. We observe that CPL clearly improves upon the baseline. Notably, CPL improves the performance in all groups: +2.2 for the rare classes, +4.0 for the common classes, and +4.2 for the frequent classes.\nAdaptive Pseudo-label Mining We then examine the effectiveness of APM. By comparing the first and third rows in Table 2, we can conclude that APM alone is also beneficial to the performance, yielding clear gains of 2.8 AP Fix r and 2.6 AP Fix c . Finally, by combining CPL and APM (the last row), the performance can be further boosted, suggesting that the two modules are complementary to each other for long-tailed and semi-supervised object detection. We observe that CPL+APM brings a non-trivial improvement of 1.2% to the rare classes compared with using CPL only. The predictions on rare classes often have smaller confidence so the class-specific design in APM is essential for handling the long-tailed issue.\nHyper-parameter ϵ k As discussed in Section 3.2, our confidence thresholds τ k are adaptively adjusted and governed by a hyper-parameter ϵ k . In Table 3, we show the effects of using different values for ϵ k to update the per-class thresholds. Overall, the performance is insensitive to different values of ϵ k , with ϵ k = {1.0, 1.5, 2.0} achieving the best performance.\nHyper-parameter K The parameter K denotes the number of detection heads. We try different values of K, and the results are shown in Table 4. We observe that from k = 1 to 3, increasing the number of heads will improve the overall performance at the cost of training speed. The performance of rare and common classes will drop if we continue to increase the k from 3 to 4 or 5, probably due to the over-fitting and undesired memorizing effects of few-shot classes as we increase the model capacity. In this study, we choose to follow previous cascade methods [5] that use K = 3 heads.\nConfirmation Bias Recall that we use the ensemble teacher to train each detection head instead of using each individual prediction to mitigate confirmation bias. To understand how our design tackles the problem, we print the pseudo-label accuracy obtained during training for each detection head and their ensemble. Specifically, we use 30% of the LVIS training set as the labeled set and the remaining 70% as the unlabeled set. Note that the annotations for the unlabeled data are used only to calculate the pseudo-label accuracy. The results obtained at the 60k-th, 120k-th and 180kth iteration are shown in Table 5. It is clear that the pseudo-label accuracy numbers for individual heads are consistently lower than that of the ensemble throughout the course of training, confirming that using ensemble predictions is the optimal choice.\nHyper-parameter λ u To examine the effect of unsupervised loss weights λ u , we vary the unsupervised loss weight λ u from 0.5 to 2.0 on LVIS [16] dataset. As Table 3: Ablation study on the selection of the confidence parameter ϵ. We observe that the ϵ works the best with progressive values (ϵ 1 < ϵ 2 < ϵ 3 ). Table 4: Ablation study on the number of detector heads K. We also report the training time (seconds) per iteration in the last column. shown in Table 6, we observe that the model performs best with our default choice λ u = 1.0.\nϵ 1 ϵ 2 ϵ 3 AP Fix AP Fix\nK AP Fix AP Fix r AP Fix c AP Fix f T train1"},{"figure_ref":[],"heading":"Burn-in Stage","publication_ref":["b57","b42","b63"],"table_ref":[],"text":"As mentioned at the beginning of Section 4, we set a 'burn-in' stage to pre-train the detector on the labeled data before training on unlabeled data. Similar to previous works [58,43,64], such a 'burn-in' stage is used to stabilize initialization results in the early stage of training. In Figure 3 (b), we provide the mAP comparison of the CascadeMatch with and without the burn-in stage during the training. We observed that the model achieves higher mAP in the early stage with the burn-in stage and converges into better endpoints compared with the counterparts."},{"figure_ref":[],"heading":"Main Results","publication_ref":["b25","b57","b42","b79","b7","b61","b4","b59","b9"],"table_ref":[],"text":"Baselines In this section, we compare our method against the supervised baseline (without using the unlabeled data) and state-of-the-art semi-supervised learning methods on the LVIS v1.0 and COCO-LT datasets. We select four representative semi-supervised detection algorithms to compare with: 1) CSD [26] is a consistency regularization-based algorithm that forces the detector to make identical predictions under different augmentations. 2) STAC [58] is a pseudolabeling-based method that uses an off-line supervised model as a teacher to extract pseudo-labels. 3) Unbiased Teacher [43] and 4) Soft Teacher [80] are also a pseudo-labeling-based method that uses the exponential moving average (EMA) ensemble to provide a strong teacher model. Soft Teacher uses extra box jit-tering augmentation to further boost the performance. 5) LabelMatch [8] introduces a re-distribution mean teacher based on the KL divergence distribution between teacher and student models. Unbiased Teacher, Soft Teacher and LabelMatch are strong baselines so the comparison with them can well demonstrate the effectiveness of our approach. We use the open-source code provided by the authors and re-train the model on the LVIS v1.0 and COCO-LT datasets, respectively. All baselines and our approach use the Equalization Loss v2 (EQL v2) [62] as the default classification loss. EQL v2 improves the model's recognition ability by downweighting negative gradients for rare classes.\nResults on LVIS v1.0 Table 7 shows the results on LVIS. When using Cascade R-CNN and ResNet50 as the backbone, our approach improves AP Fix from the supervised baseline's 26.3 to 30.5, achieving 4.2 mAP improvement. Compared with LabelMatch, which is the strongest baseline, CascadeMatch still maintains clear advantages. Overall, the results presented in the experiments validate the effectiveness of the cascade pseudolabeling design and the adaptive pseudo-label mining mechanism. 7: Comparisons of mAP against the supervised baseline and different semi-supervised methods on LVIS v1.0 validation set We select two different frameworks: Cascade R-CNN [5] and Sparse R-CNN [60] with different backbones as the supervised baseline. The symbols AP Fix r , AP Fix c , and AP Fix f refer to the Fixed mAP [10] of overall, rare, common, and frequent class groups. The '12e' and '30e' schedules refer to 12 and 30 epochs, respectively. We report the average results over three runs with different random seeds. "},{"figure_ref":[],"heading":"Results on COCO-LT As shown in","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Computation Budgets","publication_ref":[],"table_ref":["tab_8"],"text":"We report the training memory, training time, and inference time against the supervised baseline and different semi-supervised methods, as shown in Table 9. All the methods are based on the Cascade-RCNN framework with the ResNet50-FPN backbone and report on one Nvidia V100 GPU. We can see that when compared with the supervised baseline, CSD has an increased memory footprint and training time because of the extra steps during training like data augmentation and forward pass on unlabeled data. For pseudo-labeling methods, like Unbiased Teacher and LabelMatch, the training cost further increases with the generation of pseudo-labels. Our Cas-cadeMatch method shares similar memory and training time as Unbiased Teacher, thus is comparable to recent semi-supervised methods in terms of the training cost. We also find all these methods (including ours) have negligible overhead in the inference stage, with almost the same inference time as the supervised learning baseline."},{"figure_ref":["fig_5","fig_5"],"heading":"Qualitative Results","publication_ref":["b15","b91","b86","b70","b4","b86","b70"],"table_ref":[],"text":"We show some pseudo-labeling visualization results under the semi-supervised object detection (SSOD) setting in Figure 4. Since we set a progressive confidence threshold τ from stage 1 to 3, we observe that stage 1 focuses on generating redundant pseudo labels with high recall and some false positive results (in purple). In contrast, stage 3 prefers high precision pseudo labels, but some prediction results may be missed. The ensemble of pseudo label predictions is of high quality and controls the precision-recall tradeoff well. According to the quantitative results in Table 7 and the qualitative results shown in Figure 4, we can conclude that CascadeMatch benefits from more accurate pseudo-labels it estimates for the unlabeled data. In SAOD, some images are only partially annotated, meaning that not all instances in an image are identified by bounding boxes. Such a phenomenon is in fact common in existing large-vocabulary datasets like the previously used LVIS [16] dataset. Unidentified instances are simply treated as background in existing semi-supervised approaches. As a consequence, no supervision will be given to the model with respect to those instances. Different from SSOD, the goal in SAOD is to identify instances with missing labels from the training set.\nExperimental Setup We use LVIS as the benchmark dataset. CascadeMatch is compared with Federated Loss [92], which serves as a strong baseline in this setting. Concretely, Federated Loss ignores losses of potentially missing categories and thus uses only a subset of classes for training. To facilitate evaluation, we follow previous studies [87,71] to build a modified LVIS dataset where a certain percentage of annotations within each category are randomly erased. We choose the 20% and 40% as the percentage numbers. The baseline model is the combination of Cascade R-CNN [5] and Federated Loss. Noted that it is common to select 50% erasing ratio [87,71] for balanced datasets. However, for long-tailed datasets erasing 50% annotations would lead to significantly fewer annotations for rare classes (23.73% of rare classes will have zero an-notations). We chose the 20% and 40% ratios to cover different scenarios (95.54% and 88.76% of rare classes are preserved that have at least one annotation)."},{"figure_ref":[],"heading":"Results","publication_ref":[],"table_ref":["tab_9","tab_9"],"text":"We experimented with the 20% and 40% missing ratios on our modified LVIS dataset. The results are reported in Table 10 where the checkmark symbol means that CascadeMatch is applied to the model. In both settings, we observe a clear margin between CascadeMatch and the baseline: +1.8% and +2.0% gains in terms of overall AP under the settings of 20% and 40% missing ratios, respectively. Notably, the gains are more apparent for the rare classes, with +3.3% and +2.9% gains for the two settings, respectively. The quantitative results shown in Table 10 strongly demonstrate the ability of CascadeMatch in dealing with the SAOD problem."},{"figure_ref":[],"heading":"Qualitative Results","publication_ref":[],"table_ref":[],"text":"We also show the visualization results of the pseudo-labeling under the sparselyannotated object detection (SAOD) setting in Figure 5. The first column refers to the ground truth labels from the original LVIS dataset. The second column shows our modified sparsely-annotated LVIS dataset where some annotations are randomly removed with a 40% missing rate and serves as the training set under the SAOD setting. The third column contains the prediction results of CascadeMatch. We observe that CascadeMatch can recover some labels. Since the original LVIS datasets is sparsely-annotated, CascadeMatch can also detect objects whose labels are missing in the original LVIS dataset. The qualitative results in Figure 5 explain the excellent performance of CascadeMatch on the SAOD task."},{"figure_ref":[],"heading":"Limitation","publication_ref":["b40","b50","b38","b4","b66","b5","b22"],"table_ref":[],"text":"The trade-off between speed and performance is one of the key research problems in the area of object detection [41,51,39,5,67,6]. It has been widely acknowledged that achieving a perfect speed-performance trade-off is extremely difficult [23]. To obtain a highperformance detector, one has to sacrifice on the speed, and vice versa. In this work, our CascadeMatch processes data in a cascade manner, which leads to longer training time and slower inference speed compared to the single-stage detector counterpart. However, given that the majority of computation takes place in the backbone while the detection heads are generally \"lightweight\" (as they only consist of a few fully connected layers), the lower speed is outweighed by the improvements in performance. To further improve the efficiency in real-world deployment, one could apply model compression techniques to reduce the model size, and LVIS (Original) LVIS (Missing 40%) Prediction (Ours)\nFig. 5: The pseudo labels generated on the LVIS training dataset under the sparsely-annotated object detection setting (SAOD) setting. In the third column, green color refers to the predicted results that can be found in the ground truth of the first column; purple color refers to predicted results that are also missing in the original LVIS dataset (Zoom in for best view).\ndesign more lightweight architectures for the cascade detection heads."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"Our research addresses an important but largely under-studied problem in object detection, concerning both long-tailed data distributions and semi-supervised learning. The proposed approach, CascadeMatch, carefully integrates pseudo-labeling, coupled with a cascade design and an adaptive threshold tuning mechanism, into a variety of backbones and detection frameworks, such as the widely used region proposal-based detectors and more recent fully end-to-end detectors.\nThe results strongly demonstrate that CascadeMatch is a better design than existing state-of-the-art semisupervised detectors in handling long-tailed datasets such as LVIS and COCO-LT. The capability to cope with the sparsely-annotated object detection problem is also well justified."},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"Acknowledgement This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). It is also partly supported by the NTU NAP grant and Singapore MOE AcRF Tier 2 (MOE-T2EP20120-0001)."},{"figure_ref":[],"heading":"Data Availability Statements","publication_ref":[],"table_ref":[],"text":"The datasets analysed during this study are all publicly available for the research purpose -the LVIS and COCO datasets."}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_0\",\n \"fig_0\",\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b40\",\n \"b50\",\n \"b38\",\n \"b66\",\n \"b5\",\n \"b57\",\n \"b42\",\n \"b90\",\n \"b62\",\n \"b63\",\n \"b0\",\n \"b73\",\n \"b81\",\n \"b2\",\n \"b78\",\n \"b58\",\n \"b52\",\n \"b39\",\n \"b15\",\n \"b69\",\n \"b42\",\n \"b15\",\n \"b42\",\n \"b15\",\n \"b69\",\n \"b42\",\n \"b50\",\n \"b4\",\n \"b59\",\n \"b72\",\n \"b76\",\n \"b86\",\n \"b70\",\n \"b91\",\n \"b91\"\n ],\n \"table_ref\": [\n \"tab_9\"\n ],\n \"text\": \"Though object detection has been significantly advanced in the supervised learning domain by neural network-based detectors [41,51,39,67,6]), there is still a large room for improvement in semi-supervised object detection (SSOD). In practice, SSOD is desirable because annotating bounding boxes and their object classes are both costly and time-consuming. Most existing semi-supervised object detectors [58,43,91,63,64,1,74,82]) are learned by estimated pseudo-labels, which are assigned to bounding box proposals and filtered by a single fixed confidence threshold. Such a combination of pseudo-labeling and confidence thresholdsbased filtering has been largely inspired by research on semi-supervised image classification [3,79,59,53]).\\nMost existing studies are conducted on the COCO dataset [40] that has curated categories and highly balanced data distributions. However, real-world problems are much more challenging than what the COCO dataset represents in that data distributions are often long-tailed, i.e., a majority of classes have only a few labeled images, which could easily result in an extremely biased detector. In recent years, the research community has paid increasing attention to long-tailed object detection, with several relevant datasets released, such as LVIS [16] and COCO-LT [70]. However, to our knowledge, none of the existing studies has been devoted to long-tailed object detection in the semi-supervised setting, a more challenging yet practical problem.\\nImplementing semi-supervised object detection algorithms on long-tailed datasets is not trivial. By train- ing a state-of-the-art semi-supervised detector, i.e., Unbiased Teacher [43], using a long-tailed LVIS [16] dataset, we identify the following three major problems. First, a fixed confidence threshold often fails to provide a good trade-off between precision and recall. The shortcoming is evidenced in Figure 1(a), which shows none of the commonly used thresholds gives the best performance in both the AP and AR metrics, e.g., a fixed threshold of 0.6 returns the highest recall but has the lowest precision. Second, by digging deeper into the distribution of prediction scores, we observe that the model's predictions are biased toward the frequent classes (see Figure 1(b)). Finally, we identify the reason why using a fixed threshold leads to low confidenceand hence low prediction accuracy-on the common and rare classes: the model's exposure to these classes during training is substantially reduced compared to that to the frequent classes (see Figure 1(c)).\\nTo overcome these problems, we propose Cascade-Match, a novel pseudo-labeling-based approach to addressing long-tailed and semi-supervised object detection. Specifically, CascadeMatch features a cascade pseudo-labeling (CPL) design, which contains multistage detection heads. To control the precision-recall trade-off, we set progressive confidence thresholds for detection heads to focus on different parts. The early detection head is assigned a small confidence threshold to improve recall, while the subsequent heads are assigned larger confidence thresholds to ensure precision. The use of multiple heads also allows the unique chance for us to deal with confirmation bias -a phenomenon where a model is iteratively reinforced by incorrect pseudo labels produced by itself. In particular, we show the possibility of using ensemble predictions from all detection heads as the teacher's supervision signal to obtain more reliable pseudo labels for training each individual detection head. To deal with the issue of biased prediction score distributions to frequent classes, we propose an adaptive pseudo-label mining mechanism (APM) that automatically identifies suitable class-wise threshold values from data with minimal human intervention. As shown in Figure 1(c), with the APM module, our approach can retain more pseudolabels for common and rare classes than the previous SOTA approach [43], boosting the performance for classes with small sample sizes.\\nWe present comprehensive experiments on two challenging long-tailed object detection datasets, namely LVIS v1.0 [16] and COCO-LT [70], under the SSOD setting. Overall, CascadeMatch achieves the best performance on both datasets in all metrics. Notably, on LVIS, CascadeMatch improves upon the most competitive method, i.e., Unbiased Teacher [43], by 2.3% and 1.8% AP Fix in the rare and common classes, which confirm the effectiveness of our design for long-tailed data. Importantly, CascadeMatch is general and obtains consistent improvements across a variety of detection architectures, covering both anchor-based R-CNN detectors [51,5] and the recent Sparse R-CNN detector [60] with the Pyramid Vision Transformer encoder (PVT) [73] (Table 7). We also conduct various ablation studies to confirm the effectiveness of each of our proposed modules.\\nWe also apply CascadeMatch to another challenging sparsely-annotated object detection (SAOD) setting [77,87,71,92] where training data are only partially annotated and contain missing annotated instances. Again, CascadeMatch yields considerable improvements over the supervised-only baseline and a state-of-the-art method [92] (Table 10). Finally, we provide several qualitative results and analyses to show that our proposed CascadeMatch method generates high-quality pseudo labels on both SSOD and SAOD settings.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b53\",\n \"b45\",\n \"b64\",\n \"b68\",\n \"b54\",\n \"b25\",\n \"b12\",\n \"b35\",\n \"b57\",\n \"b62\",\n \"b26\",\n \"b90\",\n \"b81\",\n \"b79\",\n \"b85\",\n \"b41\",\n \"b8\",\n \"b7\",\n \"b31\",\n \"b44\",\n \"b14\",\n \"b33\",\n \"b41\",\n \"b25\",\n \"b26\",\n \"b57\",\n \"b42\",\n \"b63\",\n \"b81\",\n \"b79\",\n \"b85\",\n \"b8\",\n \"b7\",\n \"b63\",\n \"b90\",\n \"b63\",\n \"b42\",\n \"b38\",\n \"b73\",\n \"b90\",\n \"b33\",\n \"b7\",\n \"b8\",\n \"b41\",\n \"b39\",\n \"b15\",\n \"b25\",\n \"b57\",\n \"b42\",\n \"b79\",\n \"b50\",\n \"b38\",\n \"b4\",\n \"b66\",\n \"b5\",\n \"b59\",\n \"b88\",\n \"b15\",\n \"b56\",\n \"b21\",\n \"b75\",\n \"b60\",\n \"b49\",\n \"b71\",\n \"b61\",\n \"b87\",\n \"b67\",\n \"b11\",\n \"b6\",\n \"b91\",\n \"b32\",\n \"b19\",\n \"b83\",\n \"b34\",\n \"b13\",\n \"b15\",\n \"b36\",\n \"b69\",\n \"b60\",\n \"b61\",\n \"b67\",\n \"b84\",\n \"b55\",\n \"b2\",\n \"b3\",\n \"b58\",\n \"b89\",\n \"b80\",\n \"b2\",\n \"b78\",\n \"b58\",\n \"b78\",\n \"b1\",\n \"b29\",\n \"b24\",\n \"b77\",\n \"b46\",\n \"b58\",\n \"b48\",\n \"b28\",\n \"b65\",\n \"b42\",\n \"b0\",\n \"b58\",\n \"b0\",\n \"b52\",\n \"b16\",\n \"b47\",\n \"b23\",\n \"b27\",\n \"b82\",\n \"b74\",\n \"b30\",\n \"b10\",\n \"b46\",\n \"b18\",\n \"b20\"\n ],\n \"table_ref\": [],\n \"text\": \"Semi-Supervised Object Detection has been a topical research area due to its importance to practical applications [54,46,65,69,55,26,13,36,58,63,27,91,82,80,86,42,9,8,32,45,15,34,42]. Various semi-supervised object detectors have been proposed in the literature, and many of them borrow ideas from the semi-supervised learning (SSL) community.\\nIn CSD [26] and ISD [27], consistency regularization is applied to the mined bounding boxes for unlabeled images. STAC [58] uses strong data augmentation for self-training.\\nRecently, pseudo-labeling-based methods have shown promising results on several benchmark datasets, which are attributed to a stronger teacher model trained by, e.g., a weighted EMA ensemble [43,64,82,80,86,9,8], a data ensemble [64], or advanced data augmentation [91,64]. To overcome the confirmation bias, Unbiased Teacher [43] employs focal loss [39] to reduce the weights on overconfident pseudo labels, while others use uncertainty modeling [74] or co-training [91] as the countermeasure. Li, et al. [34] propose dynamic thresholding for each class based on both localization and classification confidence. LabelMatch [8] introduces a re-distribution mean teacher based on the KL divergence distribution between teacher and student models. DSL [9] assigns pixel-wise pseudo-labels for anchor-free detectors. Unbiased Teacherv2 [42] introduces a new pseudolabeling mechanism based on the relative uncertainties of teacher and student models.\\nIt is worth noting that most existing methods are designed for class-balanced datasets like MS COCO [40], while their capabilities to handle longtailed datasets like LVIS [16] have been largely understudied-to our knowledge, none of existing research has specifically investigated long-tailed object detection in the SSL setting. Instead, the majority of existing SSL algorithms are evaluated on class-balanced datasets [26,58,43,80]. Our work takes the first step toward a unified approach to solving unlabeled data and the longtailed object detection problem, which we hope to inspire more work to tackle this challenging setting.\\nLong-tailed Object Detection Though object detection has witnessed significant progress in recent years [51,39,5,67,6,60], how to deal with the long-tailed problem remains an open question [89]. Most existing methods fall into two groups: data resampling [16,57,22,76] and loss re-weighting [61,50,72,62,88,68,12,7,92,33,20]. Some recent works [84,35,14] suggest that data augmentation is useful for long-tailed recognition. In terms of data re-sampling, Repeated Factor Sampling (RFS) [16] assigns high sampling rates to images of rare classes. A couple of studies [37,70] have suggested using different sampling schemes in decoupled training stages. When it comes to data re-weighting, a representative method is equalization loss [61,62], which raises the weights for rare classes based on inverse class frequency. Seesaw Loss [68] automatically adjusts class-specific loss weights based on a statistical ratio between the positive and negative gradients computed for each class. MosaicOS [85] is one of the early studies that uses weakly-supervised learning to help long-tailed detection. Their study assumes the availability of weaklyannotated class labels. In contrast, we take a pure semisupervised setting without assuming any annotations in the unlabeled set. In our work, we first investigate how to exploit unlabeled data to improve the performance of detectors trained on long-tailed datasets.\\nSemi-Supervised Learning (SSL) Numerous SSL methods are based on consistency learning [56,3,4,59,90,81], which forces a model's predictions on two different views of the same instance to be similar. Recent state-of-the-art consistency learning methods like MixMatch [3], UDA [79] and FixMatch [59] introduce strong data augmentations [79] to the learning paradigm-they use predictions on weakly augmented images as the target to train the model to produce similar outputs given the strongly augmented views of the same images.\\nAnother research direction related to our work is pseudo-labeling [2,30,25,78,47], which is typically based on a teacher-student architecture: a teacher model's predictions are used as the target to train a student model. The teacher model can be either a pretrained model [59] or an exponential moving average of the student model [49,29,66,43]. Some studies [1] have also demonstrated that using the student model being trained to produce the target can reach decent performance-the trick is to inject strong noise to the student model, such as applying strong data augmentations to the input [59].\\nA common issue encountered in pseudo-labeling methods is confirmation bias [1], which is caused by a constant feed of incorrect pseudo labels with high confidence to the model. And such a vicious cycle would reinforce since the model will become increasingly inaccurate and subsequently provide more erroneous pseudo labels. To mitigate the issue of confirmation bias, existing methods have tried using an uncertainty-based metric [53] to modulate the confidence threshold or using the co-training framework [17,48] that simultaneously trains two neural networks each giving pseudo labels to the other. In this work, to prevent each detection head from overfitting its own prediction errors, the pseudo labels to train each detection head are formed by the ensemble predictions of multiple detection heads. This strategy is new in the literature.\\nIt is worth noting that most aforementioned algorithms are evaluated on class-balanced datasets while only very few recent works apply SSL for long-tailed image classification [24,28,83,75,31,11,47] or semantic segmentation [19,21]. The detection task requires predicting both the class labels and object locations, which is much harder than the classification-only task. The pseudo-labeling-based semi-supervised methods are unable to predict high-quality pseudo labels for detection task as accurately as for classification task, in the presence of class imbalance. This motivates us to improve the pseudo-labeling quality for semi-supervised and long-tailed detection using a cascade mechanism.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Our Approach: CascadeMatch\",\n \"publication_ref\": [\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"Problem Definition Given a labeled dataset D l = {(x, y * , b * )} with x, y * and b * denoting image, label and bounding box, respectively,1 and an unlabeled dataset D u = {x}, the goal is to learn a robust object detector using both D l and D u . We further consider the issue of long-tailed distribution [16], which is common in real-world data but have been largely unexplored in existing semi-supervised object detection methods. More specifically, let n i and n j denote the number of images for class i and j respectively, and assume i is a frequent class while j is a rare class. In a long-tailed scenario, we might have n i ≫ n j .\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"An Overview\",\n \"publication_ref\": [\n \"b4\",\n \"b59\",\n \"b17\",\n \"b43\"\n ],\n \"table_ref\": [],\n \"text\": \"A brief overview of the main paradigm of our proposed CascadeMatch is illustrated in Figure 2. CascadeMatch features a cascade pseudolabeling (CPL) design and an adaptive pseudo-label mining (APM) mechanism. The former aims to generate pseudo-labels and filter out low-quality labels in a cascade fashion to improve the trade-off between precision and recall, while the latter aims to automate threshold tuning. CascadeMatch only modifies a detector's head structure and thus can be seen as a plug-andplay module that fits into most existing object detectors including the popular anchor-based R-CNN series like Cascade R-CNN [5] or more recent end-to-end detectors like Sparse R-CNN [60]. CascadeMatch can also take either CNNs [18] or Transformers [44] as the backbone. Discussion A cascade structure benefits from the \\\"divide and conquer\\\" concept, where each stage is ded-icated to a specific sub-task. This notion of cascading has been found practical and useful in many computer vision systems. For the detection task, finding an accurate IoU threshold to separate the positive and negative region proposals is impossible. To allow a better precision-recall trade-off, Cascade R-CNN uses the cascade structure to progressively increase the IoU threshold for different stages. Recall that pseudo labeling faces a similar dilemma in pinpointing a single confidence threshold to separate the valid pseudo-labels and noisy background region proposals. It is thus natural for Cas-cadeMatch to use the cascade structure with a set of progressive confidence thresholds. Note that the confidence threshold of CascadeMatch is class-specific and self-adaptive. We will provide the details in Section 3.2.\\nBelow we provide the technical details of the two key components in CascadeMatch, namely cascade pseudolabeling (Section 3.1) and adaptive pseudo-label mining (Section 3.2). For clarity, in Section 3.1 we first present CascadeMatch in an anchor-based framework and later explain the modifications needed for an end-to-end detector.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Cascade Pseudo-Labeling\",\n \"publication_ref\": [\n \"b50\",\n \"b4\",\n \"b17\",\n \"b50\",\n \"b4\",\n \"b4\",\n \"b59\"\n ],\n \"table_ref\": [],\n \"text\": \"Model Architecture For an anchor-based framework [51,5], the CascadeMatch-based detector starts with a CNN as the backbone for feature extraction, e.g., ResNet50 [18], which is then followed by a region proposal network (RPN) [51] for generating object proposals. See Figure 2(a) for the architecture.\\nThe detector has K heads following the Cascade R-CNN [5] pipeline. The parameter K controls the tradeoff between performance and efficiency, which can be adjusted by practitioners based on their needs. Increasing the number of heads will improve the performance at the cost of speed. In the paper, we followed previous cascade methods [5,60] to use K = 3 heads. We will provide the ablation studies of varying the value of K in Table 4 of Section 4.1. Formally, given an image x, the first-stage detection head predicts for an object proposal b 0 (generated by the RPN) a class probability distribution p 1 (y|x, b 0 ) and the bounding box offsets b 1 . Then, the second-stage detection head predicts another probability p 2 (y|x, b 1 ) using the refined bounding box from the first stage;2 and so on and so forth.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Labeled Losses\",\n \"publication_ref\": [\n \"b50\",\n \"b57\",\n \"b42\",\n \"b90\",\n \"b62\",\n \"b63\",\n \"b73\",\n \"b0\"\n ],\n \"table_ref\": [\n \"tab_3\"\n ],\n \"text\": \"With labeled data D l = {(x, y * , b * )}, we train each detection head using the classification loss Cls(•, •) (for proposal classification) and the bounding box regression loss Reg(•, •) [51]. Formally, we have\\nℓ labeled cls = (x,y * )∼D l K k=1 Cls(y * , p k (y|x, b k-1 )),(1)\\nℓ labeled reg = (x,b * )∼D l K k=1 Reg(b * , b k ). (2\\n)\\nUnlabeled Losses To cope with unlabeled images, we adopt a pseudo-labeling approach with a teacherstudent architecture where the teacher's estimations on unlabeled data are given to the student as supervision. Such a paradigm has been widely used in previous semisupervised methods [58,43,91,63,64,74]. Different from previous methods, we focus on tackling the confirmation bias issue [1] when designing our architecture. We observe that the ensemble predictions are more accurate than using each individual prediction (please refer to Table 5 of Section 4.1 for more details), so we use the ensemble predictions from all detection heads as the teacher supervision signal (teacher module in Figure . 2 (a)). Formally, given an unlabeled image x ∼ D u , the ensemble prediction p t is computed as\\np t = 1 K K k=1 p k (y|x, b k-1 ) and b t = 1 K K k=1 b k ,(3)\\nwhere K is the number of heads. Let q t = max(p t ) be the confidence and qt = arg max(p t ) the pseudo label, we compute the classification loss and the bounding box regression loss for unlabeled data using\\nℓ unlabeled cls = x∼Du K k=1 1(q t ≥ τ qt k ) Cls(q t , p k (y|x, b k-1 )),(4)\\nℓ unlabeled reg = x∼Du K k=1 1(q t ≥ τ qt k ) Reg(b t , b k ),(5)\\nwhere τ qt k is a self-adaptive confidence threshold specific to class qt . We detail the design of class-specific selfadaptive thresholds in Section 3.2.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Training\",\n \"publication_ref\": [\n \"b50\",\n \"b59\",\n \"b5\",\n \"b38\",\n \"b51\"\n ],\n \"table_ref\": [],\n \"text\": \"Similar to most region-based object detectors, our CascadeMatch model is learned using four losses: a region-of-interest (ROI) classification loss\\nℓ roi cls = ℓ labeled cls + λ u • ℓ unlabeled cls , an ROI regression loss ℓ roi reg = ℓ labeled reg + λ u • ℓ unlabeled reg\\n, and two other losses for the RPN, i.e., the objectness classification loss ℓ rpn cls and the proposal regression loss ℓ rpn reg , as defined in [51]. The loss parameter λ u controls the weight between the supervised term ℓ l cls and the unsupervised term ℓ u cls . By default, we set the unsupervised loss weight λ u = 1.0.\\nTransfer to End-to-End Object Detector Cas-cadeMatch is readily applicable to an end-to-end detector. We use Sparse R-CNN [60] as an example. Two main modifications are required: 1) Since region proposals are learned from a set of embedding queries as in DETR [6], we do not need an RPN and the RPN loss ℓ rpn ; 2) The classification loss is replaced by the focal loss [39] while the regression loss is replaced by L1 and GIoU loss [52]. We show the universality of Cascade-Match on anchor-based detector (i.e., Cascade R-CNN) and an end-to-end detector (i.e., Sparse R-CNN) in the experiments, see Table 7.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Adaptive Pseudo-label Mining\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Determining a confidence threshold for pseudo labels is a non-trivial task, not to mention that each class requires a specific threshold to overcome the classimbalance issue-many-shot classes may need a higher threshold while few-shot classes may favor a lower threshold. Moreover, predictive confidence typically increases as the model observes more data (see Figure 3) (a), and therefore, dynamic thresholds are more desirable.\\nTo solve the aforementioned problems, we propose an Adaptive Pseudo-label Mining (APM) module, which is an automatic selection mechanism for predicted pseudo-labels. Specifically, at each iteration, we first aggregate the ensemble predictions made on each ground-truth class using the labeled proposals (see The formulation above is simple but meaningful. In particular, since the predictive confidence values for each class are updated every iteration, the mean µ c will increase gradually, which naturally makes τ c k selfadaptive to the learning process without extra designs. By increasing ϵ k moderately in different stages, we maintain the progressive pattern of confidence threshold for different stages (e.g., τ 1 < τ 2 < • • • < τ K ) for any class. In this work, we choose ϵ k ∈ {1, 1.5, 2} for the three stages. The ablation study is provided in Table 3 of Section 4.1. In the experiments, we show that the progressive design is useful to control the precision and recall trade-off.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [\n \"b15\",\n \"b69\",\n \"b60\",\n \"b61\",\n \"b36\",\n \"b69\",\n \"b9\",\n \"b9\",\n \"b0\",\n \"b19\",\n \"b19\",\n \"b4\",\n \"b37\",\n \"b17\",\n \"b59\",\n \"b72\",\n \"b42\",\n \"b57\",\n \"b42\",\n \"b63\"\n ],\n \"table_ref\": [\n \"tab_0\"\n ],\n \"text\": \"Datasets We evaluate our approach on two longtailed object detection datasets: LVIS v1.0 [16] and COCO-LT [70]. LVIS v1.0 widely serves as a testbed for the long-tailed object detection task [61,62,37,70 Metrics We adopt the recently proposed Fixed AP (denoted by AP Fix ) metric [10], which does not restrict the number of predictions per image and can better characterize the long-tailed object detection performance. Following Dave et al. [10], we adopt the following notations for the metrics of different class groups: AP Fix r for rare classes, AP Fix c for common classes, and AP Fix f for frequent classes. For COCO-LT dataset, the symbols AP 1 , AP 2 , AP 3 and AP 4 correspond to the bins of [1,20), [20,400), [400, 8000) and [8000, -) (i.e., number of training instances).\\nImplementation Details For the anchor-based detector, we employ the two-stage detector, Cascade R-CNN [5] with the FPN [38] neck. ResNet50 [18] pretrained from ImageNet is used as the CNN backbone. For the end-to-end detector, we adopt Sparse R-CNN [60] with the Pyramid Vision Transformer (PvT) [73] encoder. All settings for the parameters, such as learning rate, are kept the same as previous work [43]. We list the value of our used hyperparameters in Table 1. All models are trained with the standard SGD optimizer on 8 GPUs. Similar to previous methods [58,43,64], we also have a \\\"burn-in\\\" stage to stabilize training. Specifically, we pre-train the detector using labeled data first for several iterations, and then include unlabeled data in the training process. With burn-in Without burn-in Fig. 3: (a) Visualization of predictive confidence scores throughout training. We find that the predicted scores have the increasing tendency, which motivates us to propose the Adaptive Pseudo-label Mining (APM) module that using dynamic thresholds. (b) Impact of the burn-in stage. Clearly, the burn-in stage improves the performance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ablation Studies\",\n \"publication_ref\": [\n \"b4\",\n \"b15\"\n ],\n \"table_ref\": [\n \"tab_1\",\n \"tab_1\",\n \"tab_3\",\n \"tab_4\"\n ],\n \"text\": \"Before discussing the main results of long-tailed and semi-supervised object detection, we investigate the effects of the two key components of CascadeMatch, i.e., the cascade pseudo-labeling (CPL) and adaptive pseudo-label mining (APM), as well as some hyperparameters. The experiments are conducted on the LVIS v1.0 validation dataset.\\nCascade Pseudo-Labeling The results are detailed in Table 2. We first examine the effect of the cascade pseudo-labeling module. The top row contains the results of the supervised baseline, while the second row corresponds to the combination of the baseline and CPL. We observe that CPL clearly improves upon the baseline. Notably, CPL improves the performance in all groups: +2.2 for the rare classes, +4.0 for the common classes, and +4.2 for the frequent classes.\\nAdaptive Pseudo-label Mining We then examine the effectiveness of APM. By comparing the first and third rows in Table 2, we can conclude that APM alone is also beneficial to the performance, yielding clear gains of 2.8 AP Fix r and 2.6 AP Fix c . Finally, by combining CPL and APM (the last row), the performance can be further boosted, suggesting that the two modules are complementary to each other for long-tailed and semi-supervised object detection. We observe that CPL+APM brings a non-trivial improvement of 1.2% to the rare classes compared with using CPL only. The predictions on rare classes often have smaller confidence so the class-specific design in APM is essential for handling the long-tailed issue.\\nHyper-parameter ϵ k As discussed in Section 3.2, our confidence thresholds τ k are adaptively adjusted and governed by a hyper-parameter ϵ k . In Table 3, we show the effects of using different values for ϵ k to update the per-class thresholds. Overall, the performance is insensitive to different values of ϵ k , with ϵ k = {1.0, 1.5, 2.0} achieving the best performance.\\nHyper-parameter K The parameter K denotes the number of detection heads. We try different values of K, and the results are shown in Table 4. We observe that from k = 1 to 3, increasing the number of heads will improve the overall performance at the cost of training speed. The performance of rare and common classes will drop if we continue to increase the k from 3 to 4 or 5, probably due to the over-fitting and undesired memorizing effects of few-shot classes as we increase the model capacity. In this study, we choose to follow previous cascade methods [5] that use K = 3 heads.\\nConfirmation Bias Recall that we use the ensemble teacher to train each detection head instead of using each individual prediction to mitigate confirmation bias. To understand how our design tackles the problem, we print the pseudo-label accuracy obtained during training for each detection head and their ensemble. Specifically, we use 30% of the LVIS training set as the labeled set and the remaining 70% as the unlabeled set. Note that the annotations for the unlabeled data are used only to calculate the pseudo-label accuracy. The results obtained at the 60k-th, 120k-th and 180kth iteration are shown in Table 5. It is clear that the pseudo-label accuracy numbers for individual heads are consistently lower than that of the ensemble throughout the course of training, confirming that using ensemble predictions is the optimal choice.\\nHyper-parameter λ u To examine the effect of unsupervised loss weights λ u , we vary the unsupervised loss weight λ u from 0.5 to 2.0 on LVIS [16] dataset. As Table 3: Ablation study on the selection of the confidence parameter ϵ. We observe that the ϵ works the best with progressive values (ϵ 1 < ϵ 2 < ϵ 3 ). Table 4: Ablation study on the number of detector heads K. We also report the training time (seconds) per iteration in the last column. shown in Table 6, we observe that the model performs best with our default choice λ u = 1.0.\\nϵ 1 ϵ 2 ϵ 3 AP Fix AP Fix\\nK AP Fix AP Fix r AP Fix c AP Fix f T train1\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Burn-in Stage\",\n \"publication_ref\": [\n \"b57\",\n \"b42\",\n \"b63\"\n ],\n \"table_ref\": [],\n \"text\": \"As mentioned at the beginning of Section 4, we set a 'burn-in' stage to pre-train the detector on the labeled data before training on unlabeled data. Similar to previous works [58,43,64], such a 'burn-in' stage is used to stabilize initialization results in the early stage of training. In Figure 3 (b), we provide the mAP comparison of the CascadeMatch with and without the burn-in stage during the training. We observed that the model achieves higher mAP in the early stage with the burn-in stage and converges into better endpoints compared with the counterparts.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Main Results\",\n \"publication_ref\": [\n \"b25\",\n \"b57\",\n \"b42\",\n \"b79\",\n \"b7\",\n \"b61\",\n \"b4\",\n \"b59\",\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"Baselines In this section, we compare our method against the supervised baseline (without using the unlabeled data) and state-of-the-art semi-supervised learning methods on the LVIS v1.0 and COCO-LT datasets. We select four representative semi-supervised detection algorithms to compare with: 1) CSD [26] is a consistency regularization-based algorithm that forces the detector to make identical predictions under different augmentations. 2) STAC [58] is a pseudolabeling-based method that uses an off-line supervised model as a teacher to extract pseudo-labels. 3) Unbiased Teacher [43] and 4) Soft Teacher [80] are also a pseudo-labeling-based method that uses the exponential moving average (EMA) ensemble to provide a strong teacher model. Soft Teacher uses extra box jit-tering augmentation to further boost the performance. 5) LabelMatch [8] introduces a re-distribution mean teacher based on the KL divergence distribution between teacher and student models. Unbiased Teacher, Soft Teacher and LabelMatch are strong baselines so the comparison with them can well demonstrate the effectiveness of our approach. We use the open-source code provided by the authors and re-train the model on the LVIS v1.0 and COCO-LT datasets, respectively. All baselines and our approach use the Equalization Loss v2 (EQL v2) [62] as the default classification loss. EQL v2 improves the model's recognition ability by downweighting negative gradients for rare classes.\\nResults on LVIS v1.0 Table 7 shows the results on LVIS. When using Cascade R-CNN and ResNet50 as the backbone, our approach improves AP Fix from the supervised baseline's 26.3 to 30.5, achieving 4.2 mAP improvement. Compared with LabelMatch, which is the strongest baseline, CascadeMatch still maintains clear advantages. Overall, the results presented in the experiments validate the effectiveness of the cascade pseudolabeling design and the adaptive pseudo-label mining mechanism. 7: Comparisons of mAP against the supervised baseline and different semi-supervised methods on LVIS v1.0 validation set We select two different frameworks: Cascade R-CNN [5] and Sparse R-CNN [60] with different backbones as the supervised baseline. The symbols AP Fix r , AP Fix c , and AP Fix f refer to the Fixed mAP [10] of overall, rare, common, and frequent class groups. The '12e' and '30e' schedules refer to 12 and 30 epochs, respectively. We report the average results over three runs with different random seeds. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results on COCO-LT As shown in\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Computation Budgets\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_8\"\n ],\n \"text\": \"We report the training memory, training time, and inference time against the supervised baseline and different semi-supervised methods, as shown in Table 9. All the methods are based on the Cascade-RCNN framework with the ResNet50-FPN backbone and report on one Nvidia V100 GPU. We can see that when compared with the supervised baseline, CSD has an increased memory footprint and training time because of the extra steps during training like data augmentation and forward pass on unlabeled data. For pseudo-labeling methods, like Unbiased Teacher and LabelMatch, the training cost further increases with the generation of pseudo-labels. Our Cas-cadeMatch method shares similar memory and training time as Unbiased Teacher, thus is comparable to recent semi-supervised methods in terms of the training cost. We also find all these methods (including ours) have negligible overhead in the inference stage, with almost the same inference time as the supervised learning baseline.\"\n },\n {\n \"figure_ref\": [\n \"fig_5\",\n \"fig_5\"\n ],\n \"heading\": \"Qualitative Results\",\n \"publication_ref\": [\n \"b15\",\n \"b91\",\n \"b86\",\n \"b70\",\n \"b4\",\n \"b86\",\n \"b70\"\n ],\n \"table_ref\": [],\n \"text\": \"We show some pseudo-labeling visualization results under the semi-supervised object detection (SSOD) setting in Figure 4. Since we set a progressive confidence threshold τ from stage 1 to 3, we observe that stage 1 focuses on generating redundant pseudo labels with high recall and some false positive results (in purple). In contrast, stage 3 prefers high precision pseudo labels, but some prediction results may be missed. The ensemble of pseudo label predictions is of high quality and controls the precision-recall tradeoff well. According to the quantitative results in Table 7 and the qualitative results shown in Figure 4, we can conclude that CascadeMatch benefits from more accurate pseudo-labels it estimates for the unlabeled data. In SAOD, some images are only partially annotated, meaning that not all instances in an image are identified by bounding boxes. Such a phenomenon is in fact common in existing large-vocabulary datasets like the previously used LVIS [16] dataset. Unidentified instances are simply treated as background in existing semi-supervised approaches. As a consequence, no supervision will be given to the model with respect to those instances. Different from SSOD, the goal in SAOD is to identify instances with missing labels from the training set.\\nExperimental Setup We use LVIS as the benchmark dataset. CascadeMatch is compared with Federated Loss [92], which serves as a strong baseline in this setting. Concretely, Federated Loss ignores losses of potentially missing categories and thus uses only a subset of classes for training. To facilitate evaluation, we follow previous studies [87,71] to build a modified LVIS dataset where a certain percentage of annotations within each category are randomly erased. We choose the 20% and 40% as the percentage numbers. The baseline model is the combination of Cascade R-CNN [5] and Federated Loss. Noted that it is common to select 50% erasing ratio [87,71] for balanced datasets. However, for long-tailed datasets erasing 50% annotations would lead to significantly fewer annotations for rare classes (23.73% of rare classes will have zero an-notations). We chose the 20% and 40% ratios to cover different scenarios (95.54% and 88.76% of rare classes are preserved that have at least one annotation).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_9\",\n \"tab_9\"\n ],\n \"text\": \"We experimented with the 20% and 40% missing ratios on our modified LVIS dataset. The results are reported in Table 10 where the checkmark symbol means that CascadeMatch is applied to the model. In both settings, we observe a clear margin between CascadeMatch and the baseline: +1.8% and +2.0% gains in terms of overall AP under the settings of 20% and 40% missing ratios, respectively. Notably, the gains are more apparent for the rare classes, with +3.3% and +2.9% gains for the two settings, respectively. The quantitative results shown in Table 10 strongly demonstrate the ability of CascadeMatch in dealing with the SAOD problem.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Qualitative Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We also show the visualization results of the pseudo-labeling under the sparselyannotated object detection (SAOD) setting in Figure 5. The first column refers to the ground truth labels from the original LVIS dataset. The second column shows our modified sparsely-annotated LVIS dataset where some annotations are randomly removed with a 40% missing rate and serves as the training set under the SAOD setting. The third column contains the prediction results of CascadeMatch. We observe that CascadeMatch can recover some labels. Since the original LVIS datasets is sparsely-annotated, CascadeMatch can also detect objects whose labels are missing in the original LVIS dataset. The qualitative results in Figure 5 explain the excellent performance of CascadeMatch on the SAOD task.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitation\",\n \"publication_ref\": [\n \"b40\",\n \"b50\",\n \"b38\",\n \"b4\",\n \"b66\",\n \"b5\",\n \"b22\"\n ],\n \"table_ref\": [],\n \"text\": \"The trade-off between speed and performance is one of the key research problems in the area of object detection [41,51,39,5,67,6]. It has been widely acknowledged that achieving a perfect speed-performance trade-off is extremely difficult [23]. To obtain a highperformance detector, one has to sacrifice on the speed, and vice versa. In this work, our CascadeMatch processes data in a cascade manner, which leads to longer training time and slower inference speed compared to the single-stage detector counterpart. However, given that the majority of computation takes place in the backbone while the detection heads are generally \\\"lightweight\\\" (as they only consist of a few fully connected layers), the lower speed is outweighed by the improvements in performance. To further improve the efficiency in real-world deployment, one could apply model compression techniques to reduce the model size, and LVIS (Original) LVIS (Missing 40%) Prediction (Ours)\\nFig. 5: The pseudo labels generated on the LVIS training dataset under the sparsely-annotated object detection setting (SAOD) setting. In the third column, green color refers to the predicted results that can be found in the ground truth of the first column; purple color refers to predicted results that are also missing in the original LVIS dataset (Zoom in for best view).\\ndesign more lightweight architectures for the cascade detection heads.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Our research addresses an important but largely under-studied problem in object detection, concerning both long-tailed data distributions and semi-supervised learning. The proposed approach, CascadeMatch, carefully integrates pseudo-labeling, coupled with a cascade design and an adaptive threshold tuning mechanism, into a variety of backbones and detection frameworks, such as the widely used region proposal-based detectors and more recent fully end-to-end detectors.\\nThe results strongly demonstrate that CascadeMatch is a better design than existing state-of-the-art semisupervised detectors in handling long-tailed datasets such as LVIS and COCO-LT. The capability to cope with the sparsely-annotated object detection problem is also well justified.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Acknowledgement This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). It is also partly supported by the NTU NAP grant and Singapore MOE AcRF Tier 2 (MOE-T2EP20120-0001).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Data Availability Statements\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The datasets analysed during this study are all publicly available for the research purpose -the LVIS and COCO datasets.\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-05-24"},"doi":{"kind":"string","value":""},"references":{"kind":"list like","value":[{"authors":"Eric Arazo","journal":"","ref_id":"b0","title":"Pseudo-labeling and confirmation bias in deep semi-supervised learning","year":"2020"},{"authors":"Philip Bachman; Ouais Alsharif; Doina Precup","journal":"NeurIPS","ref_id":"b1","title":"Learning with pseudo-ensembles","year":"2014"},{"authors":"David Berthelot","journal":"NeurIPS","ref_id":"b2","title":"MixMatch: A holistic approach to semi-supervised learning","year":"2019"},{"authors":"David Berthelot","journal":"ICLR","ref_id":"b3","title":"Remixmatch: Semisupervised learning with distribution alignment and augmentation anchoring","year":"2020"},{"authors":"Zhaowei Cai; Nuno Vasconcelos","journal":"TPAMI","ref_id":"b4","title":"Cascade R-CNN: high quality object detection and instance segmentation","year":"2019"},{"authors":"Nicolas Carion","journal":"","ref_id":"b5","title":"End-to-end object detection with transformers","year":"2020"},{"authors":"Nadine Chang","journal":"","ref_id":"b6","title":"Image-Level or Object-Level? A Tale of Two Resampling Strategies for Long-Tailed Detection","year":"2021"},{"authors":"Binbin Chen","journal":"","ref_id":"b7","title":"Label Matching Semi-Supervised Object Detection","year":"2022"},{"authors":"Binghui Chen","journal":"","ref_id":"b8","title":"Dense Learning based Semi-Supervised Object Detection","year":"2022"},{"authors":"Achal Dave","journal":"","ref_id":"b9","title":"Evaluating Large-Vocabulary Object Detectors: The Devil is in the Details","year":"2021"},{"authors":"Fan Yue","journal":"","ref_id":"b10","title":"CoSSL: Co-Learning of Representation and Classifier for Imbalanced Semi-Supervised Learning","year":"2022"},{"authors":"Chengjian Feng; Yujie Zhong; Weilin Huang","journal":"","ref_id":"b11","title":"Exploring Classification Equilibrium in Long-Tailed Object Detection","year":"2021"},{"authors":"Jiyang Gao","journal":"","ref_id":"b12","title":"NOTE-RCNN: Noise tolerant ensemble rcnn for semi-supervised object detection","year":"2019"},{"authors":"Golnaz Ghiasi","journal":"","ref_id":"b13","title":"Simple copy-paste is a strong data augmentation method for instance segmentation","year":"2021"},{"authors":"Qiushan Guo","journal":"","ref_id":"b14","title":"Scale-Equivalent Distillation for Semi-Supervised Object Detection","year":"2022"},{"authors":"Agrim Gupta; Piotr Dollar; Ross Girshick","journal":"","ref_id":"b15","title":"LVIS: A dataset for large vocabulary instance segmentation","year":"2019"},{"authors":" Han","journal":"NeurIPS","ref_id":"b16","title":"Co-teaching: Robust Training of Deep Neural Networks with Extremely Noisy Labels","year":"2018"},{"authors":"Kaiming He","journal":"","ref_id":"b17","title":"Deep residual learning for image recognition","year":"2016"},{"authors":"Ruifei He; Jihan Yang; Xiaojuan Qi","journal":"","ref_id":"b18","title":"Redistributing Biased Pseudo Labels for Semisupervised Semantic Segmentation: A Baseline Investigation","year":"2021"},{"authors":"Yin-Yin He","journal":"","ref_id":"b19","title":"Relieving Long-tailed Instance Segmentation via Pairwise Class Balance","year":"2022"},{"authors":"Hanzhe Hu","journal":"NeurIPS","ref_id":"b20","title":"Semi-Supervised Semantic Segmentation via Adaptive Equalization Learning","year":"2021"},{"authors":"Xinting Hu","journal":"","ref_id":"b21","title":"Learning to Segment the Tail","year":"2020"},{"authors":"Jonathan Huang","journal":"","ref_id":"b22","title":"Speed/accuracy tradeoffs for modern convolutional object detectors","year":"2017"},{"authors":"Minsung Hyun; Jisoo Jeong; Nojun Kwak","journal":"","ref_id":"b23","title":"Class-imbalanced semi-supervised learning","year":"2020"},{"authors":"Ahmet Iscen","journal":"","ref_id":"b24","title":"Label propagation for deep semi-supervised learning","year":"2019"},{"authors":"Jisoo Jeong","journal":"NeurIPS","ref_id":"b25","title":"Consistency-based Semisupervised Learning for Object detection","year":"2019"},{"authors":"Jisoo Jeong","journal":"","ref_id":"b26","title":"Interpolation-based semisupervised learning for object detection","year":"2021"},{"authors":"Jaehyung Kim","journal":"NeurIPS","ref_id":"b27","title":"Distribution Aligning Refinery of Pseudo-label for Imbalanced Semisupervised Learning","year":"2020"},{"authors":"Samuli Laine; Timo Aila","journal":"ICLR","ref_id":"b28","title":"Temporal ensembling for semi-supervised learning","year":"2017"},{"authors":"Dong-Hyun Lee","journal":"","ref_id":"b29","title":"Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks","year":"2013"},{"authors":"Hyuck Lee; Seungjae Shin; Heeyoung Kim","journal":"NeurIPS","ref_id":"b30","title":"ABC: Auxiliary Balanced Classifier for Class-imbalanced Semi-supervised Learning","year":"2021"},{"authors":"Aoxue Li; Peng Yuan; Zhenguo Li","journal":"","ref_id":"b31","title":"Semi-Supervised Object Detection via Multi-Instance Alignment With Global Class Prototypes","year":"2022"},{"authors":"Bo Li","journal":"","ref_id":"b32","title":"Equalized focal loss for dense longtailed object detection","year":"2022"},{"authors":"Hengduo Li","journal":"","ref_id":"b33","title":"Rethinking pseudo labels for semi-supervised object detection","year":"2022"},{"authors":"Shuang Li","journal":"","ref_id":"b34","title":"MetaSAug: Meta Semantic Augmentation for Long-Tailed Visual Recognition","year":"2021"},{"authors":"Yandong Li","journal":"","ref_id":"b35","title":"Improving Object Detection with Selective Self-supervised Self-training","year":"2020"},{"authors":"Yu Li","journal":"","ref_id":"b36","title":"Overcoming Classifier Imbalance for Long-Tail Object Detection With Balanced Group Softmax","year":"2020"},{"authors":"Tsung-Yi Lin","journal":"","ref_id":"b37","title":"Feature pyramid networks for object detection","year":"2017"},{"authors":"Tsung-Yi Lin","journal":"","ref_id":"b38","title":"Focal loss for dense object detection","year":"2017"},{"authors":"Tsung-Yi Lin","journal":"","ref_id":"b39","title":"Microsoft COCO: Common objects in context","year":"2014"},{"authors":"Wei Liu","journal":"","ref_id":"b40","title":"SSD: Single shot multibox detector","year":"2016"},{"authors":"Yen-Cheng Liu; Chih-Yao Ma; Zsolt Kira","journal":"","ref_id":"b41","title":"Unbiased Teacher v2: Semi-Supervised Object Detection for Anchor-Free and Anchor-Based Detectors","year":"2022"},{"authors":"Yen-Cheng Liu","journal":"ICLR","ref_id":"b42","title":"Unbiased teacher for semisupervised object detection","year":"2021"},{"authors":"Ze Liu","journal":"","ref_id":"b43","title":"Swin transformer: Hierarchical vision transformer using shifted windows","year":"2021"},{"authors":"Mi Peng","journal":"","ref_id":"b44","title":"Active Teacher for Semi-Supervised Object Detection","year":"2022"},{"authors":"Ishan Misra; Abhinav Shrivastava; Martial Hebert","journal":"","ref_id":"b45","title":"Watch and learn: Semi-supervised learning for object detectors from video","year":"2015"},{"authors":"Youngtaek Oh; Dong-Jin Kim; In So Kweon","journal":"","ref_id":"b46","title":"Distribution-aware semantics-oriented pseudolabel for imbalanced semi-supervised learning","year":"2022"},{"authors":"Siyuan Qiao","journal":"","ref_id":"b47","title":"Deep co-training for semisupervised image recognition","year":"2018"},{"authors":"Antti Rasmus","journal":"NeurIPS","ref_id":"b48","title":"Semi-supervised learning with ladder networks","year":"2016"},{"authors":"Jiawei Ren","journal":"NeurIPS","ref_id":"b49","title":"Balanced Meta-Softmax for Long-Tailed Visual Recognition","year":"2020"},{"authors":"Shaoqing Ren","journal":"NeurIPS","ref_id":"b50","title":"Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks","year":"2015"},{"authors":"Hamid Rezatofighi","journal":"","ref_id":"b51","title":"Generalized intersection over union: A metric and a loss for bounding box regression","year":"2019"},{"authors":"Mamshad Nayeem; Rizve ","journal":"ICLR","ref_id":"b52","title":"In defense of pseudo-labeling: An uncertainty-aware pseudolabel selection framework for semi-supervised learning","year":"2021"},{"authors":"Chuck Rosenberg; Martial Hebert; Henry Schneiderman","journal":"","ref_id":"b53","title":"Semi-supervised self-training of object detection models","year":"2005"},{"authors":"Aruni Roychowdhury","journal":"","ref_id":"b54","title":"Automatic adaptation of object detectors to new domains using selftraining","year":"2019"},{"authors":"Mehdi Sajjadi; Mehran Javanmardi; Tolga Tasdizen","journal":"NeurIPS","ref_id":"b55","title":"Regularization with stochastic transformations and perturbations for deep semisupervised learning","year":"2016"},{"authors":"Li Shen; Zhouchen Lin; Qingming Huang","journal":"","ref_id":"b56","title":"Relay backpropagation for effective learning of deep convolutional neural networks","year":"2016"},{"authors":"Kihyuk Sohn","journal":"","ref_id":"b57","title":"A simple semi-supervised learning framework for object detection","year":"2020"},{"authors":"Kihyuk Sohn","journal":"NeurIPS","ref_id":"b58","title":"FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence","year":"2020"},{"authors":"Peize Sun","journal":"","ref_id":"b59","title":"Sparse R-CNN: End-to-end object detection with learnable proposals","year":"2021"},{"authors":"Jingru Tan","journal":"","ref_id":"b60","title":"Equalization Loss for Long-Tailed Object Recognition","year":"2020"},{"authors":"Jingru Tan","journal":"","ref_id":"b61","title":"Equalization Loss v2: A New Gradient Balance Approach for Long-tailed Object Detection","year":"2021"},{"authors":"Peng Tang","journal":"","ref_id":"b62","title":"Proposal learning for semisupervised object detection","year":"2021"},{"authors":"Yihe Tang","journal":"","ref_id":"b63","title":"Humble Teachers Teach Better Students for Semi-Supervised Object Detection","year":"2021"},{"authors":"Yuxing Tang","journal":"","ref_id":"b64","title":"Large scale semi-supervised object detection using visual and semantic knowledge transfer","year":"2016"},{"authors":"Antti Tarvainen; Harri Valpola","journal":"NeurIPS","ref_id":"b65","title":"Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results","year":"2017"},{"authors":"Zhi Tian","journal":"","ref_id":"b66","title":"FCOS: Fully convolutional onestage object detection","year":"2019"},{"authors":"Jiaqi Wang","journal":"","ref_id":"b67","title":"Seesaw Loss for Long-Tailed Instance Segmentation","year":"2021"},{"authors":"Keze Wang","journal":"","ref_id":"b68","title":"Towards human-machine cooperation: Self-supervised sample mining for object detection","year":"2018"},{"authors":"Tao Wang","journal":"","ref_id":"b69","title":"The Devil is in Classification: A Simple Framework for Long-tail Instance Segmentation","year":"2020"},{"authors":"Tiancai Wang","journal":"AAAI","ref_id":"b70","title":"Co-mining: Self-Supervised Learning for Sparsely Annotated Object Detection","year":"2021"},{"authors":"Tong Wang","journal":"","ref_id":"b71","title":"Adaptive Class Suppression Loss for Long-Tail Object Detection","year":"2021"},{"authors":"Wenhai Wang","journal":"","ref_id":"b72","title":"Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions","year":"2021"},{"authors":"Zhenyu Wang","journal":"","ref_id":"b73","title":"Data-Uncertainty Guided Multi-Phase Learning for Semi-Supervised Object Detection","year":"2021"},{"authors":"Chen Wei","journal":"","ref_id":"b74","title":"CReST: A class-rebalancing self-training framework for imbalanced semisupervised learning","year":"2021"},{"authors":"Jialian Wu","journal":"ACM MM","ref_id":"b75","title":"Forest R-CNN: Largevocabulary long-tailed object detection and instance segmentation","year":"2020"},{"authors":"Zhe Wu","journal":"","ref_id":"b76","title":"Soft sampling for robust object detection","year":"2019"},{"authors":"Qizhe Xie","journal":"","ref_id":"b77","title":"Self-training with noisy student improves imagenet classification","year":"2020"},{"authors":"Qizhe Xie","journal":"NeurIPS","ref_id":"b78","title":"Unsupervised Data Augmentation for Consistency Training","year":"2020"},{"authors":"Mengde Xu","journal":"","ref_id":"b79","title":"End-to-End Semi-Supervised Object Detection with Soft Teacher","year":"2021"},{"authors":"Fan Yang","journal":"","ref_id":"b80","title":"Class-Aware Contrastive Semi-Supervised Learning","year":"2022"},{"authors":"Qize Yang","journal":"","ref_id":"b81","title":"Interactive self-training with mean teachers for semi-supervised object detection","year":"2021"},{"authors":"Yuzhe Yang; Zhi Xu","journal":"NeurIPS","ref_id":"b82","title":"Rethinking the Value of Labels for Improving Class-Imbalanced Learning","year":"2020"},{"authors":"Yuhang Zang; Chen Huang; Chen Change Loy","journal":"","ref_id":"b83","title":"FASA: Feature Augmentation and Sampling Adaptation for Long-Tailed Instance Segmentation","year":"2021"},{"authors":"Cheng Zhang","journal":"","ref_id":"b84","title":"MosaicOS: A Simple and Effective Use of Object-Centric Images for Long-Tailed Object Detection","year":"2021"},{"authors":"Fangyuan Zhang; Tianxiang Pan; Bin Wang","journal":"","ref_id":"b85","title":"Semi-supervised object detection with adaptive class-rebalancing self-training","year":"2022"},{"authors":"Han Zhang","journal":"ICASSP","ref_id":"b86","title":"Solving Missing-Annotation Object Detection with Background Recalibration Loss","year":"2020"},{"authors":"Songyang Zhang","journal":"","ref_id":"b87","title":"Distribution Alignment: A Unified Framework for Long-tail Visual Recognition","year":"2021"},{"authors":"Yifan Zhang","journal":"","ref_id":"b88","title":"Deep Long-Tailed Learning: A Survey","year":"2021"},{"authors":"Mingkai Zheng","journal":"","ref_id":"b89","title":"SimMatch: Semisupervised Learning with Similarity Matching","year":"2022"},{"authors":"Qiang Zhou","journal":"","ref_id":"b90","title":"Instant-Teaching: An Endto-End Semi-Supervised Object Detection Framework","year":"2021"},{"authors":"Xingyi Zhou; Vladlen Koltun; Philipp Krähenbühl","journal":"","ref_id":"b91","title":"Probabilistic two-stage detection","year":"2021"}],"string":"[\n {\n \"authors\": \"Eric Arazo\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Pseudo-labeling and confirmation bias in deep semi-supervised learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Philip Bachman; Ouais Alsharif; Doina Precup\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b1\",\n \"title\": \"Learning with pseudo-ensembles\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"David Berthelot\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b2\",\n \"title\": \"MixMatch: A holistic approach to semi-supervised learning\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"David Berthelot\",\n \"journal\": \"ICLR\",\n \"ref_id\": \"b3\",\n \"title\": \"Remixmatch: Semisupervised learning with distribution alignment and augmentation anchoring\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Zhaowei Cai; Nuno Vasconcelos\",\n \"journal\": \"TPAMI\",\n \"ref_id\": \"b4\",\n \"title\": \"Cascade R-CNN: high quality object detection and instance segmentation\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Nicolas Carion\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"End-to-end object detection with transformers\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Nadine Chang\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Image-Level or Object-Level? A Tale of Two Resampling Strategies for Long-Tailed Detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Binbin Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Label Matching Semi-Supervised Object Detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Binghui Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Dense Learning based Semi-Supervised Object Detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Achal Dave\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Evaluating Large-Vocabulary Object Detectors: The Devil is in the Details\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Fan Yue\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"CoSSL: Co-Learning of Representation and Classifier for Imbalanced Semi-Supervised Learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Chengjian Feng; Yujie Zhong; Weilin Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Exploring Classification Equilibrium in Long-Tailed Object Detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jiyang Gao\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"NOTE-RCNN: Noise tolerant ensemble rcnn for semi-supervised object detection\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Golnaz Ghiasi\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Simple copy-paste is a strong data augmentation method for instance segmentation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Qiushan Guo\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Scale-Equivalent Distillation for Semi-Supervised Object Detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Agrim Gupta; Piotr Dollar; Ross Girshick\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"LVIS: A dataset for large vocabulary instance segmentation\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \" Han\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b16\",\n \"title\": \"Co-teaching: Robust Training of Deep Neural Networks with Extremely Noisy Labels\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Kaiming He\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Deep residual learning for image recognition\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Ruifei He; Jihan Yang; Xiaojuan Qi\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Redistributing Biased Pseudo Labels for Semisupervised Semantic Segmentation: A Baseline Investigation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yin-Yin He\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Relieving Long-tailed Instance Segmentation via Pairwise Class Balance\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hanzhe Hu\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b20\",\n \"title\": \"Semi-Supervised Semantic Segmentation via Adaptive Equalization Learning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Xinting Hu\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Learning to Segment the Tail\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jonathan Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Speed/accuracy tradeoffs for modern convolutional object detectors\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Minsung Hyun; Jisoo Jeong; Nojun Kwak\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Class-imbalanced semi-supervised learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ahmet Iscen\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Label propagation for deep semi-supervised learning\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Jisoo Jeong\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b25\",\n \"title\": \"Consistency-based Semisupervised Learning for Object detection\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Jisoo Jeong\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Interpolation-based semisupervised learning for object detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jaehyung Kim\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b27\",\n \"title\": \"Distribution Aligning Refinery of Pseudo-label for Imbalanced Semisupervised Learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Samuli Laine; Timo Aila\",\n \"journal\": \"ICLR\",\n \"ref_id\": \"b28\",\n \"title\": \"Temporal ensembling for semi-supervised learning\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Dong-Hyun Lee\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Hyuck Lee; Seungjae Shin; Heeyoung Kim\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b30\",\n \"title\": \"ABC: Auxiliary Balanced Classifier for Class-imbalanced Semi-supervised Learning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Aoxue Li; Peng Yuan; Zhenguo Li\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Semi-Supervised Object Detection via Multi-Instance Alignment With Global Class Prototypes\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Bo Li\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Equalized focal loss for dense longtailed object detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hengduo Li\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Rethinking pseudo labels for semi-supervised object detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Shuang Li\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"MetaSAug: Meta Semantic Augmentation for Long-Tailed Visual Recognition\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yandong Li\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Improving Object Detection with Selective Self-supervised Self-training\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yu Li\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Overcoming Classifier Imbalance for Long-Tail Object Detection With Balanced Group Softmax\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Tsung-Yi Lin\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Feature pyramid networks for object detection\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Tsung-Yi Lin\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Focal loss for dense object detection\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Tsung-Yi Lin\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Microsoft COCO: Common objects in context\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Wei Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"SSD: Single shot multibox detector\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Yen-Cheng Liu; Chih-Yao Ma; Zsolt Kira\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Unbiased Teacher v2: Semi-Supervised Object Detection for Anchor-Free and Anchor-Based Detectors\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yen-Cheng Liu\",\n \"journal\": \"ICLR\",\n \"ref_id\": \"b42\",\n \"title\": \"Unbiased teacher for semisupervised object detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ze Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"Swin transformer: Hierarchical vision transformer using shifted windows\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Mi Peng\",\n \"journal\": \"\",\n \"ref_id\": \"b44\",\n \"title\": \"Active Teacher for Semi-Supervised Object Detection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ishan Misra; Abhinav Shrivastava; Martial Hebert\",\n \"journal\": \"\",\n \"ref_id\": \"b45\",\n \"title\": \"Watch and learn: Semi-supervised learning for object detectors from video\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Youngtaek Oh; Dong-Jin Kim; In So Kweon\",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"Distribution-aware semantics-oriented pseudolabel for imbalanced semi-supervised learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Siyuan Qiao\",\n \"journal\": \"\",\n \"ref_id\": \"b47\",\n \"title\": \"Deep co-training for semisupervised image recognition\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Antti Rasmus\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b48\",\n \"title\": \"Semi-supervised learning with ladder networks\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Jiawei Ren\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b49\",\n \"title\": \"Balanced Meta-Softmax for Long-Tailed Visual Recognition\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Shaoqing Ren\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b50\",\n \"title\": \"Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Hamid Rezatofighi\",\n \"journal\": \"\",\n \"ref_id\": \"b51\",\n \"title\": \"Generalized intersection over union: A metric and a loss for bounding box regression\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Mamshad Nayeem; Rizve \",\n \"journal\": \"ICLR\",\n \"ref_id\": \"b52\",\n \"title\": \"In defense of pseudo-labeling: An uncertainty-aware pseudolabel selection framework for semi-supervised learning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Chuck Rosenberg; Martial Hebert; Henry Schneiderman\",\n \"journal\": \"\",\n \"ref_id\": \"b53\",\n \"title\": \"Semi-supervised self-training of object detection models\",\n \"year\": \"2005\"\n },\n {\n \"authors\": \"Aruni Roychowdhury\",\n \"journal\": \"\",\n \"ref_id\": \"b54\",\n \"title\": \"Automatic adaptation of object detectors to new domains using selftraining\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Mehdi Sajjadi; Mehran Javanmardi; Tolga Tasdizen\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b55\",\n \"title\": \"Regularization with stochastic transformations and perturbations for deep semisupervised learning\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Li Shen; Zhouchen Lin; Qingming Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b56\",\n \"title\": \"Relay backpropagation for effective learning of deep convolutional neural networks\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Kihyuk Sohn\",\n \"journal\": \"\",\n \"ref_id\": \"b57\",\n \"title\": \"A simple semi-supervised learning framework for object detection\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Kihyuk Sohn\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b58\",\n \"title\": \"FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Peize Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b59\",\n \"title\": \"Sparse R-CNN: End-to-end object detection with learnable proposals\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jingru Tan\",\n \"journal\": \"\",\n \"ref_id\": \"b60\",\n \"title\": \"Equalization Loss for Long-Tailed Object Recognition\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jingru Tan\",\n \"journal\": \"\",\n \"ref_id\": \"b61\",\n \"title\": \"Equalization Loss v2: A New Gradient Balance Approach for Long-tailed Object Detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Peng Tang\",\n \"journal\": \"\",\n \"ref_id\": \"b62\",\n \"title\": \"Proposal learning for semisupervised object detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yihe Tang\",\n \"journal\": \"\",\n \"ref_id\": \"b63\",\n \"title\": \"Humble Teachers Teach Better Students for Semi-Supervised Object Detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yuxing Tang\",\n \"journal\": \"\",\n \"ref_id\": \"b64\",\n \"title\": \"Large scale semi-supervised object detection using visual and semantic knowledge transfer\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Antti Tarvainen; Harri Valpola\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b65\",\n \"title\": \"Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Zhi Tian\",\n \"journal\": \"\",\n \"ref_id\": \"b66\",\n \"title\": \"FCOS: Fully convolutional onestage object detection\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Jiaqi Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b67\",\n \"title\": \"Seesaw Loss for Long-Tailed Instance Segmentation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Keze Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b68\",\n \"title\": \"Towards human-machine cooperation: Self-supervised sample mining for object detection\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Tao Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b69\",\n \"title\": \"The Devil is in Classification: A Simple Framework for Long-tail Instance Segmentation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Tiancai Wang\",\n \"journal\": \"AAAI\",\n \"ref_id\": \"b70\",\n \"title\": \"Co-mining: Self-Supervised Learning for Sparsely Annotated Object Detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Tong Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b71\",\n \"title\": \"Adaptive Class Suppression Loss for Long-Tail Object Detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Wenhai Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b72\",\n \"title\": \"Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zhenyu Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b73\",\n \"title\": \"Data-Uncertainty Guided Multi-Phase Learning for Semi-Supervised Object Detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Chen Wei\",\n \"journal\": \"\",\n \"ref_id\": \"b74\",\n \"title\": \"CReST: A class-rebalancing self-training framework for imbalanced semisupervised learning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jialian Wu\",\n \"journal\": \"ACM MM\",\n \"ref_id\": \"b75\",\n \"title\": \"Forest R-CNN: Largevocabulary long-tailed object detection and instance segmentation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Zhe Wu\",\n \"journal\": \"\",\n \"ref_id\": \"b76\",\n \"title\": \"Soft sampling for robust object detection\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Qizhe Xie\",\n \"journal\": \"\",\n \"ref_id\": \"b77\",\n \"title\": \"Self-training with noisy student improves imagenet classification\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Qizhe Xie\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b78\",\n \"title\": \"Unsupervised Data Augmentation for Consistency Training\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Mengde Xu\",\n \"journal\": \"\",\n \"ref_id\": \"b79\",\n \"title\": \"End-to-End Semi-Supervised Object Detection with Soft Teacher\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Fan Yang\",\n \"journal\": \"\",\n \"ref_id\": \"b80\",\n \"title\": \"Class-Aware Contrastive Semi-Supervised Learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Qize Yang\",\n \"journal\": \"\",\n \"ref_id\": \"b81\",\n \"title\": \"Interactive self-training with mean teachers for semi-supervised object detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yuzhe Yang; Zhi Xu\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b82\",\n \"title\": \"Rethinking the Value of Labels for Improving Class-Imbalanced Learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yuhang Zang; Chen Huang; Chen Change Loy\",\n \"journal\": \"\",\n \"ref_id\": \"b83\",\n \"title\": \"FASA: Feature Augmentation and Sampling Adaptation for Long-Tailed Instance Segmentation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Cheng Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b84\",\n \"title\": \"MosaicOS: A Simple and Effective Use of Object-Centric Images for Long-Tailed Object Detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Fangyuan Zhang; Tianxiang Pan; Bin Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b85\",\n \"title\": \"Semi-supervised object detection with adaptive class-rebalancing self-training\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Han Zhang\",\n \"journal\": \"ICASSP\",\n \"ref_id\": \"b86\",\n \"title\": \"Solving Missing-Annotation Object Detection with Background Recalibration Loss\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Songyang Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b87\",\n \"title\": \"Distribution Alignment: A Unified Framework for Long-tail Visual Recognition\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yifan Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b88\",\n \"title\": \"Deep Long-Tailed Learning: A Survey\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Mingkai Zheng\",\n \"journal\": \"\",\n \"ref_id\": \"b89\",\n \"title\": \"SimMatch: Semisupervised Learning with Similarity Matching\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Qiang Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b90\",\n \"title\": \"Instant-Teaching: An Endto-End Semi-Supervised Object Detection Framework\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Xingyi Zhou; Vladlen Koltun; Philipp Krähenbühl\",\n \"journal\": \"\",\n \"ref_id\": \"b91\",\n \"title\": \"Probabilistic two-stage detection\",\n \"year\": \"2021\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[5,42.11,408.78,238.07,31.61],"formula_id":"formula_0","formula_text":"ℓ labeled cls = (x,y * )∼D l K k=1 Cls(y * , p k (y|x, b k-1 )),(1)"},{"formula_coordinates":[5,42.11,446.09,233.83,31.61],"formula_id":"formula_1","formula_text":"ℓ labeled reg = (x,b * )∼D l K k=1 Reg(b * , b k ). (2"},{"formula_coordinates":[5,275.94,456.5,4.24,8.74],"formula_id":"formula_2","formula_text":")"},{"formula_coordinates":[5,42.11,690.17,238.07,30.55],"formula_id":"formula_3","formula_text":"p t = 1 K K k=1 p k (y|x, b k-1 ) and b t = 1 K K k=1 b k ,(3)"},{"formula_coordinates":[5,297.19,408.49,238.07,42.76],"formula_id":"formula_4","formula_text":"ℓ unlabeled cls = x∼Du K k=1 1(q t ≥ τ qt k ) Cls(q t , p k (y|x, b k-1 )),(4)"},{"formula_coordinates":[5,297.19,458.04,238.07,30.55],"formula_id":"formula_5","formula_text":"ℓ unlabeled reg = x∼Du K k=1 1(q t ≥ τ qt k ) Reg(b t , b k ),(5)"},{"formula_coordinates":[5,297.19,579.89,238.07,24.65],"formula_id":"formula_6","formula_text":"ℓ roi cls = ℓ labeled cls + λ u • ℓ unlabeled cls , an ROI regression loss ℓ roi reg = ℓ labeled reg + λ u • ℓ unlabeled reg"},{"formula_coordinates":[8,312.37,144.59,110.29,9.23],"formula_id":"formula_7","formula_text":"ϵ 1 ϵ 2 ϵ 3 AP Fix AP Fix"},{"formula_coordinates":[8,66,273.48,141.85,22.26],"formula_id":"formula_8","formula_text":"K AP Fix AP Fix r AP Fix c AP Fix f T train1"}],"string":"[\n {\n \"formula_coordinates\": [\n 5,\n 42.11,\n 408.78,\n 238.07,\n 31.61\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"ℓ labeled cls = (x,y * )∼D l K k=1 Cls(y * , p k (y|x, b k-1 )),(1)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 42.11,\n 446.09,\n 233.83,\n 31.61\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"ℓ labeled reg = (x,b * )∼D l K k=1 Reg(b * , b k ). (2\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 275.94,\n 456.5,\n 4.24,\n 8.74\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 42.11,\n 690.17,\n 238.07,\n 30.55\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"p t = 1 K K k=1 p k (y|x, b k-1 ) and b t = 1 K K k=1 b k ,(3)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 297.19,\n 408.49,\n 238.07,\n 42.76\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"ℓ unlabeled cls = x∼Du K k=1 1(q t ≥ τ qt k ) Cls(q t , p k (y|x, b k-1 )),(4)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 297.19,\n 458.04,\n 238.07,\n 30.55\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"ℓ unlabeled reg = x∼Du K k=1 1(q t ≥ τ qt k ) Reg(b t , b k ),(5)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 297.19,\n 579.89,\n 238.07,\n 24.65\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"ℓ roi cls = ℓ labeled cls + λ u • ℓ unlabeled cls , an ROI regression loss ℓ roi reg = ℓ labeled reg + λ u • ℓ unlabeled reg\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 312.37,\n 144.59,\n 110.29,\n 9.23\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"ϵ 1 ϵ 2 ϵ 3 AP Fix AP Fix\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 66,\n 273.48,\n 141.85,\n 22.26\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"K AP Fix AP Fix r AP Fix c AP Fix f T train1\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"Semi-Supervised and Long-Tailed Object Detection with CascadeMatch"},"abstract":{"kind":"string","value":"This paper focuses on long-tailed object detection in the semi-supervised learning setting, which poses realistic challenges, but has rarely been studied in the literature. We propose a novel pseudolabeling-based detector called CascadeMatch. Our detector features a cascade network architecture, which has multi-stage detection heads with progressive confidence thresholds. To avoid manually tuning the thresholds, we design a new adaptive pseudo-label mining mechanism to automatically identify suitable values from data. To mitigate confirmation bias, where a model is negatively reinforced by incorrect pseudolabels produced by itself, each detection head is trained by the ensemble pseudo-labels of all detection heads. Experiments on two long-tailed datasets, i.e., LVIS and COCO-LT, demonstrate that CascadeMatch surpasses existing state-of-the-art semi-supervised approachesacross a wide range of detection architectures-in handling long-tailed object detection. For instance, Cas-cadeMatch outperforms Unbiased Teacher by 1.9 AP Fix on LVIS when using a ResNet50-based Cascade R-CNN structure, and by 1.7 AP Fix when using Sparse R-CNN with a Transformer encoder. We also show that CascadeMatch can even handle the challenging"},"authors":{"kind":"string","value":"Yuhang Zang; Kaiyang Zhou; Chen Huang; Chen Change Loy"},"figures":{"kind":"list like","value":[{"figure_caption":"Fig. 1 :1Fig. 1: Motivation of our research. (a) The Average Precision (AP) and Average Recall (AR) curves, obtained using different fixed confidence thresholds (denoted by τ ). Clearly, none of the chosen thresholds gives the best trade-off. (b) The distribution of prediction scores for a long-tailed dataset, which shows a high degree of imbalance between the three class groups. (c) Sorted number of samples per class seen by the model during training. CascadeMatch retains much more pseudo-labeled samples than Unbiased Teacher with respect to the common and rare classes.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Fig. 2 :2Fig. 2: The pipeline of our approach. (a): Overview of CascadeMatch's cascade pseudo-labeling module. The supervision signal for unlabeled data corresponds to the ensembled pseudo label. Confidence thresholds, {τ k } k∈1,...,3 , are independently computed for each stage via our adaptive pseudo-label mining module. (b): Computation of the adaptive pseudo-label mining module. The classification confidence values predicted for each class c ∈ {1, . . . , C} on labeled proposals are aggregated in the per-class queue. For class c, the confidence value distribution is estimated where the mean µ c and the standard deviation σ c are used to determine the class-specific thresholdτ c k at the k-th cascade stage.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 2(a)), and then select a threshold such that a certain percentage of the confidence values can pass through. The challenge lies in how to select the threshold with minimal human intervention. We automate the selection process by (1) computing the mean µ c and the standard deviation σ c based on the confidence values for each class, and (2) setting the class-specific threshold τ c k for stage-k as τ c k = µ c + σ c * ϵ k . An illustration is shown in Figure 2(b).","figure_data":"","figure_id":"fig_2","figure_label":"","figure_type":"figure"},{"figure_caption":"Fig. 4 :4Fig. 4: The pseudo labels generated on the LVIS training dataset under the semi-supervised object detection setting (SSOD) setting. The green color refers to the true-positive predicted results; purple color refers to false-positive detection results (Zoom in for best view).","figure_data":"","figure_id":"fig_5","figure_label":"4","figure_type":"figure"},{"figure_caption":", List of hyper-parameters used for different detectors.","figure_data":"Hyper-parameterDetectorValueOptimizerSGDLearning RateCascade R-CNN0.01Weight Decay0.0001OptimizerAdamWLearning RateSparse R-CNN0.000025Weight Decay0.0001Input Image Size[1333, 800]Batch Size for Labeled DataBoth16Batch Size for Unlabeled Data1622, 88, 68, 12, 7, 92]. Three class groups are defined inLVIS v1.0: rare [1, 10), common [10, 100), and frequent[100, -) based on the number of images that contain atleast one instance of the corresponding class. COCO-LT [70] is used to demonstrate the generalizability ofour approach.","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"Ablation studies on 1) cascade pseudo-labeling (CPL) and 2) adaptive pseudo-label mining (APM). The top row refers to the supervised learning baseline without using the unlabeled data.","figure_data":"CPL APM AP Fix AP Fix rAP Fix cAP Fix f✗✗26.319.725.330.3✓✗30.121.929.334.5✗✓28.922.527.932.8✓✓30.523.129.734.7","figure_id":"tab_1","figure_label":"2","figure_type":"table"},{"figure_caption":"Comparison of pseudo-label accuracy. The ensemble results is more accurate than each single head. See Figure4for visualization.","figure_data":"Iter.60k120k 180kHead 032.851.567.3Head 150.562.473.2Head 255.171.084.1Ensemble 66.4 79.5 88.9","figure_id":"tab_3","figure_label":"5","figure_type":"table"},{"figure_caption":"Ablation study on the loss","figure_data":"function weight balancing parameterℓ u cls . We select ℓ u cls = 1.0 that worksthe best.λ u AP Fix AP Fix rAP Fix cAP Fix f0.530.020.928.236.11.529.921.228.335.61.0 30.521.428.936.42.029.420.427.935.1","figure_id":"tab_4","figure_label":"6","figure_type":"table"},{"figure_caption":"","figure_data":", an ab-","figure_id":"tab_5","figure_label":"8","figure_type":"table"},{"figure_caption":"Results on COCO-LT validation set set. The symbols AP 1 , AP 2 , AP 3 and AP 4 denote the bin of[1,20),[20, 400), [400, 8000), [8000, -) training instances. The symbol 'UT' is the abbreviation of the Unbiased Teacher[43] algorithm.","figure_data":"MethodAPAP 1 AP 2 AP 3 AP 4Supervised25.42.5 16.2 29.9 33.7CSD25.9 (+0.5) 2.0 15.2 32.1 34.0STAC26.4 (+1.0) 2.2 16.3 32.4 34.1UT26.7 (+1.3) 2.2 18.0 31.8 34.3Ours27.8 (+2.4) 4.0 20.4 32.4 34.5","figure_id":"tab_7","figure_label":"8","figure_type":"table"},{"figure_caption":"Comparisons of training memory (MB), training time T train (sec/iter) and inference time T test (sec/iter) on the LVIS dataset.","figure_data":"MethodMemoryT trainT testSupervised58890.22480.2694CSD64520.33100.2767STAC68010.41100.2702Unbiased Teacher73660.46160.2761Soft Teacher80290.45890.2718LabelMatch82400.49180.2698Ours74320.47330.2734CNN framework, CascadeMatch outperforms UnbiasedTeacher by 1.9 AP Fix rand 1.6 AP Fix c . With Sparse R-CNN and the Transformer encoder, CascadeMatch alsogains clear improvements: 1.7 AP Fix and 2.9 AP Fix r .Such results show that our proposed method is generalto various architectures.","figure_id":"tab_8","figure_label":"9","figure_type":"table"},{"figure_caption":"Experiment results under the Sparsely annotated object detection (SAOD) setting where missing labels exist in the training set. We follow previous studies[87,71] to build a modified LVIS dataset where we randomly erase the annotations by 20% and 40% per object category.","figure_data":"Missing RatioOursAPAP rAP cAP f40%✗ ✓22.5 24.210.4 13.720.9 22.429.6 30.920%✗ ✓24.7 26.714.3 17.222.7 25.131.4 32.84.3 Sparsely Annotated Object DetectionBackground The standard semi-supervised learn-ing setting in object detection assumes that trainingimages are fully annotated. A more realistic settingthat has received increasing attention from the com-munity is sparsely annotated object detection [77, 87,71, 92], or SAOD. In the previous experiments, we haveshown that CascadeMatch performs favorably againstthe baselines with clear improvements. In this section,we unveil how CascadeMatch fares under the SAODsetting.","figure_id":"tab_9","figure_label":"10","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Fig. 1 :1Fig. 1: Motivation of our research. (a) The Average Precision (AP) and Average Recall (AR) curves, obtained using different fixed confidence thresholds (denoted by τ ). Clearly, none of the chosen thresholds gives the best trade-off. (b) The distribution of prediction scores for a long-tailed dataset, which shows a high degree of imbalance between the three class groups. (c) Sorted number of samples per class seen by the model during training. CascadeMatch retains much more pseudo-labeled samples than Unbiased Teacher with respect to the common and rare classes.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 2 :2Fig. 2: The pipeline of our approach. (a): Overview of CascadeMatch's cascade pseudo-labeling module. The supervision signal for unlabeled data corresponds to the ensembled pseudo label. Confidence thresholds, {τ k } k∈1,...,3 , are independently computed for each stage via our adaptive pseudo-label mining module. (b): Computation of the adaptive pseudo-label mining module. The classification confidence values predicted for each class c ∈ {1, . . . , C} on labeled proposals are aggregated in the per-class queue. For class c, the confidence value distribution is estimated where the mean µ c and the standard deviation σ c are used to determine the class-specific thresholdτ c k at the k-th cascade stage.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2(a)), and then select a threshold such that a certain percentage of the confidence values can pass through. The challenge lies in how to select the threshold with minimal human intervention. We automate the selection process by (1) computing the mean µ c and the standard deviation σ c based on the confidence values for each class, and (2) setting the class-specific threshold τ c k for stage-k as τ c k = µ c + σ c * ϵ k . An illustration is shown in Figure 2(b).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 4 :4Fig. 4: The pseudo labels generated on the LVIS training dataset under the semi-supervised object detection setting (SSOD) setting. The green color refers to the true-positive predicted results; purple color refers to false-positive detection results (Zoom in for best view).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \", List of hyper-parameters used for different detectors.\",\n \"figure_data\": \"Hyper-parameterDetectorValueOptimizerSGDLearning RateCascade R-CNN0.01Weight Decay0.0001OptimizerAdamWLearning RateSparse R-CNN0.000025Weight Decay0.0001Input Image Size[1333, 800]Batch Size for Labeled DataBoth16Batch Size for Unlabeled Data1622, 88, 68, 12, 7, 92]. Three class groups are defined inLVIS v1.0: rare [1, 10), common [10, 100), and frequent[100, -) based on the number of images that contain atleast one instance of the corresponding class. COCO-LT [70] is used to demonstrate the generalizability ofour approach.\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablation studies on 1) cascade pseudo-labeling (CPL) and 2) adaptive pseudo-label mining (APM). The top row refers to the supervised learning baseline without using the unlabeled data.\",\n \"figure_data\": \"CPL APM AP Fix AP Fix rAP Fix cAP Fix f✗✗26.319.725.330.3✓✗30.121.929.334.5✗✓28.922.527.932.8✓✓30.523.129.734.7\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison of pseudo-label accuracy. The ensemble results is more accurate than each single head. See Figure4for visualization.\",\n \"figure_data\": \"Iter.60k120k 180kHead 032.851.567.3Head 150.562.473.2Head 255.171.084.1Ensemble 66.4 79.5 88.9\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablation study on the loss\",\n \"figure_data\": \"function weight balancing parameterℓ u cls . We select ℓ u cls = 1.0 that worksthe best.λ u AP Fix AP Fix rAP Fix cAP Fix f0.530.020.928.236.11.529.921.228.335.61.0 30.521.428.936.42.029.420.427.935.1\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \", an ab-\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"8\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Results on COCO-LT validation set set. The symbols AP 1 , AP 2 , AP 3 and AP 4 denote the bin of[1,20),[20, 400), [400, 8000), [8000, -) training instances. The symbol 'UT' is the abbreviation of the Unbiased Teacher[43] algorithm.\",\n \"figure_data\": \"MethodAPAP 1 AP 2 AP 3 AP 4Supervised25.42.5 16.2 29.9 33.7CSD25.9 (+0.5) 2.0 15.2 32.1 34.0STAC26.4 (+1.0) 2.2 16.3 32.4 34.1UT26.7 (+1.3) 2.2 18.0 31.8 34.3Ours27.8 (+2.4) 4.0 20.4 32.4 34.5\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"8\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparisons of training memory (MB), training time T train (sec/iter) and inference time T test (sec/iter) on the LVIS dataset.\",\n \"figure_data\": \"MethodMemoryT trainT testSupervised58890.22480.2694CSD64520.33100.2767STAC68010.41100.2702Unbiased Teacher73660.46160.2761Soft Teacher80290.45890.2718LabelMatch82400.49180.2698Ours74320.47330.2734CNN framework, CascadeMatch outperforms UnbiasedTeacher by 1.9 AP Fix rand 1.6 AP Fix c . With Sparse R-CNN and the Transformer encoder, CascadeMatch alsogains clear improvements: 1.7 AP Fix and 2.9 AP Fix r .Such results show that our proposed method is generalto various architectures.\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"9\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Experiment results under the Sparsely annotated object detection (SAOD) setting where missing labels exist in the training set. We follow previous studies[87,71] to build a modified LVIS dataset where we randomly erase the annotations by 20% and 40% per object category.\",\n \"figure_data\": \"Missing RatioOursAPAP rAP cAP f40%✗ ✓22.5 24.210.4 13.720.9 22.429.6 30.920%✗ ✓24.7 26.714.3 17.222.7 25.131.4 32.84.3 Sparsely Annotated Object DetectionBackground The standard semi-supervised learn-ing setting in object detection assumes that trainingimages are fully annotated. A more realistic settingthat has received increasing attention from the com-munity is sparsely annotated object detection [77, 87,71, 92], or SAOD. In the previous experiments, we haveshown that CascadeMatch performs favorably againstthe baselines with clear improvements. In this section,we unveil how CascadeMatch fares under the SAODsetting.\",\n \"figure_id\": \"tab_9\",\n \"figure_label\": \"10\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Supporting Evidence\", \"Citation\": \"[40]\", \"Explanation\": \"The cited work, the COCO dataset, is used as a benchmark for evaluating the performance of semi-supervised object detectors in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work, LVIS, is a long-tailed dataset that has been used in the research community to study object detection. The citing paper builds upon this dataset to explore the challenges of long-tailed object detection in the semi-supervised setting.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[70]\", \"Explanation\": \"The cited work, COCO-LT, is another long-tailed dataset that has been used in the research community to study object detection. The citing paper also uses this dataset to address the problem of long-tailed object detection in the semi-supervised setting.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[43]\", \"Explanation\": \"The cited work, Unbiased Teacher, is a state-of-the-art semi-supervised detector that the citing paper uses to train a model on a long-tailed LVIS dataset. The method and techniques employed in the cited work form the basis for the research conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[43]\", \"Explanation\": \"The cited work provides a method for identifying class-wise threshold values in data, which the citing paper adopts to improve the performance of their approach in retaining more pseudolabels for common and rare classes.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[51,5]\", \"Explanation\": \"The cited works provide the detection architectures that the citing paper adopts in their research on object detection.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[60]\", \"Explanation\": \"The cited work introduces the Sparse R-CNN detector that the citing paper uses in their research on object detection.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[73]\", \"Explanation\": \"The cited work provides the PVT encoder that the citing paper utilizes in their research on object detection.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[77,87,71,92]\", \"Explanation\": \"The cited works are related to the challenging SAOD setting, which the citing paper further explores in their research on object detection.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[26]\", \"Explanation\": \"The cited work introduces the concept of consistency regularization for unlabeled images, which the citing paper adopts in the development of their semi-supervised object detection method.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[27]\", \"Explanation\": \"The cited work also contributes to the development of the semi-supervised object detection method by applying consistency regularization to mined bounding boxes for unlabeled images.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[58]\", \"Explanation\": \"The cited work introduces the concept of strong data augmentation for self-training, which the citing paper uses in the development of their semi-supervised object detection method.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[43]\", \"Explanation\": \"The cited work introduces the concept of focal loss to address the confirmation bias in pseudo-labeling-based methods, which the citing paper extends to further improve the performance of their semi-supervised object detection method.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[64]\", \"Explanation\": \"The cited work contributes to the development of pseudo-labeling-based methods by using a weighted EMA ensemble, a data ensemble, and advanced data augmentation, which the citing paper builds upon to further improve the performance of their semi-supervised object detection method.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[91]\", \"Explanation\": \"The cited work introduces the concept of co-training as a countermeasure to the confirmation bias in pseudo-labeling-based methods, which the citing paper extends to further improve the performance of their semi-supervised object detection method.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[34]\", \"Explanation\": \"The cited work by Li et al. proposes a dynamic thresholding method for class-based confidence levels, which the citing paper adopts in their research to improve the accuracy of object detection.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[8]\", \"Explanation\": \"The cited work by LabelMatch introduces a re-distribution mean teacher model based on the KL divergence distribution between teacher and student models, which the citing paper utilizes in their research to improve the performance of re-distribution in SSL.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[9]\", \"Explanation\": \"The cited work by DSL assigns pixel-wise pseudo-labels for anchor-free detectors, which the citing paper uses as a data source in their research to improve the performance of SSL in long-tailed object detection.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[42]\", \"Explanation\": \"The cited work by Unbiased Teacherv2 introduces a new pseudo-labeling mechanism based on the relative uncertainties of teacher and student models, which the citing paper utilizes as a data source in their research to improve the performance of SSL in long-tailed object detection.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work by LVIS introduces a long-tailed dataset for object detection, which the citing paper extends the research to study the performance of SSL in long-tailed object detection on this dataset.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work, Repeated Factor Sampling (RFS), is a data re-sampling method that the citing paper adopts in their research to address the long-tailed problem in object detection.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[37]\", \"Explanation\": \"The cited work suggests using different sampling schemes in decoupled training stages, which the citing paper utilizes as a data re-sampling method in their research to address the long-tailed problem in object detection.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[70]\", \"Explanation\": \"The cited work is another study that has suggested using different sampling schemes in decoupled training stages, which the citing paper utilizes as a data re-sampling method in their research to address the long-tailed problem in object detection.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[3]\", \"Explanation\": \"The cited work, MixMatch, is used as a basis for the data augmentation techniques employed in the citing paper to improve the performance of detectors trained on long-tailed datasets.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[3]\", \"Explanation\": \"The cited work, MixMatch, is further extended in the citing paper to address the issue of long-tailed data in the training of detectors.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[59]\", \"Explanation\": \"The cited work, FixMatch, is used as a methodological basis for the consistency learning methods employed in the citing paper to improve the performance of detectors trained on long-tailed datasets.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[85]\", \"Explanation\": \"The cited work, MosaicOS, is acknowledged as a study that uses weakly-supervised learning to help long-tailed detection in the context of the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[59]\", \"Explanation\": \"The cited work provides a pretrained model that the citing paper adopts to serve as the teacher model in the training process.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[49,29,66,43]\", \"Explanation\": \"The cited works introduce the use of an exponential moving average of the student model in the training process, which the citing paper builds upon to improve the performance of the model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[1]\", \"Explanation\": \"The cited work highlights the use of strong data augmentations in the training process to inject noise and improve the performance of the model, which the citing paper incorporates into its research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[53]\", \"Explanation\": \"The cited work introduces an uncertainty-based metric to modulate the confidence threshold in the training process, which the citing paper builds upon to address the issue of confirmation bias.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[17,48]\", \"Explanation\": \"The cited works present the co-training framework for training two neural networks in the training process, which the citing paper adopts to mitigate the issue of confirmation bias.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[59]\", \"Explanation\": \"The cited work introduces the use of strong data augmentations in the training process, which the citing paper extends by applying the same strategy to the training of multiple detection heads to prevent overfitting.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[24,28,83,75,31,11,47]\", \"Explanation\": \"The cited works provide evidence of the challenge of applying SSL for long-tailed image classification in the detection task, which motivates the citing paper to address this issue using a cascade mechanism.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[19,21]\", \"Explanation\": \"The cited works on semantic segmentation provide a basis for the citing paper to extend the use of SSL to the detection task, exploring new dimensions and variables in the process.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work is mentioned in the context of long-tailed distribution, which is a common issue in real-world data that has not been explored in existing semi-supervised object detection methods. The cited work is likely a dataset or a study that provides information on the long-tailed distribution in data.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[5]\", \"Explanation\": \"The cited work, Cascade R-CNN, is a popular anchor-based object detection method that the citing paper builds upon to develop the CascadeMatch module for object detection.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[60]\", \"Explanation\": \"The cited work, Sparse R-CNN, is an end-to-end object detection method that the citing paper uses as a base to develop the CascadeMatch module for object detection.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[18]\", \"Explanation\": \"The cited work is a type of neural network architecture that the citing paper uses as the backbone for the CascadeMatch module in object detection.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[44]\", \"Explanation\": \"The cited work is a type of neural network architecture that the citing paper uses as the backbone for the CascadeMatch module in object detection.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[51]\", \"Explanation\": \"The cited work provides the RPN (region proposal network) for generating object proposals, which the citing paper adopts in its model architecture for feature extraction in the first stage of the detector.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[5]\", \"Explanation\": \"The cited work provides the Cascade R-CNN pipeline for the detector, which the citing paper follows in its model architecture to generate object proposals and class probability distributions in the first stage.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[18]\", \"Explanation\": \"The cited work provides the backbone CNN (ResNet50) for feature extraction in the model architecture, which the citing paper uses for the first stage of the detector.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[51]\", \"Explanation\": \"The cited work provides the classification and regression loss functions that the citing paper uses in training the detection heads.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[51]\", \"Explanation\": \"The cited work provides the definition of the two losses used in the CascadeMatch model, i.e., the region-of-interest (ROI) classification loss and the ROI regression loss, which the citing paper adopts in its research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[60]\", \"Explanation\": \"The citing paper extends the application of CascadeMatch to an end-to-end detector by using the Sparse R-CNN model as an example, which builds upon the original work of CascadeMatch.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work, DETR, is the source of the embedding queries used in the region proposal learning process in the citing paper. The method of learning region proposals is adopted from DETR in the cited work.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[39]\", \"Explanation\": \"The cited work, focal loss, is the data source for the replacement of the classification loss in the citing paper. The focal loss is used to replace the classification loss in the experiments conducted in the cited work.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[52]\", \"Explanation\": \"The cited work, L1 and GIoU loss, is the methodological basis for the replacement of the regression loss in the citing paper. The L1 and GIoU loss are used to replace the regression loss in the experiments conducted in the cited work.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"Table 7\", \"Explanation\": \"The cited work, Table 7, is an extension or continuation of the experiments conducted in the citing paper. The experiments are shown to be conducted on both anchor-based and end-to-end detectors, which is a continuation of the research from the cited work.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[10]\", \"Explanation\": \"The cited work by Dave et al. introduces the Fixed AP metric, which the citing paper adopts to evaluate the long-tailed object detection performance in a more accurate and comprehensive manner.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[5]\", \"Explanation\": \"The cited work provides the basis for the choice of the number of heads in the model, which is used in the citing paper to improve the performance of the detection heads.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(LVIS training set)\", \"Explanation\": \"The LVIS training set is used as a data source in the study conducted in the citing paper to train the detection heads and evaluate the performance of the model.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Table 5)\", \"Explanation\": \"The results obtained in Table 5 are an extension of the study conducted in the cited work, showing the accuracy of the pseudo-labels obtained during training for each detection head and their ensemble.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work, LVIS dataset, is used as a data source for the experiments conducted in the citing paper to evaluate the effect of the unsupervised loss weight on the model performance.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[58,43,64]\", \"Explanation\": \"The cited works are used to establish a 'burn-in' stage in the training process, which is adopted by the citing paper to stabilize initialization results in the early stage of training.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[26]\", \"Explanation\": \"The cited work CSD is used as a basis for the comparison in the citing paper, as it is a representative algorithm in the field of consistency regularization-based methods.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[58]\", \"Explanation\": \"The cited work STAC is used to compare with the proposed method, as it is a well-known pseudolabeling-based method that uses an off-line supervised model as a teacher to extract pseudo-labels.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[43]\", \"Explanation\": \"The cited work Unbiased Teacher is used to compare with the proposed method, as it is another pseudo-labeling-based method that uses the exponential moving average (EMA) ensemble to provide a strong teacher model.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[80]\", \"Explanation\": \"The cited work Soft Teacher is used to compare with the proposed method, as it is a strong baseline that uses extra box jittering augmentation to further boost the performance in the field of semi-supervised detection algorithms.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[8]\", \"Explanation\": \"The cited work LabelMatch is used to compare with the proposed method, as it introduces a re-distribution mean teacher based on the KL divergence distribution between teacher and student models, which is a strong baseline in the field of semi-supervised detection algorithms.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[5]\", \"Explanation\": \"The cited work, Cascade R-CNN, is used as the framework in the citing paper to implement the CascadeMatch approach for semi-supervised object detection on the LVIS v1.0 dataset.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[60]\", \"Explanation\": \"The cited work, Sparse R-CNN, is also used as a framework in the citing paper to implement the CascadeMatch approach for semi-supervised object detection on the LVIS v1.0 dataset.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[10]\", \"Explanation\": \"The cited work provides a method for calculating Fixed mAP, which the citing paper adopts in its research to evaluate the performance of the model in terms of mAP.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work, the LVIS dataset, is used as a data source for the unlabeled data in the semi-supervised object detection setting. The dataset is used to provide a large vocabulary of images for the model to work with, which is essential for the study conducted in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[92]\", \"Explanation\": \"The cited work, Federated Loss, serves as a strong baseline in the context of identifying instances with missing labels in the training set, as compared to the method used in the citing paper, CascadeMatch.\"}]"}}},{"rowIdx":75,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b26","b30","b11","b49","b3","b50","b8"],"table_ref":[],"text":"A core challenge in deploying NLP systems lies in managing temporal misalignment, where a model that is trained on data collected in the past is evaluated on data from the present (Lazaridou et al., 2021). Temporal misalignment causes performance degradation in a variety of NLP tasks (Luu et al., 2021;Dhingra et al., 2022;Zhang and Choi, 2021). This is particularly true for knowledge-intensive tasks, such as open-retrieval question answering (QA) (Chen et al., 2017), where models must make predictions based on world knowledge which can rapidly change. Furthermore, such issues are only exacerbated as the paradigm for creating NLP systems continues to shift toward relying on large pretrained models (Zhang et al., 2022;Chowdhery et al., 2022) that are prohibitively expensive to retrain and prone to reciting outdated facts."},{"figure_ref":[],"heading":"Question Asked in t q =2021:","publication_ref":[],"table_ref":[],"text":"What's the tallest building in the world of all time?\nWho sang the American Anthem at the Super Bowl?"},{"figure_ref":[],"heading":"QA System trained in t M =2018:","publication_ref":[],"table_ref":[],"text":"Pink sang the American…"},{"figure_ref":["fig_0"],"heading":"Retrieved Evidence from t E =2018","publication_ref":["b33","b36","b22","b14","b28","b47","b52","b24","b41","b22","b49"],"table_ref":[],"text":"The Burj Khalifa is the tallest building… In the example on the left, the temporal misalignment between when the system was trained and evaluated has no affect on the answer. On the right, the answer has changed, causing the system to output an outdated answer with high confidence. To account for this, we apply our fact duration prediction system to adjust the system's confidence accordingly.\nPrior work has attempted to address these issues by updating the knowledge stored within the parameters of an existing pretrained model (Cao et al., 2021;Mitchell et al., 2022;Onoe et al., 2023). Another line of work has proposed using retrieval-based systems, which utilize a nonparametric corpus of facts that can be updated over time (Karpukhin et al., 2020;Guu et al., 2020;Lewis et al., 2021). Both methods, however, are incomplete solutions as they rely on an oracle to identify which facts need to be updated and to continuously curate a corpus of up-to-date facts.\nGiven the difficulty of keeping existing models up-to-date, we propose an alternative solution where we abstain from presenting facts that we predict are out of date. 1 To accomplish this, we introduce fact duration prediction, the task of predicting how frequently a given fact changes, and establish several classification and regressionbased baselines. We also explore large-scale sources of distant supervision for our task, including fact durations extracted from temporal knowledge bases (Chen et al., 2021b) and durationrelated news text (Yang et al., 2020). We provide rich discussion on this challenging task, exploring the relationship between fact duration prediction and temporal commonsense (Zhou et al., 2020).\nWe provide two sets of evaluations for our fact duration prediction systems. First, as intrinsic evaluations, we report how close our systems' duration estimates are to ground truth labels. We find that models that are trained with only distant supervision can predict the duration of 65% of temporally dependent facts from real search queries in Natu-ralQuestions (Kwiatkowski et al., 2019) to within 3 years, compared to 11% from a simple averageduration baseline. Second, in extrinsic evaluations, we measure how systems' duration estimates can improve an open-retrieval QA system's predictions under temporal misalignment. We mainly focus on improving calibration (as depicted in Figure 1). Our approach can reduce expected calibration error by 50-60% over using system confidence alone on two QA systems (Roberts et al., 2020;Karpukhin et al., 2020) on the SituatedQA dataset (Zhang and Choi, 2021).\nLastly, we also explore other ways of applying our fact duration systems in QA. We experiment with adaptive inference in ensembled open/closedbook QA systems, using duration prediction to decide when retrieval is necessary due to temporal misalignment. We also apply fact duration prediction in a scenario where retrieval is performed over heterogeneous corpus containing both outdated and recent articles, and systems must weigh the relevance of an article against its recency. In summation, we present the first focused study on mitigating temporal misalignment in QA through estimating the duration of facts."},{"figure_ref":["fig_0"],"heading":"Settings","publication_ref":["b30","b49","b24","b27"],"table_ref":["tab_1"],"text":"We aim to address temporal misalignment (Luu et al., 2021) in knowledge-intensive tasks, such as open-retrieval QA. Figure 1 illustrates our setting. We assume a QA model that is developed in the past and is evaluated on a query from a later date. This system suffers from temporal misalignment, return- atedQA). Confidence estimates are taken from calibration models that have been trained for each QA system. All models are trained on 2018 answers from NQ-Open.\nOn the bottom, we compare against an oracle system which zeros the confidence of predictions whose answers have changed between 2018 and 2021.\ning outdated answers for some questions whose answer has changed in the meantime. Table 1 reports the QA performance of existing systems on SituatedQA (Zhang and Choi, 2021), a subset of questions from NQ-Open (Kwiatkowski et al., 2019;Lee et al., 2019) that has been reannotated with the correct answer as of 2021. In this dataset, 48% of questions are updated within the temporal gap (2018 to 2021). We can see that the current models, without considering temporal misalignment, experience performance degradation on both answer accuracy (EM) and calibration. 2In this table, we also explore using an oracle that identifies which answers have changed and zeroes the QA system's confidence in such predictions. While this does not change the system's accuracy, it helps models identify incorrect predictions, improving calibration metrics across the board. In real-world scenarios, however, we do not know which facts are outdated. Thus, in this work we build a fact duration model which predicts facts that are likely outdated and use it to adjust the confidence of the QA model. We introduce our fact duration prediction and QA settings in detail below."},{"figure_ref":[],"heading":"Fact Duration Prediction","publication_ref":[],"table_ref":[],"text":"We define the fact duration prediction task as follows: given a fact f , systems must predict its duration d, the amount of time that the fact remained true for. We consider datasets that represent facts in a variety of formats: QA pairs, statements, knowledge-base relations. For modeling purposes, we convert all facts to statements. For example, the fact f =\"The last Summer Olympic Games were held in Athens.\" has a duration of d = 4 years.\nError Metrics We evaluate fact duration systems by measuring error compared to the gold reference duration: Year MAE is the mean absolute error in their predictions in years and Log-Sec MSE is mean squared error in log-seconds."},{"figure_ref":[],"heading":"QA under Temporal Misalignment","publication_ref":["b21","b44","b0"],"table_ref":[],"text":"The open-retrieval QA task is defined as follows: Given a question q i , a system must produce the corresponding answer a, possibly relying on retrieved knowledge from an evidence corpus E. When taking temporal misalignment into consideration, several critical timestamps can affect performance:\n• Model Training Date (t M ): When the training data for M was collected or annotated. • Evidence Date (t E ): When E was authored. 3• Query Date (t q ): When q was asked.\nFor studying QA under temporal misalignment, we further specify that systems must produce appropriate answer at the time of the query a tq . For example, the question q =\"Where are the next Summer Olympics?\" asked at t q = 2006 has answer a 2006 =\"Beijing\". We define the magnitude of the temporal misalignment (m) to be the amount of time between a a model's training date and the query date (m = t M -t q ). We will compare this with the duration of the fact being asked d = f (q, a tq ). If m > d, we should lower the confidence of the model on this question.\nFor simplicity, we do not take an answer's start date into account. Ideally, determining whether a given QA pair (q, a) has gone out of date should also consider the answer's start date (t s ) and a model's training date (t m ), and confidence can be lowered if t s + d < t m + m. While we expect this approximation to have less of an impact when settings where the misalignment period is small with respect to the distribution of durations, we perform error analysis on examples where considering start date hurts performance in Appendix C.\nCalibration Metrics Even without temporal misalignment, models will not always know the correct answer. Well calibrated model predictions, however, allow us to identify low-confidence predictions and avoid presenting users with incorrect information (Kamath et al., 2020). Under temporal misalignment, calibration further requires identifying which predictions should receive reduced confidence because the answer has likely changed. We consider following calibration metrics:\n• AUROC: Area under the ROC curve evaluates a calibration system's performance at classifying correct and incorrect predictions over all possible confidence thresholds (Tran et al., 2022).\n• Expected Calibration Error (ECE): Computed by ordering predictions by estimated confidence then partitioning into 10 equally sized buckets. ECE is then macro-averaged absolute error each bucket's average confidence and accuracy.\n• Risk Control (RC@XX): Uncertainty estimates are often used for selective-prediction, where models withhold low-confidence predictions below some threshold (< τ ), where τ is set to achieve a target accuracy (XX%) on some evaluation set. We measure how well τ generalizes to a new dataset (Angelopoulos et al., 2022). To compute RC@XX, we set τ based on predictions from t M , then compute the accuracy on predictions from t q with confidence ≥ τ . In the ideal case, the difference (|∆|) between RC@XX and XX should be zero."},{"figure_ref":[],"heading":"Data","publication_ref":[],"table_ref":[],"text":"We first describe the datasets used for evaluation, split by task. We then describe our two large-scale sources for distant supervision. Appendix B contains further prepossessing details and examples."},{"figure_ref":[],"heading":"Evaluation Datasets","publication_ref":["b49","b24","b23","b29","b21","b40","b34","b51","b47"],"table_ref":[],"text":"QA under Misalignment Our primary evaluations are on SituatedQA (Zhang and Choi, 2021), a dataset of questions from NQ-Open (Kwiatkowski et al., 2019) with temporally or geographically dependent answers. We use the temporally-dependent subset, where each question has been annotated with a brief timeline of answers that includes the correct answer as of 2021, the prior answer, and the dates when each answer started to be true. We evaluate misalignment between t M = 2018 and t q = 2021 using the answers from NQ-Open for a 2018 and answers from SituatedQA as a 2021 .\nWhile several recent works have proposed new datasets for studying temporal shifts in QA (Kasai et al., 2022;Livska et al., 2022), these works focus on questions about new events, where answers do not necessarily change (e.g., \"How much was the deal between Elon Musk and Twitter worth?\"). We do not study such shifts in the input distribution over time. We, instead, study methods for managing the shift in the output distribution (i.e., answers changing over time). Adjusting model confidence due to changes in input distribution has been explored (Kamath et al., 2020); however, to the best of our knowledge, this is the first work on calibrating over shifts in output distribution in QA.\nFact Duration Following suit with the QA evaluations above, we also evaluate fact duration prediction on SituatedQA. To generate fact-duration pairs, we use the annotated previous answer as of 2021, converting the question/answer pair into statement using an existing T5-based (Raffel et al., 2020) conversion model (Chen et al., 2021a). We then use distance between the 2021 and previous answer's start date as the fact's duration, d.\nTemporal Commonsense Temporal commonsense focuses on inferences about generic events (e.g., identifying that glaciers move over centuries and a college tours last hours). In contrast, fact duration prediction requires making inferences about specific entities. For instance, determining the duration of an answer to a question like \"Who does Lebron James plays for?\" requires entity knowledge to determine that Lebron James is a basketball player and commonsense knowledge to determine that basketball players often change teams every few years. Previous work (Onoe et al., 2021) has demonstrated the non-trivial nature of combining entity-specific and commonsense knowledge.\nDue to the differences described above, we do not use temporal commonsense datasets for evaluating fact duration prediction. We, however, still evaluate on them to explore how these tasks compare. In particular, we evaluate our fact duration systems on the event duration subset of MCTACO (Zhou et al., 2019). Each MCTACO example consists of a multiple-choice question about the duration of some event in a context sentence, which we convert into duration statements. We evalute using the metrics proposed by the original authors. Following Yang et al. (2020), we select all multiple choice options whose duration falls within a tuned threshold of the predicted duration. EM measures accuracy, evaluating whether the gold and predicted answer sets exactly match. F1 measures the average F1 between the gold and predicted answer sets."},{"figure_ref":[],"heading":"Distant Supervision Sources","publication_ref":["b47"],"table_ref":[],"text":"Temporal Knowledge Bases have been used in numerous prior works for studying how facts change over time. TimeQA (Chen et al., 2021b) is one such work that curates a dataset of 70 different temporally-dependent relations from Wikidata and uses handcrafted templates to convert into decontextualized QA pairs, where the question specifies a time period. To convert this dataset into factduration pairs (f, d), we first convert their QA pairs into a factual statements by removing the date and using a QA-to-statement conversion model (Chen et al., 2021a). We then determine the duration of each facts to be the length of time between the start date of one answer to the question and the next.\nNews Text contains a vast array of facts and rich temporal information. Time-Aware Pretraining dataset (TimePre) (Yang et al., 2020) curates such texts from CNN and Daily Mail news articles using regular expressions to match for durationspecifying phrases (e.g., \"Crystal Palace goalkeeper Julian Speroni has ended the uncertainty over his future by signing a new 12 month contract to stay at Selhurst Park.\"). Pretraining on this dataset has previously been shown to improve performance on temporal commonsense tasks."},{"figure_ref":["fig_1","fig_1"],"heading":"Dataset Summary","publication_ref":["b47","b10","b42"],"table_ref":["tab_9"],"text":"Table 2 reports data statistics and Figure 2 presents the distribution of durations from each dataset. While most facts in SituatedQA and TimeQA change over the course of months to decades, facts in MCTACO and TimePre cover a wider range. Here, we describe our fact duration prediction systems. We include two simple lowerbound baselines: Random samples a duration and Average uses the average duration from each dataset.\nFollowing prior work on temporal common sense reasoning (Yang et al., 2020), we develop BERT-based (Devlin et al., 2018) models. 4 We frame fact duration prediction as cloze questions to more closely match the system's pretraining objective (Schick and Schütze, 2020). To this end, we append \", lasting [MASK][MASK]\" onto each fact, eliciting the model to fill in the masked tokens with a duration. We use two mask tokens as typically duration information requires at least two tokens, one for value and another for unit. For our TimePre and MCTACO datasets, we similarly replace the target durations with two mask tokens. Table 8 in Appendix B contains examples. Predictions are made by averaging the encoded representations of the two \"[MASK]\" tokens, then using this representation as an input to a single hidden layer network. Using this same representation, we train two models with regression-based and classification-based learning objectives described below.\nClassification Model frames the task as a 13-way classification task where each class corresponds to a duration in Figure 2. 5 We train using cross entropy loss, selecting the closest duration class as the pseudo-gold label for each fact. Because this model can only predict a limited set of durations, we report its upperbound by always selecting the class closest to the gold duration.\nRegression Model uses the mean squared error loss in units of log seconds, where the output from the hidden layer predicts a scalar value."},{"figure_ref":["fig_2","fig_2","fig_2","fig_1"],"heading":"Results","publication_ref":[],"table_ref":[],"text":"We experiment with training on TimeQA and TimePre individually and on the union of both datasets. Figure 3 reports duration prediction and temporal commonsense performance. Overall, we find that our trained systems outperform simple random and average baselines on SituatedQA. This is indicative of strong generalizability from our distantly-supervised fact-duration systems, even when baselines benefit from access to the gold label distribution. We also provide a histogram of errors from our systems in Figure 3 where we can see that over 60% of our classification-based system's predictions are within 3 years of the gold duration, while predicting the exact duration remains challenging. Below, we reflect on the impact of our different modeling choices and research questions.\nRegression vs. Classification Regression-based models tend to outperform their classificationbased counterparts. The instances where this is not true can be attributed to an insufficient amount of training data. In Figure 3, we can see the different types of errors each model makes. The classification system predicts duration within 1 year more frequently, but the regression system predicts duration within 4 years more frequently.\nSupervision from KB vs. News Text We find that training on temporal knowledge-base relations (TimeQA) alone vastly outperforms training on news text (TimePre) alone for fact-duration prediction; however, the opposite is true when comparing performance on temporal commonsense (MC-TACO). Training on both datasets tends to improve our regression-based system, but yields mixed results for our classification-based system. We hypothesize that the closeness in label distribution (see Figure 2) between the training and evaluation sets impacts the performance significantly."},{"figure_ref":[],"heading":"Fact Duration vs. Temporal Commonsense","publication_ref":[],"table_ref":[],"text":"While fact duration prediction and temporal commonsense are conceptually related, we find that strong performance on either task does not necessarily transfer to the other. As discussed above, this can be attributed to differences in label distributions; however, label distribution also serves as a proxy variable for the type of information being queried for in either task. Commonsense knowledge primarily differentiates events that take place over different orders of magnitude of time (e.g., seconds versus years). Differentiating whether an event takes place over finer-grained ranges (e.g., one versus two years), however, cannot be resolved with commonsense knowledge alone, and further require fact retrieval. We find that NQ contains queries for facts that change over a smaller range of durations (between 1-10 years), and, therefore, commonsense knowledge alone is insufficient."},{"figure_ref":[],"heading":"Calibrating QA under Temporal Misalignment","publication_ref":[],"table_ref":[],"text":"Here, we return our motivating use-case of using fact duration prediction to calibrate openretrieval QA systems under temporal misalignment.\nWe assume an access to base calibration system, c(q, a) ∈ [0, 1] that has the same training date as the QA system its calibrating. We then use fact duration to generate misalignment-aware confidence score c m through simple post-hoc augmentation, scaling down the original confidence score by a discount factor based on the degree of misalignment and the predicted fact duration. We compute this factor differently for each of our fact duration systems.\n• Classification: Here, the system's output is a probability distribution over different duration classes, p(d|q, a). We set the discount factor to be the CDF of this distribution evaluated at m:6 c m = c(q, a) d≥m P (d|q, a). • Regression: Here, the output is a single predicted duration d. We set the discount factor to the binary value indicating whether or not the misalignment period has exceeded the predicted duration: c m = c(q, a)1{d > m} As classification systems predict a distribution over fact durations, we are able to use the CDF of this distribution to make granular adjustments to confidence over time. In contrast, our regression systems predict a single value for fact duration, and confidence adjustments over time are abrupt, leaving confidence unchanged or setting it to zero."},{"figure_ref":[],"heading":"Models","publication_ref":["b41","b22","b49"],"table_ref":[],"text":"Base QA and Calibration Systems We experiment with three QA systems throughout our study: • 1 ⃝ T5: We use T5-large (Roberts et al., 2020) which has been trained with salient-spanmasking and closed-book QA on NQ-Open.\n• 2 ⃝ DPR (t e =2018): We use DPR (Karpukhin et al., 2020), an open-book system which retrieves passages from a t e = 2018 Wikipedia snapshot and is also trained on NQ-Open.\n• N ⃝ DPR (t e =2021): We use the same model as 2 ⃝, but swap the retrieval corpus with an updated Wikipedia snapshot that matches query times- We first adjust confidence Per-Example, which is our full system. We then adjust confidence Uniformly across all examples, such that the net decrease in confidence across the entire test set is equivalent.\ntamp (t e = 2021) following Zhang and Choi (2021), which showed partial success in returning up-to-date answers. For each QA system, we train a calibrator that predicts the correctness of the QA system's answer. We follow Zhang et al. (2021) for the design and input features to calibrator, using the model's predicted likelihood and encoded representations of the input (details in Appendix A).\nFact Duration Systems For both our regression and classification based models, we use the systems trained over both with TimeQA. We also include results using an oracle fact duration system, which zeroes the confidence for all questions that have been updated since the training date."},{"figure_ref":["fig_2"],"heading":"Results","publication_ref":[],"table_ref":["tab_5","tab_6"],"text":"Table 3 reports the results from our calibration experiments. Both QA models suffer from temporal degradation, and zero-ing out the confidence of outdated facts with oracle information improves the calibration performance. Using model prediction durations shows similar gains. Both re-gression and classification duration predictions lower the confidence of models, improving calibration metrics across the board. We find that our classification-based model consistently outperforms our regression-based model on our calibration task, despite the opposite being true for our fact-duration evaluations. We attribute this behavior to our classification-based system's error distribution, as it gets more examples correct to within 1 year (Figure 3). Classification-based system also hedge over different duration classes by predicting a distribution, which we use to compute the CDF.\nRetrieval-Based QA: Update or Adjust In Table 3, we compare the performance of DPR with static and hot-swapped retrieval corpora from t e = 2018 and t e = 2021. While updating the retrieval corpus improves EM accuracy, adjusting model confidence using fact duration on a static corpus performs better on all calibration metrics. This suggests that, when users care about having accurate confidence estimates or seeing only highconfidence predictions, confidence adjustments can be more beneficial than swapping retrieval corpus.\nAblations: Per-Example vs Uniform Adjustment We compare our system, which adjusts confidence on a per-example basis, against one that uniformly decreases confidence by the same value υ across the entire test set: c m = max(c(q, a)-υ, 0). These ablations still depend on our fact-duration systems to determine the υ such that the total confidence over the entire test set is the same for both methods. Table 4 reports the results from this ablation study. We find that uniformly adjusting confidence improves ECE, which is expected given the decrease in the QA systems EM accuracy after misalignment. We find, however, that our perexample adjustment methods outperform uniform confidence adjustments."},{"figure_ref":[],"heading":"Comparisons against Prompted LLMs","publication_ref":["b43","b20","b9"],"table_ref":[],"text":"While we primarily focus on calibrating fine-tuned QA models, recent work has also explored calibrating prompted large language models (LLMs) for QA (Si et al., 2022;Kadavath et al., 2022;Cole et al., 2023). Furthermore, recent general-purpose LLMs (e.g., ChatGPT) have demonstrated the ability to abstain from answering questions on the basis that their knowledge cutoff date is too far in the past; however, it is not publicly known how these systems exhibit this behavior.\nIn this experiment, we investigate one such system, GPT-4 (OpenAI, 2023), and its ability to abstain from answering questions with rapidly changing answers. We prompt GPT-4 to answer questions from and find that it abstains from answering 86% of all questions (95% of questions whose answers have been updated between 2018 and 2021 and on 79% of examples that have not). Overall, this behavior suggests that GPT-4 overestimates how frequently it must to abstain from answering user queries. Furthermore, GPT-4 does not provide how frequently the answer is expected to change. Collectively, GPT-4's tendency to over-abstain and its lack of transparency limits its usefulness to users. In contrast, our approach provides users with an duration estimate indicating why a prediction may not be trustworthy. Further experimental details and example outputs are reported in Appendix D."},{"figure_ref":[],"heading":"Beyond Calibration: Adaptive Inference","publication_ref":["b31"],"table_ref":["tab_7","tab_7"],"text":"In this section, we explore using our misalignmentaware confidence scores to decide how to answer a question. Below, we motivate and describe two adaptive inference scenarios where systems may choose between two methods for answering a question using their fact duration predictions. Hybrid: Closed + Open ( 1 ⃝+ N ⃝): Besides the computational benefits from not always having to use retrieval, forgoing retrieval for popular questions can also improve answer accuracy (Mallen et al., 2022). We use our fact duration predictions to decide when retrieval is necessary: we first predict an answer using T5 and run our fact duration prediction system using this answer. We then use the CDF of the predicted duration distribution to determine whether it is at least 50% likely that the fact has changed: d≤m P (d|q, a) ≥ 0.5. If so, we then run retrieval with DPR using the updated corpus t e = 2021 and present the predicted answer. We report our results in the first row of Table 5, which shows that this outperforms either system on its own, while running retrieval on less than half of all examples. Two Corpora: Relevancy vs. Recency ( 2 ⃝+ N ⃝): While most work for QA have focused on retrieving over Wikipedia, many questions require retrieving over other sources such as news or web text. One challenge in moving away from Wikipedia lies in managing temporal heterogeneity across different articles. Unlike Wikipedia, news and web articles are generally not maintained to stay current, requiring retrieval-based QA systems to identify out-of-date information in articles. Systems that retrieve such resources must consider the trade-off between the recency versus relevancy of an article. In these experiments, we experiment with using fact duration prediction as a method for weighing this trade-off in retrieval.\nOur experimental setup is as follows: instead of computing misalignment from the model's training date, we compute relative to when the article was authored (m = 3 years for 2018 Wikipedia and m = 0 years for 2021 Wikipedia). After performing inference using both corpora, we re-rank answers according to their misalignment-adjusted confidence estimates. We report results in Table 5. We find that our method is able to recover comparable performance to always using up-to-date articles, while using it just under half the time."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b53","b39","b13","b1","b12","b46","b4","b21","b49","b45","b11","b33","b32","b35","b17","b25"],"table_ref":[],"text":"Commonsense and Temporal Reasoning Recent works have proposed forecasting bench-marks (Zou et al., 2022;Jin et al., 2021a) related to our fact duration prediction task. While our task asks models to predict when a fact will change, these forecasting tasks ask how a fact will change. Qin et al. (2021) studies temporal commonsense reasoning in dialogues. Quantitative reasoning has been explored in other works as quantitative relations between nouns (Forbes and Choi, 2017;Bagherinezhad et al., 2016), distributions over quantitative attributes Elazar et al. (2019), and representing numbers in language models (Wallace et al., 2019).\nCalibration Abstaining from providing a QA system's answers has been explored in several recent works. Chen et al. (2022) examines instances where knowledge conflicts exist between a model's memorized knowledge and retrieved documents. As the authors note, such instances often arise due to temporal misalignment. Prior work (Kamath et al., 2020;Zhang et al., 2021;Varshney and Baral, 2023) has explored abstaining from answering questions by predicting whether or not the test question comes from the same training distribution of the QA system. While fact duration also predicts a shift in distribution, fact duration focuses on predicting a shift in a question's output distribution of answers instead of a shift in input distribution of questions; therefore, these two systems are addressing orthogonal challenges in robustness to distribution shift and are complementary.\nKeeping Systems Up-to-Date Several works have explored continuing pretraining to address temporal misalignment in pretrained models (Dhingra et al., 2022;Jin et al., 2021b). Other works have explored editing specific facts into models (Cao et al., 2021;Mitchell et al., 2022;Meng et al., 2022). These works, however, have only focused on synthetic settings and assume access to the updated facts. Furthermore, such systems have yet to be successfully applied to new benchmarks for measuring whether language models have acquired emergent information (Onoe et al., 2022;Padmanabhan et al., 2023). Recent works on retrievalbased QA systems have found improved adaptation when updated with up-to-date retrieval corpora (Izacard et al., 2022;Lazaridou et al., 2022)."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"We improve QA calibration under temporal misalignment by introducing the fact duration pre-diction task, alongside several datasets and baseline systems for it. Future work may build upon this evaluation framework to further improve QA calibration under temporal misalignment. For instance, future work may examine modeling different classes distributions of fact duration distributions, like modeling whether a fact changes after a regular, periodic time interval."},{"figure_ref":[],"heading":"Limitations","publication_ref":[],"table_ref":["tab_5"],"text":"We only evaluate temporal misalignment between 2018 and 2021, a three-year time difference, on a relatively small scale SituatedQA dataset (N=322). This is mainly due to a lack of benchmark that supports studying temporal misalignment. Exploring this in more diverse setup, including different languages, text domains, wide range of temporal gaps, would be fruitful direction for future work.\nAs is the case with all systems that attempt to faithfully relay world knowledge, treating model predictions as fact runs the risk of propagating misinformation. While the goal of our fact duration prediction systems is to prevent models from reciting outdated facts, it does not always succeed and facts may change earlier than expected. Even though a given fact may be expected to only change once every decade, an improbable outcome may occur and the fact changes after only a year. In such an event, our misalignment-aware calibration system may erroneously maintain high confidence in the outdated answer. Furthermore, our system, as it stands, does not take the answer start date into account. Our system also can make errors due to changes in the typical duration of a given fact. For instance \"What's the world's tallest building?\" changes more frequently over time as the rate of technological advances also increases. We provide examples of such system errors in Appendix C.\nTable 7: Additional calibration results on SituatedQA, comparing different methods of adjusting model confidence for temporal misalignment using the output of our classification-based fact-duration system. All other settings are the same as in Table 3.\nsimply assume that all such question answer pairs instances have a duration of 1 month."},{"figure_ref":[],"heading":"B.2 Temporal Commonsense Datasets Preprocessing","publication_ref":["b47"],"table_ref":[],"text":"As we noted above, each MCTACO example consists of a multiple-choice question about the duration of some event in a provided context sentence. During prepossessing, we use the same question conversion model as above to transform each QA pair into a statement and prepend the context sentence onto each question. We use the metrics proposed by the original authors, and we select all multiple choice options whose duration falls within some absolute threshold of predicted duration, measured in log seconds (Yang et al., 2020). This threshold is selected based on development set performance."},{"figure_ref":[],"heading":"C Additional Results","publication_ref":["b16"],"table_ref":[],"text":"Different Pretrained Models for Fact Duration Prediction In Table 6, we report our results from DeBERTa-v3-base (He et al., 2021) on our fact duration prediction system. We also experiment with using the large variants of both BERT and DeBERTa, but do not find substantial improvement."},{"figure_ref":[],"heading":"Adjusting Confidence with Expected Duration","publication_ref":[],"table_ref":[],"text":"In addition adjusting confidence using the CDF of the predicted duration distribution from our classification-based system, we also experiment with using the expected duration as our discounting factor. We incorporate this by zeroing the confidence estimate if the expected duration is exceeded by the degree of misalignment: f (q, a) = 1{ d d • P (d|q, a) > m}.\nIn Table 7, we report additional results on our calibration evaluation. We include calibration performance of our best performing fact duration models finetuned on SituatedQA: trained only on Sit-uatedQA for our classification-based model and first trained on TimeQA + TA Pretrain for our regression-based model.\nError Analysis In Table 9, we highlight sampled errors from our fact duration system and discuss their causes and impact."},{"figure_ref":[],"heading":"D ChatGPT and GPT-4 outputs","publication_ref":[],"table_ref":["tab_11"],"text":"Table 10 includes two examples of ChatGPT informing users that the answers to a given question may have changed. It, however, does not provide users with an estimate of how likely it has changed, or how often the answer is expected to change. This lack of a duration estimate results in lesser transparency and interpretatbility for users. To get results on SituatedQA, we prompt GPT-4 with the following system prompt (recommended by their documentation): \"You are ChatGPT, a large language model trained by OpenAI. Answer as con- 9: Example fact duration predictions errors from SituatedQA. We include the answers from t M = 2018 and t q = 2021, and the gold / predicted durations ({G / P} Dur). In the first example, we see that our system correctly identified volcanic eruptions as an infrequent event; however, our system misses that the answer has changed within the misalignment period because it doesn't have access to the answer start date. In the second example, our system fails no longer changes since the show ended. In the last example, our model underestimates the fact's duration, but still correctly identifies that the answer has changed. cisely as possible. Knowledge cutoff: September 2021. Current date: May 12, 2023.\" We then present GPT-4 with the user's question from Situat-edQA. We determine whether a systems abstains from a given prediction if it references its knowledge cutoff from the prompt September 2021 or if it mentions real-time information in its generated output."},{"figure_ref":[],"heading":"E Licensing","publication_ref":[],"table_ref":[],"text":"Dataset Licensing TimeQA is licensed under the BSD 3-Clause License. SituatedQA and Wikipedia are both distributed under the CC BY-SA 4.0 license. MCTACO and TA-Pretrain are both unlicensed."},{"figure_ref":[],"heading":"Model Licensing","publication_ref":[],"table_ref":[],"text":"The transformers library, BERT model, T5 model, and question conversion model are all licensed under the Apache License v2.0. DPR is licensed under the Attribution-NonCommercial 4.0 International license. Our use User: Who is the current US president? ChatGPT: As an AI language model, I don't have realtime information or the ability to browse the internet. My knowledge was last updated in September 2021, and at that time, the President of the United States was Joseph R. Biden Jr. However, please note that the political landscape can change, and there may have been a new election or other developments since then. To obtain the most accurate and up-to-date information, I recommend checking reliable sources such as official government websites or reputable news outlets.\nUser: Who won the last NBA Champtionship? ChatGPT: As an AI language model, I don't have realtime data, and my knowledge was last updated in September 2021. Therefore, I cannot provide information on the winner of the most recent NBA Championship. However, as of my last update, the Milwaukee Bucks were the reigning NBA Champions, having won the championship in the 2020-2021 season. I recommend checking reliable news sources or conducting a quick online search to find the current champion. of all these resources are in compliance with their licenses."},{"figure_ref":[],"heading":"F Ethical Considerations","publication_ref":[],"table_ref":[],"text":"We do not collect any data in this paper. While we do generate a dataset, we visually inspect generated examples and do not find any cases of harmful or offensive content. The existing datasets used in this work have previously been vetted by their authors prior work. We also note that our work is only applied to English datasets and models."},{"figure_ref":[],"heading":"Acknowledgements","publication_ref":[],"table_ref":[],"text":"The work was partially supported by Google Research Award, a gift from HomeDepot, and a grant from UT Machine Learning Lab. This work was in part supported by Cisco Research. Any opinions, findings and conclusions, or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of Cisco Research. The authors would like to thank the members of the UT Austin NLP community and Jordan Boyd-Graber for helpful discussions."},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"Our data and code are released publicly at https://github.com/ mikejqzhang"},{"figure_ref":[],"heading":"A Implementation Details","publication_ref":["b49"],"table_ref":[],"text":"A.1 QA Models\nWe use the T5 7 and DPR 8 checkpoints that have been finetuned on NQ-Open's training set from the transformers library model hub. 9 For DPR, we use the retrieval corpora from December 20, 2018 for t E = 2018 and February 20, 2021 for t E = 2021, following Zhang and Choi (2021).\n7 https://huggingface.co/google/t5-large-ssm-nqo 8 https://huggingface.co/facebook/dpr-reader-single-nqbase 9 https://huggingface.co/models"},{"figure_ref":[],"heading":"A.2 Calibration Models","publication_ref":["b6","b49","b49"],"table_ref":[],"text":"We implement our trained calibration systems using XGBoost (Chen and Guestrin, 2016) using the features for T5 and DPR outlined in Zhang et al. (2021). For T5, we concatenate (1) the averaged, encoded representations of the input question, and\n(2) the model likelihood. For DPR, we concatenate (1) the averaged, encoded representations of the input question and selected passage, (2) the averaged, encoded representations of the start and end tokens of the selected answer span, and (3) the likelihood of the answer span, computed as the product of the likelihoods of selecting the start index, end index, and passage index. We train our systems on NQ-Open on a randomly sampled 60/40 training and development splits, following Zhang et al. (2021). We use a maximum depth of 10 and experiment with several values for the learning rate {0.01, 0.1, 0.2, 0.5} and column sub-sampling ratio {0.0, 0.1, . . . , 0.9}, which we keep the for sampling by tree, level, and node. We train with early stopping after 10 epochs without improvement and select the best performing system as evaluated on the development split."},{"figure_ref":[],"heading":"A.3 Fact Duration Prediction Models","publication_ref":[],"table_ref":[],"text":"We use BERT-base from the transformers library for all duration prediction baselines, trained with a batch size in {32, 64} and learning rate in {1e -5, 5e -5}. We train until convergence and select the best checkpoint as determined by development set performance. Due to computational resource constraints, we do not further tune hyperparameters. All models are trained once and results reflect a single run. All experiments took were performed Quadro RTX 8000 gpus and required less than one week's worth of GPU hours."},{"figure_ref":[],"heading":"B Datasets","publication_ref":[],"table_ref":[],"text":"We provide examples from each dataset and our prepossessing pipeline in Table 8. We provide futher preprocessing details below."},{"figure_ref":[],"heading":"B.1 Fact Duration Dataset Preprocessing","publication_ref":[],"table_ref":[],"text":"In TimeQA, several examples have answers that are simply the empty string. We remove all such examples from our preprocessed dataset. In SituatedQA, several examples have answers that begin and end in the same year, without further annotation determining the exact number of days or months. We "}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b26\",\n \"b30\",\n \"b11\",\n \"b49\",\n \"b3\",\n \"b50\",\n \"b8\"\n ],\n \"table_ref\": [],\n \"text\": \"A core challenge in deploying NLP systems lies in managing temporal misalignment, where a model that is trained on data collected in the past is evaluated on data from the present (Lazaridou et al., 2021). Temporal misalignment causes performance degradation in a variety of NLP tasks (Luu et al., 2021;Dhingra et al., 2022;Zhang and Choi, 2021). This is particularly true for knowledge-intensive tasks, such as open-retrieval question answering (QA) (Chen et al., 2017), where models must make predictions based on world knowledge which can rapidly change. Furthermore, such issues are only exacerbated as the paradigm for creating NLP systems continues to shift toward relying on large pretrained models (Zhang et al., 2022;Chowdhery et al., 2022) that are prohibitively expensive to retrain and prone to reciting outdated facts.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Question Asked in t q =2021:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"What's the tallest building in the world of all time?\\nWho sang the American Anthem at the Super Bowl?\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"QA System trained in t M =2018:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Pink sang the American…\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Retrieved Evidence from t E =2018\",\n \"publication_ref\": [\n \"b33\",\n \"b36\",\n \"b22\",\n \"b14\",\n \"b28\",\n \"b47\",\n \"b52\",\n \"b24\",\n \"b41\",\n \"b22\",\n \"b49\"\n ],\n \"table_ref\": [],\n \"text\": \"The Burj Khalifa is the tallest building… In the example on the left, the temporal misalignment between when the system was trained and evaluated has no affect on the answer. On the right, the answer has changed, causing the system to output an outdated answer with high confidence. To account for this, we apply our fact duration prediction system to adjust the system's confidence accordingly.\\nPrior work has attempted to address these issues by updating the knowledge stored within the parameters of an existing pretrained model (Cao et al., 2021;Mitchell et al., 2022;Onoe et al., 2023). Another line of work has proposed using retrieval-based systems, which utilize a nonparametric corpus of facts that can be updated over time (Karpukhin et al., 2020;Guu et al., 2020;Lewis et al., 2021). Both methods, however, are incomplete solutions as they rely on an oracle to identify which facts need to be updated and to continuously curate a corpus of up-to-date facts.\\nGiven the difficulty of keeping existing models up-to-date, we propose an alternative solution where we abstain from presenting facts that we predict are out of date. 1 To accomplish this, we introduce fact duration prediction, the task of predicting how frequently a given fact changes, and establish several classification and regressionbased baselines. We also explore large-scale sources of distant supervision for our task, including fact durations extracted from temporal knowledge bases (Chen et al., 2021b) and durationrelated news text (Yang et al., 2020). We provide rich discussion on this challenging task, exploring the relationship between fact duration prediction and temporal commonsense (Zhou et al., 2020).\\nWe provide two sets of evaluations for our fact duration prediction systems. First, as intrinsic evaluations, we report how close our systems' duration estimates are to ground truth labels. We find that models that are trained with only distant supervision can predict the duration of 65% of temporally dependent facts from real search queries in Natu-ralQuestions (Kwiatkowski et al., 2019) to within 3 years, compared to 11% from a simple averageduration baseline. Second, in extrinsic evaluations, we measure how systems' duration estimates can improve an open-retrieval QA system's predictions under temporal misalignment. We mainly focus on improving calibration (as depicted in Figure 1). Our approach can reduce expected calibration error by 50-60% over using system confidence alone on two QA systems (Roberts et al., 2020;Karpukhin et al., 2020) on the SituatedQA dataset (Zhang and Choi, 2021).\\nLastly, we also explore other ways of applying our fact duration systems in QA. We experiment with adaptive inference in ensembled open/closedbook QA systems, using duration prediction to decide when retrieval is necessary due to temporal misalignment. We also apply fact duration prediction in a scenario where retrieval is performed over heterogeneous corpus containing both outdated and recent articles, and systems must weigh the relevance of an article against its recency. In summation, we present the first focused study on mitigating temporal misalignment in QA through estimating the duration of facts.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Settings\",\n \"publication_ref\": [\n \"b30\",\n \"b49\",\n \"b24\",\n \"b27\"\n ],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"We aim to address temporal misalignment (Luu et al., 2021) in knowledge-intensive tasks, such as open-retrieval QA. Figure 1 illustrates our setting. We assume a QA model that is developed in the past and is evaluated on a query from a later date. This system suffers from temporal misalignment, return- atedQA). Confidence estimates are taken from calibration models that have been trained for each QA system. All models are trained on 2018 answers from NQ-Open.\\nOn the bottom, we compare against an oracle system which zeros the confidence of predictions whose answers have changed between 2018 and 2021.\\ning outdated answers for some questions whose answer has changed in the meantime. Table 1 reports the QA performance of existing systems on SituatedQA (Zhang and Choi, 2021), a subset of questions from NQ-Open (Kwiatkowski et al., 2019;Lee et al., 2019) that has been reannotated with the correct answer as of 2021. In this dataset, 48% of questions are updated within the temporal gap (2018 to 2021). We can see that the current models, without considering temporal misalignment, experience performance degradation on both answer accuracy (EM) and calibration. 2In this table, we also explore using an oracle that identifies which answers have changed and zeroes the QA system's confidence in such predictions. While this does not change the system's accuracy, it helps models identify incorrect predictions, improving calibration metrics across the board. In real-world scenarios, however, we do not know which facts are outdated. Thus, in this work we build a fact duration model which predicts facts that are likely outdated and use it to adjust the confidence of the QA model. We introduce our fact duration prediction and QA settings in detail below.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Fact Duration Prediction\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We define the fact duration prediction task as follows: given a fact f , systems must predict its duration d, the amount of time that the fact remained true for. We consider datasets that represent facts in a variety of formats: QA pairs, statements, knowledge-base relations. For modeling purposes, we convert all facts to statements. For example, the fact f =\\\"The last Summer Olympic Games were held in Athens.\\\" has a duration of d = 4 years.\\nError Metrics We evaluate fact duration systems by measuring error compared to the gold reference duration: Year MAE is the mean absolute error in their predictions in years and Log-Sec MSE is mean squared error in log-seconds.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"QA under Temporal Misalignment\",\n \"publication_ref\": [\n \"b21\",\n \"b44\",\n \"b0\"\n ],\n \"table_ref\": [],\n \"text\": \"The open-retrieval QA task is defined as follows: Given a question q i , a system must produce the corresponding answer a, possibly relying on retrieved knowledge from an evidence corpus E. When taking temporal misalignment into consideration, several critical timestamps can affect performance:\\n• Model Training Date (t M ): When the training data for M was collected or annotated. • Evidence Date (t E ): When E was authored. 3• Query Date (t q ): When q was asked.\\nFor studying QA under temporal misalignment, we further specify that systems must produce appropriate answer at the time of the query a tq . For example, the question q =\\\"Where are the next Summer Olympics?\\\" asked at t q = 2006 has answer a 2006 =\\\"Beijing\\\". We define the magnitude of the temporal misalignment (m) to be the amount of time between a a model's training date and the query date (m = t M -t q ). We will compare this with the duration of the fact being asked d = f (q, a tq ). If m > d, we should lower the confidence of the model on this question.\\nFor simplicity, we do not take an answer's start date into account. Ideally, determining whether a given QA pair (q, a) has gone out of date should also consider the answer's start date (t s ) and a model's training date (t m ), and confidence can be lowered if t s + d < t m + m. While we expect this approximation to have less of an impact when settings where the misalignment period is small with respect to the distribution of durations, we perform error analysis on examples where considering start date hurts performance in Appendix C.\\nCalibration Metrics Even without temporal misalignment, models will not always know the correct answer. Well calibrated model predictions, however, allow us to identify low-confidence predictions and avoid presenting users with incorrect information (Kamath et al., 2020). Under temporal misalignment, calibration further requires identifying which predictions should receive reduced confidence because the answer has likely changed. We consider following calibration metrics:\\n• AUROC: Area under the ROC curve evaluates a calibration system's performance at classifying correct and incorrect predictions over all possible confidence thresholds (Tran et al., 2022).\\n• Expected Calibration Error (ECE): Computed by ordering predictions by estimated confidence then partitioning into 10 equally sized buckets. ECE is then macro-averaged absolute error each bucket's average confidence and accuracy.\\n• Risk Control (RC@XX): Uncertainty estimates are often used for selective-prediction, where models withhold low-confidence predictions below some threshold (< τ ), where τ is set to achieve a target accuracy (XX%) on some evaluation set. We measure how well τ generalizes to a new dataset (Angelopoulos et al., 2022). To compute RC@XX, we set τ based on predictions from t M , then compute the accuracy on predictions from t q with confidence ≥ τ . In the ideal case, the difference (|∆|) between RC@XX and XX should be zero.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Data\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We first describe the datasets used for evaluation, split by task. We then describe our two large-scale sources for distant supervision. Appendix B contains further prepossessing details and examples.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Evaluation Datasets\",\n \"publication_ref\": [\n \"b49\",\n \"b24\",\n \"b23\",\n \"b29\",\n \"b21\",\n \"b40\",\n \"b34\",\n \"b51\",\n \"b47\"\n ],\n \"table_ref\": [],\n \"text\": \"QA under Misalignment Our primary evaluations are on SituatedQA (Zhang and Choi, 2021), a dataset of questions from NQ-Open (Kwiatkowski et al., 2019) with temporally or geographically dependent answers. We use the temporally-dependent subset, where each question has been annotated with a brief timeline of answers that includes the correct answer as of 2021, the prior answer, and the dates when each answer started to be true. We evaluate misalignment between t M = 2018 and t q = 2021 using the answers from NQ-Open for a 2018 and answers from SituatedQA as a 2021 .\\nWhile several recent works have proposed new datasets for studying temporal shifts in QA (Kasai et al., 2022;Livska et al., 2022), these works focus on questions about new events, where answers do not necessarily change (e.g., \\\"How much was the deal between Elon Musk and Twitter worth?\\\"). We do not study such shifts in the input distribution over time. We, instead, study methods for managing the shift in the output distribution (i.e., answers changing over time). Adjusting model confidence due to changes in input distribution has been explored (Kamath et al., 2020); however, to the best of our knowledge, this is the first work on calibrating over shifts in output distribution in QA.\\nFact Duration Following suit with the QA evaluations above, we also evaluate fact duration prediction on SituatedQA. To generate fact-duration pairs, we use the annotated previous answer as of 2021, converting the question/answer pair into statement using an existing T5-based (Raffel et al., 2020) conversion model (Chen et al., 2021a). We then use distance between the 2021 and previous answer's start date as the fact's duration, d.\\nTemporal Commonsense Temporal commonsense focuses on inferences about generic events (e.g., identifying that glaciers move over centuries and a college tours last hours). In contrast, fact duration prediction requires making inferences about specific entities. For instance, determining the duration of an answer to a question like \\\"Who does Lebron James plays for?\\\" requires entity knowledge to determine that Lebron James is a basketball player and commonsense knowledge to determine that basketball players often change teams every few years. Previous work (Onoe et al., 2021) has demonstrated the non-trivial nature of combining entity-specific and commonsense knowledge.\\nDue to the differences described above, we do not use temporal commonsense datasets for evaluating fact duration prediction. We, however, still evaluate on them to explore how these tasks compare. In particular, we evaluate our fact duration systems on the event duration subset of MCTACO (Zhou et al., 2019). Each MCTACO example consists of a multiple-choice question about the duration of some event in a context sentence, which we convert into duration statements. We evalute using the metrics proposed by the original authors. Following Yang et al. (2020), we select all multiple choice options whose duration falls within a tuned threshold of the predicted duration. EM measures accuracy, evaluating whether the gold and predicted answer sets exactly match. F1 measures the average F1 between the gold and predicted answer sets.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Distant Supervision Sources\",\n \"publication_ref\": [\n \"b47\"\n ],\n \"table_ref\": [],\n \"text\": \"Temporal Knowledge Bases have been used in numerous prior works for studying how facts change over time. TimeQA (Chen et al., 2021b) is one such work that curates a dataset of 70 different temporally-dependent relations from Wikidata and uses handcrafted templates to convert into decontextualized QA pairs, where the question specifies a time period. To convert this dataset into factduration pairs (f, d), we first convert their QA pairs into a factual statements by removing the date and using a QA-to-statement conversion model (Chen et al., 2021a). We then determine the duration of each facts to be the length of time between the start date of one answer to the question and the next.\\nNews Text contains a vast array of facts and rich temporal information. Time-Aware Pretraining dataset (TimePre) (Yang et al., 2020) curates such texts from CNN and Daily Mail news articles using regular expressions to match for durationspecifying phrases (e.g., \\\"Crystal Palace goalkeeper Julian Speroni has ended the uncertainty over his future by signing a new 12 month contract to stay at Selhurst Park.\\\"). Pretraining on this dataset has previously been shown to improve performance on temporal commonsense tasks.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_1\"\n ],\n \"heading\": \"Dataset Summary\",\n \"publication_ref\": [\n \"b47\",\n \"b10\",\n \"b42\"\n ],\n \"table_ref\": [\n \"tab_9\"\n ],\n \"text\": \"Table 2 reports data statistics and Figure 2 presents the distribution of durations from each dataset. While most facts in SituatedQA and TimeQA change over the course of months to decades, facts in MCTACO and TimePre cover a wider range. Here, we describe our fact duration prediction systems. We include two simple lowerbound baselines: Random samples a duration and Average uses the average duration from each dataset.\\nFollowing prior work on temporal common sense reasoning (Yang et al., 2020), we develop BERT-based (Devlin et al., 2018) models. 4 We frame fact duration prediction as cloze questions to more closely match the system's pretraining objective (Schick and Schütze, 2020). To this end, we append \\\", lasting [MASK][MASK]\\\" onto each fact, eliciting the model to fill in the masked tokens with a duration. We use two mask tokens as typically duration information requires at least two tokens, one for value and another for unit. For our TimePre and MCTACO datasets, we similarly replace the target durations with two mask tokens. Table 8 in Appendix B contains examples. Predictions are made by averaging the encoded representations of the two \\\"[MASK]\\\" tokens, then using this representation as an input to a single hidden layer network. Using this same representation, we train two models with regression-based and classification-based learning objectives described below.\\nClassification Model frames the task as a 13-way classification task where each class corresponds to a duration in Figure 2. 5 We train using cross entropy loss, selecting the closest duration class as the pseudo-gold label for each fact. Because this model can only predict a limited set of durations, we report its upperbound by always selecting the class closest to the gold duration.\\nRegression Model uses the mean squared error loss in units of log seconds, where the output from the hidden layer predicts a scalar value.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\",\n \"fig_2\",\n \"fig_2\",\n \"fig_1\"\n ],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We experiment with training on TimeQA and TimePre individually and on the union of both datasets. Figure 3 reports duration prediction and temporal commonsense performance. Overall, we find that our trained systems outperform simple random and average baselines on SituatedQA. This is indicative of strong generalizability from our distantly-supervised fact-duration systems, even when baselines benefit from access to the gold label distribution. We also provide a histogram of errors from our systems in Figure 3 where we can see that over 60% of our classification-based system's predictions are within 3 years of the gold duration, while predicting the exact duration remains challenging. Below, we reflect on the impact of our different modeling choices and research questions.\\nRegression vs. Classification Regression-based models tend to outperform their classificationbased counterparts. The instances where this is not true can be attributed to an insufficient amount of training data. In Figure 3, we can see the different types of errors each model makes. The classification system predicts duration within 1 year more frequently, but the regression system predicts duration within 4 years more frequently.\\nSupervision from KB vs. News Text We find that training on temporal knowledge-base relations (TimeQA) alone vastly outperforms training on news text (TimePre) alone for fact-duration prediction; however, the opposite is true when comparing performance on temporal commonsense (MC-TACO). Training on both datasets tends to improve our regression-based system, but yields mixed results for our classification-based system. We hypothesize that the closeness in label distribution (see Figure 2) between the training and evaluation sets impacts the performance significantly.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Fact Duration vs. Temporal Commonsense\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"While fact duration prediction and temporal commonsense are conceptually related, we find that strong performance on either task does not necessarily transfer to the other. As discussed above, this can be attributed to differences in label distributions; however, label distribution also serves as a proxy variable for the type of information being queried for in either task. Commonsense knowledge primarily differentiates events that take place over different orders of magnitude of time (e.g., seconds versus years). Differentiating whether an event takes place over finer-grained ranges (e.g., one versus two years), however, cannot be resolved with commonsense knowledge alone, and further require fact retrieval. We find that NQ contains queries for facts that change over a smaller range of durations (between 1-10 years), and, therefore, commonsense knowledge alone is insufficient.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Calibrating QA under Temporal Misalignment\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Here, we return our motivating use-case of using fact duration prediction to calibrate openretrieval QA systems under temporal misalignment.\\nWe assume an access to base calibration system, c(q, a) ∈ [0, 1] that has the same training date as the QA system its calibrating. We then use fact duration to generate misalignment-aware confidence score c m through simple post-hoc augmentation, scaling down the original confidence score by a discount factor based on the degree of misalignment and the predicted fact duration. We compute this factor differently for each of our fact duration systems.\\n• Classification: Here, the system's output is a probability distribution over different duration classes, p(d|q, a). We set the discount factor to be the CDF of this distribution evaluated at m:6 c m = c(q, a) d≥m P (d|q, a). • Regression: Here, the output is a single predicted duration d. We set the discount factor to the binary value indicating whether or not the misalignment period has exceeded the predicted duration: c m = c(q, a)1{d > m} As classification systems predict a distribution over fact durations, we are able to use the CDF of this distribution to make granular adjustments to confidence over time. In contrast, our regression systems predict a single value for fact duration, and confidence adjustments over time are abrupt, leaving confidence unchanged or setting it to zero.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Models\",\n \"publication_ref\": [\n \"b41\",\n \"b22\",\n \"b49\"\n ],\n \"table_ref\": [],\n \"text\": \"Base QA and Calibration Systems We experiment with three QA systems throughout our study: • 1 ⃝ T5: We use T5-large (Roberts et al., 2020) which has been trained with salient-spanmasking and closed-book QA on NQ-Open.\\n• 2 ⃝ DPR (t e =2018): We use DPR (Karpukhin et al., 2020), an open-book system which retrieves passages from a t e = 2018 Wikipedia snapshot and is also trained on NQ-Open.\\n• N ⃝ DPR (t e =2021): We use the same model as 2 ⃝, but swap the retrieval corpus with an updated Wikipedia snapshot that matches query times- We first adjust confidence Per-Example, which is our full system. We then adjust confidence Uniformly across all examples, such that the net decrease in confidence across the entire test set is equivalent.\\ntamp (t e = 2021) following Zhang and Choi (2021), which showed partial success in returning up-to-date answers. For each QA system, we train a calibrator that predicts the correctness of the QA system's answer. We follow Zhang et al. (2021) for the design and input features to calibrator, using the model's predicted likelihood and encoded representations of the input (details in Appendix A).\\nFact Duration Systems For both our regression and classification based models, we use the systems trained over both with TimeQA. We also include results using an oracle fact duration system, which zeroes the confidence for all questions that have been updated since the training date.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_5\",\n \"tab_6\"\n ],\n \"text\": \"Table 3 reports the results from our calibration experiments. Both QA models suffer from temporal degradation, and zero-ing out the confidence of outdated facts with oracle information improves the calibration performance. Using model prediction durations shows similar gains. Both re-gression and classification duration predictions lower the confidence of models, improving calibration metrics across the board. We find that our classification-based model consistently outperforms our regression-based model on our calibration task, despite the opposite being true for our fact-duration evaluations. We attribute this behavior to our classification-based system's error distribution, as it gets more examples correct to within 1 year (Figure 3). Classification-based system also hedge over different duration classes by predicting a distribution, which we use to compute the CDF.\\nRetrieval-Based QA: Update or Adjust In Table 3, we compare the performance of DPR with static and hot-swapped retrieval corpora from t e = 2018 and t e = 2021. While updating the retrieval corpus improves EM accuracy, adjusting model confidence using fact duration on a static corpus performs better on all calibration metrics. This suggests that, when users care about having accurate confidence estimates or seeing only highconfidence predictions, confidence adjustments can be more beneficial than swapping retrieval corpus.\\nAblations: Per-Example vs Uniform Adjustment We compare our system, which adjusts confidence on a per-example basis, against one that uniformly decreases confidence by the same value υ across the entire test set: c m = max(c(q, a)-υ, 0). These ablations still depend on our fact-duration systems to determine the υ such that the total confidence over the entire test set is the same for both methods. Table 4 reports the results from this ablation study. We find that uniformly adjusting confidence improves ECE, which is expected given the decrease in the QA systems EM accuracy after misalignment. We find, however, that our perexample adjustment methods outperform uniform confidence adjustments.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Comparisons against Prompted LLMs\",\n \"publication_ref\": [\n \"b43\",\n \"b20\",\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"While we primarily focus on calibrating fine-tuned QA models, recent work has also explored calibrating prompted large language models (LLMs) for QA (Si et al., 2022;Kadavath et al., 2022;Cole et al., 2023). Furthermore, recent general-purpose LLMs (e.g., ChatGPT) have demonstrated the ability to abstain from answering questions on the basis that their knowledge cutoff date is too far in the past; however, it is not publicly known how these systems exhibit this behavior.\\nIn this experiment, we investigate one such system, GPT-4 (OpenAI, 2023), and its ability to abstain from answering questions with rapidly changing answers. We prompt GPT-4 to answer questions from and find that it abstains from answering 86% of all questions (95% of questions whose answers have been updated between 2018 and 2021 and on 79% of examples that have not). Overall, this behavior suggests that GPT-4 overestimates how frequently it must to abstain from answering user queries. Furthermore, GPT-4 does not provide how frequently the answer is expected to change. Collectively, GPT-4's tendency to over-abstain and its lack of transparency limits its usefulness to users. In contrast, our approach provides users with an duration estimate indicating why a prediction may not be trustworthy. Further experimental details and example outputs are reported in Appendix D.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Beyond Calibration: Adaptive Inference\",\n \"publication_ref\": [\n \"b31\"\n ],\n \"table_ref\": [\n \"tab_7\",\n \"tab_7\"\n ],\n \"text\": \"In this section, we explore using our misalignmentaware confidence scores to decide how to answer a question. Below, we motivate and describe two adaptive inference scenarios where systems may choose between two methods for answering a question using their fact duration predictions. Hybrid: Closed + Open ( 1 ⃝+ N ⃝): Besides the computational benefits from not always having to use retrieval, forgoing retrieval for popular questions can also improve answer accuracy (Mallen et al., 2022). We use our fact duration predictions to decide when retrieval is necessary: we first predict an answer using T5 and run our fact duration prediction system using this answer. We then use the CDF of the predicted duration distribution to determine whether it is at least 50% likely that the fact has changed: d≤m P (d|q, a) ≥ 0.5. If so, we then run retrieval with DPR using the updated corpus t e = 2021 and present the predicted answer. We report our results in the first row of Table 5, which shows that this outperforms either system on its own, while running retrieval on less than half of all examples. Two Corpora: Relevancy vs. Recency ( 2 ⃝+ N ⃝): While most work for QA have focused on retrieving over Wikipedia, many questions require retrieving over other sources such as news or web text. One challenge in moving away from Wikipedia lies in managing temporal heterogeneity across different articles. Unlike Wikipedia, news and web articles are generally not maintained to stay current, requiring retrieval-based QA systems to identify out-of-date information in articles. Systems that retrieve such resources must consider the trade-off between the recency versus relevancy of an article. In these experiments, we experiment with using fact duration prediction as a method for weighing this trade-off in retrieval.\\nOur experimental setup is as follows: instead of computing misalignment from the model's training date, we compute relative to when the article was authored (m = 3 years for 2018 Wikipedia and m = 0 years for 2021 Wikipedia). After performing inference using both corpora, we re-rank answers according to their misalignment-adjusted confidence estimates. We report results in Table 5. We find that our method is able to recover comparable performance to always using up-to-date articles, while using it just under half the time.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b53\",\n \"b39\",\n \"b13\",\n \"b1\",\n \"b12\",\n \"b46\",\n \"b4\",\n \"b21\",\n \"b49\",\n \"b45\",\n \"b11\",\n \"b33\",\n \"b32\",\n \"b35\",\n \"b17\",\n \"b25\"\n ],\n \"table_ref\": [],\n \"text\": \"Commonsense and Temporal Reasoning Recent works have proposed forecasting bench-marks (Zou et al., 2022;Jin et al., 2021a) related to our fact duration prediction task. While our task asks models to predict when a fact will change, these forecasting tasks ask how a fact will change. Qin et al. (2021) studies temporal commonsense reasoning in dialogues. Quantitative reasoning has been explored in other works as quantitative relations between nouns (Forbes and Choi, 2017;Bagherinezhad et al., 2016), distributions over quantitative attributes Elazar et al. (2019), and representing numbers in language models (Wallace et al., 2019).\\nCalibration Abstaining from providing a QA system's answers has been explored in several recent works. Chen et al. (2022) examines instances where knowledge conflicts exist between a model's memorized knowledge and retrieved documents. As the authors note, such instances often arise due to temporal misalignment. Prior work (Kamath et al., 2020;Zhang et al., 2021;Varshney and Baral, 2023) has explored abstaining from answering questions by predicting whether or not the test question comes from the same training distribution of the QA system. While fact duration also predicts a shift in distribution, fact duration focuses on predicting a shift in a question's output distribution of answers instead of a shift in input distribution of questions; therefore, these two systems are addressing orthogonal challenges in robustness to distribution shift and are complementary.\\nKeeping Systems Up-to-Date Several works have explored continuing pretraining to address temporal misalignment in pretrained models (Dhingra et al., 2022;Jin et al., 2021b). Other works have explored editing specific facts into models (Cao et al., 2021;Mitchell et al., 2022;Meng et al., 2022). These works, however, have only focused on synthetic settings and assume access to the updated facts. Furthermore, such systems have yet to be successfully applied to new benchmarks for measuring whether language models have acquired emergent information (Onoe et al., 2022;Padmanabhan et al., 2023). Recent works on retrievalbased QA systems have found improved adaptation when updated with up-to-date retrieval corpora (Izacard et al., 2022;Lazaridou et al., 2022).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We improve QA calibration under temporal misalignment by introducing the fact duration pre-diction task, alongside several datasets and baseline systems for it. Future work may build upon this evaluation framework to further improve QA calibration under temporal misalignment. For instance, future work may examine modeling different classes distributions of fact duration distributions, like modeling whether a fact changes after a regular, periodic time interval.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_5\"\n ],\n \"text\": \"We only evaluate temporal misalignment between 2018 and 2021, a three-year time difference, on a relatively small scale SituatedQA dataset (N=322). This is mainly due to a lack of benchmark that supports studying temporal misalignment. Exploring this in more diverse setup, including different languages, text domains, wide range of temporal gaps, would be fruitful direction for future work.\\nAs is the case with all systems that attempt to faithfully relay world knowledge, treating model predictions as fact runs the risk of propagating misinformation. While the goal of our fact duration prediction systems is to prevent models from reciting outdated facts, it does not always succeed and facts may change earlier than expected. Even though a given fact may be expected to only change once every decade, an improbable outcome may occur and the fact changes after only a year. In such an event, our misalignment-aware calibration system may erroneously maintain high confidence in the outdated answer. Furthermore, our system, as it stands, does not take the answer start date into account. Our system also can make errors due to changes in the typical duration of a given fact. For instance \\\"What's the world's tallest building?\\\" changes more frequently over time as the rate of technological advances also increases. We provide examples of such system errors in Appendix C.\\nTable 7: Additional calibration results on SituatedQA, comparing different methods of adjusting model confidence for temporal misalignment using the output of our classification-based fact-duration system. All other settings are the same as in Table 3.\\nsimply assume that all such question answer pairs instances have a duration of 1 month.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.2 Temporal Commonsense Datasets Preprocessing\",\n \"publication_ref\": [\n \"b47\"\n ],\n \"table_ref\": [],\n \"text\": \"As we noted above, each MCTACO example consists of a multiple-choice question about the duration of some event in a provided context sentence. During prepossessing, we use the same question conversion model as above to transform each QA pair into a statement and prepend the context sentence onto each question. We use the metrics proposed by the original authors, and we select all multiple choice options whose duration falls within some absolute threshold of predicted duration, measured in log seconds (Yang et al., 2020). This threshold is selected based on development set performance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C Additional Results\",\n \"publication_ref\": [\n \"b16\"\n ],\n \"table_ref\": [],\n \"text\": \"Different Pretrained Models for Fact Duration Prediction In Table 6, we report our results from DeBERTa-v3-base (He et al., 2021) on our fact duration prediction system. We also experiment with using the large variants of both BERT and DeBERTa, but do not find substantial improvement.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Adjusting Confidence with Expected Duration\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In addition adjusting confidence using the CDF of the predicted duration distribution from our classification-based system, we also experiment with using the expected duration as our discounting factor. We incorporate this by zeroing the confidence estimate if the expected duration is exceeded by the degree of misalignment: f (q, a) = 1{ d d • P (d|q, a) > m}.\\nIn Table 7, we report additional results on our calibration evaluation. We include calibration performance of our best performing fact duration models finetuned on SituatedQA: trained only on Sit-uatedQA for our classification-based model and first trained on TimeQA + TA Pretrain for our regression-based model.\\nError Analysis In Table 9, we highlight sampled errors from our fact duration system and discuss their causes and impact.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D ChatGPT and GPT-4 outputs\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_11\"\n ],\n \"text\": \"Table 10 includes two examples of ChatGPT informing users that the answers to a given question may have changed. It, however, does not provide users with an estimate of how likely it has changed, or how often the answer is expected to change. This lack of a duration estimate results in lesser transparency and interpretatbility for users. To get results on SituatedQA, we prompt GPT-4 with the following system prompt (recommended by their documentation): \\\"You are ChatGPT, a large language model trained by OpenAI. Answer as con- 9: Example fact duration predictions errors from SituatedQA. We include the answers from t M = 2018 and t q = 2021, and the gold / predicted durations ({G / P} Dur). In the first example, we see that our system correctly identified volcanic eruptions as an infrequent event; however, our system misses that the answer has changed within the misalignment period because it doesn't have access to the answer start date. In the second example, our system fails no longer changes since the show ended. In the last example, our model underestimates the fact's duration, but still correctly identifies that the answer has changed. cisely as possible. Knowledge cutoff: September 2021. Current date: May 12, 2023.\\\" We then present GPT-4 with the user's question from Situat-edQA. We determine whether a systems abstains from a given prediction if it references its knowledge cutoff from the prompt September 2021 or if it mentions real-time information in its generated output.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"E Licensing\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Dataset Licensing TimeQA is licensed under the BSD 3-Clause License. SituatedQA and Wikipedia are both distributed under the CC BY-SA 4.0 license. MCTACO and TA-Pretrain are both unlicensed.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Model Licensing\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The transformers library, BERT model, T5 model, and question conversion model are all licensed under the Apache License v2.0. DPR is licensed under the Attribution-NonCommercial 4.0 International license. Our use User: Who is the current US president? ChatGPT: As an AI language model, I don't have realtime information or the ability to browse the internet. My knowledge was last updated in September 2021, and at that time, the President of the United States was Joseph R. Biden Jr. However, please note that the political landscape can change, and there may have been a new election or other developments since then. To obtain the most accurate and up-to-date information, I recommend checking reliable sources such as official government websites or reputable news outlets.\\nUser: Who won the last NBA Champtionship? ChatGPT: As an AI language model, I don't have realtime data, and my knowledge was last updated in September 2021. Therefore, I cannot provide information on the winner of the most recent NBA Championship. However, as of my last update, the Milwaukee Bucks were the reigning NBA Champions, having won the championship in the 2020-2021 season. I recommend checking reliable news sources or conducting a quick online search to find the current champion. of all these resources are in compliance with their licenses.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"F Ethical Considerations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We do not collect any data in this paper. While we do generate a dataset, we visually inspect generated examples and do not find any cases of harmful or offensive content. The existing datasets used in this work have previously been vetted by their authors prior work. We also note that our work is only applied to English datasets and models.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgements\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The work was partially supported by Google Research Award, a gift from HomeDepot, and a grant from UT Machine Learning Lab. This work was in part supported by Cisco Research. Any opinions, findings and conclusions, or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of Cisco Research. The authors would like to thank the members of the UT Austin NLP community and Jordan Boyd-Graber for helpful discussions.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Our data and code are released publicly at https://github.com/ mikejqzhang\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Implementation Details\",\n \"publication_ref\": [\n \"b49\"\n ],\n \"table_ref\": [],\n \"text\": \"A.1 QA Models\\nWe use the T5 7 and DPR 8 checkpoints that have been finetuned on NQ-Open's training set from the transformers library model hub. 9 For DPR, we use the retrieval corpora from December 20, 2018 for t E = 2018 and February 20, 2021 for t E = 2021, following Zhang and Choi (2021).\\n7 https://huggingface.co/google/t5-large-ssm-nqo 8 https://huggingface.co/facebook/dpr-reader-single-nqbase 9 https://huggingface.co/models\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.2 Calibration Models\",\n \"publication_ref\": [\n \"b6\",\n \"b49\",\n \"b49\"\n ],\n \"table_ref\": [],\n \"text\": \"We implement our trained calibration systems using XGBoost (Chen and Guestrin, 2016) using the features for T5 and DPR outlined in Zhang et al. (2021). For T5, we concatenate (1) the averaged, encoded representations of the input question, and\\n(2) the model likelihood. For DPR, we concatenate (1) the averaged, encoded representations of the input question and selected passage, (2) the averaged, encoded representations of the start and end tokens of the selected answer span, and (3) the likelihood of the answer span, computed as the product of the likelihoods of selecting the start index, end index, and passage index. We train our systems on NQ-Open on a randomly sampled 60/40 training and development splits, following Zhang et al. (2021). We use a maximum depth of 10 and experiment with several values for the learning rate {0.01, 0.1, 0.2, 0.5} and column sub-sampling ratio {0.0, 0.1, . . . , 0.9}, which we keep the for sampling by tree, level, and node. We train with early stopping after 10 epochs without improvement and select the best performing system as evaluated on the development split.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.3 Fact Duration Prediction Models\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We use BERT-base from the transformers library for all duration prediction baselines, trained with a batch size in {32, 64} and learning rate in {1e -5, 5e -5}. We train until convergence and select the best checkpoint as determined by development set performance. Due to computational resource constraints, we do not further tune hyperparameters. All models are trained once and results reflect a single run. All experiments took were performed Quadro RTX 8000 gpus and required less than one week's worth of GPU hours.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B Datasets\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We provide examples from each dataset and our prepossessing pipeline in Table 8. We provide futher preprocessing details below.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.1 Fact Duration Dataset Preprocessing\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In TimeQA, several examples have answers that are simply the empty string. We remove all such examples from our preprocessed dataset. In SituatedQA, several examples have answers that begin and end in the same year, without further annotation determining the exact number of days or months. We \"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":""},"doi":{"kind":"string","value":"10.18653/v1/2020.emnlp-main.550"},"references":{"kind":"list like","value":[{"authors":"Stephen Anastasios N Angelopoulos; Adam Bates; Lihua Fisch; Tal Lei; Schuster","journal":"","ref_id":"b0","title":"Conformal risk control","year":"2022"},{"authors":"Hessam Bagherinezhad; Hannaneh Hajishirzi; Yejin Choi; Ali Farhadi","journal":"","ref_id":"b1","title":"Are elephants bigger than butterflies? reasoning about sizes of objects","year":"2016"},{"authors":"Nicola De Cao; W Aziz; Ivan Titov","journal":"","ref_id":"b2","title":"Editing factual knowledge in language models","year":"2021"},{"authors":"Danqi Chen; Adam Fisch; Jason Weston; Antoine Bordes","journal":"","ref_id":"b3","title":"Reading wikipedia to answer opendomain questions","year":"2017"},{"authors":"Hung-Ting Chen; Michael Jq Zhang; Eunsol Choi","journal":"","ref_id":"b4","title":"Rich knowledge sources bring complex knowledge conflicts: Recalibrating models to reflect conflicting evidence","year":"2022"},{"authors":"Jifan Chen; Eunsol Choi; Greg Durrett","journal":"","ref_id":"b5","title":"Can nli models verify qa systems' predictions?","year":"2021"},{"authors":"Tianqi Chen; Carlos Guestrin","journal":"","ref_id":"b6","title":"Xgboost: A scalable tree boosting system","year":"2016"},{"authors":"Wenhu Chen; Xinyi Wang; William Yang; Wang ","journal":"","ref_id":"b7","title":"A dataset for answering time-sensitive questions","year":"2021"},{"authors":"Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann; Parker Schuh; Kensen Shi; Sasha Tsvyashchenko; Joshua Maynez; Abhishek Rao; Parker Barnes; Yi Tay; Noam M Shazeer; Emily Vinodkumar Prabhakaran; Nan Reif; Benton C Du; Reiner Hutchinson; James Pope; Jacob Bradbury; Michael Austin; Guy Isard; Pengcheng Gur-Ari; Toju Yin; Anselm Duke; Sanjay Levskaya; Sunipa Ghemawat; Henryk Dev; Xavier Michalewski; Vedant García; Kevin Misra; Liam Robinson; Denny Fedus; Daphne Zhou; David Ippolito; Hyeontaek Luan; Barret Lim; Alexander Zoph; Ryan Spiridonov; David Sepassi; Shivani Dohan; Mark Agrawal; Andrew M Omernick; Thanumalayan Dai; Marie Sankaranarayana Pillai; Aitor Pellat; Erica Lewkowycz; Rewon Moreira; Oleksandr Child; Katherine Polozov; Zongwei Lee; Xuezhi Zhou; Brennan Wang; Mark Saeta; Orhan Díaz; Michele Firat; Jason Catasta; Kathleen S Wei; Douglas Meier-Hellstern; Jeff Eck; Slav Dean; Noah Petrov; Fiedel","journal":"","ref_id":"b8","title":"Palm: Scaling language modeling with pathways","year":"2022"},{"authors":" Jeremy R Cole; J Q Michael; Daniel Zhang; Julian Gillick; Bhuwan Martin Eisenschlos; Jacob Dhingra; Eisenstein","journal":"","ref_id":"b9","title":"Selectively answering ambiguous questions","year":"2023"},{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"","ref_id":"b10","title":"Bert: Pre-training of deep bidirectional transformers for language understanding","year":"2018"},{"authors":"Bhuwan Dhingra; Jeremy R Cole; Julian Martin Eisenschlos; Daniel Gillick; Jacob Eisenstein; William W Cohen","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b11","title":"Time-aware language models as temporal knowledge bases","year":"2022"},{"authors":"A Yanai Elazar; Deepak Mahabal; Tania Ramachandran; Dan Bedrax-Weiss; Roth","journal":"","ref_id":"b12","title":"How large are lions? inducing distributions over quantitative attributes","year":"2019"},{"authors":"Maxwell Forbes; Yejin Choi","journal":"","ref_id":"b13","title":"Verb physics: Relative physical knowledge of actions and objects","year":"2017"},{"authors":"Kelvin Guu; Kenton Lee; Zora Tung; Panupong Pasupat; Mingwei Chang","journal":"","ref_id":"b14","title":"Retrieval augmented language model pre-training","year":"2020"},{"authors":" Pmlr","journal":"","ref_id":"b15","title":"","year":""},{"authors":"Pengcheng He; Jianfeng Gao; Weizhu Chen","journal":"","ref_id":"b16","title":"Debertav3: Improving deberta using electra-style pretraining with gradient-disentangled embedding sharing","year":"2021"},{"authors":"Gautier Izacard; Patrick Lewis; Maria Lomeli; Lucas Hosseini; Fabio Petroni; Timo Schick; Jane Dwivedi-Yu; Armand Joulin; Sebastian Riedel; Edouard Grave","journal":"","ref_id":"b17","title":"Few-shot Learning with Retrieval Augmented Language Models","year":"2022"},{"authors":"Woojeong Jin; Suji Kim; Rahul Khanna; Dong-Ho Lee; Fred Morstatter; A G Galstyan; Xiang Ren","journal":"","ref_id":"b18","title":"Forecastqa: A question answering challenge for event forecasting with temporal text data","year":"2021"},{"authors":"Xisen Jin; Dejiao Zhang; Henghui Zhu; Wei Xiao; Shang-Wen; Xiaokai Li; Andrew Wei; Xiang Arnold; Ren","journal":"","ref_id":"b19","title":"Lifelong pretraining: Continually adapting language models to emerging corpora","year":"2021"},{"authors":"Saurav Kadavath; Tom Conerly; Amanda Askell; T J Henighan; Dawn Drain; Ethan Perez; Nicholas Schiefer; Zachary Dodds; Nova Dassarma; Eli Tran-Johnson; Scott Johnston; Sheer El-Showk; Andy Jones; Nelson Elhage; Tristan Hume; Anna Chen; Yuntao Bai; Sam Bowman; Stanislav Fort; Deep Ganguli; Danny Hernandez; Josh Jacobson; John Kernion; Shauna Kravec; Liane Lovitt; Kamal Ndousse; Catherine Olsson; Sam Ringer; Dario Amodei; Tom B Brown; Jack Clark; Nicholas Joseph; Benjamin Mann; Sam Mccandlish; Christopher Olah; Jared Kaplan","journal":"","ref_id":"b20","title":"Language models (mostly) know what they know","year":"2022"},{"authors":"Amita Kamath; Robin Jia; Percy Liang","journal":"","ref_id":"b21","title":"Selective question answering under domain shift","year":"2020"},{"authors":"Vladimir Karpukhin; Barlas Oguz; Sewon Min; Patrick Lewis; Ledell Wu; Sergey Edunov; Danqi Chen; Wen-Tau Yih","journal":"Association for Computational Linguistics","ref_id":"b22","title":"Dense passage retrieval for opendomain question answering","year":"2020"},{"authors":"Jungo Kasai; Keisuke Sakaguchi; Yoichi Takahashi; Ronan Le Bras; Akari Asai; Xinyan Yu; Dragomir Radev; Noah A Smith; Yejin Choi; Kentarou Inui","journal":"","ref_id":"b23","title":"Realtime qa: What's the answer right now?","year":"2022"},{"authors":"Tom Kwiatkowski; Jennimaria Palomaki; Olivia Redfield; Michael Collins; Ankur Parikh; Chris Alberti; Danielle Epstein; Illia Polosukhin; Jacob Devlin; Kenton Lee; Kristina Toutanova; Llion Jones; Matthew Kelcey; Ming-Wei Chang; Andrew M Dai; Jakob Uszkoreit; Quoc Le; Slav Petrov","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b24","title":"Natural questions: A benchmark for question answering research","year":"2019"},{"authors":"Angeliki Lazaridou; Elena Gribovskaya; Wojciech Stokowiec; Nikolai Grigorev","journal":"","ref_id":"b25","title":"Internetaugmented language models through few-shot prompting for open-domain question answering","year":"2022"},{"authors":"Angeliki Lazaridou; Adhiguna Kuncoro; Elena Gribovskaya; Devang Agrawal; Adam Liska; Tayfun Terzi; Mai Gimenez; Cyprien De Masson D'autume; Tomás Kociský; Sebastian Ruder; Dani Yogatama; Kris Cao; Susannah Young; Phil Blunsom","journal":"","ref_id":"b26","title":"Mind the gap: Assessing temporal generalization in neural language models","year":"2021"},{"authors":"Kenton Lee; Ming-Wei Chang; Kristina Toutanova","journal":"","ref_id":"b27","title":"Latent retrieval for weakly supervised open domain question answering","year":"2019"},{"authors":"Patrick Lewis; Yuxiang Wu; Linqing Liu; Pasquale Minervini; Heinrich Küttler; Aleksandra Piktus; Pontus Stenetorp; Sebastian Riedel","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b28","title":"Paq: 65 million probably-asked questions and what you can do with them","year":"2021"},{"authors":"Adam Livska; Elena Tom'avs Kovcisk'y; Tayfun Gribovskaya; Eren Terzi; Devang Sezener; Cyprien Agrawal; Tim De Masson D'autume; Manzil Scholtes; Susannah Zaheer; Ellen Young; Sophia Gilsenan-Mcmahon; Phil Austin; Angeliki Blunsom; Lazaridou","journal":"","ref_id":"b29","title":"Streamingqa: A benchmark for adaptation to new knowledge over time in question answering models","year":"2022"},{"authors":"Kelvin Luu; Daniel Khashabi; Suchin Gururangan; Karishma Mandyam; Noah A Smith","journal":"","ref_id":"b30","title":"Time waits for no one! analysis and challenges of temporal misalignment","year":"2021"},{"authors":"Alex Mallen; Akari Asai; Victor Zhong; Rajarshi Das; Hannaneh Hajishirzi; Daniel Khashabi","journal":"","ref_id":"b31","title":"When not to trust language models: Investigating effectiveness and limitations of parametric and non-parametric memories","year":"2022"},{"authors":"Kevin Meng; David Bau; Alex Andonian; Yonatan Belinkov","journal":"","ref_id":"b32","title":"Locating and Editing Factual Associations in GPT","year":"2022"},{"authors":"Eric Mitchell; Charles Lin; Antoine Bosselut; Chelsea Finn; Christopher D Manning","journal":"","ref_id":"b33","title":"Fast model editing at scale","year":"2022"},{"authors":"Yasumasa Onoe; J Q Michael; Eunsol Zhang; Greg Choi; Durrett","journal":"","ref_id":"b34","title":"Creak: A dataset for commonsense reasoning over entity knowledge","year":"2021"},{"authors":"Yasumasa Onoe; J Q Michael; Eunsol Zhang; Greg Choi; Durrett","journal":"","ref_id":"b35","title":"Entity cloze by date: What lms know about unseen entities","year":"2022"},{"authors":"Yasumasa Onoe; J Q Michael; Zhang; Greg Shankar Padmanabhan; Eunsol Durrett; Choi","journal":"","ref_id":"b36","title":"Can lms learn new entities from descriptions? challenges in propagating injected knowledge","year":"2023"},{"authors":" Openai","journal":"","ref_id":"b37","title":"","year":"2023"},{"authors":"Yasumasa Shankar Padmanabhan; Onoe; J Q Michael; Greg Zhang; Eunsol Durrett; Choi","journal":"","ref_id":"b38","title":"Propagating knowledge updates to lms through distillation","year":"2023"},{"authors":"Lianhui Qin; Aditya Gupta; Shyam Upadhyay; Luheng He; Yejin Choi; Manaal Faruqui","journal":"","ref_id":"b39","title":"Timedial: Temporal commonsense reasoning in dialog","year":"2021"},{"authors":"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu","journal":"J. Mach. Learn. Res","ref_id":"b40","title":"Exploring the limits of transfer learning with a unified text-to-text transformer","year":"2020"},{"authors":"Adam Roberts; Colin Raffel; Noam M Shazeer","journal":"","ref_id":"b41","title":"How much knowledge can you pack into the parameters of a language model?","year":"2020"},{"authors":"Timo Schick; Hinrich Schütze","journal":"","ref_id":"b42","title":"Exploiting cloze questions for few shot text classification and natural language inference","year":"2020"},{"authors":"Chenglei Si; Zhe Gan; Zhengyuan Yang; Shuohang Wang; Jianfeng Wang; Jordan Boyd-Graber; Lijuan Wang","journal":"","ref_id":"b43","title":"Prompting gpt-3 to be reliable","year":"2022"},{"authors":"Dustin Tran; Jeremiah Liu; Michael W Dusenberry; Du Phan; Mark Collier; Jie Ren; Kehang Han; Zi Wang; Zelda Mariet; Huiyi Hu","journal":"","ref_id":"b44","title":"Plex: Towards reliability using pretrained large model extensions","year":"2022"},{"authors":"Neeraj Varshney; Chitta Baral","journal":"","ref_id":"b45","title":"Postabstention: Towards reliably re-attempting the abstained instances in qa","year":"2023"},{"authors":"Eric Wallace; Yizhong Wang; Sujian Li; Sameer Singh; Matt Gardner","journal":"","ref_id":"b46","title":"Do NLP models know numbers? probing numeracy in embeddings","year":"2019"},{"authors":"Zonglin Yang; X Du; Alexander M Rush; Claire Cardie","journal":"","ref_id":"b47","title":"Improving event duration prediction via time-aware pre-training","year":"2020"},{"authors":"J Q Michael; Eunsol Zhang; Choi","journal":"","ref_id":"b48","title":"Situatedqa: Incorporating extra-linguistic contexts into qa","year":"2021"},{"authors":"Shujian Zhang; Chengyue Gong; Eunsol Choi","journal":"","ref_id":"b49","title":"Knowing more about questions can help: Improving calibration in question answering","year":"2021"},{"authors":"Susan Zhang; Stephen Roller; Naman Goyal; Mikel Artetxe; Moya Chen; Shuohui Chen; Christopher Dewan; Mona Diab; Xian Li; Xi Victoria Lin; Todor Mihaylov; Myle Ott; Sam Shleifer; Kurt Shuster; Daniel Simig; Punit Singh Koura; Anjali Sridhar; Tianlu Wang; Luke Zettlemoyer","journal":"","ref_id":"b50","title":"Opt: Open pre-trained transformer language models","year":"2022"},{"authors":"Ben Zhou; Daniel Khashabi; Qiang Ning; Dan Roth","journal":"","ref_id":"b51","title":"going on a vacation\" takes longer than \"going for a walk","year":"2019"},{"authors":"Ben Zhou; Qiang Ning; Daniel Khashabi; Dan Roth","journal":"","ref_id":"b52","title":"Temporal common sense acquisition with minimal supervision","year":"2020"},{"authors":"Andy Zou; Tristan Xiao; Ryan Jia; Joe Kwon; Mantas Mazeika; Richard Li; Dawn Song; Jacob Steinhardt; Owain Evans; Dan Hendrycks","journal":"","ref_id":"b53","title":"Forecasting future world events with neural networks","year":"2022"}],"string":"[\n {\n \"authors\": \"Stephen Anastasios N Angelopoulos; Adam Bates; Lihua Fisch; Tal Lei; Schuster\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Conformal risk control\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hessam Bagherinezhad; Hannaneh Hajishirzi; Yejin Choi; Ali Farhadi\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Are elephants bigger than butterflies? reasoning about sizes of objects\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Nicola De Cao; W Aziz; Ivan Titov\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Editing factual knowledge in language models\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Danqi Chen; Adam Fisch; Jason Weston; Antoine Bordes\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Reading wikipedia to answer opendomain questions\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Hung-Ting Chen; Michael Jq Zhang; Eunsol Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Rich knowledge sources bring complex knowledge conflicts: Recalibrating models to reflect conflicting evidence\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jifan Chen; Eunsol Choi; Greg Durrett\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Can nli models verify qa systems' predictions?\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Tianqi Chen; Carlos Guestrin\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Xgboost: A scalable tree boosting system\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Wenhu Chen; Xinyi Wang; William Yang; Wang \",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"A dataset for answering time-sensitive questions\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann; Parker Schuh; Kensen Shi; Sasha Tsvyashchenko; Joshua Maynez; Abhishek Rao; Parker Barnes; Yi Tay; Noam M Shazeer; Emily Vinodkumar Prabhakaran; Nan Reif; Benton C Du; Reiner Hutchinson; James Pope; Jacob Bradbury; Michael Austin; Guy Isard; Pengcheng Gur-Ari; Toju Yin; Anselm Duke; Sanjay Levskaya; Sunipa Ghemawat; Henryk Dev; Xavier Michalewski; Vedant García; Kevin Misra; Liam Robinson; Denny Fedus; Daphne Zhou; David Ippolito; Hyeontaek Luan; Barret Lim; Alexander Zoph; Ryan Spiridonov; David Sepassi; Shivani Dohan; Mark Agrawal; Andrew M Omernick; Thanumalayan Dai; Marie Sankaranarayana Pillai; Aitor Pellat; Erica Lewkowycz; Rewon Moreira; Oleksandr Child; Katherine Polozov; Zongwei Lee; Xuezhi Zhou; Brennan Wang; Mark Saeta; Orhan Díaz; Michele Firat; Jason Catasta; Kathleen S Wei; Douglas Meier-Hellstern; Jeff Eck; Slav Dean; Noah Petrov; Fiedel\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Palm: Scaling language modeling with pathways\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \" Jeremy R Cole; J Q Michael; Daniel Zhang; Julian Gillick; Bhuwan Martin Eisenschlos; Jacob Dhingra; Eisenstein\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Selectively answering ambiguous questions\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Bert: Pre-training of deep bidirectional transformers for language understanding\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Bhuwan Dhingra; Jeremy R Cole; Julian Martin Eisenschlos; Daniel Gillick; Jacob Eisenstein; William W Cohen\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b11\",\n \"title\": \"Time-aware language models as temporal knowledge bases\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"A Yanai Elazar; Deepak Mahabal; Tania Ramachandran; Dan Bedrax-Weiss; Roth\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"How large are lions? inducing distributions over quantitative attributes\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Maxwell Forbes; Yejin Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Verb physics: Relative physical knowledge of actions and objects\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Kelvin Guu; Kenton Lee; Zora Tung; Panupong Pasupat; Mingwei Chang\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Retrieval augmented language model pre-training\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \" Pmlr\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Pengcheng He; Jianfeng Gao; Weizhu Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Debertav3: Improving deberta using electra-style pretraining with gradient-disentangled embedding sharing\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Gautier Izacard; Patrick Lewis; Maria Lomeli; Lucas Hosseini; Fabio Petroni; Timo Schick; Jane Dwivedi-Yu; Armand Joulin; Sebastian Riedel; Edouard Grave\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Few-shot Learning with Retrieval Augmented Language Models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Woojeong Jin; Suji Kim; Rahul Khanna; Dong-Ho Lee; Fred Morstatter; A G Galstyan; Xiang Ren\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Forecastqa: A question answering challenge for event forecasting with temporal text data\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Xisen Jin; Dejiao Zhang; Henghui Zhu; Wei Xiao; Shang-Wen; Xiaokai Li; Andrew Wei; Xiang Arnold; Ren\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Lifelong pretraining: Continually adapting language models to emerging corpora\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Saurav Kadavath; Tom Conerly; Amanda Askell; T J Henighan; Dawn Drain; Ethan Perez; Nicholas Schiefer; Zachary Dodds; Nova Dassarma; Eli Tran-Johnson; Scott Johnston; Sheer El-Showk; Andy Jones; Nelson Elhage; Tristan Hume; Anna Chen; Yuntao Bai; Sam Bowman; Stanislav Fort; Deep Ganguli; Danny Hernandez; Josh Jacobson; John Kernion; Shauna Kravec; Liane Lovitt; Kamal Ndousse; Catherine Olsson; Sam Ringer; Dario Amodei; Tom B Brown; Jack Clark; Nicholas Joseph; Benjamin Mann; Sam Mccandlish; Christopher Olah; Jared Kaplan\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Language models (mostly) know what they know\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Amita Kamath; Robin Jia; Percy Liang\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Selective question answering under domain shift\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Vladimir Karpukhin; Barlas Oguz; Sewon Min; Patrick Lewis; Ledell Wu; Sergey Edunov; Danqi Chen; Wen-Tau Yih\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b22\",\n \"title\": \"Dense passage retrieval for opendomain question answering\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jungo Kasai; Keisuke Sakaguchi; Yoichi Takahashi; Ronan Le Bras; Akari Asai; Xinyan Yu; Dragomir Radev; Noah A Smith; Yejin Choi; Kentarou Inui\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Realtime qa: What's the answer right now?\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tom Kwiatkowski; Jennimaria Palomaki; Olivia Redfield; Michael Collins; Ankur Parikh; Chris Alberti; Danielle Epstein; Illia Polosukhin; Jacob Devlin; Kenton Lee; Kristina Toutanova; Llion Jones; Matthew Kelcey; Ming-Wei Chang; Andrew M Dai; Jakob Uszkoreit; Quoc Le; Slav Petrov\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b24\",\n \"title\": \"Natural questions: A benchmark for question answering research\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Angeliki Lazaridou; Elena Gribovskaya; Wojciech Stokowiec; Nikolai Grigorev\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Internetaugmented language models through few-shot prompting for open-domain question answering\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Angeliki Lazaridou; Adhiguna Kuncoro; Elena Gribovskaya; Devang Agrawal; Adam Liska; Tayfun Terzi; Mai Gimenez; Cyprien De Masson D'autume; Tomás Kociský; Sebastian Ruder; Dani Yogatama; Kris Cao; Susannah Young; Phil Blunsom\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Mind the gap: Assessing temporal generalization in neural language models\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Kenton Lee; Ming-Wei Chang; Kristina Toutanova\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Latent retrieval for weakly supervised open domain question answering\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Patrick Lewis; Yuxiang Wu; Linqing Liu; Pasquale Minervini; Heinrich Küttler; Aleksandra Piktus; Pontus Stenetorp; Sebastian Riedel\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b28\",\n \"title\": \"Paq: 65 million probably-asked questions and what you can do with them\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Adam Livska; Elena Tom'avs Kovcisk'y; Tayfun Gribovskaya; Eren Terzi; Devang Sezener; Cyprien Agrawal; Tim De Masson D'autume; Manzil Scholtes; Susannah Zaheer; Ellen Young; Sophia Gilsenan-Mcmahon; Phil Austin; Angeliki Blunsom; Lazaridou\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Streamingqa: A benchmark for adaptation to new knowledge over time in question answering models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Kelvin Luu; Daniel Khashabi; Suchin Gururangan; Karishma Mandyam; Noah A Smith\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Time waits for no one! analysis and challenges of temporal misalignment\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Alex Mallen; Akari Asai; Victor Zhong; Rajarshi Das; Hannaneh Hajishirzi; Daniel Khashabi\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"When not to trust language models: Investigating effectiveness and limitations of parametric and non-parametric memories\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Kevin Meng; David Bau; Alex Andonian; Yonatan Belinkov\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Locating and Editing Factual Associations in GPT\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Eric Mitchell; Charles Lin; Antoine Bosselut; Chelsea Finn; Christopher D Manning\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Fast model editing at scale\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yasumasa Onoe; J Q Michael; Eunsol Zhang; Greg Choi; Durrett\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Creak: A dataset for commonsense reasoning over entity knowledge\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yasumasa Onoe; J Q Michael; Eunsol Zhang; Greg Choi; Durrett\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Entity cloze by date: What lms know about unseen entities\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yasumasa Onoe; J Q Michael; Zhang; Greg Shankar Padmanabhan; Eunsol Durrett; Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Can lms learn new entities from descriptions? challenges in propagating injected knowledge\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \" Openai\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Yasumasa Shankar Padmanabhan; Onoe; J Q Michael; Greg Zhang; Eunsol Durrett; Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Propagating knowledge updates to lms through distillation\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Lianhui Qin; Aditya Gupta; Shyam Upadhyay; Luheng He; Yejin Choi; Manaal Faruqui\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Timedial: Temporal commonsense reasoning in dialog\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu\",\n \"journal\": \"J. Mach. Learn. Res\",\n \"ref_id\": \"b40\",\n \"title\": \"Exploring the limits of transfer learning with a unified text-to-text transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Adam Roberts; Colin Raffel; Noam M Shazeer\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"How much knowledge can you pack into the parameters of a language model?\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Timo Schick; Hinrich Schütze\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Exploiting cloze questions for few shot text classification and natural language inference\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Chenglei Si; Zhe Gan; Zhengyuan Yang; Shuohang Wang; Jianfeng Wang; Jordan Boyd-Graber; Lijuan Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"Prompting gpt-3 to be reliable\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Dustin Tran; Jeremiah Liu; Michael W Dusenberry; Du Phan; Mark Collier; Jie Ren; Kehang Han; Zi Wang; Zelda Mariet; Huiyi Hu\",\n \"journal\": \"\",\n \"ref_id\": \"b44\",\n \"title\": \"Plex: Towards reliability using pretrained large model extensions\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Neeraj Varshney; Chitta Baral\",\n \"journal\": \"\",\n \"ref_id\": \"b45\",\n \"title\": \"Postabstention: Towards reliably re-attempting the abstained instances in qa\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Eric Wallace; Yizhong Wang; Sujian Li; Sameer Singh; Matt Gardner\",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"Do NLP models know numbers? probing numeracy in embeddings\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Zonglin Yang; X Du; Alexander M Rush; Claire Cardie\",\n \"journal\": \"\",\n \"ref_id\": \"b47\",\n \"title\": \"Improving event duration prediction via time-aware pre-training\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"J Q Michael; Eunsol Zhang; Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b48\",\n \"title\": \"Situatedqa: Incorporating extra-linguistic contexts into qa\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Shujian Zhang; Chengyue Gong; Eunsol Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b49\",\n \"title\": \"Knowing more about questions can help: Improving calibration in question answering\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Susan Zhang; Stephen Roller; Naman Goyal; Mikel Artetxe; Moya Chen; Shuohui Chen; Christopher Dewan; Mona Diab; Xian Li; Xi Victoria Lin; Todor Mihaylov; Myle Ott; Sam Shleifer; Kurt Shuster; Daniel Simig; Punit Singh Koura; Anjali Sridhar; Tianlu Wang; Luke Zettlemoyer\",\n \"journal\": \"\",\n \"ref_id\": \"b50\",\n \"title\": \"Opt: Open pre-trained transformer language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ben Zhou; Daniel Khashabi; Qiang Ning; Dan Roth\",\n \"journal\": \"\",\n \"ref_id\": \"b51\",\n \"title\": \"going on a vacation\\\" takes longer than \\\"going for a walk\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Ben Zhou; Qiang Ning; Daniel Khashabi; Dan Roth\",\n \"journal\": \"\",\n \"ref_id\": \"b52\",\n \"title\": \"Temporal common sense acquisition with minimal supervision\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Andy Zou; Tristan Xiao; Ryan Jia; Joe Kwon; Mantas Mazeika; Richard Li; Dawn Song; Jacob Steinhardt; Owain Evans; Dan Hendrycks\",\n \"journal\": \"\",\n \"ref_id\": \"b53\",\n \"title\": \"Forecasting future world events with neural networks\",\n \"year\": \"2022\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[],"string":"[]","isConversational":false},"title":{"kind":"string","value":"Mitigating Temporal Misalignment by Discarding Outdated Facts"},"abstract":{"kind":"string","value":"While large language models are able to retain vast amounts of world knowledge seen during pretraining, such knowledge is prone to going out of date and is nontrivial to update. Furthermore, these models are often used under temporal misalignment, tasked with answering questions about the present, despite having only been trained on data collected in the past. To mitigate the effects of temporal misalignment, we propose fact duration prediction: the task of predicting how long a given fact will remain true. In our experiments, we demonstrate that identifying which facts are prone to rapid change can help models avoid reciting outdated information and determine which predictions require seeking out up-to-date knowledge sources. We also show how modeling fact duration improves calibration for knowledgeintensive tasks, such as open-retrieval question answering, under temporal misalignment, by discarding volatile facts."},"authors":{"kind":"string","value":"Michael J Q Zhang; Eunsol Choi"},"figures":{"kind":"list like","value":[{"figure_caption":"FactFigure 1 :1Figure1: We depict the critical timestamps at play in open-retrieval QA systems. In the example on the left, the temporal misalignment between when the system was trained and evaluated has no affect on the answer. On the right, the answer has changed, causing the system to output an outdated answer with high confidence. To account for this, we apply our fact duration prediction system to adjust the system's confidence accordingly.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Duration statistics on each dataset's development set. Columns represent different duration classes used by our classification model, with units abbreviated as Seconds, Minutes, Days, Weeks, Months, Years, Decades, and Centuries. Cells contain the % of examples in each dataset in the column's duration class.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Fact Duration Prediction Results. On the left, we report our full results, with performance split by model type and training data. Performance on SituatedQA and TimeQA are given as the mean average error in years (Y) and mean squared error in years in log-seconds (LS), the same as the regression system training loss. On the right, we depict error histograms evaluated on SituatedQA, with systems trained on TimeQA.","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"NQ-Open QA performance evaluated on answers from 2018 (from NQ-Open) and 2021 (from Situ-","figure_data":"","figure_id":"tab_1","figure_label":"1","figure_type":"table"},{"figure_caption":"Dataset statistics for our QA misalignment calibration and duration prediction tasks. We report the number of examples used in our QA calibration experiments along with how many examples have answers that have changed/unchanged between 2018 and 2021.","figure_data":"","figure_id":"tab_3","figure_label":"2","figure_type":"table"},{"figure_caption":"↓) (Y, ↓) (LS, ↓) (Y, ↓)","figure_data":"SituatedQATimeQAMCTACOModelTraining DataMSE MAEMSE MAEEMF1(LS, (↑)(↑)Random-7.14 13.442.226.96 20.3 38.1Average-5.42 10.611.655.41 23.8 41.3Oracle0.284.180.121.668.0 37.2ClassificationTimeQA TimePre4.20 40.518.45 8.771.55 9.364.80 20.3 41.3 7.85 28.3 57.7Time(QA+Pre)6.288.371.704.66 28.3 57.2TimeQA3.758.400.974.22 21.2 41.0RegressionTimePre38.488.9932.105.70 33.1 57.8Time(QA+Pre)3.588.150.974.19 31.8 57.8","figure_id":"tab_4","figure_label":"","figure_type":"table"},{"figure_caption":"Results for calibrating QA under temporal misalignment on SituatedQA. All systems' training dates are 2018 and evaluation dates are 2021. We report each system's EM accuracy, evaluated against the answers from 2018 and 2021. We also report how much model confidence changes on average (Avg Conf % ∆) with each adjustment method (for DPR with t e = 2021 we compare average confidence against using t e = 2018).","figure_data":"QA Model2018 → 2021 EMDur. ModelAUROC ↑ ECE ↓ RC@55 (|∆| ↓) Avg Conf % ∆N / A0.7660.26527.2 (27.8)0.01 ⃝ T536.0 → 17.4Oracle0.7490.11647.8 (7.2)-25.8Regression0.7090.18532.4 (22.6)-23.2Classification0.7650.13129.3 (25.7)-15.1N / A0.6290.43322.4 (32.6)0.0Oracle0.7080.17238.7 (16.3)-36.92 ⃝ DPR (te = 2018)37.9 → 17.1Regression0.6010.26826.1 (28.9)-34.0Classification0.6540.23543.5 (11.5)-20.9N ⃝ DPR (te = 2021)--→ 19.6N / A0.6360.37025.4 (29.6)-3.8ModelAdj.AUROCECE RC@55 (|∆|)1 ⃝ T5Uniform Per-Ex0.757 0.180 0.765 0.13130.5 (24.5) 29.3 (25.7)2 ⃝ DPRUniform Per-Ex0.627 0.259 0.654 0.23525.6 (29.4) 43.5 (11.5)","figure_id":"tab_5","figure_label":"3","figure_type":"table"},{"figure_caption":"Ablating per-example calibration:","figure_data":"","figure_id":"tab_6","figure_label":"4","figure_type":"table"},{"figure_caption":"Adaptive inference for temporal misalignment results: we use our duration prediction to decide whether to use the prediction from model with newer corpus (DPR t e = 2021). In this data (SituatedQA), 48.8% of examples requires up-to-date knowledge.","figure_data":"Inference EnsembleEM%1 ⃝ T517.40.02 ⃝ DPR te = 201817.10.0N ⃝ DPR te = 202119.6 100.01 ⃝ T5 / N ⃝ DPR te = 202120.545.72 ⃝ DPR te = 2018 / N ⃝ DPR te = 2021 19.345.0","figure_id":"tab_7","figure_label":"5","figure_type":"table"},{"figure_caption":"Who are the judges on Asia Got Talent? / A: Vanness Wu / Start: 2015, End: 2017 MI: Vanness Wu is the judge on Asia Got Talent , lasting [MASK] [MASK] . / TD: 2 Years MCTACO Context: About 30% of Ratners's profit already is derived from the U.S. Q: How long did it take to make profit? / A: 3 Months MI: About 30% of Ratners's profit already is derived from the U.S. It took [MASK] [MASK] to make profit. / TD: 3 Months TimeQA Subj: Patrick Burns (businessman) Rel: Lives in Obj: Oshawa, Ontario Start: 1856 End: 1878 MI: Patrick Burns (businessman) lived in Oshawa, Ontario , lasting [MASK] [MASK] . / TD: 22 years TA-Pretrain MI: Jorge Ramos has been the face of Univision's News broadcast for [MASK] [MASK] . TD: 24 Years","figure_data":"DatasetExampleSituatedQAQ:","figure_id":"tab_8","figure_label":"","figure_type":"table"},{"figure_caption":"Fact duration prediction input examples. We standardize formats to predict target duration (TD) from the masked input (MI). The top row(s) in each cell represents the original data, and the bottom row shows our setting.","figure_data":"Quesiton {2018 Answer / 2021 Answer}{G / P} DurWhen did the last volcano erupt in Iceland? 7 Y / 10 Y(2010 / March 19, 2021)How many episodes of Touching Evil are 1 W / 1 Ythere? (16 / 16)Who got the most passing yards in the3 Y / 1 YNFL? (Peyton Manning / Drew Brees)","figure_id":"tab_9","figure_label":"8","figure_type":"table"},{"figure_caption":"","figure_data":"","figure_id":"tab_10","figure_label":"","figure_type":"table"},{"figure_caption":"Example ChatGPT outputs to two temporally dependent questions with different answer durations. Predictions are take from the May 12, 2023 ChatGPT version.","figure_data":"","figure_id":"tab_11","figure_label":"10","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"FactFigure 1 :1Figure1: We depict the critical timestamps at play in open-retrieval QA systems. In the example on the left, the temporal misalignment between when the system was trained and evaluated has no affect on the answer. On the right, the answer has changed, causing the system to output an outdated answer with high confidence. To account for this, we apply our fact duration prediction system to adjust the system's confidence accordingly.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Duration statistics on each dataset's development set. Columns represent different duration classes used by our classification model, with units abbreviated as Seconds, Minutes, Days, Weeks, Months, Years, Decades, and Centuries. Cells contain the % of examples in each dataset in the column's duration class.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Fact Duration Prediction Results. On the left, we report our full results, with performance split by model type and training data. Performance on SituatedQA and TimeQA are given as the mean average error in years (Y) and mean squared error in years in log-seconds (LS), the same as the regression system training loss. On the right, we depict error histograms evaluated on SituatedQA, with systems trained on TimeQA.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"NQ-Open QA performance evaluated on answers from 2018 (from NQ-Open) and 2021 (from Situ-\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Dataset statistics for our QA misalignment calibration and duration prediction tasks. We report the number of examples used in our QA calibration experiments along with how many examples have answers that have changed/unchanged between 2018 and 2021.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"↓) (Y, ↓) (LS, ↓) (Y, ↓)\",\n \"figure_data\": \"SituatedQATimeQAMCTACOModelTraining DataMSE MAEMSE MAEEMF1(LS, (↑)(↑)Random-7.14 13.442.226.96 20.3 38.1Average-5.42 10.611.655.41 23.8 41.3Oracle0.284.180.121.668.0 37.2ClassificationTimeQA TimePre4.20 40.518.45 8.771.55 9.364.80 20.3 41.3 7.85 28.3 57.7Time(QA+Pre)6.288.371.704.66 28.3 57.2TimeQA3.758.400.974.22 21.2 41.0RegressionTimePre38.488.9932.105.70 33.1 57.8Time(QA+Pre)3.588.150.974.19 31.8 57.8\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Results for calibrating QA under temporal misalignment on SituatedQA. All systems' training dates are 2018 and evaluation dates are 2021. We report each system's EM accuracy, evaluated against the answers from 2018 and 2021. We also report how much model confidence changes on average (Avg Conf % ∆) with each adjustment method (for DPR with t e = 2021 we compare average confidence against using t e = 2018).\",\n \"figure_data\": \"QA Model2018 → 2021 EMDur. ModelAUROC ↑ ECE ↓ RC@55 (|∆| ↓) Avg Conf % ∆N / A0.7660.26527.2 (27.8)0.01 ⃝ T536.0 → 17.4Oracle0.7490.11647.8 (7.2)-25.8Regression0.7090.18532.4 (22.6)-23.2Classification0.7650.13129.3 (25.7)-15.1N / A0.6290.43322.4 (32.6)0.0Oracle0.7080.17238.7 (16.3)-36.92 ⃝ DPR (te = 2018)37.9 → 17.1Regression0.6010.26826.1 (28.9)-34.0Classification0.6540.23543.5 (11.5)-20.9N ⃝ DPR (te = 2021)--→ 19.6N / A0.6360.37025.4 (29.6)-3.8ModelAdj.AUROCECE RC@55 (|∆|)1 ⃝ T5Uniform Per-Ex0.757 0.180 0.765 0.13130.5 (24.5) 29.3 (25.7)2 ⃝ DPRUniform Per-Ex0.627 0.259 0.654 0.23525.6 (29.4) 43.5 (11.5)\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablating per-example calibration:\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Adaptive inference for temporal misalignment results: we use our duration prediction to decide whether to use the prediction from model with newer corpus (DPR t e = 2021). In this data (SituatedQA), 48.8% of examples requires up-to-date knowledge.\",\n \"figure_data\": \"Inference EnsembleEM%1 ⃝ T517.40.02 ⃝ DPR te = 201817.10.0N ⃝ DPR te = 202119.6 100.01 ⃝ T5 / N ⃝ DPR te = 202120.545.72 ⃝ DPR te = 2018 / N ⃝ DPR te = 2021 19.345.0\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Who are the judges on Asia Got Talent? / A: Vanness Wu / Start: 2015, End: 2017 MI: Vanness Wu is the judge on Asia Got Talent , lasting [MASK] [MASK] . / TD: 2 Years MCTACO Context: About 30% of Ratners's profit already is derived from the U.S. Q: How long did it take to make profit? / A: 3 Months MI: About 30% of Ratners's profit already is derived from the U.S. It took [MASK] [MASK] to make profit. / TD: 3 Months TimeQA Subj: Patrick Burns (businessman) Rel: Lives in Obj: Oshawa, Ontario Start: 1856 End: 1878 MI: Patrick Burns (businessman) lived in Oshawa, Ontario , lasting [MASK] [MASK] . / TD: 22 years TA-Pretrain MI: Jorge Ramos has been the face of Univision's News broadcast for [MASK] [MASK] . TD: 24 Years\",\n \"figure_data\": \"DatasetExampleSituatedQAQ:\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Fact duration prediction input examples. We standardize formats to predict target duration (TD) from the masked input (MI). The top row(s) in each cell represents the original data, and the bottom row shows our setting.\",\n \"figure_data\": \"Quesiton {2018 Answer / 2021 Answer}{G / P} DurWhen did the last volcano erupt in Iceland? 7 Y / 10 Y(2010 / March 19, 2021)How many episodes of Touching Evil are 1 W / 1 Ythere? (16 / 16)Who got the most passing yards in the3 Y / 1 YNFL? (Peyton Manning / Drew Brees)\",\n \"figure_id\": \"tab_9\",\n \"figure_label\": \"8\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_10\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Example ChatGPT outputs to two temporally dependent questions with different answer durations. Predictions are take from the May 12, 2023 ChatGPT version.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_11\",\n \"figure_label\": \"10\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(Lazaridou et al., 2021)\", \"Explanation\": \"The cited work highlights the issue of temporal misalignment in deploying NLP systems, which the citing paper addresses by proposing methods to manage the misalignment.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Luu et al., 2021)\", \"Explanation\": \"The cited work provides evidence of the performance degradation caused by temporal misalignment in NLP tasks, which the citing paper further discusses.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Dhingra et al., 2022)\", \"Explanation\": \"The cited work highlights the issue of knowledge-intensive tasks being particularly affected by temporal misalignment, which the citing paper further elaborates on.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Zhang and Choi, 2021)\", \"Explanation\": \"The cited work provides further evidence of the performance degradation caused by temporal misalignment in NLP tasks, which the citing paper builds upon.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Chen et al., 2017)\", \"Explanation\": \"The cited work introduces the concept of open-retrieval question answering (QA), which the citing paper uses as a specific example of a knowledge-intensive task affected by temporal misalignment.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhang et al., 2022)\", \"Explanation\": \"The cited work discusses the shift towards using large pretrained models in NLP systems, which the citing paper further extends by highlighting the challenges and issues associated with this shift.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Chowdhery et al., 2022)\", \"Explanation\": \"The cited work highlights the issues of retraining and outdated facts in large pretrained models, which the citing paper builds upon by discussing the need for better management of temporal misalignment in NLP systems.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Cao et al., 2021)\", \"Explanation\": \"The cited work by Cao et al. provides a method for updating the knowledge stored within the parameters of a pretrained model, which the citing paper builds upon to address the issue of temporal misalignment in answering factual questions.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Mitchell et al., 2022)\", \"Explanation\": \"The cited work by Mitchell et al. also proposes a method for updating the knowledge stored within a pretrained model, which the citing paper uses to address the issue of temporal misalignment in answering factual questions.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Onoe et al., 2023)\", \"Explanation\": \"The cited work by Onoe et al. provides another method for updating the knowledge stored within a pretrained model, which the citing paper uses to address the issue of temporal misalignment in answering factual questions.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Karpukhin et al., 2020)\", \"Explanation\": \"The cited work by Karpukhin et al. proposes a retrieval-based system that utilizes a nonparametric corpus of facts to answer factual questions, which the citing paper builds upon to address the issue of temporal misalignment in answering factual questions.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Guu et al., 2020)\", \"Explanation\": \"The cited work by Guu et al. also proposes a retrieval-based system that utilizes a nonparametric corpus of facts to answer factual questions, which the citing paper builds upon to address the issue of temporal misalignment in answering factual questions.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Lewis et al., 2021)\", \"Explanation\": \"The cited work by Lewis et al. provides another retrieval-based system that utilizes a nonparametric corpus of facts to answer factual questions, which the citing paper builds upon to address the issue of temporal misalignment in answering factual questions.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Chen et al., 2021b)\", \"Explanation\": \"The cited work provides a source of fact durations that the citing paper utilizes in their research on fact duration prediction.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Yang et al., 2020)\", \"Explanation\": \"The cited work is a source of duration-related news text that the citing paper uses in their research on fact duration prediction.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhou et al., 2020)\", \"Explanation\": \"The cited work on temporal commonsense provides a basis for the discussion in the citing paper on the relationship between fact duration prediction and temporal commonsense.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Kwiatkowski et al., 2019)\", \"Explanation\": \"The cited work provides the NaturalQuestions dataset that the citing paper uses in their intrinsic evaluations of fact duration prediction systems.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhang and Choi, 2021)\", \"Explanation\": \"The cited work on the SituatedQA dataset serves as a basis for the citing paper to focus on improving calibration in QA systems and reducing expected calibration error by 50-60%. The study on temporal misalignment in QA is also an extension of the research on fact duration prediction in the cited work.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Luu et al., 2021)\", \"Explanation\": \"The cited work by Luu et al. provides the setting for temporal misalignment in knowledge-intensive tasks, which the citing paper aims to address.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhang and Choi, 2021)\", \"Explanation\": \"The cited work by Zhang and Choi introduces the SituatedQA dataset, which the citing paper uses to compare the performance of existing systems in addressing temporal misalignment in QA tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kamath et al., 2020)\", \"Explanation\": \"The cited work by Kamath et al. provides a method for identifying low-confidence predictions in models, which the citing paper adopts to improve the accuracy of their predictions in the context of temporal misalignment.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Angelopoulos et al., 2022)\", \"Explanation\": \"The cited work by Angelopoulos et al. (2022) is used to set the threshold (\\u03c4) for selective prediction in the citing paper, which is a method for controlling risk in uncertainty estimates.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang and Choi, 2021)\", \"Explanation\": \"The cited work, SituatedQA, is the primary evaluation dataset used in the citing paper for studying temporal shifts in question answering.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Kwiatkowski et al., 2019)\", \"Explanation\": \"The cited work, NQ-Open, provides the question annotations used in the temporally-dependent subset of the evaluation dataset in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Kamath et al., 2020)\", \"Explanation\": \"The cited work by Kamath et al. has explored adjusting model confidence due to changes in input distribution, which the citing paper builds upon to address the issue of calibrating over shifts in output distribution in QA.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Chen et al., 2021a)\", \"Explanation\": \"The cited work by Chen et al. provides a T5-based conversion model for generating fact-duration pairs, which the citing paper further utilizes to evaluate fact duration prediction in the context of QA evaluations on SituatedQA.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhou et al., 2019)\", \"Explanation\": \"The cited work by Zhou et al. provides the event duration subset of MCTACO, which the citing paper uses to evaluate the fact duration prediction system.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Chen et al., 2021a)\", \"Explanation\": \"The cited work is used as a data source to convert the decontextualized QA pairs into factual statements for the study conducted in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Yang et al., 2020)\", \"Explanation\": \"The cited work is extended in the citing paper by curating news texts from CNN and Daily Mail news articles to create a dataset for time-aware pretraining, which is used to improve performance on temporal commonsense tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Yang et al., 2020)\", \"Explanation\": \"The cited work on temporal common sense reasoning provides a framework and methods for fact duration prediction, which the citing paper builds upon in their development of BERT-based models for the task.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Si et al., 2022)\", \"Explanation\": \"The cited work by Si et al. provides a method for calibrating prompted large language models (LLMs) for QA, which the citing paper builds upon to explore the same research area.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kadavath et al., 2022)\", \"Explanation\": \"The cited work by Kadavath et al. also contributes to the calibration of LLMs for QA, providing a methodological basis for the research conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Cole et al., 2023)\", \"Explanation\": \"The cited work by Cole et al. further expands on the calibration of LLMs for QA, providing additional methods and techniques that the citing paper may have adopted or adapted.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(OpenAI, 2023)\", \"Explanation\": \"The cited work by OpenAI introduces the GPT-4 system, which the citing paper uses as a data source for investigating the ability of LLMs to abstain from answering questions with rapidly changing answers.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Mallen et al., 2022)\", \"Explanation\": \"The cited work by Mallen et al. (2022) provides a method for improving answer accuracy by forgoing retrieval for popular questions, which the citing paper adopts in their hybrid approach of using T5 to predict answers and then using fact duration predictions to determine when retrieval is necessary.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Zou et al., 2022)\", \"Explanation\": \"The cited work by Zou et al. provides a benchmark for forecasting tasks related to the fact duration prediction task in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Jin et al., 2021a)\", \"Explanation\": \"The cited work by Jin et al. also contributes to the forecasting benchmark for the fact duration prediction task in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Qin et al., 2021)\", \"Explanation\": \"The cited work by Qin et al. extends the research on temporal commonsense reasoning in dialogues, which the citing paper further explores in the context of fact duration prediction.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Forbes and Choi, 2017)\", \"Explanation\": \"The cited work by Forbes and Choi extends the research on quantitative relations between nouns, which the citing paper further builds upon in the context of fact duration prediction.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Bagherinezhad et al., 2016)\", \"Explanation\": \"The cited work by Bagherinezhad et al. also extends the research on quantitative relations between nouns, which the citing paper further builds upon in the context of fact duration prediction.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Elazar et al., 2019)\", \"Explanation\": \"The cited work by Elazar et al. studies distributions over quantitative attributes, which the citing paper further builds upon in the context of fact duration prediction.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Wallace et al., 2019)\", \"Explanation\": \"The cited work by Wallace et al. explores representing numbers in language models, which the citing paper further builds upon in the context of fact duration prediction.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Chen et al., 2022)\", \"Explanation\": \"The cited work by Chen et al. examines instances where knowledge conflicts exist between a model's memorized knowledge and retrieved documents, which the citing paper further discusses in the context of temporal misalignment.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kamath et al., 2020;Zhang et al., 2021;Varshney and Baral, 2023)\", \"Explanation\": \"The cited works have explored the use of predicting whether or not the test question comes from the same training distribution of the QA system, which the citing paper adopts as a method for addressing robustness to distribution shift in QA systems.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"Dhingra et al., 2022;Jin et al., 2021b\", \"Explanation\": \"The cited works have explored continuing pretraining to address temporal misalignment in pretrained models, which the citing paper extends by exploring the use of this method in addressing the challenge of distribution shift in QA systems.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"Cao et al., 2021;Mitchell et al., 2022;Meng et al., 2022\", \"Explanation\": \"The cited works have explored editing specific facts into models, which the citing paper extends by exploring the use of this method in addressing the challenge of distribution shift in QA systems in synthetic settings.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Onoe et al., 2022)\", \"Explanation\": \"The cited work provides a benchmark for measuring the acquisition of emergent information in language models, which the citing paper uses to assess the performance of their system.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Padmanabhan et al., 2023)\", \"Explanation\": \"The cited work is a new benchmark for measuring the acquisition of emergent information in language models, which the citing paper uses to compare the performance of their system.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Izacard et al., 2022)\", \"Explanation\": \"The cited work on retrieval-based QA systems has found improved adaptation when updated with up-to-date retrieval corpora, which the citing paper extends by applying the same method to their system.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Lazaridou et al., 2022)\", \"Explanation\": \"The cited work on retrieval-based QA systems has found improved adaptation when updated with up-to-date retrieval corpora, which the citing paper extends by applying the same method to their system.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Yang et al., 2020)\", \"Explanation\": \"The cited work provides the metrics used to evaluate the performance of the prepossessing model in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(He et al., 2021)\", \"Explanation\": \"The cited work by He et al. (2021) provides the DeBERTa-v3-base model that the citing paper uses in their fact duration prediction system.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chen and Guestrin, 2016)\", \"Explanation\": \"The cited work by Chen and Guestrin (2016) provides the XGBoost algorithm that the citing paper uses to implement their trained calibration systems for T5 and DPR features.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2021)\", \"Explanation\": \"The cited work by Zhang et al. (2021) serves as the data source for the training of the calibration systems, providing the features and the NQ-Open dataset used in the study.\"}]"}}},{"rowIdx":76,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_0"],"heading":"Introduction","publication_ref":["b10","b1","b18","b26","b11","b14","b22","b6","b3","b17","b25","b6","b17","b22","b10","b3","b17","b19"],"table_ref":[],"text":"Identification of user intentions, a problem known as intent classification (IC), plays an important role in task-oriented dialogue (TOD) systems. However, it is challenging for TOD developers to collect data and re-train models when designing new intent classes. Recent studies have aimed to tackle this challenge by applying zero-and few-shot text classification methods and leveraging the semantics of intent label names (Liu et al., 2019a;Krone et al., 2020;Burnyshev et al., 2021;Mueller et al., 2022;Zhang et al., 2022;Lamanov et al., 2022;Liu et al., 2022) . Dopierre et al. (2021a) compare various classification methods on few-shot IC tasks and find that (PIE). Given an utterance x 1 from pre-training corpus, we generate a pseudo intent name y pseudo 1 using labels from the intent role labeling (IRL) tagger. Our PIE model is then optimized by pulling the gold utterance x gold 1 , gold intent y 1 , and pseudo intent y pseudo 1 close to the input utterance x 1 in the embedding space.\nPrototypical Networks (Snell et al., 2017) consistently show strong performance when combined with transformer-based text encoders. Prototypical Networks use text encoders to construct class representations and retrieve correct classes given queries based on a similarity metric. Dopierre et al. (2021a) also stress that few-shot learning techniques and text encoders can have an orthogonal impact on classification performance. Thus, although some studies have focused on improving learning techniques for few-shot IC tasks (Dopierre et al., 2021b;Chen et al., 2022), better text encoder selection should also be considered as an important research direction. Ma et al. (2022) observe that sentence encoders pre-trained on paraphrase or natural language inference datasets serve as strong text encoders for Prototypical Networks. However, existing sentence encoders are not explicitly designed to produce representations for utterances that are similar to their intent names. Therefore, their abilities are limited in zero-and few-shot settings where predictions may heavily rely on the semantics of intent names. Pre-training encoders to align user utterances with intent names can mitigate this issue; however, it is typically expensive to obtain annotations for a diverse intent set.\nIn this paper, we propose a novel pre-training method for zero-and few-shot IC tasks (Figure 1). Specifically, we adopt intent role labeling (IRL) (Zeng et al., 2021), which is an approach for identifying and assigning roles to words or phrases that are relevant to user intents in sentences. Once we obtain the IRL predictions, we convert them to the pseudo intent names of query utterances and use them to pre-train the encoder in a contrastive learning fashion. This intent-aware contrastive learning aims to not only align utterances with their pseudo intent names in the semantic embedding space, but also to encourage the encoder to pay attention to the intent-relevant spans that are important for distinguishing intents. To the best of our knowledge, this work is the first to extract key information from utterances and use it as pseudo labels for pre-training intent-aware text encoders.\nThe contributions of our work are as follows:\n• First, we propose an algorithm for generating pseudo intent names from utterances across several dialogue datasets and publicly release the associated datasets.\n• Second, by applying intent-aware contrastive learning on gold and pseudo intent names, we build Pre-trained Intent-aware Encoder (PIE), which is designed to align encodings of utterances with their intent names.\n• Finally, experiments on four IC datasets demonstrate that the proposed model outperforms the state-of-the-art work (Dopierre et al., 2021b;Ma et al., 2022) by up to 5.4% and 4.0% on N-way zero-and one-shot settings, respectively.\n2 Background: Prototypical Networks for Intent Classification\nPrototypical Networks (Snell et al., 2017) is a metalearning approach that enables classifiers to quickly adapt to unseen classes when only a few labeled examples are available. Several studies have demonstrated the effectiveness of Prototypical Networks when building intent classifiers with a few example utterances (Krone et al., 2020;Dopierre et al., 2021a;Chen et al., 2022). They first define a fewshot IC task, also known as an episode in the metalearning context, with K example utterances from N intent classes (i.e., K×N utterances in a single episode). At the training time, the intent classifiers are optimized on a series of these episodes. Example utterances for each intent class are called a support set, and are encoded and averaged to produce a class representation, called a prototype. This can be formulated as follows:\nc n = 1 K x n,i ∈Sn f ϕ (x n,i )(1)\nwhere S n denotes the support set of the n-th intent class, x n,i denotes the i-th labeled example of the support set S n , f ϕ (•) denotes a trainable encoder, and c n denotes the n-th prototype. At the inference time, the task is to map the query utterance representation to the closest prototype in a metric space (e.g., Euclidean) among the N prototypes. When there are N intent classes and each intent class has K example utterances, this setting is called N-way K-shot intent classification. Ma et al. (2022) suggest that leveraging intent names as additional support examples is beneficial in few-shot IC tasks because the semantics of intent names can give additional hints to example utterances. When intents are used as additional support examples, the new prototype representations can be formulated as follows:\nc label n = 1 K + 1 [[ x n,i ∈Sn f ϕ (x n,i )] + f ϕ (y n )] (2)\nwhere y n is the intent name of the utterance in the nth support set, and c label n is the n-th prototype using intent names as support. By using intents as support examples, it is possible to classify input utterances without example utterances in a zero-shot fashion. Specifically, the prototypes in Equation ( 2) can be calculated as c label n = f ϕ (y n ) based solely on intent names, which facilitates the zero-shot IC.\nTo pre-train an encoder f ϕ that works robustly in zero-or few-shot IC settings, a variety of predefined intent names are required. Because annotating them is expensive, we opt to automatically generate pseudo intent names from utterances in our pre-training data. To annotate pseudo intents, we employ a tagging method, intent role labeling (IRL). IRL can be considered similar to semantic role labeling (SRL), which is a task of assigning general semantic roles to words or phrases in sentences (Palmer et al., 2010). However, IRL focuses on providing an extractive summary of the intent expressed in a user's utterance, annotating important roles with respect to the goal of the user rather than a predicate. Specifically, it tags words or phrases that are key to interpret intent.\nIRL was first introduced by Zeng et al. ( 2021) for discovering intents from utterances, but their tagger focuses only on Chinese utterances. In this section, we outline the process of building the IRL tagger from scratch. We provide a description of how we annotate IRL training data on English utterances (Section 3.1), the training procedure for the IRL tagger (Section 3.2), and the utilization of IRL predictions for generating pseudo intent names to pre-train our model (Section 3.3)."},{"figure_ref":[],"heading":"Annotating Intent Roles","publication_ref":["b20"],"table_ref":[],"text":"We define six intent role labels, Action, Argument, Request, Query, Slot, and Problem, for extracting intent-relevant spans from utterances. Action is a word or phrase (typically a verb or verb phrase) that describes the main action relevant to an intent in an utterance. Argument is an argument of an action, or entity/event that is important to interpreting an intent. Request indicates a request for something such as a question or information-seeking verb. Query indicates the expected type of answer to a question or request for information, or a requested entity to be obtained or searched for. Slot is an optional/variable value provided by the speaker that does not impact the interpretation of an intent. Finally, Problem describes some problematic states or events, and typically makes an implicit request.\nBased on these definitions, we manually annotate IRL labels on a subset of utterances from SGD (Rastogi et al., 2020). "},{"figure_ref":[],"heading":"Training the IRL Tagger","publication_ref":[],"table_ref":[],"text":"Using manually curated IRL annotations, we formulate IRL as a sequence tagging problem. Specifically, we assign each token in an utterance with one of the 13 IRL labels under the Beginning-Inside-Outside (BIO) scheme (e.g., B-Action, I-Action, or O). The IRL model is, then, trained to predict the correct IRL labels of the tokens using the cross entropy loss. We use RoBERTa-base (Liu et al., 2019c) as the initial model for the IRL tagger. "},{"figure_ref":[],"heading":"Generating Pseudo Intents","publication_ref":["b23"],"table_ref":[],"text":"After obtaining the IRL tagger, we leverage it to predict IRL labels for tokens in utterances from pre-training corpus described in Section 5.2. To generate pseudo intent names, we simply concatenate all spans that have been predicted as IRL labels in each utterance. 4 Intent-Aware Contrastive Learning\nWe aim to build an encoder that produces similar representations between utterances and the corresponding intent names. In this section, we introduce the intent-aware contrastive learning approach using triples of an utterance, gold intent, and pseudo intent from various dialogue datasets.\nOur training objective is designed to align the representations of utterances and their intent names in the semantic embedding space. For this purpose, we use the InfoNCE loss (van den Oord et al., 2018), which pulls positive pairs close to each other and pushes away negative pairs. The loss for the i-th sample x i is formulated as follows:\nℓ(x i , y) = exp sim(f ϕ (x i ), f ϕ (y i )) N k exp sim(f ϕ (x i ), f ϕ (y k )) ,(3)\nwhere y = ⟨y 1 , y 2 , . . . , y N ⟩ are pairs of the input x i with a batch size of N , and sim(•) denotes the cosine similarity between two embeddings. Again, f ϕ (•) denotes any text encoder that represents intent names or utterances in the embedding space. Note that pairs that are not positive in a batch are treated as negative pairs. We here define three types of positive pairs, two of which are supervised and one is semi-supervised. The first of the supervised positive pairs is between the input utterances and their gold intent names annotated in the pre-training datasets. The equation used is as follows:\nL gold_intent = - 1 N N i ℓ(x i , y gold ),(4)\nwhere\ny gold i\nis the a gold intent name of x i .\nThe second supervised positive pair is between the input utterances and their gold utterances. We define gold utterances as randomly sampled utterances that share the same gold intent names as the input utterances:\nL gold_utterance = - 1 N N i ℓ(x i , x gold ),(5)\nwhere\nx gold i\nis the gold utterance of x i . Finally, the semi-supervised positive pairs are between the input utterances and their pseudo intent names:\nL pseudo = - 1 N N i ℓ(x i , y pseudo ),(6)\nwhere y pseudo i denotes the pseudo intent name of x i constructed by the IRL tagger, as described in Section 3.3\nOur final loss is a combination of these three losses as follows:\nL = L gold_intent + L gold_utterance + λL pseudo , (7)\nwhere λ is the weight term of the semi-supervised loss term."},{"figure_ref":[],"heading":"Experiment","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Baselines","publication_ref":["b6","b17","b9","b21","b17","b8","b4","b7","b20","b24"],"table_ref":["tab_4","tab_6"],"text":"To evaluate the effectiveness of our proposed PIE model, we compare it with the following stateof-the-art approaches for few-shot IC tasks: Pro-toNet (Dopierre et al., 2021a) and ProtAugment (Dopierre et al., 2021b) as fine-tuning methods, and SBERT Paraphrase (Ma et al., 2022) as a pre-trained text encoder.\nProtoNet is a meta-training approach that finetunes encoders by using a series of episodes constructed on task-specific training sets. ProtAugment is an advanced method derived from Pro-toNet, which augments paraphrased utterances within episodes to mitigate overfitting caused by the biased distribution introduced by a limited number of training examples. The authors of ProtoNet and ProtAugment perform additional pre-training BERT-base-cased (110M) using training utterances and the language model objective, and use it as their initial model. We refer to this model as BERT TAPT , inspired by task-adaptive pre-training (TAPT) (Gururangan et al., 2020).\nSBERT Paraphrase is a text encoder pre-trained on large-scale paraphrase text pairs (Reimers and Gurevych, 2019). Ma et al. (2022) discover that this pre-trained text encoder can produce good utterance embeddings without any fine-tuning on task-specific datasets. Although the authors leverage SBERT Paraphrase solely at the inference stage of Prototypical Networks, we conduct additional experiments by fine-tuning the encoder using Pro-toNet and ProtAugment as baselines. Note that we reproduce the performance of SBERT Paraphrase using paraphrase-mpnet-base-v2 2 (110M), which has the same number of parameters as BERT-basecased, for a fair comparison. We collect four dialogue datasets to pre-train and one to validate our encoder: TOP (Gupta et al., 2018), TOPv2 (Chen et al., 2020), DSTC11-T2 (Gung et al., 2023), SGD (Rastogi et al., 2020), and MultiWOZ 2.2 (Zang et al., 2020). Dialogues in SGD and MultiWOZ 2.2 datasets consist of multi-turn utterances, and these utterances are often ambiguous when context around them is not given (e.g., 'Can you suggest something else?' is labeled as 'LookupMusic'). To minimize this am-2 https://huggingface.co/sentencetransformers/paraphrase-mpnet-base-v2 biguity, we use the first-turn utterance of each dialogue from these datasets. Furthermore, the number of utterances between intents in the raw datasets is highly imbalanced. To alleviate this imbalance, we set the maximum number of utterances per intent of the TOP and DSTC11-T2 datasets to 1000 and the SGD and MultiWOZ 2.2 datasets to 100. We then annotate the IRL labels on the utterances using the IRL tagger. Based on the IRL predictions, we filter utterances when no Action, Argument, or Query labels are detected, because they are likely to lack information for interpreting user intents. Finally, we treat MultiWOZ 2.2 as the validation set for tuning the hyperparameters of the pre-training stage. Table 4 summarizes the statistics of the datasets. We evaluate our PIE model and baseline models on four IC datasets (Table 5)."},{"figure_ref":[],"heading":"Pre-training Datasets","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Downstream Datasets","publication_ref":["b2","b12"],"table_ref":["tab_7"],"text":"Banking77 (Casanueva et al., 2020) is an IC dataset in the banking domain. As there are many overlapping tokens between intent names (e.g. 'verify top up', 'top up limits', and 'pending top up'), fine-grained understanding is required when correctly classifying intents for this dataset. HWU64 (Liu et al., 2019b) is a dataset in 21 different domains, such as alarm and calendar for a home assistant robot. Liu54 (Liu et al., 2019b) is a dataset collected from Amazon Mechanical Turk, and workers designed utterances for given intents. Clinc150 (Larson et al., 2019) is a dataset that includes a wide range of intents from ten different domains such as 'small talk' and 'travel'. Before proceeding, we examine the number of intent names that overlap between the pre-training data and downstream data (Table 6). When comparing the intent names, we first apply stemming ('restaurant reservation' → 'restaur reserve') and arrange the tokens in alphabetical order ('restaur reserv' → 'reserv restaur') for each intent name. This approach aims to maximize the recall of overlapping intent names. Consequently, we find that only 11 out of 345 intent names from the downstream data overlaps (e.g, 'reserve restaurant' intent in the pre-training data and 'restaurant reservation' intent in the downstream data)."},{"figure_ref":[],"heading":"Implementation Details","publication_ref":[],"table_ref":[],"text":"Here, we describe detailed information when pretraining our PIE model and employing it for zeroand few-shot IC tasks.\nWe use paraphrase-mpnet-base-v2, the same encoder used in the SBERT Paraphrase baseline, as an initial model for further pre-training in our approach. The hyperparameters are tuned based on the validation set described in Section 5.2. As a result, we set the training epochs to 1, the learning rate to 1e-6, the batch size to 50, and λ to 2.\nAfter pre-training, we apply the model to zeroand few-shot IC tasks in the 5-way and N-way settings. The baselines we compare with only experiment with a 5-way setting where the task is predicting the correct intent from among five candidate classes. We further include an N-way setting, where N can be much larger than five because, in practice, it is often required to assign more than five intents in building TOD systems. When evaluating models on the N-way setting, we use all the intent classes in the test set as candidate intents, for example, 27-way for Banking77 and 50-way for Clinc150. We set K, which is the number of examples per intent in an episode, to 0 and 1 to experiment with zero-and one-shot IC. Finally, we treat intent labels as examples when creating prototypes for each intent. This enables experiments in the zero-shot setting and enhances performance in the few-shot setting. To denote the usage of labels as examples, we append 'L-' prefixes to the method names (e.g., L-PIE). 7 shows 5-way K-shot IC performance on four test sets. The results demonstrate that PIE achieves an average accuracy of 88.5%, surpassing SBERT Paraphrase , which is considered the strongest baseline model, by 2.8% in the one-shot setting. This highlights the effectiveness of our pre-training strategy. Additionally, where intent labels are used as examples, our L-PIE model achieves 89.1% and 93.7% in the zero-shot and one-shot settings, respectively, consistently outperforming L-SBERT Paraphrase by 2.9% and 1.8%. It is worth noting that the L-PIE model also significantly outperforms L-BERT TAPT + ProtoNet, which fine-tunes an encoder on the target datasets, by a substantial margin of 4.0% and 2.9%. This shows that our proposed approach builds an effective intent classifier that performs well even prior to fine-tuning on task-specific data. Our L-PIE model shows further improvement when fine-tuned with Pro-tAugment, outperforming the strongest baseline L-SBERT Paraphrase + ProtAugment by 1.0% in zeroshot IC."},{"figure_ref":[],"heading":"Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"5-way K-shot intent classification Table","publication_ref":[],"table_ref":["tab_0"],"text":"Table 8: N-way K-shot intent classification performance of pre-trained models with and without fine-tuning on four test set. Averaged accuracies and standard deviations across five class splits are reported. The 'L-' prefixes indicate the use of intent label names when creating prototypes. Highest scores are boldfaced. Fine-tuning is done in the 5-way setting due to memory constraints.3 N-way K-shot intent classification We showcase the performance of our PIE model and the baselines in a more challenging and practical scenario (Table 8). In this scenario, the intent for user utterances needs to be classified among a significantly larger number of intent classes (e.g., 10× for Clinc150). The results show that our L-PIE model achieves 73.3% and 82.0% in zero-and one-shot settings, respectively, outperforming the baseline L-SBERT Paraphrase by 5.4% and 4.0%. These performance improvements are significantly higher than those observed in the 5-way K-shot IC task. This indicates that our PIE model performs well in practical scenarios, as stated above. We leverage dialogue datasets for building the PIE model as described in Section 5.2. Here, we perform an ablation study over the pre-training datasets on N-way K-shot IC tasks (Table 9). The result shows that using the TOP (+TOPv2) dataset, which has 31K utterances, 61 gold intents, and 23K pseudo intents, improves the performance the most over L-SBERT Paraphrase (indicated as None in Table 9). Specifically, there is an improvement of 4.8% and 3.7% in the zero-and one-shot settings, respectively. Although using other datasets, such as SGD or DSTC11-T2, does not improve the performance in comparison with using the TOP dataset, we observe that merging them further improves the overall performance on downstream tasks. As described in Section 4, our intent-aware contrastive loss comprise three sub-losses, L gold_intent , L gold_utterance , and L pseudo . To see the benefit of using these losses during pre-training, we ablate each loss function in the N-way K-shot IC tasks on HWU64 and Clinc150 (Table 10). The results indicate that two sub-losses, L gold_intent and L gold_utterance show relatively marignal improvements. However, it is noteworthy that L pseudo serves as the key sub-loss for PIE, highlighting the effectiveness of using pseudo intents. Specifically, removing L pseudo from the final loss results in up to 1.8% and 1.1% degradation in performance in the zero-and one-shot settings, respectively."},{"figure_ref":[],"heading":"Analysis","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_2","fig_3"],"heading":"Varying K and N","publication_ref":[],"table_ref":["tab_7","tab_0"],"text":"We visualize the performance of the PIE model in challenging N-way K-shot IC task settings where the number of example utterances K or the number of candidate intent classes N varies. Plots of performance at varying K (Figure 2) show that our model has consistently higher performance than the baselines, and the performance improvement of our model is the largest when K is small (e.g., K=0). Plots of performance at varying N (Figure 3) show that the performance improvement of our model increases as the number of intents N increases (i.e. increasing from N=5 to N=50). These visualizations reveal that the PIE model can be utilized in more practical and realistic settings where many user intents are used for the TOD system and only a few utterances are available. As shown in Table 6, there are a few overlapping intent names between pre-training data and downstream data (except for Liu54). These coincidental overlaps can hinder an accurate evaluation of the generalization ability of our model. To understand the impact of intent overlaps, we also measure the performance using only non-overlapping intents (Table 11). We observe that the impact is marginal enough to be neglected, and surprisingly, removing overlapping intents rather can lead to better performance on Banking77 and HWU64. Through further analysis, we discover that this is partly because of the bias towards pairs of utterances and intent annotated in pre-training datasets. For example, an utterance 'please play my favorite song' in pre-training data has an intent 'play music'. Our model then incorrectly predicted 'play music' for a test utterance 'that song is my favorite', where the correct intent is 'music likeness'."},{"figure_ref":[],"heading":"Impact of Overlapping Intents","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Conclusions","publication_ref":[],"table_ref":[],"text":"In this work, we propose a pre-training method that leverages pseudo intent names constructed using an IRL tagger in a semi-supervised manner, followed by intent-aware pre-training (PIE). Experiments on four intent classification datasets show that our model achieves state-of-the-art performance on all datasets, outperforming the strongest sentence encoder baseline by up to 5.4% and 4.0% in N-way zero-and one-shot settings, respectively. Our analysis shows that PIE performs robustly compared to the baselines in challenging and practical settings with a large number of classes and small number of support examples. In future work, we will explore the use of IRL and our PIE model in multi-label intent classification or out-of-scope detection tasks."},{"figure_ref":[],"heading":"Limitations","publication_ref":[],"table_ref":[],"text":"One limitation of our method is that while it leverages annotations from the IRL tagger, the detection of spans for certain labels, such as 'Problem,' is not accurate enough (38.1% F1 score). This is likely due to the relatively short number of annotations of this type in the training set (45 annotations). To mitigate this limitation, we could consider annotating more instances of this label or implementing techniques for handling imbalanced labels.\nAnother limitation is that we currently treat all IRL labels equally when constructing pseudo intents. However, the importance of each label in interpreting intent can vary. To address this, we plan to investigate treating different labels differently when pre-training the encoder (e.g. by giving more weight to 'Action' and 'Argument' labels and less weight to 'Slot' labels)."},{"figure_ref":[],"heading":"Ethics Statement","publication_ref":[],"table_ref":[],"text":"Our proposed method is for enhancing zero-and few-shot intent classification, and it does not raise any ethical concerns. We believe that this research has valuable merits that can lead to more reliable task-oriented dialogue systems. All experiments in this study were carried out using publicly available datasets."},{"figure_ref":[],"heading":"Acknowledgements","publication_ref":[],"table_ref":[],"text":"We gratefully acknowledge the members of AWS AI for providing valuable feedback on the project. We would also like to thank the anonymous reviewers for their insightful comments. The part of Mujeen Sung's graduate study and, accordingly, this work was supported by National Research Foundation of Korea (NRF-2023R1A2C3004176)."},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"* Work performed during an internship at AWS AI Labs."}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b10\",\n \"b1\",\n \"b18\",\n \"b26\",\n \"b11\",\n \"b14\",\n \"b22\",\n \"b6\",\n \"b3\",\n \"b17\",\n \"b25\",\n \"b6\",\n \"b17\",\n \"b22\",\n \"b10\",\n \"b3\",\n \"b17\",\n \"b19\"\n ],\n \"table_ref\": [],\n \"text\": \"Identification of user intentions, a problem known as intent classification (IC), plays an important role in task-oriented dialogue (TOD) systems. However, it is challenging for TOD developers to collect data and re-train models when designing new intent classes. Recent studies have aimed to tackle this challenge by applying zero-and few-shot text classification methods and leveraging the semantics of intent label names (Liu et al., 2019a;Krone et al., 2020;Burnyshev et al., 2021;Mueller et al., 2022;Zhang et al., 2022;Lamanov et al., 2022;Liu et al., 2022) . Dopierre et al. (2021a) compare various classification methods on few-shot IC tasks and find that (PIE). Given an utterance x 1 from pre-training corpus, we generate a pseudo intent name y pseudo 1 using labels from the intent role labeling (IRL) tagger. Our PIE model is then optimized by pulling the gold utterance x gold 1 , gold intent y 1 , and pseudo intent y pseudo 1 close to the input utterance x 1 in the embedding space.\\nPrototypical Networks (Snell et al., 2017) consistently show strong performance when combined with transformer-based text encoders. Prototypical Networks use text encoders to construct class representations and retrieve correct classes given queries based on a similarity metric. Dopierre et al. (2021a) also stress that few-shot learning techniques and text encoders can have an orthogonal impact on classification performance. Thus, although some studies have focused on improving learning techniques for few-shot IC tasks (Dopierre et al., 2021b;Chen et al., 2022), better text encoder selection should also be considered as an important research direction. Ma et al. (2022) observe that sentence encoders pre-trained on paraphrase or natural language inference datasets serve as strong text encoders for Prototypical Networks. However, existing sentence encoders are not explicitly designed to produce representations for utterances that are similar to their intent names. Therefore, their abilities are limited in zero-and few-shot settings where predictions may heavily rely on the semantics of intent names. Pre-training encoders to align user utterances with intent names can mitigate this issue; however, it is typically expensive to obtain annotations for a diverse intent set.\\nIn this paper, we propose a novel pre-training method for zero-and few-shot IC tasks (Figure 1). Specifically, we adopt intent role labeling (IRL) (Zeng et al., 2021), which is an approach for identifying and assigning roles to words or phrases that are relevant to user intents in sentences. Once we obtain the IRL predictions, we convert them to the pseudo intent names of query utterances and use them to pre-train the encoder in a contrastive learning fashion. This intent-aware contrastive learning aims to not only align utterances with their pseudo intent names in the semantic embedding space, but also to encourage the encoder to pay attention to the intent-relevant spans that are important for distinguishing intents. To the best of our knowledge, this work is the first to extract key information from utterances and use it as pseudo labels for pre-training intent-aware text encoders.\\nThe contributions of our work are as follows:\\n• First, we propose an algorithm for generating pseudo intent names from utterances across several dialogue datasets and publicly release the associated datasets.\\n• Second, by applying intent-aware contrastive learning on gold and pseudo intent names, we build Pre-trained Intent-aware Encoder (PIE), which is designed to align encodings of utterances with their intent names.\\n• Finally, experiments on four IC datasets demonstrate that the proposed model outperforms the state-of-the-art work (Dopierre et al., 2021b;Ma et al., 2022) by up to 5.4% and 4.0% on N-way zero-and one-shot settings, respectively.\\n2 Background: Prototypical Networks for Intent Classification\\nPrototypical Networks (Snell et al., 2017) is a metalearning approach that enables classifiers to quickly adapt to unseen classes when only a few labeled examples are available. Several studies have demonstrated the effectiveness of Prototypical Networks when building intent classifiers with a few example utterances (Krone et al., 2020;Dopierre et al., 2021a;Chen et al., 2022). They first define a fewshot IC task, also known as an episode in the metalearning context, with K example utterances from N intent classes (i.e., K×N utterances in a single episode). At the training time, the intent classifiers are optimized on a series of these episodes. Example utterances for each intent class are called a support set, and are encoded and averaged to produce a class representation, called a prototype. This can be formulated as follows:\\nc n = 1 K x n,i ∈Sn f ϕ (x n,i )(1)\\nwhere S n denotes the support set of the n-th intent class, x n,i denotes the i-th labeled example of the support set S n , f ϕ (•) denotes a trainable encoder, and c n denotes the n-th prototype. At the inference time, the task is to map the query utterance representation to the closest prototype in a metric space (e.g., Euclidean) among the N prototypes. When there are N intent classes and each intent class has K example utterances, this setting is called N-way K-shot intent classification. Ma et al. (2022) suggest that leveraging intent names as additional support examples is beneficial in few-shot IC tasks because the semantics of intent names can give additional hints to example utterances. When intents are used as additional support examples, the new prototype representations can be formulated as follows:\\nc label n = 1 K + 1 [[ x n,i ∈Sn f ϕ (x n,i )] + f ϕ (y n )] (2)\\nwhere y n is the intent name of the utterance in the nth support set, and c label n is the n-th prototype using intent names as support. By using intents as support examples, it is possible to classify input utterances without example utterances in a zero-shot fashion. Specifically, the prototypes in Equation ( 2) can be calculated as c label n = f ϕ (y n ) based solely on intent names, which facilitates the zero-shot IC.\\nTo pre-train an encoder f ϕ that works robustly in zero-or few-shot IC settings, a variety of predefined intent names are required. Because annotating them is expensive, we opt to automatically generate pseudo intent names from utterances in our pre-training data. To annotate pseudo intents, we employ a tagging method, intent role labeling (IRL). IRL can be considered similar to semantic role labeling (SRL), which is a task of assigning general semantic roles to words or phrases in sentences (Palmer et al., 2010). However, IRL focuses on providing an extractive summary of the intent expressed in a user's utterance, annotating important roles with respect to the goal of the user rather than a predicate. Specifically, it tags words or phrases that are key to interpret intent.\\nIRL was first introduced by Zeng et al. ( 2021) for discovering intents from utterances, but their tagger focuses only on Chinese utterances. In this section, we outline the process of building the IRL tagger from scratch. We provide a description of how we annotate IRL training data on English utterances (Section 3.1), the training procedure for the IRL tagger (Section 3.2), and the utilization of IRL predictions for generating pseudo intent names to pre-train our model (Section 3.3).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Annotating Intent Roles\",\n \"publication_ref\": [\n \"b20\"\n ],\n \"table_ref\": [],\n \"text\": \"We define six intent role labels, Action, Argument, Request, Query, Slot, and Problem, for extracting intent-relevant spans from utterances. Action is a word or phrase (typically a verb or verb phrase) that describes the main action relevant to an intent in an utterance. Argument is an argument of an action, or entity/event that is important to interpreting an intent. Request indicates a request for something such as a question or information-seeking verb. Query indicates the expected type of answer to a question or request for information, or a requested entity to be obtained or searched for. Slot is an optional/variable value provided by the speaker that does not impact the interpretation of an intent. Finally, Problem describes some problematic states or events, and typically makes an implicit request.\\nBased on these definitions, we manually annotate IRL labels on a subset of utterances from SGD (Rastogi et al., 2020). \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Training the IRL Tagger\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Using manually curated IRL annotations, we formulate IRL as a sequence tagging problem. Specifically, we assign each token in an utterance with one of the 13 IRL labels under the Beginning-Inside-Outside (BIO) scheme (e.g., B-Action, I-Action, or O). The IRL model is, then, trained to predict the correct IRL labels of the tokens using the cross entropy loss. We use RoBERTa-base (Liu et al., 2019c) as the initial model for the IRL tagger. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Generating Pseudo Intents\",\n \"publication_ref\": [\n \"b23\"\n ],\n \"table_ref\": [],\n \"text\": \"After obtaining the IRL tagger, we leverage it to predict IRL labels for tokens in utterances from pre-training corpus described in Section 5.2. To generate pseudo intent names, we simply concatenate all spans that have been predicted as IRL labels in each utterance. 4 Intent-Aware Contrastive Learning\\nWe aim to build an encoder that produces similar representations between utterances and the corresponding intent names. In this section, we introduce the intent-aware contrastive learning approach using triples of an utterance, gold intent, and pseudo intent from various dialogue datasets.\\nOur training objective is designed to align the representations of utterances and their intent names in the semantic embedding space. For this purpose, we use the InfoNCE loss (van den Oord et al., 2018), which pulls positive pairs close to each other and pushes away negative pairs. The loss for the i-th sample x i is formulated as follows:\\nℓ(x i , y) = exp sim(f ϕ (x i ), f ϕ (y i )) N k exp sim(f ϕ (x i ), f ϕ (y k )) ,(3)\\nwhere y = ⟨y 1 , y 2 , . . . , y N ⟩ are pairs of the input x i with a batch size of N , and sim(•) denotes the cosine similarity between two embeddings. Again, f ϕ (•) denotes any text encoder that represents intent names or utterances in the embedding space. Note that pairs that are not positive in a batch are treated as negative pairs. We here define three types of positive pairs, two of which are supervised and one is semi-supervised. The first of the supervised positive pairs is between the input utterances and their gold intent names annotated in the pre-training datasets. The equation used is as follows:\\nL gold_intent = - 1 N N i ℓ(x i , y gold ),(4)\\nwhere\\ny gold i\\nis the a gold intent name of x i .\\nThe second supervised positive pair is between the input utterances and their gold utterances. We define gold utterances as randomly sampled utterances that share the same gold intent names as the input utterances:\\nL gold_utterance = - 1 N N i ℓ(x i , x gold ),(5)\\nwhere\\nx gold i\\nis the gold utterance of x i . Finally, the semi-supervised positive pairs are between the input utterances and their pseudo intent names:\\nL pseudo = - 1 N N i ℓ(x i , y pseudo ),(6)\\nwhere y pseudo i denotes the pseudo intent name of x i constructed by the IRL tagger, as described in Section 3.3\\nOur final loss is a combination of these three losses as follows:\\nL = L gold_intent + L gold_utterance + λL pseudo , (7)\\nwhere λ is the weight term of the semi-supervised loss term.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiment\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Baselines\",\n \"publication_ref\": [\n \"b6\",\n \"b17\",\n \"b9\",\n \"b21\",\n \"b17\",\n \"b8\",\n \"b4\",\n \"b7\",\n \"b20\",\n \"b24\"\n ],\n \"table_ref\": [\n \"tab_4\",\n \"tab_6\"\n ],\n \"text\": \"To evaluate the effectiveness of our proposed PIE model, we compare it with the following stateof-the-art approaches for few-shot IC tasks: Pro-toNet (Dopierre et al., 2021a) and ProtAugment (Dopierre et al., 2021b) as fine-tuning methods, and SBERT Paraphrase (Ma et al., 2022) as a pre-trained text encoder.\\nProtoNet is a meta-training approach that finetunes encoders by using a series of episodes constructed on task-specific training sets. ProtAugment is an advanced method derived from Pro-toNet, which augments paraphrased utterances within episodes to mitigate overfitting caused by the biased distribution introduced by a limited number of training examples. The authors of ProtoNet and ProtAugment perform additional pre-training BERT-base-cased (110M) using training utterances and the language model objective, and use it as their initial model. We refer to this model as BERT TAPT , inspired by task-adaptive pre-training (TAPT) (Gururangan et al., 2020).\\nSBERT Paraphrase is a text encoder pre-trained on large-scale paraphrase text pairs (Reimers and Gurevych, 2019). Ma et al. (2022) discover that this pre-trained text encoder can produce good utterance embeddings without any fine-tuning on task-specific datasets. Although the authors leverage SBERT Paraphrase solely at the inference stage of Prototypical Networks, we conduct additional experiments by fine-tuning the encoder using Pro-toNet and ProtAugment as baselines. Note that we reproduce the performance of SBERT Paraphrase using paraphrase-mpnet-base-v2 2 (110M), which has the same number of parameters as BERT-basecased, for a fair comparison. We collect four dialogue datasets to pre-train and one to validate our encoder: TOP (Gupta et al., 2018), TOPv2 (Chen et al., 2020), DSTC11-T2 (Gung et al., 2023), SGD (Rastogi et al., 2020), and MultiWOZ 2.2 (Zang et al., 2020). Dialogues in SGD and MultiWOZ 2.2 datasets consist of multi-turn utterances, and these utterances are often ambiguous when context around them is not given (e.g., 'Can you suggest something else?' is labeled as 'LookupMusic'). To minimize this am-2 https://huggingface.co/sentencetransformers/paraphrase-mpnet-base-v2 biguity, we use the first-turn utterance of each dialogue from these datasets. Furthermore, the number of utterances between intents in the raw datasets is highly imbalanced. To alleviate this imbalance, we set the maximum number of utterances per intent of the TOP and DSTC11-T2 datasets to 1000 and the SGD and MultiWOZ 2.2 datasets to 100. We then annotate the IRL labels on the utterances using the IRL tagger. Based on the IRL predictions, we filter utterances when no Action, Argument, or Query labels are detected, because they are likely to lack information for interpreting user intents. Finally, we treat MultiWOZ 2.2 as the validation set for tuning the hyperparameters of the pre-training stage. Table 4 summarizes the statistics of the datasets. We evaluate our PIE model and baseline models on four IC datasets (Table 5).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Pre-training Datasets\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Downstream Datasets\",\n \"publication_ref\": [\n \"b2\",\n \"b12\"\n ],\n \"table_ref\": [\n \"tab_7\"\n ],\n \"text\": \"Banking77 (Casanueva et al., 2020) is an IC dataset in the banking domain. As there are many overlapping tokens between intent names (e.g. 'verify top up', 'top up limits', and 'pending top up'), fine-grained understanding is required when correctly classifying intents for this dataset. HWU64 (Liu et al., 2019b) is a dataset in 21 different domains, such as alarm and calendar for a home assistant robot. Liu54 (Liu et al., 2019b) is a dataset collected from Amazon Mechanical Turk, and workers designed utterances for given intents. Clinc150 (Larson et al., 2019) is a dataset that includes a wide range of intents from ten different domains such as 'small talk' and 'travel'. Before proceeding, we examine the number of intent names that overlap between the pre-training data and downstream data (Table 6). When comparing the intent names, we first apply stemming ('restaurant reservation' → 'restaur reserve') and arrange the tokens in alphabetical order ('restaur reserv' → 'reserv restaur') for each intent name. This approach aims to maximize the recall of overlapping intent names. Consequently, we find that only 11 out of 345 intent names from the downstream data overlaps (e.g, 'reserve restaurant' intent in the pre-training data and 'restaurant reservation' intent in the downstream data).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Implementation Details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Here, we describe detailed information when pretraining our PIE model and employing it for zeroand few-shot IC tasks.\\nWe use paraphrase-mpnet-base-v2, the same encoder used in the SBERT Paraphrase baseline, as an initial model for further pre-training in our approach. The hyperparameters are tuned based on the validation set described in Section 5.2. As a result, we set the training epochs to 1, the learning rate to 1e-6, the batch size to 50, and λ to 2.\\nAfter pre-training, we apply the model to zeroand few-shot IC tasks in the 5-way and N-way settings. The baselines we compare with only experiment with a 5-way setting where the task is predicting the correct intent from among five candidate classes. We further include an N-way setting, where N can be much larger than five because, in practice, it is often required to assign more than five intents in building TOD systems. When evaluating models on the N-way setting, we use all the intent classes in the test set as candidate intents, for example, 27-way for Banking77 and 50-way for Clinc150. We set K, which is the number of examples per intent in an episode, to 0 and 1 to experiment with zero-and one-shot IC. Finally, we treat intent labels as examples when creating prototypes for each intent. This enables experiments in the zero-shot setting and enhances performance in the few-shot setting. To denote the usage of labels as examples, we append 'L-' prefixes to the method names (e.g., L-PIE). 7 shows 5-way K-shot IC performance on four test sets. The results demonstrate that PIE achieves an average accuracy of 88.5%, surpassing SBERT Paraphrase , which is considered the strongest baseline model, by 2.8% in the one-shot setting. This highlights the effectiveness of our pre-training strategy. Additionally, where intent labels are used as examples, our L-PIE model achieves 89.1% and 93.7% in the zero-shot and one-shot settings, respectively, consistently outperforming L-SBERT Paraphrase by 2.9% and 1.8%. It is worth noting that the L-PIE model also significantly outperforms L-BERT TAPT + ProtoNet, which fine-tunes an encoder on the target datasets, by a substantial margin of 4.0% and 2.9%. This shows that our proposed approach builds an effective intent classifier that performs well even prior to fine-tuning on task-specific data. Our L-PIE model shows further improvement when fine-tuned with Pro-tAugment, outperforming the strongest baseline L-SBERT Paraphrase + ProtAugment by 1.0% in zeroshot IC.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"5-way K-shot intent classification Table\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_0\"\n ],\n \"text\": \"Table 8: N-way K-shot intent classification performance of pre-trained models with and without fine-tuning on four test set. Averaged accuracies and standard deviations across five class splits are reported. The 'L-' prefixes indicate the use of intent label names when creating prototypes. Highest scores are boldfaced. Fine-tuning is done in the 5-way setting due to memory constraints.3 N-way K-shot intent classification We showcase the performance of our PIE model and the baselines in a more challenging and practical scenario (Table 8). In this scenario, the intent for user utterances needs to be classified among a significantly larger number of intent classes (e.g., 10× for Clinc150). The results show that our L-PIE model achieves 73.3% and 82.0% in zero-and one-shot settings, respectively, outperforming the baseline L-SBERT Paraphrase by 5.4% and 4.0%. These performance improvements are significantly higher than those observed in the 5-way K-shot IC task. This indicates that our PIE model performs well in practical scenarios, as stated above. We leverage dialogue datasets for building the PIE model as described in Section 5.2. Here, we perform an ablation study over the pre-training datasets on N-way K-shot IC tasks (Table 9). The result shows that using the TOP (+TOPv2) dataset, which has 31K utterances, 61 gold intents, and 23K pseudo intents, improves the performance the most over L-SBERT Paraphrase (indicated as None in Table 9). Specifically, there is an improvement of 4.8% and 3.7% in the zero-and one-shot settings, respectively. Although using other datasets, such as SGD or DSTC11-T2, does not improve the performance in comparison with using the TOP dataset, we observe that merging them further improves the overall performance on downstream tasks. As described in Section 4, our intent-aware contrastive loss comprise three sub-losses, L gold_intent , L gold_utterance , and L pseudo . To see the benefit of using these losses during pre-training, we ablate each loss function in the N-way K-shot IC tasks on HWU64 and Clinc150 (Table 10). The results indicate that two sub-losses, L gold_intent and L gold_utterance show relatively marignal improvements. However, it is noteworthy that L pseudo serves as the key sub-loss for PIE, highlighting the effectiveness of using pseudo intents. Specifically, removing L pseudo from the final loss results in up to 1.8% and 1.1% degradation in performance in the zero-and one-shot settings, respectively.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_2\",\n \"fig_3\"\n ],\n \"heading\": \"Varying K and N\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_7\",\n \"tab_0\"\n ],\n \"text\": \"We visualize the performance of the PIE model in challenging N-way K-shot IC task settings where the number of example utterances K or the number of candidate intent classes N varies. Plots of performance at varying K (Figure 2) show that our model has consistently higher performance than the baselines, and the performance improvement of our model is the largest when K is small (e.g., K=0). Plots of performance at varying N (Figure 3) show that the performance improvement of our model increases as the number of intents N increases (i.e. increasing from N=5 to N=50). These visualizations reveal that the PIE model can be utilized in more practical and realistic settings where many user intents are used for the TOD system and only a few utterances are available. As shown in Table 6, there are a few overlapping intent names between pre-training data and downstream data (except for Liu54). These coincidental overlaps can hinder an accurate evaluation of the generalization ability of our model. To understand the impact of intent overlaps, we also measure the performance using only non-overlapping intents (Table 11). We observe that the impact is marginal enough to be neglected, and surprisingly, removing overlapping intents rather can lead to better performance on Banking77 and HWU64. Through further analysis, we discover that this is partly because of the bias towards pairs of utterances and intent annotated in pre-training datasets. For example, an utterance 'please play my favorite song' in pre-training data has an intent 'play music'. Our model then incorrectly predicted 'play music' for a test utterance 'that song is my favorite', where the correct intent is 'music likeness'.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Impact of Overlapping Intents\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this work, we propose a pre-training method that leverages pseudo intent names constructed using an IRL tagger in a semi-supervised manner, followed by intent-aware pre-training (PIE). Experiments on four intent classification datasets show that our model achieves state-of-the-art performance on all datasets, outperforming the strongest sentence encoder baseline by up to 5.4% and 4.0% in N-way zero-and one-shot settings, respectively. Our analysis shows that PIE performs robustly compared to the baselines in challenging and practical settings with a large number of classes and small number of support examples. In future work, we will explore the use of IRL and our PIE model in multi-label intent classification or out-of-scope detection tasks.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"One limitation of our method is that while it leverages annotations from the IRL tagger, the detection of spans for certain labels, such as 'Problem,' is not accurate enough (38.1% F1 score). This is likely due to the relatively short number of annotations of this type in the training set (45 annotations). To mitigate this limitation, we could consider annotating more instances of this label or implementing techniques for handling imbalanced labels.\\nAnother limitation is that we currently treat all IRL labels equally when constructing pseudo intents. However, the importance of each label in interpreting intent can vary. To address this, we plan to investigate treating different labels differently when pre-training the encoder (e.g. by giving more weight to 'Action' and 'Argument' labels and less weight to 'Slot' labels).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ethics Statement\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Our proposed method is for enhancing zero-and few-shot intent classification, and it does not raise any ethical concerns. We believe that this research has valuable merits that can lead to more reliable task-oriented dialogue systems. All experiments in this study were carried out using publicly available datasets.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgements\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We gratefully acknowledge the members of AWS AI for providing valuable feedback on the project. We would also like to thank the anonymous reviewers for their insightful comments. The part of Mujeen Sung's graduate study and, accordingly, this work was supported by National Research Foundation of Korea (NRF-2023R1A2C3004176).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"* Work performed during an internship at AWS AI Labs.\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-11-14"},"doi":{"kind":"string","value":"10.18653/v1/2020.nlp4convai-1.5"},"references":{"kind":"list like","value":[{"authors":"","journal":"L-BERT TAPT","ref_id":"b0","title":"","year":""},{"authors":"References Pavel Burnyshev; A Bout; Anfrey Bout; Valentin Malykh; Irina Piontkovskaya","journal":"","ref_id":"b1","title":"Infobert: Zeroshot approach to natural language understanding using contextualized word embedding","year":"2021"},{"authors":"Iñigo Casanueva; Tadas Temčinas; Daniela Gerz; Matthew Henderson; Ivan Vulić","journal":"","ref_id":"b2","title":"Efficient intent detection with dual sentence encoders","year":"2020"},{"authors":"Junfan Chen; Richong Zhang; Yongyi Mao; Jie Xu","journal":"","ref_id":"b3","title":"Contrastnet: A contrastive learning framework for few-shot text classification","year":"2022"},{"authors":"Xilun Chen; Asish Ghoshal; Yashar Mehdad; Luke Zettlemoyer; Sonal Gupta","journal":"","ref_id":"b4","title":"Low-resource domain adaptation for compositional task-oriented semantic parsing","year":"2020"},{"authors":"Thomas Dopierre; Christophe Gravier; Wilfried Logerais","journal":"","ref_id":"b5","title":"a. A neural few-shot text classification reality check","year":"2021"},{"authors":"Thomas Dopierre; Christophe Gravier; Wilfried Logerais","journal":"","ref_id":"b6","title":"Protaugment: Intent detection meta-learning through unsupervised diverse paraphrasing","year":"2021"},{"authors":"James Gung; Raphael Shu; Emily Moeng; Wesley Rose; Salvatore Romeo; Yassine Benajiba; Arshit Gupta; Saab Mansour; Yi Zhang","journal":"","ref_id":"b7","title":"Intent induction from conversations for task-oriented dialogue track at dstc 11","year":"2023"},{"authors":"Sonal Gupta; Rushin Shah; Mrinal Mohit; Anuj Kumar; Mike Lewis","journal":"","ref_id":"b8","title":"Semantic parsing for task oriented dialog using hierarchical representations","year":"2018"},{"authors":"Suchin Gururangan; Ana Marasović; Swabha Swayamdipta; Kyle Lo; Iz Beltagy; Doug Downey; Noah A Smith","journal":"","ref_id":"b9","title":"Don't stop pretraining: Adapt language models to domains and tasks","year":"2020"},{"authors":"Jason Krone; Yi Zhang; Mona Diab","journal":"","ref_id":"b10","title":"Learning to classify intents and slot labels given a handful of examples","year":"2020"},{"authors":"Dmitry Lamanov; Pavel Burnyshev; Katya Artemova; Valentin Malykh; Andrey Bout; Irina Piontkovskaya","journal":"","ref_id":"b11","title":"Template-based approach to zeroshot intent recognition","year":"2022"},{"authors":"Stefan Larson; Anish Mahendran; Joseph J Peper; Christopher Clarke; Andrew Lee; Parker Hill; Jonathan K Kummerfeld; Kevin Leach; Michael A Laurenzano; Lingjia Tang; Jason Mars","journal":"","ref_id":"b12","title":"An evaluation dataset for intent classification and out-ofscope prediction","year":"2019"},{"authors":"Han Liu; Xiaotong Zhang; Lu Fan; Xuandi Fu; Qimai Li; Xiao-Ming Wu; Albert Y S Lam ; A","journal":"","ref_id":"b13","title":"Reconstructing capsule networks for zero-shot intent classification","year":"2019"},{"authors":"Han Liu; Siyang Zhao; Xiaotong Zhang; Feng Zhang; Junjie Sun; Hong Yu; Xianchao Zhang","journal":"","ref_id":"b14","title":"A simple meta-learning paradigm for zero-shot intent classification with mixture attention mechanism","year":"2022"},{"authors":"Xingkun Liu; Arash Eshghi; Pawel Swietojanski; Verena Rieser","journal":"","ref_id":"b15","title":"Benchmarking natural language understanding services for building conversational agents","year":"2019"},{"authors":"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b16","title":"Roberta: A robustly optimized bert pretraining approach","year":"2019"},{"authors":"Tingting Ma; Qianhui Wu; Zhiwei Yu; Tiejun Zhao; Chin-Yew Lin","journal":"","ref_id":"b17","title":"On the effectiveness of sentence encoding for intent detection meta-learning","year":"2022"},{"authors":"Aaron Mueller; Jason Krone; Salvatore Romeo; Saab Mansour; Elman Mansimov; Yi Zhang; Dan Roth","journal":"","ref_id":"b18","title":"Label semantic aware pre-training for few-shot text classification","year":"2022"},{"authors":"Martha Palmer; Daniel Gildea; Nianwen Xue","journal":"SLHLT","ref_id":"b19","title":"Semantic role labeling","year":"2010"},{"authors":"Abhinav Rastogi; Xiaoxue Zang; Srinivas Sunkara; Raghav Gupta; Pranav Khaitan","journal":"","ref_id":"b20","title":"Towards scalable multi-domain conversational agents: The schema-guided dialogue dataset","year":"2020"},{"authors":"Nils Reimers; Iryna Gurevych","journal":"","ref_id":"b21","title":"Sentence-BERT: Sentence embeddings using Siamese BERTnetworks","year":"2019"},{"authors":"Jake Snell; Kevin Swersky; Richard Zemel","journal":"","ref_id":"b22","title":"Prototypical networks for few-shot learning","year":"2017"},{"authors":"Aaron Van Den Oord; Yazhe Li; Oriol Vinyals","journal":"","ref_id":"b23","title":"Representation learning with contrastive predictive coding","year":"2018"},{"authors":"Xiaoxue Zang; Abhinav Rastogi; Srinivas Sunkara; Raghav Gupta; Jianguo Zhang; Jindong Chen","journal":"","ref_id":"b24","title":"MultiWOZ 2.2 : A dialogue dataset with additional annotation corrections and state tracking baselines","year":"2020"},{"authors":"Zengfeng Zeng; Dan Ma; Haiqing Yang; Zhen Gou; Jianping Shen","journal":"","ref_id":"b25","title":"Automatic intent-slot induction for dialogue systems","year":"2021"},{"authors":"Yiwen Zhang; Caixia Yuan; Xiaojie Wang; Ziwei Bai; Yongbin Liu","journal":"","ref_id":"b26","title":"Learn to adapt for generalized zero-shot text classification","year":"2022"}],"string":"[\n {\n \"authors\": \"\",\n \"journal\": \"L-BERT TAPT\",\n \"ref_id\": \"b0\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"References Pavel Burnyshev; A Bout; Anfrey Bout; Valentin Malykh; Irina Piontkovskaya\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Infobert: Zeroshot approach to natural language understanding using contextualized word embedding\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Iñigo Casanueva; Tadas Temčinas; Daniela Gerz; Matthew Henderson; Ivan Vulić\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Efficient intent detection with dual sentence encoders\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Junfan Chen; Richong Zhang; Yongyi Mao; Jie Xu\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Contrastnet: A contrastive learning framework for few-shot text classification\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xilun Chen; Asish Ghoshal; Yashar Mehdad; Luke Zettlemoyer; Sonal Gupta\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Low-resource domain adaptation for compositional task-oriented semantic parsing\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Thomas Dopierre; Christophe Gravier; Wilfried Logerais\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"a. A neural few-shot text classification reality check\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Thomas Dopierre; Christophe Gravier; Wilfried Logerais\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Protaugment: Intent detection meta-learning through unsupervised diverse paraphrasing\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"James Gung; Raphael Shu; Emily Moeng; Wesley Rose; Salvatore Romeo; Yassine Benajiba; Arshit Gupta; Saab Mansour; Yi Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Intent induction from conversations for task-oriented dialogue track at dstc 11\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Sonal Gupta; Rushin Shah; Mrinal Mohit; Anuj Kumar; Mike Lewis\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Semantic parsing for task oriented dialog using hierarchical representations\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Suchin Gururangan; Ana Marasović; Swabha Swayamdipta; Kyle Lo; Iz Beltagy; Doug Downey; Noah A Smith\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Don't stop pretraining: Adapt language models to domains and tasks\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jason Krone; Yi Zhang; Mona Diab\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Learning to classify intents and slot labels given a handful of examples\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Dmitry Lamanov; Pavel Burnyshev; Katya Artemova; Valentin Malykh; Andrey Bout; Irina Piontkovskaya\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Template-based approach to zeroshot intent recognition\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Stefan Larson; Anish Mahendran; Joseph J Peper; Christopher Clarke; Andrew Lee; Parker Hill; Jonathan K Kummerfeld; Kevin Leach; Michael A Laurenzano; Lingjia Tang; Jason Mars\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"An evaluation dataset for intent classification and out-ofscope prediction\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Han Liu; Xiaotong Zhang; Lu Fan; Xuandi Fu; Qimai Li; Xiao-Ming Wu; Albert Y S Lam ; A\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Reconstructing capsule networks for zero-shot intent classification\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Han Liu; Siyang Zhao; Xiaotong Zhang; Feng Zhang; Junjie Sun; Hong Yu; Xianchao Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"A simple meta-learning paradigm for zero-shot intent classification with mixture attention mechanism\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xingkun Liu; Arash Eshghi; Pawel Swietojanski; Verena Rieser\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Benchmarking natural language understanding services for building conversational agents\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Roberta: A robustly optimized bert pretraining approach\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Tingting Ma; Qianhui Wu; Zhiwei Yu; Tiejun Zhao; Chin-Yew Lin\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"On the effectiveness of sentence encoding for intent detection meta-learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Aaron Mueller; Jason Krone; Salvatore Romeo; Saab Mansour; Elman Mansimov; Yi Zhang; Dan Roth\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Label semantic aware pre-training for few-shot text classification\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Martha Palmer; Daniel Gildea; Nianwen Xue\",\n \"journal\": \"SLHLT\",\n \"ref_id\": \"b19\",\n \"title\": \"Semantic role labeling\",\n \"year\": \"2010\"\n },\n {\n \"authors\": \"Abhinav Rastogi; Xiaoxue Zang; Srinivas Sunkara; Raghav Gupta; Pranav Khaitan\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Towards scalable multi-domain conversational agents: The schema-guided dialogue dataset\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Nils Reimers; Iryna Gurevych\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Sentence-BERT: Sentence embeddings using Siamese BERTnetworks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Jake Snell; Kevin Swersky; Richard Zemel\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Prototypical networks for few-shot learning\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Aaron Van Den Oord; Yazhe Li; Oriol Vinyals\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Representation learning with contrastive predictive coding\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Xiaoxue Zang; Abhinav Rastogi; Srinivas Sunkara; Raghav Gupta; Jianguo Zhang; Jindong Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"MultiWOZ 2.2 : A dialogue dataset with additional annotation corrections and state tracking baselines\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Zengfeng Zeng; Dan Ma; Haiqing Yang; Zhen Gou; Jianping Shen\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Automatic intent-slot induction for dialogue systems\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yiwen Zhang; Caixia Yuan; Xiaojie Wang; Ziwei Bai; Yongbin Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Learn to adapt for generalized zero-shot text classification\",\n \"year\": \"2022\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[2,360.78,345.38,164.36,30.47],"formula_id":"formula_0","formula_text":"c n = 1 K x n,i ∈Sn f ϕ (x n,i )(1)"},{"formula_coordinates":[2,313.41,625.12,211.73,30.47],"formula_id":"formula_1","formula_text":"c label n = 1 K + 1 [[ x n,i ∈Sn f ϕ (x n,i )] + f ϕ (y n )] (2)"},{"formula_coordinates":[4,77.95,488.14,211.91,29.81],"formula_id":"formula_2","formula_text":"ℓ(x i , y) = exp sim(f ϕ (x i ), f ϕ (y i )) N k exp sim(f ϕ (x i ), f ϕ (y k )) ,(3)"},{"formula_coordinates":[4,105.03,714.31,184.84,33.71],"formula_id":"formula_3","formula_text":"L gold_intent = - 1 N N i ℓ(x i , y gold ),(4)"},{"formula_coordinates":[4,99.85,760.35,19.91,15.47],"formula_id":"formula_4","formula_text":"y gold i"},{"formula_coordinates":[4,334.6,354.71,190.54,33.71],"formula_id":"formula_5","formula_text":"L gold_utterance = - 1 N N i ℓ(x i , x gold ),(5)"},{"formula_coordinates":[4,335.13,398.28,20.41,15.47],"formula_id":"formula_6","formula_text":"x gold i"},{"formula_coordinates":[4,342.97,460.9,182.18,33.71],"formula_id":"formula_7","formula_text":"L pseudo = - 1 N N i ℓ(x i , y pseudo ),(6)"},{"formula_coordinates":[4,312.4,584.54,212.74,10.82],"formula_id":"formula_8","formula_text":"L = L gold_intent + L gold_utterance + λL pseudo , (7)"}],"string":"[\n {\n \"formula_coordinates\": [\n 2,\n 360.78,\n 345.38,\n 164.36,\n 30.47\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"c n = 1 K x n,i ∈Sn f ϕ (x n,i )(1)\"\n },\n {\n \"formula_coordinates\": [\n 2,\n 313.41,\n 625.12,\n 211.73,\n 30.47\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"c label n = 1 K + 1 [[ x n,i ∈Sn f ϕ (x n,i )] + f ϕ (y n )] (2)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 77.95,\n 488.14,\n 211.91,\n 29.81\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"ℓ(x i , y) = exp sim(f ϕ (x i ), f ϕ (y i )) N k exp sim(f ϕ (x i ), f ϕ (y k )) ,(3)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 105.03,\n 714.31,\n 184.84,\n 33.71\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"L gold_intent = - 1 N N i ℓ(x i , y gold ),(4)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 99.85,\n 760.35,\n 19.91,\n 15.47\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"y gold i\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 334.6,\n 354.71,\n 190.54,\n 33.71\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"L gold_utterance = - 1 N N i ℓ(x i , x gold ),(5)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 335.13,\n 398.28,\n 20.41,\n 15.47\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"x gold i\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 342.97,\n 460.9,\n 182.18,\n 33.71\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"L pseudo = - 1 N N i ℓ(x i , y pseudo ),(6)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 312.4,\n 584.54,\n 212.74,\n 10.82\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"L = L gold_intent + L gold_utterance + λL pseudo , (7)\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"Pre-training Intent-Aware Encoders for Zero-and Few-Shot Intent Classification"},"abstract":{"kind":"string","value":"Intent classification (IC) plays an important role in task-oriented dialogue systems. However, IC models often generalize poorly when training without sufficient annotated examples for each user intent. We propose a novel pretraining method for text encoders that uses contrastive learning with intent psuedo-labels to produce embeddings that are well-suited for IC tasks, reducing the need for manual annotations. By applying this pre-training strategy, we also introduce Pre-trained Intent-aware Encoder (PIE), which is designed to align encodings of utterances with their intent names. Specifically, we first train a tagger to identify key phrases within utterances that are crucial for interpreting intents. We then use these extracted phrases to create examples for pretraining a text encoder in a contrastive manner. As a result, our PIE model achieves up to 5.4% and 4.0% higher accuracy than the previous state-of-the-art text encoder for the N-way zeroand one-shot settings on four IC datasets. 1 "},"authors":{"kind":"string","value":"Mujeen Sung; James Gung; Elman Mansimov; Nikolaos Pappas; Raphael Shu; Salvatore Romeo; Yi Zhang; Vittorio Castelli"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: Overview of pre-training the intent-aware encoder (PIE). Given an utterance x 1 from pre-training corpus, we generate a pseudo intent name y pseudo","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Performance on N-way K-shot intent classification with varying K. PA refers to ProtAugment.","figure_data":"","figure_id":"fig_2","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Performance on N-way 0-shot intent classification with varying N.","figure_data":"","figure_id":"fig_3","figure_label":"3","figure_type":"figure"},{"figure_caption":"Table 1 shows the statistics and examples of each IRL label from 3,879 utterances. To train and evaluate the IRL tagger, we split annotations into training, valida-tion, and test sets with 3,121 / 379 / 379 utterances, respectively, which is approximately an 80:10:10 ratio. Statistics and examples of each IRL label from 3,879 utterances.","figure_data":"LabelCount ExampleAction2,163 I want to book ACT a flightArgument2,011 I want to book a flight ARGRequest3,002 Can you show REQ me my account balanceQuery3,247 Can you show me my account balance QRYSlot2,030Can you show me my account balance for my checking account SLTProblem45 I'm starting to get hungry PRB","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"","figure_data":"Action89.386.988.1Argument85.885.485.6Request90.993.592.1Query92.295.894.0Slot82.884.683.7Problem35.641.738.1","figure_id":"tab_1","figure_label":"2","figure_type":"table"},{"figure_caption":"Precision, recall, and F1 scores of each IRL label on the test set.","figure_data":"","figure_id":"tab_2","figure_label":"2","figure_type":"table"},{"figure_caption":"like to open ACT a savings SLT account ARG please open savings account So I need to sign up ACT for a a savings SLT account ARG sign up savings account Can you buy ACT me some movie tickets ARG buy movie tickets buy movie tickets I am looking to book ACT movie tickets ARG book movie tickets I am looking to purchase ACT movie tickets ARG purchase movie tickets Delete ACT this song ARG from playlist ARG delete song playlist remove from playlist music Erase ACT this track ARG from the playlist ARG erase track playlist Could you remove ACT this song ARG permanently remove song Some examples of IRL predictions (boldfaced) from utterances, extracted pseudo intent names, and gold intent names annotated in the original dataset.","figure_data":"Utterances","figure_id":"tab_3","figure_label":"3","figure_type":"table"},{"figure_caption":"Pre-training datasets for the PIE model.","figure_data":"DatasetUtterancesGold IntentsPseudo IntentsTOP (+v2)31,1116123,711TrainDSTC11-T2 SGD4,459 3,561148 443,304 2,647Total39,11725229,577ValMultiWOZ 2.258610-","figure_id":"tab_4","figure_label":"4","figure_type":"table"},{"figure_caption":"The statistics of four IC datasets.","figure_data":"","figure_id":"tab_6","figure_label":"5","figure_type":"table"},{"figure_caption":"Number of overlapping intent names between the pre-training data and downstream data.Table7: 5-way K-shot intent classification performance of pre-trained models with and without fine-tuning on four test sets. Averaged accuracies and standard deviations across five class splits are reported. The 'L-' prefixes indicate the use of intent label names when creating prototypes, enabling zero-shot evaluation. Highest scores are boldfaced.","figure_data":"MethodFine-tuningBanking77 K=0 K=1HWU64 K=0 K=1K=0Liu54K=1Clinc150 K=0 K=1Average K=0 K=1BERT TAPT-50.8 ±1.5-52.1 ±1.1-52.7 ±2.6-50.7 ±4.1-51.6SBERT Paraphrase--83.6 ±1.8-82.2 ±1.0-82.3 ±2.2-94.7 ±0.6-85.7PIE (Ours)-86.1 ±1.3-86.0 ±1.7-85.6 ±1.9-96.2 ±0.5-88.5L-BERT TAPT27.7 ±1.841.0 ±1.538.5 ±2.453.6 ±1.551.7 ±4.660.0 ±4.139.5 ±1.451.8 ±3.239.3 51.6L-SBERT Paraphrase-86.9 ±1.990.9 ±0.683.6 ±1.889.0 ±1.479.9 ±3.889.8 ±1.994.3 ±1.197.7 ±0.486.2 91.9L-PIE (Ours)88.3 ±2.292.4 ±0.787.7 ±2.692.2 ±1.583.8 ±3.791.7 ±1.496.5 ±0.898.3 ±0.489.1 93.7L-BERT TAPT85.7 ±2.391.5 ±1.081.4 ±1.386.6 ±1.380.3 ±3.188.2 ±1.293.0 ±1.996.9 ±0.685.1 90.8L-SBERT ParaphraseProtoNet90.9 ±1.994.5 ±0.585.8 ±2.891.1 ±1.683.7 ±4.392.1 ±1.697.1 ±0.798.6 ±0.289.4 94.0L-PIE (Ours)90.7 ±2.294.3 ±0.786.6 ±4.092.1 ±1.785.0 ±3.992.4 ±1.597.3 ±0.398.6 ±0.389.9 94.4L-BERT TAPT89.2 ±2.193.4 ±0.587.0 ±2.689.8 ±1.183.0 ±4.690.9 ±0.995.3 ±1.097.7 ±0.288.6 92.9L-SBERT ParaphraseProtAugment 92.3 ±1.194.8 ±0.487.3 ±2.591.7 ±1.684.1 ±3.392.5 ±1.497.0 ±0.798.5 ±0.290.2 94.4L-PIE (Ours)92.4 ±1.094.8 ±0.488.8 ±3.092.4 ±1.686.0 ±3.392.9 ±1.397.6 ±0.498.7 ±0.291.2 94.7","figure_id":"tab_7","figure_label":"6","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: Overview of pre-training the intent-aware encoder (PIE). Given an utterance x 1 from pre-training corpus, we generate a pseudo intent name y pseudo\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Performance on N-way K-shot intent classification with varying K. PA refers to ProtAugment.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Performance on N-way 0-shot intent classification with varying N.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Table 1 shows the statistics and examples of each IRL label from 3,879 utterances. To train and evaluate the IRL tagger, we split annotations into training, valida-tion, and test sets with 3,121 / 379 / 379 utterances, respectively, which is approximately an 80:10:10 ratio. Statistics and examples of each IRL label from 3,879 utterances.\",\n \"figure_data\": \"LabelCount ExampleAction2,163 I want to book ACT a flightArgument2,011 I want to book a flight ARGRequest3,002 Can you show REQ me my account balanceQuery3,247 Can you show me my account balance QRYSlot2,030Can you show me my account balance for my checking account SLTProblem45 I'm starting to get hungry PRB\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"Action89.386.988.1Argument85.885.485.6Request90.993.592.1Query92.295.894.0Slot82.884.683.7Problem35.641.738.1\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Precision, recall, and F1 scores of each IRL label on the test set.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"like to open ACT a savings SLT account ARG please open savings account So I need to sign up ACT for a a savings SLT account ARG sign up savings account Can you buy ACT me some movie tickets ARG buy movie tickets buy movie tickets I am looking to book ACT movie tickets ARG book movie tickets I am looking to purchase ACT movie tickets ARG purchase movie tickets Delete ACT this song ARG from playlist ARG delete song playlist remove from playlist music Erase ACT this track ARG from the playlist ARG erase track playlist Could you remove ACT this song ARG permanently remove song Some examples of IRL predictions (boldfaced) from utterances, extracted pseudo intent names, and gold intent names annotated in the original dataset.\",\n \"figure_data\": \"Utterances\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Pre-training datasets for the PIE model.\",\n \"figure_data\": \"DatasetUtterancesGold IntentsPseudo IntentsTOP (+v2)31,1116123,711TrainDSTC11-T2 SGD4,459 3,561148 443,304 2,647Total39,11725229,577ValMultiWOZ 2.258610-\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"The statistics of four IC datasets.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Number of overlapping intent names between the pre-training data and downstream data.Table7: 5-way K-shot intent classification performance of pre-trained models with and without fine-tuning on four test sets. Averaged accuracies and standard deviations across five class splits are reported. The 'L-' prefixes indicate the use of intent label names when creating prototypes, enabling zero-shot evaluation. Highest scores are boldfaced.\",\n \"figure_data\": \"MethodFine-tuningBanking77 K=0 K=1HWU64 K=0 K=1K=0Liu54K=1Clinc150 K=0 K=1Average K=0 K=1BERT TAPT-50.8 ±1.5-52.1 ±1.1-52.7 ±2.6-50.7 ±4.1-51.6SBERT Paraphrase--83.6 ±1.8-82.2 ±1.0-82.3 ±2.2-94.7 ±0.6-85.7PIE (Ours)-86.1 ±1.3-86.0 ±1.7-85.6 ±1.9-96.2 ±0.5-88.5L-BERT TAPT27.7 ±1.841.0 ±1.538.5 ±2.453.6 ±1.551.7 ±4.660.0 ±4.139.5 ±1.451.8 ±3.239.3 51.6L-SBERT Paraphrase-86.9 ±1.990.9 ±0.683.6 ±1.889.0 ±1.479.9 ±3.889.8 ±1.994.3 ±1.197.7 ±0.486.2 91.9L-PIE (Ours)88.3 ±2.292.4 ±0.787.7 ±2.692.2 ±1.583.8 ±3.791.7 ±1.496.5 ±0.898.3 ±0.489.1 93.7L-BERT TAPT85.7 ±2.391.5 ±1.081.4 ±1.386.6 ±1.380.3 ±3.188.2 ±1.293.0 ±1.996.9 ±0.685.1 90.8L-SBERT ParaphraseProtoNet90.9 ±1.994.5 ±0.585.8 ±2.891.1 ±1.683.7 ±4.392.1 ±1.697.1 ±0.798.6 ±0.289.4 94.0L-PIE (Ours)90.7 ±2.294.3 ±0.786.6 ±4.092.1 ±1.785.0 ±3.992.4 ±1.597.3 ±0.398.6 ±0.389.9 94.4L-BERT TAPT89.2 ±2.193.4 ±0.587.0 ±2.689.8 ±1.183.0 ±4.690.9 ±0.995.3 ±1.097.7 ±0.288.6 92.9L-SBERT ParaphraseProtAugment 92.3 ±1.194.8 ±0.487.3 ±2.591.7 ±1.684.1 ±3.392.5 ±1.497.0 ±0.798.5 ±0.290.2 94.4L-PIE (Ours)92.4 ±1.094.8 ±0.488.8 ±3.092.4 ±1.686.0 ±3.392.9 ±1.397.6 ±0.498.7 ±0.291.2 94.7\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2019a)\", \"Explanation\": \"The cited work by Liu et al. provides a method for few-shot text classification that the citing paper adopts to tackle the challenge of data collection and re-training models in task-oriented dialogue systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Krone et al., 2020)\", \"Explanation\": \"The cited work by Krone et al. contributes a method for few-shot text classification that the citing paper uses to address the challenge of data collection and re-training models in task-oriented dialogue systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Burnyshev et al., 2021)\", \"Explanation\": \"The cited work by Burnyshev et al. provides a method for few-shot text classification that the citing paper adopts to improve the design of new intent classes in task-oriented dialogue systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Mueller et al., 2022)\", \"Explanation\": \"The cited work by Mueller et al. contributes a method for few-shot text classification that the citing paper uses to enhance the design of new intent classes in task-oriented dialogue systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2022)\", \"Explanation\": \"The cited work by Zhang et al. provides a method for few-shot text classification that the citing paper adopts to improve the design of new intent classes in task-oriented dialogue systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lamanov et al., 2022)\", \"Explanation\": \"The cited work by Lamanov et al. contributes a method for few-shot text classification that the citing paper uses to enhance the design of new intent classes in task-oriented dialogue systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The cited work by Liu et al. provides a method for few-shot text classification that the citing paper adopts to improve the design of new intent classes in task-oriented dialogue systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Snell et al., 2017)\", \"Explanation\": \"The cited work on Prototypical Networks is used as a methodological basis for the PIE model in the citing paper, which utilizes text encoders to construct class representations and retrieve correct classes based on a similarity metric.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Dopierre et al., 2021a)\", \"Explanation\": \"The cited work on the impact of few-shot learning techniques and text encoders on classification performance provides supporting evidence for the research direction of better text encoder selection in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ma et al., 2022)\", \"Explanation\": \"The cited work on sentence encoders pre-trained on paraphrase and natural language inference datasets serves as a continuation of the research on text encoder selection for Prototypical Networks in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zeng et al., 2021)\", \"Explanation\": \"The cited work introduces the concept of intent role labeling (IRL), which the citing paper adopts in their pre-training method for zero-and few-shot IC tasks. This method is used to identify and assign roles to words or phrases in sentences that are relevant to user intents, which is a key element in the pre-training process.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Snell et al., 2017)\", \"Explanation\": \"The cited work introduces the Prototypical Networks approach, which the citing paper builds upon to develop a metalearning approach for intent classification with a few example utterances.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ma et al., 2019)\", \"Explanation\": \"The cited work by Ma et al. (2019) provides a method for optimizing intent classifiers in a fewshot learning setting, which the citing paper adopts in their research on intent classification.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Palmer et al., 2010)\", \"Explanation\": \"The cited work on semantic role labeling (SRL) provides a method for assigning general semantic roles to words or phrases in sentences, which the citing paper adapts to the task of intent role labeling (IRL) for discovering intents in user utterances.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Rastogi et al., 2020)\", \"Explanation\": \"The cited work provides a dataset of utterances that the citing paper uses to manually annotate intent role labels, which serves as a foundational element for the study conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(van den Oord et al., 2018)\", \"Explanation\": \"The cited work introduces the InfoNCE loss function, which the citing paper adopts in their training objective to align representations in the semantic embedding space.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Dopierre et al., 2021a)\", \"Explanation\": \"The cited work provides the Pro-toNet method, which the citing paper uses as a fine-tuning approach for few-shot IC tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Dopierre et al., 2021b)\", \"Explanation\": \"The cited work introduces the ProtAugment method, which the citing paper adopts to augment paraphrased utterances within episodes to mitigate overfitting in the few-shot IC task.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ma et al., 2022)\", \"Explanation\": \"The cited work provides the SBERT Paraphrase model, which the citing paper uses as a pre-trained text encoder in the few-shot IC task.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Reimers and Gurevych, 2019)\", \"Explanation\": \"The cited work provides a pre-trained text encoder that the citing paper uses to produce utterance embeddings for Prototypical Networks and Pro-toNet and ProtAugment baselines.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ma et al., 2022)\", \"Explanation\": \"The cited work discovers the use of SBERT Paraphrase in the inference stage of Prototypical Networks, which the citing paper further extends by fine-tuning the encoder using Pro-toNet and ProtAugment as baselines.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Gupta et al., 2018), (Chen et al., 2020), (Gung et al., 2023), (Rastogi et al., 2020), (Zang et al., 2020)\", \"Explanation\": \"The cited works provide the dialogue datasets that the citing paper uses to pre-train and validate the SBERT Paraphrase text encoder.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Casanueva et al., 2020)\", \"Explanation\": \"The cited work provides the Banking77 dataset, which is a crucial data source for the research conducted in the citing paper on understanding intent names in the banking domain.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Liu et al., 2019b)\", \"Explanation\": \"The cited work provides the HWU64 and Liu54 datasets, which are used in the research conducted in the citing paper to study a wide range of domains in the home assistant robot and Amazon Mechanical Turk contexts.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Larson et al., 2019)\", \"Explanation\": \"The Clinc150 dataset is cited as a data source for the research conducted in the citing paper, which includes a diverse range of intents from different domains such as small talk and travel.\"}]"}}},{"rowIdx":77,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b10","b7","b15","b17","b14","b12","b0","b7","b15","b17","b14","b12","b0","b17","b5","b7"],"table_ref":[],"text":"Neural radiance fields (NeRF) [11] have recently emerged as a new 3D volumetric representation capable of synthesizing photo-realistic novel views from multi-view images. Many dynamic NeRFs are also proposed to reconstruct scenes with moving or deforming objects, using space-time 4D [8,16] or explicit deformation [18,15] representations. However, most of the dynamic NeRFs focus on reconstructing the scene given the full training videos of a dynamic event. Although capable of playing back the dynamic scene after the event, they cannot stream the event while it is happening. In this paper, we introduce a new \"on-the-fly training\" approach to train a dynamic NeRF concurrently with the acquisition of the training frames, as the dynamic event unfolds. For the purpose of streaming the dynamic scene, the model only needs to render the current radiance field anytime during the training. With the recent success in NeRF acceleration to train [13,1] a small scene in seconds, it becomes a promising possibility to further speed up the dynamic scene on-the-fly training to an interactive speed. This enables the streaming of dynamic scenes with a wide variety of applications in social media, VR/AR, and gaming industries.\nTraining a dynamic NeRF on-the-fly is considerably different from training it after the dynamic event. For training after the event, the NeRF is trained to fit the 4D space-time radiance field with the available full training videos. In contrast, for training on-the-fly, NeRF is trained to represent the current radiance field given the previous reconstruction and the available video frames up to the current time step. As the dynamic event continues, new frames will become available and the NeRF will be updated. Training with each frame from scratch is very time-consuming, so the key to efficient on-the-fly dynamic NeRF training is to effectively utilize the radiance fields estimated from the previous frames for the faster convergence of the current radiance field. By leveraging the correspondence and transition across frames, we propose: 1) a NeRF representation conditioned on multi-view projected colors for faster convergence, and 2) a transition and update to the occupancy grid used for efficient sampling.\nOur proposed NeRF representation conditioned on multi-view projected colors is designed for fast convergence when training the radiance fields of consecutive frames on-the-fly. Most of the existing dynamic NeRFs explicitly represent the motion using a temporal input dimension [8,16] in NeRF or a temporal deformation field [18,15]. Due to the lack of direct supervision or known dynamics, these models with temporal input often extrapolate poorly to the unseen time and require many iterations of optimization for each frame. Instead of being conditioned on time, we propose a model that is conditioned on the projected colors from the training views. This is based on the observation that the projected color of a corresponding 3D point often stays unchanged in consecutive frames. With guidance from the invariant projected colors, the NeRF model is implicitly aware of the point correspondence across consecutive frames. Consequently, the NeRF model can effectively utilize the radiance fields from the previous frames to render for the current frame with the implicit correspondence. The experiments suggest that our proposed model has excellent temporal extrapolation capability and requires minimum number of optimization iterations for each new frame. As a result, our on-the-fly training speed of the model is significantly improved.\nAdditionally, we introduce a method of transiting and updating the occupancy grid used for efficient point sampling. Since most of the 3D scenes are empty spaces, the occupancy grid has been used in static NeRFs [13,1] to reduce the number of sampled points for acceleration. To adapt the occupancy grid to on-the-fly dynamic training, we consider the occupancy grid as a probability of the 3D voxels occupied by any object. To probabilistically model the motions in the scene, a transition function is applied to occupancy probability at the start of each frame optimization and later updated with new observations. The updated occupancy grid can then be used to sample points only in the occupied areas anytime during the on-the-fly training.\nWe evaluate our method on synthetic D-NeRF dataset [18], real-world MeetRoom [6], and DyNeRF [8] dataset. When trained and rendered on-the-fly, our method achieves significant acceleration compared to the state-of-the-art algorithms while maintaining a comparable rendering quality. Particularly, our method can train and render 6 frames per second (FPS) on the synthetic D-NeRF dataset. We summarize our contributions as follows: 1) we introduce and formally formulate the new setting of training dynamic NeRF on-the-fly. 2) We propose a projected color-guided on-the-fly dynamic NeRF and a transiting occupancy grid for efficient on-the-fly training. 3) We achieve 10× on-the-fly training and rendering acceleration in synthetic scenes and 3× acceleration in real-world scenes compared to the state-of-the-art."},{"figure_ref":[],"heading":"Related Works","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Accelerated Dynamic Neural Radiance Fields","publication_ref":["b10","b22","b25","b17","b7","b14","b24","b3","b12","b5","b25","b15","b19"],"table_ref":[],"text":"Following the success in representing static 3D scenes with Neural Radiance Fields(NeRFs) [11,23,26], many works [18,8,15,25] have been exploring representing dynamic scenes with NeRFs as well. Similar to the early static NeRFs, they usually take at least hours to train for a single scene. To improve the training efficiency of dynamic NeRFs, some works try to migrate the acceleration methods used for static NeRFs to dynamic NeRFs. TiNeuVox [4] follows the voxel and Multi-Layer Perceptron(MLP) architecture as in [13] to represent static canonical space, and uses another MLP to capture temporal deformation. StreamRF [6] represents the scene at the first frame using the Plenoxel [26], and learns the subsequent changes to this voxel grid. [16,20] directly expand the 3D representation used in static NeRF to 4D with a time dimension to represent dynamic scenes. However, most of the existing works only focus on joint training of the dynamic scenes, instead of on-the-fly training introduced in our work. Their per frame training speed is also not fast enough for interactive on-the-fly applications."},{"figure_ref":[],"heading":"Image Based Rendering","publication_ref":["b23","b1","b4","b18","b9","b21","b20","b8"],"table_ref":[],"text":"Instead of representing a 3D scene as a NeRF conditioned on spatial coordinates and viewing direction, some works rely on additional projected colors/features on the training views to improve generalization or robustness. IBRNet [24], MVSNeRF [2] and many following works [5,19,10] construct a cost volume of a dynamic scene based on the image features of the nearby views to learn a blending weights or the density and color output. These models are designed to be generalizable to unseen scenes as they do not rely on the spatial coordinates input. LLFF [22,21] aggregates features along the multi-view epipolar line to render view dependent effect. Recently, DynIBaR [9] applies this technique to dynamic NeRF by aggregating the projected image features across frames, after warpping a point using a motion trajectory learned from past and future frames. However, this method requires time-consuming optimization of the per-frame motion trajectory and cannot be applied to on-the-fly training as the future frames are not known."},{"figure_ref":[],"heading":"Problem Formulation","publication_ref":["b10","b17","b14","b7","b15","b13","b16"],"table_ref":[],"text":"In this section, we first briefly describe the NeRF preliminaries necessary to understand our formulation. We then formally define our new on-the-fly dynamic NeRF training and followed by simplifying the problem formulation in the form of a Hidden Markov model (HMM).\nNeRF Preliminaries. Static neural radiance fields [11] represent a 3D scene implicitly with a continuous volumetric function F : (x, d) → (σ, c) that maps the spatial position x ∈ R 3 and viewing direction d ∈ R 3 to the volume density σ ∈ R and RGB color c ∈ R 3 . To synthesize the pixel color Ĉ(r) on any 2D images, volume rendering is used to aggregate the color of N points with interval δ along the ray r shooting from the pixel:\nĈ(r) = N -1 i=0 T i (1 -exp(-σ i δ i ))c i , T i = exp(- i-1 j=0 σ j δ j ).(1)\nDynamic NeRFs usually include an additional time dimension t in the input, through either a timevarying deformation field D : (x, t) → (x ′ ) that maps the spatial coordinates x to its canonical space correspondence x ′ [18,15], or directly expanding the 3D NeRF model to a 4D variant [8,16]. During the training stage, all K frames of the training images C 0:K (r) from all time t 0:K are used to jointly minimize the rendering loss:\nF ′ : (x, d, t) → (σ, c)\nL = K t=0 r∈R || Ĉt (r) -C t (r)|| 2 2 .(2)\nOn-the-fly Dynamic NeRF. We introduce a new on-the-fly training of the dynamic NeRF suitable for streaming dynamic scenes. Instead of training after the dynamic event, we train the NeRF concurrently while the dynamic event unfolds. Furthermore, the NeRF trained at time t k renders novel views only at time t k . Given the images up to the current time step C 0:k and the radiance field estimated up to the last time step F 0:k-1 , the goal of on-the-fly dynamic NeRF training is to find the radiance field function F k at time t k that minimizes the rendering loss at the current time step:\nF k (x, d) = argmax F k P (F k | C 0:k , F 0:k-1 ) = argmin F k r∈R || Ĉk (r) -C k (r)|| 2 2 .\n(3)\nHowever, estimating the radiance field F k at the current time step conditioned on all the previous images C 0:k and radiance field F 0:k-1 is not scalable as time t k increases. To mitigate this growth in complexity, we apply the first order Markov assumption to simplify the probability model by assuming conditional independence\nF k ⊥ ⊥ {C 0:k-1 , F 0:k-2 } | {C k , F k-1 }.\nThis simplifies the estimation of the current radiance field as:\nF k = argmax F k P (F k | C k , F k-1 ).(4)\nTaking the radiance fields F 0:K as hidden states and images C 0:K as observations, we can formulate the on-the-fly training as the process of estimating the hidden states in a Hidden Markov model (HMM). The emission function\nP (C k | F k )\nis the process of volumetric rendering that renders 2D images from 3D radiance fields. The transition function\nP (F k | F k-1\n) is the radiance field deformation or motion between two consecutive time steps.\nBased on the formulation above, the key to efficient on-the-fly training is to maintain low complexity of the update\nP (F k | C k , F k-1\n) to the radiance field at each time step. To this end, we propose a projected color-guided NeRF that is implicitly aware of point correspondence that requires minimum optimization when transiting from F k-1 to F k for fast training. Furthermore, we also introduce a simple transition function P (G k |G k-1 ) to the occupancy grid G used for efficient sampling.\nRemarks. Note that we limit the scope of this work to dynamic scenes captured by multi-view forward-facing cameras based on realistic considerations. Although the reconstruction of dynamic scenes from a monocular camera is less demanding on the hardware, it requires the photographer to keep on moving the camera [14]. This is cumbersome in prolonged streaming scenarios. 360-degree inward-facing cameras can be used to reconstruct from all angles. Nonetheless, this often requires dozens of cameras and a much bigger space [17]. It is difficult for most streamers to acquire such professional setups. Consequently, we focus on scenes captured by static multi-view forward-facing cameras that are most aligned with the setups used in the current streaming industry."},{"figure_ref":["fig_0"],"heading":"Method","publication_ref":[],"table_ref":[],"text":"As mentioned in the previous section, the goal of on-the-fly training is to estimate the current radiance field F k based on the last radiance field F k-1 and the current training images C k . In practice, this can be achieved by optimizing the radiance field model from the last step with the current training images.\nTo achieve highly efficient on-the-fly training, we can either reduce the number of optimization iterations needed for each time step or reduce the time spent on each iteration without scarifying the performance. Based on our HMM-based on-the-fly training paradigm, we propose: 1) a dynamic NeRF guided by projected colors, and 2) an occupancy grid transition and update strategy, to achieve these two goals respectively (Fig. 1)."},{"figure_ref":["fig_1","fig_1"],"heading":"Dynamic NeRF Guided by Projected Colors","publication_ref":[],"table_ref":[],"text":"According to the HMM formulation of on-the-fly training, the optimization at each time step is effectively learning a probability update to the last step model given the current observations:\nF k = P (F k | F k-1 , C k )F k-1 .\nIn the context of dynamic NeRFs, this can usually be achieved by estimating a deformation field or point correspondence function.\nExisting deformation-based dynamic NeRFs try to learn a deformation field D : (x, t k ) → (x ′ ) that captures the point correspondence through joint training. However, our experiments indicate that this process itself takes too long for efficient on-the-fly training. We postulate that this is because the Figure 2: We illustrate the trajectory of a blue ball moving on a (x, y) over time(left). When mapping it to the \"space-time\" coordinate system, the point correspondence can be hard to track due to the irregular trajectory(middle, 2D space represented with x + y for simpler visualization). When mapping it to the \"spacecolor\" coordinate system, the correspondence can be easily tracked due to the invariant color(right).\ntemporal deformation field does not have a clear pattern in most cases. We illustrate this problem with a toy example as shown in Fig. 2. In this example, we track the center of the 2D moving ball of a constant color. When tracking the trajectory of the ball in \"space-time\" input space, it is difficult to accurately predict the location of the ball at the next time step. This is similar to the temporal deformation field used by many existing dynamic NeRFs. The deformation parameters need to be optimized at each time step based on the indirect RGB supervision. In the case of 3D neural radiance field reconstruction, it is very time-consuming to optimize for all points in the space.\nTo efficiently utilize the point correspondence across frames, we change the time input t in the space-time dynamic NeRF model F (x, d, t) to the multi-view projected color mean cp (x, k) and variance Var(c p (x, k)) at frame k to become:\nF k : (x, d, cp (x, k), Var(c p (x, k))) → (σ, c).(5)\nThe projected colors c p (x, k) are defined as the pixel colors of the training images C k,cam of all cameras M through 3D to 2D projection with camera projection matrix P cam :\nc p (x, k) = {C k,cam (P cam • x) | cam ∈ M}.(6)\nAs the model is conditioned on the spatial coordinates and projected colors, it circumvents the difficulty of estimating the point correspondence with irregularity in the spatial and temporal dimensions. Instead, it is much easier to estimate the correspondence in the spatial and color input space due to the invariance of the point color under motion. As illustrated in Fig. 2, the trajectory of the point in the spatial and color input space is much more regular than that in the spatial and temporal input space. Although the projected colors of 3D points are not perfectly invariant due to factors such as reflection, lighting, and occlusion, we observe that the projected colors stay largely similar across consecutive frames.\nTo verify that our method is better at estimating point correspondence on-the-fly, we demonstrate a simple extrapolation experiment in Fig. 3. Specifically, we train the different dynamic NeRFs on-thefly up to the (k -1)th frame and then extrapolate to render the kth frame without any training on the kth frame. As shown in Fig. 3, the model operating in the spatial and temporal input dimensions extrapolates poorly as it is not aware of the point correspondence with untrained time. Our model operating in the spatial and projected color input dimensions can extrapolate well based on the point correspondence hinted by the projected colors. With the implicit point correspondence and the good extrapolation ability, our model requires a low number of iterations for each new frame and thus is very fast when training on-the-fly."},{"figure_ref":["fig_2"],"heading":"Occupancy Grid Transition and Update","publication_ref":["b15","b12"],"table_ref":[],"text":"Occupancy grids are often used in static NeRFs to reduce the number of points sampled by caching whether a voxel is occupied. Formally, the occupancy grid is a 3D voxel grid G = {max(σ(x)) | ∀x ∈ V cur } 3 , where max(σ(x)) represents the maximum volume density of all points in the respective voxel V cur . When sampling points on a ray, only points within the voxel above a certain volume density threshold are kept. It is obvious that this cannot be directly applied to dynamic scenes since the volume density of the 3D space changes over time. One way of applying this approach in dynamic scene reconstruction is to maintain a space-time 4D occupancy grid [16]. Unfortunately, this does not improve the sampling efficiency when training on-the-fly as the 3D occupancy at t k-1 does not affect the occupancy grid at t k . When training the new frame k, the occupancy grid at the current time t k is the same as being initialized from scratch. To tackle this problem, we follow our Hidden Markov Model formulation to apply a simple transition function and belief update to the occupancy grid. We consider the occupancy grid at any time t k as a 3D probability function\nG k = {P (max(σ k (x)) > 0) | ∀x ∈ V cur } 3 ,\nrepresenting the chance of any point present in the voxel with positive volume density at the current time step. Since the occupancy grid G k is constantly updated throughout the per-frame training, we use G j k to denote the occupancy grid after j iterations of optimization where j ∈ [0, J]. At the start of each new frame optimization, we apply a transition function to this occupancy grid for the possible motions of the objects in the 3D space. This transition function takes the form of a simple 3D convolution kernel S because the occupancy grid is a 3D tensor, such that:\nG 0 k = P (G k | G k-1 ) • G J k-1 = S * G J k-1 .(7)\nSince we formulate the problem with the Markov assumption for higher efficiency, we have little information about the actual motion based on the previous frames. Thus, we assign the kernel with a simple 3D Gaussian function to represent the probability of the motion.\nAfter the transition function is applied to the occupancy grid at the start of the optimization for each frame, it needs to be updated with the new observations. Similar to the existing occupancy grid methods, we apply a simple Bayesian update to the occupancy grid probabilities using the volume density output σ(x) of the NeRF model:\nG j k = P (σ(x) | G j-1 k ) • G j-1 k , P (σ(x) | G) = 1, if σ(x) < G(x) σ(x), otherwise .(8)\nNote that the global update method used in Instant-NGP [13] can also be used to update the occupancy grid. However, the global sampling is not suitable for a constantly changing occupancy grid. The global update method randomly samples points in each occupancy voxel to update the occupancy grid. It ignores the previous occupancy values stored and the transition across frames entirely. Since the model takes much fewer iterations to optimize for each new frame when trained on-the-fly, the random sampling in each voxel can miss the occupied region and leave a blank cube in the rendered image, as shown in Fig. 4.\nFrames "},{"figure_ref":[],"heading":"Experiments","publication_ref":["b17","b5","b7"],"table_ref":[],"text":"We demonstrate the efficient on-the-fly training capability of our method on both synthetic and realworld novel view synthesis datasets. We have explained in Sec. 3 that we focus on scenes captured by multi-view forward-facing cameras since it is the most realistic setting in streaming applications.\nFor the synthetic dataset, we evaluate on the widely used D-NeRF [18] dataset. However, the original D-NeRF dataset is captured with unrealistic teleporting cameras, and thus we render a forward-facing version using Blender for training and testing instead. This dataset will be released and more details are included in the supplementary. For the real-world dataset, we evaluate on the MeetRoom [6] and DyNeRF [8] datasets, which are both captured with multi-view forward-facing cameras. All results of our method are reported for models trained with a single RTX3090 GPU. Some rendered videos are included in the supplementary."},{"figure_ref":["fig_3","fig_4"],"heading":"Synthetic Dataset","publication_ref":["b3","b26","b3","b3"],"table_ref":[],"text":"On the synthetic dataset, we implement our method on top of the TiNeuVox [4] and evaluate the improvements in on-the-fly training speed and rendering quality. We also compare the performance of our method against the reported results of many other baseline models as shown in Tab. 1, but most of these models are trained jointly instead of trained on-the-fly. Our model is trained on the first frame for 200 iterations (for around 2 seconds), and then optimized for just 10 iterations per frame on-the-fly. Compared to the baseline methods, our model has significantly faster training speed (12 FPS) and rendering speed(15 FPS), while achieving a superior rendering quality measured in PSNR and LPIPS [27]. Our model can be trained and rendered at a total of 6.78 FPS, which makes it possible to be used for many interactive applications.\nAs shown in Fig. 5, we also demonstrate some qualitative results of our method compared to TiNeuVox [4]. We compare the training speed and rendering quality when trained to a comparable quality or under similar time constraints. Our model can achieve comparable novel view synthesis quality with significantly faster on-the-fly training, or render with much superior quality under the same time constraint. We present an ablation study in Tab. 2 to better analyze the effectiveness of the projected color guidance component and occupancy grid transition component proposed. We remove the projected color input and the occupancy transition function one by one for evaluation. All ablation models are trained for the same 10 iterations per frame. The ablation results suggest that both components contribute towards a fast convergence during on-the-fly training. The occupancy transition slightly reduces the average training speed as more points are sampled at the start of per-frame optimization due to the transited occupancy grid, but significantly improves the quality.\nTo better compare the performance of our proposed methods and the baseline model under different time constraints, we illustrate the plot showing their rendering quality (PSNR) against their training time in Fig. 6. Our proposed OD-NeRFs perform significantly better than the TiNeuVox [4] baseline when at very strict time constraints. As the training time increases, the rendering quality gap between our method and the baseline reduces. It is worth noting that the performance of the proposed model deteriorates slightly when given more than around 0.15 seconds per frame to train, possibly caused by over-fitting to the last frame. "},{"figure_ref":["fig_5","fig_6"],"heading":"Real-World Datasets","publication_ref":["b6","b12","b5","b5","b12","b5","b7"],"table_ref":[],"text":"For the real-world datasets, our model is trained on the first frame for 6000 iterations and later 100 iterations per frame. We compare the training speed and rendering quality of our method implemented on top of the NerfAcc [7] implementation of InstantNGP [13] against various baseline models used for dynamic and static NeRFs. Although some of the models (e.g. StreamRF [6]) do not claim the on-the-fly training ability, they are compatible with the on-the-fly training proposed in our paper. As shown in Tab. 3, our model can be trained on-the-fly significantly faster than the baseline models while maintaining a similar rendering quality.\nWe also present some of the qualitative results of our model compared to StreamRF [6] and Instant-NGP [13] as shown in Fig. 7 on MeetRoom [6] dataset and in Fig. 8 on DyNeRF [8] "},{"figure_ref":[],"heading":"Limitations","publication_ref":[],"table_ref":[],"text":"The implicit correspondence of our projected color-guided NeRF relies on the relative invariance of the projected color of a point. However, this invariance can be violated with specular surfaces and occluded points. It may be possible to filter out the outlier projected colors caused by specularity and occlusion, or explicitly detect occlusion. However, this process may incur significant computation costs and can be further analyzed in future works."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"We introduced a new on-the-fly dynamic NeRF training setting, where the radiance field is trained and rendered frame by frame. To tackle the efficient on-the-fly training challenge, we propose a projected color-guided dynamic NeRF conditioned on the spatial and color input to efficiently optimize the radiance field with implicit point correspondence. We also propose a transition and update function to the occupancy grid for efficient point sampling in space. The experiment results in both synthetic and real-world datasets indicate the superior on-the-fly training speed of our method while maintaining a comparable rendering quality."},{"figure_ref":[],"heading":"Appendices A Qualitative Result Videos","publication_ref":[],"table_ref":[],"text":"We include a few videos rendered by our model and baselines in the supplementary zip file."},{"figure_ref":[],"heading":"B Implementation Details","publication_ref":["b3","b12"],"table_ref":[],"text":"We implement our model on top of the TiNeuVox [4] for the synthetic dataset and the InstantNGP [13] for the real-world dataset. We describe the changes we have made to the models in this section."},{"figure_ref":[],"heading":"B.1 Synthetic Dataset Model","publication_ref":["b3","b12","b11"],"table_ref":[],"text":"As we have mentioned in the main paper, we remove the temporal components of the model because of its poor extrapolation capability. More specifically, the temporal deformation model and the temporal information enhancement are removed. Instead, the mean and variance projected color of the sampled point is concatenated with its spatial feature as the input to the NeRF Multi-Layer Perceptron(MLP). We also replace the multi-scale voxel used in TiNeuVox [4] with the hash voxel used in InstantNGP [13], implemented with tiny-cuda-nn [12]. We observe that this hashed voxel can better capture the details, but converge slower than the original multi-scale voxel. Hence we added the 2-second warm-up for the first frame as mentioned in the main paper. The rest of the model structure is following the TiNeuVox-S version published.\nSince the original TiNeuVox sample uniformly along the ray instead of sampling based on the occupancy grid, we implement a rejection sampling based on our transited and updated occupancy grid. The rejection sampling filters the uniform samples based on the occupancy grid and a fixed interval. The ith sample x i along the ray is rejected if the occupancy value from the occupancy grid σ occ (x i ) is smaller than a density threshold σ min and it is not on a fixed interval R:\nreject x i if (σ occ (x i ) < σ min ) and ¬(i mod R ≡ R//2).(9)\nWe fix the interval R with a value of 20, and gradually decrease the threshold σ min over the optimization process from 1(at the frame 1) to 0.05(at frame 10) for better convergence at the start."},{"figure_ref":[],"heading":"B.2 Real-world Dataset Model","publication_ref":["b6","b12","b6","b12","b5","b7","b5","b7","b12"],"table_ref":[],"text":"The NerfAcc [7] implementation of InstantNGP [13] is used as the code base for our implementation. We concatenate the mean and variance of the projected colors of the sampled point with its spatial feature queried from the hash voxel grid, before inputting them into the NeRF MLP. We also notice that the NerfAcc [7] implementation of InstantNGP [13] does not converge very well on the forward facing dynamic dataset of MeetRoom [6] and DyNeRF [8], even for the first frame static scene. It could be because of the large planer background with constant colors, like walls and tables. Hence, we implement a simple depth smoothness regularization based on patch sampling. For any 3 × 3 patch of ray sampled, the depth regularization loss is calculated as:\nL depth = std(d f ar /d 3x3 ) std(c 3x3 ) ,(10)\nwhere std represents the standard deviation, d 3x3 represents the depth values of the patch, c 3x3 represents the ground truth color of the patch and d f ar represents the far plane depth. This depth smoothness loss penalizes local large inverse depth variation when the color variation is small. The loss is added to the total loss with a weight of 1e -4 for MeetRoom [6] dataset and 1e -6 for DyNeRF [8] dataset.\nSince the InstantNGP [13] model already has a sampling strategy based on the occupancy grid, we only update the occupancy grid itself during the training and do not change the sampling strategy itself."},{"figure_ref":[],"heading":"C Modifications to D-NeRF Dataset","publication_ref":["b17"],"table_ref":[],"text":"As we have mentioned in main paper, we use a multi-view forward facing camera version of the D-NeRF [18] "},{"figure_ref":[],"heading":"D Additional Ablation Results","publication_ref":[],"table_ref":[],"text":"To better analyze the effectiveness of the projected color mean and variance individually, we illustrate an additional ablation study using only the mean or variance of the projected color in Tab. 4. This quantitative comparison suggests that the mean of projected color increases the performance significantly. Using the projected color variance further improves the performance slightly."},{"figure_ref":["fig_7","fig_16"],"heading":"E Additional Qualitative Results","publication_ref":[],"table_ref":[],"text":"We present some additional qualitative results for both the synthetic and real-world dataset. For the synthetic dataset, we present qualitative comparisons of models trained to a comparable rendering quality but different time(Fig. 9, 10, 11 and 12), and trained under similar time constraint but with different rendering quality(Fig. 13, 14, 15 and 16). For the real-world dataset, we show more rendering results of different scenes(Fig. 17 "}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b10\",\n \"b7\",\n \"b15\",\n \"b17\",\n \"b14\",\n \"b12\",\n \"b0\",\n \"b7\",\n \"b15\",\n \"b17\",\n \"b14\",\n \"b12\",\n \"b0\",\n \"b17\",\n \"b5\",\n \"b7\"\n ],\n \"table_ref\": [],\n \"text\": \"Neural radiance fields (NeRF) [11] have recently emerged as a new 3D volumetric representation capable of synthesizing photo-realistic novel views from multi-view images. Many dynamic NeRFs are also proposed to reconstruct scenes with moving or deforming objects, using space-time 4D [8,16] or explicit deformation [18,15] representations. However, most of the dynamic NeRFs focus on reconstructing the scene given the full training videos of a dynamic event. Although capable of playing back the dynamic scene after the event, they cannot stream the event while it is happening. In this paper, we introduce a new \\\"on-the-fly training\\\" approach to train a dynamic NeRF concurrently with the acquisition of the training frames, as the dynamic event unfolds. For the purpose of streaming the dynamic scene, the model only needs to render the current radiance field anytime during the training. With the recent success in NeRF acceleration to train [13,1] a small scene in seconds, it becomes a promising possibility to further speed up the dynamic scene on-the-fly training to an interactive speed. This enables the streaming of dynamic scenes with a wide variety of applications in social media, VR/AR, and gaming industries.\\nTraining a dynamic NeRF on-the-fly is considerably different from training it after the dynamic event. For training after the event, the NeRF is trained to fit the 4D space-time radiance field with the available full training videos. In contrast, for training on-the-fly, NeRF is trained to represent the current radiance field given the previous reconstruction and the available video frames up to the current time step. As the dynamic event continues, new frames will become available and the NeRF will be updated. Training with each frame from scratch is very time-consuming, so the key to efficient on-the-fly dynamic NeRF training is to effectively utilize the radiance fields estimated from the previous frames for the faster convergence of the current radiance field. By leveraging the correspondence and transition across frames, we propose: 1) a NeRF representation conditioned on multi-view projected colors for faster convergence, and 2) a transition and update to the occupancy grid used for efficient sampling.\\nOur proposed NeRF representation conditioned on multi-view projected colors is designed for fast convergence when training the radiance fields of consecutive frames on-the-fly. Most of the existing dynamic NeRFs explicitly represent the motion using a temporal input dimension [8,16] in NeRF or a temporal deformation field [18,15]. Due to the lack of direct supervision or known dynamics, these models with temporal input often extrapolate poorly to the unseen time and require many iterations of optimization for each frame. Instead of being conditioned on time, we propose a model that is conditioned on the projected colors from the training views. This is based on the observation that the projected color of a corresponding 3D point often stays unchanged in consecutive frames. With guidance from the invariant projected colors, the NeRF model is implicitly aware of the point correspondence across consecutive frames. Consequently, the NeRF model can effectively utilize the radiance fields from the previous frames to render for the current frame with the implicit correspondence. The experiments suggest that our proposed model has excellent temporal extrapolation capability and requires minimum number of optimization iterations for each new frame. As a result, our on-the-fly training speed of the model is significantly improved.\\nAdditionally, we introduce a method of transiting and updating the occupancy grid used for efficient point sampling. Since most of the 3D scenes are empty spaces, the occupancy grid has been used in static NeRFs [13,1] to reduce the number of sampled points for acceleration. To adapt the occupancy grid to on-the-fly dynamic training, we consider the occupancy grid as a probability of the 3D voxels occupied by any object. To probabilistically model the motions in the scene, a transition function is applied to occupancy probability at the start of each frame optimization and later updated with new observations. The updated occupancy grid can then be used to sample points only in the occupied areas anytime during the on-the-fly training.\\nWe evaluate our method on synthetic D-NeRF dataset [18], real-world MeetRoom [6], and DyNeRF [8] dataset. When trained and rendered on-the-fly, our method achieves significant acceleration compared to the state-of-the-art algorithms while maintaining a comparable rendering quality. Particularly, our method can train and render 6 frames per second (FPS) on the synthetic D-NeRF dataset. We summarize our contributions as follows: 1) we introduce and formally formulate the new setting of training dynamic NeRF on-the-fly. 2) We propose a projected color-guided on-the-fly dynamic NeRF and a transiting occupancy grid for efficient on-the-fly training. 3) We achieve 10× on-the-fly training and rendering acceleration in synthetic scenes and 3× acceleration in real-world scenes compared to the state-of-the-art.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Works\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Accelerated Dynamic Neural Radiance Fields\",\n \"publication_ref\": [\n \"b10\",\n \"b22\",\n \"b25\",\n \"b17\",\n \"b7\",\n \"b14\",\n \"b24\",\n \"b3\",\n \"b12\",\n \"b5\",\n \"b25\",\n \"b15\",\n \"b19\"\n ],\n \"table_ref\": [],\n \"text\": \"Following the success in representing static 3D scenes with Neural Radiance Fields(NeRFs) [11,23,26], many works [18,8,15,25] have been exploring representing dynamic scenes with NeRFs as well. Similar to the early static NeRFs, they usually take at least hours to train for a single scene. To improve the training efficiency of dynamic NeRFs, some works try to migrate the acceleration methods used for static NeRFs to dynamic NeRFs. TiNeuVox [4] follows the voxel and Multi-Layer Perceptron(MLP) architecture as in [13] to represent static canonical space, and uses another MLP to capture temporal deformation. StreamRF [6] represents the scene at the first frame using the Plenoxel [26], and learns the subsequent changes to this voxel grid. [16,20] directly expand the 3D representation used in static NeRF to 4D with a time dimension to represent dynamic scenes. However, most of the existing works only focus on joint training of the dynamic scenes, instead of on-the-fly training introduced in our work. Their per frame training speed is also not fast enough for interactive on-the-fly applications.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Image Based Rendering\",\n \"publication_ref\": [\n \"b23\",\n \"b1\",\n \"b4\",\n \"b18\",\n \"b9\",\n \"b21\",\n \"b20\",\n \"b8\"\n ],\n \"table_ref\": [],\n \"text\": \"Instead of representing a 3D scene as a NeRF conditioned on spatial coordinates and viewing direction, some works rely on additional projected colors/features on the training views to improve generalization or robustness. IBRNet [24], MVSNeRF [2] and many following works [5,19,10] construct a cost volume of a dynamic scene based on the image features of the nearby views to learn a blending weights or the density and color output. These models are designed to be generalizable to unseen scenes as they do not rely on the spatial coordinates input. LLFF [22,21] aggregates features along the multi-view epipolar line to render view dependent effect. Recently, DynIBaR [9] applies this technique to dynamic NeRF by aggregating the projected image features across frames, after warpping a point using a motion trajectory learned from past and future frames. However, this method requires time-consuming optimization of the per-frame motion trajectory and cannot be applied to on-the-fly training as the future frames are not known.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Problem Formulation\",\n \"publication_ref\": [\n \"b10\",\n \"b17\",\n \"b14\",\n \"b7\",\n \"b15\",\n \"b13\",\n \"b16\"\n ],\n \"table_ref\": [],\n \"text\": \"In this section, we first briefly describe the NeRF preliminaries necessary to understand our formulation. We then formally define our new on-the-fly dynamic NeRF training and followed by simplifying the problem formulation in the form of a Hidden Markov model (HMM).\\nNeRF Preliminaries. Static neural radiance fields [11] represent a 3D scene implicitly with a continuous volumetric function F : (x, d) → (σ, c) that maps the spatial position x ∈ R 3 and viewing direction d ∈ R 3 to the volume density σ ∈ R and RGB color c ∈ R 3 . To synthesize the pixel color Ĉ(r) on any 2D images, volume rendering is used to aggregate the color of N points with interval δ along the ray r shooting from the pixel:\\nĈ(r) = N -1 i=0 T i (1 -exp(-σ i δ i ))c i , T i = exp(- i-1 j=0 σ j δ j ).(1)\\nDynamic NeRFs usually include an additional time dimension t in the input, through either a timevarying deformation field D : (x, t) → (x ′ ) that maps the spatial coordinates x to its canonical space correspondence x ′ [18,15], or directly expanding the 3D NeRF model to a 4D variant [8,16]. During the training stage, all K frames of the training images C 0:K (r) from all time t 0:K are used to jointly minimize the rendering loss:\\nF ′ : (x, d, t) → (σ, c)\\nL = K t=0 r∈R || Ĉt (r) -C t (r)|| 2 2 .(2)\\nOn-the-fly Dynamic NeRF. We introduce a new on-the-fly training of the dynamic NeRF suitable for streaming dynamic scenes. Instead of training after the dynamic event, we train the NeRF concurrently while the dynamic event unfolds. Furthermore, the NeRF trained at time t k renders novel views only at time t k . Given the images up to the current time step C 0:k and the radiance field estimated up to the last time step F 0:k-1 , the goal of on-the-fly dynamic NeRF training is to find the radiance field function F k at time t k that minimizes the rendering loss at the current time step:\\nF k (x, d) = argmax F k P (F k | C 0:k , F 0:k-1 ) = argmin F k r∈R || Ĉk (r) -C k (r)|| 2 2 .\\n(3)\\nHowever, estimating the radiance field F k at the current time step conditioned on all the previous images C 0:k and radiance field F 0:k-1 is not scalable as time t k increases. To mitigate this growth in complexity, we apply the first order Markov assumption to simplify the probability model by assuming conditional independence\\nF k ⊥ ⊥ {C 0:k-1 , F 0:k-2 } | {C k , F k-1 }.\\nThis simplifies the estimation of the current radiance field as:\\nF k = argmax F k P (F k | C k , F k-1 ).(4)\\nTaking the radiance fields F 0:K as hidden states and images C 0:K as observations, we can formulate the on-the-fly training as the process of estimating the hidden states in a Hidden Markov model (HMM). The emission function\\nP (C k | F k )\\nis the process of volumetric rendering that renders 2D images from 3D radiance fields. The transition function\\nP (F k | F k-1\\n) is the radiance field deformation or motion between two consecutive time steps.\\nBased on the formulation above, the key to efficient on-the-fly training is to maintain low complexity of the update\\nP (F k | C k , F k-1\\n) to the radiance field at each time step. To this end, we propose a projected color-guided NeRF that is implicitly aware of point correspondence that requires minimum optimization when transiting from F k-1 to F k for fast training. Furthermore, we also introduce a simple transition function P (G k |G k-1 ) to the occupancy grid G used for efficient sampling.\\nRemarks. Note that we limit the scope of this work to dynamic scenes captured by multi-view forward-facing cameras based on realistic considerations. Although the reconstruction of dynamic scenes from a monocular camera is less demanding on the hardware, it requires the photographer to keep on moving the camera [14]. This is cumbersome in prolonged streaming scenarios. 360-degree inward-facing cameras can be used to reconstruct from all angles. Nonetheless, this often requires dozens of cameras and a much bigger space [17]. It is difficult for most streamers to acquire such professional setups. Consequently, we focus on scenes captured by static multi-view forward-facing cameras that are most aligned with the setups used in the current streaming industry.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Method\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"As mentioned in the previous section, the goal of on-the-fly training is to estimate the current radiance field F k based on the last radiance field F k-1 and the current training images C k . In practice, this can be achieved by optimizing the radiance field model from the last step with the current training images.\\nTo achieve highly efficient on-the-fly training, we can either reduce the number of optimization iterations needed for each time step or reduce the time spent on each iteration without scarifying the performance. Based on our HMM-based on-the-fly training paradigm, we propose: 1) a dynamic NeRF guided by projected colors, and 2) an occupancy grid transition and update strategy, to achieve these two goals respectively (Fig. 1).\"\n },\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_1\"\n ],\n \"heading\": \"Dynamic NeRF Guided by Projected Colors\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"According to the HMM formulation of on-the-fly training, the optimization at each time step is effectively learning a probability update to the last step model given the current observations:\\nF k = P (F k | F k-1 , C k )F k-1 .\\nIn the context of dynamic NeRFs, this can usually be achieved by estimating a deformation field or point correspondence function.\\nExisting deformation-based dynamic NeRFs try to learn a deformation field D : (x, t k ) → (x ′ ) that captures the point correspondence through joint training. However, our experiments indicate that this process itself takes too long for efficient on-the-fly training. We postulate that this is because the Figure 2: We illustrate the trajectory of a blue ball moving on a (x, y) over time(left). When mapping it to the \\\"space-time\\\" coordinate system, the point correspondence can be hard to track due to the irregular trajectory(middle, 2D space represented with x + y for simpler visualization). When mapping it to the \\\"spacecolor\\\" coordinate system, the correspondence can be easily tracked due to the invariant color(right).\\ntemporal deformation field does not have a clear pattern in most cases. We illustrate this problem with a toy example as shown in Fig. 2. In this example, we track the center of the 2D moving ball of a constant color. When tracking the trajectory of the ball in \\\"space-time\\\" input space, it is difficult to accurately predict the location of the ball at the next time step. This is similar to the temporal deformation field used by many existing dynamic NeRFs. The deformation parameters need to be optimized at each time step based on the indirect RGB supervision. In the case of 3D neural radiance field reconstruction, it is very time-consuming to optimize for all points in the space.\\nTo efficiently utilize the point correspondence across frames, we change the time input t in the space-time dynamic NeRF model F (x, d, t) to the multi-view projected color mean cp (x, k) and variance Var(c p (x, k)) at frame k to become:\\nF k : (x, d, cp (x, k), Var(c p (x, k))) → (σ, c).(5)\\nThe projected colors c p (x, k) are defined as the pixel colors of the training images C k,cam of all cameras M through 3D to 2D projection with camera projection matrix P cam :\\nc p (x, k) = {C k,cam (P cam • x) | cam ∈ M}.(6)\\nAs the model is conditioned on the spatial coordinates and projected colors, it circumvents the difficulty of estimating the point correspondence with irregularity in the spatial and temporal dimensions. Instead, it is much easier to estimate the correspondence in the spatial and color input space due to the invariance of the point color under motion. As illustrated in Fig. 2, the trajectory of the point in the spatial and color input space is much more regular than that in the spatial and temporal input space. Although the projected colors of 3D points are not perfectly invariant due to factors such as reflection, lighting, and occlusion, we observe that the projected colors stay largely similar across consecutive frames.\\nTo verify that our method is better at estimating point correspondence on-the-fly, we demonstrate a simple extrapolation experiment in Fig. 3. Specifically, we train the different dynamic NeRFs on-thefly up to the (k -1)th frame and then extrapolate to render the kth frame without any training on the kth frame. As shown in Fig. 3, the model operating in the spatial and temporal input dimensions extrapolates poorly as it is not aware of the point correspondence with untrained time. Our model operating in the spatial and projected color input dimensions can extrapolate well based on the point correspondence hinted by the projected colors. With the implicit point correspondence and the good extrapolation ability, our model requires a low number of iterations for each new frame and thus is very fast when training on-the-fly.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Occupancy Grid Transition and Update\",\n \"publication_ref\": [\n \"b15\",\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"Occupancy grids are often used in static NeRFs to reduce the number of points sampled by caching whether a voxel is occupied. Formally, the occupancy grid is a 3D voxel grid G = {max(σ(x)) | ∀x ∈ V cur } 3 , where max(σ(x)) represents the maximum volume density of all points in the respective voxel V cur . When sampling points on a ray, only points within the voxel above a certain volume density threshold are kept. It is obvious that this cannot be directly applied to dynamic scenes since the volume density of the 3D space changes over time. One way of applying this approach in dynamic scene reconstruction is to maintain a space-time 4D occupancy grid [16]. Unfortunately, this does not improve the sampling efficiency when training on-the-fly as the 3D occupancy at t k-1 does not affect the occupancy grid at t k . When training the new frame k, the occupancy grid at the current time t k is the same as being initialized from scratch. To tackle this problem, we follow our Hidden Markov Model formulation to apply a simple transition function and belief update to the occupancy grid. We consider the occupancy grid at any time t k as a 3D probability function\\nG k = {P (max(σ k (x)) > 0) | ∀x ∈ V cur } 3 ,\\nrepresenting the chance of any point present in the voxel with positive volume density at the current time step. Since the occupancy grid G k is constantly updated throughout the per-frame training, we use G j k to denote the occupancy grid after j iterations of optimization where j ∈ [0, J]. At the start of each new frame optimization, we apply a transition function to this occupancy grid for the possible motions of the objects in the 3D space. This transition function takes the form of a simple 3D convolution kernel S because the occupancy grid is a 3D tensor, such that:\\nG 0 k = P (G k | G k-1 ) • G J k-1 = S * G J k-1 .(7)\\nSince we formulate the problem with the Markov assumption for higher efficiency, we have little information about the actual motion based on the previous frames. Thus, we assign the kernel with a simple 3D Gaussian function to represent the probability of the motion.\\nAfter the transition function is applied to the occupancy grid at the start of the optimization for each frame, it needs to be updated with the new observations. Similar to the existing occupancy grid methods, we apply a simple Bayesian update to the occupancy grid probabilities using the volume density output σ(x) of the NeRF model:\\nG j k = P (σ(x) | G j-1 k ) • G j-1 k , P (σ(x) | G) = 1, if σ(x) < G(x) σ(x), otherwise .(8)\\nNote that the global update method used in Instant-NGP [13] can also be used to update the occupancy grid. However, the global sampling is not suitable for a constantly changing occupancy grid. The global update method randomly samples points in each occupancy voxel to update the occupancy grid. It ignores the previous occupancy values stored and the transition across frames entirely. Since the model takes much fewer iterations to optimize for each new frame when trained on-the-fly, the random sampling in each voxel can miss the occupied region and leave a blank cube in the rendered image, as shown in Fig. 4.\\nFrames \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [\n \"b17\",\n \"b5\",\n \"b7\"\n ],\n \"table_ref\": [],\n \"text\": \"We demonstrate the efficient on-the-fly training capability of our method on both synthetic and realworld novel view synthesis datasets. We have explained in Sec. 3 that we focus on scenes captured by multi-view forward-facing cameras since it is the most realistic setting in streaming applications.\\nFor the synthetic dataset, we evaluate on the widely used D-NeRF [18] dataset. However, the original D-NeRF dataset is captured with unrealistic teleporting cameras, and thus we render a forward-facing version using Blender for training and testing instead. This dataset will be released and more details are included in the supplementary. For the real-world dataset, we evaluate on the MeetRoom [6] and DyNeRF [8] datasets, which are both captured with multi-view forward-facing cameras. All results of our method are reported for models trained with a single RTX3090 GPU. Some rendered videos are included in the supplementary.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\",\n \"fig_4\"\n ],\n \"heading\": \"Synthetic Dataset\",\n \"publication_ref\": [\n \"b3\",\n \"b26\",\n \"b3\",\n \"b3\"\n ],\n \"table_ref\": [],\n \"text\": \"On the synthetic dataset, we implement our method on top of the TiNeuVox [4] and evaluate the improvements in on-the-fly training speed and rendering quality. We also compare the performance of our method against the reported results of many other baseline models as shown in Tab. 1, but most of these models are trained jointly instead of trained on-the-fly. Our model is trained on the first frame for 200 iterations (for around 2 seconds), and then optimized for just 10 iterations per frame on-the-fly. Compared to the baseline methods, our model has significantly faster training speed (12 FPS) and rendering speed(15 FPS), while achieving a superior rendering quality measured in PSNR and LPIPS [27]. Our model can be trained and rendered at a total of 6.78 FPS, which makes it possible to be used for many interactive applications.\\nAs shown in Fig. 5, we also demonstrate some qualitative results of our method compared to TiNeuVox [4]. We compare the training speed and rendering quality when trained to a comparable quality or under similar time constraints. Our model can achieve comparable novel view synthesis quality with significantly faster on-the-fly training, or render with much superior quality under the same time constraint. We present an ablation study in Tab. 2 to better analyze the effectiveness of the projected color guidance component and occupancy grid transition component proposed. We remove the projected color input and the occupancy transition function one by one for evaluation. All ablation models are trained for the same 10 iterations per frame. The ablation results suggest that both components contribute towards a fast convergence during on-the-fly training. The occupancy transition slightly reduces the average training speed as more points are sampled at the start of per-frame optimization due to the transited occupancy grid, but significantly improves the quality.\\nTo better compare the performance of our proposed methods and the baseline model under different time constraints, we illustrate the plot showing their rendering quality (PSNR) against their training time in Fig. 6. Our proposed OD-NeRFs perform significantly better than the TiNeuVox [4] baseline when at very strict time constraints. As the training time increases, the rendering quality gap between our method and the baseline reduces. It is worth noting that the performance of the proposed model deteriorates slightly when given more than around 0.15 seconds per frame to train, possibly caused by over-fitting to the last frame. \"\n },\n {\n \"figure_ref\": [\n \"fig_5\",\n \"fig_6\"\n ],\n \"heading\": \"Real-World Datasets\",\n \"publication_ref\": [\n \"b6\",\n \"b12\",\n \"b5\",\n \"b5\",\n \"b12\",\n \"b5\",\n \"b7\"\n ],\n \"table_ref\": [],\n \"text\": \"For the real-world datasets, our model is trained on the first frame for 6000 iterations and later 100 iterations per frame. We compare the training speed and rendering quality of our method implemented on top of the NerfAcc [7] implementation of InstantNGP [13] against various baseline models used for dynamic and static NeRFs. Although some of the models (e.g. StreamRF [6]) do not claim the on-the-fly training ability, they are compatible with the on-the-fly training proposed in our paper. As shown in Tab. 3, our model can be trained on-the-fly significantly faster than the baseline models while maintaining a similar rendering quality.\\nWe also present some of the qualitative results of our model compared to StreamRF [6] and Instant-NGP [13] as shown in Fig. 7 on MeetRoom [6] dataset and in Fig. 8 on DyNeRF [8] \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The implicit correspondence of our projected color-guided NeRF relies on the relative invariance of the projected color of a point. However, this invariance can be violated with specular surfaces and occluded points. It may be possible to filter out the outlier projected colors caused by specularity and occlusion, or explicitly detect occlusion. However, this process may incur significant computation costs and can be further analyzed in future works.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We introduced a new on-the-fly dynamic NeRF training setting, where the radiance field is trained and rendered frame by frame. To tackle the efficient on-the-fly training challenge, we propose a projected color-guided dynamic NeRF conditioned on the spatial and color input to efficiently optimize the radiance field with implicit point correspondence. We also propose a transition and update function to the occupancy grid for efficient point sampling in space. The experiment results in both synthetic and real-world datasets indicate the superior on-the-fly training speed of our method while maintaining a comparable rendering quality.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Appendices A Qualitative Result Videos\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We include a few videos rendered by our model and baselines in the supplementary zip file.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B Implementation Details\",\n \"publication_ref\": [\n \"b3\",\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"We implement our model on top of the TiNeuVox [4] for the synthetic dataset and the InstantNGP [13] for the real-world dataset. We describe the changes we have made to the models in this section.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.1 Synthetic Dataset Model\",\n \"publication_ref\": [\n \"b3\",\n \"b12\",\n \"b11\"\n ],\n \"table_ref\": [],\n \"text\": \"As we have mentioned in the main paper, we remove the temporal components of the model because of its poor extrapolation capability. More specifically, the temporal deformation model and the temporal information enhancement are removed. Instead, the mean and variance projected color of the sampled point is concatenated with its spatial feature as the input to the NeRF Multi-Layer Perceptron(MLP). We also replace the multi-scale voxel used in TiNeuVox [4] with the hash voxel used in InstantNGP [13], implemented with tiny-cuda-nn [12]. We observe that this hashed voxel can better capture the details, but converge slower than the original multi-scale voxel. Hence we added the 2-second warm-up for the first frame as mentioned in the main paper. The rest of the model structure is following the TiNeuVox-S version published.\\nSince the original TiNeuVox sample uniformly along the ray instead of sampling based on the occupancy grid, we implement a rejection sampling based on our transited and updated occupancy grid. The rejection sampling filters the uniform samples based on the occupancy grid and a fixed interval. The ith sample x i along the ray is rejected if the occupancy value from the occupancy grid σ occ (x i ) is smaller than a density threshold σ min and it is not on a fixed interval R:\\nreject x i if (σ occ (x i ) < σ min ) and ¬(i mod R ≡ R//2).(9)\\nWe fix the interval R with a value of 20, and gradually decrease the threshold σ min over the optimization process from 1(at the frame 1) to 0.05(at frame 10) for better convergence at the start.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.2 Real-world Dataset Model\",\n \"publication_ref\": [\n \"b6\",\n \"b12\",\n \"b6\",\n \"b12\",\n \"b5\",\n \"b7\",\n \"b5\",\n \"b7\",\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"The NerfAcc [7] implementation of InstantNGP [13] is used as the code base for our implementation. We concatenate the mean and variance of the projected colors of the sampled point with its spatial feature queried from the hash voxel grid, before inputting them into the NeRF MLP. We also notice that the NerfAcc [7] implementation of InstantNGP [13] does not converge very well on the forward facing dynamic dataset of MeetRoom [6] and DyNeRF [8], even for the first frame static scene. It could be because of the large planer background with constant colors, like walls and tables. Hence, we implement a simple depth smoothness regularization based on patch sampling. For any 3 × 3 patch of ray sampled, the depth regularization loss is calculated as:\\nL depth = std(d f ar /d 3x3 ) std(c 3x3 ) ,(10)\\nwhere std represents the standard deviation, d 3x3 represents the depth values of the patch, c 3x3 represents the ground truth color of the patch and d f ar represents the far plane depth. This depth smoothness loss penalizes local large inverse depth variation when the color variation is small. The loss is added to the total loss with a weight of 1e -4 for MeetRoom [6] dataset and 1e -6 for DyNeRF [8] dataset.\\nSince the InstantNGP [13] model already has a sampling strategy based on the occupancy grid, we only update the occupancy grid itself during the training and do not change the sampling strategy itself.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C Modifications to D-NeRF Dataset\",\n \"publication_ref\": [\n \"b17\"\n ],\n \"table_ref\": [],\n \"text\": \"As we have mentioned in main paper, we use a multi-view forward facing camera version of the D-NeRF [18] \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D Additional Ablation Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To better analyze the effectiveness of the projected color mean and variance individually, we illustrate an additional ablation study using only the mean or variance of the projected color in Tab. 4. This quantitative comparison suggests that the mean of projected color increases the performance significantly. Using the projected color variance further improves the performance slightly.\"\n },\n {\n \"figure_ref\": [\n \"fig_7\",\n \"fig_16\"\n ],\n \"heading\": \"E Additional Qualitative Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We present some additional qualitative results for both the synthetic and real-world dataset. For the synthetic dataset, we present qualitative comparisons of models trained to a comparable rendering quality but different time(Fig. 9, 10, 11 and 12), and trained under similar time constraint but with different rendering quality(Fig. 13, 14, 15 and 16). For the real-world dataset, we show more rendering results of different scenes(Fig. 17 \"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":""},"doi":{"kind":"string","value":""},"references":{"kind":"list like","value":[{"authors":"Anpei Chen; Zexiang Xu; Andreas Geiger; Jingyi Yu; Hao Su","journal":"","ref_id":"b0","title":"Tensorf: Tensorial radiance fields","year":"2022"},{"authors":"Anpei Chen; Zexiang Xu; Fuqiang Zhao; Xiaoshuai Zhang; Fanbo Xiang; Jingyi Yu; Hao Su","journal":"","ref_id":"b1","title":"Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo","year":"2021"},{"authors":"Boyang Deng; Jonathan T Barron; Pratul P Srinivasan","journal":"","ref_id":"b2","title":"JaxNeRF: an efficient JAX implementation of NeRF","year":"2020"},{"authors":"Jiemin Fang; Taoran Yi; Xinggang Wang; Lingxi Xie; Xiaopeng Zhang; Wenyu Liu; Matthias Nießner; Qi Tian","journal":"","ref_id":"b3","title":"Fast dynamic radiance fields with time-aware neural voxels","year":"2022"},{"authors":"Xin Huang; Qi Zhang; Ying Feng; Xiaoyu Li; Xuan Wang; Qing Wang","journal":"","ref_id":"b4","title":"Local implicit ray function for generalizable radiance field representation","year":"2023"},{"authors":"Lingzhi Li; Zhen Shen; Zhongshu Wang; Li Shen; Ping Tan","journal":"","ref_id":"b5","title":"Streaming radiance fields for 3d video synthesis","year":"2022"},{"authors":"Ruilong Li; Hang Gao; Matthew Tancik; Angjoo Kanazawa","journal":"","ref_id":"b6","title":"Nerfacc: Efficient sampling accelerates nerfs","year":"2023"},{"authors":"Tianye Li; Mira Slavcheva; Michael Zollhoefer; Simon Green; Christoph Lassner; Changil Kim; Tanner Schmidt; Steven Lovegrove; Michael Goesele; Richard Newcombe","journal":"","ref_id":"b7","title":"Neural 3d video synthesis from multi-view video","year":"2022"},{"authors":"Zhengqi Li; Qianqian Wang; Forrester Cole; Richard Tucker; Noah Snavely","journal":"","ref_id":"b8","title":"Dynibar: Neural dynamic image-based rendering","year":"2022"},{"authors":"Xiaoxiao Long; Cheng Lin; Peng Wang; Taku Komura; Wenping Wang","journal":"Springer","ref_id":"b9","title":"Sparseneus: Fast generalizable neural surface reconstruction from sparse views","year":"2022"},{"authors":"Ben Mildenhall; P Pratul; Matthew Srinivasan; Jonathan T Tancik; Ravi Barron; Ren Ramamoorthi; Ng","journal":"Communications of the ACM","ref_id":"b10","title":"Nerf: Representing scenes as neural radiance fields for view synthesis","year":"2021"},{"authors":"Thomas Müller","journal":"","ref_id":"b11","title":"tiny-cuda-nn","year":"2021"},{"authors":"Thomas Müller; Alex Evans; Christoph Schied; Alexander Keller","journal":"ACM Trans. Graph","ref_id":"b12","title":"Instant neural graphics primitives with a multiresolution hash encoding","year":"2022-07"},{"authors":"Keunhong Park; Utkarsh Sinha; Jonathan T Barron; Sofien Bouaziz; Dan B Goldman; Steven M Seitz; Ricardo Martin-Brualla","journal":"","ref_id":"b13","title":"Nerfies: Deformable neural radiance fields","year":"2021"},{"authors":"Keunhong Park; Utkarsh Sinha; Peter Hedman; Jonathan T Barron; Sofien Bouaziz; Dan B Goldman; Ricardo Martin-Brualla; Steven M Seitz","journal":"ACM Trans. Graph","ref_id":"b14","title":"Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields","year":"2021-12"},{"authors":"Sungheon Park; Minjung Son; Seokhwan Jang; Young Chun Ahn; Ji-Yeon Kim; Nahyup Kang","journal":"","ref_id":"b15","title":"Temporal interpolation is all you need for dynamic neural radiance fields","year":"2023"},{"authors":"Sida Peng; Yuanqing Zhang; Yinghao Xu; Qianqian Wang; Qing Shuai; Hujun Bao; Xiaowei Zhou","journal":"","ref_id":"b16","title":"Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans","year":"2021"},{"authors":"Albert Pumarola; Enric Corona; Gerard Pons-Moll; Francesc Moreno-Noguer","journal":"","ref_id":"b17","title":"D-nerf: Neural radiance fields for dynamic scenes","year":"2020"},{"authors":"Yufan Ren; Fangjinhua Wang; Tong Zhang; Marc Pollefeys; Sabine Süsstrunk","journal":"","ref_id":"b18","title":"Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction","year":"2022"},{"authors":"Sara Fridovich; -Keil ; Giacomo Meanti; Frederik Rahbaek Warburg; Benjamin Recht; Angjoo Kanazawa","journal":"","ref_id":"b19","title":"K-planes: Explicit radiance fields in space, time, and appearance","year":"2023"},{"authors":"Mohammed Suhail; Carlos Esteves; Leonid Sigal; Ameesh Makadia","journal":"Springer","ref_id":"b20","title":"Generalizable patch-based neural rendering","year":"2022"},{"authors":"Mohammed Suhail; Carlos Esteves; Leonid Sigal; Ameesh Makadia","journal":"","ref_id":"b21","title":"Light field neural rendering","year":"2022"},{"authors":"Peng Wang; Lingjie Liu; Yuan Liu; Christian Theobalt; Taku Komura; Wenping Wang","journal":"NeurIPS","ref_id":"b22","title":"Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction","year":"2021"},{"authors":"Qianqian Wang; Zhicheng Wang; Kyle Genova; Pratul Srinivasan; Howard Zhou; Jonathan T Barron; Ricardo Martin-Brualla; Noah Snavely; Thomas Funkhouser","journal":"","ref_id":"b23","title":"Ibrnet: Learning multi-view image-based rendering","year":"2021"},{"authors":"Zhiwen Yan; Chen Li; Gim Hee; Lee ","journal":"","ref_id":"b24","title":"Nerf-ds: Neural radiance fields for dynamic specular objects","year":"2023"},{"authors":"Alex Yu; Sara Fridovich-Keil; Matthew Tancik; Qinhong Chen; Benjamin Recht; Angjoo Kanazawa","journal":"","ref_id":"b25","title":"Plenoxels: Radiance fields without neural networks","year":"2021"},{"authors":"Richard Zhang; Phillip Isola; Alexei A Efros; Eli Shechtman; Oliver Wang","journal":"","ref_id":"b26","title":"The unreasonable effectiveness of deep features as a perceptual metric","year":"2018"}],"string":"[\n {\n \"authors\": \"Anpei Chen; Zexiang Xu; Andreas Geiger; Jingyi Yu; Hao Su\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Tensorf: Tensorial radiance fields\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Anpei Chen; Zexiang Xu; Fuqiang Zhao; Xiaoshuai Zhang; Fanbo Xiang; Jingyi Yu; Hao Su\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Boyang Deng; Jonathan T Barron; Pratul P Srinivasan\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"JaxNeRF: an efficient JAX implementation of NeRF\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jiemin Fang; Taoran Yi; Xinggang Wang; Lingxi Xie; Xiaopeng Zhang; Wenyu Liu; Matthias Nießner; Qi Tian\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Fast dynamic radiance fields with time-aware neural voxels\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xin Huang; Qi Zhang; Ying Feng; Xiaoyu Li; Xuan Wang; Qing Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Local implicit ray function for generalizable radiance field representation\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Lingzhi Li; Zhen Shen; Zhongshu Wang; Li Shen; Ping Tan\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Streaming radiance fields for 3d video synthesis\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ruilong Li; Hang Gao; Matthew Tancik; Angjoo Kanazawa\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Nerfacc: Efficient sampling accelerates nerfs\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Tianye Li; Mira Slavcheva; Michael Zollhoefer; Simon Green; Christoph Lassner; Changil Kim; Tanner Schmidt; Steven Lovegrove; Michael Goesele; Richard Newcombe\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Neural 3d video synthesis from multi-view video\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Zhengqi Li; Qianqian Wang; Forrester Cole; Richard Tucker; Noah Snavely\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Dynibar: Neural dynamic image-based rendering\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xiaoxiao Long; Cheng Lin; Peng Wang; Taku Komura; Wenping Wang\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b9\",\n \"title\": \"Sparseneus: Fast generalizable neural surface reconstruction from sparse views\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ben Mildenhall; P Pratul; Matthew Srinivasan; Jonathan T Tancik; Ravi Barron; Ren Ramamoorthi; Ng\",\n \"journal\": \"Communications of the ACM\",\n \"ref_id\": \"b10\",\n \"title\": \"Nerf: Representing scenes as neural radiance fields for view synthesis\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Thomas Müller\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"tiny-cuda-nn\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Thomas Müller; Alex Evans; Christoph Schied; Alexander Keller\",\n \"journal\": \"ACM Trans. Graph\",\n \"ref_id\": \"b12\",\n \"title\": \"Instant neural graphics primitives with a multiresolution hash encoding\",\n \"year\": \"2022-07\"\n },\n {\n \"authors\": \"Keunhong Park; Utkarsh Sinha; Jonathan T Barron; Sofien Bouaziz; Dan B Goldman; Steven M Seitz; Ricardo Martin-Brualla\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Nerfies: Deformable neural radiance fields\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Keunhong Park; Utkarsh Sinha; Peter Hedman; Jonathan T Barron; Sofien Bouaziz; Dan B Goldman; Ricardo Martin-Brualla; Steven M Seitz\",\n \"journal\": \"ACM Trans. Graph\",\n \"ref_id\": \"b14\",\n \"title\": \"Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields\",\n \"year\": \"2021-12\"\n },\n {\n \"authors\": \"Sungheon Park; Minjung Son; Seokhwan Jang; Young Chun Ahn; Ji-Yeon Kim; Nahyup Kang\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Temporal interpolation is all you need for dynamic neural radiance fields\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Sida Peng; Yuanqing Zhang; Yinghao Xu; Qianqian Wang; Qing Shuai; Hujun Bao; Xiaowei Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Albert Pumarola; Enric Corona; Gerard Pons-Moll; Francesc Moreno-Noguer\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"D-nerf: Neural radiance fields for dynamic scenes\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yufan Ren; Fangjinhua Wang; Tong Zhang; Marc Pollefeys; Sabine Süsstrunk\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Sara Fridovich; -Keil ; Giacomo Meanti; Frederik Rahbaek Warburg; Benjamin Recht; Angjoo Kanazawa\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"K-planes: Explicit radiance fields in space, time, and appearance\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Mohammed Suhail; Carlos Esteves; Leonid Sigal; Ameesh Makadia\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b20\",\n \"title\": \"Generalizable patch-based neural rendering\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Mohammed Suhail; Carlos Esteves; Leonid Sigal; Ameesh Makadia\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Light field neural rendering\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Peng Wang; Lingjie Liu; Yuan Liu; Christian Theobalt; Taku Komura; Wenping Wang\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b22\",\n \"title\": \"Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Qianqian Wang; Zhicheng Wang; Kyle Genova; Pratul Srinivasan; Howard Zhou; Jonathan T Barron; Ricardo Martin-Brualla; Noah Snavely; Thomas Funkhouser\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Ibrnet: Learning multi-view image-based rendering\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zhiwen Yan; Chen Li; Gim Hee; Lee \",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Nerf-ds: Neural radiance fields for dynamic specular objects\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Alex Yu; Sara Fridovich-Keil; Matthew Tancik; Qinhong Chen; Benjamin Recht; Angjoo Kanazawa\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Plenoxels: Radiance fields without neural networks\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Richard Zhang; Phillip Isola; Alexei A Efros; Eli Shechtman; Oliver Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"The unreasonable effectiveness of deep features as a perceptual metric\",\n \"year\": \"2018\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,183.78,556.48,320.89,30.32],"formula_id":"formula_0","formula_text":"Ĉ(r) = N -1 i=0 T i (1 -exp(-σ i δ i ))c i , T i = exp(- i-1 j=0 σ j δ j ).(1)"},{"formula_coordinates":[3,108,624.92,96.94,10.31],"formula_id":"formula_1","formula_text":"F ′ : (x, d, t) → (σ, c)"},{"formula_coordinates":[3,240.76,653.71,263.91,30.48],"formula_id":"formula_2","formula_text":"L = K t=0 r∈R || Ĉt (r) -C t (r)|| 2 2 .(2)"},{"formula_coordinates":[4,216.82,124.75,178.37,44.96],"formula_id":"formula_3","formula_text":"F k (x, d) = argmax F k P (F k | C 0:k , F 0:k-1 ) = argmin F k r∈R || Ĉk (r) -C k (r)|| 2 2 ."},{"formula_coordinates":[4,258.14,215.19,163.31,9.65],"formula_id":"formula_4","formula_text":"F k ⊥ ⊥ {C 0:k-1 , F 0:k-2 } | {C k , F k-1 }."},{"formula_coordinates":[4,238.74,243.12,265.93,17.19],"formula_id":"formula_5","formula_text":"F k = argmax F k P (F k | C k , F k-1 ).(4)"},{"formula_coordinates":[4,241.39,289.26,50.16,9.65],"formula_id":"formula_6","formula_text":"P (C k | F k )"},{"formula_coordinates":[4,359.83,300.17,55.1,9.65],"formula_id":"formula_7","formula_text":"P (F k | F k-1"},{"formula_coordinates":[4,164.14,338.38,69.58,9.65],"formula_id":"formula_8","formula_text":"P (F k | C k , F k-1"},{"formula_coordinates":[4,108,664.08,123.66,9.65],"formula_id":"formula_9","formula_text":"F k = P (F k | F k-1 , C k )F k-1 ."},{"formula_coordinates":[5,213.9,401.09,290.77,9.79],"formula_id":"formula_10","formula_text":"F k : (x, d, cp (x, k), Var(c p (x, k))) → (σ, c).(5)"},{"formula_coordinates":[5,218.29,448.37,286.38,9.84],"formula_id":"formula_11","formula_text":"c p (x, k) = {C k,cam (P cam • x) | cam ∈ M}.(6)"},{"formula_coordinates":[6,145.4,381.1,191.49,11.38],"formula_id":"formula_12","formula_text":"G k = {P (max(σ k (x)) > 0) | ∀x ∈ V cur } 3 ,"},{"formula_coordinates":[6,220.64,504.12,284.03,12.69],"formula_id":"formula_13","formula_text":"G 0 k = P (G k | G k-1 ) • G J k-1 = S * G J k-1 .(7)"},{"formula_coordinates":[6,149.99,614.75,354.68,22.05],"formula_id":"formula_14","formula_text":"G j k = P (σ(x) | G j-1 k ) • G j-1 k , P (σ(x) | G) = 1, if σ(x) < G(x) σ(x), otherwise .(8)"},{"formula_coordinates":[12,182.19,387.46,322.48,9.68],"formula_id":"formula_15","formula_text":"reject x i if (σ occ (x i ) < σ min ) and ¬(i mod R ≡ R//2).(9)"},{"formula_coordinates":[12,251.84,544.4,252.83,23.22],"formula_id":"formula_16","formula_text":"L depth = std(d f ar /d 3x3 ) std(c 3x3 ) ,(10)"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 183.78,\n 556.48,\n 320.89,\n 30.32\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"Ĉ(r) = N -1 i=0 T i (1 -exp(-σ i δ i ))c i , T i = exp(- i-1 j=0 σ j δ j ).(1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 108,\n 624.92,\n 96.94,\n 10.31\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"F ′ : (x, d, t) → (σ, c)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 240.76,\n 653.71,\n 263.91,\n 30.48\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"L = K t=0 r∈R || Ĉt (r) -C t (r)|| 2 2 .(2)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 216.82,\n 124.75,\n 178.37,\n 44.96\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"F k (x, d) = argmax F k P (F k | C 0:k , F 0:k-1 ) = argmin F k r∈R || Ĉk (r) -C k (r)|| 2 2 .\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 258.14,\n 215.19,\n 163.31,\n 9.65\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"F k ⊥ ⊥ {C 0:k-1 , F 0:k-2 } | {C k , F k-1 }.\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 238.74,\n 243.12,\n 265.93,\n 17.19\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"F k = argmax F k P (F k | C k , F k-1 ).(4)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 241.39,\n 289.26,\n 50.16,\n 9.65\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"P (C k | F k )\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 359.83,\n 300.17,\n 55.1,\n 9.65\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"P (F k | F k-1\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 164.14,\n 338.38,\n 69.58,\n 9.65\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"P (F k | C k , F k-1\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 108,\n 664.08,\n 123.66,\n 9.65\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"F k = P (F k | F k-1 , C k )F k-1 .\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 213.9,\n 401.09,\n 290.77,\n 9.79\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"F k : (x, d, cp (x, k), Var(c p (x, k))) → (σ, c).(5)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 218.29,\n 448.37,\n 286.38,\n 9.84\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"c p (x, k) = {C k,cam (P cam • x) | cam ∈ M}.(6)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 145.4,\n 381.1,\n 191.49,\n 11.38\n ],\n \"formula_id\": \"formula_12\",\n \"formula_text\": \"G k = {P (max(σ k (x)) > 0) | ∀x ∈ V cur } 3 ,\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 220.64,\n 504.12,\n 284.03,\n 12.69\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"G 0 k = P (G k | G k-1 ) • G J k-1 = S * G J k-1 .(7)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 149.99,\n 614.75,\n 354.68,\n 22.05\n ],\n \"formula_id\": \"formula_14\",\n \"formula_text\": \"G j k = P (σ(x) | G j-1 k ) • G j-1 k , P (σ(x) | G) = 1, if σ(x) < G(x) σ(x), otherwise .(8)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 182.19,\n 387.46,\n 322.48,\n 9.68\n ],\n \"formula_id\": \"formula_15\",\n \"formula_text\": \"reject x i if (σ occ (x i ) < σ min ) and ¬(i mod R ≡ R//2).(9)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 251.84,\n 544.4,\n 252.83,\n 23.22\n ],\n \"formula_id\": \"formula_16\",\n \"formula_text\": \"L depth = std(d f ar /d 3x3 ) std(c 3x3 ) ,(10)\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"OD-NeRF: Efficient Training of On-the-Fly Dynamic Neural Radiance Fields"},"abstract":{"kind":"string","value":"Dynamic neural radiance fields (dynamic NeRFs) have demonstrated impressive results in novel view synthesis on 3D dynamic scenes. However, they often require complete video sequences for training followed by novel view synthesis, which is similar to playing back the recording of a dynamic 3D scene. In contrast, we propose OD-NeRF to efficiently train and render dynamic NeRFs on-the-fly which instead is capable of streaming the dynamic scene. When training on-the-fly, the training frames become available sequentially and the model is trained and rendered frame-by-frame. The key challenge of efficient on-the-fly training is how to utilize the radiance field estimated from the previous frames effectively. To tackle this challenge, we propose: 1) a NeRF model conditioned on the multi-view projected colors to implicitly track correspondence between the current and previous frames, and 2) a transition and update algorithm that leverages the occupancy grid from the last frame to sample efficiently at the current frame. Our algorithm can achieve an interactive speed of 6FPS training and rendering on synthetic dynamic scenes on-the-fly, and a significant speed-up compared to the state-of-the-art on real-world dynamic scenes."},"authors":{"kind":"string","value":"Zhiwen Yan; Chen Li; Gim Hee Lee"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: We introduce the on-the-fly training(left) of dynamic NeRFs and the OD-NeRF model(right). In on-the-fly training, the dynamic NeRF is trained based on the current and previous training frames to synthesize novel views for the current time step. Our OD-NeRF leverages the projected colors(orange arrow) to track implicit correspondence for fast on-the-fly convergence, and transition and update to the occupancy grid(green arrows) for efficient sampling.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: We compare the extrapolation ability of dynamic NeRF with \"space-time\" input and with \"space-color\" input. The two images on the right illustrate the novel view synthesized for time t k by the respective model trained with frames up to t k-1 . The green channel represents the rendered image, and the purple challenge represents the ground truth for frame k. The model with \"space-time\" input extrapolates poorly and renders the image lagging behind the dynamic ground truth, while our model with \"space-color\" input extrapolates well without any training on frame k.","figure_data":"","figure_id":"fig_1","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Updating the occupancy grid by sampling random points in each voxel can miss the moving occupied space and causes the entire voxel skipped during rendering. Our transit and update method does not have this issue as it updates the grid at the occupied point.","figure_data":"","figure_id":"fig_2","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: Qualitative results of our model compared to TiNeuVox on the D-NeRF dataset, by training to a comparable rendering quality(left) and training under a similar time constraint.","figure_data":"","figure_id":"fig_3","figure_label":"5","figure_type":"figure"},{"figure_caption":"Figure 6 :6Figure 6: Ablation comparison of rendering quality with different training time.","figure_data":"","figure_id":"fig_4","figure_label":"6","figure_type":"figure"},{"figure_caption":"Figure 7 :7Figure 7: Qualitative results of models with different training speed on the MeetRoom dataset.","figure_data":"","figure_id":"fig_5","figure_label":"7","figure_type":"figure"},{"figure_caption":"Figure 8 :8Figure 8: Qualitative results of models with different training speed on the DyNeRF dataset.","figure_data":"","figure_id":"fig_6","figure_label":"8","figure_type":"figure"},{"figure_caption":"Figure 9 :9Figure 9: Qualitative results of the \"T-Rex\" scene in the D-NeRF dataset, where the two models are trained to a comparable rendering quality.","figure_data":"","figure_id":"fig_7","figure_label":"9","figure_type":"figure"},{"figure_caption":"and 18).","figure_data":"","figure_id":"fig_8","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 10 :10Figure 10: Qualitative results of the \"Hook\" scene in the D-NeRF dataset, where the two models are trained to a comparable rendering quality.","figure_data":"","figure_id":"fig_9","figure_label":"10","figure_type":"figure"},{"figure_caption":"Figure 11 :11Figure 11: Qualitative results of the \"Jumping-Jack\" scene in the D-NeRF dataset, where the two models are trained to a comparable rendering quality.","figure_data":"","figure_id":"fig_10","figure_label":"11","figure_type":"figure"},{"figure_caption":"Figure 12 :12Figure 12: Qualitative results of the \"Lego\" scene in the D-NeRF dataset, where the two models are trained to a comparable rendering quality.","figure_data":"","figure_id":"fig_11","figure_label":"12","figure_type":"figure"},{"figure_caption":"Figure 13 :13Figure 13: Qualitative results of the \"T-Rex\" scene in the D-NeRF dataset, where the two models are trained with a similar time constraint.","figure_data":"","figure_id":"fig_12","figure_label":"13","figure_type":"figure"},{"figure_caption":"Figure 14 :14Figure 14: Qualitative results of the \"Hook\" scene in the D-NeRF dataset, where the two models are trained with a similar time constraint.","figure_data":"","figure_id":"fig_13","figure_label":"14","figure_type":"figure"},{"figure_caption":"Figure 15 :15Figure 15: Qualitative results of the \"Jumping-Jack\" scene in the D-NeRF dataset, where the two models are trained with a similar time constraint.","figure_data":"","figure_id":"fig_14","figure_label":"15","figure_type":"figure"},{"figure_caption":"Figure 16 :16Figure 16: Qualitative results of the \"Lego\" scene in the D-NeRF dataset, where the two models are trained with a similar time constraint.","figure_data":"","figure_id":"fig_15","figure_label":"16","figure_type":"figure"},{"figure_caption":"Figure 17 :17Figure 17: Qualitative results of the \"discussion\" scene in the MeetRoom dataset.","figure_data":"","figure_id":"fig_16","figure_label":"17","figure_type":"figure"},{"figure_caption":"Figure 18 :18Figure 18: Qualitative results of the \"vrheadset\" scene in the MeetRoom dataset.","figure_data":"","figure_id":"fig_17","figure_label":"18","figure_type":"figure"},{"figure_caption":"","figure_data":"","figure_id":"","figure_label":"","figure_type":"figure"},{"figure_caption":"Quantitative results of speed and novel view synthesis qualities on the D-NeRF dataset.","figure_data":"Per Second (FPS)↑Render QualityMethodTrainRenderTotalPSNR↑ LPIPS↓D-NeRF[18]*0.0013NA<0.001330.500.07K-Plane[20]*0.04NA<0.0431.67NATiNeuVox-S[4]*0.21NA<0.2130.750.07TiNeuVox-B[4]*0.08NA<0.0832.670.04TempInterp[16]*0.21NA<0.2129.840.06TiNeuVox[4]1.272.800.8730.570.07Ours12.1115.686.7832.870.04*Result reported for joint training instead of on-the-fly training. NA if renderingspeed not reported.","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"Quantitative ablation study of each component proposed for our model.","figure_data":"","figure_id":"tab_2","figure_label":"2","figure_type":"table"},{"figure_caption":"dataset. Our model is free of many common artifacts present in the rendered results of the baseline model while maintaining a significantly faster on-the-fly training speed.","figure_data":"MeetRoomDyNeRFTime(seconds per frame)↓Time(seconds per frame)↓MethodTrain RenderTotalPSNR↑MethodTrain RenderTotalPSNR↑-----DyNeRF[8]* 15600671566729.58JaxNeRF[3]* 28380402842027.11K-Plane[20]*21.6NA>21.631.63Plenoxels[26]8400.184027.15Plenoxels[26] 13800.12138028.68LLFF[22]1800.000318022.88LLFF[22]4800.00448023.23StreamRF[6]10.20.110.326.72StreamRF[6]15.00.115.128.26InstNGP[13]12.90.813.622.82InstNGP[13]12.61.013.624.30Ours2.31.23.527.32Ours2.72.45.027.52","figure_id":"tab_3","figure_label":"","figure_type":"table"},{"figure_caption":"Quantitative results on real-world MeetRoom and DyNeRF dataset.","figure_data":"","figure_id":"tab_4","figure_label":"3","figure_type":"table"},{"figure_caption":"Quantitative ablation study of each component proposed for our model.","figure_data":"Render Quality","figure_id":"tab_5","figure_label":"4","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: We introduce the on-the-fly training(left) of dynamic NeRFs and the OD-NeRF model(right). In on-the-fly training, the dynamic NeRF is trained based on the current and previous training frames to synthesize novel views for the current time step. Our OD-NeRF leverages the projected colors(orange arrow) to track implicit correspondence for fast on-the-fly convergence, and transition and update to the occupancy grid(green arrows) for efficient sampling.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: We compare the extrapolation ability of dynamic NeRF with \\\"space-time\\\" input and with \\\"space-color\\\" input. The two images on the right illustrate the novel view synthesized for time t k by the respective model trained with frames up to t k-1 . The green channel represents the rendered image, and the purple challenge represents the ground truth for frame k. The model with \\\"space-time\\\" input extrapolates poorly and renders the image lagging behind the dynamic ground truth, while our model with \\\"space-color\\\" input extrapolates well without any training on frame k.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Updating the occupancy grid by sampling random points in each voxel can miss the moving occupied space and causes the entire voxel skipped during rendering. Our transit and update method does not have this issue as it updates the grid at the occupied point.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: Qualitative results of our model compared to TiNeuVox on the D-NeRF dataset, by training to a comparable rendering quality(left) and training under a similar time constraint.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 :6Figure 6: Ablation comparison of rendering quality with different training time.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 7 :7Figure 7: Qualitative results of models with different training speed on the MeetRoom dataset.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"7\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 8 :8Figure 8: Qualitative results of models with different training speed on the DyNeRF dataset.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"8\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 9 :9Figure 9: Qualitative results of the \\\"T-Rex\\\" scene in the D-NeRF dataset, where the two models are trained to a comparable rendering quality.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"9\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"and 18).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_8\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 10 :10Figure 10: Qualitative results of the \\\"Hook\\\" scene in the D-NeRF dataset, where the two models are trained to a comparable rendering quality.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_9\",\n \"figure_label\": \"10\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 11 :11Figure 11: Qualitative results of the \\\"Jumping-Jack\\\" scene in the D-NeRF dataset, where the two models are trained to a comparable rendering quality.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_10\",\n \"figure_label\": \"11\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 12 :12Figure 12: Qualitative results of the \\\"Lego\\\" scene in the D-NeRF dataset, where the two models are trained to a comparable rendering quality.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_11\",\n \"figure_label\": \"12\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 13 :13Figure 13: Qualitative results of the \\\"T-Rex\\\" scene in the D-NeRF dataset, where the two models are trained with a similar time constraint.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_12\",\n \"figure_label\": \"13\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 14 :14Figure 14: Qualitative results of the \\\"Hook\\\" scene in the D-NeRF dataset, where the two models are trained with a similar time constraint.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_13\",\n \"figure_label\": \"14\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 15 :15Figure 15: Qualitative results of the \\\"Jumping-Jack\\\" scene in the D-NeRF dataset, where the two models are trained with a similar time constraint.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_14\",\n \"figure_label\": \"15\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 16 :16Figure 16: Qualitative results of the \\\"Lego\\\" scene in the D-NeRF dataset, where the two models are trained with a similar time constraint.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_15\",\n \"figure_label\": \"16\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 17 :17Figure 17: Qualitative results of the \\\"discussion\\\" scene in the MeetRoom dataset.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_16\",\n \"figure_label\": \"17\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 18 :18Figure 18: Qualitative results of the \\\"vrheadset\\\" scene in the MeetRoom dataset.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_17\",\n \"figure_label\": \"18\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"\",\n \"figure_id\": \"\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Quantitative results of speed and novel view synthesis qualities on the D-NeRF dataset.\",\n \"figure_data\": \"Per Second (FPS)↑Render QualityMethodTrainRenderTotalPSNR↑ LPIPS↓D-NeRF[18]*0.0013NA<0.001330.500.07K-Plane[20]*0.04NA<0.0431.67NATiNeuVox-S[4]*0.21NA<0.2130.750.07TiNeuVox-B[4]*0.08NA<0.0832.670.04TempInterp[16]*0.21NA<0.2129.840.06TiNeuVox[4]1.272.800.8730.570.07Ours12.1115.686.7832.870.04*Result reported for joint training instead of on-the-fly training. NA if renderingspeed not reported.\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Quantitative ablation study of each component proposed for our model.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"dataset. Our model is free of many common artifacts present in the rendered results of the baseline model while maintaining a significantly faster on-the-fly training speed.\",\n \"figure_data\": \"MeetRoomDyNeRFTime(seconds per frame)↓Time(seconds per frame)↓MethodTrain RenderTotalPSNR↑MethodTrain RenderTotalPSNR↑-----DyNeRF[8]* 15600671566729.58JaxNeRF[3]* 28380402842027.11K-Plane[20]*21.6NA>21.631.63Plenoxels[26]8400.184027.15Plenoxels[26] 13800.12138028.68LLFF[22]1800.000318022.88LLFF[22]4800.00448023.23StreamRF[6]10.20.110.326.72StreamRF[6]15.00.115.128.26InstNGP[13]12.90.813.622.82InstNGP[13]12.61.013.624.30Ours2.31.23.527.32Ours2.72.45.027.52\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Quantitative results on real-world MeetRoom and DyNeRF dataset.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Quantitative ablation study of each component proposed for our model.\",\n \"figure_data\": \"Render Quality\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"[11]\", \"Explanation\": \"The cited work introduces the concept of neural radiance fields (NeRF), which serves as the basis for the research conducted in the citing paper to develop a new 3D volumetric representation capable of synthesizing photo-realistic novel views from multi-view images.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[8,16]\", \"Explanation\": \"The cited works focus on reconstructing scenes with moving or deforming objects using space-time 4D representations, which the citing paper extends by proposing a new approach to train a dynamic NeRF concurrently with the acquisition of training frames as the dynamic event unfolds.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[18,15]\", \"Explanation\": \"The cited works present explicit deformation representations for dynamic NeRFs, which the citing paper further extends by introducing a new on-the-fly training approach to train a dynamic NeRF concurrently with the acquisition of training frames during the dynamic event.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[13,1]\", \"Explanation\": \"The cited works demonstrate the success in NeRF acceleration to train a small scene in seconds, which the citing paper leverages to propose a new approach for on-the-fly training of a dynamic NeRF to stream the dynamic scene in an interactive speed.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[8,16]\", \"Explanation\": \"The cited works provide the basis for the use of a temporal input dimension in NeRF models, which the citing paper builds upon in their own research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[18,15]\", \"Explanation\": \"The cited works introduce a temporal deformation field in NeRF models, which the citing paper extends by exploring new dimensions and variables in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[18,15]\", \"Explanation\": \"The cited works provide the data source for the use of a temporal deformation field in NeRF models, which the citing paper utilizes in their research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[18,15]\", \"Explanation\": \"The cited works provide a method of transiting and updating the occupancy grid for efficient point sampling, which the citing paper extends by introducing a new method in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work provides the occupancy grid method for static NeRFs, which the citing paper adopts to reduce the number of sampled points for acceleration in dynamic NeRFs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[1]\", \"Explanation\": \"The cited work provides the occupancy grid method for static NeRFs, which the citing paper adopts to adapt the occupancy grid to on-the-fly dynamic training in the new setting of training dynamic NeRF on-the-fly.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[18]\", \"Explanation\": \"The cited work provides the D-NeRF dataset for evaluation, which the citing paper uses to test the performance of the method in a real-world setting.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work provides the MeetRoom dataset for evaluation, which the citing paper uses to test the performance of the method in a real-world setting.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[8]\", \"Explanation\": \"The cited work provides the DyNeRF dataset for evaluation, which the citing paper uses to test the performance of the method in a real-world setting.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[11,23,26]\", \"Explanation\": \"The cited works on static 3D scene representation with NeRFs serve as a basis for the development of dynamic NeRFs in the citing paper, which aims to improve the training efficiency of dynamic NeRFs for on-the-fly applications.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[18,8,15,25]\", \"Explanation\": \"The cited works on dynamic scene representation with NeRFs provide the data and methods used in the citing paper to further explore the representation of dynamic scenes with NeRFs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work on static NeRFs in voxel and MLP architecture serves as the methodological basis for the development of the dynamic NeRF in the citing paper, which is used to represent the static canonical space in the dynamic NeRF.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[26]\", \"Explanation\": \"The cited work on the Plenoxel representation in the first frame of the scene in StreamRF is used as a methodological basis in the citing paper to learn the subsequent changes in the dynamic NeRF.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16,20]\", \"Explanation\": \"The cited works on expanding the 3D representation in static NeRF to 4D with a time dimension for dynamic scene representation in the citing paper build upon the methods used in static NeRFs to develop a new approach for dynamic NeRFs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[24]\", \"Explanation\": \"The cited work IBRNet is used as a methodological basis for constructing a cost volume of a dynamic scene based on image features of nearby views in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[2]\", \"Explanation\": \"The cited work MVSNeRF is used as a methodological basis for learning a blending weights or density and color output in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[5,19,10]\", \"Explanation\": \"The cited works are used as methodological basis for constructing a cost volume of a dynamic scene based on image features of nearby views in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[22,21]\", \"Explanation\": \"The cited works LLFF are used as methodological basis for aggregating features along the multi-view epipolar line to render view dependent effect in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[9]\", \"Explanation\": \"The cited work DynIBaR is used as methodological basis for aggregating projected image features across frames in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[11]\", \"Explanation\": \"The cited work on static neural radiance fields provides the foundational methodology for the on-the-fly dynamic NeRF training in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[18,15]\", \"Explanation\": \"The cited works introduce a time-varying deformation field D that maps spatial coordinates to their canonical space correspondence, which the citing paper adopts in the input of dynamic NeRFs to include an additional time dimension.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[8,16]\", \"Explanation\": \"The cited works expand the 3D NeRF model to a 4D variant, which the citing paper also adopts in the input of dynamic NeRFs to include an additional time dimension.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[1]\", \"Explanation\": \"The cited work provides the training images C 0:K (r) from all time t 0:K for the training stage of the dynamic NeRF model, which the citing paper uses to minimize the rendering loss during the training process.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[1]\", \"Explanation\": \"The cited work introduces a new on-the-fly training of the dynamic NeRF model suitable for streaming dynamic scenes, which the citing paper further extends by training the NeRF concurrently while the dynamic event unfolds.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work provides a method for reconstructing dynamic scenes from a monocular camera, which the citing paper adopts in their research to address the limitations of static multi-view forward-facing cameras.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[17]\", \"Explanation\": \"The cited work discusses the use of 360-degree inward-facing cameras for scene reconstruction, which the citing paper builds upon to further explore the challenges and limitations of static multi-view forward-facing cameras in the streaming industry.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work introduces the concept of a space-time 4D occupancy grid, which the citing paper adopts in their research on dynamic scene reconstruction to improve sampling efficiency in on-the-fly training.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work, Instant-NGP, provides a global update method for updating the occupancy grid in a constant state of change. The citing paper adopts this method to update the occupancy grid in their research on on-the-fly training of NeRF models.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[18]\", \"Explanation\": \"The cited work, D-NeRF, is the source of the synthetic dataset used in the evaluation of the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work, MeetRoom, is the source of the real-world dataset used in the evaluation of the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[8]\", \"Explanation\": \"The cited work, DyNeRF, is the source of the real-world dataset used in the evaluation of the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[4]\", \"Explanation\": \"The cited work, TiNeuVox, serves as the base model for implementing the method in the citing paper, providing the foundation for the improvements in on-the-fly training speed and rendering quality.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[4]\", \"Explanation\": \"The cited work, TiNeuVox, serves as the baseline model for comparison in the study conducted in the citing paper. The comparison highlights the performance of the proposed method in terms of rendering quality and training time under different time constraints.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work, StreamRF, is used as a baseline for comparison in the training and rendering quality of the model proposed in the citing paper. The citing paper extends the research by comparing the model against StreamRF and providing a more detailed analysis of the results.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work, Instant-NGP, is the implementation used in the model proposed in the citing paper. The citing paper acknowledges the origin of the model and its contribution to the research conducted in the study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[8]\", \"Explanation\": \"The cited work, DyNeRF, is used as a dataset for the qualitative results presented in the citing paper. The citing paper leverages the data from DyNeRF to showcase the performance of the model in a real-world scenario.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[4]\", \"Explanation\": \"The cited work, TiNeuVox, serves as the base model for the implementation in the citing paper, providing the necessary methods and techniques for generating synthetic data.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work, InstantNGP, is the base model used for the real-world dataset in the citing paper, providing the necessary methods and techniques for data collection and analysis.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[4]\", \"Explanation\": \"The cited work, TiNeuVox, provides the model structure and implementation details that the citing paper follows in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[12]\", \"Explanation\": \"The cited work, tiny-cuda-nn, is used as a tool to implement the hash voxel in the model structure, which is a data source for the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work, InstantNGP, is used to implement the hash voxel in the model structure, which is a methodological basis for the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[7]\", \"Explanation\": \"The cited work provides the code base for the implementation of the citing paper, which serves as the methodological basis for the study conducted in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work of InstantNGP is extended in the citing paper by incorporating the mean and variance of projected colors of sampled points with spatial features from hash voxel grid into the NeRF MLP.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work of MeetRoom is used as a dynamic dataset for the study conducted in the citing paper, providing a real-world scenario for testing the performance of the model.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[8]\", \"Explanation\": \"The cited work of DyNeRF is used as a dynamic dataset for the study conducted in the citing paper, providing a real-world scenario for testing the performance of the model.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[10]\", \"Explanation\": \"The cited work of depth smoothness regularization based on patch sampling provides supporting evidence for the implementation of the model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work, MeetRoom, provides a depth smoothness loss that the citing paper adopts in their research to penalize local large inverse depth variation when the color variation is small.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[8]\", \"Explanation\": \"The DyNeRF dataset is used as a data source in the citing paper to train the model and evaluate its performance.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work, InstantNGP, has a sampling strategy based on the occupancy grid that the citing paper updates during training, but does not change the sampling strategy itself.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[18]\", \"Explanation\": \"The cited work, D-NeRF, serves as the data source for the multi-view forward facing camera version used in the citing paper.\"}]"}}},{"rowIdx":78,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b4","b22","b18","b14","b36","b5","b13","b8","b17"],"table_ref":[],"text":"Document summarization aims to compress text material while retaining its most salient information. With the increasing amount of publicly available text data, the significance of automated summarization techniques has amplified. Recent advancements in summarization systems, leveraging neural networks and pre-trained language models, have demonstrated notable progress (Cheng and Lapata, 2016;Nallapati et al., 2016;Liu and Lapata, 2019;Lewis et al., 2019;Zhang et al., 2020). The above summarization systems are all built end-toend and follow a one-shot paradigm that generates summaries in a single step. On the contrary, humans often write text through an evolutionary pro- cess characterized by multiple iterations of drafting and editing (Faltings et al., 2020).\nThese end-to-end summarization systems encounter multiple challenges. Firstly, they frequently suffer from the issue of hallucination, resulting in the generation of ungrammatical or factually incorrect content (Kryscinski et al., 2020;Zhang et al., 2022b). Secondly, these systems are often optimized using imperfect reference summaries, and widely adopted evaluation metrics like ROUGE may not accurately assess summary quality. Thirdly, most of these systems lack controllability, as they only produce a single generic summary conditionally on the input document. In practice, instead of a single condensed version of the entire document, generating summaries that cater to specific aspects or queries would be more beneficial to meet the diverse requirements of users.\nThe emergence of advanced instruction-tuned large language models (LLMs), such as ChatGPT2 , has presented exciting possibilities for summarization systems by exhibiting strong zero-shot performance in various downstream tasks. A recent study by (Goyal et al., 2022) compared GPT-3 with traditional fine-tuning methods and found that despite lower ROUGE scores, human annotators preferred the GPT-3 generated summaries. Another comprehensive analysis by (Zhang et al., 2023d) focused on large language models for news summarization and revealed that the quality of generated summaries is already on par with those created by humans. Furthermore, Liu et al. (2023) demonstrated the utilization of LLMs like GPT-4 as an effective natural language generation evaluator, showing a higher correlation with humans in the summarization task compared to previous reference-based methods.\nThe advent of LLMs also introduces new opportunities for summarization beyond the traditional one-shot generation setting. In this paper, we introduce SummIt, a framework that leverages large language models for iterative text summarization. Instead of generating summaries in a single step, our framework enables the model to iteratively refine the generated summary through self-evaluation and feedback, resembling the human process of drafting and revising summaries. According to our experiments, the rationale generation and summary refinement in SummIt can be guided effectively with in-context learning, eliminating the need for supervised training or reinforcement learning processes. Additionally, we explore the potential benefits of incorporating knowledge and topic extractors to enhance summary faithfulness and controllability. We instantiate SummIt with ChatGPT as the backbone, and the automatic evaluation results on three benchmark datasets demonstrate the effectiveness of SummIt in improving summary quality, faithfulness, and controllability within only a few iterations. Furthermore, we conduct a human evaluation to validate the iterative refinements quality and identify a potential over-correction issue.\nWe summarize the contributions of this paper as follows:\n• We propose SummIt, a novel framework for iterative text summarization. SummIt enables the iterative refinement of generated summaries by incorporating self-evaluation and feedback mechanisms. In addition, we propose to incorporate knowledge and topic extractors to further improve the faithfulness and controllability of SummIt.\n• We conduct experiments on three summarization benchmark datasets, and empirical results from automatic evaluation demonstrate the effectiveness of our proposed framework in summary refinement.\n• A human evaluation is conducted to investigate the impact of self-evaluation-guided summary refinement. The results revealed a potential issue of over-correction: while the large language model effectively refines the summary based on the feedback rationale, it exhibits a bias toward its own evaluation criteria, rather than aligning closely with human judgment.\n2 Related Work"},{"figure_ref":[],"heading":"Text Summarization","publication_ref":["b30","b18","b39","b14","b36","b13","b19"],"table_ref":[],"text":"Recent years have witnessed significant advancements in text summarization systems with the development of deep neural networks and pre-trained language models. Automatic summarization methods can be broadly categorized into extractive and abstractive approaches. Extractive summarization involves the direct extraction of sentences from the source text to form summaries (Xu et al., 2019;Liu and Lapata, 2019;Zhong et al., 2020;Zhang et al., 2022aZhang et al., , 2023b)), while abstractive approaches conditionally generate summaries using a sequence-tosequence (seq2seq) framework (Lewis et al., 2019;Zhang et al., 2020). Existing approaches mentioned above generate summaries in a one-shot manner, and their outputs may not always align with user expectations and may contain hallucinated content (Kryscinski et al., 2020). To address the limitation, Liu et al. (2022) proposes to automatically correct factual inconsistencies in generated summaries with generated human feedback. In contrast, our SummIt framework enables iterative summary refinement with self-evaluation and feedback, eliminating the need for costly human annotations. Additionally, we propose the integration of knowledge and topic extractors to further enhance summary faithfulness and controllability.\nYou are a summarizer that follows the output pattern. You revise the summary based on the given instructions. You follow all the instructions without commenting on them.\nRefine: [Revise Suggestions] Revise the summary. Follow all the suggestions and you an not make more comments. [Format Instructions] You are a summary evaluator that gives scores for the summaries with revise suggestions. Your suggestions can be: 1. Add the information of <insert> 2.Remove the information of <insert> 3. Rephrase the information of <insert> 4. Shorten the summary 5. Keep the summary unchanged If you think there's no further revision is needed, you must add \"<STOP>\" at the end."},{"figure_ref":[],"heading":"Knowledge Extractor Topic Extractor","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Summarizer Evaluator","publication_ref":[],"table_ref":[],"text":"Refined Summary Evaluation Rationale Refined Summary"},{"figure_ref":[],"heading":"Source Document","publication_ref":[],"table_ref":[],"text":"Figure 2: The overall framework of our proposed iterative text summarization system. The evaluator generates an evaluation rationale based on the current summary, and the summarizer then refines the summary accordingly. The knowledge and topic extractors retrieve information from the source document to guide the process."},{"figure_ref":[],"heading":"Summarization with Large Language Models","publication_ref":["b24","b3","b8","b17","b6","b20","b16","b37"],"table_ref":[],"text":"Recent years have seen a surge in training largescale language models (LLM) on large amounts of text, such as GPT (Radford et al., 2019;Brown et al., 2020). Several studies have explored the application of LLMs in the context of text summarization. For instance, Goyal et al. (2022) compared the performance of GPT-3-generated summaries with traditional fine-tuning methods, finding that although the former achieved slightly lower ROUGE scores, human evaluators expressed a preference for them. Similarly, Zhang et al. (2023d) reported that LLM-generated summaries were on par with human-written summaries in the news domain. Zhang et al. (2023c) benchmarked the performance of ChatGPT on extractive summarization and proposes to improve summary faithfulness with an extract-then-generate pipeline. On the other hand, prior works have also leveraged LLMs for summarization evaluation (Liu et al., 2023;Fu et al., 2023;Luo et al., 2023), demonstrating that LLM-based metrics outperform all previous evaluation metrics like ROUGE (Lin, 2004) and BertScore (Zhang et al., 2019) by a significant margin in terms of correlation with human evaluations."},{"figure_ref":[],"heading":"Text Editing","publication_ref":["b11","b1","b5","b28","b21"],"table_ref":[],"text":"Our work is also closely related to the task of text editing. Traditional editing models are trained to solve specific tasks, such as information updating (Iso et al., 2020), Wikipedia edit (Reid and Neu-big, 2022), and grammar error correction (Awasthi et al., 2019). Recent works also formulate text editing as an interactive task, such as commandbased editing systems (Faltings et al., 2020), and interactive editing systems (Schick et al., 2022). 2023) propose a reinforcement learningbased approach to generate natural language feedback for correcting generation errors. Concurrent work Madaan et al. (2023) presents a similar generation pipeline that enhances initial outputs through iterative feedback using a single LLM for short text generation tasks. In contrast, our SummIt framework differs from these approaches as it specifically focuses on the conditional generation task of summarization, with an emphasis on improving summary faithfulness and controllability. Additionally, we empirically observe that separating the summarizer and evaluator into different LLMs, each employing different in-context guidance leads to improved performance in our framework."},{"figure_ref":[],"heading":"Methods","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Iterative Summarization","publication_ref":[],"table_ref":[],"text":"The overall architecture of our iterative text summarization system SummIt is shown in Figure 2.\nThe system consists of two major components, a summarizer that generates and refines the summary, and an evaluator that generates feedback rationale.\nSummarizer: The summarizer is in charge of generating the initial summary and revising a summary conditioned on the given explanations and source document. We instantiate the summarizer with an instruction-tuned language model S.\nFormally, given the input source document x, the initial summary y 0 generation process can be represented as:\np S (y 0 | x) = m t=1 p S y 0 t | y 0 <t , x ,(1)\n, where y 0 <t denotes the generated tokens, y 0 t refers to the t-th summary token, and m denotes the summary length.\nAfter obtaining the i-step self-evaluation feedback e i from the evaluator E, the summarizer will refine the summary accordingly and then generates refined summary y (i+1) as: p S (y (i+1) | x, e i ).\nEvaluator: The evaluator is another instance of language model E that generates summary quality evaluation and corresponding explanations e i for the i-th iteration as: p E (e i | x, y i ).\nStopping Criteria: The evaluator gives a quality assessment of the generated summary and then outputs the rationale for the evaluation as feedback. The summarizer receives model evaluation and feedback from the evaluator, subsequently refining the summary based on this input.\nThis iterative process can be repeated until 1) the evaluator determines that no further refinement is required or 2) fulfills rule-based stopping criteria, such as reaching a maximum iteration number."},{"figure_ref":[],"heading":"In-context Learning","publication_ref":["b3"],"table_ref":[],"text":"Since the summarizer and evaluator in SummIt are not fine-tuned with supervised data or trained reinforcement learning rewards, it would be beneficial to guide the explanation and summary generation process with the desired format or template. Recent studies have shown that large language models have strong few-shot performance on various downstream tasks, known as in-context learning (ICL) (Brown et al., 2020).\nThe standard ICL prompts a language model, M , with a set of exemplar source-summary pairs, C = {(x 1 , y 1 )...(x m , y m )}, and generates summary y by concatenating the exemplar sourcesummary pairs and input document as prompt:\np M (y | x, C).\nWe also use in-context learning to guide our iterative summarization system, where we use \"document-reference summary\" pairs as the context for the summarizer S, and use \"documentreference summary-human written explanation\" triplets as the context for the evaluator E. We empirically find that in-context learning could improve the efficacy of our system."},{"figure_ref":[],"heading":"Summary Faithfulness and Controllability","publication_ref":["b13","b10","b40","b20"],"table_ref":[],"text":"In practical applications, the faithfulness of the generated summary holds significant importance, alongside the overall quality of the summary (Kryscinski et al., 2020). Previous research has demonstrated the effectiveness of leveraging knowledge extraction from source documents to enhance the faithfulness of generated summaries (Huang et al., 2020;Zhu et al., 2020). Building upon these insights, we propose integrating a knowledge extractor into our iterative summarization system.\nKnowledge Extractor: In particular, we utilize OpenIE3 , which extracts knowledge k in the form of triplets from the source document. During each iteration, the summarizer (S) is guided to refine the summary in accordance with the extracted knowledge, represented as: p S (y (i+1) | x, e i , k). Moreover, the evaluator (E) can be directed to factor in faithfulness when delivering feedback, denoted as p E (e i | x, y i , k), as LLMs have shown to be efficient faithfulness evaluators (Luo et al., 2023).\nFurthermore, real-world applications often require the generation of summaries tailored to specific aspects or queries, rather than a single generic summary of the entire document. Our iterative summarization framework offers enhanced controllability for aspect-based summarization tasks.\nTopic Extractor: Given an aspect-oriented query q, we prompt both summarizer S and evaluator E to initially extract relevant snippets, each containing less than 5 words, from the source document x. Following the extraction, these components then proceed to either generate or assess the summary by taking into account the extracted snippets. The iterative nature of our framework further facilitates the controllable summary generation, allowing for the easy transformation of generic summaries into topic-focused summaries based on the user's preferences."},{"figure_ref":[],"heading":"Prompt Format","publication_ref":["b26","b5"],"table_ref":[],"text":"We utilize both system prompts and user prompts following the OpenAI API in our system implementations. The full prompts used in the experiments can be found in Table 8 andTable 9. Notably, we empirically find that pre-defining the possible edit operations for the evaluator improves the system performance significantly since it avoids free-form edits to the summary by the large language model. Thus, we adopt the five types of text editing operations commonly used in text editing systems (Reid and Neubig, 2022;Faltings et al., 2020). We specifically require the evaluator to generate feedback based on the source document and summary at this iteration with the following five types of possible refinement operations:\n• Add: Add the information of <insert> "},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":"In this section, we validate our SummIt framework on three benchmark summarization datasets. We employ both automatic metrics and human assessment to evaluate the quality 4.2, faithfulness 4.3, and controllability 4.4 of the generated summaries."},{"figure_ref":[],"heading":"Experiment Settings","publication_ref":["b16","b17","b13","b7","b27","b12","b8","b15"],"table_ref":[],"text":"Datasets: We conduct experiments on the following three publicly available benchmark datasets, as presented in Evaluation metrics: For summary quality, we use ROUGE scores (Lin, 2004) and G-Eval (Liu et al., 2023) as the automatic metrics. We report ROUGE-1, ROUGE-2, and ROUGE-L scores, which respectively measure the overlap of unigrams, bigrams, and the longest common sequence between the generated summary and the reference summary. G-Eval is an LLM-based matrix with a scale ranging from 1 to 5. G-Eval uses LLM with chain-of-thoughts (CoT) and a form-filling paradigm to assess the quality of NLG outputs. It shows the highest correlation with humans compared to other summarization quality metrics. For summary faithfulness, we use FactCC (Kryscinski et al., 2020) and DAE (Defining arc entailment) (Goyal and Durrett, 2020) as our evaluation metrics. FactCC is a weakly supervised BERT-based model metric that verifies factual consistency through rule-based transformations applied to source document sentences. It shows a high correlation in assessing summary faithfulness with human judgments. DAE decomposes entailment at the level of dependency arcs, examining the semantic relationships within the generated output and input. Rather than focusing on aggregate decisions, DAE measures the semantic relationship manifested by individual dependency arcs in the generated output supported by the input.\nFor the controllability of query-focused summarization, we use BM25 (Robertson et al., 2009) and DPR (Karpukhin et al., 2020) to measure the similarity between the query and the summary with both sparse and dense evaluations. BM25 is a probabilistic retrieval function that ranks documents based on query term frequency. DPR leverages dense vector representations for scalable retrieval, embedding both questions and passages into fixedlength vector spaces for nuanced similarity calculations. In line with previous research findings that have emphasized the inclination of human annotators towards summaries generated by LLM models, even in the presence of comparatively lower ROUGE scores (Goyal et al., 2022), we further validate the effectiveness of SummIt through a dedicated human study. Specifically, we use 1) five-point Likert scale ratings (Likert, 1932) covering summary coherence, fluency, relevance, consistency, conciseness, and overall evaluation, and 2) human preference test, where annotators are shown summaries of the same source document from all five summarization systems and then asked to select their most preferred summary or summaries."},{"figure_ref":[],"heading":"Model","publication_ref":[],"table_ref":[],"text":"We evaluated the performance using 1000 ran-dom samples from CNN/DM and XSum test sets, with seed 101, and the full NEWTS test set. Our prompts were refined with a 50-example development set. The detailed experimental setup is provided in Appendix A."},{"figure_ref":[],"heading":"Generic Summary Quality Evaluation","publication_ref":["b36","b14","b25","b8"],"table_ref":["tab_2","tab_2","tab_3"],"text":"The automatic evaluation results for generic summarization quality are shown in Table 2. We use previous pre-trained language models, including PEGASUS (Zhang et al., 2020), BART (Lewis et al., 2019), and T5 (Raffel et al., 2020) as baseline models. We compare our framework SummIt with these baseline models under a zero-shot setting for a fair comparison.\nIt is observed that SummIt has inferior ROUGE scores compared to fine-tuning approaches on CN-N/DM, while exhibiting significantly higher LLMbased evaluation metric G-Eval. On the other hand, it outperforms all baseline methods on the XSum dataset. Compared to the output of ChatGPT, the summaries of ChatGPT after our iterative refinement see a consistent improvement in the G-Eval score. The results are consistent with the previous conclusions in (Zhang et al., 2023d), where large language model summary outputs receive lower ROUGE scores due to the low quality of reference summaries.\nIn addition to the zero-shot setting, we investigate the effects of in-context learning for SummIt, as shown in the lower block of Table 2. The results consistently demonstrate that incorporating incontext learning significantly enhances the model's performance on ROUGE and G-Eval scores. This observation underscores the substantial few-shot capabilities of SummIt, showcasing its ability to adapt effectively and generate high-quality summaries in contexts with very few examples.\nTo further verify the summary quality, we conduct a human study to evaluate the overall quality of the summaries as shown in Table 3. According to the five-point Likert scale ratings, the summaries of ChatGPT and SummIt consistently outperform pre-trained language model results. The iterative refinement of SummIt also provides consistent improvements, which align with the G-Eval results obtained from the automatic evaluation. We also conducted a human preference study, where summaries from all models were presented to human annotators. They were tasked to select the best summary, without any prior knowledge of the origin of each summary. Consistent with the findings in (Goyal et al., 2022), the results reveal a clear preference among human annotators for summaries generated by large language models (LLMs) for both CNN (86%) and BBC (62%) style summaries. We also notice the summaries of ChatGPT after our iterative refinement (SummIt) show a significant improvement in human preference, with 18% and 14% percent improvements on CNN/DM and XSum datasets. The results demonstrate the effectiveness of refining generic summaries of our framework. "},{"figure_ref":[],"heading":"Summary Faithfulness Evaluation","publication_ref":[],"table_ref":["tab_4"],"text":"To evaluate the efficacy of the SummIt framework in enhancing summary faithfulness with the knowledge extractor, we conducted additional experiments, as presented in Table 4. The findings demonstrate that our framework's iterative approach to refining summaries yields significant improvements in summary faithfulness, as indicated by both FactCC and DAE results. Furthermore, the integration of a knowledge extractor such as OpenIE further enhances the level of faithfulness.\nThe LLM-based evaluation score G-Eval also indicates a higher level of satisfaction with the refined summaries when guided by the extracted knowledge triplets. In conclusion, our study reveals that iterative refinements with the incorporation of the knowledge extractor effectively enhance summary faithfulness without compromising the quality of the summaries."},{"figure_ref":[],"heading":"Query-focued Summarization Controlability Evaluation","publication_ref":[],"table_ref":["tab_5"],"text":"We utilize the query-based summarization dataset NEWTS as our testbed to demonstrate the controllability ability of SummIt. The results obtained, as depicted in Table 5, highlight the framework's capability to align the focus of a generic summary with the specific topic of interest or query provided by the user. We also observe improved G-Eval evaluation scores by directing the summary generation process toward the intended topic. Furthermore, we evaluate the controllability of the summarization systems by quantifying the sim- ilarity between the query and the generated summary. Both BM25 and DPR are employed as similarity metrics, and we consistently observe enhancements after the iterative refinement process. This observation serves as evidence that SummIt effectively refines the summary to align with the topics specified in the query."},{"figure_ref":[],"heading":"Analysis","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Ablation Studies","publication_ref":[],"table_ref":["tab_7"],"text":"Table 6 shows the results of the ablation study by removing refinement operations. The ablation study is conducted on the CNN/DM dataset under the zero-shot settings. According to the results, each option contributes to the success of our method, and the add operation affects the ROUGE score most, while the simplify operation affects the GPTevaluation scores the most. Without adding the operation, the information in the iterative process will only decrease, resulting in less n-gram overlap. On the other hand, without the simplify and remove operations, the redundant information results in low G-Eval scores."},{"figure_ref":["fig_3"],"heading":"Over-correction Issue","publication_ref":["b17"],"table_ref":[],"text":"A recent work (Liu et al., 2023) Evaluation-Iter1: The summary effectively conveys the main point of the article, but it could be shortened for conciseness. Consider removing the specific hospitals mentioned and rephrasing the sentence about hospitals having to choose between emergency care and non-urgent surgery to make it more concise.\nSummary-Iter2: Hospitals in Wales may have to prioritize emergency care over non-urgent surgery during peak winter months, according to Dr. Andrew Goodall. Some surgical procedures have already been cancelled until after Christmas. more iterations... Table 7: An example of iterative summary refinement from the XSum dataset. The revision between the two iterations and their corresponding comments are presented in the same color. The blue color refers to the rephrase revision and the orange color refers to the remove operation.\n1) Does the refinement actually improve the summary? Another potential issue of SummIt would be:\n2) Does the refinement actually follow the rationale feedback from the evaluator?\nTo address these two concerns and provide further validation for the step-wise summary refinement in SummIt, we conducted the corresponding human evaluations. Specifically, we asked expert human annotators to label 1) whether these edits resulted in improvements to the summary based on human judgment and 2) whether the edits made by the summarizer align with the feedback provided in the last step by the evaluator.\nThe results of the human evaluation, presented in Figure 3, indicate that approximately 90% of the edits performed by the summarizer adhered to the provided feedback as intended on both datasets. However, only around 50 -60% of these edits after 2 or more iterations were deemed beneficial according to human judgment, whereas the evaluator in SummIt still asks to perform the refinements. We also notice a clear trend that the percentage of beneficial refinements decreases as the iteration number goes up. The finding shows an Over-correction problem: the LLM may demand itself to continuously refine the summary based on its own evaluation criteria, rather than adhering to the true evaluation criteria of good summaries by humans.\nThis finding highlights the need for better stopping criteria in developing iterative summarization systems, and we argue that incorporating humanin-the-loop may be a potential solution. We leave this for future work."},{"figure_ref":[],"heading":"Case Study","publication_ref":[],"table_ref":[],"text":"We show an example of iterative summary refinement in Table 7. The evaluator provides a detailed rationale for the summary and the summarizer can refine the summary accordingly. The full example can be found in Appendix C."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"In this paper, we propose a new framework for text summarization by iteratively refining summaries with model feedback. Our framework is fully built upon large language models and doesn't require supervised training or reinforcement learning alignment. We also demonstrate that the system improves faithfulness and controllability by incorporating knowledge and topic extractors. We conduct extensive experiments and analyses on three benchmark datasets and experimental results show that our iterative summarization system outperforms the one-shot generation setting systems with LLM, which demonstrates the effectiveness of our method. Our human evaluation finds that the summary refinement by our framework can clearly follow the self-evaluation feedback, but is highly biased toward its own evaluation criteria, rather than human judgment. We believe the potential issue could be addressed with human-in-the-loop feedback. We hope the insights gained from this work can guide future research in building more powerful LLM-based summarization systems."},{"figure_ref":[],"heading":"Limitations","publication_ref":["b8"],"table_ref":[],"text":"Instead of conducting experiments on the entire test set, we randomly sample 1000 examples from each dataset test set due to budget limits. Previous research efforts (Goyal et al., 2022;Zhang et al., 2023d) have also been limited in their testing of GPT-3 on a small number of instances.\nWe only use gpt-3.5-turbo model from openAI API as an instance of large language models. The focus of the paper is to explore the iterative summarization framework with LLM, but not compare different open and closed LLMs."},{"figure_ref":[],"heading":"Appendix","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"A Experimental Setup","publication_ref":[],"table_ref":[],"text":"We use the official checkpoints of the baseline models BART, T5, and PEGASUS from Huggingface. We use gpt-3.5-turbo model 4 as the backbone LLM for both the generation and evaluation of summaries, keeping the temperature parameter at 0 to ensure reproducibility.\nAs for the datasets, we randomly sample 1000 samples with random seed 101 from the test set for both CNN/DM and XSum datasets and use the full test set for the NEWTS dataset. We also tune the LLM optimal prompt and hyperparameters on a dev set of 50 examples. Each discovery experiment was run three times, and the average result was used to mitigate the instability of small datasets."},{"figure_ref":[],"heading":"B Prompts","publication_ref":[],"table_ref":[],"text":"Here we list prompts used in our experiments for extracted and generated summaries in Table 8 andTable 9. Note that according to OpenAI's document, the model could receive two categories of prompts: system prompt and user prompt, where the system prompt functions as the global instruction to initialize the model and the user prompt as the question proposed by users. In our experiment, we leverage both prompts to guide the model and select the best prompts on a dev set of 50 examples."},{"figure_ref":[],"heading":"Model","publication_ref":[],"table_ref":[],"text":"System Prompt Summarizer You are a summarizer that follows the output pattern. You revise the summary based on the given instructions. You follow all the instructions without commenting on them. Make sure the summary is concise and accurate."},{"figure_ref":[],"heading":"Evaluator","publication_ref":[],"table_ref":[],"text":"You are a summary evaluator that follows the output pattern. You give scores for the summaries as well as revise suggestions. Your score should be corresponding to your suggestions. You suggestions can be: 1. Add the information of [] 2. Remove the information of [] 3. Rephrase the information of [] 4. Shorten the summary. 5. Do nothing.\nOnly ask for the information that appeared in the document. If you find the summary is too long, ask for a shorter summary. Keep the summary short and concise. If you think there's no further revision is needed, you must add \"<STOP>\" at the end of your output at the end of the comment. Give precise and clear suggestions.\nTable 8: System prompts of the summarizer and the evaluator for all settings."},{"figure_ref":[],"heading":"C Example summaries","publication_ref":[],"table_ref":[],"text":"Here we show the full example of the iterative summarization and rationale feedback in Table 9: User prompts of summarizer and evaluator for different settings."},{"figure_ref":[],"heading":"Document","publication_ref":[],"table_ref":[],"text":"Media playback is unsupported on your device16 december 2014 last updated at 08:58 gmtdr andrew goodall said a choice may have to be made between emergency care and non-urgent surgery .in an extended interview with bbc wales health correspondent owain clarke , dr goodall said hospitals might make the decision not to schedule surgery in the first place to focus on \" front door pressures \" during the peak winter months .two hospitals , morriston in swansea and the princess of wales in bridgend , have already cancelled some surgical procedures until after christmas ."},{"figure_ref":[],"heading":"Reference","publication_ref":[],"table_ref":[],"text":"The new chief executive of the welsh nhs has told bbc wales winter pressures could force hospitals to cancel operations."},{"figure_ref":[],"heading":"Summary-Iter1","publication_ref":[],"table_ref":[],"text":"Hospitals in Wales may have to choose between emergency care and non-urgent surgery during peak winter months, according to Dr. Andrew Goodall. He suggested that hospitals may decide not to schedule surgery in order to focus on front door pressures. Two hospitals in Swansea and Bridgend have already cancelled some surgical procedures until after Christmas. Evaluation-Iter1 The summary effectively conveys the main point of the article, but it could be shortened for conciseness. Consider removing the specific hospitals mentioned and rephrasing the sentence about hospitals having to choose between emergency care and non-urgent surgery to make it more concise."},{"figure_ref":[],"heading":"Summary-Iter2","publication_ref":[],"table_ref":[],"text":"Hospitals in Wales may have to prioritize emergency care over non-urgent surgery during peak winter months, according to Dr. Andrew Goodall. Some surgical procedures have already been cancelled until after Christmas. Evaluation-Iter2 The revised summary effectively conveys the main point of the article in a concise manner.\nHowever, it could benefit from adding more specific information about the reason for the cancellations and the potential impact on patients. Consider adding information about the reason for the cancellations and the potential impact on patients to make the summary more informative."},{"figure_ref":[],"heading":"Summary-Iter3","publication_ref":[],"table_ref":[],"text":"Dr. Andrew Goodall warns that hospitals in Wales may have to prioritize emergency care over non-urgent surgery during peak winter months, which has already led to some surgical procedures being cancelled until after Christmas. This decision is made to focus on front door pressuresänd ensure that emergency care is not compromised. The cancellations may cause inconvenience to patients, but it is necessary to ensure that emergency care is not overwhelmed. Evaluation-Iter3 The revised summary effectively conveys the main point of the article in a concise and informative manner. It includes specific information about the reason for the cancellations and the potential impact on patients, as well as the importance of prioritizing emergency care. The only suggestion would be to consider shortening the summary slightly for even greater conciseness."},{"figure_ref":[],"heading":"Summary-Iter4","publication_ref":[],"table_ref":[],"text":"Hospitals in Wales may have to prioritize emergency care over non-urgent surgery during peak winter months, leading to some cancellations until after Christmas. Dr. Andrew Goodall emphasizes the importance of focusing on front door pressures ẗo ensure that emergency care is not compromised. While the cancellations may cause inconvenience to patients, it is necessary to prevent emergency care from being overwhelmed. Evaluation-Iter4 The revised summary effectively conveys the main point of the article in a concise and informative manner. It includes specific information about the reason for the cancellations and the potential impact on patients, as well as the importance of prioritizing emergency care. The summary is well-written and does not require any further revision. "}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b4\",\n \"b22\",\n \"b18\",\n \"b14\",\n \"b36\",\n \"b5\",\n \"b13\",\n \"b8\",\n \"b17\"\n ],\n \"table_ref\": [],\n \"text\": \"Document summarization aims to compress text material while retaining its most salient information. With the increasing amount of publicly available text data, the significance of automated summarization techniques has amplified. Recent advancements in summarization systems, leveraging neural networks and pre-trained language models, have demonstrated notable progress (Cheng and Lapata, 2016;Nallapati et al., 2016;Liu and Lapata, 2019;Lewis et al., 2019;Zhang et al., 2020). The above summarization systems are all built end-toend and follow a one-shot paradigm that generates summaries in a single step. On the contrary, humans often write text through an evolutionary pro- cess characterized by multiple iterations of drafting and editing (Faltings et al., 2020).\\nThese end-to-end summarization systems encounter multiple challenges. Firstly, they frequently suffer from the issue of hallucination, resulting in the generation of ungrammatical or factually incorrect content (Kryscinski et al., 2020;Zhang et al., 2022b). Secondly, these systems are often optimized using imperfect reference summaries, and widely adopted evaluation metrics like ROUGE may not accurately assess summary quality. Thirdly, most of these systems lack controllability, as they only produce a single generic summary conditionally on the input document. In practice, instead of a single condensed version of the entire document, generating summaries that cater to specific aspects or queries would be more beneficial to meet the diverse requirements of users.\\nThe emergence of advanced instruction-tuned large language models (LLMs), such as ChatGPT2 , has presented exciting possibilities for summarization systems by exhibiting strong zero-shot performance in various downstream tasks. A recent study by (Goyal et al., 2022) compared GPT-3 with traditional fine-tuning methods and found that despite lower ROUGE scores, human annotators preferred the GPT-3 generated summaries. Another comprehensive analysis by (Zhang et al., 2023d) focused on large language models for news summarization and revealed that the quality of generated summaries is already on par with those created by humans. Furthermore, Liu et al. (2023) demonstrated the utilization of LLMs like GPT-4 as an effective natural language generation evaluator, showing a higher correlation with humans in the summarization task compared to previous reference-based methods.\\nThe advent of LLMs also introduces new opportunities for summarization beyond the traditional one-shot generation setting. In this paper, we introduce SummIt, a framework that leverages large language models for iterative text summarization. Instead of generating summaries in a single step, our framework enables the model to iteratively refine the generated summary through self-evaluation and feedback, resembling the human process of drafting and revising summaries. According to our experiments, the rationale generation and summary refinement in SummIt can be guided effectively with in-context learning, eliminating the need for supervised training or reinforcement learning processes. Additionally, we explore the potential benefits of incorporating knowledge and topic extractors to enhance summary faithfulness and controllability. We instantiate SummIt with ChatGPT as the backbone, and the automatic evaluation results on three benchmark datasets demonstrate the effectiveness of SummIt in improving summary quality, faithfulness, and controllability within only a few iterations. Furthermore, we conduct a human evaluation to validate the iterative refinements quality and identify a potential over-correction issue.\\nWe summarize the contributions of this paper as follows:\\n• We propose SummIt, a novel framework for iterative text summarization. SummIt enables the iterative refinement of generated summaries by incorporating self-evaluation and feedback mechanisms. In addition, we propose to incorporate knowledge and topic extractors to further improve the faithfulness and controllability of SummIt.\\n• We conduct experiments on three summarization benchmark datasets, and empirical results from automatic evaluation demonstrate the effectiveness of our proposed framework in summary refinement.\\n• A human evaluation is conducted to investigate the impact of self-evaluation-guided summary refinement. The results revealed a potential issue of over-correction: while the large language model effectively refines the summary based on the feedback rationale, it exhibits a bias toward its own evaluation criteria, rather than aligning closely with human judgment.\\n2 Related Work\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Text Summarization\",\n \"publication_ref\": [\n \"b30\",\n \"b18\",\n \"b39\",\n \"b14\",\n \"b36\",\n \"b13\",\n \"b19\"\n ],\n \"table_ref\": [],\n \"text\": \"Recent years have witnessed significant advancements in text summarization systems with the development of deep neural networks and pre-trained language models. Automatic summarization methods can be broadly categorized into extractive and abstractive approaches. Extractive summarization involves the direct extraction of sentences from the source text to form summaries (Xu et al., 2019;Liu and Lapata, 2019;Zhong et al., 2020;Zhang et al., 2022aZhang et al., , 2023b)), while abstractive approaches conditionally generate summaries using a sequence-tosequence (seq2seq) framework (Lewis et al., 2019;Zhang et al., 2020). Existing approaches mentioned above generate summaries in a one-shot manner, and their outputs may not always align with user expectations and may contain hallucinated content (Kryscinski et al., 2020). To address the limitation, Liu et al. (2022) proposes to automatically correct factual inconsistencies in generated summaries with generated human feedback. In contrast, our SummIt framework enables iterative summary refinement with self-evaluation and feedback, eliminating the need for costly human annotations. Additionally, we propose the integration of knowledge and topic extractors to further enhance summary faithfulness and controllability.\\nYou are a summarizer that follows the output pattern. You revise the summary based on the given instructions. You follow all the instructions without commenting on them.\\nRefine: [Revise Suggestions] Revise the summary. Follow all the suggestions and you an not make more comments. [Format Instructions] You are a summary evaluator that gives scores for the summaries with revise suggestions. Your suggestions can be: 1. Add the information of <insert> 2.Remove the information of <insert> 3. Rephrase the information of <insert> 4. Shorten the summary 5. Keep the summary unchanged If you think there's no further revision is needed, you must add \\\"<STOP>\\\" at the end.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Knowledge Extractor Topic Extractor\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Summarizer Evaluator\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Refined Summary Evaluation Rationale Refined Summary\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Source Document\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Figure 2: The overall framework of our proposed iterative text summarization system. The evaluator generates an evaluation rationale based on the current summary, and the summarizer then refines the summary accordingly. The knowledge and topic extractors retrieve information from the source document to guide the process.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Summarization with Large Language Models\",\n \"publication_ref\": [\n \"b24\",\n \"b3\",\n \"b8\",\n \"b17\",\n \"b6\",\n \"b20\",\n \"b16\",\n \"b37\"\n ],\n \"table_ref\": [],\n \"text\": \"Recent years have seen a surge in training largescale language models (LLM) on large amounts of text, such as GPT (Radford et al., 2019;Brown et al., 2020). Several studies have explored the application of LLMs in the context of text summarization. For instance, Goyal et al. (2022) compared the performance of GPT-3-generated summaries with traditional fine-tuning methods, finding that although the former achieved slightly lower ROUGE scores, human evaluators expressed a preference for them. Similarly, Zhang et al. (2023d) reported that LLM-generated summaries were on par with human-written summaries in the news domain. Zhang et al. (2023c) benchmarked the performance of ChatGPT on extractive summarization and proposes to improve summary faithfulness with an extract-then-generate pipeline. On the other hand, prior works have also leveraged LLMs for summarization evaluation (Liu et al., 2023;Fu et al., 2023;Luo et al., 2023), demonstrating that LLM-based metrics outperform all previous evaluation metrics like ROUGE (Lin, 2004) and BertScore (Zhang et al., 2019) by a significant margin in terms of correlation with human evaluations.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Text Editing\",\n \"publication_ref\": [\n \"b11\",\n \"b1\",\n \"b5\",\n \"b28\",\n \"b21\"\n ],\n \"table_ref\": [],\n \"text\": \"Our work is also closely related to the task of text editing. Traditional editing models are trained to solve specific tasks, such as information updating (Iso et al., 2020), Wikipedia edit (Reid and Neu-big, 2022), and grammar error correction (Awasthi et al., 2019). Recent works also formulate text editing as an interactive task, such as commandbased editing systems (Faltings et al., 2020), and interactive editing systems (Schick et al., 2022). 2023) propose a reinforcement learningbased approach to generate natural language feedback for correcting generation errors. Concurrent work Madaan et al. (2023) presents a similar generation pipeline that enhances initial outputs through iterative feedback using a single LLM for short text generation tasks. In contrast, our SummIt framework differs from these approaches as it specifically focuses on the conditional generation task of summarization, with an emphasis on improving summary faithfulness and controllability. Additionally, we empirically observe that separating the summarizer and evaluator into different LLMs, each employing different in-context guidance leads to improved performance in our framework.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Methods\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Iterative Summarization\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The overall architecture of our iterative text summarization system SummIt is shown in Figure 2.\\nThe system consists of two major components, a summarizer that generates and refines the summary, and an evaluator that generates feedback rationale.\\nSummarizer: The summarizer is in charge of generating the initial summary and revising a summary conditioned on the given explanations and source document. We instantiate the summarizer with an instruction-tuned language model S.\\nFormally, given the input source document x, the initial summary y 0 generation process can be represented as:\\np S (y 0 | x) = m t=1 p S y 0 t | y 0 <t , x ,(1)\\n, where y 0 <t denotes the generated tokens, y 0 t refers to the t-th summary token, and m denotes the summary length.\\nAfter obtaining the i-step self-evaluation feedback e i from the evaluator E, the summarizer will refine the summary accordingly and then generates refined summary y (i+1) as: p S (y (i+1) | x, e i ).\\nEvaluator: The evaluator is another instance of language model E that generates summary quality evaluation and corresponding explanations e i for the i-th iteration as: p E (e i | x, y i ).\\nStopping Criteria: The evaluator gives a quality assessment of the generated summary and then outputs the rationale for the evaluation as feedback. The summarizer receives model evaluation and feedback from the evaluator, subsequently refining the summary based on this input.\\nThis iterative process can be repeated until 1) the evaluator determines that no further refinement is required or 2) fulfills rule-based stopping criteria, such as reaching a maximum iteration number.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"In-context Learning\",\n \"publication_ref\": [\n \"b3\"\n ],\n \"table_ref\": [],\n \"text\": \"Since the summarizer and evaluator in SummIt are not fine-tuned with supervised data or trained reinforcement learning rewards, it would be beneficial to guide the explanation and summary generation process with the desired format or template. Recent studies have shown that large language models have strong few-shot performance on various downstream tasks, known as in-context learning (ICL) (Brown et al., 2020).\\nThe standard ICL prompts a language model, M , with a set of exemplar source-summary pairs, C = {(x 1 , y 1 )...(x m , y m )}, and generates summary y by concatenating the exemplar sourcesummary pairs and input document as prompt:\\np M (y | x, C).\\nWe also use in-context learning to guide our iterative summarization system, where we use \\\"document-reference summary\\\" pairs as the context for the summarizer S, and use \\\"documentreference summary-human written explanation\\\" triplets as the context for the evaluator E. We empirically find that in-context learning could improve the efficacy of our system.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Summary Faithfulness and Controllability\",\n \"publication_ref\": [\n \"b13\",\n \"b10\",\n \"b40\",\n \"b20\"\n ],\n \"table_ref\": [],\n \"text\": \"In practical applications, the faithfulness of the generated summary holds significant importance, alongside the overall quality of the summary (Kryscinski et al., 2020). Previous research has demonstrated the effectiveness of leveraging knowledge extraction from source documents to enhance the faithfulness of generated summaries (Huang et al., 2020;Zhu et al., 2020). Building upon these insights, we propose integrating a knowledge extractor into our iterative summarization system.\\nKnowledge Extractor: In particular, we utilize OpenIE3 , which extracts knowledge k in the form of triplets from the source document. During each iteration, the summarizer (S) is guided to refine the summary in accordance with the extracted knowledge, represented as: p S (y (i+1) | x, e i , k). Moreover, the evaluator (E) can be directed to factor in faithfulness when delivering feedback, denoted as p E (e i | x, y i , k), as LLMs have shown to be efficient faithfulness evaluators (Luo et al., 2023).\\nFurthermore, real-world applications often require the generation of summaries tailored to specific aspects or queries, rather than a single generic summary of the entire document. Our iterative summarization framework offers enhanced controllability for aspect-based summarization tasks.\\nTopic Extractor: Given an aspect-oriented query q, we prompt both summarizer S and evaluator E to initially extract relevant snippets, each containing less than 5 words, from the source document x. Following the extraction, these components then proceed to either generate or assess the summary by taking into account the extracted snippets. The iterative nature of our framework further facilitates the controllable summary generation, allowing for the easy transformation of generic summaries into topic-focused summaries based on the user's preferences.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Prompt Format\",\n \"publication_ref\": [\n \"b26\",\n \"b5\"\n ],\n \"table_ref\": [],\n \"text\": \"We utilize both system prompts and user prompts following the OpenAI API in our system implementations. The full prompts used in the experiments can be found in Table 8 andTable 9. Notably, we empirically find that pre-defining the possible edit operations for the evaluator improves the system performance significantly since it avoids free-form edits to the summary by the large language model. Thus, we adopt the five types of text editing operations commonly used in text editing systems (Reid and Neubig, 2022;Faltings et al., 2020). We specifically require the evaluator to generate feedback based on the source document and summary at this iteration with the following five types of possible refinement operations:\\n• Add: Add the information of <insert> \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we validate our SummIt framework on three benchmark summarization datasets. We employ both automatic metrics and human assessment to evaluate the quality 4.2, faithfulness 4.3, and controllability 4.4 of the generated summaries.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiment Settings\",\n \"publication_ref\": [\n \"b16\",\n \"b17\",\n \"b13\",\n \"b7\",\n \"b27\",\n \"b12\",\n \"b8\",\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"Datasets: We conduct experiments on the following three publicly available benchmark datasets, as presented in Evaluation metrics: For summary quality, we use ROUGE scores (Lin, 2004) and G-Eval (Liu et al., 2023) as the automatic metrics. We report ROUGE-1, ROUGE-2, and ROUGE-L scores, which respectively measure the overlap of unigrams, bigrams, and the longest common sequence between the generated summary and the reference summary. G-Eval is an LLM-based matrix with a scale ranging from 1 to 5. G-Eval uses LLM with chain-of-thoughts (CoT) and a form-filling paradigm to assess the quality of NLG outputs. It shows the highest correlation with humans compared to other summarization quality metrics. For summary faithfulness, we use FactCC (Kryscinski et al., 2020) and DAE (Defining arc entailment) (Goyal and Durrett, 2020) as our evaluation metrics. FactCC is a weakly supervised BERT-based model metric that verifies factual consistency through rule-based transformations applied to source document sentences. It shows a high correlation in assessing summary faithfulness with human judgments. DAE decomposes entailment at the level of dependency arcs, examining the semantic relationships within the generated output and input. Rather than focusing on aggregate decisions, DAE measures the semantic relationship manifested by individual dependency arcs in the generated output supported by the input.\\nFor the controllability of query-focused summarization, we use BM25 (Robertson et al., 2009) and DPR (Karpukhin et al., 2020) to measure the similarity between the query and the summary with both sparse and dense evaluations. BM25 is a probabilistic retrieval function that ranks documents based on query term frequency. DPR leverages dense vector representations for scalable retrieval, embedding both questions and passages into fixedlength vector spaces for nuanced similarity calculations. In line with previous research findings that have emphasized the inclination of human annotators towards summaries generated by LLM models, even in the presence of comparatively lower ROUGE scores (Goyal et al., 2022), we further validate the effectiveness of SummIt through a dedicated human study. Specifically, we use 1) five-point Likert scale ratings (Likert, 1932) covering summary coherence, fluency, relevance, consistency, conciseness, and overall evaluation, and 2) human preference test, where annotators are shown summaries of the same source document from all five summarization systems and then asked to select their most preferred summary or summaries.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Model\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We evaluated the performance using 1000 ran-dom samples from CNN/DM and XSum test sets, with seed 101, and the full NEWTS test set. Our prompts were refined with a 50-example development set. The detailed experimental setup is provided in Appendix A.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Generic Summary Quality Evaluation\",\n \"publication_ref\": [\n \"b36\",\n \"b14\",\n \"b25\",\n \"b8\"\n ],\n \"table_ref\": [\n \"tab_2\",\n \"tab_2\",\n \"tab_3\"\n ],\n \"text\": \"The automatic evaluation results for generic summarization quality are shown in Table 2. We use previous pre-trained language models, including PEGASUS (Zhang et al., 2020), BART (Lewis et al., 2019), and T5 (Raffel et al., 2020) as baseline models. We compare our framework SummIt with these baseline models under a zero-shot setting for a fair comparison.\\nIt is observed that SummIt has inferior ROUGE scores compared to fine-tuning approaches on CN-N/DM, while exhibiting significantly higher LLMbased evaluation metric G-Eval. On the other hand, it outperforms all baseline methods on the XSum dataset. Compared to the output of ChatGPT, the summaries of ChatGPT after our iterative refinement see a consistent improvement in the G-Eval score. The results are consistent with the previous conclusions in (Zhang et al., 2023d), where large language model summary outputs receive lower ROUGE scores due to the low quality of reference summaries.\\nIn addition to the zero-shot setting, we investigate the effects of in-context learning for SummIt, as shown in the lower block of Table 2. The results consistently demonstrate that incorporating incontext learning significantly enhances the model's performance on ROUGE and G-Eval scores. This observation underscores the substantial few-shot capabilities of SummIt, showcasing its ability to adapt effectively and generate high-quality summaries in contexts with very few examples.\\nTo further verify the summary quality, we conduct a human study to evaluate the overall quality of the summaries as shown in Table 3. According to the five-point Likert scale ratings, the summaries of ChatGPT and SummIt consistently outperform pre-trained language model results. The iterative refinement of SummIt also provides consistent improvements, which align with the G-Eval results obtained from the automatic evaluation. We also conducted a human preference study, where summaries from all models were presented to human annotators. They were tasked to select the best summary, without any prior knowledge of the origin of each summary. Consistent with the findings in (Goyal et al., 2022), the results reveal a clear preference among human annotators for summaries generated by large language models (LLMs) for both CNN (86%) and BBC (62%) style summaries. We also notice the summaries of ChatGPT after our iterative refinement (SummIt) show a significant improvement in human preference, with 18% and 14% percent improvements on CNN/DM and XSum datasets. The results demonstrate the effectiveness of refining generic summaries of our framework. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Summary Faithfulness Evaluation\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\"\n ],\n \"text\": \"To evaluate the efficacy of the SummIt framework in enhancing summary faithfulness with the knowledge extractor, we conducted additional experiments, as presented in Table 4. The findings demonstrate that our framework's iterative approach to refining summaries yields significant improvements in summary faithfulness, as indicated by both FactCC and DAE results. Furthermore, the integration of a knowledge extractor such as OpenIE further enhances the level of faithfulness.\\nThe LLM-based evaluation score G-Eval also indicates a higher level of satisfaction with the refined summaries when guided by the extracted knowledge triplets. In conclusion, our study reveals that iterative refinements with the incorporation of the knowledge extractor effectively enhance summary faithfulness without compromising the quality of the summaries.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Query-focued Summarization Controlability Evaluation\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_5\"\n ],\n \"text\": \"We utilize the query-based summarization dataset NEWTS as our testbed to demonstrate the controllability ability of SummIt. The results obtained, as depicted in Table 5, highlight the framework's capability to align the focus of a generic summary with the specific topic of interest or query provided by the user. We also observe improved G-Eval evaluation scores by directing the summary generation process toward the intended topic. Furthermore, we evaluate the controllability of the summarization systems by quantifying the sim- ilarity between the query and the generated summary. Both BM25 and DPR are employed as similarity metrics, and we consistently observe enhancements after the iterative refinement process. This observation serves as evidence that SummIt effectively refines the summary to align with the topics specified in the query.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ablation Studies\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_7\"\n ],\n \"text\": \"Table 6 shows the results of the ablation study by removing refinement operations. The ablation study is conducted on the CNN/DM dataset under the zero-shot settings. According to the results, each option contributes to the success of our method, and the add operation affects the ROUGE score most, while the simplify operation affects the GPTevaluation scores the most. Without adding the operation, the information in the iterative process will only decrease, resulting in less n-gram overlap. On the other hand, without the simplify and remove operations, the redundant information results in low G-Eval scores.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Over-correction Issue\",\n \"publication_ref\": [\n \"b17\"\n ],\n \"table_ref\": [],\n \"text\": \"A recent work (Liu et al., 2023) Evaluation-Iter1: The summary effectively conveys the main point of the article, but it could be shortened for conciseness. Consider removing the specific hospitals mentioned and rephrasing the sentence about hospitals having to choose between emergency care and non-urgent surgery to make it more concise.\\nSummary-Iter2: Hospitals in Wales may have to prioritize emergency care over non-urgent surgery during peak winter months, according to Dr. Andrew Goodall. Some surgical procedures have already been cancelled until after Christmas. more iterations... Table 7: An example of iterative summary refinement from the XSum dataset. The revision between the two iterations and their corresponding comments are presented in the same color. The blue color refers to the rephrase revision and the orange color refers to the remove operation.\\n1) Does the refinement actually improve the summary? Another potential issue of SummIt would be:\\n2) Does the refinement actually follow the rationale feedback from the evaluator?\\nTo address these two concerns and provide further validation for the step-wise summary refinement in SummIt, we conducted the corresponding human evaluations. Specifically, we asked expert human annotators to label 1) whether these edits resulted in improvements to the summary based on human judgment and 2) whether the edits made by the summarizer align with the feedback provided in the last step by the evaluator.\\nThe results of the human evaluation, presented in Figure 3, indicate that approximately 90% of the edits performed by the summarizer adhered to the provided feedback as intended on both datasets. However, only around 50 -60% of these edits after 2 or more iterations were deemed beneficial according to human judgment, whereas the evaluator in SummIt still asks to perform the refinements. We also notice a clear trend that the percentage of beneficial refinements decreases as the iteration number goes up. The finding shows an Over-correction problem: the LLM may demand itself to continuously refine the summary based on its own evaluation criteria, rather than adhering to the true evaluation criteria of good summaries by humans.\\nThis finding highlights the need for better stopping criteria in developing iterative summarization systems, and we argue that incorporating humanin-the-loop may be a potential solution. We leave this for future work.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Case Study\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We show an example of iterative summary refinement in Table 7. The evaluator provides a detailed rationale for the summary and the summarizer can refine the summary accordingly. The full example can be found in Appendix C.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we propose a new framework for text summarization by iteratively refining summaries with model feedback. Our framework is fully built upon large language models and doesn't require supervised training or reinforcement learning alignment. We also demonstrate that the system improves faithfulness and controllability by incorporating knowledge and topic extractors. We conduct extensive experiments and analyses on three benchmark datasets and experimental results show that our iterative summarization system outperforms the one-shot generation setting systems with LLM, which demonstrates the effectiveness of our method. Our human evaluation finds that the summary refinement by our framework can clearly follow the self-evaluation feedback, but is highly biased toward its own evaluation criteria, rather than human judgment. We believe the potential issue could be addressed with human-in-the-loop feedback. We hope the insights gained from this work can guide future research in building more powerful LLM-based summarization systems.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [\n \"b8\"\n ],\n \"table_ref\": [],\n \"text\": \"Instead of conducting experiments on the entire test set, we randomly sample 1000 examples from each dataset test set due to budget limits. Previous research efforts (Goyal et al., 2022;Zhang et al., 2023d) have also been limited in their testing of GPT-3 on a small number of instances.\\nWe only use gpt-3.5-turbo model from openAI API as an instance of large language models. The focus of the paper is to explore the iterative summarization framework with LLM, but not compare different open and closed LLMs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Appendix\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Experimental Setup\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We use the official checkpoints of the baseline models BART, T5, and PEGASUS from Huggingface. We use gpt-3.5-turbo model 4 as the backbone LLM for both the generation and evaluation of summaries, keeping the temperature parameter at 0 to ensure reproducibility.\\nAs for the datasets, we randomly sample 1000 samples with random seed 101 from the test set for both CNN/DM and XSum datasets and use the full test set for the NEWTS dataset. We also tune the LLM optimal prompt and hyperparameters on a dev set of 50 examples. Each discovery experiment was run three times, and the average result was used to mitigate the instability of small datasets.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B Prompts\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Here we list prompts used in our experiments for extracted and generated summaries in Table 8 andTable 9. Note that according to OpenAI's document, the model could receive two categories of prompts: system prompt and user prompt, where the system prompt functions as the global instruction to initialize the model and the user prompt as the question proposed by users. In our experiment, we leverage both prompts to guide the model and select the best prompts on a dev set of 50 examples.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Model\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"System Prompt Summarizer You are a summarizer that follows the output pattern. You revise the summary based on the given instructions. You follow all the instructions without commenting on them. Make sure the summary is concise and accurate.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Evaluator\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"You are a summary evaluator that follows the output pattern. You give scores for the summaries as well as revise suggestions. Your score should be corresponding to your suggestions. You suggestions can be: 1. Add the information of [] 2. Remove the information of [] 3. Rephrase the information of [] 4. Shorten the summary. 5. Do nothing.\\nOnly ask for the information that appeared in the document. If you find the summary is too long, ask for a shorter summary. Keep the summary short and concise. If you think there's no further revision is needed, you must add \\\"<STOP>\\\" at the end of your output at the end of the comment. Give precise and clear suggestions.\\nTable 8: System prompts of the summarizer and the evaluator for all settings.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C Example summaries\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Here we show the full example of the iterative summarization and rationale feedback in Table 9: User prompts of summarizer and evaluator for different settings.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Document\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Media playback is unsupported on your device16 december 2014 last updated at 08:58 gmtdr andrew goodall said a choice may have to be made between emergency care and non-urgent surgery .in an extended interview with bbc wales health correspondent owain clarke , dr goodall said hospitals might make the decision not to schedule surgery in the first place to focus on \\\" front door pressures \\\" during the peak winter months .two hospitals , morriston in swansea and the princess of wales in bridgend , have already cancelled some surgical procedures until after christmas .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Reference\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The new chief executive of the welsh nhs has told bbc wales winter pressures could force hospitals to cancel operations.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Summary-Iter1\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Hospitals in Wales may have to choose between emergency care and non-urgent surgery during peak winter months, according to Dr. Andrew Goodall. He suggested that hospitals may decide not to schedule surgery in order to focus on front door pressures. Two hospitals in Swansea and Bridgend have already cancelled some surgical procedures until after Christmas. Evaluation-Iter1 The summary effectively conveys the main point of the article, but it could be shortened for conciseness. Consider removing the specific hospitals mentioned and rephrasing the sentence about hospitals having to choose between emergency care and non-urgent surgery to make it more concise.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Summary-Iter2\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Hospitals in Wales may have to prioritize emergency care over non-urgent surgery during peak winter months, according to Dr. Andrew Goodall. Some surgical procedures have already been cancelled until after Christmas. Evaluation-Iter2 The revised summary effectively conveys the main point of the article in a concise manner.\\nHowever, it could benefit from adding more specific information about the reason for the cancellations and the potential impact on patients. Consider adding information about the reason for the cancellations and the potential impact on patients to make the summary more informative.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Summary-Iter3\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Dr. Andrew Goodall warns that hospitals in Wales may have to prioritize emergency care over non-urgent surgery during peak winter months, which has already led to some surgical procedures being cancelled until after Christmas. This decision is made to focus on front door pressuresänd ensure that emergency care is not compromised. The cancellations may cause inconvenience to patients, but it is necessary to ensure that emergency care is not overwhelmed. Evaluation-Iter3 The revised summary effectively conveys the main point of the article in a concise and informative manner. It includes specific information about the reason for the cancellations and the potential impact on patients, as well as the importance of prioritizing emergency care. The only suggestion would be to consider shortening the summary slightly for even greater conciseness.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Summary-Iter4\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Hospitals in Wales may have to prioritize emergency care over non-urgent surgery during peak winter months, leading to some cancellations until after Christmas. Dr. Andrew Goodall emphasizes the importance of focusing on front door pressures ẗo ensure that emergency care is not compromised. While the cancellations may cause inconvenience to patients, it is necessary to prevent emergency care from being overwhelmed. Evaluation-Iter4 The revised summary effectively conveys the main point of the article in a concise and informative manner. It includes specific information about the reason for the cancellations and the potential impact on patients, as well as the importance of prioritizing emergency care. The summary is well-written and does not require any further revision. \"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-10-09"},"doi":{"kind":"string","value":"10.18653/v1/D19-1435"},"references":{"kind":"list like","value":[{"authors":"Afra Feyza Akyürek; Ekin Akyürek; Aman Madaan; Ashwin Kalyan; Peter Clark; Derry Wijaya; Niket Tandon","journal":"","ref_id":"b0","title":"Rl4f: Generating natural language feedback with reinforcement learning for repairing model outputs","year":"2023"},{"authors":"Abhijeet Awasthi; Sunita Sarawagi; Rasna Goyal; Sabyasachi Ghosh; Vihari Piratla","journal":"Association for Computational Linguistics","ref_id":"b1","title":"Parallel iterative edit models for local sequence transduction","year":"2019"},{"authors":"Seyed Ali Bahrainian; Sheridan Feucht; Carsten Eickhoff","journal":"","ref_id":"b2","title":"Newts: A corpus for news topic-focused summarization","year":"2022"},{"authors":"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell","journal":"Advances in neural information processing systems","ref_id":"b3","title":"Language models are few-shot learners","year":"2020"},{"authors":"Jianpeng Cheng; Mirella Lapata","journal":"","ref_id":"b4","title":"Neural summarization by extracting sentences and words","year":"2016"},{"authors":"Felix Faltings; Michel Galley; Gerold Hintz; Chris Brockett; Chris Quirk; Jianfeng Gao; Bill Dolan","journal":"","ref_id":"b5","title":"Text editing by command","year":"2020"},{"authors":"Jinlan Fu; See-Kiong Ng; Zhengbao Jiang; Pengfei Liu","journal":"","ref_id":"b6","title":"Gptscore: Evaluate as you desire","year":"2023"},{"authors":"Tanya Goyal; Greg Durrett","journal":"","ref_id":"b7","title":"Evaluating factuality in generation with dependency-level entailment","year":"2020"},{"authors":"Tanya Goyal; Junyi ; Jessy Li; Greg Durrett","journal":"","ref_id":"b8","title":"News summarization and evaluation in the era of gpt-3","year":"2022"},{"authors":"Karl Moritz Hermann; Tomas Kocisky; Edward Grefenstette; Lasse Espeholt; Will Kay; Mustafa Suleyman; Phil Blunsom","journal":"","ref_id":"b9","title":"Teaching machines to read and comprehend","year":"2015"},{"authors":"Luyang Huang; Lingfei Wu; Lu Wang","journal":"","ref_id":"b10","title":"Knowledge graph-augmented abstractive summarization with semantic-driven cloze reward","year":"2020"},{"authors":"Hayate Iso; Chao Qiao; Hang Li","journal":"Association for Computational Linguistics","ref_id":"b11","title":"Fact-based Text Editing","year":"2020"},{"authors":"Vladimir Karpukhin; Barlas Oguz; Sewon Min; Patrick Lewis; Ledell Wu; Sergey Edunov; Danqi Chen; Wen-Tau Yih","journal":"Association for Computational Linguistics","ref_id":"b12","title":"Dense passage retrieval for opendomain question answering","year":"2020"},{"authors":"Wojciech Kryscinski; Bryan Mccann; Caiming Xiong; Richard Socher","journal":"Association for Computational Linguistics","ref_id":"b13","title":"Evaluating the factual consistency of abstractive text summarization","year":"2020"},{"authors":"Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Ves Stoyanov; Luke Zettlemoyer","journal":"","ref_id":"b14","title":"Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension","year":"2019"},{"authors":"Rensis Likert","journal":"Archives of psychology","ref_id":"b15","title":"A technique for the measurement of attitudes","year":"1932"},{"authors":"Chin-Yew Lin","journal":"","ref_id":"b16","title":"Rouge: A package for automatic evaluation of summaries","year":"2004"},{"authors":"Yang Liu; Dan Iter; Yichong Xu; Shuohang Wang; Ruochen Xu; Chenguang Zhu","journal":"","ref_id":"b17","title":"Gpteval: Nlg evaluation using gpt-4 with better human alignment","year":"2023"},{"authors":"Yang Liu; Mirella Lapata","journal":"","ref_id":"b18","title":"Text summarization with pretrained encoders","year":"2019"},{"authors":"Yixin Liu; Budhaditya Deb; Milagro Teruel; Aaron L Halfaker; Dragomir R Radev; Ahmed Hassan; Awadallah ","journal":"","ref_id":"b19","title":"On improving summarization factual consistency from natural language feedback","year":"2022"},{"authors":"Zheheng Luo; Qianqian Xie; Sophia Ananiadou","journal":"","ref_id":"b20","title":"Chatgpt as a factual inconsistency evaluator for abstractive text summarization","year":"2023"},{"authors":"Aman Madaan; Niket Tandon; Prakhar Gupta; Skyler Hallinan; Luyu Gao; Sarah Wiegreffe; Uri Alon; Nouha Dziri; Shrimai Prabhumoye; Yiming Yang","journal":"","ref_id":"b21","title":"Self-refine: Iterative refinement with self-feedback","year":"2023"},{"authors":"Ramesh Nallapati; Bowen Zhou; Caglar Gulcehre; Bing Xiang","journal":"","ref_id":"b22","title":"Abstractive text summarization using sequence-to-sequence rnns and beyond","year":"2016"},{"authors":"Shashi Narayan; Shay B Cohen; Mirella Lapata","journal":"","ref_id":"b23","title":"Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization","year":"2018"},{"authors":"Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever","journal":"OpenAI blog","ref_id":"b24","title":"Language models are unsupervised multitask learners","year":"2019"},{"authors":"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu","journal":"The Journal of Machine Learning Research","ref_id":"b25","title":"Exploring the limits of transfer learning with a unified text-to-text transformer","year":"2020"},{"authors":"Machel Reid; Graham Neubig","journal":"","ref_id":"b26","title":"Learning to model editing processes","year":"2022"},{"authors":"Stephen Robertson; Hugo Zaragoza","journal":"Foundations and Trends® in Information Retrieval","ref_id":"b27","title":"The probabilistic relevance framework: Bm25 and beyond","year":"2009"},{"authors":"Timo Schick; Jane Dwivedi-Yu; Zhengbao Jiang; Fabio Petroni; Patrick Lewis; Gautier Izacard; Qingfei You; Christoforos Nalmpantis; Edouard Grave; Sebastian Riedel","journal":"","ref_id":"b28","title":"Peer: A collaborative language model","year":"2022"},{"authors":"Sean Welleck; Ximing Lu; Peter West; Faeze Brahman; Tianxiao Shen; Daniel Khashabi; Yejin Choi","journal":"","ref_id":"b29","title":"Generating sequences by learning to self-correct","year":"2022"},{"authors":"Jiacheng Xu; Zhe Gan; Yu Cheng; Jingjing Liu","journal":"","ref_id":"b30","title":"Discourse-aware neural extractive model for text summarization","year":"2019"},{"authors":"Haopeng Zhang; Hayate Iso; Sairam Gurajada; Nikita Bhutani","journal":"","ref_id":"b31","title":"Xatu: A fine-grained instruction-based benchmark for explainable text updates","year":"2023"},{"authors":"Haopeng Zhang; Xiao Liu; Jiawei Zhang; ; ","journal":"Association for Computational Linguistics","ref_id":"b32","title":"HEGEL: Hypergraph transformer for long document summarization","year":"2022"},{"authors":"Haopeng Zhang; Xiao Liu; Jiawei Zhang","journal":"Association for Computational Linguistics","ref_id":"b33","title":"DiffuSum: Generation enhanced extractive summarization with diffusion","year":"2023"},{"authors":"Haopeng Zhang; Xiao Liu; Jiawei Zhang","journal":"","ref_id":"b34","title":"Extractive summarization via chatgpt for faithful summary generation","year":"2023"},{"authors":"Haopeng Zhang; Semih Yavuz; Wojciech Kryscinski; Kazuma Hashimoto; Yingbo Zhou","journal":"Association for Computational Linguistics","ref_id":"b35","title":"Improving the faithfulness of abstractive summarization via entity coverage control","year":"2022"},{"authors":"Jingqing Zhang; Yao Zhao; Mohammad Ahmad Saleh; Peter J Liu","journal":"","ref_id":"b36","title":"Pegasus: Pretraining with extracted gap-sentences for abstractive summarization by sequence-to-sequence models","year":"2020"},{"authors":"Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi","journal":"","ref_id":"b37","title":"Bertscore: Evaluating text generation with bert","year":"2019"},{"authors":"Tianyi Zhang; Faisal Ladhak; Esin Durmus; Percy Liang; Kathleen Mckeown; Tatsunori B Hashimoto","journal":"","ref_id":"b38","title":"Benchmarking large language models for news summarization","year":"2023"},{"authors":"Ming Zhong; Pengfei Liu; Yiran Chen; Danqing Wang; Xipeng Qiu; Xuanjing Huang","journal":"","ref_id":"b39","title":"Extractive summarization as text matching","year":"2020"},{"authors":"Chenguang Zhu; William Hinthorn; Ruochen Xu; Qingkai Zeng; Michael Zeng; Xuedong Huang; Meng Jiang","journal":"","ref_id":"b40","title":"Enhancing factual consistency of abstractive summarization","year":"2020"}],"string":"[\n {\n \"authors\": \"Afra Feyza Akyürek; Ekin Akyürek; Aman Madaan; Ashwin Kalyan; Peter Clark; Derry Wijaya; Niket Tandon\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Rl4f: Generating natural language feedback with reinforcement learning for repairing model outputs\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Abhijeet Awasthi; Sunita Sarawagi; Rasna Goyal; Sabyasachi Ghosh; Vihari Piratla\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b1\",\n \"title\": \"Parallel iterative edit models for local sequence transduction\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Seyed Ali Bahrainian; Sheridan Feucht; Carsten Eickhoff\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Newts: A corpus for news topic-focused summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b3\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jianpeng Cheng; Mirella Lapata\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Neural summarization by extracting sentences and words\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Felix Faltings; Michel Galley; Gerold Hintz; Chris Brockett; Chris Quirk; Jianfeng Gao; Bill Dolan\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Text editing by command\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jinlan Fu; See-Kiong Ng; Zhengbao Jiang; Pengfei Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Gptscore: Evaluate as you desire\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Tanya Goyal; Greg Durrett\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Evaluating factuality in generation with dependency-level entailment\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Tanya Goyal; Junyi ; Jessy Li; Greg Durrett\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"News summarization and evaluation in the era of gpt-3\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Karl Moritz Hermann; Tomas Kocisky; Edward Grefenstette; Lasse Espeholt; Will Kay; Mustafa Suleyman; Phil Blunsom\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Teaching machines to read and comprehend\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Luyang Huang; Lingfei Wu; Lu Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Knowledge graph-augmented abstractive summarization with semantic-driven cloze reward\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Hayate Iso; Chao Qiao; Hang Li\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b11\",\n \"title\": \"Fact-based Text Editing\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Vladimir Karpukhin; Barlas Oguz; Sewon Min; Patrick Lewis; Ledell Wu; Sergey Edunov; Danqi Chen; Wen-Tau Yih\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b12\",\n \"title\": \"Dense passage retrieval for opendomain question answering\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Wojciech Kryscinski; Bryan Mccann; Caiming Xiong; Richard Socher\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b13\",\n \"title\": \"Evaluating the factual consistency of abstractive text summarization\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Mike Lewis; Yinhan Liu; Naman Goyal; Marjan Ghazvininejad; Abdelrahman Mohamed; Omer Levy; Ves Stoyanov; Luke Zettlemoyer\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Rensis Likert\",\n \"journal\": \"Archives of psychology\",\n \"ref_id\": \"b15\",\n \"title\": \"A technique for the measurement of attitudes\",\n \"year\": \"1932\"\n },\n {\n \"authors\": \"Chin-Yew Lin\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Rouge: A package for automatic evaluation of summaries\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Yang Liu; Dan Iter; Yichong Xu; Shuohang Wang; Ruochen Xu; Chenguang Zhu\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Gpteval: Nlg evaluation using gpt-4 with better human alignment\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Yang Liu; Mirella Lapata\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Text summarization with pretrained encoders\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yixin Liu; Budhaditya Deb; Milagro Teruel; Aaron L Halfaker; Dragomir R Radev; Ahmed Hassan; Awadallah \",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"On improving summarization factual consistency from natural language feedback\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Zheheng Luo; Qianqian Xie; Sophia Ananiadou\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Chatgpt as a factual inconsistency evaluator for abstractive text summarization\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Aman Madaan; Niket Tandon; Prakhar Gupta; Skyler Hallinan; Luyu Gao; Sarah Wiegreffe; Uri Alon; Nouha Dziri; Shrimai Prabhumoye; Yiming Yang\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Self-refine: Iterative refinement with self-feedback\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Ramesh Nallapati; Bowen Zhou; Caglar Gulcehre; Bing Xiang\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Abstractive text summarization using sequence-to-sequence rnns and beyond\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Shashi Narayan; Shay B Cohen; Mirella Lapata\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever\",\n \"journal\": \"OpenAI blog\",\n \"ref_id\": \"b24\",\n \"title\": \"Language models are unsupervised multitask learners\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu\",\n \"journal\": \"The Journal of Machine Learning Research\",\n \"ref_id\": \"b25\",\n \"title\": \"Exploring the limits of transfer learning with a unified text-to-text transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Machel Reid; Graham Neubig\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Learning to model editing processes\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Stephen Robertson; Hugo Zaragoza\",\n \"journal\": \"Foundations and Trends® in Information Retrieval\",\n \"ref_id\": \"b27\",\n \"title\": \"The probabilistic relevance framework: Bm25 and beyond\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Timo Schick; Jane Dwivedi-Yu; Zhengbao Jiang; Fabio Petroni; Patrick Lewis; Gautier Izacard; Qingfei You; Christoforos Nalmpantis; Edouard Grave; Sebastian Riedel\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Peer: A collaborative language model\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Sean Welleck; Ximing Lu; Peter West; Faeze Brahman; Tianxiao Shen; Daniel Khashabi; Yejin Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Generating sequences by learning to self-correct\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jiacheng Xu; Zhe Gan; Yu Cheng; Jingjing Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Discourse-aware neural extractive model for text summarization\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Haopeng Zhang; Hayate Iso; Sairam Gurajada; Nikita Bhutani\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Xatu: A fine-grained instruction-based benchmark for explainable text updates\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Haopeng Zhang; Xiao Liu; Jiawei Zhang; ; \",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b32\",\n \"title\": \"HEGEL: Hypergraph transformer for long document summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Haopeng Zhang; Xiao Liu; Jiawei Zhang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b33\",\n \"title\": \"DiffuSum: Generation enhanced extractive summarization with diffusion\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Haopeng Zhang; Xiao Liu; Jiawei Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Extractive summarization via chatgpt for faithful summary generation\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Haopeng Zhang; Semih Yavuz; Wojciech Kryscinski; Kazuma Hashimoto; Yingbo Zhou\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b35\",\n \"title\": \"Improving the faithfulness of abstractive summarization via entity coverage control\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jingqing Zhang; Yao Zhao; Mohammad Ahmad Saleh; Peter J Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Pegasus: Pretraining with extracted gap-sentences for abstractive summarization by sequence-to-sequence models\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Bertscore: Evaluating text generation with bert\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Tianyi Zhang; Faisal Ladhak; Esin Durmus; Percy Liang; Kathleen Mckeown; Tatsunori B Hashimoto\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Benchmarking large language models for news summarization\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Ming Zhong; Pengfei Liu; Yiran Chen; Danqing Wang; Xipeng Qiu; Xuanjing Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Extractive summarization as text matching\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Chenguang Zhu; William Hinthorn; Ruochen Xu; Qingkai Zeng; Michael Zeng; Xuedong Huang; Meng Jiang\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Enhancing factual consistency of abstractive summarization\",\n \"year\": \"2020\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[4,103.17,243.19,186.69,33.58],"formula_id":"formula_0","formula_text":"p S (y 0 | x) = m t=1 p S y 0 t | y 0 <t , x ,(1)"},{"formula_coordinates":[4,306.14,244.68,62.6,10.72],"formula_id":"formula_1","formula_text":"p M (y | x, C)."}],"string":"[\n {\n \"formula_coordinates\": [\n 4,\n 103.17,\n 243.19,\n 186.69,\n 33.58\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"p S (y 0 | x) = m t=1 p S y 0 t | y 0 <t , x ,(1)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 306.14,\n 244.68,\n 62.6,\n 10.72\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"p M (y | x, C).\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"SummIt: Iterative Text Summarization via ChatGPT"},"abstract":{"kind":"string","value":"Text summarization systems have made significant progress in recent years, but typically generate summaries in one single step. However, the one-shot summarization setting is sometimes inadequate, as the generated summary may contain hallucinations or overlook essential details related to the reader's interests. This paper addresses this limitation by proposing SummIt, an iterative text summarization framework based on large language models like Chat-GPT. Our framework enables the model to refine the generated summary iteratively through self-evaluation and feedback, resembling humans' iterative process when drafting and revising summaries. Furthermore, we explore the potential benefits of integrating knowledge and topic extractors into the framework to enhance summary faithfulness and controllability. We automatically evaluate the performance of our framework on three benchmark summarization datasets. We also conduct a human evaluation to validate the effectiveness of the iterative refinements and identify a potential issue of overcorrection."},"authors":{"kind":"string","value":"Haopeng Zhang; Xiao Liu; Jiawei Zhang"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: An illustration of the iterative summarization process. The summarizer continuously refines the summary according to self-feedback from the evaluator at each iteration.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Zhang et al. (2023a) also proposed a benchmark for fine-grained instruction-based editing. Recently, Welleck et al. (2022) introduced a selfcorrective learning framework that incorporates a corrector into the language model to facilitate selfcorrection during sequence generation. Akyürek et al. (","figure_data":"","figure_id":"fig_1","figure_label":"","figure_type":"figure"},{"figure_caption":"•Remove: Remove the information of <insert> from the summary • Rephrase: Rephrase the information of <in-sert> in the summary • Simplify: Shorten the summary • Keep: Keep the summary unchanged","figure_data":"","figure_id":"fig_2","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure3: Human evaluation to justify the refinement behavior of SummIt. The top plot refers to the human justification of the ratio that the summary is improved at each iteration and the bottom plot indicates the ratio that the summarizer follows the evaluator's evaluation rationale.","figure_data":"","figure_id":"fig_3","figure_label":"3","figure_type":"figure"},{"figure_caption":"Detailed statistics of the experimental datasets. Doc # words and Sum # words refer to the average word number in the source document and summary.","figure_data":"Dataset#TestDoc #wordsSum #words#SumXSum11,334430.223.31CNN/DM 11,489766.158.21NEWTS600738.570.12","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"","figure_data":", ensuring they are con-sistent with previous fine-tuning approaches: 1)","figure_id":"tab_1","figure_label":"1","figure_type":"table"},{"figure_caption":"This table presents results from experiments conducted on the CNN/DM and XSum datasets under both zero-shot and few-shot settings. A random sample of 1, 000 data points was taken from each dataset for evaluation. G-Eval represents the score evaluated by the ChatGPT evaluator in our framework.","figure_data":"CNN/DMXSumR1R2RLG-EvalR1R2RLG-EvalZero-shot settingPEGASUS ZS 32.90 13.28 29.383.2319.27 3.00 12.723.52BART ZS32.83 13.30 29.643.4219.26 3.30 14.673.49T5 ZS39.68 17.24 26.283.4719.66 2.91 15.313.55ChatGPT39.44 16.14 29.833.4621.61 5.98 17.603.47SummIt (ours) 36.50 13.49 26.764.3321.92 5.93 17.624.24Few-shot settingChatGPT40.00 16.39 30.023.5723.96 7.36 19.363.57SummIt (ours) 37.29 13.60 26.874.3522.04 6.20 17.464.32ModelCoherence Fluency Relevance Consistency Conciseness Overall Human PrefCNN/DMBART3.924.164.003.123.643.240.04T53.724.244.323.523.843.680.10PEGASUS3.203.533.332.871.851.630.00ChatGPT4.204.364.284.013.924.010.34SummIt4.244.504.294.123.844.090.52XSumBART3.974.304.133.303.933.840.30T53.844.324.023.633.843.250.08PEGASUS3.134.103.522.872.032.410.00ChatGPT4.034.404.303.933.873.920.24SummIt4.044.354.284.053.723.960.38","figure_id":"tab_2","figure_label":"2","figure_type":"table"},{"figure_caption":"Human study results on generic summary quality. The first five columns include Likert scale ratings and the last column is the human preference results.","figure_data":"","figure_id":"tab_3","figure_label":"3","figure_type":"table"},{"figure_caption":"Experimental results of incorporating knowledge extractor on summary quality and faithfulness on XSum dataset. -IE refers to the model integrated with OpenIE.","figure_data":"R1R2RLG-Eval FactCC DAEChatGPT21.61 5.98 17.603.4728.0010.34SummIt21.92 5.93 17.624.2436.0033.02ChatGPT-IE 22.01 5.11 17.063.8551.6893.68SummIt-IE19.72 3.85 15.364.9547.2490.36R1R2RLG-Eval BM25 DPRChatGPT30.018.9427.031.0633.0977.22ChatGPT-Topic 33.24 10.20 29.881.1636.2078.77SummIt-Topic30.458.4827.194.7439.1182.41","figure_id":"tab_4","figure_label":"4","figure_type":"table"},{"figure_caption":"Experimental results on NEWTS dataset to test the controllability of our framework. -Topic indicates a model that is prompted to extract topic-related snippets before generating a summary.","figure_data":"","figure_id":"tab_5","figure_label":"5","figure_type":"table"},{"figure_caption":"highlights the potential issue of LLM-based evaluators having a bias towards the LLM outputs, which raises the doubt:","figure_data":"R1R2RLG-EvalSummIt36.50 13.49 26.764.33-w/o Add33.01 11.55 24.713.98-w/o Remove36.46 13.44 26.553.64-w/o Rephrase 34.71 12.12 26.313.82-w/o Simplify33.49 12.33 25.763.55-w/o Keep33.87 13.03 25.703.94","figure_id":"tab_6","figure_label":"","figure_type":"table"},{"figure_caption":"Ablation Study on Iterative Refinement OperationsSummary-Iter1: Hospitals in Wales may have to choose between emergency care and non-urgent surgery during peak winter months, according to Dr. Andrew Goodall. He suggested that hospitals may decide not to schedule surgery in order to focus on \"front door pressures.\" Two hospitals in Swansea and Bridgend have already cancelled some surgical procedures until after Christmas.","figure_data":"","figure_id":"tab_7","figure_label":"6","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: An illustration of the iterative summarization process. The summarizer continuously refines the summary according to self-feedback from the evaluator at each iteration.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Zhang et al. (2023a) also proposed a benchmark for fine-grained instruction-based editing. Recently, Welleck et al. (2022) introduced a selfcorrective learning framework that incorporates a corrector into the language model to facilitate selfcorrection during sequence generation. Akyürek et al. (\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"•Remove: Remove the information of <insert> from the summary • Rephrase: Rephrase the information of <in-sert> in the summary • Simplify: Shorten the summary • Keep: Keep the summary unchanged\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure3: Human evaluation to justify the refinement behavior of SummIt. The top plot refers to the human justification of the ratio that the summary is improved at each iteration and the bottom plot indicates the ratio that the summarizer follows the evaluator's evaluation rationale.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Detailed statistics of the experimental datasets. Doc # words and Sum # words refer to the average word number in the source document and summary.\",\n \"figure_data\": \"Dataset#TestDoc #wordsSum #words#SumXSum11,334430.223.31CNN/DM 11,489766.158.21NEWTS600738.570.12\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \", ensuring they are con-sistent with previous fine-tuning approaches: 1)\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"This table presents results from experiments conducted on the CNN/DM and XSum datasets under both zero-shot and few-shot settings. A random sample of 1, 000 data points was taken from each dataset for evaluation. G-Eval represents the score evaluated by the ChatGPT evaluator in our framework.\",\n \"figure_data\": \"CNN/DMXSumR1R2RLG-EvalR1R2RLG-EvalZero-shot settingPEGASUS ZS 32.90 13.28 29.383.2319.27 3.00 12.723.52BART ZS32.83 13.30 29.643.4219.26 3.30 14.673.49T5 ZS39.68 17.24 26.283.4719.66 2.91 15.313.55ChatGPT39.44 16.14 29.833.4621.61 5.98 17.603.47SummIt (ours) 36.50 13.49 26.764.3321.92 5.93 17.624.24Few-shot settingChatGPT40.00 16.39 30.023.5723.96 7.36 19.363.57SummIt (ours) 37.29 13.60 26.874.3522.04 6.20 17.464.32ModelCoherence Fluency Relevance Consistency Conciseness Overall Human PrefCNN/DMBART3.924.164.003.123.643.240.04T53.724.244.323.523.843.680.10PEGASUS3.203.533.332.871.851.630.00ChatGPT4.204.364.284.013.924.010.34SummIt4.244.504.294.123.844.090.52XSumBART3.974.304.133.303.933.840.30T53.844.324.023.633.843.250.08PEGASUS3.134.103.522.872.032.410.00ChatGPT4.034.404.303.933.873.920.24SummIt4.044.354.284.053.723.960.38\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Human study results on generic summary quality. The first five columns include Likert scale ratings and the last column is the human preference results.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Experimental results of incorporating knowledge extractor on summary quality and faithfulness on XSum dataset. -IE refers to the model integrated with OpenIE.\",\n \"figure_data\": \"R1R2RLG-Eval FactCC DAEChatGPT21.61 5.98 17.603.4728.0010.34SummIt21.92 5.93 17.624.2436.0033.02ChatGPT-IE 22.01 5.11 17.063.8551.6893.68SummIt-IE19.72 3.85 15.364.9547.2490.36R1R2RLG-Eval BM25 DPRChatGPT30.018.9427.031.0633.0977.22ChatGPT-Topic 33.24 10.20 29.881.1636.2078.77SummIt-Topic30.458.4827.194.7439.1182.41\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Experimental results on NEWTS dataset to test the controllability of our framework. -Topic indicates a model that is prompted to extract topic-related snippets before generating a summary.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"highlights the potential issue of LLM-based evaluators having a bias towards the LLM outputs, which raises the doubt:\",\n \"figure_data\": \"R1R2RLG-EvalSummIt36.50 13.49 26.764.33-w/o Add33.01 11.55 24.713.98-w/o Remove36.46 13.44 26.553.64-w/o Rephrase 34.71 12.12 26.313.82-w/o Simplify33.49 12.33 25.763.55-w/o Keep33.87 13.03 25.703.94\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablation Study on Iterative Refinement OperationsSummary-Iter1: Hospitals in Wales may have to choose between emergency care and non-urgent surgery during peak winter months, according to Dr. Andrew Goodall. He suggested that hospitals may decide not to schedule surgery in order to focus on \\\"front door pressures.\\\" Two hospitals in Swansea and Bridgend have already cancelled some surgical procedures until after Christmas.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(Cheng and Lapata, 2016)\", \"Explanation\": \"The cited work by Cheng and Lapata (2016) provides a methodological basis for the development of end-to-end summarization systems by introducing the use of neural networks and pre-trained language models in summarization systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lewis et al., 2019)\", \"Explanation\": \"The cited work by Lewis et al. (2019) provides a methodological basis for the development of end-to-end summarization systems by introducing the use of pre-trained language models in summarization systems.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhang et al., 2020)\", \"Explanation\": \"The cited work by Zhang et al. (2020) extends the research on end-to-end summarization systems by exploring the use of pre-trained language models in the development of such systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Goyal et al., 2022)\", \"Explanation\": \"The cited work by Goyal et al. compares GPT-3 with traditional fine-tuning methods, providing a methodological basis for the citing paper to assess the performance of large language models in summarization tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2023d)\", \"Explanation\": \"The cited work by Zhang et al. focuses on large language models for news summarization, providing a methodological basis for the citing paper to analyze the quality of generated summaries in this context.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Liu et al., n.d.)\", \"Explanation\": \"The cited work by Liu et al. is an extension of the research on large language models for news summarization, providing further insights into the performance of these models in this area.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xu et al., 2019)\", \"Explanation\": \"The cited work by Xu et al. provides a method for extractive summarization that the citing paper builds upon in their research on text summarization systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu and Lapata, 2019)\", \"Explanation\": \"The cited work by Liu and Lapata provides a method for extractive summarization that the citing paper adopts in their research on text summarization systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhong et al., 2020)\", \"Explanation\": \"The cited work by Zhong et al. provides a method for extractive summarization that the citing paper uses in their research on text summarization systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2022a)\", \"Explanation\": \"The cited work by Zhang et al. provides a method for extractive summarization that the citing paper utilizes in their research on text summarization systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2023b)\", \"Explanation\": \"The cited work by Zhang et al. provides a method for extractive summarization that the citing paper builds upon in their research on text summarization systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lewis et al., 2019)\", \"Explanation\": \"The cited work by Lewis et al. provides a method for abstractive summarization that the citing paper adopts in their research on text summarization systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2020)\", \"Explanation\": \"The cited work by Zhang et al. provides a method for abstractive summarization that the citing paper uses in their research on text summarization systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kryscinski et al., 2020)\", \"Explanation\": \"The cited work by Kryscinski et al. provides a method for addressing the limitations of one-shot summary generation, which the citing paper builds upon in their research on text summarization systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The cited work by Liu et al. provides a method for automatically correcting factual inconsistencies in generated summaries, which the citing paper adopts in their research on text summarization systems.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Goyal et al., 2022)\", \"Explanation\": \"The cited work by Goyal et al. (2022) provides a comparison of the performance of LLMs in text summarization, which serves as a methodological basis for the citing paper to build upon in their own research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2023d)\", \"Explanation\": \"The cited work by Zhang et al. (2023d) reports on the performance of LLMs in text summarization, which the citing paper uses to benchmark the performance of their own research in the same domain.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2023c)\", \"Explanation\": \"The cited work by Zhang et al. (2023c) proposes an extract-then-generate pipeline for improving summary faithfulness in text summarization, which the citing paper builds upon in their own research to improve the performance of their LLM-generated summaries.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Lin, 2004)\", \"Explanation\": \"The cited work by Lin (2004) provides the basis for the evaluation metrics used in the citing paper, which are ROUGE and BertScore. The cited work establishes the standard for evaluation metrics in the field of summarization.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Liu et al., 2023)\", \"Explanation\": \"The cited work by Liu et al. (2023) extends the research on LLM-based metrics for summarization evaluation by demonstrating the superiority of these metrics over previous evaluation methods like ROUGE and BertScore in terms of correlation with human evaluations.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Fu et al., 2023)\", \"Explanation\": \"The cited work by Fu et al. (2023) further extends the research on LLM-based metrics for summarization evaluation by showing the performance of these metrics in terms of correlation with human evaluations.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Luo et al., 2023)\", \"Explanation\": \"The cited work by Luo et al. (2023) also builds upon the research on LLM-based metrics for summarization evaluation by demonstrating the effectiveness of these metrics in terms of correlation with human evaluations.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Schick et al., 2022)\", \"Explanation\": \"The cited work is an extension of the task of text editing, specifically focusing on interactive editing systems for text generation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Iso et al., 2020)\", \"Explanation\": \"The cited work provides a method for information updating in text editing models, which the citing paper may adopt or adapt in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Awasthi et al., 2019)\", \"Explanation\": \"The cited work is a data source for grammar error correction in text editing models, which the citing paper may utilize in their research or analysis.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Faltings et al., 2020)\", \"Explanation\": \"The cited work presents a method for command-based editing systems in text generation, which the citing paper may build upon in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Madaan et al., 2023)\", \"Explanation\": \"The cited work presents a similar approach to the generation pipeline in the citing paper, with a focus on short text generation tasks and iterative feedback for error correction.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. (2020) introduces the concept of in-context learning (ICL), which the citing paper adopts to guide the explanation and summary generation process in their iterative summarization system.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kryscinski et al., 2020)\", \"Explanation\": \"The cited work highlights the importance of faithfulness in generated summaries, which serves as a methodological basis for the citing paper to focus on improving the quality of generated summaries.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Huang et al., 2020)\", \"Explanation\": \"The cited work demonstrates the effectiveness of leveraging knowledge extraction in enhancing the faithfulness of generated summaries, which the citing paper builds upon to propose integrating a knowledge extractor into the iterative summarization system.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhu et al., 2020)\", \"Explanation\": \"The cited work also highlights the effectiveness of knowledge extraction in improving the faithfulness of generated summaries, which the citing paper extends by proposing to integrate a knowledge extractor into the iterative summarization system.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(OpenIE3)\", \"Explanation\": \"The cited work is a knowledge extractor that the citing paper utilizes to extract knowledge from source documents, serving as a data source for the proposed iterative summarization system.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Luo et al., 2023)\", \"Explanation\": \"The cited work demonstrates the efficiency of LLMs in evaluating faithfulness, which the citing paper leverages as a methodological basis for directing the evaluator to factor in faithfulness in delivering feedback in the iterative summarization system.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Reid and Neubig, 2022)\", \"Explanation\": \"The cited work provides a set of text editing operations that the citing paper adopts in the system implementation to improve the performance of the system.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Faltings et al., 2020)\", \"Explanation\": \"The cited work also provides a set of text editing operations that the citing paper uses in the system implementation to improve the performance of the system.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Lin, 2004)\", \"Explanation\": \"The cited work by Lin (2004) provides the definition of ROUGE scores, which the citing paper uses as a measure of summary quality in their evaluation metrics.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Liu et al., 2023)\", \"Explanation\": \"The cited work by Liu et al. (2023) introduces the G-Eval metric, which the citing paper uses in their evaluation of summary quality.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Kryscinski et al., 2020)\", \"Explanation\": \"The cited work by Kryscinski et al. (2020) presents the FactCC model metric, which the citing paper uses to evaluate summary faithfulness in their evaluation metrics.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Goyal and Durrett, 2020)\", \"Explanation\": \"The cited work by Goyal and Durrett (2020) introduces the DAE metric, which the citing paper uses in their evaluation of summary faithfulness.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Robertson et al., 2009)\", \"Explanation\": \"The cited work by Robertson et al. (2009) provides a probabilistic retrieval function that is used in the citing paper to measure the similarity between the query and the summary in query-focused summarization.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Karpukhin et al., 2020)\", \"Explanation\": \"The cited work by Karpukhin et al. (2020) introduces the dense vector representation of questions and passages for measuring similarity in the citing paper, which is used to assess the controllability of query-focused summarization.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Goyal et al., 2022)\", \"Explanation\": \"The cited work by Goyal et al. (2022) highlights the human annotators' preference for summaries generated by LLM models, which the citing paper further validates through a dedicated human study to assess the effectiveness of SummIt.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2020)\", \"Explanation\": \"The cited work, PEGASUS, is used as a baseline model in the citing paper to compare the performance of SummIt in a zero-shot setting for generic summarization quality evaluation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lewis et al., 2019)\", \"Explanation\": \"The cited work, BART, is used as a baseline model in the citing paper to compare the performance of SummIt in a zero-shot setting for generic summarization quality evaluation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Raffel et al., 2020)\", \"Explanation\": \"The cited work, T5, is used as a baseline model in the citing paper to compare the performance of SummIt in a zero-shot setting for generic summarization quality evaluation.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2023d)\", \"Explanation\": \"The cited work provides the results of a previous study on the low quality of reference summaries in large language model summary outputs, which is used as a reference in the citing paper to support the analysis of the performance of SummIt in the zero-shot setting.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Goyal et al., 2022)\", \"Explanation\": \"The human preference study conducted in the cited work provides evidence that large language models (LLMs) generate high-quality summaries that are preferred by human annotators.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2023)\", \"Explanation\": \"The cited work provides a method for evaluating summary quality in the context of iterative summary refinement, which the citing paper adopts in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Goyal et al., 2022)\", \"Explanation\": \"The cited work by Goyal et al. provides the dataset used for the experiments conducted in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2023d)\", \"Explanation\": \"The cited work by Zhang et al. is a research effort that has also limited the testing of GPT-3 on a small number of instances, which the citing paper also follows in their testing.\"}]"}}},{"rowIdx":79,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b25","b4","b1","b10","b15","b2","b3","b20","b14","b36","b0","b37"],"table_ref":[],"text":"Autonomous driving is a rapidly developing field with immense potential to improve transportation safety and efficiency with advancements in sensor technologies and computer vision. As the increasing maturity of traditional perceptual techniques such as 3D object detection (Liu et al. 2023;Jiao et al. 2023b) and tracking (Chen et al. 2023), autonomous driving systems are progressing towards enhanced interpretability and flexible human-car interactivity. In this context, visual question answering (VQA) (Antol et al. 2015) can play a critical role. On one hand, VQA possesses interactive and entertainment, enabling passengers to perceive their surroundings through language and enhancing the user experience of intelligent driving systems. On the other hand, users can verify the correctness of perception Figure 1: NuScenes-QA is a multi-modal, multi-frame, outdoor dataset that differs significantly from other VQA benchmarks in terms of visual data. system through question answering, fortifying their trust in its capabilities.\nDespite the notable progress made by the VQA community, models trained on existing VQA datasets (Goyal et al. 2017;Hudson and Manning 2019) have limitations in addressing the complexities of autonomous driving scenario. This limitation is primarily caused by the difference in visual data between self-driving scenario and existing VQA benchmarks. For instance, to answer question like \"Are there any moving pedestrians in front of the stopped bus?\", it is necessary to locate and identify the bus, pedestrians, and their status accurately. This requires the model to effectively leverage the complementary information from images and point clouds to understand complex scenes and capture object dynamics from multiple frames of data streams. Therefore, it is essential to explore VQA in the context of multi-modal, multi-frame and outdoor scenario. However, existing VQA benchmarks cannot satisfy all these conditions simultaneously, as illustrated in Fig. 1. For instance, although 3D-QA (Azuma et al. 2022) and the self-driving scenario both focus on understanding the structure and spatial relationships of objects, 3D-QA is limited to single-modal (i.e., point cloud), arXiv:2305.14836v2 [cs.CV] 20 Feb 2024 single-frame, and static indoor scenes. The same goes for other benchmarks. To bridge this gap, we construct the first VQA benchmark specifically designed for autonomous driving scenario, named NuScenes-QA. NuScenes-QA is different from all other existing VQA benchmarks in terms of visual data characteristics, presenting new challenges for both VQA and autonomous driving community.\nThe proposed NuScenes-QA is built upon nuScenes (Caesar et al. 2020), which is a popular 3D perception dataset for autonomous driving. We automatically annotate the question-answer pairs using the CLEVR benchmark (Johnson et al. 2017) as inspiration. To be specific, we consider each keyframe annotated in nuScenes as a \"scene\" and construct a related scene graph. The objects and their attributes are regarded as the nodes in the graph, while the relative spatial relationships between objects are regarded as the edges, which are calculated based on the 3D bounding boxes annotated in nuScenes. Additionally, we design different types of question templates manually, including counting, comparison, and existence, etc. Based on these constructed templates and scene graphs, we sample different parameters to instantiate the templates, and use the scene graph to infer the correct answers, thus automatically generating question-answer pairs. Eventually, we obtained a total of 460K questionanswer pairs on 34K scenes from the annotated nuScenes training and validation split, with 377K pairs for training and 83K for testing.\nIn addition to the dataset, we also develop baseline models using the existing 3D perception (Huang et al. 2021;Yin, Zhou, and Krahenbuhl 2021;Jiao et al. 2023a) and visual question answering (Anderson et al. 2018;Yu et al. 2019) techniques. These models fall into three categories: image-based, point cloud-based, and multi-modal fusionbased. The 3D detection models are used to extract visual features and provide object proposals, which are then combined with question features and fed into the question answering model for answer decoding. While our experiments show that these models outperform the question-only blind model, their performance still significantly lags behind models that use ground truth object labels as inputs. This indicates that combining existing technologies is not sufficient for intricate street views understanding. Thus, NuScenes-QA poses a new challenge, urging further research in this realm.\nOverall, our contributions can be summarized as follows:\n• We introduce a novel visual question answering task in autonomous driving scenario, which evaluates current deep learning based models' ability to understand and reason with complex visual data in multi-modal, multiframe, and outdoor scenes. To facilitate this task, we contribute a large-scale dataset, NuScenes-QA, consisting of 34K complex autonomous driving scenes and 460K question-answer pairs. • We establish several baseline models and extensively evaluate the performance of existing techniques for this task. Additionally, we conduct ablation experiments to analyze specific techniques that are relevant to this task, which provide a foundation for future research."},{"figure_ref":[],"heading":"Related Works","publication_ref":["b10","b20","b15","b16","b23","b26","b0","b22"],"table_ref":[],"text":"Visual Question Answering\nThere are various datasets available for VQA, including image-based datasets such as VQA2.0 (Goyal et al. 2017), CLEVR (Johnson et al. 2017), and GQA (Hudson and Manning 2019), as well as video-based datasets such as TGIF-QA (Jang et al. 2017) and TVQA (Lei et al. 2018). For the image-based VQA, earlier works (Lu et al. 2016;Anderson et al. 2018;Qian et al. 2022a) typically use CNNs to extract image features, and RNNs to process the question. Then, joint embddings of vision and language obtained through concatenation or other operations (Kim, Jun, and Zhang 2018) "},{"figure_ref":[],"heading":"3D Visual Question Answering","publication_ref":["b35","b5","b2"],"table_ref":[],"text":"3D Visual Question Answering (3D-QA) is a novel task in the VQA field that focuses on answering questions about 3D scenes represented by point cloud. Unlike traditional VQA tasks, 3D-QA requires models to understand the geometric structure and the spatial relations of objects in a indoor scene. Recently, many 3D-QA datasets have been constructed. For example, the 3DQA dataset (Ye et al. 2022), which is based on ScanNet (Dai et al. 2017), has manually annotated 6K question-answer pairs. Similarly, ScanQA (Azuma et al. 2022) has utilized a question generation model along with manual editing to annotate 41K pairs on the same visual data. Despite these advancements, current 3D-QA models face limitations in solving more complex autonomous driving scenario, which involve multi-modalities, multi-frames, and outdoor scenes. Step ①"},{"figure_ref":[],"heading":"Vision-Language Tasks in Autonomous Driving","publication_ref":[],"table_ref":[],"text":"Step ②"},{"figure_ref":[],"heading":"Question Template Construction","publication_ref":[],"table_ref":[],"text":"What is the <A2> <O2> to the <R> of the <A> <O>? Are there any <A2> <O2>s to the <R> of the <A> <O>?\nStep "},{"figure_ref":[],"heading":"NuScenes-QA Dataset","publication_ref":[],"table_ref":[],"text":"Our primary contribution is the NuScenes-QA dataset, which we will introduce in detail in this section. We provide a comprehensive overview of the dataset construction, including scene graph development, question template design, question-answer pair generation, and post-processing.\nIn addition, we analyze the statistical characteristics of the NuScenes-QA dataset, such as the distribution of question types, lengths, and answers."},{"figure_ref":["fig_0","fig_0","fig_0"],"heading":"Data Construction","publication_ref":["b20","b21"],"table_ref":[],"text":"For question-answer pairs generation, we adapted an automated method inspired by CLEVR (Johnson et al. 2017). This method requires two types of structured data: scene graphs generated from 3D annotations, containing object category, position, and relationships; alongside manually crafted question templates that specify the question type, expected answer type, and reasoning required to answer it. By combining these structured data, we automatically generate question-answer pairs. These pairs are then filtered and validated through post-processing programs to construct the complete dataset. Fig. 2 illustrates the overall data construction pipeline.\nScene Graph Construction A scene graph (Johnson et al. 2015) is defined as an abstract representation of a visual scene, where nodes in the graph represent objects in the scene, and edges represent relationships between objects.\nIn nuScenes, the collected data is annotated with a frequency of 2Hz, and each annotated frame is referred as a \"keyframe\". We consider each keyframe as a \"scene\" in NuScenes-QA. The existing annotations include object categories and their attributes in the scene, as well as the 3D bounding boxes of the objects. These annotated objects and their attributes are directly used as nodes in the graph. However, relationships between objects are not provided in the original annotations, so we developed a rule for calculating object relationships. Given that spatial position relationships are crucial in autonomous driving scenario, we define six relationships between objects, namely front, back, front left, front right, back left, and back right. To determine object relationships, we first project 3D bounding boxes onto the Bird's-Eye-View (BEV). Subsequently, we calculate the angle between the vector connecting the centers of two bounding boxes and the forward direction of the ego-car. The formula is given by\nθ = cos -1 (B 1 [: 2] -B 2 [: 2]) • V ego [: 2] ∥B 1 [: 2] -B 2 [: 2]∥∥V ego [: 2]∥ ,(1)\nwhere B i = [x, y, z, x size , y size , z size , φ] is the 3D bounding box of object i, and\nV ego = [v x , v y , v z ]\nrepresents the speed of the ego car. Based on the angle range, the relationship between two objects is defined as\nrelation = f ront if -30 • < θ <= 30 • f ront lef t if 30 • < θ <= 90 • f ront right if -90 • < θ <= -30 • back lef t if 90 • < θ <= 150 • back rigth if -150 • < θ <= -90 • back else.\n(2) We define the forward direction of the car as 0 • and counterclockwise angle as positive. At this point, we can convert the annotations of nuScenes into the scene graphs we need, as illustrated in step one of Fig. 2.\nQuestion Template Design We devised templates manually for question generation. For instance, the question \"What is the moving thing to the front left of the stopped bus?\" can be abstracted as the template \"What is the <A2><O2>to the <R>of the <A1><O1>?\", with <A>, <O>, and <R> as parameters for instantiation, representing attribute, object, and relationship, respectively. Additionally, we can express the same semantic with another form like \"There is a <A2><O2>to the <R>of the <A1><O1>, what is it?\". Ultimately, NuScenes-QA holds 66 diverse question templates, divided into 5 question types: existence, counting, query-object, query-status, and comparison. In addition, to better evaluate the models reasoning performance, we also divide the questions into zero-hop and one-hop. Specifically, zero-hop questions require no reasoning between objects, e.g., \"What is the status of the <A><O>?\". One-hop questions involve one step spatial reasoning, e.g., \"What is the status of the <A2><O2>to the <R>of the <A1><O1>?\". Comprehensive template details are available in the supplementary material.\nQ&A Pair Generation and Filtering Given the scene graphs and question templates, instantiating a questionanswer pair is straightforward: we select a template, assign parameter values through depth-first search, and deduce the ground truth answer on the scene graph. Moreover, we dismiss ill-posed or degenerate questions. For instance, the question is ill-posed if the scene do not contain any cars or pedestrians when <O1>==pedestrian and <O2>==car is assigned for the template \"What is the status of the <O2>to the <R>of the <A1><O1>?\".\nIt is important to note that post-processing, as depicted in step 4 of Fig. 2, addresses numerous unsuitable expressions. For example, we added the ego-car as an object in the scene, it is referred to as \"me\" in questions. This led to some inappropriate instances like \"the me\" or \"there is a me\" when <O> is assigned \"me\". We revised such expressions. In addition, during the instantiation, inappropriate <A>+<O> combinations like \"standing cars\" and \"parked pedestrians\" were eliminated through rules. Also, we removed questions with counting answers greater than 10 to balance the answer distribution."},{"figure_ref":[],"heading":"Data Statistics","publication_ref":[],"table_ref":[],"text":"In total, NuScenes-QA provides 459,941 question-answer pairs across 34,149 visual scenes, with 376,604 questions from 28,130 scenes for training, and 83,337 questions from 6,019 scenes for testing. To the best of our knowledge, NuScenes-QA is currently the largest 3D related question answering dataset. Detailed comparison of 3D-QA datasets can be found in supplementary materials. "},{"figure_ref":[],"heading":"Method","publication_ref":[],"table_ref":[],"text":"Along with the proposed dataset, we provide several baselines based on existing 3D detection and VQA techniques."},{"figure_ref":[],"heading":"Task Definition","publication_ref":[],"table_ref":[],"text":"Given a visual scene S, and a question Q, the task of visual question answering aims to select an answer â from the answer space A = {a i } N i=1 that best answers the question. Therefore, the task can be formulated as:\nâ = arg max a∈A P (a | S, Q).\n(3)\nFor NuScenes-QA, visual scene data encompass multi-view images I, point clouds P , and any frames I i and P i before the current frame in the data sequences. We can further decompose the Eq. 3 into:\nP (a | S, Q) = P (a | I, P, Q) I = {I i , T -t < i ≤ T } (4) P = {P i , T -t < i ≤ T },\nwhere T is the index of current frame and t is the number of previous frame used in the model. It is also possible to use only single modality or single frame data for prediction."},{"figure_ref":["fig_3"],"heading":"Framework Overview","publication_ref":[],"table_ref":[],"text":"The overall framework of our proposed baseline is illustrated in Fig. 4 and mainly consists of three components. The first is the feature extraction backbone, which includes both image and point cloud feature extractor. The second part is the region proposal module for object embedding, and the last component is the QA-head for answer prediction.\nInitially, the surrounded-view images and point clouds are fed into the feature extraction backbone, with features projected onto the Bird's-Eye-View (BEV). Subsequently, 3D bounding boxes inferred by a pre-trained detection model are used to crop and pool object features. Finally, the QAmodel takes the question features and the object features as input for cross-modal interaction to predict the answer."},{"figure_ref":["fig_3"],"heading":"Input Embedding","publication_ref":["b13","b12","b24","b29","b41","b11","b32"],"table_ref":[],"text":"Question Embedding For a question Q = {w i } nq i=1 that contains n q words, we first tokenize it and initialize the tokens with pre-trained GloVe (Pennington, Socher, and Manning 2014) embeddings. The sequence is then fed into a single-layer biLSTM (Hochreiter and Schmidhuber 1997) for word-level context encoding. Each word feature w i is represented by the concatenation of the forward and backward hidden states of the biLSTM, denoted as:\nw i = [ → h i ; ← h i ] ∈ R d ,(5)\nand the question embedding is represented as Q ∈ R nq×d .\nVisual Feature Extraction We adopt leading-edge 3D detection techniques for visual feature extraction. As shown in Fig. 4, it entails two branches: image stream and point cloud stream. For multi-view images, ResNet (He et al. 2016) with FPN (Lin et al. 2017) is used as the backbone for multi-scale feature extraction. Then, in order to make the feature spatialaware, we estimate the depth of the pixels in the images and lift them to 3D virtual points with a view transformer inspired by LSS (Philion and Fidler 2020). Finally, pooling along the Z-axis compresses the feature in voxel space, producing the BEV featmap M I ∈ R H×W ×dm .\nFor point clouds, we first partition 3D space into voxels, transforming raw point clouds into binary voxel grids (Zhou and Tuzel 2018). Subsequently, 3D sparse convolutional neural network (Graham, Engelcke, and Van Der Maaten 2018) is applied to the voxel grid for feature representation. Similar to the image features mentioned earlier, Z-axis pooling yields the point cloud BEV featmap M P ∈ R H×W ×dm . Combining M I and M P , we can aggregate them to obtain multi-modal featmap M ∈ R H×W ×dm .\nObject Embedding Following 2D detection works (Ren et al. 2015), we crop and pool the features in bounding boxes as the object embedding. However, unlike standard 2D bounding boxes aligned with the coordinate axis in images, projecting 3D boxes to BEV yields rotated boxes unsuited for standard RoI Pooling. To this end, we make some modifications. Firstly, we project the 3D box B = [x, y, z, x size , y size , z size , φ] into the BEV featmap:\nx m = x -R pc F v × F o ,(6)\nwhere,F v , F o and R pc represent the voxel factor, out size factor of the backbone, and the point cloud range, respectively. All box parameters follow the Eq. 6 to transform into BEV space except the heading angle φ. Then, based on the center and size of the box, we can easily calculate the four vertices V = {x i , y i } 3 i=0 . Secondly, we calculate the rotated vertex V ′ via the heading angle φ:\nx ′ i y ′ i = cos φ -sin φ sin φ cos φ x i y i(7)\nFinally, we use cross product algorithm to identify pixel membership in the rotated rectangle. Then, we perform mean pooling on the features of all the pixels inside the rectangle to obtain the object embedding O ∈ R N ×dm . Algorithm details can be found in supplementary materials."},{"figure_ref":[],"heading":"Answer Head and Training","publication_ref":["b37"],"table_ref":[],"text":"We adopt the classical VQA model MCAN (Yu et al. 2019) as our answer head. It leverages stacked self-attention layers to model the language and visual context independently, along with stacked cross-attention layers for cross-modal feature interaction. The fused features are then projected to the answer space for prediction via basic MLP layers.\nDuring the training phase, we extract the object embeddings using a pre-trained 3D detection model offline. And answer head is trained with the standard cross-entropy loss. "},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":"To validate the challenge of NuScenes-QA, we assess baseline performance in various configurations: camera-only or LiDAR-only single-modality models, camera-lidar fusion models, and diverse answering heads. We conduct ablation studies on crucial steps of the baseline, including BEV feature cropping and pooling strategies, as well as the influence of detected 3D bounding boxes. Furthermore, visualization samples are showcased in supplementary material."},{"figure_ref":[],"heading":"Evaluation Metrics","publication_ref":["b1","b2"],"table_ref":[],"text":"Questions in NuScenes-QA span 5 categories based on query format: 1) Exist, querying the existence of a object in the scene; 2) Count, object counting under specified conditions; 3) Object, object recognition based on language description; 4) Status, querying the status of a specified object; 5) Comparison, specified objects or their status comparison. Additionally, questions are also divided into two groups based their complexity of reasoning: zero-hop (denoted as H0) and one-hop (denoted as H1). We adopt the Top-1 accuracy as our evaluation metric, follow the practice of many other VQA works (Antol et al. 2015;Azuma et al. 2022), and evaluate the performance of different question types separately."},{"figure_ref":[],"heading":"Implementation Details","publication_ref":["b14","b36"],"table_ref":[],"text":"For the feature extraction backbone, we use the pre-trained detection model following the original settings (Huang et al. 2021;Yin, Zhou, and Krahenbuhl 2021;Jiao et al. 2023a). The dimension of the QA model d m is set to 512, and MCAN adopts a 6-layer encoder-decoder version. As for training, we used the Adam optimizer with an initial learning rate of 1e-4 and half decaying every 2 epochs. All experiments are conducted with a batch size of 256 on 2 NVIDIA GeForce RTX 3090 GPUs. More details can be found in supplementary material."},{"figure_ref":[],"heading":"Quantitative Results","publication_ref":["b14","b0","b37"],"table_ref":["tab_3"],"text":"Compared Methods As mentioned earlier, our task can be divided into three settings: camera-only, LiDAR-only, camera+LiDAR. To explore the impact of different modalities on the question-answering performance, we select representative backbone for each setting. We choose BEVDet (Huang et al. 2021) for camera-only setting, which proposed a novel paradigm of explicitly encoding the perspectiveview features into the BEV space. CenterPoint (Yin, Zhou, and Krahenbuhl 2021) is selected for LiDAR-only setting. It introduced a center-based object keypoint detector and has shown excellent performance in both detection accuracy and speed. For the multi-modal model, we opt for MSMD-Fusion (Jiao et al. 2023a), which leverages depth and finegrained LiDAR-camera interaction, achieving state-of-theart results on the nuScenes detection benchmark for single model.\nRegarding the QA-head, we select two classic models, BUTD (Anderson et al. 2018) and MCAN (Yu et al. 2019). BUTD advocates for computing bottom-up and top-down attention on salient regions of the image. MCAN stacks selfattention and cross-attention modules for vision-language feature interaction. To establish the upper bound of the QA models, we employ perfect perceptual results, i.e., groundtruth object labels. Specifically, we use GloVe for objects and their status embedding, noted as GroundTruth in Table 1. Additionally, we design a Q-Only baseline to investigate the impact of language bias. Q-Only can be considered as a blind model that ignores the visual input."},{"figure_ref":[],"heading":"Results and Discussions","publication_ref":["b39"],"table_ref":["tab_3","tab_3","tab_3","tab_3"],"text":"According to the results shown in Table 1, we have the following observations that are worth discussing.\n1. It is evident that visual data play a critical role in the performance of our task. When comparing the Q-Only baseline to others, we find that it only achieves an accuracy of 53.4%, which is significantly lower than that of other models. For instance, MSMDFusion+MCAN performs 7% better. This indicates that model cannot achieve good performance solely rely on language shortcuts, but needs to leverage rich visual information.\n2. Referring to the bottom part of Table 1, we can see that the LiDAR-based CenterPoint outperforms the camerabased BEVDet, achieving accuracy of 57.9% and 59.5%, respectively. We attribute this performance gap to the task characteristics. Images possess detailed texture information, point clouds excel in structure and spatial representation. Our proposed NuScenes-QA emphasizes more on the understanding of structure and spatial relationships of objects. On the other hand, the fusion-based model MSMDFusion attains the best performance with an accuracy of 60.4%, demonstrating the camera and LiDAR data are complementary. Further work can explore how to better exploit the complementary information of multi-modal data. Of course, our baselines still have a long way to go compared to the GroundTruth (achieving an accuracy of 84.3%).\n3. According to Table 1, QA-head has a significant impact on the performance. With the same detection backbone, we observed that the QA-head based on MCAN outperforms BUTD by a large margin. For example, the overall accuracy of CenterPoint+MCAN is 59.5%, 1.4% higher than CenterPoint+BUTD. A dedicated QA-head designed for NuScenes-QA may lead to a greater improvement. We leave this as future work.\n4. In a horizontal comparison of Table 1, it is not difficult to find that counting is the most difficult among all question types. Our best baseline model achieved just 23.2% accuracy, much lower than other question types. Counting is historically tough in visual question answering, and some explorations (Zhang, Hare, and Prügel-Bennett 2018) have been made in traditional 2D-QA. Future efforts could involve counting modules to enhance its performance."},{"figure_ref":[],"heading":"Ablation Studies","publication_ref":[],"table_ref":["tab_4","tab_5","tab_6"],"text":"To validate effectiveness of different operations in our baselines, we conduct extensive ablation experiments on the NuScenes-QA test split using the CenterPoint+MCAN baseline combination.\nEffects of Bounding Boxes Most 2D and 3D VQA models fuse the visual feature with object bounding box in the object embedding stage, making it position-aware. We follow this paradigm and evaluate the impact of 3D bounding boxes in our NuScenes-QA. Specifically, we project the 7-dimensional box B = [x, y, z, x size , y size , z size , φ] obtained from the detection model onto the same dimension as the object embeddings using MLP, and concatenate the two features as the final input for the QA head. As shown in Table 2, we are surprised to find that the performance varies significantly on different data. Adding box features for ground truth can increase the model's accuracy from 70.8% to 84.3%, a significant improvement of 13.5%. However, adding the detected boxes slightly decreased performance by 0.6%, which is counterproductive. We speculate that this phenomenon may be caused by two reasons. On one hand, the current 3D detection models are still immature, and the noise in the detected boxes hurts the QA model. adding box features is not significant.\nBEV Feature Crop Strategy As mentioned earlier, due to the non-parallelism between the 3D boxes and the BEV coordinate axes, we cannot perform standard RoI pooling as in traditional 2D images. Therefore, we use cross product algorithm to determine pixels inside the rotated box for feature cropping. In addition to this method, we can also use a simpler approach, which directly uses the circumscribed rectangle of the rotated box parallel to the coordinate axes as the cropping region. Table 3 shows the performance comparison of these two crop strategy, where the circumscribed box is slightly inferior to the rotated box. The reason for this is that NuScenes-QA contains many elongated objects, such as bus and truck. These objects occupy a small area in the BEV space, but their circumscribed rectangles have a large range, making the object features over smoothing.\nBEV Feature Pooling Strategy In terms of the feature pooling strategy for the cropped regions, we compared the classic Max Pooling and Mean Pooling operations. As illustrated in Table 4, Max Pooling achieved an accuracy of 58.9% under the same conditions, which is 0.6% lower than Mean Pooling. We speculate that this difference may be due to the fact that Max Pooling focuses on the texture features within the region, while Mean Pooling preserves the overall features. Our proposed NuScenes-QA mainly tests the model's ability of understanding the structure of objects and their spatial relationships in street views, and relatively ignores the texture of the objects. Thus, Mean Pooling has a slight advantage over Max Pooling."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"In this paper, we apply VQA to the context of autonomous driving. We construct NuScenes-QA, the first large-scale multi-modal VQA benchmark for autonomous driving scenario. NuScenes-QA are generated automatically based on visual scene graphs and question templates, containing 34K scenes and 460K question-answer pairs. Alongside a series of baseline models, comprehensive experiments establish a solid foundation for future research. We strongly hope that NuScenes-QA can invigorate the evolution of multi-modal VQA and propel advancements in autonomous driving. "},{"figure_ref":[],"heading":"Question Templates","publication_ref":[],"table_ref":[],"text":"As described in the main paper, we employ manually designed question templates to generate questions programmatically. For instance, the question \"Are there any bicycles to the back left of the moving car?\" has a template of \"Are there any <A2> <O2>s to the <R> of the <A> <O>?\", where <A>, <O>, and <R> denote status, object, and relation, respectively. Table 6 enumerates all the templates based on different question types. Among them, one-hop questions involve reasoning about relations between objects, while zero-hop questions are relatively simpler. We design a total of 5 question types and 16 different semantic question templates. To increase question diversity, we created different variations of each template, such as \"How many <A2> <O2>s are to the <R> of the <A> <O>?\" can also be expressed as \"There is a <A> <O>; how many <A2> <O2>s are to the <R> of it?\" In the end, we get a total of 66 different question templates. In future work, we can further enrich the question templates to enhance the diversity of questions."},{"figure_ref":["fig_4","fig_1"],"heading":"Question Distribution of the First Four Words","publication_ref":[],"table_ref":[],"text":"In Figure 5, we visualize the distribution of the first four words in the questions, from which we can draw two observations. First, our data distribution is balanced, as already verified in Figure 3. Second, our questions encompass a diverse range of visual semantics. To answer them, not only the object categories such as pedestrian, motorcycle are demanded, but also their status, such as moving or parked. The semantic complexities in our questions also presents a considerable challenge for models."},{"figure_ref":[],"heading":"Visualization Examples","publication_ref":[],"table_ref":[],"text":"In order to get a deeper insights of the difficulty of the NuScenes-QA dataset and to validate the performance of proposed baseline models, we select some samples from the test split for visualization, as shown in Figure 6 and Figure 7. We present the point clouds, surround-view images, questions, ground truth answers, and the predictions of 6 different baseline models.\nThe visual data of these examples (with LiDAR point clouds on the left, surround-view images in the middle) provide compelling evidence of the challenges inherent in NuScenes-QA. First, the point clouds and images are in different coordinate systems, which poses great difficulties for the model to comprehend spatial relationships and fuse multi-modal information. Second, the street-view scenes often contain a lot of background noise, which can significantly interfere with the model's ability to locate and recognize foreground objects relevant to the question. Thirdly, some scenes contain dense and diverse objects, making the reasoning for questions in these complex scenes prone to errors.\nFigure 6 shows some successful cases, demonstrating the impressive performance of the baseline models. Inevitably, there are also failed cases, as shown in Figure 7. By comparing the prediction results of different baselines, we find that multi-modal fusion models usually outperform single-modal models, which confirms the complementarity between point clouds and images."},{"figure_ref":[],"heading":"Discussion","publication_ref":[],"table_ref":[],"text":"Limitations As the first work in this field, there are still many shortcomings in NuScenes-QA that need to be im-"},{"figure_ref":[],"heading":"Acknowledgments","publication_ref":[],"table_ref":[],"text":"This work was supported in part by National Natural Science Foundation of China Project (No. 62072116) and Shanghai Science and Technology Program [Project No. 21JC1400600]."},{"figure_ref":[],"heading":"Dataset Visual Modality","publication_ref":["b6","b35","b2","b34","b40","b27"],"table_ref":[],"text":"Multi Frame Scenario Collection # Scenes #Amount EQA (Das et al. 2018) image indoor template 767 1.5k 3DQA (Ye et al. 2022) point cloud indoor human 806 10k ScanQA (Azuma et al. 2022) point cloud indoor auto + human 800 41k CLEVR3D (Yan et al. 2021) point cloud indoor template 1,129 60.1k FE-3DGQA (Zhao et al. 2022) point cloud indoor human 800 20k SQA3D (Ma et al. 2023) image Are there any other <A2><O2>s that in the same status as the <A><O>? Is there another <A2><O2>that has the same status as the <A><O>? Are there any other <A2><O2>s of the same status as the <A><O>? Is there another <A2><O2>of the same status as the <A><O>?\n1 Are there any <A2><O2>s to the <R>of the <A><O>? There is a <A><O>; are there any <A2><O2>s to the <R>of it? Are there any other <A3><O3>s that in the same status as the <A2><O2>[that is] to the <R>of the <A><O>? Is there another <A3><O3>that has the same status as the <A2><O2>[that is] to the <R>of the <A><O>? Are there any other <A3><O3>s of the same status as the <A2><O2>[that is] to the <R>of the <A><O>? Is there another <A3><O3>of the same status as the <A2><O2>[that is] to the <R>of the <A><O>? Is the status of the <A><O>the same as the <A2><O2>? Do the <A><O>and the <A2><O2>have the same status? Is the <A><O>the same status as the <A2><O2>? Does the <A><O>have the same status as the <A2><O2>?\n1\nIs the status of the <A2><O2>[that is] to the <R>of the <A><O>the same as the <A3><O3>? Does the <A2><O2>[that is] to the <R>of the <A><O>have the same status as the <A3><O3>? Do the <A2><O2>[that is] to the <R>of the <A><O>and the <A3><O3>have the same status? There is a <A2><O2>[that is] to the <R>of the <A><O>; does it have the same status as the <A3><O3>?\nThere is a <A2><O2>[that is] to the <R>of the <A><O>; is it the same status as the <A3><O3>?\nThere is a <A2><O2>[that is] to the <R>of the <A><O>; is its status the same as the <A3><O3>?\nIs the status of the <A><O>the same as the <A3><O3>[that is] to the <R>of the <A2><O2>? Does the <A><O>have the same status as the <A3><O3>[that is] to the <R>of the <A2><O2>? Do the <A><O>and the <A3><O3>[that is] to the <R>of the <A2><O2>have the same status? There is a <A><O>; does it have the same status as the <A3><O3>[that is] to the <R>of the <A2><O2>? There is a <A><O>; is its status the same as the <A3><O3>[that is] to the <R>of the <A2><O2>? There is a <A><O>; is it the same status as the <A3><O3>[that is] to the <R>of the <A2><O2>? Is the status of the <A2><O2>[that is] to the <R>of the <A><O>the same as the <A4><O4>[that is] to the <R2>of the <A3><O3>? Does the <A2><O2>[that is] to the <R>of the <A><O>have the same status as the <A4><O4>[that is] to the <R2>of the <A3><O3>? Do the <A2><O2>[that is] to the <R>of the <A><O>and the <A4><O4>[that is] to the <R2>of the <A3><O3>have the same status? There is a <A2><O2>[that is] to the <R>of the <A><O>; does it have the same status as the <A4><O4>[that is] to the <R2>of the <A3><O3>? There is a <A2><O2>[that is] to the <R>of the <A><O>; is its status the same as the <A4><O4>[that is] to the <R2>of the <A3><O3>? There is a <A2><O2>[that is] to the <R>of the <A><O>; is it the same status as the <A4><O4>[that is] to the <R2>of the <A3><O3>?\nTable 6: Question templates.\nproved. First, our questions focus on foreground objects and their relationships, but ignore the connection between objects and the background. For example, we can ask the agent which parking spots ahead are available, which is also important for autonomous driving. Second, our questions focus on the model's understanding of the visual scene, which may be too naive for humans. Meanwhile, traditional image or video based question answering is gradually moving towards higher levels of reasoning such as causal inference. Third, we define the relationships between objects using spatial positioning only, lacking many rich semantic relationships. Finally, although NuScenes-QA is large scale, our questions may not be sufficiently diverse in language due to templates based generation. In summary, there is still significant room for improvement in NuScenes-QA."},{"figure_ref":[],"heading":"Future Work","publication_ref":["b2"],"table_ref":[],"text":"In future work, we can gradually improve NuScenes-QA in the following aspects. Firstly, we can add object localization tasks, like ScanQA (Azuma et al. 2022), to increase the model's interpretability. Secondly, we can utilize crowd sourcing to manually annotate the scene graphs, including new relationships, object status, and more. This can enrich the semantics of questions and increase the diversity of language. Lastly, we can consider integrating with perception tasks such as tracking to increase its practical value. "}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b25\",\n \"b4\",\n \"b1\",\n \"b10\",\n \"b15\",\n \"b2\",\n \"b3\",\n \"b20\",\n \"b14\",\n \"b36\",\n \"b0\",\n \"b37\"\n ],\n \"table_ref\": [],\n \"text\": \"Autonomous driving is a rapidly developing field with immense potential to improve transportation safety and efficiency with advancements in sensor technologies and computer vision. As the increasing maturity of traditional perceptual techniques such as 3D object detection (Liu et al. 2023;Jiao et al. 2023b) and tracking (Chen et al. 2023), autonomous driving systems are progressing towards enhanced interpretability and flexible human-car interactivity. In this context, visual question answering (VQA) (Antol et al. 2015) can play a critical role. On one hand, VQA possesses interactive and entertainment, enabling passengers to perceive their surroundings through language and enhancing the user experience of intelligent driving systems. On the other hand, users can verify the correctness of perception Figure 1: NuScenes-QA is a multi-modal, multi-frame, outdoor dataset that differs significantly from other VQA benchmarks in terms of visual data. system through question answering, fortifying their trust in its capabilities.\\nDespite the notable progress made by the VQA community, models trained on existing VQA datasets (Goyal et al. 2017;Hudson and Manning 2019) have limitations in addressing the complexities of autonomous driving scenario. This limitation is primarily caused by the difference in visual data between self-driving scenario and existing VQA benchmarks. For instance, to answer question like \\\"Are there any moving pedestrians in front of the stopped bus?\\\", it is necessary to locate and identify the bus, pedestrians, and their status accurately. This requires the model to effectively leverage the complementary information from images and point clouds to understand complex scenes and capture object dynamics from multiple frames of data streams. Therefore, it is essential to explore VQA in the context of multi-modal, multi-frame and outdoor scenario. However, existing VQA benchmarks cannot satisfy all these conditions simultaneously, as illustrated in Fig. 1. For instance, although 3D-QA (Azuma et al. 2022) and the self-driving scenario both focus on understanding the structure and spatial relationships of objects, 3D-QA is limited to single-modal (i.e., point cloud), arXiv:2305.14836v2 [cs.CV] 20 Feb 2024 single-frame, and static indoor scenes. The same goes for other benchmarks. To bridge this gap, we construct the first VQA benchmark specifically designed for autonomous driving scenario, named NuScenes-QA. NuScenes-QA is different from all other existing VQA benchmarks in terms of visual data characteristics, presenting new challenges for both VQA and autonomous driving community.\\nThe proposed NuScenes-QA is built upon nuScenes (Caesar et al. 2020), which is a popular 3D perception dataset for autonomous driving. We automatically annotate the question-answer pairs using the CLEVR benchmark (Johnson et al. 2017) as inspiration. To be specific, we consider each keyframe annotated in nuScenes as a \\\"scene\\\" and construct a related scene graph. The objects and their attributes are regarded as the nodes in the graph, while the relative spatial relationships between objects are regarded as the edges, which are calculated based on the 3D bounding boxes annotated in nuScenes. Additionally, we design different types of question templates manually, including counting, comparison, and existence, etc. Based on these constructed templates and scene graphs, we sample different parameters to instantiate the templates, and use the scene graph to infer the correct answers, thus automatically generating question-answer pairs. Eventually, we obtained a total of 460K questionanswer pairs on 34K scenes from the annotated nuScenes training and validation split, with 377K pairs for training and 83K for testing.\\nIn addition to the dataset, we also develop baseline models using the existing 3D perception (Huang et al. 2021;Yin, Zhou, and Krahenbuhl 2021;Jiao et al. 2023a) and visual question answering (Anderson et al. 2018;Yu et al. 2019) techniques. These models fall into three categories: image-based, point cloud-based, and multi-modal fusionbased. The 3D detection models are used to extract visual features and provide object proposals, which are then combined with question features and fed into the question answering model for answer decoding. While our experiments show that these models outperform the question-only blind model, their performance still significantly lags behind models that use ground truth object labels as inputs. This indicates that combining existing technologies is not sufficient for intricate street views understanding. Thus, NuScenes-QA poses a new challenge, urging further research in this realm.\\nOverall, our contributions can be summarized as follows:\\n• We introduce a novel visual question answering task in autonomous driving scenario, which evaluates current deep learning based models' ability to understand and reason with complex visual data in multi-modal, multiframe, and outdoor scenes. To facilitate this task, we contribute a large-scale dataset, NuScenes-QA, consisting of 34K complex autonomous driving scenes and 460K question-answer pairs. • We establish several baseline models and extensively evaluate the performance of existing techniques for this task. Additionally, we conduct ablation experiments to analyze specific techniques that are relevant to this task, which provide a foundation for future research.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Works\",\n \"publication_ref\": [\n \"b10\",\n \"b20\",\n \"b15\",\n \"b16\",\n \"b23\",\n \"b26\",\n \"b0\",\n \"b22\"\n ],\n \"table_ref\": [],\n \"text\": \"Visual Question Answering\\nThere are various datasets available for VQA, including image-based datasets such as VQA2.0 (Goyal et al. 2017), CLEVR (Johnson et al. 2017), and GQA (Hudson and Manning 2019), as well as video-based datasets such as TGIF-QA (Jang et al. 2017) and TVQA (Lei et al. 2018). For the image-based VQA, earlier works (Lu et al. 2016;Anderson et al. 2018;Qian et al. 2022a) typically use CNNs to extract image features, and RNNs to process the question. Then, joint embddings of vision and language obtained through concatenation or other operations (Kim, Jun, and Zhang 2018) \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"3D Visual Question Answering\",\n \"publication_ref\": [\n \"b35\",\n \"b5\",\n \"b2\"\n ],\n \"table_ref\": [],\n \"text\": \"3D Visual Question Answering (3D-QA) is a novel task in the VQA field that focuses on answering questions about 3D scenes represented by point cloud. Unlike traditional VQA tasks, 3D-QA requires models to understand the geometric structure and the spatial relations of objects in a indoor scene. Recently, many 3D-QA datasets have been constructed. For example, the 3DQA dataset (Ye et al. 2022), which is based on ScanNet (Dai et al. 2017), has manually annotated 6K question-answer pairs. Similarly, ScanQA (Azuma et al. 2022) has utilized a question generation model along with manual editing to annotate 41K pairs on the same visual data. Despite these advancements, current 3D-QA models face limitations in solving more complex autonomous driving scenario, which involve multi-modalities, multi-frames, and outdoor scenes. Step ①\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Vision-Language Tasks in Autonomous Driving\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Step ②\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Question Template Construction\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"What is the <A2> <O2> to the <R> of the <A> <O>? Are there any <A2> <O2>s to the <R> of the <A> <O>?\\nStep \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"NuScenes-QA Dataset\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Our primary contribution is the NuScenes-QA dataset, which we will introduce in detail in this section. We provide a comprehensive overview of the dataset construction, including scene graph development, question template design, question-answer pair generation, and post-processing.\\nIn addition, we analyze the statistical characteristics of the NuScenes-QA dataset, such as the distribution of question types, lengths, and answers.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_0\",\n \"fig_0\"\n ],\n \"heading\": \"Data Construction\",\n \"publication_ref\": [\n \"b20\",\n \"b21\"\n ],\n \"table_ref\": [],\n \"text\": \"For question-answer pairs generation, we adapted an automated method inspired by CLEVR (Johnson et al. 2017). This method requires two types of structured data: scene graphs generated from 3D annotations, containing object category, position, and relationships; alongside manually crafted question templates that specify the question type, expected answer type, and reasoning required to answer it. By combining these structured data, we automatically generate question-answer pairs. These pairs are then filtered and validated through post-processing programs to construct the complete dataset. Fig. 2 illustrates the overall data construction pipeline.\\nScene Graph Construction A scene graph (Johnson et al. 2015) is defined as an abstract representation of a visual scene, where nodes in the graph represent objects in the scene, and edges represent relationships between objects.\\nIn nuScenes, the collected data is annotated with a frequency of 2Hz, and each annotated frame is referred as a \\\"keyframe\\\". We consider each keyframe as a \\\"scene\\\" in NuScenes-QA. The existing annotations include object categories and their attributes in the scene, as well as the 3D bounding boxes of the objects. These annotated objects and their attributes are directly used as nodes in the graph. However, relationships between objects are not provided in the original annotations, so we developed a rule for calculating object relationships. Given that spatial position relationships are crucial in autonomous driving scenario, we define six relationships between objects, namely front, back, front left, front right, back left, and back right. To determine object relationships, we first project 3D bounding boxes onto the Bird's-Eye-View (BEV). Subsequently, we calculate the angle between the vector connecting the centers of two bounding boxes and the forward direction of the ego-car. The formula is given by\\nθ = cos -1 (B 1 [: 2] -B 2 [: 2]) • V ego [: 2] ∥B 1 [: 2] -B 2 [: 2]∥∥V ego [: 2]∥ ,(1)\\nwhere B i = [x, y, z, x size , y size , z size , φ] is the 3D bounding box of object i, and\\nV ego = [v x , v y , v z ]\\nrepresents the speed of the ego car. Based on the angle range, the relationship between two objects is defined as\\nrelation = f ront if -30 • < θ <= 30 • f ront lef t if 30 • < θ <= 90 • f ront right if -90 • < θ <= -30 • back lef t if 90 • < θ <= 150 • back rigth if -150 • < θ <= -90 • back else.\\n(2) We define the forward direction of the car as 0 • and counterclockwise angle as positive. At this point, we can convert the annotations of nuScenes into the scene graphs we need, as illustrated in step one of Fig. 2.\\nQuestion Template Design We devised templates manually for question generation. For instance, the question \\\"What is the moving thing to the front left of the stopped bus?\\\" can be abstracted as the template \\\"What is the <A2><O2>to the <R>of the <A1><O1>?\\\", with <A>, <O>, and <R> as parameters for instantiation, representing attribute, object, and relationship, respectively. Additionally, we can express the same semantic with another form like \\\"There is a <A2><O2>to the <R>of the <A1><O1>, what is it?\\\". Ultimately, NuScenes-QA holds 66 diverse question templates, divided into 5 question types: existence, counting, query-object, query-status, and comparison. In addition, to better evaluate the models reasoning performance, we also divide the questions into zero-hop and one-hop. Specifically, zero-hop questions require no reasoning between objects, e.g., \\\"What is the status of the <A><O>?\\\". One-hop questions involve one step spatial reasoning, e.g., \\\"What is the status of the <A2><O2>to the <R>of the <A1><O1>?\\\". Comprehensive template details are available in the supplementary material.\\nQ&A Pair Generation and Filtering Given the scene graphs and question templates, instantiating a questionanswer pair is straightforward: we select a template, assign parameter values through depth-first search, and deduce the ground truth answer on the scene graph. Moreover, we dismiss ill-posed or degenerate questions. For instance, the question is ill-posed if the scene do not contain any cars or pedestrians when <O1>==pedestrian and <O2>==car is assigned for the template \\\"What is the status of the <O2>to the <R>of the <A1><O1>?\\\".\\nIt is important to note that post-processing, as depicted in step 4 of Fig. 2, addresses numerous unsuitable expressions. For example, we added the ego-car as an object in the scene, it is referred to as \\\"me\\\" in questions. This led to some inappropriate instances like \\\"the me\\\" or \\\"there is a me\\\" when <O> is assigned \\\"me\\\". We revised such expressions. In addition, during the instantiation, inappropriate <A>+<O> combinations like \\\"standing cars\\\" and \\\"parked pedestrians\\\" were eliminated through rules. Also, we removed questions with counting answers greater than 10 to balance the answer distribution.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Data Statistics\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In total, NuScenes-QA provides 459,941 question-answer pairs across 34,149 visual scenes, with 376,604 questions from 28,130 scenes for training, and 83,337 questions from 6,019 scenes for testing. To the best of our knowledge, NuScenes-QA is currently the largest 3D related question answering dataset. Detailed comparison of 3D-QA datasets can be found in supplementary materials. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Method\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Along with the proposed dataset, we provide several baselines based on existing 3D detection and VQA techniques.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Task Definition\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Given a visual scene S, and a question Q, the task of visual question answering aims to select an answer â from the answer space A = {a i } N i=1 that best answers the question. Therefore, the task can be formulated as:\\nâ = arg max a∈A P (a | S, Q).\\n(3)\\nFor NuScenes-QA, visual scene data encompass multi-view images I, point clouds P , and any frames I i and P i before the current frame in the data sequences. We can further decompose the Eq. 3 into:\\nP (a | S, Q) = P (a | I, P, Q) I = {I i , T -t < i ≤ T } (4) P = {P i , T -t < i ≤ T },\\nwhere T is the index of current frame and t is the number of previous frame used in the model. It is also possible to use only single modality or single frame data for prediction.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Framework Overview\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The overall framework of our proposed baseline is illustrated in Fig. 4 and mainly consists of three components. The first is the feature extraction backbone, which includes both image and point cloud feature extractor. The second part is the region proposal module for object embedding, and the last component is the QA-head for answer prediction.\\nInitially, the surrounded-view images and point clouds are fed into the feature extraction backbone, with features projected onto the Bird's-Eye-View (BEV). Subsequently, 3D bounding boxes inferred by a pre-trained detection model are used to crop and pool object features. Finally, the QAmodel takes the question features and the object features as input for cross-modal interaction to predict the answer.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Input Embedding\",\n \"publication_ref\": [\n \"b13\",\n \"b12\",\n \"b24\",\n \"b29\",\n \"b41\",\n \"b11\",\n \"b32\"\n ],\n \"table_ref\": [],\n \"text\": \"Question Embedding For a question Q = {w i } nq i=1 that contains n q words, we first tokenize it and initialize the tokens with pre-trained GloVe (Pennington, Socher, and Manning 2014) embeddings. The sequence is then fed into a single-layer biLSTM (Hochreiter and Schmidhuber 1997) for word-level context encoding. Each word feature w i is represented by the concatenation of the forward and backward hidden states of the biLSTM, denoted as:\\nw i = [ → h i ; ← h i ] ∈ R d ,(5)\\nand the question embedding is represented as Q ∈ R nq×d .\\nVisual Feature Extraction We adopt leading-edge 3D detection techniques for visual feature extraction. As shown in Fig. 4, it entails two branches: image stream and point cloud stream. For multi-view images, ResNet (He et al. 2016) with FPN (Lin et al. 2017) is used as the backbone for multi-scale feature extraction. Then, in order to make the feature spatialaware, we estimate the depth of the pixels in the images and lift them to 3D virtual points with a view transformer inspired by LSS (Philion and Fidler 2020). Finally, pooling along the Z-axis compresses the feature in voxel space, producing the BEV featmap M I ∈ R H×W ×dm .\\nFor point clouds, we first partition 3D space into voxels, transforming raw point clouds into binary voxel grids (Zhou and Tuzel 2018). Subsequently, 3D sparse convolutional neural network (Graham, Engelcke, and Van Der Maaten 2018) is applied to the voxel grid for feature representation. Similar to the image features mentioned earlier, Z-axis pooling yields the point cloud BEV featmap M P ∈ R H×W ×dm . Combining M I and M P , we can aggregate them to obtain multi-modal featmap M ∈ R H×W ×dm .\\nObject Embedding Following 2D detection works (Ren et al. 2015), we crop and pool the features in bounding boxes as the object embedding. However, unlike standard 2D bounding boxes aligned with the coordinate axis in images, projecting 3D boxes to BEV yields rotated boxes unsuited for standard RoI Pooling. To this end, we make some modifications. Firstly, we project the 3D box B = [x, y, z, x size , y size , z size , φ] into the BEV featmap:\\nx m = x -R pc F v × F o ,(6)\\nwhere,F v , F o and R pc represent the voxel factor, out size factor of the backbone, and the point cloud range, respectively. All box parameters follow the Eq. 6 to transform into BEV space except the heading angle φ. Then, based on the center and size of the box, we can easily calculate the four vertices V = {x i , y i } 3 i=0 . Secondly, we calculate the rotated vertex V ′ via the heading angle φ:\\nx ′ i y ′ i = cos φ -sin φ sin φ cos φ x i y i(7)\\nFinally, we use cross product algorithm to identify pixel membership in the rotated rectangle. Then, we perform mean pooling on the features of all the pixels inside the rectangle to obtain the object embedding O ∈ R N ×dm . Algorithm details can be found in supplementary materials.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Answer Head and Training\",\n \"publication_ref\": [\n \"b37\"\n ],\n \"table_ref\": [],\n \"text\": \"We adopt the classical VQA model MCAN (Yu et al. 2019) as our answer head. It leverages stacked self-attention layers to model the language and visual context independently, along with stacked cross-attention layers for cross-modal feature interaction. The fused features are then projected to the answer space for prediction via basic MLP layers.\\nDuring the training phase, we extract the object embeddings using a pre-trained 3D detection model offline. And answer head is trained with the standard cross-entropy loss. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To validate the challenge of NuScenes-QA, we assess baseline performance in various configurations: camera-only or LiDAR-only single-modality models, camera-lidar fusion models, and diverse answering heads. We conduct ablation studies on crucial steps of the baseline, including BEV feature cropping and pooling strategies, as well as the influence of detected 3D bounding boxes. Furthermore, visualization samples are showcased in supplementary material.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Evaluation Metrics\",\n \"publication_ref\": [\n \"b1\",\n \"b2\"\n ],\n \"table_ref\": [],\n \"text\": \"Questions in NuScenes-QA span 5 categories based on query format: 1) Exist, querying the existence of a object in the scene; 2) Count, object counting under specified conditions; 3) Object, object recognition based on language description; 4) Status, querying the status of a specified object; 5) Comparison, specified objects or their status comparison. Additionally, questions are also divided into two groups based their complexity of reasoning: zero-hop (denoted as H0) and one-hop (denoted as H1). We adopt the Top-1 accuracy as our evaluation metric, follow the practice of many other VQA works (Antol et al. 2015;Azuma et al. 2022), and evaluate the performance of different question types separately.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Implementation Details\",\n \"publication_ref\": [\n \"b14\",\n \"b36\"\n ],\n \"table_ref\": [],\n \"text\": \"For the feature extraction backbone, we use the pre-trained detection model following the original settings (Huang et al. 2021;Yin, Zhou, and Krahenbuhl 2021;Jiao et al. 2023a). The dimension of the QA model d m is set to 512, and MCAN adopts a 6-layer encoder-decoder version. As for training, we used the Adam optimizer with an initial learning rate of 1e-4 and half decaying every 2 epochs. All experiments are conducted with a batch size of 256 on 2 NVIDIA GeForce RTX 3090 GPUs. More details can be found in supplementary material.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Quantitative Results\",\n \"publication_ref\": [\n \"b14\",\n \"b0\",\n \"b37\"\n ],\n \"table_ref\": [\n \"tab_3\"\n ],\n \"text\": \"Compared Methods As mentioned earlier, our task can be divided into three settings: camera-only, LiDAR-only, camera+LiDAR. To explore the impact of different modalities on the question-answering performance, we select representative backbone for each setting. We choose BEVDet (Huang et al. 2021) for camera-only setting, which proposed a novel paradigm of explicitly encoding the perspectiveview features into the BEV space. CenterPoint (Yin, Zhou, and Krahenbuhl 2021) is selected for LiDAR-only setting. It introduced a center-based object keypoint detector and has shown excellent performance in both detection accuracy and speed. For the multi-modal model, we opt for MSMD-Fusion (Jiao et al. 2023a), which leverages depth and finegrained LiDAR-camera interaction, achieving state-of-theart results on the nuScenes detection benchmark for single model.\\nRegarding the QA-head, we select two classic models, BUTD (Anderson et al. 2018) and MCAN (Yu et al. 2019). BUTD advocates for computing bottom-up and top-down attention on salient regions of the image. MCAN stacks selfattention and cross-attention modules for vision-language feature interaction. To establish the upper bound of the QA models, we employ perfect perceptual results, i.e., groundtruth object labels. Specifically, we use GloVe for objects and their status embedding, noted as GroundTruth in Table 1. Additionally, we design a Q-Only baseline to investigate the impact of language bias. Q-Only can be considered as a blind model that ignores the visual input.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results and Discussions\",\n \"publication_ref\": [\n \"b39\"\n ],\n \"table_ref\": [\n \"tab_3\",\n \"tab_3\",\n \"tab_3\",\n \"tab_3\"\n ],\n \"text\": \"According to the results shown in Table 1, we have the following observations that are worth discussing.\\n1. It is evident that visual data play a critical role in the performance of our task. When comparing the Q-Only baseline to others, we find that it only achieves an accuracy of 53.4%, which is significantly lower than that of other models. For instance, MSMDFusion+MCAN performs 7% better. This indicates that model cannot achieve good performance solely rely on language shortcuts, but needs to leverage rich visual information.\\n2. Referring to the bottom part of Table 1, we can see that the LiDAR-based CenterPoint outperforms the camerabased BEVDet, achieving accuracy of 57.9% and 59.5%, respectively. We attribute this performance gap to the task characteristics. Images possess detailed texture information, point clouds excel in structure and spatial representation. Our proposed NuScenes-QA emphasizes more on the understanding of structure and spatial relationships of objects. On the other hand, the fusion-based model MSMDFusion attains the best performance with an accuracy of 60.4%, demonstrating the camera and LiDAR data are complementary. Further work can explore how to better exploit the complementary information of multi-modal data. Of course, our baselines still have a long way to go compared to the GroundTruth (achieving an accuracy of 84.3%).\\n3. According to Table 1, QA-head has a significant impact on the performance. With the same detection backbone, we observed that the QA-head based on MCAN outperforms BUTD by a large margin. For example, the overall accuracy of CenterPoint+MCAN is 59.5%, 1.4% higher than CenterPoint+BUTD. A dedicated QA-head designed for NuScenes-QA may lead to a greater improvement. We leave this as future work.\\n4. In a horizontal comparison of Table 1, it is not difficult to find that counting is the most difficult among all question types. Our best baseline model achieved just 23.2% accuracy, much lower than other question types. Counting is historically tough in visual question answering, and some explorations (Zhang, Hare, and Prügel-Bennett 2018) have been made in traditional 2D-QA. Future efforts could involve counting modules to enhance its performance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ablation Studies\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\",\n \"tab_5\",\n \"tab_6\"\n ],\n \"text\": \"To validate effectiveness of different operations in our baselines, we conduct extensive ablation experiments on the NuScenes-QA test split using the CenterPoint+MCAN baseline combination.\\nEffects of Bounding Boxes Most 2D and 3D VQA models fuse the visual feature with object bounding box in the object embedding stage, making it position-aware. We follow this paradigm and evaluate the impact of 3D bounding boxes in our NuScenes-QA. Specifically, we project the 7-dimensional box B = [x, y, z, x size , y size , z size , φ] obtained from the detection model onto the same dimension as the object embeddings using MLP, and concatenate the two features as the final input for the QA head. As shown in Table 2, we are surprised to find that the performance varies significantly on different data. Adding box features for ground truth can increase the model's accuracy from 70.8% to 84.3%, a significant improvement of 13.5%. However, adding the detected boxes slightly decreased performance by 0.6%, which is counterproductive. We speculate that this phenomenon may be caused by two reasons. On one hand, the current 3D detection models are still immature, and the noise in the detected boxes hurts the QA model. adding box features is not significant.\\nBEV Feature Crop Strategy As mentioned earlier, due to the non-parallelism between the 3D boxes and the BEV coordinate axes, we cannot perform standard RoI pooling as in traditional 2D images. Therefore, we use cross product algorithm to determine pixels inside the rotated box for feature cropping. In addition to this method, we can also use a simpler approach, which directly uses the circumscribed rectangle of the rotated box parallel to the coordinate axes as the cropping region. Table 3 shows the performance comparison of these two crop strategy, where the circumscribed box is slightly inferior to the rotated box. The reason for this is that NuScenes-QA contains many elongated objects, such as bus and truck. These objects occupy a small area in the BEV space, but their circumscribed rectangles have a large range, making the object features over smoothing.\\nBEV Feature Pooling Strategy In terms of the feature pooling strategy for the cropped regions, we compared the classic Max Pooling and Mean Pooling operations. As illustrated in Table 4, Max Pooling achieved an accuracy of 58.9% under the same conditions, which is 0.6% lower than Mean Pooling. We speculate that this difference may be due to the fact that Max Pooling focuses on the texture features within the region, while Mean Pooling preserves the overall features. Our proposed NuScenes-QA mainly tests the model's ability of understanding the structure of objects and their spatial relationships in street views, and relatively ignores the texture of the objects. Thus, Mean Pooling has a slight advantage over Max Pooling.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we apply VQA to the context of autonomous driving. We construct NuScenes-QA, the first large-scale multi-modal VQA benchmark for autonomous driving scenario. NuScenes-QA are generated automatically based on visual scene graphs and question templates, containing 34K scenes and 460K question-answer pairs. Alongside a series of baseline models, comprehensive experiments establish a solid foundation for future research. We strongly hope that NuScenes-QA can invigorate the evolution of multi-modal VQA and propel advancements in autonomous driving. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Question Templates\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"As described in the main paper, we employ manually designed question templates to generate questions programmatically. For instance, the question \\\"Are there any bicycles to the back left of the moving car?\\\" has a template of \\\"Are there any <A2> <O2>s to the <R> of the <A> <O>?\\\", where <A>, <O>, and <R> denote status, object, and relation, respectively. Table 6 enumerates all the templates based on different question types. Among them, one-hop questions involve reasoning about relations between objects, while zero-hop questions are relatively simpler. We design a total of 5 question types and 16 different semantic question templates. To increase question diversity, we created different variations of each template, such as \\\"How many <A2> <O2>s are to the <R> of the <A> <O>?\\\" can also be expressed as \\\"There is a <A> <O>; how many <A2> <O2>s are to the <R> of it?\\\" In the end, we get a total of 66 different question templates. In future work, we can further enrich the question templates to enhance the diversity of questions.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\",\n \"fig_1\"\n ],\n \"heading\": \"Question Distribution of the First Four Words\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In Figure 5, we visualize the distribution of the first four words in the questions, from which we can draw two observations. First, our data distribution is balanced, as already verified in Figure 3. Second, our questions encompass a diverse range of visual semantics. To answer them, not only the object categories such as pedestrian, motorcycle are demanded, but also their status, such as moving or parked. The semantic complexities in our questions also presents a considerable challenge for models.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Visualization Examples\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In order to get a deeper insights of the difficulty of the NuScenes-QA dataset and to validate the performance of proposed baseline models, we select some samples from the test split for visualization, as shown in Figure 6 and Figure 7. We present the point clouds, surround-view images, questions, ground truth answers, and the predictions of 6 different baseline models.\\nThe visual data of these examples (with LiDAR point clouds on the left, surround-view images in the middle) provide compelling evidence of the challenges inherent in NuScenes-QA. First, the point clouds and images are in different coordinate systems, which poses great difficulties for the model to comprehend spatial relationships and fuse multi-modal information. Second, the street-view scenes often contain a lot of background noise, which can significantly interfere with the model's ability to locate and recognize foreground objects relevant to the question. Thirdly, some scenes contain dense and diverse objects, making the reasoning for questions in these complex scenes prone to errors.\\nFigure 6 shows some successful cases, demonstrating the impressive performance of the baseline models. Inevitably, there are also failed cases, as shown in Figure 7. By comparing the prediction results of different baselines, we find that multi-modal fusion models usually outperform single-modal models, which confirms the complementarity between point clouds and images.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Discussion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Limitations As the first work in this field, there are still many shortcomings in NuScenes-QA that need to be im-\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This work was supported in part by National Natural Science Foundation of China Project (No. 62072116) and Shanghai Science and Technology Program [Project No. 21JC1400600].\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Dataset Visual Modality\",\n \"publication_ref\": [\n \"b6\",\n \"b35\",\n \"b2\",\n \"b34\",\n \"b40\",\n \"b27\"\n ],\n \"table_ref\": [],\n \"text\": \"Multi Frame Scenario Collection # Scenes #Amount EQA (Das et al. 2018) image indoor template 767 1.5k 3DQA (Ye et al. 2022) point cloud indoor human 806 10k ScanQA (Azuma et al. 2022) point cloud indoor auto + human 800 41k CLEVR3D (Yan et al. 2021) point cloud indoor template 1,129 60.1k FE-3DGQA (Zhao et al. 2022) point cloud indoor human 800 20k SQA3D (Ma et al. 2023) image Are there any other <A2><O2>s that in the same status as the <A><O>? Is there another <A2><O2>that has the same status as the <A><O>? Are there any other <A2><O2>s of the same status as the <A><O>? Is there another <A2><O2>of the same status as the <A><O>?\\n1 Are there any <A2><O2>s to the <R>of the <A><O>? There is a <A><O>; are there any <A2><O2>s to the <R>of it? Are there any other <A3><O3>s that in the same status as the <A2><O2>[that is] to the <R>of the <A><O>? Is there another <A3><O3>that has the same status as the <A2><O2>[that is] to the <R>of the <A><O>? Are there any other <A3><O3>s of the same status as the <A2><O2>[that is] to the <R>of the <A><O>? Is there another <A3><O3>of the same status as the <A2><O2>[that is] to the <R>of the <A><O>? Is the status of the <A><O>the same as the <A2><O2>? Do the <A><O>and the <A2><O2>have the same status? Is the <A><O>the same status as the <A2><O2>? Does the <A><O>have the same status as the <A2><O2>?\\n1\\nIs the status of the <A2><O2>[that is] to the <R>of the <A><O>the same as the <A3><O3>? Does the <A2><O2>[that is] to the <R>of the <A><O>have the same status as the <A3><O3>? Do the <A2><O2>[that is] to the <R>of the <A><O>and the <A3><O3>have the same status? There is a <A2><O2>[that is] to the <R>of the <A><O>; does it have the same status as the <A3><O3>?\\nThere is a <A2><O2>[that is] to the <R>of the <A><O>; is it the same status as the <A3><O3>?\\nThere is a <A2><O2>[that is] to the <R>of the <A><O>; is its status the same as the <A3><O3>?\\nIs the status of the <A><O>the same as the <A3><O3>[that is] to the <R>of the <A2><O2>? Does the <A><O>have the same status as the <A3><O3>[that is] to the <R>of the <A2><O2>? Do the <A><O>and the <A3><O3>[that is] to the <R>of the <A2><O2>have the same status? There is a <A><O>; does it have the same status as the <A3><O3>[that is] to the <R>of the <A2><O2>? There is a <A><O>; is its status the same as the <A3><O3>[that is] to the <R>of the <A2><O2>? There is a <A><O>; is it the same status as the <A3><O3>[that is] to the <R>of the <A2><O2>? Is the status of the <A2><O2>[that is] to the <R>of the <A><O>the same as the <A4><O4>[that is] to the <R2>of the <A3><O3>? Does the <A2><O2>[that is] to the <R>of the <A><O>have the same status as the <A4><O4>[that is] to the <R2>of the <A3><O3>? Do the <A2><O2>[that is] to the <R>of the <A><O>and the <A4><O4>[that is] to the <R2>of the <A3><O3>have the same status? There is a <A2><O2>[that is] to the <R>of the <A><O>; does it have the same status as the <A4><O4>[that is] to the <R2>of the <A3><O3>? There is a <A2><O2>[that is] to the <R>of the <A><O>; is its status the same as the <A4><O4>[that is] to the <R2>of the <A3><O3>? There is a <A2><O2>[that is] to the <R>of the <A><O>; is it the same status as the <A4><O4>[that is] to the <R2>of the <A3><O3>?\\nTable 6: Question templates.\\nproved. First, our questions focus on foreground objects and their relationships, but ignore the connection between objects and the background. For example, we can ask the agent which parking spots ahead are available, which is also important for autonomous driving. Second, our questions focus on the model's understanding of the visual scene, which may be too naive for humans. Meanwhile, traditional image or video based question answering is gradually moving towards higher levels of reasoning such as causal inference. Third, we define the relationships between objects using spatial positioning only, lacking many rich semantic relationships. Finally, although NuScenes-QA is large scale, our questions may not be sufficiently diverse in language due to templates based generation. In summary, there is still significant room for improvement in NuScenes-QA.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Future Work\",\n \"publication_ref\": [\n \"b2\"\n ],\n \"table_ref\": [],\n \"text\": \"In future work, we can gradually improve NuScenes-QA in the following aspects. Firstly, we can add object localization tasks, like ScanQA (Azuma et al. 2022), to increase the model's interpretability. Secondly, we can utilize crowd sourcing to manually annotate the scene graphs, including new relationships, object status, and more. This can enrich the semantics of questions and increase the diversity of language. Lastly, we can consider integrating with perception tasks such as tracking to increase its practical value. \"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":""},"doi":{"kind":"string","value":""},"references":{"kind":"list like","value":[{"authors":"P Anderson; X He; C Buehler; D Teney; M Johnson; S Gould; L Zhang","journal":"","ref_id":"b0","title":"Bottom-up and top-down attention for image captioning and visual question answering","year":"2018"},{"authors":"S Antol; A Agrawal; J Lu; M Mitchell; D Batra; C L Zitnick; D Parikh","journal":"","ref_id":"b1","title":"Vqa: Visual question answering","year":"2015"},{"authors":"D Azuma; T Miyanishi; S Kurita; M Kawanabe","journal":"","ref_id":"b2","title":"ScanQA: 3D question answering for spatial scene understanding","year":"2022"},{"authors":"H Caesar; V Bankiti; A H Lang; S Vora; V E Liong; Q Xu; A Krishnan; Y Pan; G Baldan; O Beijbom","journal":"","ref_id":"b3","title":"nuscenes: A multimodal dataset for autonomous driving","year":"2020"},{"authors":"Y Chen; J Liu; X Zhang; X Qi; J Jia","journal":"","ref_id":"b4","title":"Vox-elNeXt: Fully Sparse VoxelNet for 3D Object Detection and Tracking","year":"2023"},{"authors":"A Dai; A X Chang; M Savva; M Halber; T Funkhouser; M Nießner","journal":"","ref_id":"b5","title":"Scannet: Richly-annotated 3d reconstructions of indoor scenes","year":"2017"},{"authors":"A Das; S Datta; G Gkioxari; S Lee; D Parikh; D Batra","journal":"","ref_id":"b6","title":"Embodied question answering","year":"2018"},{"authors":"T Deruyttere; S Vandenhende; D Grujicic; L Van Gool; M.-F Moens","journal":"","ref_id":"b7","title":"Talk2car: Taking control of your self-driving car","year":"2019"},{"authors":"W Dongming; H Wencheng; W Tiancai; D Xingping; Z Xiangyu; J Shen","journal":"","ref_id":"b8","title":"Referring Multi-Object Tracking","year":"2023"},{"authors":"A Geiger; P Lenz; R Urtasun","journal":"IEEE","ref_id":"b9","title":"Are we ready for autonomous driving? the kitti vision benchmark suite","year":"2012"},{"authors":"Y Goyal; T Khot; D Summers-Stay; D Batra; D Parikh","journal":"","ref_id":"b10","title":"Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering","year":"2017"},{"authors":"B Graham; M Engelcke; L Van Der Maaten","journal":"","ref_id":"b11","title":"3d semantic segmentation with submanifold sparse convolutional networks","year":"2018"},{"authors":"K He; X Zhang; S Ren; J Sun","journal":"","ref_id":"b12","title":"Deep residual learning for image recognition","year":"2016"},{"authors":"S Hochreiter; J Schmidhuber","journal":"Neural computation","ref_id":"b13","title":"Long short-term memory","year":"1997"},{"authors":"J Huang; G Huang; Z Zhu; Y Ye; D Du","journal":"","ref_id":"b14","title":"Bevdet: High-performance multi-camera 3d object detection in bird-eye-view","year":"2021"},{"authors":"D A Hudson; C D Manning","journal":"","ref_id":"b15","title":"GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering","year":"2019"},{"authors":"Y Jang; Y Song; Y Yu; Y Kim; G Kim","journal":"","ref_id":"b16","title":"Tgifqa: Toward spatio-temporal reasoning in visual question answering","year":"2017"},{"authors":"J Jiang; Z Chen; H Lin; X Zhao; Y Gao","journal":"","ref_id":"b17","title":"Divide and conquer: Question-guided spatio-temporal contextual attention for video question answering","year":"2020"},{"authors":"Y Jiao; Z Jie; S Chen; J Chen; L Ma; Y.-G Jiang","journal":"","ref_id":"b18","title":"MSMDFusion: Fusing LiDAR and Camera at Multiple Scales with Multi-Depth Seeds for 3D Object Detection","year":"2023"},{"authors":"Y Jiao; Z Jie; S Chen; L Cheng; J Chen; L Ma; Y.-G Jiang","journal":"","ref_id":"b19","title":"Instance-aware Multi-Camera 3D Object Detection with Structural Priors Mining and Self-Boosting Learning","year":"2023"},{"authors":"J Johnson; B Hariharan; L Van Der Maaten; L Fei-Fei; C Lawrence Zitnick; R Girshick","journal":"","ref_id":"b20","title":"Clevr: A diagnostic dataset for compositional language and elementary visual reasoning","year":"2017"},{"authors":"J Johnson; R Krishna; M Stark; L.-J Li; D Shamma; M Bernstein; L Fei-Fei","journal":"","ref_id":"b21","title":"Image retrieval using scene graphs","year":"2015"},{"authors":"J.-H Kim; J Jun; B.-T Zhang","journal":"Advances in neural information processing systems","ref_id":"b22","title":"Bilinear attention networks","year":"2018"},{"authors":"J Lei; L Yu; M Bansal; T L Berg","journal":"","ref_id":"b23","title":"Tvqa: Localized, compositional video question answering","year":"2018"},{"authors":"T.-Y Lin; P Dollár; R Girshick; K He; B Hariharan; S Belongie","journal":"","ref_id":"b24","title":"Feature pyramid networks for object detection","year":"2017"},{"authors":"Z Liu; H Tang; A Amini; X Yang; H Mao; D Rus; S Han","journal":"","ref_id":"b25","title":"BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation","year":"2023"},{"authors":"J Lu; J Yang; D Batra; D Parikh","journal":"Advances in neural information processing systems","ref_id":"b26","title":"Hierarchical question-image co-attention for visual question answering","year":"2016"},{"authors":"X Ma; S Yong; Z Zheng; Q Li; Y Liang; S.-C Zhu; S Huang","journal":"","ref_id":"b27","title":"SQA3D: Situated Question Answering in 3D Scenes","year":"2023"},{"authors":"J Pennington; R Socher; C D Manning","journal":"","ref_id":"b28","title":"Glove: Global vectors for word representation","year":"2014"},{"authors":"J Philion; S Fidler","journal":"","ref_id":"b29","title":"Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d","year":"2020-08-23"},{"authors":"T Qian; J Chen; S Chen; B Wu; Y.-G Jiang","journal":"IEEE Transactions on Multimedia","ref_id":"b30","title":"Scene graph refinement network for visual question answering","year":"2022"},{"authors":"T Qian; R Cui; J Chen; P Peng; X Guo; Y.-G Jiang","journal":"","ref_id":"b31","title":"Locate before Answering: Answer Guided Question Localization for Video Question Answering","year":"2022"},{"authors":"S Ren; K He; R Girshick; J Sun","journal":"Advances in neural information processing systems","ref_id":"b32","title":"Faster r-cnn: Towards real-time object detection with region proposal networks","year":"2015"},{"authors":"H Tan; M Bansal","journal":"","ref_id":"b33","title":"Lxmert: Learning crossmodality encoder representations from transformers","year":"2019"},{"authors":"X Yan; Z Yuan; Y Du; Y Liao; Y Guo; Z Li; S Cui","journal":"","ref_id":"b34","title":"CLEVR3D: Compositional language and elementary visual reasoning for question answering in 3D realworld scenes","year":"2021"},{"authors":"S Ye; D Chen; S Han; J Liao","journal":"IEEE Transactions on Visualization and Computer Graphics","ref_id":"b35","title":"3D question answering","year":"2022"},{"authors":"T Yin; X Zhou; P Krahenbuhl","journal":"","ref_id":"b36","title":"Centerbased 3d object detection and tracking","year":"2021"},{"authors":"Z Yu; J Yu; Y Cui; D Tao; Q Tian","journal":"","ref_id":"b37","title":"Deep modular co-attention networks for visual question answering","year":"2019"},{"authors":"P Zhang; X Li; X Hu; J Yang; L Zhang; L Wang; Y Choi; J Gao","journal":"","ref_id":"b38","title":"Vinvl: Revisiting visual representations in vision-language models","year":"2021"},{"authors":"Y Zhang; J Hare; A Prügel-Bennett","journal":"","ref_id":"b39","title":"Learning to count objects in natural images for visual question answering","year":"2018"},{"authors":"L Zhao; D Cai; J Zhang; L Sheng; D Xu; R Zheng; Y Zhao; L Wang; X Fan","journal":"IEEE Transactions on Circuits and Systems for Video Technology","ref_id":"b40","title":"Towards Explainable 3D Grounded Visual Question Answering: A New Benchmark and Strong Baseline","year":"2022"},{"authors":"Y Zhou; O Tuzel","journal":"","ref_id":"b41","title":"Voxelnet: End-to-end learning for point cloud based 3d object detection","year":"2018"},{"authors":"L Zhu; Z Xu; Y Yang; A G Hauptmann","journal":"International Journal of Computer Vision","ref_id":"b42","title":"Uncovering the temporal context for video question answering","year":"2017"}],"string":"[\n {\n \"authors\": \"P Anderson; X He; C Buehler; D Teney; M Johnson; S Gould; L Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Bottom-up and top-down attention for image captioning and visual question answering\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"S Antol; A Agrawal; J Lu; M Mitchell; D Batra; C L Zitnick; D Parikh\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Vqa: Visual question answering\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"D Azuma; T Miyanishi; S Kurita; M Kawanabe\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"ScanQA: 3D question answering for spatial scene understanding\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"H Caesar; V Bankiti; A H Lang; S Vora; V E Liong; Q Xu; A Krishnan; Y Pan; G Baldan; O Beijbom\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"nuscenes: A multimodal dataset for autonomous driving\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Y Chen; J Liu; X Zhang; X Qi; J Jia\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Vox-elNeXt: Fully Sparse VoxelNet for 3D Object Detection and Tracking\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"A Dai; A X Chang; M Savva; M Halber; T Funkhouser; M Nießner\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Scannet: Richly-annotated 3d reconstructions of indoor scenes\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"A Das; S Datta; G Gkioxari; S Lee; D Parikh; D Batra\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Embodied question answering\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"T Deruyttere; S Vandenhende; D Grujicic; L Van Gool; M.-F Moens\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Talk2car: Taking control of your self-driving car\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"W Dongming; H Wencheng; W Tiancai; D Xingping; Z Xiangyu; J Shen\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Referring Multi-Object Tracking\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"A Geiger; P Lenz; R Urtasun\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b9\",\n \"title\": \"Are we ready for autonomous driving? the kitti vision benchmark suite\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Y Goyal; T Khot; D Summers-Stay; D Batra; D Parikh\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"B Graham; M Engelcke; L Van Der Maaten\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"3d semantic segmentation with submanifold sparse convolutional networks\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"K He; X Zhang; S Ren; J Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Deep residual learning for image recognition\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"S Hochreiter; J Schmidhuber\",\n \"journal\": \"Neural computation\",\n \"ref_id\": \"b13\",\n \"title\": \"Long short-term memory\",\n \"year\": \"1997\"\n },\n {\n \"authors\": \"J Huang; G Huang; Z Zhu; Y Ye; D Du\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Bevdet: High-performance multi-camera 3d object detection in bird-eye-view\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"D A Hudson; C D Manning\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Y Jang; Y Song; Y Yu; Y Kim; G Kim\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Tgifqa: Toward spatio-temporal reasoning in visual question answering\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"J Jiang; Z Chen; H Lin; X Zhao; Y Gao\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Divide and conquer: Question-guided spatio-temporal contextual attention for video question answering\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Y Jiao; Z Jie; S Chen; J Chen; L Ma; Y.-G Jiang\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"MSMDFusion: Fusing LiDAR and Camera at Multiple Scales with Multi-Depth Seeds for 3D Object Detection\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Y Jiao; Z Jie; S Chen; L Cheng; J Chen; L Ma; Y.-G Jiang\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Instance-aware Multi-Camera 3D Object Detection with Structural Priors Mining and Self-Boosting Learning\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"J Johnson; B Hariharan; L Van Der Maaten; L Fei-Fei; C Lawrence Zitnick; R Girshick\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Clevr: A diagnostic dataset for compositional language and elementary visual reasoning\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"J Johnson; R Krishna; M Stark; L.-J Li; D Shamma; M Bernstein; L Fei-Fei\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Image retrieval using scene graphs\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"J.-H Kim; J Jun; B.-T Zhang\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b22\",\n \"title\": \"Bilinear attention networks\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"J Lei; L Yu; M Bansal; T L Berg\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Tvqa: Localized, compositional video question answering\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"T.-Y Lin; P Dollár; R Girshick; K He; B Hariharan; S Belongie\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Feature pyramid networks for object detection\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Z Liu; H Tang; A Amini; X Yang; H Mao; D Rus; S Han\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"J Lu; J Yang; D Batra; D Parikh\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b26\",\n \"title\": \"Hierarchical question-image co-attention for visual question answering\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"X Ma; S Yong; Z Zheng; Q Li; Y Liang; S.-C Zhu; S Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"SQA3D: Situated Question Answering in 3D Scenes\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"J Pennington; R Socher; C D Manning\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Glove: Global vectors for word representation\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"J Philion; S Fidler\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d\",\n \"year\": \"2020-08-23\"\n },\n {\n \"authors\": \"T Qian; J Chen; S Chen; B Wu; Y.-G Jiang\",\n \"journal\": \"IEEE Transactions on Multimedia\",\n \"ref_id\": \"b30\",\n \"title\": \"Scene graph refinement network for visual question answering\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"T Qian; R Cui; J Chen; P Peng; X Guo; Y.-G Jiang\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Locate before Answering: Answer Guided Question Localization for Video Question Answering\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"S Ren; K He; R Girshick; J Sun\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b32\",\n \"title\": \"Faster r-cnn: Towards real-time object detection with region proposal networks\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"H Tan; M Bansal\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Lxmert: Learning crossmodality encoder representations from transformers\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"X Yan; Z Yuan; Y Du; Y Liao; Y Guo; Z Li; S Cui\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"CLEVR3D: Compositional language and elementary visual reasoning for question answering in 3D realworld scenes\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"S Ye; D Chen; S Han; J Liao\",\n \"journal\": \"IEEE Transactions on Visualization and Computer Graphics\",\n \"ref_id\": \"b35\",\n \"title\": \"3D question answering\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"T Yin; X Zhou; P Krahenbuhl\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Centerbased 3d object detection and tracking\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Z Yu; J Yu; Y Cui; D Tao; Q Tian\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Deep modular co-attention networks for visual question answering\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"P Zhang; X Li; X Hu; J Yang; L Zhang; L Wang; Y Choi; J Gao\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Vinvl: Revisiting visual representations in vision-language models\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Y Zhang; J Hare; A Prügel-Bennett\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Learning to count objects in natural images for visual question answering\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"L Zhao; D Cai; J Zhang; L Sheng; D Xu; R Zheng; Y Zhao; L Wang; X Fan\",\n \"journal\": \"IEEE Transactions on Circuits and Systems for Video Technology\",\n \"ref_id\": \"b40\",\n \"title\": \"Towards Explainable 3D Grounded Visual Question Answering: A New Benchmark and Strong Baseline\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Y Zhou; O Tuzel\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Voxelnet: End-to-end learning for point cloud based 3d object detection\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"L Zhu; Z Xu; Y Yang; A G Hauptmann\",\n \"journal\": \"International Journal of Computer Vision\",\n \"ref_id\": \"b42\",\n \"title\": \"Uncovering the temporal context for video question answering\",\n \"year\": \"2017\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,352.97,376.64,205.03,23.23],"formula_id":"formula_0","formula_text":"θ = cos -1 (B 1 [: 2] -B 2 [: 2]) • V ego [: 2] ∥B 1 [: 2] -B 2 [: 2]∥∥V ego [: 2]∥ ,(1)"},{"formula_coordinates":[3,420.54,418.37,78.09,9.65],"formula_id":"formula_1","formula_text":"V ego = [v x , v y , v z ]"},{"formula_coordinates":[3,326.96,456.32,221.89,76.91],"formula_id":"formula_2","formula_text":"relation = f ront if -30 • < θ <= 30 • f ront lef t if 30 • < θ <= 90 • f ront right if -90 • < θ <= -30 • back lef t if 90 • < θ <= 150 • back rigth if -150 • < θ <= -90 • back else."},{"formula_coordinates":[4,386.54,418.05,104.57,16.52],"formula_id":"formula_3","formula_text":"â = arg max a∈A P (a | S, Q)."},{"formula_coordinates":[4,379.51,493.65,178.49,37.55],"formula_id":"formula_4","formula_text":"P (a | S, Q) = P (a | I, P, Q) I = {I i , T -t < i ≤ T } (4) P = {P i , T -t < i ≤ T },"},{"formula_coordinates":[5,131.51,441.88,160.99,16.52],"formula_id":"formula_5","formula_text":"w i = [ → h i ; ← h i ] ∈ R d ,(5)"},{"formula_coordinates":[5,405.82,382.27,152.18,23.23],"formula_id":"formula_6","formula_text":"x m = x -R pc F v × F o ,(6)"},{"formula_coordinates":[5,381.1,495.41,176.9,26.65],"formula_id":"formula_7","formula_text":"x ′ i y ′ i = cos φ -sin φ sin φ cos φ x i y i(7)"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 352.97,\n 376.64,\n 205.03,\n 23.23\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"θ = cos -1 (B 1 [: 2] -B 2 [: 2]) • V ego [: 2] ∥B 1 [: 2] -B 2 [: 2]∥∥V ego [: 2]∥ ,(1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 420.54,\n 418.37,\n 78.09,\n 9.65\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"V ego = [v x , v y , v z ]\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 326.96,\n 456.32,\n 221.89,\n 76.91\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"relation = f ront if -30 • < θ <= 30 • f ront lef t if 30 • < θ <= 90 • f ront right if -90 • < θ <= -30 • back lef t if 90 • < θ <= 150 • back rigth if -150 • < θ <= -90 • back else.\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 386.54,\n 418.05,\n 104.57,\n 16.52\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"â = arg max a∈A P (a | S, Q).\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 379.51,\n 493.65,\n 178.49,\n 37.55\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"P (a | S, Q) = P (a | I, P, Q) I = {I i , T -t < i ≤ T } (4) P = {P i , T -t < i ≤ T },\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 131.51,\n 441.88,\n 160.99,\n 16.52\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"w i = [ → h i ; ← h i ] ∈ R d ,(5)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 405.82,\n 382.27,\n 152.18,\n 23.23\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"x m = x -R pc F v × F o ,(6)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 381.1,\n 495.41,\n 176.9,\n 26.65\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"x ′ i y ′ i = cos φ -sin φ sin φ cos φ x i y i(7)\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"NuScenes-QA: A Multi-Modal Visual Question Answering Benchmark for Autonomous Driving Scenario"},"abstract":{"kind":"string","value":"We introduce a novel visual question answering (VQA) task in the context of autonomous driving, aiming to answer natural language questions based on street-view clues. Compared to traditional VQA tasks, VQA in autonomous driving scenario presents more challenges. Firstly, the raw visual data are multi-modal, including images and point clouds captured by camera and LiDAR, respectively. Secondly, the data are multi-frame due to the continuous, real-time acquisition. Thirdly, the outdoor scenes exhibit both moving foreground and static background. Existing VQA benchmarks fail to adequately address these complexities. To bridge this gap, we propose NuScenes-QA, the first benchmark for VQA in the autonomous driving scenario, encompassing 34K visual scenes and 460K question-answer pairs. Specifically, we leverage existing 3D detection annotations to generate scene graphs and design question templates manually. Subsequently, the question-answer pairs are generated programmatically based on these templates. Comprehensive statistics prove that our NuScenes-QA is a balanced large-scale benchmark with diverse question formats. Built upon it, we develop a series of baselines that employ advanced 3D detection and VQA techniques. Our extensive experiments highlight the challenges posed by this new task. Codes and dataset are available at https://github.com/qiantianwen/NuScenes-QA."},"authors":{"kind":"string","value":"Tianwen Qian; Jingjing Chen; Linhai Zhuo; Yang Jiao; Yu-Gang Jiang"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 2 :2Figure 2: Data construction flow of NuScenes-QA. First, the scene graphs are generated using the annotated object labels and 3D bounding boxes. Then, we design question templates manually, and instantiate the question-answer pairs with them. Finally, the generated data are filtered based on certain rules.","figure_data":"","figure_id":"fig_0","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Statistical distributions of questions and answers in the NuScenes-QA training split.","figure_data":"","figure_id":"fig_1","figure_label":"3","figure_type":"figure"},{"figure_caption":"Fig. 33depicts various statistical distributions of NuScenes-QA. Fig. 3(a) showcases a broad spectrum of question lengths (5 to 35 words). Fig. 3(b) and 3(c) present answer and question category distributions, revealing the balance of NuScenes-QA. A balanced dataset can prevent models from learning answer biases or language shortcuts, which are common in many other VQA benchmarks (Antol et al. 2015; Azuma et al. 2022).","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Framework of baseline. The multi-view images and point clouds are first processed by the feature extraction backbone to obtain BEV features. Then, the objects embeddings are cropped based on the detected 3D bounding boxes. Finally, these objects features are fed into the question-answering head along with the given question for answer decoding.","figure_data":"","figure_id":"fig_3","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: Question distribution in terms of the first four words.","figure_data":"","figure_id":"fig_4","figure_label":"5","figure_type":"figure"},{"figure_caption":"","figure_data":"","figure_id":"","figure_label":"","figure_type":"figure"},{"figure_caption":"Results of different models on the NuScenes-QA test set. We evaluate top-1 accuracy across the overall test split and different question types. H0 denotes zero-hop and H1 denotes one-hop. C denotes camera, L denotes LiDAR.","figure_data":"ModelsH0Exist H1AllH0Count H1AllH0Object H1AllH0Status H1AllComparison H0 H1 AllAccQ-Only81.7 77.9 79.6 17.8 16.5 17.2 59.4 38.9 42.0 57.2 48.3 51.3 79.5 65.7 66.9 53.4BEVDet+BUTD87.2 80.6 83.7 21.7 20.0 20.9 69.4 45.2 48.8 55.0 50.5 52.0 76.1 66.8 67.7 57.0CenterPoint+BUTD87.7 81.1 84.1 21.9 20.7 21.3 70.2 45.6 49.2 62.8 52.4 55.9 81.6 68.0 69.2 58.1MSMDFusion+BUTD 89.4 81.4 85.1 25.3 21.3 23.2 73.3 48.7 52.3 67.4 55.4 59.5 81.6 67.2 68.5 59.8GroundTruth+BUTD98.9 87.2 92.6 76.8 38.7 57.5 99.7 71.9 76.0 98.8 81.9 87.6 98.1 76.1 78.1 79.2BEVDet+MCAN87.2 81.7 84.2 21.8 19.2 20.4 73.0 47.4 51.2 64.1 49.9 54.7 75.1 66.7 67.4 57.9CenterPoint+MCAN87.7 82.3 84.8 22.5 19.1 20.8 71.3 49.0 52.3 66.6 56.3 59.8 82.4 68.8 70.0 59.5MSMDFusion+MCAN 89.0 82.3 85.4 23.4 21.1 22.2 75.3 50.6 54.3 69.0 56.2 60.6 78.8 68.8 69.7 60.4GroundTruth+MCAN 99.6 95.5 97.4 52.7 39.9 46.2 99.7 86.2 88.2 99.3 95.4 96.8 99.7 90.2 91.0 84.3","figure_id":"tab_3","figure_label":"1","figure_type":"table"},{"figure_caption":"Ablation comparison between model trained with and without detection boxes feature.","figure_data":"VariantsQuestion Types Exi. Cou. Obj. Sta. Com.AllDet w/o boxes 84.8 20.8 52.3 59.8 70.0 59.5Det w/ boxes 84.3 21.7 53.0 57.7 67.2 58.9GT w/o boxes 91.2 38.8 61.1 80.3 76.8 70.8GT w/ boxes97.4 46.2 88.2 96.8 91.0 84.3","figure_id":"tab_4","figure_label":"2","figure_type":"table"},{"figure_caption":"On the other hand, the point cloud represented by XYZ itself has great position expression ability, and the gain from Ablation comparison of BEV feature crop strategies.","figure_data":"CropQuestion Types Exi. Cou. Obj. Sta. Com.AllCir. Box 84.0 21.8 52.3 60.2 65.2 58.8Rot. Box 84.8 20.8 52.3 59.8 70.0 59.5PoolingQuestion Types Exi. Cou. Obj. Sta. Com.AllMax84.2 20.7 51.6 58.0 69.7 58.9Mean84.8 20.8 52.3 59.8 70.0 59.5","figure_id":"tab_5","figure_label":"3","figure_type":"table"},{"figure_caption":"Ablation comparison of BEV feature pooling strategies.","figure_data":"","figure_id":"tab_6","figure_label":"4","figure_type":"table"},{"figure_caption":"","figure_data":"summarizes the differences between our NuScenes-QA and other related 3D VQA datasets. Firstly, to the bestof our knowledge, NuScenes-QA is currently the largest3D VQA datasets, with 34k diverse visual scenes and 460kquestion-answer pairs, averaging 13.5 pairs per scene. Incontrast, other datasets (Das et al. 2018; Ye et al. 2022;Azuma et al. 2022; Yan et al. 2021; Zhao et al. 2022; Maet al. 2023) typically have less than 1000 scenes due to thedifficulty of acquiring 3D data. Secondly, in terms of visualmodality, our NuScenes-QA is multimodal, comprising ofimages and point clouds, posing higher demands on visualreasoning while increasing the dataset's complexity. Addi-tionally, our visual data is multi-frame, requiring temporalinformation mining. Thirdly, Our dataset focuses on outdoorscenes with a combination of dynamic objects and staticbackground, posing higher challenges for answering ques-tions that require perceiving and reasoning about dynamicobjects.In summary, compared to other 3D VQA datasets,NuScenes-QA stands out in terms of scale, data modality,and content, making it an important research resource thatcan advance the research and development of 3D visualquestion answering.","figure_id":"tab_7","figure_label":"5","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Data construction flow of NuScenes-QA. First, the scene graphs are generated using the annotated object labels and 3D bounding boxes. Then, we design question templates manually, and instantiate the question-answer pairs with them. Finally, the generated data are filtered based on certain rules.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Statistical distributions of questions and answers in the NuScenes-QA training split.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 33depicts various statistical distributions of NuScenes-QA. Fig. 3(a) showcases a broad spectrum of question lengths (5 to 35 words). Fig. 3(b) and 3(c) present answer and question category distributions, revealing the balance of NuScenes-QA. A balanced dataset can prevent models from learning answer biases or language shortcuts, which are common in many other VQA benchmarks (Antol et al. 2015; Azuma et al. 2022).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Framework of baseline. The multi-view images and point clouds are first processed by the feature extraction backbone to obtain BEV features. Then, the objects embeddings are cropped based on the detected 3D bounding boxes. Finally, these objects features are fed into the question-answering head along with the given question for answer decoding.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: Question distribution in terms of the first four words.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"\",\n \"figure_id\": \"\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Results of different models on the NuScenes-QA test set. We evaluate top-1 accuracy across the overall test split and different question types. H0 denotes zero-hop and H1 denotes one-hop. C denotes camera, L denotes LiDAR.\",\n \"figure_data\": \"ModelsH0Exist H1AllH0Count H1AllH0Object H1AllH0Status H1AllComparison H0 H1 AllAccQ-Only81.7 77.9 79.6 17.8 16.5 17.2 59.4 38.9 42.0 57.2 48.3 51.3 79.5 65.7 66.9 53.4BEVDet+BUTD87.2 80.6 83.7 21.7 20.0 20.9 69.4 45.2 48.8 55.0 50.5 52.0 76.1 66.8 67.7 57.0CenterPoint+BUTD87.7 81.1 84.1 21.9 20.7 21.3 70.2 45.6 49.2 62.8 52.4 55.9 81.6 68.0 69.2 58.1MSMDFusion+BUTD 89.4 81.4 85.1 25.3 21.3 23.2 73.3 48.7 52.3 67.4 55.4 59.5 81.6 67.2 68.5 59.8GroundTruth+BUTD98.9 87.2 92.6 76.8 38.7 57.5 99.7 71.9 76.0 98.8 81.9 87.6 98.1 76.1 78.1 79.2BEVDet+MCAN87.2 81.7 84.2 21.8 19.2 20.4 73.0 47.4 51.2 64.1 49.9 54.7 75.1 66.7 67.4 57.9CenterPoint+MCAN87.7 82.3 84.8 22.5 19.1 20.8 71.3 49.0 52.3 66.6 56.3 59.8 82.4 68.8 70.0 59.5MSMDFusion+MCAN 89.0 82.3 85.4 23.4 21.1 22.2 75.3 50.6 54.3 69.0 56.2 60.6 78.8 68.8 69.7 60.4GroundTruth+MCAN 99.6 95.5 97.4 52.7 39.9 46.2 99.7 86.2 88.2 99.3 95.4 96.8 99.7 90.2 91.0 84.3\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablation comparison between model trained with and without detection boxes feature.\",\n \"figure_data\": \"VariantsQuestion Types Exi. Cou. Obj. Sta. Com.AllDet w/o boxes 84.8 20.8 52.3 59.8 70.0 59.5Det w/ boxes 84.3 21.7 53.0 57.7 67.2 58.9GT w/o boxes 91.2 38.8 61.1 80.3 76.8 70.8GT w/ boxes97.4 46.2 88.2 96.8 91.0 84.3\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"On the other hand, the point cloud represented by XYZ itself has great position expression ability, and the gain from Ablation comparison of BEV feature crop strategies.\",\n \"figure_data\": \"CropQuestion Types Exi. Cou. Obj. Sta. Com.AllCir. Box 84.0 21.8 52.3 60.2 65.2 58.8Rot. Box 84.8 20.8 52.3 59.8 70.0 59.5PoolingQuestion Types Exi. Cou. Obj. Sta. Com.AllMax84.2 20.7 51.6 58.0 69.7 58.9Mean84.8 20.8 52.3 59.8 70.0 59.5\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablation comparison of BEV feature pooling strategies.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"summarizes the differences between our NuScenes-QA and other related 3D VQA datasets. Firstly, to the bestof our knowledge, NuScenes-QA is currently the largest3D VQA datasets, with 34k diverse visual scenes and 460kquestion-answer pairs, averaging 13.5 pairs per scene. Incontrast, other datasets (Das et al. 2018; Ye et al. 2022;Azuma et al. 2022; Yan et al. 2021; Zhao et al. 2022; Maet al. 2023) typically have less than 1000 scenes due to thedifficulty of acquiring 3D data. Secondly, in terms of visualmodality, our NuScenes-QA is multimodal, comprising ofimages and point clouds, posing higher demands on visualreasoning while increasing the dataset's complexity. Addi-tionally, our visual data is multi-frame, requiring temporalinformation mining. Thirdly, Our dataset focuses on outdoorscenes with a combination of dynamic objects and staticbackground, posing higher challenges for answering ques-tions that require perceiving and reasoning about dynamicobjects.In summary, compared to other 3D VQA datasets,NuScenes-QA stands out in terms of scale, data modality,and content, making it an important research resource thatcan advance the research and development of 3D visualquestion answering.\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Data Source\", \"Citation\": \"(Liu et al. 2023)\", \"Explanation\": \"The cited work by Liu et al. is a source of data for the autonomous driving system, providing information on 3D object detection that is used in the system.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Jiao et al. 2023b)\", \"Explanation\": \"The cited work by Jiao et al. is another data source for the autonomous driving system, providing information on 3D object detection that is utilized in the system.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Chen et al. 2023)\", \"Explanation\": \"The cited work by Chen et al. is a data source for the autonomous driving system, providing information on tracking that is incorporated into the system.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Antol et al. 2015)\", \"Explanation\": \"The cited work by Antol et al. is a data source for the autonomous driving system, providing information on visual question answering that is utilized in the system.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Goyal et al.)\", \"Explanation\": \"The cited work by Goyal et al. is a data source for the autonomous driving system, providing information on existing VQA datasets that the system is trained on.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Hudson and Manning 2019)\", \"Explanation\": \"The cited work provides a method for understanding the complexities of autonomous driving scenario, which the citing paper adopts to address the limitations in existing VQA benchmarks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Azuma et al. 2022)\", \"Explanation\": \"The cited work, 3D-QA, focuses on understanding the structure and spatial relationships of objects in a self-driving scenario, which the citing paper further extends to address the limitations in existing VQA benchmarks.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Caesar et al. 2020)\", \"Explanation\": \"The cited work, nuScenes, is the data source for the construction of the NuScenes-QA benchmark, providing the keyframes and annotations used in the development of the benchmark.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Huang et al. 2021)\", \"Explanation\": \"The cited work provides the nuScenes training and validation data split, which the citing paper uses to obtain a total of 460K question-answer pairs for training and testing.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Yin, Zhou, and Krahenbuhl 2021)\", \"Explanation\": \"The cited work contributes the point cloud data to the citing paper, which is used to extract visual features and provide object proposals for the question answering model.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Jiao et al. 2023a)\", \"Explanation\": \"The cited work provides the image-based models and visual question answering techniques that the citing paper uses to develop baseline models for the question answering task.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Anderson et al. 2018)\", \"Explanation\": \"The cited work contributes the image-based models and question answering techniques that the citing paper uses to develop baseline models for the question answering task.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Yu et al. 2019)\", \"Explanation\": \"The cited work provides the point cloud-based models and question answering techniques that the citing paper uses to develop baseline models for the question answering task.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Goyal et al. 2017)\", \"Explanation\": \"The cited work VQA2.0 is a dataset that the citing paper uses as a source of data for their research on visual question answering.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Johnson et al. 2017)\", \"Explanation\": \"The cited work CLEVR is another dataset that the citing paper uses as a data source for their research on visual question answering.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Hudson and Manning 2019)\", \"Explanation\": \"The cited work GQA is yet another dataset that the citing paper uses as a data source for their research on visual question answering.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Jang et al. 2017)\", \"Explanation\": \"The cited work TGIF-QA is a video-based dataset that the citing paper uses as a data source for their research on visual question answering.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Lei et al. 2018)\", \"Explanation\": \"The cited work TVQA is another video-based dataset that the citing paper uses as a data source for their research on visual question answering.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ye et al. 2022)\", \"Explanation\": \"The cited work, 3DQA dataset, serves as a foundational data source for the citing paper in the field of 3D Visual Question Answering (3D-QA), providing a dataset of question-answer pairs for indoor scenes based on ScanNet.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Azuma et al. 2022)\", \"Explanation\": \"The cited work, ScanQA dataset, is also a data source for the citing paper in the field of 3D-QA, providing a larger dataset of question-answer pairs for indoor scenes that has been generated using a question generation model and manual editing.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Dai et al. 2017)\", \"Explanation\": \"The cited work, ScanNet dataset, serves as a methodological basis for the construction of the 3DQA and ScanQA datasets in the field of 3D-QA, as it provides the visual data used in the datasets for indoor scenes.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Johnson et al. 2015)\", \"Explanation\": \"The cited work provides the definition of a scene graph, which is used as a foundational element in the data construction pipeline for the question-answer pairs generation in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Pennington, Socher, and Manning 2014)\", \"Explanation\": \"The cited work provides the pre-trained GloVe embeddings that are used in the word-level context encoding in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Hochreiter and Schmidhuber 1997)\", \"Explanation\": \"The cited work introduces the biLSTM model, which is used in the word-level context encoding in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(He et al. 2016)\", \"Explanation\": \"The cited work introduces the ResNet with FPN backbone for multi-scale feature extraction, which is used in the image stream branch of the visual feature extraction in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lin et al. 2017)\", \"Explanation\": \"The cited work introduces the FPN technique, which is used in the image stream branch of the visual feature extraction in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Philion and Fidler 2020)\", \"Explanation\": \"The cited work provides the view transformer method for estimating the depth of pixels in images and lifting them to 3D virtual points, which the citing paper adopts in their research to make the feature spatial-aware.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhou and Tuzel 2018)\", \"Explanation\": \"The cited work provides the method of partitioning 3D space into voxels and transforming raw point clouds into binary voxel grids, which the citing paper utilizes in their research to partition space and represent point clouds.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Graham, Engelcke, and Van Der Maaten 2018)\", \"Explanation\": \"The cited work provides the 3D sparse convolutional neural network method for feature representation in voxel grids, which the citing paper applies in their research to process the point cloud data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Yu et al. 2019)\", \"Explanation\": \"The cited work provides the pre-trained 3D detection model that the citing paper uses to extract object embeddings for training the answer head.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Antol et al. 2015)\", \"Explanation\": \"The cited work by Antol et al. provides a standard evaluation metric for VQA works, which the citing paper adopts to measure the performance of their research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Azuma et al. 2022)\", \"Explanation\": \"The cited work by Azuma et al. is also referenced in the context of evaluating the performance of VQA works, further supporting the use of the Top-1 accuracy as the evaluation metric.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Huang et al. 2021)\", \"Explanation\": \"The cited work provides the pre-trained detection model that the citing paper uses as a feature extraction backbone for their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Yin, Zhou, and Krahenbuhl 2021)\", \"Explanation\": \"The cited work provides the pre-trained detection model that the citing paper uses as a feature extraction backbone for their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Jiao et al. 2023a)\", \"Explanation\": \"The cited work provides the pre-trained detection model that the citing paper uses as a feature extraction backbone for their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Anderson et al. 2018)\", \"Explanation\": \"The cited work, BUTD, serves as a methodological basis for the citing paper by advocating for the use of bottom-up and top-down attention on salient regions of the image in the question-answering task.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Yu et al. 2019)\", \"Explanation\": \"The cited work, MCAN, also serves as a methodological basis for the citing paper by providing a method for computing attention on salient regions of the image in the question-answering task.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Das et al. 2018)\", \"Explanation\": \"The cited work provides a dataset of image indoor templates with a large number of samples (767) to support the study of EQA in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Ye et al. 2022)\", \"Explanation\": \"The cited work contributes a point cloud indoor human dataset with a large number of samples (10k) to support the study of 3DQA in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Azuma et al. 2022)\", \"Explanation\": \"The cited work provides a point cloud indoor auto + human dataset with a large number of samples (41k) to support the study of ScanQA in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Yan et al. 2021)\", \"Explanation\": \"The cited work contributes a point cloud indoor template dataset with a large number of samples (1,129) to support the study of CLEVR3D in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Zhao et al. 2022)\", \"Explanation\": \"The cited work provides a point cloud indoor human dataset with a large number of samples (20k) to support the study of FE-3DGQA in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Ma et al. 2023)\", \"Explanation\": \"The cited work contributes an image dataset with a large number of samples (60.1k) to support the study of SQA3D in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Azuma et al. 2022)\", \"Explanation\": \"The cited work, ScanQA, is mentioned as a potential extension to the NuScenes-QA dataset in the future. The citing paper suggests that object localization tasks can be added to the dataset to increase the model's interpretability and improve the data quality.\"}]"}}},{"rowIdx":80,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b31","b8","b4","b18","b32","b34","b33","b30","b7","b3","b36","b5","b3","b15","b3","b26","b14","b37","b21","b37","b21","b26","b17","b26","b37","b21"],"table_ref":[],"text":"Transformer as an emerging model for natural language processing [31] has attracted much attention in computer vision. So far, a couple of vision Transformers have been proposed and made tremendous success in promising superior performance in a variety of applications compared with convolution neural network based deep learning frameworks [7,29,3,1,18,32,13,23,34]. At the same time, a major problem arises: The heavy computational load prevents * * Corresponding author: suyang@fudan.edu.cn such models from being applied to edge computing-based applications. Therefore, a recent trend has been shifted to develop light weighted models of vision Transformer (ViT). It is known that self-attention is the major bottleneck to incur dense computations in a Transformer as it requires permutation to couple tokens. Accordingly, the recent efforts were devoted to the following trials: (1) Enforce the self-attention to be confined in a neighborhood around each token such that fewer tokens will be involved in updating each token. The methods falling in this category include Swin Transformer [23], Pale Transformer [33], HaloNet [30], and CSWin Transformer [6]. These methods are based on such an assumption that tokens spatially far away are not semantically correlated, but this does not always hold true. Moreover, since the neighborhood to confine selfattention is predefined, not machine learning based, it may sometimes not be coherent to practice. (2) Another solution aims to modify the self-attention operations internally [36,4,2,15]. By changing the computing order in selfattention while incorporating the combination of multiple heads into the self-attention, the complexity of Hydra Attention [2] could be made relatively low provided no nonlinear component is contained in the self-attention module, which is a strong constraint to prevent such a solution from being applied broadly. (3) On account of the O(N 2 d) complexity of self-attention, where N is the token number and d the feature dimension, a straightforward way is to reduce the number of tokens fed to self-attention instead of the effort to modify self-attention itself. One methodology is to group similar tokens into clusters via unsupervised learning and let each cluster act as a higher-level abstractive representation to take part in the self-attention [39,20]. Here, the difficulty lies in the quality control of clustering, which may lead to not semantically meaningful representations, and thus affect the final decision negatively. The other kind of solution aims to reduce the token number by applying tokens filter explicitly or based on certain heuristics. In [26], a couple of token filters realized using multi-layer perceptron (MLP) are incorporated into some middle layers of ViT as gating functions, which are trained end-to-end with the backbone [29,14], such that the tokens resulting from one self-attention layer can be selectively forwarded to the subsequent self-attention layers. In [37], an early stop criterion based on the accumulated token value at the first dimension is proposed. In [21], token importance is assumed to be its attentive weight correlated to class token. However, the complex coupling layer by layer brings in uncertainty to the attentive weights in terms of correlating to class token, so gradual token filtering has to be applied while the less attentive tokens are also preserved to aid further testing.\nIn sum, these token filtering methods miss to address the following issue: They are based on heuristics [37,21] or enclosed in the end-to-end training with backbone [26], so the rationality of discarding some tokens selectively is not straightforward. In other words, due to the heuristic and less explainable nature of these methods, they are unable to foresee the impact of a token on the final decision explicitly. Therefore it is impossible for them to filter out all irrelevant tokens from the very beginning and token filtering has to be done gradually in a layer-wise manner, which results in unpredictable token filtering on the fly, not favored by parallel computing.\nThis study aims to solve the aforementioned problem by proposing a ranking method to measure how relevant a token is in regard to the final decision. Based on such a measure, then, we proceed to train a binary classifier as a token filter with learnable parameters generalized from the whole training corpus, such that we can filter out irrelevant tokens from the very beginning prior to self-attention. For this sake, we propose a measure referred to as delta loss (DL) to evaluate how much the loss changes once masking the token of interest, where the naive Transformer can act as the agent to score the difference of loss caused by with or without a token of interest. The mechanism is similar to a wrapper in the sense of classical feature selection [17]. Then, we label the tokens resulting in big DL values as positive instances since masking them will have a significant impact on the final decision. Further, we train a MLP based binary classifier using the labeled tokens based on their DL values. Finally, we apply such a token filter on each token, prior to all the subsequent Transformer blocks, and fine tune the whole pipeline end-to-end. As a result, the irrelevant tokens can be discarded from the very beginning, which is a one-pass process in contrast to reducing token numbers gradually [26,37,21].\nThe contribution of this work is as follows:\n(1) In the context of light weighted ViT, it is the first time that token filtering is proposed from a feature selection point of view to rank the relevance of each token in regard to the final decision. Hence, whether a token makes sense for the final decision becomes predictable from the very beginning, which can prevent irrelevant tokens from taking part in selfattention to the best extent. As a one-pass filter deployed at the very beginning prior to self-attention, it can lead to higher efficiency with even fewer token dropout compared with gradual token dropout throughout the pipeline.\n(2) We propose a new metric referred to as delta loss to weigh the importance of each token in terms of affecting the final decision and then force the token classifier to optimize its performance on the pseudo labels quantized from the DL values.\n(3) The only change compared to the original ViT is applying a MLP as the pre-filter for binary classification, which is fine turned with backbone, so the deployment is quite simple compared with the state-of-the-art (SOTA) methods, which rely on training from scratch.\n(4) The experiments show that the proposed method promises SOTA performance in terms of both precision and efficiency in an overall sense."},{"figure_ref":[],"heading":"Related works","publication_ref":["b31","b8","b4","b18","b6","b12","b28","b38","b24","b35","b40","b25","b9","b36","b5","b3","b3","b33","b30","b7","b26","b37","b21","b19","b27","b10","b22","b16","b17"],"table_ref":[],"text":"Vision Transformer. Transformer is initially applied in natural language processing (NLP) [31]. ViT [7] is the first work extending Transformer to computer vision by using no-overlapping image patches for image classification such that no convolution operation is needed. It shows comparable performance to convolution neural networks (CNN) on large-scale datasets. To perform well on various vision tasks, however, ViT and some of its following variants [3,18,1] require large-scale data and long training time for pre-training. DeiT [29] improved the training configuration by using a novel distillation method and proposed a Transformer architecture that can be trained only with Ima-geNet1K [5], whose performance is even better than ViT.\nEfficient Transformer. Although Transformer has recently led to great success in computer vision, it suffers from dense computations arising from self-attention, which is also the major mechanism to grant the promising performance in various down-streaming applications. Therefore, recent efforts are focused on proposing various methods to reduce the self-attention caused by dense computations. Provided there are N tokens of d dimension corresponding with the image patches, the self-attention to correlate every couple from the permutation of the N tokens will result in O(N 2 d) complexity in a simple updating round. For deploying Transformer on edge devices, a variety of simplified models have been proposed, aiming to reduce parameters and operations, for example, parameter pruning [12,28], low-rank factorization [38], and knowledge distillation [24,35]. Yet, these strategies for acceleration are limited in that they still rely on CNN, which deviates from the original design of Transformer, that is, facilitating deep learning with a new working mechanism other than CNN.\nOne way for rendering light weighted vision Transformer is to simplify the layers of Transformer [40,25,8], but its benefit is limited since the major complexity arises from self-attention, not layer stack. So, some other efforts are focused on altering the internal operations of Transformer to make self-attention more efficient [36,4,2]. As for Hydra Attention [2], the computing order insider selfattention is reorganized while the conbination of multiple heads is incorporated into self-attention to reduce the complexity. Nevertheless, it is workable only when no nonlinear component such as SoftMax is applied in self-attention, which limits its applications. Some other methods try to alleviate the computations of self-attention by reducing the number of tokens. One way is to enforce the computation of self-attention to be conducted in a predefined local region, for instance, Swin Transformer [23], Pale Transformer [33], HaloNet [30], and CSWin Transformer [6]. These methods are based on the assumption that image patches located far from each other are not semantically relevant, but this only partially holds true. Besides, since determining the local context does not rely on machine learning, it cannot be adaptive to various real scenarios end-to-end. Another solution is grouping similar tokens together to obtain more abstractive sparse token representations from clustering. The self-attention confined to such highly abstractive representations can thus be made efficient. TCFormer [39] fuses the tokens in the same cluster into a new one utilizing a weighted average, and the tokens involved in self-attention can then be reduced layer by layer. When tackling high-resolution images, Liang et al. [20] leverage clustering in the first few layers to reduce the number of tokens and reconstruct them in the last few layers. Thus, the dense computations on self-attention can be avoided in the middle layers. The limit for the clustering based methods is: They simply merge similar tokens but ignore the quality control of token clustering in case some clusters might be spanned by less homogeneous tokens.\nSince the aforementioned approaches suffer from hard quality control or lack of machine learning, this gives rise to another methodology, which aims to filter out tokens gradually throughout the pipeline of ViT. Dynamic ViT [26] incorporates a couple of learnable neural networks to the middle layers of ViT as the gating structure to make tokens gradually sparser throughout a relatively long course. A-ViT [37] calculates the accumulated halting probability of each token by using the feature values resulting from each Transformer layer, which gradually reduces the number of tokens without adding any additional modules, but could result in suddenly halted computing on a token, in general, not favored when scheduling parallel computing. E-ViT [21] assumes that top-k attentive weights correspond with relevant tokens but it still preserves irrelevant tokens throughout the whole pipeline to undergo a gradual token dropout procedure. The reason is: Token impact cannot be related to final decision in an explicit way due to the complex interlayer coupling between tokens when back tracing each to-ken's correlation to class token. Besides, every trail of the hyper parameter k in preserving selectively the top-k attentive tokens will lead to a new-round training from scratch. A common limit of the aforementioned approaches is: All such works rely on the running results of the backbone for token filtering, as it is impossible for them to foresee the token-caused effect on the final decision from the very beginning. In view of such a limit, we propose a new method from a feature selection point of view to conduct token filtering from the very beginning prior to self-attention to filter out truly irrelevant tokens.\nFeature selection. In the literature on deep learning, Le et al. [19] proposed a feature selector by adding a sparse one-to-one linear layer. It directly uses network weight as the feature weight, so it is sensitive to noise. Roy et al. [27] used the activation potential as a measure for feature selection at each single input dimension but is limited to specific DNNs. Since then, the interest has been turned to the data with a specific structure, which relies more on the progress of traditional data feature selection methods [10,22]. AFS [11] proposes to transform feature weight generation into a mode that can be solved by using an attention mechanism. Takumi et al. [16] proposed a method that harnesses feature partition in SoftMax loss function for effectively learning the discriminative features. However, these methods are focused on reducing feature map or selecting channels of CNN rather than Transformer. We are the first to use the delta loss value as an indicator for identifying relevant Transformer tokens from a wrapper-based feature selection point of view [17] by testing their impact on the final decision once masked."},{"figure_ref":[],"heading":"DL-ViT","publication_ref":["b12"],"table_ref":[],"text":"We propose a metric referred to as delta loss to weigh how vital a token is. In detail, we mask a token at first and then compute its impact on the loss, say, the change of cross entropy with/without such token for the final decision. If masking a token leads to big DL, it means that such a token does affect the final decision much, which should be preserved to take part in the subsequent self-attentions. Vice versa, if the loss does not change much with/without a token, such a token should be discarded due to its less importance to the decision. Correspondingly, a plausible trick arising from the aforementioned scheme is: The Transformer itself can act as the agent to score the importance of each token via DL without any further machine learning required in this phase. By using the DL scores to label the tokens in the training corpus as positive or negative, we can then train a binary classifier to check whether the tokens of an input image should be preserved to take part in the subsequent self-attentions, where the classifier is implemented using MLP. Finally, we preset the simple MLP module prior to the backbone Transformer, and fine tune the whole pipeline, where preset a token filter as such is the only change in the architecture.\nIn the following, we describe the two phases of the delta loss based efficient vision Transformer (DL-ViT): Evaluating token importance with delta loss to train the token filter and then fine tuning the entire network after incorporating the token filter. After an image passes through the embedding layer of the vision Transformer, the non-overlapping image patches are encoded into tokens denoted as:\nX = x i ∈ R d |i = 1, 2. . . , N ,(1)\nwhere N is the total number of tokens and d the embedding dimension. After masking the i-th token x i ∈ R d , we get the tokens in the following form:\nX i = {x 1 , . . . x i-1 , ∅, x i+1 , . . . , x N } ,(2)\nwhere ∅ means replacing the i-th token with zeros (masking). Then, we feed X and X i to the Transformer, respectively, to obtain the corresponding prediction results:\nŷ = T ransf ormer(X),(3)\nŷi = T ransf ormer(X i ).(4)\nBased on the previous prediction results, we calculate the cross-entropy loss of either case in reference to the ground truth y as follows:\nL = CrossEntropy(ŷ, y),(5)\nL i = CrossEntropy(ŷ i , y).(6)\nIt is known that the value of loss measures how close the prediction result approaches the ground truth, where a lower value corresponds with closer to the ground truth. Let:\n∆L i = L -L i .(7)\nObviously, if the delta loss defined in Eq. ( 7) is positive, it means that masking the i-th token makes the decision closer to the ground truth since masking as such causes a lower cross-entropy value in contrast to the original case. In such a case, discarding the token should not affect but benefit the decision of ViT, and a bigger delta loss corresponds with a better change on the decision. So, we quantize the delta loss measure to mark whether the current token should be discarded or not, formulated as:\nlabel(x i ) = 0, L -L i ≤ ρ 1, L -L i > ρ(8)\nwhere 0 means leaving the token out, 1 preserving the token, and ρ the only hyperparameter to control the significance of the pseudo labeling.\nAfter labeling all the tokens in the training corpus, we can then proceed to learn the generalizable law to distinguish positive token examples from negative ones in a population sense, which leads to a binary classifier realized using MLP for token filtering, acting to determine whether each token should be preserved to the next phase of the pipeline or not. So far, there is still a critical problem to be tackled, that is, some similar tokens may lead to contradicting results in terms of delta loss. This is quite common when two images share some similar patches locally but are quite different in an overall sense. Such semantically ambiguous local patches impose difficulty on token filter training, so we attach the profile featuring the whole image to each token as context, namely, global feature, to solve this problem. That is, we not only use the tokens with original embedding but also apply adaptive average pooling (AAP) over all tokens of an image to obtain the global feature of the image, acting as the context to make each token distinguishable from the others. Thus, the overall descriptor for each token becomes:\nx ′ i = [x i , x global ]. (9) x global = AAP (X) = 1 N Σ N k=1 x k .(10)\nConsequently, x ′ i instead of x is fed to the token selection module for training and inference:\np i = Sigmoid(M LP (x ′ i )). (11\n)\nDuring training, we first fix all parameters of the pretrained backbone Transformer for token labeling, and then, train the MLP only. Here, we use binary-cross-entropy loss to train the network: (12) where p i is the prediction from MLP, and label(x i ) the pseudo label calculated from Eq. ( 8). Label = Label ∪ label(x i ) 15: end for 16: return Label As shown in Fig. 2, before entering the Transformer, all tokens must go through the token selection module that will output the decision of keeping or discarding the token. During fine tuning, we train both the token selection module and the DeiT end-to-end, based on the cross-entropy loss: x global = AAP (X)\nL M LP = BinaryCrossEntropy(p i , label(x i )),\nL f inetune = CrossEntropy(ŷ, y),(13)\n5:\nfor i = 1, ..., N do 6:\nx\n′ i = [x i , x global ]\n7:\np i = Sigmoid(M LP (x ′ i ))\n8:\nL M LP = BinaryCrossEntropy(p i , label(x i ))\n9:\nBack-propagation updating W 10:\nend for 11: until no more descent on L M LP 12: return W Figure 2. The overall fine tuning and inference process of the proposed approach. All the tokens enter the selection module in turn to decide whether they should be passed to the subsequent pipeline of Transformer according to the predicted probability, after which the number of the preserve tokens will remain unchanged in the rest pipeline.\nwhere y is the ground truth and ŷ the output of the whole network.\nDuring fine tuning, in order to make it easy to parallelize the computation, we do not delete tokens directly but replace them with zeros to prevent them from affecting subsequent operations. Such a token masking strategy makes the computational cost of the training iterations similar to those of the original vision Transformer. During inferring, we throw the masked tokens out of the subsequent calculations in order to examine the actual acceleration resulting from the token selection mechanism."},{"figure_ref":["fig_3"],"heading":"Experiments","publication_ref":["b6"],"table_ref":[],"text":"Data: We evaluate our method for image classification on the 1000-class ImageNet1K ILSVRC 2012 dataset [5], Distribution of DL values: In order to examine the rationality of our method intuitively, we visualize the distribution of all DL values in Fig. 3. We find that small DL values around 0 dominate the majority of the distribution, which reveals the fact that only a part of the image patches are significantly relevant to classification. So considerable tokens with smaller DL values can be discarded."},{"figure_ref":["fig_4","fig_5","fig_6"],"heading":"Intuitive insight from statistics","publication_ref":[],"table_ref":[],"text":"Fig. 4 depicts the average DL value of the tokens at each patch resulting from DeiT-T [29] on the training set of Ima-geNet1K. We find that most semantically important patches are on the center of an image, the rest of which should fall into the outliers to be eliminated more frequently by our method.\nQualitative analysis. Fig. 5 visualizes the masks on the ImageNet1K validation set resulting from DL-ViT, where the dark portion appearing mostly along the edge of an image are the less contributive patches for classification, namely, the outliers favored by the token filter to activate elimination. Sometimes, the token selection module eliminates not only the background of the image, but also the confusing portion that may cause classification errors. For example, the third image in the last row of Fig. 6 prefers eliminating the patches unrelated to the dog."},{"figure_ref":["fig_3","fig_7","fig_8"],"heading":"Comparison to baselines","publication_ref":["b37","b26"],"table_ref":[],"text":"We compare our method with the baselines in Table 1 in terms of efficiency and precision, where we set ρ to 0.002 and 0.001 for DL-ViT-T and DL-ViT-S, respectively. At the cost of sacrificing only 0.3% and 0.2% accuracy compared with that of the backbone, we cut down 46% and 15% FLOPs of DeiT-T and DeiT-S, respectively. Moreover our method performs best to make DeiT-T more efficient and more precise compared with the baselines, where the Floating-point Operations (FLOPs) metric is measured by FlopCountAnalysis [9].\nAs E-ViT is an exception that only reports comparison on ViT-S, we follow A-ViT [37] and Dynamic-ViT [26] to report the performance on both ViT-S and ViT-T. Regarding ViT-S, no method performs best on all metrics, where E-ViT runs faster but is inferior to DL-ViT on top-1 precision. Except for the highest top-1 precision on both benchmarks, DL-ViT promises the state-of-the-art (SOTA) performance in an overall sense if taking into account both benchmarks. Since E-ViT misses to compare with all the baselines on ViT-T, except for the performance, we compare it with DL-ViT in a methodological sense to allow a more comprehensive insight: (1) We evaluate token importance via delta loss while E-ViT leverages top-k attention weights as token importance; (2) We filter out irrelevant tokens from the very beginning but E-ViT does this gradually and preserve both important and less important tokens in the whole pipeline. That is, E-ViT cannot foresee the impact of each token on the final decision at the beginning but DL-ViT can. (3) E-ViT modifies the self-attention, and the whole pipeline has to be changed wherever attention is applied, so it has to train from scratch, which is too expensive compared with the fine tune as adopted in our framework. As we change nothing in ViT, the pseudo labeling is performed by using naive ViT without any training. Besides, MLP is a two-class classifier, whose training is not tough. In this sense, the change on the architecture is minor. (4) In DL-ViT, ρ controls the significance of pseudo labeling, where the heuristics to choose its value lies in the statistics of DL values as shown in Fig. 3. For E-ViT, determining k is not easy in that there is no explicit heuristic to foresee its impact on the overall performance, and every trail will lead to a new-round computation-intensive training from scratch. Besides, the layer-varying token importance accounts for why layer-wise token dropout has to be done gradually. In Fig. 7 and Fig. 8, we compare DL-ViT with the baselines under different settings. It is obvious that our model can achieve a good trade-off between efficiency and precision.\nIn addition to FLOPs, we also evaluate the image throughput of our model on a single NVIDIA RTX 3090 GPU with batch size fixed to 64, and for GPU warming up, 512 forward passes are conducted. The experiment demonstrates that our DL-ViT can accelerate the inference by 15% ∼ 41%."},{"figure_ref":[],"heading":"Ablation study","publication_ref":[],"table_ref":["tab_0","tab_1"],"text":"In our method, the full configuration of a solution is subject to the following factors: The backbone for token importance evaluation, the threshold ρ to control the annotation on DL values, MLP, and the local/global feature applied to it. As shown in Table 2, a high threshold value of ρ can filter out more tokens, resulting in higher efficiency, but a too high one will cause degradation in precision. So, there is a compromise to determine the value of ρ, where we let ρ = 0.002 for DL-ViT-T. Note that when ρ = 0.002, the accuracy of using only local features is even higher than that of DeiT-T [29], at the cost of sacrificing FLOPs. Yet, we incorporate global feature as our primary solution due to its promising overall performance. Note that both cases lead to varying accuracy when ρ changes from 0.001 to 0.003, but the FLOPs with global feature remain stably low. Table 3 shows that the proposed model degrades in terms of precision if replacing the pre-training of MLP with the random initialization, and the backbone based on random token filtering also leads to inferior performance. This indicates that our token filtering scheme does contribute to making the DeiT-T efficient while preserving its precision to the best extent."},{"figure_ref":[],"heading":"Conclusions","publication_ref":[],"table_ref":[],"text":"We develop an efficient vision transformer with token impact prediction such that token filtering can be deployed at the very beginning prior to self-attention, where the back-bone Transformer is used as an agent/wrapper to rank the impact in terms of the difference of loss caused by masking a token of interest. It is the first time to develop a lightweighted model from a feature selection point of view with explicit insight into token's relevance to the decision. A MLP for token filtering is the only added module, which acts as a two-class classifier with minor change on the overall architecture, and its training is not tough. The present solution is a one-pass filter. In the future, we will investigate into the relevance of tokens at middle layers to the final decision to further improve the efficiency."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b31\",\n \"b8\",\n \"b4\",\n \"b18\",\n \"b32\",\n \"b34\",\n \"b33\",\n \"b30\",\n \"b7\",\n \"b3\",\n \"b36\",\n \"b5\",\n \"b3\",\n \"b15\",\n \"b3\",\n \"b26\",\n \"b14\",\n \"b37\",\n \"b21\",\n \"b37\",\n \"b21\",\n \"b26\",\n \"b17\",\n \"b26\",\n \"b37\",\n \"b21\"\n ],\n \"table_ref\": [],\n \"text\": \"Transformer as an emerging model for natural language processing [31] has attracted much attention in computer vision. So far, a couple of vision Transformers have been proposed and made tremendous success in promising superior performance in a variety of applications compared with convolution neural network based deep learning frameworks [7,29,3,1,18,32,13,23,34]. At the same time, a major problem arises: The heavy computational load prevents * * Corresponding author: suyang@fudan.edu.cn such models from being applied to edge computing-based applications. Therefore, a recent trend has been shifted to develop light weighted models of vision Transformer (ViT). It is known that self-attention is the major bottleneck to incur dense computations in a Transformer as it requires permutation to couple tokens. Accordingly, the recent efforts were devoted to the following trials: (1) Enforce the self-attention to be confined in a neighborhood around each token such that fewer tokens will be involved in updating each token. The methods falling in this category include Swin Transformer [23], Pale Transformer [33], HaloNet [30], and CSWin Transformer [6]. These methods are based on such an assumption that tokens spatially far away are not semantically correlated, but this does not always hold true. Moreover, since the neighborhood to confine selfattention is predefined, not machine learning based, it may sometimes not be coherent to practice. (2) Another solution aims to modify the self-attention operations internally [36,4,2,15]. By changing the computing order in selfattention while incorporating the combination of multiple heads into the self-attention, the complexity of Hydra Attention [2] could be made relatively low provided no nonlinear component is contained in the self-attention module, which is a strong constraint to prevent such a solution from being applied broadly. (3) On account of the O(N 2 d) complexity of self-attention, where N is the token number and d the feature dimension, a straightforward way is to reduce the number of tokens fed to self-attention instead of the effort to modify self-attention itself. One methodology is to group similar tokens into clusters via unsupervised learning and let each cluster act as a higher-level abstractive representation to take part in the self-attention [39,20]. Here, the difficulty lies in the quality control of clustering, which may lead to not semantically meaningful representations, and thus affect the final decision negatively. The other kind of solution aims to reduce the token number by applying tokens filter explicitly or based on certain heuristics. In [26], a couple of token filters realized using multi-layer perceptron (MLP) are incorporated into some middle layers of ViT as gating functions, which are trained end-to-end with the backbone [29,14], such that the tokens resulting from one self-attention layer can be selectively forwarded to the subsequent self-attention layers. In [37], an early stop criterion based on the accumulated token value at the first dimension is proposed. In [21], token importance is assumed to be its attentive weight correlated to class token. However, the complex coupling layer by layer brings in uncertainty to the attentive weights in terms of correlating to class token, so gradual token filtering has to be applied while the less attentive tokens are also preserved to aid further testing.\\nIn sum, these token filtering methods miss to address the following issue: They are based on heuristics [37,21] or enclosed in the end-to-end training with backbone [26], so the rationality of discarding some tokens selectively is not straightforward. In other words, due to the heuristic and less explainable nature of these methods, they are unable to foresee the impact of a token on the final decision explicitly. Therefore it is impossible for them to filter out all irrelevant tokens from the very beginning and token filtering has to be done gradually in a layer-wise manner, which results in unpredictable token filtering on the fly, not favored by parallel computing.\\nThis study aims to solve the aforementioned problem by proposing a ranking method to measure how relevant a token is in regard to the final decision. Based on such a measure, then, we proceed to train a binary classifier as a token filter with learnable parameters generalized from the whole training corpus, such that we can filter out irrelevant tokens from the very beginning prior to self-attention. For this sake, we propose a measure referred to as delta loss (DL) to evaluate how much the loss changes once masking the token of interest, where the naive Transformer can act as the agent to score the difference of loss caused by with or without a token of interest. The mechanism is similar to a wrapper in the sense of classical feature selection [17]. Then, we label the tokens resulting in big DL values as positive instances since masking them will have a significant impact on the final decision. Further, we train a MLP based binary classifier using the labeled tokens based on their DL values. Finally, we apply such a token filter on each token, prior to all the subsequent Transformer blocks, and fine tune the whole pipeline end-to-end. As a result, the irrelevant tokens can be discarded from the very beginning, which is a one-pass process in contrast to reducing token numbers gradually [26,37,21].\\nThe contribution of this work is as follows:\\n(1) In the context of light weighted ViT, it is the first time that token filtering is proposed from a feature selection point of view to rank the relevance of each token in regard to the final decision. Hence, whether a token makes sense for the final decision becomes predictable from the very beginning, which can prevent irrelevant tokens from taking part in selfattention to the best extent. As a one-pass filter deployed at the very beginning prior to self-attention, it can lead to higher efficiency with even fewer token dropout compared with gradual token dropout throughout the pipeline.\\n(2) We propose a new metric referred to as delta loss to weigh the importance of each token in terms of affecting the final decision and then force the token classifier to optimize its performance on the pseudo labels quantized from the DL values.\\n(3) The only change compared to the original ViT is applying a MLP as the pre-filter for binary classification, which is fine turned with backbone, so the deployment is quite simple compared with the state-of-the-art (SOTA) methods, which rely on training from scratch.\\n(4) The experiments show that the proposed method promises SOTA performance in terms of both precision and efficiency in an overall sense.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related works\",\n \"publication_ref\": [\n \"b31\",\n \"b8\",\n \"b4\",\n \"b18\",\n \"b6\",\n \"b12\",\n \"b28\",\n \"b38\",\n \"b24\",\n \"b35\",\n \"b40\",\n \"b25\",\n \"b9\",\n \"b36\",\n \"b5\",\n \"b3\",\n \"b3\",\n \"b33\",\n \"b30\",\n \"b7\",\n \"b26\",\n \"b37\",\n \"b21\",\n \"b19\",\n \"b27\",\n \"b10\",\n \"b22\",\n \"b16\",\n \"b17\"\n ],\n \"table_ref\": [],\n \"text\": \"Vision Transformer. Transformer is initially applied in natural language processing (NLP) [31]. ViT [7] is the first work extending Transformer to computer vision by using no-overlapping image patches for image classification such that no convolution operation is needed. It shows comparable performance to convolution neural networks (CNN) on large-scale datasets. To perform well on various vision tasks, however, ViT and some of its following variants [3,18,1] require large-scale data and long training time for pre-training. DeiT [29] improved the training configuration by using a novel distillation method and proposed a Transformer architecture that can be trained only with Ima-geNet1K [5], whose performance is even better than ViT.\\nEfficient Transformer. Although Transformer has recently led to great success in computer vision, it suffers from dense computations arising from self-attention, which is also the major mechanism to grant the promising performance in various down-streaming applications. Therefore, recent efforts are focused on proposing various methods to reduce the self-attention caused by dense computations. Provided there are N tokens of d dimension corresponding with the image patches, the self-attention to correlate every couple from the permutation of the N tokens will result in O(N 2 d) complexity in a simple updating round. For deploying Transformer on edge devices, a variety of simplified models have been proposed, aiming to reduce parameters and operations, for example, parameter pruning [12,28], low-rank factorization [38], and knowledge distillation [24,35]. Yet, these strategies for acceleration are limited in that they still rely on CNN, which deviates from the original design of Transformer, that is, facilitating deep learning with a new working mechanism other than CNN.\\nOne way for rendering light weighted vision Transformer is to simplify the layers of Transformer [40,25,8], but its benefit is limited since the major complexity arises from self-attention, not layer stack. So, some other efforts are focused on altering the internal operations of Transformer to make self-attention more efficient [36,4,2]. As for Hydra Attention [2], the computing order insider selfattention is reorganized while the conbination of multiple heads is incorporated into self-attention to reduce the complexity. Nevertheless, it is workable only when no nonlinear component such as SoftMax is applied in self-attention, which limits its applications. Some other methods try to alleviate the computations of self-attention by reducing the number of tokens. One way is to enforce the computation of self-attention to be conducted in a predefined local region, for instance, Swin Transformer [23], Pale Transformer [33], HaloNet [30], and CSWin Transformer [6]. These methods are based on the assumption that image patches located far from each other are not semantically relevant, but this only partially holds true. Besides, since determining the local context does not rely on machine learning, it cannot be adaptive to various real scenarios end-to-end. Another solution is grouping similar tokens together to obtain more abstractive sparse token representations from clustering. The self-attention confined to such highly abstractive representations can thus be made efficient. TCFormer [39] fuses the tokens in the same cluster into a new one utilizing a weighted average, and the tokens involved in self-attention can then be reduced layer by layer. When tackling high-resolution images, Liang et al. [20] leverage clustering in the first few layers to reduce the number of tokens and reconstruct them in the last few layers. Thus, the dense computations on self-attention can be avoided in the middle layers. The limit for the clustering based methods is: They simply merge similar tokens but ignore the quality control of token clustering in case some clusters might be spanned by less homogeneous tokens.\\nSince the aforementioned approaches suffer from hard quality control or lack of machine learning, this gives rise to another methodology, which aims to filter out tokens gradually throughout the pipeline of ViT. Dynamic ViT [26] incorporates a couple of learnable neural networks to the middle layers of ViT as the gating structure to make tokens gradually sparser throughout a relatively long course. A-ViT [37] calculates the accumulated halting probability of each token by using the feature values resulting from each Transformer layer, which gradually reduces the number of tokens without adding any additional modules, but could result in suddenly halted computing on a token, in general, not favored when scheduling parallel computing. E-ViT [21] assumes that top-k attentive weights correspond with relevant tokens but it still preserves irrelevant tokens throughout the whole pipeline to undergo a gradual token dropout procedure. The reason is: Token impact cannot be related to final decision in an explicit way due to the complex interlayer coupling between tokens when back tracing each to-ken's correlation to class token. Besides, every trail of the hyper parameter k in preserving selectively the top-k attentive tokens will lead to a new-round training from scratch. A common limit of the aforementioned approaches is: All such works rely on the running results of the backbone for token filtering, as it is impossible for them to foresee the token-caused effect on the final decision from the very beginning. In view of such a limit, we propose a new method from a feature selection point of view to conduct token filtering from the very beginning prior to self-attention to filter out truly irrelevant tokens.\\nFeature selection. In the literature on deep learning, Le et al. [19] proposed a feature selector by adding a sparse one-to-one linear layer. It directly uses network weight as the feature weight, so it is sensitive to noise. Roy et al. [27] used the activation potential as a measure for feature selection at each single input dimension but is limited to specific DNNs. Since then, the interest has been turned to the data with a specific structure, which relies more on the progress of traditional data feature selection methods [10,22]. AFS [11] proposes to transform feature weight generation into a mode that can be solved by using an attention mechanism. Takumi et al. [16] proposed a method that harnesses feature partition in SoftMax loss function for effectively learning the discriminative features. However, these methods are focused on reducing feature map or selecting channels of CNN rather than Transformer. We are the first to use the delta loss value as an indicator for identifying relevant Transformer tokens from a wrapper-based feature selection point of view [17] by testing their impact on the final decision once masked.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"DL-ViT\",\n \"publication_ref\": [\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"We propose a metric referred to as delta loss to weigh how vital a token is. In detail, we mask a token at first and then compute its impact on the loss, say, the change of cross entropy with/without such token for the final decision. If masking a token leads to big DL, it means that such a token does affect the final decision much, which should be preserved to take part in the subsequent self-attentions. Vice versa, if the loss does not change much with/without a token, such a token should be discarded due to its less importance to the decision. Correspondingly, a plausible trick arising from the aforementioned scheme is: The Transformer itself can act as the agent to score the importance of each token via DL without any further machine learning required in this phase. By using the DL scores to label the tokens in the training corpus as positive or negative, we can then train a binary classifier to check whether the tokens of an input image should be preserved to take part in the subsequent self-attentions, where the classifier is implemented using MLP. Finally, we preset the simple MLP module prior to the backbone Transformer, and fine tune the whole pipeline, where preset a token filter as such is the only change in the architecture.\\nIn the following, we describe the two phases of the delta loss based efficient vision Transformer (DL-ViT): Evaluating token importance with delta loss to train the token filter and then fine tuning the entire network after incorporating the token filter. After an image passes through the embedding layer of the vision Transformer, the non-overlapping image patches are encoded into tokens denoted as:\\nX = x i ∈ R d |i = 1, 2. . . , N ,(1)\\nwhere N is the total number of tokens and d the embedding dimension. After masking the i-th token x i ∈ R d , we get the tokens in the following form:\\nX i = {x 1 , . . . x i-1 , ∅, x i+1 , . . . , x N } ,(2)\\nwhere ∅ means replacing the i-th token with zeros (masking). Then, we feed X and X i to the Transformer, respectively, to obtain the corresponding prediction results:\\nŷ = T ransf ormer(X),(3)\\nŷi = T ransf ormer(X i ).(4)\\nBased on the previous prediction results, we calculate the cross-entropy loss of either case in reference to the ground truth y as follows:\\nL = CrossEntropy(ŷ, y),(5)\\nL i = CrossEntropy(ŷ i , y).(6)\\nIt is known that the value of loss measures how close the prediction result approaches the ground truth, where a lower value corresponds with closer to the ground truth. Let:\\n∆L i = L -L i .(7)\\nObviously, if the delta loss defined in Eq. ( 7) is positive, it means that masking the i-th token makes the decision closer to the ground truth since masking as such causes a lower cross-entropy value in contrast to the original case. In such a case, discarding the token should not affect but benefit the decision of ViT, and a bigger delta loss corresponds with a better change on the decision. So, we quantize the delta loss measure to mark whether the current token should be discarded or not, formulated as:\\nlabel(x i ) = 0, L -L i ≤ ρ 1, L -L i > ρ(8)\\nwhere 0 means leaving the token out, 1 preserving the token, and ρ the only hyperparameter to control the significance of the pseudo labeling.\\nAfter labeling all the tokens in the training corpus, we can then proceed to learn the generalizable law to distinguish positive token examples from negative ones in a population sense, which leads to a binary classifier realized using MLP for token filtering, acting to determine whether each token should be preserved to the next phase of the pipeline or not. So far, there is still a critical problem to be tackled, that is, some similar tokens may lead to contradicting results in terms of delta loss. This is quite common when two images share some similar patches locally but are quite different in an overall sense. Such semantically ambiguous local patches impose difficulty on token filter training, so we attach the profile featuring the whole image to each token as context, namely, global feature, to solve this problem. That is, we not only use the tokens with original embedding but also apply adaptive average pooling (AAP) over all tokens of an image to obtain the global feature of the image, acting as the context to make each token distinguishable from the others. Thus, the overall descriptor for each token becomes:\\nx ′ i = [x i , x global ]. (9) x global = AAP (X) = 1 N Σ N k=1 x k .(10)\\nConsequently, x ′ i instead of x is fed to the token selection module for training and inference:\\np i = Sigmoid(M LP (x ′ i )). (11\\n)\\nDuring training, we first fix all parameters of the pretrained backbone Transformer for token labeling, and then, train the MLP only. Here, we use binary-cross-entropy loss to train the network: (12) where p i is the prediction from MLP, and label(x i ) the pseudo label calculated from Eq. ( 8). Label = Label ∪ label(x i ) 15: end for 16: return Label As shown in Fig. 2, before entering the Transformer, all tokens must go through the token selection module that will output the decision of keeping or discarding the token. During fine tuning, we train both the token selection module and the DeiT end-to-end, based on the cross-entropy loss: x global = AAP (X)\\nL M LP = BinaryCrossEntropy(p i , label(x i )),\\nL f inetune = CrossEntropy(ŷ, y),(13)\\n5:\\nfor i = 1, ..., N do 6:\\nx\\n′ i = [x i , x global ]\\n7:\\np i = Sigmoid(M LP (x ′ i ))\\n8:\\nL M LP = BinaryCrossEntropy(p i , label(x i ))\\n9:\\nBack-propagation updating W 10:\\nend for 11: until no more descent on L M LP 12: return W Figure 2. The overall fine tuning and inference process of the proposed approach. All the tokens enter the selection module in turn to decide whether they should be passed to the subsequent pipeline of Transformer according to the predicted probability, after which the number of the preserve tokens will remain unchanged in the rest pipeline.\\nwhere y is the ground truth and ŷ the output of the whole network.\\nDuring fine tuning, in order to make it easy to parallelize the computation, we do not delete tokens directly but replace them with zeros to prevent them from affecting subsequent operations. Such a token masking strategy makes the computational cost of the training iterations similar to those of the original vision Transformer. During inferring, we throw the masked tokens out of the subsequent calculations in order to examine the actual acceleration resulting from the token selection mechanism.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Experiments\",\n \"publication_ref\": [\n \"b6\"\n ],\n \"table_ref\": [],\n \"text\": \"Data: We evaluate our method for image classification on the 1000-class ImageNet1K ILSVRC 2012 dataset [5], Distribution of DL values: In order to examine the rationality of our method intuitively, we visualize the distribution of all DL values in Fig. 3. We find that small DL values around 0 dominate the majority of the distribution, which reveals the fact that only a part of the image patches are significantly relevant to classification. So considerable tokens with smaller DL values can be discarded.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\",\n \"fig_5\",\n \"fig_6\"\n ],\n \"heading\": \"Intuitive insight from statistics\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Fig. 4 depicts the average DL value of the tokens at each patch resulting from DeiT-T [29] on the training set of Ima-geNet1K. We find that most semantically important patches are on the center of an image, the rest of which should fall into the outliers to be eliminated more frequently by our method.\\nQualitative analysis. Fig. 5 visualizes the masks on the ImageNet1K validation set resulting from DL-ViT, where the dark portion appearing mostly along the edge of an image are the less contributive patches for classification, namely, the outliers favored by the token filter to activate elimination. Sometimes, the token selection module eliminates not only the background of the image, but also the confusing portion that may cause classification errors. For example, the third image in the last row of Fig. 6 prefers eliminating the patches unrelated to the dog.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\",\n \"fig_7\",\n \"fig_8\"\n ],\n \"heading\": \"Comparison to baselines\",\n \"publication_ref\": [\n \"b37\",\n \"b26\"\n ],\n \"table_ref\": [],\n \"text\": \"We compare our method with the baselines in Table 1 in terms of efficiency and precision, where we set ρ to 0.002 and 0.001 for DL-ViT-T and DL-ViT-S, respectively. At the cost of sacrificing only 0.3% and 0.2% accuracy compared with that of the backbone, we cut down 46% and 15% FLOPs of DeiT-T and DeiT-S, respectively. Moreover our method performs best to make DeiT-T more efficient and more precise compared with the baselines, where the Floating-point Operations (FLOPs) metric is measured by FlopCountAnalysis [9].\\nAs E-ViT is an exception that only reports comparison on ViT-S, we follow A-ViT [37] and Dynamic-ViT [26] to report the performance on both ViT-S and ViT-T. Regarding ViT-S, no method performs best on all metrics, where E-ViT runs faster but is inferior to DL-ViT on top-1 precision. Except for the highest top-1 precision on both benchmarks, DL-ViT promises the state-of-the-art (SOTA) performance in an overall sense if taking into account both benchmarks. Since E-ViT misses to compare with all the baselines on ViT-T, except for the performance, we compare it with DL-ViT in a methodological sense to allow a more comprehensive insight: (1) We evaluate token importance via delta loss while E-ViT leverages top-k attention weights as token importance; (2) We filter out irrelevant tokens from the very beginning but E-ViT does this gradually and preserve both important and less important tokens in the whole pipeline. That is, E-ViT cannot foresee the impact of each token on the final decision at the beginning but DL-ViT can. (3) E-ViT modifies the self-attention, and the whole pipeline has to be changed wherever attention is applied, so it has to train from scratch, which is too expensive compared with the fine tune as adopted in our framework. As we change nothing in ViT, the pseudo labeling is performed by using naive ViT without any training. Besides, MLP is a two-class classifier, whose training is not tough. In this sense, the change on the architecture is minor. (4) In DL-ViT, ρ controls the significance of pseudo labeling, where the heuristics to choose its value lies in the statistics of DL values as shown in Fig. 3. For E-ViT, determining k is not easy in that there is no explicit heuristic to foresee its impact on the overall performance, and every trail will lead to a new-round computation-intensive training from scratch. Besides, the layer-varying token importance accounts for why layer-wise token dropout has to be done gradually. In Fig. 7 and Fig. 8, we compare DL-ViT with the baselines under different settings. It is obvious that our model can achieve a good trade-off between efficiency and precision.\\nIn addition to FLOPs, we also evaluate the image throughput of our model on a single NVIDIA RTX 3090 GPU with batch size fixed to 64, and for GPU warming up, 512 forward passes are conducted. The experiment demonstrates that our DL-ViT can accelerate the inference by 15% ∼ 41%.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ablation study\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_0\",\n \"tab_1\"\n ],\n \"text\": \"In our method, the full configuration of a solution is subject to the following factors: The backbone for token importance evaluation, the threshold ρ to control the annotation on DL values, MLP, and the local/global feature applied to it. As shown in Table 2, a high threshold value of ρ can filter out more tokens, resulting in higher efficiency, but a too high one will cause degradation in precision. So, there is a compromise to determine the value of ρ, where we let ρ = 0.002 for DL-ViT-T. Note that when ρ = 0.002, the accuracy of using only local features is even higher than that of DeiT-T [29], at the cost of sacrificing FLOPs. Yet, we incorporate global feature as our primary solution due to its promising overall performance. Note that both cases lead to varying accuracy when ρ changes from 0.001 to 0.003, but the FLOPs with global feature remain stably low. Table 3 shows that the proposed model degrades in terms of precision if replacing the pre-training of MLP with the random initialization, and the backbone based on random token filtering also leads to inferior performance. This indicates that our token filtering scheme does contribute to making the DeiT-T efficient while preserving its precision to the best extent.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We develop an efficient vision transformer with token impact prediction such that token filtering can be deployed at the very beginning prior to self-attention, where the back-bone Transformer is used as an agent/wrapper to rank the impact in terms of the difference of loss caused by masking a token of interest. It is the first time to develop a lightweighted model from a feature selection point of view with explicit insight into token's relevance to the decision. A MLP for token filtering is the only added module, which acts as a two-class classifier with minor change on the overall architecture, and its training is not tough. The present solution is a one-pass filter. In the future, we will investigate into the relevance of tokens at middle layers to the final decision to further improve the efficiency.\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-05-24"},"doi":{"kind":"string","value":""},"references":{"kind":"list like","value":[{"authors":"","journal":"FLOPs ↓ Throughput ↑ ViT-B","ref_id":"b0","title":"Model Efficiency Top1 Acc.(%) ↑ Resolution #Params","year":""},{"authors":"","journal":"","ref_id":"b1","title":"Comparison with baselines. Except for E-ViT, which undergoes training of 300 epochs, the other models are trained with 100 epochs. Note that Dynamic-ViT-S turns out from the resolution of 384 × 384","year":""},{"authors":"Gedas Bertasius; Heng Wang; Lorenzo Torresani","journal":"ICML","ref_id":"b2","title":"Is space-time attention all you need for video understanding","year":"2021"},{"authors":"Daniel Bolya; Cheng-Yang Fu; Xiaoliang Dai; Peizhao Zhang; Judy Hoffman","journal":"","ref_id":"b3","title":"Hydra attention: Efficient attention with many heads","year":"2022"},{"authors":"Nicolas Carion; Francisco Massa; Gabriel Synnaeve; Nicolas Usunier; Alexander Kirillov; Sergey Zagoruyko","journal":"Springer","ref_id":"b4","title":"End-toend object detection with transformers","year":"2020"},{"authors":"Krzysztof Choromanski; Valerii Likhosherstov; David Dohan; Xingyou Song; Andreea Gane; Tamas Sarlos; Peter Hawkins; Jared Davis; Afroz Mohiuddin; Lukasz Kaiser","journal":"","ref_id":"b5","title":"Rethinking attention with performers","year":"2020"},{"authors":"Jia Deng; Wei Dong; Richard Socher; Li-Jia Li; Kai Li; Li Fei-Fei","journal":"Ieee","ref_id":"b6","title":"Imagenet: A large-scale hierarchical image database","year":"2009"},{"authors":"Xiaoyi Dong; Jianmin Bao; Dongdong Chen; Weiming Zhang; Nenghai Yu; Lu Yuan; Dong Chen; Baining Guo","journal":"","ref_id":"b7","title":"Cswin transformer: A general vision transformer backbone with cross-shaped windows","year":"2022"},{"authors":"Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly","journal":"","ref_id":"b8","title":"An image is worth 16x16 words: Transformers for image recognition at scale","year":"2020"},{"authors":"Maha Elbayad; Jiatao Gu; Edouard Grave; Michael Auli","journal":"","ref_id":"b9","title":"Depth-adaptive transformer","year":"2019"},{"authors":"Nina Grgić-Hlača; Muhammad Bilal Zafar; Krishna P Gummadi; Adrian Weller","journal":"","ref_id":"b10","title":"Beyond distributive fairness in algorithmic decision making: Feature selection for procedurally fair learning","year":"2018"},{"authors":"Ning Gui; Danni Ge; Ziyin Hu","journal":"","ref_id":"b11","title":"Afs: An attention-based mechanism for supervised feature selection","year":"2019"},{"authors":"Yihui He; Xiangyu Zhang; Jian Sun","journal":"","ref_id":"b12","title":"Channel pruning for accelerating very deep neural networks","year":"2017"},{"authors":"A Drew; Larry Hudson; Zitnick","journal":"PMLR","ref_id":"b13","title":"Generative adversarial transformers","year":"2021"},{"authors":"Zihang Jiang; Qibin Hou; Li Yuan; Daquan Zhou; Xiaojie Jin; Anran Wang; Jiashi Feng","journal":"","ref_id":"b14","title":"Token labeling: Training a 85.4% top-1 accuracy vision transformer with 56m parameters on imagenet","year":"2021"},{"authors":"Nikita Kitaev; Lukasz Kaiser; Anselm Levskaya","journal":"Learning","ref_id":"b15","title":"Reformer: The efficient transformer","year":"2020"},{"authors":"Takumi Kobayashi","journal":"","ref_id":"b16","title":"Group softmax loss with discriminative feature grouping","year":"2021"},{"authors":"Ron Kohavi; George H John","journal":"Artificial intelligence","ref_id":"b17","title":"Wrappers for feature subset selection","year":"1997"},{"authors":"Xin Lai; Jianhui Liu; Li Jiang; Liwei Wang; Hengshuang Zhao; Shu Liu; Xiaojuan Qi; Jiaya Jia","journal":"","ref_id":"b18","title":"Stratified transformer for 3d point cloud segmentation","year":"2022"},{"authors":"Yifeng Li; Chih-Yu Chen; Wyeth W Wasserman","journal":"Journal of Computational Biology","ref_id":"b19","title":"Deep feature selection: theory and application to identify enhancers and promoters","year":"2016"},{"authors":"Weicong Liang; Yuhui Yuan; Henghui Ding; Xiao Luo; Weihong Lin; Ding Jia; Zheng Zhang; Chao Zhang; Han Hu","journal":"","ref_id":"b20","title":"Expediting large-scale vision transformer for dense prediction without fine-tuning","year":"2022"},{"authors":"Youwei Liang; Chongjian Ge; Zhan Tong; Yibing Song; Jue Wang; Pengtao Xie","journal":"","ref_id":"b21","title":"Not all patches are what you need: Expediting vision transformers via token reorganizations","year":"2022"},{"authors":"Hongfu Liu; Haiyi Mao; Yun Fu","journal":"IEEE","ref_id":"b22","title":"Robust multi-view feature selection","year":"2016"},{"authors":"Ze Liu; Yutong Lin; Yue Cao; Han Hu; Yixuan Wei; Zheng Zhang; Stephen Lin; Baining Guo","journal":"","ref_id":"b23","title":"Swin transformer: Hierarchical vision transformer using shifted windows","year":"2021"},{"authors":"Ping Luo; Zhenyao Zhu; Ziwei Liu; Xiaogang Wang; Xiaoou Tang","journal":"","ref_id":"b24","title":"Face model compression by distilling knowledge from neurons","year":"2016"},{"authors":"Zizheng Pan; Bohan Zhuang; Haoyu He; Jing Liu; Jianfei Cai","journal":"","ref_id":"b25","title":"Less is more: Pay less attention in vision transformers","year":"2022"},{"authors":"Yongming Rao; Wenliang Zhao; Benlin Liu; Jiwen Lu; Jie Zhou; Cho-Jui Hsieh","journal":"NeurIPS","ref_id":"b26","title":"Dynamicvit: Efficient vision transformers with dynamic token sparsification","year":"2008"},{"authors":"Debaditya Roy; K Sri Rama; C Murty; Mohan Krishna","journal":"IEEE","ref_id":"b27","title":"Feature selection using deep neural networks","year":"2015"},{"authors":"Maying Shen; Pavlo Molchanov; Hongxu Yin; Jose M Alvarez","journal":"","ref_id":"b28","title":"When to prune? a policy towards early structural pruning","year":"2022"},{"authors":"Hugo Touvron; Matthieu Cord; Matthijs Douze; Francisco Massa; Alexandre Sablayrolles; Hervé Jégou","journal":"PMLR","ref_id":"b29","title":"Training data-efficient image transformers & distillation through attention","year":"2021"},{"authors":"Ashish Vaswani; Prajit Ramachandran; Aravind Srinivas; Niki Parmar; Blake Hechtman; Jonathon Shlens","journal":"","ref_id":"b30","title":"Scaling local self-attention for parameter efficient visual backbones","year":"2021"},{"authors":"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin","journal":"Advances in neural information processing systems","ref_id":"b31","title":"Attention is all you need","year":"2017"},{"authors":"Wenhai Wang; Enze Xie; Xiang Li; Deng-Ping Fan; Kaitao Song; Ding Liang; Tong Lu; Ping Luo; Ling Shao","journal":"","ref_id":"b32","title":"Pyramid vision transformer: A versatile backbone for dense prediction without convolutions","year":"2021"},{"authors":"Sitong Wu; Tianyi Wu; Haoru Tan; Guodong Guo","journal":"","ref_id":"b33","title":"Pale transformer: A general vision transformer backbone with pale-shaped attention","year":"2022"},{"authors":"Enze Xie; Wenhai Wang; Zhiding Yu; Animashree Anandkumar; Jose M Alvarez; Ping Luo","journal":"","ref_id":"b34","title":"Segformer: Simple and efficient design for semantic segmentation with transformers. neural information processing systems","year":"2021"},{"authors":"Pengtao Xie; Xuefeng Du","journal":"","ref_id":"b35","title":"Performance-aware mutual knowledge distillation for improving neural architecture search","year":"2022"},{"authors":"Chenglin Yang; Yilin Wang; Jianming Zhang; He Zhang; Zijun Wei; Zhe Lin; Alan Yuille","journal":"","ref_id":"b36","title":"Lite vision transformer with enhanced self-attention","year":"2022"},{"authors":"Arash Hongxu Yin; Jose M Vahdat; Arun Alvarez; Jan Mallya; Pavlo Kautz; Molchanov","journal":"","ref_id":"b37","title":"A-vit: Adaptive tokens for efficient vision transformer","year":"2022"},{"authors":"Xiyu Yu; Tongliang Liu; Xinchao Wang; Dacheng Tao","journal":"","ref_id":"b38","title":"On compressing deep models by low rank and sparse decomposition","year":"2017"},{"authors":"Wang Zeng; Sheng Jin; Wentao Liu; Chen Qian; Ping Luo; Wanli Ouyang; Xiaogang Wang","journal":"","ref_id":"b39","title":"Not all tokens are equal: Human-centric visual analysis via token clustering transformer","year":"2022"},{"authors":"Wangchunshu Zhou; Canwen Xu; Tao Ge; Julian Mcauley; Ke Xu; Furu Wei","journal":"Advances in Neural Information Processing Systems","ref_id":"b40","title":"Bert loses patience: Fast and robust inference with early exit","year":"2020"}],"string":"[\n {\n \"authors\": \"\",\n \"journal\": \"FLOPs ↓ Throughput ↑ ViT-B\",\n \"ref_id\": \"b0\",\n \"title\": \"Model Efficiency Top1 Acc.(%) ↑ Resolution #Params\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Comparison with baselines. Except for E-ViT, which undergoes training of 300 epochs, the other models are trained with 100 epochs. Note that Dynamic-ViT-S turns out from the resolution of 384 × 384\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Gedas Bertasius; Heng Wang; Lorenzo Torresani\",\n \"journal\": \"ICML\",\n \"ref_id\": \"b2\",\n \"title\": \"Is space-time attention all you need for video understanding\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Daniel Bolya; Cheng-Yang Fu; Xiaoliang Dai; Peizhao Zhang; Judy Hoffman\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Hydra attention: Efficient attention with many heads\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Nicolas Carion; Francisco Massa; Gabriel Synnaeve; Nicolas Usunier; Alexander Kirillov; Sergey Zagoruyko\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b4\",\n \"title\": \"End-toend object detection with transformers\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Krzysztof Choromanski; Valerii Likhosherstov; David Dohan; Xingyou Song; Andreea Gane; Tamas Sarlos; Peter Hawkins; Jared Davis; Afroz Mohiuddin; Lukasz Kaiser\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Rethinking attention with performers\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jia Deng; Wei Dong; Richard Socher; Li-Jia Li; Kai Li; Li Fei-Fei\",\n \"journal\": \"Ieee\",\n \"ref_id\": \"b6\",\n \"title\": \"Imagenet: A large-scale hierarchical image database\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Xiaoyi Dong; Jianmin Bao; Dongdong Chen; Weiming Zhang; Nenghai Yu; Lu Yuan; Dong Chen; Baining Guo\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Cswin transformer: A general vision transformer backbone with cross-shaped windows\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"An image is worth 16x16 words: Transformers for image recognition at scale\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Maha Elbayad; Jiatao Gu; Edouard Grave; Michael Auli\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Depth-adaptive transformer\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Nina Grgić-Hlača; Muhammad Bilal Zafar; Krishna P Gummadi; Adrian Weller\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Beyond distributive fairness in algorithmic decision making: Feature selection for procedurally fair learning\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Ning Gui; Danni Ge; Ziyin Hu\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Afs: An attention-based mechanism for supervised feature selection\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yihui He; Xiangyu Zhang; Jian Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Channel pruning for accelerating very deep neural networks\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"A Drew; Larry Hudson; Zitnick\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b13\",\n \"title\": \"Generative adversarial transformers\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zihang Jiang; Qibin Hou; Li Yuan; Daquan Zhou; Xiaojie Jin; Anran Wang; Jiashi Feng\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Token labeling: Training a 85.4% top-1 accuracy vision transformer with 56m parameters on imagenet\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Nikita Kitaev; Lukasz Kaiser; Anselm Levskaya\",\n \"journal\": \"Learning\",\n \"ref_id\": \"b15\",\n \"title\": \"Reformer: The efficient transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Takumi Kobayashi\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Group softmax loss with discriminative feature grouping\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ron Kohavi; George H John\",\n \"journal\": \"Artificial intelligence\",\n \"ref_id\": \"b17\",\n \"title\": \"Wrappers for feature subset selection\",\n \"year\": \"1997\"\n },\n {\n \"authors\": \"Xin Lai; Jianhui Liu; Li Jiang; Liwei Wang; Hengshuang Zhao; Shu Liu; Xiaojuan Qi; Jiaya Jia\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Stratified transformer for 3d point cloud segmentation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yifeng Li; Chih-Yu Chen; Wyeth W Wasserman\",\n \"journal\": \"Journal of Computational Biology\",\n \"ref_id\": \"b19\",\n \"title\": \"Deep feature selection: theory and application to identify enhancers and promoters\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Weicong Liang; Yuhui Yuan; Henghui Ding; Xiao Luo; Weihong Lin; Ding Jia; Zheng Zhang; Chao Zhang; Han Hu\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Expediting large-scale vision transformer for dense prediction without fine-tuning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Youwei Liang; Chongjian Ge; Zhan Tong; Yibing Song; Jue Wang; Pengtao Xie\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Not all patches are what you need: Expediting vision transformers via token reorganizations\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hongfu Liu; Haiyi Mao; Yun Fu\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b22\",\n \"title\": \"Robust multi-view feature selection\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Ze Liu; Yutong Lin; Yue Cao; Han Hu; Yixuan Wei; Zheng Zhang; Stephen Lin; Baining Guo\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Swin transformer: Hierarchical vision transformer using shifted windows\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ping Luo; Zhenyao Zhu; Ziwei Liu; Xiaogang Wang; Xiaoou Tang\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Face model compression by distilling knowledge from neurons\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Zizheng Pan; Bohan Zhuang; Haoyu He; Jing Liu; Jianfei Cai\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Less is more: Pay less attention in vision transformers\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yongming Rao; Wenliang Zhao; Benlin Liu; Jiwen Lu; Jie Zhou; Cho-Jui Hsieh\",\n \"journal\": \"NeurIPS\",\n \"ref_id\": \"b26\",\n \"title\": \"Dynamicvit: Efficient vision transformers with dynamic token sparsification\",\n \"year\": \"2008\"\n },\n {\n \"authors\": \"Debaditya Roy; K Sri Rama; C Murty; Mohan Krishna\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b27\",\n \"title\": \"Feature selection using deep neural networks\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Maying Shen; Pavlo Molchanov; Hongxu Yin; Jose M Alvarez\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"When to prune? a policy towards early structural pruning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hugo Touvron; Matthieu Cord; Matthijs Douze; Francisco Massa; Alexandre Sablayrolles; Hervé Jégou\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b29\",\n \"title\": \"Training data-efficient image transformers & distillation through attention\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ashish Vaswani; Prajit Ramachandran; Aravind Srinivas; Niki Parmar; Blake Hechtman; Jonathon Shlens\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Scaling local self-attention for parameter efficient visual backbones\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Łukasz Kaiser; Illia Polosukhin\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b31\",\n \"title\": \"Attention is all you need\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Wenhai Wang; Enze Xie; Xiang Li; Deng-Ping Fan; Kaitao Song; Ding Liang; Tong Lu; Ping Luo; Ling Shao\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Pyramid vision transformer: A versatile backbone for dense prediction without convolutions\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Sitong Wu; Tianyi Wu; Haoru Tan; Guodong Guo\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Pale transformer: A general vision transformer backbone with pale-shaped attention\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Enze Xie; Wenhai Wang; Zhiding Yu; Animashree Anandkumar; Jose M Alvarez; Ping Luo\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Segformer: Simple and efficient design for semantic segmentation with transformers. neural information processing systems\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Pengtao Xie; Xuefeng Du\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Performance-aware mutual knowledge distillation for improving neural architecture search\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Chenglin Yang; Yilin Wang; Jianming Zhang; He Zhang; Zijun Wei; Zhe Lin; Alan Yuille\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Lite vision transformer with enhanced self-attention\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Arash Hongxu Yin; Jose M Vahdat; Arun Alvarez; Jan Mallya; Pavlo Kautz; Molchanov\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"A-vit: Adaptive tokens for efficient vision transformer\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xiyu Yu; Tongliang Liu; Xinchao Wang; Dacheng Tao\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"On compressing deep models by low rank and sparse decomposition\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Wang Zeng; Sheng Jin; Wentao Liu; Chen Qian; Ping Luo; Wanli Ouyang; Xiaogang Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Not all tokens are equal: Human-centric visual analysis via token clustering transformer\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Wangchunshu Zhou; Canwen Xu; Tao Ge; Julian Mcauley; Ke Xu; Furu Wei\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b40\",\n \"title\": \"Bert loses patience: Fast and robust inference with early exit\",\n \"year\": \"2020\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[4,101.83,425.68,184.54,11.72],"formula_id":"formula_0","formula_text":"X = x i ∈ R d |i = 1, 2. . . , N ,(1)"},{"formula_coordinates":[4,90.03,489.24,196.34,9.65],"formula_id":"formula_1","formula_text":"X i = {x 1 , . . . x i-1 , ∅, x i+1 , . . . , x N } ,(2)"},{"formula_coordinates":[4,119.08,549.72,167.28,8.96],"formula_id":"formula_2","formula_text":"ŷ = T ransf ormer(X),(3)"},{"formula_coordinates":[4,116.33,569.01,170.03,9.65],"formula_id":"formula_3","formula_text":"ŷi = T ransf ormer(X i ).(4)"},{"formula_coordinates":[4,113.38,626.96,172.98,8.96],"formula_id":"formula_4","formula_text":"L = CrossEntropy(ŷ, y),(5)"},{"formula_coordinates":[4,110.24,645.25,176.12,9.65],"formula_id":"formula_5","formula_text":"L i = CrossEntropy(ŷ i , y).(6)"},{"formula_coordinates":[4,136.35,704.2,150.01,9.65],"formula_id":"formula_6","formula_text":"∆L i = L -L i .(7)"},{"formula_coordinates":[4,364.41,430.28,180.71,24],"formula_id":"formula_7","formula_text":"label(x i ) = 0, L -L i ≤ ρ 1, L -L i > ρ(8)"},{"formula_coordinates":[5,98.08,95.13,188.28,44.92],"formula_id":"formula_8","formula_text":"x ′ i = [x i , x global ]. (9) x global = AAP (X) = 1 N Σ N k=1 x k .(10)"},{"formula_coordinates":[5,112.43,177.88,169.78,12.69],"formula_id":"formula_9","formula_text":"p i = Sigmoid(M LP (x ′ i )). (11"},{"formula_coordinates":[5,282.21,180.27,4.15,8.64],"formula_id":"formula_10","formula_text":")"},{"formula_coordinates":[5,60.92,271.97,198.04,9.65],"formula_id":"formula_11","formula_text":"L M LP = BinaryCrossEntropy(p i , label(x i )),"},{"formula_coordinates":[5,96.82,704.2,189.54,9.65],"formula_id":"formula_12","formula_text":"L f inetune = CrossEntropy(ŷ, y),(13)"},{"formula_coordinates":[5,361.38,206.16,62.78,12.32],"formula_id":"formula_13","formula_text":"′ i = [x i , x global ]"},{"formula_coordinates":[5,355.69,218.12,108.85,12.32],"formula_id":"formula_14","formula_text":"p i = Sigmoid(M LP (x ′ i ))"}],"string":"[\n {\n \"formula_coordinates\": [\n 4,\n 101.83,\n 425.68,\n 184.54,\n 11.72\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"X = x i ∈ R d |i = 1, 2. . . , N ,(1)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 90.03,\n 489.24,\n 196.34,\n 9.65\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"X i = {x 1 , . . . x i-1 , ∅, x i+1 , . . . , x N } ,(2)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 119.08,\n 549.72,\n 167.28,\n 8.96\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"ŷ = T ransf ormer(X),(3)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 116.33,\n 569.01,\n 170.03,\n 9.65\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"ŷi = T ransf ormer(X i ).(4)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 113.38,\n 626.96,\n 172.98,\n 8.96\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"L = CrossEntropy(ŷ, y),(5)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 110.24,\n 645.25,\n 176.12,\n 9.65\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"L i = CrossEntropy(ŷ i , y).(6)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 136.35,\n 704.2,\n 150.01,\n 9.65\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"∆L i = L -L i .(7)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 364.41,\n 430.28,\n 180.71,\n 24\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"label(x i ) = 0, L -L i ≤ ρ 1, L -L i > ρ(8)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 98.08,\n 95.13,\n 188.28,\n 44.92\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"x ′ i = [x i , x global ]. (9) x global = AAP (X) = 1 N Σ N k=1 x k .(10)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 112.43,\n 177.88,\n 169.78,\n 12.69\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"p i = Sigmoid(M LP (x ′ i )). (11\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 282.21,\n 180.27,\n 4.15,\n 8.64\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 60.92,\n 271.97,\n 198.04,\n 9.65\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"L M LP = BinaryCrossEntropy(p i , label(x i )),\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 96.82,\n 704.2,\n 189.54,\n 9.65\n ],\n \"formula_id\": \"formula_12\",\n \"formula_text\": \"L f inetune = CrossEntropy(ŷ, y),(13)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 361.38,\n 206.16,\n 62.78,\n 12.32\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"′ i = [x i , x global ]\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 355.69,\n 218.12,\n 108.85,\n 12.32\n ],\n \"formula_id\": \"formula_14\",\n \"formula_text\": \"p i = Sigmoid(M LP (x ′ i ))\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"Predicting Token Impact Towards Efficient Vision Transformer"},"abstract":{"kind":"string","value":"Token filtering to reduce irrelevant tokens prior to selfattention is a straightforward way to enable efficient vision Transformer. This is the first work to view token filtering from a feature selection perspective, where we weigh the importance of a token according to how much it can change the loss once masked. If the loss changes greatly after masking a token of interest, it means that such a token has a significant impact on the final decision and is thus relevant. Otherwise, the token is less important for the final decision, so it can be filtered out. After applying the token filtering module generalized from the whole training data, the token number fed to the self-attention module can be obviously reduced in the inference phase, leading to much fewer computations in all the subsequent self-attention layers. The token filter can be realized using a very simple network, where we utilize multi-layer perceptron. Except for the uniqueness of performing token filtering only once from the very beginning prior to self-attention, the other core feature making our method different from the other token filters lies in the predictability of token impact from a feature selection point of view. The experiments show that the proposed method provides an efficient way to approach a light weighted model after optimized with a backbone by means of fine tune, which is easy to be deployed in comparison with the existing methods based on training from scratch."},"authors":{"kind":"string","value":"Hong Wang; Su Yang; Xiaoke Huang; Weishan Zhang"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 .1Figure 1. The overall training process: The two branches of the vision Transformer are in fact the same one, whose parameters are fixed during training. The delta loss and ρ refer to Eq. (7) and Eq. (8).","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Algorithm 11Fig. 1 depicts how to train the token selection module with delta loss. Algorithm 1 and Algorithm 2 describe respectively how to label tokens with naive DeiT [29] and how to train the selection module. Token labeling with naive DeiT [29] Input: X = x i ∈ R d |i = 1, 2. . . , N , and the corresponding ground truth y. Output: Label = {label(x i )|i = 1, 2. . . , N }. 1: Label = ∅ 2: Set ρ to control the significance of pseudo labeling. 3: ŷ = T ransf ormer(X) 4: L = CrossEntropy(ŷ, y) 5: for i = 1, 2, ..., N do 6: X i = {x 1 , . . . x i-1 , ∅, x i+1 , . . . , x N } 7: ŷi = T ransf ormer(X i ) 8:L i = CrossEntropy(ŷ i , y) 9:if L -L i ≤ ρ then","figure_data":"","figure_id":"fig_1","figure_label":"1","figure_type":"figure"},{"figure_caption":"Algorithm 2 : repeat 3 :23Token filter training Input: Batch of images with tokens and the corresponding pseudo labels in the form of X = x i ∈ R d |i = 1, 2. . . , N and Label = {label(x i )|i = 1, 2. . . , N }. Output: Parameters W of the MLP token filter. 1: Random initialization of W 2Load (X,Label) of one image in the batch in turn 4:","figure_data":"","figure_id":"fig_2","figure_label":"23","figure_type":"figure"},{"figure_caption":"Figure 3 .3Figure 3. Distribution of all the DL values on ImageNet1K training set.","figure_data":"","figure_id":"fig_3","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 .4Figure 4. Average DL of every image patch obtained from DeiT-T [29]. Darker color corresponds with lower DL value.","figure_data":"","figure_id":"fig_4","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 .5Figure 5. The average of masks predicted by our token selection module on ImageNet1K validation set.","figure_data":"","figure_id":"fig_5","figure_label":"5","figure_type":"figure"},{"figure_caption":"Figure 6 .6Figure 6. Original image (left) and the masked image (right) resulting from DL-ViT on the ImageNet1K set. The left two columns are the validation set, and the right two columns the training set.","figure_data":"","figure_id":"fig_6","figure_label":"6","figure_type":"figure"},{"figure_caption":"Figure 7 .7Figure 7. The tiny model complexity (FLOPs) and top-1 accuracy trade-offs on ImageNet.","figure_data":"","figure_id":"fig_7","figure_label":"7","figure_type":"figure"},{"figure_caption":"Figure 8 .8Figure 8. The small model complexity (FLOPs) and top-1 accuracy trade-offs on ImageNet.","figure_data":"","figure_id":"fig_8","figure_label":"8","figure_type":"figure"},{"figure_caption":"Performance of DL-ViT-T subject to local/global feature and threshold ρ.","figure_data":"ThresholdTop-1 Acc. (%)↑Top-5 Acc. (%)↑FLOPs ↓Throughput (images/s)↑Top-1 Acc. (%)↑Top-5 Acc. (%)↑FLOPs ↓Throughput (images/s)↑DeiT-T [29]71.490.81.3G323171.490.81.3G3231DL-ViT-T with local featureDL-ViT-T with local and global feature0.00172.090.80.8G477166.586.10.6G46900.00273.191.51.0G393771.189.90.7G45650.00362.483.50.3G599670.089.10.6G5527StrategyMetric#Params.↓ FLOPs ↓ Top1 Acc.(%) ↑DL-ViT-T5.7M0.7G71.1DeiT-T 0 [29]5.7M0.7G68.2DL-ViT-T 05.7M0.4G56.6","figure_id":"tab_0","figure_label":"2","figure_type":"table"},{"figure_caption":"Comparison with DeiT using random token discard (DeiT-T0 in the second row) and DL-ViT without pre-training but randomly initializing the MLP (DL-ViT-T0 in the third row).","figure_data":"","figure_id":"tab_1","figure_label":"3","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 .1Figure 1. The overall training process: The two branches of the vision Transformer are in fact the same one, whose parameters are fixed during training. The delta loss and ρ refer to Eq. (7) and Eq. (8).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Algorithm 11Fig. 1 depicts how to train the token selection module with delta loss. Algorithm 1 and Algorithm 2 describe respectively how to label tokens with naive DeiT [29] and how to train the selection module. Token labeling with naive DeiT [29] Input: X = x i ∈ R d |i = 1, 2. . . , N , and the corresponding ground truth y. Output: Label = {label(x i )|i = 1, 2. . . , N }. 1: Label = ∅ 2: Set ρ to control the significance of pseudo labeling. 3: ŷ = T ransf ormer(X) 4: L = CrossEntropy(ŷ, y) 5: for i = 1, 2, ..., N do 6: X i = {x 1 , . . . x i-1 , ∅, x i+1 , . . . , x N } 7: ŷi = T ransf ormer(X i ) 8:L i = CrossEntropy(ŷ i , y) 9:if L -L i ≤ ρ then\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Algorithm 2 : repeat 3 :23Token filter training Input: Batch of images with tokens and the corresponding pseudo labels in the form of X = x i ∈ R d |i = 1, 2. . . , N and Label = {label(x i )|i = 1, 2. . . , N }. Output: Parameters W of the MLP token filter. 1: Random initialization of W 2Load (X,Label) of one image in the batch in turn 4:\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"23\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 .3Figure 3. Distribution of all the DL values on ImageNet1K training set.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 .4Figure 4. Average DL of every image patch obtained from DeiT-T [29]. Darker color corresponds with lower DL value.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 .5Figure 5. The average of masks predicted by our token selection module on ImageNet1K validation set.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 .6Figure 6. Original image (left) and the masked image (right) resulting from DL-ViT on the ImageNet1K set. The left two columns are the validation set, and the right two columns the training set.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 7 .7Figure 7. The tiny model complexity (FLOPs) and top-1 accuracy trade-offs on ImageNet.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"7\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 8 .8Figure 8. The small model complexity (FLOPs) and top-1 accuracy trade-offs on ImageNet.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_8\",\n \"figure_label\": \"8\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Performance of DL-ViT-T subject to local/global feature and threshold ρ.\",\n \"figure_data\": \"ThresholdTop-1 Acc. (%)↑Top-5 Acc. (%)↑FLOPs ↓Throughput (images/s)↑Top-1 Acc. (%)↑Top-5 Acc. (%)↑FLOPs ↓Throughput (images/s)↑DeiT-T [29]71.490.81.3G323171.490.81.3G3231DL-ViT-T with local featureDL-ViT-T with local and global feature0.00172.090.80.8G477166.586.10.6G46900.00273.191.51.0G393771.189.90.7G45650.00362.483.50.3G599670.089.10.6G5527StrategyMetric#Params.↓ FLOPs ↓ Top1 Acc.(%) ↑DL-ViT-T5.7M0.7G71.1DeiT-T 0 [29]5.7M0.7G68.2DL-ViT-T 05.7M0.4G56.6\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison with DeiT using random token discard (DeiT-T0 in the second row) and DL-ViT without pre-training but randomly initializing the MLP (DL-ViT-T0 in the third row).\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"[23]\", \"Explanation\": \"The cited work, Swin Transformer, is a method that the citing paper adopts to reduce the computational load in vision Transformer models by confining self-attention to a local neighborhood.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[33]\", \"Explanation\": \"The cited work, Pale Transformer, is another method that the citing paper mentions as a way to reduce the computational load in vision Transformer models by limiting self-attention to a local region.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[30]\", \"Explanation\": \"The cited work, HaloNet, is a method that the citing paper discusses as a way to reduce the computational load in vision Transformer models by limiting self-attention to a local neighborhood.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work, CSWin Transformer, is a method that the citing paper mentions as a way to reduce the computational load in vision Transformer models by limiting self-attention to a local region.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[2]\", \"Explanation\": \"The cited work introduces the Hydra Attention method, which the citing paper adopts to change the computing order in self-attention and incorporate multiple heads in the self-attention module to reduce complexity.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[39]\", \"Explanation\": \"The cited work provides a methodology for grouping similar tokens into clusters via unsupervised learning, which the citing paper utilizes in its research to group tokens and act as abstractive representations in self-attention.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[26]\", \"Explanation\": \"The cited work introduces a method of using multi-layer perceptron (MLP) as a gating function in ViT to selectively forward tokens to subsequent self-attention layers, which the citing paper adopts in their research to improve the quality of clustering and reduce the token number.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[37]\", \"Explanation\": \"The cited work proposes an early stop criterion based on the accumulated token value at the first dimension, which the citing paper builds upon to further improve the quality of token filtering in their research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[21]\", \"Explanation\": \"The cited work assumes token importance to be its attentive weight correlated to class token, but the complex coupling layer by layer brings in uncertainty to the weights. The citing paper extends the research by applying gradual token filtering to preserve less attentive tokens and aid further testing.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[37,21]\", \"Explanation\": \"The cited works provide the basis for the token filtering methods used in the citing paper, but the methods are based on heuristics and are not fully explainable, making it difficult to understand the impact of token filtering on the final decision.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[26]\", \"Explanation\": \"The cited work is used as a basis for the end-to-end training of the token filtering method in the citing paper, but the method is not fully explainable and may result in unpredictable token filtering on the fly.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[26]\", \"Explanation\": \"The cited work is used as a basis for the end-to-end training of the token filtering method in the citing paper, but the method is extended to include a token filter with learnable parameters that can be generalized from the whole training corpus to filter out irrelevant tokens from the very beginning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[17]\", \"Explanation\": \"The cited work on feature selection provides the basis for the token filtering mechanism proposed in the citing paper, which is used to evaluate the relevance of tokens in a ViT model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[31]\", \"Explanation\": \"The cited work on natural language processing (NLP) provides the foundation for the use of Transformer in computer vision by introducing the concept of no-overlapping image patches for image classification.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[7]\", \"Explanation\": \"The cited work on ViT is an extension of the use of Transformer in computer vision, as it is the first to apply the concept to the field and use no-overlapping image patches for image classification.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[29]\", \"Explanation\": \"The cited work on DeiT is a continuation of the use of Transformer in computer vision, as it improves the training configuration and proposes a new architecture that can be trained with ImageNet1K for better performance.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[2]\", \"Explanation\": \"The cited work Hydra Attention is referenced in the citing paper to provide a new method for reorganizing the computing order in self-attention and incorporating multiple heads into self-attention to reduce complexity.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[23]\", \"Explanation\": \"The cited work, Swin Transformer, provides a method for reducing the computational complexity of self-attention by enforcing the computation to be conducted in a predefined local region. This method is adopted in the citing paper to alleviate the computations of self-attention in a more efficient way.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[33]\", \"Explanation\": \"The cited work, Pale Transformer, is another method for reducing the computational complexity of self-attention by reducing the number of tokens through enforcing the computation of self-attention in a predefined local region. This method is also adopted in the citing paper to address the issue of high computational complexity in self-attention.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[30]\", \"Explanation\": \"The cited work, HaloNet, is another method for reducing the computational complexity of self-attention by reducing the number of tokens through enforcing the computation of self-attention in a predefined local region. This method is also adopted in the citing paper to address the issue of high computational complexity in self-attention.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work, CSWin Transformer, is another method for reducing the computational complexity of self-attention by reducing the number of tokens through enforcing the computation of self-attention in a predefined local region. This method is also adopted in the citing paper to address the issue of high computational complexity in self-attention.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[39]\", \"Explanation\": \"The cited work, TCFormer, is a method for reducing the computational complexity of self-attention by fusing the tokens in the same cluster into a new one utilizing a weighted average. This method is adopted in the citing paper to address the issue of high computational complexity in self-attention in a more efficient way.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[20]\", \"Explanation\": \"The cited work introduces a clustering-based approach to reduce the number of tokens in the first few layers of ViT, which the citing paper adopts to improve the efficiency of dense computations in the middle layers.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[26]\", \"Explanation\": \"The cited work introduces a dynamic gating structure in the middle layers of ViT to make tokens gradually sparser, which the citing paper incorporates to improve the efficiency of self-attention computations.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[37]\", \"Explanation\": \"The cited work calculates the accumulated halting probability of each token in ViT using feature values from each Transformer layer, which the citing paper adopts to reduce the number of tokens without adding additional modules.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[21]\", \"Explanation\": \"The cited work assumes that top-k attentive weights correspond with relevant tokens in ViT and preserves irrelevant tokens throughout the pipeline for a gradual token dropout procedure, which the citing paper adopts to improve the efficiency of token processing.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[19]\", \"Explanation\": \"The work by Le et al. proposes a feature selector that is used as a method for feature selection in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[27]\", \"Explanation\": \"The work by Roy et al. provides a method for feature selection at each single input dimension, which is adopted in the citing paper to address feature selection limitations in DNNs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[11]\", \"Explanation\": \"The cited work introduces the use of an attention mechanism for feature weight generation, which the citing paper adopts in their research to address the specific structure of the data.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work proposes a method that harnesses feature partition in SoftMax loss function for effective learning of discriminative features, which the citing paper leverages in their research to improve the performance of the Transformer model.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[17]\", \"Explanation\": \"The cited work tests the impact of masking certain Transformer tokens on the final decision, which the citing paper builds upon to develop a wrapper-based feature selection method for identifying relevant tokens in the context of data with a specific structure.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[5]\", \"Explanation\": \"The cited work, ImageNet1K ILSVRC 2012 dataset, serves as the data source for the evaluation of the method presented in the citing paper for image classification.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[9]\", \"Explanation\": \"The cited work provides a method for measuring the number of floating-point operations (FLOPs) in DeiT-T and DeiT-S, which the citing paper uses to compare the efficiency of their method with the baselines.\"}]"}}},{"rowIdx":81,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_0"],"heading":"Introduction","publication_ref":["b32","b40","b28","b38","b1","b15","b29","b39","b16","b0","b13","b5","b30","b5","b17","b17"],"table_ref":[],"text":"Multi-modal models focus on jointly learning representations from multiple modalities, such as vision and language. Many tasks require the integration of information of vision and language, including image captioning (Vinyals et al., 2015), natural language visual reasoning (Zhou et al., 2017;Suhr et al., 2019), and cross-modal retrieval (Zhen et al., 2019). Multi-modal learning captures the interaction between different modalities, allowing the resulting representations to be used in multimedia applications that enhance human-computer interaction.\nRecently, pre-trained vision-language models (PVLMs; Chen et al. 2020;Lu et al. 2019;Tan and Bansal 2019) have achieved significant advances in multi-modal tasks. However, the data which PVLMs learn from is mostly for high-resource languages such as English. The resulting models rely on large amounts of training data for good performance, and often the models acquire biases that mean they perform poorly in low-resource languages such as Indonesian or Swahili. To address this, several multilingual PVLMs have been proposed (Zhou et al., 2021;Ni et al., 2021). A number of studies have used multilingual multimodal datasets (Bugliarello et al., 2022;Liu et al., 2021) and Figure 1 shows two examples from such datasets. The authors of these datasets used them to evaluate current famous PVLMs and demonstrated they do not perform well in low-resource cross-lingual transfer settings.\nIn this paper, we conjecture that meta-learning can mitigate this issue. This is a learning approach that enables machine learning models to adapt quickly to new tasks by learning the learning algorithm itself. Model-agnostic Meta-learning (MAML; Finn et al. 2017) is one of the most widely used meta-learning frameworks. It is based on gradient-descent optimization, does not require multiple models or complex settings, and can be used for a range of models. In previous work (Verma et al., 2020;Finn et al., 2017;Nooralahzadeh et al., 2020), MAML-based methods have been shown to be useful in low-resource and cross-lingual transfer scenarios, including both few-shot and zero-shot cross-lingual tasks. However, prior work has only attempted to use MAML for cross-lingual transfer in text-only tasks (Nooralahzadeh et al., 2020).\nInspired by previous works about using MAML for natural language tasks, this paper proposes XVL-MAML, a novel variant of MAML that addresses the limitations of previous PVLMs in vision-language tasks for low-resource crosslingual transfer. Our framework combines a traditional supervised loss for learning down-stream tasks with a contrastive loss to encourage the alignment between modalities, resulting in a crosslingual, multi-modal MAML optimization procedure.\nThe intuition underlying our method is that a contrastive loss can align representations of different modalities, and MAML allows the model to generalize quickly to unseen tasks (languages, in our case). We show that XVL-MAML can lead to significant improvements in PVLM performance for cross-lingual transfer. We also find that using contrastive learning in a MAML framework on its own can bring improvements in PVLM performance in unsupervised settings.\nIn sum, our contributions are as follows:\n(1) We propose a novel framework called XVL-MAML which is the first meta-learning method specialized for vision-language cross-lingual transfer, and doesn't require the translation or pre-training data.\n(2) We show that using only contrastive learning in the MAML framework in an unsupervised setting can also be useful. (3) We demonstrate that our proposed framework can boost the performance of current PVLMs across 14 languages and four tasks in both zero-shot learning and few-shot learning. (4) We conduct an ablation study to verify the effect of contrastive learning in both supervised and unsupervised settings and present an analysis across languages and tasks."},{"figure_ref":[],"heading":"Related Work","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Multilingual Vision-and-Language Methods and Tasks","publication_ref":["b4","b7","b26","b19","b13","b27","b16","b13","b1","b3","b2","b39","b0","b19","b34","b21"],"table_ref":[],"text":"Recent work has investigated vision-and-language cross-lingual transfer tasks. Elliott et al. (2016) proposed Multi30K, an image description dataset which contains descriptions in multiple languages.\nPrevious methods (Gella et al., 2017;Rotman et al., 2018) propose ways of bridging languages through images, but they mainly focus on imagetext retrieval and only consider high-resource languages such as English and German. Pfeiffer et al. (2022) built a multilingual visual question answer-ing dataset xGQA. Liu et al. (2021) proposed a multilingual version of the grounded visual reasoning dataset MaRVL, which follow the same setting as the natural language visual reasoning dataset NLVR2 (Su et al., 2019), but considers both cross-lingual transfer and domain shift between languages.\nSeveral pre-trained models are recently proposed for vision-and-language cross-lingual transfer. Ni et al. (2021) proposed M3P, a transformer-based pre-trained model that maps the same concepts in different modalities and languages into a common semantic space. Similar to M3P, Liu et al. (2021) extended UNITER (Chen et al., 2020), proposing mUNITER based on M-BERT (Devlin et al., 2019), and xUNITER based on XLM-R (Conneau et al., 2020). Zhou et al. (2021) proposed UC2, a model using a data augmentation method based on machine translation for cross-lingual cross-modal pre-training. Although pre-training methods have proven powerful across multiple tasks, they require large amounts of training data and show a clear performance gap between English and low-resource languages on the IGLUE benchmark (Bugliarello et al., 2022).\nRecently, some adapter-based efficient tuning methods (Pfeiffer et al., 2022;Wang et al., 2023) and translation augmented methods (Qiu et al., 2022) were proposed for multilingual multimodal tasks. But they still require a large amount of data or machine translated data for training. Our method, in contrast, only requires a small amount of auxiliary data."},{"figure_ref":[],"heading":"Meta-Learning","publication_ref":["b31","b5","b24","b36","b10","b33","b5","b10","b18","b8","b17"],"table_ref":[],"text":"Meta-learning has been increasingly popular in machine learning. Whereas conventional machine learning methods learn by data points, metalearning learns by tasks. Previous meta-learning work (Vinyals et al., 2016;Finn et al., 2017) focused on adapting to new tasks quickly. But metalearning can be applied to other scenarios as well, including semi-supervised learning (Ren et al., 2018), multi-task learning (Yu et al., 2020), and domain generalization (Li et al., 2018).\nPrior work has also explored the effectiveness of meta-learning in NLP: Wang et al. (2021) applied meta-learning in semantic parsing for domain generalization based on MAML (Finn et al., 2017;Li et al., 2018). Obamuyide and Vlachos (2019) leveraged meta-learning under limited su-pervision in a relation classification task. Recently, there have been some applications using MAML in cross-lingual transfer: Gu et al. (2018) and Nooralahzadeh et al. (2020) regard languages as tasks in their meta-learning framework. In contrast to these existing approaches, which explore text-only scenarios, we are the first to utilize metalearning for cross-lingual transfer in multi-modal tasks.\n3 Meta-learning for Vision-and-language Cross-lingual Transfer\nWe first formally define the problem of vision-and-Language cross-lingual transfer in the context of zero-shot and few-shot scenarios in Section 3.1. Then, we introduce our overall fine-tuning framework in Section 3.2. And we introduce the contrastive learning used for vision-and-language tasks in Section 3.3. Finally, we introduce our XVL-MAML algorithm in Section 3.4."},{"figure_ref":[],"heading":"Problem Definition","publication_ref":["b0"],"table_ref":[],"text":"Following the multilingual vision-language IGLUE benchmark (Bugliarello et al., 2022), we formulate the problem of cross-lingual transfer learning in vision-and-language scenarios. For understanding tasks, the input is a pair of an image V and text U , and the output Y is the result inferred by the multimodal model. We can thus formulate this problem as computing P θ (Y |V, U ), where θ are the parameters of the PVLMs. During training, the image-text pairs come from datasets D s in a set of source languages, and our aim is to perform well on datasets D t for the same task in the target languages. For the zero-shot setup, the pre-trained model fine-tuned on D s is directly used in inference on D t for unseen target languages. For the few-shot setup, after training on D s , the model is continually fine-tuned on several shots of the training set of D t and then evaluated on the development set of D t ."},{"figure_ref":[],"heading":"Overall Fine-tuning Framework For Cross-lingual Transfer","publication_ref":["b0"],"table_ref":[],"text":"The pipeline of our proposed meta-learning finetuning framework can be divided into three parts:\n1. Fine-tune the pre-trained vision-language model on data of the down-stream task in English 2. Fine-tune the model on data in the auxiliary language (one language other than English) using our proposed XVL-MAML algorithm.\n3. Evaluate the fine-tuned model on data in the target languages (languages other than English and the auxiliary language).\nThe traditional cross-lingual transfer learning procedure described in Bugliarello et al. (2022) only includes part 1 and 3. In part 3, if the setting is zero-shot, the model is evaluated on data in the target language directly, but if the setting is few-shot, the model continues to be fine-tuned on few-shot data in the target languages and is then evaluated. The difference between our framework and the traditional procedure is the additional finetuning step of part 2. We will describe it specifically in Section 3.4, but before that, we will introduce contrastive learning for vision-and-language tasks."},{"figure_ref":[],"heading":"Contrastive Learning for Vision-and-language Tasks","publication_ref":["b37","b22","b37"],"table_ref":[],"text":"The vision-and-language contrastive learning loss proposed by Zhang et al. (2020) has proven effective in medical image scenarios and is used as the pre-training objective function of CLIP (Radford et al., 2021). It can be regarded as an auxiliary task for representation learning, aiming to enable models to gain better aligned multi-modal representation for downstream tasks. In the contrastive learning scheme, a batch of embeddings of images encoded by the model can be written as I = {I 1 , ..., I N }, and a batch of embeddings of texts encoded by the model can be written as \nT = {T 1 , ..., T N },\nU = I • W ⊤ 1 (1) V = T • W ⊤ 2 (2)\nWhere U and V represent the batch of image-text pairs. Then the cosine similarity of each pair can be computed as\n⟨U i , V j ⟩ = U ⊤ i V j ∥U i ∥∥V j ∥ .\nThe objective is to maximize the similarity of matched image-text pairs and minimize the similarity of others. So the image-text contrastive loss can be formulated as follows:\nL 1 i = -log exp(⟨U i , V i ⟩) N K=1 exp(⟨U i , V k ⟩)(3)\nFollowing Zhang et al. (2020), the contrastive loss should be symmetric for each modality, and the text-image contrastive loss is:\nL 2 i = -log exp(⟨V i , U i ⟩) N K=1 exp(⟨V i , U k ⟩) (4)\nThe final contrastive loss of this batch of paired data is then:\nL CL = N i=1 (L 1 i + L 2 i ) (5)\nWhere L CL is the overall contrastive loss. When we minimize L CL , we maximize the similarity of image-text pairs which are positive examples."},{"figure_ref":[],"heading":"XVL-MAML","publication_ref":["b17"],"table_ref":[],"text":"Inspired by the effectiveness of MAML for quickly adapting to new tasks, we propose a novel variant of the MAML algorithm specialized for crosslingual transfer in vision and language tasks, called XVL-MAML. Specifically, we first integrate contrastive learning into the MAML algorithm, making it specialized for the visual-language task of crosslingual transfer learning. Our intuition is that we can use MAML with a contrastive loss as its learning objective for quickly adapting vision-language alignment to new languages. In this framework, the alignment between image and text in a specific language can be regarded as a task. Inspired by Nooralahzadeh et al. (2020), we use the data of one auxiliary language for fine-tuning, but with a contrastive loss as objective function in the MAML algorithm. Specifically, we sample a batch of support data B s and a batch of query data B q in the data in auxiliary language A for each virtual task T . Assuming the parameters of the model are θ and the contrastive loss on the support data is L CL (θ) Bs , then the parameters of the model can be updated by one step of gradient descent:\nθ ′ = θ -α∇ θ L CL (θ) Bs (6)\nFollowing the MAML algorithm, our final objective for this task is to minimize L CL (θ ′ ) Bq on the query data B q using gradient descent:\nθ ← θ -β∇ θ L CL (θ ′ ) Bq (7) θ ← θ -β∇ θ L CL (θ -α∇ θ L CL (θ) Bs ) Bq (8)\nOptimized using this method, pre-trained visionlanguage models can quickly adapt to new tasks in other languages without using any annotation in the auxiliary language for downstream tasks, so we will refer to this as an unsupervised scenario.\nIn supervised scenarios, where the downstream tasks labels in the auxiliary language are available, we combine the loss of the downstream task L with the vision-language contrastive loss L CL by adding them together. So during fine-tuning, Equation ( 8) is modified to:\nθ ← θ -β(∇ θ L(θ ′′ ) Bq + λ∇ θ L CL (θ ′ ) Bq ) (9)\nWhere the temporary parameters optimized for one step by the downstream task loss L on the support set B s is θ ′′ , β is the meta-learning rate, and λ is the scale factor of contrastive learning. By simply adding the gradients of the downstream task and contrastive learning in the meta-update, the model learns downstream tasks and vision-language alignment simultaneously for cross-lingual transfer."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":"In this section, we introduce both the base PVLMs we use for vision-language cross-lingual transfer, as well as the datasets and metrics we use to evaluate our proposed method. Then we describe how the experiments were conducted and discuss the results."},{"figure_ref":[],"heading":"Base models","publication_ref":["b13","b39","b1","b25","b0"],"table_ref":[],"text":"In this paper, we choose xUNITER (Liu et al., 2021) and UC2 (Zhou et al., 2021) as our base models, as they use different pre-training methods.\nThen we applied XVL-MAML to both models to show that this method works across different models.\nxUNITER is a multilingual version of the UNITER model (Chen et al., 2020). It has a similar architecture to UNITER and uses Faster-RCNN (Ren et al., 2015) 64.4 (+1.9) 29.9 (+0.9) 57.0 (+0.6) 21.3 (+1.6) 18.7 (+1.7)\nTable 1: Zero-shot performance (accuracy) of four baseline models only fine-tuned on English data (Baseline) and two models fine-tuned by our meta-learning method (Ours) on four IGLUE datasets (Bugliarello et al., 2022).\nUC2 uses a similar model architecture as UNITER, but different pre-training methods. Specifically, UC2 augments pre-training on English data by constructing a multilingual corpus via machine translation and then uses this augmented data for pre-training. It also proposes the Visual Translation Language Modeling (VTLM) pre-training method, which uses the image as a pivot to learn the relationship between parallel texts in two languages and their corresponding images."},{"figure_ref":["fig_0"],"heading":"Datasets and Metrics","publication_ref":["b0","b19","b13","b20","b12","b28","b11"],"table_ref":[],"text":"We use datasets for four tasks from the IGLUE benchmark (Bugliarello et al., 2022), which includes xGQA (Pfeiffer et al., 2022), MaRVL (Liu et al., 2021), XVNLI, and xFlickr&Co (Plummer et al., 2015;Lin et al., 2014). We show examples from MaRVL and XVNLI in Figure 1. Following the convention in IGLUE, the evaluation metric is accuracy for all tasks except cross-modal retrieval, which uses Recall@1. The task format of these four datasets are described below:\n• MaRVL is a multicutural vision-language reasoning dataset, following the format of English NLVR2 (Suhr et al., 2019) which namely to judge whether a sentence is correct or not for a pair of images.\n• XVNLI is a multilingual version of visual natural language inference task, which requires models to predict the relationships between premise and hypothesis based on a given image.\n• xGQA is a multilingual grounded question answering task based on GQA (Hudson and Manning, 2019) and machine translated question-answer pairs.\n• xFlickr&CO is a multilingual image-text retrieval dataset collected from Flickr30k (Plum-mer et al., 2015) and COCO (Lin et al., 2015) 4"},{"figure_ref":[],"heading":".3 Implementation and Hyperparameters","publication_ref":["b14","b28","b35","b0","b13","b0","b13","b13"],"table_ref":[],"text":"We conduct all experiments based on the Visiolinguistic Transformer Architectures framework VOLTA on four 2080Ti GPUs. We implement the MAML algorithm using the Higher library. We use the AdamW (Loshchilov and Hutter, 2018) optimizer to fine-tune all models in PyTorch.\nFine-tuning on English Data Before evaluating models on data in low-resource languages, we firstly fine-tune the pre-trained models on the corresponding English datasets: GQA (Hudson and Manning, 2019), NLVR2 (Suhr et al., 2019), SNLI-VE (Xie et al., 2019), andFlickr30k (Plummer et al., 2015) for xGQA, MaRVL, XVNLI, and xFlickr&Co, respectively, using the procedure of Bugliarello et al. (2022) and Liu et al. (2021). We follow the setting in IGLUE (Bugliarello et al., 2022) and also use the IGLUE hyper-parameters for each task when fine-tuning. We save the parameters of models in each epoch, then pick the best performing model for each task as the initialized parameters θ for the meta-learning fine-tuning stage.\nFine-tuning with Meta-learning For the XVL-MAML algorithm, the size of the support set and the query set is 64. We explore learning rates 5 × 10 -5 , 1 × 10 -5 , 5 × 10 -6 , 1 × 10 -6 for both UC2 and xUNITER, and find the best learning rate is 5 × 10 -6 for both the normal fine-tuning stage and the meta-update of MAML. For the inner learning rate of XVL-MAML, we explore learning rates 5 × 10 -6 , 5 × 10 -5 , 5 × 10 -4 and 5 × 10 -3 , and find that 5 × 10 -4 is the best inner learning rate.\nFor the proposed meta-learning framework, we find that models overfit after 300 iterations in most situations (for each iterations, we sample a batch of data as support set and a batch as query set), Table 2: Zero-shot performance (accuracy/consistency) of two baseline models fine-tuned only on English data (Base) and then fine-tuned by our meta-learning method (Ours) on the MaRVL dataset (Liu et al., 2021), where the definition of consistency following Liu et al. (2021). Columns indicate target languages. The avg column gives the average performance across all target languages in this row. zh → X means the auxiliary language is Chinese, and the target languages is other low-resource languages X. We also show the average and maximum performance across all auxiliary languages for each target language.\nso we set the number of iterations to 400 for all our experiments, and evaluate the performance of models for each 25 iterations to guarantee that we can pick the model with best performance of each setting for evaluation."},{"figure_ref":[],"heading":"Results and Discussion","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_3"],"heading":"Zero-shot","publication_ref":[],"table_ref":[],"text":"We report the results of the baseline models and the results for fine-tuning them using our metalearning framework in Table 1. In our setting, baseline model means that the PVLM is only fine-tuned on the English datasets. For simplicity, we report the averaged results of all combinations of target languages and auxiliary languages for each model and task. We set the value of λ in Equation ( 8) to 2 × 10 -2 for xUNITER and 5 × 10 -2 for UC2 to gain the best performance.\nThe results in the Table 1 indicate the effectiveness of our meta-learning framework and show that our method can boost the zero-shot performance of UC2 and XUNITER on all four datasets in IGLUE. Note that Table 1 shows average performance across all languages. The performance for individual languages can vary, and is shown in detail in Appendix A, Table 4. We also show the differences in improvements when using different auxiliary languages for different target languages in Figure 5."},{"figure_ref":[],"heading":"Few-shot","publication_ref":["b0"],"table_ref":[],"text":"We also conduct few-shot experiments following the setting in IGLUE (Bugliarello et al., 2022) Table 3: Ablation study in the unsupervised setting and supervised setting. The labels of the down-stream task data in the auxiliary language are not given in unsupervised setting and provided in supervised setting.\nboth xUNITER and UC2 on XVNLI and MaRVL.\nThe results are shown in Figure 2, where the horizontal axis represents the number of shots, and the vertical axis represents the accuracy score. The leftmost point of the horizontal axis is zero, which represents the performance in the zero-shot setup.\nThe blue points and lines show the performance of our method. The yellow points and lines represent the performance of the baseline. We have performed five runs and the interval represents the standard error. It is clear that in all four figures, our method achieves better performance across all shots. And it is worth noting that although there is a slight increase from the performance of zero-shot to one-shot, our proposed method, without seeing any data in the target languages, outperforms the baselines in the few-shot setting, except for UC2 on MaRVL. In other words, only a few instances of training data in target languages are not enough to eliminate the advantage of our method. This demonstrates that while our method requires training data in one auxiliary language, there is no need for few-shot data in the target languages."},{"figure_ref":[],"heading":"Ablation Study and Further Analysis","publication_ref":[],"table_ref":[],"text":"In this section, we conduct a series of ablation studies which investigate the effect of each part of our proposed meta-learning framework. We have performed five runs for each setting and reported the average and standard error to estimate significant differences.\nThe Effect of Contrastive Learning We investigate the effect of contrastive learning in our metalearning fine-tuning framework. Specifically, we fine-tune the model only using a contrastive learning loss in the MAML algorithm (called as \"XVL-MAML (w/o down-stream)\" in Table 3), where the labels of down-stream task data are not given. We evaluate the performance of UC2 and xUNITER on the XVNLI dataset in this setting and reported them in unsupervised setting part of Table 3. The results indicate that using contrastive learning solely in the MAML algorithm can improve performance. It provides evidence for the hypothesis that contrastive learning can enable models to learn alignments of modalities in cross-lingual transfer, resulting in better representations.\nWe also compare the performance of the model in the supervised setting where labels of data in auxiliary language are available; hence in the XVL-MAML algorithm, both contrastive loss and down-stream task loss are used. Then we remove the contrastive learning loss in XVL-MAML, only keeping the down-stream task loss. We compare the performance of these two settings in Table 3 to show the effectiveness of the contrastive learning loss in XVL-MAML in the supervised setting. In the \"Supervised Setting\" part of Table 3, the first row is XVL-MAML without contrastive learning loss, which means only using down-stream task loss when fine-tuning, and the second row is normal XVL-MAML using both contrastive loss and down-stream task loss.\nMoreover, we show the difference in performance in each target language separately in Figure 3. Contrastive learning can bring improvements for most of the target languages, especially those whose performance is relatively low when not using contrastive learning. For example, in the leftmost plot, performance in zh, ta, and sw is relatively lower than tr in the baseline, but gains significant improvements when using our method. The similar effect can be seen in other three plots and Table 2.\nDiverse down-stream tasks We report the results of experiments using four diverse multilingual vision-and-language understanding tasks in Table 1. Our method can bring clear improvements across all tasks for both UC2 and xUNITER, indicating that the approach generalises across tasks. Furthermore, these four IGLUE tasks also differ in the distribution of language families and domains, which indicates our method can be useful across language families and domains. Moreover, our method can significantly boost the performance of xUNITER even in the challenging MaRVL dataset which encompasses five diverse language families and cultures, improving accuracy by 4.4 points."},{"figure_ref":["fig_3"],"heading":"Diverse languages","publication_ref":[],"table_ref":[],"text":"We also investigate the difference of performance between languages. Specifically, we take the MaRVL dataset as an example and report results in Table 2, which lists the performance when using Chinese (zh) as the auxiliary language for meta-learning, and the average and maximum performance across all auxiliary languages for each target language respectively. In most situations, our method results in clear improvements. We then visualize the improvements of xUNITER when using different auxiliary languages for different target languages on MaRVL and XVNLI in Figure 5. The improvements we see for MaRVL (which range from 0.44 to 5.4) are smaller than for XVNLI (which range from 2.8 to 6.4), and one possible reason is that the language families of MaRVL are more diverse than those of XVNLI. But in general, our method improves performance for all combinations of auxiliary and target languages, even when they come from different language families. This further indicates that our method is language-agnostic."},{"figure_ref":[],"heading":"Example Predictions","publication_ref":[],"table_ref":[],"text":"We show some examples of inputs and predictions for baseline and our method in Figure 4. We use xUNITER to predict the Chinese part of the MaRVL dataset. We have selected two examples where the baseline prediction is incorrect, but our method predicts correctly (the rightmost two examples), and two examples where both our method and baseline method predict correctly (the leftmost two examples). In the two rightmost examples, the label is \"True\", but the baseline predicts \"False\". We find that in these two examples, the same concepts (\"church\" and \"drum\") described in related texts have different visual features, which makes it more difficult for models to identify them. In the left two examples, however, the concepts (panda and roses) described in the text do not have diverse or obscure visual features when they appear in the images. Therefore, based on these cases, we can surmise that the meta-learning framework makes the model more adaptive to diverse information and resulting in better generalization capabilities when mapping between texts and images.\nels for new languages in vision-and-language tasks. Our framework combines contrastive learning and downstream task supervised learning. We verify the effectiveness of our approach in both supervised and unsupervised settings. The key strength of our method is that we leverage contrastive learning in the MAML procedure so that models can quickly learn to align representations from different modalities and adapt them to unseen languages.\nExperimental results demonstrate that our proposed meta-learning framework significantly improves the performance of models in vision-andlanguage cross-lingual transfer both in zero-shot and few-shot setups. We applied our method to two representative PVLMs, UC2 and xUNITER, and verified its effectiveness on four datasets in the IGLUE benchmark in 14 languages. We also conducted an ablation study to explore the effect of contrastive learning, and analysed the effect of different languages and tasks."},{"figure_ref":[],"heading":"Limitations","publication_ref":[],"table_ref":[],"text":"Our proposed method applies contrastive learning to samples of image-text pairs. The alignments induced in this fashion work best if there is a concept or an object that is both depicted in the image and referred to in the sentence. If this is not the case, then the method may end up learning incorrect alignments; this includes cases where the image or the sentence contain multiple objects or concepts, not all of which can be aligned. To address this limitation, future work should explore how to construct better positive and negative samples and how to enable learning at a more fine-grained level. Besides, current famous PVLMs are encoder-only models, which is different with recent decoder-only LLMs, so meta-learning methods for multi-modal multilingual LLMs is worth to explore as a future work."},{"figure_ref":[],"heading":"Ethics Statement","publication_ref":["b0"],"table_ref":[],"text":"The use of the IGLUE benchmark in our paper is consistent with its intended use. We have checked the datasets for offensive content by sampling and visualizing examples. There are 14 languages in the datasets we use, we list them in Table 4. More detailed information about the IGLUE dataset can be found in (Bugliarello et al., 2022)."},{"figure_ref":[],"heading":"A Appendix","publication_ref":[],"table_ref":[],"text":"This is a section in the appendix. "}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b32\",\n \"b40\",\n \"b28\",\n \"b38\",\n \"b1\",\n \"b15\",\n \"b29\",\n \"b39\",\n \"b16\",\n \"b0\",\n \"b13\",\n \"b5\",\n \"b30\",\n \"b5\",\n \"b17\",\n \"b17\"\n ],\n \"table_ref\": [],\n \"text\": \"Multi-modal models focus on jointly learning representations from multiple modalities, such as vision and language. Many tasks require the integration of information of vision and language, including image captioning (Vinyals et al., 2015), natural language visual reasoning (Zhou et al., 2017;Suhr et al., 2019), and cross-modal retrieval (Zhen et al., 2019). Multi-modal learning captures the interaction between different modalities, allowing the resulting representations to be used in multimedia applications that enhance human-computer interaction.\\nRecently, pre-trained vision-language models (PVLMs; Chen et al. 2020;Lu et al. 2019;Tan and Bansal 2019) have achieved significant advances in multi-modal tasks. However, the data which PVLMs learn from is mostly for high-resource languages such as English. The resulting models rely on large amounts of training data for good performance, and often the models acquire biases that mean they perform poorly in low-resource languages such as Indonesian or Swahili. To address this, several multilingual PVLMs have been proposed (Zhou et al., 2021;Ni et al., 2021). A number of studies have used multilingual multimodal datasets (Bugliarello et al., 2022;Liu et al., 2021) and Figure 1 shows two examples from such datasets. The authors of these datasets used them to evaluate current famous PVLMs and demonstrated they do not perform well in low-resource cross-lingual transfer settings.\\nIn this paper, we conjecture that meta-learning can mitigate this issue. This is a learning approach that enables machine learning models to adapt quickly to new tasks by learning the learning algorithm itself. Model-agnostic Meta-learning (MAML; Finn et al. 2017) is one of the most widely used meta-learning frameworks. It is based on gradient-descent optimization, does not require multiple models or complex settings, and can be used for a range of models. In previous work (Verma et al., 2020;Finn et al., 2017;Nooralahzadeh et al., 2020), MAML-based methods have been shown to be useful in low-resource and cross-lingual transfer scenarios, including both few-shot and zero-shot cross-lingual tasks. However, prior work has only attempted to use MAML for cross-lingual transfer in text-only tasks (Nooralahzadeh et al., 2020).\\nInspired by previous works about using MAML for natural language tasks, this paper proposes XVL-MAML, a novel variant of MAML that addresses the limitations of previous PVLMs in vision-language tasks for low-resource crosslingual transfer. Our framework combines a traditional supervised loss for learning down-stream tasks with a contrastive loss to encourage the alignment between modalities, resulting in a crosslingual, multi-modal MAML optimization procedure.\\nThe intuition underlying our method is that a contrastive loss can align representations of different modalities, and MAML allows the model to generalize quickly to unseen tasks (languages, in our case). We show that XVL-MAML can lead to significant improvements in PVLM performance for cross-lingual transfer. We also find that using contrastive learning in a MAML framework on its own can bring improvements in PVLM performance in unsupervised settings.\\nIn sum, our contributions are as follows:\\n(1) We propose a novel framework called XVL-MAML which is the first meta-learning method specialized for vision-language cross-lingual transfer, and doesn't require the translation or pre-training data.\\n(2) We show that using only contrastive learning in the MAML framework in an unsupervised setting can also be useful. (3) We demonstrate that our proposed framework can boost the performance of current PVLMs across 14 languages and four tasks in both zero-shot learning and few-shot learning. (4) We conduct an ablation study to verify the effect of contrastive learning in both supervised and unsupervised settings and present an analysis across languages and tasks.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Multilingual Vision-and-Language Methods and Tasks\",\n \"publication_ref\": [\n \"b4\",\n \"b7\",\n \"b26\",\n \"b19\",\n \"b13\",\n \"b27\",\n \"b16\",\n \"b13\",\n \"b1\",\n \"b3\",\n \"b2\",\n \"b39\",\n \"b0\",\n \"b19\",\n \"b34\",\n \"b21\"\n ],\n \"table_ref\": [],\n \"text\": \"Recent work has investigated vision-and-language cross-lingual transfer tasks. Elliott et al. (2016) proposed Multi30K, an image description dataset which contains descriptions in multiple languages.\\nPrevious methods (Gella et al., 2017;Rotman et al., 2018) propose ways of bridging languages through images, but they mainly focus on imagetext retrieval and only consider high-resource languages such as English and German. Pfeiffer et al. (2022) built a multilingual visual question answer-ing dataset xGQA. Liu et al. (2021) proposed a multilingual version of the grounded visual reasoning dataset MaRVL, which follow the same setting as the natural language visual reasoning dataset NLVR2 (Su et al., 2019), but considers both cross-lingual transfer and domain shift between languages.\\nSeveral pre-trained models are recently proposed for vision-and-language cross-lingual transfer. Ni et al. (2021) proposed M3P, a transformer-based pre-trained model that maps the same concepts in different modalities and languages into a common semantic space. Similar to M3P, Liu et al. (2021) extended UNITER (Chen et al., 2020), proposing mUNITER based on M-BERT (Devlin et al., 2019), and xUNITER based on XLM-R (Conneau et al., 2020). Zhou et al. (2021) proposed UC2, a model using a data augmentation method based on machine translation for cross-lingual cross-modal pre-training. Although pre-training methods have proven powerful across multiple tasks, they require large amounts of training data and show a clear performance gap between English and low-resource languages on the IGLUE benchmark (Bugliarello et al., 2022).\\nRecently, some adapter-based efficient tuning methods (Pfeiffer et al., 2022;Wang et al., 2023) and translation augmented methods (Qiu et al., 2022) were proposed for multilingual multimodal tasks. But they still require a large amount of data or machine translated data for training. Our method, in contrast, only requires a small amount of auxiliary data.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Meta-Learning\",\n \"publication_ref\": [\n \"b31\",\n \"b5\",\n \"b24\",\n \"b36\",\n \"b10\",\n \"b33\",\n \"b5\",\n \"b10\",\n \"b18\",\n \"b8\",\n \"b17\"\n ],\n \"table_ref\": [],\n \"text\": \"Meta-learning has been increasingly popular in machine learning. Whereas conventional machine learning methods learn by data points, metalearning learns by tasks. Previous meta-learning work (Vinyals et al., 2016;Finn et al., 2017) focused on adapting to new tasks quickly. But metalearning can be applied to other scenarios as well, including semi-supervised learning (Ren et al., 2018), multi-task learning (Yu et al., 2020), and domain generalization (Li et al., 2018).\\nPrior work has also explored the effectiveness of meta-learning in NLP: Wang et al. (2021) applied meta-learning in semantic parsing for domain generalization based on MAML (Finn et al., 2017;Li et al., 2018). Obamuyide and Vlachos (2019) leveraged meta-learning under limited su-pervision in a relation classification task. Recently, there have been some applications using MAML in cross-lingual transfer: Gu et al. (2018) and Nooralahzadeh et al. (2020) regard languages as tasks in their meta-learning framework. In contrast to these existing approaches, which explore text-only scenarios, we are the first to utilize metalearning for cross-lingual transfer in multi-modal tasks.\\n3 Meta-learning for Vision-and-language Cross-lingual Transfer\\nWe first formally define the problem of vision-and-Language cross-lingual transfer in the context of zero-shot and few-shot scenarios in Section 3.1. Then, we introduce our overall fine-tuning framework in Section 3.2. And we introduce the contrastive learning used for vision-and-language tasks in Section 3.3. Finally, we introduce our XVL-MAML algorithm in Section 3.4.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Problem Definition\",\n \"publication_ref\": [\n \"b0\"\n ],\n \"table_ref\": [],\n \"text\": \"Following the multilingual vision-language IGLUE benchmark (Bugliarello et al., 2022), we formulate the problem of cross-lingual transfer learning in vision-and-language scenarios. For understanding tasks, the input is a pair of an image V and text U , and the output Y is the result inferred by the multimodal model. We can thus formulate this problem as computing P θ (Y |V, U ), where θ are the parameters of the PVLMs. During training, the image-text pairs come from datasets D s in a set of source languages, and our aim is to perform well on datasets D t for the same task in the target languages. For the zero-shot setup, the pre-trained model fine-tuned on D s is directly used in inference on D t for unseen target languages. For the few-shot setup, after training on D s , the model is continually fine-tuned on several shots of the training set of D t and then evaluated on the development set of D t .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Overall Fine-tuning Framework For Cross-lingual Transfer\",\n \"publication_ref\": [\n \"b0\"\n ],\n \"table_ref\": [],\n \"text\": \"The pipeline of our proposed meta-learning finetuning framework can be divided into three parts:\\n1. Fine-tune the pre-trained vision-language model on data of the down-stream task in English 2. Fine-tune the model on data in the auxiliary language (one language other than English) using our proposed XVL-MAML algorithm.\\n3. Evaluate the fine-tuned model on data in the target languages (languages other than English and the auxiliary language).\\nThe traditional cross-lingual transfer learning procedure described in Bugliarello et al. (2022) only includes part 1 and 3. In part 3, if the setting is zero-shot, the model is evaluated on data in the target language directly, but if the setting is few-shot, the model continues to be fine-tuned on few-shot data in the target languages and is then evaluated. The difference between our framework and the traditional procedure is the additional finetuning step of part 2. We will describe it specifically in Section 3.4, but before that, we will introduce contrastive learning for vision-and-language tasks.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Contrastive Learning for Vision-and-language Tasks\",\n \"publication_ref\": [\n \"b37\",\n \"b22\",\n \"b37\"\n ],\n \"table_ref\": [],\n \"text\": \"The vision-and-language contrastive learning loss proposed by Zhang et al. (2020) has proven effective in medical image scenarios and is used as the pre-training objective function of CLIP (Radford et al., 2021). It can be regarded as an auxiliary task for representation learning, aiming to enable models to gain better aligned multi-modal representation for downstream tasks. In the contrastive learning scheme, a batch of embeddings of images encoded by the model can be written as I = {I 1 , ..., I N }, and a batch of embeddings of texts encoded by the model can be written as \\nT = {T 1 , ..., T N },\\nU = I • W ⊤ 1 (1) V = T • W ⊤ 2 (2)\\nWhere U and V represent the batch of image-text pairs. Then the cosine similarity of each pair can be computed as\\n⟨U i , V j ⟩ = U ⊤ i V j ∥U i ∥∥V j ∥ .\\nThe objective is to maximize the similarity of matched image-text pairs and minimize the similarity of others. So the image-text contrastive loss can be formulated as follows:\\nL 1 i = -log exp(⟨U i , V i ⟩) N K=1 exp(⟨U i , V k ⟩)(3)\\nFollowing Zhang et al. (2020), the contrastive loss should be symmetric for each modality, and the text-image contrastive loss is:\\nL 2 i = -log exp(⟨V i , U i ⟩) N K=1 exp(⟨V i , U k ⟩) (4)\\nThe final contrastive loss of this batch of paired data is then:\\nL CL = N i=1 (L 1 i + L 2 i ) (5)\\nWhere L CL is the overall contrastive loss. When we minimize L CL , we maximize the similarity of image-text pairs which are positive examples.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"XVL-MAML\",\n \"publication_ref\": [\n \"b17\"\n ],\n \"table_ref\": [],\n \"text\": \"Inspired by the effectiveness of MAML for quickly adapting to new tasks, we propose a novel variant of the MAML algorithm specialized for crosslingual transfer in vision and language tasks, called XVL-MAML. Specifically, we first integrate contrastive learning into the MAML algorithm, making it specialized for the visual-language task of crosslingual transfer learning. Our intuition is that we can use MAML with a contrastive loss as its learning objective for quickly adapting vision-language alignment to new languages. In this framework, the alignment between image and text in a specific language can be regarded as a task. Inspired by Nooralahzadeh et al. (2020), we use the data of one auxiliary language for fine-tuning, but with a contrastive loss as objective function in the MAML algorithm. Specifically, we sample a batch of support data B s and a batch of query data B q in the data in auxiliary language A for each virtual task T . Assuming the parameters of the model are θ and the contrastive loss on the support data is L CL (θ) Bs , then the parameters of the model can be updated by one step of gradient descent:\\nθ ′ = θ -α∇ θ L CL (θ) Bs (6)\\nFollowing the MAML algorithm, our final objective for this task is to minimize L CL (θ ′ ) Bq on the query data B q using gradient descent:\\nθ ← θ -β∇ θ L CL (θ ′ ) Bq (7) θ ← θ -β∇ θ L CL (θ -α∇ θ L CL (θ) Bs ) Bq (8)\\nOptimized using this method, pre-trained visionlanguage models can quickly adapt to new tasks in other languages without using any annotation in the auxiliary language for downstream tasks, so we will refer to this as an unsupervised scenario.\\nIn supervised scenarios, where the downstream tasks labels in the auxiliary language are available, we combine the loss of the downstream task L with the vision-language contrastive loss L CL by adding them together. So during fine-tuning, Equation ( 8) is modified to:\\nθ ← θ -β(∇ θ L(θ ′′ ) Bq + λ∇ θ L CL (θ ′ ) Bq ) (9)\\nWhere the temporary parameters optimized for one step by the downstream task loss L on the support set B s is θ ′′ , β is the meta-learning rate, and λ is the scale factor of contrastive learning. By simply adding the gradients of the downstream task and contrastive learning in the meta-update, the model learns downstream tasks and vision-language alignment simultaneously for cross-lingual transfer.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we introduce both the base PVLMs we use for vision-language cross-lingual transfer, as well as the datasets and metrics we use to evaluate our proposed method. Then we describe how the experiments were conducted and discuss the results.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Base models\",\n \"publication_ref\": [\n \"b13\",\n \"b39\",\n \"b1\",\n \"b25\",\n \"b0\"\n ],\n \"table_ref\": [],\n \"text\": \"In this paper, we choose xUNITER (Liu et al., 2021) and UC2 (Zhou et al., 2021) as our base models, as they use different pre-training methods.\\nThen we applied XVL-MAML to both models to show that this method works across different models.\\nxUNITER is a multilingual version of the UNITER model (Chen et al., 2020). It has a similar architecture to UNITER and uses Faster-RCNN (Ren et al., 2015) 64.4 (+1.9) 29.9 (+0.9) 57.0 (+0.6) 21.3 (+1.6) 18.7 (+1.7)\\nTable 1: Zero-shot performance (accuracy) of four baseline models only fine-tuned on English data (Baseline) and two models fine-tuned by our meta-learning method (Ours) on four IGLUE datasets (Bugliarello et al., 2022).\\nUC2 uses a similar model architecture as UNITER, but different pre-training methods. Specifically, UC2 augments pre-training on English data by constructing a multilingual corpus via machine translation and then uses this augmented data for pre-training. It also proposes the Visual Translation Language Modeling (VTLM) pre-training method, which uses the image as a pivot to learn the relationship between parallel texts in two languages and their corresponding images.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Datasets and Metrics\",\n \"publication_ref\": [\n \"b0\",\n \"b19\",\n \"b13\",\n \"b20\",\n \"b12\",\n \"b28\",\n \"b11\"\n ],\n \"table_ref\": [],\n \"text\": \"We use datasets for four tasks from the IGLUE benchmark (Bugliarello et al., 2022), which includes xGQA (Pfeiffer et al., 2022), MaRVL (Liu et al., 2021), XVNLI, and xFlickr&Co (Plummer et al., 2015;Lin et al., 2014). We show examples from MaRVL and XVNLI in Figure 1. Following the convention in IGLUE, the evaluation metric is accuracy for all tasks except cross-modal retrieval, which uses Recall@1. The task format of these four datasets are described below:\\n• MaRVL is a multicutural vision-language reasoning dataset, following the format of English NLVR2 (Suhr et al., 2019) which namely to judge whether a sentence is correct or not for a pair of images.\\n• XVNLI is a multilingual version of visual natural language inference task, which requires models to predict the relationships between premise and hypothesis based on a given image.\\n• xGQA is a multilingual grounded question answering task based on GQA (Hudson and Manning, 2019) and machine translated question-answer pairs.\\n• xFlickr&CO is a multilingual image-text retrieval dataset collected from Flickr30k (Plum-mer et al., 2015) and COCO (Lin et al., 2015) 4\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \".3 Implementation and Hyperparameters\",\n \"publication_ref\": [\n \"b14\",\n \"b28\",\n \"b35\",\n \"b0\",\n \"b13\",\n \"b0\",\n \"b13\",\n \"b13\"\n ],\n \"table_ref\": [],\n \"text\": \"We conduct all experiments based on the Visiolinguistic Transformer Architectures framework VOLTA on four 2080Ti GPUs. We implement the MAML algorithm using the Higher library. We use the AdamW (Loshchilov and Hutter, 2018) optimizer to fine-tune all models in PyTorch.\\nFine-tuning on English Data Before evaluating models on data in low-resource languages, we firstly fine-tune the pre-trained models on the corresponding English datasets: GQA (Hudson and Manning, 2019), NLVR2 (Suhr et al., 2019), SNLI-VE (Xie et al., 2019), andFlickr30k (Plummer et al., 2015) for xGQA, MaRVL, XVNLI, and xFlickr&Co, respectively, using the procedure of Bugliarello et al. (2022) and Liu et al. (2021). We follow the setting in IGLUE (Bugliarello et al., 2022) and also use the IGLUE hyper-parameters for each task when fine-tuning. We save the parameters of models in each epoch, then pick the best performing model for each task as the initialized parameters θ for the meta-learning fine-tuning stage.\\nFine-tuning with Meta-learning For the XVL-MAML algorithm, the size of the support set and the query set is 64. We explore learning rates 5 × 10 -5 , 1 × 10 -5 , 5 × 10 -6 , 1 × 10 -6 for both UC2 and xUNITER, and find the best learning rate is 5 × 10 -6 for both the normal fine-tuning stage and the meta-update of MAML. For the inner learning rate of XVL-MAML, we explore learning rates 5 × 10 -6 , 5 × 10 -5 , 5 × 10 -4 and 5 × 10 -3 , and find that 5 × 10 -4 is the best inner learning rate.\\nFor the proposed meta-learning framework, we find that models overfit after 300 iterations in most situations (for each iterations, we sample a batch of data as support set and a batch as query set), Table 2: Zero-shot performance (accuracy/consistency) of two baseline models fine-tuned only on English data (Base) and then fine-tuned by our meta-learning method (Ours) on the MaRVL dataset (Liu et al., 2021), where the definition of consistency following Liu et al. (2021). Columns indicate target languages. The avg column gives the average performance across all target languages in this row. zh → X means the auxiliary language is Chinese, and the target languages is other low-resource languages X. We also show the average and maximum performance across all auxiliary languages for each target language.\\nso we set the number of iterations to 400 for all our experiments, and evaluate the performance of models for each 25 iterations to guarantee that we can pick the model with best performance of each setting for evaluation.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results and Discussion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Zero-shot\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We report the results of the baseline models and the results for fine-tuning them using our metalearning framework in Table 1. In our setting, baseline model means that the PVLM is only fine-tuned on the English datasets. For simplicity, we report the averaged results of all combinations of target languages and auxiliary languages for each model and task. We set the value of λ in Equation ( 8) to 2 × 10 -2 for xUNITER and 5 × 10 -2 for UC2 to gain the best performance.\\nThe results in the Table 1 indicate the effectiveness of our meta-learning framework and show that our method can boost the zero-shot performance of UC2 and XUNITER on all four datasets in IGLUE. Note that Table 1 shows average performance across all languages. The performance for individual languages can vary, and is shown in detail in Appendix A, Table 4. We also show the differences in improvements when using different auxiliary languages for different target languages in Figure 5.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Few-shot\",\n \"publication_ref\": [\n \"b0\"\n ],\n \"table_ref\": [],\n \"text\": \"We also conduct few-shot experiments following the setting in IGLUE (Bugliarello et al., 2022) Table 3: Ablation study in the unsupervised setting and supervised setting. The labels of the down-stream task data in the auxiliary language are not given in unsupervised setting and provided in supervised setting.\\nboth xUNITER and UC2 on XVNLI and MaRVL.\\nThe results are shown in Figure 2, where the horizontal axis represents the number of shots, and the vertical axis represents the accuracy score. The leftmost point of the horizontal axis is zero, which represents the performance in the zero-shot setup.\\nThe blue points and lines show the performance of our method. The yellow points and lines represent the performance of the baseline. We have performed five runs and the interval represents the standard error. It is clear that in all four figures, our method achieves better performance across all shots. And it is worth noting that although there is a slight increase from the performance of zero-shot to one-shot, our proposed method, without seeing any data in the target languages, outperforms the baselines in the few-shot setting, except for UC2 on MaRVL. In other words, only a few instances of training data in target languages are not enough to eliminate the advantage of our method. This demonstrates that while our method requires training data in one auxiliary language, there is no need for few-shot data in the target languages.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ablation Study and Further Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we conduct a series of ablation studies which investigate the effect of each part of our proposed meta-learning framework. We have performed five runs for each setting and reported the average and standard error to estimate significant differences.\\nThe Effect of Contrastive Learning We investigate the effect of contrastive learning in our metalearning fine-tuning framework. Specifically, we fine-tune the model only using a contrastive learning loss in the MAML algorithm (called as \\\"XVL-MAML (w/o down-stream)\\\" in Table 3), where the labels of down-stream task data are not given. We evaluate the performance of UC2 and xUNITER on the XVNLI dataset in this setting and reported them in unsupervised setting part of Table 3. The results indicate that using contrastive learning solely in the MAML algorithm can improve performance. It provides evidence for the hypothesis that contrastive learning can enable models to learn alignments of modalities in cross-lingual transfer, resulting in better representations.\\nWe also compare the performance of the model in the supervised setting where labels of data in auxiliary language are available; hence in the XVL-MAML algorithm, both contrastive loss and down-stream task loss are used. Then we remove the contrastive learning loss in XVL-MAML, only keeping the down-stream task loss. We compare the performance of these two settings in Table 3 to show the effectiveness of the contrastive learning loss in XVL-MAML in the supervised setting. In the \\\"Supervised Setting\\\" part of Table 3, the first row is XVL-MAML without contrastive learning loss, which means only using down-stream task loss when fine-tuning, and the second row is normal XVL-MAML using both contrastive loss and down-stream task loss.\\nMoreover, we show the difference in performance in each target language separately in Figure 3. Contrastive learning can bring improvements for most of the target languages, especially those whose performance is relatively low when not using contrastive learning. For example, in the leftmost plot, performance in zh, ta, and sw is relatively lower than tr in the baseline, but gains significant improvements when using our method. The similar effect can be seen in other three plots and Table 2.\\nDiverse down-stream tasks We report the results of experiments using four diverse multilingual vision-and-language understanding tasks in Table 1. Our method can bring clear improvements across all tasks for both UC2 and xUNITER, indicating that the approach generalises across tasks. Furthermore, these four IGLUE tasks also differ in the distribution of language families and domains, which indicates our method can be useful across language families and domains. Moreover, our method can significantly boost the performance of xUNITER even in the challenging MaRVL dataset which encompasses five diverse language families and cultures, improving accuracy by 4.4 points.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Diverse languages\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We also investigate the difference of performance between languages. Specifically, we take the MaRVL dataset as an example and report results in Table 2, which lists the performance when using Chinese (zh) as the auxiliary language for meta-learning, and the average and maximum performance across all auxiliary languages for each target language respectively. In most situations, our method results in clear improvements. We then visualize the improvements of xUNITER when using different auxiliary languages for different target languages on MaRVL and XVNLI in Figure 5. The improvements we see for MaRVL (which range from 0.44 to 5.4) are smaller than for XVNLI (which range from 2.8 to 6.4), and one possible reason is that the language families of MaRVL are more diverse than those of XVNLI. But in general, our method improves performance for all combinations of auxiliary and target languages, even when they come from different language families. This further indicates that our method is language-agnostic.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Example Predictions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We show some examples of inputs and predictions for baseline and our method in Figure 4. We use xUNITER to predict the Chinese part of the MaRVL dataset. We have selected two examples where the baseline prediction is incorrect, but our method predicts correctly (the rightmost two examples), and two examples where both our method and baseline method predict correctly (the leftmost two examples). In the two rightmost examples, the label is \\\"True\\\", but the baseline predicts \\\"False\\\". We find that in these two examples, the same concepts (\\\"church\\\" and \\\"drum\\\") described in related texts have different visual features, which makes it more difficult for models to identify them. In the left two examples, however, the concepts (panda and roses) described in the text do not have diverse or obscure visual features when they appear in the images. Therefore, based on these cases, we can surmise that the meta-learning framework makes the model more adaptive to diverse information and resulting in better generalization capabilities when mapping between texts and images.\\nels for new languages in vision-and-language tasks. Our framework combines contrastive learning and downstream task supervised learning. We verify the effectiveness of our approach in both supervised and unsupervised settings. The key strength of our method is that we leverage contrastive learning in the MAML procedure so that models can quickly learn to align representations from different modalities and adapt them to unseen languages.\\nExperimental results demonstrate that our proposed meta-learning framework significantly improves the performance of models in vision-andlanguage cross-lingual transfer both in zero-shot and few-shot setups. We applied our method to two representative PVLMs, UC2 and xUNITER, and verified its effectiveness on four datasets in the IGLUE benchmark in 14 languages. We also conducted an ablation study to explore the effect of contrastive learning, and analysed the effect of different languages and tasks.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Our proposed method applies contrastive learning to samples of image-text pairs. The alignments induced in this fashion work best if there is a concept or an object that is both depicted in the image and referred to in the sentence. If this is not the case, then the method may end up learning incorrect alignments; this includes cases where the image or the sentence contain multiple objects or concepts, not all of which can be aligned. To address this limitation, future work should explore how to construct better positive and negative samples and how to enable learning at a more fine-grained level. Besides, current famous PVLMs are encoder-only models, which is different with recent decoder-only LLMs, so meta-learning methods for multi-modal multilingual LLMs is worth to explore as a future work.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ethics Statement\",\n \"publication_ref\": [\n \"b0\"\n ],\n \"table_ref\": [],\n \"text\": \"The use of the IGLUE benchmark in our paper is consistent with its intended use. We have checked the datasets for offensive content by sampling and visualizing examples. There are 14 languages in the datasets we use, we list them in Table 4. More detailed information about the IGLUE dataset can be found in (Bugliarello et al., 2022).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Appendix\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This is a section in the appendix. \"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-10-24"},"doi":{"kind":"string","value":"10.18653/v1/D17-1303"},"references":{"kind":"list like","value":[{"authors":"Emanuele Bugliarello; Fangyu Liu; Jonas Pfeiffer; Siva Reddy; Desmond Elliott; Edoardo ; Maria Ponti; Ivan Vulić","journal":"","ref_id":"b0","title":"Iglue: A benchmark for transfer learning across modalities, tasks, and languages","year":"2022"},{"authors":"Yen-Chun Chen; Linjie Li; Licheng Yu; Ahmed El Kholy; Faisal Ahmed; Zhe Gan; Yu Cheng; Jingjing Liu","journal":"Springer","ref_id":"b1","title":"Uniter: Universal image-text representation learning","year":"2020"},{"authors":"Alexis Conneau; Kartikay Khandelwal; Naman Goyal; Vishrav Chaudhary; Guillaume Wenzek; Francisco Guzmán; Édouard Grave; Myle Ott; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b2","title":"Unsupervised cross-lingual representation learning at scale","year":"2020"},{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"","ref_id":"b3","title":"Bert: Pre-training of deep bidirectional transformers for language understanding","year":"2019"},{"authors":"Desmond Elliott; Stella Frank; Khalil Sima'an; Lucia Specia","journal":"","ref_id":"b4","title":"Multi30k: Multilingual english-german image descriptions","year":"2016"},{"authors":"Chelsea Finn; Pieter Abbeel; Sergey Levine","journal":"","ref_id":"b5","title":"Model-agnostic meta-learning for fast adaptation of deep networks","year":"2017"},{"authors":" Pmlr","journal":"","ref_id":"b6","title":"","year":""},{"authors":"Spandana Gella; Rico Sennrich; Frank Keller; Mirella Lapata","journal":"Association for Computational Linguistics","ref_id":"b7","title":"Image pivoting for learning multilingual multimodal representations","year":"2017"},{"authors":"Jiatao Gu; Yong Wang; Yun Chen; O K Victor; Kyunghyun Li; Cho","journal":"Association for Computational Linguistics","ref_id":"b8","title":"Meta-learning for lowresource neural machine translation","year":"2018"},{"authors":"A Drew; Christopher D Hudson; Manning","journal":"","ref_id":"b9","title":"Gqa: A new dataset for real-world visual reasoning and compositional question answering","year":"2019"},{"authors":"Da Li; Yongxin Yang; Yi-Zhe Song; Timothy Hospedales","journal":"","ref_id":"b10","title":"Learning to generalize: Metalearning for domain generalization","year":"2018"},{"authors":"Tsung-Yi Lin; Michael Maire; Serge Belongie; Lubomir Bourdev; Ross Girshick; James Hays; Pietro Perona; Deva Ramanan; C Lawrence Zitnick; Piotr Dollár","journal":"","ref_id":"b11","title":"Microsoft coco: Common objects in context","year":"2015"},{"authors":"Tsung-Yi Lin; Michael Maire; Serge Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence; Zitnick ","journal":"Springer","ref_id":"b12","title":"Microsoft coco: Common objects in context","year":"2014"},{"authors":"Fangyu Liu; Emanuele Bugliarello; Maria Edoardo; Siva Ponti; Nigel Reddy; Desmond Collier; Elliott","journal":"","ref_id":"b13","title":"Visually grounded reasoning across languages and cultures","year":"2021"},{"authors":"Ilya Loshchilov; Frank Hutter","journal":"","ref_id":"b14","title":"Decoupled weight decay regularization","year":"2018"},{"authors":"Jiasen Lu; Dhruv Batra; Devi Parikh; Stefan Lee","journal":"Advances in neural information processing systems","ref_id":"b15","title":"Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks","year":"2019"},{"authors":"Minheng Ni; Haoyang Huang; Lin Su; Edward Cui; Taroon Bharti; Lijuan Wang; Dongdong Zhang; Nan Duan","journal":"","ref_id":"b16","title":"M3p: Learning universal representations via multitask multilingual multimodal pretraining","year":"2021"},{"authors":"Farhad Nooralahzadeh; Giannis Bekoulis; Johannes Bjerva; Isabelle Augenstein","journal":"","ref_id":"b17","title":"Zero-shot cross-lingual transfer with meta learning","year":"2020"},{"authors":"Abiola Obamuyide; Andreas Vlachos","journal":"","ref_id":"b18","title":"Modelagnostic meta-learning for relation classification with limited supervision","year":"2019"},{"authors":"Jonas Pfeiffer; Gregor Geigle; Aishwarya Kamath; Jan-Martin Steitz; Stefan Roth; Ivan Vulić; Iryna Gurevych","journal":"","ref_id":"b19","title":"xgqa: Cross-lingual visual question answering","year":"2022"},{"authors":"Liwei Bryan A Plummer; Chris M Wang; Juan C Cervantes; Julia Caicedo; Svetlana Hockenmaier; Lazebnik","journal":"","ref_id":"b20","title":"Flickr30k entities: Collecting region-to-phrase correspondences for richer imageto-sentence models","year":"2015"},{"authors":"Chen Qiu; Dan Oneata; Emanuele Bugliarello; Stella Frank; Desmond Elliott","journal":"","ref_id":"b21","title":"Multilingual multimodal learning with machine translated text","year":"2022"},{"authors":"Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark","journal":"","ref_id":"b22","title":"Learning transferable visual models from natural language supervision","year":"2021"},{"authors":" Pmlr","journal":"","ref_id":"b23","title":"","year":""},{"authors":"Mengye Ren; Eleni Triantafillou; Sachin Ravi; Jake Snell; Kevin Swersky; Joshua B Tenenbaum; Hugo Larochelle; Richard S Zemel","journal":"","ref_id":"b24","title":"Metalearning for semi-supervised few-shot classification","year":"2018"},{"authors":"Kaiming Shaoqing Ren; Ross He; Jian Girshick; Sun","journal":"Advances in neural information processing systems","ref_id":"b25","title":"Faster r-cnn: Towards real-time object detection with region proposal networks","year":"2015"},{"authors":"Guy Rotman; Ivan Vulić; Roi Reichart","journal":"Association for Computational Linguistics","ref_id":"b26","title":"Bridging languages through images with deep partial canonical correlation analysis","year":"2018"},{"authors":"Weijie Su; Xizhou Zhu; Yue Cao; Bin Li; Lewei Lu; Furu Wei; Jifeng Dai","journal":"","ref_id":"b27","title":"Vl-bert: Pre-training of generic visual-linguistic representations","year":"2019"},{"authors":"Alane Suhr; Stephanie Zhou; Ally Zhang; Iris Zhang; Huajun Bai; Yoav Artzi","journal":"","ref_id":"b28","title":"A corpus for reasoning about natural language grounded in photographs","year":"2019"},{"authors":"Hao Tan; Mohit Bansal","journal":"","ref_id":"b29","title":"Lxmert: Learning cross-modality encoder representations from transformers","year":"2019"},{"authors":"Kumar Vinay; Dhanajit Verma; Piyush Brahma; Rai","journal":"","ref_id":"b30","title":"Meta-learning for generalized zero-shot learning","year":"2020"},{"authors":"Oriol Vinyals; Charles Blundell; Timothy Lillicrap; Daan Wierstra","journal":"Advances in neural information processing systems","ref_id":"b31","title":"Matching networks for one shot learning","year":"2016"},{"authors":"Oriol Vinyals; Alexander Toshev; Samy Bengio; Dumitru Erhan","journal":"","ref_id":"b32","title":"Show and tell: A neural image caption generator","year":"2015"},{"authors":"Bailin Wang; Mirella Lapata; Ivan Titov","journal":"","ref_id":"b33","title":"Meta-learning for domain generalization in semantic parsing","year":"2021"},{"authors":"Ying Wang; Jonas Pfeiffer; Nicolas Carion; Yann Lecun; Aishwarya Kamath","journal":"","ref_id":"b34","title":"Adapting grounded visual question answering models to low resource languages","year":"2023"},{"authors":"Ning Xie; Farley Lai; Derek Doran; Asim Kadav","journal":"","ref_id":"b35","title":"Visual entailment: A novel task for fine-grained image understanding","year":"2019"},{"authors":"Tianhe Yu; Saurabh Kumar; Abhishek Gupta; Sergey Levine; Karol Hausman; Chelsea Finn","journal":"Advances in Neural Information Processing Systems","ref_id":"b36","title":"Gradient surgery for multi-task learning","year":"2020"},{"authors":"Yuhao Zhang; Hang Jiang; Yasuhide Miura; Christopher D Manning; Curtis P Langlotz","journal":"","ref_id":"b37","title":"Contrastive learning of medical visual representations from paired images and text","year":"2020"},{"authors":"Liangli Zhen; Peng Hu; Xu Wang; Dezhong Peng","journal":"","ref_id":"b38","title":"Deep supervised cross-modal retrieval","year":"2019"},{"authors":"Mingyang Zhou; Luowei Zhou; Shuohang Wang; Yu Cheng; Linjie Li; Zhou Yu; Jingjing Liu","journal":"","ref_id":"b39","title":"Uc2: Universal cross-lingual cross-modal vision-and-language pre-training","year":"2021"},{"authors":"Stephanie Zhou; Alane Suhr; Yoav Artzi","journal":"","ref_id":"b40","title":"Visual reasoning with natural language","year":"2017"}],"string":"[\n {\n \"authors\": \"Emanuele Bugliarello; Fangyu Liu; Jonas Pfeiffer; Siva Reddy; Desmond Elliott; Edoardo ; Maria Ponti; Ivan Vulić\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Iglue: A benchmark for transfer learning across modalities, tasks, and languages\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yen-Chun Chen; Linjie Li; Licheng Yu; Ahmed El Kholy; Faisal Ahmed; Zhe Gan; Yu Cheng; Jingjing Liu\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b1\",\n \"title\": \"Uniter: Universal image-text representation learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Alexis Conneau; Kartikay Khandelwal; Naman Goyal; Vishrav Chaudhary; Guillaume Wenzek; Francisco Guzmán; Édouard Grave; Myle Ott; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Unsupervised cross-lingual representation learning at scale\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Bert: Pre-training of deep bidirectional transformers for language understanding\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Desmond Elliott; Stella Frank; Khalil Sima'an; Lucia Specia\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Multi30k: Multilingual english-german image descriptions\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Chelsea Finn; Pieter Abbeel; Sergey Levine\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Model-agnostic meta-learning for fast adaptation of deep networks\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \" Pmlr\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Spandana Gella; Rico Sennrich; Frank Keller; Mirella Lapata\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b7\",\n \"title\": \"Image pivoting for learning multilingual multimodal representations\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Jiatao Gu; Yong Wang; Yun Chen; O K Victor; Kyunghyun Li; Cho\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b8\",\n \"title\": \"Meta-learning for lowresource neural machine translation\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"A Drew; Christopher D Hudson; Manning\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Gqa: A new dataset for real-world visual reasoning and compositional question answering\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Da Li; Yongxin Yang; Yi-Zhe Song; Timothy Hospedales\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Learning to generalize: Metalearning for domain generalization\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Tsung-Yi Lin; Michael Maire; Serge Belongie; Lubomir Bourdev; Ross Girshick; James Hays; Pietro Perona; Deva Ramanan; C Lawrence Zitnick; Piotr Dollár\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Microsoft coco: Common objects in context\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Tsung-Yi Lin; Michael Maire; Serge Belongie; James Hays; Pietro Perona; Deva Ramanan; Piotr Dollár; C Lawrence; Zitnick \",\n \"journal\": \"Springer\",\n \"ref_id\": \"b12\",\n \"title\": \"Microsoft coco: Common objects in context\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Fangyu Liu; Emanuele Bugliarello; Maria Edoardo; Siva Ponti; Nigel Reddy; Desmond Collier; Elliott\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Visually grounded reasoning across languages and cultures\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ilya Loshchilov; Frank Hutter\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Decoupled weight decay regularization\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Jiasen Lu; Dhruv Batra; Devi Parikh; Stefan Lee\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b15\",\n \"title\": \"Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Minheng Ni; Haoyang Huang; Lin Su; Edward Cui; Taroon Bharti; Lijuan Wang; Dongdong Zhang; Nan Duan\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"M3p: Learning universal representations via multitask multilingual multimodal pretraining\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Farhad Nooralahzadeh; Giannis Bekoulis; Johannes Bjerva; Isabelle Augenstein\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Zero-shot cross-lingual transfer with meta learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Abiola Obamuyide; Andreas Vlachos\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Modelagnostic meta-learning for relation classification with limited supervision\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Jonas Pfeiffer; Gregor Geigle; Aishwarya Kamath; Jan-Martin Steitz; Stefan Roth; Ivan Vulić; Iryna Gurevych\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"xgqa: Cross-lingual visual question answering\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Liwei Bryan A Plummer; Chris M Wang; Juan C Cervantes; Julia Caicedo; Svetlana Hockenmaier; Lazebnik\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Flickr30k entities: Collecting region-to-phrase correspondences for richer imageto-sentence models\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Chen Qiu; Dan Oneata; Emanuele Bugliarello; Stella Frank; Desmond Elliott\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Multilingual multimodal learning with machine translated text\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Learning transferable visual models from natural language supervision\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \" Pmlr\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Mengye Ren; Eleni Triantafillou; Sachin Ravi; Jake Snell; Kevin Swersky; Joshua B Tenenbaum; Hugo Larochelle; Richard S Zemel\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Metalearning for semi-supervised few-shot classification\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Kaiming Shaoqing Ren; Ross He; Jian Girshick; Sun\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b25\",\n \"title\": \"Faster r-cnn: Towards real-time object detection with region proposal networks\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Guy Rotman; Ivan Vulić; Roi Reichart\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b26\",\n \"title\": \"Bridging languages through images with deep partial canonical correlation analysis\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Weijie Su; Xizhou Zhu; Yue Cao; Bin Li; Lewei Lu; Furu Wei; Jifeng Dai\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Vl-bert: Pre-training of generic visual-linguistic representations\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Alane Suhr; Stephanie Zhou; Ally Zhang; Iris Zhang; Huajun Bai; Yoav Artzi\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"A corpus for reasoning about natural language grounded in photographs\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Hao Tan; Mohit Bansal\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Lxmert: Learning cross-modality encoder representations from transformers\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Kumar Vinay; Dhanajit Verma; Piyush Brahma; Rai\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Meta-learning for generalized zero-shot learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Oriol Vinyals; Charles Blundell; Timothy Lillicrap; Daan Wierstra\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b31\",\n \"title\": \"Matching networks for one shot learning\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Oriol Vinyals; Alexander Toshev; Samy Bengio; Dumitru Erhan\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Show and tell: A neural image caption generator\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Bailin Wang; Mirella Lapata; Ivan Titov\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Meta-learning for domain generalization in semantic parsing\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ying Wang; Jonas Pfeiffer; Nicolas Carion; Yann Lecun; Aishwarya Kamath\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Adapting grounded visual question answering models to low resource languages\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Ning Xie; Farley Lai; Derek Doran; Asim Kadav\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Visual entailment: A novel task for fine-grained image understanding\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Tianhe Yu; Saurabh Kumar; Abhishek Gupta; Sergey Levine; Karol Hausman; Chelsea Finn\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b36\",\n \"title\": \"Gradient surgery for multi-task learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yuhao Zhang; Hang Jiang; Yasuhide Miura; Christopher D Manning; Curtis P Langlotz\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Contrastive learning of medical visual representations from paired images and text\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Liangli Zhen; Peng Hu; Xu Wang; Dezhong Peng\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Deep supervised cross-modal retrieval\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Mingyang Zhou; Luowei Zhou; Shuohang Wang; Yu Cheng; Linjie Li; Zhou Yu; Jingjing Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Uc2: Universal cross-lingual cross-modal vision-and-language pre-training\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Stephanie Zhou; Alane Suhr; Yoav Artzi\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Visual reasoning with natural language\",\n \"year\": \"2017\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,306.14,500.51,83.33,10.69],"formula_id":"formula_0","formula_text":"T = {T 1 , ..., T N },"},{"formula_coordinates":[3,386.27,606.41,138.87,33.18],"formula_id":"formula_1","formula_text":"U = I • W ⊤ 1 (1) V = T • W ⊤ 2 (2)"},{"formula_coordinates":[3,363.19,671.31,89.22,19.18],"formula_id":"formula_2","formula_text":"⟨U i , V j ⟩ = U ⊤ i V j ∥U i ∥∥V j ∥ ."},{"formula_coordinates":[3,341.5,750.25,183.64,27.8],"formula_id":"formula_3","formula_text":"L 1 i = -log exp(⟨U i , V i ⟩) N K=1 exp(⟨U i , V k ⟩)(3)"},{"formula_coordinates":[4,106.22,122.31,183.65,27.8],"formula_id":"formula_4","formula_text":"L 2 i = -log exp(⟨V i , U i ⟩) N K=1 exp(⟨V i , U k ⟩) (4)"},{"formula_coordinates":[4,132.18,191.19,157.69,31.85],"formula_id":"formula_5","formula_text":"L CL = N i=1 (L 1 i + L 2 i ) (5)"},{"formula_coordinates":[4,126.49,629.99,163.38,14.72],"formula_id":"formula_6","formula_text":"θ ′ = θ -α∇ θ L CL (θ) Bs (6)"},{"formula_coordinates":[4,81.9,704.04,207.96,37.56],"formula_id":"formula_7","formula_text":"θ ← θ -β∇ θ L CL (θ ′ ) Bq (7) θ ← θ -β∇ θ L CL (θ -α∇ θ L CL (θ) Bs ) Bq (8)"},{"formula_coordinates":[4,315.35,206.96,209.79,14.72],"formula_id":"formula_8","formula_text":"θ ← θ -β(∇ θ L(θ ′′ ) Bq + λ∇ θ L CL (θ ′ ) Bq ) (9)"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 306.14,\n 500.51,\n 83.33,\n 10.69\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"T = {T 1 , ..., T N },\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 386.27,\n 606.41,\n 138.87,\n 33.18\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"U = I • W ⊤ 1 (1) V = T • W ⊤ 2 (2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 363.19,\n 671.31,\n 89.22,\n 19.18\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"⟨U i , V j ⟩ = U ⊤ i V j ∥U i ∥∥V j ∥ .\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 341.5,\n 750.25,\n 183.64,\n 27.8\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"L 1 i = -log exp(⟨U i , V i ⟩) N K=1 exp(⟨U i , V k ⟩)(3)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 106.22,\n 122.31,\n 183.65,\n 27.8\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"L 2 i = -log exp(⟨V i , U i ⟩) N K=1 exp(⟨V i , U k ⟩) (4)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 132.18,\n 191.19,\n 157.69,\n 31.85\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"L CL = N i=1 (L 1 i + L 2 i ) (5)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 126.49,\n 629.99,\n 163.38,\n 14.72\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"θ ′ = θ -α∇ θ L CL (θ) Bs (6)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 81.9,\n 704.04,\n 207.96,\n 37.56\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"θ ← θ -β∇ θ L CL (θ ′ ) Bq (7) θ ← θ -β∇ θ L CL (θ -α∇ θ L CL (θ) Bs ) Bq (8)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 315.35,\n 206.96,\n 209.79,\n 14.72\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"θ ← θ -β(∇ θ L(θ ′′ ) Bq + λ∇ θ L CL (θ ′ ) Bq ) (9)\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"Meta-learning For Vision-and-language Cross-lingual Transfer"},"abstract":{"kind":"string","value":"Current pre-trained vision-language models (PVLMs) achieve excellent performance on a range of multi-modal datasets. Recent work aims at building multilingual versions of such models, and a range of multilingual multimodal datasets have been introduced for this purpose. However, current PVLMs typically perform poorly on such datasets when used for zero-shot or few-shot cross-lingual transfer, especially for low-resource languages. To alleviate this problem, we propose a novel meta-learning fine-tuning framework. Our framework makes it possible to rapidly adapt PVLMs to new languages by using Modelagnostic Meta-learning (MAML) in a novel cross-lingual multi-modal manner. Experiments show that this new method boosts the performance of current PVLMs in both zero-shot and few-shot settings on four different visionlanguage tasks across 14 languages."},"authors":{"kind":"string","value":"Hanxu Hu; Frank Keller"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: Examples in IGLUE (Bugliarello et al., 2022) benchmark. The left example comes from MaRVL (Liu et al., 2021) dataset, and the right example comes from XVNLI dataset proposed in IGLUE.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :Figure 3 :23Figure 2: Average few-shot performance (accuracy) across all languages of two baseline models on the XVNLI and MaRVL datasets. The horizontal axis represents the number of shots in the training data.","figure_data":"","figure_id":"fig_1","figure_label":"23","figure_type":"figure"},{"figure_caption":"Figure 4: Examples from the Chinese part of the MaRVL dataset and predictions of the baseline and ours method.","figure_data":"","figure_id":"fig_2","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: Improvements of zero-shot performance by fine-tuning xUNITER on different auxiliary languages then evaluating on different target languages using our proposed framework compared with baseline. The left heatmap is for MaRVL, and the right is for XVNLI. Rows correspond to auxiliary and columns correspond to target languages.","figure_data":"","figure_id":"fig_3","figure_label":"5","figure_type":"figure"},{"figure_caption":"as a feature extractor for images.","figure_data":"xFlickr&CoMethodModelXNVLIxGQAMaRVLIRTRmUNITER53.710.053.78.18.9BaselinexUNITER UC259.0 62.520.8 29.056.0 56.413.8 19.712.5 17.0M3P58.228.256.012.911.9OursxUNITER UC263.0 (+4.0) 22.5 (+1.7) 59.4 (+4.4) 16.3 (+2.5) 14.2 (+1.7)The image features are pooled and reshaped asvectors with the same dimensions as text embed-dings. UNITER has four pre-training methods:Masked Language Modelling (MLM), MaskedRegion Modelling (MRM), Image-Text Matching","figure_id":"tab_1","figure_label":"","figure_type":"table"},{"figure_caption":"for","figure_data":"Unsupervised SettingMethod/ModelsUC2xUNITERBaseline62.5±0.159.1±0.1XVL-MAML(w/o down-stream) 63.1±0.160.8±0.1Supervised SettingMethod/ModelsUC2xUNITERXVL-MAML(w/o contrastive)63.8±0.161.6±0.1XVL-MAML64.4±0.162.9±0.1","figure_id":"tab_3","figure_label":"","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: Examples in IGLUE (Bugliarello et al., 2022) benchmark. The left example comes from MaRVL (Liu et al., 2021) dataset, and the right example comes from XVNLI dataset proposed in IGLUE.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :Figure 3 :23Figure 2: Average few-shot performance (accuracy) across all languages of two baseline models on the XVNLI and MaRVL datasets. The horizontal axis represents the number of shots in the training data.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"23\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4: Examples from the Chinese part of the MaRVL dataset and predictions of the baseline and ours method.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: Improvements of zero-shot performance by fine-tuning xUNITER on different auxiliary languages then evaluating on different target languages using our proposed framework compared with baseline. The left heatmap is for MaRVL, and the right is for XVNLI. Rows correspond to auxiliary and columns correspond to target languages.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"as a feature extractor for images.\",\n \"figure_data\": \"xFlickr&CoMethodModelXNVLIxGQAMaRVLIRTRmUNITER53.710.053.78.18.9BaselinexUNITER UC259.0 62.520.8 29.056.0 56.413.8 19.712.5 17.0M3P58.228.256.012.911.9OursxUNITER UC263.0 (+4.0) 22.5 (+1.7) 59.4 (+4.4) 16.3 (+2.5) 14.2 (+1.7)The image features are pooled and reshaped asvectors with the same dimensions as text embed-dings. UNITER has four pre-training methods:Masked Language Modelling (MLM), MaskedRegion Modelling (MRM), Image-Text Matching\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"for\",\n \"figure_data\": \"Unsupervised SettingMethod/ModelsUC2xUNITERBaseline62.5±0.159.1±0.1XVL-MAML(w/o down-stream) 63.1±0.160.8±0.1Supervised SettingMethod/ModelsUC2xUNITERXVL-MAML(w/o contrastive)63.8±0.161.6±0.1XVL-MAML64.4±0.162.9±0.1\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(Chen et al., 2020)\", \"Explanation\": \"The cited work by Chen et al. provides a pre-trained vision-language model that the citing paper adopts in their research to perform multi-modal tasks.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Lu et al., 2019)\", \"Explanation\": \"The cited work by Lu et al. is a pre-trained vision-language model that the citing paper uses as a data source in their research on multi-modal tasks.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Tan and Bansal 2019)\", \"Explanation\": \"The cited work by Tan and Bansal is a pre-trained vision-language model that the citing paper utilizes as a data source in their research on multi-modal tasks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhou et al., 2021)\", \"Explanation\": \"The cited work by Zhou et al. proposes a multilingual pre-trained vision-language model that the citing paper extends by exploring the use of the model in low-resource languages such as Indonesian and Swahili.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ni et al., 2021)\", \"Explanation\": \"The cited work by Ni et al. also contributes to the field of multilingual pre-trained vision-language models, which the citing paper further builds upon in their research on multi-modal tasks in low-resource languages.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Bugliarello et al., 2022)\", \"Explanation\": \"The cited work provides a dataset that the citing paper uses to evaluate the performance of PVLMs in low-resource cross-lingual transfer settings.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2021)\", \"Explanation\": \"The cited work also provides a dataset that the citing paper uses to evaluate the performance of PVLMs in low-resource cross-lingual transfer settings.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Verma et al., 2020)\", \"Explanation\": \"The cited work builds upon the use of MAML-based methods in low-resource and cross-lingual transfer scenarios, including both few-shot and zero-shot cross-lingual tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Finn et al., 2017)\", \"Explanation\": \"The cited work introduces the MAML framework, which the citing paper uses in its research on low-resource and cross-lingual transfer scenarios.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Nooralahzadeh et al., 2020)\", \"Explanation\": \"The cited work further demonstrates the usefulness of MAML-based methods in low-resource and cross-lingual transfer scenarios, including both few-shot and zero-shot cross-lingual tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Nooralahzadeh et al., 2020)\", \"Explanation\": \"The cited work by Nooralahzadeh et al. provides the basis for the use of MAML in cross-lingual transfer tasks in text-only settings, which the citing paper extends to vision-language tasks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"Previous works about using MAML for natural language tasks\", \"Explanation\": \"The citing paper builds upon the work of using MAML in natural language tasks by proposing a novel variant of MAML for vision-language tasks in low-resource cross-lingual transfer.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"The cited work on contrastive loss for alignment between modalities\", \"Explanation\": \"The cited work provides evidence that a contrastive loss can be used to align representations of different modalities, which the citing paper leverages in their cross-lingual, multi-modal MAML optimization procedure.\"}, {\"Category\": \"Data Source\", \"Citation\": \"The cited work on MAML for generalizing quickly to unseen tasks\", \"Explanation\": \"The cited work on MAML for generalizing quickly to unseen tasks provides the data source for the citing paper to demonstrate the improvements in PVLM performance for cross-lingual transfer using XVL-MAML.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Elliott et al., 2016)\", \"Explanation\": \"The cited work by Elliott et al. (2016) provides the Multi30K dataset, which serves as a foundational source of image descriptions in multiple languages for the research conducted in the citing paper on vision-and-language cross-lingual transfer tasks.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Gella et al., 2017;Rotman et al., 2018)\", \"Explanation\": \"The cited works by Gella et al. (2017) and Rotman et al. (2018) offer methods for bridging languages through images, which the citing paper uses as a data source in their research on image-text retrieval in high-resource languages such as English and German.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Pfeiffer et al., 2022)\", \"Explanation\": \"The cited work by Pfeiffer et al. (2022) builds a multilingual visual question answering dataset xGQA, which the citing paper extends by further exploring the cross-lingual transfer and domain shift in the context of visual question answering.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2021)\", \"Explanation\": \"The cited work by Liu et al. (2021) proposes a multilingual version of the MaRVL dataset for grounded visual reasoning, which the citing paper adopts as a methodological basis for their research on cross-lingual transfer and domain shift in the NLVR2 natural language visual reasoning dataset.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ni et al., 2021)\", \"Explanation\": \"The cited work by Ni et al. (2021) presents a transformer-based pre-trained model M3P for vision-and-language cross-lingual transfer, which the citing paper uses as a methodological basis for their research on mapping concepts in different modalities and languages into a common semantic space.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chen et al., 2020)\", \"Explanation\": \"The cited work, UNITER, is the basis for the extension proposed in the citing paper, mUNITER and xUNITER, which are based on the M-BERT and XLM-R models, respectively.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Conneau et al., 2020)\", \"Explanation\": \"The cited work, XLM-R, is the data source for the extension proposed in the citing paper, xUNITER.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhou et al., 2021)\", \"Explanation\": \"The cited work, UC2, proposes a data augmentation method for cross-lingual cross-modal pre-training, which the citing paper builds upon to develop a new model for the same task.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Bugliarello et al., 2022)\", \"Explanation\": \"The cited work provides evidence that pre-training methods require large amounts of data and show a performance gap between English and low-resource languages on the IGLUE benchmark, which the citing paper acknowledges in discussing the challenges of multilingual multimodal tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Pfeiffer et al., 2022)\", \"Explanation\": \"The cited work proposes efficient tuning methods for multilingual multimodal tasks, which the citing paper may have adopted or adapted in developing its own methods for the same task.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang et al., 2023)\", \"Explanation\": \"The cited work proposes translation augmented methods for multilingual multimodal tasks, which the citing paper may have used as a basis for its own methods in the same area.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Qiu et al., 2022)\", \"Explanation\": \"The cited work discusses translation augmented methods for multilingual multimodal tasks, which the citing paper may have used as a data source in developing its own methods in the same area.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Vinyals et al., 2016)\", \"Explanation\": \"The cited work by Vinyals et al. (2016) provides a foundational method for metalearning that the citing paper builds upon in their research on adapting to new tasks quickly.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Finn et al., 2017)\", \"Explanation\": \"The cited work by Finn et al. (2017) introduces the MAML method, which the citing paper uses in their research on metalearning for domain generalization.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Li et al., 2018)\", \"Explanation\": \"The cited work by Li et al. (2018) extends the research on metalearning by exploring its application in multi-task learning and domain generalization, which the citing paper further builds upon in their work.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ren et al., 2018)\", \"Explanation\": \"The cited work by Ren et al. (2018) is a data source for the study on metalearning in semi-supervised learning, which the citing paper utilizes in their research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Yu et al., 2020)\", \"Explanation\": \"The cited work by Yu et al. (2020) extends the research on metalearning by exploring its application in multi-task learning, which the citing paper further builds upon in their work.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Gu et al., 2018)\", \"Explanation\": \"The cited work by Gu et al. (2018) extends the research on metalearning in cross-lingual transfer by regarding languages as tasks, which the citing paper further builds upon in their work.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Nooralahzadeh et al., 2020)\", \"Explanation\": \"The cited work by Nooralahzadeh et al. (2020) extends the research on metalearning in cross-lingual transfer by regarding languages as tasks, which the citing paper further builds upon in their work.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Obamuyide and Vlachos, 2019)\", \"Explanation\": \"The cited work by Obamuyide and Vlachos (2019) is a data source for the study on metalearning under limited supervision in relation classification, which the citing paper utilizes in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wang et al., 2021)\", \"Explanation\": \"The cited work by Wang et al. (2021) is a data source for the study on metalearning in semantic parsing for domain generalization based on MAML, which the citing paper utilizes in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Bugliarello et al., 2022)\", \"Explanation\": \"The cited work provides the framework for the cross-lingual transfer learning problem in vision-and-language scenarios, which the citing paper builds upon to formulate the problem in a specific way.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Bugliarello et al., 2022)\", \"Explanation\": \"The cited work by Bugliarello et al. (2022) provides a traditional cross-lingual transfer learning procedure that the citing paper builds upon in the design of their proposed meta-learning finetuning framework.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2020)\", \"Explanation\": \"The vision-and-language contrastive learning loss proposed by Zhang et al. (2020) serves as the pre-training objective function of CLIP (Radford et al., 2021), providing a methodological basis for the representation learning in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2020)\", \"Explanation\": \"The cited work by Zhang et al. (2020) provides a method for symmetric contrastive loss in image-text pairs, which the citing paper adopts in their research to ensure the symmetry of the loss function.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Nooralahzadeh et al., 2020)\", \"Explanation\": \"The cited work by Nooralahzadeh et al. (2020) provides a method for fine-tuning a model using a contrastive loss as the objective function in the MAML algorithm, which the citing paper adopts in their research on crosslingual transfer learning in vision and language tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2021)\", \"Explanation\": \"The cited work, xUNITER, is used as a base model in the citing paper to explore the performance of the meta-learning method in a specific model architecture.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhou et al., 2021)\", \"Explanation\": \"The cited work, UC2, is also used as a base model in the citing paper to demonstrate the effectiveness of the meta-learning method across different pre-training methods.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ren et al., 2015)\", \"Explanation\": \"The cited work, Faster-RCNN, is used as a pre-training method in the xUNITER model, providing a specific data source for the pre-training process.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Bugliarello et al., 2022)\", \"Explanation\": \"The cited work provides the IGLUE benchmark datasets for the four tasks used in the citing paper, which serve as the basis for the research conducted in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Pfeiffer et al., 2022)\", \"Explanation\": \"The cited work is the source of the xGQA dataset used in the citing paper for the cross-modal retrieval task.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Liu et al., 2021)\", \"Explanation\": \"The cited work is the source of the MaRVL dataset used in the citing paper for the cross-modal reasoning task.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Lin et al., 2014)\", \"Explanation\": \"The cited work is the source of the xFlickr&Co dataset used in the citing paper for the cross-modal retrieval task.\"}, {\"Category\": \"Data Source\", \"Citation\": \"Lin et al., 2015\", \"Explanation\": \"The cited work, COCO, is a data source for the xFlickr&CO dataset used in the citing paper for image-text retrieval research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Bugliarello et al., 2022)\", \"Explanation\": \"The cited work provides the fine-tuning procedure that the citing paper follows in their research on English datasets for xGQA, MaRVL, XVNLI, and xFlickr&Co.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Bugliarello et al., 2022)\", \"Explanation\": \"The cited work is acknowledged for providing the data used in the research on English datasets for xGQA, MaRVL, XVNLI, and xFlickr&Co.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Liu et al., 2021)\", \"Explanation\": \"The cited work is acknowledged for providing the data used in the research on English datasets for xGQA, MaRVL, XVNLI, and xFlickr&Co.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Loshchilov and Hutter, 2018)\", \"Explanation\": \"The cited work provides the AdamW optimizer used in the research to fine-tune all models in PyTorch.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2021)\", \"Explanation\": \"The cited work by Liu et al. provides the definition of consistency used in the citing paper for evaluating the performance of the proposed meta-learning framework.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Bugliarello et al., 2022)\", \"Explanation\": \"The cited work by Bugliarello et al. provides the setting for the few-shot experiments conducted in the citing paper, which includes the IGLUE dataset and the use of xUNITER and UC2 on the XVNLI and MaRVL tasks.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Bugliarello et al., 2022)\", \"Explanation\": \"The cited work provides the detailed information about the IGLUE dataset, which the citing paper uses in its research.\"}]"}}},{"rowIdx":82,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b4","b5","b33","b16","b28","b27","b16","b26"],"table_ref":[],"text":"Predicting and modeling sequences of events has become more sophisticated over the past decade. Early work mined narrative schemas that were limited in representational power: initially sequences of predicate-role pairs (Chambers and Jurafsky, 2008), then generalized to predicate-argument structures (Chambers and Jurafsky, 2009), which continue to be used in neural approaches (Weber et al., 2018;Koupaee et al., 2021). Recently, language modeling provides a very flexible interface for predicting tokens given context and has been applied to event prediction (Rudinger et al., 2015;Pichotta and Mooney, 2016;Koupaee et al., 2021) and cloze tasks (Paperno et al., 2016). Event List 10 things that each happen (1) before; (2) during; and (3) after an international conflict? Before an international conflict, there are several things that can happen: 1."},{"figure_ref":[],"heading":"What happens in international conflicts?","publication_ref":[],"table_ref":[],"text":"Let me draft an event schema…\nDiplomatic efforts may be made in an attempt to resolve the situation peacefully.\n2. The conflict may be discussed in the United Nations Security Council.\n3. Sanctions may be imposed on the country involved in the conflict.\n[…]\nDuring an international conflict, there are several things that can happen: 1. The conflict may be escalated 2. The conflict may be resolved peacefully"},{"figure_ref":[],"heading":"Large language model","publication_ref":["b40","b29","b0","b18","b1","b25","b13","b22","b33","b18","b9","b39","b8","b34","b11","b37","b17"],"table_ref":[],"text":"Figure 1: Overview of our system. A large language model can generate lightly structured lists of events, which themselves may have complex predicateargument structure.\nschemas can only compete with language modeling approaches if they are high-quality and specific enough to provide strong predictions. Ultimately, event schemas might enable explainable forecasting (Zou et al., 2022) grounded in an expert-curated knowledge structure. This paper attempts to bridge this gap by constructing natural language represents of event knowledge from language models. Past efforts like ATOMIC (Sap et al., 2019) and COMET (Bosselut et al., 2019) show that structured repositories of knowledge and the ability to extend them can help enable predictions about the world. We follow in their vein and construct collections of events we call light schemas. These are less structured than graph-based schemas anchored in event ontologies (Li et al., 2021). Our chief aim is to have high recall over a set of events in a domain to serve as a draft for curation of a more structured schema.\nWe generate these schemas using language models like GPT-3.5 (Brown et al., 2020;Ouyang et al., 2022) and Flan-T5 (Chung et al., 2022). As shown in Figure 1, these models have strong abilities to surface events characteristic to a particular domain (e.g., international conflict), including typical arguments for those events. Although our schemas are ontology-free, they implicitly have a certain \"style\" associated with the natural language expressions of their events. We explore both zero-shot and few-shot (specifically one-shot) prediction of schemas.\nUnderstanding the event coverage of our schemas requires comparing them to schemas built by human curators. Evaluation of schematic knowledge and what it predicts have typically been restricted to cloze tasks (Granroth-Wilding and Clark, 2016;Modi et al., 2017;Weber et al., 2018) or events in certain coarse ontologies (Li et al., 2021), but these do not directly evaluate schema representations themselves. Recent past work uses measures very tied to lexical expression of schema predicates (Dror et al., 2022;Zhang et al., 2023), but these are most appropriate for schemas in closed ontologies.\nInstead, we evaluate our schema generation using textual entailment methods (Dagan et al., 2005;Williams et al., 2018), following a similar application of these methods to evaluate groundedness of summaries (Falke et al., 2019;Zhang and Bansal, 2021;Laban et al., 2022). We use entailment to compare our drafted schemas to two sources of ground-truth schemas annotated by human annotators. Specifically, we investigate whether an event we generate entails an event in the ground-truth schema as a measure of recall; we also explore bidirectional entailment (is there mutual entailment between the events?) for a more precise measure.\nThrough human study, we validate that our entailment-based evaluation is reliable. Our results show that large language models can generate schemas that have substantial overlap with groundtruth schemas written by curators. One-shot prediction allows us to emulate the stylistic features of target schemas and attain varying levels of specificity with respect to arguments of predicates. We compare different methods and find that drawing multiple samples from these models can further improve recall.\nOur main contributions are (1) We analyze the performance of current text generation models (GPT-3.5 and Flan-T5) for the task of generating lightly organized event schemas in a completely training-free regime. (2) We show promising results of using textual entailment as a metric to automatically measure event coverage by generated schemas. (3) We show that one-shot prediction can be used to achieve stylistic control over schema generation, suggesting a way to adapt approaches for different desired output formats."},{"figure_ref":[],"heading":"Methods","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Preliminaries","publication_ref":[],"table_ref":[],"text":"Our schemas are anchored to domains d. An example of d shown in Figure 1 is international conflict; these may be broad topics or more specific scenarios like roadside bombing attack. In this work, they are not anchored to specific entity participants. We define a schema S = (s 1 , . . . , s n ) as an ordered1 collection of sentences expressing events. The s i are sentences expressing events at a moderate level of generality; they are typically short descriptions and do not involve specific named entities. However, we do not structurally constrain their form. We refer to the style of the schema as a collection of surface-level factors including the average length in words of the s i and the specificity of the events.\nWe explore two classes of models in this work. First, zero-shot models have the form P (S | v c (d)); they condition on a verbalization v of domain d, parameterized by a strategy c. For example, the prompt in Figure 1 has d = international conflict and the verbalizer List 10 things that each happen (1) before; (2) during; and\n(3) after [d]...Before an [d], there are several things that can happen: 1.. This verbalizer is designed to produce a certain pattern of temporal information; in this case, the answer from the model separates into events occurring before, during, and after the conflict. Other verbalizers we explore look at aspects like cause and effect; a full list of verbalizers is included in the Appendix A.\nThe verbalizers give us control over attributes c; however, they do not necessarily allow us to specify a target style for the schema. We find that each model has certain styles it tends to generate in across a range of verbalizers.\nWe also explore one-shot models P (S | v(d), S demo ) that condition on a schema demonstration as well as a verbalizer of the domain. Note that S demo is a hand-authored schema (or post-edited output of the model) coming from a separate domain d ′ . We give examples of the prompts we use in Appendix A."},{"figure_ref":[],"heading":"Models Considered","publication_ref":["b1","b25","b7","b1","b14"],"table_ref":[],"text":"GPT-3.5 text-davinci-003 We experiment with the most capable of the OpenAI GPT-3.5 models (Brown et al., 2020). According to (Ope-nAI, 2022), text-davinci-003 is an instructiontuned model (Ouyang et al., 2022) using reinforcement learning on models of human preference judgments.\nFlan-T5 We also experiment with (Chung et al., 2022). We use the XXL variant, which is 11B parameters. This allows us to see what is achievable with a smaller instruction-tuned model that can be more easily and cheaply run.\nWe qualitatively observed that flan-t5-xxl does not perform well on temporally-aided complex prompt as described in A. Hence we simplify the prompt into three independent prompts: 1. List events that occur before ... 2. List events that occur during ... 3. List events that occur after ... The outputs generated are minimally postprocessed if necessary to extract the events generated.\nOlder GPT-3 variants We also tried using older variants of GPT-3 model (Brown et al., 2020) such as text-davinci-base, however the generations consisted of a lot of redundancies and required a lot of human curation to extract relevant information from the output, refer appendix D. For this reason, we exclude the base GPT-3 model text-davinci-base from our main results.\nInference hyperparameters For all models, we decode using nucleus sampling (Holtzman et al., 2020) with hyperparameters top-p=1.0 and temperature set to 0.7. We do not perform any model training and use off-the-shelf models for our analysis. The GPT-3.5 variants are accessed through Ope-nAI's API with estimate compute cost amounting to less than $100. To run inference for flan-t5-xxl we host the model on a p3.16xlarge AWS instance."},{"figure_ref":[],"heading":"Evaluation via Textual Entailment","publication_ref":["b19","b20"],"table_ref":["tab_6"],"text":"Inspection of our schemas (see Figure 1, Table 6) shows that they are very high quality. As we are using state-of-the-art autoregressive Transformer models, the fluency, coherence of the sequence of events, and linguistic quality of each individual event are very high and do not need to be the focus of our evaluation. Instead, the main question is to what extent the events we have cover the important events in the target domain; they may fail to do so as a result of reporting biases in text. We can compare these to human-written schemas; however, because our schemas are in natural language, we will need a sophisticated comparison function in order to do so. Here, we turn to textual entailment.\nOur evaluation focuses on comparing a predicted schema Ŝ with a ground-truth, human-annotated schema S * . S * is considered to contain events that we want to see represented in predicted schemas. Note that S * may not be exhaustive; that is, an event not in S * may still be considered of high quality and relevant. Therefore, our evaluation will focus on recall.\nWe use textual entailment models of the form E : (s 1 , s 2 ) → R to judge whether two sentences are matching events. An entailment model computes a distribution over three classes {entailment, neutral, contradiction}. We set E to return the probability of s 1 entailing s 2 , ignoring the distinction between neutral and contradiction. Intuitively, a sentence like protests break out in the city should entail (and be entailed by) a sufficiently similar event like civil unrest in the capital. While there may be minor variation in the argument (e.g., city vs. capital), the notion of entailment still approximately captures the appropriate similarity for this task.\nOur goal is to compare an event s to an entire schema S. The recall score r for an event s ∈ S * is then given by\nr(s, Ŝ) = max ŝ∈ Ŝ (max(E(s, ŝ), E(ŝ, s)))(1)\nmaxing over the events in the predicted schema.\nAs the level of specificity between the gold events and predicted events can sway in either direction, we run the entailment model in both directions (e.g. gold event entails predicted event or vice versa).\nWe consider two variants of this procedure: any-directional entailment where we use the entailment model as described above, and bidirectional entailment where we modify the score to be min{E(s, ŝ), E(ŝ, s)}. This places a stronger requirement that the two statements be equivalent.\nEntailment Models Used We test our generated schemas using the textual entailment model roberta-large-wanli by (Liu et al., 2022) trained on the WANLI dataset. This model uses RoBERTa-large (Liu et al., 2019) architecture and has 345M parameters."},{"figure_ref":[],"heading":"Experimental Setup","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Gold Schema","publication_ref":["b10"],"table_ref":[],"text":"We conduct experiments on the gold schemas from two datasets: RESIN-11 (Du et al., 2022) and Cu-ratedSchemas, described below. The domains included in our dataset are international conflict, natural disaster, IED attacks, disease outbreak, mass shooting, and kidnapping. We sample these domains as they are available in both datasets and gives us the opportunity to test various interesting aspects of schema datasets such as varying coverage, and style of event descriptions. More details on both dataset can be found in Appendix E. For the published RESIN-11 Schema, we modify the event structure into a natural language sentence as described in Appendix E.2.\nWe also use a separate set of schemas we call the CuratedSchemas set. These schemas were annotated by ourselves and our collaborators independently of the RESIN schemas. Appendix E.1 describes these."},{"figure_ref":[],"heading":"Language Model Schema Generation","publication_ref":[],"table_ref":[],"text":"We predominatly test the schema drafting performance of GPT-3 variant text-davinci-003 and Flan-T5 variant flan-t5-xxl. Each prompt is used to generate 3 generations. We report statistics and event recall on average of 3 generations in the 5 section.\nWe can also over-generate predictions using diverse prompts and achieve a higher event recall with the possibility of generating incorrect or redundant events. To experiment this, we craft 3 different prompts and sample 3 generations from each using the text-davinci-003. We call this approach prompt union. More details on prompt union can be found in the appendix A.\nA key aspect of natural language event schemas is stylistic variations across datasets and language models that generate it. As discussed in 2, we use one-shot prompts to guide the models to generate outputs similar to the target dataset style."},{"figure_ref":[],"heading":"Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Schema Generation Performance","publication_ref":["b9","b9"],"table_ref":[],"text":"Table 1 shows the results of several schema generation approaches measured against the gold schemas from RESIN-11 and CuratedSchemas. The metric used to measure the recall is the any-directional entailment as described in Section 3. We report the mean and standard deviation of event recall across three sampled generations for each prompt. Along with the event recall, we also report the number of events predicted by each model, which gives signal about the precision of their event generation.\nZero-shot generation performance is high. Table 1 highlights that both text-davinci-003 and flan-t5-xxl show an average of 0.39 and 0.3958 event coverage with respect to the gold schemas. Discussions on human agreement of the entailment model judgments and influence of generation style are deferred to Sections 6 and 5.2. We also report the average number of event generated by both models for each prompt which are in the range of 15-30, indicating that we do not over-generate events for each domain to increase the recall performance. Finally, the overlap of generated events with both the human curated gold schemas (RESIN-11 and CuratedSchemas) is substantial and reflects on the potential of language models in drafting complex schemas with sufficient coverage.\nDrawing more samples from this model can increase recall further. Complex event schemas of domains like disease outbreak or natural disasters can have varying actors and topics that cannot be exhaustively sampled from a single prompt. For instance, What happens after a disease outbreak? can have various responses talking about either legal proceedings against organisations who are held accountable for a disease outbreak OR research on preventing future outbreaks of the disease -both responses are valid but cover various aspects of the complex event. This result can be emulated by using diversity in prompts generation to generate events affecting different participants from the event. In Table 2 we compare event coverage results from a single prompt versus taking a union across a larger number of prompts and samples. We see that taking a union of generations from various prompts leads to a substantial boost to the event recall with the caveat that we generate larger number of events. Table 1: Event recall of zero-shot and one-shot performance performance of different language models measured against human curated gold schemas from two datasets. We use any-directional entailment. One-shot results are substantially better for certain domains and lead to generation of more events. However, all systems are able to generate a substantial number of matching events across the domains of interest.\nPrompting models with complex prompts leads to significantly higher performance than past work Dror et al. (2022) explore using language models to generate documents which can be used to construct complex schemas for events. Their work studies generation of direct step-by-step schemas using prompts such as What are the steps involved in topic? 1.. We generate responses from this prompt template to extract natural language events. Our work is not a direct comparison to (Dror et al., 2022) as their focus is predominantly using language models to generate documents for downstream event schema induction pipelines. However, we only adopt their direct step-by-step schema generation prompt and argue that using complex prompts can lead to better event coverage in comparison to simpler prompts such as listing steps in an event. This result is highlighted by comparing the results of text-davinci-003 and Dror et al.\nin Table 1.\nOverall, we see promising results on event schema drafting performance with language models with minimal human intervention and ability to automatically evaluate against gold schemas."},{"figure_ref":[],"heading":"Stylistic and Coverage Differences","publication_ref":[],"table_ref":["tab_2"],"text":"In this section, we investigate the various differences that can occur in natural language schemas derived from different sources.\nThere are stylistic differences between event schema datasets and generations. Our gold datasets are derived from two independent sources and have stylistic differences in the method of representing natural language events. In Table 3 we show the average length of these prompts. We see that the mean length of sentences measured by word count varies between 3.57 to 6.29 among the datasets and LM generated schemas. Some stylistic influence can be achieved by one-shot prompting as noted in the case of word count difference between zero-shot and one-shot outputs of text-davinci-003. We also show this qualitatively in Section 7.\nOne-shot prompts for style-matching with gold schemas We can much better match the style of schemas by providing them as demonstrations in one-shot prompts. Specifically, for generating a schema for domain d, we formulate one-shot prompts as shown in appendix A from three domains\nx, ∀x / ∈ [d].\nInter-dataset Agreement To further confirm that the schemas we have differ, Table 2: Diverse and instructive prompts improve the coverage of schema generation. We compare the single prompt version used in 1 against using a prompt union method which uses three prompts to over-generate events and improve recall. This result shows that the method can be potentially used to increase event coverage with gold schemas while compromising precision. result conflates two things: the performance of the entailment model (discussed more in Section 6) and the meaningful differences in events between the two schemas. However, on inspection, stylistic attributes are responsible for both, as certain more specific events in RESIN have no analogue in CuratedSchemas due to the different styles. Entailment reflects this even though it is not reliable on every case. "},{"figure_ref":[],"heading":"Human Evaluation of Entailment","publication_ref":[],"table_ref":["tab_7","tab_8"],"text":"Our recall values in Table 2 are high enough to establish the utility of our approach. Most events can theoretically be matched to some other event in our generated dataset.\nTo confirm whether the entailment systems are making correct decisions, we conduct a precisionfocused human evaluation of the automatic entail-ment decisions. The objective was to assess how reliably the entailment models aligned with our actual judgments regarding the equivalence between events. To gather annotations for this evaluation, we used Amazon Mechanical Turk (AMT) and enlisted the participation of randomly selected human annotators. We presented them with 216 sampled event pairs from all domains consisting of gold and predicted events that are matched by the anydirectional entailment model as described in 3. The annotators were then asked to indicate their agreement with each match. Each task is annotated by three unique annotators to measure overall consensus. Further details on the task setup can be found in Appendix B.\nThe results of our human agreement study are shown in Table 5. The event match performance of entailment models across two datasets and all event domains achieves a majority vote agreement of 0.55 with the entailment judgments. However, at least one annotator agrees with the event match 75% of the time. We also measure the Krippendorff's Alpha to measure the inter-annotator agreement. The alpha score for our task is 0.43, which is considered moderate agreement, but does reflect the subjectivity of the task.\nWe argue that not all of the entailment mistakes labeled as such truly represent errors. For instance, the any-directional entailment model matches the prediction \"Implementation of preventative measures\" to two gold events: \"people maintain physical distancing to prevent disease spread\" and \"people are vaccinated against the disease.\" Although the level of specificity differs between the two, we argue that any-directional entailment can be a reasonable candidate for automatic metric while serving the purpose of assigning soft matches between gold and predicted events. Cases like this are often marked as not equivalent by Turkers, but we argue that the entailment judgment is still a reliable method for assessing recall.\nFor a highly precise evaluation protocol, bidirectional entailment can be a suitable candidate, however, as this is a very strict metric, the recall achieved by this evaluation protocol is significantly lower (see Table 7 in the Appendix).\nWe also conduct a internal human evaluation of the entailment metric at a granular level in Appendix F.\nThe performance of entailment depends on stylistic matching Table 8 highlights that hu-man agreement of anydirectional entailment improves across all domains for davinci-003 when the schemas are generated with one-shot prompts compared to zero-shot. This signifies that one-shot prompts are beneficial in guiding the language models to generate schemas of a specific style."},{"figure_ref":[],"heading":"Qualitative Analysis","publication_ref":[],"table_ref":["tab_2","tab_6"],"text":"While the length analysis in Table 3 shows differences between various domains and schema sources, the stylistic differences go beyond length in ways that are hard to precisely quantify. We show examples from the disease-outbreak domain in Table 6 to highlight these differences and qualitatively depict the variation in the writing style of events across human-curated datasets (RESIN-11 and CuratedSchemas) and generations from language models (text-davinci-001 and flan-t5-xxl) in zero-shot and one-shot settings.\nWe see that event samples from AltShemas are more formal and shorter in size as compared to RESIN-11 which have a high length variance and are more natural language like. This style also differs from the zero-shot generations from davinci-003 and flan-t5-xxl. A controlled generation using one-shot prompts derived from RESIN-11 schema can be used to attempt to match the event description style of the gold schemas."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b3","b6","b23","b31","b36","b18","b10","b32","b12","b39","b9","b35","b30","b38","b21","b37"],"table_ref":[],"text":"Event-centric modeling and schema induction Methods performing schema induction can be categorized into simple and complex schema induction. Simple schema induction methods rely on identifying event triggers and participants and do no incorporate the relationships between events (Chambers, 2013;Cheung et al., 2013;Nguyen et al., 2015;Sha et al., 2016;Yuan et al., 2018). Recent work (Li et al., 2021;Du et al., 2022) focuses on generating complex schemas that incorporate temporal as well as event argument relationships but assume availability of large amount of event relevant corpora. Existing event datasets such as MAVEN (Wang et al., 2020) and event-centric knowledge bases such as Event-Wiki (Ge et al., 2018), but working with these datasets naturally restricts a system designer to a fixed ontology.\nClosest to our work, Zhang et al. (2023) also generate schemas including a GPT prompting stage. However, they follow this with a stage of grounding to an ontology, sidestepping the challenges with evaluation we tackle in this work and losing the ability to homogenize between two different sources of schemas. Dror et al. (2022) use language models to generate large number of source documents about a topic that can be used to extract events and relations to build schemas in a zeroshot manner. However, their method uses language models to generate documents containing relevant information which is further used to extract events using event extraction methods. In this work, we provide a way to both generate and automatically evaluate light event schemas in natural language making the process less dependent on traditional event extraction and schema matching pipleines.\nTextual Entailment Natural Language Inference research focuses on establishing entailment between a premise and a hypothesis pair. Although most of the previous work focuses on sentence level hypothesis and premise pair, recent datasets such as DocNLI (Yin et al., 2021) and ContractNLI (Koreeda and Manning, 2021) push the boundaries to extend NLI models to longer multi-sentence inputs and real-work datasets. Schuster et al. (2022) explore the utility of NLI models on longer inputs using a \"stretching\" form of aggregation, namely maxing over possible alignments to a document. It is common to see similarity models being used to judge similarity between two sentences in ROUGE, BLEU and BERTScore (Zhang et al., 2020). However, recent works recommend the usage of NLI models as evaluation metrics for Ab-stract Summarization (Maynez et al., 2020) as they capture the faithfulness and factuality of summaries better than standard metrics. (Zhang and Bansal, 2021) explore the usage of NLI models to automate evaluation of summarization tasks which can also benefit automated best model checkpointing. In this work, we explore using NLI as a metric for schema coverage matching directly in natural language."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"In this paper, we explored the ability of language models to draft light event schemas. In both zeroand one-shot settings, we showed that large language models can generate coherent, varied sequences of events in natural language that overlap substantially with human-curated events across several domains of interest. We show that textual entailment can be used to evaluate these matches. We believe our work can pave the way for future efforts looking at how explicit knowledge like schemas can be used in tandem with large language models to make predictions. Streamlining the ability to generate schemas and then curate them with human intervention will be an important step to scaling this method to work across many domains."},{"figure_ref":[],"heading":"Limitations","publication_ref":["b18"],"table_ref":[],"text":"The schemas we produce in this work are, by choice, lighter weight than representations used in some prior work. Past work (Li et al., 2021) has explored schemas with graph-structured ordering. While these schemas can express a richer set of partial ordering and mutual exclusion relationships between events, they are both cumbersome to produce and relatively little work has shown the ability to use them to perform complex inferences. Our view is that more complex structural relationships should also be specified in natural language for maximal compatibility with prediction based on large language models; we use this for future work. Human curation can also be used to impart these features for use in downstream applications.\nA second limitation is that the robustness of event recall evaluation using textual entailment is dependent on the stylistic similarities between generated and gold schemas. While we analyze this in the paper, stronger textual entailment systems down the road can potentially be useful to improve the precision of our performance estimates further.\nFinally, we note that schema-mediated prediction with neural models is an emerging and ongoing area of research. Therefore, there are not standard systems we can plug our schemas into for downstream evaluation. Nevertheless, we believe that these knowledge structures can be intrinsically evaluated, and high quality representations will pave the way for future work in this area."},{"figure_ref":[],"heading":"A Prompts","publication_ref":[],"table_ref":[],"text":"Zero-shot Prompts The text-davinci-003 zero-shot experiments predominantly use the prompt below. The domains d that we consider are natural disaster, disease outbreak, international conflict, mass shooting, IED attack, and kidnapping Temporally-aided prompt \"List 10 things that each happen (1) before; (2) during; and (3) after a [d]?\nBefore a [d], there are several things that can happen:"},{"figure_ref":[],"heading":"1.\"","publication_ref":[],"table_ref":[],"text":"Prompt Union This approach uses three different prompt templates as shown below. In addition to the previously described prompt above, the two additional prompts that we sample for prompt union experiments are: For generating events for one domain we sample from such prompts from three other domains and generate the output."},{"figure_ref":["fig_2"],"heading":"B Annotation Task Description","publication_ref":[],"table_ref":[],"text":"The task description provided to Turk workers on Amazon Mechanical Turk is shown in Figure 3."},{"figure_ref":[],"heading":"C Comparison of any-directional and bi-directional entailment over event recall","publication_ref":[],"table_ref":[],"text":"We compare the event recall values predicted by anydirectional and bidirectional entailment as a metric in 7."},{"figure_ref":["fig_3"],"heading":"D Example of davinci-base output","publication_ref":[],"table_ref":[],"text":"Figure 4 shows a representative example for output from davinic-base using zero-shot prompt. "},{"figure_ref":[],"heading":"E Dataset Description","publication_ref":[],"table_ref":["tab_9"],"text":"In this work we experiment with a subset of two datasets: 1) RESIN-11 and 2) CuratedSchemas. dataset statistics of the subsets are described in Table 9."},{"figure_ref":["fig_5"],"heading":"E.1 CuratedSchemas","publication_ref":[],"table_ref":[],"text":"CuratedSchemas was produced over a period of roughly 1 year. Around 10 distinct annotators, all experts in natural language processing, contributed events to build up these schemas over several stages. Multiple rounds of feedback were collected based on a downstream prediction task and human assessment of outputs from that task, not conducted by the authors of this paper. The schemas balance precision and recall, aiming to cover each domain thoroughly but not include events that are too lowlevel or would never conceivably be reported in text. Additional layers of structure are presented in the CuratedSchemas dataset, but here we only evaluate against the collection of abstract events. An example of events from the Disease Outbreak domain of CuratedSchemas is shown in Figure 6."},{"figure_ref":[],"heading":"E.2 RESIN Schemas","publication_ref":[],"table_ref":[],"text":"The natural language description of the event is derived from the published RESIN schemas using the fields of qnode, qlabel, description, and/or ta1explanation as applicable. An example of the events from the Disease Outbreak domain derived from schemas published in RESIN-11 is shown in Figure 7."},{"figure_ref":["fig_4"],"heading":"F Granular Human Agreement Study","publication_ref":[],"table_ref":["tab_8"],"text":"We perform a granular human agreement study to investigate the trends and impacts that various domains, prompting methods, and language models have on the performance of entailment models for automatic evaluation. This study is performed by the authors of the paper and is illustrated in Table 8.\nFigure 5 shows examples of our annotation across different domains for human agreement with the pairings predicted by the any-directional entailment model between gold and generated schema events. "},{"figure_ref":[],"heading":"Dataset","publication_ref":[],"table_ref":[],"text":""}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b4\",\n \"b5\",\n \"b33\",\n \"b16\",\n \"b28\",\n \"b27\",\n \"b16\",\n \"b26\"\n ],\n \"table_ref\": [],\n \"text\": \"Predicting and modeling sequences of events has become more sophisticated over the past decade. Early work mined narrative schemas that were limited in representational power: initially sequences of predicate-role pairs (Chambers and Jurafsky, 2008), then generalized to predicate-argument structures (Chambers and Jurafsky, 2009), which continue to be used in neural approaches (Weber et al., 2018;Koupaee et al., 2021). Recently, language modeling provides a very flexible interface for predicting tokens given context and has been applied to event prediction (Rudinger et al., 2015;Pichotta and Mooney, 2016;Koupaee et al., 2021) and cloze tasks (Paperno et al., 2016). Event List 10 things that each happen (1) before; (2) during; and (3) after an international conflict? Before an international conflict, there are several things that can happen: 1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"What happens in international conflicts?\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Let me draft an event schema…\\nDiplomatic efforts may be made in an attempt to resolve the situation peacefully.\\n2. The conflict may be discussed in the United Nations Security Council.\\n3. Sanctions may be imposed on the country involved in the conflict.\\n[…]\\nDuring an international conflict, there are several things that can happen: 1. The conflict may be escalated 2. The conflict may be resolved peacefully\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Large language model\",\n \"publication_ref\": [\n \"b40\",\n \"b29\",\n \"b0\",\n \"b18\",\n \"b1\",\n \"b25\",\n \"b13\",\n \"b22\",\n \"b33\",\n \"b18\",\n \"b9\",\n \"b39\",\n \"b8\",\n \"b34\",\n \"b11\",\n \"b37\",\n \"b17\"\n ],\n \"table_ref\": [],\n \"text\": \"Figure 1: Overview of our system. A large language model can generate lightly structured lists of events, which themselves may have complex predicateargument structure.\\nschemas can only compete with language modeling approaches if they are high-quality and specific enough to provide strong predictions. Ultimately, event schemas might enable explainable forecasting (Zou et al., 2022) grounded in an expert-curated knowledge structure. This paper attempts to bridge this gap by constructing natural language represents of event knowledge from language models. Past efforts like ATOMIC (Sap et al., 2019) and COMET (Bosselut et al., 2019) show that structured repositories of knowledge and the ability to extend them can help enable predictions about the world. We follow in their vein and construct collections of events we call light schemas. These are less structured than graph-based schemas anchored in event ontologies (Li et al., 2021). Our chief aim is to have high recall over a set of events in a domain to serve as a draft for curation of a more structured schema.\\nWe generate these schemas using language models like GPT-3.5 (Brown et al., 2020;Ouyang et al., 2022) and Flan-T5 (Chung et al., 2022). As shown in Figure 1, these models have strong abilities to surface events characteristic to a particular domain (e.g., international conflict), including typical arguments for those events. Although our schemas are ontology-free, they implicitly have a certain \\\"style\\\" associated with the natural language expressions of their events. We explore both zero-shot and few-shot (specifically one-shot) prediction of schemas.\\nUnderstanding the event coverage of our schemas requires comparing them to schemas built by human curators. Evaluation of schematic knowledge and what it predicts have typically been restricted to cloze tasks (Granroth-Wilding and Clark, 2016;Modi et al., 2017;Weber et al., 2018) or events in certain coarse ontologies (Li et al., 2021), but these do not directly evaluate schema representations themselves. Recent past work uses measures very tied to lexical expression of schema predicates (Dror et al., 2022;Zhang et al., 2023), but these are most appropriate for schemas in closed ontologies.\\nInstead, we evaluate our schema generation using textual entailment methods (Dagan et al., 2005;Williams et al., 2018), following a similar application of these methods to evaluate groundedness of summaries (Falke et al., 2019;Zhang and Bansal, 2021;Laban et al., 2022). We use entailment to compare our drafted schemas to two sources of ground-truth schemas annotated by human annotators. Specifically, we investigate whether an event we generate entails an event in the ground-truth schema as a measure of recall; we also explore bidirectional entailment (is there mutual entailment between the events?) for a more precise measure.\\nThrough human study, we validate that our entailment-based evaluation is reliable. Our results show that large language models can generate schemas that have substantial overlap with groundtruth schemas written by curators. One-shot prediction allows us to emulate the stylistic features of target schemas and attain varying levels of specificity with respect to arguments of predicates. We compare different methods and find that drawing multiple samples from these models can further improve recall.\\nOur main contributions are (1) We analyze the performance of current text generation models (GPT-3.5 and Flan-T5) for the task of generating lightly organized event schemas in a completely training-free regime. (2) We show promising results of using textual entailment as a metric to automatically measure event coverage by generated schemas. (3) We show that one-shot prediction can be used to achieve stylistic control over schema generation, suggesting a way to adapt approaches for different desired output formats.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Methods\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Preliminaries\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Our schemas are anchored to domains d. An example of d shown in Figure 1 is international conflict; these may be broad topics or more specific scenarios like roadside bombing attack. In this work, they are not anchored to specific entity participants. We define a schema S = (s 1 , . . . , s n ) as an ordered1 collection of sentences expressing events. The s i are sentences expressing events at a moderate level of generality; they are typically short descriptions and do not involve specific named entities. However, we do not structurally constrain their form. We refer to the style of the schema as a collection of surface-level factors including the average length in words of the s i and the specificity of the events.\\nWe explore two classes of models in this work. First, zero-shot models have the form P (S | v c (d)); they condition on a verbalization v of domain d, parameterized by a strategy c. For example, the prompt in Figure 1 has d = international conflict and the verbalizer List 10 things that each happen (1) before; (2) during; and\\n(3) after [d]...Before an [d], there are several things that can happen: 1.. This verbalizer is designed to produce a certain pattern of temporal information; in this case, the answer from the model separates into events occurring before, during, and after the conflict. Other verbalizers we explore look at aspects like cause and effect; a full list of verbalizers is included in the Appendix A.\\nThe verbalizers give us control over attributes c; however, they do not necessarily allow us to specify a target style for the schema. We find that each model has certain styles it tends to generate in across a range of verbalizers.\\nWe also explore one-shot models P (S | v(d), S demo ) that condition on a schema demonstration as well as a verbalizer of the domain. Note that S demo is a hand-authored schema (or post-edited output of the model) coming from a separate domain d ′ . We give examples of the prompts we use in Appendix A.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Models Considered\",\n \"publication_ref\": [\n \"b1\",\n \"b25\",\n \"b7\",\n \"b1\",\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"GPT-3.5 text-davinci-003 We experiment with the most capable of the OpenAI GPT-3.5 models (Brown et al., 2020). According to (Ope-nAI, 2022), text-davinci-003 is an instructiontuned model (Ouyang et al., 2022) using reinforcement learning on models of human preference judgments.\\nFlan-T5 We also experiment with (Chung et al., 2022). We use the XXL variant, which is 11B parameters. This allows us to see what is achievable with a smaller instruction-tuned model that can be more easily and cheaply run.\\nWe qualitatively observed that flan-t5-xxl does not perform well on temporally-aided complex prompt as described in A. Hence we simplify the prompt into three independent prompts: 1. List events that occur before ... 2. List events that occur during ... 3. List events that occur after ... The outputs generated are minimally postprocessed if necessary to extract the events generated.\\nOlder GPT-3 variants We also tried using older variants of GPT-3 model (Brown et al., 2020) such as text-davinci-base, however the generations consisted of a lot of redundancies and required a lot of human curation to extract relevant information from the output, refer appendix D. For this reason, we exclude the base GPT-3 model text-davinci-base from our main results.\\nInference hyperparameters For all models, we decode using nucleus sampling (Holtzman et al., 2020) with hyperparameters top-p=1.0 and temperature set to 0.7. We do not perform any model training and use off-the-shelf models for our analysis. The GPT-3.5 variants are accessed through Ope-nAI's API with estimate compute cost amounting to less than $100. To run inference for flan-t5-xxl we host the model on a p3.16xlarge AWS instance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Evaluation via Textual Entailment\",\n \"publication_ref\": [\n \"b19\",\n \"b20\"\n ],\n \"table_ref\": [\n \"tab_6\"\n ],\n \"text\": \"Inspection of our schemas (see Figure 1, Table 6) shows that they are very high quality. As we are using state-of-the-art autoregressive Transformer models, the fluency, coherence of the sequence of events, and linguistic quality of each individual event are very high and do not need to be the focus of our evaluation. Instead, the main question is to what extent the events we have cover the important events in the target domain; they may fail to do so as a result of reporting biases in text. We can compare these to human-written schemas; however, because our schemas are in natural language, we will need a sophisticated comparison function in order to do so. Here, we turn to textual entailment.\\nOur evaluation focuses on comparing a predicted schema Ŝ with a ground-truth, human-annotated schema S * . S * is considered to contain events that we want to see represented in predicted schemas. Note that S * may not be exhaustive; that is, an event not in S * may still be considered of high quality and relevant. Therefore, our evaluation will focus on recall.\\nWe use textual entailment models of the form E : (s 1 , s 2 ) → R to judge whether two sentences are matching events. An entailment model computes a distribution over three classes {entailment, neutral, contradiction}. We set E to return the probability of s 1 entailing s 2 , ignoring the distinction between neutral and contradiction. Intuitively, a sentence like protests break out in the city should entail (and be entailed by) a sufficiently similar event like civil unrest in the capital. While there may be minor variation in the argument (e.g., city vs. capital), the notion of entailment still approximately captures the appropriate similarity for this task.\\nOur goal is to compare an event s to an entire schema S. The recall score r for an event s ∈ S * is then given by\\nr(s, Ŝ) = max ŝ∈ Ŝ (max(E(s, ŝ), E(ŝ, s)))(1)\\nmaxing over the events in the predicted schema.\\nAs the level of specificity between the gold events and predicted events can sway in either direction, we run the entailment model in both directions (e.g. gold event entails predicted event or vice versa).\\nWe consider two variants of this procedure: any-directional entailment where we use the entailment model as described above, and bidirectional entailment where we modify the score to be min{E(s, ŝ), E(ŝ, s)}. This places a stronger requirement that the two statements be equivalent.\\nEntailment Models Used We test our generated schemas using the textual entailment model roberta-large-wanli by (Liu et al., 2022) trained on the WANLI dataset. This model uses RoBERTa-large (Liu et al., 2019) architecture and has 345M parameters.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experimental Setup\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Gold Schema\",\n \"publication_ref\": [\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"We conduct experiments on the gold schemas from two datasets: RESIN-11 (Du et al., 2022) and Cu-ratedSchemas, described below. The domains included in our dataset are international conflict, natural disaster, IED attacks, disease outbreak, mass shooting, and kidnapping. We sample these domains as they are available in both datasets and gives us the opportunity to test various interesting aspects of schema datasets such as varying coverage, and style of event descriptions. More details on both dataset can be found in Appendix E. For the published RESIN-11 Schema, we modify the event structure into a natural language sentence as described in Appendix E.2.\\nWe also use a separate set of schemas we call the CuratedSchemas set. These schemas were annotated by ourselves and our collaborators independently of the RESIN schemas. Appendix E.1 describes these.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Language Model Schema Generation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We predominatly test the schema drafting performance of GPT-3 variant text-davinci-003 and Flan-T5 variant flan-t5-xxl. Each prompt is used to generate 3 generations. We report statistics and event recall on average of 3 generations in the 5 section.\\nWe can also over-generate predictions using diverse prompts and achieve a higher event recall with the possibility of generating incorrect or redundant events. To experiment this, we craft 3 different prompts and sample 3 generations from each using the text-davinci-003. We call this approach prompt union. More details on prompt union can be found in the appendix A.\\nA key aspect of natural language event schemas is stylistic variations across datasets and language models that generate it. As discussed in 2, we use one-shot prompts to guide the models to generate outputs similar to the target dataset style.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Schema Generation Performance\",\n \"publication_ref\": [\n \"b9\",\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"Table 1 shows the results of several schema generation approaches measured against the gold schemas from RESIN-11 and CuratedSchemas. The metric used to measure the recall is the any-directional entailment as described in Section 3. We report the mean and standard deviation of event recall across three sampled generations for each prompt. Along with the event recall, we also report the number of events predicted by each model, which gives signal about the precision of their event generation.\\nZero-shot generation performance is high. Table 1 highlights that both text-davinci-003 and flan-t5-xxl show an average of 0.39 and 0.3958 event coverage with respect to the gold schemas. Discussions on human agreement of the entailment model judgments and influence of generation style are deferred to Sections 6 and 5.2. We also report the average number of event generated by both models for each prompt which are in the range of 15-30, indicating that we do not over-generate events for each domain to increase the recall performance. Finally, the overlap of generated events with both the human curated gold schemas (RESIN-11 and CuratedSchemas) is substantial and reflects on the potential of language models in drafting complex schemas with sufficient coverage.\\nDrawing more samples from this model can increase recall further. Complex event schemas of domains like disease outbreak or natural disasters can have varying actors and topics that cannot be exhaustively sampled from a single prompt. For instance, What happens after a disease outbreak? can have various responses talking about either legal proceedings against organisations who are held accountable for a disease outbreak OR research on preventing future outbreaks of the disease -both responses are valid but cover various aspects of the complex event. This result can be emulated by using diversity in prompts generation to generate events affecting different participants from the event. In Table 2 we compare event coverage results from a single prompt versus taking a union across a larger number of prompts and samples. We see that taking a union of generations from various prompts leads to a substantial boost to the event recall with the caveat that we generate larger number of events. Table 1: Event recall of zero-shot and one-shot performance performance of different language models measured against human curated gold schemas from two datasets. We use any-directional entailment. One-shot results are substantially better for certain domains and lead to generation of more events. However, all systems are able to generate a substantial number of matching events across the domains of interest.\\nPrompting models with complex prompts leads to significantly higher performance than past work Dror et al. (2022) explore using language models to generate documents which can be used to construct complex schemas for events. Their work studies generation of direct step-by-step schemas using prompts such as What are the steps involved in topic? 1.. We generate responses from this prompt template to extract natural language events. Our work is not a direct comparison to (Dror et al., 2022) as their focus is predominantly using language models to generate documents for downstream event schema induction pipelines. However, we only adopt their direct step-by-step schema generation prompt and argue that using complex prompts can lead to better event coverage in comparison to simpler prompts such as listing steps in an event. This result is highlighted by comparing the results of text-davinci-003 and Dror et al.\\nin Table 1.\\nOverall, we see promising results on event schema drafting performance with language models with minimal human intervention and ability to automatically evaluate against gold schemas.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Stylistic and Coverage Differences\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"In this section, we investigate the various differences that can occur in natural language schemas derived from different sources.\\nThere are stylistic differences between event schema datasets and generations. Our gold datasets are derived from two independent sources and have stylistic differences in the method of representing natural language events. In Table 3 we show the average length of these prompts. We see that the mean length of sentences measured by word count varies between 3.57 to 6.29 among the datasets and LM generated schemas. Some stylistic influence can be achieved by one-shot prompting as noted in the case of word count difference between zero-shot and one-shot outputs of text-davinci-003. We also show this qualitatively in Section 7.\\nOne-shot prompts for style-matching with gold schemas We can much better match the style of schemas by providing them as demonstrations in one-shot prompts. Specifically, for generating a schema for domain d, we formulate one-shot prompts as shown in appendix A from three domains\\nx, ∀x / ∈ [d].\\nInter-dataset Agreement To further confirm that the schemas we have differ, Table 2: Diverse and instructive prompts improve the coverage of schema generation. We compare the single prompt version used in 1 against using a prompt union method which uses three prompts to over-generate events and improve recall. This result shows that the method can be potentially used to increase event coverage with gold schemas while compromising precision. result conflates two things: the performance of the entailment model (discussed more in Section 6) and the meaningful differences in events between the two schemas. However, on inspection, stylistic attributes are responsible for both, as certain more specific events in RESIN have no analogue in CuratedSchemas due to the different styles. Entailment reflects this even though it is not reliable on every case. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Human Evaluation of Entailment\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_7\",\n \"tab_8\"\n ],\n \"text\": \"Our recall values in Table 2 are high enough to establish the utility of our approach. Most events can theoretically be matched to some other event in our generated dataset.\\nTo confirm whether the entailment systems are making correct decisions, we conduct a precisionfocused human evaluation of the automatic entail-ment decisions. The objective was to assess how reliably the entailment models aligned with our actual judgments regarding the equivalence between events. To gather annotations for this evaluation, we used Amazon Mechanical Turk (AMT) and enlisted the participation of randomly selected human annotators. We presented them with 216 sampled event pairs from all domains consisting of gold and predicted events that are matched by the anydirectional entailment model as described in 3. The annotators were then asked to indicate their agreement with each match. Each task is annotated by three unique annotators to measure overall consensus. Further details on the task setup can be found in Appendix B.\\nThe results of our human agreement study are shown in Table 5. The event match performance of entailment models across two datasets and all event domains achieves a majority vote agreement of 0.55 with the entailment judgments. However, at least one annotator agrees with the event match 75% of the time. We also measure the Krippendorff's Alpha to measure the inter-annotator agreement. The alpha score for our task is 0.43, which is considered moderate agreement, but does reflect the subjectivity of the task.\\nWe argue that not all of the entailment mistakes labeled as such truly represent errors. For instance, the any-directional entailment model matches the prediction \\\"Implementation of preventative measures\\\" to two gold events: \\\"people maintain physical distancing to prevent disease spread\\\" and \\\"people are vaccinated against the disease.\\\" Although the level of specificity differs between the two, we argue that any-directional entailment can be a reasonable candidate for automatic metric while serving the purpose of assigning soft matches between gold and predicted events. Cases like this are often marked as not equivalent by Turkers, but we argue that the entailment judgment is still a reliable method for assessing recall.\\nFor a highly precise evaluation protocol, bidirectional entailment can be a suitable candidate, however, as this is a very strict metric, the recall achieved by this evaluation protocol is significantly lower (see Table 7 in the Appendix).\\nWe also conduct a internal human evaluation of the entailment metric at a granular level in Appendix F.\\nThe performance of entailment depends on stylistic matching Table 8 highlights that hu-man agreement of anydirectional entailment improves across all domains for davinci-003 when the schemas are generated with one-shot prompts compared to zero-shot. This signifies that one-shot prompts are beneficial in guiding the language models to generate schemas of a specific style.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Qualitative Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\",\n \"tab_6\"\n ],\n \"text\": \"While the length analysis in Table 3 shows differences between various domains and schema sources, the stylistic differences go beyond length in ways that are hard to precisely quantify. We show examples from the disease-outbreak domain in Table 6 to highlight these differences and qualitatively depict the variation in the writing style of events across human-curated datasets (RESIN-11 and CuratedSchemas) and generations from language models (text-davinci-001 and flan-t5-xxl) in zero-shot and one-shot settings.\\nWe see that event samples from AltShemas are more formal and shorter in size as compared to RESIN-11 which have a high length variance and are more natural language like. This style also differs from the zero-shot generations from davinci-003 and flan-t5-xxl. A controlled generation using one-shot prompts derived from RESIN-11 schema can be used to attempt to match the event description style of the gold schemas.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b3\",\n \"b6\",\n \"b23\",\n \"b31\",\n \"b36\",\n \"b18\",\n \"b10\",\n \"b32\",\n \"b12\",\n \"b39\",\n \"b9\",\n \"b35\",\n \"b30\",\n \"b38\",\n \"b21\",\n \"b37\"\n ],\n \"table_ref\": [],\n \"text\": \"Event-centric modeling and schema induction Methods performing schema induction can be categorized into simple and complex schema induction. Simple schema induction methods rely on identifying event triggers and participants and do no incorporate the relationships between events (Chambers, 2013;Cheung et al., 2013;Nguyen et al., 2015;Sha et al., 2016;Yuan et al., 2018). Recent work (Li et al., 2021;Du et al., 2022) focuses on generating complex schemas that incorporate temporal as well as event argument relationships but assume availability of large amount of event relevant corpora. Existing event datasets such as MAVEN (Wang et al., 2020) and event-centric knowledge bases such as Event-Wiki (Ge et al., 2018), but working with these datasets naturally restricts a system designer to a fixed ontology.\\nClosest to our work, Zhang et al. (2023) also generate schemas including a GPT prompting stage. However, they follow this with a stage of grounding to an ontology, sidestepping the challenges with evaluation we tackle in this work and losing the ability to homogenize between two different sources of schemas. Dror et al. (2022) use language models to generate large number of source documents about a topic that can be used to extract events and relations to build schemas in a zeroshot manner. However, their method uses language models to generate documents containing relevant information which is further used to extract events using event extraction methods. In this work, we provide a way to both generate and automatically evaluate light event schemas in natural language making the process less dependent on traditional event extraction and schema matching pipleines.\\nTextual Entailment Natural Language Inference research focuses on establishing entailment between a premise and a hypothesis pair. Although most of the previous work focuses on sentence level hypothesis and premise pair, recent datasets such as DocNLI (Yin et al., 2021) and ContractNLI (Koreeda and Manning, 2021) push the boundaries to extend NLI models to longer multi-sentence inputs and real-work datasets. Schuster et al. (2022) explore the utility of NLI models on longer inputs using a \\\"stretching\\\" form of aggregation, namely maxing over possible alignments to a document. It is common to see similarity models being used to judge similarity between two sentences in ROUGE, BLEU and BERTScore (Zhang et al., 2020). However, recent works recommend the usage of NLI models as evaluation metrics for Ab-stract Summarization (Maynez et al., 2020) as they capture the faithfulness and factuality of summaries better than standard metrics. (Zhang and Bansal, 2021) explore the usage of NLI models to automate evaluation of summarization tasks which can also benefit automated best model checkpointing. In this work, we explore using NLI as a metric for schema coverage matching directly in natural language.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we explored the ability of language models to draft light event schemas. In both zeroand one-shot settings, we showed that large language models can generate coherent, varied sequences of events in natural language that overlap substantially with human-curated events across several domains of interest. We show that textual entailment can be used to evaluate these matches. We believe our work can pave the way for future efforts looking at how explicit knowledge like schemas can be used in tandem with large language models to make predictions. Streamlining the ability to generate schemas and then curate them with human intervention will be an important step to scaling this method to work across many domains.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [\n \"b18\"\n ],\n \"table_ref\": [],\n \"text\": \"The schemas we produce in this work are, by choice, lighter weight than representations used in some prior work. Past work (Li et al., 2021) has explored schemas with graph-structured ordering. While these schemas can express a richer set of partial ordering and mutual exclusion relationships between events, they are both cumbersome to produce and relatively little work has shown the ability to use them to perform complex inferences. Our view is that more complex structural relationships should also be specified in natural language for maximal compatibility with prediction based on large language models; we use this for future work. Human curation can also be used to impart these features for use in downstream applications.\\nA second limitation is that the robustness of event recall evaluation using textual entailment is dependent on the stylistic similarities between generated and gold schemas. While we analyze this in the paper, stronger textual entailment systems down the road can potentially be useful to improve the precision of our performance estimates further.\\nFinally, we note that schema-mediated prediction with neural models is an emerging and ongoing area of research. Therefore, there are not standard systems we can plug our schemas into for downstream evaluation. Nevertheless, we believe that these knowledge structures can be intrinsically evaluated, and high quality representations will pave the way for future work in this area.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Prompts\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Zero-shot Prompts The text-davinci-003 zero-shot experiments predominantly use the prompt below. The domains d that we consider are natural disaster, disease outbreak, international conflict, mass shooting, IED attack, and kidnapping Temporally-aided prompt \\\"List 10 things that each happen (1) before; (2) during; and (3) after a [d]?\\nBefore a [d], there are several things that can happen:\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"1.\\\"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Prompt Union This approach uses three different prompt templates as shown below. In addition to the previously described prompt above, the two additional prompts that we sample for prompt union experiments are: For generating events for one domain we sample from such prompts from three other domains and generate the output.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"B Annotation Task Description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The task description provided to Turk workers on Amazon Mechanical Turk is shown in Figure 3.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C Comparison of any-directional and bi-directional entailment over event recall\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We compare the event recall values predicted by anydirectional and bidirectional entailment as a metric in 7.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"D Example of davinci-base output\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Figure 4 shows a representative example for output from davinic-base using zero-shot prompt. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"E Dataset Description\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_9\"\n ],\n \"text\": \"In this work we experiment with a subset of two datasets: 1) RESIN-11 and 2) CuratedSchemas. dataset statistics of the subsets are described in Table 9.\"\n },\n {\n \"figure_ref\": [\n \"fig_5\"\n ],\n \"heading\": \"E.1 CuratedSchemas\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"CuratedSchemas was produced over a period of roughly 1 year. Around 10 distinct annotators, all experts in natural language processing, contributed events to build up these schemas over several stages. Multiple rounds of feedback were collected based on a downstream prediction task and human assessment of outputs from that task, not conducted by the authors of this paper. The schemas balance precision and recall, aiming to cover each domain thoroughly but not include events that are too lowlevel or would never conceivably be reported in text. Additional layers of structure are presented in the CuratedSchemas dataset, but here we only evaluate against the collection of abstract events. An example of events from the Disease Outbreak domain of CuratedSchemas is shown in Figure 6.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"E.2 RESIN Schemas\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The natural language description of the event is derived from the published RESIN schemas using the fields of qnode, qlabel, description, and/or ta1explanation as applicable. An example of the events from the Disease Outbreak domain derived from schemas published in RESIN-11 is shown in Figure 7.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"F Granular Human Agreement Study\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_8\"\n ],\n \"text\": \"We perform a granular human agreement study to investigate the trends and impacts that various domains, prompting methods, and language models have on the performance of entailment models for automatic evaluation. This study is performed by the authors of the paper and is illustrated in Table 8.\\nFigure 5 shows examples of our annotation across different domains for human agreement with the pairings predicted by the any-directional entailment model between gold and generated schema events. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Dataset\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-05-24"},"doi":{"kind":"string","value":"10.18653/v1/P19-1470"},"references":{"kind":"list like","value":[{"authors":"Antoine Bosselut; Hannah Rashkin; Maarten Sap; Chaitanya Malaviya; Asli Celikyilmaz; Yejin Choi","journal":"Association for Computational Linguistics","ref_id":"b0","title":"COMET: Commonsense transformers for automatic knowledge graph construction","year":"2019"},{"authors":"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel Ziegler; Jeffrey Wu; Clemens Winter; Chris Hesse; Mark Chen; Eric Sigler; Mateusz Litwin; Scott Gray; Benjamin Chess; Jack Clark; Christopher Berner; Sam Mccandlish; Alec Radford; Ilya Sutskever; Dario Amodei","journal":"","ref_id":"b1","title":"Language models are few-shot learners","year":"2020"},{"authors":"Curran Associates; Inc ","journal":"","ref_id":"b2","title":"","year":""},{"authors":"Nathanael Chambers","journal":"","ref_id":"b3","title":"Event schema induction with a probabilistic entity-driven model","year":"2013"},{"authors":"Nathanael Chambers; Dan Jurafsky","journal":"Association for Computational Linguistics","ref_id":"b4","title":"Unsupervised learning of narrative event chains","year":"2008"},{"authors":"Nathanael Chambers; Dan Jurafsky","journal":"Association for Computational Linguistics","ref_id":"b5","title":"Unsupervised learning of narrative schemas and their participants","year":"2009"},{"authors":"Jackie Chi; Kit Cheung; Hoifung Poon; Lucy Vanderwende","journal":"","ref_id":"b6","title":"Probabilistic frame induction","year":"2013"},{"authors":"Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Yunxuan Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma; Albert Webson; Shane Shixiang; Zhuyun Gu; Mirac Dai; Xinyun Suzgun; Aakanksha Chen; Alex Chowdhery; Marie Castro-Ros; Kevin Pellat; Dasha Robinson; Sharan Valter; Gaurav Narang; Adams Mishra; Vincent Yu; Yanping Zhao; Andrew Huang; Hongkun Dai; Slav Yu; Ed H Petrov; Jeff Chi; Jacob Dean; Adam Devlin; Denny Roberts; Quoc V Zhou; Jason Le; Wei","journal":"","ref_id":"b7","title":"Scaling Instruction-Finetuned Language Models","year":"2022"},{"authors":"Ido Dagan; Oren Glickman; Bernardo Magnini","journal":"Springer-Verlag","ref_id":"b8","title":"The PASCAL Recognising Textual Entailment Challenge","year":"2005"},{"authors":"Rotem Dror; Haoyu Wang; Dan Roth","journal":"","ref_id":"b9","title":"Zeroshot on-the-fly event schema induction","year":"2022"},{"authors":"Xinya Du; Zixuan Zhang; Sha Li; Pengfei Yu; Hongwei Wang; Tuan Lai; Xudong Lin; Ziqi Wang; Iris Liu; Ben Zhou","journal":"","ref_id":"b10","title":"Resin-11: Schema-guided event prediction for 11 newsworthy scenarios","year":"2022"},{"authors":"Tobias Falke; Leonardo F R Ribeiro; Ajie Prasetya; Ido Utama; Iryna Dagan; Gurevych","journal":"Association for Computational Linguistics","ref_id":"b11","title":"Ranking generated summaries by correctness: An interesting but challenging application for natural language inference","year":"2019"},{"authors":"Tao Ge; Lei Cui; Baobao Chang; Zhifang Sui; Furu Wei; Ming Zhou","journal":"","ref_id":"b12","title":"Eventwiki: a knowledge base of major events","year":"2018"},{"authors":"Mark Granroth; -Wilding ; Stephen Clark","journal":"","ref_id":"b13","title":"What Happens Next? Event Prediction Using a Compositional Neural Network Model","year":"2016"},{"authors":"Ari Holtzman; Jan Buys; Li Du; Maxwell Forbes; Yejin Choi","journal":"","ref_id":"b14","title":"The Curious Case of Neural Text Degeneration","year":"2020"},{"authors":"Yuta Koreeda; Christopher Manning","journal":"Association for Computational Linguistics","ref_id":"b15","title":"Con-tractNLI: A dataset for document-level natural language inference for contracts","year":"2021"},{"authors":"Mahnaz Koupaee; Greg Durrett; Nathanael Chambers; Niranjan Balasubramanian","journal":"Association for Computational Linguistics","ref_id":"b16","title":"Don't let discourse confine your model: Sequence perturbations for improved event language models","year":"2021"},{"authors":"Philippe Laban; Tobias Schnabel; Paul N Bennett; Marti A Hearst","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b17","title":"SummaC: Re-visiting NLIbased models for inconsistency detection in summarization","year":"2022"},{"authors":"Manling Li; Sha Li; Zhenhailong Wang; Lifu Huang; Kyunghyun Cho; Heng Ji; Jiawei Han; Clare Voss","journal":"Association for Computational Linguistics","ref_id":"b18","title":"The future is not one-dimensional: Complex event schema induction by graph modeling for event prediction","year":"2021"},{"authors":"Alisa Liu; Swabha Swayamdipta; Noah A Smith; Yejin Choi","journal":"Association for Computational Linguistics","ref_id":"b19","title":"WANLI: Worker and AI collaboration for natural language inference dataset creation","year":"2022"},{"authors":"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b20","title":"RoBERTa: A robustly optimized BERT pretraining approach","year":"2019"},{"authors":"Joshua Maynez; Shashi Narayan; Bernd Bohnet; Ryan Mcdonald","journal":"","ref_id":"b21","title":"On faithfulness and factuality in abstractive summarization","year":"2020"},{"authors":"Ashutosh Modi; Ivan Titov; Vera Demberg; Asad Sayeed; Manfred Pinkal","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b22","title":"Modeling semantic expectation: Using script knowledge for referent prediction","year":"2017"},{"authors":"Kiem-Hieu Nguyen; Xavier Tannier; Olivier Ferret; Romaric Besançon","journal":"","ref_id":"b23","title":"Generative event schema induction with entity disambiguation","year":"2015"},{"authors":" Openai","journal":"","ref_id":"b24","title":"Model Index for Researchers","year":"2022"},{"authors":"Long Ouyang; Jeff Wu; Xu Jiang; Diogo Almeida; Carroll L Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray","journal":"","ref_id":"b25","title":"Training language models to follow instructions with human feedback","year":"2022"},{"authors":"Denis Paperno; Germán Kruszewski; Angeliki Lazaridou; Ngoc Quan Pham; Raffaella Bernardi; Sandro Pezzelle; Marco Baroni; Gemma Boleda; Raquel Fernández","journal":"Association for Computational Linguistics","ref_id":"b26","title":"The LAMBADA dataset: Word prediction requiring a broad discourse context","year":"2016"},{"authors":"Karl Pichotta; Raymond J Mooney","journal":"Association for Computational Linguistics","ref_id":"b27","title":"Using sentence-level LSTM language models for script inference","year":"2016"},{"authors":"Rachel Rudinger; Pushpendre Rastogi; Francis Ferraro; Benjamin Van Durme","journal":"","ref_id":"b28","title":"Script induction as language modeling","year":"2015"},{"authors":"Maarten Sap; Le Ronan; Emily Bras; Chandra Allaway; Nicholas Bhagavatula; Hannah Lourie; Brendan Rashkin; Noah A Roof; Yejin Smith; Choi","journal":"","ref_id":"b29","title":"ATOMIC: An Atlas of Machine Commonsense for If-Then Reasoning","year":"2019"},{"authors":"Tal Schuster; Sihao Chen; Senaka Buthpitiya; Alex Fabrikant; Donald Metzler","journal":"","ref_id":"b30","title":"Stretching Sentence-pair NLI Models to Reason over Long Documents and Clusters","year":"2022"},{"authors":"Lei Sha; Sujian Li; Baobao Chang; Zhifang Sui","journal":"","ref_id":"b31","title":"Joint learning templates and slots for event schema induction","year":"2016"},{"authors":"Xiaozhi Wang; Ziqi Wang; Xu Han; Wangyi Jiang; Rong Han; Zhiyuan Liu; Juanzi Li; Peng Li; Yankai Lin; Jie Zhou","journal":"","ref_id":"b32","title":"MAVEN: A massive general domain event detection dataset","year":"2020"},{"authors":"Noah Weber; Niranjan Balasubramanian; Nathanael Chambers","journal":"","ref_id":"b33","title":"Event Representations with Tensorbased Compositions","year":"2018"},{"authors":"Adina Williams; Nikita Nangia; Samuel Bowman","journal":"Association for Computational Linguistics","ref_id":"b34","title":"A broad-coverage challenge corpus for sentence understanding through inference","year":"2018"},{"authors":"Wenpeng Yin; Dragomir Radev; Caiming Xiong","journal":"","ref_id":"b35","title":"DocNLI: A large-scale dataset for documentlevel natural language inference","year":"2021"},{"authors":"Quan Yuan; Xiang Ren; Wenqi He; Chao Zhang; Xinhe Geng; Lifu Huang; Heng Ji; Chin-Yew Lin; Jiawei Han","journal":"","ref_id":"b36","title":"Open-schema event profiling for massive news corpora","year":"2018"},{"authors":"Shiyue Zhang; Mohit Bansal","journal":"Association for Computational Linguistics","ref_id":"b37","title":"Finding a balanced degree of automation for summary evaluation","year":"2021"},{"authors":"Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi","journal":"","ref_id":"b38","title":"BERTScore: Evaluating Text Generation with BERT","year":"2020"},{"authors":"Tianyi Zhang; Isaac Tham; Zhaoyi Hou; Jiaxuan Ren; Liyang Zhou; Hainiu Xu; Li Zhang; Lara J Martin; Rotem Dror; Sha Li; Heng Ji; Martha Palmer; Susan Brown; Reece Suchocki; Chris Callison-Burch","journal":"","ref_id":"b39","title":"Human-in-the-Loop Schema Induction","year":"2023"},{"authors":"Andy Zou; Tristan Xiao; Ryan Jia; Joe Kwon; Mantas Mazeika; Richard Li; Dawn Song; Jacob Steinhardt; Owain Evans; Dan Hendrycks","journal":"","ref_id":"b40","title":"Forecasting Future World Events with Neural Networks","year":"2022"}],"string":"[\n {\n \"authors\": \"Antoine Bosselut; Hannah Rashkin; Maarten Sap; Chaitanya Malaviya; Asli Celikyilmaz; Yejin Choi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b0\",\n \"title\": \"COMET: Commonsense transformers for automatic knowledge graph construction\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel Ziegler; Jeffrey Wu; Clemens Winter; Chris Hesse; Mark Chen; Eric Sigler; Mateusz Litwin; Scott Gray; Benjamin Chess; Jack Clark; Christopher Berner; Sam Mccandlish; Alec Radford; Ilya Sutskever; Dario Amodei\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Curran Associates; Inc \",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Nathanael Chambers\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Event schema induction with a probabilistic entity-driven model\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Nathanael Chambers; Dan Jurafsky\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b4\",\n \"title\": \"Unsupervised learning of narrative event chains\",\n \"year\": \"2008\"\n },\n {\n \"authors\": \"Nathanael Chambers; Dan Jurafsky\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b5\",\n \"title\": \"Unsupervised learning of narrative schemas and their participants\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Jackie Chi; Kit Cheung; Hoifung Poon; Lucy Vanderwende\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Probabilistic frame induction\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Yunxuan Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma; Albert Webson; Shane Shixiang; Zhuyun Gu; Mirac Dai; Xinyun Suzgun; Aakanksha Chen; Alex Chowdhery; Marie Castro-Ros; Kevin Pellat; Dasha Robinson; Sharan Valter; Gaurav Narang; Adams Mishra; Vincent Yu; Yanping Zhao; Andrew Huang; Hongkun Dai; Slav Yu; Ed H Petrov; Jeff Chi; Jacob Dean; Adam Devlin; Denny Roberts; Quoc V Zhou; Jason Le; Wei\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Scaling Instruction-Finetuned Language Models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ido Dagan; Oren Glickman; Bernardo Magnini\",\n \"journal\": \"Springer-Verlag\",\n \"ref_id\": \"b8\",\n \"title\": \"The PASCAL Recognising Textual Entailment Challenge\",\n \"year\": \"2005\"\n },\n {\n \"authors\": \"Rotem Dror; Haoyu Wang; Dan Roth\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Zeroshot on-the-fly event schema induction\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xinya Du; Zixuan Zhang; Sha Li; Pengfei Yu; Hongwei Wang; Tuan Lai; Xudong Lin; Ziqi Wang; Iris Liu; Ben Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Resin-11: Schema-guided event prediction for 11 newsworthy scenarios\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tobias Falke; Leonardo F R Ribeiro; Ajie Prasetya; Ido Utama; Iryna Dagan; Gurevych\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b11\",\n \"title\": \"Ranking generated summaries by correctness: An interesting but challenging application for natural language inference\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Tao Ge; Lei Cui; Baobao Chang; Zhifang Sui; Furu Wei; Ming Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Eventwiki: a knowledge base of major events\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Mark Granroth; -Wilding ; Stephen Clark\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"What Happens Next? Event Prediction Using a Compositional Neural Network Model\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Ari Holtzman; Jan Buys; Li Du; Maxwell Forbes; Yejin Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"The Curious Case of Neural Text Degeneration\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yuta Koreeda; Christopher Manning\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b15\",\n \"title\": \"Con-tractNLI: A dataset for document-level natural language inference for contracts\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Mahnaz Koupaee; Greg Durrett; Nathanael Chambers; Niranjan Balasubramanian\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b16\",\n \"title\": \"Don't let discourse confine your model: Sequence perturbations for improved event language models\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Philippe Laban; Tobias Schnabel; Paul N Bennett; Marti A Hearst\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b17\",\n \"title\": \"SummaC: Re-visiting NLIbased models for inconsistency detection in summarization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Manling Li; Sha Li; Zhenhailong Wang; Lifu Huang; Kyunghyun Cho; Heng Ji; Jiawei Han; Clare Voss\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b18\",\n \"title\": \"The future is not one-dimensional: Complex event schema induction by graph modeling for event prediction\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Alisa Liu; Swabha Swayamdipta; Noah A Smith; Yejin Choi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b19\",\n \"title\": \"WANLI: Worker and AI collaboration for natural language inference dataset creation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"RoBERTa: A robustly optimized BERT pretraining approach\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Joshua Maynez; Shashi Narayan; Bernd Bohnet; Ryan Mcdonald\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"On faithfulness and factuality in abstractive summarization\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ashutosh Modi; Ivan Titov; Vera Demberg; Asad Sayeed; Manfred Pinkal\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b22\",\n \"title\": \"Modeling semantic expectation: Using script knowledge for referent prediction\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Kiem-Hieu Nguyen; Xavier Tannier; Olivier Ferret; Romaric Besançon\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Generative event schema induction with entity disambiguation\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \" Openai\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Model Index for Researchers\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Long Ouyang; Jeff Wu; Xu Jiang; Diogo Almeida; Carroll L Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Training language models to follow instructions with human feedback\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Denis Paperno; Germán Kruszewski; Angeliki Lazaridou; Ngoc Quan Pham; Raffaella Bernardi; Sandro Pezzelle; Marco Baroni; Gemma Boleda; Raquel Fernández\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b26\",\n \"title\": \"The LAMBADA dataset: Word prediction requiring a broad discourse context\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Karl Pichotta; Raymond J Mooney\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b27\",\n \"title\": \"Using sentence-level LSTM language models for script inference\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Rachel Rudinger; Pushpendre Rastogi; Francis Ferraro; Benjamin Van Durme\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Script induction as language modeling\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Maarten Sap; Le Ronan; Emily Bras; Chandra Allaway; Nicholas Bhagavatula; Hannah Lourie; Brendan Rashkin; Noah A Roof; Yejin Smith; Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"ATOMIC: An Atlas of Machine Commonsense for If-Then Reasoning\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Tal Schuster; Sihao Chen; Senaka Buthpitiya; Alex Fabrikant; Donald Metzler\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Stretching Sentence-pair NLI Models to Reason over Long Documents and Clusters\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Lei Sha; Sujian Li; Baobao Chang; Zhifang Sui\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Joint learning templates and slots for event schema induction\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Xiaozhi Wang; Ziqi Wang; Xu Han; Wangyi Jiang; Rong Han; Zhiyuan Liu; Juanzi Li; Peng Li; Yankai Lin; Jie Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"MAVEN: A massive general domain event detection dataset\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Noah Weber; Niranjan Balasubramanian; Nathanael Chambers\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Event Representations with Tensorbased Compositions\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Adina Williams; Nikita Nangia; Samuel Bowman\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b34\",\n \"title\": \"A broad-coverage challenge corpus for sentence understanding through inference\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Wenpeng Yin; Dragomir Radev; Caiming Xiong\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"DocNLI: A large-scale dataset for documentlevel natural language inference\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Quan Yuan; Xiang Ren; Wenqi He; Chao Zhang; Xinhe Geng; Lifu Huang; Heng Ji; Chin-Yew Lin; Jiawei Han\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Open-schema event profiling for massive news corpora\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Shiyue Zhang; Mohit Bansal\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b37\",\n \"title\": \"Finding a balanced degree of automation for summary evaluation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"BERTScore: Evaluating Text Generation with BERT\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Tianyi Zhang; Isaac Tham; Zhaoyi Hou; Jiaxuan Ren; Liyang Zhou; Hainiu Xu; Li Zhang; Lara J Martin; Rotem Dror; Sha Li; Heng Ji; Martha Palmer; Susan Brown; Reece Suchocki; Chris Callison-Burch\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Human-in-the-Loop Schema Induction\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Andy Zou; Tristan Xiao; Ryan Jia; Joe Kwon; Mantas Mazeika; Richard Li; Dawn Song; Jacob Steinhardt; Owain Evans; Dan Hendrycks\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Forecasting Future World Events with Neural Networks\",\n \"year\": \"2022\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,323.24,620.91,201.9,21.12],"formula_id":"formula_0","formula_text":"r(s, Ŝ) = max ŝ∈ Ŝ (max(E(s, ŝ), E(ŝ, s)))(1)"},{"formula_coordinates":[5,334.93,699.92,50.39,9.57],"formula_id":"formula_1","formula_text":"x, ∀x / ∈ [d]."}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 323.24,\n 620.91,\n 201.9,\n 21.12\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"r(s, Ŝ) = max ŝ∈ Ŝ (max(E(s, ŝ), E(ŝ, s)))(1)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 334.93,\n 699.92,\n 50.39,\n 9.57\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"x, ∀x / ∈ [d].\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"Drafting Event Schemas using Language Models"},"abstract":{"kind":"string","value":"Past work has studied event prediction and event language modeling, sometimes mediated through structured representations of knowledge in the form of event schemas. Such schemas can lead to explainable predictions and forecasting of unseen events given incomplete information. In this work, we look at the process of creating such schemas to describe complex events. We use large language models (LLMs) to draft schemas directly in natural language, which can be further refined by human curators as necessary. Our focus is on whether we can achieve sufficient diversity and recall of key events and whether we can produce the schemas in a sufficiently descriptive style. We show that large language models are able to achieve moderate recall against schemas taken from two different datasets, with even better results when multiple prompts and multiple samples are combined. Moreover, we show that textual entailment methods can be used for both matching schemas to instances of events as well as evaluating overlap between gold and predicted schemas. Our method paves the way for easier distillation of event knowledge from large language model into schemas."},"authors":{"kind":"string","value":"Anisha Gunjal; Greg Durrett"},"figures":{"kind":"list like","value":[{"figure_caption":"\"List causes and events that can happen over the course of a [d]? Causes of a [d]: 1.\" Causes and temporally-aided prompt \"List causes and events that can happen before, during and after a [d]? Causes of a [d]: 1.\" One-Shot Prompts Figure 2 depicts a sample of the one-shot prompts we use to generate events.","figure_data":"","figure_id":"fig_0","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Example of one-shot prompt used to generate events.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Task Description for Human Agreement Annotation for Event Matching on AMT","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Examples of output from davinci-base.","figure_data":"","figure_id":"fig_3","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: Examples of annotation for human agreement with the predictions of the any-directional entailment model.","figure_data":"","figure_id":"fig_4","figure_label":"5","figure_type":"figure"},{"figure_caption":"Figure 6 :6Figure 6: Example of events in the Disease Outbreak Schema from CuratedSchemas dataset","figure_data":"","figure_id":"fig_5","figure_label":"6","figure_type":"figure"},{"figure_caption":"","figure_data":"DomainRESIN-11 CuratedSchemas text-davinci-003 flan-t5-xxl text-davinci-003 zero-shot zero-shot one-shotNatural Disaster4.473.365.684.756.68International Conflict5.272.552.864.814.89Mass Shooting5.243.328.66.626.02Disease Outbreak7.574.455.087.15.95Kidnapping7.084.55.076.676.11IED4.873.245.486.878.11Mean5.753.575.466.146.29Gold SchemaAvg. Event RecallRESIN0.62CuratedSchemas 0.46","figure_id":"tab_2","figure_label":"3","figure_type":"table"},{"figure_caption":"Measuring the overlap between events in the gold schemas: RESIN and CuratedSchemas. We use any-directional entailment to get an estimate of the overlap between two distinct human-curated schemas.","figure_data":"","figure_id":"tab_3","figure_label":"4","figure_type":"table"},{"figure_caption":"Examples of output events from the different gold annotations we use and large language models.","figure_data":"","figure_id":"tab_6","figure_label":"6","figure_type":"table"},{"figure_caption":"Event recall difference between anydirectional and bidirectional entailment.","figure_data":"DomainGold SchemaAny-directional Bi-directionalDisease OutbreakRESIN0.380.05CuratedSchemas 0.290.06KidnappingRESIN0.420.05CuratedSchemas 0.540.02International Conflict RESIN0.460.07CuratedSchemas 0.550.11Natural DisasterRESIN0.560.06CuratedSchemas 0.40.06IEDRESIN0.530.08CuratedSchemas 0.420.09","figure_id":"tab_7","figure_label":"7","figure_type":"table"},{"figure_caption":"Granular Human Agreement Study","figure_data":"DomainGold LabelPredictionHuman AgreementIED AttackRESINCuratedSchemas0.74IED AttackCuratedSchemasRESIN0.77Disease Outbreak RESINCuratedSchemas0.38Disease Outbreak CuratedSchemasRESIN0.8All DomainsRESIN+CuratedSchemas davinci-003 (zero-shot ) 0.7All DomainsRESIN+CuratedSchemas flan-t5-xxl (zero-shot)0.61All DomainsRESIN+CuratedSchemas davinci-003 (one-shot)0.79","figure_id":"tab_8","figure_label":"8","figure_type":"table"},{"figure_caption":"Domain Count Average No. of Events Avg. Length of Events Dataset Statistics for RESIN-11 and CuratedSchemas used in this work.","figure_data":"RESIN-11634.335.75CuratedSchemas 638.333.57","figure_id":"tab_9","figure_label":"9","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"\\\"List causes and events that can happen over the course of a [d]? Causes of a [d]: 1.\\\" Causes and temporally-aided prompt \\\"List causes and events that can happen before, during and after a [d]? Causes of a [d]: 1.\\\" One-Shot Prompts Figure 2 depicts a sample of the one-shot prompts we use to generate events.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Example of one-shot prompt used to generate events.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Task Description for Human Agreement Annotation for Event Matching on AMT\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Examples of output from davinci-base.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: Examples of annotation for human agreement with the predictions of the any-directional entailment model.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 :6Figure 6: Example of events in the Disease Outbreak Schema from CuratedSchemas dataset\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"DomainRESIN-11 CuratedSchemas text-davinci-003 flan-t5-xxl text-davinci-003 zero-shot zero-shot one-shotNatural Disaster4.473.365.684.756.68International Conflict5.272.552.864.814.89Mass Shooting5.243.328.66.626.02Disease Outbreak7.574.455.087.15.95Kidnapping7.084.55.076.676.11IED4.873.245.486.878.11Mean5.753.575.466.146.29Gold SchemaAvg. Event RecallRESIN0.62CuratedSchemas 0.46\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Measuring the overlap between events in the gold schemas: RESIN and CuratedSchemas. We use any-directional entailment to get an estimate of the overlap between two distinct human-curated schemas.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Examples of output events from the different gold annotations we use and large language models.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Event recall difference between anydirectional and bidirectional entailment.\",\n \"figure_data\": \"DomainGold SchemaAny-directional Bi-directionalDisease OutbreakRESIN0.380.05CuratedSchemas 0.290.06KidnappingRESIN0.420.05CuratedSchemas 0.540.02International Conflict RESIN0.460.07CuratedSchemas 0.550.11Natural DisasterRESIN0.560.06CuratedSchemas 0.40.06IEDRESIN0.530.08CuratedSchemas 0.420.09\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"7\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Granular Human Agreement Study\",\n \"figure_data\": \"DomainGold LabelPredictionHuman AgreementIED AttackRESINCuratedSchemas0.74IED AttackCuratedSchemasRESIN0.77Disease Outbreak RESINCuratedSchemas0.38Disease Outbreak CuratedSchemasRESIN0.8All DomainsRESIN+CuratedSchemas davinci-003 (zero-shot ) 0.7All DomainsRESIN+CuratedSchemas flan-t5-xxl (zero-shot)0.61All DomainsRESIN+CuratedSchemas davinci-003 (one-shot)0.79\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"8\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Domain Count Average No. of Events Avg. Length of Events Dataset Statistics for RESIN-11 and CuratedSchemas used in this work.\",\n \"figure_data\": \"RESIN-11634.335.75CuratedSchemas 638.333.57\",\n \"figure_id\": \"tab_9\",\n \"figure_label\": \"9\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(Chambers and Jurafsky, 2008)\", \"Explanation\": \"The cited work by Chambers and Jurafsky in 2008 introduced the concept of mining narrative schemas using predicate-role pairs, which has been adopted in later works for event prediction and cloze tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chambers and Jurafsky, 2009)\", \"Explanation\": \"The cited work by Chambers and Jurafsky in 2009 further extended the use of narrative schemas to predicate-argument structures, which have been used in neural approaches for event prediction and cloze tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Rudinger et al., 2015)\", \"Explanation\": \"The cited work by Rudinger et al. in 2015 applied language modeling to event prediction tasks, providing a flexible interface for predicting tokens given context.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Paperno et al., 2016)\", \"Explanation\": \"The cited work by Paperno et al. in 2016 also used language modeling for event prediction and cloze tasks, further demonstrating the effectiveness of this approach.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Pichotta and Mooney, 2016)\", \"Explanation\": \"The cited work by Pichotta and Mooney in 2016 also applied language modeling to event prediction tasks, providing a useful method for predicting tokens given context.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Weber et al., 2018)\", \"Explanation\": \"The cited work by Weber et al. in 2018 also used neural approaches for event prediction tasks, further demonstrating the effectiveness of this method.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Koupaee et al., 2021)\", \"Explanation\": \"The cited work by Koupaee et al. in 2021 also applied neural approaches to event prediction tasks, providing a useful method for predicting tokens given context.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Sap et al., 2019)\", \"Explanation\": \"The cited work ATOMIC provides a structured repository of knowledge that can help enable predictions about the world, which is a key aspect of the citing paper in constructing natural language representations of event knowledge from language models.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Bosselut et al., 2019)\", \"Explanation\": \"The cited work COMET also contributes to the construction of natural language representations of event knowledge by providing a method for extending structured repositories of knowledge, which is a crucial aspect of the research in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Li et al., 2021)\", \"Explanation\": \"The cited work on event ontologies by Li et al. (2021) is a data source for the construction of more structured event schemas, which the citing paper aims to achieve by constructing light schemas with high recall over a set of events in a domain.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Granroth-Wilding and Clark, 2016)\", \"Explanation\": \"The cited work by Granroth-Wilding and Clark provides a method for evaluating schematic knowledge in the form of cloze tasks, which the citing paper adopts to evaluate the event coverage of their schemas.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Modi et al., 2017)\", \"Explanation\": \"The cited work by Modi et al. also contributes a method for evaluating schematic knowledge, which the citing paper may have used to further refine or augment the evaluation process.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Weber et al., 2018)\", \"Explanation\": \"The cited work by Weber et al. may have provided another method for evaluating schematic knowledge, which the citing paper may have considered or incorporated in their evaluation process.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Li et al., 2021)\", \"Explanation\": \"The cited work by Li et al. may have provided a dataset or a specific ontology that the citing paper used in their evaluation of event coverage in schemas.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Dror et al., 2022;Zhang et al., 2023)\", \"Explanation\": \"The cited works provide measures that are tied to lexical expression of schema predicates, which the citing paper adopts to evaluate schema generation in closed ontologies.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Dagan et al., 2005;Williams et al., 2018)\", \"Explanation\": \"The cited works introduce textual entailment methods for evaluating groundedness of summaries, which the citing paper extends to compare drafted schemas to ground-truth schemas in a more precise manner.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Falke et al., 2019;Zhang and Bansal, 2021;Laban et al., 2022)\", \"Explanation\": \"The cited works provide a specific application of textual entailment methods to evaluate groundedness of summaries, which the citing paper utilizes as a data source for evaluating schema generation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. (2020) provides the most capable model for the OpenAI GPT-3.5, which the citing paper uses in their experiments to study the capabilities of the model.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ope-nAI, 2022)\", \"Explanation\": \"The cited work by Ope-nAI (2022) serves as the data source for the instruction-tuned model used in the experiment, as it provides the information on the model and its training process.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ouyang et al., 2022)\", \"Explanation\": \"The cited work by Ouyang et al. (2022) contributes to the instruction-tuning process of the model, as it discusses the use of reinforcement learning to improve the model performance based on human preference judgments.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chung et al., 2022)\", \"Explanation\": \"The cited work by Chung et al. (2022) provides the instruction-tuned model (flan-t5-xxl) used in the experiment, which allows the study of a smaller and more cost-effective model for instruction-tuning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(A)\", \"Explanation\": \"The cited work in A provides a qualitative observation of the performance of the model in temporally-aided complex prompt tasks. The citing paper simplifies the prompt into three independent prompts to improve the model performance and generate events that are minimally postprocessed for analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work is the text-davinci-base model, which the citing paper uses to conduct their analysis but excludes from the main results due to the presence of redundancies and the need for human curation to extract relevant information.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Liu et al., 2019)\", \"Explanation\": \"The cited work provides the RoBERTa-large architecture used in the entailment model employed in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Du et al., 2022)\", \"Explanation\": \"The RESIN-11 dataset is used as a data source for the experiments conducted in the citing paper, providing a specific set of gold schemas to work with.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Curated Schemas)\", \"Explanation\": \"The Curated Schemas set is a separate data source used in the experiments, providing a new set of annotated schemas for analysis and study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Dror et al., 2022)\", \"Explanation\": \"The cited work by Dror et al. provides a method for generating complex schemas for events using language models, which the citing paper adopts in their research to improve the performance of event generation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chambers, 2013)\", \"Explanation\": \"The cited work by Chambers (2013) is used as a methodological basis for identifying event triggers and participants in simple schema induction methods.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Cheung et al., 2013)\", \"Explanation\": \"The cited work by Cheung et al. (2013) is also used as a methodological basis for identifying event triggers and participants in simple schema induction methods.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Nguyen et al., 2015)\", \"Explanation\": \"The cited work by Nguyen et al. (2015) is used as a methodological basis for identifying event triggers and participants in simple schema induction methods.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Sha et al., 2016)\", \"Explanation\": \"The cited work by Sha et al. (2016) is used as a methodological basis for identifying event triggers and participants in simple schema induction methods.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Yuan et al., 2018)\", \"Explanation\": \"The cited work by Yuan et al. (2018) is used as a methodological basis for identifying event triggers and participants in simple schema induction methods.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Li et al., 2021)\", \"Explanation\": \"The cited work by Li et al. (2021) focuses on generating complex schemas that incorporate temporal and event argument relationships, which serves as a methodological basis for the research conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Du et al., 2022)\", \"Explanation\": \"The cited work by Du et al. (2022) also focuses on generating complex schemas that incorporate temporal and event argument relationships, which provides a methodological basis for the research conducted in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhang et al., 2023)\", \"Explanation\": \"The cited work by Zhang et al. (2023) generates schemas including a GPT prompting stage, which extends the research conducted in the citing paper by incorporating a similar approach for schema generation.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ge et al., 2018)\", \"Explanation\": \"The cited work by Ge et al. (2018) introduces the Event-Wiki dataset, which serves as a data source for the research conducted in the citing paper on event-centric knowledge bases.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wang et al., 2020)\", \"Explanation\": \"The cited work by Wang et al. (2020) introduces the MAVEN dataset, which serves as a data source for the research conducted in the citing paper on event-centric modeling and schema induction.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2020)\", \"Explanation\": \"The cited work by Zhang et al. provides a method of using similarity models to judge similarity between sentences in ROUGE, BLEU, and BERTScore, which the citing paper adopts in their research on evaluating the quality of summaries using NLI models.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Maynez et al., 2020)\", \"Explanation\": \"The cited work by Maynez et al. recommends the usage of NLI models for evaluating the quality of summaries in Ab-stract Summarization tasks. The citing paper builds upon this recommendation by exploring the usage of NLI models in automated evaluation of summarization tasks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhang and Bansal, 2021)\", \"Explanation\": \"The cited work by Zhang and Bansal explores the usage of NLI models for automated evaluation of summarization tasks. The citing paper further extends this research by exploring the use of NLI as a metric for schema coverage matching in natural language.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Li et al., 2021)\", \"Explanation\": \"The cited work explores the use of graph-structured ordering in schemas, which the citing paper adopts to express a richer set of partial ordering and mutual exclusion relationships between events in their work.\"}, {\"Category\": \"Data Source\", \"Citation\": \"The paper\", \"Explanation\": \"The citing paper uses human curation to impart complex structural relationships in natural language for maximum compatibility with prediction based on large language models.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"Future work\", \"Explanation\": \"The citing paper views the more complex structural relationships as a potential area for future research, with the goal of improving the performance of schema-mediated prediction with neural models.\"}]"}}},{"rowIdx":83,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"I. INTRODUCTION","publication_ref":["b0","b1","b2","b4","b5","b6","b7","b10","b11","b12","b13","b14","b15","b16","b6","b17","b18","b19","b11","b12","b7","b10","b9","b8","b13","b14"],"table_ref":[],"text":"Modern face recognition (FR) systems achieve excellent results even with large-scale recognition problems, as long as the appearance variability of the facial images is reasonably constrained. However, the performance in constrained scenarios does not always translate to real-world scenarios where out-of-distribution data, often of poor quality, still presents a challenge for the majority of existing FR models [1], [2].\nFace image quality assessment (FIQA) techniques aim to assist FR models in such challenging scenarios by providing additional information on the quality of facial images. This quality information can then be used to either reject lowquality samples that typically lead to false match errors or design robust quality-aware face recognition techniques. Thus, different from general purpose image quality assessment (IQA) methods [3]- [5] that commonly measure the perceived visual quality of images by examining explicit image characteristics, such as sharpness, lighting conditions and resolution, FIQA techniques typically try to capture the utility (or fitness) of the given face image for the recognition task [6]. In other words, they measure the usefulness of the sample for face recognition.\nSeveral groups of FIQA techniques that differ slightly in their approach have been proposed so far in the literature [7].\nSupported by ARRS: P2-0250(B), J2-2501(A), Junior Researcher grants.\nThe majority of recent techniques learns quality-estimation networks using (reference) quality information inferred from a large database of face images [8]- [11]. Another notable group of FIQA techniques estimates quality based only on the information present in the input image and the characteristics of the targeted FR system [12], [13]. More recently, approaches have also appeared that incorporate quality estimation directly into the FR process [14], [15], paving the way towards qualityaware face recognition.\nWhile most of the existing FIQA techniques perform well enough to distinguish between high-quality and low-quality facial images, correctly ranking face images of similar quality remains an open problem. The correct (optimal) ordering does not depend solely on the input face images, but also on the targeted FR model. Each model may, in a sense, perceive the quality of individual samples differently due to different model-specific biases introduced by the learning process and the data used for training [16], [17]. This observation also suggests that FIQA techniques, that are not FR model specific, can not determine the correct order for all possible FR models. For this reason, we propose in this paper a novel optimization approach, that attempts to improve the predictive power of any given FIQA approach by incorporating quality information obtained by a particular FR model into the quality scores generated by the selected FIQA approach. Thus, the main contributions of this papers are:\n• A novel optimization approach that incorporates modelspecific quality information into the quality scores produced by existing FIQA techniques with the goal of improving FIQA performance. • An in-depth evaluation of the proposed optimization approach over six FIQA techniques, five datasets, three recognition models and in two settings that demonstrates significant performance gains in most situations.\nII. RELATED WORK In this section, we briefly review previous FIQA research that can be broadly categorized into three groups: (i) analytical, (ii) regression and (iii) model-based techniques. More in-depth information on face quality assessment can be found in the comprehensive survey paper by Schlett et al. [7].\nAnalytical FIQA techniques are mostly unsupervised and rely solely on the information that can be extracted directly from the given input sample. Techniques from this group typically focus on the visual quality of the facial images and, as a result, often exhibit limited performance. The method proposed by Gao et al. [18], for example, attempts to extract quality information based on facial symmetry estimation only. Zhang et al. [19] try to quantify quality based on image illumination information, while Lijun et al. [20] combine multiple cues, such as occlusions, blur and pose for the qualityestimation task. Different from these methods, two analytical FIQA techniques have been proposed recently that in addition to the characteristics of the input image also consider the targeted FR system during the quality estimation task. The first, SER-FIQ by Terhörst et al. [12], uses the properties of dropout layers to quantify quality, while FaceQAN, by Babnik et al. [13], exploits adversarial examples for quality assessment. Both methods were shown to yield state-of-the-art performance for various FR models and different benchmarks.\nRegression-based FIQA techniques are the most numerous and usually learn a quality estimation (regression) model to predict quality scores based on some pseudo (ground-truth) quality labels. FaceQNet [8], for example, trains a ResNet50 model using labels obtained by embedding comparisons with the highest quality image of each subject. Here, the highest quality images are determined using an external quality compliance tool. A similar approach, called PCNet [11], trains a quality-regression network on mated-image pairs, with the goal of predicting the similarity of the image pair. LightQNet [10] builds on the ideas introduced with PCNet, but additionally relies on a so-called Identification Quality (IQ) loss, while SDD-FIQA [9] considers both mated and nonmated similarity scores between a large number of samples to determine the final reference quality for the regression task.\nModel-based FIQA techniques are less common and usually try to combine face recognition and quality assessment in a single quality-aware face recognition task. The main goal of these techniques is to simultaneously produce, for a given sample, its embedding and an estimate of the sample's quality. For example, the approach presented by Shi and Jain [14], estimates a mean and variance vector for each given input sample, where the mean vector represents the embedding, while the variance provides the corresponding uncertainty and can be interpreted as a sample quality estimate. MagFace [15], a similar approach by Meng et al., uses a modified version of the commonly used ArcFace loss, called MagFace loss, which is able to generate quality-aware embeddings, by incorporating quality information into the magnitude of the embedding itself.\nThe method we propose cannot be clearly assigned to one of the above groups, because it relies on an already existing FIQA approach (from any of the three groups) to generate reference quality scores. In a sense, it distills FIQA knowledge from any existing technique. However, if treated as a black-box, the proposed FIQA approach can be thought of as a regressionbased technique, as it trains a regression model using quality labels extracted from a large database."},{"figure_ref":["fig_0"],"heading":"III. METHODOLOGY","publication_ref":[],"table_ref":[],"text":"State-of-the-art FIQA techniques are able to efficiently discriminate between images of distinctly different qualities, yet may not be able to properly distinguish between images of similar quality. To exacerbate this problem, the relative ordering of images of similar quality may additionally depend on the targeted FR model, which not all FIQA techniques take into account. Because face quality assessment aims to quantify the utility of face images for a given FR model, the slight variations in the biases present in modern FR systems may result in different (optimal) quality scores for different FR models. For this reason, we propose in this paper an approach that aims to incorporate FR model-specific quality information into (some initial) quality scores, with the goal of improving the fine-grained performance of existing FIQA techniques. The overall pipeline of the proposed approach, shown in Fig. 1, consists of two main steps: (i) label optimization and (ii) transfer learning. The label-optimization step aims to incorporate additional quality-related information into the baseline quality labels, precomputed using a selected (existing) FIQA approach. The optimized quality labels are then used in a transfer-learning scheme, that uses a pre-trained FR model, extended with a quality-regression head."},{"figure_ref":[],"heading":"A. Method Overview","publication_ref":[],"table_ref":[],"text":"Let Q and M denote a given FIQA method and a pretrained FR model that produce quality scores q I = Q(I) and embeddings e I = M (I), respectively, for an arbitrary input face image I, and {I i } N i=1 denote a large facial image database consisting of N distinct images. The goal of our approach is to train a regression-based quality-estimation model\nQ * = H(M (I))\n, where H represents a quality-regression head, that outperforms the initial FIQA method Q. The model Q * is trained on optimized quality labels {q * i } N i=1 generated by the proposed optimization scheme O. The method relies on information obtained from mated image pairs of the face database {I i } N i=1 . Details on the procedure are given below."},{"figure_ref":[],"heading":"B. Initialization","publication_ref":[],"table_ref":[],"text":"We first extract initial quality scores q i = Q(I i ) and embeddings e i = M (I i ) from all images of the given face image database {I i } N i=1 using the selected FIQA method Q i=1 we first generate mated image pairs. From the image pairs, we compute the pair similarity distribution Xs using the cosine similarity of the image embedings. At the same time, we also construct the quality distribution Xq from the given quality scores {q I i } N i=1 . The mean similarity index id s I i\n, calculated as the average index of all image pairs from I, is then used to update the quality index id Xq (q I i ), using the equation presented above. and chosen FR model M . This initialization step is conducted once and provides the input data for the label optimization and consequently the transfer learning procedures."},{"figure_ref":[],"heading":"C. Label Optimization","publication_ref":["b7","b10"],"table_ref":[],"text":"Looking at past research [8]- [11], we observe that quality information is often inferred from mated image comparisons, where the term mated images refers to two unique images of the same individual. We, therefore, follow this insight and use such information in our optimization approach as well. By computing the similarity of mated image pairs in the embedding space of the given FR model M , we are also able to include FR-specific quality estimates in the optimization.\nSelecting mated image pairs. Large-scale databases contain a significant amount of images for each individual, where many of the images may be nearly identical. Selecting all possible mated pairs, can, therefore, introduce database specific biases into our approach and adversely affect performance.\nTo avoid such issues, we propose a technique for sampling mated image pairs based on clustering. We use a clustering procedure to find groups of similar images and to identify the most informative (and least redundant) mated image pairs. We cluster the embedding space \nE k = {e k i } N k i=1 corresponding\n′ ̸ = c, c ′ ∈ [1, C] to form mated pairs (I k c , I k c ′ )\n. By repeating this process for each image of every individual, we obtain the final mated image pairs for the label-optimization procedure G = {(I i , I j ) l } L l=1 , where i ̸ = j and L = N • (c -1). Optimizing prior quality scores. We aim at optimizing the initial quality labels {q Ii } N i=1 using information provided by the average pair similarity sim Ii of each image. In other words, if an image has a low quality score, yet its average pair similarity is high, we want to increase its quality. Conversely, if the opposite is true, we want to decrease it. The design of the optimization procedure is based on the assumption that the initial quality scores already provide a reasonable estimate of the true quality. We, therefore, try to retain the overall quality distribution over the face database. As a result, we simply rearrange the order of the images in the original quality score distribution generated by the selected FIQA technique Q instead of computing new optimal quality scores that could differ significantly from the initial estimates.\nFrom the list of genuine image pairs G, we first calculate the cosine similarity of all image embedding pairs, i.e.:\nsim cos (e Ii , e Ij ) = e Ii • e Ij ∥e Ii ∥ • ∥e Ij ∥ ,(1)\nwhere e Ii and e Ij denote embeddings of images I i and I j .\nWe then construct the distribution of the computed similarity scores X s ∼ sim cos (e i , e j ), ∀(I i , I j ) ∈ G, by sorting all the pairs according to their calculated similarity score. From the distribution X s we compute for each image I i its average pair index,\nid s Ii = 1 |I| I id Xs (I i , I j ),(2)\nwhere id Xs (•) is a function, that for a given pair (I i , I j ) returns the index of sim cos (e i , e j ) within the similarity distribution X s and I represents the set of all image pairs (I i , I j ), where the quality q Ii is lower then q Ij . The latter follows from the fact that the quality of an image pair is computed as q(I i , I j ) = min(I i , I j ), i.e., it depends only on the image with the lower quality. In addition, we construct a quality score distribution X q ∼ {q Ii } N i=1 , by sorting the quality scores of all images within the given database. The average pair indices and the distribution X q are then used to compute the optimized quality indices id Xq (q * Ii ) = id Xq (q Ii ) + θ • (id s Ii -id Xq (q Ii )),\nwhere θ is an open hyperparameter that controls the degree of change for the indices, and id Xq (•) is a function that returns, for some quality q, its index within the distribution X q . Final steps. To avoid bias from randomly selecting mated pairs, we also repeat the entire process R times, and average the final optimized quality indices, id(q\n* Ii ) = 1 R R r=1 id\nXq r (q * Ii ) for all images. The images are then sorted by the calculated optimized quality indices id(q * Ii ) and assigned the quality score according to the output of the sorted list and the original quality score distribution X q ."},{"figure_ref":[],"heading":"D. Transfer Learning","publication_ref":[],"table_ref":[],"text":"One of the main goals of FIQA techniques is to improve the stability and performance of FR systems. We propose to use a pre-trained state-of-the-art FR model for quality prediction, as it efficiently extracts identity information from given facial images. Moreover, the embeddings generated by state-of-theart FR models already contain some information about the quality of the input image. Formally, from an FR model M , we construct a quality regression model H • M , where H represents a regression head. The regression head H attempts to extract the quality of the input image q i = H(e Ii ) from the embedding e Ii = M (I i ) and is learned through an L 1 loss applied over the optimized labels. To improve the transferlearning process, we normalize the optimized quality scores to the interval of [0, 1]."},{"figure_ref":[],"heading":"IV. EXPERIMENTS AND RESULTS","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_3"],"heading":"A. Experimental Setup","publication_ref":["b20","b21","b12","b14","b11","b8","b10","b22","b23","b24","b25","b26","b27","b28","b29","b6","b12","b14","b21"],"table_ref":[],"text":"Training Database. To train the proposed approach, a largescale database of diverse facial images with rich appearance variability is needed. To this end, we select the VGGFace2 database [21], which contains over 3 million images of more than 9000 individuals. Images in the database vary in terms of facial pose, lighting conditions, image resolution, occlusions, and other similar factors that greatly affect the overall quality and appearance of the facial images, as also illustrated in Fig. 3 for three individuals (in columns) from the database.\nEvaluation Setting. We use six state-of-the-art FIQA methods as baselines to evaluate the proposed optimization scheme, i.e., CR-FIQA [22], FaceQAN [13], MagFace [15], SER-FIQ [12], SDD-FIQA [9] and PCNet [11]. The baselines and the learned quality-regression networks are evaluated on five commonly used benchmark databases: XQLFW [23], CPLFW [24], CFP-FP [25], CALFW [26] and LFW [27]. As the pre-trained FR model, we use ArcFace [28] with a ResNet100 architecture, trained on the MS1MV3 database using an angular margin loss. For the performance evaluation we consider two different scenarios: (i) the same-model scenario, where we use the ArcFace model for both quality-score prediction and generation of the performance indicators, and (ii) the cross-model scenario where ArcFace is used for quality assessment, and the CurricularFace [29] and ElasticFace-Cos+ [30] models are utilized to evaluate performance. Both of the test models are based on the ResNet100 architecture, but CurricularFace was trained on MS1MV2, while ElasticFace was trained with CASIA-WebFaces and MS1MV2.\nPerformance Evaluation. The performance of a FIQA technique directly correlates with its ability to properly rank images of similar quality. Therefore, to evaluate our approach, we follow standard FIQA evaluation methodology and use Error-versus-Reject-Characteristic (ERC) curves as the basis for performance reporting [7], [13], [15], [22]. ERC curves Implementation Details. When clustering the embedding space of each individual within the VGGFace2 database, we decide to set the number of clusters C to 20. Consequently, we generate C -1 = 19, mated image pairs for each image, which means that each individual list of mated pairs consists of approximately 60 million pairs. For the hyperparameter θ we use a relatively small value of 0.001, since the goal is to optimize the already computed baseline quality scores. We repeat the whole process 10 times and average the final results."},{"figure_ref":["fig_4"],"heading":"B. Results","publication_ref":[],"table_ref":["tab_0","tab_3","tab_0"],"text":"Before presenting results, we note that SER-FIQ was used in the construction of the XQLFW database, so any results that combine the two are excluded from the presented analysis.\nSame-Model Results. Table I shows the AUC values produced directly with the original FIQA methods (labeled Baseline) as well as the AUC scores of the quality-regression network trained using our optimized labels (marked Optimized). For readability purposes, the AUC scores are multiplied by 10 3 and rounded to one decimal place. We observe that in most cases the results of our approach are better than those of the underlying FIQA approaches. The only exception to this observation is CR-FIQA, where a concrete improvement is observed only for the hardest of the considered datasets, i.e., XQLFW, while the results for the remaining datasets are mostly close, but deteriorate drastically for CPLFW. For all other methods the results consistently improve, with occasional outliers on the CALFW or CPLFW benchmarks.\nCross-Model Results. methods, CR-FIQA and FaceQAN do not show a clear edge for either the baseline or optimized results. While for the hardest benchmark, XQLFW, the optimized variant always performs better than the baseline variant, the opposite is true for CALFW, which contains cross-age image data. For all other FIQA approaches, the proposed optimization method yields better results, and outperforms the baselines in all cases except for PCNet on CALFW. The results are consistent for both the ElasticFace and CurricularFace model. Cross-Model vs. Same-Model Results. Comparing the cross-model with the same-model results, many similarities can be observed. The performance benefit due to the optimization approach is relatively unconvincing for CR-FIQA, while the results for all other methods are mirrored between the two evaluation schemes. The biggest difference is seen for FaceQAN, where the proposed method performs comparably worse in the cross-model evaluation setting.\nQualitative Analysis. If we look more closely at how the proposed approach works, we see that the distribution of the initial quality scores remains the same under the optimization scheme. This is because the method only rearranges the order of the images and assigns them quality scores from the prior distribution. However, a potential problem with this approach, is that the quality scores of images in higher density areas of the distribution, are harder to change than the quality scores of the images in the lower density areas. This phenomenon is well illustrated in Fig. 4, where for each of the FIQA methods used, a histogram of the prior quality scores over VGGFace2 is presented together with a scatter plot, where each point represents the prior quality of a given image on the x-axis and the optimized quality score on the y-axis. Note how the quality scores in areas of lower density seem to change drastically, while almost no movement is observed in higher density areas.\nAblation Study. To demonstrate how the optimization of the quality labels affects the final results, we present in Table III AUC scores obtained with a quality-regression network trained with the initial (unoptimized) quality labels as well as the performance gain(-)/loss(+) due to the optimization procedure (in brackets). We use the two most difficult benchmarks: CPLFW and XQLFW, as well as the LFW benchmark for this ablation study. From the presented results, we see that the effectiveness of the optimization in the same-model scenario, i.e. with ArcFace, to a certain extent depends on the chosen FIQA technique. For CR-FIQA and SER-FIQ the results do not really seem to favour the optimization approach, as most of the performance gains observed in Table I appear to be a consequence of the transfer learning step. On the cross-model side, the results for both ElasticFace and CurricularFace seem to be more in favour of the optimized labels, with only a few counterexamples on the LFW database.\nRun-time performance. Because we use a regression-based model trained with the optimized quality scores, the run-time performance of our approach is (approximately) the same regardless of the initial FIQA method used as the basis for the reference quality scores. Thus, the proposed transfer learning step can also be seen as a knowledge distillation procedure that allows us to retain the performance of a given FIQA technique while ensuring a (approximately) fixed run-time complexity, as evidenced by the run-times in Table IV -computed on a desktop PC with an Intel i9-10900KF (3.70GHz) CPU and a Nvidia 3090 GPU with 24GB of video RAM. "},{"figure_ref":[],"heading":"V. CONCLUSION","publication_ref":[],"table_ref":[],"text":"We presented a novel optimization approach, that aims to improve the performance of modern FIQA approaches. A thorough evaluation was performed using multiple state-ofthe-art FIQA methods, datasets and FR models. The results of the evaluation showed significant performance improvements in most cases when using the optimization scheme both in the same-model and cross-model setting. As part of our future work, we plan to incorporate multiple sources of quality scores into the optimization procedure to benefit from the complementary quality description provided by different FIQA techniques."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"I. INTRODUCTION\",\n \"publication_ref\": [\n \"b0\",\n \"b1\",\n \"b2\",\n \"b4\",\n \"b5\",\n \"b6\",\n \"b7\",\n \"b10\",\n \"b11\",\n \"b12\",\n \"b13\",\n \"b14\",\n \"b15\",\n \"b16\",\n \"b6\",\n \"b17\",\n \"b18\",\n \"b19\",\n \"b11\",\n \"b12\",\n \"b7\",\n \"b10\",\n \"b9\",\n \"b8\",\n \"b13\",\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"Modern face recognition (FR) systems achieve excellent results even with large-scale recognition problems, as long as the appearance variability of the facial images is reasonably constrained. However, the performance in constrained scenarios does not always translate to real-world scenarios where out-of-distribution data, often of poor quality, still presents a challenge for the majority of existing FR models [1], [2].\\nFace image quality assessment (FIQA) techniques aim to assist FR models in such challenging scenarios by providing additional information on the quality of facial images. This quality information can then be used to either reject lowquality samples that typically lead to false match errors or design robust quality-aware face recognition techniques. Thus, different from general purpose image quality assessment (IQA) methods [3]- [5] that commonly measure the perceived visual quality of images by examining explicit image characteristics, such as sharpness, lighting conditions and resolution, FIQA techniques typically try to capture the utility (or fitness) of the given face image for the recognition task [6]. In other words, they measure the usefulness of the sample for face recognition.\\nSeveral groups of FIQA techniques that differ slightly in their approach have been proposed so far in the literature [7].\\nSupported by ARRS: P2-0250(B), J2-2501(A), Junior Researcher grants.\\nThe majority of recent techniques learns quality-estimation networks using (reference) quality information inferred from a large database of face images [8]- [11]. Another notable group of FIQA techniques estimates quality based only on the information present in the input image and the characteristics of the targeted FR system [12], [13]. More recently, approaches have also appeared that incorporate quality estimation directly into the FR process [14], [15], paving the way towards qualityaware face recognition.\\nWhile most of the existing FIQA techniques perform well enough to distinguish between high-quality and low-quality facial images, correctly ranking face images of similar quality remains an open problem. The correct (optimal) ordering does not depend solely on the input face images, but also on the targeted FR model. Each model may, in a sense, perceive the quality of individual samples differently due to different model-specific biases introduced by the learning process and the data used for training [16], [17]. This observation also suggests that FIQA techniques, that are not FR model specific, can not determine the correct order for all possible FR models. For this reason, we propose in this paper a novel optimization approach, that attempts to improve the predictive power of any given FIQA approach by incorporating quality information obtained by a particular FR model into the quality scores generated by the selected FIQA approach. Thus, the main contributions of this papers are:\\n• A novel optimization approach that incorporates modelspecific quality information into the quality scores produced by existing FIQA techniques with the goal of improving FIQA performance. • An in-depth evaluation of the proposed optimization approach over six FIQA techniques, five datasets, three recognition models and in two settings that demonstrates significant performance gains in most situations.\\nII. RELATED WORK In this section, we briefly review previous FIQA research that can be broadly categorized into three groups: (i) analytical, (ii) regression and (iii) model-based techniques. More in-depth information on face quality assessment can be found in the comprehensive survey paper by Schlett et al. [7].\\nAnalytical FIQA techniques are mostly unsupervised and rely solely on the information that can be extracted directly from the given input sample. Techniques from this group typically focus on the visual quality of the facial images and, as a result, often exhibit limited performance. The method proposed by Gao et al. [18], for example, attempts to extract quality information based on facial symmetry estimation only. Zhang et al. [19] try to quantify quality based on image illumination information, while Lijun et al. [20] combine multiple cues, such as occlusions, blur and pose for the qualityestimation task. Different from these methods, two analytical FIQA techniques have been proposed recently that in addition to the characteristics of the input image also consider the targeted FR system during the quality estimation task. The first, SER-FIQ by Terhörst et al. [12], uses the properties of dropout layers to quantify quality, while FaceQAN, by Babnik et al. [13], exploits adversarial examples for quality assessment. Both methods were shown to yield state-of-the-art performance for various FR models and different benchmarks.\\nRegression-based FIQA techniques are the most numerous and usually learn a quality estimation (regression) model to predict quality scores based on some pseudo (ground-truth) quality labels. FaceQNet [8], for example, trains a ResNet50 model using labels obtained by embedding comparisons with the highest quality image of each subject. Here, the highest quality images are determined using an external quality compliance tool. A similar approach, called PCNet [11], trains a quality-regression network on mated-image pairs, with the goal of predicting the similarity of the image pair. LightQNet [10] builds on the ideas introduced with PCNet, but additionally relies on a so-called Identification Quality (IQ) loss, while SDD-FIQA [9] considers both mated and nonmated similarity scores between a large number of samples to determine the final reference quality for the regression task.\\nModel-based FIQA techniques are less common and usually try to combine face recognition and quality assessment in a single quality-aware face recognition task. The main goal of these techniques is to simultaneously produce, for a given sample, its embedding and an estimate of the sample's quality. For example, the approach presented by Shi and Jain [14], estimates a mean and variance vector for each given input sample, where the mean vector represents the embedding, while the variance provides the corresponding uncertainty and can be interpreted as a sample quality estimate. MagFace [15], a similar approach by Meng et al., uses a modified version of the commonly used ArcFace loss, called MagFace loss, which is able to generate quality-aware embeddings, by incorporating quality information into the magnitude of the embedding itself.\\nThe method we propose cannot be clearly assigned to one of the above groups, because it relies on an already existing FIQA approach (from any of the three groups) to generate reference quality scores. In a sense, it distills FIQA knowledge from any existing technique. However, if treated as a black-box, the proposed FIQA approach can be thought of as a regressionbased technique, as it trains a regression model using quality labels extracted from a large database.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"III. METHODOLOGY\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"State-of-the-art FIQA techniques are able to efficiently discriminate between images of distinctly different qualities, yet may not be able to properly distinguish between images of similar quality. To exacerbate this problem, the relative ordering of images of similar quality may additionally depend on the targeted FR model, which not all FIQA techniques take into account. Because face quality assessment aims to quantify the utility of face images for a given FR model, the slight variations in the biases present in modern FR systems may result in different (optimal) quality scores for different FR models. For this reason, we propose in this paper an approach that aims to incorporate FR model-specific quality information into (some initial) quality scores, with the goal of improving the fine-grained performance of existing FIQA techniques. The overall pipeline of the proposed approach, shown in Fig. 1, consists of two main steps: (i) label optimization and (ii) transfer learning. The label-optimization step aims to incorporate additional quality-related information into the baseline quality labels, precomputed using a selected (existing) FIQA approach. The optimized quality labels are then used in a transfer-learning scheme, that uses a pre-trained FR model, extended with a quality-regression head.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A. Method Overview\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Let Q and M denote a given FIQA method and a pretrained FR model that produce quality scores q I = Q(I) and embeddings e I = M (I), respectively, for an arbitrary input face image I, and {I i } N i=1 denote a large facial image database consisting of N distinct images. The goal of our approach is to train a regression-based quality-estimation model\\nQ * = H(M (I))\\n, where H represents a quality-regression head, that outperforms the initial FIQA method Q. The model Q * is trained on optimized quality labels {q * i } N i=1 generated by the proposed optimization scheme O. The method relies on information obtained from mated image pairs of the face database {I i } N i=1 . Details on the procedure are given below.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B. Initialization\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We first extract initial quality scores q i = Q(I i ) and embeddings e i = M (I i ) from all images of the given face image database {I i } N i=1 using the selected FIQA method Q i=1 we first generate mated image pairs. From the image pairs, we compute the pair similarity distribution Xs using the cosine similarity of the image embedings. At the same time, we also construct the quality distribution Xq from the given quality scores {q I i } N i=1 . The mean similarity index id s I i\\n, calculated as the average index of all image pairs from I, is then used to update the quality index id Xq (q I i ), using the equation presented above. and chosen FR model M . This initialization step is conducted once and provides the input data for the label optimization and consequently the transfer learning procedures.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C. Label Optimization\",\n \"publication_ref\": [\n \"b7\",\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"Looking at past research [8]- [11], we observe that quality information is often inferred from mated image comparisons, where the term mated images refers to two unique images of the same individual. We, therefore, follow this insight and use such information in our optimization approach as well. By computing the similarity of mated image pairs in the embedding space of the given FR model M , we are also able to include FR-specific quality estimates in the optimization.\\nSelecting mated image pairs. Large-scale databases contain a significant amount of images for each individual, where many of the images may be nearly identical. Selecting all possible mated pairs, can, therefore, introduce database specific biases into our approach and adversely affect performance.\\nTo avoid such issues, we propose a technique for sampling mated image pairs based on clustering. We use a clustering procedure to find groups of similar images and to identify the most informative (and least redundant) mated image pairs. We cluster the embedding space \\nE k = {e k i } N k i=1 corresponding\\n′ ̸ = c, c ′ ∈ [1, C] to form mated pairs (I k c , I k c ′ )\\n. By repeating this process for each image of every individual, we obtain the final mated image pairs for the label-optimization procedure G = {(I i , I j ) l } L l=1 , where i ̸ = j and L = N • (c -1). Optimizing prior quality scores. We aim at optimizing the initial quality labels {q Ii } N i=1 using information provided by the average pair similarity sim Ii of each image. In other words, if an image has a low quality score, yet its average pair similarity is high, we want to increase its quality. Conversely, if the opposite is true, we want to decrease it. The design of the optimization procedure is based on the assumption that the initial quality scores already provide a reasonable estimate of the true quality. We, therefore, try to retain the overall quality distribution over the face database. As a result, we simply rearrange the order of the images in the original quality score distribution generated by the selected FIQA technique Q instead of computing new optimal quality scores that could differ significantly from the initial estimates.\\nFrom the list of genuine image pairs G, we first calculate the cosine similarity of all image embedding pairs, i.e.:\\nsim cos (e Ii , e Ij ) = e Ii • e Ij ∥e Ii ∥ • ∥e Ij ∥ ,(1)\\nwhere e Ii and e Ij denote embeddings of images I i and I j .\\nWe then construct the distribution of the computed similarity scores X s ∼ sim cos (e i , e j ), ∀(I i , I j ) ∈ G, by sorting all the pairs according to their calculated similarity score. From the distribution X s we compute for each image I i its average pair index,\\nid s Ii = 1 |I| I id Xs (I i , I j ),(2)\\nwhere id Xs (•) is a function, that for a given pair (I i , I j ) returns the index of sim cos (e i , e j ) within the similarity distribution X s and I represents the set of all image pairs (I i , I j ), where the quality q Ii is lower then q Ij . The latter follows from the fact that the quality of an image pair is computed as q(I i , I j ) = min(I i , I j ), i.e., it depends only on the image with the lower quality. In addition, we construct a quality score distribution X q ∼ {q Ii } N i=1 , by sorting the quality scores of all images within the given database. The average pair indices and the distribution X q are then used to compute the optimized quality indices id Xq (q * Ii ) = id Xq (q Ii ) + θ • (id s Ii -id Xq (q Ii )),\\nwhere θ is an open hyperparameter that controls the degree of change for the indices, and id Xq (•) is a function that returns, for some quality q, its index within the distribution X q . Final steps. To avoid bias from randomly selecting mated pairs, we also repeat the entire process R times, and average the final optimized quality indices, id(q\\n* Ii ) = 1 R R r=1 id\\nXq r (q * Ii ) for all images. The images are then sorted by the calculated optimized quality indices id(q * Ii ) and assigned the quality score according to the output of the sorted list and the original quality score distribution X q .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D. Transfer Learning\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"One of the main goals of FIQA techniques is to improve the stability and performance of FR systems. We propose to use a pre-trained state-of-the-art FR model for quality prediction, as it efficiently extracts identity information from given facial images. Moreover, the embeddings generated by state-of-theart FR models already contain some information about the quality of the input image. Formally, from an FR model M , we construct a quality regression model H • M , where H represents a regression head. The regression head H attempts to extract the quality of the input image q i = H(e Ii ) from the embedding e Ii = M (I i ) and is learned through an L 1 loss applied over the optimized labels. To improve the transferlearning process, we normalize the optimized quality scores to the interval of [0, 1].\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"IV. EXPERIMENTS AND RESULTS\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"A. Experimental Setup\",\n \"publication_ref\": [\n \"b20\",\n \"b21\",\n \"b12\",\n \"b14\",\n \"b11\",\n \"b8\",\n \"b10\",\n \"b22\",\n \"b23\",\n \"b24\",\n \"b25\",\n \"b26\",\n \"b27\",\n \"b28\",\n \"b29\",\n \"b6\",\n \"b12\",\n \"b14\",\n \"b21\"\n ],\n \"table_ref\": [],\n \"text\": \"Training Database. To train the proposed approach, a largescale database of diverse facial images with rich appearance variability is needed. To this end, we select the VGGFace2 database [21], which contains over 3 million images of more than 9000 individuals. Images in the database vary in terms of facial pose, lighting conditions, image resolution, occlusions, and other similar factors that greatly affect the overall quality and appearance of the facial images, as also illustrated in Fig. 3 for three individuals (in columns) from the database.\\nEvaluation Setting. We use six state-of-the-art FIQA methods as baselines to evaluate the proposed optimization scheme, i.e., CR-FIQA [22], FaceQAN [13], MagFace [15], SER-FIQ [12], SDD-FIQA [9] and PCNet [11]. The baselines and the learned quality-regression networks are evaluated on five commonly used benchmark databases: XQLFW [23], CPLFW [24], CFP-FP [25], CALFW [26] and LFW [27]. As the pre-trained FR model, we use ArcFace [28] with a ResNet100 architecture, trained on the MS1MV3 database using an angular margin loss. For the performance evaluation we consider two different scenarios: (i) the same-model scenario, where we use the ArcFace model for both quality-score prediction and generation of the performance indicators, and (ii) the cross-model scenario where ArcFace is used for quality assessment, and the CurricularFace [29] and ElasticFace-Cos+ [30] models are utilized to evaluate performance. Both of the test models are based on the ResNet100 architecture, but CurricularFace was trained on MS1MV2, while ElasticFace was trained with CASIA-WebFaces and MS1MV2.\\nPerformance Evaluation. The performance of a FIQA technique directly correlates with its ability to properly rank images of similar quality. Therefore, to evaluate our approach, we follow standard FIQA evaluation methodology and use Error-versus-Reject-Characteristic (ERC) curves as the basis for performance reporting [7], [13], [15], [22]. ERC curves Implementation Details. When clustering the embedding space of each individual within the VGGFace2 database, we decide to set the number of clusters C to 20. Consequently, we generate C -1 = 19, mated image pairs for each image, which means that each individual list of mated pairs consists of approximately 60 million pairs. For the hyperparameter θ we use a relatively small value of 0.001, since the goal is to optimize the already computed baseline quality scores. We repeat the whole process 10 times and average the final results.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"B. Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_0\",\n \"tab_3\",\n \"tab_0\"\n ],\n \"text\": \"Before presenting results, we note that SER-FIQ was used in the construction of the XQLFW database, so any results that combine the two are excluded from the presented analysis.\\nSame-Model Results. Table I shows the AUC values produced directly with the original FIQA methods (labeled Baseline) as well as the AUC scores of the quality-regression network trained using our optimized labels (marked Optimized). For readability purposes, the AUC scores are multiplied by 10 3 and rounded to one decimal place. We observe that in most cases the results of our approach are better than those of the underlying FIQA approaches. The only exception to this observation is CR-FIQA, where a concrete improvement is observed only for the hardest of the considered datasets, i.e., XQLFW, while the results for the remaining datasets are mostly close, but deteriorate drastically for CPLFW. For all other methods the results consistently improve, with occasional outliers on the CALFW or CPLFW benchmarks.\\nCross-Model Results. methods, CR-FIQA and FaceQAN do not show a clear edge for either the baseline or optimized results. While for the hardest benchmark, XQLFW, the optimized variant always performs better than the baseline variant, the opposite is true for CALFW, which contains cross-age image data. For all other FIQA approaches, the proposed optimization method yields better results, and outperforms the baselines in all cases except for PCNet on CALFW. The results are consistent for both the ElasticFace and CurricularFace model. Cross-Model vs. Same-Model Results. Comparing the cross-model with the same-model results, many similarities can be observed. The performance benefit due to the optimization approach is relatively unconvincing for CR-FIQA, while the results for all other methods are mirrored between the two evaluation schemes. The biggest difference is seen for FaceQAN, where the proposed method performs comparably worse in the cross-model evaluation setting.\\nQualitative Analysis. If we look more closely at how the proposed approach works, we see that the distribution of the initial quality scores remains the same under the optimization scheme. This is because the method only rearranges the order of the images and assigns them quality scores from the prior distribution. However, a potential problem with this approach, is that the quality scores of images in higher density areas of the distribution, are harder to change than the quality scores of the images in the lower density areas. This phenomenon is well illustrated in Fig. 4, where for each of the FIQA methods used, a histogram of the prior quality scores over VGGFace2 is presented together with a scatter plot, where each point represents the prior quality of a given image on the x-axis and the optimized quality score on the y-axis. Note how the quality scores in areas of lower density seem to change drastically, while almost no movement is observed in higher density areas.\\nAblation Study. To demonstrate how the optimization of the quality labels affects the final results, we present in Table III AUC scores obtained with a quality-regression network trained with the initial (unoptimized) quality labels as well as the performance gain(-)/loss(+) due to the optimization procedure (in brackets). We use the two most difficult benchmarks: CPLFW and XQLFW, as well as the LFW benchmark for this ablation study. From the presented results, we see that the effectiveness of the optimization in the same-model scenario, i.e. with ArcFace, to a certain extent depends on the chosen FIQA technique. For CR-FIQA and SER-FIQ the results do not really seem to favour the optimization approach, as most of the performance gains observed in Table I appear to be a consequence of the transfer learning step. On the cross-model side, the results for both ElasticFace and CurricularFace seem to be more in favour of the optimized labels, with only a few counterexamples on the LFW database.\\nRun-time performance. Because we use a regression-based model trained with the optimized quality scores, the run-time performance of our approach is (approximately) the same regardless of the initial FIQA method used as the basis for the reference quality scores. Thus, the proposed transfer learning step can also be seen as a knowledge distillation procedure that allows us to retain the performance of a given FIQA technique while ensuring a (approximately) fixed run-time complexity, as evidenced by the run-times in Table IV -computed on a desktop PC with an Intel i9-10900KF (3.70GHz) CPU and a Nvidia 3090 GPU with 24GB of video RAM. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"V. CONCLUSION\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We presented a novel optimization approach, that aims to improve the performance of modern FIQA approaches. A thorough evaluation was performed using multiple state-ofthe-art FIQA methods, datasets and FR models. The results of the evaluation showed significant performance improvements in most cases when using the optimization scheme both in the same-model and cross-model setting. As part of our future work, we plan to incorporate multiple sources of quality scores into the optimization procedure to benefit from the complementary quality description provided by different FIQA techniques.\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-05-24"},"doi":{"kind":"string","value":""},"references":{"kind":"list like","value":[{"authors":"M Wang; W Deng","journal":"Neurocomputing","ref_id":"b0","title":"Deep Face Recognition: A Survey","year":"2021"},{"authors":"P Li; L Prieto; D Mery; P J Flynn","journal":"IEEE Transactions on Information Forensics and Security","ref_id":"b1","title":"On Low-Resolution Face Recognition in the Wild: Comparisons and New Techniques","year":"2019"},{"authors":"U Sara; M Akter; M S Uddin","journal":"Journal of Computer and Communications","ref_id":"b2","title":"Image Quality Assessment Through FSIM, SSIM, MSE and PSNR-a Comparative Study","year":"2019"},{"authors":"W Zhang; K Ma; J Yan; D Deng; Z Wang","journal":"IEEE Transactions on Circuits and Systems for Video Technology","ref_id":"b3","title":"Blind Image Quality Assessment Using a Deep Bilinear Convolutional Neural Network","year":"2018"},{"authors":"W Zhang; K Ma; G Zhai; X Yang","journal":"IEEE","ref_id":"b4","title":"Learning to Blindly Assess Image Quality in the Laboratory and Wild","year":"2020"},{"authors":"","journal":"International Organization for Standardization","ref_id":"b5","title":"Biometrics","year":"2016"},{"authors":"T Schlett; C Rathgeb; O Henniger; J Galbally; J Fierrez; C Busch","journal":"ACM Computing Surveys","ref_id":"b6","title":"Face Image Quality Assessment: A Literature Survey","year":"2022"},{"authors":"J Hernandez-Ortega; J Galbally; J Fierrez; R Haraksim; L Beslay","journal":"","ref_id":"b7","title":"FaceQnet: Quality Assessment for Face Recognition Based on Deep Learning","year":"2019"},{"authors":"O Fu-Zhao; X Chen; R Zhang; Y Huang; S Li; J Li; Y Li; L Cao; W Yuan-Gen","journal":"","ref_id":"b8","title":"SDD-FIQA: Unsupervised Face Image Quality Assessment with Similarity Distribution Distance","year":"2021"},{"authors":"K Chen; T Yi; Q Lv","journal":"IEEE Signal Processing Letters","ref_id":"b9","title":"LightQNet: Lightweight Deep Face Quality Assessment for Risk-Controlled Face Recognition","year":"2021"},{"authors":"W Xie; J Byrne; A Zisserman","journal":"","ref_id":"b10","title":"Inducing Predictive Uncertainty Estimation for Face Verification","year":"2020"},{"authors":"P Terhörst; J N Kolf; N Damer; F Kirchbuchner; A Kuijper","journal":"","ref_id":"b11","title":"SER-FIQ: Unsupervised Estimation of Face Image Quality Based on Stochastic Embedding Robustness","year":"2020"},{"authors":"Ž Babnik; P Peer; V Štruc","journal":"","ref_id":"b12","title":"FaceQAN: Face Image Quality Assessment Through Adversarial Noise Exploration","year":"2022"},{"authors":"Y Shi; A K Jain","journal":"","ref_id":"b13","title":"Probabilistic Face Embeddings","year":"2019"},{"authors":"Q Meng; S Zhao; Z Huang; F Zhou","journal":"","ref_id":"b14","title":"Magface: A Universal Representation for Face Recognition and Quality Assessment","year":"2021"},{"authors":"J G Cavazos; P J Phillips; C D Castillo; A J O'toole","journal":"IEEE Transactions on Biometrics, Behavior, and Identity Science","ref_id":"b15","title":"Accuracy Comparison Across Face Recognition Algorithms: Where are we on Measuring Race Bias?","year":"2020"},{"authors":"J P Robinson; G Livitz; Y Henon; C Qin; Y Fu; S Timoner","journal":"","ref_id":"b16","title":"Face Recognition: Too Bias, or Not Too Bias?","year":"2020"},{"authors":"X Gao; S Z Li; R Liu; P Zhang","journal":"","ref_id":"b17","title":"Standardization of Face Image Sample Quality","year":"2007"},{"authors":"L Zhang; L Zhang; L Li","journal":"Springer","ref_id":"b18","title":"Illumination Quality Assessment for Face Images: A Benchmark and a Convolutional Neural Networks Based Model","year":"2017"},{"authors":"Z Lijun; S Xiaohu; Y Fei; D Pingling; Z Xiangdong; S Yu","journal":"","ref_id":"b19","title":"Multi-Branch Face Quality Assessment for Face Recognition","year":"2019"},{"authors":"Q Cao; L Shen; W Xie; O M Parkhi; A Zisserman","journal":"","ref_id":"b20","title":"VGGFace2: A Dataset for Recognising Faces Across Pose and Age","year":"2018"},{"authors":"F Boutros; M Fang; M Klemt; B Fu; N Damer","journal":"","ref_id":"b21","title":"Cr-fiqa: face image quality assessment by learning sample relative classifiability","year":"2021"},{"authors":"M Knoche; S Hoermann; G Rigoll","journal":"","ref_id":"b22","title":"Cross-Quality LFW: A Database for Analyzing Cross-Resolution Image Face Recognition in Unconstrained Environments","year":"2021"},{"authors":"T Zheng; W Deng","journal":"","ref_id":"b23","title":"Cross-Pose LFW: A Database for Studying Cross-Pose Face Recognition in Unconstrained Environments","year":"2018"},{"authors":"S Sengupta; J C Cheng; C D Castillo; V M Patel; R Chellappa; D W Jacobs","journal":"","ref_id":"b24","title":"Frontal to Profile Face Verification in the Wild","year":"2016"},{"authors":"T Zheng; W Deng; J Hu","journal":"CoRR","ref_id":"b25","title":"Cross-Age LFW: A Database for Studying Cross-Age Face Recognition in Unconstrained Environments","year":"2017"},{"authors":"G B Huang; M Ramesh; T Berg; E Learned-Miller","journal":"","ref_id":"b26","title":"Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments","year":"2007-10"},{"authors":"J Deng; J Guo; N Xue; S Zafeiriou","journal":"","ref_id":"b27","title":"Arcface: Additive Angular Margin Loss for Deep Face Recognition","year":"2019"},{"authors":"Y Huang; Y Wang; Y Tai; X Liu; P Shen; S Li; J Li; F Huang","journal":"","ref_id":"b28","title":"CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition","year":"2020"},{"authors":"F Boutros; N Damer; F Kirchbuchner; A Kuijper","journal":"","ref_id":"b29","title":"ElasticFace: Elastic Margin Loss for Deep Face Recognition","year":"2022"}],"string":"[\n {\n \"authors\": \"M Wang; W Deng\",\n \"journal\": \"Neurocomputing\",\n \"ref_id\": \"b0\",\n \"title\": \"Deep Face Recognition: A Survey\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"P Li; L Prieto; D Mery; P J Flynn\",\n \"journal\": \"IEEE Transactions on Information Forensics and Security\",\n \"ref_id\": \"b1\",\n \"title\": \"On Low-Resolution Face Recognition in the Wild: Comparisons and New Techniques\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"U Sara; M Akter; M S Uddin\",\n \"journal\": \"Journal of Computer and Communications\",\n \"ref_id\": \"b2\",\n \"title\": \"Image Quality Assessment Through FSIM, SSIM, MSE and PSNR-a Comparative Study\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"W Zhang; K Ma; J Yan; D Deng; Z Wang\",\n \"journal\": \"IEEE Transactions on Circuits and Systems for Video Technology\",\n \"ref_id\": \"b3\",\n \"title\": \"Blind Image Quality Assessment Using a Deep Bilinear Convolutional Neural Network\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"W Zhang; K Ma; G Zhai; X Yang\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b4\",\n \"title\": \"Learning to Blindly Assess Image Quality in the Laboratory and Wild\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"International Organization for Standardization\",\n \"ref_id\": \"b5\",\n \"title\": \"Biometrics\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"T Schlett; C Rathgeb; O Henniger; J Galbally; J Fierrez; C Busch\",\n \"journal\": \"ACM Computing Surveys\",\n \"ref_id\": \"b6\",\n \"title\": \"Face Image Quality Assessment: A Literature Survey\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"J Hernandez-Ortega; J Galbally; J Fierrez; R Haraksim; L Beslay\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"FaceQnet: Quality Assessment for Face Recognition Based on Deep Learning\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"O Fu-Zhao; X Chen; R Zhang; Y Huang; S Li; J Li; Y Li; L Cao; W Yuan-Gen\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"SDD-FIQA: Unsupervised Face Image Quality Assessment with Similarity Distribution Distance\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"K Chen; T Yi; Q Lv\",\n \"journal\": \"IEEE Signal Processing Letters\",\n \"ref_id\": \"b9\",\n \"title\": \"LightQNet: Lightweight Deep Face Quality Assessment for Risk-Controlled Face Recognition\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"W Xie; J Byrne; A Zisserman\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Inducing Predictive Uncertainty Estimation for Face Verification\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"P Terhörst; J N Kolf; N Damer; F Kirchbuchner; A Kuijper\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"SER-FIQ: Unsupervised Estimation of Face Image Quality Based on Stochastic Embedding Robustness\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ž Babnik; P Peer; V Štruc\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"FaceQAN: Face Image Quality Assessment Through Adversarial Noise Exploration\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Y Shi; A K Jain\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Probabilistic Face Embeddings\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Q Meng; S Zhao; Z Huang; F Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Magface: A Universal Representation for Face Recognition and Quality Assessment\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"J G Cavazos; P J Phillips; C D Castillo; A J O'toole\",\n \"journal\": \"IEEE Transactions on Biometrics, Behavior, and Identity Science\",\n \"ref_id\": \"b15\",\n \"title\": \"Accuracy Comparison Across Face Recognition Algorithms: Where are we on Measuring Race Bias?\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"J P Robinson; G Livitz; Y Henon; C Qin; Y Fu; S Timoner\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Face Recognition: Too Bias, or Not Too Bias?\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"X Gao; S Z Li; R Liu; P Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Standardization of Face Image Sample Quality\",\n \"year\": \"2007\"\n },\n {\n \"authors\": \"L Zhang; L Zhang; L Li\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b18\",\n \"title\": \"Illumination Quality Assessment for Face Images: A Benchmark and a Convolutional Neural Networks Based Model\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Z Lijun; S Xiaohu; Y Fei; D Pingling; Z Xiangdong; S Yu\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Multi-Branch Face Quality Assessment for Face Recognition\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Q Cao; L Shen; W Xie; O M Parkhi; A Zisserman\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"VGGFace2: A Dataset for Recognising Faces Across Pose and Age\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"F Boutros; M Fang; M Klemt; B Fu; N Damer\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Cr-fiqa: face image quality assessment by learning sample relative classifiability\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"M Knoche; S Hoermann; G Rigoll\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Cross-Quality LFW: A Database for Analyzing Cross-Resolution Image Face Recognition in Unconstrained Environments\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"T Zheng; W Deng\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Cross-Pose LFW: A Database for Studying Cross-Pose Face Recognition in Unconstrained Environments\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"S Sengupta; J C Cheng; C D Castillo; V M Patel; R Chellappa; D W Jacobs\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Frontal to Profile Face Verification in the Wild\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"T Zheng; W Deng; J Hu\",\n \"journal\": \"CoRR\",\n \"ref_id\": \"b25\",\n \"title\": \"Cross-Age LFW: A Database for Studying Cross-Age Face Recognition in Unconstrained Environments\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"G B Huang; M Ramesh; T Berg; E Learned-Miller\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments\",\n \"year\": \"2007-10\"\n },\n {\n \"authors\": \"J Deng; J Guo; N Xue; S Zafeiriou\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Arcface: Additive Angular Margin Loss for Deep Face Recognition\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Y Huang; Y Wang; Y Tai; X Liu; P Shen; S Li; J Li; F Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"F Boutros; N Damer; F Kirchbuchner; A Kuijper\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"ElasticFace: Elastic Margin Loss for Deep Face Recognition\",\n \"year\": \"2022\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[2,311.98,617,71.2,10.31],"formula_id":"formula_0","formula_text":"Q * = H(M (I))"},{"formula_coordinates":[3,169.29,490.68,119.19,13.38],"formula_id":"formula_1","formula_text":"E k = {e k i } N k i=1 corresponding"},{"formula_coordinates":[3,53.28,571.06,188.67,12.47],"formula_id":"formula_2","formula_text":"′ ̸ = c, c ′ ∈ [1, C] to form mated pairs (I k c , I k c ′ )"},{"formula_coordinates":[3,370.12,298.52,192.91,23.83],"formula_id":"formula_3","formula_text":"sim cos (e Ii , e Ij ) = e Ii • e Ij ∥e Ii ∥ • ∥e Ij ∥ ,(1)"},{"formula_coordinates":[3,382.13,388.52,180.91,26.8],"formula_id":"formula_4","formula_text":"id s Ii = 1 |I| I id Xs (I i , I j ),(2)"},{"formula_coordinates":[3,313.17,625.47,249.87,27.07],"formula_id":"formula_6","formula_text":"* Ii ) = 1 R R r=1 id"}],"string":"[\n {\n \"formula_coordinates\": [\n 2,\n 311.98,\n 617,\n 71.2,\n 10.31\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"Q * = H(M (I))\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 169.29,\n 490.68,\n 119.19,\n 13.38\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"E k = {e k i } N k i=1 corresponding\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 53.28,\n 571.06,\n 188.67,\n 12.47\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"′ ̸ = c, c ′ ∈ [1, C] to form mated pairs (I k c , I k c ′ )\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 370.12,\n 298.52,\n 192.91,\n 23.83\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"sim cos (e Ii , e Ij ) = e Ii • e Ij ∥e Ii ∥ • ∥e Ij ∥ ,(1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 382.13,\n 388.52,\n 180.91,\n 26.8\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"id s Ii = 1 |I| I id Xs (I i , I j ),(2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 313.17,\n 625.47,\n 249.87,\n 27.07\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"* Ii ) = 1 R R r=1 id\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"Optimization-Based Improvement of Face Image Quality Assessment Techniques"},"abstract":{"kind":"string","value":"Contemporary face recognition (FR) models achieve near-ideal recognition performance in constrained settings, yet do not fully translate the performance to unconstrained (realworld) scenarios. To help improve the performance and stability of FR systems in such unconstrained settings, face image quality assessment (FIQA) techniques try to infer sample-quality information from the input face images that can aid with the recognition process. While existing FIQA techniques are able to efficiently capture the differences between high and low quality images, they typically cannot fully distinguish between images of similar quality, leading to lower performance in many scenarios. To address this issue, we present in this paper a supervised quality-label optimization approach, aimed at improving the performance of existing FIQA techniques. The developed optimization procedure infuses additional information (computed with a selected FR model) into the initial quality scores generated with a given FIQA technique to produce better estimates of the \"actual\" image quality. We evaluate the proposed approach in comprehensive experiments with six state-of-the-art FIQA approaches (CR-FIQA, FaceQAN, SER-FIQ, PCNet, MagFace, SDD-FIQA) on five commonly used benchmarks (LFW, CFP-FP, CPLFW, CALFW, XQLFW) using three targeted FR models (ArcFace, ElasticFace, CurricularFace) with highly encouraging results."},"authors":{"kind":"string","value":"Žiga Babnik; Naser Damer; Vitomir Štruc"},"figures":{"kind":"list like","value":[{"figure_caption":"Fig. 1 .1Fig.1. Overview of the proposed method that consists of: Label Optimization and Transfer Learning. The label-optimization step incorporates information extracted from mated image pairs into quality scores precomputed with an existing FIQA technique. The transfer-learning step is then used to train a FR model, extended with a regression head, on the optimized quality-scores. The learned regressor is finally utilized for quality estimation.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Fig. 2 .2Fig. 2. Overview of Label Optimization.We present a visualization of the proposed optimization scheme. Based on the embeddings {e I i } N k","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"to images of each individual k = 1, ..., K present in the database using K-Means, where N = k N k . The algorithm initializes C cluster centers by randomly sampling the given data points and iteratively corrects them using nearby examples. For each image I k c of the k-th individual belonging to cluster c ∈ [1, C], we randomly select images from all other clusters c","figure_data":"","figure_id":"fig_2","figure_label":"","figure_type":"figure"},{"figure_caption":"Fig. 3 .3Fig. 3. Example VGGFace2 images. Images of three distinct individuals are shown, illustrating the amount of variability present in the database.","figure_data":"","figure_id":"fig_3","figure_label":"3","figure_type":"figure"},{"figure_caption":"Fig. 4 .4Fig.4. Qualitative analysis of the proposed approach. For each FIQA method, we show the prior distribution of the quality scores of the VGGFace2 database, and an associated scatter plot showing the changes in the quality scores due to our optimization approach.","figure_data":"","figure_id":"fig_4","figure_label":"4","figure_type":"figure"},{"figure_caption":"SAME-MODEL PERFORMANCE COMPARISON. COMPARISON OF AUC SCORES BETWEEN BASELINE FIQA METHODS (Baseline) AND OUR PROPOSED OPTIMIZATION APPROACH (Optimized) USING ARCFACE FOR QUALITY ESTIMATION AND PERFORMANCE EVALUATION. BEST RESULTS ARE HIGHLIGHTED IN GREEN.","figure_data":"MethodsAUC@FMR1e-3[×10 -3 ](↓) LFW CFP-FP CPLFW CALFW XQLFWCR-FIQABaseline Optimized1.7 1.81.3 1.386.5 109.373.3 73.7115.7 105.7FaceQANBaseline Optimized1.5 1.21.9 1.8112.6 82.872.1 73.9134.6 98.5SER-FIQBaseline Optimized2.9 2.42.7 2.0100.0 102.273.0 66.6/PCNetBaseline Optimized1.7 0.92.3 1.788.5 87.659.3 60.2139.7 116.4MagFaceBaseline Optimized2.0 1.73.7 1.7118.6 125.374.4 72.3294.9 127.6SDD-FIQABaseline Optimized1.8 1.82.6 1.791.4 80.687.5 66.6190.1 135.3measure the False Non-Match Rate (FNMR) at a predefinedFalse Match Rate (FMR), typically fixed at 0.001, at variouslow-quality image drop (also unconsidered) rates. Specifically,we report the Area Under the ERC Curves (AUC) as our mainperformance indicator, where smaller values indicate betterperformance.","figure_id":"tab_0","figure_label":"I","figure_type":"table"},{"figure_caption":"Table II again shows the AUC values of both the baseline and our (optimized) regressionbased FIQA techniques, but this time computed for the crossmodel scenario, where the FR model used for estimating the quality of the input images differs from the FR model used for performance reporting. Looking at the individual","figure_data":"","figure_id":"tab_1","figure_label":"","figure_type":"table"},{"figure_caption":"CROSS-MODEL PERFORMANCE COMPARISON. COMPARISON OF AUC SCORES (AUC@FMR1E-3[×10 -3 ](↓)) BETWEEN THE BASELINE FIQA METHODS (Baseline) AND THE PROPOSED OPTIMIZATION APPROACH (Optimized), USING THE ARCFACE MODEL FOR QUALITY ESTIMATION AND ELASTICFACE AND CURRICULARFACE FOR PERFORMANCE SCORING. BEST RESULTS ARE HIGHLIGHTED IN GREEN .","figure_data":"MethodsElasticFace LFW CFP-FP CPLFW CALFW XQLFWCurricularFace LFW CFP-FP CPLFW CALFW XQLFWCR-FIQABaseline Optimized2.1 2.32.1 2.062.0 62.774.9 76.1315.8 305.51.8 2.01.9 1.957.3 56.176.6 78.2281.0 272.2FaceQANBaseline Optimized1.9 2.02.6 2.766.4 65.077.7 79.7329.6 287.11.6 1.52.2 2.370.7 63.782.6 83.7323.3 263.5SER-FIQBaseline Optimized3.8 3.23.9 3.079.4 77.574.9 69.7/3.2 2.53.4 2.569.6 69.278.1 71.8/PCNetBaseline Optimized3.2 2.03.5 2.574.4 69.562.0 62.7312.2 295.22.1 1.33.0 2.366.5 62.064.3 64.7300.3 272.6MagFaceBaseline Optimized2.6 2.14.9 2.579.5 68.674.9 73.9601.8 306.62.1 2.04.4 2.281.5 65.179.9 77.6593.2 260.4SDD-FIQABaseline Optimized3.1 2.53.8 2.573.1 63.479.9 68.7480.8 292.82.1 2.03.3 2.164.8 55.777.4 70.4438.6 268.3","figure_id":"tab_2","figure_label":"II","figure_type":"table"},{"figure_caption":"OF THE ABLATION STUDY. SHOWN ARE AUC SCORES (AUC@FMR1E-3[×10 -3 ](↓)) GENERATED BY THE REGRESSION-BASED FIQA METHODS TRAINED WITH THE INITIAL/PRIOR QUALITY LABELS AND THE PERFORMANCE GAIN(-)/LOSS(+) IN TERMS OF AUC CHANGE DUE TO THE LABEL OPTIMIZATION PROCEDURE (IN BRACKETS). RESULTS ARE PRESENTED FOR THE SAME-MODEL (ARCFACE) AND CROSS-MODEL (ELASTICFACE AND CURRICULARFACE) SETTINGS. GAINS ARE MARKED GREEN, LOSSES RED.","figure_data":"MethodsLFWArcFace CPLFWXQLFWLFWElasticFace CPLFWXQLFWLFWCurricularFace CPLFWXQLFWCR-FIQA1.3 (+0.5)84.0 (+1.7)102.8 (+2.9)2.1 (+0.2) 65.0 (-2.3)280.4 (+24.9)1.6 (+0.4) 64.2 (-4.1)273.3 (-1.1)FaceQAN1.4 (-0.2)85.7 (-2.9)109.8 (-11.3)2.1 (-0.1) 65.5 (-0.5)298.2 (-11.1)1.7 (-0.2) 65.2 (-1.5)273.0 (-9.5)SER-FIQ2.2 (+0.2)101.9 (+0.3)/3.2 (-0.0) 79.5 (-2.0)/2.4 (+0.1) 72.8 (-3.6)/PCNet0.9 (-0.0)93.8 (-6.2)211.3 (-94.9)2.5 (-0.5) 78.6 (-9.1)406.6 (-111.4)1.2 (+0.1) 69.2 (-7.2)360.1 (-87.5)MagFace1.8 (-0.1)117.2 (+8.1) 195.5 (-67.9)2.0 (+0.1) 69.8 (-1.2)492.3 (-185.7)2.1 (-0.1) 66.1 (-1.0)501.8 (-241.4)SDD-FIQA2.2 (-0.4)80.9 (-0.3)164.8 (-29.5)2.9 (-0.4) 64.3 (-0.9)393.8 (-101.0)2.4 (-0.4) 57.5 (-1.8)337.1 (-68.8)","figure_id":"tab_3","figure_label":"III","figure_type":"table"},{"figure_caption":"TABLE IV RUN-TIME PERFORMANCE. SHOWN IN [µS].± 5.0 346.5 ± 9.0 125.2 ± 19.5 14.5 ± 1.7 11.5 ± 4.9 5.7 ± 5.0 Ours 11.4 ± 3.0 11.3 ± 5.0 11.3 ± 5.0 10.6 ± 4.9 11.3 ± 5.0 11.3 ± 5.0","figure_data":"MethodCR-FIQAFaceQANSER-FIQPCNetMagFaceSDD-FIQAOriginal 11.5 Speed-up † +0.8%+3000%+1100%+36%+1.7%-50%","figure_id":"tab_4","figure_label":"","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Fig. 1 .1Fig.1. Overview of the proposed method that consists of: Label Optimization and Transfer Learning. The label-optimization step incorporates information extracted from mated image pairs into quality scores precomputed with an existing FIQA technique. The transfer-learning step is then used to train a FR model, extended with a regression head, on the optimized quality-scores. The learned regressor is finally utilized for quality estimation.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 2 .2Fig. 2. Overview of Label Optimization.We present a visualization of the proposed optimization scheme. Based on the embeddings {e I i } N k\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"to images of each individual k = 1, ..., K present in the database using K-Means, where N = k N k . The algorithm initializes C cluster centers by randomly sampling the given data points and iteratively corrects them using nearby examples. For each image I k c of the k-th individual belonging to cluster c ∈ [1, C], we randomly select images from all other clusters c\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 3 .3Fig. 3. Example VGGFace2 images. Images of three distinct individuals are shown, illustrating the amount of variability present in the database.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 4 .4Fig.4. Qualitative analysis of the proposed approach. For each FIQA method, we show the prior distribution of the quality scores of the VGGFace2 database, and an associated scatter plot showing the changes in the quality scores due to our optimization approach.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"SAME-MODEL PERFORMANCE COMPARISON. COMPARISON OF AUC SCORES BETWEEN BASELINE FIQA METHODS (Baseline) AND OUR PROPOSED OPTIMIZATION APPROACH (Optimized) USING ARCFACE FOR QUALITY ESTIMATION AND PERFORMANCE EVALUATION. BEST RESULTS ARE HIGHLIGHTED IN GREEN.\",\n \"figure_data\": \"MethodsAUC@FMR1e-3[×10 -3 ](↓) LFW CFP-FP CPLFW CALFW XQLFWCR-FIQABaseline Optimized1.7 1.81.3 1.386.5 109.373.3 73.7115.7 105.7FaceQANBaseline Optimized1.5 1.21.9 1.8112.6 82.872.1 73.9134.6 98.5SER-FIQBaseline Optimized2.9 2.42.7 2.0100.0 102.273.0 66.6/PCNetBaseline Optimized1.7 0.92.3 1.788.5 87.659.3 60.2139.7 116.4MagFaceBaseline Optimized2.0 1.73.7 1.7118.6 125.374.4 72.3294.9 127.6SDD-FIQABaseline Optimized1.8 1.82.6 1.791.4 80.687.5 66.6190.1 135.3measure the False Non-Match Rate (FNMR) at a predefinedFalse Match Rate (FMR), typically fixed at 0.001, at variouslow-quality image drop (also unconsidered) rates. Specifically,we report the Area Under the ERC Curves (AUC) as our mainperformance indicator, where smaller values indicate betterperformance.\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"I\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Table II again shows the AUC values of both the baseline and our (optimized) regressionbased FIQA techniques, but this time computed for the crossmodel scenario, where the FR model used for estimating the quality of the input images differs from the FR model used for performance reporting. Looking at the individual\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"CROSS-MODEL PERFORMANCE COMPARISON. COMPARISON OF AUC SCORES (AUC@FMR1E-3[×10 -3 ](↓)) BETWEEN THE BASELINE FIQA METHODS (Baseline) AND THE PROPOSED OPTIMIZATION APPROACH (Optimized), USING THE ARCFACE MODEL FOR QUALITY ESTIMATION AND ELASTICFACE AND CURRICULARFACE FOR PERFORMANCE SCORING. BEST RESULTS ARE HIGHLIGHTED IN GREEN .\",\n \"figure_data\": \"MethodsElasticFace LFW CFP-FP CPLFW CALFW XQLFWCurricularFace LFW CFP-FP CPLFW CALFW XQLFWCR-FIQABaseline Optimized2.1 2.32.1 2.062.0 62.774.9 76.1315.8 305.51.8 2.01.9 1.957.3 56.176.6 78.2281.0 272.2FaceQANBaseline Optimized1.9 2.02.6 2.766.4 65.077.7 79.7329.6 287.11.6 1.52.2 2.370.7 63.782.6 83.7323.3 263.5SER-FIQBaseline Optimized3.8 3.23.9 3.079.4 77.574.9 69.7/3.2 2.53.4 2.569.6 69.278.1 71.8/PCNetBaseline Optimized3.2 2.03.5 2.574.4 69.562.0 62.7312.2 295.22.1 1.33.0 2.366.5 62.064.3 64.7300.3 272.6MagFaceBaseline Optimized2.6 2.14.9 2.579.5 68.674.9 73.9601.8 306.62.1 2.04.4 2.281.5 65.179.9 77.6593.2 260.4SDD-FIQABaseline Optimized3.1 2.53.8 2.573.1 63.479.9 68.7480.8 292.82.1 2.03.3 2.164.8 55.777.4 70.4438.6 268.3\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"II\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"OF THE ABLATION STUDY. SHOWN ARE AUC SCORES (AUC@FMR1E-3[×10 -3 ](↓)) GENERATED BY THE REGRESSION-BASED FIQA METHODS TRAINED WITH THE INITIAL/PRIOR QUALITY LABELS AND THE PERFORMANCE GAIN(-)/LOSS(+) IN TERMS OF AUC CHANGE DUE TO THE LABEL OPTIMIZATION PROCEDURE (IN BRACKETS). RESULTS ARE PRESENTED FOR THE SAME-MODEL (ARCFACE) AND CROSS-MODEL (ELASTICFACE AND CURRICULARFACE) SETTINGS. GAINS ARE MARKED GREEN, LOSSES RED.\",\n \"figure_data\": \"MethodsLFWArcFace CPLFWXQLFWLFWElasticFace CPLFWXQLFWLFWCurricularFace CPLFWXQLFWCR-FIQA1.3 (+0.5)84.0 (+1.7)102.8 (+2.9)2.1 (+0.2) 65.0 (-2.3)280.4 (+24.9)1.6 (+0.4) 64.2 (-4.1)273.3 (-1.1)FaceQAN1.4 (-0.2)85.7 (-2.9)109.8 (-11.3)2.1 (-0.1) 65.5 (-0.5)298.2 (-11.1)1.7 (-0.2) 65.2 (-1.5)273.0 (-9.5)SER-FIQ2.2 (+0.2)101.9 (+0.3)/3.2 (-0.0) 79.5 (-2.0)/2.4 (+0.1) 72.8 (-3.6)/PCNet0.9 (-0.0)93.8 (-6.2)211.3 (-94.9)2.5 (-0.5) 78.6 (-9.1)406.6 (-111.4)1.2 (+0.1) 69.2 (-7.2)360.1 (-87.5)MagFace1.8 (-0.1)117.2 (+8.1) 195.5 (-67.9)2.0 (+0.1) 69.8 (-1.2)492.3 (-185.7)2.1 (-0.1) 66.1 (-1.0)501.8 (-241.4)SDD-FIQA2.2 (-0.4)80.9 (-0.3)164.8 (-29.5)2.9 (-0.4) 64.3 (-0.9)393.8 (-101.0)2.4 (-0.4) 57.5 (-1.8)337.1 (-68.8)\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"III\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"TABLE IV RUN-TIME PERFORMANCE. SHOWN IN [µS].± 5.0 346.5 ± 9.0 125.2 ± 19.5 14.5 ± 1.7 11.5 ± 4.9 5.7 ± 5.0 Ours 11.4 ± 3.0 11.3 ± 5.0 11.3 ± 5.0 10.6 ± 4.9 11.3 ± 5.0 11.3 ± 5.0\",\n \"figure_data\": \"MethodCR-FIQAFaceQANSER-FIQPCNetMagFaceSDD-FIQAOriginal 11.5 Speed-up † +0.8%+3000%+1100%+36%+1.7%-50%\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Supporting Evidence\", \"Citation\": \"[1], [2]\", \"Explanation\": \"The cited works highlight the challenges of out-of-distribution data in real-world scenarios, which motivates the need for face image quality assessment techniques in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[3]- [5]\", \"Explanation\": \"The cited works on image quality assessment provide a basis for the design of face image quality assessment techniques in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work on face image quality assessment techniques extends the research in the citing paper by focusing on the utility of samples for face recognition.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[7]\", \"Explanation\": \"The cited works in the last group provide a summary of the different approaches to face image quality assessment techniques, which the citing paper builds upon in its own research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[8]- [11]\", \"Explanation\": \"The cited works provide a basis for learning quality-estimation networks in the citing paper, which is a method used to infer quality information from a large database of face images.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[12], [13]\", \"Explanation\": \"The cited works are a notable group of techniques that estimate quality based on the information present in the input image and the characteristics of the targeted FR system, providing a methodological basis for the research in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[14], [15]\", \"Explanation\": \"The cited works have also appeared that incorporate quality estimation directly into the FR process, which the citing paper extends by exploring the possibility of quality-aware face recognition.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[16], [17]\", \"Explanation\": \"The cited works suggest that the correct ordering of face images for FR models depends on the data used for training, which the citing paper acknowledges as a data source for the research conducted.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[7]\", \"Explanation\": \"The cited work by Schlett et al. provides a comprehensive survey on face quality assessment techniques, which serves as a methodological basis for the proposed optimization approach in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[12]\", \"Explanation\": \"The method proposed by Terh\\u00f6rst et al. in SER-FIQ is used as a basis for the quality estimation task in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[13]\", \"Explanation\": \"The method proposed by Babnik et al. in FaceQAN is also used as a basis for the quality assessment in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[8]\", \"Explanation\": \"The FaceQNet model is referenced as a data source for the quality labels used in the quality estimation task in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[11]\", \"Explanation\": \"The cited work introduces the concept of quality-regression networks, which the citing paper builds upon in the development of the PCNet model for image quality prediction.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[10]\", \"Explanation\": \"The cited work introduces the LightQNet model, which the citing paper builds upon by incorporating the Identification Quality (IQ) loss in the development of the SDD-FIQA model for image quality assessment.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[9]\", \"Explanation\": \"The cited work presents the SDD-FIQA model, which the citing paper builds upon by considering both mated and nonmated similarity scores in the development of the final reference quality for the regression task.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work by Shi and Jain introduces the idea of estimating mean and variance vectors for face recognition and quality assessment, which the citing paper builds upon in the development of a model that produces embeddings and quality estimates for input samples.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[15]\", \"Explanation\": \"The cited work by Meng et al. introduces the MagFace loss function, which is a modified version of the ArcFace loss that incorporates quality information into the magnitude of the embedding. The citing paper uses this method to generate quality-aware embeddings, providing a foundational technique for the research conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[8]- [11]\", \"Explanation\": \"The cited works provide the insight of using mated image comparisons to infer quality information, which the citing paper adopts in their optimization approach.\"}, {\"Category\": \"Data Source\", \"Citation\": \"Large-scale databases\", \"Explanation\": \"The large-scale databases used in the study are a crucial data source for the sampling of mated image pairs in the optimization approach.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"Clustering procedure\", \"Explanation\": \"The citing paper extends the research by proposing a technique for sampling mated image pairs based on clustering, which builds upon the clustering procedure used in the cited works.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[21]\", \"Explanation\": \"The cited work, VGGFace2 database, is used as the training data for the proposed approach in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[22]\", \"Explanation\": \"The cited work, CR-FIQA, is a baseline method used in the evaluation of the proposed optimization scheme in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work, FaceQAN, is a baseline method used in the evaluation of the proposed optimization scheme in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[15]\", \"Explanation\": \"The cited work, MagFace, is a baseline method used in the evaluation of the proposed optimization scheme in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[12]\", \"Explanation\": \"The cited work, SER-FIQ, is a baseline method used in the evaluation of the proposed optimization scheme in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[9]\", \"Explanation\": \"The cited work, SDD-FIQA, is a baseline method used in the evaluation of the proposed optimization scheme in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[11]\", \"Explanation\": \"The cited work, PCNet, is a baseline method used in the evaluation of the proposed optimization scheme in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[23]\", \"Explanation\": \"The cited work, XQLFW database, is a benchmark database used in the evaluation of the baselines in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[24]\", \"Explanation\": \"The cited work, CPLFW database, is a benchmark database used in the evaluation of the baselines in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[25]\", \"Explanation\": \"The cited work, CFP-FP database, is a benchmark database used in the evaluation of the baselines in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[26]\", \"Explanation\": \"The cited work, CALFW database, is a benchmark database used in the evaluation of the baselines in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[27]\", \"Explanation\": \"The cited work, LFW database, is a benchmark database used in the evaluation of the baselines in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[28]\", \"Explanation\": \"The cited work, ArcFace, is used as the pre-trained model for quality-score prediction in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[29]\", \"Explanation\": \"The CurricularFace model is used in the cross-model scenario to evaluate performance in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[30]\", \"Explanation\": \"The ElasticFace-Cos+ model is also used in the cross-model scenario to evaluate performance in the citing paper.\"}]"}}},{"rowIdx":84,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_0","fig_0","fig_0"],"heading":"Introduction","publication_ref":["b12","b34","b24","b11","b54","b42","b42","b43"],"table_ref":[],"text":"Recent advances in NLP primarily focus on the English language (Blasi et al., 2022). Due to the lack of sufficient training data in most of the world's languages (Yu et al., 2022), prior work explores direct transfer of pretrained language models to new languages after fine-tuning on resource-rich languages (zero-shot cross-lingual transfer; Hu et al. 2020b). Transferring after training a model on a few examples (few-shot cross-lingual transfer) often boosts performance, especially in languages that are distant from the source language (Lauscher et al., 2020;Hedderich et al., 2020).\nIn English, zero-or few-shot learning via incontext learning is an active area of research (Beltagy et al., 2022;Schick and Schütze, 2021a; Shin et al., 2020). In this learning paradigm, one prompts a large language model (LLM) with fewshot demonstrations or natural language instructions to adapt to a new task, without any parameter updates. Yet, few-shot transfer across languages is still under-explored (Lin et al., 2021) in a wide range of tasks and languages. Moreover, it is unclear how effectively in-context learning performs in comparison to widely-used fine-tuning-based transfer methods under a comparable setup.\nThis work introduces a new benchmark called BUFFET: Benchmark of Unified Format FEwshot Transfer Evaluation (Figure 1) to enable rigorous evaluations and advance research on few-shot cross-lingual transfer. Similar to a rich buffet, BUF-FET curates a diverse mix of tasks: 15 different tasks-including classification, structured prediction, and natural language generation-across 54 languages. BUFFET has several unique characteristics that are not present in prior multi-task multilingual benchmarks (summarized in Table 1):\n• BUFFET provides a fixed set of few-shot examples for training and validation, allowing for fair comparisons across LMs and transfer methods. • BUFFET includes datasets annotated in each language or covering under-represented languages, which are often not included in existing multi-task benchmarks. • BUFFET combines diverse tasks into a unified text-to-text format and provides a set of English and machine-translated instructions for each task, removing the burdens of task-specific architecture changes or prompt engineering.\nUsing this new benchmark, we extensively evaluate the current state-of-the-art multilingual large language models (LLMs), including mT5 (Xue et al., 2021), mT0 (Muennighoff et al., 2022), BLOOM (Scao et al., 2022), BLOOMZ (Muennighoff et al., 2022), and ChatGPT (Ouyang et al., 2022), using both fine-tuning and in-context learning approaches. In particular, BUFFET enables us to investigate the following research questions: (RQ1) Is in-context learning competitive with fine-tuning in few-shot cross-lingual transfer? Notably, given the same small numbers of examples in the target languages, in-context learning on LLMs (including ChatGPT, the most powerful model we evaluate in this work) often underperforms much smaller specialized mT5-base models, as shown in Figure 1 (bottom left). (RQ2) How well do different transfer methods perform across tasks and languages? The performance gap between in-context learning-based baselines and fine-tuning-based baselines is more significant in under-represented languages (Figure 1 bottom center). On NLI in indigenous languages of the Americas, ChatGPT or mT0-11B using in-context learning performs barely above random, while 580 million parameter mt5-base fine-tuned models retain strong performance. On the contrary, these LLMs perform well on generative tasks where a smaller task-specific model struggles, demonstrating their superiority in generating fluent text for diverse languages without abundant training data. (RQ3) How does the choice of transfer setup affect different transfer strategies? BUFFET also enables us to perform an in-depth and extensive analysis of the effects of diverse demonstrations and instructions on the downstream transfer quality. Our observations indicate that the choice of few-shot training examples has a substantial influence on a model's performance, particularly, Multi-ling. Few-S Gen. Low-R\nXTREME ✓ XTREME-R ✓ XGLUE ✓ ✓ CrossFit ✓ ✓ MEGA* ✓ ✓ BUFFET ✓ ✓ ✓ ✓\nTable 1: Comparison of the existing benchmarks based on their multilinguality (Multi-ling.), few-shot task formulation (Few-S), availability of generative tasks (Gen.), and coverage of low-resource languages (Low-R). * indicates concurrent work.\nwith greater variability in in-context learning, compared to fine-tuning. We note that optimal transfer settings may differ across models. For example, instruction-tuned models often face challenges in effectively utilizing few-shot samples and their performance deteriorates as the number of demonstrations increases, possibly because they are optimized for the zero-shot instruction-tuned training scheme. This highlights the need for a standardized benchmark to facilitate fair comparisons and further studies to assess such transfer dynamics in non-English data. Grounded in our analysis, we suggest avenues for future research in few-shot cross-lingual transfer for both dataset creation and model development. Our data and code are available online. 12 Background and Related Work"},{"figure_ref":[],"heading":"Problem Formulation","publication_ref":["b12","b30","b34"],"table_ref":[],"text":"Due to the lack of annotated training data in many languages (Blasi et al., 2022;Yu et al., 2022;Joshi et al., 2020), transferring models trained on resource-rich languages (e.g., English) to other languages has been actively studied in multilingual NLP. In this paper, our main focus is on fewshot cross-lingual transfer (Lauscher et al., 2020), where a model is adapted using only a limited number of training or validation examples in the target language L. Another popular paradigm is zeroshot cross-lingual transfer (Artetxe et al., 2020a;Hu et al., 2020b) from English, where a model has access to training sets or instructions in English but not in the target language.\nVarious transfer methods have been investigated in the field, including the in-context learning methods (Section 2.3). Yet, limited research explores different transfer methods under comparable condi-tions. With our new benchmark, BUFFET, we facilitate fair comparisons between models and learning methods, establishing a basis for studying the dynamics of few-shot transfer across various languages (Section 2.2)."},{"figure_ref":[],"heading":"Benchmarks for Cross-lingual Transfer","publication_ref":["b49","b36","b56","b59","b3"],"table_ref":[],"text":"To enable a scalable and rigorous evaluation across multiple tasks, prior work has proposed multi-task benchmarks that unify diverse existing datasets. XTREME (Hu et al., 2020b), XTREME-R (Ruder et al., 2021) and XGLUE (Liang et al., 2020) focus on zero-shot transfer of models fine-tuned on English datasets. Despite English-based few-shot evaluation benchmarks, such as CrossFit (Ye et al., 2021), in few-shot cross-lingual transfer, we lack a standardized evaluation benchmark to facilitate the comparison of models and learning methods at scale. BUFFET provides the first large-scale few-shot cross-lingual transfer suits to address the gap. Importantly, to mitigate the effects of the highperformance variance in few-shot cross-lingual transfer (Zhao et al., 2021), we curate and aggregate results from multiple fixed k-shot training instances for each task and language. Concurrent with our work, MEGA (Ahuja et al., 2023) conducts experiments of few-shot cross-lingual transfer with a focus on classification and question answering tasks. BUFFET unifies diverse tasks including both discriminative and generative tasks. We also include datasets covering languages underrepresented in prior work (e.g., African and indigenous languages). Table 1 summarizes the key differences between BUFFET and prior benchmarks."},{"figure_ref":[],"heading":"Methods for Cross-lingual Transfer","publication_ref":["b20","b54","b52","b34","b24","b13","b42","b60","b10"],"table_ref":[],"text":"Fine-tuning-based approaches. Multilingual pre-trained models (Devlin et al., 2019;Xue et al., 2021;Conneau et al., 2020a) have the ability to adapt to new languages with no or few training instances in a target language (Conneau et al., 2020b;Hu et al., 2020b;Wu and Dredze, 2019). Lauscher et al. (2020) andHedderich et al. (2020) report that particularly in languages that are distant from the source language, further fine-tuning model on few-shot samples greatly improves performance.\nCross-lingual in-context learning. In-context learning (Brown et al., 2020) (Schick and Schütze, 2021b;Min et al., 2022b), most prior work focuses only on English. Recent work (Lin et al., 2021;Muennighoff et al., 2022) introduces pre-trained LMs trained on more multilingual pre-trained corpora or translated datasets and shows improved results. While prior evaluations often focus on classification or translation tasks (Zhu et al., 2023;Vilar et al., 2022), more recently Shi et al. (2023), evaluate the use of instructions, demonstrations, and rationales in different languages across multiple reasoning tasks. However, how much LLMs with respect to in-context learning compete with the aforementioned fine-tuned approaches in a comparable setup and at scale has yet to be investigated, as they often use a large number of training examples in target languages (Bang et al., 2023). We demonstrate even with a small number of training examples, fine-tuning methods are competitive with in-context learning for cross-lingual transfer."},{"figure_ref":[],"heading":"Benchmark: BUFFET","publication_ref":[],"table_ref":["tab_2"],"text":"We introduce a new standardized few-shot cross-lingual evaluation benchmark: BUFFET (Benchmark of Unified Format Few-shot Transfer Evaluation). BUFFET unifies diverse NLP tasks and provides fixed sets of few-shot samples per task to facilitate consistent comparisons (Table 2)."},{"figure_ref":[],"heading":"Design Principles","publication_ref":["b49","b36","b46","b15"],"table_ref":["tab_7","tab_2"],"text":"We create the BUFFET benchmark to establish a rigorous and equitable evaluation framework for few-shot cross-lingual transfer across a broad range of tasks and languages. We adhere to the following design principles with our benchmark.\nStandardized few-shot samples. BUFFET provides three different training and validation sets of k-shots (e.g., k = 32) per task for a nonclassification task, or per class for a classification task, for each language.\nTask diversity. Existing cross-lingual benchmarks often focus on classification or retrieval (Hu et al., 2020b;Ruder et al., 2021;Liang et al., 2020). BUFFET encompasses a broad range of task types, such as classification, generation, extraction, and structured prediction tasks. By converting all tasks into the same text-to-text format, we eliminate the need for task-specific model modifications or template conversions. 5.\nBeyond evaluations on translated data. Prior few-or zero-shot evaluations were often conducted on widely-used datasets translated from English (e.g., XNLI; Conneau et al. 2018, XCOPA;Ponti et al. 2020). Those datasets might exhibit undesired biases, such as translation artifacts or unnatural topic distributions (Clark et al., 2020;Artetxe et al., 2020b). We collect both translation-based datasets and datasets that are annotated directly in each language (Table 2, Data curation)."},{"figure_ref":[],"heading":"BUFFET Construction Process","publication_ref":["b56","b47"],"table_ref":[],"text":"Following Ye et al. (2021), we unify all tasks into the same text-to-text format, where a model is expected to directly generate the desired outputs given diverse inputs (Raffel et al., 2020). For each dataset in BUFFET, we unify instance representations of instruction, k-shot instances for training and validation. Each training instance consists of an input and output. Figure 2 shows an overview. Section 3.2.1 provides the outline of the unification, and Section 3.2.2 provides a task-specific process."},{"figure_ref":[],"heading":"Unification Process","publication_ref":["b50","b9","b41","b19"],"table_ref":[],"text":"Few-shot instance selection. By default, we use all of the languages included in the original datasets. Among multiple annotated instructions, we sample the first instruction for a similar task that suits our text-to-text scheme. For some tasks, we modify the original instruction to make labels consistent with the names used in BUFFET4 or to remove taskspecific dependencies in the input data field. See Appendix Table 6 for the full list of instructions.\nInstruction translation. Despite rapid progress of instruction-tuning in English LLMs (Wei et al., 2022;Sanh et al., 2022;Mishra et al., 2022;Wang et al., 2022b), cross-lingual setups still lag behind due to a lack of instructions in the target languages.\nPrior work often translates instructions for the target tasks (Lin et al., 2021;Shi et al., 2023). We provide translated instructions for 15 datasets in 54 target languages, translated by NLLB (Costa-jussà et al., 2022), and manually translate the instructions into five languages.5 "},{"figure_ref":[],"heading":"Tasks and Dataset Curation","publication_ref":["b17","b21","b33","b45","b55","b32","b1","b46","b42","b15","b44","b23","b53"],"table_ref":[],"text":"We first select eight popular NLP tasks and, for each task, we identify available datasets using a careful survey of multilingual datasets by Yu et al. (2022). Appendix Table 6 shows examples.\nNatural language inference. Natural Language Inference (NLI) involves determining the logical relationship (i.e., entailment, contradiction, neutral) between two text fragments, i.e., a premise and a hypothesis. In addition to the widely used XNLI (Conneau et al., 2018), we gather NLI datasets that are annotated in each language or designed to cover extremely under-represented languages: AMER-ICASNLI (Ebrahimi et al., 2022), PARSINLU-ENTAILMENT (Khashabi et al., 2021), KLUE-NLI (Park et al., 2021), and OCNLI (Hu et al., 2020a). We use 16 examples for each class.\nParaphrase detection. Paraphrase detection is the task of identifying whether two sentences have/do not have the same meaning (duplicate or not duplicated).\nWe adopt PAWS-X (Yang et al., 2019) and include 16 shots for each class as few-shot training and validation data.\nSentiment analysis. Binary sentiment analysis identifies whether a text (e.g., a product review from Amazon) expresses positive or negative sentiment towards a topic. We use the MULTILINGUAL AMAZON REVIEW DATASET (Keung et al., 2020) and INDICNLU-SENTIMENT (Aggarwal et al., 2022). For the former, we discard the neutral class (the reviews with a score of 3) and assign reviews with scores of 4 and 5 to the positive class and reviews with scores of 1 and 2 to the negative class. For both datasets, we sample 16 demonstrations per class.\nCommonsense reasoning. We use two commonsense reasoning datasets, XCOPA (Ponti et al., 2020) and XWINOGRAD (Muennighoff et al., 2022). Given a sentence and two options, a model selects one of the option labels, (A) or (B), based on which is better suited to the given context. Due to the smaller scale of the datasets, we sample 16 and 8 training instances in total for XCOPA and XWINOGRAD, respectively.\nQuestion answering. Question Answering (QA) is the task of answering a question given a paragraph, where the answer is a sub-span of the paragraph. We use TYDIQA-GOLDP (Clark et al., 2020), which we refer to as TYDIQA for simplicity. Due to the longer average input length, we limit the number of exemplars to 8. Named entity recognition. Named Entity Recognition (NER) is a representative sequence labeling task, where a system detects and classifies named entities in an input sentence. We adopt WIKIANN (Pan et al., 2017) and MASAKHANER (Adelani et al., 2021). Though WIKIANN covers 216 languages, we exclude languages that are covered only by WIKIANN or XLSUM due to the aforementioned issues. We convert the task into a text-to-text format, where given an input sentence, a model extracts all named entities with named entity tags:6 <location>, <person>, and <organization>. 7 We use 32 instances overall for few-shot transfer.\nSummarization. We use the XLSUM (Hasan et al., 2021) dataset to benchmark models' ability to generate a summary given a news article. Due to the context window limit, we use only 1 shot for training in this task.\nQuestion generation. Question generation generates a question according to a given input passage and a corresponding answer (Xiao et al., 2021). We convert the TYDIQA-GOLDP dataset into a question generation task, which we refer to TYDIQA-QG. Given the gold paragraph and an answer, the system generates the original question. We use 8 examples for few-shot training."},{"figure_ref":[],"heading":"BUFFET Evaluation","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Evaluation Metrics","publication_ref":["b15","b36"],"table_ref":[],"text":"Table 2 (Metric) lists task-specific metrics. To mitigate the variance from different few-shot samples, for each language included in each task, we take the average of a model's performance given three different sets of k-shot instances. Subsequently, each dataset score is calculated as a macro-average of the per-language score (Clark et al., 2020). Finally, following Liang et al. (2020), we take two separate average scores: (a) Avg. class score of all classification and QA tasks, and (b) Avg. generation score of all generation tasks."},{"figure_ref":[],"heading":"BUFFET-Light","publication_ref":[],"table_ref":[],"text":"Conducting a comprehensive evaluation covering a wide range of languages and tasks in BUFFET, while undoubtedly necessary, can be a timeconsuming process. We introduce BUFFET-light, which contains a representative subset of languages and tasks for a rapid assessment even in resourcelimited scenarios. We carefully select languages and datasets to ensure that we cover a diverse range of languages and output formats, assuming limited resources. See the overview of BUFFET-light in Appendix Section A.2."},{"figure_ref":[],"heading":"Benchmarking LMs on BUFFET","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Transfer Methods","publication_ref":["b30"],"table_ref":["tab_5"],"text":"In this study, we investigate various transfer methods with and without parameter updates. To assess the benefit of k-shot training examples in the target language, we also conduct experiments on zero-shot transfer methods. We assume that the model can optionally use instructions in the target language or another language, or full training sets in a high-resource language like English. This assumption is reasonable given the abundance of labeled datasets in high-resource languages (Yu et al., 2022;Joshi et al., 2020) and the cheaper costs of instruction annotations. Table 3 provides an overview of different approaches, categorized according to the optional inputs they use during training or inference.\nFine-tuning (methods with parameter updates).\nWe explore several transfer approaches that require parameter updates.\n• Target fine-tuning (TARGET FT) trains models on few-shot samples for each language. • English fine-tuning (ENGLISH FT) trains models on a source language (i.e., English) only and uses no target language data. • English+Target fine-tuning (ENG.+TGT. FT) first trains models on large-scale English datasets and then fine-tunes models on few-shot samples of target languages."},{"figure_ref":[],"heading":"In-context learning (methods without updates).","publication_ref":[],"table_ref":[],"text":"We explore several in-context learning methods.\n• English in-context learning (ENGLISH ICL) uses English instructions and demonstrations in the target languages. • Target ICL (TARGET ICL) uses both instructions and demonstrations in the target language. • Zero-shot English In-context learning (Z-EICL) uses only English instructions without demonstrations (neither in English nor in the target language), as in zero-shot transfer.\nUnlike in English, where abundant instructions and instance annotations are available, for many languages we often lack annotated instructions (Wang et al., 2022b). We use machine-translated instructions in BUFFET as the main baseline."},{"figure_ref":[],"heading":"Language Models","publication_ref":["b20","b13","b42","b43","b50","b54","b54","b42","b43"],"table_ref":["tab_5"],"text":"A key aspect of language models is their pretraining strategies. In addition to conventional pretraining using unlabeled corpora (Devlin et al., 2019;Brown et al., 2020), instruction-tuning has been actively studied; this approach trains an LLM on a massive number of tasks with instructions (Muennighoff et al., 2022;Ouyang et al., 2022;Wei et al., 2022). In this work, we evaluate six diverse models pretrained with different strategies (Table 3).\nModels for fine-tuning. Due to the high costs of fine-tuning for every k-shot setting, we experiment with an efficient yet competitive mT5-base with 580 million parameters (Xue et al., 2021).\nModels for in-context learning. We experiment with BLOOM-7B (7 billion parameters; Scao et al., 2022) and mT5-xxl (13 billion parameters; Xue et al., 2021). We also experiment with their instruction-tuned variants: BLOOMZ-7B and mT0xxl (Muennighoff et al., 2022), as well as the current state-of-the-art ChatGPT (gpt-3.5-turbo; Ouyang et al. 2022). Note that these models are trained on some of the datasets included in BUFFET. We do not exclude such overlapping datasets, but we indicate such seen tasks with * in the main result table. 8"},{"figure_ref":[],"heading":"Experiment Details","publication_ref":["b48","b51","b58","b23","b7"],"table_ref":[],"text":"Fine-tuning. In all settings, we fine-tune models on few-shot samples for 300 epochs for TARGET FT and 200 epochs for ENG.+TGT. FT. When finetuning LMs on large-scale English datasets (for both ENG.+TGT. FT and ENGLISH FT), we train for three epochs. We use representative English datasets following Hu et al. (2020b): SQUAD (Rajpurkar et al., 2016) for QA, MNLI (Williams et al., 2017) for NLI, PAWS (Zhang et al., 2019) for paraphrase detection, XLSUM (Hasan et al., 2021) for summarization, COPA (Arun and Balakrishnan, 2018) for XCOPA, WINOGRAD for XWINOGRAD, the AMAZON MULTILINGUAL REVIEW English 8 It is unclear which datasets ChatGPT is trained on.\nset for sentiment analysis, and the TYDIQA-QG English set for question generation.\nIn-context learning. We prompt LLMs with instructions and k-shot demonstrations available in BUFFET. Different models have different maximum context window sizes: mT0 only accepts up to 1024 tokens, while BLOOMZ and ChatGPT accept up to 2048 and 4096, respectively. We add training instances up to the maximum token length for each model and discard instances that do not fit the context window. We found that mT0 often performs well-given zero or smaller numbers of fewshot samples. We use 4-shots for mT0 ENGLISH ICL and TARGET ICL by default. We use greedy decoding for predictions. For tasks with a fixed set of pre-specified answer candidates, we compute the probability of option tokens by iterating all options except for ChatGPT without access to token probabilities. Due to the high inference costs, we evaluate ChatGPT only on BUFFET-Light, 5 Results and Analysis"},{"figure_ref":["fig_1"],"heading":"Main Results","publication_ref":["b42","b50","b41","b42","b29"],"table_ref":["tab_6","tab_25"],"text":"Table 4 shows aggregated results of fine-tuned and in-context learning-based LMs on BUFFET. We show full experiment results on each task in the Appendix. Below, we summarize the key findings.\nLLMs with in-context learning often lag behind much smaller fine-tuned models. While in-context learning has shown remarkable performance in English, our comparison shows that fewshot cross-lingual transfer via in-context learning remains challenging; ENGLISH ICL using BLOOM, BLOOMZ (7 billion) and mT0 (13 billion) often under-perform mt5-base (580 million) fine-tuned on English datasets (ENGLISH FT or ENG.+TGT. FT). However, when abundant English task data is not available, mT5-based finetuning methods (TARGET FT, or ENG.+TGT. FT on XCOPA and XWINOGRAD) often perform poorly and are outperformed by ENGLISH ICL or TARGET ICL baselines. This implies that when lacking task-specific training data, prompting LLMs can be more effective.\nInstruction-tuning helps in zero-shot but may not generalize for few-shot settings. 10. The blue-colored rows are instruction-tuned models, and we added * symbols next to the scores for the tasks on which the models have been trained. \"Random\" shows random baseline performance.\nBold fonts indicate the best results for each task, among the models that are not directly trained on the task. When ChatGPT achieves the best results, we also note the second-best number from the models that are not trained on the task, acknowledging the possibility that ChatGPT may have encountered a similar task during training.\nand BLOOMZ-7B Z-EICL, demonstrate significantly better performance compared to their noninstruction tuned counterparts, namely mT5-xxl and BLOOM-7B Z-EICL, with margins of 12.7 and 23.9 points in Avg. class, respectively. It is worth noting that while the performance improvements on seen tasks contribute to these gains (indicated by *), mT0-xxl Z-EICL exhibits substantial advancements on unfamiliar tasks. This further confirms the effectiveness of instructiontuning in zero-shot transfer, as discussed in prior studies (Muennighoff et al., 2022;Wei et al., 2022;Mishra et al., 2022).\nHowever, our study also highlights a surprising performance deterioration when moving from zeroshot to few-shot settings for instruction-tuned models: across tasks, mT0 performs worse in few-shot settings than in zero-shot settings (ENGLISH ICL v.s. Z EICL). BLOOMZ shows performance gains from few-shot demonstrations; BLOOMZ E ICL achieves 44.3, outperforming BLOOMZ Z EICL by 5 points in Avg. class score. Yet, it also ex-hibits large performance declines on the tasks that are used during their instruction-tuning (TYDIQA, PAWS-X). Our hypothesis is that such instructiontuned models are optimized to execute a new task solely based on an instruction, with no prior demonstrations (Muennighoff et al., 2022), and may struggle to learn in context from few-shot demonstrations. We conduct controlled experiments in Section 5.2 for further analysis.\nZero-or few-shot transfer remains challenging in under-represented languages. transferability of the model is often constrained in understudied languages.\nIn NER and NLI tasks, a noticeable decrease occurs in performance from high-resource to low-resource languages. It's important to note that several languages included in MASAKHANER or AMERICAS NLI are not part of the pretraining process. Models such as mT5 ENGLISH FT or ChatGPT ENGLISH ICL exhibit strong performance in high-resource languages. However, their performance significantly drops in less-represented languages. For instance in Aymara (aym), ChatGPT achieves slightly higher performance than a random baseline, outperformed by mT5 ENG.+TGT. FT by 13%. mT5 ENG.+TGT. FT also significantly outperforms mT5 ENGLISH FT in lower-resource languages, as indicated by the performance gap between the orange and blue lines in Figure 3. Notably, mT5 ENG.+TGT. FT outperforms mT5 ENGLISH FT by 30% in Hausa on MasakhaNER. This indicates that fine-tuning with only k instances in target languages can still greatly helps in less-represented languages.\nWe also observe performance drops in Finnish, Korean, and Russian for BLOOM and BLOOMZ in TYDIQA. Finnish, Korean, and Russian are excluded from BLOOM pretraining,10 which we attribute to these performance drops. Conversely, mT5 fine-tuning-based methods consistently display strong performance across languages. Interestingly, in Bengali, which is often considered less represented, BLOOMZ achieves performance comparable to mT5 fine-tuned models. We also observe the same trends in BLOOMZ. These results suggest pretraining setup may strongly affect downstream task performance even after instruction tuning.\nChatGPT has strong generation capabilities but requires careful instruction design. As discussed, though ChatGPT significantly outperforms other LLMs with in-context learning, its performance often lags behind fine-tuning-based methods in some discriminative tasks, particularly in lessrepresented languages. ChatGPT, however, significantly outperforms fine-tuned models on tasks that require target language generations (e.g., question generation, QA) with the exception of summarization (XLSUM). On XLSUM, we found that Chat-GPT often generates semantically correct summarizations in English rather than in the input article language, resulting in low ROUGE-2 scores. We do not observe that phenomenon in other LLMs (e.g., BLOOMZ); we show some ChatGPT output examples in the Appendix Table 25. Though more prompt engineering can boost ChatGPT's performance in summarization (Huang et al., 2023), we use the same prompts throughout the evaluations for a fair comparison. We also observe that when instructions are given in the target language, Chat-GPT often outputs a summary in the language, as shown in improved XLSUM performance in Chat-GPT TARGET ICL."},{"figure_ref":["fig_4"],"heading":"Analysis","publication_ref":["b59","b50","b13"],"table_ref":["tab_6"],"text":"Performance variance among different k shots. Figure 4 shows model performance across the three different k-shots and reveals a significant performance disparity in many of the tasks and languages. We observe the significant variance in fine-tuning-based transfer across different k-shot samples, confirming (Zhao et al., 2021). Importantly, we show that in-context learning is even more sensitive to different demonstrations than fewshot fine-tuning. For instance, for AMAZON RE-VIEW, the standard deviation for BLOOM E-CIL TyDI-Ar TyDI-Fi TyDI-Be TyDI-Ko WikiANN-Fr WikiANN-Vi WikiANN-Sw WikiANN-Te Indic-As Indic-Hi Indic-Gu Indic-Ta Amazon-De Amazon-Es Amazon-Zh Amazon-Ja and mT5 ENG.+TGT. fine-tuning is 2.2 and 0.2, respectively. We also analyze whether a demonstration set k that achieves the best performance with a model also leads to the optimal performance for another model. Specifically, we compare the best kshots for each task and language for BLOOM and BLOOMZ English ICL. We found that in 49.7% of the cases, their optimal k-shot demonstrations differ. These results emphasize the difficulty of comparing model performance in the absence of standardized k-shot samples. On the bright side, these results provide insights into potential approaches for identifying optimal demonstrations that can enhance few-shot ICL performance.\nThe effects of varying number of k. Figure 5 demonstrates the impact of increasing the number of few-shot samples for in-context learning and fine-tuning, on four tasks: TYDIQA, TYDIQA-QG, WIKIANN, and AMAZON REVIEW. Full results on the four tasks in a subset of the languages are available in Appendix D.3. Specifically, we vary the number of few-shot demonstrations, in-cluding 1, 4, and 8 (for the tasks with more than 8 shots), and assess the performance of BLOOM EN-GLISH ICL, BLOOMZ ENGLISH ICL, mT0 EN-GLISH ICL and mT5 ENG.+TGT. FT.\nIncreasing the number of few-shot examples has a notable positive impact on fine-tuning (mT5 fine-tuning) across different tasks. Similarly, noninstruction-tuned BLOOM also benefits from the inclusion of few-shot samples on most of the tasks. However, for instruction-tuned models (mT0 and BLOOMZ), we observe a significant decline in performance when additional demonstrations are added, which aligns with the findings in Table 4. Specifically, on mT0, we observe that the zero-shot performance surpasses the few-shot performance on TYDIQA and AMAZON REVIEW. Surprisingly, even on previously unseen tasks such as TYDIQA-QG and WIKIANN, the addition of more than four demonstrations leads to a significant decline in performance.\nIt is worth noting that mT0 and BLOOMZ were exclusively trained with instructions and did not utilize demonstrations during training (Muennighoff , 2022). We hypothesize that this training approach may cause the models to overfit the zeroshot instruction-based in-context learning scenario, thereby hindering their ability to effectively learn in-context information through few-shot demonstrations. Wei et al. (2022) also find that while fewshot demonstrations mitigate high variance of the zero-shot inference with instructions only, the optimal zero-shot performance with the best template often outperforms the best few-shot performance.\nEffects of model scaling on few-shot in-context cross-lingual transfer. Figure 6 shows BLOOM-560 million, 1 billion, and 7 billion performance on a subset of the tasks. The transfer method is ENGLISH ICL. As the model scales, the overall performance on few-shot in-context learning significantly improves, as found in English (Brown et al., 2020), indicating that models' cross-lingual fewshot transfer performance via in-context learning may improve as the model size increases. These findings are consistent with the results reported by Lin et al. ( 2021) on a set of classification tasks."},{"figure_ref":[],"heading":"Conclusion and Discussion","publication_ref":["b14","b43","b22","b26","b2","b38","b31","b15","b8","b25","b42","b3","b35","b8"],"table_ref":[],"text":"In this work, we introduce BUFFET, a few-shot cross-lingual transfer benchmark that encompasses a diverse range of discriminative and generative tasks across a variety of typologically distinct languages. Through our comprehensive evaluation, involving six different transfer methods and various LLMs, we offer valuable insights into the strengths and limitations of these transfer methods and LLMs. Our analysis reveals that while LLMs utilizing in-context learning excel in generation tasks, they are often surpassed by smaller fine-tuned models specifically trained for target tasks. Furthermore, our findings highlight significant performance variations dependent on different transfer setups (e.g., demonstrations).\nMoving forward, our findings suggest the following exciting opportunities for future research in the field of few-shot learning transfer across diverse languages:\nImprove multilingual instruction tuning. Although instruction tuning can be beneficial for both zero-shot transfer, certain models, such as mT0, may become overly specialized for zero-shot instruction-tuning scenarios, leading to lower average few-shot performance than the optimal zeroshot performance. Although these models demonstrate impressive zero-shot performance, even on tasks they haven't encountered before (such as XCOPA), they face challenges when it comes to tasks that involve generating outputs in less commonly used formats (like structured predictions). We believe that developing multilingual instructionfollowing models capable of effectively utilizing both instructions and demonstrations is crucial. Recent studies demonstrate that incorporating both instructions and demonstrations during instructiontuning on English data can enhance the model's performance (Chung et al., 2022), allowing it to learn within context (Min et al., 2022a). This type of training may potentially mitigate the issue of overfitting to specific formats. Hence, it is necessary to explore various instruction-tuning setups to further improve few-shot in-context learning, with a focus on cross-lingual transfer.\nAdditionally, while high-quality humantranslated instructions are effective, numerous instruction repositories are still dominated by English instructions.\nTherefore, community efforts to increase the availability of multilingual instructions may assist in the development of more generalizable multilingual large-language models. Overcome data scarcity using LLMs. Our research reveals that smaller task-specific fine-tuned models, with intermediate training in English, can still outperform ChatGPT on discriminative tasks that require strict output formats. Conversely, Chat-GPT outperforms fine-tuned models on tasks that necessitate more open-ended generations, such as question generation. In recent studies, Instruct-GPT (Ouyang et al., 2022) has exhibited the ability to generate high-quality generations in English, even outperforming humans on some tasks (Goyal et al., 2022). This impressive capacity for flexible generations has prompted active investiga-tions into generating training instances from such LLMs, which have predominantly focused on English (Wang et al., 2022a;Honovich et al., 2022). Some preliminary attempts have been made to explore task-specific data generation in certain target tasks, such as question answering (Agrawal et al., 2022). However, there remains limited exploration on how to generate diverse task instructions and outputs for a variety of typologically diverse languages. We believe that using LLMs to generate data offers a promising solution to obtaining more annotated data for under-represented languages.\nUnderstand transfer dynamics in cross-lingual in-context learning. The impact of various instructions and demonstrations has been extensively examined in the context of English in-context learning, highlighting critical concerns such as sensitivity to prompt order (Lu et al., 2022) and/or motivating methods for identifying optimal demonstrations (Su et al., 2022). This research has found that demonstrations or instructions that are optimal for one model may not necessarily result in the best performance for another model. We anticipate that our benchmark will inspire and assist in further research into the relationship between language and instruction/demonstration for cross-lingual incontext learning.\nFairness beyond languages: underrepresented variants, dialects, and cross-cultural NLP. Many of the diverse world languages are often excluded in widely used cross-lingual evaluation benchmarks, where recent papers show strong cross-lingual transfer capabilities. However, through our comprehensive analysis, we have discovered that even the most advanced LLMs currently available still face difficulties when dealing with less-represented languages. The most competitive instruction-tuned models, ChatGPT or mT0, show significant performance declines when it comes to indigenous languages, reaching a level akin to a random baseline.\nWe advocate for conducting more studies on diverse local languages, including under-represented languages and their dialects, as emphasized in previous works such as Aji et al. (2022); Kakwani et al. (2020). We note that datasets in such languages are often translated from English (Yu et al., 2022), which may introduce translation biases (Artetxe et al., 2020b) and fail to capture the linguistic nuances and interests of native speakers (Clark et al., 2020;Asai et al., 2021). To address these challenges, it is important that further work be done to develop cross-cultural Natural Language Processing (Hershcovich et al., 2022).\nExpand evaluations to complex tasks. Most recent research on multilingual in-context learning predominantly focuses on discriminative tasks (Muennighoff et al., 2022;Ahuja et al., 2023) or translation tasks (Lin et al., 2021). Further exploration can expand these evaluations to more diverse and complex tasks, such as MTOP (Li et al., 2021) or MGMS8K (Shi et al., 2023), or knowledgeintensive tasks (Asai et al., 2021) as new multilingual benchmarks are developed."},{"figure_ref":[],"heading":"Limitations","publication_ref":[],"table_ref":[],"text":"As the first step toward standardized evaluation for few-shot cross-lingual transfer, BUFFET focuses on popular discriminative tasks and some generative tasks. It does not include many datasets that require complex reasoning tasks, as noted above. Since our main focus is to benchmark different LLMs and learning methods in a comparable format, we do not explore sophisticated prompting methods, which can further boost performance. We anticipate that BUFFET will encourage the LLM community to explore new methods to further improve in-context learning beyond English. We use instructions translated by the NLLB (Costajussà et al., 2022) for TARGET ICL; such machinetranslated instructions are prone to errors, especially in less-represented languages, that can affect the final performance."},{"figure_ref":[],"heading":"Ethics Statement","publication_ref":[],"table_ref":["tab_7","tab_9"],"text":"While there has been significant research on incontext learning with LLMs, most of the focus has been on the English language. This raises questions about the applicability of findings from English few-shot NLP to few-shot cross-lingual transfer scenarios. To address this gap, BUFFET aims to provide a comprehensive and less biased evaluation framework. However, it is important to note that our benchmark dataset currently covers only 57 out of the approximately 6,000 world languages. Moreover, we do not specifically focus on finer-grained language varieties and dialects that are commonly spoken by underrepresented populations. In light of these limitations, we encourage future research to explore the effectiveness and limitations of widelyused transfer methods in a more diverse range of languages. This will help us gain a deeper understanding of the generalizability of transfer learning techniques across different linguistic contexts.\nList of the languages. We show the list of the 55 languages included in BUFFET in Table 5. BUFFET covers 25 different language families, and also shows geographical diversities.\nExamples. Table 6 shows the input and output examples in BUFFET. We reformulate all o the tasks with diverse formats into the same text-to-text format.\nInstructions. The full list of the instructions written in English is available in Table 7."},{"figure_ref":[],"heading":"List of the datasets with languages included.","publication_ref":[],"table_ref":["tab_10"],"text":"Table 8 shows the full list of the datasets with language names included in BUFFET."},{"figure_ref":[],"heading":"A.2 BUFFET-Light","publication_ref":[],"table_ref":["tab_11"],"text":"Task and language decisions. The goal of building the BUFFET-Light subset is to enable quick multilingual evaluation without losing the language and task diversity in the original BUFFET. To this end, we filter BUFFET so that we evaluate between 3 and 7 languages per task, and each language is included in at most three tasks. 11 This design choice allows us to consider 31 diverse languages across all tasks in BUFFET while reducing the number of evaluation settings by 66%.\nFinal list of BUFFET-light. The full list of tasks and languages in BUFFET are in Table 9."},{"figure_ref":[],"heading":"B More Experimental Details","publication_ref":["b54"],"table_ref":[],"text":"Fine-tuning. For ENGLISH FT, we limit the number of English training samples to 100,000 and fine-tune mt5-base (Xue et al., 2021) for 3 epochs. For the ENGLISH FT baseline, we transfer this model directly to new languages, while for ENG.+TGT. FT, we initialize the model checkpoint with the trained model weight and further fine-tune a model on few-shot samples for 300 epochs. In-context learning. We set the maximum token length to 15 except for XLSUM and TYDIQA-QG.\nFor XLSUM, we set the maximum token length to 100, and for TYDIQA-QG, we set the maximum token length to 50. We use greedy decoding throughout the experiments. For BLOOM-based model evaluations, we use a single RTX-100 GPU \nᄅ ᅩ ᆺᄃ ᅦᄂ ᅳ ᆫ ᄋ ᅵᄇ ᅥ ᆫ ᄉ ᅡ ᆼᄇ ᅡ ᆫᄀ ᅵ ᄎ ᅢᄋ ᅭ ᆼᄀ ᅪ ᄀ ᅪ ᆫᄅ ᅧ ᆫᄒ ᅢ ᄀ ᅮᄌ ᅵ ᆨᄌ ᅡᄃ ᅳ ᆯᄋ ᅦᄀ ᅦ ᄉ ᅵ ᆯᄌ ᅵ ᆯᄌ ᅥ ᆨᄋ ᅵ ᆫ ᄃ ᅩᄋ ᅮ ᆷᄋ ᅵ ᄃ ᅬ ᆯ ᄉ ᅮ ᄋ ᅵ ᆻᄃ ᅩᄅ ᅩ ᆨ ᄃ ᅡᄋ ᅣ ᆼᄒ ᅡ ᆫ ᄇ ᅡ ᆼᄇ ᅥ ᆸᄋ ᅳᄅ ᅩ ᄌ ᅥ ᆼᄇ ᅩᄌ ᅦᄀ ᅩ ᆼ ᄒ ᅪ ᆯᄃ ᅩ ᆼᄋ ᅳ ᆯ ᄀ ᅡ ᆼᄒ ᅪᄒ ᅡ ᆯ ᄀ ᅨᄒ ᅬ ᆨᄋ ᅵᄃ ᅡ. [SEP] hypothesis: ᄅ ᅩ ᆺᄃ ᅦᄂ ᅳ ᆫ ᄋ ᅥᄄ ᅥᄒ ᅡ ᆫ ᄌ ᅥ ᆼᄇ ᅩᄃ ᅩ ᄌ ᅦᄀ ᅩ ᆼᄒ ᅡᄌ ᅵ ᄋ ᅡ ᆭᄋ ᅳ ᆯ ᄀ ᅨᄒ ᅬ ᆨᄋ ᅵᄃ ᅡ.\ncontradiction Table 6: The input and output examples in BUFFET. We show one example from one dataset per task. Due to the long input length, we do not include a summarization example in this table.\nwith 24 GB GPU memory. We use int8bit quantization to avoid GPU out-of-memory errors. To evaluate mT5 and mT0, we use TPU v3-8."},{"figure_ref":[],"heading":"C Detailed BUFFET Results","publication_ref":[],"table_ref":[],"text":"This section includes the full list of the experimental results. Overall results on the full BUFFET are available in Table 10, and Figure 4 summarize overall performance across the eight tasks, on the BUFFET-Light subset. The overall trends on BUFFET-light remain the same as the original BUFFET. This indicates BUFFET-Light is a reliable and more efficient alternative for holistic evaluations for few-shot cross-lingual transfer. Below, we present the performance breakdown for each dataset. PAWS-X In this task you are given a sentence pair that has high lexical overlap. If the sentences have the same meaning and are just paraphrases of each other label them as duplicate, if not label them as not_duplicate"},{"figure_ref":[],"heading":"SENTIMENT","publication_ref":[],"table_ref":[],"text":"In this task, you're given a review from Amazon. Your task is to generate a rating for the product. The rating is extremely negative, negative, neutral, positive, and really positive."},{"figure_ref":[],"heading":"XCOPA","publication_ref":[],"table_ref":[],"text":"In this task you are given a premise and two alternatives (A) and (B). You must choose the alternative that is more plausibly the cause or effect of the situation described by the premise.\nXWINOGRAD Replace the _ in the above sentence with the correct option QA Read the given passage and answer a question about the information present in the passage."},{"figure_ref":[],"heading":"NER","publication_ref":[],"table_ref":[],"text":"Given the following sentence, indicate the name entities (i.e., the real-world objects such as a person, location, organization, etc. that can be denoted with a proper name) such as \"New York Times\". For each word of a named-entity, indicate their type \"location\" or \"organization\" or \"person\".\nSUMMARIZATION In this task, you are given an article. Your task is to summarize the article in a sentence."},{"figure_ref":[],"heading":"QG","publication_ref":[],"table_ref":[],"text":"This task is about reading the given passage and constructing a question about the information present in the passage. , it, ja, vi, qu, be, gu, et, th, or, kn, fi, gn, ru, el, ur, es, hi, te, as, sw, pa, bg, ml, tr, fa, id, ko, mr, de, ar, bn, zh MASAKHANER amh, hau, ibo, kin, luo, pcm, swa, wol, yor SUMMARIZATION XLSUM (english, ) ta, vi, id, tr, ja, th, bn, ar, en, es, fa, zh, sw QG TYDIQA-QG (en,) ar, be, fi, id, sw, ko, ru, te Table 10: Overall experiment results on BUFFET. The blue-colored rows are instruction-tuned models, and we added * symbols next to the scores for the tasks on which the models have been trained. \"Random\" shows random baseline performance. Bold fonts indicate the best results for each task, among the models that are not directly trained on the task. When ChatGPT achieves the best results, we also note the second-best number from the models that are not trained on the task, acknowledging the possibility that ChatGPT may have encountered a similar task during training."},{"figure_ref":[],"heading":"C.1 NLI","publication_ref":[],"table_ref":["tab_13","tab_2","tab_14"],"text":"Table 11 shows the full results on AMERICASNLI.\nTable 12 shows the full results on XNLI. Table 13 presents the full results on the other three entailment datasets annotated in each language, KLU-ENLI, OCNLI, and PARSINLUENTAILMENT. On XNLI, ENGLISH FT (zero-shot transfer) shows strong performance and often outperforms ENG.+TGT. FT (few-shot transfer). Among ICL baselines, mT0 ZICL shows the best macro performance on XNLI. However, on AMERICASNLI, all methods struggle, while ENG.+TGT. FT shows the best macro performance on AMERICAS NLI. The performance gap between ENGLISH FT and ENG.+TGT. FT get significantly larger, with the largest gap in Aymara (5.5%). Despite its strong performance on XNLI, mT0 ZICL struggles in AMERICAS NLI (33.7% on average).\nWhile mT0 ZICL shows robust performance across XNLI languages, ChatGPT shows a large performance gap between higher-resource languages and low-resource languages (57% in Greek v.s. 33% Urdu)."},{"figure_ref":[],"heading":"C.2 Paraphrase Detection","publication_ref":["b42"],"table_ref":[],"text":"The results on PAWS-X results are available in Table14. ENG. FT shows the best performance on this task among non-instruction-tuned models. We hypothesize that as the languages included in PAWS-X are all relatively well-represented languages and the task is relatively simple, ENG. FT, which is not trained in the target languages, can achieve high performance. mT0 ZICL shows quite high performance, likely because the model is trained on PAWS-X (Muennighoff et al., 2022)."},{"figure_ref":[],"heading":"C.3 Sentiment Analysis","publication_ref":[],"table_ref":["tab_16"],"text":"The experimental results on AMAZON REVIEW MULTI and INDIC SENTIMENT are available in Tables 15 and16. On both datasets, all models yield high accuracy across languages, except for mT5 ZEICL."},{"figure_ref":[],"heading":"C.4 Commonsense","publication_ref":[],"table_ref":["tab_2","tab_18"],"text":"XCOPA. The experimental results on XCOPA are available in 33.7 33.5 34.6 33.2 33.6 33.3 33.5 33.3 34.3 34.0 33.6 TICL mT5 33.3 33.3 32.8 33.3 33.6 33.2 33.2 33.3 33.3 33.3 33.3 TICL BLOOMZ 33.4 33.3 33.5 33.7 33.3 33.3 33.8 33.4 33.3 33.3 33.3 TICL mT0 33.4 33.6 33.3 33.3 33.3 33.3 33.3 33.3 33.3 33.3 Table 12: Model performance on XNLI. We report the average of the three few-shot samples.\nlanguages. ChatGPT achieves particularly notable performance in Italian (91.2%). On the other hand, all of the fine-tuning-based methods struggle, as the small size of the source datasets in English. This result indicates that for a task that lacks a large-scale training dataset even in high-resource Table 16: Model performance on INDIC SENTIMENT. We report the average of the three few-shot samples.\nlanguages, LLMs using in-context learning may often result in higher performance. We observed that mT0 ENGLISH FT faces difficulties when applied to XCOPA. This could be attributed to the limited size of the XCOPA English set, which might not provide enough data for a smaller mT5-base model to acquire comprehensive task knowledge.\nXWINOGRAD. The experimental results on XWINOGRAD are available in Table 18. Similar to XCOPA, on XWINOGRAD, fine-tuning-based methods often struggle, while in-context learning with competitive LLMs yields strong performance."},{"figure_ref":[],"heading":"C.5 Question Answering","publication_ref":[],"table_ref":[],"text":"TYDIQA experimental results are available in Table 19. Both the fine-tuning and ICL methods exhibit commendable performance on this particular task. It is intriguing to note that both mT0 and BLOOMZ demonstrate relatively lower efficacy in Korean, Finnish, and Russian. This can be at-tributed to the fact that these languages were not included during the pretraining phase."},{"figure_ref":[],"heading":"C.6 Named Entity Recognition","publication_ref":[],"table_ref":["tab_20","tab_2"],"text":"WIKIANN. Table 20 contains the results for WIKIANN. We specifically present the few-shot results since we discovered that zero-shot baselines consistently exhibit extremely poor performance, often close to zero, primarily because generating the answer in the precise output format proves to be challenging.\nIt's important to acknowledge that the BUFFET-Light WIKIANN subset comprises languages that are relatively high-resource, which could potentially lead to an overestimation of ChatGPT's performance. When comparing the best fine-tuning method with ChatGPT in the BUFFET-light languages, they generally perform competitively, with the exception of Finnish. Table 22: Model performance on TYDIQA-QG. We report the average of the three few-shot samples.\nTARGET ICL shows large improvements from EN-GLISH ICL, which has not been observed in other tasks. When instructions in the target language are given, ChatGPT almost always generates a summary in the target language. Among non-instruction-tuned models, ENG.+TGT. FT yields the highest average performance. It should be noted that mT0 and BLOOMZ are trained on XLSUM. Nevertheless, their performance in some languages remains low."},{"figure_ref":[],"heading":"D More Analysis","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_6"],"heading":"D.1 Performance across Languages","publication_ref":[],"table_ref":[],"text":"Figure 8 shows performance across languages on the three tasks, NLI, NER, and QA, adding two more LLMs: BLOOMZ and mT0."},{"figure_ref":[],"heading":"D.2 Variances of Different k-shots","publication_ref":[],"table_ref":[],"text":"In Section 5, we show that different sets of demonstrations can cause significant performance differ-ences. We provide the full visualization results across different tasks."},{"figure_ref":[],"heading":"D.3 Variances of the Varying Number of k","publication_ref":[],"table_ref":[],"text":"We provide the full experimental results with a different number of k. We evaluate BLOOM EN-GLISH ICL, BLOOMZ ENGLISH ICL and mT5-ENG.+TGT. FINE-TUNING and mT0 ENGLISH ICL experimental results on AMAZON REVIEW, TYDIQA, TYDIQA-AG, WIKIANN,and in Figures 9,10,11 and 12,respectively. AMAZON REVIEW. On AMAZON REVIEW, All models except for BLOOM (pretraining only) show competitive zero-shot performance. BLOOM EN-GLISH ICL benefits from few-shot demonstrations while mT0 ENGLISH ICL exhibit performance deterioration as adding more demonstrations across languages.\nTYDIQA. Among ENGLISH ICL baselines, mT0 shows strong performance up to four demon- strations, although their performance gets really low once more demonstrations are added. Similar deterioration happens in BLOOMZ. On the contrary, BLOOM performance improves as more shots are added. Despite using only 32 shots.\nTYDIQA-QG. Unlike in AMAZON REVIEW or TYDIQA, BLOOMZ ENGLISH ICL shows performance improvements with more demonstrations in Arabic and Bengali, reaching the highest QG performance in Bengali with four demonstrations. On the contrary, both BLOOM and BLOOMZ show poor QG performance in Korean and Russian, pos-sibly due to the lack of those languages during pretraining.\nWIKIANN. On WikiANN, all of the models gain performance improvements by adding at least one demonstration, possibly due to the difficulty of learning the exact output format expected given the instruction only. As in other datasets, mT0 reaches its highest performance with four demonstrations. mT5 ENG.+TGT. FT exhibits performance drops with one shot, possibly due to their overfit to the single example."},{"figure_ref":[],"heading":"D.4 Variances of Different Instructions","publication_ref":[],"table_ref":["tab_24"],"text":"We investigate the effectiveness of different English instructions on question generation tasks for TYDIQA in 0 and 4-shot setting using mT0 and BLOOM as base models in Table 24. We compare four relevant instructions and one irrelevant instruction (an instruction for AMAZON REVIEW).\nIn a zero-shot setting, instruction does not make much difference for both instruction-tuned and noninstruction-tuned models, since irrelevant instructions are sometimes better than the relevant prompt.\nIn a four-shot setting, whether the instruction is relevant does not make a huge difference for BLOOM, and we observed that random seeds impact the performance more, yet the performances do suffer a sharp loss if we are using irrelevant instruction in the instruction-tuned model. We also discovered that different models might favor different instructions for different languages, for example, in Swahili, 4-shot BLOOM favors the first instruction, while mT0 favors the fourth instruction."},{"figure_ref":[],"heading":"D.5 Qualitative Results for Generation Tasks","publication_ref":[],"table_ref":["tab_25"],"text":"Table 25 shows some qualitative results of Chat-GPT ENGLISH ICL and TARGET TCL on XLSUM and TYDIQA. Given English instruction, ChatGPT often generates summaries in English, rather than in the article language. On the other hand, such cross-lingual generation behaviors don't occur in QA tasks, and the model's predictions with TAR-GET ICL and ENGLISH ICL exhibit high overlap with each other. We hypothesize that ChatGPT's this cross-lingual summarization behavior can be related to their private training corpus, and future work can further investigate this issue. This task is about reading the given passage and constructing a question about the information present in the passage. Construct a question in such a way that (i) it is unambiguous, (ii) it is answered from the passage, (iii) its answer is unique (iv) its answer is a continuous text span from the paragraph. Avoid creating questions that (i) can be answered correctly without actually understanding the paragraph and (ii) uses the same words or phrases given in the passage.\n5 In this task, you are given a review from Amazon. Your task is to generate a rating for the product on a scale of 1-5 based on the review. The rating means -2: extremely poor, 1: poor, 0: neutral, 1: good, 2: extremely good, 3: mixed and -3: no sentiment expressed. "},{"figure_ref":[],"heading":"Acknowledgements","publication_ref":[],"table_ref":[],"text":"This research was supported by NSF IIS-2044660, ONR N00014-18-1-2826, ONR MURI N00014-18-1-2670, DARPA MCS program through NIWC Pacific (N66001-19-2-4031), and Allen Distinguished Award. AA is supported by the IBM fellowship. We are grateful to Orevaoghene Ahia for her help with ChatGPT evaluations. We thank our volunteer translators, Joongwon Kim, Usharani Injeti, and Sven Dorkenwald, for their help with translating instructions into different languages. Finally, we extend our appreciation to Jonathan H. Clark, Orevaoghene Ahia, Sandy Kaplan, and UW NLP researchers for their feedback on this draft. Shailaja Keyur Sampat, Siddhartha Mishra, Sujan Reddy A, Sumanta Patro, Tanay Dixit, and Xudong Shen. 2022b. Super-NaturalInstructions: Generalization via declarative instructions on 1600+ NLP tasks.\nIn Proceedings of Empirical Methods in Natural Language Processing."},{"figure_ref":[],"heading":"Appendix A Benchmark Details","publication_ref":[],"table_ref":[],"text":"A.1 BUFFET Constructions This section will provide further details of the BUFFET construction process. "}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_0\",\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b12\",\n \"b34\",\n \"b24\",\n \"b11\",\n \"b54\",\n \"b42\",\n \"b42\",\n \"b43\"\n ],\n \"table_ref\": [],\n \"text\": \"Recent advances in NLP primarily focus on the English language (Blasi et al., 2022). Due to the lack of sufficient training data in most of the world's languages (Yu et al., 2022), prior work explores direct transfer of pretrained language models to new languages after fine-tuning on resource-rich languages (zero-shot cross-lingual transfer; Hu et al. 2020b). Transferring after training a model on a few examples (few-shot cross-lingual transfer) often boosts performance, especially in languages that are distant from the source language (Lauscher et al., 2020;Hedderich et al., 2020).\\nIn English, zero-or few-shot learning via incontext learning is an active area of research (Beltagy et al., 2022;Schick and Schütze, 2021a; Shin et al., 2020). In this learning paradigm, one prompts a large language model (LLM) with fewshot demonstrations or natural language instructions to adapt to a new task, without any parameter updates. Yet, few-shot transfer across languages is still under-explored (Lin et al., 2021) in a wide range of tasks and languages. Moreover, it is unclear how effectively in-context learning performs in comparison to widely-used fine-tuning-based transfer methods under a comparable setup.\\nThis work introduces a new benchmark called BUFFET: Benchmark of Unified Format FEwshot Transfer Evaluation (Figure 1) to enable rigorous evaluations and advance research on few-shot cross-lingual transfer. Similar to a rich buffet, BUF-FET curates a diverse mix of tasks: 15 different tasks-including classification, structured prediction, and natural language generation-across 54 languages. BUFFET has several unique characteristics that are not present in prior multi-task multilingual benchmarks (summarized in Table 1):\\n• BUFFET provides a fixed set of few-shot examples for training and validation, allowing for fair comparisons across LMs and transfer methods. • BUFFET includes datasets annotated in each language or covering under-represented languages, which are often not included in existing multi-task benchmarks. • BUFFET combines diverse tasks into a unified text-to-text format and provides a set of English and machine-translated instructions for each task, removing the burdens of task-specific architecture changes or prompt engineering.\\nUsing this new benchmark, we extensively evaluate the current state-of-the-art multilingual large language models (LLMs), including mT5 (Xue et al., 2021), mT0 (Muennighoff et al., 2022), BLOOM (Scao et al., 2022), BLOOMZ (Muennighoff et al., 2022), and ChatGPT (Ouyang et al., 2022), using both fine-tuning and in-context learning approaches. In particular, BUFFET enables us to investigate the following research questions: (RQ1) Is in-context learning competitive with fine-tuning in few-shot cross-lingual transfer? Notably, given the same small numbers of examples in the target languages, in-context learning on LLMs (including ChatGPT, the most powerful model we evaluate in this work) often underperforms much smaller specialized mT5-base models, as shown in Figure 1 (bottom left). (RQ2) How well do different transfer methods perform across tasks and languages? The performance gap between in-context learning-based baselines and fine-tuning-based baselines is more significant in under-represented languages (Figure 1 bottom center). On NLI in indigenous languages of the Americas, ChatGPT or mT0-11B using in-context learning performs barely above random, while 580 million parameter mt5-base fine-tuned models retain strong performance. On the contrary, these LLMs perform well on generative tasks where a smaller task-specific model struggles, demonstrating their superiority in generating fluent text for diverse languages without abundant training data. (RQ3) How does the choice of transfer setup affect different transfer strategies? BUFFET also enables us to perform an in-depth and extensive analysis of the effects of diverse demonstrations and instructions on the downstream transfer quality. Our observations indicate that the choice of few-shot training examples has a substantial influence on a model's performance, particularly, Multi-ling. Few-S Gen. Low-R\\nXTREME ✓ XTREME-R ✓ XGLUE ✓ ✓ CrossFit ✓ ✓ MEGA* ✓ ✓ BUFFET ✓ ✓ ✓ ✓\\nTable 1: Comparison of the existing benchmarks based on their multilinguality (Multi-ling.), few-shot task formulation (Few-S), availability of generative tasks (Gen.), and coverage of low-resource languages (Low-R). * indicates concurrent work.\\nwith greater variability in in-context learning, compared to fine-tuning. We note that optimal transfer settings may differ across models. For example, instruction-tuned models often face challenges in effectively utilizing few-shot samples and their performance deteriorates as the number of demonstrations increases, possibly because they are optimized for the zero-shot instruction-tuned training scheme. This highlights the need for a standardized benchmark to facilitate fair comparisons and further studies to assess such transfer dynamics in non-English data. Grounded in our analysis, we suggest avenues for future research in few-shot cross-lingual transfer for both dataset creation and model development. Our data and code are available online. 12 Background and Related Work\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Problem Formulation\",\n \"publication_ref\": [\n \"b12\",\n \"b30\",\n \"b34\"\n ],\n \"table_ref\": [],\n \"text\": \"Due to the lack of annotated training data in many languages (Blasi et al., 2022;Yu et al., 2022;Joshi et al., 2020), transferring models trained on resource-rich languages (e.g., English) to other languages has been actively studied in multilingual NLP. In this paper, our main focus is on fewshot cross-lingual transfer (Lauscher et al., 2020), where a model is adapted using only a limited number of training or validation examples in the target language L. Another popular paradigm is zeroshot cross-lingual transfer (Artetxe et al., 2020a;Hu et al., 2020b) from English, where a model has access to training sets or instructions in English but not in the target language.\\nVarious transfer methods have been investigated in the field, including the in-context learning methods (Section 2.3). Yet, limited research explores different transfer methods under comparable condi-tions. With our new benchmark, BUFFET, we facilitate fair comparisons between models and learning methods, establishing a basis for studying the dynamics of few-shot transfer across various languages (Section 2.2).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Benchmarks for Cross-lingual Transfer\",\n \"publication_ref\": [\n \"b49\",\n \"b36\",\n \"b56\",\n \"b59\",\n \"b3\"\n ],\n \"table_ref\": [],\n \"text\": \"To enable a scalable and rigorous evaluation across multiple tasks, prior work has proposed multi-task benchmarks that unify diverse existing datasets. XTREME (Hu et al., 2020b), XTREME-R (Ruder et al., 2021) and XGLUE (Liang et al., 2020) focus on zero-shot transfer of models fine-tuned on English datasets. Despite English-based few-shot evaluation benchmarks, such as CrossFit (Ye et al., 2021), in few-shot cross-lingual transfer, we lack a standardized evaluation benchmark to facilitate the comparison of models and learning methods at scale. BUFFET provides the first large-scale few-shot cross-lingual transfer suits to address the gap. Importantly, to mitigate the effects of the highperformance variance in few-shot cross-lingual transfer (Zhao et al., 2021), we curate and aggregate results from multiple fixed k-shot training instances for each task and language. Concurrent with our work, MEGA (Ahuja et al., 2023) conducts experiments of few-shot cross-lingual transfer with a focus on classification and question answering tasks. BUFFET unifies diverse tasks including both discriminative and generative tasks. We also include datasets covering languages underrepresented in prior work (e.g., African and indigenous languages). Table 1 summarizes the key differences between BUFFET and prior benchmarks.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Methods for Cross-lingual Transfer\",\n \"publication_ref\": [\n \"b20\",\n \"b54\",\n \"b52\",\n \"b34\",\n \"b24\",\n \"b13\",\n \"b42\",\n \"b60\",\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"Fine-tuning-based approaches. Multilingual pre-trained models (Devlin et al., 2019;Xue et al., 2021;Conneau et al., 2020a) have the ability to adapt to new languages with no or few training instances in a target language (Conneau et al., 2020b;Hu et al., 2020b;Wu and Dredze, 2019). Lauscher et al. (2020) andHedderich et al. (2020) report that particularly in languages that are distant from the source language, further fine-tuning model on few-shot samples greatly improves performance.\\nCross-lingual in-context learning. In-context learning (Brown et al., 2020) (Schick and Schütze, 2021b;Min et al., 2022b), most prior work focuses only on English. Recent work (Lin et al., 2021;Muennighoff et al., 2022) introduces pre-trained LMs trained on more multilingual pre-trained corpora or translated datasets and shows improved results. While prior evaluations often focus on classification or translation tasks (Zhu et al., 2023;Vilar et al., 2022), more recently Shi et al. (2023), evaluate the use of instructions, demonstrations, and rationales in different languages across multiple reasoning tasks. However, how much LLMs with respect to in-context learning compete with the aforementioned fine-tuned approaches in a comparable setup and at scale has yet to be investigated, as they often use a large number of training examples in target languages (Bang et al., 2023). We demonstrate even with a small number of training examples, fine-tuning methods are competitive with in-context learning for cross-lingual transfer.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Benchmark: BUFFET\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"We introduce a new standardized few-shot cross-lingual evaluation benchmark: BUFFET (Benchmark of Unified Format Few-shot Transfer Evaluation). BUFFET unifies diverse NLP tasks and provides fixed sets of few-shot samples per task to facilitate consistent comparisons (Table 2).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Design Principles\",\n \"publication_ref\": [\n \"b49\",\n \"b36\",\n \"b46\",\n \"b15\"\n ],\n \"table_ref\": [\n \"tab_7\",\n \"tab_2\"\n ],\n \"text\": \"We create the BUFFET benchmark to establish a rigorous and equitable evaluation framework for few-shot cross-lingual transfer across a broad range of tasks and languages. We adhere to the following design principles with our benchmark.\\nStandardized few-shot samples. BUFFET provides three different training and validation sets of k-shots (e.g., k = 32) per task for a nonclassification task, or per class for a classification task, for each language.\\nTask diversity. Existing cross-lingual benchmarks often focus on classification or retrieval (Hu et al., 2020b;Ruder et al., 2021;Liang et al., 2020). BUFFET encompasses a broad range of task types, such as classification, generation, extraction, and structured prediction tasks. By converting all tasks into the same text-to-text format, we eliminate the need for task-specific model modifications or template conversions. 5.\\nBeyond evaluations on translated data. Prior few-or zero-shot evaluations were often conducted on widely-used datasets translated from English (e.g., XNLI; Conneau et al. 2018, XCOPA;Ponti et al. 2020). Those datasets might exhibit undesired biases, such as translation artifacts or unnatural topic distributions (Clark et al., 2020;Artetxe et al., 2020b). We collect both translation-based datasets and datasets that are annotated directly in each language (Table 2, Data curation).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"BUFFET Construction Process\",\n \"publication_ref\": [\n \"b56\",\n \"b47\"\n ],\n \"table_ref\": [],\n \"text\": \"Following Ye et al. (2021), we unify all tasks into the same text-to-text format, where a model is expected to directly generate the desired outputs given diverse inputs (Raffel et al., 2020). For each dataset in BUFFET, we unify instance representations of instruction, k-shot instances for training and validation. Each training instance consists of an input and output. Figure 2 shows an overview. Section 3.2.1 provides the outline of the unification, and Section 3.2.2 provides a task-specific process.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Unification Process\",\n \"publication_ref\": [\n \"b50\",\n \"b9\",\n \"b41\",\n \"b19\"\n ],\n \"table_ref\": [],\n \"text\": \"Few-shot instance selection. By default, we use all of the languages included in the original datasets. Among multiple annotated instructions, we sample the first instruction for a similar task that suits our text-to-text scheme. For some tasks, we modify the original instruction to make labels consistent with the names used in BUFFET4 or to remove taskspecific dependencies in the input data field. See Appendix Table 6 for the full list of instructions.\\nInstruction translation. Despite rapid progress of instruction-tuning in English LLMs (Wei et al., 2022;Sanh et al., 2022;Mishra et al., 2022;Wang et al., 2022b), cross-lingual setups still lag behind due to a lack of instructions in the target languages.\\nPrior work often translates instructions for the target tasks (Lin et al., 2021;Shi et al., 2023). We provide translated instructions for 15 datasets in 54 target languages, translated by NLLB (Costa-jussà et al., 2022), and manually translate the instructions into five languages.5 \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Tasks and Dataset Curation\",\n \"publication_ref\": [\n \"b17\",\n \"b21\",\n \"b33\",\n \"b45\",\n \"b55\",\n \"b32\",\n \"b1\",\n \"b46\",\n \"b42\",\n \"b15\",\n \"b44\",\n \"b23\",\n \"b53\"\n ],\n \"table_ref\": [],\n \"text\": \"We first select eight popular NLP tasks and, for each task, we identify available datasets using a careful survey of multilingual datasets by Yu et al. (2022). Appendix Table 6 shows examples.\\nNatural language inference. Natural Language Inference (NLI) involves determining the logical relationship (i.e., entailment, contradiction, neutral) between two text fragments, i.e., a premise and a hypothesis. In addition to the widely used XNLI (Conneau et al., 2018), we gather NLI datasets that are annotated in each language or designed to cover extremely under-represented languages: AMER-ICASNLI (Ebrahimi et al., 2022), PARSINLU-ENTAILMENT (Khashabi et al., 2021), KLUE-NLI (Park et al., 2021), and OCNLI (Hu et al., 2020a). We use 16 examples for each class.\\nParaphrase detection. Paraphrase detection is the task of identifying whether two sentences have/do not have the same meaning (duplicate or not duplicated).\\nWe adopt PAWS-X (Yang et al., 2019) and include 16 shots for each class as few-shot training and validation data.\\nSentiment analysis. Binary sentiment analysis identifies whether a text (e.g., a product review from Amazon) expresses positive or negative sentiment towards a topic. We use the MULTILINGUAL AMAZON REVIEW DATASET (Keung et al., 2020) and INDICNLU-SENTIMENT (Aggarwal et al., 2022). For the former, we discard the neutral class (the reviews with a score of 3) and assign reviews with scores of 4 and 5 to the positive class and reviews with scores of 1 and 2 to the negative class. For both datasets, we sample 16 demonstrations per class.\\nCommonsense reasoning. We use two commonsense reasoning datasets, XCOPA (Ponti et al., 2020) and XWINOGRAD (Muennighoff et al., 2022). Given a sentence and two options, a model selects one of the option labels, (A) or (B), based on which is better suited to the given context. Due to the smaller scale of the datasets, we sample 16 and 8 training instances in total for XCOPA and XWINOGRAD, respectively.\\nQuestion answering. Question Answering (QA) is the task of answering a question given a paragraph, where the answer is a sub-span of the paragraph. We use TYDIQA-GOLDP (Clark et al., 2020), which we refer to as TYDIQA for simplicity. Due to the longer average input length, we limit the number of exemplars to 8. Named entity recognition. Named Entity Recognition (NER) is a representative sequence labeling task, where a system detects and classifies named entities in an input sentence. We adopt WIKIANN (Pan et al., 2017) and MASAKHANER (Adelani et al., 2021). Though WIKIANN covers 216 languages, we exclude languages that are covered only by WIKIANN or XLSUM due to the aforementioned issues. We convert the task into a text-to-text format, where given an input sentence, a model extracts all named entities with named entity tags:6 <location>, <person>, and <organization>. 7 We use 32 instances overall for few-shot transfer.\\nSummarization. We use the XLSUM (Hasan et al., 2021) dataset to benchmark models' ability to generate a summary given a news article. Due to the context window limit, we use only 1 shot for training in this task.\\nQuestion generation. Question generation generates a question according to a given input passage and a corresponding answer (Xiao et al., 2021). We convert the TYDIQA-GOLDP dataset into a question generation task, which we refer to TYDIQA-QG. Given the gold paragraph and an answer, the system generates the original question. We use 8 examples for few-shot training.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"BUFFET Evaluation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Evaluation Metrics\",\n \"publication_ref\": [\n \"b15\",\n \"b36\"\n ],\n \"table_ref\": [],\n \"text\": \"Table 2 (Metric) lists task-specific metrics. To mitigate the variance from different few-shot samples, for each language included in each task, we take the average of a model's performance given three different sets of k-shot instances. Subsequently, each dataset score is calculated as a macro-average of the per-language score (Clark et al., 2020). Finally, following Liang et al. (2020), we take two separate average scores: (a) Avg. class score of all classification and QA tasks, and (b) Avg. generation score of all generation tasks.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"BUFFET-Light\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Conducting a comprehensive evaluation covering a wide range of languages and tasks in BUFFET, while undoubtedly necessary, can be a timeconsuming process. We introduce BUFFET-light, which contains a representative subset of languages and tasks for a rapid assessment even in resourcelimited scenarios. We carefully select languages and datasets to ensure that we cover a diverse range of languages and output formats, assuming limited resources. See the overview of BUFFET-light in Appendix Section A.2.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Benchmarking LMs on BUFFET\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Transfer Methods\",\n \"publication_ref\": [\n \"b30\"\n ],\n \"table_ref\": [\n \"tab_5\"\n ],\n \"text\": \"In this study, we investigate various transfer methods with and without parameter updates. To assess the benefit of k-shot training examples in the target language, we also conduct experiments on zero-shot transfer methods. We assume that the model can optionally use instructions in the target language or another language, or full training sets in a high-resource language like English. This assumption is reasonable given the abundance of labeled datasets in high-resource languages (Yu et al., 2022;Joshi et al., 2020) and the cheaper costs of instruction annotations. Table 3 provides an overview of different approaches, categorized according to the optional inputs they use during training or inference.\\nFine-tuning (methods with parameter updates).\\nWe explore several transfer approaches that require parameter updates.\\n• Target fine-tuning (TARGET FT) trains models on few-shot samples for each language. • English fine-tuning (ENGLISH FT) trains models on a source language (i.e., English) only and uses no target language data. • English+Target fine-tuning (ENG.+TGT. FT) first trains models on large-scale English datasets and then fine-tunes models on few-shot samples of target languages.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"In-context learning (methods without updates).\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We explore several in-context learning methods.\\n• English in-context learning (ENGLISH ICL) uses English instructions and demonstrations in the target languages. • Target ICL (TARGET ICL) uses both instructions and demonstrations in the target language. • Zero-shot English In-context learning (Z-EICL) uses only English instructions without demonstrations (neither in English nor in the target language), as in zero-shot transfer.\\nUnlike in English, where abundant instructions and instance annotations are available, for many languages we often lack annotated instructions (Wang et al., 2022b). We use machine-translated instructions in BUFFET as the main baseline.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Language Models\",\n \"publication_ref\": [\n \"b20\",\n \"b13\",\n \"b42\",\n \"b43\",\n \"b50\",\n \"b54\",\n \"b54\",\n \"b42\",\n \"b43\"\n ],\n \"table_ref\": [\n \"tab_5\"\n ],\n \"text\": \"A key aspect of language models is their pretraining strategies. In addition to conventional pretraining using unlabeled corpora (Devlin et al., 2019;Brown et al., 2020), instruction-tuning has been actively studied; this approach trains an LLM on a massive number of tasks with instructions (Muennighoff et al., 2022;Ouyang et al., 2022;Wei et al., 2022). In this work, we evaluate six diverse models pretrained with different strategies (Table 3).\\nModels for fine-tuning. Due to the high costs of fine-tuning for every k-shot setting, we experiment with an efficient yet competitive mT5-base with 580 million parameters (Xue et al., 2021).\\nModels for in-context learning. We experiment with BLOOM-7B (7 billion parameters; Scao et al., 2022) and mT5-xxl (13 billion parameters; Xue et al., 2021). We also experiment with their instruction-tuned variants: BLOOMZ-7B and mT0xxl (Muennighoff et al., 2022), as well as the current state-of-the-art ChatGPT (gpt-3.5-turbo; Ouyang et al. 2022). Note that these models are trained on some of the datasets included in BUFFET. We do not exclude such overlapping datasets, but we indicate such seen tasks with * in the main result table. 8\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiment Details\",\n \"publication_ref\": [\n \"b48\",\n \"b51\",\n \"b58\",\n \"b23\",\n \"b7\"\n ],\n \"table_ref\": [],\n \"text\": \"Fine-tuning. In all settings, we fine-tune models on few-shot samples for 300 epochs for TARGET FT and 200 epochs for ENG.+TGT. FT. When finetuning LMs on large-scale English datasets (for both ENG.+TGT. FT and ENGLISH FT), we train for three epochs. We use representative English datasets following Hu et al. (2020b): SQUAD (Rajpurkar et al., 2016) for QA, MNLI (Williams et al., 2017) for NLI, PAWS (Zhang et al., 2019) for paraphrase detection, XLSUM (Hasan et al., 2021) for summarization, COPA (Arun and Balakrishnan, 2018) for XCOPA, WINOGRAD for XWINOGRAD, the AMAZON MULTILINGUAL REVIEW English 8 It is unclear which datasets ChatGPT is trained on.\\nset for sentiment analysis, and the TYDIQA-QG English set for question generation.\\nIn-context learning. We prompt LLMs with instructions and k-shot demonstrations available in BUFFET. Different models have different maximum context window sizes: mT0 only accepts up to 1024 tokens, while BLOOMZ and ChatGPT accept up to 2048 and 4096, respectively. We add training instances up to the maximum token length for each model and discard instances that do not fit the context window. We found that mT0 often performs well-given zero or smaller numbers of fewshot samples. We use 4-shots for mT0 ENGLISH ICL and TARGET ICL by default. We use greedy decoding for predictions. For tasks with a fixed set of pre-specified answer candidates, we compute the probability of option tokens by iterating all options except for ChatGPT without access to token probabilities. Due to the high inference costs, we evaluate ChatGPT only on BUFFET-Light, 5 Results and Analysis\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Main Results\",\n \"publication_ref\": [\n \"b42\",\n \"b50\",\n \"b41\",\n \"b42\",\n \"b29\"\n ],\n \"table_ref\": [\n \"tab_6\",\n \"tab_25\"\n ],\n \"text\": \"Table 4 shows aggregated results of fine-tuned and in-context learning-based LMs on BUFFET. We show full experiment results on each task in the Appendix. Below, we summarize the key findings.\\nLLMs with in-context learning often lag behind much smaller fine-tuned models. While in-context learning has shown remarkable performance in English, our comparison shows that fewshot cross-lingual transfer via in-context learning remains challenging; ENGLISH ICL using BLOOM, BLOOMZ (7 billion) and mT0 (13 billion) often under-perform mt5-base (580 million) fine-tuned on English datasets (ENGLISH FT or ENG.+TGT. FT). However, when abundant English task data is not available, mT5-based finetuning methods (TARGET FT, or ENG.+TGT. FT on XCOPA and XWINOGRAD) often perform poorly and are outperformed by ENGLISH ICL or TARGET ICL baselines. This implies that when lacking task-specific training data, prompting LLMs can be more effective.\\nInstruction-tuning helps in zero-shot but may not generalize for few-shot settings. 10. The blue-colored rows are instruction-tuned models, and we added * symbols next to the scores for the tasks on which the models have been trained. \\\"Random\\\" shows random baseline performance.\\nBold fonts indicate the best results for each task, among the models that are not directly trained on the task. When ChatGPT achieves the best results, we also note the second-best number from the models that are not trained on the task, acknowledging the possibility that ChatGPT may have encountered a similar task during training.\\nand BLOOMZ-7B Z-EICL, demonstrate significantly better performance compared to their noninstruction tuned counterparts, namely mT5-xxl and BLOOM-7B Z-EICL, with margins of 12.7 and 23.9 points in Avg. class, respectively. It is worth noting that while the performance improvements on seen tasks contribute to these gains (indicated by *), mT0-xxl Z-EICL exhibits substantial advancements on unfamiliar tasks. This further confirms the effectiveness of instructiontuning in zero-shot transfer, as discussed in prior studies (Muennighoff et al., 2022;Wei et al., 2022;Mishra et al., 2022).\\nHowever, our study also highlights a surprising performance deterioration when moving from zeroshot to few-shot settings for instruction-tuned models: across tasks, mT0 performs worse in few-shot settings than in zero-shot settings (ENGLISH ICL v.s. Z EICL). BLOOMZ shows performance gains from few-shot demonstrations; BLOOMZ E ICL achieves 44.3, outperforming BLOOMZ Z EICL by 5 points in Avg. class score. Yet, it also ex-hibits large performance declines on the tasks that are used during their instruction-tuning (TYDIQA, PAWS-X). Our hypothesis is that such instructiontuned models are optimized to execute a new task solely based on an instruction, with no prior demonstrations (Muennighoff et al., 2022), and may struggle to learn in context from few-shot demonstrations. We conduct controlled experiments in Section 5.2 for further analysis.\\nZero-or few-shot transfer remains challenging in under-represented languages. transferability of the model is often constrained in understudied languages.\\nIn NER and NLI tasks, a noticeable decrease occurs in performance from high-resource to low-resource languages. It's important to note that several languages included in MASAKHANER or AMERICAS NLI are not part of the pretraining process. Models such as mT5 ENGLISH FT or ChatGPT ENGLISH ICL exhibit strong performance in high-resource languages. However, their performance significantly drops in less-represented languages. For instance in Aymara (aym), ChatGPT achieves slightly higher performance than a random baseline, outperformed by mT5 ENG.+TGT. FT by 13%. mT5 ENG.+TGT. FT also significantly outperforms mT5 ENGLISH FT in lower-resource languages, as indicated by the performance gap between the orange and blue lines in Figure 3. Notably, mT5 ENG.+TGT. FT outperforms mT5 ENGLISH FT by 30% in Hausa on MasakhaNER. This indicates that fine-tuning with only k instances in target languages can still greatly helps in less-represented languages.\\nWe also observe performance drops in Finnish, Korean, and Russian for BLOOM and BLOOMZ in TYDIQA. Finnish, Korean, and Russian are excluded from BLOOM pretraining,10 which we attribute to these performance drops. Conversely, mT5 fine-tuning-based methods consistently display strong performance across languages. Interestingly, in Bengali, which is often considered less represented, BLOOMZ achieves performance comparable to mT5 fine-tuned models. We also observe the same trends in BLOOMZ. These results suggest pretraining setup may strongly affect downstream task performance even after instruction tuning.\\nChatGPT has strong generation capabilities but requires careful instruction design. As discussed, though ChatGPT significantly outperforms other LLMs with in-context learning, its performance often lags behind fine-tuning-based methods in some discriminative tasks, particularly in lessrepresented languages. ChatGPT, however, significantly outperforms fine-tuned models on tasks that require target language generations (e.g., question generation, QA) with the exception of summarization (XLSUM). On XLSUM, we found that Chat-GPT often generates semantically correct summarizations in English rather than in the input article language, resulting in low ROUGE-2 scores. We do not observe that phenomenon in other LLMs (e.g., BLOOMZ); we show some ChatGPT output examples in the Appendix Table 25. Though more prompt engineering can boost ChatGPT's performance in summarization (Huang et al., 2023), we use the same prompts throughout the evaluations for a fair comparison. We also observe that when instructions are given in the target language, Chat-GPT often outputs a summary in the language, as shown in improved XLSUM performance in Chat-GPT TARGET ICL.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"Analysis\",\n \"publication_ref\": [\n \"b59\",\n \"b50\",\n \"b13\"\n ],\n \"table_ref\": [\n \"tab_6\"\n ],\n \"text\": \"Performance variance among different k shots. Figure 4 shows model performance across the three different k-shots and reveals a significant performance disparity in many of the tasks and languages. We observe the significant variance in fine-tuning-based transfer across different k-shot samples, confirming (Zhao et al., 2021). Importantly, we show that in-context learning is even more sensitive to different demonstrations than fewshot fine-tuning. For instance, for AMAZON RE-VIEW, the standard deviation for BLOOM E-CIL TyDI-Ar TyDI-Fi TyDI-Be TyDI-Ko WikiANN-Fr WikiANN-Vi WikiANN-Sw WikiANN-Te Indic-As Indic-Hi Indic-Gu Indic-Ta Amazon-De Amazon-Es Amazon-Zh Amazon-Ja and mT5 ENG.+TGT. fine-tuning is 2.2 and 0.2, respectively. We also analyze whether a demonstration set k that achieves the best performance with a model also leads to the optimal performance for another model. Specifically, we compare the best kshots for each task and language for BLOOM and BLOOMZ English ICL. We found that in 49.7% of the cases, their optimal k-shot demonstrations differ. These results emphasize the difficulty of comparing model performance in the absence of standardized k-shot samples. On the bright side, these results provide insights into potential approaches for identifying optimal demonstrations that can enhance few-shot ICL performance.\\nThe effects of varying number of k. Figure 5 demonstrates the impact of increasing the number of few-shot samples for in-context learning and fine-tuning, on four tasks: TYDIQA, TYDIQA-QG, WIKIANN, and AMAZON REVIEW. Full results on the four tasks in a subset of the languages are available in Appendix D.3. Specifically, we vary the number of few-shot demonstrations, in-cluding 1, 4, and 8 (for the tasks with more than 8 shots), and assess the performance of BLOOM EN-GLISH ICL, BLOOMZ ENGLISH ICL, mT0 EN-GLISH ICL and mT5 ENG.+TGT. FT.\\nIncreasing the number of few-shot examples has a notable positive impact on fine-tuning (mT5 fine-tuning) across different tasks. Similarly, noninstruction-tuned BLOOM also benefits from the inclusion of few-shot samples on most of the tasks. However, for instruction-tuned models (mT0 and BLOOMZ), we observe a significant decline in performance when additional demonstrations are added, which aligns with the findings in Table 4. Specifically, on mT0, we observe that the zero-shot performance surpasses the few-shot performance on TYDIQA and AMAZON REVIEW. Surprisingly, even on previously unseen tasks such as TYDIQA-QG and WIKIANN, the addition of more than four demonstrations leads to a significant decline in performance.\\nIt is worth noting that mT0 and BLOOMZ were exclusively trained with instructions and did not utilize demonstrations during training (Muennighoff , 2022). We hypothesize that this training approach may cause the models to overfit the zeroshot instruction-based in-context learning scenario, thereby hindering their ability to effectively learn in-context information through few-shot demonstrations. Wei et al. (2022) also find that while fewshot demonstrations mitigate high variance of the zero-shot inference with instructions only, the optimal zero-shot performance with the best template often outperforms the best few-shot performance.\\nEffects of model scaling on few-shot in-context cross-lingual transfer. Figure 6 shows BLOOM-560 million, 1 billion, and 7 billion performance on a subset of the tasks. The transfer method is ENGLISH ICL. As the model scales, the overall performance on few-shot in-context learning significantly improves, as found in English (Brown et al., 2020), indicating that models' cross-lingual fewshot transfer performance via in-context learning may improve as the model size increases. These findings are consistent with the results reported by Lin et al. ( 2021) on a set of classification tasks.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion and Discussion\",\n \"publication_ref\": [\n \"b14\",\n \"b43\",\n \"b22\",\n \"b26\",\n \"b2\",\n \"b38\",\n \"b31\",\n \"b15\",\n \"b8\",\n \"b25\",\n \"b42\",\n \"b3\",\n \"b35\",\n \"b8\"\n ],\n \"table_ref\": [],\n \"text\": \"In this work, we introduce BUFFET, a few-shot cross-lingual transfer benchmark that encompasses a diverse range of discriminative and generative tasks across a variety of typologically distinct languages. Through our comprehensive evaluation, involving six different transfer methods and various LLMs, we offer valuable insights into the strengths and limitations of these transfer methods and LLMs. Our analysis reveals that while LLMs utilizing in-context learning excel in generation tasks, they are often surpassed by smaller fine-tuned models specifically trained for target tasks. Furthermore, our findings highlight significant performance variations dependent on different transfer setups (e.g., demonstrations).\\nMoving forward, our findings suggest the following exciting opportunities for future research in the field of few-shot learning transfer across diverse languages:\\nImprove multilingual instruction tuning. Although instruction tuning can be beneficial for both zero-shot transfer, certain models, such as mT0, may become overly specialized for zero-shot instruction-tuning scenarios, leading to lower average few-shot performance than the optimal zeroshot performance. Although these models demonstrate impressive zero-shot performance, even on tasks they haven't encountered before (such as XCOPA), they face challenges when it comes to tasks that involve generating outputs in less commonly used formats (like structured predictions). We believe that developing multilingual instructionfollowing models capable of effectively utilizing both instructions and demonstrations is crucial. Recent studies demonstrate that incorporating both instructions and demonstrations during instructiontuning on English data can enhance the model's performance (Chung et al., 2022), allowing it to learn within context (Min et al., 2022a). This type of training may potentially mitigate the issue of overfitting to specific formats. Hence, it is necessary to explore various instruction-tuning setups to further improve few-shot in-context learning, with a focus on cross-lingual transfer.\\nAdditionally, while high-quality humantranslated instructions are effective, numerous instruction repositories are still dominated by English instructions.\\nTherefore, community efforts to increase the availability of multilingual instructions may assist in the development of more generalizable multilingual large-language models. Overcome data scarcity using LLMs. Our research reveals that smaller task-specific fine-tuned models, with intermediate training in English, can still outperform ChatGPT on discriminative tasks that require strict output formats. Conversely, Chat-GPT outperforms fine-tuned models on tasks that necessitate more open-ended generations, such as question generation. In recent studies, Instruct-GPT (Ouyang et al., 2022) has exhibited the ability to generate high-quality generations in English, even outperforming humans on some tasks (Goyal et al., 2022). This impressive capacity for flexible generations has prompted active investiga-tions into generating training instances from such LLMs, which have predominantly focused on English (Wang et al., 2022a;Honovich et al., 2022). Some preliminary attempts have been made to explore task-specific data generation in certain target tasks, such as question answering (Agrawal et al., 2022). However, there remains limited exploration on how to generate diverse task instructions and outputs for a variety of typologically diverse languages. We believe that using LLMs to generate data offers a promising solution to obtaining more annotated data for under-represented languages.\\nUnderstand transfer dynamics in cross-lingual in-context learning. The impact of various instructions and demonstrations has been extensively examined in the context of English in-context learning, highlighting critical concerns such as sensitivity to prompt order (Lu et al., 2022) and/or motivating methods for identifying optimal demonstrations (Su et al., 2022). This research has found that demonstrations or instructions that are optimal for one model may not necessarily result in the best performance for another model. We anticipate that our benchmark will inspire and assist in further research into the relationship between language and instruction/demonstration for cross-lingual incontext learning.\\nFairness beyond languages: underrepresented variants, dialects, and cross-cultural NLP. Many of the diverse world languages are often excluded in widely used cross-lingual evaluation benchmarks, where recent papers show strong cross-lingual transfer capabilities. However, through our comprehensive analysis, we have discovered that even the most advanced LLMs currently available still face difficulties when dealing with less-represented languages. The most competitive instruction-tuned models, ChatGPT or mT0, show significant performance declines when it comes to indigenous languages, reaching a level akin to a random baseline.\\nWe advocate for conducting more studies on diverse local languages, including under-represented languages and their dialects, as emphasized in previous works such as Aji et al. (2022); Kakwani et al. (2020). We note that datasets in such languages are often translated from English (Yu et al., 2022), which may introduce translation biases (Artetxe et al., 2020b) and fail to capture the linguistic nuances and interests of native speakers (Clark et al., 2020;Asai et al., 2021). To address these challenges, it is important that further work be done to develop cross-cultural Natural Language Processing (Hershcovich et al., 2022).\\nExpand evaluations to complex tasks. Most recent research on multilingual in-context learning predominantly focuses on discriminative tasks (Muennighoff et al., 2022;Ahuja et al., 2023) or translation tasks (Lin et al., 2021). Further exploration can expand these evaluations to more diverse and complex tasks, such as MTOP (Li et al., 2021) or MGMS8K (Shi et al., 2023), or knowledgeintensive tasks (Asai et al., 2021) as new multilingual benchmarks are developed.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"As the first step toward standardized evaluation for few-shot cross-lingual transfer, BUFFET focuses on popular discriminative tasks and some generative tasks. It does not include many datasets that require complex reasoning tasks, as noted above. Since our main focus is to benchmark different LLMs and learning methods in a comparable format, we do not explore sophisticated prompting methods, which can further boost performance. We anticipate that BUFFET will encourage the LLM community to explore new methods to further improve in-context learning beyond English. We use instructions translated by the NLLB (Costajussà et al., 2022) for TARGET ICL; such machinetranslated instructions are prone to errors, especially in less-represented languages, that can affect the final performance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ethics Statement\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_7\",\n \"tab_9\"\n ],\n \"text\": \"While there has been significant research on incontext learning with LLMs, most of the focus has been on the English language. This raises questions about the applicability of findings from English few-shot NLP to few-shot cross-lingual transfer scenarios. To address this gap, BUFFET aims to provide a comprehensive and less biased evaluation framework. However, it is important to note that our benchmark dataset currently covers only 57 out of the approximately 6,000 world languages. Moreover, we do not specifically focus on finer-grained language varieties and dialects that are commonly spoken by underrepresented populations. In light of these limitations, we encourage future research to explore the effectiveness and limitations of widelyused transfer methods in a more diverse range of languages. This will help us gain a deeper understanding of the generalizability of transfer learning techniques across different linguistic contexts.\\nList of the languages. We show the list of the 55 languages included in BUFFET in Table 5. BUFFET covers 25 different language families, and also shows geographical diversities.\\nExamples. Table 6 shows the input and output examples in BUFFET. We reformulate all o the tasks with diverse formats into the same text-to-text format.\\nInstructions. The full list of the instructions written in English is available in Table 7.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"List of the datasets with languages included.\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_10\"\n ],\n \"text\": \"Table 8 shows the full list of the datasets with language names included in BUFFET.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.2 BUFFET-Light\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_11\"\n ],\n \"text\": \"Task and language decisions. The goal of building the BUFFET-Light subset is to enable quick multilingual evaluation without losing the language and task diversity in the original BUFFET. To this end, we filter BUFFET so that we evaluate between 3 and 7 languages per task, and each language is included in at most three tasks. 11 This design choice allows us to consider 31 diverse languages across all tasks in BUFFET while reducing the number of evaluation settings by 66%.\\nFinal list of BUFFET-light. The full list of tasks and languages in BUFFET are in Table 9.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B More Experimental Details\",\n \"publication_ref\": [\n \"b54\"\n ],\n \"table_ref\": [],\n \"text\": \"Fine-tuning. For ENGLISH FT, we limit the number of English training samples to 100,000 and fine-tune mt5-base (Xue et al., 2021) for 3 epochs. For the ENGLISH FT baseline, we transfer this model directly to new languages, while for ENG.+TGT. FT, we initialize the model checkpoint with the trained model weight and further fine-tune a model on few-shot samples for 300 epochs. In-context learning. We set the maximum token length to 15 except for XLSUM and TYDIQA-QG.\\nFor XLSUM, we set the maximum token length to 100, and for TYDIQA-QG, we set the maximum token length to 50. We use greedy decoding throughout the experiments. For BLOOM-based model evaluations, we use a single RTX-100 GPU \\nᄅ ᅩ ᆺᄃ ᅦᄂ ᅳ ᆫ ᄋ ᅵᄇ ᅥ ᆫ ᄉ ᅡ ᆼᄇ ᅡ ᆫᄀ ᅵ ᄎ ᅢᄋ ᅭ ᆼᄀ ᅪ ᄀ ᅪ ᆫᄅ ᅧ ᆫᄒ ᅢ ᄀ ᅮᄌ ᅵ ᆨᄌ ᅡᄃ ᅳ ᆯᄋ ᅦᄀ ᅦ ᄉ ᅵ ᆯᄌ ᅵ ᆯᄌ ᅥ ᆨᄋ ᅵ ᆫ ᄃ ᅩᄋ ᅮ ᆷᄋ ᅵ ᄃ ᅬ ᆯ ᄉ ᅮ ᄋ ᅵ ᆻᄃ ᅩᄅ ᅩ ᆨ ᄃ ᅡᄋ ᅣ ᆼᄒ ᅡ ᆫ ᄇ ᅡ ᆼᄇ ᅥ ᆸᄋ ᅳᄅ ᅩ ᄌ ᅥ ᆼᄇ ᅩᄌ ᅦᄀ ᅩ ᆼ ᄒ ᅪ ᆯᄃ ᅩ ᆼᄋ ᅳ ᆯ ᄀ ᅡ ᆼᄒ ᅪᄒ ᅡ ᆯ ᄀ ᅨᄒ ᅬ ᆨᄋ ᅵᄃ ᅡ. [SEP] hypothesis: ᄅ ᅩ ᆺᄃ ᅦᄂ ᅳ ᆫ ᄋ ᅥᄄ ᅥᄒ ᅡ ᆫ ᄌ ᅥ ᆼᄇ ᅩᄃ ᅩ ᄌ ᅦᄀ ᅩ ᆼᄒ ᅡᄌ ᅵ ᄋ ᅡ ᆭᄋ ᅳ ᆯ ᄀ ᅨᄒ ᅬ ᆨᄋ ᅵᄃ ᅡ.\\ncontradiction Table 6: The input and output examples in BUFFET. We show one example from one dataset per task. Due to the long input length, we do not include a summarization example in this table.\\nwith 24 GB GPU memory. We use int8bit quantization to avoid GPU out-of-memory errors. To evaluate mT5 and mT0, we use TPU v3-8.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C Detailed BUFFET Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This section includes the full list of the experimental results. Overall results on the full BUFFET are available in Table 10, and Figure 4 summarize overall performance across the eight tasks, on the BUFFET-Light subset. The overall trends on BUFFET-light remain the same as the original BUFFET. This indicates BUFFET-Light is a reliable and more efficient alternative for holistic evaluations for few-shot cross-lingual transfer. Below, we present the performance breakdown for each dataset. PAWS-X In this task you are given a sentence pair that has high lexical overlap. If the sentences have the same meaning and are just paraphrases of each other label them as duplicate, if not label them as not_duplicate\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"SENTIMENT\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this task, you're given a review from Amazon. Your task is to generate a rating for the product. The rating is extremely negative, negative, neutral, positive, and really positive.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"XCOPA\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this task you are given a premise and two alternatives (A) and (B). You must choose the alternative that is more plausibly the cause or effect of the situation described by the premise.\\nXWINOGRAD Replace the _ in the above sentence with the correct option QA Read the given passage and answer a question about the information present in the passage.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"NER\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Given the following sentence, indicate the name entities (i.e., the real-world objects such as a person, location, organization, etc. that can be denoted with a proper name) such as \\\"New York Times\\\". For each word of a named-entity, indicate their type \\\"location\\\" or \\\"organization\\\" or \\\"person\\\".\\nSUMMARIZATION In this task, you are given an article. Your task is to summarize the article in a sentence.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"QG\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This task is about reading the given passage and constructing a question about the information present in the passage. , it, ja, vi, qu, be, gu, et, th, or, kn, fi, gn, ru, el, ur, es, hi, te, as, sw, pa, bg, ml, tr, fa, id, ko, mr, de, ar, bn, zh MASAKHANER amh, hau, ibo, kin, luo, pcm, swa, wol, yor SUMMARIZATION XLSUM (english, ) ta, vi, id, tr, ja, th, bn, ar, en, es, fa, zh, sw QG TYDIQA-QG (en,) ar, be, fi, id, sw, ko, ru, te Table 10: Overall experiment results on BUFFET. The blue-colored rows are instruction-tuned models, and we added * symbols next to the scores for the tasks on which the models have been trained. \\\"Random\\\" shows random baseline performance. Bold fonts indicate the best results for each task, among the models that are not directly trained on the task. When ChatGPT achieves the best results, we also note the second-best number from the models that are not trained on the task, acknowledging the possibility that ChatGPT may have encountered a similar task during training.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.1 NLI\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_13\",\n \"tab_2\",\n \"tab_14\"\n ],\n \"text\": \"Table 11 shows the full results on AMERICASNLI.\\nTable 12 shows the full results on XNLI. Table 13 presents the full results on the other three entailment datasets annotated in each language, KLU-ENLI, OCNLI, and PARSINLUENTAILMENT. On XNLI, ENGLISH FT (zero-shot transfer) shows strong performance and often outperforms ENG.+TGT. FT (few-shot transfer). Among ICL baselines, mT0 ZICL shows the best macro performance on XNLI. However, on AMERICASNLI, all methods struggle, while ENG.+TGT. FT shows the best macro performance on AMERICAS NLI. The performance gap between ENGLISH FT and ENG.+TGT. FT get significantly larger, with the largest gap in Aymara (5.5%). Despite its strong performance on XNLI, mT0 ZICL struggles in AMERICAS NLI (33.7% on average).\\nWhile mT0 ZICL shows robust performance across XNLI languages, ChatGPT shows a large performance gap between higher-resource languages and low-resource languages (57% in Greek v.s. 33% Urdu).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.2 Paraphrase Detection\",\n \"publication_ref\": [\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"The results on PAWS-X results are available in Table14. ENG. FT shows the best performance on this task among non-instruction-tuned models. We hypothesize that as the languages included in PAWS-X are all relatively well-represented languages and the task is relatively simple, ENG. FT, which is not trained in the target languages, can achieve high performance. mT0 ZICL shows quite high performance, likely because the model is trained on PAWS-X (Muennighoff et al., 2022).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.3 Sentiment Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_16\"\n ],\n \"text\": \"The experimental results on AMAZON REVIEW MULTI and INDIC SENTIMENT are available in Tables 15 and16. On both datasets, all models yield high accuracy across languages, except for mT5 ZEICL.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.4 Commonsense\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\",\n \"tab_18\"\n ],\n \"text\": \"XCOPA. The experimental results on XCOPA are available in 33.7 33.5 34.6 33.2 33.6 33.3 33.5 33.3 34.3 34.0 33.6 TICL mT5 33.3 33.3 32.8 33.3 33.6 33.2 33.2 33.3 33.3 33.3 33.3 TICL BLOOMZ 33.4 33.3 33.5 33.7 33.3 33.3 33.8 33.4 33.3 33.3 33.3 TICL mT0 33.4 33.6 33.3 33.3 33.3 33.3 33.3 33.3 33.3 33.3 Table 12: Model performance on XNLI. We report the average of the three few-shot samples.\\nlanguages. ChatGPT achieves particularly notable performance in Italian (91.2%). On the other hand, all of the fine-tuning-based methods struggle, as the small size of the source datasets in English. This result indicates that for a task that lacks a large-scale training dataset even in high-resource Table 16: Model performance on INDIC SENTIMENT. We report the average of the three few-shot samples.\\nlanguages, LLMs using in-context learning may often result in higher performance. We observed that mT0 ENGLISH FT faces difficulties when applied to XCOPA. This could be attributed to the limited size of the XCOPA English set, which might not provide enough data for a smaller mT5-base model to acquire comprehensive task knowledge.\\nXWINOGRAD. The experimental results on XWINOGRAD are available in Table 18. Similar to XCOPA, on XWINOGRAD, fine-tuning-based methods often struggle, while in-context learning with competitive LLMs yields strong performance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.5 Question Answering\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"TYDIQA experimental results are available in Table 19. Both the fine-tuning and ICL methods exhibit commendable performance on this particular task. It is intriguing to note that both mT0 and BLOOMZ demonstrate relatively lower efficacy in Korean, Finnish, and Russian. This can be at-tributed to the fact that these languages were not included during the pretraining phase.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C.6 Named Entity Recognition\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_20\",\n \"tab_2\"\n ],\n \"text\": \"WIKIANN. Table 20 contains the results for WIKIANN. We specifically present the few-shot results since we discovered that zero-shot baselines consistently exhibit extremely poor performance, often close to zero, primarily because generating the answer in the precise output format proves to be challenging.\\nIt's important to acknowledge that the BUFFET-Light WIKIANN subset comprises languages that are relatively high-resource, which could potentially lead to an overestimation of ChatGPT's performance. When comparing the best fine-tuning method with ChatGPT in the BUFFET-light languages, they generally perform competitively, with the exception of Finnish. Table 22: Model performance on TYDIQA-QG. We report the average of the three few-shot samples.\\nTARGET ICL shows large improvements from EN-GLISH ICL, which has not been observed in other tasks. When instructions in the target language are given, ChatGPT almost always generates a summary in the target language. Among non-instruction-tuned models, ENG.+TGT. FT yields the highest average performance. It should be noted that mT0 and BLOOMZ are trained on XLSUM. Nevertheless, their performance in some languages remains low.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D More Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_6\"\n ],\n \"heading\": \"D.1 Performance across Languages\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Figure 8 shows performance across languages on the three tasks, NLI, NER, and QA, adding two more LLMs: BLOOMZ and mT0.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D.2 Variances of Different k-shots\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In Section 5, we show that different sets of demonstrations can cause significant performance differ-ences. We provide the full visualization results across different tasks.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D.3 Variances of the Varying Number of k\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We provide the full experimental results with a different number of k. We evaluate BLOOM EN-GLISH ICL, BLOOMZ ENGLISH ICL and mT5-ENG.+TGT. FINE-TUNING and mT0 ENGLISH ICL experimental results on AMAZON REVIEW, TYDIQA, TYDIQA-AG, WIKIANN,and in Figures 9,10,11 and 12,respectively. AMAZON REVIEW. On AMAZON REVIEW, All models except for BLOOM (pretraining only) show competitive zero-shot performance. BLOOM EN-GLISH ICL benefits from few-shot demonstrations while mT0 ENGLISH ICL exhibit performance deterioration as adding more demonstrations across languages.\\nTYDIQA. Among ENGLISH ICL baselines, mT0 shows strong performance up to four demon- strations, although their performance gets really low once more demonstrations are added. Similar deterioration happens in BLOOMZ. On the contrary, BLOOM performance improves as more shots are added. Despite using only 32 shots.\\nTYDIQA-QG. Unlike in AMAZON REVIEW or TYDIQA, BLOOMZ ENGLISH ICL shows performance improvements with more demonstrations in Arabic and Bengali, reaching the highest QG performance in Bengali with four demonstrations. On the contrary, both BLOOM and BLOOMZ show poor QG performance in Korean and Russian, pos-sibly due to the lack of those languages during pretraining.\\nWIKIANN. On WikiANN, all of the models gain performance improvements by adding at least one demonstration, possibly due to the difficulty of learning the exact output format expected given the instruction only. As in other datasets, mT0 reaches its highest performance with four demonstrations. mT5 ENG.+TGT. FT exhibits performance drops with one shot, possibly due to their overfit to the single example.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D.4 Variances of Different Instructions\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_24\"\n ],\n \"text\": \"We investigate the effectiveness of different English instructions on question generation tasks for TYDIQA in 0 and 4-shot setting using mT0 and BLOOM as base models in Table 24. We compare four relevant instructions and one irrelevant instruction (an instruction for AMAZON REVIEW).\\nIn a zero-shot setting, instruction does not make much difference for both instruction-tuned and noninstruction-tuned models, since irrelevant instructions are sometimes better than the relevant prompt.\\nIn a four-shot setting, whether the instruction is relevant does not make a huge difference for BLOOM, and we observed that random seeds impact the performance more, yet the performances do suffer a sharp loss if we are using irrelevant instruction in the instruction-tuned model. We also discovered that different models might favor different instructions for different languages, for example, in Swahili, 4-shot BLOOM favors the first instruction, while mT0 favors the fourth instruction.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D.5 Qualitative Results for Generation Tasks\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_25\"\n ],\n \"text\": \"Table 25 shows some qualitative results of Chat-GPT ENGLISH ICL and TARGET TCL on XLSUM and TYDIQA. Given English instruction, ChatGPT often generates summaries in English, rather than in the article language. On the other hand, such cross-lingual generation behaviors don't occur in QA tasks, and the model's predictions with TAR-GET ICL and ENGLISH ICL exhibit high overlap with each other. We hypothesize that ChatGPT's this cross-lingual summarization behavior can be related to their private training corpus, and future work can further investigate this issue. This task is about reading the given passage and constructing a question about the information present in the passage. Construct a question in such a way that (i) it is unambiguous, (ii) it is answered from the passage, (iii) its answer is unique (iv) its answer is a continuous text span from the paragraph. Avoid creating questions that (i) can be answered correctly without actually understanding the paragraph and (ii) uses the same words or phrases given in the passage.\\n5 In this task, you are given a review from Amazon. Your task is to generate a rating for the product on a scale of 1-5 based on the review. The rating means -2: extremely poor, 1: poor, 0: neutral, 1: good, 2: extremely good, 3: mixed and -3: no sentiment expressed. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgements\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This research was supported by NSF IIS-2044660, ONR N00014-18-1-2826, ONR MURI N00014-18-1-2670, DARPA MCS program through NIWC Pacific (N66001-19-2-4031), and Allen Distinguished Award. AA is supported by the IBM fellowship. We are grateful to Orevaoghene Ahia for her help with ChatGPT evaluations. We thank our volunteer translators, Joongwon Kim, Usharani Injeti, and Sven Dorkenwald, for their help with translating instructions into different languages. Finally, we extend our appreciation to Jonathan H. Clark, Orevaoghene Ahia, Sandy Kaplan, and UW NLP researchers for their feedback on this draft. Shailaja Keyur Sampat, Siddhartha Mishra, Sujan Reddy A, Sumanta Patro, Tanay Dixit, and Xudong Shen. 2022b. Super-NaturalInstructions: Generalization via declarative instructions on 1600+ NLP tasks.\\nIn Proceedings of Empirical Methods in Natural Language Processing.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Appendix A Benchmark Details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"A.1 BUFFET Constructions This section will provide further details of the BUFFET construction process. \"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":""},"doi":{"kind":"string","value":"10.1162/tacl_a_00416"},"references":{"kind":"list like","value":[{"authors":"David Ifeoluwa Adelani; Jade Abbott; Graham Neubig; D' Daniel; Julia Souza; Constantine Kreutzer; Chester Lignos; Happy Palen-Michel; Shruti Buzaaba; Sebastian Rijhwani; Stephen Ruder; Israel Mayhew; Shamsuddeen H Abebe Azime; Chris Chinenye Muhammad; Joyce Emezue; Perez Nakatumba-Nabende; Aremu Ogayo; Catherine Anuoluwapo; Derguene Gitau; Jesujoba Mbaye; Seid Alabi; Tajuddeen Muhie Yimam; Ignatius Rabiu Gwadabe; Ezeani; Andre Rubungo; Jonathan Niyongabo; Verrah Mukiibi; Iroro Otiende; Davis Orife; Samba David; Tosin Ngom; Paul Adewumi; Mofetoluwa Rayson; Gerald Adeyemi; Emmanuel Muriuki; Chiamaka Anebi; Nkiruka Chukwuneke; Eric Odu; Samuel Peter Wairagala; Clemencia Oyerinde; Tobius Siro; Temilola Saul Bateesa; Yvonne Oloyede; Victor Wambui; Deborah Akinode; Maurice Nabagereka; Ayodele Katusiime; Awokoya; Mboup Mouhamadane; Dibora Gebreyohannes; Henok Tilaye; Kelechi Nwaike; Degaga Wolde; Abdoulaye Faye; Blessing Sibanda; Orevaoghene Ahia; F P Bonaventure; Kelechi Dossou; Thierno Ogueji; Diop Ibrahima; Abdoulaye Diallo; Adewale Akinfaderin; Tendai Marengereke; Salomey Osei","journal":"","ref_id":"b0","title":"MasakhaNER: Named Entity Recognition for African Languages","year":"2021"},{"authors":"Divyanshu Aggarwal; Vivek Gupta; Anoop Kunchukuttan","journal":"Proceedings of Empirical Methods in Natural Language Processing","ref_id":"b1","title":"IndicXNLI: Evaluating multilingual inference for Indian languages","year":"2022"},{"authors":"Priyanka Agrawal; Chris Alberti; Fantine Huot; Joshua Maynez; Ji Ma; Sebastian Ruder; Kuzman Ganchev; Dipanjan Das; Mirella Lapata","journal":"","ref_id":"b2","title":"Qameleon: Multilingual qa with only 5 examples","year":"2022"},{"authors":"Kabir Ahuja; Rishav Hada; Millicent Ochieng; Prachi Jain; Harshita Diddee; Samuel Maina; Tanuja Ganu; Sameer Segal; Maxamed Axmed; Kalika Bali","journal":"","ref_id":"b3","title":"Mega: Multilingual evaluation of generative ai","year":"2023"},{"authors":"Alham Fikri Aji; Genta Indra Winata; Fajri Koto; Samuel Cahyawijaya; Ade Romadhony; Rahmad Mahendra; Kemal Kurniawan; David Moeljadi; Radityo Eko Prasojo; Timothy Baldwin; Jey Han Lau; Sebastian Ruder","journal":"","ref_id":"b4","title":"One country, 700+ languages: NLP challenges for underrepresented languages and dialects in Indonesia","year":"2022"},{"authors":"Mikel Artetxe; Sebastian Ruder; Dani Yogatama","journal":"","ref_id":"b5","title":"On the cross-lingual transferability of monolingual representations","year":"2020"},{"authors":"Mikel Artetxe; Sebastian Ruder; Dani Yogatama; Gorka Labaka; Eneko Agirre","journal":"","ref_id":"b6","title":"A call for more rigor in unsupervised cross-lingual learning","year":"2020"},{"authors":"Venkat Arun; Hari Balakrishnan","journal":"","ref_id":"b7","title":"Copa: Practical Delay-Based congestion control for the internet","year":"2018"},{"authors":"Akari Asai; Jungo Kasai; Jonathan Clark; Kenton Lee; Eunsol Choi; Hannaneh Hajishirzi","journal":"","ref_id":"b8","title":"XOR QA: Cross-lingual open-retrieval question answering","year":"2021"},{"authors":"H Stephen; Victor Bach; Zheng-Xin Sanh; Albert Yong; Colin Webson; Raffel; V Nihal; Abheesht Nayak; Taewoon Sharma; M Kim; Thibault Saiful Bari; Fevry","journal":"","ref_id":"b9","title":"Promptsource: An integrated development environment and repository for natural language prompts","year":"2022"},{"authors":"Yejin Bang; Samuel Cahyawijaya; Nayeon Lee; Wenliang Dai; Dan Su; Bryan Wilie; Holy Lovenia; Ziwei Ji; Tiezheng Yu; Willy Chung","journal":"","ref_id":"b10","title":"A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity","year":"2023"},{"authors":"Iz Beltagy; Arman Cohan; Robert Logan; I V ; Sewon Min; Sameer Singh","journal":"","ref_id":"b11","title":"Zero-and few-shot NLP with pretrained language models","year":"2022"},{"authors":"Damian Blasi; Antonios Anastasopoulos; Graham Neubig","journal":"","ref_id":"b12","title":"Systematic inequalities in language technology performance across the world's languages","year":"2022"},{"authors":"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell","journal":"","ref_id":"b13","title":"Language models are few-shot learners","year":"2020"},{"authors":"Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma","journal":"","ref_id":"b14","title":"Scaling instruction-finetuned language models","year":"2022"},{"authors":"Jonathan H Clark; Eunsol Choi; Michael Collins; Dan Garrette; Tom Kwiatkowski; Vitaly Nikolaev; Jennimaria Palomaki","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b15","title":"TyDi QA: A benchmark for information-seeking question answering in typologically diverse languages","year":"2020"},{"authors":"Alexis Conneau; Kartikay Khandelwal; Naman Goyal; Vishrav Chaudhary; Guillaume Wenzek; Francisco Guzmán; Edouard Grave; Myle Ott; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b16","title":"Unsupervised cross-lingual representation learning at scale","year":"2020"},{"authors":"Alexis Conneau; Ruty Rinott; Guillaume Lample; Adina Williams; Samuel Bowman; Holger Schwenk; Veselin Stoyanov","journal":"","ref_id":"b17","title":"XNLI: Evaluating crosslingual sentence representations","year":"2018"},{"authors":"Alexis Conneau; Shijie Wu; Haoran Li; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b18","title":"Emerging cross-lingual structure in pretrained language models","year":"2020"},{"authors":"James Marta R Costa-Jussà; Onur Cross; Maha Çelebi; Kenneth Elbayad; Kevin Heafield; Elahe Heffernan; Janice Kalbassi; Daniel Lam; Jean Licht; Maillard","journal":"","ref_id":"b19","title":"No language left behind: Scaling humancentered machine translation","year":"2022"},{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"","ref_id":"b20","title":"Bert: Pre-training of deep bidirectional transformers for language understanding","year":"2019"},{"authors":"Abteen Ebrahimi; Manuel Mager; Arturo Oncevay; Vishrav Chaudhary; Luis Chiruzzo; Angela Fan; John Ortega; Ricardo Ramos; Annette Rios; Ivan Vladimir Meza; Gustavo Ruiz; Elisabeth Giménez-Lugo; Graham Mager; Alexis Neubig; Rolando Palmer; Thang Coto-Solano; Katharina Vu; Kann","journal":"","ref_id":"b21","title":"AmericasNLI: Evaluating zero-shot natural language understanding of pretrained multilingual models in truly low-resource languages","year":"2022"},{"authors":"Tanya Goyal; Junyi ; Jessy Li; Greg Durrett","journal":"","ref_id":"b22","title":"News summarization and evaluation in the era of gpt-3","year":"2022"},{"authors":"Tahmid Hasan; Abhik Bhattacharjee; Md Saiful Islam; Kazi Mubasshir; Yuan-Fang Li; Yong-Bin Kang; M Sohel Rahman; Rifat Shahriyar","journal":"","ref_id":"b23","title":"XLsum: Large-scale multilingual abstractive summarization for 44 languages","year":"2021"},{"authors":"A Michael; David Hedderich; Dawei Adelani; Jesujoba Zhu; Udia Alabi; Dietrich Markus; Klakow","journal":"","ref_id":"b24","title":"Transfer learning and distant supervision for multilingual transformer models: A study on African languages","year":"2020"},{"authors":"Daniel Hershcovich; Stella Frank; Heather Lent; Mostafa Miryam De Lhoneux; Stephanie Abdou; Emanuele Brandl; Laura Cabello Bugliarello; Ilias Piqueras; Ruixiang Chalkidis; Constanza Cui; Katerina Fierro; Phillip Margatina; Anders Rust; Søgaard","journal":"","ref_id":"b25","title":"Challenges and strategies in crosscultural NLP","year":"2022"},{"authors":"Or Honovich; Thomas Scialom; Omer Levy; Timo Schick","journal":"","ref_id":"b26","title":"Unnatural instructions: Tuning language models with (almost) no human labor","year":"2022"},{"authors":"Hai Hu; Kyle Richardson; Liang Xu; Lu Li; Sandra Kübler; Lawrence Moss","journal":"","ref_id":"b27","title":"OCNLI: Original Chinese Natural Language Inference","year":"2020"},{"authors":"Junjie Hu; Sebastian Ruder; Aditya Siddhant; Graham Neubig; Orhan Firat; Melvin Johnson","journal":"","ref_id":"b28","title":"Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalisation","year":"2020"},{"authors":"Haoyang Huang; Tianyi Tang; Dongdong Zhang; Wayne Xin Zhao; Ting Song; Yan Xia; Furu Wei","journal":"","ref_id":"b29","title":"Not all languages are created equal in llms: Improving multilingual capability by cross-lingualthought prompting","year":"2023"},{"authors":"Pratik Joshi; Sebastin Santy; Amar Budhiraja; Kalika Bali; Monojit Choudhury","journal":"","ref_id":"b30","title":"The state and fate of linguistic diversity and inclusion in the NLP world","year":"2020"},{"authors":"Divyanshu Kakwani; Anoop Kunchukuttan; Satish Golla; N C Gokul; Avik Bhattacharyya; M Mitesh; Pratyush Khapra; Kumar","journal":"","ref_id":"b31","title":"IndicNLPSuite: Monolingual corpora, evaluation benchmarks and pre-trained multilingual language models for Indian languages","year":"2020"},{"authors":"Phillip Keung; Yichao Lu; György Szarvas; Noah A Smith","journal":"","ref_id":"b32","title":"The multilingual Amazon reviews corpus","year":"2020"},{"authors":"Daniel Khashabi; Arman Cohan; Siamak Shakeri; Pedram Hosseini; Pouya Pezeshkpour; Malihe Alikhani; Moin Aminnaseri; Marzieh Bitaab; Faeze Brahman; Sarik Ghazarian; Mozhdeh Gheini; Arman Kabiri; Rabeeh Karimi Mahabagdi; Omid Memarrast; Ahmadreza Mosallanezhad; Erfan Noury; Shahab Raji; Mohammad Sadegh Rasooli; Sepideh Sadeghi; Erfan Sadeqi Azer; Niloofar Safi Samghabadi; Mahsa Shafaei; Saber Sheybani; Ali Tazarv; Yadollah Yaghoobzadeh","journal":"","ref_id":"b33","title":"ParsiNLU: A Suite of Language Understanding Challenges for Persian","year":"2021"},{"authors":"Anne Lauscher; Ivan Ravishankar; Goran Vulić; Glavaš","journal":"","ref_id":"b34","title":"From zero to hero: On the limitations of zero-shot language transfer with multilingual Transformers","year":"2020"},{"authors":"Haoran Li; Abhinav Arora; Shuohui Chen; Anchit Gupta; Sonal Gupta; Yashar Mehdad","journal":"","ref_id":"b35","title":"MTOP: A comprehensive multilingual task-oriented semantic parsing benchmark","year":"2021"},{"authors":"Yaobo Liang; Nan Duan; Yeyun Gong; Ning Wu; Fenfei Guo; Weizhen Qi; Ming Gong; Linjun Shou; Daxin Jiang; Guihong Cao; Xiaodong Fan; Ruofei Zhang; Rahul Agrawal; Edward Cui; Sining Wei; Taroon Bharti; Ying Qiao; Jiun-Hung Chen; Winnie Wu; Shuguang Liu; Fan Yang; Daniel Campos; Rangan Majumder; Ming Zhou","journal":"","ref_id":"b36","title":"XGLUE: A new benchmark datasetfor cross-lingual pre-training, understanding and generation","year":"2020"},{"authors":"Victoria Xi; Todor Lin; Mikel Mihaylov; Tianlu Artetxe; Shuohui Wang; Daniel Chen; Myle Simig; Naman Ott; Shruti Goyal; Jingfei Bhosale; Ramakanth Du; Sam Pasunuru; Punit Shleifer; Vishrav Singh Koura; Brian O' Chaudhary; Jeff Horo; Luke Wang; Zornitsa Zettlemoyer; Mona Kozareva; Veselin Diab; Xian Stoyanov; Li","journal":"","ref_id":"b37","title":"Few-shot learning with multilingual language models","year":"2021"},{"authors":"Yao Lu; Max Bartolo; Alastair Moore; Sebastian Riedel; Pontus Stenetorp","journal":"","ref_id":"b38","title":"Fantastically ordered prompts and where to find them: Overcoming fewshot prompt order sensitivity","year":"2022"},{"authors":"Sewon Min; Mike Lewis; Luke Zettlemoyer; Hannaneh Hajishirzi","journal":"","ref_id":"b39","title":"MetaICL: Learning to learn in context","year":"2022"},{"authors":"Sewon Min; Xinxi Lyu; Ari Holtzman; Mikel Artetxe; Mike Lewis; Hannaneh Hajishirzi; Luke Zettlemoyer","journal":"","ref_id":"b40","title":"Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of Empirical Methods in Natural Language Processing","year":"2022"},{"authors":"Swaroop Mishra; Daniel Khashabi; Chitta Baral; Hannaneh Hajishirzi","journal":"","ref_id":"b41","title":"Cross-task generalization via natural language crowdsourcing instructions","year":"2022"},{"authors":"Niklas Muennighoff; Thomas Wang; Lintang Sutawika; Adam Roberts; Stella Biderman; Teven Le Scao; M Saiful Bari; Sheng Shen; Zheng-Xin Yong; Hailey Schoelkopf","journal":"","ref_id":"b42","title":"Crosslingual generalization through multitask finetuning","year":"2022"},{"authors":"Long Ouyang; Jeffrey Wu; Xu Jiang; Diogo Almeida; Carroll Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray","journal":"","ref_id":"b43","title":"Training language models to follow instructions with human feedback","year":"2022"},{"authors":"Xiaoman Pan; Boliang Zhang; Jonathan May; Joel Nothman; Kevin Knight; Heng Ji","journal":"","ref_id":"b44","title":"Cross-lingual name tagging and linking for 282 languages","year":"2017"},{"authors":"Sungjoon Park; Jihyung Moon; Sungdong Kim; Won Ik Cho; Jiyoon Han; Jangwon Park; Chisung Song; Junseong Kim; Yongsook Song; Taehwan Oh","journal":"","ref_id":"b45","title":"Klue: Korean language understanding evaluation","year":"2021"},{"authors":"Maria Edoardo; Goran Ponti; Olga Glavaš; Qianchu Majewska; Ivan Liu; Anna Vulić; Korhonen","journal":"","ref_id":"b46","title":"XCOPA: A multilingual dataset for causal commonsense reasoning","year":"2020"},{"authors":"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu","journal":"Journal of Machine Learning Research","ref_id":"b47","title":"Exploring the limits of transfer learning with a unified text-to-text transformer","year":"2020"},{"authors":"Pranav Rajpurkar; Jian Zhang; Konstantin Lopyrev; Percy Liang","journal":"","ref_id":"b48","title":"SQuAD: 100,000+ questions for machine comprehension of text","year":"2016"},{"authors":"Sebastian Ruder; Noah Constant; Jan Botha; Aditya Siddhant; Orhan Firat; Jinlan Fu; Pengfei Liu; Junjie Hu; Dan Garrette; Graham Neubig; Melvin Johnson","journal":"","ref_id":"b49","title":"XTREME-R: Towards more challenging and nuanced multilingual evaluation","year":"2021"},{"authors":"Jason Wei; Maarten Bosma; Y Vincent; Kelvin Zhao; Adams Wei Guu; Brian Yu; Nan Lester; Andrew M Du; Quoc V Dai; Le","journal":"","ref_id":"b50","title":"Finetuned language models are zero-shot learners","year":"2022"},{"authors":"Adina Williams; Nikita Nangia; Samuel R Bowman","journal":"","ref_id":"b51","title":"A broad-coverage challenge corpus for sentence understanding through inference","year":"2017"},{"authors":"Shijie Wu; Mark Dredze","journal":"","ref_id":"b52","title":"Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT","year":"2019"},{"authors":"Dongling Xiao; Han Zhang; Yukun Li; Yu Sun; Hua Hao Tian; Haifeng Wu; Wang","journal":"","ref_id":"b53","title":"Ernie-gen: An enhanced multi-flow pre-training and fine-tuning framework for natural language generation","year":"2021"},{"authors":"Linting Xue; Noah Constant; Adam Roberts; Mihir Kale; Rami Al-Rfou; Aditya Siddhant; Aditya Barua; Colin Raffel","journal":"","ref_id":"b54","title":"mt5: A massively multilingual pre-trained text-to-text transformer","year":"2021"},{"authors":"Yinfei Yang; Yuan Zhang; Chris Tar; Jason Baldridge","journal":"","ref_id":"b55","title":"PAWS-X: A cross-lingual adversarial dataset for paraphrase identification","year":"2019"},{"authors":"Qinyuan Ye; Bill Yuchen Lin; Xiang Ren","journal":"","ref_id":"b56","title":"CrossFit: A few-shot learning challenge for crosstask generalization in NLP","year":"2021"},{"authors":"Xinyan Velocity; Yu ; Trina Chatterjee; Akari Asai; Junjie Hu; Eunsol Choi","journal":"Proceedings of Findings of Empirical Methods in Natural Language Processing","ref_id":"b57","title":"Beyond counting datasets: A survey of multilingual dataset construction and necessary resources","year":"2022"},{"authors":"Yuan Zhang; Jason Baldridge; Luheng He","journal":"","ref_id":"b58","title":"Paws: Paraphrase adversaries from word scrambling","year":"2019"},{"authors":"Mengjie Zhao; Yi Zhu; Ehsan Shareghi; Ivan Vulić; Roi Reichart; Anna Korhonen; Hinrich Schütze","journal":"","ref_id":"b59","title":"A closer look at few-shot crosslingual transfer: The choice of shots matters","year":"2021"},{"authors":"Wenhao Zhu; Hongyi Liu; Qingxiu Dong; Jingjing Xu; Lingpeng Kong; Jiajun Chen; Li ; Shujian Huang","journal":"","ref_id":"b60","title":"Multilingual machine translation with large language models: Empirical results and analysis","year":"2023"}],"string":"[\n {\n \"authors\": \"David Ifeoluwa Adelani; Jade Abbott; Graham Neubig; D' Daniel; Julia Souza; Constantine Kreutzer; Chester Lignos; Happy Palen-Michel; Shruti Buzaaba; Sebastian Rijhwani; Stephen Ruder; Israel Mayhew; Shamsuddeen H Abebe Azime; Chris Chinenye Muhammad; Joyce Emezue; Perez Nakatumba-Nabende; Aremu Ogayo; Catherine Anuoluwapo; Derguene Gitau; Jesujoba Mbaye; Seid Alabi; Tajuddeen Muhie Yimam; Ignatius Rabiu Gwadabe; Ezeani; Andre Rubungo; Jonathan Niyongabo; Verrah Mukiibi; Iroro Otiende; Davis Orife; Samba David; Tosin Ngom; Paul Adewumi; Mofetoluwa Rayson; Gerald Adeyemi; Emmanuel Muriuki; Chiamaka Anebi; Nkiruka Chukwuneke; Eric Odu; Samuel Peter Wairagala; Clemencia Oyerinde; Tobius Siro; Temilola Saul Bateesa; Yvonne Oloyede; Victor Wambui; Deborah Akinode; Maurice Nabagereka; Ayodele Katusiime; Awokoya; Mboup Mouhamadane; Dibora Gebreyohannes; Henok Tilaye; Kelechi Nwaike; Degaga Wolde; Abdoulaye Faye; Blessing Sibanda; Orevaoghene Ahia; F P Bonaventure; Kelechi Dossou; Thierno Ogueji; Diop Ibrahima; Abdoulaye Diallo; Adewale Akinfaderin; Tendai Marengereke; Salomey Osei\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"MasakhaNER: Named Entity Recognition for African Languages\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Divyanshu Aggarwal; Vivek Gupta; Anoop Kunchukuttan\",\n \"journal\": \"Proceedings of Empirical Methods in Natural Language Processing\",\n \"ref_id\": \"b1\",\n \"title\": \"IndicXNLI: Evaluating multilingual inference for Indian languages\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Priyanka Agrawal; Chris Alberti; Fantine Huot; Joshua Maynez; Ji Ma; Sebastian Ruder; Kuzman Ganchev; Dipanjan Das; Mirella Lapata\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Qameleon: Multilingual qa with only 5 examples\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Kabir Ahuja; Rishav Hada; Millicent Ochieng; Prachi Jain; Harshita Diddee; Samuel Maina; Tanuja Ganu; Sameer Segal; Maxamed Axmed; Kalika Bali\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Mega: Multilingual evaluation of generative ai\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Alham Fikri Aji; Genta Indra Winata; Fajri Koto; Samuel Cahyawijaya; Ade Romadhony; Rahmad Mahendra; Kemal Kurniawan; David Moeljadi; Radityo Eko Prasojo; Timothy Baldwin; Jey Han Lau; Sebastian Ruder\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"One country, 700+ languages: NLP challenges for underrepresented languages and dialects in Indonesia\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Mikel Artetxe; Sebastian Ruder; Dani Yogatama\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"On the cross-lingual transferability of monolingual representations\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Mikel Artetxe; Sebastian Ruder; Dani Yogatama; Gorka Labaka; Eneko Agirre\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"A call for more rigor in unsupervised cross-lingual learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Venkat Arun; Hari Balakrishnan\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Copa: Practical Delay-Based congestion control for the internet\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Akari Asai; Jungo Kasai; Jonathan Clark; Kenton Lee; Eunsol Choi; Hannaneh Hajishirzi\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"XOR QA: Cross-lingual open-retrieval question answering\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"H Stephen; Victor Bach; Zheng-Xin Sanh; Albert Yong; Colin Webson; Raffel; V Nihal; Abheesht Nayak; Taewoon Sharma; M Kim; Thibault Saiful Bari; Fevry\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Promptsource: An integrated development environment and repository for natural language prompts\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yejin Bang; Samuel Cahyawijaya; Nayeon Lee; Wenliang Dai; Dan Su; Bryan Wilie; Holy Lovenia; Ziwei Ji; Tiezheng Yu; Willy Chung\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Iz Beltagy; Arman Cohan; Robert Logan; I V ; Sewon Min; Sameer Singh\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Zero-and few-shot NLP with pretrained language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Damian Blasi; Antonios Anastasopoulos; Graham Neubig\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Systematic inequalities in language technology performance across the world's languages\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Chung Hyung Won; Le Hou; Shayne Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Scaling instruction-finetuned language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jonathan H Clark; Eunsol Choi; Michael Collins; Dan Garrette; Tom Kwiatkowski; Vitaly Nikolaev; Jennimaria Palomaki\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b15\",\n \"title\": \"TyDi QA: A benchmark for information-seeking question answering in typologically diverse languages\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Alexis Conneau; Kartikay Khandelwal; Naman Goyal; Vishrav Chaudhary; Guillaume Wenzek; Francisco Guzmán; Edouard Grave; Myle Ott; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Unsupervised cross-lingual representation learning at scale\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Alexis Conneau; Ruty Rinott; Guillaume Lample; Adina Williams; Samuel Bowman; Holger Schwenk; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"XNLI: Evaluating crosslingual sentence representations\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Alexis Conneau; Shijie Wu; Haoran Li; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Emerging cross-lingual structure in pretrained language models\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"James Marta R Costa-Jussà; Onur Cross; Maha Çelebi; Kenneth Elbayad; Kevin Heafield; Elahe Heffernan; Janice Kalbassi; Daniel Lam; Jean Licht; Maillard\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"No language left behind: Scaling humancentered machine translation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Bert: Pre-training of deep bidirectional transformers for language understanding\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Abteen Ebrahimi; Manuel Mager; Arturo Oncevay; Vishrav Chaudhary; Luis Chiruzzo; Angela Fan; John Ortega; Ricardo Ramos; Annette Rios; Ivan Vladimir Meza; Gustavo Ruiz; Elisabeth Giménez-Lugo; Graham Mager; Alexis Neubig; Rolando Palmer; Thang Coto-Solano; Katharina Vu; Kann\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"AmericasNLI: Evaluating zero-shot natural language understanding of pretrained multilingual models in truly low-resource languages\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tanya Goyal; Junyi ; Jessy Li; Greg Durrett\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"News summarization and evaluation in the era of gpt-3\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tahmid Hasan; Abhik Bhattacharjee; Md Saiful Islam; Kazi Mubasshir; Yuan-Fang Li; Yong-Bin Kang; M Sohel Rahman; Rifat Shahriyar\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"XLsum: Large-scale multilingual abstractive summarization for 44 languages\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"A Michael; David Hedderich; Dawei Adelani; Jesujoba Zhu; Udia Alabi; Dietrich Markus; Klakow\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Transfer learning and distant supervision for multilingual transformer models: A study on African languages\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Daniel Hershcovich; Stella Frank; Heather Lent; Mostafa Miryam De Lhoneux; Stephanie Abdou; Emanuele Brandl; Laura Cabello Bugliarello; Ilias Piqueras; Ruixiang Chalkidis; Constanza Cui; Katerina Fierro; Phillip Margatina; Anders Rust; Søgaard\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Challenges and strategies in crosscultural NLP\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Or Honovich; Thomas Scialom; Omer Levy; Timo Schick\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Unnatural instructions: Tuning language models with (almost) no human labor\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hai Hu; Kyle Richardson; Liang Xu; Lu Li; Sandra Kübler; Lawrence Moss\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"OCNLI: Original Chinese Natural Language Inference\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Junjie Hu; Sebastian Ruder; Aditya Siddhant; Graham Neubig; Orhan Firat; Melvin Johnson\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalisation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Haoyang Huang; Tianyi Tang; Dongdong Zhang; Wayne Xin Zhao; Ting Song; Yan Xia; Furu Wei\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Not all languages are created equal in llms: Improving multilingual capability by cross-lingualthought prompting\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Pratik Joshi; Sebastin Santy; Amar Budhiraja; Kalika Bali; Monojit Choudhury\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"The state and fate of linguistic diversity and inclusion in the NLP world\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Divyanshu Kakwani; Anoop Kunchukuttan; Satish Golla; N C Gokul; Avik Bhattacharyya; M Mitesh; Pratyush Khapra; Kumar\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"IndicNLPSuite: Monolingual corpora, evaluation benchmarks and pre-trained multilingual language models for Indian languages\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Phillip Keung; Yichao Lu; György Szarvas; Noah A Smith\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"The multilingual Amazon reviews corpus\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Daniel Khashabi; Arman Cohan; Siamak Shakeri; Pedram Hosseini; Pouya Pezeshkpour; Malihe Alikhani; Moin Aminnaseri; Marzieh Bitaab; Faeze Brahman; Sarik Ghazarian; Mozhdeh Gheini; Arman Kabiri; Rabeeh Karimi Mahabagdi; Omid Memarrast; Ahmadreza Mosallanezhad; Erfan Noury; Shahab Raji; Mohammad Sadegh Rasooli; Sepideh Sadeghi; Erfan Sadeqi Azer; Niloofar Safi Samghabadi; Mahsa Shafaei; Saber Sheybani; Ali Tazarv; Yadollah Yaghoobzadeh\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"ParsiNLU: A Suite of Language Understanding Challenges for Persian\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Anne Lauscher; Ivan Ravishankar; Goran Vulić; Glavaš\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"From zero to hero: On the limitations of zero-shot language transfer with multilingual Transformers\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Haoran Li; Abhinav Arora; Shuohui Chen; Anchit Gupta; Sonal Gupta; Yashar Mehdad\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"MTOP: A comprehensive multilingual task-oriented semantic parsing benchmark\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yaobo Liang; Nan Duan; Yeyun Gong; Ning Wu; Fenfei Guo; Weizhen Qi; Ming Gong; Linjun Shou; Daxin Jiang; Guihong Cao; Xiaodong Fan; Ruofei Zhang; Rahul Agrawal; Edward Cui; Sining Wei; Taroon Bharti; Ying Qiao; Jiun-Hung Chen; Winnie Wu; Shuguang Liu; Fan Yang; Daniel Campos; Rangan Majumder; Ming Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"XGLUE: A new benchmark datasetfor cross-lingual pre-training, understanding and generation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Victoria Xi; Todor Lin; Mikel Mihaylov; Tianlu Artetxe; Shuohui Wang; Daniel Chen; Myle Simig; Naman Ott; Shruti Goyal; Jingfei Bhosale; Ramakanth Du; Sam Pasunuru; Punit Shleifer; Vishrav Singh Koura; Brian O' Chaudhary; Jeff Horo; Luke Wang; Zornitsa Zettlemoyer; Mona Kozareva; Veselin Diab; Xian Stoyanov; Li\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Few-shot learning with multilingual language models\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yao Lu; Max Bartolo; Alastair Moore; Sebastian Riedel; Pontus Stenetorp\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Fantastically ordered prompts and where to find them: Overcoming fewshot prompt order sensitivity\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Sewon Min; Mike Lewis; Luke Zettlemoyer; Hannaneh Hajishirzi\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"MetaICL: Learning to learn in context\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Sewon Min; Xinxi Lyu; Ari Holtzman; Mikel Artetxe; Mike Lewis; Hannaneh Hajishirzi; Luke Zettlemoyer\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of Empirical Methods in Natural Language Processing\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Swaroop Mishra; Daniel Khashabi; Chitta Baral; Hannaneh Hajishirzi\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Cross-task generalization via natural language crowdsourcing instructions\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Niklas Muennighoff; Thomas Wang; Lintang Sutawika; Adam Roberts; Stella Biderman; Teven Le Scao; M Saiful Bari; Sheng Shen; Zheng-Xin Yong; Hailey Schoelkopf\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Crosslingual generalization through multitask finetuning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Long Ouyang; Jeffrey Wu; Xu Jiang; Diogo Almeida; Carroll Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray\",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"Training language models to follow instructions with human feedback\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xiaoman Pan; Boliang Zhang; Jonathan May; Joel Nothman; Kevin Knight; Heng Ji\",\n \"journal\": \"\",\n \"ref_id\": \"b44\",\n \"title\": \"Cross-lingual name tagging and linking for 282 languages\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Sungjoon Park; Jihyung Moon; Sungdong Kim; Won Ik Cho; Jiyoon Han; Jangwon Park; Chisung Song; Junseong Kim; Yongsook Song; Taehwan Oh\",\n \"journal\": \"\",\n \"ref_id\": \"b45\",\n \"title\": \"Klue: Korean language understanding evaluation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Maria Edoardo; Goran Ponti; Olga Glavaš; Qianchu Majewska; Ivan Liu; Anna Vulić; Korhonen\",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"XCOPA: A multilingual dataset for causal commonsense reasoning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu\",\n \"journal\": \"Journal of Machine Learning Research\",\n \"ref_id\": \"b47\",\n \"title\": \"Exploring the limits of transfer learning with a unified text-to-text transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Pranav Rajpurkar; Jian Zhang; Konstantin Lopyrev; Percy Liang\",\n \"journal\": \"\",\n \"ref_id\": \"b48\",\n \"title\": \"SQuAD: 100,000+ questions for machine comprehension of text\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Sebastian Ruder; Noah Constant; Jan Botha; Aditya Siddhant; Orhan Firat; Jinlan Fu; Pengfei Liu; Junjie Hu; Dan Garrette; Graham Neubig; Melvin Johnson\",\n \"journal\": \"\",\n \"ref_id\": \"b49\",\n \"title\": \"XTREME-R: Towards more challenging and nuanced multilingual evaluation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jason Wei; Maarten Bosma; Y Vincent; Kelvin Zhao; Adams Wei Guu; Brian Yu; Nan Lester; Andrew M Du; Quoc V Dai; Le\",\n \"journal\": \"\",\n \"ref_id\": \"b50\",\n \"title\": \"Finetuned language models are zero-shot learners\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Adina Williams; Nikita Nangia; Samuel R Bowman\",\n \"journal\": \"\",\n \"ref_id\": \"b51\",\n \"title\": \"A broad-coverage challenge corpus for sentence understanding through inference\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Shijie Wu; Mark Dredze\",\n \"journal\": \"\",\n \"ref_id\": \"b52\",\n \"title\": \"Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Dongling Xiao; Han Zhang; Yukun Li; Yu Sun; Hua Hao Tian; Haifeng Wu; Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b53\",\n \"title\": \"Ernie-gen: An enhanced multi-flow pre-training and fine-tuning framework for natural language generation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Linting Xue; Noah Constant; Adam Roberts; Mihir Kale; Rami Al-Rfou; Aditya Siddhant; Aditya Barua; Colin Raffel\",\n \"journal\": \"\",\n \"ref_id\": \"b54\",\n \"title\": \"mt5: A massively multilingual pre-trained text-to-text transformer\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yinfei Yang; Yuan Zhang; Chris Tar; Jason Baldridge\",\n \"journal\": \"\",\n \"ref_id\": \"b55\",\n \"title\": \"PAWS-X: A cross-lingual adversarial dataset for paraphrase identification\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Qinyuan Ye; Bill Yuchen Lin; Xiang Ren\",\n \"journal\": \"\",\n \"ref_id\": \"b56\",\n \"title\": \"CrossFit: A few-shot learning challenge for crosstask generalization in NLP\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Xinyan Velocity; Yu ; Trina Chatterjee; Akari Asai; Junjie Hu; Eunsol Choi\",\n \"journal\": \"Proceedings of Findings of Empirical Methods in Natural Language Processing\",\n \"ref_id\": \"b57\",\n \"title\": \"Beyond counting datasets: A survey of multilingual dataset construction and necessary resources\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yuan Zhang; Jason Baldridge; Luheng He\",\n \"journal\": \"\",\n \"ref_id\": \"b58\",\n \"title\": \"Paws: Paraphrase adversaries from word scrambling\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Mengjie Zhao; Yi Zhu; Ehsan Shareghi; Ivan Vulić; Roi Reichart; Anna Korhonen; Hinrich Schütze\",\n \"journal\": \"\",\n \"ref_id\": \"b59\",\n \"title\": \"A closer look at few-shot crosslingual transfer: The choice of shots matters\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Wenhao Zhu; Hongyi Liu; Qingxiu Dong; Jingjing Xu; Lingpeng Kong; Jiajun Chen; Li ; Shujian Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b60\",\n \"title\": \"Multilingual machine translation with large language models: Empirical results and analysis\",\n \"year\": \"2023\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[2,315.93,91.73,189.85,58.29],"formula_id":"formula_0","formula_text":"XTREME ✓ XTREME-R ✓ XGLUE ✓ ✓ CrossFit ✓ ✓ MEGA* ✓ ✓ BUFFET ✓ ✓ ✓ ✓"},{"formula_coordinates":[19,214.94,452.98,226.77,38.85],"formula_id":"formula_1","formula_text":"ᄅ ᅩ ᆺᄃ ᅦᄂ ᅳ ᆫ ᄋ ᅵᄇ ᅥ ᆫ ᄉ ᅡ ᆼᄇ ᅡ ᆫᄀ ᅵ ᄎ ᅢᄋ ᅭ ᆼᄀ ᅪ ᄀ ᅪ ᆫᄅ ᅧ ᆫᄒ ᅢ ᄀ ᅮᄌ ᅵ ᆨᄌ ᅡᄃ ᅳ ᆯᄋ ᅦᄀ ᅦ ᄉ ᅵ ᆯᄌ ᅵ ᆯᄌ ᅥ ᆨᄋ ᅵ ᆫ ᄃ ᅩᄋ ᅮ ᆷᄋ ᅵ ᄃ ᅬ ᆯ ᄉ ᅮ ᄋ ᅵ ᆻᄃ ᅩᄅ ᅩ ᆨ ᄃ ᅡᄋ ᅣ ᆼᄒ ᅡ ᆫ ᄇ ᅡ ᆼᄇ ᅥ ᆸᄋ ᅳᄅ ᅩ ᄌ ᅥ ᆼᄇ ᅩᄌ ᅦᄀ ᅩ ᆼ ᄒ ᅪ ᆯᄃ ᅩ ᆼᄋ ᅳ ᆯ ᄀ ᅡ ᆼᄒ ᅪᄒ ᅡ ᆯ ᄀ ᅨᄒ ᅬ ᆨᄋ ᅵᄃ ᅡ. [SEP] hypothesis: ᄅ ᅩ ᆺᄃ ᅦᄂ ᅳ ᆫ ᄋ ᅥᄄ ᅥᄒ ᅡ ᆫ ᄌ ᅥ ᆼᄇ ᅩᄃ ᅩ ᄌ ᅦᄀ ᅩ ᆼᄒ ᅡᄌ ᅵ ᄋ ᅡ ᆭᄋ ᅳ ᆯ ᄀ ᅨᄒ ᅬ ᆨᄋ ᅵᄃ ᅡ."}],"string":"[\n {\n \"formula_coordinates\": [\n 2,\n 315.93,\n 91.73,\n 189.85,\n 58.29\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"XTREME ✓ XTREME-R ✓ XGLUE ✓ ✓ CrossFit ✓ ✓ MEGA* ✓ ✓ BUFFET ✓ ✓ ✓ ✓\"\n },\n {\n \"formula_coordinates\": [\n 19,\n 214.94,\n 452.98,\n 226.77,\n 38.85\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"ᄅ ᅩ ᆺᄃ ᅦᄂ ᅳ ᆫ ᄋ ᅵᄇ ᅥ ᆫ ᄉ ᅡ ᆼᄇ ᅡ ᆫᄀ ᅵ ᄎ ᅢᄋ ᅭ ᆼᄀ ᅪ ᄀ ᅪ ᆫᄅ ᅧ ᆫᄒ ᅢ ᄀ ᅮᄌ ᅵ ᆨᄌ ᅡᄃ ᅳ ᆯᄋ ᅦᄀ ᅦ ᄉ ᅵ ᆯᄌ ᅵ ᆯᄌ ᅥ ᆨᄋ ᅵ ᆫ ᄃ ᅩᄋ ᅮ ᆷᄋ ᅵ ᄃ ᅬ ᆯ ᄉ ᅮ ᄋ ᅵ ᆻᄃ ᅩᄅ ᅩ ᆨ ᄃ ᅡᄋ ᅣ ᆼᄒ ᅡ ᆫ ᄇ ᅡ ᆼᄇ ᅥ ᆸᄋ ᅳᄅ ᅩ ᄌ ᅥ ᆼᄇ ᅩᄌ ᅦᄀ ᅩ ᆼ ᄒ ᅪ ᆯᄃ ᅩ ᆼᄋ ᅳ ᆯ ᄀ ᅡ ᆼᄒ ᅪᄒ ᅡ ᆯ ᄀ ᅨᄒ ᅬ ᆨᄋ ᅵᄃ ᅡ. [SEP] hypothesis: ᄅ ᅩ ᆺᄃ ᅦᄂ ᅳ ᆫ ᄋ ᅥᄄ ᅥᄒ ᅡ ᆫ ᄌ ᅥ ᆼᄇ ᅩᄃ ᅩ ᄌ ᅦᄀ ᅩ ᆼᄒ ᅡᄌ ᅵ ᄋ ᅡ ᆭᄋ ᅳ ᆯ ᄀ ᅨᄒ ᅬ ᆨᄋ ᅵᄃ ᅡ.\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"BUFFET: Benchmarking Large Language Models for Few-shot Cross-lingual Transfer"},"abstract":{"kind":"string","value":"Despite remarkable advancements in few-shot generalization in natural language processing, most models are developed and evaluated primarily in English. To facilitate research on few-shot cross-lingual transfer, we introduce a new benchmark, called BUFFET, which unifies 15 diverse tasks across 54 languages in a sequence-to-sequence format and provides a fixed set of few-shot examples and instructions. BUFFET is designed to establish a rigorous and equitable evaluation framework for fewshot cross-lingual transfer across a broad range of tasks and languages. Using BUFFET, we perform thorough evaluations of state-of-theart multilingual large language models with different transfer methods, namely in-context learning and fine-tuning. Our findings reveal significant room for improvement in few-shot in-context cross-lingual transfer. In particular, ChatGPT with in-context learning often performs worse than much smaller mT5-base models fine-tuned on English task data and few-shot in-language examples. Our analysis suggests various avenues for future research in few-shot cross-lingual transfer, such as improved pretraining, understanding, and future evaluations."},"authors":{"kind":"string","value":"Akari Asai; Sneha Kudugunta; Xinyan Velocity; Terra Blevins; Hila Gonen; Machel Reid; Yulia Tsvetkov; Sebastian Ruder; Hannaneh Hajishirzi; Victor Sanh; Albert Webson; Colin Raffel; Stephen Bach; Lintang Sutawika; Zaid Alyafeai; Antoine Chaffin; Arnaud Stiegler; Arun Raja; Manan Dey; Saiful Bari; Canwen Xu; Urmish Thakker; Shanya Sharma; Eliza Szczechla; Taewoon Kim; Gunjan Chhablani; Nihal Nayak; Debajyoti Datta; Jonathan Chang; Tian-Jian Jiang; Han Wang; Matteo Manica; Sheng Shen; Zheng Xin Yong; Harshit Pandey; Rachel Bawden; Thomas Wang; Tr- Ishala Neeraj; Jos Rozen; Abheesht Sharma; An- Drea Santilli; Thibault Fevry; Jason Fries; Freda Shi; Mirac Suzgun; Markus Freitag; Xuezhi Wang; Suraj Srivats; Soroush Vosoughi; Hyung Won; Taylor Shin; Yasaman Razeghi; Robert L Logan; Eric Wallace; Sameer Singh; Autoprompt; Hongjin Su; Jungo Kasai; Chen Henry Wu; Weijia Shi; Tianlu Wang; Jiayi Xin; Rui Zhang; Mari Ostendorf; Luke Zettlemoyer; Noah A Smith; Selec; David Vilar; Colin Cherry; Jiaming Luo; Viresh Ratnakar; Yizhong Wang; Yeganeh Kordi; Swaroop Mishra; Al- Isa Liu; Daniel Khashabi; Han- Naneh 2022a Hajishirzi; Self; Pegah Alipoormo- Labashi; Amirreza Mirzaei; Atharva Naik; Arjun Ashok; Selvan Dhanasekaran; Anjana Arunkumar; David Stap; Eshaan Pathak; Giannis Karamanolakis; Haizhi Lai; Ishan Puro- Hit; Ishani Mondal; Jacob Anderson; Kirby Kuznia; Krima Doshi; Kuntal Kumar Pal; Maitreya Patel; Mehrad Moradshahi; Mihir Parmar; Mirali Purohit; Neeraj Varshney; Rohitha Kaza; Pulkit Verma; Singh Puri; Rushang Karia; Savan Doshi"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: BUFFET includes unified diverse tasks in the same format, covering many typologically diverse languages. It enables a fair comparison across models, transfer methods, and languages and facilitates largescale analysis across different learning setups.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Model performance across three tasks, NLI, NER, and QA, displayed for various languages. The languages are sorted based on token availability in mC4, with the left side representing high-resource languages. All methods show performance deteriorations in lower-resource languages (right side), with larger drops in ENGLISH-ICL methods. Additional fine-tuning in target languages is more effective in less-represented languages.","figure_data":"","figure_id":"fig_1","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :Figure 5 :45Figure 4: Model performance across different k-shot demonstrations for QA (TYDIQA), NER (WIKIANN), and sentiment analysis (INDICSENTIMENT, AMAZONREVIEW). Each circle indicates performance given different k-shot demonstrations. There's a significant performance gap caused by the choice of demonstrations, which is often larger in ICL methods.","figure_data":"","figure_id":"fig_3","figure_label":"45","figure_type":"figure"},{"figure_caption":"Figure 6 :6Figure 6: Model scaling experimental results. We conduct experiments on four sub-tasks and use three BLOOM models, BLOOM-560M, 1B, and 7B.","figure_data":"","figure_id":"fig_4","figure_label":"6","figure_type":"figure"},{"figure_caption":"Figure 7 :7Figure 7: Overall results on BUFFET-Light.","figure_data":"","figure_id":"fig_5","figure_label":"7","figure_type":"figure"},{"figure_caption":"Figure 8 :8Figure 8: Model performance across three tasks, NLI, NER, and QA, displayed for various languages. The languages are sorted based on token availability in mC4, with the left side representing high-resource languages. All methods show performance deterioration in lower-resource languages (right side), with larger drops in ENGLISH-ICL methods. Additional fine-tuning in target languages is more effective in less-represented languages.","figure_data":"","figure_id":"fig_6","figure_label":"8","figure_type":"figure"},{"figure_caption":"Figure 9 :Figure 10 :Figure 11 :Figure 12 :9101112Figure 9: Effects of demonstrations on Amazon Review. The x-axis indicates the number of training instances used during the transfer.","figure_data":"","figure_id":"fig_7","figure_label":"9101112","figure_type":"figure"},{"figure_caption":"The eight target tasks built upon 15 existing datasets in BUFFET. |L| indicates the number of languages, and k indicates the total number of training instances. We include datasets that are diverse in terms of output format, tasks, and domains. We also include datasets that are curated by translation, in-language annotation (in-language) and automatically aligned (aligned) followingYu et al. (2022).","figure_data":"","figure_id":"tab_2","figure_label":"2","figure_type":"table"},{"figure_caption":"Comparison","figure_data":"of different few-shot and zero-shot transfer methods, based on the resources theyuse. The top section requires parameter updates viafine-tuning (FT), and the bottom uses ICL with no up-dates. k = k-shot examples; N = full training data; ✓=instruction language. The bottom half lists the mod-els evaluated in this work. The blue-colored rows areinstruction-tuned models.","figure_id":"tab_5","figure_label":"3","figure_type":"table"},{"figure_caption":"Overall","figure_data":"OutputClassificationMultiple ChoiceSpanStr.GenerationAvg.Tasks NLI Sent. PWX XCPA XWGDTyDiNER QG Summ. classgenRandom33.350.050.050.050.0-----TGT. FTmT5 35.067.247.744.148.85.233.43.22.540.72.9ENG. FTmT5 49.989.877.50.00.066.839.03.86.255.55.0ENG.+TGT.mT5 51.891.077.849.548.569.547.8 12.511.861.212.2ENG. ICLBLOOM 32.181.742.250.251.054.724.29.33.445.06.4mT5 35.750.042.250.447.50.20.00.00.431.70.2BLOOMZ 31.5 86.3* 48.5*50.454.265.8*25.5 13.58.3*47.510.9mT0 36.2 72.1* 60.6*50.560.373.6*7.916.13.4*46.39.7ChatGPT † 54.591.168.676.773.368.145.4 21.25.464.613.3TGT. ICLBLOOM 27.980.546.549.951.811.823.4 11.23.640.47.4mT5 35.750.042.249.845.20.20.00.00.431.50.2BLOOMZ 32.0 61.7* 52.5*49.755.563.1*23.49.18.0*43.48.5mT0 36.2 72.1* 60.6*50.560.373.6*7.916.13.4*46.39.7ChatGPT † 48.291.568.274.373.468.044.8 21.111.462.716.3Z-EICLBLOOM 33.337.242.350.047.14.30.014.06.329.210.1mT5 35.149.842.250.755.52.20.00.14.832.50.6BLOOMZ 33.5 51.6* 57.8*51.851.083.2*11.29.54.3*41.96.9mT0 48.5 90.0* 90.6*63.861.080.1*0.010.212.0*56.411.1","figure_id":"tab_6","figure_label":"4","figure_type":"table"},{"figure_caption":"List of all languages in BUFFET.","figure_data":"Language nameLanguage familycodeAmharicAfro-AsiaticamhArabicAfro-AsiaticarAssameseIndo-EuropeanasAymaraaymaran languages aymBelarusianIndo-EuropeanbeBengaliIndo-EuropeanbnBoroSino-TibetanbrxBulgarianIndo-EuropeanbgBribriChibchanbzdChineseSino-TibetanzhAsháninkaArawakancniEstonianUralicetFinnishUralicfiFrenchIndo-EuropeanfrGermanIndo-EuropeandeGuaraniTupiangnGujaratiIndo-EuropeanguHaitianFrench CreolehtHausaNiger-CongohauWixarikaUto-AztecanhchHindiIndo-EuropeanhiIgboNiger-CongoiboIndonesianAustronesianidItalianIndo-EuropeanitJapaneseJaponicjaKannadaDravidianknKinyarwandaNiger-CongokinKoreanKoreanickoLuoNilo SaharanluoMaithiliIndo-EuropeanmaiMalayalamDravidianmlMarathiIndo-EuropeanmrModern GreekIndo-EuropeanelNahuatlUto-AztecannahOriya (macrolanguage) Indo-EuropeanorOtomíOto-MangueanotoPanjabiIndo-EuropeanpaNigerianPidginEnglish CreolepcmPersianIndo-EuropeanfaPortugueseIndo-EuropeanptQuechuaothersquRussianIndo-EuropeanruShipibo-KoniboPanoanshpSpanishIndo-EuropeanesSwahilNiger-CongoswTamilDravidiantaRarámuriUto-AztecantarTeluguDravidianteThaiKra-DaithTurkishTurkictrUrduIndo-EuropeanurVietnameseAustroasiaticviWolofNiger-Congo\"wolYorùbáNiger-Congoyor","figure_id":"tab_7","figure_label":"5","figure_type":"table"},{"figure_caption":"The list of English instructions for each task in BUFFET.","figure_data":"TaskDatasetLanguagesNLIAMERICAS NLIaym, bzd, cni, gn, hch, nah, too, quy, shp, tarKLUE NLIkoOCNLIzhPARSI NLU ENTAILMENT faXNLIar, bg, de, el, en, es, fr, hi, ru, sw, th, tr, ur, vi, zhPARAPHRASE DETECTION PAWS(en,) de, es, fr, ja, ko, zhSENTIMENTAMAZON REVIEW(en), de, es, fr, ja, zhANALYSISINDIC SENTIMENTas, bn, brx, gu, hi, kn, mai, ml, mr, or, pa, ta, te, urCOMMONSENSEXCOPAet, ht, it, id, qu, sw, zh, ta, th, tr, viCOMMONSENSEXWINOGRAD(en,) ja, pt, ru, zhQATYDIQA(en,) ar, be, fi, id, sw, ko, ru, teNERWIKIANN( en,) ta, fr","figure_id":"tab_9","figure_label":"7","figure_type":"table"},{"figure_caption":"The list of datasets with language lists in BUFFET.","figure_data":"TaskDatasetLanguagesNLIAMERICAS NLIaym, cni, hchKLUE NLIkoPARSI NLU ENTAILMENT faXNLIbg, el, hi, sw, urParaphrase Detection PAWS-Xde, es, ja, ko, zhSentimentAMAZON REVIEWde, fr, ja, zhAnalysisINDIC SENTIMENTbn, ta, urCommonsenseXCOPAet, it, ta, th, trXWINOGRADpt, ruQATYDIQAbe, id, swNERWIKIANNbe, bg, el, et, fi, itMASAKHANERyorSummarizationXLSUMbn, fa, es, id, tr, viQGTYDIQA-QGar, fi, ko, ru, te","figure_id":"tab_10","figure_label":"8","figure_type":"table"},{"figure_caption":"The subset of datasets and languages included in BUFFET-Light.","figure_data":"OutputClassificationMultiple ChoiceSpanStr.GenerationAvg.Tasks NLI Sent.Para. XCPA XWGDQANERQGSumm. classgenRandom33.350.050.050.050.0-----TGT. FTmT5 34.667.247.246.750.08.330.83.42.840.23.1ENG. FTmT5 46.089.778.60.00.062.930.84.24.048.24.1ENG.+TGT.mT5 48.890.477.949.949.066.743.512.28.458.810.0ENG. ICLBLOOM 33.685.342.450.050.839.225.011.62.444.07.0mT5 34.550.043.250.049.20.31.60.00.332.10.1BLOOMZ 33.0 87.2* 49.5*50.552.144.5*20.013.99.0*44.311.4mT0 33.6 79.9* 61.1*57.159.669.0*7.915.31.5*45.68.4ChatGPT † 54.591.168.676.773.368.145.421.264.613.3TGT. ICLBLOOM 31.785.345.950.151.77.025.212.84.741.28.7mT5 34.450.043.150.047.30.20.20.00.331.70.1BLOOMZ 32.1 64.7* 51.7*50.553.143.7*19.112.010.9*40.611.4mT0 38.1 70.6* 60.9*52.857.970.8*8.514.61.8*45.78.2ChatGPT † 48.291.568.274.373.468.044.821.111.462.716.3Z-EICLBLOOM 32.335.842.350.146.43.10.016.44.128.810.0mT5 34.250.042.450.146.42.00.00.11.332.50.7BLOOMZ 34.0 51.6* 58.0*50.150.965.2*7.610.22.9*39.36.6mT0 49.1 90.2* 91.2*64.164.575.2*0.010.38.5*56.09.4","figure_id":"tab_11","figure_label":"9","figure_type":"table"},{"figure_caption":"On XCOPA, ChatGPT and mT0 (Z EICL) yield high performance across","figure_data":"Transfer + ModelMacro aymbzdcnignhchnahotoquyshptarTarget FT35.9 36.0 35.5 35.5 35.7 32.7 37.5 35.2 35.4 37.6 37.8English FT42.6 40.7 44.9 43.3 46.8 38.0 42.5 41.6 46.1 43.2 39.2English Target FT45.1 46.2 48.6 45.0 49.7 38.8 46.8 44.2 46.4 42.5 43.0EICL BLOOM33.7 33.4 34.6 33.2 34.1 33.3 33.5 33.4 34.3 34.0 33.6EICL mT533.3 33.3 32.8 33.3 33.3 33.2 33.2 33.2 33.3 33.3 33.3EICL BLOOMZ33.3 33.1 33.5 33.7 33.3 33.3 33.8 32.0 33.3 33.3 33.3EICL mT033.3 33.3 33.2 33.3 33.3 33.4 33.3 33.3 33.4 33.3 32.9EICL ChatGPT36.3 33.6-40.9-34.3-----TICL BLOOM","figure_id":"tab_12","figure_label":"17","figure_type":"table"},{"figure_caption":"Model performance on AMERICASNLI. We report the average of the three few-shot samples.","figure_data":"33.3","figure_id":"tab_13","figure_label":"11","figure_type":"table"},{"figure_caption":"Model performance on KLUENLI, OCNLI and PARSINLUENTAILMENT. We report the average of the three few-shot samples.","figure_data":"Transfer + ModelKLUENLI PARSINLUENTAILMENT OCNLITarget FT34.034.634.0English FT57.951.132.5English Target FT61.150.538.6EICL BLOOM33.828.938.9EICL mT533.340.431.0EICL BLOOMZ31.928.838.2EICL mT034.330.036.7EICL ChatGPT64.862.3-TICL BLOOM33.428.838.2TICL mT533.340.430.5TICL BLOOMZ33.829.032.1TICL mT043.137.438.6TICL ChatGPT56.550.2-ZICL BLOOM33.837.432.0ZICL mT532.431.937.6ZICL BLOOMZ32.431.937.6ZICL mT056.955.250.6Transfer + ModelMacrodeesfrjakozhTarget FT47.2 47.5 48.8 47.1 48.1 44.2 47.3English FT78.6 79.9 83.5 84.0 74.5 74.3 75.5English Target FT77.9 79.9 82.6 81.0 73.1 73.9 77.0EICL BLOOM42.4 41.5 42.3 43.0 42.7 42.0 42.8EICL mT543.2 41.5 42.4 47.7 42.7 42.0 42.6EICL BLOOMZ49.5 58.9 58.9 57.7 34.5 29.5 57.8EICL mT061.1 78.7 57.6 57.8 57.3 58.0 57.4EICL ChatGPT68.6 73.5 72.0-67.4 60.1 69.8TICL BLOOM45.9 49.3 42.3 42.4 42.9 54.9 43.0TICL mT543.1 41.5 46.4 43.0 42.7 42.0 42.6TICL BLOOMZ51.7 47.4 56.4 51.3 48.8 55.6 50.4TICL mT060.9 67.9 68.1 57.0 57.3 58.0 57.4TICL ChatGPT68.5 71.9 71.5-67.0 62.8 69.1ZICL BLOOM42.4 41.6 42.4 42.9 43.0 42.0 42.7ZICL mT558.0 58.0 57.8 58.6 57.7 58.1 57.5ZICL BLOOMZ58.0 58.0 57.8 58.6 57.7 58.1 57.5ZICL mT091.2 91.5 95.5 94.3 87.5 87.9 90.8","figure_id":"tab_14","figure_label":"13","figure_type":"table"},{"figure_caption":"Model performance on PAWSX. We report the average of the three few-shot samples.","figure_data":"Transfer + ModelMacrodezhesfrjaTarget FT76.3 72.9 77.1 76.1 82.3 73.1English FT91.9 94.2 84.5 93.8 95.1 91.8English Target FT92.4 93.6 87.6 93.4 94.9 92.3EICL BLOOM83.4 82.0 84.9 92.8 88.0 69.2EICL mT550.2 49.4 50.6 50.9 50.6 49.8EICL BLOOMZ81.5 75.7 80.2 93.8 93.5 64.3EICL mT079.8 88.7 70.6 81.8 89.6 68.5EICL ChatGPT85.8 94.3 87.5-96.1 65.0TICL BLOOM84.2 87.3 85.7 92.8 84.2 70.9TICL mT550.2 49.4 50.6 50.9 50.6 49.8TICL BLOOMZ64.9 57.1 71.2 79.2 61.5 55.5TICL mT072.2 88.9 51.3 58.9 85.1 76.8TICL ChatGPT89.7 94.4 85.5-95.6 83.2ZICL BLOOM50.3 49.4 50.6 50.9 50.7 49.8ZICL mT545.1 48.5 49.6 39.9 37.0 50.4ZICL BLOOMZ15.6 23.9 18.46.09.6 19.8ZICL mT087.3 90.5 72.7 90.8 93.0 89.5","figure_id":"tab_15","figure_label":"14","figure_type":"table"},{"figure_caption":"Model performance on AMAZON REVIEWS MULTI. We report the average of the three few-shot samples.","figure_data":"Transfer + ModelMacroasbnbrxguhiTarget FT58.2 61.4 55.8 62.6 56.7 64.1English FT87.4 85.0 87.4 89.4 88.4 91.6English Target FT88.4 84.6 90.2 90.6 89.7 93.0EICL BLOOM87.2 83.7 87.6 91.2 86.1 92.0EICL mT549.8 49.8 49.8 49.8 49.8 49.8EICL BLOOMZ93.0 89.6 94.2 94.9 93.1 95.6EICL mT079.9 73.6 88.4 81.3 80.2 81.1EICL ChatGPT89.3-91.8---TICL BLOOM86.5 83.1 86.7 91.2 84.1 92.6TICL mT549.8 49.8 49.8 49.8 49.8 49.8TICL BLOOMZ64.5 67.0 61.2 94.9 52.8 56.5TICL mT069.0 87.4 82.9 50.1 78.2 68.3TICL ChatGPT89.7-92.6---ZICL BLOOM49.7 49.8 49.8 49.8 49.8 49.8ZICL mT526.5 24.4 24.4 24.8 26.0 26.1ZICL BLOOMZ64.5 67.0 61.2 94.9 52.8 56.5ZICL mT093.2 90.5 93.7 94.3 92.2 95.3Transfer + ModelknmaimlmrorpatateurTarget FT59.5 62.6 45.8 60.4 62.7 48.9 57.8 55.0 60.8English FT88.4 89.4 86.9 86.1 77.2 90.4 87.0 86.7 90.3English Target FT 89.6 90.6 86.4 86.2 77.9 91.6 87.4 88.5 91.1EICL BLOOM83.0 91.2 85.8 88.9 85.8 89.0 85.0 86.0 85.1EICL mT549.8 49.8 49.8 49.8 49.8 49.8 49.8 49.8 49.8EICL BLOOMZ92.7 94.9 91.8 92.4 93.8 94.2 90.6 90.5 93.5EICL mT074.8 71.6 83.2 81.6 78.3 88.1 86.7 78.0 71.7EICL ChatGPT------82.3-93.9TICL BLOOM81.8 91.2 84.0 88.2 85.0 88.2 85.3 85.1 84.1TICL mT549.8 49.8 49.8 49.8 49.8 49.8 49.8 49.8 49.8TICL BLOOMZ49.7 94.9 66.3 58.3 59.2 57.3 68.2 50.3 66.9TICL mT072.1 49.7 84.4 79.7 66.1 68.8 55.3 58.7 64.9TICL ChatGPT------83.9-92.4ZICL BLOOM49.8 49.8 49.3 49.8 49.8 49.8 49.6 49.8 48.7ZICL mT526.8 24.8 29.0 20.7 22.4 32.4 25.4 28.9 34.5ZICL BLOOMZ26.8 24.8 29.0 20.7 22.4 32.4 25.4 28.9 34.5ZICL mT093.5 94.3 92.0 92.8 91.2 95.2 92.3 92.9 94.6","figure_id":"tab_16","figure_label":"15","figure_type":"table"},{"figure_caption":"Model performance on XCOPA. We report the average of the three few-shot samples.","figure_data":"Transfer + ModelMacroethtitidquswzhtathtrviTarget FT46.7 50.0 50.1 48.3 50.5 50.4 32.5 49.8 49.3 49.4 33.9 50.0English FT0.00.00.00.00.00.00.00.00.00.00.00.0English Target FT49.9 50.3 49.9 49.6 49.2 50.5 50.4 50.4 49.2 50.7 49.5 49.4EICL BLOOM50.0 51.5 49.0 49.9 50.0 50.6 50.0 50.1 49.5 50.0 49.9 50.0EICL mT550.0 50.0 49.9 50.7 50.0 49.5 49.8 49.9 50.7 50.0 50.0 50.0EICL BLOOMZ50.5 50.7 51.2 50.9 50.0 52.7 49.9 50.0 50.1 49.8 49.8 50.0EICL mT057.1 60.7 60.6 53.4 59.8 50.0 61.6 64.1 51.9 54.1 54.1 58.1EICL ChatGPT76.7 87.6-91.2----54.6 62.6 87.4-TICL BLOOM50.1 49.8 50.4 50.4 49.0 48.8 52.2 50.6 49.6 50.0 49.8 50.2TICL mT550.0 49.9 50.0 49.9 50.0 50.0 49.9 50.0 50.0 50.0 49.5 50.9TICL BLOOMZ50.5 45.6 50.8 50.4 53.4 47.4 49.8 51.8 53.2 50.0 49.4 53.4TICL mT052.8 50.4 51.9 51.0 51.9 50.6 53.7 50.5 50.1 50.6 54.3 65.5TICL ChatGPT74.4 89.2-91.6----49.5 55.7 86.2-ZICL BLOOM50.9 51.8 48.8 51.2 51.4 50.6 51.2 53.6 52.4 48.2 49.8 50.6ZICL mT550.1 49.8 50.4 50.4 49.0 48.8 52.2 50.6 49.6 50.0 49.8 50.2ZICL BLOOMZ50.1 48.6 50.2 52.4 47.4 50.8 45.2 46.8 54.8 50.6 52.8 51.0ZICL mT064.1 64.0 62.2 66.2 70.0 48.8 66.2 71.8 61.0 63.0 65.0 67.2Transfer + ModelMacrojpptruzhTarget FT50.0 48.4 50.3 49.9 51.4English FT0.00.00.00.00.0English Target FT49.0 48.4 48.4 48.8 50.6EICL BLOOM50.8 49.6 48.0 54.0 51.5EICL mT549.2 48.4 49.5 47.4 51.3EICL BLOOMZ52.1 52.6 50.3 55.3 50.1EICL mT059.6 62.2 57.7 53.2 65.2EICL ChatGPT73.3-74.1 72.5-TICL BLOOM51.7 52.2 50.2 54.3 50.1TICL mT547.3 48.4 46.2 44.4 50.3TICL BLOOMZ53.1 52.7 54.5 55.3 50.0TICL mT057.9 54.9 57.2 62.9 56.5TICL ChatGPT71.6-70.4 72.8-ZICL BLOOM53.7 51.9 54.4 56.7 51.9ZICL mT546.4 47.4 48.5 45.7 44.2ZICL BLOOMZ50.9 51.9 51.9 50.2 49.6ZICL mT064.5 68.7 59.8 62.2 67.3","figure_id":"tab_17","figure_label":"17","figure_type":"table"},{"figure_caption":"Model performance on XWINOGRAD We report the average of the three few-shot samples.","figure_data":"Transfer + ModelMacroarbefiidswkoruteTarget FT8.38.16.19.16.45.57.59.2 14.7English FT62.9 61.0 63.2 65.3 69.2 67.9 57.1 56.3 63.5English Target FT66.7 65.9 68.0 63.6 70.0 69.3 60.6 65.1 70.7EICL BLOOM39.2 43.8 58.2 20.6 47.0 57.5 23.2 32.7 30.4EICL mT50.30.70.10.40.20.30.00.30.0EICL BLOOMZ44.5 45.3 67.7 18.9 61.0 73.7 12.4 19.6 57.6EICL mT069.0 54.0 75.8 68.9 68.8 75.5 68.1 53.7 86.7EICL ChatGPT70.8-58.9-76.5 77.0---TICL BLOOM7.0 13.2 11.91.7 19.14.50.71.33.7TICL mT50.20.40.10.20.60.2-0.3-TICL BLOOMZ43.7 44.7 63.7 17.5 60.3 71.5 12.1 20.3 59.3TICL mT070.8 58.7 75.8 66.9 72.1 78.3 72.1 65.9 76.6TICL ChatGPT66.7-46.0-76.7 77.4---ZICL BLOOM2.02.21.13.13.22.31.01.51.7ZICL mT565.2 80.0 86.37.3 81.3 82.0 44.7 55.0 85.1ZICL BLOOMZ65.2 80.0 86.37.3 81.3 82.0 44.7 55.0 85.1ZICL mT075.2 71.8 84.4 67.3 77.3 78.6 68.3 65.0 88.9","figure_id":"tab_18","figure_label":"18","figure_type":"table"},{"figure_caption":"Model performance on TYDIQA. We report the average of the three few-shot samples.","figure_data":"Transfer + ModelMacrotafritjavibeguetthTarget FT43.7 0.2 59.0 55.5 43.9 58.3 63.5 26.0 54.4 23.7English FT52.2 0.8 78.2 79.4 56.1 80.5 73.9 24.0 60.5 10.7English Target FT56.6 0.8 78.1 76.8 55.7 75.9 76.8 37.0 76.0 25.6EICL BLOOM32.8 0.6 51.6 51.0 22.1 53.8 25.6 22.3 37.01.7EICL mT51.6 0.00.00.00.00.03.30.30.00.0EICL BLOOMZ22.4 0.5 37.1 43.4 15.6 36.8 15.4 13.0 29.60.3EICL mT015.8 0.1 13.8 13.09.1 22.9 11.06.0 24.11.4EICL ChatGPT77.6--81.8--78.2-78.2-TICL BLOOM32.8 0.7 52.5 50.2 20.8 53.5 24.4 24.0 34.01.0TICL mT50.3 0.00.00.10.00.10.21.30.01.7TICL BLOOMZ20.7 0.6 37.3 39.8 15.0 32.1 13.58.7 25.10.2TICL mT015.8 0.1 13.8 13.09.1 22.9 11.06.0 24.11.4TICL ChatGPT76.8--82.3--78.4-76.9-Transfer + ModelorknfignruelureshiteasTarget FT36.5 12.5 55.5 60.3 50.1 59.0 68.4 54.9 42.4 7.0 25.3English FT35.5 11.0 64.2 71.0 60.4 73.4 79.6 75.7 47.9 6.6 26.0English Target FT 40.0 22.5 74.8 68.0 67.8 74.4 79.1 78.3 53.7 9.5 28.3EICL BLOOM22.06.0 39.5 47.3 26.1 20.4 70.7 55.2 40.2 5.6 22.7EICL mT50.01.30.00.00.00.0 10.10.0 10.0 0.00.7EICL BLOOMZ10.05.7 31.8 28.0 19.7 15.8 41.7 37.5 30.9 4.2 16.0EICL mT016.33.3 15.2 24.3 15.1 12.8 47.1 20.3 18.7 3.3 10.0EICL ChatGPT--81.5--72.4-----TICL BLOOM25.36.7 37.6 49.0 26.2 19.7 71.7 55.6 39.9 5.3 24.0TICL mT50.30.00.00.00.00.00.00.01.8 0.01.0TICL BLOOMZ6.54.0 26.5 24.7 17.4 13.0 47.3 41.1 26.5 3.8 13.0TICL mT016.33.3 15.2 24.3 15.1 12.8 47.1 20.3 18.7 3.3 10.0TICL ChatGPT--81.9--69.3-----Transfer + ModelswpabgmltrfaidkomrdearbnzhTarget FT57.5 29.7 54.2 19.7 55.4 48.0 64.2 36.1 34.8 51.2 40.6 43.0 49.9English FT61.0 35.5 67.0 21.4 64.5 60.5 81.6 36.2 36.6 75.1 52.9 48.7 66.6English Target FT 75.3 42.3 67.1 24.5 79.5 57.6 80.7 57.7 44.7 73.2 52.9 47.7 65.2EICL BLOOM60.3 26.3 30.9 14.0 39.4 28.6 61.2 12.0 28.4 41.7 43.9 34.9 38.7EICL mT50.00.70.00.00.00.00.30.00.46.7 16.73.70.0EICL BLOOMZ34.9 15.0 22.75.0 34.6 14.7 31.79.8 22.6 26.4 21.0 36.0 31.3EICL mT024.3 10.0 14.75.0 20.2 21.4 23.4 11.2 12.3 15.7 23.0 23.9 27.7EICL ChatGPT--73.3----------TICL BLOOM58.8 26.7 29.6 14.4 39.6 27.8 61.4 10.6 27.9 43.3 44.6 36.8 38.3TICL mT50.40.50.10.40.3TICL BLOOMZ26.8 14.0 19.74.2 31.3 14.7 35.28.1 20.4 22.4 23.6 36.2 31.0TICL mT024.3 10.0 14.75.0 20.2 21.4 23.4 11.2 12.3 15.7 23.0 23.9 27.7TICL ChatGPT--72.0----------","figure_id":"tab_19","figure_label":"19","figure_type":"table"},{"figure_caption":"Model performance on WIKIANN. We report the average of the three few-shot samples.","figure_data":"MASAKHANER. Results on MASAKHANERC.7 Generationare available at Table 21. In this benchmark, all ICL methods, including ChatGPT, encounter difficul-ties, whereas TARGET FT and ENG.+TGT. FT con-sistently demonstrates strong performance across various languages. Notably, by incorporating an additional 32 training examples, ENG.+TGT. FT achieves a significant 34% improvement in perfor-TYDIQA-QG. The experimental results for TYDIQA-QG are available in Table 22. On this task, ChatGPT and mT0 ENGLISH ICL show su-perior performance than smaller fine-tuned models, demonstrating their competitiveness in generating fluent text in target languages.mance for Hausa. These remarkable enhancementsXLSUM. XLSUM results are available in Ta-underscore the effectiveness of fine-tuning a spe-ble 23. Despite strong generation capability, Chat-cialized model on a limited set of training samplesGPT ENGLISH ICL performance remains low. Wein target languages.found that when instructed in English, ChatGPToften generates summaries in English, not in thearticle language. We haven't observed such be-haviors on other tasks or other LLMs. ChatGPT","figure_id":"tab_20","figure_label":"20","figure_type":"table"},{"figure_caption":"Model performance on MASAKHANER. We report the average of the three few-shot samples.","figure_data":"Transfer + ModelMacroarbefiidswkoruteTarget FT3.42.74.12.54.43.22.82.15.8English FT4.22.13.55.16.25.13.04.74.2English Target FT12.2 11.57.3 15.8 14.1 13.17.98.9 18.8EICL BLOOM11.6 18.3 10.4 10.8 16.1 15.21.33.7 17.4EICL mT50.00.00.00.00.00.00.00.00.1EICL BLOOMZ13.9 19.5 14.27.8 23.6 23.10.72.1 20.3EICL mT015.3 25.8 10.33.7 19.6 12.34.16.2 40.1EICL ChatGPT17.8 30.6-28.2--0.72.6 26.9TICL BLOOM12.8 18.19.6 10.0 15.7 14.97.79.2 16.8TICL mT50.00.00.00.00.00.00.00.00.0TICL BLOOMZ12.0 16.0 10.75.0 20.0 21.11.95.2 15.9TICL mT014.6 17.79.16.6 18.3 12.05.18.5 39.3TICL ChatGPT19.2 24.0-27.5--14.8 17.6 12.2ZICL BLOOM0.10.10.10.00.10.10.10.10.0ZICL mT516.5 30.6 15.55.2 24.5 21.83.04.6 26.8ZICL BLOOMZ1.72.42.11.72.52.21.00.91.2ZICL mT010.34.9 13.73.5 12.35.41.92.0 39.1","figure_id":"tab_21","figure_label":"21","figure_type":"table"},{"figure_caption":"Model performance on XLSUM","figure_data":"Transfer + ModelMacro Tamil Vietnamese Swahili IndonesianTarget FT2.80.811.02.01.7English FT4.00.118.47.84.9English Target FT8.410.924.78.87.8EICL BLOOM2.40.19.04.63.8EICL mT50.30.01.70.40.2EICL BLOOMZ9.018.612.31.63.3EICL mT01.80.010.45.31.0EICL ChatGPT5.4-19.5-4.9TICL BLOOM4.713.910.34.63.1TICL mT50.30.01.70.30.3TICL BLOOMZ10.94.612.91.215.7TICL mT01.80.010.45.31.0TICL ChatGPT11.4-19.5-7.2ZICL BLOOM4.10.110.79.09.5ZICL mT51.30.54.81.10.7ZICL BLOOMZ4.30.00.00.09.5ZICL mT08.51.126.918.316.8","figure_id":"tab_22","figure_label":"23","figure_type":"table"},{"figure_caption":"The performance (in BLEU score) for different instructions for TYDIQA-QG. lang stands for the full language name, e.g., Arabic. We modify the instructions from BUFFET, PromptSource(Bach et al., 2022), and SuperNaturalInstructions (Wang et al., 2022b). Manajer Chelsea Jose Mourinho memberi waktu 48 jam untuk Wayne Rooney memutuskan apakah dia ingin meninggalkan Manchester United dan pindah ke Stamford Bridge. (Google Translate: Chelsea manager Jose Mourinho has given Wayne Rooney 48 hours to decide whether he wants to leave Manchester United and move to Stamford Bridge.) ENGLISH ICL: Manchester United striker Wayne Rooney has been given 48 hours to resolve speculation over his future, with Chelsea expressing interest in the player. [en] TARGET ICL:Mourinho memberi ultimatum kepada Wayne Rooney untuk mengakhiri spekulasi kepindahannya dalam waktu 48 jam, sementara MU dikabarkan sedang mengincar Gareth Bale. [id] XLSUM-Turkish Suriye'nin Rakka vilayetine baglı Tel Abyad kent merkezinde I Şİ D ile YPG güçleri arasında şiddetli çatışmalar sürerken, Akçakale sınırı sıgınmacılara yeniden açıldı. I Şİ D'e karşı savaşan YPG güçleri de Tal Abyad'a girdi. ENGLISH ICL: YPG forces have cut off a vital ISIS supply route from Tel Abyad to Raqqa, and have taken control of two neighborhoods in Tel Abyad. [en]","figure_data":"TargetGold.Pred.XLSUM-IndonesianTARGET ICL: YPGńin ele geçirdigi, çatışmaların devamettigi ve sivillerin tehlikede oldugu\" bilgisini verdi. [tr]ENGLISH ICL: sehemu za chakula pamoja na wanga,TYDIQA-'Sehemu ya chakula pamoja na wanga, protini naprotini na vitamini.[sw]Swahilivitamin","figure_id":"tab_24","figure_label":"24","figure_type":"table"},{"figure_caption":"ChatGPT outputs for XLSUM and TYDIQA Output. [lang] indicate the language codes of the output text languages.","figure_data":"","figure_id":"tab_25","figure_label":"25","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: BUFFET includes unified diverse tasks in the same format, covering many typologically diverse languages. It enables a fair comparison across models, transfer methods, and languages and facilitates largescale analysis across different learning setups.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Model performance across three tasks, NLI, NER, and QA, displayed for various languages. The languages are sorted based on token availability in mC4, with the left side representing high-resource languages. All methods show performance deteriorations in lower-resource languages (right side), with larger drops in ENGLISH-ICL methods. Additional fine-tuning in target languages is more effective in less-represented languages.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :Figure 5 :45Figure 4: Model performance across different k-shot demonstrations for QA (TYDIQA), NER (WIKIANN), and sentiment analysis (INDICSENTIMENT, AMAZONREVIEW). Each circle indicates performance given different k-shot demonstrations. There's a significant performance gap caused by the choice of demonstrations, which is often larger in ICL methods.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"45\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 :6Figure 6: Model scaling experimental results. We conduct experiments on four sub-tasks and use three BLOOM models, BLOOM-560M, 1B, and 7B.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 7 :7Figure 7: Overall results on BUFFET-Light.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"7\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 8 :8Figure 8: Model performance across three tasks, NLI, NER, and QA, displayed for various languages. The languages are sorted based on token availability in mC4, with the left side representing high-resource languages. All methods show performance deterioration in lower-resource languages (right side), with larger drops in ENGLISH-ICL methods. Additional fine-tuning in target languages is more effective in less-represented languages.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"8\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 9 :Figure 10 :Figure 11 :Figure 12 :9101112Figure 9: Effects of demonstrations on Amazon Review. The x-axis indicates the number of training instances used during the transfer.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"9101112\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"The eight target tasks built upon 15 existing datasets in BUFFET. |L| indicates the number of languages, and k indicates the total number of training instances. We include datasets that are diverse in terms of output format, tasks, and domains. We also include datasets that are curated by translation, in-language annotation (in-language) and automatically aligned (aligned) followingYu et al. (2022).\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison\",\n \"figure_data\": \"of different few-shot and zero-shot transfer methods, based on the resources theyuse. The top section requires parameter updates viafine-tuning (FT), and the bottom uses ICL with no up-dates. k = k-shot examples; N = full training data; ✓=instruction language. The bottom half lists the mod-els evaluated in this work. The blue-colored rows areinstruction-tuned models.\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Overall\",\n \"figure_data\": \"OutputClassificationMultiple ChoiceSpanStr.GenerationAvg.Tasks NLI Sent. PWX XCPA XWGDTyDiNER QG Summ. classgenRandom33.350.050.050.050.0-----TGT. FTmT5 35.067.247.744.148.85.233.43.22.540.72.9ENG. FTmT5 49.989.877.50.00.066.839.03.86.255.55.0ENG.+TGT.mT5 51.891.077.849.548.569.547.8 12.511.861.212.2ENG. ICLBLOOM 32.181.742.250.251.054.724.29.33.445.06.4mT5 35.750.042.250.447.50.20.00.00.431.70.2BLOOMZ 31.5 86.3* 48.5*50.454.265.8*25.5 13.58.3*47.510.9mT0 36.2 72.1* 60.6*50.560.373.6*7.916.13.4*46.39.7ChatGPT † 54.591.168.676.773.368.145.4 21.25.464.613.3TGT. ICLBLOOM 27.980.546.549.951.811.823.4 11.23.640.47.4mT5 35.750.042.249.845.20.20.00.00.431.50.2BLOOMZ 32.0 61.7* 52.5*49.755.563.1*23.49.18.0*43.48.5mT0 36.2 72.1* 60.6*50.560.373.6*7.916.13.4*46.39.7ChatGPT † 48.291.568.274.373.468.044.8 21.111.462.716.3Z-EICLBLOOM 33.337.242.350.047.14.30.014.06.329.210.1mT5 35.149.842.250.755.52.20.00.14.832.50.6BLOOMZ 33.5 51.6* 57.8*51.851.083.2*11.29.54.3*41.96.9mT0 48.5 90.0* 90.6*63.861.080.1*0.010.212.0*56.411.1\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"List of all languages in BUFFET.\",\n \"figure_data\": \"Language nameLanguage familycodeAmharicAfro-AsiaticamhArabicAfro-AsiaticarAssameseIndo-EuropeanasAymaraaymaran languages aymBelarusianIndo-EuropeanbeBengaliIndo-EuropeanbnBoroSino-TibetanbrxBulgarianIndo-EuropeanbgBribriChibchanbzdChineseSino-TibetanzhAsháninkaArawakancniEstonianUralicetFinnishUralicfiFrenchIndo-EuropeanfrGermanIndo-EuropeandeGuaraniTupiangnGujaratiIndo-EuropeanguHaitianFrench CreolehtHausaNiger-CongohauWixarikaUto-AztecanhchHindiIndo-EuropeanhiIgboNiger-CongoiboIndonesianAustronesianidItalianIndo-EuropeanitJapaneseJaponicjaKannadaDravidianknKinyarwandaNiger-CongokinKoreanKoreanickoLuoNilo SaharanluoMaithiliIndo-EuropeanmaiMalayalamDravidianmlMarathiIndo-EuropeanmrModern GreekIndo-EuropeanelNahuatlUto-AztecannahOriya (macrolanguage) Indo-EuropeanorOtomíOto-MangueanotoPanjabiIndo-EuropeanpaNigerianPidginEnglish CreolepcmPersianIndo-EuropeanfaPortugueseIndo-EuropeanptQuechuaothersquRussianIndo-EuropeanruShipibo-KoniboPanoanshpSpanishIndo-EuropeanesSwahilNiger-CongoswTamilDravidiantaRarámuriUto-AztecantarTeluguDravidianteThaiKra-DaithTurkishTurkictrUrduIndo-EuropeanurVietnameseAustroasiaticviWolofNiger-Congo\\\"wolYorùbáNiger-Congoyor\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"The list of English instructions for each task in BUFFET.\",\n \"figure_data\": \"TaskDatasetLanguagesNLIAMERICAS NLIaym, bzd, cni, gn, hch, nah, too, quy, shp, tarKLUE NLIkoOCNLIzhPARSI NLU ENTAILMENT faXNLIar, bg, de, el, en, es, fr, hi, ru, sw, th, tr, ur, vi, zhPARAPHRASE DETECTION PAWS(en,) de, es, fr, ja, ko, zhSENTIMENTAMAZON REVIEW(en), de, es, fr, ja, zhANALYSISINDIC SENTIMENTas, bn, brx, gu, hi, kn, mai, ml, mr, or, pa, ta, te, urCOMMONSENSEXCOPAet, ht, it, id, qu, sw, zh, ta, th, tr, viCOMMONSENSEXWINOGRAD(en,) ja, pt, ru, zhQATYDIQA(en,) ar, be, fi, id, sw, ko, ru, teNERWIKIANN( en,) ta, fr\",\n \"figure_id\": \"tab_9\",\n \"figure_label\": \"7\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"The list of datasets with language lists in BUFFET.\",\n \"figure_data\": \"TaskDatasetLanguagesNLIAMERICAS NLIaym, cni, hchKLUE NLIkoPARSI NLU ENTAILMENT faXNLIbg, el, hi, sw, urParaphrase Detection PAWS-Xde, es, ja, ko, zhSentimentAMAZON REVIEWde, fr, ja, zhAnalysisINDIC SENTIMENTbn, ta, urCommonsenseXCOPAet, it, ta, th, trXWINOGRADpt, ruQATYDIQAbe, id, swNERWIKIANNbe, bg, el, et, fi, itMASAKHANERyorSummarizationXLSUMbn, fa, es, id, tr, viQGTYDIQA-QGar, fi, ko, ru, te\",\n \"figure_id\": \"tab_10\",\n \"figure_label\": \"8\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"The subset of datasets and languages included in BUFFET-Light.\",\n \"figure_data\": \"OutputClassificationMultiple ChoiceSpanStr.GenerationAvg.Tasks NLI Sent.Para. XCPA XWGDQANERQGSumm. classgenRandom33.350.050.050.050.0-----TGT. FTmT5 34.667.247.246.750.08.330.83.42.840.23.1ENG. FTmT5 46.089.778.60.00.062.930.84.24.048.24.1ENG.+TGT.mT5 48.890.477.949.949.066.743.512.28.458.810.0ENG. ICLBLOOM 33.685.342.450.050.839.225.011.62.444.07.0mT5 34.550.043.250.049.20.31.60.00.332.10.1BLOOMZ 33.0 87.2* 49.5*50.552.144.5*20.013.99.0*44.311.4mT0 33.6 79.9* 61.1*57.159.669.0*7.915.31.5*45.68.4ChatGPT † 54.591.168.676.773.368.145.421.264.613.3TGT. ICLBLOOM 31.785.345.950.151.77.025.212.84.741.28.7mT5 34.450.043.150.047.30.20.20.00.331.70.1BLOOMZ 32.1 64.7* 51.7*50.553.143.7*19.112.010.9*40.611.4mT0 38.1 70.6* 60.9*52.857.970.8*8.514.61.8*45.78.2ChatGPT † 48.291.568.274.373.468.044.821.111.462.716.3Z-EICLBLOOM 32.335.842.350.146.43.10.016.44.128.810.0mT5 34.250.042.450.146.42.00.00.11.332.50.7BLOOMZ 34.0 51.6* 58.0*50.150.965.2*7.610.22.9*39.36.6mT0 49.1 90.2* 91.2*64.164.575.2*0.010.38.5*56.09.4\",\n \"figure_id\": \"tab_11\",\n \"figure_label\": \"9\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"On XCOPA, ChatGPT and mT0 (Z EICL) yield high performance across\",\n \"figure_data\": \"Transfer + ModelMacro aymbzdcnignhchnahotoquyshptarTarget FT35.9 36.0 35.5 35.5 35.7 32.7 37.5 35.2 35.4 37.6 37.8English FT42.6 40.7 44.9 43.3 46.8 38.0 42.5 41.6 46.1 43.2 39.2English Target FT45.1 46.2 48.6 45.0 49.7 38.8 46.8 44.2 46.4 42.5 43.0EICL BLOOM33.7 33.4 34.6 33.2 34.1 33.3 33.5 33.4 34.3 34.0 33.6EICL mT533.3 33.3 32.8 33.3 33.3 33.2 33.2 33.2 33.3 33.3 33.3EICL BLOOMZ33.3 33.1 33.5 33.7 33.3 33.3 33.8 32.0 33.3 33.3 33.3EICL mT033.3 33.3 33.2 33.3 33.3 33.4 33.3 33.3 33.4 33.3 32.9EICL ChatGPT36.3 33.6-40.9-34.3-----TICL BLOOM\",\n \"figure_id\": \"tab_12\",\n \"figure_label\": \"17\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Model performance on AMERICASNLI. We report the average of the three few-shot samples.\",\n \"figure_data\": \"33.3\",\n \"figure_id\": \"tab_13\",\n \"figure_label\": \"11\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Model performance on KLUENLI, OCNLI and PARSINLUENTAILMENT. We report the average of the three few-shot samples.\",\n \"figure_data\": \"Transfer + ModelKLUENLI PARSINLUENTAILMENT OCNLITarget FT34.034.634.0English FT57.951.132.5English Target FT61.150.538.6EICL BLOOM33.828.938.9EICL mT533.340.431.0EICL BLOOMZ31.928.838.2EICL mT034.330.036.7EICL ChatGPT64.862.3-TICL BLOOM33.428.838.2TICL mT533.340.430.5TICL BLOOMZ33.829.032.1TICL mT043.137.438.6TICL ChatGPT56.550.2-ZICL BLOOM33.837.432.0ZICL mT532.431.937.6ZICL BLOOMZ32.431.937.6ZICL mT056.955.250.6Transfer + ModelMacrodeesfrjakozhTarget FT47.2 47.5 48.8 47.1 48.1 44.2 47.3English FT78.6 79.9 83.5 84.0 74.5 74.3 75.5English Target FT77.9 79.9 82.6 81.0 73.1 73.9 77.0EICL BLOOM42.4 41.5 42.3 43.0 42.7 42.0 42.8EICL mT543.2 41.5 42.4 47.7 42.7 42.0 42.6EICL BLOOMZ49.5 58.9 58.9 57.7 34.5 29.5 57.8EICL mT061.1 78.7 57.6 57.8 57.3 58.0 57.4EICL ChatGPT68.6 73.5 72.0-67.4 60.1 69.8TICL BLOOM45.9 49.3 42.3 42.4 42.9 54.9 43.0TICL mT543.1 41.5 46.4 43.0 42.7 42.0 42.6TICL BLOOMZ51.7 47.4 56.4 51.3 48.8 55.6 50.4TICL mT060.9 67.9 68.1 57.0 57.3 58.0 57.4TICL ChatGPT68.5 71.9 71.5-67.0 62.8 69.1ZICL BLOOM42.4 41.6 42.4 42.9 43.0 42.0 42.7ZICL mT558.0 58.0 57.8 58.6 57.7 58.1 57.5ZICL BLOOMZ58.0 58.0 57.8 58.6 57.7 58.1 57.5ZICL mT091.2 91.5 95.5 94.3 87.5 87.9 90.8\",\n \"figure_id\": \"tab_14\",\n \"figure_label\": \"13\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Model performance on PAWSX. We report the average of the three few-shot samples.\",\n \"figure_data\": \"Transfer + ModelMacrodezhesfrjaTarget FT76.3 72.9 77.1 76.1 82.3 73.1English FT91.9 94.2 84.5 93.8 95.1 91.8English Target FT92.4 93.6 87.6 93.4 94.9 92.3EICL BLOOM83.4 82.0 84.9 92.8 88.0 69.2EICL mT550.2 49.4 50.6 50.9 50.6 49.8EICL BLOOMZ81.5 75.7 80.2 93.8 93.5 64.3EICL mT079.8 88.7 70.6 81.8 89.6 68.5EICL ChatGPT85.8 94.3 87.5-96.1 65.0TICL BLOOM84.2 87.3 85.7 92.8 84.2 70.9TICL mT550.2 49.4 50.6 50.9 50.6 49.8TICL BLOOMZ64.9 57.1 71.2 79.2 61.5 55.5TICL mT072.2 88.9 51.3 58.9 85.1 76.8TICL ChatGPT89.7 94.4 85.5-95.6 83.2ZICL BLOOM50.3 49.4 50.6 50.9 50.7 49.8ZICL mT545.1 48.5 49.6 39.9 37.0 50.4ZICL BLOOMZ15.6 23.9 18.46.09.6 19.8ZICL mT087.3 90.5 72.7 90.8 93.0 89.5\",\n \"figure_id\": \"tab_15\",\n \"figure_label\": \"14\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Model performance on AMAZON REVIEWS MULTI. We report the average of the three few-shot samples.\",\n \"figure_data\": \"Transfer + ModelMacroasbnbrxguhiTarget FT58.2 61.4 55.8 62.6 56.7 64.1English FT87.4 85.0 87.4 89.4 88.4 91.6English Target FT88.4 84.6 90.2 90.6 89.7 93.0EICL BLOOM87.2 83.7 87.6 91.2 86.1 92.0EICL mT549.8 49.8 49.8 49.8 49.8 49.8EICL BLOOMZ93.0 89.6 94.2 94.9 93.1 95.6EICL mT079.9 73.6 88.4 81.3 80.2 81.1EICL ChatGPT89.3-91.8---TICL BLOOM86.5 83.1 86.7 91.2 84.1 92.6TICL mT549.8 49.8 49.8 49.8 49.8 49.8TICL BLOOMZ64.5 67.0 61.2 94.9 52.8 56.5TICL mT069.0 87.4 82.9 50.1 78.2 68.3TICL ChatGPT89.7-92.6---ZICL BLOOM49.7 49.8 49.8 49.8 49.8 49.8ZICL mT526.5 24.4 24.4 24.8 26.0 26.1ZICL BLOOMZ64.5 67.0 61.2 94.9 52.8 56.5ZICL mT093.2 90.5 93.7 94.3 92.2 95.3Transfer + ModelknmaimlmrorpatateurTarget FT59.5 62.6 45.8 60.4 62.7 48.9 57.8 55.0 60.8English FT88.4 89.4 86.9 86.1 77.2 90.4 87.0 86.7 90.3English Target FT 89.6 90.6 86.4 86.2 77.9 91.6 87.4 88.5 91.1EICL BLOOM83.0 91.2 85.8 88.9 85.8 89.0 85.0 86.0 85.1EICL mT549.8 49.8 49.8 49.8 49.8 49.8 49.8 49.8 49.8EICL BLOOMZ92.7 94.9 91.8 92.4 93.8 94.2 90.6 90.5 93.5EICL mT074.8 71.6 83.2 81.6 78.3 88.1 86.7 78.0 71.7EICL ChatGPT------82.3-93.9TICL BLOOM81.8 91.2 84.0 88.2 85.0 88.2 85.3 85.1 84.1TICL mT549.8 49.8 49.8 49.8 49.8 49.8 49.8 49.8 49.8TICL BLOOMZ49.7 94.9 66.3 58.3 59.2 57.3 68.2 50.3 66.9TICL mT072.1 49.7 84.4 79.7 66.1 68.8 55.3 58.7 64.9TICL ChatGPT------83.9-92.4ZICL BLOOM49.8 49.8 49.3 49.8 49.8 49.8 49.6 49.8 48.7ZICL mT526.8 24.8 29.0 20.7 22.4 32.4 25.4 28.9 34.5ZICL BLOOMZ26.8 24.8 29.0 20.7 22.4 32.4 25.4 28.9 34.5ZICL mT093.5 94.3 92.0 92.8 91.2 95.2 92.3 92.9 94.6\",\n \"figure_id\": \"tab_16\",\n \"figure_label\": \"15\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Model performance on XCOPA. We report the average of the three few-shot samples.\",\n \"figure_data\": \"Transfer + ModelMacroethtitidquswzhtathtrviTarget FT46.7 50.0 50.1 48.3 50.5 50.4 32.5 49.8 49.3 49.4 33.9 50.0English FT0.00.00.00.00.00.00.00.00.00.00.00.0English Target FT49.9 50.3 49.9 49.6 49.2 50.5 50.4 50.4 49.2 50.7 49.5 49.4EICL BLOOM50.0 51.5 49.0 49.9 50.0 50.6 50.0 50.1 49.5 50.0 49.9 50.0EICL mT550.0 50.0 49.9 50.7 50.0 49.5 49.8 49.9 50.7 50.0 50.0 50.0EICL BLOOMZ50.5 50.7 51.2 50.9 50.0 52.7 49.9 50.0 50.1 49.8 49.8 50.0EICL mT057.1 60.7 60.6 53.4 59.8 50.0 61.6 64.1 51.9 54.1 54.1 58.1EICL ChatGPT76.7 87.6-91.2----54.6 62.6 87.4-TICL BLOOM50.1 49.8 50.4 50.4 49.0 48.8 52.2 50.6 49.6 50.0 49.8 50.2TICL mT550.0 49.9 50.0 49.9 50.0 50.0 49.9 50.0 50.0 50.0 49.5 50.9TICL BLOOMZ50.5 45.6 50.8 50.4 53.4 47.4 49.8 51.8 53.2 50.0 49.4 53.4TICL mT052.8 50.4 51.9 51.0 51.9 50.6 53.7 50.5 50.1 50.6 54.3 65.5TICL ChatGPT74.4 89.2-91.6----49.5 55.7 86.2-ZICL BLOOM50.9 51.8 48.8 51.2 51.4 50.6 51.2 53.6 52.4 48.2 49.8 50.6ZICL mT550.1 49.8 50.4 50.4 49.0 48.8 52.2 50.6 49.6 50.0 49.8 50.2ZICL BLOOMZ50.1 48.6 50.2 52.4 47.4 50.8 45.2 46.8 54.8 50.6 52.8 51.0ZICL mT064.1 64.0 62.2 66.2 70.0 48.8 66.2 71.8 61.0 63.0 65.0 67.2Transfer + ModelMacrojpptruzhTarget FT50.0 48.4 50.3 49.9 51.4English FT0.00.00.00.00.0English Target FT49.0 48.4 48.4 48.8 50.6EICL BLOOM50.8 49.6 48.0 54.0 51.5EICL mT549.2 48.4 49.5 47.4 51.3EICL BLOOMZ52.1 52.6 50.3 55.3 50.1EICL mT059.6 62.2 57.7 53.2 65.2EICL ChatGPT73.3-74.1 72.5-TICL BLOOM51.7 52.2 50.2 54.3 50.1TICL mT547.3 48.4 46.2 44.4 50.3TICL BLOOMZ53.1 52.7 54.5 55.3 50.0TICL mT057.9 54.9 57.2 62.9 56.5TICL ChatGPT71.6-70.4 72.8-ZICL BLOOM53.7 51.9 54.4 56.7 51.9ZICL mT546.4 47.4 48.5 45.7 44.2ZICL BLOOMZ50.9 51.9 51.9 50.2 49.6ZICL mT064.5 68.7 59.8 62.2 67.3\",\n \"figure_id\": \"tab_17\",\n \"figure_label\": \"17\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Model performance on XWINOGRAD We report the average of the three few-shot samples.\",\n \"figure_data\": \"Transfer + ModelMacroarbefiidswkoruteTarget FT8.38.16.19.16.45.57.59.2 14.7English FT62.9 61.0 63.2 65.3 69.2 67.9 57.1 56.3 63.5English Target FT66.7 65.9 68.0 63.6 70.0 69.3 60.6 65.1 70.7EICL BLOOM39.2 43.8 58.2 20.6 47.0 57.5 23.2 32.7 30.4EICL mT50.30.70.10.40.20.30.00.30.0EICL BLOOMZ44.5 45.3 67.7 18.9 61.0 73.7 12.4 19.6 57.6EICL mT069.0 54.0 75.8 68.9 68.8 75.5 68.1 53.7 86.7EICL ChatGPT70.8-58.9-76.5 77.0---TICL BLOOM7.0 13.2 11.91.7 19.14.50.71.33.7TICL mT50.20.40.10.20.60.2-0.3-TICL BLOOMZ43.7 44.7 63.7 17.5 60.3 71.5 12.1 20.3 59.3TICL mT070.8 58.7 75.8 66.9 72.1 78.3 72.1 65.9 76.6TICL ChatGPT66.7-46.0-76.7 77.4---ZICL BLOOM2.02.21.13.13.22.31.01.51.7ZICL mT565.2 80.0 86.37.3 81.3 82.0 44.7 55.0 85.1ZICL BLOOMZ65.2 80.0 86.37.3 81.3 82.0 44.7 55.0 85.1ZICL mT075.2 71.8 84.4 67.3 77.3 78.6 68.3 65.0 88.9\",\n \"figure_id\": \"tab_18\",\n \"figure_label\": \"18\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Model performance on TYDIQA. We report the average of the three few-shot samples.\",\n \"figure_data\": \"Transfer + ModelMacrotafritjavibeguetthTarget FT43.7 0.2 59.0 55.5 43.9 58.3 63.5 26.0 54.4 23.7English FT52.2 0.8 78.2 79.4 56.1 80.5 73.9 24.0 60.5 10.7English Target FT56.6 0.8 78.1 76.8 55.7 75.9 76.8 37.0 76.0 25.6EICL BLOOM32.8 0.6 51.6 51.0 22.1 53.8 25.6 22.3 37.01.7EICL mT51.6 0.00.00.00.00.03.30.30.00.0EICL BLOOMZ22.4 0.5 37.1 43.4 15.6 36.8 15.4 13.0 29.60.3EICL mT015.8 0.1 13.8 13.09.1 22.9 11.06.0 24.11.4EICL ChatGPT77.6--81.8--78.2-78.2-TICL BLOOM32.8 0.7 52.5 50.2 20.8 53.5 24.4 24.0 34.01.0TICL mT50.3 0.00.00.10.00.10.21.30.01.7TICL BLOOMZ20.7 0.6 37.3 39.8 15.0 32.1 13.58.7 25.10.2TICL mT015.8 0.1 13.8 13.09.1 22.9 11.06.0 24.11.4TICL ChatGPT76.8--82.3--78.4-76.9-Transfer + ModelorknfignruelureshiteasTarget FT36.5 12.5 55.5 60.3 50.1 59.0 68.4 54.9 42.4 7.0 25.3English FT35.5 11.0 64.2 71.0 60.4 73.4 79.6 75.7 47.9 6.6 26.0English Target FT 40.0 22.5 74.8 68.0 67.8 74.4 79.1 78.3 53.7 9.5 28.3EICL BLOOM22.06.0 39.5 47.3 26.1 20.4 70.7 55.2 40.2 5.6 22.7EICL mT50.01.30.00.00.00.0 10.10.0 10.0 0.00.7EICL BLOOMZ10.05.7 31.8 28.0 19.7 15.8 41.7 37.5 30.9 4.2 16.0EICL mT016.33.3 15.2 24.3 15.1 12.8 47.1 20.3 18.7 3.3 10.0EICL ChatGPT--81.5--72.4-----TICL BLOOM25.36.7 37.6 49.0 26.2 19.7 71.7 55.6 39.9 5.3 24.0TICL mT50.30.00.00.00.00.00.00.01.8 0.01.0TICL BLOOMZ6.54.0 26.5 24.7 17.4 13.0 47.3 41.1 26.5 3.8 13.0TICL mT016.33.3 15.2 24.3 15.1 12.8 47.1 20.3 18.7 3.3 10.0TICL ChatGPT--81.9--69.3-----Transfer + ModelswpabgmltrfaidkomrdearbnzhTarget FT57.5 29.7 54.2 19.7 55.4 48.0 64.2 36.1 34.8 51.2 40.6 43.0 49.9English FT61.0 35.5 67.0 21.4 64.5 60.5 81.6 36.2 36.6 75.1 52.9 48.7 66.6English Target FT 75.3 42.3 67.1 24.5 79.5 57.6 80.7 57.7 44.7 73.2 52.9 47.7 65.2EICL BLOOM60.3 26.3 30.9 14.0 39.4 28.6 61.2 12.0 28.4 41.7 43.9 34.9 38.7EICL mT50.00.70.00.00.00.00.30.00.46.7 16.73.70.0EICL BLOOMZ34.9 15.0 22.75.0 34.6 14.7 31.79.8 22.6 26.4 21.0 36.0 31.3EICL mT024.3 10.0 14.75.0 20.2 21.4 23.4 11.2 12.3 15.7 23.0 23.9 27.7EICL ChatGPT--73.3----------TICL BLOOM58.8 26.7 29.6 14.4 39.6 27.8 61.4 10.6 27.9 43.3 44.6 36.8 38.3TICL mT50.40.50.10.40.3TICL BLOOMZ26.8 14.0 19.74.2 31.3 14.7 35.28.1 20.4 22.4 23.6 36.2 31.0TICL mT024.3 10.0 14.75.0 20.2 21.4 23.4 11.2 12.3 15.7 23.0 23.9 27.7TICL ChatGPT--72.0----------\",\n \"figure_id\": \"tab_19\",\n \"figure_label\": \"19\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Model performance on WIKIANN. We report the average of the three few-shot samples.\",\n \"figure_data\": \"MASAKHANER. Results on MASAKHANERC.7 Generationare available at Table 21. In this benchmark, all ICL methods, including ChatGPT, encounter difficul-ties, whereas TARGET FT and ENG.+TGT. FT con-sistently demonstrates strong performance across various languages. Notably, by incorporating an additional 32 training examples, ENG.+TGT. FT achieves a significant 34% improvement in perfor-TYDIQA-QG. The experimental results for TYDIQA-QG are available in Table 22. On this task, ChatGPT and mT0 ENGLISH ICL show su-perior performance than smaller fine-tuned models, demonstrating their competitiveness in generating fluent text in target languages.mance for Hausa. These remarkable enhancementsXLSUM. XLSUM results are available in Ta-underscore the effectiveness of fine-tuning a spe-ble 23. Despite strong generation capability, Chat-cialized model on a limited set of training samplesGPT ENGLISH ICL performance remains low. Wein target languages.found that when instructed in English, ChatGPToften generates summaries in English, not in thearticle language. We haven't observed such be-haviors on other tasks or other LLMs. ChatGPT\",\n \"figure_id\": \"tab_20\",\n \"figure_label\": \"20\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Model performance on MASAKHANER. We report the average of the three few-shot samples.\",\n \"figure_data\": \"Transfer + ModelMacroarbefiidswkoruteTarget FT3.42.74.12.54.43.22.82.15.8English FT4.22.13.55.16.25.13.04.74.2English Target FT12.2 11.57.3 15.8 14.1 13.17.98.9 18.8EICL BLOOM11.6 18.3 10.4 10.8 16.1 15.21.33.7 17.4EICL mT50.00.00.00.00.00.00.00.00.1EICL BLOOMZ13.9 19.5 14.27.8 23.6 23.10.72.1 20.3EICL mT015.3 25.8 10.33.7 19.6 12.34.16.2 40.1EICL ChatGPT17.8 30.6-28.2--0.72.6 26.9TICL BLOOM12.8 18.19.6 10.0 15.7 14.97.79.2 16.8TICL mT50.00.00.00.00.00.00.00.00.0TICL BLOOMZ12.0 16.0 10.75.0 20.0 21.11.95.2 15.9TICL mT014.6 17.79.16.6 18.3 12.05.18.5 39.3TICL ChatGPT19.2 24.0-27.5--14.8 17.6 12.2ZICL BLOOM0.10.10.10.00.10.10.10.10.0ZICL mT516.5 30.6 15.55.2 24.5 21.83.04.6 26.8ZICL BLOOMZ1.72.42.11.72.52.21.00.91.2ZICL mT010.34.9 13.73.5 12.35.41.92.0 39.1\",\n \"figure_id\": \"tab_21\",\n \"figure_label\": \"21\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Model performance on XLSUM\",\n \"figure_data\": \"Transfer + ModelMacro Tamil Vietnamese Swahili IndonesianTarget FT2.80.811.02.01.7English FT4.00.118.47.84.9English Target FT8.410.924.78.87.8EICL BLOOM2.40.19.04.63.8EICL mT50.30.01.70.40.2EICL BLOOMZ9.018.612.31.63.3EICL mT01.80.010.45.31.0EICL ChatGPT5.4-19.5-4.9TICL BLOOM4.713.910.34.63.1TICL mT50.30.01.70.30.3TICL BLOOMZ10.94.612.91.215.7TICL mT01.80.010.45.31.0TICL ChatGPT11.4-19.5-7.2ZICL BLOOM4.10.110.79.09.5ZICL mT51.30.54.81.10.7ZICL BLOOMZ4.30.00.00.09.5ZICL mT08.51.126.918.316.8\",\n \"figure_id\": \"tab_22\",\n \"figure_label\": \"23\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"The performance (in BLEU score) for different instructions for TYDIQA-QG. lang stands for the full language name, e.g., Arabic. We modify the instructions from BUFFET, PromptSource(Bach et al., 2022), and SuperNaturalInstructions (Wang et al., 2022b). Manajer Chelsea Jose Mourinho memberi waktu 48 jam untuk Wayne Rooney memutuskan apakah dia ingin meninggalkan Manchester United dan pindah ke Stamford Bridge. (Google Translate: Chelsea manager Jose Mourinho has given Wayne Rooney 48 hours to decide whether he wants to leave Manchester United and move to Stamford Bridge.) ENGLISH ICL: Manchester United striker Wayne Rooney has been given 48 hours to resolve speculation over his future, with Chelsea expressing interest in the player. [en] TARGET ICL:Mourinho memberi ultimatum kepada Wayne Rooney untuk mengakhiri spekulasi kepindahannya dalam waktu 48 jam, sementara MU dikabarkan sedang mengincar Gareth Bale. [id] XLSUM-Turkish Suriye'nin Rakka vilayetine baglı Tel Abyad kent merkezinde I Şİ D ile YPG güçleri arasında şiddetli çatışmalar sürerken, Akçakale sınırı sıgınmacılara yeniden açıldı. I Şİ D'e karşı savaşan YPG güçleri de Tal Abyad'a girdi. ENGLISH ICL: YPG forces have cut off a vital ISIS supply route from Tel Abyad to Raqqa, and have taken control of two neighborhoods in Tel Abyad. [en]\",\n \"figure_data\": \"TargetGold.Pred.XLSUM-IndonesianTARGET ICL: YPGńin ele geçirdigi, çatışmaların devamettigi ve sivillerin tehlikede oldugu\\\" bilgisini verdi. [tr]ENGLISH ICL: sehemu za chakula pamoja na wanga,TYDIQA-'Sehemu ya chakula pamoja na wanga, protini naprotini na vitamini.[sw]Swahilivitamin\",\n \"figure_id\": \"tab_24\",\n \"figure_label\": \"24\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"ChatGPT outputs for XLSUM and TYDIQA Output. [lang] indicate the language codes of the output text languages.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_25\",\n \"figure_label\": \"25\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Supporting Evidence\", \"Citation\": \"(Blasi et al., 2022)\", \"Explanation\": \"The cited work by Blasi et al. (2022) highlights the focus of recent advances in NLP on the English language, which serves as a foundational point for the citing paper to build upon.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Yu et al., 2022)\", \"Explanation\": \"The cited work by Yu et al. (2022) mentions the lack of sufficient training data in most languages, which motivates the need for research in cross-lingual transfer learning to address the issue.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Hu et al., 2020b)\", \"Explanation\": \"The cited work by Hu et al. (2020b) explores the direct transfer of pretrained language models to new languages after fine-tuning on resource-rich languages, which the citing paper extends by discussing the potential benefits of this approach in cross-lingual transfer learning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Lauscher et al., 2020)\", \"Explanation\": \"The cited work by Lauscher et al. (2020) highlights the use of few-shot cross-lingual transfer to improve performance in languages that are distant from the source language, which the citing paper further explores in the context of English in-context learning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Hedderich et al., 2020)\", \"Explanation\": \"The cited work by Hedderich et al. (2020) discusses the use of few-shot cross-lingual transfer to boost performance in languages that are distant from the source language, which the citing paper extends by providing a more in-depth analysis of the approach in the context of English in-context learning.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Beltagy et al., 2022)\", \"Explanation\": \"The cited work by Beltagy et al. (2022) highlights the use of in-context learning in English for zero-or few-shot learning, which the citing paper further elaborates on in the context of in-context learning in English.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Schick and Sch\\u00fctze, 2021a)\", \"Explanation\": \"The cited work by Schick and Sch\\u00fctze (2021a) discusses the use of in-context learning in English for zero-or few-shot learning, which the citing paper builds upon to provide a more comprehensive overview of the research in this area.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Shin et al., 2020)\", \"Explanation\": \"The cited work by Shin et al. (2020) highlights the use of in-context learning in English for zero-or few-shot learning, which the citing paper further expands upon to provide a more in-depth analysis of the research in this area.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Lin et al., 2021)\", \"Explanation\": \"The cited work by Lin et al. provides a benchmark for few-shot cross-lingual transfer in a range of tasks and languages, which serves as a data source for the citing paper to build upon in their research on the same topic.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Xue et al., 2021)\", \"Explanation\": \"The cited work, mT5, is mentioned as a model that the citing paper uses for fine-tuning in the context of evaluating the state-of-the-art in multilingual large language models.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Muennighoff et al., 2022)\", \"Explanation\": \"The cited work, mT0, is mentioned as a model that the citing paper uses for fine-tuning in the context of evaluating the state-of-the-art in multilingual large language models.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Scao et al., 2022)\", \"Explanation\": \"The cited work, BLOOM, is mentioned as a model that the citing paper uses for fine-tuning in the context of evaluating the state-of-the-art in multilingual large language models.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Muennighoff et al., 2022)\", \"Explanation\": \"The cited work, BLOOMZ, is mentioned as a model that the citing paper uses for fine-tuning in the context of evaluating the state-of-the-art in multilingual large language models.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Ouyang et al., 2022)\", \"Explanation\": \"The cited work, ChatGPT, is mentioned as a model that the citing paper uses for in-context learning in the context of evaluating the state-of-the-art in multilingual large language models.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Blasi et al., 2022)\", \"Explanation\": \"The cited work by Blasi et al. provides a dataset that the citing paper uses in their research on the lack of annotated training data in many languages.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Yu et al., 2022)\", \"Explanation\": \"The cited work by Yu et al. contributes a dataset that the citing paper uses in their research on the lack of annotated training data in many languages.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Joshi et al., 2020)\", \"Explanation\": \"The cited work by Joshi et al. provides a dataset that the citing paper uses in their research on the lack of annotated training data in many languages.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lauscher et al., 2020)\", \"Explanation\": \"The cited work by Lauscher et al. serves as a methodological basis for the citing paper, as it focuses on the few-shot cross-lingual transfer paradigm that the paper is interested in.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Artetxe et al., 2020a)\", \"Explanation\": \"The cited work by Artetxe et al. provides a methodological basis for the citing paper, as it studies the zero-shot cross-lingual transfer from English that the paper is interested in.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Hu et al., 2020b)\", \"Explanation\": \"The cited work by Hu et al. contributes a methodological basis for the citing paper, as it studies the zero-shot cross-lingual transfer from English that the paper is interested in.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Hu et al., 2020b)\", \"Explanation\": \"The cited work, XTREME, is a multi-task benchmark that unifies diverse existing datasets for zero-shot transfer of models fine-tuned on English datasets, providing a foundational dataset for the citing paper to build upon.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ruder et al., 2021)\", \"Explanation\": \"The cited work, XTREME-R, is a multi-task benchmark that focuses on zero-shot transfer of models fine-tuned on English datasets, providing a data source for the citing paper to utilize in its research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Liang et al., 2020)\", \"Explanation\": \"The cited work, XGLUE, is a multi-task benchmark that focuses on zero-shot transfer of models fine-tuned on English datasets, providing a data source for the citing paper to use in its research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ye et al., 2021)\", \"Explanation\": \"The cited work, CrossFit, is a few-shot evaluation benchmark for English-based few-shot transfer, providing a data source for the citing paper to use in its research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhao et al., 2021)\", \"Explanation\": \"The cited work highlights the high performance variance in few-shot cross-lingual transfer, providing a data source for the citing paper to consider in its research and analysis.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ahuja et al., 2023)\", \"Explanation\": \"The cited work, MEGA, conducts experiments of few-shot cross-lingual transfer with a focus on classification and question answering tasks, extending the research in the citing paper to include a broader range of tasks.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Devlin et al., 2019)\", \"Explanation\": \"The cited work by Devlin et al. (2019) provides a foundational pre-trained model that enables the citing paper to adapt to new languages with no or few training instances in a target language.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Conneau et al., 2020a)\", \"Explanation\": \"The cited work by Conneau et al. (2020a) contributes to the citing paper by providing a pre-trained model that can adapt to new languages with no or few training instances in a target language.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Conneau et al., 2020b)\", \"Explanation\": \"The cited work by Conneau et al. (2020b) further supports the claim that multilingual pre-trained models can adapt to new languages with no or few training instances in a target language.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Hu et al., 2020b)\", \"Explanation\": \"The cited work by Hu et al. (2020b) provides evidence that multilingual pre-trained models can adapt to new languages with no or few training instances in a target language.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Wu and Dredze, 2019)\", \"Explanation\": \"The cited work by Wu and Dredze (2019) contributes to the citing paper by demonstrating the ability of multilingual pre-trained models to adapt to new languages with no or few training instances in a target language.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Lauscher et al., 2020)\", \"Explanation\": \"The cited work by Lauscher et al. (2020) provides evidence that fine-tuning multilingual pre-trained models on few-shot samples can improve performance in languages that are distant from the source language.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Hedderich et al., 2020)\", \"Explanation\": \"The cited work by Hedderich et al. (2020) further supports the claim that fine-tuning multilingual pre-trained models on few-shot samples can improve performance in languages that are distant from the source language.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Lin et al., 2021)\", \"Explanation\": \"The cited work by Lin et al. (2021) introduces pre-trained LMs trained on more multilingual pre-trained corpora or translated datasets and shows improved results in in-context learning.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Muennighoff et al., 2022)\", \"Explanation\": \"The cited work by Muennighoff et al. (2022) further supports the claim that in-context learning with pre-trained LMs trained on more multilingual pre-trained corpora or translated datasets can lead to improved results.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Shi et al., 2023)\", \"Explanation\": \"The cited work by Shi et al. (2023) provides a comprehensive evaluation of the use of instructions, demonstrations, and rationales in different languages across multiple reasoning tasks, which serves as a foundational reference for the citing paper to build upon in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Hu et al., 2020b)\", \"Explanation\": \"The cited work provides a dataset that the citing paper uses in their research to establish a cross-lingual benchmark for few-shot evaluations.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ruder et al., 2021)\", \"Explanation\": \"The cited work provides a dataset that the citing paper uses in their research to establish a cross-lingual benchmark for few-shot evaluations.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Liang et al., 2020)\", \"Explanation\": \"The cited work provides a dataset that the citing paper uses in their research to establish a cross-lingual benchmark for few-shot evaluations.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ponti et al., 2020)\", \"Explanation\": \"The cited work provides a dataset that the citing paper uses in their research to establish a cross-lingual benchmark for few-shot evaluations.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Conneau et al., 2018)\", \"Explanation\": \"The cited work provides a dataset that the citing paper uses in their research to establish a cross-lingual benchmark for few-shot evaluations.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Clark et al., 2020)\", \"Explanation\": \"The cited work by Clark et al. is the source of the undesired biases mentioned in the citing paper, which are translation artifacts and unnatural topic distributions.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Artetxe et al., 2020b)\", \"Explanation\": \"The cited work by Artetxe et al. is the source of the data that is used to collect translation-based datasets in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Raffel et al., 2020)\", \"Explanation\": \"The cited work by Raffel et al. serves as the basis for the text-to-text format used in the citing paper to unify all tasks into a common format for model training and evaluation.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Lin et al., 2021)\", \"Explanation\": \"The cited work by Lin et al. provides a method for translating instructions in target languages, which the citing paper adopts to translate instructions for 15 datasets in 54 target languages.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Shi et al., 2023)\", \"Explanation\": \"The cited work by Shi et al. also contributes to the translation of instructions in target languages, which the citing paper uses to translate the instructions into five languages.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Costa-juss\\u00e0 et al., 2022)\", \"Explanation\": \"The cited work by Costa-juss\\u00e0 et al. provides a translation model (NLLB) that the citing paper uses to translate instructions for 15 datasets in 54 target languages.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Conneau et al., 2018)\", \"Explanation\": \"The cited work by Conneau et al. (2018) is the source of the XNLI dataset, which is used in the Natural Language Inference task in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ebrahimi et al., 2022)\", \"Explanation\": \"The cited work by Ebrahimi et al. (2022) is the source of the AMER-ICASNLI dataset, which is used in the Natural Language Inference task in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Khashabi et al., 2021)\", \"Explanation\": \"The cited work by Khashabi et al. (2021) is the source of the PARSINLU-ENTAILMENT dataset, which is used in the Natural Language Inference task in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Park et al., 2021)\", \"Explanation\": \"The cited work by Park et al. (2021) is the source of the KLUE-NLI dataset, which is used in the Natural Language Inference task in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Hu et al., 2020a)\", \"Explanation\": \"The cited work by Hu et al. (2020a) is the source of the OCNLI dataset, which is used in the Natural Language Inference task in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Yang et al., 2019)\", \"Explanation\": \"The cited work provides the PAWS-X dataset used in the citing paper for few-shot training and validation data in sentiment analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Keung et al., 2020)\", \"Explanation\": \"The cited work provides the MULTILINGUAL AMAZON REVIEW DATASET used in the citing paper for binary sentiment analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Aggarwal et al., 2022)\", \"Explanation\": \"The cited work provides the INDICNLU-SENTIMENT dataset used in the citing paper for binary sentiment analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ponti et al., 2020)\", \"Explanation\": \"The cited work provides the XCOPA dataset used in the citing paper for commonsense reasoning.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Muennighoff et al., 2022)\", \"Explanation\": \"The cited work provides the XWINOGRAD dataset used in the citing paper for commonsense reasoning.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Clark et al., 2020)\", \"Explanation\": \"The cited work provides the dataset used for the question answering task in the citing paper, which is referred to as TYDIQA for simplicity.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Pan et al., 2017)\", \"Explanation\": \"The cited work provides the dataset used for the named entity recognition task in the citing paper, which is referred to as WIKIANN.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Adelani et al., 2021)\", \"Explanation\": \"The cited work provides the dataset used for the named entity recognition task in the citing paper, which is referred to as MASAKHANER.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Clark et al., 2020)\", \"Explanation\": \"The cited work provides the method for calculating the average score of a model given three different sets of k-shot instances, which the citing paper uses to calculate the per-language score for each dataset.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liang et al., 2020)\", \"Explanation\": \"The cited work provides the method for taking two average scores, which the citing paper adopts in its research to calculate the average class score of all classification and QA tasks and the average generation score of all generation tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Yu et al., 2022)\", \"Explanation\": \"The cited work by Yu et al. (2022) provides a wealth of labeled datasets in high-resource languages, which the citing paper leverages to support its research on transfer methods in low-resource languages.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Joshi et al., 2020)\", \"Explanation\": \"The cited work by Joshi et al. (2020) contributes to the citing paper by providing a cost-effective method for instruction annotation, which is used in the study of transfer methods in low-resource languages.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Devlin et al., 2019)\", \"Explanation\": \"The cited work by Devlin et al. provides the unlabeled corpora used in the pretraining strategy of language models.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. is a key reference for the conventional pretraining strategy of language models using unlabeled corpora.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Muennighoff et al., 2022)\", \"Explanation\": \"The cited work by Muennighoff et al. extends the study of instruction-tuning in language models by training LLMs on a massive number of tasks with instructions.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ouyang et al., 2022)\", \"Explanation\": \"The cited work by Ouyang et al. also contributes to the study of instruction-tuning in language models by training LLMs on a large number of tasks with instructions.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Wei et al., 2022)\", \"Explanation\": \"The cited work by Wei et al. further extends the study of instruction-tuning in language models by training LLMs on a large number of tasks with instructions.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xue et al., 2021)\", \"Explanation\": \"The cited work by Xue et al. provides the mT5-base model with 580 million parameters that is used in the fine-tuning experiments in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Scao et al., 2022)\", \"Explanation\": \"The cited work by Scao et al. introduces the BLOOM-7B model with 7 billion parameters that is used in the in-context learning experiments in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xue et al., 2021)\", \"Explanation\": \"The cited work by Xue et al. also provides the mT5-xxl model with 13 billion parameters that is used in the in-context learning experiments in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Muennighoff et al., 2022)\", \"Explanation\": \"The cited work provides the instruction-tuned variants of BLOOMZ-7B and mT0xxl that the citing paper uses in their experiments.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ouyang et al. 2022)\", \"Explanation\": \"The cited work is the current state-of-the-art ChatGPT (gpt-3.5-turbo) that the citing paper includes in their experiments.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Hu et al., 2020b)\", \"Explanation\": \"The cited work by Hu et al. provides a set of representative English datasets that the citing paper uses in its research on fine-tuning LMs for various tasks, including QA, NLI, paraphrase detection, summarization, XCOPA, XWINOGRAD, sentiment analysis, and question generation.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Muennighoff et al., 2022)\", \"Explanation\": \"The cited work by Muennighoff et al. provides a discussion on the effectiveness of instruction-tuning in zero-shot transfer, which the citing paper builds upon in their study of the same topic.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wei et al., 2022)\", \"Explanation\": \"The cited work by Wei et al. provides a method for instruction-tuning that the citing paper adopts in their research on the same topic.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Mishra et al., 2022)\", \"Explanation\": \"The cited work by Mishra et al. extends the research on instruction-tuning by exploring new dimensions and variables in the field.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Huang et al., 2023)\", \"Explanation\": \"The cited work by Huang et al. provides a method for prompt engineering in Chat-GPT that can improve its performance in summarization tasks. The citing paper adopts this method in their research to achieve better results in the evaluation of Chat-GPT.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhao et al., 2021)\", \"Explanation\": \"The cited work by Zhao et al. (2021) is used to support the claim that in-context learning is more sensitive to different demonstration samples than few-shot fine-tuning, which the citing paper builds upon in its research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. (2020) provides a method for in-context learning in few-shot cross-lingual transfer, which the citing paper adopts to improve the performance of the model in few-shot in-context learning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Lin et al., 2021)\", \"Explanation\": \"The cited work by Lin et al. (2021) reports on the results of classification tasks in few-shot in-context learning, which the citing paper extends by showing the effect of model scaling on the performance of the model in this context.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chung et al., 2022)\", \"Explanation\": \"The cited work by Chung et al. (2022) demonstrates the effectiveness of incorporating both instructions and demonstrations during instruction-tuning on English data, which the citing paper builds upon to improve the performance of their model in instruction-following.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Min et al., 2022a)\", \"Explanation\": \"The cited work by Min et al. (2022a) explores the issue of overfitting to specific formats in instruction-following models, which the citing paper further extends by focusing on cross-lingual transfer to improve few-shot in-context learning.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Chung et al., 2022)\", \"Explanation\": \"The cited work by Chung et al. (2022) provides a dataset of high-quality human-translated instructions that the citing paper utilizes in their research to improve the performance of their model in instruction-following.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ouyang et al., 2022)\", \"Explanation\": \"The cited work by Ouyang et al. has demonstrated the ability of Instruct-GPT to generate high-quality generations in English, which has prompted the citing paper to further explore the use of LLMs in data generation for under-represented languages.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Goyal et al., 2022)\", \"Explanation\": \"The cited work by Goyal et al. has shown that Instruct-GPT can outperform humans in some tasks, which has further motivated the citing paper to investigate the use of LLMs in data generation for a variety of typologically diverse languages.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang et al., 2022a)\", \"Explanation\": \"The cited work by Wang et al. has provided a method for generating task-specific data in certain target tasks, which the citing paper has adopted in its own research on data generation for under-represented languages.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Honovich et al., 2022)\", \"Explanation\": \"The cited work by Honovich et al. has contributed to the field of data generation in under-represented languages, which the citing paper has leveraged in its own research on the use of LLMs in data generation for a variety of typologically diverse languages.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Agrawal et al., 2022)\", \"Explanation\": \"The cited work by Agrawal et al. has explored the use of LLMs in data generation for question answering in certain target tasks, which the citing paper has extended to a wider range of typologically diverse languages.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Lu et al., 2022)\", \"Explanation\": \"The cited work by Lu et al. has examined the impact of prompt order in English in-context learning, which the citing paper has further extended to a variety of typologically diverse languages in its research on the use of LLMs in data generation.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Su et al., 2022)\", \"Explanation\": \"The cited work by Su et al. has proposed methods for identifying optimal demonstrations in English in-context learning, which the citing paper has extended to a variety of typologically diverse languages in its research on the use of LLMs in data generation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Aji et al., 2022)\", \"Explanation\": \"The cited work emphasizes the need for more studies on diverse local languages, including under-represented languages and their dialects, which the citing paper adopts in their research to address the challenges faced by LLMs in dealing with less-represented languages.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Kakwani et al., 2020)\", \"Explanation\": \"The cited work highlights the need for more research on under-represented languages and their dialects, which the citing paper extends by advocating for conducting more studies in this area to address the challenges faced by LLMs in dealing with less-represented languages.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Muennighoff et al., 2022)\", \"Explanation\": \"The cited work by Muennighoff et al. focuses on discriminative tasks, which the citing paper extends by expanding the evaluation to more diverse and complex tasks such as MTOP and MGMS8K.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Lin et al., 2021)\", \"Explanation\": \"The cited work by Lin et al. focuses on translation tasks, which the citing paper extends by expanding the evaluation to more diverse and complex tasks such as MTOP and MGMS8K.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Li et al., 2021)\", \"Explanation\": \"The cited work by Li et al. focuses on MTOP tasks, which the citing paper extends by expanding the evaluation to more diverse and complex tasks such as MTOP and MGMS8K.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Shi et al., 2023)\", \"Explanation\": \"The cited work by Shi et al. focuses on MGMS8K tasks, which the citing paper extends by expanding the evaluation to more diverse and complex tasks such as MTOP and MGMS8K.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Asai et al., 2021)\", \"Explanation\": \"The cited work by Asai et al. focuses on knowledge-intensive tasks, which the citing paper extends by expanding the evaluation to more diverse and complex tasks such as MTOP and MGMS8K.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Xue et al., 2021)\", \"Explanation\": \"The cited work provides the model checkpoint used in the fine-tuning process of the English model, which is a foundational element for the research conducted in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Muennighoff et al., 2022)\", \"Explanation\": \"The cited work is the source of the mT0 ZICL model used in the citing paper for training on PAWS-X data.\"}]"}}},{"rowIdx":85,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b37","b17","b32","b4","b61","b42","b54","b44","b57","b31"],"table_ref":[],"text":"Large language models (LLMs) have evolved considerably in size, architecture, and usage patterns since the introduction of ELMo (Peters et al., 2018), BERT (Devlin et al., 2019), and RoBERTa (Liu et al., 2019). BERT-style pretrained LLMs allow for rapid specialization to different tasks via supervised finetuning but often contain minimal zeroshot capability. Conversely, modern LLMs such as GPT-3 (Brown et al., 2020), PaLM (Chowdhery 1 https://github.com/anonymous et al., 2022), and LLaMA (Zhang et al., 2023) are challenging to exploit for task-specific finetuning but have proven highly capable at gradient-free specialization, e.g., in-context learning with zero-or few-shot data. Such practice is now the norm for using contemporary LLMs.\nWhile language models have grown in size, scale, and zero-shot capability, techniques for compressing such artifacts to smaller and more efficient models have not kept pace. Most works on compressing models via pruning (Sajjad et al., 2020;Voita et al., 2019;Chen et al., 2020b,a) or knowledge distillation (Sanh et al., 2019;Wang et al., 2020b;Liang et al., 2023) show success only for BERT-style models requiring task-specific finetuning. Whether these techniques can be adapted to contemporary GPT-style LLMs and evaluated with zero-shot prompting remains an open question.\nThe pipeline for compressing larger models should preserve the generality and zero-shot promptability of the original and avoid any taskspecific finetuning. Intuitively, this is a more challenging goal because task-specific compression often works at the expense of a model's generality. In contrast, our goal is to preserve the larger model's abilities. It is unclear if the methods designed for the former are also suitable for the latter.\nIn this work, we consider the open question of how to distill GPT-style large language models. We study the efficacy of state-of-the-art task-agnostic distillation methods with zero-shot evaluation on 12 diverse tasks. Results show that our simple teacher-free approach with continued pretraining matches or outperforms existing baselines while being 1.5x more computationally efficient.\nIn further experiments, we compare our bestcompressed model to an equally-sized model trained from scratch. Our zero-shot evaluation shows that the compressed model underperforms compared to a model trained from scratch. This opposes the trend from task-specialized compression, where training large then compressing is preferred over training a smaller model from scratch. Here we club together task-specific and task-agnostic compression with end-task finetuning under taskspecialized compression. We offer analysis and explanation for this phenomenon rooted in how much data is available for pretraining or finetuning stages. Our contributions are:\n• A characterization of how the benefits of distillation poorly transfer to a zero-shot setup.\n• A simple alternative compression method which is both performant (on perplexity and end-tasks) and efficient (1.5x faster training) at the LLM scale.\n• Insight into how typical compression techniques transfer poorly at scale with promising new angles for future methods to consider.\nWe argue that LLM compression methods should focus on zero-shot evaluation to encourage the development of new methods that maintain model generality. Our findings highlight that existing compression methods transfer poorly to large-scale general compression settings, and there is a larger gap between current methods and upperbound performance. We advocate for the future development of compression methods for LLMs to close this gap and promote model reuse and training efficiency."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b45","b20","b26","b20","b54","b34","b24","b33","b26","b20","b54","b33","b34","b49","b52","b36","b51","b59","b44","b25","b50","b31","b24","b33","b59","b31"],"table_ref":[],"text":"Pruning seeks to identify sub-networks within larger models which yield equivalent performance with increased computational efficiency. The parts of the model are can be pruned using a heuristic or scoring function (Sanh et al., 2020;Li et al., 2021a). Pruning can be described in three steps: (i) model initialization; (ii) training to specialize for relevant data; (iii) evaluation against the initial model. Pruning literature can broadly be divided between structured and unstructured methods.\nStructured pruning approaches (Fan et al., 2019;Kao et al., 2022) aim to prune larger blocks within a model while considering architecture nuances for improving inference speed. Structured pruning can be coarse-or fine-grained. Coarse-grained pruning removes entire model layers (Fan et al., 2019). In the case of LMs, fine-grain pruning methods prunes atomic components like attentionheads (Voita et al., 2019;Michel et al., 2019;Li et al., 2021b) or hidden layers (Hou et al., 2020;Chen et al., 2020b;McCarley et al., 2019). Another line of structured-pruning approach (Kao et al., 2022) aims to prune two out of every four weights in a matrix for 2:4 sparsity speedups on A100s. Unstructured pruning follows the lottery ticket hypothesis (Chen et al., 2020a), where the weights of a network are iteratively pruned without concern for structure-based speedups.\nThe other classification approach for pruning is based on task-agnostic vs. task-specific pruning. Task-agnostic pruning approaches (Chen et al., 2020b;Fan et al., 2019) prune a model on pretraining data and then add task specialization as a second step. Task-specific pruning (Voita et al., 2019;McCarley et al., 2019;Michel et al., 2019) methods specialize their models on end-task data during the pruning process.\nWe acknowledge a related line of work in quantization (Dettmers et al., 2022, inter alia) for model compression; however, quantization methods are orthogonal to our approach and can be applied after a model has been compressed using our method.\nKnowledge Distillation can be similarly divided between task-specific (Sun et al., 2019;Turc et al., 2019;Mukherjee et al., 2021;Tang et al., 2019;Xia et al., 2022) and task-agnostic methods (Sanh et al., 2019;Jiao et al., 2019;Sun et al., 2020;Wang et al., 2020b,a;Liang et al., 2023). For existing distillation literature on BERT-style models, we use task-specialized distillation for both methods. The reason for this is apparent in the case of task-specific methods, which distill task information from a teacher model into the student. However, even task-agnostic methods are specialized on end-tasks by finetuning following distillation on pretraining data, and hence the classification term for both categories.\nThere are methods that follow pruning with a distillation phase (Hou et al., 2020;McCarley et al., 2019). More recently, methods like Co-Fi (Xia et al., 2022) and Homotopic-distillation (Liang et al., 2023) combine the pruning phase with distillation by applying distillation losses during the pruning process."},{"figure_ref":[],"heading":"Training (Compressed) LLMs","publication_ref":["b25","b50","b31","b44"],"table_ref":[],"text":"We focus on commonly available distillation methodologies to study how standard practices (Sanh et al., 2019, inter alia) apply to our setting of compressing decoder-only LLMs. Unlike prior work, our compressed model is evaluated with zeroshot tasks without any end-task finetuning. This setting presents a more challenging situation than previously explored by distillation literature, wherein the methods benefit from the final stage of supervised task-specific finetuning. When it's assumed that supervised finetuning follows compression, the standard for compressed model quality can be much lower. For taskspecialized compression (Jiao et al., 2019;Sun et al., 2020;Liang et al., 2023;Sanh et al., 2019), reduced model capacity is needed compared to when models need to maintain generality. Also, parameters updated during the finetuning phase can compensate for task-specific information lost during compression. When optimizing for in-context learning, zero-shot evaluations will likely be more sensitive to removing parameters.\nOur principal research question is: How can we effectively compress GPT-style models to maintain zero-shot promptability? Within this question, we study the efficacy of distillation without finetuning for our class of models (described in Section 4). We consider how to initialize, prune, and distill a student model from an LLM. We describe our initialization strategies, distillation and language modeling objectives, and our straightforward methodology for learning a compressed model with continued pretraining."},{"figure_ref":[],"heading":"Language Modeling Objective","publication_ref":["b38","b53","b1"],"table_ref":[],"text":"Let X be a sequence of tokens composed of (x 1 , x 2 , ..., x n ). A language model computes the probability of the sequence by factorizing over the probability of each next token prediction given the prior context tokens:\np(X) = N i=1 p (x i |x 1 , x 2 , ..., x i-1 )(1)\nWe limit the scope of our work to causal autoregressive decoders (Radford et al., 2018), composed of Transformer decoder layers (Vaswani et al., 2017). This model, parameterized by θ, maximizes the log-likelihood of the next token prediction. The model is trained over N sequences from a pretraining corpus using a truncated window of ≤ k context tokens (Bengio et al., 2003). Equa-tion 2 defines this learning objective.\nL LM = log p θ (X) = N i=1 p(x i |x i-k , ..., x i-1 ; θ) (2)"},{"figure_ref":[],"heading":"Distillation Objectives","publication_ref":["b44","b57","b25","b27","b57","b25","b31","b31"],"table_ref":[],"text":"Task-agnostic distillation literature for \"BERTstyle\" models (Sanh et al., 2019;Wang et al., 2020b;Jiao et al., 2019) begins with a single-step layer-pruned initialization of the student. Following this, knowledge distillation uses one or more distillation objectives to align the output and/or intermediate states between the student and teacher. We first evaluate a \"vanilla\" distillation technique as a baseline, training a student with the language modeling objective combined with the distillation objective in Equation 3. L distill computes the Kullback-Leibler divergence (Kullback and Leibler, 1951) between the student logits, f s , and teacher logits, f t for a softmax output x i scaled by temperature τ temp .\nL distill = KL (f s (x i ) , f t (x i ))(3)\nWe consider the ideas introduced in miniLM (Wang et al., 2020b) and miniLMv2 (Wang et al., 2020a) for our second distillation baseline. Using the idea of relation transfer between Q, K, and V matrix products, we implement a Q -K and V -V relation transfer between the final layers of the student (s) and teacher (t). Equation 4outlines this signal for query-key relation transfer. Value-value relation transfer is equivalent but replaces Q and K with V . d is the dimensionality of each model. Using only two relation transfers follows the memoryperformance tradeoff from miniLMv2, and we adapt it to a causal-decoder setup by masking. The final loss function for this baseline contains the LM objective, the distillation objective from Eq. 3, and the two relation transfer KL-divergence terms.\nLQK = KL softmax QsKs √ ds , softmax QtKt √ dt(4)\nOur third and final baseline draws from tiny-BERT (Jiao et al., 2019) and homotopic distillation (Liang et al., 2023). This baseline aligns the intermediate states (L hid ), embeddings (L emb ), and attention outputs (L att ) at each layer between the student and the teacher using a mean-squared error (MSE) objective. These objectives work along the language modeling loss and KL-divergence-based distillation objective from Eq. 3. Following Liang et al. (2023), we align each student layer to the original teacher layer before pruning. These loss functions are summarized in Equations 5-7 where H k and A k are the hidden states and attention output respectively for layer k for teacher (t) and student (s). E i are the respective embedding tables. Since we do not prune in the hidden dimensions of the student, there is no need to learn a projection to match the dimensionality between the models. Equation 8 is the complete loss function.\nL hid (θ s , θ t ) = K k=1 MSE H k t , H k s (5) L emb (θ s , θ t ) = MSE (E t , E s ) (6) L att (θ s , θ t ) = K k=1 MSE A k t , A k s (7) L Σ = L LM + L distill + L hid + L emb + L att (8)"},{"figure_ref":["fig_0"],"heading":"Truncated Initialization Strategy","publication_ref":["b44"],"table_ref":[],"text":"The goal of initialization in a compression pipeline is choosing a subset of larger model parameters to retain maximum information about the data. We follow conventional distillation work (Sanh et al., 2019) in using a layer-level pruning strategy. This produces a truncated initialization of a subset of teacher layers for the student. Choosing where and when to prune layers is a critical design decision to maximize performance recovery after pruning. Figure 1 outlines the configurations we consider studying where to prune half the model layers. Considering when to enact pruning primarily concerns either pruning all layers at initialization or incrementally removing layers periodically during continued pretraining (described in Section 3.4). Our intuition here is that incremental layer removal yields improved training stability. We present results and ablations for these two questions in Section 5."},{"figure_ref":[],"heading":"Continued Pretraining","publication_ref":[],"table_ref":[],"text":"While we explore distillation in our setup, the study of introducing additional student-teacher alignment is somewhat saturated. Unlike prior work, we observe that the additional distillation signals (outlined above) do not significantly contribute to the student learning process. In fact, in some instances, they prove to be detrimental. We hypothesize that this supervision over-constrains the student to be unable to adequately model the data and recover performance lost during layer pruning. Therefore, we consider the alternative, continued pretraining of the student without teacher supervision. We remove all distillation objectives and instead continue pretraining the student, initialized with a truncated subset of layers, with the same corpus and objective as the teacher model. Our proposal resembles domain adaptive pretraining (Gururangan et al., 2020); however, our goal of task-agnostic compression requires domain-agnostic continued pretraining to maintain model generality. Section 5 highlights that this strategy is superior to distillation objectives given a limited token budget.\n4 Experimental Setup"},{"figure_ref":[],"heading":"Base Models","publication_ref":["b62","b61","b9","b39","b46","b0","b14","b31","b40"],"table_ref":["tab_0"],"text":"We pretrain our own baseline models for exact control over token budgets and model configurations2 . However, our proposed compression method can be applied to any publicly available decoder-only model checkpoint e.g., Zhang et al. (2022Zhang et al. ( , 2023)). We follow the PaLM architecture (Chowdhery et al., 2022) owing to improved throughput efficiency. Specifically, the attention and feed-forward network (FFN) modules are parallel instead of sequential (Radford et al., 2019). SwiGLU activation (Shazeer, 2020) is used in the FFN module. Multihead attention uses the equivalent Flash-Attention (Dao et al., 2022, FA) implementation. The first layer of the FFN module and the layers generating attention query, key, and value are fused. Similarly, the second layer of the FFN module and the feed-forward layer after the attention operation are fused. The LayerNorm (Ba et al., 2016) is before the first fused feed-forward layer. The query and the key vectors are passed through additional layer normalization layers for increased training stability following Dehghani et al. (2023). This block structure is repeated with skip connections to form our decoder-only Transformer architecture.\nUsing the language modeling objective, we train two core baselines with 300 million (M) and 1.1 billion (B) parameters. We also consider additional baseline models containing half the number of layers (180M and 610M parameters). We use Lion optimizer (Chen et al., 2023) in all our pretraining experiments with β parameters set to (0.9, 0.95). Weight decay is 1 × 10 -4 but is omitted for bias and normalization parameters. 3 We schedule learning rate warm-up from the start to 2000 steps (4B tokens) and then use a cosine decay schedule until training stops. The final learning rate is 10% of the peak value. Table 1 summarizes all baseline model configurations and their respective pretraining token budgets. For pretraining, we use the C4 dataset (Raffel et al., 2020) of 160B web-crawled tokens. Pretrained models of all sizes are trained for 1 epoch on the C4 dataset."},{"figure_ref":[],"heading":"Compression and Distillation Baselines","publication_ref":[],"table_ref":[],"text":"To simulate the typical practice of a compressed model continuing to train with a pretrained optimizer state, we seed the model with training for 1B additional tokens to \"warm-start\" the optimizer. This seed phase includes a linear warm-up of the learning rate for 400 steps (800M tokens). After this, we begin layer pruning according to the truncation configuration, controlling when and where we prune layers from the model. A layer that needs to be removed in the future is frozen from the start of the optimization process.\nWe use the Lion optimizer with the same batch size, weight decay, and β as the above pretraining configuration. However, the peak learning rate is fixed at 10% of the peak learning rate during the"},{"figure_ref":[],"heading":"Category Task Metric","publication_ref":["b3","b60","b43","b10"],"table_ref":[],"text":"Common Sense Reasoning PIQA (Bisk et al., 2019) ACC (LN) Hellaswag (Zellers et al., 2019) ACC (LN) Winogrande (Sakaguchi et al., 2019) ACC BoolQ (Clark et al., 2019) ACC (LN)"},{"figure_ref":[],"heading":"Science Question Answering","publication_ref":["b35","b58","b11"],"table_ref":[],"text":"OpenBookQA (Mihaylov et al., 2018) ACC (LN) SciQ (Welbl et al., 2017) ACC Arc-Easy (Clark et al., 2018) ACC Arc-Challenge (Clark et al., pretraining phase. For the 300M model, this is 3 × 10 -5 ; for the 1.1B model, this is 2 × 10 -5 .\nFor distillation baselines, we use the same truncated initialization strategy for student models as our compression method. Each loss component of the total distillation loss is normalized such that the total magnitude for distillation loss matches the cross-entropy loss. We note that some forms of knowledge distillation cannot exploit recent modeling optimizations, such as flash-attention. Aligning QK T outputs for distillation at every layer requires the additional computation of O(n 2 ) matrix products, significantly slowing down training."},{"figure_ref":[],"heading":"Downstream Evaluation","publication_ref":["b21","b22"],"table_ref":["tab_2"],"text":"We evaluate our model on 12 tasks from 5 categories: common sense reasoning, science question answering, causal reasoning, natural language inference, and paraphrase identification. All tasks are evaluated in a zero-shot setting by providing the language model with a prompt from Eleuther-AI evaluation harness (Gao et al., 2021) and a possible completion. We score the model output for each completion. The completion with the highest likelihood is the prediction to compute task accuracy. The completion likelihood can be normalized by either the character count in the completion (Gao et al., 2021, length normalized accuracy) or by the probability of completion conditioned on the domain premise (Holtzman et al., 2021;Brown et al., 2020, domain Table 3: Results for our model and distillation baselines from Sec. 3.2 for language model perplexity and all tasks outlined in Table 2. All models are evaluated zero-shot with no finetuning. For extrinsic tasks, higher is better, while lower is better for perplexity (ppl.). \"Pre-compression\" is the larger teacher model evaluated on the same task suite.\nOur method yields superior perplexity compared to distillation and is competitive on most tasks."},{"figure_ref":[],"heading":"Compression Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Compression for generality and zero-shot promptability","publication_ref":[],"table_ref":["tab_5"],"text":"This section compares our method against the distillation baselines mentioned in Section 3.2. We show perplexity results on C4 validation set and zero-shot performance on 12 end-tasks. All baselines and our method are applied to two core model baselines mentioned in Section 4.1: a 300M parameter model and a 1.1B parameter model, both trained until 160B tokens. In addition, we provide the end-task performance of the checkpoints themselves for comparison against distillation baselines and our method. All compression methods in this section remove 12/24 layers in the decoder-only base models. All methods are trained on additional 20B tokens from the C4 corpus. As shown in Table 3, our method matches the best distillation baseline on perplexity and remains competitive on most end-tasks. Table 4 summarizes training statistics like computation required by each method in terms of FLOPs or the wall-clock time consumed. This result shows that our method is at par with the distillation baselines in our implementation with significantly less compute and wall-clock time. Section 7 discusses our hypothesis on why we expect our distillation-free method to work well despite being trained on a less-informative training signal without a teacher."},{"figure_ref":["fig_0"],"heading":"Where to Prune a Model?","publication_ref":[],"table_ref":["tab_2"],"text":"Our baseline models contain 24 decoder layers.\nTo determine which layers we should prune for the best compression performance, we define five layer pruning configurations shown in Figure 1, each removing 12 out of the 24 decoder layers of the 300M and 1.1B models. In all these pruning configurations, we always keep the first and the last layers because they interact with the embedding table. We made this design choice based on early experiments. Table 5 summarizes the results of this ablation. We report the perplexity score on the C4 validation set and the average task accuracy across 12 tasks in our evaluation suite (Table 2).\nFor the base 300M model, pruning configurations of max-gap and both perform the best out of the five possible configurations. For the 1.1B model, pruning layers from the input configuration yielded the best results for both reported metrics. The output pruning configuration resulted in the worst performance across model sizes, suggesting that pruning layers towards the output side of the model should be avoided. Given these results, we use the pruning configuration of max-gap for all our 300M model experiments and the configuration of input for all our 1.1B model experiments."},{"figure_ref":["fig_0"],"heading":"When to Prune a Model?","publication_ref":[],"table_ref":["tab_2"],"text":"To answer the question of when we should prune layers from our model while continuing to train it with language modeling loss, we can decide in one of two ways: either remove selected layers at once or remove them one by one, each after a fixed number of training tokens. We run this experiment in four configurations to see if increasing the Table 5: Influence on average task performance and C4 validation perplexity of different truncated initialization strategies from Figure 1 for models of size 300M and 1.1B. The average task performance score is across 12 tasks listed in Table 2, and higher numbers are better. For perplexity scores, lower is better.\ngap between each layer pruning increases training stability or model performance. The four configurations are: dropping all layers at once (0M token gap between pruning each layer) or pruning them after 100M, 500M, and 1B training tokens each. We run this experiment for the 300M and 1.1B model sizes.\nWe prune 12/24 layers from our decoder-only models for this ablation, each at a training token gap mentioned in one of the four configurations above. The result of this experiment is summarized in Fig- ure 2. Pruning layers one by one with an increasing token budget between each layer pruning does not benefit the average task accuracy or C4 validation perplexity. In fact, there is a marginal preference to prune layers as early into the training as possible. Hence, we decided to prune all 12/24 layers simultaneously for other experiments."},{"figure_ref":["fig_2"],"heading":"Upweighting Distillation Loss","publication_ref":[],"table_ref":["tab_5","tab_2"],"text":"Our vanilla distillation baseline described in Section 3.2 performs at par with our proposed approach of model pruning, followed by continued pretraining. Any distillation setup adds a teacher and a distillation loss component at the minimum to the compression pipeline. Because of this, the vanilla distillation baseline is at least 1.6x slower than our proposed method in terms of theoretical FLOPs and wall-clock time (Table 4). To justify the additional compute requirements of a distillation setup, we want to verify further the importance of the teacher and the distillation loss component. We maintain the setup from our vanilla distillation baseline for this experiment and append a coefficient to the KL-divergence loss term. The new loss function for distillation becomes L Σ = L LM + λ * L distill , where L distill is same as Eq. 3. We ablate the value of λ on our 1.1B model and report results on the C4 validation perplexity and the average task accuracy in Figure 3. It is evident from this experiment that increasing the value of λ results in a decrease in the end-task per- 2) and perplexity on the C4 validation set for model sizes 300M and 1.1B comparing schedules for when to prune layers during continued pretraining. We find a marginal performance degradation as we remove layers one by one further apart during continued pretraining.\nformance and an increase in C4 perplexity. Thus, this experiment highlights the importance of the language modeling component in our vanilla distillation baseline and strengthens the argument that appending the continued pretraining setup with a distillation component to get similar performance is redundant."},{"figure_ref":[],"heading":"Removing Language Modeling","publication_ref":["b57","b31"],"table_ref":["tab_2"],"text":"Our other two distillation baselines described in Section 3.2 constrain the teacher-teacher setup in more ways than just using a KL-divergence term for distillation. The miniLM-KD (Wang et al., 2020b) baseline adds constraints on the attention map and intermediate representations of the final layer between the teacher and the student model on top of the loss terms in the vanilla distillation setup. The homotopic-KD (Liang et al., 2023) baseline takes a step further and adds constraints on each layer between the teacher and the student on attention maps and intermediate representations. These setups are well designed to align student outputs to the teacher perfectly; thus, we ablate on whether they even require the language modeling component for the student.\nIn this experiment, we maintain the same setup of our miniLM and homotopic distillation base- 2) and C4 validation perplexity. lines without the language modeling loss term. We show the results in Figure 4 for our 1.1B model. Both distillation baselines lose some end-task performance by removing the language modeling loss term. However, the language modeling task performance on the C4 validation set suffers significantly without the language modeling loss term on the student."},{"figure_ref":["fig_3","fig_4"],"heading":"Train Large then Compress, or Not","publication_ref":["b30","b47"],"table_ref":["tab_2"],"text":"Existing literature (Li et al., 2020) suggests that training a large model and then compressing it usually works better than training a smaller model Figure 4: Experiment demonstrating the effectiveness of language modeling loss component in compound loss functions of miniLM-KD (Sec. 3.2) and homotopic-KD distillation (Sec. 3.2) baselines. We report the average task performance across 12 tasks (Table 2) and perplexity on the C4 validation set. from scratch. In this section, we study this observation in the context of zero-shot evaluation of the compressed model. As mentioned in Section 4.1, in addition to our 300M/1.1B baselines, we train two half-sized models (180M/610M) from scratch on the same pretraining token budget as their larger counterparts.\nFigure 5 compares the performance of our source LLM models (\"full @ max-budget\"), half-sized pretrained models to the same token budget (\"half @ max-budget\"), compressed models from the source LLM with a token budget of 20B (\"compressed @ 20B\") and half-sized pretrained models with the token budget of our compression method (\"half @ 20B\"). Our compressed models outperform the half-sized models pretrained from scratch when the token budget is fixed at the level of the compression budget. As we further pretrain our half-sized models, they overtake the comparable compressed model in both zero-shot task evaluation and language modeling perplexity.\nFigure 6 shows how a billion-sized compressed model starts to flatten and is overtaken by a pretrained half-sized model. This suggests that while continued pretraining is a good way to recover some of the model's zero-shot abilities, truncated initialization due to layer pruning is not very effective at preserving these abilities in the first place.\nThe ideal truncated initialization method should satisfy two requirements: (1) loss after truncation is the same (or as close as possible) to the loss before truncation, (2) as we continue training the truncated model, its learning curve continues improving at the same rate as if it was trained from scratch. Using the terminogloy from Shen et al. (2022), it should be \"loss-preserving\" and \"trainingdynamics-preserving\". Although achieving both requirements might not be possible, they define the upper bound. We continue to explore the space of pruning strategies for future work."},{"figure_ref":[],"heading":"Discussion","publication_ref":[],"table_ref":[],"text":"The results in sections 5 and 6 showed that existing methods for task-specialized compression are not necessarily the best in the zero-shot evaluation setting. We found that simple continued pretraining outperforms student/teacher distillation and that training a smaller model from scratch is better than training a larger model and then compressing. This section summarizes a few hypotheses explaining these findings and discusses why zero-shot evaluation favors different compression methods."},{"figure_ref":[],"heading":"Non-ideal truncated initialization","publication_ref":["b47"],"table_ref":[],"text":"To preserve a model's zero-shot promptability, we need a truncated initialization method that's both losspreserving and training-dynamics preserving Shen et al. (2022). However, the results in section 6 show that the truncated initialization methods cause the compressed model to lose too much of its zero-shot abilities."},{"figure_ref":[],"heading":"Size of training data","publication_ref":[],"table_ref":[],"text":"We hypothesize that the reason distillation does not lead to improved zeroshot performance is related to the training data size. In task-specific finetuning setup, where the task data is relatively limited, the larger model can learn more generalizable representations from the data and can provide a more informative training signal to the smaller student model, compared to what the student model can learn itself from data. In contrast, in the zero-shot evaluation setup, where we have unlimited pretraining data, distilling from a larger model does not provide much advantage when we continue to learn from billions of tokens.\nAligning the pretraining objective Distillation uses a collection of training objectives that constrain the smaller model in various ways. This is a reasonable setup when the goal is to imitate the larger model as much as possible. However, in an infinite data regime with the goal of training a zero-shot model, a single language modeling objective is better aligned with the goal. It is more capable of eliciting zero-shot behavior compared to distillation."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"We examine compression techniques to preserve the zero-shot promptability of large language models. We proposed a task-agnostic compression methodology, with no teacher, using pruning for truncated initialization of the compressed model, and continued pretraining as an alternative to endtask finetuning. On a diverse zero-shot evaluation suite and perplexity, our method is comparable in performance to a range of distillation baselines while offering 1.5× compute efficient training. We also compare compression to an \"oracle\" setup of an equally sized model pretrained for the same token budget as the larger teacher model. We highlight that a broad range of compression methods underperform compared to an uncompressed model with more training. This surprising result that distillation is not the \"silver bullet\" here, as it has been in prior works, leads us to call for future research into closing this gap. We release our code and evaluation setup to encourage future research and discussion on improved compression strategies."},{"figure_ref":[],"heading":"Limitations","publication_ref":[],"table_ref":[],"text":"While our work is in the spirit of reducing model size and improving efficiency -we require significant computational resources for our experiments demanding both high energy usage and processing power. Experiments such as the teacher model pretraining and \"oracle\" student model upper-bound demand upto 5 days of training time using 32xA100 GPUs with a high bandwidth interconnect. Therefore, reproducing our experiments are only reasonably tractable with commensurate GPU resources which may be infeasible for some researchers.\nAdditionally, we demonstrate our findings compared to a 'vanilla' distillation approach and recently published alternatives in our decoder-only setup. We take this approach to report how the most typical distillation strategy can be ported to a contemporary LLM. Our findings do not indicate that distillation is potentially fruitful for GPT-style models, however, our work is limited in that there may exist some atypical distillation strategy with even better performance. We encourage future work and discussion of how these methods can be improved in this regard."},{"figure_ref":[],"heading":"Ethics Statement","publication_ref":["b18","b23","b48"],"table_ref":[],"text":"We report all pretraining experiments with the widely used C4 corpus. This corpus has been found to contain harmful artifacts and biases (Dodge et al., 2021) which our models may inherit, however, the study of this phenomena is outside of the scope of our work but may inform future study. Model compression has been linked to increased bias and toxicity in a model (Hooker et al., 2020) but it is currently unclear how such effects extend to our setting; particularly as we expose the student to the same corpus as the teacher. Further study is needed in this area to examine how compression influences biases in increasingly large language models (Solaiman et al., 2023)."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b37\",\n \"b17\",\n \"b32\",\n \"b4\",\n \"b61\",\n \"b42\",\n \"b54\",\n \"b44\",\n \"b57\",\n \"b31\"\n ],\n \"table_ref\": [],\n \"text\": \"Large language models (LLMs) have evolved considerably in size, architecture, and usage patterns since the introduction of ELMo (Peters et al., 2018), BERT (Devlin et al., 2019), and RoBERTa (Liu et al., 2019). BERT-style pretrained LLMs allow for rapid specialization to different tasks via supervised finetuning but often contain minimal zeroshot capability. Conversely, modern LLMs such as GPT-3 (Brown et al., 2020), PaLM (Chowdhery 1 https://github.com/anonymous et al., 2022), and LLaMA (Zhang et al., 2023) are challenging to exploit for task-specific finetuning but have proven highly capable at gradient-free specialization, e.g., in-context learning with zero-or few-shot data. Such practice is now the norm for using contemporary LLMs.\\nWhile language models have grown in size, scale, and zero-shot capability, techniques for compressing such artifacts to smaller and more efficient models have not kept pace. Most works on compressing models via pruning (Sajjad et al., 2020;Voita et al., 2019;Chen et al., 2020b,a) or knowledge distillation (Sanh et al., 2019;Wang et al., 2020b;Liang et al., 2023) show success only for BERT-style models requiring task-specific finetuning. Whether these techniques can be adapted to contemporary GPT-style LLMs and evaluated with zero-shot prompting remains an open question.\\nThe pipeline for compressing larger models should preserve the generality and zero-shot promptability of the original and avoid any taskspecific finetuning. Intuitively, this is a more challenging goal because task-specific compression often works at the expense of a model's generality. In contrast, our goal is to preserve the larger model's abilities. It is unclear if the methods designed for the former are also suitable for the latter.\\nIn this work, we consider the open question of how to distill GPT-style large language models. We study the efficacy of state-of-the-art task-agnostic distillation methods with zero-shot evaluation on 12 diverse tasks. Results show that our simple teacher-free approach with continued pretraining matches or outperforms existing baselines while being 1.5x more computationally efficient.\\nIn further experiments, we compare our bestcompressed model to an equally-sized model trained from scratch. Our zero-shot evaluation shows that the compressed model underperforms compared to a model trained from scratch. This opposes the trend from task-specialized compression, where training large then compressing is preferred over training a smaller model from scratch. Here we club together task-specific and task-agnostic compression with end-task finetuning under taskspecialized compression. We offer analysis and explanation for this phenomenon rooted in how much data is available for pretraining or finetuning stages. Our contributions are:\\n• A characterization of how the benefits of distillation poorly transfer to a zero-shot setup.\\n• A simple alternative compression method which is both performant (on perplexity and end-tasks) and efficient (1.5x faster training) at the LLM scale.\\n• Insight into how typical compression techniques transfer poorly at scale with promising new angles for future methods to consider.\\nWe argue that LLM compression methods should focus on zero-shot evaluation to encourage the development of new methods that maintain model generality. Our findings highlight that existing compression methods transfer poorly to large-scale general compression settings, and there is a larger gap between current methods and upperbound performance. We advocate for the future development of compression methods for LLMs to close this gap and promote model reuse and training efficiency.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b45\",\n \"b20\",\n \"b26\",\n \"b20\",\n \"b54\",\n \"b34\",\n \"b24\",\n \"b33\",\n \"b26\",\n \"b20\",\n \"b54\",\n \"b33\",\n \"b34\",\n \"b49\",\n \"b52\",\n \"b36\",\n \"b51\",\n \"b59\",\n \"b44\",\n \"b25\",\n \"b50\",\n \"b31\",\n \"b24\",\n \"b33\",\n \"b59\",\n \"b31\"\n ],\n \"table_ref\": [],\n \"text\": \"Pruning seeks to identify sub-networks within larger models which yield equivalent performance with increased computational efficiency. The parts of the model are can be pruned using a heuristic or scoring function (Sanh et al., 2020;Li et al., 2021a). Pruning can be described in three steps: (i) model initialization; (ii) training to specialize for relevant data; (iii) evaluation against the initial model. Pruning literature can broadly be divided between structured and unstructured methods.\\nStructured pruning approaches (Fan et al., 2019;Kao et al., 2022) aim to prune larger blocks within a model while considering architecture nuances for improving inference speed. Structured pruning can be coarse-or fine-grained. Coarse-grained pruning removes entire model layers (Fan et al., 2019). In the case of LMs, fine-grain pruning methods prunes atomic components like attentionheads (Voita et al., 2019;Michel et al., 2019;Li et al., 2021b) or hidden layers (Hou et al., 2020;Chen et al., 2020b;McCarley et al., 2019). Another line of structured-pruning approach (Kao et al., 2022) aims to prune two out of every four weights in a matrix for 2:4 sparsity speedups on A100s. Unstructured pruning follows the lottery ticket hypothesis (Chen et al., 2020a), where the weights of a network are iteratively pruned without concern for structure-based speedups.\\nThe other classification approach for pruning is based on task-agnostic vs. task-specific pruning. Task-agnostic pruning approaches (Chen et al., 2020b;Fan et al., 2019) prune a model on pretraining data and then add task specialization as a second step. Task-specific pruning (Voita et al., 2019;McCarley et al., 2019;Michel et al., 2019) methods specialize their models on end-task data during the pruning process.\\nWe acknowledge a related line of work in quantization (Dettmers et al., 2022, inter alia) for model compression; however, quantization methods are orthogonal to our approach and can be applied after a model has been compressed using our method.\\nKnowledge Distillation can be similarly divided between task-specific (Sun et al., 2019;Turc et al., 2019;Mukherjee et al., 2021;Tang et al., 2019;Xia et al., 2022) and task-agnostic methods (Sanh et al., 2019;Jiao et al., 2019;Sun et al., 2020;Wang et al., 2020b,a;Liang et al., 2023). For existing distillation literature on BERT-style models, we use task-specialized distillation for both methods. The reason for this is apparent in the case of task-specific methods, which distill task information from a teacher model into the student. However, even task-agnostic methods are specialized on end-tasks by finetuning following distillation on pretraining data, and hence the classification term for both categories.\\nThere are methods that follow pruning with a distillation phase (Hou et al., 2020;McCarley et al., 2019). More recently, methods like Co-Fi (Xia et al., 2022) and Homotopic-distillation (Liang et al., 2023) combine the pruning phase with distillation by applying distillation losses during the pruning process.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Training (Compressed) LLMs\",\n \"publication_ref\": [\n \"b25\",\n \"b50\",\n \"b31\",\n \"b44\"\n ],\n \"table_ref\": [],\n \"text\": \"We focus on commonly available distillation methodologies to study how standard practices (Sanh et al., 2019, inter alia) apply to our setting of compressing decoder-only LLMs. Unlike prior work, our compressed model is evaluated with zeroshot tasks without any end-task finetuning. This setting presents a more challenging situation than previously explored by distillation literature, wherein the methods benefit from the final stage of supervised task-specific finetuning. When it's assumed that supervised finetuning follows compression, the standard for compressed model quality can be much lower. For taskspecialized compression (Jiao et al., 2019;Sun et al., 2020;Liang et al., 2023;Sanh et al., 2019), reduced model capacity is needed compared to when models need to maintain generality. Also, parameters updated during the finetuning phase can compensate for task-specific information lost during compression. When optimizing for in-context learning, zero-shot evaluations will likely be more sensitive to removing parameters.\\nOur principal research question is: How can we effectively compress GPT-style models to maintain zero-shot promptability? Within this question, we study the efficacy of distillation without finetuning for our class of models (described in Section 4). We consider how to initialize, prune, and distill a student model from an LLM. We describe our initialization strategies, distillation and language modeling objectives, and our straightforward methodology for learning a compressed model with continued pretraining.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Language Modeling Objective\",\n \"publication_ref\": [\n \"b38\",\n \"b53\",\n \"b1\"\n ],\n \"table_ref\": [],\n \"text\": \"Let X be a sequence of tokens composed of (x 1 , x 2 , ..., x n ). A language model computes the probability of the sequence by factorizing over the probability of each next token prediction given the prior context tokens:\\np(X) = N i=1 p (x i |x 1 , x 2 , ..., x i-1 )(1)\\nWe limit the scope of our work to causal autoregressive decoders (Radford et al., 2018), composed of Transformer decoder layers (Vaswani et al., 2017). This model, parameterized by θ, maximizes the log-likelihood of the next token prediction. The model is trained over N sequences from a pretraining corpus using a truncated window of ≤ k context tokens (Bengio et al., 2003). Equa-tion 2 defines this learning objective.\\nL LM = log p θ (X) = N i=1 p(x i |x i-k , ..., x i-1 ; θ) (2)\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Distillation Objectives\",\n \"publication_ref\": [\n \"b44\",\n \"b57\",\n \"b25\",\n \"b27\",\n \"b57\",\n \"b25\",\n \"b31\",\n \"b31\"\n ],\n \"table_ref\": [],\n \"text\": \"Task-agnostic distillation literature for \\\"BERTstyle\\\" models (Sanh et al., 2019;Wang et al., 2020b;Jiao et al., 2019) begins with a single-step layer-pruned initialization of the student. Following this, knowledge distillation uses one or more distillation objectives to align the output and/or intermediate states between the student and teacher. We first evaluate a \\\"vanilla\\\" distillation technique as a baseline, training a student with the language modeling objective combined with the distillation objective in Equation 3. L distill computes the Kullback-Leibler divergence (Kullback and Leibler, 1951) between the student logits, f s , and teacher logits, f t for a softmax output x i scaled by temperature τ temp .\\nL distill = KL (f s (x i ) , f t (x i ))(3)\\nWe consider the ideas introduced in miniLM (Wang et al., 2020b) and miniLMv2 (Wang et al., 2020a) for our second distillation baseline. Using the idea of relation transfer between Q, K, and V matrix products, we implement a Q -K and V -V relation transfer between the final layers of the student (s) and teacher (t). Equation 4outlines this signal for query-key relation transfer. Value-value relation transfer is equivalent but replaces Q and K with V . d is the dimensionality of each model. Using only two relation transfers follows the memoryperformance tradeoff from miniLMv2, and we adapt it to a causal-decoder setup by masking. The final loss function for this baseline contains the LM objective, the distillation objective from Eq. 3, and the two relation transfer KL-divergence terms.\\nLQK = KL softmax QsKs √ ds , softmax QtKt √ dt(4)\\nOur third and final baseline draws from tiny-BERT (Jiao et al., 2019) and homotopic distillation (Liang et al., 2023). This baseline aligns the intermediate states (L hid ), embeddings (L emb ), and attention outputs (L att ) at each layer between the student and the teacher using a mean-squared error (MSE) objective. These objectives work along the language modeling loss and KL-divergence-based distillation objective from Eq. 3. Following Liang et al. (2023), we align each student layer to the original teacher layer before pruning. These loss functions are summarized in Equations 5-7 where H k and A k are the hidden states and attention output respectively for layer k for teacher (t) and student (s). E i are the respective embedding tables. Since we do not prune in the hidden dimensions of the student, there is no need to learn a projection to match the dimensionality between the models. Equation 8 is the complete loss function.\\nL hid (θ s , θ t ) = K k=1 MSE H k t , H k s (5) L emb (θ s , θ t ) = MSE (E t , E s ) (6) L att (θ s , θ t ) = K k=1 MSE A k t , A k s (7) L Σ = L LM + L distill + L hid + L emb + L att (8)\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Truncated Initialization Strategy\",\n \"publication_ref\": [\n \"b44\"\n ],\n \"table_ref\": [],\n \"text\": \"The goal of initialization in a compression pipeline is choosing a subset of larger model parameters to retain maximum information about the data. We follow conventional distillation work (Sanh et al., 2019) in using a layer-level pruning strategy. This produces a truncated initialization of a subset of teacher layers for the student. Choosing where and when to prune layers is a critical design decision to maximize performance recovery after pruning. Figure 1 outlines the configurations we consider studying where to prune half the model layers. Considering when to enact pruning primarily concerns either pruning all layers at initialization or incrementally removing layers periodically during continued pretraining (described in Section 3.4). Our intuition here is that incremental layer removal yields improved training stability. We present results and ablations for these two questions in Section 5.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Continued Pretraining\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"While we explore distillation in our setup, the study of introducing additional student-teacher alignment is somewhat saturated. Unlike prior work, we observe that the additional distillation signals (outlined above) do not significantly contribute to the student learning process. In fact, in some instances, they prove to be detrimental. We hypothesize that this supervision over-constrains the student to be unable to adequately model the data and recover performance lost during layer pruning. Therefore, we consider the alternative, continued pretraining of the student without teacher supervision. We remove all distillation objectives and instead continue pretraining the student, initialized with a truncated subset of layers, with the same corpus and objective as the teacher model. Our proposal resembles domain adaptive pretraining (Gururangan et al., 2020); however, our goal of task-agnostic compression requires domain-agnostic continued pretraining to maintain model generality. Section 5 highlights that this strategy is superior to distillation objectives given a limited token budget.\\n4 Experimental Setup\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Base Models\",\n \"publication_ref\": [\n \"b62\",\n \"b61\",\n \"b9\",\n \"b39\",\n \"b46\",\n \"b0\",\n \"b14\",\n \"b31\",\n \"b40\"\n ],\n \"table_ref\": [\n \"tab_0\"\n ],\n \"text\": \"We pretrain our own baseline models for exact control over token budgets and model configurations2 . However, our proposed compression method can be applied to any publicly available decoder-only model checkpoint e.g., Zhang et al. (2022Zhang et al. ( , 2023)). We follow the PaLM architecture (Chowdhery et al., 2022) owing to improved throughput efficiency. Specifically, the attention and feed-forward network (FFN) modules are parallel instead of sequential (Radford et al., 2019). SwiGLU activation (Shazeer, 2020) is used in the FFN module. Multihead attention uses the equivalent Flash-Attention (Dao et al., 2022, FA) implementation. The first layer of the FFN module and the layers generating attention query, key, and value are fused. Similarly, the second layer of the FFN module and the feed-forward layer after the attention operation are fused. The LayerNorm (Ba et al., 2016) is before the first fused feed-forward layer. The query and the key vectors are passed through additional layer normalization layers for increased training stability following Dehghani et al. (2023). This block structure is repeated with skip connections to form our decoder-only Transformer architecture.\\nUsing the language modeling objective, we train two core baselines with 300 million (M) and 1.1 billion (B) parameters. We also consider additional baseline models containing half the number of layers (180M and 610M parameters). We use Lion optimizer (Chen et al., 2023) in all our pretraining experiments with β parameters set to (0.9, 0.95). Weight decay is 1 × 10 -4 but is omitted for bias and normalization parameters. 3 We schedule learning rate warm-up from the start to 2000 steps (4B tokens) and then use a cosine decay schedule until training stops. The final learning rate is 10% of the peak value. Table 1 summarizes all baseline model configurations and their respective pretraining token budgets. For pretraining, we use the C4 dataset (Raffel et al., 2020) of 160B web-crawled tokens. Pretrained models of all sizes are trained for 1 epoch on the C4 dataset.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Compression and Distillation Baselines\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To simulate the typical practice of a compressed model continuing to train with a pretrained optimizer state, we seed the model with training for 1B additional tokens to \\\"warm-start\\\" the optimizer. This seed phase includes a linear warm-up of the learning rate for 400 steps (800M tokens). After this, we begin layer pruning according to the truncation configuration, controlling when and where we prune layers from the model. A layer that needs to be removed in the future is frozen from the start of the optimization process.\\nWe use the Lion optimizer with the same batch size, weight decay, and β as the above pretraining configuration. However, the peak learning rate is fixed at 10% of the peak learning rate during the\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Category Task Metric\",\n \"publication_ref\": [\n \"b3\",\n \"b60\",\n \"b43\",\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"Common Sense Reasoning PIQA (Bisk et al., 2019) ACC (LN) Hellaswag (Zellers et al., 2019) ACC (LN) Winogrande (Sakaguchi et al., 2019) ACC BoolQ (Clark et al., 2019) ACC (LN)\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Science Question Answering\",\n \"publication_ref\": [\n \"b35\",\n \"b58\",\n \"b11\"\n ],\n \"table_ref\": [],\n \"text\": \"OpenBookQA (Mihaylov et al., 2018) ACC (LN) SciQ (Welbl et al., 2017) ACC Arc-Easy (Clark et al., 2018) ACC Arc-Challenge (Clark et al., pretraining phase. For the 300M model, this is 3 × 10 -5 ; for the 1.1B model, this is 2 × 10 -5 .\\nFor distillation baselines, we use the same truncated initialization strategy for student models as our compression method. Each loss component of the total distillation loss is normalized such that the total magnitude for distillation loss matches the cross-entropy loss. We note that some forms of knowledge distillation cannot exploit recent modeling optimizations, such as flash-attention. Aligning QK T outputs for distillation at every layer requires the additional computation of O(n 2 ) matrix products, significantly slowing down training.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Downstream Evaluation\",\n \"publication_ref\": [\n \"b21\",\n \"b22\"\n ],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"We evaluate our model on 12 tasks from 5 categories: common sense reasoning, science question answering, causal reasoning, natural language inference, and paraphrase identification. All tasks are evaluated in a zero-shot setting by providing the language model with a prompt from Eleuther-AI evaluation harness (Gao et al., 2021) and a possible completion. We score the model output for each completion. The completion with the highest likelihood is the prediction to compute task accuracy. The completion likelihood can be normalized by either the character count in the completion (Gao et al., 2021, length normalized accuracy) or by the probability of completion conditioned on the domain premise (Holtzman et al., 2021;Brown et al., 2020, domain Table 3: Results for our model and distillation baselines from Sec. 3.2 for language model perplexity and all tasks outlined in Table 2. All models are evaluated zero-shot with no finetuning. For extrinsic tasks, higher is better, while lower is better for perplexity (ppl.). \\\"Pre-compression\\\" is the larger teacher model evaluated on the same task suite.\\nOur method yields superior perplexity compared to distillation and is competitive on most tasks.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Compression Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Compression for generality and zero-shot promptability\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_5\"\n ],\n \"text\": \"This section compares our method against the distillation baselines mentioned in Section 3.2. We show perplexity results on C4 validation set and zero-shot performance on 12 end-tasks. All baselines and our method are applied to two core model baselines mentioned in Section 4.1: a 300M parameter model and a 1.1B parameter model, both trained until 160B tokens. In addition, we provide the end-task performance of the checkpoints themselves for comparison against distillation baselines and our method. All compression methods in this section remove 12/24 layers in the decoder-only base models. All methods are trained on additional 20B tokens from the C4 corpus. As shown in Table 3, our method matches the best distillation baseline on perplexity and remains competitive on most end-tasks. Table 4 summarizes training statistics like computation required by each method in terms of FLOPs or the wall-clock time consumed. This result shows that our method is at par with the distillation baselines in our implementation with significantly less compute and wall-clock time. Section 7 discusses our hypothesis on why we expect our distillation-free method to work well despite being trained on a less-informative training signal without a teacher.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Where to Prune a Model?\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"Our baseline models contain 24 decoder layers.\\nTo determine which layers we should prune for the best compression performance, we define five layer pruning configurations shown in Figure 1, each removing 12 out of the 24 decoder layers of the 300M and 1.1B models. In all these pruning configurations, we always keep the first and the last layers because they interact with the embedding table. We made this design choice based on early experiments. Table 5 summarizes the results of this ablation. We report the perplexity score on the C4 validation set and the average task accuracy across 12 tasks in our evaluation suite (Table 2).\\nFor the base 300M model, pruning configurations of max-gap and both perform the best out of the five possible configurations. For the 1.1B model, pruning layers from the input configuration yielded the best results for both reported metrics. The output pruning configuration resulted in the worst performance across model sizes, suggesting that pruning layers towards the output side of the model should be avoided. Given these results, we use the pruning configuration of max-gap for all our 300M model experiments and the configuration of input for all our 1.1B model experiments.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"When to Prune a Model?\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"To answer the question of when we should prune layers from our model while continuing to train it with language modeling loss, we can decide in one of two ways: either remove selected layers at once or remove them one by one, each after a fixed number of training tokens. We run this experiment in four configurations to see if increasing the Table 5: Influence on average task performance and C4 validation perplexity of different truncated initialization strategies from Figure 1 for models of size 300M and 1.1B. The average task performance score is across 12 tasks listed in Table 2, and higher numbers are better. For perplexity scores, lower is better.\\ngap between each layer pruning increases training stability or model performance. The four configurations are: dropping all layers at once (0M token gap between pruning each layer) or pruning them after 100M, 500M, and 1B training tokens each. We run this experiment for the 300M and 1.1B model sizes.\\nWe prune 12/24 layers from our decoder-only models for this ablation, each at a training token gap mentioned in one of the four configurations above. The result of this experiment is summarized in Fig- ure 2. Pruning layers one by one with an increasing token budget between each layer pruning does not benefit the average task accuracy or C4 validation perplexity. In fact, there is a marginal preference to prune layers as early into the training as possible. Hence, we decided to prune all 12/24 layers simultaneously for other experiments.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Upweighting Distillation Loss\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_5\",\n \"tab_2\"\n ],\n \"text\": \"Our vanilla distillation baseline described in Section 3.2 performs at par with our proposed approach of model pruning, followed by continued pretraining. Any distillation setup adds a teacher and a distillation loss component at the minimum to the compression pipeline. Because of this, the vanilla distillation baseline is at least 1.6x slower than our proposed method in terms of theoretical FLOPs and wall-clock time (Table 4). To justify the additional compute requirements of a distillation setup, we want to verify further the importance of the teacher and the distillation loss component. We maintain the setup from our vanilla distillation baseline for this experiment and append a coefficient to the KL-divergence loss term. The new loss function for distillation becomes L Σ = L LM + λ * L distill , where L distill is same as Eq. 3. We ablate the value of λ on our 1.1B model and report results on the C4 validation perplexity and the average task accuracy in Figure 3. It is evident from this experiment that increasing the value of λ results in a decrease in the end-task per- 2) and perplexity on the C4 validation set for model sizes 300M and 1.1B comparing schedules for when to prune layers during continued pretraining. We find a marginal performance degradation as we remove layers one by one further apart during continued pretraining.\\nformance and an increase in C4 perplexity. Thus, this experiment highlights the importance of the language modeling component in our vanilla distillation baseline and strengthens the argument that appending the continued pretraining setup with a distillation component to get similar performance is redundant.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Removing Language Modeling\",\n \"publication_ref\": [\n \"b57\",\n \"b31\"\n ],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"Our other two distillation baselines described in Section 3.2 constrain the teacher-teacher setup in more ways than just using a KL-divergence term for distillation. The miniLM-KD (Wang et al., 2020b) baseline adds constraints on the attention map and intermediate representations of the final layer between the teacher and the student model on top of the loss terms in the vanilla distillation setup. The homotopic-KD (Liang et al., 2023) baseline takes a step further and adds constraints on each layer between the teacher and the student on attention maps and intermediate representations. These setups are well designed to align student outputs to the teacher perfectly; thus, we ablate on whether they even require the language modeling component for the student.\\nIn this experiment, we maintain the same setup of our miniLM and homotopic distillation base- 2) and C4 validation perplexity. lines without the language modeling loss term. We show the results in Figure 4 for our 1.1B model. Both distillation baselines lose some end-task performance by removing the language modeling loss term. However, the language modeling task performance on the C4 validation set suffers significantly without the language modeling loss term on the student.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\",\n \"fig_4\"\n ],\n \"heading\": \"Train Large then Compress, or Not\",\n \"publication_ref\": [\n \"b30\",\n \"b47\"\n ],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"Existing literature (Li et al., 2020) suggests that training a large model and then compressing it usually works better than training a smaller model Figure 4: Experiment demonstrating the effectiveness of language modeling loss component in compound loss functions of miniLM-KD (Sec. 3.2) and homotopic-KD distillation (Sec. 3.2) baselines. We report the average task performance across 12 tasks (Table 2) and perplexity on the C4 validation set. from scratch. In this section, we study this observation in the context of zero-shot evaluation of the compressed model. As mentioned in Section 4.1, in addition to our 300M/1.1B baselines, we train two half-sized models (180M/610M) from scratch on the same pretraining token budget as their larger counterparts.\\nFigure 5 compares the performance of our source LLM models (\\\"full @ max-budget\\\"), half-sized pretrained models to the same token budget (\\\"half @ max-budget\\\"), compressed models from the source LLM with a token budget of 20B (\\\"compressed @ 20B\\\") and half-sized pretrained models with the token budget of our compression method (\\\"half @ 20B\\\"). Our compressed models outperform the half-sized models pretrained from scratch when the token budget is fixed at the level of the compression budget. As we further pretrain our half-sized models, they overtake the comparable compressed model in both zero-shot task evaluation and language modeling perplexity.\\nFigure 6 shows how a billion-sized compressed model starts to flatten and is overtaken by a pretrained half-sized model. This suggests that while continued pretraining is a good way to recover some of the model's zero-shot abilities, truncated initialization due to layer pruning is not very effective at preserving these abilities in the first place.\\nThe ideal truncated initialization method should satisfy two requirements: (1) loss after truncation is the same (or as close as possible) to the loss before truncation, (2) as we continue training the truncated model, its learning curve continues improving at the same rate as if it was trained from scratch. Using the terminogloy from Shen et al. (2022), it should be \\\"loss-preserving\\\" and \\\"trainingdynamics-preserving\\\". Although achieving both requirements might not be possible, they define the upper bound. We continue to explore the space of pruning strategies for future work.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Discussion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The results in sections 5 and 6 showed that existing methods for task-specialized compression are not necessarily the best in the zero-shot evaluation setting. We found that simple continued pretraining outperforms student/teacher distillation and that training a smaller model from scratch is better than training a larger model and then compressing. This section summarizes a few hypotheses explaining these findings and discusses why zero-shot evaluation favors different compression methods.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Non-ideal truncated initialization\",\n \"publication_ref\": [\n \"b47\"\n ],\n \"table_ref\": [],\n \"text\": \"To preserve a model's zero-shot promptability, we need a truncated initialization method that's both losspreserving and training-dynamics preserving Shen et al. (2022). However, the results in section 6 show that the truncated initialization methods cause the compressed model to lose too much of its zero-shot abilities.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Size of training data\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We hypothesize that the reason distillation does not lead to improved zeroshot performance is related to the training data size. In task-specific finetuning setup, where the task data is relatively limited, the larger model can learn more generalizable representations from the data and can provide a more informative training signal to the smaller student model, compared to what the student model can learn itself from data. In contrast, in the zero-shot evaluation setup, where we have unlimited pretraining data, distilling from a larger model does not provide much advantage when we continue to learn from billions of tokens.\\nAligning the pretraining objective Distillation uses a collection of training objectives that constrain the smaller model in various ways. This is a reasonable setup when the goal is to imitate the larger model as much as possible. However, in an infinite data regime with the goal of training a zero-shot model, a single language modeling objective is better aligned with the goal. It is more capable of eliciting zero-shot behavior compared to distillation.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We examine compression techniques to preserve the zero-shot promptability of large language models. We proposed a task-agnostic compression methodology, with no teacher, using pruning for truncated initialization of the compressed model, and continued pretraining as an alternative to endtask finetuning. On a diverse zero-shot evaluation suite and perplexity, our method is comparable in performance to a range of distillation baselines while offering 1.5× compute efficient training. We also compare compression to an \\\"oracle\\\" setup of an equally sized model pretrained for the same token budget as the larger teacher model. We highlight that a broad range of compression methods underperform compared to an uncompressed model with more training. This surprising result that distillation is not the \\\"silver bullet\\\" here, as it has been in prior works, leads us to call for future research into closing this gap. We release our code and evaluation setup to encourage future research and discussion on improved compression strategies.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"While our work is in the spirit of reducing model size and improving efficiency -we require significant computational resources for our experiments demanding both high energy usage and processing power. Experiments such as the teacher model pretraining and \\\"oracle\\\" student model upper-bound demand upto 5 days of training time using 32xA100 GPUs with a high bandwidth interconnect. Therefore, reproducing our experiments are only reasonably tractable with commensurate GPU resources which may be infeasible for some researchers.\\nAdditionally, we demonstrate our findings compared to a 'vanilla' distillation approach and recently published alternatives in our decoder-only setup. We take this approach to report how the most typical distillation strategy can be ported to a contemporary LLM. Our findings do not indicate that distillation is potentially fruitful for GPT-style models, however, our work is limited in that there may exist some atypical distillation strategy with even better performance. We encourage future work and discussion of how these methods can be improved in this regard.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ethics Statement\",\n \"publication_ref\": [\n \"b18\",\n \"b23\",\n \"b48\"\n ],\n \"table_ref\": [],\n \"text\": \"We report all pretraining experiments with the widely used C4 corpus. This corpus has been found to contain harmful artifacts and biases (Dodge et al., 2021) which our models may inherit, however, the study of this phenomena is outside of the scope of our work but may inform future study. Model compression has been linked to increased bias and toxicity in a model (Hooker et al., 2020) but it is currently unclear how such effects extend to our setting; particularly as we expose the student to the same corpus as the teacher. Further study is needed in this area to examine how compression influences biases in increasingly large language models (Solaiman et al., 2023).\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-11-19"},"doi":{"kind":"string","value":"10.18653/v1/2021.emnlp-main.98"},"references":{"kind":"list like","value":[{"authors":"Jimmy Ba; Jamie Ryan Kiros; Geoffrey E Hinton","journal":"","ref_id":"b0","title":"Layer normalization","year":"2016"},{"authors":"Yoshua Bengio; Réjean Ducharme; Pascal Vincent; Christian Janvin","journal":"J. Mach. Learn. Res","ref_id":"b1","title":"A neural probabilistic language model","year":"2003"},{"authors":"Stella Biderman; Hailey Schoelkopf; Quentin Anthony; Herbie Bradley; O' Kyle; Eric Brien; Mohammad Hallahan; Shivanshu Aflah Khan; Purohit; Edward Usvsn Sai Prashanth; Aviya Raff; Lintang Skowron; Oskar Sutawika; Van Der Wal","journal":"","ref_id":"b2","title":"Pythia: A suite for analyzing large language models across training and scaling","year":"2023"},{"authors":"Yonatan Bisk; Rowan Zellers; Le Ronan; Jianfeng Bras; Yejin Gao; Choi","journal":"","ref_id":"b3","title":"Piqa: Reasoning about physical commonsense in natural language","year":"2019"},{"authors":"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel Ziegler; Jeffrey Wu; Clemens Winter; Chris Hesse; Mark Chen; Eric Sigler; Mateusz Litwin; Scott Gray; Benjamin Chess; Jack Clark; Christopher Berner; Sam Mccandlish; Alec Radford; Ilya Sutskever; Dario Amodei","journal":"","ref_id":"b4","title":"Language models are few-shot learners","year":"2020"},{"authors":"Curran Associates; Inc ","journal":"","ref_id":"b5","title":"","year":""},{"authors":"Tianlong Chen; Jonathan Frankle; Shiyu Chang; Sijia Liu; Yang Zhang; Zhangyang Wang; Michael Carbin","journal":"","ref_id":"b6","title":"The lottery ticket hypothesis for pretrained bert networks","year":"2020"},{"authors":"Xiangning Chen; Chen Liang; Da Huang; Esteban Real; Kaiyuan Wang; Yao Liu; Hieu Pham; Xuanyi Dong; Thang Luong; Cho-Jui Hsieh; Yifeng Lu; Quoc V Le","journal":"","ref_id":"b7","title":"Symbolic discovery of optimization algorithms","year":"2023"},{"authors":"Xiaohan Chen; Yu Cheng; Shuohang Wang; Zhe Gan; Zhangyang Wang; Jingjing Liu","journal":"","ref_id":"b8","title":"Earlybert: Efficient bert training via early-bird lottery tickets","year":"2020"},{"authors":"Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann; Parker Schuh; Kensen Shi; Sasha Tsvyashchenko; Joshua Maynez; Abhishek Rao; Parker Barnes; Yi Tay; Noam M Shazeer; Emily Vinodkumar Prabhakaran; Nan Reif; Benton C Du; Reiner Hutchinson; James Pope; Jacob Bradbury; Michael Austin; Guy Isard; Pengcheng Gur-Ari; Toju Yin; Anselm Duke; Sanjay Levskaya; Sunipa Ghemawat; Henryk Dev; Xavier Michalewski; Vedant García; Kevin Misra; Liam Robinson; Denny Fedus; Daphne Zhou; David Ippolito; Hyeontaek Luan; Barret Lim; Alexander Zoph; Ryan Spiridonov; David Sepassi; Shivani Dohan; Mark Agrawal; Andrew M Omernick; Thanumalayan Dai; Marie Sankaranarayana Pillai; Aitor Pellat; Erica Lewkowycz; Rewon Moreira; Oleksandr Child; Katherine Polozov; Zongwei Lee; Xuezhi Zhou; Brennan Wang; Mark Saeta; Orhan Díaz; Michele Firat; Jason Catasta; Kathleen S Wei; Douglas Meier-Hellstern; Jeff Eck; Slav Dean; Noah Petrov; Fiedel","journal":"","ref_id":"b9","title":"Palm: Scaling language modeling with pathways","year":"2022"},{"authors":"Christopher Clark; Kenton Lee; Ming-Wei Chang; Tom Kwiatkowski; Michael Collins; Kristina Toutanova","journal":"","ref_id":"b10","title":"Boolq: Exploring the surprising difficulty of natural yes/no questions","year":"2019"},{"authors":"Peter Clark; Isaac Cowhey; Oren Etzioni; Tushar Khot; Ashish Sabharwal; Carissa Schoenick; Oyvind Tafjord","journal":"","ref_id":"b11","title":"Think you have solved question answering? try arc, the ai2 reasoning challenge","year":"2018"},{"authors":"Tri Dao; Dan Fu; Stefano Ermon; Atri Rudra; Christopher Ré","journal":"","ref_id":"b12","title":"Flashattention: Fast and memory-efficient exact attention with io-awareness","year":"2022"},{"authors":"Marie-Catherine De Marneffe; Mandy Simons; Judith Tonhauser","journal":"","ref_id":"b13","title":"The commitmentbank: Investigating projection in naturally occurring discourse","year":"2019"},{"authors":"Mostafa Dehghani; Josip Djolonga; Basil Mustafa; Piotr Padlewski; Jonathan Heek; Justin Gilmer; Andreas Steiner; Mathilde Caron; Robert Geirhos; Ibrahim M Alabdulmohsin; Rodolphe Jenatton; Lucas Beyer; Michael Tschannen; Anurag Arnab; Xiao Wang; Carlos Riquelme; Matthias Minderer; Joan Puigcerver; Utku Evci; Manoj Kumar; Sjoerd Van Steenkiste; Gamaleldin F Elsayed; Aravindh Mahendran; Fisher Yu; Avital Oliver; Fantine Huot; Jasmijn Bastings; Mark Collier; A ; Alexey; Yi Tay; Thomas Mensink; Alexander Kolesnikov; Filip Paveti'c; Dustin Tran; Thomas Kipf; Mario Luvci'c; Xiaohua Zhai; Daniel Keysers; Jeremiah Harmsen; Neil Houlsby","journal":"","ref_id":"b14","title":"Gritsenko, Vighnesh Birodkar, Cristina Nader Vasconcelos","year":"2023"},{"authors":"Tim Dettmers; Mike Lewis; Younes Belkada; Luke Zettlemoyer","journal":"","ref_id":"b15","title":"Gpt3.int8(): 8-bit matrix multiplication for transformers at scale","year":"2022"},{"authors":"Curran Associates; Inc ","journal":"","ref_id":"b16","title":"","year":""},{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"","ref_id":"b17","title":"Bert: Pre-training of deep bidirectional transformers for language understanding","year":"2019"},{"authors":"Jesse Dodge; Maarten Sap; Ana Marasović; William Agnew; Gabriel Ilharco; Dirk Groeneveld; Margaret Mitchell; Matt Gardner","journal":"Association for Computational Linguistics","ref_id":"b18","title":"Documenting large webtext corpora: A case study on the colossal clean crawled corpus","year":"2021"},{"authors":"William B Dolan; Chris Brockett","journal":"","ref_id":"b19","title":"Automatically constructing a corpus of sentential paraphrases","year":"2005"},{"authors":"Angela Fan; Edouard Grave; Armand Joulin","journal":"","ref_id":"b20","title":"Reducing transformer depth on demand with structured dropout","year":"2019"},{"authors":"Leo Gao; Jonathan Tow; Stella Biderman; Sid Black; Anthony Dipofi; Charles Foster; Laurence Golding; Jeffrey Hsu; Kyle Mcdonell; Niklas Muennighoff; Jason Phang; Laria Reynolds; Eric Tang; Anish Thite; Ben Wang; Kevin Wang; Andy Zou ; Suchin Gururangan; Ana Marasović; Swabha Swayamdipta; Kyle Lo; Iz Beltagy; Doug Downey; Noah A Smith","journal":"","ref_id":"b21","title":"A framework for few-shot language model evaluation","year":"2020"},{"authors":"Ari Holtzman; Peter West; Vered Schwartz; Yejin Choi; Luke Zettlemoyer","journal":"","ref_id":"b22","title":"Surface form competition: Why the highest probability answer isn't always right","year":"2021"},{"authors":"Sara Hooker; Nyalleng Moorosi; Gregory Clark; Samy Bengio; Emily Denton","journal":"","ref_id":"b23","title":"Characterising bias in compressed models","year":"2020"},{"authors":"Lu Hou; Zhiqi Huang; Lifeng Shang; Xin Jiang; Qun Liu","journal":"","ref_id":"b24","title":"Dynabert: Dynamic bert with adaptive width and depth","year":"2020"},{"authors":"Xiaoqi Jiao; Yichun Yin; Lifeng Shang; Xin Jiang; Xiao Chen; Linlin Li; Fang Wang; Qun Liu","journal":"","ref_id":"b25","title":"Tinybert: Distilling bert for natural language understanding","year":"2019"},{"authors":"Sheng-Chun Kao; Amir Yazdanbakhsh; Suvinay Subramanian; Shivani Agrawal; Utku Evci; Tushar Krishna","journal":"","ref_id":"b26","title":"Training recipe for n: M structured sparsity with decaying pruning mask","year":"2022"},{"authors":"Solomon Kullback; R A Leibler","journal":"Annals of Mathematical Statistics","ref_id":"b27","title":"On information and sufficiency","year":"1951"},{"authors":"Jiaoda Li; Ryan Cotterell; Mrinmaya Sachan","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b28","title":"a. Differentiable subset pruning of transformer heads","year":"2021"},{"authors":"Jiaoda Li; Ryan Cotterell; Mrinmaya Sachan","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b29","title":"Differentiable subset pruning of transformer heads","year":"2021"},{"authors":"Zhuohan Li; Eric Wallace; Sheng Shen; Kevin Lin; Kurt Keutzer; Dan Klein; Joseph Gonzalez","journal":"","ref_id":"b30","title":"Train large, then compress: Rethinking model size for efficient training and inference of transformers","year":"2020"},{"authors":"Chen Liang; Haoming Jiang; Zheng Li; Xianfeng Tang; Bin Yin; Tuo Zhao","journal":"","ref_id":"b31","title":"Homodistil: Homotopic task-agnostic distillation of pre-trained transformers","year":"2023"},{"authors":"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b32","title":"Roberta: A robustly optimized bert pretraining approach","year":"2019"},{"authors":"J Scott Mccarley; Rishav Chakravarti; Avirup Sil","journal":"arXiv: Computation and Language","ref_id":"b33","title":"Structured pruning of a bert-based question answering model","year":"2019"},{"authors":"Paul Michel; Omer Levy; Graham Neubig","journal":"","ref_id":"b34","title":"Are sixteen heads really better than one? In Neural Information Processing Systems","year":"2019"},{"authors":"Todor Mihaylov; Peter Clark; Tushar Khot; Ashish Sabharwal","journal":"","ref_id":"b35","title":"Can a suit of armor conduct electricity? a new dataset for open book question answering","year":"2018"},{"authors":"Subhabrata Mukherjee; Ahmed Hassan Awadallah; Jianfeng Gao","journal":"","ref_id":"b36","title":"Xtremedistiltransformers: Task transfer for task-agnostic distillation","year":"2021"},{"authors":"Matthew E Peters; Mark Neumann; Mohit Iyyer; Matt Gardner; Christopher Clark; Kenton Lee; Luke Zettlemoyer","journal":"","ref_id":"b37","title":"Deep contextualized word representations","year":"2018"},{"authors":"Alec Radford; Karthik Narasimhan; Tim Salimans; Ilya Sutskever","journal":"","ref_id":"b38","title":"Improving language understanding by generative pre-training","year":"2018"},{"authors":"Alec Radford; Jeff Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever","journal":"","ref_id":"b39","title":"Language models are unsupervised multitask learners","year":"2019"},{"authors":"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu","journal":"Journal of Machine Learning Research","ref_id":"b40","title":"Exploring the limits of transfer learning with a unified text-to-text transformer","year":"2020"},{"authors":"Melissa Roemmele; Cosmin ; Adrian Bejan; Andrew S Gordon","journal":"","ref_id":"b41","title":"Choice of plausible alternatives: An evaluation of commonsense causal reasoning","year":"2011"},{"authors":"Hassan Sajjad; Fahim Dalvi; Nadir Durrani; Preslav Nakov","journal":"","ref_id":"b42","title":"Poor man's bert: Smaller and faster transformer models","year":"2020"},{"authors":"Keisuke Sakaguchi; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi","journal":"","ref_id":"b43","title":"Winogrande: An adversarial winograd schema challenge at scale","year":"2019"},{"authors":"Victor Sanh; Lysandre Debut; Julien Chaumond; Thomas Wolf","journal":"","ref_id":"b44","title":"Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter","year":"2019"},{"authors":"Victor Sanh; Thomas Wolf; Alexander M Rush","journal":"","ref_id":"b45","title":"Movement pruning: Adaptive sparsity by finetuning","year":"2020"},{"authors":"Noam M Shazeer","journal":"","ref_id":"b46","title":"Glu variants improve transformer","year":"2020"},{"authors":"Sheng Shen; Pete Walsh; Kurt Keutzer; Jesse Dodge; Matthew E Peters; Iz Beltagy","journal":"","ref_id":"b47","title":"Staged training for transformer language models","year":"2022"},{"authors":"Irene Solaiman; Zeerak Talat; William Agnew; Lama Ahmad; Dylan Baker; Su Lin Blodgett; Hal Daumé Iii Au2; Jesse Dodge; Ellie Evans; Sara Hooker; Yacine Jernite; Alexandra Sasha Luccioni; Alberto Lusoli; Margaret Mitchell; Jessica Newman; Marie-Therese Png; Andrew Strait; Apostol Vassilev","journal":"","ref_id":"b48","title":"Evaluating the social impact of generative ai systems in systems and society","year":"2023"},{"authors":"S Sun; Yu Cheng; Zhe Gan; Jingjing Liu","journal":"","ref_id":"b49","title":"Patient knowledge distillation for bert model compression","year":"2019"},{"authors":"Zhiqing Sun; Hongkun Yu; Xiaodan Song; Renjie Liu; Yiming Yang; Denny Zhou","journal":"","ref_id":"b50","title":"Mobilebert: a compact task-agnostic bert for resource-limited devices","year":"2020"},{"authors":"Raphael Tang; Yao Lu; Linqing Liu; Lili Mou; Olga Vechtomova; Jimmy J Lin","journal":"","ref_id":"b51","title":"Distilling task-specific knowledge from bert into simple neural networks","year":"2019"},{"authors":"Iulia Turc; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"arXiv: Computation and Language","ref_id":"b52","title":"Well-read students learn better: On the importance of pre-training compact models","year":"2019"},{"authors":"Ashish Vaswani; Noam M Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin","journal":"","ref_id":"b53","title":"Attention is all you need","year":"2017"},{"authors":"Elena Voita; David Talbot; F Moiseev; Rico Sennrich; Ivan Titov","journal":"","ref_id":"b54","title":"Analyzing multi-head selfattention: Specialized heads do the heavy lifting, the rest can be pruned","year":"2019"},{"authors":"Alex Wang; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel R Bowman","journal":"","ref_id":"b55","title":"Glue: A multi-task benchmark and analysis platform for natural language understanding","year":"2018"},{"authors":"Wenhui Wang; Hangbo Bao; Shaohan Huang; Li Dong; Furu Wei","journal":"","ref_id":"b56","title":"Minilmv2: Multi-head selfattention relation distillation for compressing pretrained transformers","year":"2020"},{"authors":"Wenhui Wang; Furu Wei; Li Dong; Hangbo Bao; Nan Yang; Ming Zhou","journal":"","ref_id":"b57","title":"Minilm: Deep selfattention distillation for task-agnostic compression of pre-trained transformers","year":"2020"},{"authors":"Johannes Welbl; Nelson F Liu; Matt Gardner","journal":"","ref_id":"b58","title":"Crowdsourcing multiple choice science questions","year":"2017"},{"authors":"M Xia; Zexuan Zhong; Danqi Chen","journal":"","ref_id":"b59","title":"Structured pruning learns compact and accurate models","year":"2022"},{"authors":"Rowan Zellers; Ari Holtzman; Yonatan Bisk; Ali Farhadi; Yejin Choi","journal":"","ref_id":"b60","title":"Hellaswag: Can a machine really finish your sentence","year":"2019"},{"authors":"Renrui Zhang; Jiaming Han; Aojun Zhou; Xiangfei Hu; Shilin Yan; Pan Lu; Hongsheng Li; Peng Gao; Yu Jiao; Qiao ","journal":"","ref_id":"b61","title":"Llama-adapter: Efficient finetuning of language models with zero-init attention","year":"2023"},{"authors":"Susan Zhang; Stephen Roller; Naman Goyal; Mikel Artetxe; Moya Chen; Shuohui Chen; Christopher Dewan; Mona Diab; Xian Li; Xi Victoria Lin; Todor Mihaylov; Myle Ott; Sam Shleifer; Kurt Shuster; Daniel Simig; Punit Singh Koura; Anjali Sridhar; Tianlu Wang; Luke Zettlemoyer","journal":"","ref_id":"b62","title":"Opt: Open pre-trained transformer language models","year":"2022"}],"string":"[\n {\n \"authors\": \"Jimmy Ba; Jamie Ryan Kiros; Geoffrey E Hinton\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Layer normalization\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Yoshua Bengio; Réjean Ducharme; Pascal Vincent; Christian Janvin\",\n \"journal\": \"J. Mach. Learn. Res\",\n \"ref_id\": \"b1\",\n \"title\": \"A neural probabilistic language model\",\n \"year\": \"2003\"\n },\n {\n \"authors\": \"Stella Biderman; Hailey Schoelkopf; Quentin Anthony; Herbie Bradley; O' Kyle; Eric Brien; Mohammad Hallahan; Shivanshu Aflah Khan; Purohit; Edward Usvsn Sai Prashanth; Aviya Raff; Lintang Skowron; Oskar Sutawika; Van Der Wal\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Pythia: A suite for analyzing large language models across training and scaling\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Yonatan Bisk; Rowan Zellers; Le Ronan; Jianfeng Bras; Yejin Gao; Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Piqa: Reasoning about physical commonsense in natural language\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel Ziegler; Jeffrey Wu; Clemens Winter; Chris Hesse; Mark Chen; Eric Sigler; Mateusz Litwin; Scott Gray; Benjamin Chess; Jack Clark; Christopher Berner; Sam Mccandlish; Alec Radford; Ilya Sutskever; Dario Amodei\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Curran Associates; Inc \",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Tianlong Chen; Jonathan Frankle; Shiyu Chang; Sijia Liu; Yang Zhang; Zhangyang Wang; Michael Carbin\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"The lottery ticket hypothesis for pretrained bert networks\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Xiangning Chen; Chen Liang; Da Huang; Esteban Real; Kaiyuan Wang; Yao Liu; Hieu Pham; Xuanyi Dong; Thang Luong; Cho-Jui Hsieh; Yifeng Lu; Quoc V Le\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Symbolic discovery of optimization algorithms\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Xiaohan Chen; Yu Cheng; Shuohang Wang; Zhe Gan; Zhangyang Wang; Jingjing Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Earlybert: Efficient bert training via early-bird lottery tickets\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann; Parker Schuh; Kensen Shi; Sasha Tsvyashchenko; Joshua Maynez; Abhishek Rao; Parker Barnes; Yi Tay; Noam M Shazeer; Emily Vinodkumar Prabhakaran; Nan Reif; Benton C Du; Reiner Hutchinson; James Pope; Jacob Bradbury; Michael Austin; Guy Isard; Pengcheng Gur-Ari; Toju Yin; Anselm Duke; Sanjay Levskaya; Sunipa Ghemawat; Henryk Dev; Xavier Michalewski; Vedant García; Kevin Misra; Liam Robinson; Denny Fedus; Daphne Zhou; David Ippolito; Hyeontaek Luan; Barret Lim; Alexander Zoph; Ryan Spiridonov; David Sepassi; Shivani Dohan; Mark Agrawal; Andrew M Omernick; Thanumalayan Dai; Marie Sankaranarayana Pillai; Aitor Pellat; Erica Lewkowycz; Rewon Moreira; Oleksandr Child; Katherine Polozov; Zongwei Lee; Xuezhi Zhou; Brennan Wang; Mark Saeta; Orhan Díaz; Michele Firat; Jason Catasta; Kathleen S Wei; Douglas Meier-Hellstern; Jeff Eck; Slav Dean; Noah Petrov; Fiedel\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Palm: Scaling language modeling with pathways\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Christopher Clark; Kenton Lee; Ming-Wei Chang; Tom Kwiatkowski; Michael Collins; Kristina Toutanova\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Boolq: Exploring the surprising difficulty of natural yes/no questions\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Peter Clark; Isaac Cowhey; Oren Etzioni; Tushar Khot; Ashish Sabharwal; Carissa Schoenick; Oyvind Tafjord\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Think you have solved question answering? try arc, the ai2 reasoning challenge\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Tri Dao; Dan Fu; Stefano Ermon; Atri Rudra; Christopher Ré\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Flashattention: Fast and memory-efficient exact attention with io-awareness\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Marie-Catherine De Marneffe; Mandy Simons; Judith Tonhauser\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"The commitmentbank: Investigating projection in naturally occurring discourse\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Mostafa Dehghani; Josip Djolonga; Basil Mustafa; Piotr Padlewski; Jonathan Heek; Justin Gilmer; Andreas Steiner; Mathilde Caron; Robert Geirhos; Ibrahim M Alabdulmohsin; Rodolphe Jenatton; Lucas Beyer; Michael Tschannen; Anurag Arnab; Xiao Wang; Carlos Riquelme; Matthias Minderer; Joan Puigcerver; Utku Evci; Manoj Kumar; Sjoerd Van Steenkiste; Gamaleldin F Elsayed; Aravindh Mahendran; Fisher Yu; Avital Oliver; Fantine Huot; Jasmijn Bastings; Mark Collier; A ; Alexey; Yi Tay; Thomas Mensink; Alexander Kolesnikov; Filip Paveti'c; Dustin Tran; Thomas Kipf; Mario Luvci'c; Xiaohua Zhai; Daniel Keysers; Jeremiah Harmsen; Neil Houlsby\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Gritsenko, Vighnesh Birodkar, Cristina Nader Vasconcelos\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Tim Dettmers; Mike Lewis; Younes Belkada; Luke Zettlemoyer\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Gpt3.int8(): 8-bit matrix multiplication for transformers at scale\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Curran Associates; Inc \",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Bert: Pre-training of deep bidirectional transformers for language understanding\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Jesse Dodge; Maarten Sap; Ana Marasović; William Agnew; Gabriel Ilharco; Dirk Groeneveld; Margaret Mitchell; Matt Gardner\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b18\",\n \"title\": \"Documenting large webtext corpora: A case study on the colossal clean crawled corpus\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"William B Dolan; Chris Brockett\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Automatically constructing a corpus of sentential paraphrases\",\n \"year\": \"2005\"\n },\n {\n \"authors\": \"Angela Fan; Edouard Grave; Armand Joulin\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Reducing transformer depth on demand with structured dropout\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Leo Gao; Jonathan Tow; Stella Biderman; Sid Black; Anthony Dipofi; Charles Foster; Laurence Golding; Jeffrey Hsu; Kyle Mcdonell; Niklas Muennighoff; Jason Phang; Laria Reynolds; Eric Tang; Anish Thite; Ben Wang; Kevin Wang; Andy Zou ; Suchin Gururangan; Ana Marasović; Swabha Swayamdipta; Kyle Lo; Iz Beltagy; Doug Downey; Noah A Smith\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"A framework for few-shot language model evaluation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ari Holtzman; Peter West; Vered Schwartz; Yejin Choi; Luke Zettlemoyer\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Surface form competition: Why the highest probability answer isn't always right\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Sara Hooker; Nyalleng Moorosi; Gregory Clark; Samy Bengio; Emily Denton\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Characterising bias in compressed models\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Lu Hou; Zhiqi Huang; Lifeng Shang; Xin Jiang; Qun Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Dynabert: Dynamic bert with adaptive width and depth\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Xiaoqi Jiao; Yichun Yin; Lifeng Shang; Xin Jiang; Xiao Chen; Linlin Li; Fang Wang; Qun Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Tinybert: Distilling bert for natural language understanding\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Sheng-Chun Kao; Amir Yazdanbakhsh; Suvinay Subramanian; Shivani Agrawal; Utku Evci; Tushar Krishna\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Training recipe for n: M structured sparsity with decaying pruning mask\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Solomon Kullback; R A Leibler\",\n \"journal\": \"Annals of Mathematical Statistics\",\n \"ref_id\": \"b27\",\n \"title\": \"On information and sufficiency\",\n \"year\": \"1951\"\n },\n {\n \"authors\": \"Jiaoda Li; Ryan Cotterell; Mrinmaya Sachan\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b28\",\n \"title\": \"a. Differentiable subset pruning of transformer heads\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jiaoda Li; Ryan Cotterell; Mrinmaya Sachan\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b29\",\n \"title\": \"Differentiable subset pruning of transformer heads\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zhuohan Li; Eric Wallace; Sheng Shen; Kevin Lin; Kurt Keutzer; Dan Klein; Joseph Gonzalez\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Train large, then compress: Rethinking model size for efficient training and inference of transformers\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Chen Liang; Haoming Jiang; Zheng Li; Xianfeng Tang; Bin Yin; Tuo Zhao\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Homodistil: Homotopic task-agnostic distillation of pre-trained transformers\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Roberta: A robustly optimized bert pretraining approach\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"J Scott Mccarley; Rishav Chakravarti; Avirup Sil\",\n \"journal\": \"arXiv: Computation and Language\",\n \"ref_id\": \"b33\",\n \"title\": \"Structured pruning of a bert-based question answering model\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Paul Michel; Omer Levy; Graham Neubig\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Are sixteen heads really better than one? In Neural Information Processing Systems\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Todor Mihaylov; Peter Clark; Tushar Khot; Ashish Sabharwal\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Can a suit of armor conduct electricity? a new dataset for open book question answering\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Subhabrata Mukherjee; Ahmed Hassan Awadallah; Jianfeng Gao\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Xtremedistiltransformers: Task transfer for task-agnostic distillation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Matthew E Peters; Mark Neumann; Mohit Iyyer; Matt Gardner; Christopher Clark; Kenton Lee; Luke Zettlemoyer\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Deep contextualized word representations\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Alec Radford; Karthik Narasimhan; Tim Salimans; Ilya Sutskever\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Improving language understanding by generative pre-training\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Alec Radford; Jeff Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Language models are unsupervised multitask learners\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu\",\n \"journal\": \"Journal of Machine Learning Research\",\n \"ref_id\": \"b40\",\n \"title\": \"Exploring the limits of transfer learning with a unified text-to-text transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Melissa Roemmele; Cosmin ; Adrian Bejan; Andrew S Gordon\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Choice of plausible alternatives: An evaluation of commonsense causal reasoning\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"Hassan Sajjad; Fahim Dalvi; Nadir Durrani; Preslav Nakov\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Poor man's bert: Smaller and faster transformer models\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Keisuke Sakaguchi; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"Winogrande: An adversarial winograd schema challenge at scale\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Victor Sanh; Lysandre Debut; Julien Chaumond; Thomas Wolf\",\n \"journal\": \"\",\n \"ref_id\": \"b44\",\n \"title\": \"Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Victor Sanh; Thomas Wolf; Alexander M Rush\",\n \"journal\": \"\",\n \"ref_id\": \"b45\",\n \"title\": \"Movement pruning: Adaptive sparsity by finetuning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Noam M Shazeer\",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"Glu variants improve transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Sheng Shen; Pete Walsh; Kurt Keutzer; Jesse Dodge; Matthew E Peters; Iz Beltagy\",\n \"journal\": \"\",\n \"ref_id\": \"b47\",\n \"title\": \"Staged training for transformer language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Irene Solaiman; Zeerak Talat; William Agnew; Lama Ahmad; Dylan Baker; Su Lin Blodgett; Hal Daumé Iii Au2; Jesse Dodge; Ellie Evans; Sara Hooker; Yacine Jernite; Alexandra Sasha Luccioni; Alberto Lusoli; Margaret Mitchell; Jessica Newman; Marie-Therese Png; Andrew Strait; Apostol Vassilev\",\n \"journal\": \"\",\n \"ref_id\": \"b48\",\n \"title\": \"Evaluating the social impact of generative ai systems in systems and society\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"S Sun; Yu Cheng; Zhe Gan; Jingjing Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b49\",\n \"title\": \"Patient knowledge distillation for bert model compression\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Zhiqing Sun; Hongkun Yu; Xiaodan Song; Renjie Liu; Yiming Yang; Denny Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b50\",\n \"title\": \"Mobilebert: a compact task-agnostic bert for resource-limited devices\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Raphael Tang; Yao Lu; Linqing Liu; Lili Mou; Olga Vechtomova; Jimmy J Lin\",\n \"journal\": \"\",\n \"ref_id\": \"b51\",\n \"title\": \"Distilling task-specific knowledge from bert into simple neural networks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Iulia Turc; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"arXiv: Computation and Language\",\n \"ref_id\": \"b52\",\n \"title\": \"Well-read students learn better: On the importance of pre-training compact models\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Ashish Vaswani; Noam M Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin\",\n \"journal\": \"\",\n \"ref_id\": \"b53\",\n \"title\": \"Attention is all you need\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Elena Voita; David Talbot; F Moiseev; Rico Sennrich; Ivan Titov\",\n \"journal\": \"\",\n \"ref_id\": \"b54\",\n \"title\": \"Analyzing multi-head selfattention: Specialized heads do the heavy lifting, the rest can be pruned\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Alex Wang; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel R Bowman\",\n \"journal\": \"\",\n \"ref_id\": \"b55\",\n \"title\": \"Glue: A multi-task benchmark and analysis platform for natural language understanding\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Wenhui Wang; Hangbo Bao; Shaohan Huang; Li Dong; Furu Wei\",\n \"journal\": \"\",\n \"ref_id\": \"b56\",\n \"title\": \"Minilmv2: Multi-head selfattention relation distillation for compressing pretrained transformers\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Wenhui Wang; Furu Wei; Li Dong; Hangbo Bao; Nan Yang; Ming Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b57\",\n \"title\": \"Minilm: Deep selfattention distillation for task-agnostic compression of pre-trained transformers\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Johannes Welbl; Nelson F Liu; Matt Gardner\",\n \"journal\": \"\",\n \"ref_id\": \"b58\",\n \"title\": \"Crowdsourcing multiple choice science questions\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"M Xia; Zexuan Zhong; Danqi Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b59\",\n \"title\": \"Structured pruning learns compact and accurate models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Rowan Zellers; Ari Holtzman; Yonatan Bisk; Ali Farhadi; Yejin Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b60\",\n \"title\": \"Hellaswag: Can a machine really finish your sentence\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Renrui Zhang; Jiaming Han; Aojun Zhou; Xiangfei Hu; Shilin Yan; Pan Lu; Hongsheng Li; Peng Gao; Yu Jiao; Qiao \",\n \"journal\": \"\",\n \"ref_id\": \"b61\",\n \"title\": \"Llama-adapter: Efficient finetuning of language models with zero-init attention\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Susan Zhang; Stephen Roller; Naman Goyal; Mikel Artetxe; Moya Chen; Shuohui Chen; Christopher Dewan; Mona Diab; Xian Li; Xi Victoria Lin; Todor Mihaylov; Myle Ott; Sam Shleifer; Kurt Shuster; Daniel Simig; Punit Singh Koura; Anjali Sridhar; Tianlu Wang; Luke Zettlemoyer\",\n \"journal\": \"\",\n \"ref_id\": \"b62\",\n \"title\": \"Opt: Open pre-trained transformer language models\",\n \"year\": \"2022\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,105.69,611.93,184.17,33.71],"formula_id":"formula_0","formula_text":"p(X) = N i=1 p (x i |x 1 , x 2 , ..., x i-1 )(1)"},{"formula_coordinates":[3,310.71,97.68,214.43,44.01],"formula_id":"formula_1","formula_text":"L LM = log p θ (X) = N i=1 p(x i |x i-k , ..., x i-1 ; θ) (2)"},{"formula_coordinates":[3,344.15,389.95,181,10.77],"formula_id":"formula_2","formula_text":"L distill = KL (f s (x i ) , f t (x i ))(3)"},{"formula_coordinates":[3,309.9,662.06,215.11,29.45],"formula_id":"formula_3","formula_text":"LQK = KL softmax QsKs √ ds , softmax QtKt √ dt(4)"},{"formula_coordinates":[4,79.64,423.36,210.23,119.04],"formula_id":"formula_4","formula_text":"L hid (θ s , θ t ) = K k=1 MSE H k t , H k s (5) L emb (θ s , θ t ) = MSE (E t , E s ) (6) L att (θ s , θ t ) = K k=1 MSE A k t , A k s (7) L Σ = L LM + L distill + L hid + L emb + L att (8)"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 105.69,\n 611.93,\n 184.17,\n 33.71\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"p(X) = N i=1 p (x i |x 1 , x 2 , ..., x i-1 )(1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 310.71,\n 97.68,\n 214.43,\n 44.01\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"L LM = log p θ (X) = N i=1 p(x i |x i-k , ..., x i-1 ; θ) (2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 344.15,\n 389.95,\n 181,\n 10.77\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"L distill = KL (f s (x i ) , f t (x i ))(3)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 309.9,\n 662.06,\n 215.11,\n 29.45\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"LQK = KL softmax QsKs √ ds , softmax QtKt √ dt(4)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 79.64,\n 423.36,\n 210.23,\n 119.04\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"L hid (θ s , θ t ) = K k=1 MSE H k t , H k s (5) L emb (θ s , θ t ) = MSE (E t , E s ) (6) L att (θ s , θ t ) = K k=1 MSE A k t , A k s (7) L Σ = L LM + L distill + L hid + L emb + L att (8)\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"How To Train Your (Compressed) Large Language Model"},"abstract":{"kind":"string","value":"With the increase in the size of large language models (LLMs), we need compression methods that can reduce the model size while preserving the generality and zero-shot promptability of the model. This goal is more ambitious than the typical compression setup, which reduces the model's size at the expense of specializing it to a specific end-task. To study this, we develop a task-agnostic compression pipeline with a large-scale evaluation comprising language modeling perplexity and 12 zeroshot end-tasks. Our results show that a simple layer-wise pruning followed by continued language model pretraining matches or outperforms three existing state-of-the-art baselines while being 1.5x more computationally efficient. However, unlike typical task-specialized compression, our best-compressed model significantly underperforms a similar-sized model trained from scratch. We posit the half-sized pretrained model as an upper bound for taskagnostic compression and call for future work to bridge this gap under a reasonable token budget. Our findings highlight the inadequacy of existing compression methods for LLMs and establish a requirement for new methods that preserve a model's generality and zero-shot promptability under compression. We release our code and evaluation setup to facilitate reproducibility and help iterate on method design."},"authors":{"kind":"string","value":"Ananya Harsh Jha; Tom Sherborne; Evan Pete Walsh; Dirk Groeneveld; Emma Strubell; Iz Beltagy"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: Truncated initialization configurations for layer pruning in a decoder-only language model. Highlighted layers (green) are removed. Our method and distillation baselines remove half of the layers according to each configuration. We retain the first and last layer as these layers interact with the embedding table.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Average task accuracy across 12 tasks (Table2) and perplexity on the C4 validation set for model sizes 300M and 1.1B comparing schedules for when to prune layers during continued pretraining. We find a marginal performance degradation as we remove layers one by one further apart during continued pretraining.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: We ablate different coefficients on the KLdivergence distillation loss component of the vanilla distillation baseline (Sec. 3.2) for model size 1.1B. We report the average task score across 12 tasks in our evaluation suite (Table2) and C4 validation perplexity.","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: Comparing compressed models against pretraining an equivalent model from scratch with the training budget of only continued pretraining (20B tokens) or the full pretraining stage (max budget).","figure_data":"","figure_id":"fig_3","figure_label":"5","figure_type":"figure"},{"figure_caption":"Figure 6 :6Figure 6: Evaluation perplexity curves for our compressed model trained for 20B tokens and the \"oracle\" smaller model trained for the maximum token budget.","figure_data":"","figure_id":"fig_4","figure_label":"6","figure_type":"figure"},{"figure_caption":"Configuration for Decoder-only Transformers. Each is trained on C4 as pretraining baseline. The 2M token batch size follows Biderman et al. (2023).","figure_data":"Params Dim Heads Layers Batch SizeLRToken Budget180M102416122M6.0e-4160B300M102416242M3.0e-4160B610M204816122M2.5e-4160B1.1B204816242M2.0e-4160B","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"","figure_data":": Downstream tasks for evaluating our com-pressed models and baselines. Each task reports a dif-ferent metric: 'ACC': accuracy, 'ACC (LN)': length nor-malized accuracy, 'F1': F1 score, 'PMI-DC': domain-conditioned pointwise mutual information.","figure_id":"tab_2","figure_label":"2","figure_type":"table"},{"figure_caption":"conditional point-wise mutual information). In Table2, we list tasks in each evaluation category and respective metrics.","figure_data":"ModelMethodPIQA Hellaswag Winogrande BoolQ OBQA SciQ Arc-e Arc-c COPA RTE CB MRPC Avg (↑) C4 ppl.(↓)Pre-compression 71.045.852.361.132.081.7 52.623.473.053.8 41.181.255.816.2Vanilla-KD66.634.949.762.130.676.4 47.225.869.055.6 39.380.553.122.8300M-160BminiLM-KD66.134.649.862.229.475.2 45.622.769.056.7 39.381.052.623.4Homotopic-KD63.632.450.762.228.474.4 42.827.468.057.4 41.181.052.427.2Ours66.835.149.862.029.276.7 47.927.168.056.0 39.380.353.223.0Pre-compression 74.456.855.262.534.485.1 57.029.177.054.9 50.081.159.813.0Vanilla-KD70.644.450.962.230.081.7 53.327.170.053.4 41.180.855.517.31.1B-160BminiLM-KD70.643.752.462.228.281.0 50.926.170.054.2 41.181.355.117.5Homotopic-KD68.740.951.662.029.879.4 49.323.169.054.9 37.581.253.919.8Ours70.544.351.662.229.480.9 51.425.171.052.7 41.180.955.117.3","figure_id":"tab_3","figure_label":"","figure_type":"table"},{"figure_caption":"Training statistics for each model and baseline.","figure_data":"The 300M models train using 4xA100 80GB GPUs andthe 1.1B models use 8xA100 80GB GPUs.","figure_id":"tab_5","figure_label":"4","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: Truncated initialization configurations for layer pruning in a decoder-only language model. Highlighted layers (green) are removed. Our method and distillation baselines remove half of the layers according to each configuration. We retain the first and last layer as these layers interact with the embedding table.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Average task accuracy across 12 tasks (Table2) and perplexity on the C4 validation set for model sizes 300M and 1.1B comparing schedules for when to prune layers during continued pretraining. We find a marginal performance degradation as we remove layers one by one further apart during continued pretraining.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: We ablate different coefficients on the KLdivergence distillation loss component of the vanilla distillation baseline (Sec. 3.2) for model size 1.1B. We report the average task score across 12 tasks in our evaluation suite (Table2) and C4 validation perplexity.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: Comparing compressed models against pretraining an equivalent model from scratch with the training budget of only continued pretraining (20B tokens) or the full pretraining stage (max budget).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 :6Figure 6: Evaluation perplexity curves for our compressed model trained for 20B tokens and the \\\"oracle\\\" smaller model trained for the maximum token budget.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Configuration for Decoder-only Transformers. Each is trained on C4 as pretraining baseline. The 2M token batch size follows Biderman et al. (2023).\",\n \"figure_data\": \"Params Dim Heads Layers Batch SizeLRToken Budget180M102416122M6.0e-4160B300M102416242M3.0e-4160B610M204816122M2.5e-4160B1.1B204816242M2.0e-4160B\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \": Downstream tasks for evaluating our com-pressed models and baselines. Each task reports a dif-ferent metric: 'ACC': accuracy, 'ACC (LN)': length nor-malized accuracy, 'F1': F1 score, 'PMI-DC': domain-conditioned pointwise mutual information.\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"conditional point-wise mutual information). In Table2, we list tasks in each evaluation category and respective metrics.\",\n \"figure_data\": \"ModelMethodPIQA Hellaswag Winogrande BoolQ OBQA SciQ Arc-e Arc-c COPA RTE CB MRPC Avg (↑) C4 ppl.(↓)Pre-compression 71.045.852.361.132.081.7 52.623.473.053.8 41.181.255.816.2Vanilla-KD66.634.949.762.130.676.4 47.225.869.055.6 39.380.553.122.8300M-160BminiLM-KD66.134.649.862.229.475.2 45.622.769.056.7 39.381.052.623.4Homotopic-KD63.632.450.762.228.474.4 42.827.468.057.4 41.181.052.427.2Ours66.835.149.862.029.276.7 47.927.168.056.0 39.380.353.223.0Pre-compression 74.456.855.262.534.485.1 57.029.177.054.9 50.081.159.813.0Vanilla-KD70.644.450.962.230.081.7 53.327.170.053.4 41.180.855.517.31.1B-160BminiLM-KD70.643.752.462.228.281.0 50.926.170.054.2 41.181.355.117.5Homotopic-KD68.740.951.662.029.879.4 49.323.169.054.9 37.581.253.919.8Ours70.544.351.662.229.480.9 51.425.171.052.7 41.180.955.117.3\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Training statistics for each model and baseline.\",\n \"figure_data\": \"The 300M models train using 4xA100 80GB GPUs andthe 1.1B models use 8xA100 80GB GPUs.\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. (2020) introduces the GPT-3 model, which is a key methodological basis for the citing paper in terms of the development of large language models and the use of in-context learning for task-specific finetuning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Chowdhery 1 https://github.com/anonymous et al., 2022)\", \"Explanation\": \"The cited work by Chowdhery et al. (2022) extends the research on large language models by introducing the PaLM model, which is a further development in the field of language model compression and efficiency.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2023)\", \"Explanation\": \"The cited work by Zhang et al. (2023) is a data source for the citing paper, as it introduces the LLaMA model, which is a large language model that is used for in-context learning and task-specific finetuning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Voita et al., 2019)\", \"Explanation\": \"The cited work by Voita et al. (2019) provides a method for compressing models via pruning, which the citing paper builds upon in their study of compressing large language models.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Chen et al., 2020b,a)\", \"Explanation\": \"The cited works by Chen et al. (2020b,a) are extensions of the study on compressing models via pruning, as the citing paper further explores the same research area.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Sanh et al., 2019)\", \"Explanation\": \"The cited work by Sanh et al. (2019) introduces knowledge distillation techniques that the citing paper adopts in their study of compressing large language models.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Wang et al., 2020b)\", \"Explanation\": \"The cited work by Wang et al. (2020b) extends the study of knowledge distillation techniques, as the citing paper further explores the same research area.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Liang et al., 2023)\", \"Explanation\": \"The cited work by Liang et al. (2023) is an extension of the study on knowledge distillation techniques, as the citing paper further builds upon the research area.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Sanh et al., 2020)\", \"Explanation\": \"The cited work provides a scoring function for identifying sub-networks within larger models, which the citing paper adopts in its research on pruning methods.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Li et al., 2021a)\", \"Explanation\": \"The cited work extends the research on pruning by providing a more specific method for identifying sub-networks within larger models, which the citing paper further builds upon in its study of pruning methods.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Li et al., 2021b)\", \"Explanation\": \"The cited work is a data source for the citing paper, as it provides a method for pruning attentionheads in LMs, which the citing paper utilizes in its research on pruning methods.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Hou et al., 2020)\", \"Explanation\": \"The cited work extends the research on pruning by providing a method for pruning hidden layers in LMs, which the citing paper further builds upon in its study of pruning methods.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(McCarley et al., 2019)\", \"Explanation\": \"The cited work provides a method for pruning hidden layers in LMs, which the citing paper adopts in its research on pruning methods.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chen et al., 2020a)\", \"Explanation\": \"The cited work introduces the lottery ticket hypothesis, which the citing paper adopts in their approach to unstructured pruning of weights in a network.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Voita et al., 2019)\", \"Explanation\": \"The cited work on task-specific pruning is extended in the citing paper by adding task specialization as a second step in the pruning process.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Sun et al., 2019)\", \"Explanation\": \"The cited work by Sun et al. provides a task-specific method for knowledge distillation, which the citing paper builds upon to develop a similar approach for their research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Turc et al., 2019)\", \"Explanation\": \"The cited work by Turc et al. presents a task-specific method for knowledge distillation that contributes to the citing paper in the development of a similar approach.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Mukherjee et al., 2021)\", \"Explanation\": \"The cited work by Mukherjee et al. provides a task-specific method for knowledge distillation that the citing paper utilizes in their research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Tang et al., 2019)\", \"Explanation\": \"The cited work by Tang et al. offers a task-specific method for knowledge distillation that the citing paper references in the development of their research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Xia et al., 2022)\", \"Explanation\": \"The cited work by Xia et al. presents a task-specific method for knowledge distillation that the citing paper builds upon in their research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Sanh et al., 2019)\", \"Explanation\": \"The cited work by Sanh et al. provides a task-agnostic method for knowledge distillation that the citing paper uses in their research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Jiao et al., 2019)\", \"Explanation\": \"The cited work by Jiao et al. offers a task-agnostic method for knowledge distillation that the citing paper references in their research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Sun et al., 2020)\", \"Explanation\": \"The cited work by Sun et al. presents a task-agnostic method for knowledge distillation that the citing paper builds upon in their research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Wang et al., 2020b,a)\", \"Explanation\": \"The cited work by Wang et al. provides task-agnostic methods for knowledge distillation that the citing paper utilizes in their research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Liang et al., 2023)\", \"Explanation\": \"The cited work by Liang et al. offers a task-agnostic method for knowledge distillation that the citing paper references in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Hou et al., 2020)\", \"Explanation\": \"The cited work by Hou et al. provides a dataset that the citing paper uses in their research on pruning with distillation.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(McCarley et al., 2019)\", \"Explanation\": \"The cited work by McCarley et al. offers a dataset that the citing paper utilizes in their research on pruning with distillation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xia et al., 2022)\", \"Explanation\": \"Co-Fi is a method that combines the pruning phase with distillation by applying distillation losses during the pruning process, which the citing paper adopts as a basis for their own research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liang et al., 2023)\", \"Explanation\": \"Homotopic-distillation is another method that combines the pruning phase with distillation by applying distillation losses during the pruning process, which the citing paper also adopts as a basis for their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Sanh et al., 2019)\", \"Explanation\": \"The cited work by Sanh et al. (2019) provides a methodological basis for the compression of decoder-only LLMs in the setting of compressing models for zero-shot tasks without end-task finetuning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Radford et al., 2018)\", \"Explanation\": \"The cited work introduces the concept of causal autoregressive decoders, which the citing paper adopts in their research on language modeling.\"}, {\"Category\": \"Theoretical Foundation\", \"Citation\": \"(Vaswani et al., 2017)\", \"Explanation\": \"The cited work presents the Transformer decoder layers, which the citing paper uses as a foundational model for their research on language modeling.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Bengio et al., 2003)\", \"Explanation\": \"The cited work discusses the concept of a truncated window of context tokens in language modeling, which the citing paper utilizes in their research to limit the scope of their work.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Jiao et al., 2019)\", \"Explanation\": \"The cited work by Jiao et al. (2019) provides a method of aligning intermediate states, embeddings, and attention outputs between the student and teacher models, which the citing paper adopts in their research on relation transfer between Q, K, and V matrix products.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liang et al., 2023)\", \"Explanation\": \"The cited work by Liang et al. (2023) introduces the concept of homotopic distillation, which the citing paper uses to align the intermediate states, embeddings, and attention outputs between the student and teacher models in their research on relation transfer between Q, K, and V matrix products.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liang et al., 2023)\", \"Explanation\": \"The cited work by Liang et al. provides the method of aligning student layers to the original teacher layer before pruning, which the citing paper adopts in their research to improve the performance of the student model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Sanh et al., 2019)\", \"Explanation\": \"The cited work by Sanh et al. (2019) provides a layer-level pruning strategy for model initialization, which the citing paper adopts in their research on choosing a subset of larger model parameters to retain maximum information about the data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2022Zhang et al. ( , 2023))\", \"Explanation\": \"The cited work is acknowledged as a publicly available model checkpoint that the citing paper can use for its research on decoder-only model checkpoints.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chen et al., 2023)\", \"Explanation\": \"The cited work introduces the Lion optimizer, which the citing paper adopts in all its pretraining experiments to train the baselines with different parameter sizes.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Raffel et al., 2020)\", \"Explanation\": \"The C4 dataset is the source of the web-crawled tokens used in the pretraining of the baselines in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Bisk et al., 2019)\", \"Explanation\": \"The cited work provides the Common Sense Reasoning PIQA dataset, which the citing paper utilizes in its research on the performance of language models in the context of common sense reasoning.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zellers et al., 2019)\", \"Explanation\": \"The cited work provides the Hellaswag dataset, which the citing paper uses in its research on the performance of language models in the context of common sense reasoning.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Sakaguchi et al., 2019)\", \"Explanation\": \"The cited work provides the Winogrande dataset, which the citing paper uses in its research on the performance of language models in the context of common sense reasoning.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Clark et al., 2019)\", \"Explanation\": \"The cited work provides the BoolQ dataset, which the citing paper uses in its research on the performance of language models in the context of common sense reasoning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Mihaylov et al., 2018)\", \"Explanation\": \"The cited work by Mihaylov et al. provides a dataset called OpenBookQA that the citing paper uses in their research to train and evaluate their model.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Welbl et al., 2017)\", \"Explanation\": \"The cited work by Welbl et al. contributes a dataset called SciQ that the citing paper uses in their research to train and evaluate their model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Clark et al., 2018)\", \"Explanation\": \"The cited work by Clark et al. provides a dataset called Arc-Easy and Arc-Challenge that the citing paper uses in their research to test the performance of their model in different settings.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Gao et al., 2021)\", \"Explanation\": \"The cited work by Gao et al. provides the evaluation harness used in the citing paper to assess the performance of the language model in a zero-shot setting.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Holtzman et al., 2021)\", \"Explanation\": \"The cited work by Holtzman et al. contributes a method for normalizing the completion likelihood in the evaluation of the language model, which is used in the citing paper to compute task accuracy.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a method for conditioning the probability of completion on the domain premise, which is used in the citing paper to assess the performance of the language model in a zero-shot setting.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang et al., 2020b)\", \"Explanation\": \"The miniLM-KD baseline is cited to indicate that the citing paper adopts the method of adding constraints on attention maps and intermediate representations in the teacher-teacher setup for distillation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liang et al., 2023)\", \"Explanation\": \"The homotopic-KD baseline is cited to highlight the method of adding constraints on each layer between the teacher and the student for attention maps and intermediate representations in the distillation setup.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Li et al., 2020)\", \"Explanation\": \"The cited work by Li et al. provides the basis for the observation in the citing paper that training a large model and then compressing it works better than training a smaller model from scratch, which the citing paper builds upon in the context of zero-shot evaluation of the compressed model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Shen et al., 2022)\", \"Explanation\": \"The cited work by Shen et al. provides a terminogloy that the citing paper uses to define the upper bound for the ideal truncated initialization method in terms of loss and training dynamics.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Shen et al., 2022)\", \"Explanation\": \"The cited work by Shen et al. provides a method for preserving a model's zero-shot promptability through truncated initialization, which the citing paper adopts in its research to ensure the model retains its zero-shot abilities during compression.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Dodge et al., 2021)\", \"Explanation\": \"The cited work by Dodge et al. (2021) is used to acknowledge the presence of harmful artifacts and biases in the C4 corpus, which the citing paper uses for pretraining their language models.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Hooker et al., 2020)\", \"Explanation\": \"The cited work by Hooker et al. (2020) has linked model compression to increased bias and toxicity in a model, which the citing paper extends by exploring the effects of compression in the context of pretraining language models.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Solaiman et al., 2023)\", \"Explanation\": \"The cited work by Solaiman et al. (2023) is an extension of the study on the effects of model compression on biases in large language models, which the citing paper builds upon to examine the impact of compression in the pretraining process.\"}]"}}},{"rowIdx":86,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_0","fig_1"],"heading":"Introduction","publication_ref":["b15","b12","b42","b44","b41","b87","b75","b7","b65","b37","b27","b30","b42","b4","b34","b24","b35","b43","b81","b21","b49","b24","b70"],"table_ref":[],"text":"Pre-trained Language Models (PLMs; Devlin et al., 2019;Clark et al., 2020) fine-tuned on task-specific training sets achieve remarkable near-human performance on held-out test sets, yet struggle to generalize to examples that are distributionally different * Equal Contribution (Ma et al., 2021). The simple heuristic used in this process can result in false negative options.\nfrom their training sets (McCoy et al., 2019;Ma et al., 2019;Zhou et al., 2021;Wang et al., 2021). This discrepancy arises because fine-tuned PLMs often rely on spurious, dataset-specific correlations to learn a task rather than learning to fully leverage implicit commonsense knowledge required for reasoning (Branco et al., 2021). For reasoning systems to be effective, though, they must be robust across domains and generalize beyond the specificities of individual datasets.\nTo confront the generalization issue in commonsense reasoning tasks, the task of zero-shot commonsense Question-Answering (QA) requires models to answer questions for evaluation benchmarks without access to their corresponding training data (Shwartz et al., 2020;Li et al., 2020). Among several methods that tackle this task, the most performant ones inject commonsense knowledge from CSKBs (Hwang et al., 2021;Jiang et al., 2021) into PLMs by fine-tuning them on synthetic QA pairs transformed from commonsense knowledge triples, where the head and relation are transformed to a question, and the tail serves as a ground answer. Negative examples are randomly sampled with keyword-overlap constraints (Ma et al., 2021). Such knowledge injection benefits not only QA tasks that are derived from CSKBs, such as So-cialIQA (Sap et al., 2019b), which is derived from ATOMIC (Sap et al., 2019a), but also QA datasets in other domains (Bisk et al., 2020).\nDespite recent advancements in this area, two major challenges remain. First, manually curated CSKBs, such as ATOMIC, are incomplete (Kuo and Hsu, 2010). While consolidating multiple CSKBs can improve coverage, it remains infeasible to cover all conceivable knowledge for the vast range of entities and situations in the real world (He et al., 2022). Automatic methods for expanding CSKBs exist, such as knowledge base completion (Li et al., 2016;Malaviya et al., 2020), and knowledge distillation from large language models (West et al., 2022;Gao et al., 2023), but they either fail to provide knowledge about novel entities or only provide highly accurate yet less informative knowledge (e.g., vague adjectives, such as happy, as situation descriptors). Second, in zero-shot commonsense QA, negative examples are required for models to learn to distinguish the validity of commonsense scenarios (Chen et al., 2023a). However, existing negative QA examples are synthesized using simple heuristic-based negative sampling without considering deeper semantics, resulting in too many false negative options. For instance, in Figure 1, \"have a drink\" is also plausible in the context of \"after playing a football game.\" These questions that label plausible options as negative instances confuse the model during training, impeding its ability to discern correct commonsense knowledge.\nWe tackle both of these challenges by utilizing conceptualization. As Murphy (2004) posits, humans rely on conceptual induction to draw inferences about unseen situations without the need for memorizing specific knowledge. Conceptualization (He et al., 2022) offers a similar capability by abstracting a set of instances into concepts, which allows for the derivation of abstract commonsense knowledge associated with each concept that can be instantiated to assist reasoning on specific downstream situations. For example, in Figure 2, \"play a football game\" can be conceptualized as a tiring event, which further generalizes as abstract knowledge. The benefits of conceptualization are twofold. First, conceptualized commonsense knowledge introduces abstract knowledge through a one-step concept inference based on the original CSKB, enhancing knowledge coverage. Second, as the abstract knowledge is conditioned on the original knowledge, the recall of knowledge regarding the same head is increased, leading to more finegrained constraints for negative option sampling.\nInspired by these advantages, we propose CAR (Conceptualization-Augmented Reasoner), a simple yet effective zero-shot commonsense QA framework that leverages conceptualization to expand existing CSKBs and reduce false-negative distractors. We first augment the original CSKB with conceptualization to infuse abstract commonsense knowledge to improve knowledge coverage. Then, we propose a conceptualization-constraint sampling strategy that generates distractors with conceptlevel constraints to prevent false negative options (Section 4). Experimental results on five popular commonsense QA benchmarks demonstrate the effectiveness of CAR, which even surpasses GPT3.5 and ChatGPT (Section 5). In Section 6, we analyze why CAR works by providing human evaluations that show a significant reduction of false negative options compared to other methods. Finally, our analysis reveals that conceptualizationaugmented training examples tend to be more ambiguous (Swayamdipta et al., 2020) than those produced by prior heuristics, leading to better out-ofdomain generalization."},{"figure_ref":[],"heading":"Related Works","publication_ref":["b73","b65","b17","b5","b1","b42","b69","b31","b64","b23","b66","b67","b49","b72","b18","b22","b38","b54","b24","b6","b13","b71","b50","b80","b80","b77","b51","b63","b81","b29","b21","b43"],"table_ref":[],"text":"Zero-shot Commonsense QA. Zero-shot commonsense QA evaluates a model's reasoning generalizability on unseen QA entries without any supervision signals from the corresponding annotated training data. To tackle this task, two primary pipelines have emerged in existing works. The first paradigm employs off-the-shelf language models without changing the parameters, either using vanilla language modeling with prompts (Trinh and Le, 2018;Li et al., 2022), or with some inferencetime mechanisms specifically designed for reasoning, such as self-talk (Shwartz et al., 2020), cloze translation (Dou and Peng, 2022), and dynamic generation of reasoning sub-graphs and graph reasoning (Bosselut et al., 2021). The second pipeline leverages external CSKBs as knowledge sources to provide PLMs with additional supervision signals for further fine-tuning (Banerjee and Baral, 2020;Ma et al., 2021;Su et al., 2022). A common strategy involves converting knowledge triples in CSKBs to synthetic QA pairs by transforming the head and relation to a question, the tail to a gold answer, and (randomly) sample tails from other heads as distractors. Such fine-tuning paradigm benefits from incorporating CSKBs within different domains (Kim et al., 2022;Shi et al., 2023) and exploiting multi-hop graph structures with graph neural networks (Guan et al., 2023), and heightens the model's commonsense sensitivity in a QA context, which leads to state-of-the-art performances.\nConceptualization. Conceptualization refers to the process of abstracting a group of instances or events into a general concept (Song et al., 2011(Song et al., , 2015)). In commonsense reasoning, it simulates conceptual induction (Murphy, 2004) and enables the derivation of abstract commonsense knowledge under the specific contextualization of the original commonsense knowledge (Tenenbaum et al., 2011), which is often lacking in existing CSKBs. Around many existing works studying conceptualization (Durme et al., 2009;Gong et al., 2016;Liu et al., 2022;Peng et al., 2022), He et al. (2022) investigate it at event-level semantics and construct AbstractATOMIC, an event conceptualization benchmark and knowledge base based on ATOMIC (Sap et al., 2019a). Recently, Wang et al. (2023a) propose to conceptualize CSKBs at scale with semi-supervised learning and demonstrate abstract knowledge can enhance commonsense inference modeling (Bosselut et al., 2019;Da et al., 2021). With current works mostly investigating the problem of conceptualization itself, none of them have extrinsically evaluated the impact of conceptualization on downstream tasks, such as commonsense QA (Talmor et al., 2019) or machine reading comprehension (Nguyen et al., 2016). Data Augmentation. Data augmentation aims at generating new examples from existing data to expand the size and diversity of a training set without requiring costly data annotations (Wei and Zou, 2019). Various methods have been proposed to augment textual data, including those using random perturbation (Wei and Zou, 2019), text embeddings (Wang and Yang, 2015), lexical semantics (Niu and Bansal, 2018), back translation (Sennrich et al., 2016), and large language models (West et al., 2022;Ismayilzada and Bosselut, 2023;Gao et al., 2023) for CSKB construction. Nevertheless, text-perturbation-based augmentations do not provide new knowledge to CSKBs, and knowledge mining from large language models suffers from high typicality (e.g., favoring simple commonsense over informative yet rare commonsense) and low density, still making negative sampling subject to false negatives (Malaviya et al., 2020)."},{"figure_ref":[],"heading":"Problem Definition","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Definitions","publication_ref":["b24","b42"],"table_ref":[],"text":"Conceptualization. Formally, denote a CSKB as D with knowledge triples in the format of\nD = {(h, r, t)|h ∈ H, r ∈ R, t ∈ T }\n, where H, R, and T are the sets of heads, relations, and tails in the original CSKB. Following He et al. (2022), the conceptualized CSKB, conditioned on D, can be denoted as\nD C = {(h c , r, t)|h c ∈ H c , r ∈ R, t ∈ T }\n, where H c is the set of conceptualized head events. Specifically, each conceptualized head h c is obtained by replacing a component i ∈ h with its abstract concept c while ensuring that the formed (h c , r, t) triple is still plausible in the original context (r, t). Such (h c , r, t) triples are commonly referred to as abstract commonsense knowledge.\nZero-shot Commonsense QA. In this paper, we employ the zero-shot commonsense QA task proposed by Ma et al. (2021) to study our framework. First, the CSKB D is transformed into multiple (Q i , A i ) pairs where Q i is a natural langauge question and A i = {A i,1 , A i,2 , ..., A i,m } is a set of options with m candidates. Specifically, for a given knowledge triple (h, r, t) ∈ D, we convert h, r into Q i via natural language templates and use t as the ground answer. Additionally, we retrieve m -1 distractors from other triples sampled from D using a manually defined strategy, such as keyword overlap filtering. The objective of our task is to train a QA model from the synthetic QA sets\nD Q = {(Q i , A i )|(h i , r i , t i ) ∈ D}.\nOnce trained, the model is tested on held-out test entries (Q test , A test ) from QA benchmarks. This requires the model to perform zero-shot commonsense reasoning since the training data from the target benchmarks are unavailable to the model."},{"figure_ref":[],"heading":"Dataset","publication_ref":["b24"],"table_ref":[],"text":"We use ATOMIC (Sap et al., 2019b) as the source CSKB D. ATOMIC contains inferential commonsense knowledge, in the format of (h, r, t) triple, that is associated with commonly seen events. Specifically, the heads of ATOMIC triples are events, whereas the tail nodes are either events or attributes. For conceptualization, we use the human-annotated abstract knowledge from Abstrac-tATOMIC (He et al., 2022) to train a generative conceptualizer for acquiring D C . More details of conceptualizations and statistics of AbstractATOMIC are provided in Section 4.1 and Appendix B.1."},{"figure_ref":[],"heading":"Evaluation Benchmarks","publication_ref":["b42"],"table_ref":[],"text":"Following Ma et al. (2021) "},{"figure_ref":["fig_3"],"heading":"CAR Framework","publication_ref":[],"table_ref":[],"text":"This section introduces our proposed CAR framework. A general sketch is presented in Figure 3. Our framework can be summarized into three steps:\n(1) Conduct one-step conceptualization inference on existing triples in the CSKB to obtain abstract commonsense knowledge triples.\n(2) Transfer the triples into QA pairs and generate distractors using keywords and conceptualizations as constraints.\n(3) Train the QA model using marginal ranking loss."},{"figure_ref":[],"heading":"Conceptualization Augmentation","publication_ref":["b83","b47"],"table_ref":[],"text":"To incorporate abstract knowledge into the CSKB, we begin by augmenting the (h, r, t) ∈ D triples by conducting a one-step conceptualization inference. Initially, given a head event h, we retrieve all plausible conceptualizations C h = {c i 1 ,1 , c i 1 ,2 , ...} for all identified instances i ∈ {i 1 , i 2 , ...|i ∈ h} using entity-linking heuristics to retrieve concepts from Probase (Wu et al., 2012) and WordNet (Miller, 1995). The conceptualized head event h c is then obtained by replacing an i ∈ h with one of its retrieved conceptualization c ∈ {c i,1 , c i,2 , ...}. This is done for all identified instances and their retrieved conceptualizations, thereby constructing the set of conceptualized head events of h. Subsequently, we link the non-abstract counterpart (r, t) after h c to generate candidate abstract knowledge triples (h c , r, t), where we adopt a discriminator trained with a semi-supervised conceptualizationinstantiation framework to determine their plausibility (Wang et al., 2023a). Only plausible triples are kept to form D C . Details about the conceptualization retrieval processes and the discriminator are presented in Appendix B.1."},{"figure_ref":["fig_3","fig_3"],"heading":"Concept-Constrained QA Synthesis","publication_ref":[],"table_ref":[],"text":"To synthesize a commonsense triple (h, r, t) into a (Q i , A i ) pair, we first transfer h, r into Q i by using natural language templates and set t as the groundtruth answer A 1 . For example, the triple in Figure 3 becomes \nC h = {c i 1 ,1 , c i 1 ,2 , • • • , c i 2 ,1 , • • • }, we as- sociate a triple (h, r, t) with T h + C h to form its constraint. Only knowledge triple (h ′ , r, t ′ ) which satisfies (T h ′ + C h ′ ) ∩ (T h + C h ) =\n∅ can be sampled as a distractor candidate. This constraint requires that the two triples have no common keywords, and their instances cannot be abstracted into the same conceptualization. For example, in Figure 3, \"(PersonX is at the casino, xWant, have a drink)\" cannot be used as a distractor triple because \"casino\" can be conceptualized as \"entertainment place,\" which is the same as \"bar\" in the original triple. Finally, we sample two distractor triples for the triple (h, r, t) and use the tails of these two triples as the distractors. To guarantee that the abstract commonsense knowledge from our previous augmentation is learnable by the QA model, we synthesize both the original triple (h, r, t) and its conceptualized versions (h c , r, t) into QA pairs."},{"figure_ref":["fig_3"],"heading":"Model Training","publication_ref":["b42"],"table_ref":[],"text":"Following Ma et al. (2021), we train our QA model by fine-tuning a pre-trained Masked Language Model (MLM) using the Marginal Ranking (MR) loss. Let C represent the original context (if any), Q represent the question, and (A 1 , A 2 , ...) be the list of options. We first concatenate C, Q, and an answer option A i together via natural language prompts to generate input sequences (T 1 , T 2 , ...).\nFor example, the synthesized question with its correct answer in Figure 3 will be transformed as:\n\"PersonX arrives at the bar, as a result, PersonX want to, relax himself.\" We then repeatedly mask out a token at one time and calculate the masked loss. The final MLM score for an input sequence T ∈ {T 1 , T 2 , ...} with n tokens is:\nS(T ) = - 1 n n i=1 log P (t i |..., t i-1 , t i+1 , ...) (1)\nAfter calculating the scores S 1 , S 2 , ... for all answer candidates A 1 , A 2 , ..., we compute the marginal ranking loss based on Equation 2, where η represents the margin and y is the index of the correct answer.\nL = 1 m m i=1,i̸ =y max(0, η -S y + S i )(2)\nDuring the evaluation phase, we use the same scoring procedure to assign a score to each option and select the one whose concatenated sentence achieves the lowest score as the model's prediction."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Setup","publication_ref":["b39","b25","b65","b5","b1","b69","b31","b42","b42","b2","b9","b56","b8","b81","b48","b79","b59"],"table_ref":[],"text":"Baselines First, we use random voting (Random) and most-frequent labeling (Majority) to demonstrate the characteristics of each benchmark. Vanilla RoBERTa-Large (Liu et al., 2019), and DeBERTa-v3-Large (He et al., 2023) PLMs are used to demonstrate the power of fine-tuning. The performances of these two models under a supervised training regime are also included to show the upper bound of our results. We also include the results of several existing approaches that tackle the same task, including Self-talk (Shwartz et al., 2020), COMET-DynaGen (Bosselut et al., 2021), SMLM (Banerjee and Baral, 2020), MICO (Su et al., 2022), and the previous state-of-the-art STL-Adapter (Kim et al., 2022). Most importantly, we compare our framework with Ma et al. (2021) to validate the efficacy of conceptualization since both methods share similar model architecture and training procedures. Both RoBERTa-Large and DeBERTa-v3-Large are used as the backbones for fair comparisons. There are, in total, 534,833 synthetic QA pairs provided by Ma et al. (2021).\nWith the recent advances in Large Langauge Models (LLMs) (Bang et al., 2023;Chan et al., 2023;Qin et al., 2023), we also benchmark the performances of GPT3.5 (Brown et al., 2020 2021) on ATOMIC (marked with △). ATM C stands for the ATOMIC with abstract commonsense knowledge injected. ATM-10X stands for using ATOMIC-10X (West et al., 2022) as the source CSKB D. All baseline results are consistent with their original papers.\net al., 2022) and ChatGPT (OpenAI, 2022) as baselines. We prompt the LLM directly in a zero-shot setting, where no in-context learning (Min et al., 2022) or chain-of-thought reasoning (Wei et al., 2022) are applied. For every QA entry, the LLM is presented with a question, several choices, and a natural language command that asks it to choose the index of the correct answer directly (Robinson et al., 2022). We then parse the generated outputs to obtain the \"predictions\" of LLM by using meticulously designed rules and compare them with the ground-truth labels. More details of the baselines and LLM setups can be found in Appendix B.2 and B.3."},{"figure_ref":[],"heading":"Implementation Details","publication_ref":["b40","b42","b31"],"table_ref":[],"text":"We use accuracy as the evaluation metric and compare our framework with the following baseline methods. For conceptualization, we leverage an off-the-shelf conceptualizer from Wang et al. (2023a), which is a semi-supervised conceptualization discriminator fine-tuned on labeled conceptualization data from AbstractATOMIC and unlabeled data from ATOMIC. We use a plausibility score T = 0.9 to filter out plausible conceptualizations, which results in 440K conceptualization-aided synthetic QA pairs for training. We employ an AdamW optimizer (Loshchilov and Hutter, 2019) with a learning rate of 7e-6 and a max sequence length of 128 to accommodate QA pairs with different lengths. We select the best checkpoint according to the highest accuracy achieved on the synthetic validation QA set. Each experiment is repeated using three different random seeds, and the average performance is reported. The model is warmed up with 5% of total iterations and evaluated every 1000 global steps, while the margin η for the marginal ranking loss is set to 1, in line with the choices made by Ma et al. (2021) and Kim et al. (2022). More details about implementations can be found in Appendix B.4,"},{"figure_ref":[],"heading":"Results","publication_ref":[],"table_ref":[],"text":"The main results are reported in Table 1. For the baselines, DeBERTa-v3-Large (MR) trained on ATOMIC achieves the best performance, followed by ChatGPT. Both achieve an accuracy of more than 70% on average. "},{"figure_ref":[],"heading":"Analysis and Discussion","publication_ref":["b70"],"table_ref":[],"text":"In this section, we study the effects of conceptualization and the reasons contributing to CAR's success. First, we conduct expert evaluations on the synthetic QA pairs to study the quality and diversity of different CSKB augmentation methods in comparison with conceptualization. Second, we conduct training dynamics (Swayamdipta et al., 2020) analysis to show that conceptualization-aided QA pairs can provide more ambiguous examples helpful for training. Finally, we study the impact of filtering ATOMIC10X with different critic thresholds, the ablations of CAR, and the effect of conceptualization from an out-of-domain generalization perspective in Appendix B.5, B.7, and B.8."},{"figure_ref":[],"heading":"Comparisons With Data Augmentations","publication_ref":["b55","b47","b45","b42","b42"],"table_ref":["tab_4","tab_4","tab_4","tab_4"],"text":"To demonstrate the effectiveness of our proposed conceptualization method, we conduct comprehensive analyses with other data augmentations that expand the semantic coverage of CSKBs in a similar way as conceptualization using both expert and automatic evaluations. We use EDA, augmenting with word embedding (Word2Vec; Mikolov et al., 2013 and GLOVE;Pennington et al., 2014), contextual embedding (BERT; Devlin et al., 2019), and synonym (WordNet;Miller, 1995) score (McHugh, 2012) of 0.64, comparable to 0.62, as reported by (Ma et al., 2021).\nDiversity. First, we study whether augmentations can introduce new knowledge to the training set. We begin by calculating the average cosine similarity of each ATOMIC triple and its augmented siblings from each method according to their Sen-tenceBERT (Reimers and Gurevych, 2019) embeddings. For ATOMIC-10X, we regard the sampled knowledge as augmentations. The complement of average similarity across all ATOMIC triples serves as an automatically measured diversity (Div.). Meanwhile, we retrieve top-10 similar triples from ATOMIC for each augmented triple according to their SentenceBERT similarity. The experts are asked to annotate whether each triple can be semantically covered by their retrievals. We define the expert-evaluated diversity as the ratio of uncovered triples among 300 samples. Table 2 shows that conceptualization champions both metrics, indicating that the introduced abstract knowledge is diverse and lacking in existing CSKBs, which is helpful in expanding their knowledge coverage.\nQuality of Synthetic QA Pairs. Next, we synthesize the augmented triples into QA pairs with their head events' keywords and augmentations as constraints. We then sample 300 QA pairs for each method and ask the same experts to perform expert evaluations by annotating the correctness of each QA pair's ground-truth answer and whether the distractors could also be plausible with respect to the augmented head event. This evaluates the plausibility ratio of the augmented knowledge and the ratio of QA pairs containing false negative distractors. Table 2 shows that the majority of augmented knowledge is implausible, and they fail to enhance distractors sampling. Conceptualization, on the other hand, maintains being highly plausible and can effectively eliminate false negative distractors.\nExpert annotators also achieve a remarkable ac-curacy of 86% while working on 300 randomly sampled question-answer pairs, surpassing the 80% accuracy reported by Ma et al. (2021).\nZero-shot Commonsense QA Performances.\nFinally, we train DeBERTa-v3-Large models on the QA pairs synthesized from the concatenation of both original and augmented ATOMIC triples from each method. Only keywords of each head event are used as their constraints. The models are trained using a marginal ranking loss, as explained in Section 4.3, and evaluated on five QA benchmarks in a zero-shot manner. Performances by different methods are shown in Table 2. We observe that conceptualization outperforms all other augmentations on average and successfully improves the model's zero-shot commonsense reasoning ability.\nComparison with ATOMIC-10X. Augmenting ATOMIC10X appears to be a promising option as it contains a wealth of valuable commonsense knowledge. However, despite its diverse and highquality knowledge, Table 2 demonstrates that the model cannot leverage this information effectively. One possible explanation is that the model's performance is hindered by the significantly high number of false-negative distractors. This issue arises because the knowledge distilled from GPT-3 tends to be versatile, resulting in many tail events being general and vague. These events can be applied to a large collection of heads, which leads to false negative options. More experiments and case studies are in Appendix B.5 and C, respectively."},{"figure_ref":["fig_4","fig_4"],"heading":"Training Dynamics Analysis","publication_ref":["b70"],"table_ref":[],"text":"Training dynamics effectively assess a model's confidence and variability for individual instances when training on a large dataset. In the context of QA, we define confidence as the model's certainty when assigning the correct label to the ground-truth option compared to distractors, as indicated by the logit difference. Variability, on the other hand, refers to the fluctuation of confidence over time.\nThese insights can aid in analyzing the model's behavior when different knowledge is introduced into the training set. More explanations are in Appendix B.6.\nIn this section, we examine the impact of abstract commonsense knowledge (conceptualization) and GPT3-distilled knowledge (ATOMIC-10X) by exploring their training dynamics on two sets of data. We train three QA models on synthetic QA pairs from conceptualization-augmented ATOMIC, ATOMIC10X-augmented ATOMIC, and the original ATOMIC, which serves as the baseline. First, we randomly select the same 1,000 QA pairs synthesized from the original ATOMIC and calculate their training dynamics using these three models. The left side of Figure 4 displays the alterations caused by the two augmentation methods in comparison with the baseline. It is evident that introducing abstract commonsense knowledge through conceptualization significantly reduces the model's average variability and enhances its confidence in learning the knowledge from the original ATOMIC. In contrast, incorporating knowledge from ATOMIC-10X produces the opposite effect.\nSecond, we check the training dynamics on 1,000 randomly sampled QA pairs synthesized from abstract commonsense knowledge and another 1,000 from knowledge in ATOMIC-10X. The rightmost plots in Figure 4 reveal that, compared to ATOMIC-10X, conceptualization introduces knowledge with higher variability and lower confidence, making it more ambiguous and challenging for the model to learn. As Swayamdipta et al. (2020) suggest, such data contributes to a more robust model to out-of-distribution (OOD) data, which are downstream QA datasets in our case. Therefore, we conclude that conceptualization is superior to ATOMIC-10X as abstract knowledge, on the one hand, makes the original knowledge more easy-to-learn to aid optimization, and on the other hand, provides more ambiguous examples to boost OOD generalization.\n3URSRUWLRQRI$70 C GDWD $YJ$FFXUDF\\ "},{"figure_ref":["fig_5"],"heading":"Impact of Training Data Size","publication_ref":[],"table_ref":[],"text":"In Figure 5, we present the influence of the number of training examples against the final performance, which reveals a clear and intuitive trend of a positive correlation between the amount of training data and overall performance."},{"figure_ref":[],"heading":"Generalization to other CSKBs","publication_ref":["b68","b47","b74"],"table_ref":["tab_6"],"text":"We explore the feasibility of transferring our framework to CSKBs other than ATOMIC. We take the CWWV dataset as an example, which comprises multiple CSKBs, including ConceptNet (Speer et al., 2017), WordNet (Miller, 1995), and Wiki-Data (Vrandecic and Krötzsch, 2014). We use the off-the-shelf GPT2 conceptualizer (Wang et al., 2023a) and ChatGPT as two flexible generative conceptualizers. The generated conceptualizations are then transformed into abstract knowledge and integrated into the CWWV dataset. The experimental results are presented in Table 3, which shows an improvement of over 1% compared to all baselines leveraging CWWV as the source of knowledge, indicating CAR's generalizability to other CSKBs. More details are presented in the Appendix B.9."},{"figure_ref":[],"heading":"Conclusions","publication_ref":[],"table_ref":[],"text":"In this paper, we present CAR, a pioneering framework for zero-shot commonsense QA empowered by conceptualization. Our approach surpasses even large language models on five QA benchmarks, achieving state-of-the-art performance on average. Our analyses reveal that conceptualization can improve the sampling of negative examples, and abstract knowledge is more helpful compared with those distilled from GPT3 as it provides more ambiguous knowledge to support OOD generalization. These findings demonstrate the substantial benefits of introducing conceptualization and abstract knowledge into zero-shot commonsense reasoning."},{"figure_ref":[],"heading":"Limitations","publication_ref":["b42","b31","b17","b68","b47","b33","b74","b24","b42"],"table_ref":[],"text":"One limitation of this paper is that the proposed CAR framework has only been validated on the ATOMIC dataset. While previous works (Ma et al., 2021;Kim et al., 2022;Dou and Peng, 2022) have studied the zero-shot commonsense question answering task by consolidating multiple CSKBs, including ATOMIC (Sap et al., 2019a), Concept-Net (Speer et al., 2017), WordNet (Miller, 1995), VisualGenome (Krishna et al., 2017), and Wiki-Data (Vrandecic and Krötzsch, 2014), our work only utilizes ATOMIC (more details discussed in Appendix B.2). This was mainly due to the availability of conceptualizations for the CSKB, with only AbstractATOMIC (He et al., 2022) being available as the conceptualized expansion of ATOMIC, while other CSKBs lack such resources. Additionally, ATOMIC has been shown to play the most critical role in experimental results by Ma et al. (2021). Nonetheless, such limitation does not restrict CAR's potential to seek further improvements from incorporating other CSKBs, as conceptualization frameworks, such as CAT (Wang et al., 2023a), can be applied to other CSKBs and provide the required resources for CAR to operate. Thus, we believe CAR can overcome such limitations and still possess the potential to improve with more CSKBassociated conceptualization resources available."},{"figure_ref":[],"heading":"Ethics Statement","publication_ref":[],"table_ref":[],"text":"This paper presents CAR, a novel framework for zero-shot commonsense question answering that achieves state-of-the-art performance via conceptualization. All datasets used, including ATOMIC, AbstractATOMIC, and commonsense questionanswering benchmarks, are publicly available and shared via open-access licenses solely for research purposes, consistent with their intended usage. These datasets are anonymized and desensitized, ensuring that no data privacy issues are involved. Moreover, the CAR framework is a questionanswering system that selects the most plausible choice from a list of options and does not yield any private, offensive, biased, or sensitive information or social and political issues. The expert annotations are performed by the authors of this paper as part of their contribution, who are graduate and undergraduate students working on machine commonsense in natural language processing, and they are fully aware of the annotation protocol and the intended use of their annotations. They are well-trained with specially designed instructions and have voluntarily agreed to participate without compensation. Based on this, the authors believe that this paper does not raise any ethical concerns to the best of their knowledge."},{"figure_ref":[],"heading":"Appendices A Benchmark Descriptions","publication_ref":["b3","b71","b68","b4","b60"],"table_ref":["tab_7"],"text":"In this section, we introduce more details regarding five evaluation benchmarks.\nAbductive NLI (aNLI) (Bhagavatula et al., 2020) is a Natural Langauge Inference (NLI) benchmark that aims to infer the most plausible explanation based on a given causal situation. For each question sample, the model is required to choose the more plausible hypothesis out of two options that fit the beginning and end of a story. This benchmark evaluates the model's abductive reasoning ability, which typically requires commonsense reasoning.\nCommonsenseQA (CSQA) (Talmor et al., 2019) is a question-answering benchmark that evaluates a broad range of commonsense aspects. Each sample contains a question and five choices. The question and some choices are generated from subgraphs of ConceptNet (Speer et al., 2017) while crowdsourcing annotators also annotate some distractors. This benchmark evaluates the model's concept-level commonsense reasoning ability.\nPhysicalIQA (PIQA) (Bisk et al., 2020) is a question-answering benchmark that requires the model to select the more plausible option out of two possible continuations given a common scenario that requires physical commonsense to infer. This benchmark evaluates the model's physical commonsense reasoning ability.\nSocialIQA (SIQA) (Sap et al., 2019b) is a question-answering benchmark that requires rea- soning about social interactions. Each sample contains a context that is derived from ATOMIC (Sap et al., 2019a), a question, and three choices. The questions are automatically generated using nine templates that correspond to the nine relations in ATOMIC, and the correct answers are crowdsourced. This benchmark evaluates the model's reasoning ability for emotional and social commonsense in daily situations. WinoGrande (WG) (Sakaguchi et al., 2021) is a pronoun resolution benchmark. Each sample contains an emphasized pronoun and a short context description. The model is asked to choose the correct reference given two options. This benchmark evaluates the model's pronoun resolution ability, which is also part of commonsense knowledge.\nIn our experiments, we use the validation splits of these benchmarks as the official testing sets may not be publicly available. Detailed statistics on the number of QA pairs and the number of options per question are reported in Table 4."},{"figure_ref":[],"heading":"B Additional Explanations and Analyses","publication_ref":[],"table_ref":[],"text":"In this section, we aim to cover additional details regarding the CSKB conceptualization in CAR (Appendix B.1), implementations of our system (Appendix B.4), baselines (Appendix B.2 and B.3), experiments using ATOMIC-10X (Appendix B.5), analyses (Appendix B.6, B.7, and B.8), and generalizability experiments (Appendix B.9) that are not covered in the body text due to space constraints."},{"figure_ref":[],"heading":"B.1 Definitions and Statistics of CSKB Conceptualization","publication_ref":["b84","b85","b24","b24","b47","b26","b57"],"table_ref":["tab_9"],"text":"Conceptualization plays a crucial role in generalizable commonsense reasoning. Previous studies have demonstrated its potential in aiding commonsense inference modeling (Wang et al., 2023a) and commonsense knowledge graph construction (Yu et al., 2023;Zhang et al., 2022). In our paper, we follow the definition of conceptualization proposed by He et al. (2022) and Wang et al. (2023a) in conceptualizing an instance within an event to a concept: (1) Events: Each event represents a commonly observed occurrence that encompasses valuable subsequential or inferential commonsense knowledge. In AbstractATOMIC, the events are the head events of all triples in ATOMIC without a wildcard ('_').\n(2) Instances: Within each event, multiple instances have been identified with semantic parsing tools, representing specific components of the event that are worthy of conceptualization.\n(3) Concepts: Concepts are the conceptualization of each instance. These concepts are thus extracted from Probase/WordNet and further validated by human annotators or critic filtering models. For an event e, which is the head of an ATOMIC triple, an instance refers to either an entity within the event or the complete event itself. Multiple instances can exist within a single event, denoted as i 1 , i 2 , i 3 , . . . , i n ∈ e. A concept corresponds to the conceptualization of an instance, and multiple conceptualizations can be associated with a single instance, as demonstrated by (i 1 , c 1 , 1), (i 1 , c 1 , 2), (i 1 , c 1 , 3), ..., (i 2 , c 2 , 1), ..., (i n , c n , 1), . . . , (i n , c n , m). For instance, consider the event \"PersonX is drunk when exiting the bar.\" In this case, two instances can be identified: \"Per-sonX is drunk when exiting the bar\" and \"bar.\" The conceptualization for the instance \"PersonX is drunk when exiting the bar\" may include \"drunk\" or \"enjoyed,\" while the instance \"bar\" can be conceptualized as an \"entertainment place\" or a \"fun place.\"\nIn this paper, we leverage the AbstractATOMIC dataset, provided by He et al. (2022), as our primary source for conceptualizations. Abstrac-tATOMIC is a benchmark for conceptualized commonsense knowledge that is built upon the ATOMIC dataset (Sap et al., 2019a). It contains three folds of data, each conditioned on the original commonsense knowledge triples (h, r, t) in ATOMIC.\nIn the first fold, He et al. ( 2022) identify all possible instances {i 1 , i 2 , i 3 , • • • |i ⊆ h} in each ATOMIC head event, using syntactic parsing through a spaCy 1 parser and matching with five human-defined rules. It is important to note that, unlike traditional entity-level conceptualization benchmarks, the identified instance in Abstrac-tATOMIC can also be the entire head event i = h.\nIn the next fold, each identified instance i is heuristically matched against Probase (Wu et al., 1 https://spacy.io/ 2012) and WordNet (Miller, 1995) via Gloss-BERT (Huang et al., 2019) to find their corresponding conceptualization candidates. Human annotations are conducted to verify part of the plausibility of the conceptualization candidates. To pseudolabel unannotated conceptualizations, we use a semi-supervised conceptualization discriminator provided by Wang et al. (2023a) and set a threshold of T = 0.9 to filter out plausible conceptualizations. Additionally, we utilize a GPT2-based (Radford et al., 2019) generator, trained on the concatenation of annotated and positively pseudo-labeled conceptualizations, to generate additional conceptualizations for further expanding the size of the conceptualization bank.\nHowever, it is worth noting that such conceptualization may not yield plausible abstract knowledge when (r, t) is connected back to h c , where h c is obtained by replacing i ∈ h with its conceptualizations. This is because the process of conceptualizing a head event omits its context in (r, t). Thus, the last fold of data stores the plausibility of such abstract commonsense triples (h c , r, t), where human annotations are conducted to verify part of the triples' plausibilities. In addition, we adopt a semi-supervised instantiation discriminator, provided by Wang et al. (2023a), to pseudo-label the unannotated triples. Another threshold, T = 0.9, is set to filter out plausible abstract triples.\nIn the CAR framework, for every ATOMIC event h, we retrieve every instance i's plausible conceptualizations {c i,1 , c i,2 , • • • } from all plausible conceptualizations derived in the second fold to serve as the distractor sampling constraint. We also augment the original (h, r, t) triples with their plausible (h c , r, t) siblings from both human-annotated and pseudo-labeled triples, as explained in the last fold. These knowledge triples are then synthesized into QA pairs using our proposed method to train the model to perform general reasoning. Detailed statistics for the conceptualizations and abstract commonsense triples we finally obtained from the AbstractATOMIC dataset are reported in Table 5 and Table 6, respectively."},{"figure_ref":[],"heading":"B.2 Baseline Performances","publication_ref":["b1","b1","b31","b68","b47","b74"],"table_ref":[],"text":"For SMLM (Banerjee and Baral, 2020), we adopt the official implementation of Banerjee and Baral (2020), which employs the CSKB that exhibits the highest alignment with each task. Specifically, So-cialIQA uses ATOMIC, while CommonsenseQA uses ConceptNet. For STL-Adapter (Kim et al., 2022), only those trained on ATOMIC are used for comparison in the body text. In this paper, all baseline performances are solely based on their officially reported results in their respective papers.\nAs noted in the Limitations section, previous research in this area has primarily focused on using four CSKBs, namely ATOMIC (Sap et al., 2019a), ConceptNet (Speer et al., 2017), WordNet (Miller, 1995), and WikiData (Vrandecic and Krötzsch, 2014). In order to comprehensively benchmark our framework's performance in the field of zeroshot commonsense QA, we compare our results on ATOMIC against baseline methods that use multiple CSKBs despite the unbalanced amount of knowledge in such a comparison. Table 11 presents a full comparison of our method with all existing baselines. Notably, for models based on RoBERTa-Large, our approach trained only on abstract knowledge injected ATOMIC achieves second place in the leaderboard, falling only behind Kim et al. ( 2022) with four CSKBs. While this comparison may be unfair due to the unbalanced amount of knowledge, it provides a strong justification for the excellent performance of our system. Our DeBERTa-v3-Large-based model still surpasses all baselines on average, indicating the necessity of leveraging a strong pre-trained language model."},{"figure_ref":[],"heading":"B.3 Benchmarking Large Language Models","publication_ref":["b56","b9","b71","b60","b59","b71","b0","b16","b79"],"table_ref":[],"text":"We then discuss our method for benchmarking large language models on five commonsense QA benchmarks. The emergence of Large Language Models (LLMs), such as ChatGPT (OpenAI, 2022), has been the hot trend in recent NLP research. Numerous studies have evaluated the capability of LLMs on various NLP downstream tasks. Among them, Qin et al. (2023); Chan et al. (2023) have shown that ChatGPT can achieve competitive performance on commonsense reasoning tasks, such as CommonsenseQA (Talmor et al., 2019), Wino-Grande (Sakaguchi et al., 2021), and Commonsense Knowledge Base Population (Fang et al., 2021b,a). In this study, we aim to benchmark Chat-GPT's zero-shot performance on five QA evaluation benchmarks used in our zero-shot commonsense QA task. Following (Robinson et al., 2022), we design and leverage a batch of prompts, as shown in Table 7, to probe ChatGPT's predictions. The prompt provides ChatGPT with a question and its possible choices, along with a natural language command to control the response action of ChatGPT. We then parse the generations using our meticulously designed rules, where punctuations and irrelevant wordings will be dropped, and the first choice-letter prediction will be identified as ChatGPT's answer. Specifically, if ChatGPT hesitates and cannot make a concrete prediction, it will be counted as a wrong answer. The benchmarking results are shown in Table 1. We observe that ChatGPT demonstrates superior performance compared to GPT3.5 and excels in tasks such as CommonsenseQA (Talmor et al., 2019) and So-cialIQA (Sap et al., 2019b), potentially due to the high frequency of their questions and answers in large text corpora. However, its performance on the remaining three benchmarks is suboptimal, suggesting that they are more challenging and require more complex reasoning (Bai et al., 2023;Ding et al., 2023) and implicit commonsense knowledge to solve. This intriguing outcome warrants further investigation to determine the reasons behind it and explore methods to boost the LLM's abilities in these challenging benchmarks.\nGenerally speaking, CAR and conceptualization own the advantage over the large language model in the following aspects: (1) Smaller Model Size: The CAR framework offers models that are significantly smaller in scale compared to LLMs (0.2% of a standard 175 billion parameter GPT-3 model) while maintaining comparable performance in a zero-shot setting. Such size makes it more efficient in terms of training and deployment. In contrast, advanced prompting techniques used in LLMs require extensive computational resources, making the conceptualization-based model more versatile and accessible to researchers with limited access or resources for deploying LLMs. (2) Broader Commonsense Knowledge: Conceptualization provides a broader range of commonsense knowledge compared to current CSKBs. Integrating this type of knowledge into generative models has been shown to enhance their performance in commonsense reasoning tasks (Wang et al., 2023a). Such knowledge can also be dynamically encoded in language models during inference time (Chen et al., 2023b).\n(3) Advanced Prompting of LLMs: Conceptualization data introduces the potential for a more advanced prompting framework of reasoning with LLMs. The process of conceptualization and the instantiation of knowledge play a crucial role in reasoning. Thus, future work may consider introducing the \"chain of concept\" reasoning process to further advance the popular \"chain of thought\" reasoning paradigm (Wei et al., 2022;Wang et al., 2023b)."},{"figure_ref":[],"heading":"B.4 Implementation Details","publication_ref":["b82","b42","b40","b81","b42","b31"],"table_ref":[],"text":"We present additional implementation details for building our system. For the pre-trained language models, We use PLMs from the Huggingface Transformers2 Library (Wolf et al., 2020) as the vanilla model checkpoints. Our system relies heavily on the open-source code repository3 provided by Ma et al. (2021) for synthesizing QA pairs and training the QA models. To optimize the model, we employ an AdamW optimizer (Loshchilov and Hutter, 2019) with a learning rate of 7e-6 and a max sequence length of 128 to accommodate QA pairs with different lengths. When evaluating the model on downstream commonsense QA benchmarks, a maximum sequence length of 80 is used. We select the best checkpoint according to the highest accuracy achieved on the synthetic validation QA set. Each experiment is repeated using different random seeds three times, and the average performance is Table 8: Zero-shot evaluation results (%) on five commonsense question answering benchmarks using different critic thresholds for filtering ATOMIC-10X. The best results are bold-faced, and the second-best ones are underlined. ATM C stands for the ATOMIC with abstract commonsense knowledge injected. ATM-10X stands for using ATOMIC-10X (West et al., 2022) as the source CSKB D. ATM ATM-10X indicates the ATOMIC with sampled knowledge from ATOMIC-10X injected. Critic indicates the lower bound for filtering knowledge from ATOMIC-10X, which means that only knowledge with a critic score above the threshold will be selected. reported. To overcome the limited GPU memory issue, we utilize gradient accumulation with a gradient accumulated four steps before every descent, and each step calculates the gradient of eight data samples. The model is warmed up with 5% of total iterations and evaluated every 1000 global steps, while the margin η for the marginal ranking loss is set to 1, in line with the choices made by Ma et al. (2021) and Kim et al. (2022). The Huggingface model code for our RoBERTa-Large model is roberta-large, and the one for our DeBERTa-v3-Large model is microsoft/deberta-v3-large. All of our experiments are conducted on eight NVIDIA A100 GPUs, each with 40G graphical memory. Training a RoBERTa-based model typically requires 14G graphical memory, while training DeBERTa-based models requires 30G graphical memory."},{"figure_ref":[],"heading":"B.5 Experiments with ATOMIC-10X","publication_ref":["b81","b27","b42","b42","b14"],"table_ref":["tab_4"],"text":"ATOMIC-10X is a machine-generated corpus developed by West et al. (2022) using the symbolic knowledge distillation framework. They employed a selective distillation approach to extract knowledge from large language models like GPT-3 (Brown et al., 2020) by prompting them with head events and commonsense relations from the ATOMIC dataset. The extracted knowledge was used to train a student model to generate symbolic knowledge graphs evaluated using a separate critic model. The resulting corpus, ATOMIC-10X, surpassed the human-generated corpus ATOMIC2020 (Hwang et al., 2021) in scale, accuracy, and diversity.\nIn this section, we provide additional explanations regarding the usage of ATOMIC-10X in our paper and conduct further experiments to explore its impact on zero-shot commonsense QA. Specifically, we study the role of filtering the knowledge in ATOMIC-10X using multiple critic thresholds to acquire high-quality knowledge and improve model performance.\nWe utilize ATOMIC-10X in two distinct scenarios. First, as discussed in Section 5.2, we directly train our QA models on ATOMIC-10X without integrating other CSKBs, such as ATOMIC and AbstractATOMIC. We source all questions, answers, and distractors from ATOMIC-10X and follow its original train/dev/test partitions to divide the data. The lemmatized tokens of the head event, excluding commonly seen subjects, prepositions, and stopwords, are used as keywords for each piece of knowledge, and the original QA synthesis pipeline proposed by Ma et al. (2021) is applied. To ensure the quality of the knowledge from ATOMIC-10X, we set multiple critic thresholds to filter the dataset accordingly. The QA models are trained using marginal ranking loss on four subsets of ATOMIC-10X, each with a different critic threshold of 0.9, 0.8, 0.7, and 0.5, along with an additional trained on the complete ATOMIC-10X dataset. Finally, we evaluate these models on five commonsense QA benchmarks in a zero-shot setting and report the results in Table 8. Specifically, models trained solely on ATOMIC-10X using critic thresholds of 0.7 (RoBERTa) and 0.0 (DeBERTa) for filtering are responsible for the results reported in Table 1. We observe that even using high critic thresholds to filter the knowledge in ATOMIC-10X, the model still fails to improve beyond marginal. Meanwhile, training the models only on ATOMIC-10X fails to surpass training on ATOMIC, which indicates that the amount of knowledge is not the critical element to determining the performance. Rather, it should be the diversity and quality of knowledge, where the humanannotated knowledge from ATOMIC is superior to those machine-generated ones from ATOMIC. Nonetheless, none of the models outperform those trained on conceptualization-augmented ATOMIC using our CAR framework, which further validates the strengths of CAR.\nIn the second scenario, as discussed in Section 6.1, we utilize ATOMIC-10X as a means of augmentation to extend the original ATOMIC dataset. This is achieved by randomly selecting a specific number of knowledge triples from ATOMIC-10X, equivalent to the total number of plausible abstract commonsense knowledge in AbstractATOMIC, and merging them back into the original dataset. The triples in the resulting ATOMIC10X-augmented ATOMIC are then transformed into QA pairs and used to train our model following the original pipeline suggested by Ma et al. (2021). Similar to the previous scenario, we set four thresholds, namely 0.9, 0.8, 0.7, and 0.5, to filter the triples in ATOMIC-10X for augmentation quality control. In this way, the QA pairs' distractors can come from ATOMIC and ATOMIC-10X. The models are then trained and evaluated on five benchmarks. Their zero-shot commonsense QA evaluation results are reported in Table 8, and the best model, trained using a critic threshold of 0.8 for filtering with DeBERTa-v3-large as the backbone, is responsible for the results indicated in Table 2. Interestingly, we observe that leveraging the knowledge in ATOMIC-10X, either for direct training or augmentation, occasionally improves the model's performance on a specific benchmark. However, it fails to boost the overall performance across all benchmarks on average, which is considered a closer metric for evaluating the generalizable reasoning ability of a commonsense QA model. Thus, we come to the conclusion that ATOMIC-10X is inconsistently helpful in improving the zero-shot commonsense QA performances, with most times failing to improve, while conceptualization resolves such issues and can benefit the model across all benchmarks significantly. One potential reason is that ATOMIC-10X main contain noise that are not benefitial to the task of zero-shot commonsense QA, as demonstrated by Deng et al. (2023)."},{"figure_ref":["fig_4"],"heading":"B.6 Training Dynamic Definitions","publication_ref":["b70","b70","b70","b70"],"table_ref":[],"text":"Training dynamic, as proposed by Swayamdipta et al. (2020), refers to the analysis of a model's behavior on individual instances during training on large datasets. This analysis examines the model's confidence in predicting the true class of an instance and the variability of this confidence across epochs. To achieve this, multiple checkpoints are saved throughout a training epoch, and probability scores are derived for each data instance to calculate their training dynamics. By plotting the training dynamics of all instances on a data map, instances can be categorized into three groups: easyto-learn, ambiguous, and hard-to-learn. For instance, consider a QA pair where a model consis-tently assigns a higher logit score to the correct answer than to other distractors across multiple checkpoints during an epoch. In this scenario, the model exhibits high confidence and low variability for that specific instance, suggesting that it is easy to learn. Conversely, instances with higher variability are ambiguous to the model, and those with low confidence are difficult to learn. Experimental results by Swayamdipta et al. (2020) demonstrates that training the model with ambiguous data contributes the most to out-of-distribution generalization.\nInspired by this finding, our research investigates the role of abstract commonsense knowledge within the training set and the effects of leveraging conceptualization. Since our QA model is trained with a marginal ranking loss, as described in Section 4.3, it does not output a probability score but rather an MLM score for each option. Thus, the definition of model's confidence proposed by Swayamdipta et al. (2020) does not fit into our problem definition. To address this, we re-define the calculation of confidence to align with the model's degree of certainty in predicting an instance as the true class. Formally, denote n as the number of saved checkpoints during an epoch for computing their training dynamics and the list of m options in a (Q i , A i ) pair as A i = {A i,1 , A i,2 , ..., A i,m } with A i,j being the ground-truth answer (1 ≤ j ≤ m)). We define the confidence of the model for such a QA pair in Equation 3, where σ is the sigmoid function and S c i,d is the score of option A i,d at checkpoint c.\nC(Q i , A i ) = 1 n n c=1 σ( m d=1 (S c i,d -S c i,j ) m -1 )(3)\nIntuitively, this equation averages the gap between the ground-truth answer's score and the score of each distractor. A larger gap indicates a more confident model when choosing the answer. Variability aligns with the definition established by Swayamdipta et al. (2020). Specifically, it is calculated as the standard deviation of the score gap between the ground-truth answer and the distractors relative to the level of confidence exhibited throughout an entire epoch, as shown in Equation 4. By revisiting the plots in Figure 4, we observe that the inclusion of abstract commonsense knowledge enhances the model's confidence and reduces variability when encountering knowledge in ATOMIC. The introduction of conceptualization appears to widen the differences between the model's predicted scores for the correct answer and those for the distractors. This suggests that the correct answer is more likely to be selected, leading to an improved learning outcome. However, the introduction of knowledge from ATOMIC-10X results in a reversed trend, indicating that it does not aid in better learning ATOMIC. Furthermore, we observe that abstract knowledge derived from conceptualizations is more ambiguous to the model in the conceptualization-augmented ATOMIC, which theoretically contributes more to out-of-domain generalization. Nonetheless, ATOMIC-10X still contains some easy-to-learn knowledge that does not facilitate the model's generalization. Thus, abstract commonsense knowledge benefits zero-shot commonsense QA better than ATOMIC-10X by providing more ambiguous conceptual knowledge, which aids in making the model more generalizable.\nV(Qi, Ai) = n c=1 (σ( m d=1 (S c i,d -S c i,j ) m-1 ) -C(Qi, Ai)) 2 n (4)\nWe also plot the changes in training dynamics on different QA benchmarks, comparing models with and without the injection of abstract knowledge. The plots are shown in Figure 7. We observe that the inclusion of abstract commonsense knowledge significantly improves the models' confidence in downstream QA entries. However, the impact on the trend of variability is unclear. Nevertheless, this improvement in average confidence provides strong evidence for the model's enhancement in these downstream QA benchmarks."},{"figure_ref":[],"heading":"B.7 Ablation Study","publication_ref":["b42"],"table_ref":[],"text":"Next, we study the ablation of different components in our CAR framework to determine the impact of utilizing conceptualization through various techniques. There are two critical components that distinguish CAR from traditional zero-shot QA systems (Ma et al., 2021):\n• Conceptualization Augmentation: Augmenting the original commonsense knowledge in a CSKB with its conceptualizations to derive abstract commonsense knowledge. This knowledge is then synthesized into QA pairs, enabling the model to reason from a more generalized perspective. Without this component, abstract commonsense knowledge is not incorporated into the CSKB. Conceptualizations still remain as constraints for assisting QA pair synthesis, resulting in an approach that is similar to applying our proposed QA synthesis protocol directly to ATOMIC.\n• Concept-Constrained QA Synthesis: Constraining a question's distractors by ensuring that none of their head events share a common keyword or conceptualization with question's keywords and conceptualizations. If this component is dropped, the constraint will be downgraded, and only no sharing of common keywords between the question and distractors will be restricted. This approach introduces abstract commonsense knowledge into the CSKB and uses the original distractor generation strategy for synthesizing QA pairs.\nWe then train two batches of QA models, using RoBERTa-Large and DeBERTa-v3-Large as the backbone, by sequentially dropping the two components mentioned above one at a time. Their zeroshot performances on five commonsense QA benchmarks are reported in Table 9. From the results, it is observed that both components play important roles in CAR, with CCQS being more effective on average. This underscores the significance of eliminating false negative distractors, and conceptualization proves to be a useful tool for achieving this objective in improving the QA model's overall performance."},{"figure_ref":["fig_6"],"heading":"B.8 The Effect of Conceptualization","publication_ref":["b86","b42"],"table_ref":[],"text":"Lastly, we study the improvement in the generalizability of our framework with the aid of conceptualizations by examining the accuracy gains on questions with varying levels of semantic overlap with knowledge in ATOMIC's training split. To do so, we sort the questions in every benchmark by their average BERTScore (Zhang et al., 2020) between each individual question entry against the whole training set in the original ATOMIC. We then split the questions into two sets based on their BERTScores, with the lower BERTScore indicating a lower semantic overlap and a greater need for the model to generalize to answer the question. These questions are denoted as \"Difficult.\" Conversely, we refer to questions with high BERTScores as \"Easy.\" Then, we train two QA models following the pipeline proposed by Ma et al. (2021), with one trained on conceptualization-augmented ATOMIC and the other on ATOMIC only. We evaluate their performance on five commonsense QA benchmarks and compare the performance gains between two sets of questions in each benchmark, as shown in Figure 6. Results demonstrate that incorporating conceptualizations positively impacts accuracy, particularly for questions that deviate significantly from ATOMIC across multiple benchmarks. This indicates that augmenting ATOMIC with conceptualizations can improve the model's generalizability, particularly for questions that tend to be out-ofdistribution, requiring more relevant knowledge to answer correctly."},{"figure_ref":[],"heading":"B.9 Generalization to Other CSKBs","publication_ref":["b42","b68","b47","b74","b57"],"table_ref":["tab_11"],"text":"While our work primarily experiments with the AbstractATOMIC dataset as the conceptualization source of ATOMIC, we also aim to extend our framework to other CSKBs for a more generalizable evaluation. To address this, we follow Ma et al. (2021) and explore the feasibility of transferring our framework to the CWWV dataset, which comprises multiple CSKBs including Concept-Net (Speer et al., 2017), WordNet (Miller, 1995), and WikiData (Vrandecic and Krötzsch, 2014). To accomplish this, we train a conceptualization generator based on GPT2 (Radford et al., 2019) and utilize ChatGPT (OpenAI, 2022) as two flexible generative conceptualizers. The generated conceptualizations are then transformed into abstract knowledge and integrated into the CWWV dataset. This augmented dataset is employed to train a zeroshot commonsense QA reasoner using our proposed CAR framework. We present the experimental results and compare them with baselines in Table 10. Our observations reveal a modest improvement in an average accuracy of 1% compared to all baselines and comparable performance to GPT3.5. These results demonstrate the effectiveness of incorporating conceptualizations from other CSKBs. In future research, we suggest exploring automatic construction methods for conceptualization resources in other CSKBs and investigating their potential benefits for general commonsense reasoning."},{"figure_ref":[],"heading":"C Case Study","publication_ref":[],"table_ref":["tab_4","tab_6"],"text":"In this section, we present case studies to demonstrate the effectiveness of CAR. First, we discuss cases that illustrate the power of conceptualization augmentation, as shown in Table 12. By transforming triples into abstract commonsense knowledge, we can introduce more general knowledge into the CSKB and improve its coverage. Moreover, the newly introduced triples were missing from the original CSKB. For instance, conceptualizing \"Per-sonX plays the games together\" as an \"entertainment activity\" introduces higher-level knowledge that cannot be simply represented by the original triple. Additionally, by synthesizing both types of triples into QA pairs, the QA model can learn both types of knowledge, which can help the model perform more generalizable reasoning on out-ofdistribution commonsense QA benchmarks.\nNext, in Table 13, we present QA pairs consisting of false negative options generated using keyword constraints during their synthesis from both the original ATOMIC and ATOMIC-10X. We also demonstrate how our concept constraint resolves this issue. Through these case studies, we observe that the original distractors may contain one or even both plausible options, which is suboptimal when training a QA model. Specifically, for distractors sampled from ATOMIC-10X, we observe that several distractors are vague and general (denoted as \"?\"), which can be plausible in many contexts. For example, in various triples, adjectives like \"happy\" and verb phrases such as \"do it\" are easy to be plausible and do not serve as significant distractions. This is not desirable when training a QA model. However, by using conceptualizations as a constraint, the newly sampled distractors are all strong negatives, allowing the model to learn from such negative commonsense knowledge. This is because the distractors are sourced from triples that are more likely to be irrelevant to the original triple's context and, thus, more likely to be truly negative distractors."},{"figure_ref":[],"heading":"Acknowledgements","publication_ref":[],"table_ref":[],"text":"The authors would like to thank Haochen Shi for his help in implementing the training dynamics and the anonymous reviewers for their constructive comments. The authors of this paper were supported by the NSFC Fund (U20B2053) from the NSFC of China, the RIF (R6020-19 and R6021-20), and the GRF (16211520 and 16205322) from RGC of Hong Kong. We also thank the support from the UGC Research Matching Grants (RMGS20EG01-D, RMGS20CR11, RMGS20CR12, RMGS20EG19, RMGS20EG21, RMGS23CR05, RMGS23EG08). We also gratefully acknowledge the support of the Swiss National Science Foundation (No. 215390), Innosuisse (PFFS-21-29), the EPFL Science Seed Fund, the EPFL Center for Imaging, Sony Group Corporation, and the Allen Institute for AI."},{"figure_ref":[],"heading":" ","publication_ref":["b68","b33","b74","b47","b28"],"table_ref":[],"text":"et al., 2022)\n. All baseline results are consistent with their original papers. CWWV refers to the combination of ConceptNet (Speer et al., 2017), VisualGenome (Krishna et al., 2017), WikiData (Vrandecic and Krötzsch, 2014), and WordNet (Miller, 1995). CSKG (Ilievski et al., 2021) consists of ATOMIC (Sap et al., 2019a) and CWWV. Table 12: Case study of conceptualized triples and their synthesized QA pairs. Given an original triple from ATOMIC, we conceptualize the triple by replacing an instance with its [plausible conceptualization] to form a conceptualized triple. The conceptualized triples are then synthesized into QA pairs using the same ground-truth answer and distractors, sampled for the original triple, to train the QA model. * indicates the ground-truth answer. Table 13: Case study of the false negative options in the original QA synthesis and how our proposed conceptualization-constraint resolves such an issue. The (ground truth answer) is appended below each question. Keyword represents only using keywords as the constraint for sampling distractor, while Concept refers to using both the keywords and conceptualizations. F.Neg. refers to whether the distractor is a false negative one or not."}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_1\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b15\",\n \"b12\",\n \"b42\",\n \"b44\",\n \"b41\",\n \"b87\",\n \"b75\",\n \"b7\",\n \"b65\",\n \"b37\",\n \"b27\",\n \"b30\",\n \"b42\",\n \"b4\",\n \"b34\",\n \"b24\",\n \"b35\",\n \"b43\",\n \"b81\",\n \"b21\",\n \"b49\",\n \"b24\",\n \"b70\"\n ],\n \"table_ref\": [],\n \"text\": \"Pre-trained Language Models (PLMs; Devlin et al., 2019;Clark et al., 2020) fine-tuned on task-specific training sets achieve remarkable near-human performance on held-out test sets, yet struggle to generalize to examples that are distributionally different * Equal Contribution (Ma et al., 2021). The simple heuristic used in this process can result in false negative options.\\nfrom their training sets (McCoy et al., 2019;Ma et al., 2019;Zhou et al., 2021;Wang et al., 2021). This discrepancy arises because fine-tuned PLMs often rely on spurious, dataset-specific correlations to learn a task rather than learning to fully leverage implicit commonsense knowledge required for reasoning (Branco et al., 2021). For reasoning systems to be effective, though, they must be robust across domains and generalize beyond the specificities of individual datasets.\\nTo confront the generalization issue in commonsense reasoning tasks, the task of zero-shot commonsense Question-Answering (QA) requires models to answer questions for evaluation benchmarks without access to their corresponding training data (Shwartz et al., 2020;Li et al., 2020). Among several methods that tackle this task, the most performant ones inject commonsense knowledge from CSKBs (Hwang et al., 2021;Jiang et al., 2021) into PLMs by fine-tuning them on synthetic QA pairs transformed from commonsense knowledge triples, where the head and relation are transformed to a question, and the tail serves as a ground answer. Negative examples are randomly sampled with keyword-overlap constraints (Ma et al., 2021). Such knowledge injection benefits not only QA tasks that are derived from CSKBs, such as So-cialIQA (Sap et al., 2019b), which is derived from ATOMIC (Sap et al., 2019a), but also QA datasets in other domains (Bisk et al., 2020).\\nDespite recent advancements in this area, two major challenges remain. First, manually curated CSKBs, such as ATOMIC, are incomplete (Kuo and Hsu, 2010). While consolidating multiple CSKBs can improve coverage, it remains infeasible to cover all conceivable knowledge for the vast range of entities and situations in the real world (He et al., 2022). Automatic methods for expanding CSKBs exist, such as knowledge base completion (Li et al., 2016;Malaviya et al., 2020), and knowledge distillation from large language models (West et al., 2022;Gao et al., 2023), but they either fail to provide knowledge about novel entities or only provide highly accurate yet less informative knowledge (e.g., vague adjectives, such as happy, as situation descriptors). Second, in zero-shot commonsense QA, negative examples are required for models to learn to distinguish the validity of commonsense scenarios (Chen et al., 2023a). However, existing negative QA examples are synthesized using simple heuristic-based negative sampling without considering deeper semantics, resulting in too many false negative options. For instance, in Figure 1, \\\"have a drink\\\" is also plausible in the context of \\\"after playing a football game.\\\" These questions that label plausible options as negative instances confuse the model during training, impeding its ability to discern correct commonsense knowledge.\\nWe tackle both of these challenges by utilizing conceptualization. As Murphy (2004) posits, humans rely on conceptual induction to draw inferences about unseen situations without the need for memorizing specific knowledge. Conceptualization (He et al., 2022) offers a similar capability by abstracting a set of instances into concepts, which allows for the derivation of abstract commonsense knowledge associated with each concept that can be instantiated to assist reasoning on specific downstream situations. For example, in Figure 2, \\\"play a football game\\\" can be conceptualized as a tiring event, which further generalizes as abstract knowledge. The benefits of conceptualization are twofold. First, conceptualized commonsense knowledge introduces abstract knowledge through a one-step concept inference based on the original CSKB, enhancing knowledge coverage. Second, as the abstract knowledge is conditioned on the original knowledge, the recall of knowledge regarding the same head is increased, leading to more finegrained constraints for negative option sampling.\\nInspired by these advantages, we propose CAR (Conceptualization-Augmented Reasoner), a simple yet effective zero-shot commonsense QA framework that leverages conceptualization to expand existing CSKBs and reduce false-negative distractors. We first augment the original CSKB with conceptualization to infuse abstract commonsense knowledge to improve knowledge coverage. Then, we propose a conceptualization-constraint sampling strategy that generates distractors with conceptlevel constraints to prevent false negative options (Section 4). Experimental results on five popular commonsense QA benchmarks demonstrate the effectiveness of CAR, which even surpasses GPT3.5 and ChatGPT (Section 5). In Section 6, we analyze why CAR works by providing human evaluations that show a significant reduction of false negative options compared to other methods. Finally, our analysis reveals that conceptualizationaugmented training examples tend to be more ambiguous (Swayamdipta et al., 2020) than those produced by prior heuristics, leading to better out-ofdomain generalization.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Works\",\n \"publication_ref\": [\n \"b73\",\n \"b65\",\n \"b17\",\n \"b5\",\n \"b1\",\n \"b42\",\n \"b69\",\n \"b31\",\n \"b64\",\n \"b23\",\n \"b66\",\n \"b67\",\n \"b49\",\n \"b72\",\n \"b18\",\n \"b22\",\n \"b38\",\n \"b54\",\n \"b24\",\n \"b6\",\n \"b13\",\n \"b71\",\n \"b50\",\n \"b80\",\n \"b80\",\n \"b77\",\n \"b51\",\n \"b63\",\n \"b81\",\n \"b29\",\n \"b21\",\n \"b43\"\n ],\n \"table_ref\": [],\n \"text\": \"Zero-shot Commonsense QA. Zero-shot commonsense QA evaluates a model's reasoning generalizability on unseen QA entries without any supervision signals from the corresponding annotated training data. To tackle this task, two primary pipelines have emerged in existing works. The first paradigm employs off-the-shelf language models without changing the parameters, either using vanilla language modeling with prompts (Trinh and Le, 2018;Li et al., 2022), or with some inferencetime mechanisms specifically designed for reasoning, such as self-talk (Shwartz et al., 2020), cloze translation (Dou and Peng, 2022), and dynamic generation of reasoning sub-graphs and graph reasoning (Bosselut et al., 2021). The second pipeline leverages external CSKBs as knowledge sources to provide PLMs with additional supervision signals for further fine-tuning (Banerjee and Baral, 2020;Ma et al., 2021;Su et al., 2022). A common strategy involves converting knowledge triples in CSKBs to synthetic QA pairs by transforming the head and relation to a question, the tail to a gold answer, and (randomly) sample tails from other heads as distractors. Such fine-tuning paradigm benefits from incorporating CSKBs within different domains (Kim et al., 2022;Shi et al., 2023) and exploiting multi-hop graph structures with graph neural networks (Guan et al., 2023), and heightens the model's commonsense sensitivity in a QA context, which leads to state-of-the-art performances.\\nConceptualization. Conceptualization refers to the process of abstracting a group of instances or events into a general concept (Song et al., 2011(Song et al., , 2015)). In commonsense reasoning, it simulates conceptual induction (Murphy, 2004) and enables the derivation of abstract commonsense knowledge under the specific contextualization of the original commonsense knowledge (Tenenbaum et al., 2011), which is often lacking in existing CSKBs. Around many existing works studying conceptualization (Durme et al., 2009;Gong et al., 2016;Liu et al., 2022;Peng et al., 2022), He et al. (2022) investigate it at event-level semantics and construct AbstractATOMIC, an event conceptualization benchmark and knowledge base based on ATOMIC (Sap et al., 2019a). Recently, Wang et al. (2023a) propose to conceptualize CSKBs at scale with semi-supervised learning and demonstrate abstract knowledge can enhance commonsense inference modeling (Bosselut et al., 2019;Da et al., 2021). With current works mostly investigating the problem of conceptualization itself, none of them have extrinsically evaluated the impact of conceptualization on downstream tasks, such as commonsense QA (Talmor et al., 2019) or machine reading comprehension (Nguyen et al., 2016). Data Augmentation. Data augmentation aims at generating new examples from existing data to expand the size and diversity of a training set without requiring costly data annotations (Wei and Zou, 2019). Various methods have been proposed to augment textual data, including those using random perturbation (Wei and Zou, 2019), text embeddings (Wang and Yang, 2015), lexical semantics (Niu and Bansal, 2018), back translation (Sennrich et al., 2016), and large language models (West et al., 2022;Ismayilzada and Bosselut, 2023;Gao et al., 2023) for CSKB construction. Nevertheless, text-perturbation-based augmentations do not provide new knowledge to CSKBs, and knowledge mining from large language models suffers from high typicality (e.g., favoring simple commonsense over informative yet rare commonsense) and low density, still making negative sampling subject to false negatives (Malaviya et al., 2020).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Problem Definition\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Definitions\",\n \"publication_ref\": [\n \"b24\",\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"Conceptualization. Formally, denote a CSKB as D with knowledge triples in the format of\\nD = {(h, r, t)|h ∈ H, r ∈ R, t ∈ T }\\n, where H, R, and T are the sets of heads, relations, and tails in the original CSKB. Following He et al. (2022), the conceptualized CSKB, conditioned on D, can be denoted as\\nD C = {(h c , r, t)|h c ∈ H c , r ∈ R, t ∈ T }\\n, where H c is the set of conceptualized head events. Specifically, each conceptualized head h c is obtained by replacing a component i ∈ h with its abstract concept c while ensuring that the formed (h c , r, t) triple is still plausible in the original context (r, t). Such (h c , r, t) triples are commonly referred to as abstract commonsense knowledge.\\nZero-shot Commonsense QA. In this paper, we employ the zero-shot commonsense QA task proposed by Ma et al. (2021) to study our framework. First, the CSKB D is transformed into multiple (Q i , A i ) pairs where Q i is a natural langauge question and A i = {A i,1 , A i,2 , ..., A i,m } is a set of options with m candidates. Specifically, for a given knowledge triple (h, r, t) ∈ D, we convert h, r into Q i via natural language templates and use t as the ground answer. Additionally, we retrieve m -1 distractors from other triples sampled from D using a manually defined strategy, such as keyword overlap filtering. The objective of our task is to train a QA model from the synthetic QA sets\\nD Q = {(Q i , A i )|(h i , r i , t i ) ∈ D}.\\nOnce trained, the model is tested on held-out test entries (Q test , A test ) from QA benchmarks. This requires the model to perform zero-shot commonsense reasoning since the training data from the target benchmarks are unavailable to the model.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Dataset\",\n \"publication_ref\": [\n \"b24\"\n ],\n \"table_ref\": [],\n \"text\": \"We use ATOMIC (Sap et al., 2019b) as the source CSKB D. ATOMIC contains inferential commonsense knowledge, in the format of (h, r, t) triple, that is associated with commonly seen events. Specifically, the heads of ATOMIC triples are events, whereas the tail nodes are either events or attributes. For conceptualization, we use the human-annotated abstract knowledge from Abstrac-tATOMIC (He et al., 2022) to train a generative conceptualizer for acquiring D C . More details of conceptualizations and statistics of AbstractATOMIC are provided in Section 4.1 and Appendix B.1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Evaluation Benchmarks\",\n \"publication_ref\": [\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"Following Ma et al. (2021) \"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"CAR Framework\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This section introduces our proposed CAR framework. A general sketch is presented in Figure 3. Our framework can be summarized into three steps:\\n(1) Conduct one-step conceptualization inference on existing triples in the CSKB to obtain abstract commonsense knowledge triples.\\n(2) Transfer the triples into QA pairs and generate distractors using keywords and conceptualizations as constraints.\\n(3) Train the QA model using marginal ranking loss.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conceptualization Augmentation\",\n \"publication_ref\": [\n \"b83\",\n \"b47\"\n ],\n \"table_ref\": [],\n \"text\": \"To incorporate abstract knowledge into the CSKB, we begin by augmenting the (h, r, t) ∈ D triples by conducting a one-step conceptualization inference. Initially, given a head event h, we retrieve all plausible conceptualizations C h = {c i 1 ,1 , c i 1 ,2 , ...} for all identified instances i ∈ {i 1 , i 2 , ...|i ∈ h} using entity-linking heuristics to retrieve concepts from Probase (Wu et al., 2012) and WordNet (Miller, 1995). The conceptualized head event h c is then obtained by replacing an i ∈ h with one of its retrieved conceptualization c ∈ {c i,1 , c i,2 , ...}. This is done for all identified instances and their retrieved conceptualizations, thereby constructing the set of conceptualized head events of h. Subsequently, we link the non-abstract counterpart (r, t) after h c to generate candidate abstract knowledge triples (h c , r, t), where we adopt a discriminator trained with a semi-supervised conceptualizationinstantiation framework to determine their plausibility (Wang et al., 2023a). Only plausible triples are kept to form D C . Details about the conceptualization retrieval processes and the discriminator are presented in Appendix B.1.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\",\n \"fig_3\"\n ],\n \"heading\": \"Concept-Constrained QA Synthesis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To synthesize a commonsense triple (h, r, t) into a (Q i , A i ) pair, we first transfer h, r into Q i by using natural language templates and set t as the groundtruth answer A 1 . For example, the triple in Figure 3 becomes \\nC h = {c i 1 ,1 , c i 1 ,2 , • • • , c i 2 ,1 , • • • }, we as- sociate a triple (h, r, t) with T h + C h to form its constraint. Only knowledge triple (h ′ , r, t ′ ) which satisfies (T h ′ + C h ′ ) ∩ (T h + C h ) =\\n∅ can be sampled as a distractor candidate. This constraint requires that the two triples have no common keywords, and their instances cannot be abstracted into the same conceptualization. For example, in Figure 3, \\\"(PersonX is at the casino, xWant, have a drink)\\\" cannot be used as a distractor triple because \\\"casino\\\" can be conceptualized as \\\"entertainment place,\\\" which is the same as \\\"bar\\\" in the original triple. Finally, we sample two distractor triples for the triple (h, r, t) and use the tails of these two triples as the distractors. To guarantee that the abstract commonsense knowledge from our previous augmentation is learnable by the QA model, we synthesize both the original triple (h, r, t) and its conceptualized versions (h c , r, t) into QA pairs.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Model Training\",\n \"publication_ref\": [\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"Following Ma et al. (2021), we train our QA model by fine-tuning a pre-trained Masked Language Model (MLM) using the Marginal Ranking (MR) loss. Let C represent the original context (if any), Q represent the question, and (A 1 , A 2 , ...) be the list of options. We first concatenate C, Q, and an answer option A i together via natural language prompts to generate input sequences (T 1 , T 2 , ...).\\nFor example, the synthesized question with its correct answer in Figure 3 will be transformed as:\\n\\\"PersonX arrives at the bar, as a result, PersonX want to, relax himself.\\\" We then repeatedly mask out a token at one time and calculate the masked loss. The final MLM score for an input sequence T ∈ {T 1 , T 2 , ...} with n tokens is:\\nS(T ) = - 1 n n i=1 log P (t i |..., t i-1 , t i+1 , ...) (1)\\nAfter calculating the scores S 1 , S 2 , ... for all answer candidates A 1 , A 2 , ..., we compute the marginal ranking loss based on Equation 2, where η represents the margin and y is the index of the correct answer.\\nL = 1 m m i=1,i̸ =y max(0, η -S y + S i )(2)\\nDuring the evaluation phase, we use the same scoring procedure to assign a score to each option and select the one whose concatenated sentence achieves the lowest score as the model's prediction.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Setup\",\n \"publication_ref\": [\n \"b39\",\n \"b25\",\n \"b65\",\n \"b5\",\n \"b1\",\n \"b69\",\n \"b31\",\n \"b42\",\n \"b42\",\n \"b2\",\n \"b9\",\n \"b56\",\n \"b8\",\n \"b81\",\n \"b48\",\n \"b79\",\n \"b59\"\n ],\n \"table_ref\": [],\n \"text\": \"Baselines First, we use random voting (Random) and most-frequent labeling (Majority) to demonstrate the characteristics of each benchmark. Vanilla RoBERTa-Large (Liu et al., 2019), and DeBERTa-v3-Large (He et al., 2023) PLMs are used to demonstrate the power of fine-tuning. The performances of these two models under a supervised training regime are also included to show the upper bound of our results. We also include the results of several existing approaches that tackle the same task, including Self-talk (Shwartz et al., 2020), COMET-DynaGen (Bosselut et al., 2021), SMLM (Banerjee and Baral, 2020), MICO (Su et al., 2022), and the previous state-of-the-art STL-Adapter (Kim et al., 2022). Most importantly, we compare our framework with Ma et al. (2021) to validate the efficacy of conceptualization since both methods share similar model architecture and training procedures. Both RoBERTa-Large and DeBERTa-v3-Large are used as the backbones for fair comparisons. There are, in total, 534,833 synthetic QA pairs provided by Ma et al. (2021).\\nWith the recent advances in Large Langauge Models (LLMs) (Bang et al., 2023;Chan et al., 2023;Qin et al., 2023), we also benchmark the performances of GPT3.5 (Brown et al., 2020 2021) on ATOMIC (marked with △). ATM C stands for the ATOMIC with abstract commonsense knowledge injected. ATM-10X stands for using ATOMIC-10X (West et al., 2022) as the source CSKB D. All baseline results are consistent with their original papers.\\net al., 2022) and ChatGPT (OpenAI, 2022) as baselines. We prompt the LLM directly in a zero-shot setting, where no in-context learning (Min et al., 2022) or chain-of-thought reasoning (Wei et al., 2022) are applied. For every QA entry, the LLM is presented with a question, several choices, and a natural language command that asks it to choose the index of the correct answer directly (Robinson et al., 2022). We then parse the generated outputs to obtain the \\\"predictions\\\" of LLM by using meticulously designed rules and compare them with the ground-truth labels. More details of the baselines and LLM setups can be found in Appendix B.2 and B.3.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Implementation Details\",\n \"publication_ref\": [\n \"b40\",\n \"b42\",\n \"b31\"\n ],\n \"table_ref\": [],\n \"text\": \"We use accuracy as the evaluation metric and compare our framework with the following baseline methods. For conceptualization, we leverage an off-the-shelf conceptualizer from Wang et al. (2023a), which is a semi-supervised conceptualization discriminator fine-tuned on labeled conceptualization data from AbstractATOMIC and unlabeled data from ATOMIC. We use a plausibility score T = 0.9 to filter out plausible conceptualizations, which results in 440K conceptualization-aided synthetic QA pairs for training. We employ an AdamW optimizer (Loshchilov and Hutter, 2019) with a learning rate of 7e-6 and a max sequence length of 128 to accommodate QA pairs with different lengths. We select the best checkpoint according to the highest accuracy achieved on the synthetic validation QA set. Each experiment is repeated using three different random seeds, and the average performance is reported. The model is warmed up with 5% of total iterations and evaluated every 1000 global steps, while the margin η for the marginal ranking loss is set to 1, in line with the choices made by Ma et al. (2021) and Kim et al. (2022). More details about implementations can be found in Appendix B.4,\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The main results are reported in Table 1. For the baselines, DeBERTa-v3-Large (MR) trained on ATOMIC achieves the best performance, followed by ChatGPT. Both achieve an accuracy of more than 70% on average. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Analysis and Discussion\",\n \"publication_ref\": [\n \"b70\"\n ],\n \"table_ref\": [],\n \"text\": \"In this section, we study the effects of conceptualization and the reasons contributing to CAR's success. First, we conduct expert evaluations on the synthetic QA pairs to study the quality and diversity of different CSKB augmentation methods in comparison with conceptualization. Second, we conduct training dynamics (Swayamdipta et al., 2020) analysis to show that conceptualization-aided QA pairs can provide more ambiguous examples helpful for training. Finally, we study the impact of filtering ATOMIC10X with different critic thresholds, the ablations of CAR, and the effect of conceptualization from an out-of-domain generalization perspective in Appendix B.5, B.7, and B.8.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Comparisons With Data Augmentations\",\n \"publication_ref\": [\n \"b55\",\n \"b47\",\n \"b45\",\n \"b42\",\n \"b42\"\n ],\n \"table_ref\": [\n \"tab_4\",\n \"tab_4\",\n \"tab_4\",\n \"tab_4\"\n ],\n \"text\": \"To demonstrate the effectiveness of our proposed conceptualization method, we conduct comprehensive analyses with other data augmentations that expand the semantic coverage of CSKBs in a similar way as conceptualization using both expert and automatic evaluations. We use EDA, augmenting with word embedding (Word2Vec; Mikolov et al., 2013 and GLOVE;Pennington et al., 2014), contextual embedding (BERT; Devlin et al., 2019), and synonym (WordNet;Miller, 1995) score (McHugh, 2012) of 0.64, comparable to 0.62, as reported by (Ma et al., 2021).\\nDiversity. First, we study whether augmentations can introduce new knowledge to the training set. We begin by calculating the average cosine similarity of each ATOMIC triple and its augmented siblings from each method according to their Sen-tenceBERT (Reimers and Gurevych, 2019) embeddings. For ATOMIC-10X, we regard the sampled knowledge as augmentations. The complement of average similarity across all ATOMIC triples serves as an automatically measured diversity (Div.). Meanwhile, we retrieve top-10 similar triples from ATOMIC for each augmented triple according to their SentenceBERT similarity. The experts are asked to annotate whether each triple can be semantically covered by their retrievals. We define the expert-evaluated diversity as the ratio of uncovered triples among 300 samples. Table 2 shows that conceptualization champions both metrics, indicating that the introduced abstract knowledge is diverse and lacking in existing CSKBs, which is helpful in expanding their knowledge coverage.\\nQuality of Synthetic QA Pairs. Next, we synthesize the augmented triples into QA pairs with their head events' keywords and augmentations as constraints. We then sample 300 QA pairs for each method and ask the same experts to perform expert evaluations by annotating the correctness of each QA pair's ground-truth answer and whether the distractors could also be plausible with respect to the augmented head event. This evaluates the plausibility ratio of the augmented knowledge and the ratio of QA pairs containing false negative distractors. Table 2 shows that the majority of augmented knowledge is implausible, and they fail to enhance distractors sampling. Conceptualization, on the other hand, maintains being highly plausible and can effectively eliminate false negative distractors.\\nExpert annotators also achieve a remarkable ac-curacy of 86% while working on 300 randomly sampled question-answer pairs, surpassing the 80% accuracy reported by Ma et al. (2021).\\nZero-shot Commonsense QA Performances.\\nFinally, we train DeBERTa-v3-Large models on the QA pairs synthesized from the concatenation of both original and augmented ATOMIC triples from each method. Only keywords of each head event are used as their constraints. The models are trained using a marginal ranking loss, as explained in Section 4.3, and evaluated on five QA benchmarks in a zero-shot manner. Performances by different methods are shown in Table 2. We observe that conceptualization outperforms all other augmentations on average and successfully improves the model's zero-shot commonsense reasoning ability.\\nComparison with ATOMIC-10X. Augmenting ATOMIC10X appears to be a promising option as it contains a wealth of valuable commonsense knowledge. However, despite its diverse and highquality knowledge, Table 2 demonstrates that the model cannot leverage this information effectively. One possible explanation is that the model's performance is hindered by the significantly high number of false-negative distractors. This issue arises because the knowledge distilled from GPT-3 tends to be versatile, resulting in many tail events being general and vague. These events can be applied to a large collection of heads, which leads to false negative options. More experiments and case studies are in Appendix B.5 and C, respectively.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\",\n \"fig_4\"\n ],\n \"heading\": \"Training Dynamics Analysis\",\n \"publication_ref\": [\n \"b70\"\n ],\n \"table_ref\": [],\n \"text\": \"Training dynamics effectively assess a model's confidence and variability for individual instances when training on a large dataset. In the context of QA, we define confidence as the model's certainty when assigning the correct label to the ground-truth option compared to distractors, as indicated by the logit difference. Variability, on the other hand, refers to the fluctuation of confidence over time.\\nThese insights can aid in analyzing the model's behavior when different knowledge is introduced into the training set. More explanations are in Appendix B.6.\\nIn this section, we examine the impact of abstract commonsense knowledge (conceptualization) and GPT3-distilled knowledge (ATOMIC-10X) by exploring their training dynamics on two sets of data. We train three QA models on synthetic QA pairs from conceptualization-augmented ATOMIC, ATOMIC10X-augmented ATOMIC, and the original ATOMIC, which serves as the baseline. First, we randomly select the same 1,000 QA pairs synthesized from the original ATOMIC and calculate their training dynamics using these three models. The left side of Figure 4 displays the alterations caused by the two augmentation methods in comparison with the baseline. It is evident that introducing abstract commonsense knowledge through conceptualization significantly reduces the model's average variability and enhances its confidence in learning the knowledge from the original ATOMIC. In contrast, incorporating knowledge from ATOMIC-10X produces the opposite effect.\\nSecond, we check the training dynamics on 1,000 randomly sampled QA pairs synthesized from abstract commonsense knowledge and another 1,000 from knowledge in ATOMIC-10X. The rightmost plots in Figure 4 reveal that, compared to ATOMIC-10X, conceptualization introduces knowledge with higher variability and lower confidence, making it more ambiguous and challenging for the model to learn. As Swayamdipta et al. (2020) suggest, such data contributes to a more robust model to out-of-distribution (OOD) data, which are downstream QA datasets in our case. Therefore, we conclude that conceptualization is superior to ATOMIC-10X as abstract knowledge, on the one hand, makes the original knowledge more easy-to-learn to aid optimization, and on the other hand, provides more ambiguous examples to boost OOD generalization.\\n3URSRUWLRQRI$70 C GDWD $YJ$FFXUDF\\\\ \"\n },\n {\n \"figure_ref\": [\n \"fig_5\"\n ],\n \"heading\": \"Impact of Training Data Size\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In Figure 5, we present the influence of the number of training examples against the final performance, which reveals a clear and intuitive trend of a positive correlation between the amount of training data and overall performance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Generalization to other CSKBs\",\n \"publication_ref\": [\n \"b68\",\n \"b47\",\n \"b74\"\n ],\n \"table_ref\": [\n \"tab_6\"\n ],\n \"text\": \"We explore the feasibility of transferring our framework to CSKBs other than ATOMIC. We take the CWWV dataset as an example, which comprises multiple CSKBs, including ConceptNet (Speer et al., 2017), WordNet (Miller, 1995), and Wiki-Data (Vrandecic and Krötzsch, 2014). We use the off-the-shelf GPT2 conceptualizer (Wang et al., 2023a) and ChatGPT as two flexible generative conceptualizers. The generated conceptualizations are then transformed into abstract knowledge and integrated into the CWWV dataset. The experimental results are presented in Table 3, which shows an improvement of over 1% compared to all baselines leveraging CWWV as the source of knowledge, indicating CAR's generalizability to other CSKBs. More details are presented in the Appendix B.9.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we present CAR, a pioneering framework for zero-shot commonsense QA empowered by conceptualization. Our approach surpasses even large language models on five QA benchmarks, achieving state-of-the-art performance on average. Our analyses reveal that conceptualization can improve the sampling of negative examples, and abstract knowledge is more helpful compared with those distilled from GPT3 as it provides more ambiguous knowledge to support OOD generalization. These findings demonstrate the substantial benefits of introducing conceptualization and abstract knowledge into zero-shot commonsense reasoning.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [\n \"b42\",\n \"b31\",\n \"b17\",\n \"b68\",\n \"b47\",\n \"b33\",\n \"b74\",\n \"b24\",\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"One limitation of this paper is that the proposed CAR framework has only been validated on the ATOMIC dataset. While previous works (Ma et al., 2021;Kim et al., 2022;Dou and Peng, 2022) have studied the zero-shot commonsense question answering task by consolidating multiple CSKBs, including ATOMIC (Sap et al., 2019a), Concept-Net (Speer et al., 2017), WordNet (Miller, 1995), VisualGenome (Krishna et al., 2017), and Wiki-Data (Vrandecic and Krötzsch, 2014), our work only utilizes ATOMIC (more details discussed in Appendix B.2). This was mainly due to the availability of conceptualizations for the CSKB, with only AbstractATOMIC (He et al., 2022) being available as the conceptualized expansion of ATOMIC, while other CSKBs lack such resources. Additionally, ATOMIC has been shown to play the most critical role in experimental results by Ma et al. (2021). Nonetheless, such limitation does not restrict CAR's potential to seek further improvements from incorporating other CSKBs, as conceptualization frameworks, such as CAT (Wang et al., 2023a), can be applied to other CSKBs and provide the required resources for CAR to operate. Thus, we believe CAR can overcome such limitations and still possess the potential to improve with more CSKBassociated conceptualization resources available.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ethics Statement\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This paper presents CAR, a novel framework for zero-shot commonsense question answering that achieves state-of-the-art performance via conceptualization. All datasets used, including ATOMIC, AbstractATOMIC, and commonsense questionanswering benchmarks, are publicly available and shared via open-access licenses solely for research purposes, consistent with their intended usage. These datasets are anonymized and desensitized, ensuring that no data privacy issues are involved. Moreover, the CAR framework is a questionanswering system that selects the most plausible choice from a list of options and does not yield any private, offensive, biased, or sensitive information or social and political issues. The expert annotations are performed by the authors of this paper as part of their contribution, who are graduate and undergraduate students working on machine commonsense in natural language processing, and they are fully aware of the annotation protocol and the intended use of their annotations. They are well-trained with specially designed instructions and have voluntarily agreed to participate without compensation. Based on this, the authors believe that this paper does not raise any ethical concerns to the best of their knowledge.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Appendices A Benchmark Descriptions\",\n \"publication_ref\": [\n \"b3\",\n \"b71\",\n \"b68\",\n \"b4\",\n \"b60\"\n ],\n \"table_ref\": [\n \"tab_7\"\n ],\n \"text\": \"In this section, we introduce more details regarding five evaluation benchmarks.\\nAbductive NLI (aNLI) (Bhagavatula et al., 2020) is a Natural Langauge Inference (NLI) benchmark that aims to infer the most plausible explanation based on a given causal situation. For each question sample, the model is required to choose the more plausible hypothesis out of two options that fit the beginning and end of a story. This benchmark evaluates the model's abductive reasoning ability, which typically requires commonsense reasoning.\\nCommonsenseQA (CSQA) (Talmor et al., 2019) is a question-answering benchmark that evaluates a broad range of commonsense aspects. Each sample contains a question and five choices. The question and some choices are generated from subgraphs of ConceptNet (Speer et al., 2017) while crowdsourcing annotators also annotate some distractors. This benchmark evaluates the model's concept-level commonsense reasoning ability.\\nPhysicalIQA (PIQA) (Bisk et al., 2020) is a question-answering benchmark that requires the model to select the more plausible option out of two possible continuations given a common scenario that requires physical commonsense to infer. This benchmark evaluates the model's physical commonsense reasoning ability.\\nSocialIQA (SIQA) (Sap et al., 2019b) is a question-answering benchmark that requires rea- soning about social interactions. Each sample contains a context that is derived from ATOMIC (Sap et al., 2019a), a question, and three choices. The questions are automatically generated using nine templates that correspond to the nine relations in ATOMIC, and the correct answers are crowdsourced. This benchmark evaluates the model's reasoning ability for emotional and social commonsense in daily situations. WinoGrande (WG) (Sakaguchi et al., 2021) is a pronoun resolution benchmark. Each sample contains an emphasized pronoun and a short context description. The model is asked to choose the correct reference given two options. This benchmark evaluates the model's pronoun resolution ability, which is also part of commonsense knowledge.\\nIn our experiments, we use the validation splits of these benchmarks as the official testing sets may not be publicly available. Detailed statistics on the number of QA pairs and the number of options per question are reported in Table 4.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B Additional Explanations and Analyses\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we aim to cover additional details regarding the CSKB conceptualization in CAR (Appendix B.1), implementations of our system (Appendix B.4), baselines (Appendix B.2 and B.3), experiments using ATOMIC-10X (Appendix B.5), analyses (Appendix B.6, B.7, and B.8), and generalizability experiments (Appendix B.9) that are not covered in the body text due to space constraints.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.1 Definitions and Statistics of CSKB Conceptualization\",\n \"publication_ref\": [\n \"b84\",\n \"b85\",\n \"b24\",\n \"b24\",\n \"b47\",\n \"b26\",\n \"b57\"\n ],\n \"table_ref\": [\n \"tab_9\"\n ],\n \"text\": \"Conceptualization plays a crucial role in generalizable commonsense reasoning. Previous studies have demonstrated its potential in aiding commonsense inference modeling (Wang et al., 2023a) and commonsense knowledge graph construction (Yu et al., 2023;Zhang et al., 2022). In our paper, we follow the definition of conceptualization proposed by He et al. (2022) and Wang et al. (2023a) in conceptualizing an instance within an event to a concept: (1) Events: Each event represents a commonly observed occurrence that encompasses valuable subsequential or inferential commonsense knowledge. In AbstractATOMIC, the events are the head events of all triples in ATOMIC without a wildcard ('_').\\n(2) Instances: Within each event, multiple instances have been identified with semantic parsing tools, representing specific components of the event that are worthy of conceptualization.\\n(3) Concepts: Concepts are the conceptualization of each instance. These concepts are thus extracted from Probase/WordNet and further validated by human annotators or critic filtering models. For an event e, which is the head of an ATOMIC triple, an instance refers to either an entity within the event or the complete event itself. Multiple instances can exist within a single event, denoted as i 1 , i 2 , i 3 , . . . , i n ∈ e. A concept corresponds to the conceptualization of an instance, and multiple conceptualizations can be associated with a single instance, as demonstrated by (i 1 , c 1 , 1), (i 1 , c 1 , 2), (i 1 , c 1 , 3), ..., (i 2 , c 2 , 1), ..., (i n , c n , 1), . . . , (i n , c n , m). For instance, consider the event \\\"PersonX is drunk when exiting the bar.\\\" In this case, two instances can be identified: \\\"Per-sonX is drunk when exiting the bar\\\" and \\\"bar.\\\" The conceptualization for the instance \\\"PersonX is drunk when exiting the bar\\\" may include \\\"drunk\\\" or \\\"enjoyed,\\\" while the instance \\\"bar\\\" can be conceptualized as an \\\"entertainment place\\\" or a \\\"fun place.\\\"\\nIn this paper, we leverage the AbstractATOMIC dataset, provided by He et al. (2022), as our primary source for conceptualizations. Abstrac-tATOMIC is a benchmark for conceptualized commonsense knowledge that is built upon the ATOMIC dataset (Sap et al., 2019a). It contains three folds of data, each conditioned on the original commonsense knowledge triples (h, r, t) in ATOMIC.\\nIn the first fold, He et al. ( 2022) identify all possible instances {i 1 , i 2 , i 3 , • • • |i ⊆ h} in each ATOMIC head event, using syntactic parsing through a spaCy 1 parser and matching with five human-defined rules. It is important to note that, unlike traditional entity-level conceptualization benchmarks, the identified instance in Abstrac-tATOMIC can also be the entire head event i = h.\\nIn the next fold, each identified instance i is heuristically matched against Probase (Wu et al., 1 https://spacy.io/ 2012) and WordNet (Miller, 1995) via Gloss-BERT (Huang et al., 2019) to find their corresponding conceptualization candidates. Human annotations are conducted to verify part of the plausibility of the conceptualization candidates. To pseudolabel unannotated conceptualizations, we use a semi-supervised conceptualization discriminator provided by Wang et al. (2023a) and set a threshold of T = 0.9 to filter out plausible conceptualizations. Additionally, we utilize a GPT2-based (Radford et al., 2019) generator, trained on the concatenation of annotated and positively pseudo-labeled conceptualizations, to generate additional conceptualizations for further expanding the size of the conceptualization bank.\\nHowever, it is worth noting that such conceptualization may not yield plausible abstract knowledge when (r, t) is connected back to h c , where h c is obtained by replacing i ∈ h with its conceptualizations. This is because the process of conceptualizing a head event omits its context in (r, t). Thus, the last fold of data stores the plausibility of such abstract commonsense triples (h c , r, t), where human annotations are conducted to verify part of the triples' plausibilities. In addition, we adopt a semi-supervised instantiation discriminator, provided by Wang et al. (2023a), to pseudo-label the unannotated triples. Another threshold, T = 0.9, is set to filter out plausible abstract triples.\\nIn the CAR framework, for every ATOMIC event h, we retrieve every instance i's plausible conceptualizations {c i,1 , c i,2 , • • • } from all plausible conceptualizations derived in the second fold to serve as the distractor sampling constraint. We also augment the original (h, r, t) triples with their plausible (h c , r, t) siblings from both human-annotated and pseudo-labeled triples, as explained in the last fold. These knowledge triples are then synthesized into QA pairs using our proposed method to train the model to perform general reasoning. Detailed statistics for the conceptualizations and abstract commonsense triples we finally obtained from the AbstractATOMIC dataset are reported in Table 5 and Table 6, respectively.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.2 Baseline Performances\",\n \"publication_ref\": [\n \"b1\",\n \"b1\",\n \"b31\",\n \"b68\",\n \"b47\",\n \"b74\"\n ],\n \"table_ref\": [],\n \"text\": \"For SMLM (Banerjee and Baral, 2020), we adopt the official implementation of Banerjee and Baral (2020), which employs the CSKB that exhibits the highest alignment with each task. Specifically, So-cialIQA uses ATOMIC, while CommonsenseQA uses ConceptNet. For STL-Adapter (Kim et al., 2022), only those trained on ATOMIC are used for comparison in the body text. In this paper, all baseline performances are solely based on their officially reported results in their respective papers.\\nAs noted in the Limitations section, previous research in this area has primarily focused on using four CSKBs, namely ATOMIC (Sap et al., 2019a), ConceptNet (Speer et al., 2017), WordNet (Miller, 1995), and WikiData (Vrandecic and Krötzsch, 2014). In order to comprehensively benchmark our framework's performance in the field of zeroshot commonsense QA, we compare our results on ATOMIC against baseline methods that use multiple CSKBs despite the unbalanced amount of knowledge in such a comparison. Table 11 presents a full comparison of our method with all existing baselines. Notably, for models based on RoBERTa-Large, our approach trained only on abstract knowledge injected ATOMIC achieves second place in the leaderboard, falling only behind Kim et al. ( 2022) with four CSKBs. While this comparison may be unfair due to the unbalanced amount of knowledge, it provides a strong justification for the excellent performance of our system. Our DeBERTa-v3-Large-based model still surpasses all baselines on average, indicating the necessity of leveraging a strong pre-trained language model.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.3 Benchmarking Large Language Models\",\n \"publication_ref\": [\n \"b56\",\n \"b9\",\n \"b71\",\n \"b60\",\n \"b59\",\n \"b71\",\n \"b0\",\n \"b16\",\n \"b79\"\n ],\n \"table_ref\": [],\n \"text\": \"We then discuss our method for benchmarking large language models on five commonsense QA benchmarks. The emergence of Large Language Models (LLMs), such as ChatGPT (OpenAI, 2022), has been the hot trend in recent NLP research. Numerous studies have evaluated the capability of LLMs on various NLP downstream tasks. Among them, Qin et al. (2023); Chan et al. (2023) have shown that ChatGPT can achieve competitive performance on commonsense reasoning tasks, such as CommonsenseQA (Talmor et al., 2019), Wino-Grande (Sakaguchi et al., 2021), and Commonsense Knowledge Base Population (Fang et al., 2021b,a). In this study, we aim to benchmark Chat-GPT's zero-shot performance on five QA evaluation benchmarks used in our zero-shot commonsense QA task. Following (Robinson et al., 2022), we design and leverage a batch of prompts, as shown in Table 7, to probe ChatGPT's predictions. The prompt provides ChatGPT with a question and its possible choices, along with a natural language command to control the response action of ChatGPT. We then parse the generations using our meticulously designed rules, where punctuations and irrelevant wordings will be dropped, and the first choice-letter prediction will be identified as ChatGPT's answer. Specifically, if ChatGPT hesitates and cannot make a concrete prediction, it will be counted as a wrong answer. The benchmarking results are shown in Table 1. We observe that ChatGPT demonstrates superior performance compared to GPT3.5 and excels in tasks such as CommonsenseQA (Talmor et al., 2019) and So-cialIQA (Sap et al., 2019b), potentially due to the high frequency of their questions and answers in large text corpora. However, its performance on the remaining three benchmarks is suboptimal, suggesting that they are more challenging and require more complex reasoning (Bai et al., 2023;Ding et al., 2023) and implicit commonsense knowledge to solve. This intriguing outcome warrants further investigation to determine the reasons behind it and explore methods to boost the LLM's abilities in these challenging benchmarks.\\nGenerally speaking, CAR and conceptualization own the advantage over the large language model in the following aspects: (1) Smaller Model Size: The CAR framework offers models that are significantly smaller in scale compared to LLMs (0.2% of a standard 175 billion parameter GPT-3 model) while maintaining comparable performance in a zero-shot setting. Such size makes it more efficient in terms of training and deployment. In contrast, advanced prompting techniques used in LLMs require extensive computational resources, making the conceptualization-based model more versatile and accessible to researchers with limited access or resources for deploying LLMs. (2) Broader Commonsense Knowledge: Conceptualization provides a broader range of commonsense knowledge compared to current CSKBs. Integrating this type of knowledge into generative models has been shown to enhance their performance in commonsense reasoning tasks (Wang et al., 2023a). Such knowledge can also be dynamically encoded in language models during inference time (Chen et al., 2023b).\\n(3) Advanced Prompting of LLMs: Conceptualization data introduces the potential for a more advanced prompting framework of reasoning with LLMs. The process of conceptualization and the instantiation of knowledge play a crucial role in reasoning. Thus, future work may consider introducing the \\\"chain of concept\\\" reasoning process to further advance the popular \\\"chain of thought\\\" reasoning paradigm (Wei et al., 2022;Wang et al., 2023b).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.4 Implementation Details\",\n \"publication_ref\": [\n \"b82\",\n \"b42\",\n \"b40\",\n \"b81\",\n \"b42\",\n \"b31\"\n ],\n \"table_ref\": [],\n \"text\": \"We present additional implementation details for building our system. For the pre-trained language models, We use PLMs from the Huggingface Transformers2 Library (Wolf et al., 2020) as the vanilla model checkpoints. Our system relies heavily on the open-source code repository3 provided by Ma et al. (2021) for synthesizing QA pairs and training the QA models. To optimize the model, we employ an AdamW optimizer (Loshchilov and Hutter, 2019) with a learning rate of 7e-6 and a max sequence length of 128 to accommodate QA pairs with different lengths. When evaluating the model on downstream commonsense QA benchmarks, a maximum sequence length of 80 is used. We select the best checkpoint according to the highest accuracy achieved on the synthetic validation QA set. Each experiment is repeated using different random seeds three times, and the average performance is Table 8: Zero-shot evaluation results (%) on five commonsense question answering benchmarks using different critic thresholds for filtering ATOMIC-10X. The best results are bold-faced, and the second-best ones are underlined. ATM C stands for the ATOMIC with abstract commonsense knowledge injected. ATM-10X stands for using ATOMIC-10X (West et al., 2022) as the source CSKB D. ATM ATM-10X indicates the ATOMIC with sampled knowledge from ATOMIC-10X injected. Critic indicates the lower bound for filtering knowledge from ATOMIC-10X, which means that only knowledge with a critic score above the threshold will be selected. reported. To overcome the limited GPU memory issue, we utilize gradient accumulation with a gradient accumulated four steps before every descent, and each step calculates the gradient of eight data samples. The model is warmed up with 5% of total iterations and evaluated every 1000 global steps, while the margin η for the marginal ranking loss is set to 1, in line with the choices made by Ma et al. (2021) and Kim et al. (2022). The Huggingface model code for our RoBERTa-Large model is roberta-large, and the one for our DeBERTa-v3-Large model is microsoft/deberta-v3-large. All of our experiments are conducted on eight NVIDIA A100 GPUs, each with 40G graphical memory. Training a RoBERTa-based model typically requires 14G graphical memory, while training DeBERTa-based models requires 30G graphical memory.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.5 Experiments with ATOMIC-10X\",\n \"publication_ref\": [\n \"b81\",\n \"b27\",\n \"b42\",\n \"b42\",\n \"b14\"\n ],\n \"table_ref\": [\n \"tab_4\"\n ],\n \"text\": \"ATOMIC-10X is a machine-generated corpus developed by West et al. (2022) using the symbolic knowledge distillation framework. They employed a selective distillation approach to extract knowledge from large language models like GPT-3 (Brown et al., 2020) by prompting them with head events and commonsense relations from the ATOMIC dataset. The extracted knowledge was used to train a student model to generate symbolic knowledge graphs evaluated using a separate critic model. The resulting corpus, ATOMIC-10X, surpassed the human-generated corpus ATOMIC2020 (Hwang et al., 2021) in scale, accuracy, and diversity.\\nIn this section, we provide additional explanations regarding the usage of ATOMIC-10X in our paper and conduct further experiments to explore its impact on zero-shot commonsense QA. Specifically, we study the role of filtering the knowledge in ATOMIC-10X using multiple critic thresholds to acquire high-quality knowledge and improve model performance.\\nWe utilize ATOMIC-10X in two distinct scenarios. First, as discussed in Section 5.2, we directly train our QA models on ATOMIC-10X without integrating other CSKBs, such as ATOMIC and AbstractATOMIC. We source all questions, answers, and distractors from ATOMIC-10X and follow its original train/dev/test partitions to divide the data. The lemmatized tokens of the head event, excluding commonly seen subjects, prepositions, and stopwords, are used as keywords for each piece of knowledge, and the original QA synthesis pipeline proposed by Ma et al. (2021) is applied. To ensure the quality of the knowledge from ATOMIC-10X, we set multiple critic thresholds to filter the dataset accordingly. The QA models are trained using marginal ranking loss on four subsets of ATOMIC-10X, each with a different critic threshold of 0.9, 0.8, 0.7, and 0.5, along with an additional trained on the complete ATOMIC-10X dataset. Finally, we evaluate these models on five commonsense QA benchmarks in a zero-shot setting and report the results in Table 8. Specifically, models trained solely on ATOMIC-10X using critic thresholds of 0.7 (RoBERTa) and 0.0 (DeBERTa) for filtering are responsible for the results reported in Table 1. We observe that even using high critic thresholds to filter the knowledge in ATOMIC-10X, the model still fails to improve beyond marginal. Meanwhile, training the models only on ATOMIC-10X fails to surpass training on ATOMIC, which indicates that the amount of knowledge is not the critical element to determining the performance. Rather, it should be the diversity and quality of knowledge, where the humanannotated knowledge from ATOMIC is superior to those machine-generated ones from ATOMIC. Nonetheless, none of the models outperform those trained on conceptualization-augmented ATOMIC using our CAR framework, which further validates the strengths of CAR.\\nIn the second scenario, as discussed in Section 6.1, we utilize ATOMIC-10X as a means of augmentation to extend the original ATOMIC dataset. This is achieved by randomly selecting a specific number of knowledge triples from ATOMIC-10X, equivalent to the total number of plausible abstract commonsense knowledge in AbstractATOMIC, and merging them back into the original dataset. The triples in the resulting ATOMIC10X-augmented ATOMIC are then transformed into QA pairs and used to train our model following the original pipeline suggested by Ma et al. (2021). Similar to the previous scenario, we set four thresholds, namely 0.9, 0.8, 0.7, and 0.5, to filter the triples in ATOMIC-10X for augmentation quality control. In this way, the QA pairs' distractors can come from ATOMIC and ATOMIC-10X. The models are then trained and evaluated on five benchmarks. Their zero-shot commonsense QA evaluation results are reported in Table 8, and the best model, trained using a critic threshold of 0.8 for filtering with DeBERTa-v3-large as the backbone, is responsible for the results indicated in Table 2. Interestingly, we observe that leveraging the knowledge in ATOMIC-10X, either for direct training or augmentation, occasionally improves the model's performance on a specific benchmark. However, it fails to boost the overall performance across all benchmarks on average, which is considered a closer metric for evaluating the generalizable reasoning ability of a commonsense QA model. Thus, we come to the conclusion that ATOMIC-10X is inconsistently helpful in improving the zero-shot commonsense QA performances, with most times failing to improve, while conceptualization resolves such issues and can benefit the model across all benchmarks significantly. One potential reason is that ATOMIC-10X main contain noise that are not benefitial to the task of zero-shot commonsense QA, as demonstrated by Deng et al. (2023).\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"B.6 Training Dynamic Definitions\",\n \"publication_ref\": [\n \"b70\",\n \"b70\",\n \"b70\",\n \"b70\"\n ],\n \"table_ref\": [],\n \"text\": \"Training dynamic, as proposed by Swayamdipta et al. (2020), refers to the analysis of a model's behavior on individual instances during training on large datasets. This analysis examines the model's confidence in predicting the true class of an instance and the variability of this confidence across epochs. To achieve this, multiple checkpoints are saved throughout a training epoch, and probability scores are derived for each data instance to calculate their training dynamics. By plotting the training dynamics of all instances on a data map, instances can be categorized into three groups: easyto-learn, ambiguous, and hard-to-learn. For instance, consider a QA pair where a model consis-tently assigns a higher logit score to the correct answer than to other distractors across multiple checkpoints during an epoch. In this scenario, the model exhibits high confidence and low variability for that specific instance, suggesting that it is easy to learn. Conversely, instances with higher variability are ambiguous to the model, and those with low confidence are difficult to learn. Experimental results by Swayamdipta et al. (2020) demonstrates that training the model with ambiguous data contributes the most to out-of-distribution generalization.\\nInspired by this finding, our research investigates the role of abstract commonsense knowledge within the training set and the effects of leveraging conceptualization. Since our QA model is trained with a marginal ranking loss, as described in Section 4.3, it does not output a probability score but rather an MLM score for each option. Thus, the definition of model's confidence proposed by Swayamdipta et al. (2020) does not fit into our problem definition. To address this, we re-define the calculation of confidence to align with the model's degree of certainty in predicting an instance as the true class. Formally, denote n as the number of saved checkpoints during an epoch for computing their training dynamics and the list of m options in a (Q i , A i ) pair as A i = {A i,1 , A i,2 , ..., A i,m } with A i,j being the ground-truth answer (1 ≤ j ≤ m)). We define the confidence of the model for such a QA pair in Equation 3, where σ is the sigmoid function and S c i,d is the score of option A i,d at checkpoint c.\\nC(Q i , A i ) = 1 n n c=1 σ( m d=1 (S c i,d -S c i,j ) m -1 )(3)\\nIntuitively, this equation averages the gap between the ground-truth answer's score and the score of each distractor. A larger gap indicates a more confident model when choosing the answer. Variability aligns with the definition established by Swayamdipta et al. (2020). Specifically, it is calculated as the standard deviation of the score gap between the ground-truth answer and the distractors relative to the level of confidence exhibited throughout an entire epoch, as shown in Equation 4. By revisiting the plots in Figure 4, we observe that the inclusion of abstract commonsense knowledge enhances the model's confidence and reduces variability when encountering knowledge in ATOMIC. The introduction of conceptualization appears to widen the differences between the model's predicted scores for the correct answer and those for the distractors. This suggests that the correct answer is more likely to be selected, leading to an improved learning outcome. However, the introduction of knowledge from ATOMIC-10X results in a reversed trend, indicating that it does not aid in better learning ATOMIC. Furthermore, we observe that abstract knowledge derived from conceptualizations is more ambiguous to the model in the conceptualization-augmented ATOMIC, which theoretically contributes more to out-of-domain generalization. Nonetheless, ATOMIC-10X still contains some easy-to-learn knowledge that does not facilitate the model's generalization. Thus, abstract commonsense knowledge benefits zero-shot commonsense QA better than ATOMIC-10X by providing more ambiguous conceptual knowledge, which aids in making the model more generalizable.\\nV(Qi, Ai) = n c=1 (σ( m d=1 (S c i,d -S c i,j ) m-1 ) -C(Qi, Ai)) 2 n (4)\\nWe also plot the changes in training dynamics on different QA benchmarks, comparing models with and without the injection of abstract knowledge. The plots are shown in Figure 7. We observe that the inclusion of abstract commonsense knowledge significantly improves the models' confidence in downstream QA entries. However, the impact on the trend of variability is unclear. Nevertheless, this improvement in average confidence provides strong evidence for the model's enhancement in these downstream QA benchmarks.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.7 Ablation Study\",\n \"publication_ref\": [\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"Next, we study the ablation of different components in our CAR framework to determine the impact of utilizing conceptualization through various techniques. There are two critical components that distinguish CAR from traditional zero-shot QA systems (Ma et al., 2021):\\n• Conceptualization Augmentation: Augmenting the original commonsense knowledge in a CSKB with its conceptualizations to derive abstract commonsense knowledge. This knowledge is then synthesized into QA pairs, enabling the model to reason from a more generalized perspective. Without this component, abstract commonsense knowledge is not incorporated into the CSKB. Conceptualizations still remain as constraints for assisting QA pair synthesis, resulting in an approach that is similar to applying our proposed QA synthesis protocol directly to ATOMIC.\\n• Concept-Constrained QA Synthesis: Constraining a question's distractors by ensuring that none of their head events share a common keyword or conceptualization with question's keywords and conceptualizations. If this component is dropped, the constraint will be downgraded, and only no sharing of common keywords between the question and distractors will be restricted. This approach introduces abstract commonsense knowledge into the CSKB and uses the original distractor generation strategy for synthesizing QA pairs.\\nWe then train two batches of QA models, using RoBERTa-Large and DeBERTa-v3-Large as the backbone, by sequentially dropping the two components mentioned above one at a time. Their zeroshot performances on five commonsense QA benchmarks are reported in Table 9. From the results, it is observed that both components play important roles in CAR, with CCQS being more effective on average. This underscores the significance of eliminating false negative distractors, and conceptualization proves to be a useful tool for achieving this objective in improving the QA model's overall performance.\"\n },\n {\n \"figure_ref\": [\n \"fig_6\"\n ],\n \"heading\": \"B.8 The Effect of Conceptualization\",\n \"publication_ref\": [\n \"b86\",\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"Lastly, we study the improvement in the generalizability of our framework with the aid of conceptualizations by examining the accuracy gains on questions with varying levels of semantic overlap with knowledge in ATOMIC's training split. To do so, we sort the questions in every benchmark by their average BERTScore (Zhang et al., 2020) between each individual question entry against the whole training set in the original ATOMIC. We then split the questions into two sets based on their BERTScores, with the lower BERTScore indicating a lower semantic overlap and a greater need for the model to generalize to answer the question. These questions are denoted as \\\"Difficult.\\\" Conversely, we refer to questions with high BERTScores as \\\"Easy.\\\" Then, we train two QA models following the pipeline proposed by Ma et al. (2021), with one trained on conceptualization-augmented ATOMIC and the other on ATOMIC only. We evaluate their performance on five commonsense QA benchmarks and compare the performance gains between two sets of questions in each benchmark, as shown in Figure 6. Results demonstrate that incorporating conceptualizations positively impacts accuracy, particularly for questions that deviate significantly from ATOMIC across multiple benchmarks. This indicates that augmenting ATOMIC with conceptualizations can improve the model's generalizability, particularly for questions that tend to be out-ofdistribution, requiring more relevant knowledge to answer correctly.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.9 Generalization to Other CSKBs\",\n \"publication_ref\": [\n \"b42\",\n \"b68\",\n \"b47\",\n \"b74\",\n \"b57\"\n ],\n \"table_ref\": [\n \"tab_11\"\n ],\n \"text\": \"While our work primarily experiments with the AbstractATOMIC dataset as the conceptualization source of ATOMIC, we also aim to extend our framework to other CSKBs for a more generalizable evaluation. To address this, we follow Ma et al. (2021) and explore the feasibility of transferring our framework to the CWWV dataset, which comprises multiple CSKBs including Concept-Net (Speer et al., 2017), WordNet (Miller, 1995), and WikiData (Vrandecic and Krötzsch, 2014). To accomplish this, we train a conceptualization generator based on GPT2 (Radford et al., 2019) and utilize ChatGPT (OpenAI, 2022) as two flexible generative conceptualizers. The generated conceptualizations are then transformed into abstract knowledge and integrated into the CWWV dataset. This augmented dataset is employed to train a zeroshot commonsense QA reasoner using our proposed CAR framework. We present the experimental results and compare them with baselines in Table 10. Our observations reveal a modest improvement in an average accuracy of 1% compared to all baselines and comparable performance to GPT3.5. These results demonstrate the effectiveness of incorporating conceptualizations from other CSKBs. In future research, we suggest exploring automatic construction methods for conceptualization resources in other CSKBs and investigating their potential benefits for general commonsense reasoning.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C Case Study\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\",\n \"tab_6\"\n ],\n \"text\": \"In this section, we present case studies to demonstrate the effectiveness of CAR. First, we discuss cases that illustrate the power of conceptualization augmentation, as shown in Table 12. By transforming triples into abstract commonsense knowledge, we can introduce more general knowledge into the CSKB and improve its coverage. Moreover, the newly introduced triples were missing from the original CSKB. For instance, conceptualizing \\\"Per-sonX plays the games together\\\" as an \\\"entertainment activity\\\" introduces higher-level knowledge that cannot be simply represented by the original triple. Additionally, by synthesizing both types of triples into QA pairs, the QA model can learn both types of knowledge, which can help the model perform more generalizable reasoning on out-ofdistribution commonsense QA benchmarks.\\nNext, in Table 13, we present QA pairs consisting of false negative options generated using keyword constraints during their synthesis from both the original ATOMIC and ATOMIC-10X. We also demonstrate how our concept constraint resolves this issue. Through these case studies, we observe that the original distractors may contain one or even both plausible options, which is suboptimal when training a QA model. Specifically, for distractors sampled from ATOMIC-10X, we observe that several distractors are vague and general (denoted as \\\"?\\\"), which can be plausible in many contexts. For example, in various triples, adjectives like \\\"happy\\\" and verb phrases such as \\\"do it\\\" are easy to be plausible and do not serve as significant distractions. This is not desirable when training a QA model. However, by using conceptualizations as a constraint, the newly sampled distractors are all strong negatives, allowing the model to learn from such negative commonsense knowledge. This is because the distractors are sourced from triples that are more likely to be irrelevant to the original triple's context and, thus, more likely to be truly negative distractors.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgements\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The authors would like to thank Haochen Shi for his help in implementing the training dynamics and the anonymous reviewers for their constructive comments. The authors of this paper were supported by the NSFC Fund (U20B2053) from the NSFC of China, the RIF (R6020-19 and R6021-20), and the GRF (16211520 and 16205322) from RGC of Hong Kong. We also thank the support from the UGC Research Matching Grants (RMGS20EG01-D, RMGS20CR11, RMGS20CR12, RMGS20EG19, RMGS20EG21, RMGS23CR05, RMGS23EG08). We also gratefully acknowledge the support of the Swiss National Science Foundation (No. 215390), Innosuisse (PFFS-21-29), the EPFL Science Seed Fund, the EPFL Center for Imaging, Sony Group Corporation, and the Allen Institute for AI.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \" \",\n \"publication_ref\": [\n \"b68\",\n \"b33\",\n \"b74\",\n \"b47\",\n \"b28\"\n ],\n \"table_ref\": [],\n \"text\": \"et al., 2022)\\n. All baseline results are consistent with their original papers. CWWV refers to the combination of ConceptNet (Speer et al., 2017), VisualGenome (Krishna et al., 2017), WikiData (Vrandecic and Krötzsch, 2014), and WordNet (Miller, 1995). CSKG (Ilievski et al., 2021) consists of ATOMIC (Sap et al., 2019a) and CWWV. Table 12: Case study of conceptualized triples and their synthesized QA pairs. Given an original triple from ATOMIC, we conceptualize the triple by replacing an instance with its [plausible conceptualization] to form a conceptualized triple. The conceptualized triples are then synthesized into QA pairs using the same ground-truth answer and distractors, sampled for the original triple, to train the QA model. * indicates the ground-truth answer. Table 13: Case study of the false negative options in the original QA synthesis and how our proposed conceptualization-constraint resolves such an issue. The (ground truth answer) is appended below each question. Keyword represents only using keywords as the constraint for sampling distractor, while Concept refers to using both the keywords and conceptualizations. F.Neg. refers to whether the distractor is a false negative one or not.\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":""},"doi":{"kind":"string","value":"10.48550/arXiv.2305.19068"},"references":{"kind":"list like","value":[{"authors":"Jiaxin Bai; Xin Liu; Weiqi Wang; Chen Luo; Yangqiu Song","journal":"","ref_id":"b0","title":"Complex query answering on eventuality knowledge graph with implicit logical constraints","year":"2023"},{"authors":"Pratyay Banerjee; Chitta Baral","journal":"Association for Computational Linguistics","ref_id":"b1","title":"Selfsupervised knowledge triplet learning for zero-shot question answering","year":"2020-11-16"},{"authors":"Yejin Bang; Samuel Cahyawijaya; Nayeon Lee; Wenliang Dai; Dan Su; Bryan Wilie; Holy Lovenia; Ziwei Ji; Tiezheng Yu; Willy Chung; Quyet V Do; Yan Xu; Pascale Fung","journal":"","ref_id":"b2","title":"A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity","year":"2023"},{"authors":"Chandra Bhagavatula; Le Ronan; Chaitanya Bras; Keisuke Malaviya; Ari Sakaguchi; Hannah Holtzman; Doug Rashkin; Wen-Tau Downey; Yejin Yih; Choi","journal":"","ref_id":"b3","title":"Abductive commonsense reasoning","year":"2020-04-26"},{"authors":"Yonatan Bisk; Rowan Zellers; Le Ronan; Jianfeng Bras; Yejin Gao; Choi","journal":"AAAI Press","ref_id":"b4","title":"PIQA: reasoning about physical commonsense in natural language","year":"2020-02-07"},{"authors":"Antoine Bosselut; Ronan Le Bras; Yejin Choi","journal":"AAAI Press","ref_id":"b5","title":"Dynamic neuro-symbolic knowledge graph construction for zero-shot commonsense question answering","year":"2021-02-02"},{"authors":"Antoine Bosselut; Hannah Rashkin; Maarten Sap; Chaitanya Malaviya; Asli Celikyilmaz; Yejin Choi","journal":"Association for Computational Linguistics","ref_id":"b6","title":"COMET: commonsense transformers for automatic knowledge graph construction","year":"2019-07-28"},{"authors":"Ruben Branco; António Branco; João António Rodrigues; João Ricardo Silva","journal":"Association for Computational Linguistics","ref_id":"b7","title":"Shortcutted commonsense: Data spuriousness in deep learning of commonsense reasoning","year":"2021-07-11"},{"authors":"B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mccandlish; Ilya Radford; Dario Sutskever; Amodei","journal":"","ref_id":"b8","title":"Language models are few-shot learners","year":"2020-12-06"},{"authors":"Chunkit Chan; Jiayang Cheng; Weiqi Wang; Yuxin Jiang; Tianqing Fang; Xin Liu; Yangqiu Song","journal":"","ref_id":"b9","title":"Chatgpt evaluation on sentence level relations: A focus on temporal, causal, and discourse relations","year":"2023"},{"authors":"Jiangjie Chen; Wei Shi; Ziquan Fu; Sijie Cheng; Lei Li; Yanghua Xiao; ; ","journal":"Association for Computational Linguistics","ref_id":"b10","title":"Say what you mean! large language models speak too positively about negative commonsense knowledge","year":"2023"},{"authors":"Zeming Chen; Gail Weiss; Eric Mitchell; Asli Celikyilmaz; Antoine Bosselut","journal":"","ref_id":"b11","title":"RECKON-ING: reasoning through dynamic knowledge encoding","year":"2023"},{"authors":"Kevin Clark; Minh-Thang Luong; Quoc V Le; Christopher D Manning","journal":"","ref_id":"b12","title":"ELECTRA: pretraining text encoders as discriminators rather than generators","year":"2020-04-26"},{"authors":"Jeff Da; Le Ronan; Ximing Bras; Yejin Lu; Antoine Choi; Bosselut","journal":"","ref_id":"b13","title":"Analyzing commonsense emergence in few-shot knowledge models","year":"2021-10-04"},{"authors":"Zheye Deng; Weiqi Wang; Zhaowei Wang; Xin Liu; Yangqiu Song","journal":"","ref_id":"b14","title":"Gold: A global and local-aware denoising framework for commonsense knowledge graph noise detection","year":"2023"},{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"Association for Computational Linguistics","ref_id":"b15","title":"BERT: pre-training of deep bidirectional transformers for language understanding","year":"2019-06-02"},{"authors":"Wenxuan Ding; Shangbin Feng; Yuhan Liu; Zhaoxuan Tan; Vidhisha Balachandran; Tianxing He; Yulia Tsvetkov","journal":"","ref_id":"b16","title":"Knowledge crosswords: Geometric reasoning over structured knowledge with large language models","year":"2023"},{"authors":"Zi-Yi Dou; Nanyun Peng","journal":"AAAI Press","ref_id":"b17","title":"Zero-shot commonsense question answering with cloze translation and consistency optimization","year":"2022-02-22"},{"authors":"Benjamin Van Durme; Phillip Michalak; Lenhart K Schubert","journal":"The Association for Computer Linguistics","ref_id":"b18","title":"Deriving generalized knowledge from corpora using wordnet abstraction","year":"2009-03-30"},{"authors":"Tianqing Fang; Weiqi Wang; Sehyun Choi; Shibo Hao; Hongming Zhang; Yangqiu Song; Bin He; ; ","journal":"Association for Computational Linguistics","ref_id":"b19","title":"Benchmarking commonsense knowledge base population with an effective evaluation dataset","year":"2021-07-11"},{"authors":"Tianqing Fang; Hongming Zhang; Weiqi Wang; Yangqiu Song; Bin He","journal":"ACM","ref_id":"b20","title":"DISCOS: bridging the gap between discourse knowledge and commonsense knowledge","year":"2021-04-19"},{"authors":"Silin Gao; Beatriz Borges; Soyoung Oh; Deniz Bayazit; Saya Kanno; Hiromi Wakaki; Yuki Mitsufuji; Antoine Bosselut","journal":"Association for Computational Linguistics","ref_id":"b21","title":"Peacok: Persona commonsense knowledge for consistent and engaging narratives","year":"2023-07-09"},{"authors":"Yu Gong; Kaiqi Zhao; Kenny Qili; Zhu ","journal":"AAAI Press","ref_id":"b22","title":"Representing verbs as argument concepts","year":"2016-02-12"},{"authors":"Xin Guan; Biwei Cao; Qingqing Gao; Zheng Yin; Bo Liu; Jiuxin Cao","journal":"","ref_id":"b23","title":"Multi-hop commonsense knowledge injection framework for zeroshot commonsense question answering","year":"2023"},{"authors":"Mutian He; Tianqing Fang; Weiqi Wang; Yangqiu Song","journal":"","ref_id":"b24","title":"Acquiring and modelling abstract commonsense knowledge via conceptualization","year":"2022"},{"authors":"Pengcheng He; Jianfeng Gao; Weizhu Chen","journal":"","ref_id":"b25","title":"DeBERTav3: Improving deBERTa using ELECTRAstyle pre-training with gradient-disentangled embedding sharing","year":"2023"},{"authors":"Luyao Huang; Chi Sun; Xipeng Qiu; Xuanjing Huang","journal":"Association for Computational Linguistics","ref_id":"b26","title":"Glossbert: BERT for word sense disambiguation with gloss knowledge","year":"2019-11-03"},{"authors":"Jena D Hwang; Chandra Bhagavatula; Le Ronan; Jeff Bras; Keisuke Da; Antoine Sakaguchi; Yejin Bosselut; Choi","journal":"AAAI Press","ref_id":"b27","title":"comet-) atomic 2020: On symbolic and neural commonsense knowledge graphs","year":"2021-02-02"},{"authors":"Filip Ilievski; Pedro A Szekely; Bin Zhang","journal":"Springer","ref_id":"b28","title":"CSKG: the commonsense knowledge graph","year":"2021-06-06"},{"authors":"Mete Ismayilzada; Antoine Bosselut","journal":"Association for Computational Linguistics","ref_id":"b29","title":"kogito: A commonsense knowledge inference toolkit","year":"2023-05-02"},{"authors":"Liwei Jiang; Antoine Bosselut; Chandra Bhagavatula; Yejin Choi","journal":"Association for Computational Linguistics","ref_id":"b30","title":"i'm not mad\": Commonsense implications of negation and contradiction","year":"2021-06-06"},{"authors":"Jin Yu; Beong-Woo Kim; Youngwook Kwak; Reinald Kim; Seung-Won Kim Amplayo; Jinyoung Hwang; Yeo","journal":"Association for Computational Linguistics","ref_id":"b31","title":"Modularized transfer learning with multiple knowledge graphs for zero-shot commonsense reasoning","year":"2022-07-10"},{"authors":"Sosuke Kobayashi","journal":"Association for Computational Linguistics","ref_id":"b32","title":"Contextual augmentation: Data augmentation by words with paradigmatic relations","year":"2018-06-01"},{"authors":"Ranjay Krishna; Yuke Zhu; Oliver Groth; Justin Johnson; Kenji Hata; Joshua Kravitz; Stephanie Chen; Yannis Kalantidis; Li-Jia Li; David A Shamma; Michael S Bernstein; Li Fei-Fei","journal":"Int. J. Comput. Vis","ref_id":"b33","title":"Visual genome: Connecting language and vision using crowdsourced dense image annotations","year":"2017"},{"authors":"Yen-Ling Kuo; Jane Yung; -Jen Hsu","journal":"AAAI","ref_id":"b34","title":"Bridging common sense knowledge bases with analogy by graph similarity","year":"2010-07-11"},{"authors":"Xiang Li; Aynaz Taheri; Lifu Tu; Kevin Gimpel","journal":"The Association for Computer Linguistics","ref_id":"b35","title":"Commonsense knowledge base completion","year":"2016-08-07"},{"authors":"Lorraine Xiang; Adhiguna Li; Jordan Kuncoro; Cyprien Hoffmann; Phil De Masson D'autume; Aida Blunsom; Nematzadeh","journal":"Association for Computational Linguistics","ref_id":"b36","title":"A systematic investigation of commonsense knowledge in large language models","year":"2022-12-07"},{"authors":"Zhongli Li; Wenhui Wang; Li Dong; Furu Wei; Ke Xu","journal":"Association for Computational Linguistics","ref_id":"b37","title":"Harvesting and refining questionanswer pairs for unsupervised QA","year":"2020-07-05"},{"authors":"Jingping Liu; Tao Chen; Chao Wang; Jiaqing Liang; Lihan Chen; Yanghua Xiao; Yunwen Chen; Ke Jin","journal":"Artif. Intell","ref_id":"b38","title":"Vocsk: Verb-oriented commonsense knowledge mining with taxonomy-guided induction","year":"2022"},{"authors":"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b39","title":"Roberta: A robustly optimized BERT pretraining approach","year":"2019"},{"authors":"Ilya Loshchilov; Frank Hutter","journal":"","ref_id":"b40","title":"Decoupled weight decay regularization","year":"2019-05-06"},{"authors":"Kaixin Ma; Jonathan Francis; Quanyang Lu; Eric Nyberg; Alessandro Oltramari","journal":"","ref_id":"b41","title":"Towards generalizable neuro-symbolic systems for commonsense question answering","year":"2019"},{"authors":"Kaixin Ma; Filip Ilievski; Jonathan Francis; Yonatan Bisk; Eric Nyberg; Alessandro Oltramari","journal":"AAAI Press","ref_id":"b42","title":"Knowledge-driven data construction for zero-shot evaluation in commonsense question answering","year":"2021-02-02"},{"authors":"Chaitanya Malaviya; Chandra Bhagavatula; Antoine Bosselut; Yejin Choi","journal":"AAAI Press","ref_id":"b43","title":"Commonsense knowledge base completion with structural and semantic context","year":"2020-02-07"},{"authors":"Tom Mccoy; Ellie Pavlick; Tal Linzen","journal":"Association for Computational Linguistics","ref_id":"b44","title":"Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference","year":"2019-07-28"},{"authors":"Mary L Mchugh","journal":"Biochemia medica","ref_id":"b45","title":"Interrater reliability: the kappa statistic","year":"2012"},{"authors":"Tomás Mikolov; Kai Chen; Greg Corrado; Jeffrey Dean","journal":"","ref_id":"b46","title":"Efficient estimation of word representations in vector space","year":"2013-05-02"},{"authors":"George A Miller","journal":"Commun. ACM","ref_id":"b47","title":"Wordnet: A lexical database for english","year":"1995"},{"authors":"Sewon Min; Xinxi Lyu; Ari Holtzman; Mikel Artetxe; Mike Lewis; Hannaneh Hajishirzi; Luke Zettlemoyer","journal":"Association for Computational Linguistics","ref_id":"b48","title":"Rethinking the role of demonstrations: What makes in-context learning work?","year":"2022-12-07"},{"authors":"Gregory Murphy","journal":"MIT press","ref_id":"b49","title":"The big book of concepts","year":"2004"},{"authors":"Tri Nguyen; Mir Rosenberg; Xia Song; Jianfeng Gao; Saurabh Tiwary; Rangan Majumder; Li Deng","journal":"","ref_id":"b50","title":"MS MARCO: A human generated machine reading comprehension dataset","year":"2016-12-09"},{"authors":"Tong Niu; Mohit Bansal","journal":"Association for Computational Linguistics","ref_id":"b51","title":"Adversarial oversensitivity and over-stability strategies for dialogue models","year":"2018-10-31"},{"authors":" Openai","journal":"OpenAI","ref_id":"b52","title":"Chatgpt: Optimizing language models for dialogue","year":"2022"},{"authors":"Long Ouyang; Jeffrey Wu; Xu Jiang; Diogo Almeida; Carroll L Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray; John Schulman; Jacob Hilton; Fraser Kelton; Luke Miller; Maddie Simens; Amanda Askell; Peter Welinder; Paul F Christiano; Jan Leike; Ryan Lowe","journal":"","ref_id":"b53","title":"Training language models to follow instructions with human feedback","year":"2022"},{"authors":"Hao Peng; Xiaozhi Wang; Shengding Hu; Hailong Jin; Lei Hou; Juanzi Li; Zhiyuan Liu; Qun Liu","journal":"Association for Computational Linguistics","ref_id":"b54","title":"COPEN: probing conceptual knowledge in pretrained language models","year":"2022-12-07"},{"authors":"Jeffrey Pennington; Richard Socher; Christopher D Manning","journal":"Association for Computational Linguistics","ref_id":"b55","title":"Glove: Global vectors for word representation","year":"2014-10-25"},{"authors":"Chengwei Qin; Aston Zhang; Zhuosheng Zhang; Jiaao Chen; Michihiro Yasunaga; Diyi Yang","journal":"","ref_id":"b56","title":"Is chatgpt a general-purpose natural language processing task solver?","year":"2023"},{"authors":"Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever","journal":"OpenAI blog","ref_id":"b57","title":"Language models are unsupervised multitask learners","year":"2019"},{"authors":"Nils Reimers; Iryna Gurevych","journal":"Association for Computational Linguistics","ref_id":"b58","title":"Sentence-bert: Sentence embeddings using siamese bert-networks","year":"2019-11-03"},{"authors":"Joshua Robinson; Christopher Michael Rytting; David Wingate","journal":"","ref_id":"b59","title":"Leveraging large language models for multiple choice question answering","year":"2022"},{"authors":"Keisuke Sakaguchi; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi","journal":"Commun. ACM","ref_id":"b60","title":"Winogrande: an adversarial winograd schema challenge at scale","year":"2021"},{"authors":"Maarten Sap; Le Ronan; Emily Bras; Chandra Allaway; Nicholas Bhagavatula; Hannah Lourie; Brendan Rashkin; Noah A Roof; Yejin Smith; Choi","journal":"AAAI Press","ref_id":"b61","title":"ATOMIC: an atlas of machine commonsense for if-then reasoning","year":"2019-01-27"},{"authors":"Maarten Sap; Hannah Rashkin; Derek Chen; Ronan Le Bras; Yejin Choi","journal":"Association for Computational Linguistics","ref_id":"b62","title":"Social iqa: Commonsense reasoning about social interactions","year":"2019-11-03"},{"authors":"Rico Sennrich; Barry Haddow; Alexandra Birch","journal":"The Association for Computer Linguistics","ref_id":"b63","title":"Improving neural machine translation models with monolingual data","year":"2016-08-07"},{"authors":"Haochen Shi; Weiqi Wang; Tianqing Fang; Baixuan Xu; Wenxuan Ding; Xin Liu; Yangqiu Song","journal":"","ref_id":"b64","title":"Qadynamics: Training dynamics-driven synthetic qa diagnostic for zero-shot commonsense question answering","year":"2023"},{"authors":"Vered Shwartz; Peter West; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi","journal":"Association for Computational Linguistics","ref_id":"b65","title":"Unsupervised commonsense question answering with self-talk","year":"2020-11-16"},{"authors":"Yangqiu Song; Haixun Wang; Zhongyuan Wang; Hongsong Li; Weizhu Chen","journal":"IJCAI/AAAI","ref_id":"b66","title":"Short text conceptualization using a probabilistic knowledgebase","year":"2011-07-16"},{"authors":"Yangqiu Song; Shusen Wang; Haixun Wang","journal":"AAAI Press","ref_id":"b67","title":"Open domain short text conceptualization: A generative + descriptive modeling approach","year":"2015-07-25"},{"authors":"Robyn Speer; Joshua Chin; Catherine Havasi","journal":"AAAI Press","ref_id":"b68","title":"Conceptnet 5.5: An open multilingual graph of general knowledge","year":"2017-02-04"},{"authors":"Ying Su; Zihao Wang; Tianqing Fang; Hongming Zhang; Yangqiu Song; Tong Zhang","journal":"Association for Computational Linguistics","ref_id":"b69","title":"MICO: A multi-alternative contrastive learning framework for commonsense knowledge representation","year":"2022-12-07"},{"authors":"Swabha Swayamdipta; Roy Schwartz; Nicholas Lourie; Yizhong Wang; Hannaneh Hajishirzi; Noah A Smith; Yejin Choi","journal":"Association for Computational Linguistics","ref_id":"b70","title":"Dataset cartography: Mapping and diagnosing datasets with training dynamics","year":"2020-11-16"},{"authors":"Alon Talmor; Jonathan Herzig; Nicholas Lourie; Jonathan Berant","journal":"Association for Computational Linguistics","ref_id":"b71","title":"Commonsenseqa: A question answering challenge targeting commonsense knowledge","year":"2019-06-02"},{"authors":"Joshua B Tenenbaum; Charles Kemp; Thomas L Griffiths; Noah D Goodman","journal":"science","ref_id":"b72","title":"How to grow a mind: Statistics, structure, and abstraction","year":"2011"},{"authors":"H Trieu; Quoc V Trinh; Le","journal":"","ref_id":"b73","title":"A simple method for commonsense reasoning","year":"2018"},{"authors":"Denny Vrandecic; Markus Krötzsch","journal":"Commun. ACM","ref_id":"b74","title":"Wikidata: a free collaborative knowledgebase","year":"2014"},{"authors":"Peifeng Wang; Filip Ilievski; Muhao Chen; Xiang Ren","journal":"Association for Computational Linguistics","ref_id":"b75","title":"Do language models perform generalizable commonsense inference?","year":"2021-08-01"},{"authors":"Weiqi Wang; Tianqing Fang; Baixuan Xu; Chun Yi; Louis Bo; Yangqiu Song; Lei Chen; ; ","journal":"Association for Computational Linguistics","ref_id":"b76","title":"CAT: A contextualized conceptualization and instantiation framework for commonsense reasoning","year":"2023-07-09"},{"authors":"William Yang; Wang ; Diyi Yang","journal":"The Association for Computational Linguistics","ref_id":"b77","title":"That's so annoying!!!: A lexical and frame-semantic embedding based data augmentation approach to automatic categorization of annoying behaviors using #petpeeve tweets","year":"2015-09-17"},{"authors":"Xuezhi Wang; Jason Wei; Dale Schuurmans; Quoc V Le; Ed H Chi; Sharan Narang; Aakanksha Chowdhery; Denny Zhou","journal":"","ref_id":"b78","title":"Self-consistency improves chain of thought reasoning in language models","year":"2023-05-01"},{"authors":"Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Brian Ichter; Fei Xia; Ed H Chi; V Quoc; Denny Le; Zhou","journal":"","ref_id":"b79","title":"Chain-of-thought prompting elicits reasoning in large language models","year":"2022"},{"authors":"Jason W Wei; Kai Zou","journal":"Association for Computational Linguistics","ref_id":"b80","title":"EDA: easy data augmentation techniques for boosting performance on text classification tasks","year":"2019-11-03"},{"authors":"Peter West; Chandra Bhagavatula; Jack Hessel; Jena D Hwang; Liwei Jiang; Ronan Le Bras; Ximing Lu; Sean Welleck; Yejin Choi","journal":"Association for Computational Linguistics","ref_id":"b81","title":"Symbolic knowledge distillation: from general language models to commonsense models","year":"2022-07-10"},{"authors":"Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Rémi Louf; Morgan Funtowicz; Joe Davison; Sam Shleifer; Clara Patrick Von Platen; Yacine Ma; Julien Jernite; Canwen Plu; Teven Xu; Sylvain Le Scao; Mariama Gugger; Quentin Drame; Alexander M Lhoest; Rush","journal":"Association for Computational Linguistics","ref_id":"b82","title":"Transformers: State-of-the-art natural language processing","year":"2020-11-16"},{"authors":"Wentao Wu; Hongsong Li; Haixun Wang; Kenny Qili; Zhu ","journal":"ACM","ref_id":"b83","title":"Probase: a probabilistic taxonomy for text understanding","year":"2012-05-20"},{"authors":"Changlong Yu; Weiqi Wang; Xin Liu; Jiaxin Bai; Yangqiu Song; Zheng Li; Yifan Gao; Tianyu Cao; Bing Yin","journal":"Association for Computational Linguistics","ref_id":"b84","title":"FolkScope: Intention knowledge graph construction for E-commerce commonsense discovery","year":"2023"},{"authors":"Hongming Zhang; Xin Liu; Haojie Pan; Haowen Ke; Jiefu Ou; Tianqing Fang; Yangqiu Song","journal":"Artif. Intell","ref_id":"b85","title":"ASER: towards large-scale commonsense knowledge acquisition via higher-order selectional preference over eventualities","year":"2022"},{"authors":"Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi","journal":"","ref_id":"b86","title":"Bertscore: Evaluating text generation with BERT","year":"2020-04-26"},{"authors":"Pei Zhou; Rahul Khanna; Seyeon Lee; Bill Yuchen Lin; Daniel Ho; Jay Pujara; Xiang Ren","journal":"Association for Computational Linguistics","ref_id":"b87","title":"RICA: evaluating robust inference capabilities based on commonsense axioms","year":"2021-07-11"}],"string":"[\n {\n \"authors\": \"Jiaxin Bai; Xin Liu; Weiqi Wang; Chen Luo; Yangqiu Song\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Complex query answering on eventuality knowledge graph with implicit logical constraints\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Pratyay Banerjee; Chitta Baral\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b1\",\n \"title\": \"Selfsupervised knowledge triplet learning for zero-shot question answering\",\n \"year\": \"2020-11-16\"\n },\n {\n \"authors\": \"Yejin Bang; Samuel Cahyawijaya; Nayeon Lee; Wenliang Dai; Dan Su; Bryan Wilie; Holy Lovenia; Ziwei Ji; Tiezheng Yu; Willy Chung; Quyet V Do; Yan Xu; Pascale Fung\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Chandra Bhagavatula; Le Ronan; Chaitanya Bras; Keisuke Malaviya; Ari Sakaguchi; Hannah Holtzman; Doug Rashkin; Wen-Tau Downey; Yejin Yih; Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Abductive commonsense reasoning\",\n \"year\": \"2020-04-26\"\n },\n {\n \"authors\": \"Yonatan Bisk; Rowan Zellers; Le Ronan; Jianfeng Bras; Yejin Gao; Choi\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b4\",\n \"title\": \"PIQA: reasoning about physical commonsense in natural language\",\n \"year\": \"2020-02-07\"\n },\n {\n \"authors\": \"Antoine Bosselut; Ronan Le Bras; Yejin Choi\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b5\",\n \"title\": \"Dynamic neuro-symbolic knowledge graph construction for zero-shot commonsense question answering\",\n \"year\": \"2021-02-02\"\n },\n {\n \"authors\": \"Antoine Bosselut; Hannah Rashkin; Maarten Sap; Chaitanya Malaviya; Asli Celikyilmaz; Yejin Choi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b6\",\n \"title\": \"COMET: commonsense transformers for automatic knowledge graph construction\",\n \"year\": \"2019-07-28\"\n },\n {\n \"authors\": \"Ruben Branco; António Branco; João António Rodrigues; João Ricardo Silva\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b7\",\n \"title\": \"Shortcutted commonsense: Data spuriousness in deep learning of commonsense reasoning\",\n \"year\": \"2021-07-11\"\n },\n {\n \"authors\": \"B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mccandlish; Ilya Radford; Dario Sutskever; Amodei\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020-12-06\"\n },\n {\n \"authors\": \"Chunkit Chan; Jiayang Cheng; Weiqi Wang; Yuxin Jiang; Tianqing Fang; Xin Liu; Yangqiu Song\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Chatgpt evaluation on sentence level relations: A focus on temporal, causal, and discourse relations\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Jiangjie Chen; Wei Shi; Ziquan Fu; Sijie Cheng; Lei Li; Yanghua Xiao; ; \",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b10\",\n \"title\": \"Say what you mean! large language models speak too positively about negative commonsense knowledge\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Zeming Chen; Gail Weiss; Eric Mitchell; Asli Celikyilmaz; Antoine Bosselut\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"RECKON-ING: reasoning through dynamic knowledge encoding\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Kevin Clark; Minh-Thang Luong; Quoc V Le; Christopher D Manning\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"ELECTRA: pretraining text encoders as discriminators rather than generators\",\n \"year\": \"2020-04-26\"\n },\n {\n \"authors\": \"Jeff Da; Le Ronan; Ximing Bras; Yejin Lu; Antoine Choi; Bosselut\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Analyzing commonsense emergence in few-shot knowledge models\",\n \"year\": \"2021-10-04\"\n },\n {\n \"authors\": \"Zheye Deng; Weiqi Wang; Zhaowei Wang; Xin Liu; Yangqiu Song\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Gold: A global and local-aware denoising framework for commonsense knowledge graph noise detection\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b15\",\n \"title\": \"BERT: pre-training of deep bidirectional transformers for language understanding\",\n \"year\": \"2019-06-02\"\n },\n {\n \"authors\": \"Wenxuan Ding; Shangbin Feng; Yuhan Liu; Zhaoxuan Tan; Vidhisha Balachandran; Tianxing He; Yulia Tsvetkov\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Knowledge crosswords: Geometric reasoning over structured knowledge with large language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Zi-Yi Dou; Nanyun Peng\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b17\",\n \"title\": \"Zero-shot commonsense question answering with cloze translation and consistency optimization\",\n \"year\": \"2022-02-22\"\n },\n {\n \"authors\": \"Benjamin Van Durme; Phillip Michalak; Lenhart K Schubert\",\n \"journal\": \"The Association for Computer Linguistics\",\n \"ref_id\": \"b18\",\n \"title\": \"Deriving generalized knowledge from corpora using wordnet abstraction\",\n \"year\": \"2009-03-30\"\n },\n {\n \"authors\": \"Tianqing Fang; Weiqi Wang; Sehyun Choi; Shibo Hao; Hongming Zhang; Yangqiu Song; Bin He; ; \",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b19\",\n \"title\": \"Benchmarking commonsense knowledge base population with an effective evaluation dataset\",\n \"year\": \"2021-07-11\"\n },\n {\n \"authors\": \"Tianqing Fang; Hongming Zhang; Weiqi Wang; Yangqiu Song; Bin He\",\n \"journal\": \"ACM\",\n \"ref_id\": \"b20\",\n \"title\": \"DISCOS: bridging the gap between discourse knowledge and commonsense knowledge\",\n \"year\": \"2021-04-19\"\n },\n {\n \"authors\": \"Silin Gao; Beatriz Borges; Soyoung Oh; Deniz Bayazit; Saya Kanno; Hiromi Wakaki; Yuki Mitsufuji; Antoine Bosselut\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b21\",\n \"title\": \"Peacok: Persona commonsense knowledge for consistent and engaging narratives\",\n \"year\": \"2023-07-09\"\n },\n {\n \"authors\": \"Yu Gong; Kaiqi Zhao; Kenny Qili; Zhu \",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b22\",\n \"title\": \"Representing verbs as argument concepts\",\n \"year\": \"2016-02-12\"\n },\n {\n \"authors\": \"Xin Guan; Biwei Cao; Qingqing Gao; Zheng Yin; Bo Liu; Jiuxin Cao\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Multi-hop commonsense knowledge injection framework for zeroshot commonsense question answering\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Mutian He; Tianqing Fang; Weiqi Wang; Yangqiu Song\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Acquiring and modelling abstract commonsense knowledge via conceptualization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Pengcheng He; Jianfeng Gao; Weizhu Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"DeBERTav3: Improving deBERTa using ELECTRAstyle pre-training with gradient-disentangled embedding sharing\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Luyao Huang; Chi Sun; Xipeng Qiu; Xuanjing Huang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b26\",\n \"title\": \"Glossbert: BERT for word sense disambiguation with gloss knowledge\",\n \"year\": \"2019-11-03\"\n },\n {\n \"authors\": \"Jena D Hwang; Chandra Bhagavatula; Le Ronan; Jeff Bras; Keisuke Da; Antoine Sakaguchi; Yejin Bosselut; Choi\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b27\",\n \"title\": \"comet-) atomic 2020: On symbolic and neural commonsense knowledge graphs\",\n \"year\": \"2021-02-02\"\n },\n {\n \"authors\": \"Filip Ilievski; Pedro A Szekely; Bin Zhang\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b28\",\n \"title\": \"CSKG: the commonsense knowledge graph\",\n \"year\": \"2021-06-06\"\n },\n {\n \"authors\": \"Mete Ismayilzada; Antoine Bosselut\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b29\",\n \"title\": \"kogito: A commonsense knowledge inference toolkit\",\n \"year\": \"2023-05-02\"\n },\n {\n \"authors\": \"Liwei Jiang; Antoine Bosselut; Chandra Bhagavatula; Yejin Choi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b30\",\n \"title\": \"i'm not mad\\\": Commonsense implications of negation and contradiction\",\n \"year\": \"2021-06-06\"\n },\n {\n \"authors\": \"Jin Yu; Beong-Woo Kim; Youngwook Kwak; Reinald Kim; Seung-Won Kim Amplayo; Jinyoung Hwang; Yeo\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b31\",\n \"title\": \"Modularized transfer learning with multiple knowledge graphs for zero-shot commonsense reasoning\",\n \"year\": \"2022-07-10\"\n },\n {\n \"authors\": \"Sosuke Kobayashi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b32\",\n \"title\": \"Contextual augmentation: Data augmentation by words with paradigmatic relations\",\n \"year\": \"2018-06-01\"\n },\n {\n \"authors\": \"Ranjay Krishna; Yuke Zhu; Oliver Groth; Justin Johnson; Kenji Hata; Joshua Kravitz; Stephanie Chen; Yannis Kalantidis; Li-Jia Li; David A Shamma; Michael S Bernstein; Li Fei-Fei\",\n \"journal\": \"Int. J. Comput. Vis\",\n \"ref_id\": \"b33\",\n \"title\": \"Visual genome: Connecting language and vision using crowdsourced dense image annotations\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Yen-Ling Kuo; Jane Yung; -Jen Hsu\",\n \"journal\": \"AAAI\",\n \"ref_id\": \"b34\",\n \"title\": \"Bridging common sense knowledge bases with analogy by graph similarity\",\n \"year\": \"2010-07-11\"\n },\n {\n \"authors\": \"Xiang Li; Aynaz Taheri; Lifu Tu; Kevin Gimpel\",\n \"journal\": \"The Association for Computer Linguistics\",\n \"ref_id\": \"b35\",\n \"title\": \"Commonsense knowledge base completion\",\n \"year\": \"2016-08-07\"\n },\n {\n \"authors\": \"Lorraine Xiang; Adhiguna Li; Jordan Kuncoro; Cyprien Hoffmann; Phil De Masson D'autume; Aida Blunsom; Nematzadeh\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b36\",\n \"title\": \"A systematic investigation of commonsense knowledge in large language models\",\n \"year\": \"2022-12-07\"\n },\n {\n \"authors\": \"Zhongli Li; Wenhui Wang; Li Dong; Furu Wei; Ke Xu\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b37\",\n \"title\": \"Harvesting and refining questionanswer pairs for unsupervised QA\",\n \"year\": \"2020-07-05\"\n },\n {\n \"authors\": \"Jingping Liu; Tao Chen; Chao Wang; Jiaqing Liang; Lihan Chen; Yanghua Xiao; Yunwen Chen; Ke Jin\",\n \"journal\": \"Artif. Intell\",\n \"ref_id\": \"b38\",\n \"title\": \"Vocsk: Verb-oriented commonsense knowledge mining with taxonomy-guided induction\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Roberta: A robustly optimized BERT pretraining approach\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Ilya Loshchilov; Frank Hutter\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Decoupled weight decay regularization\",\n \"year\": \"2019-05-06\"\n },\n {\n \"authors\": \"Kaixin Ma; Jonathan Francis; Quanyang Lu; Eric Nyberg; Alessandro Oltramari\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Towards generalizable neuro-symbolic systems for commonsense question answering\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Kaixin Ma; Filip Ilievski; Jonathan Francis; Yonatan Bisk; Eric Nyberg; Alessandro Oltramari\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b42\",\n \"title\": \"Knowledge-driven data construction for zero-shot evaluation in commonsense question answering\",\n \"year\": \"2021-02-02\"\n },\n {\n \"authors\": \"Chaitanya Malaviya; Chandra Bhagavatula; Antoine Bosselut; Yejin Choi\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b43\",\n \"title\": \"Commonsense knowledge base completion with structural and semantic context\",\n \"year\": \"2020-02-07\"\n },\n {\n \"authors\": \"Tom Mccoy; Ellie Pavlick; Tal Linzen\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b44\",\n \"title\": \"Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference\",\n \"year\": \"2019-07-28\"\n },\n {\n \"authors\": \"Mary L Mchugh\",\n \"journal\": \"Biochemia medica\",\n \"ref_id\": \"b45\",\n \"title\": \"Interrater reliability: the kappa statistic\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Tomás Mikolov; Kai Chen; Greg Corrado; Jeffrey Dean\",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"Efficient estimation of word representations in vector space\",\n \"year\": \"2013-05-02\"\n },\n {\n \"authors\": \"George A Miller\",\n \"journal\": \"Commun. ACM\",\n \"ref_id\": \"b47\",\n \"title\": \"Wordnet: A lexical database for english\",\n \"year\": \"1995\"\n },\n {\n \"authors\": \"Sewon Min; Xinxi Lyu; Ari Holtzman; Mikel Artetxe; Mike Lewis; Hannaneh Hajishirzi; Luke Zettlemoyer\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b48\",\n \"title\": \"Rethinking the role of demonstrations: What makes in-context learning work?\",\n \"year\": \"2022-12-07\"\n },\n {\n \"authors\": \"Gregory Murphy\",\n \"journal\": \"MIT press\",\n \"ref_id\": \"b49\",\n \"title\": \"The big book of concepts\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Tri Nguyen; Mir Rosenberg; Xia Song; Jianfeng Gao; Saurabh Tiwary; Rangan Majumder; Li Deng\",\n \"journal\": \"\",\n \"ref_id\": \"b50\",\n \"title\": \"MS MARCO: A human generated machine reading comprehension dataset\",\n \"year\": \"2016-12-09\"\n },\n {\n \"authors\": \"Tong Niu; Mohit Bansal\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b51\",\n \"title\": \"Adversarial oversensitivity and over-stability strategies for dialogue models\",\n \"year\": \"2018-10-31\"\n },\n {\n \"authors\": \" Openai\",\n \"journal\": \"OpenAI\",\n \"ref_id\": \"b52\",\n \"title\": \"Chatgpt: Optimizing language models for dialogue\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Long Ouyang; Jeffrey Wu; Xu Jiang; Diogo Almeida; Carroll L Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray; John Schulman; Jacob Hilton; Fraser Kelton; Luke Miller; Maddie Simens; Amanda Askell; Peter Welinder; Paul F Christiano; Jan Leike; Ryan Lowe\",\n \"journal\": \"\",\n \"ref_id\": \"b53\",\n \"title\": \"Training language models to follow instructions with human feedback\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hao Peng; Xiaozhi Wang; Shengding Hu; Hailong Jin; Lei Hou; Juanzi Li; Zhiyuan Liu; Qun Liu\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b54\",\n \"title\": \"COPEN: probing conceptual knowledge in pretrained language models\",\n \"year\": \"2022-12-07\"\n },\n {\n \"authors\": \"Jeffrey Pennington; Richard Socher; Christopher D Manning\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b55\",\n \"title\": \"Glove: Global vectors for word representation\",\n \"year\": \"2014-10-25\"\n },\n {\n \"authors\": \"Chengwei Qin; Aston Zhang; Zhuosheng Zhang; Jiaao Chen; Michihiro Yasunaga; Diyi Yang\",\n \"journal\": \"\",\n \"ref_id\": \"b56\",\n \"title\": \"Is chatgpt a general-purpose natural language processing task solver?\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Alec Radford; Jeffrey Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever\",\n \"journal\": \"OpenAI blog\",\n \"ref_id\": \"b57\",\n \"title\": \"Language models are unsupervised multitask learners\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Nils Reimers; Iryna Gurevych\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b58\",\n \"title\": \"Sentence-bert: Sentence embeddings using siamese bert-networks\",\n \"year\": \"2019-11-03\"\n },\n {\n \"authors\": \"Joshua Robinson; Christopher Michael Rytting; David Wingate\",\n \"journal\": \"\",\n \"ref_id\": \"b59\",\n \"title\": \"Leveraging large language models for multiple choice question answering\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Keisuke Sakaguchi; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi\",\n \"journal\": \"Commun. ACM\",\n \"ref_id\": \"b60\",\n \"title\": \"Winogrande: an adversarial winograd schema challenge at scale\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Maarten Sap; Le Ronan; Emily Bras; Chandra Allaway; Nicholas Bhagavatula; Hannah Lourie; Brendan Rashkin; Noah A Roof; Yejin Smith; Choi\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b61\",\n \"title\": \"ATOMIC: an atlas of machine commonsense for if-then reasoning\",\n \"year\": \"2019-01-27\"\n },\n {\n \"authors\": \"Maarten Sap; Hannah Rashkin; Derek Chen; Ronan Le Bras; Yejin Choi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b62\",\n \"title\": \"Social iqa: Commonsense reasoning about social interactions\",\n \"year\": \"2019-11-03\"\n },\n {\n \"authors\": \"Rico Sennrich; Barry Haddow; Alexandra Birch\",\n \"journal\": \"The Association for Computer Linguistics\",\n \"ref_id\": \"b63\",\n \"title\": \"Improving neural machine translation models with monolingual data\",\n \"year\": \"2016-08-07\"\n },\n {\n \"authors\": \"Haochen Shi; Weiqi Wang; Tianqing Fang; Baixuan Xu; Wenxuan Ding; Xin Liu; Yangqiu Song\",\n \"journal\": \"\",\n \"ref_id\": \"b64\",\n \"title\": \"Qadynamics: Training dynamics-driven synthetic qa diagnostic for zero-shot commonsense question answering\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Vered Shwartz; Peter West; Le Ronan; Chandra Bras; Yejin Bhagavatula; Choi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b65\",\n \"title\": \"Unsupervised commonsense question answering with self-talk\",\n \"year\": \"2020-11-16\"\n },\n {\n \"authors\": \"Yangqiu Song; Haixun Wang; Zhongyuan Wang; Hongsong Li; Weizhu Chen\",\n \"journal\": \"IJCAI/AAAI\",\n \"ref_id\": \"b66\",\n \"title\": \"Short text conceptualization using a probabilistic knowledgebase\",\n \"year\": \"2011-07-16\"\n },\n {\n \"authors\": \"Yangqiu Song; Shusen Wang; Haixun Wang\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b67\",\n \"title\": \"Open domain short text conceptualization: A generative + descriptive modeling approach\",\n \"year\": \"2015-07-25\"\n },\n {\n \"authors\": \"Robyn Speer; Joshua Chin; Catherine Havasi\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b68\",\n \"title\": \"Conceptnet 5.5: An open multilingual graph of general knowledge\",\n \"year\": \"2017-02-04\"\n },\n {\n \"authors\": \"Ying Su; Zihao Wang; Tianqing Fang; Hongming Zhang; Yangqiu Song; Tong Zhang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b69\",\n \"title\": \"MICO: A multi-alternative contrastive learning framework for commonsense knowledge representation\",\n \"year\": \"2022-12-07\"\n },\n {\n \"authors\": \"Swabha Swayamdipta; Roy Schwartz; Nicholas Lourie; Yizhong Wang; Hannaneh Hajishirzi; Noah A Smith; Yejin Choi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b70\",\n \"title\": \"Dataset cartography: Mapping and diagnosing datasets with training dynamics\",\n \"year\": \"2020-11-16\"\n },\n {\n \"authors\": \"Alon Talmor; Jonathan Herzig; Nicholas Lourie; Jonathan Berant\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b71\",\n \"title\": \"Commonsenseqa: A question answering challenge targeting commonsense knowledge\",\n \"year\": \"2019-06-02\"\n },\n {\n \"authors\": \"Joshua B Tenenbaum; Charles Kemp; Thomas L Griffiths; Noah D Goodman\",\n \"journal\": \"science\",\n \"ref_id\": \"b72\",\n \"title\": \"How to grow a mind: Statistics, structure, and abstraction\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"H Trieu; Quoc V Trinh; Le\",\n \"journal\": \"\",\n \"ref_id\": \"b73\",\n \"title\": \"A simple method for commonsense reasoning\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Denny Vrandecic; Markus Krötzsch\",\n \"journal\": \"Commun. ACM\",\n \"ref_id\": \"b74\",\n \"title\": \"Wikidata: a free collaborative knowledgebase\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Peifeng Wang; Filip Ilievski; Muhao Chen; Xiang Ren\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b75\",\n \"title\": \"Do language models perform generalizable commonsense inference?\",\n \"year\": \"2021-08-01\"\n },\n {\n \"authors\": \"Weiqi Wang; Tianqing Fang; Baixuan Xu; Chun Yi; Louis Bo; Yangqiu Song; Lei Chen; ; \",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b76\",\n \"title\": \"CAT: A contextualized conceptualization and instantiation framework for commonsense reasoning\",\n \"year\": \"2023-07-09\"\n },\n {\n \"authors\": \"William Yang; Wang ; Diyi Yang\",\n \"journal\": \"The Association for Computational Linguistics\",\n \"ref_id\": \"b77\",\n \"title\": \"That's so annoying!!!: A lexical and frame-semantic embedding based data augmentation approach to automatic categorization of annoying behaviors using #petpeeve tweets\",\n \"year\": \"2015-09-17\"\n },\n {\n \"authors\": \"Xuezhi Wang; Jason Wei; Dale Schuurmans; Quoc V Le; Ed H Chi; Sharan Narang; Aakanksha Chowdhery; Denny Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b78\",\n \"title\": \"Self-consistency improves chain of thought reasoning in language models\",\n \"year\": \"2023-05-01\"\n },\n {\n \"authors\": \"Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Brian Ichter; Fei Xia; Ed H Chi; V Quoc; Denny Le; Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b79\",\n \"title\": \"Chain-of-thought prompting elicits reasoning in large language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jason W Wei; Kai Zou\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b80\",\n \"title\": \"EDA: easy data augmentation techniques for boosting performance on text classification tasks\",\n \"year\": \"2019-11-03\"\n },\n {\n \"authors\": \"Peter West; Chandra Bhagavatula; Jack Hessel; Jena D Hwang; Liwei Jiang; Ronan Le Bras; Ximing Lu; Sean Welleck; Yejin Choi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b81\",\n \"title\": \"Symbolic knowledge distillation: from general language models to commonsense models\",\n \"year\": \"2022-07-10\"\n },\n {\n \"authors\": \"Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Rémi Louf; Morgan Funtowicz; Joe Davison; Sam Shleifer; Clara Patrick Von Platen; Yacine Ma; Julien Jernite; Canwen Plu; Teven Xu; Sylvain Le Scao; Mariama Gugger; Quentin Drame; Alexander M Lhoest; Rush\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b82\",\n \"title\": \"Transformers: State-of-the-art natural language processing\",\n \"year\": \"2020-11-16\"\n },\n {\n \"authors\": \"Wentao Wu; Hongsong Li; Haixun Wang; Kenny Qili; Zhu \",\n \"journal\": \"ACM\",\n \"ref_id\": \"b83\",\n \"title\": \"Probase: a probabilistic taxonomy for text understanding\",\n \"year\": \"2012-05-20\"\n },\n {\n \"authors\": \"Changlong Yu; Weiqi Wang; Xin Liu; Jiaxin Bai; Yangqiu Song; Zheng Li; Yifan Gao; Tianyu Cao; Bing Yin\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b84\",\n \"title\": \"FolkScope: Intention knowledge graph construction for E-commerce commonsense discovery\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Hongming Zhang; Xin Liu; Haojie Pan; Haowen Ke; Jiefu Ou; Tianqing Fang; Yangqiu Song\",\n \"journal\": \"Artif. Intell\",\n \"ref_id\": \"b85\",\n \"title\": \"ASER: towards large-scale commonsense knowledge acquisition via higher-order selectional preference over eventualities\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi\",\n \"journal\": \"\",\n \"ref_id\": \"b86\",\n \"title\": \"Bertscore: Evaluating text generation with BERT\",\n \"year\": \"2020-04-26\"\n },\n {\n \"authors\": \"Pei Zhou; Rahul Khanna; Seyeon Lee; Bill Yuchen Lin; Daniel Ho; Jay Pujara; Xiang Ren\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b87\",\n \"title\": \"RICA: evaluating robust inference capabilities based on commonsense axioms\",\n \"year\": \"2021-07-11\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,306.14,387.35,168.84,9.57],"formula_id":"formula_0","formula_text":"D = {(h, r, t)|h ∈ H, r ∈ R, t ∈ T }"},{"formula_coordinates":[3,306.14,439.6,218.27,25.07],"formula_id":"formula_1","formula_text":"D C = {(h c , r, t)|h c ∈ H c , r ∈ R, t ∈ T }"},{"formula_coordinates":[3,372.88,748.07,153.44,12.58],"formula_id":"formula_2","formula_text":"D Q = {(Q i , A i )|(h i , r i , t i ) ∈ D}."},{"formula_coordinates":[5,70.87,358,220.08,51.48],"formula_id":"formula_3","formula_text":"C h = {c i 1 ,1 , c i 1 ,2 , • • • , c i 2 ,1 , • • • }, we as- sociate a triple (h, r, t) with T h + C h to form its constraint. Only knowledge triple (h ′ , r, t ′ ) which satisfies (T h ′ + C h ′ ) ∩ (T h + C h ) ="},{"formula_coordinates":[5,315.74,150.74,209.4,33.71],"formula_id":"formula_4","formula_text":"S(T ) = - 1 n n i=1 log P (t i |..., t i-1 , t i+1 , ...) (1)"},{"formula_coordinates":[5,334.79,270.96,190.35,33.98],"formula_id":"formula_5","formula_text":"L = 1 m m i=1,i̸ =y max(0, η -S y + S i )(2)"},{"formula_coordinates":[22,81.31,542.78,208.56,34.1],"formula_id":"formula_6","formula_text":"C(Q i , A i ) = 1 n n c=1 σ( m d=1 (S c i,d -S c i,j ) m -1 )(3)"},{"formula_coordinates":[22,71.02,738.21,218.71,34.75],"formula_id":"formula_7","formula_text":"V(Qi, Ai) = n c=1 (σ( m d=1 (S c i,d -S c i,j ) m-1 ) -C(Qi, Ai)) 2 n (4)"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 306.14,\n 387.35,\n 168.84,\n 9.57\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"D = {(h, r, t)|h ∈ H, r ∈ R, t ∈ T }\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 306.14,\n 439.6,\n 218.27,\n 25.07\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"D C = {(h c , r, t)|h c ∈ H c , r ∈ R, t ∈ T }\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 372.88,\n 748.07,\n 153.44,\n 12.58\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"D Q = {(Q i , A i )|(h i , r i , t i ) ∈ D}.\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 70.87,\n 358,\n 220.08,\n 51.48\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"C h = {c i 1 ,1 , c i 1 ,2 , • • • , c i 2 ,1 , • • • }, we as- sociate a triple (h, r, t) with T h + C h to form its constraint. Only knowledge triple (h ′ , r, t ′ ) which satisfies (T h ′ + C h ′ ) ∩ (T h + C h ) =\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 315.74,\n 150.74,\n 209.4,\n 33.71\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"S(T ) = - 1 n n i=1 log P (t i |..., t i-1 , t i+1 , ...) (1)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 334.79,\n 270.96,\n 190.35,\n 33.98\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"L = 1 m m i=1,i̸ =y max(0, η -S y + S i )(2)\"\n },\n {\n \"formula_coordinates\": [\n 22,\n 81.31,\n 542.78,\n 208.56,\n 34.1\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"C(Q i , A i ) = 1 n n c=1 σ( m d=1 (S c i,d -S c i,j ) m -1 )(3)\"\n },\n {\n \"formula_coordinates\": [\n 22,\n 71.02,\n 738.21,\n 218.71,\n 34.75\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"V(Qi, Ai) = n c=1 (σ( m d=1 (S c i,d -S c i,j ) m-1 ) -C(Qi, Ai)) 2 n (4)\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"CAR: Conceptualization-Augmented Reasoner for Zero-Shot Commonsense Question Answering"},"abstract":{"kind":"string","value":"The task of zero-shot commonsense question answering evaluates models on their capacity to reason about general scenarios beyond those presented in specific datasets. Existing approaches for tackling this task leverage external knowledge from CommonSense Knowledge Bases (CSKBs) by pre-training the model on synthetic QA pairs constructed from CSKBs. In these approaches, negative examples (distractors) are formulated by randomly sampling from CSKBs using fairly primitive keyword constraints. However, two bottlenecks limit these approaches: the inherent incompleteness of CSKBs limits the semantic coverage of synthetic QA pairs, and the lack of human annotations makes the sampled negative examples potentially uninformative and contradictory. To tackle these limitations above, we propose Conceptualization-Augmented Reasoner (CAR), a zero-shot commonsense questionanswering framework that fully leverages the power of conceptualization. Specifically, CAR abstracts a commonsense knowledge triple to many higher-level instances, which increases the coverage of the CSKB and expands the ground-truth answer space, reducing the likelihood of selecting false-negative distractors. Extensive experiments demonstrate that CAR more robustly generalizes to answering questions about zero-shot commonsense scenarios than existing methods, including large language models, such as GPT3.5 and Chat-GPT. Our codes, data, and model checkpoints are available at https://github.com/HKUST-KnowComp/CAR."},"authors":{"kind":"string","value":"Weiqi Wang; Tianqing Fang; Wenxuan Ding; Baixuan Xu; Xin Liu; Yangqiu Song; Antoine Bosselut"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: An example of constructing synthetic QA pairs from CSKB(Ma et al., 2021). The simple heuristic used in this process can result in false negative options.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: An example of conceptualization inference. More abstracted knowledge, such as (Do sport, xWant, take a rest), can be obtained through conceptualization.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"PersonX arrive at the bar, what does PersonX want to do? A: relax himself B: keep fit C: cry PersonX arrive at [entertainment place], what does PersonX want to do? A: relax himself B: keep fit C","figure_data":"","figure_id":"fig_2","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: An overview of the CAR framework, which shows the process of synthesizing (PersonX arrive at the bar, xWant, relax himself) into QA pairs. The triple is conceptualized first, and potential distractor triples are sampled and filtered by keyword and concept overlap. Only those triples that have no overlap are used as distractors.","figure_data":"","figure_id":"fig_3","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Analyses on training dynamics of different knowledge. The dotted lines refer to the median values.","figure_data":"","figure_id":"fig_4","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: Average accuracy achieved by models trained on our training set downsampled to several ratios.","figure_data":"","figure_id":"fig_5","figure_label":"5","figure_type":"figure"},{"figure_caption":"Figure 6 :6Figure 6: Comparison of accuracy improvement (%) with/without conceptualization-augmentation for two groups of QA entries across five benchmarks. Avg. stands for averaging across all benchmarks.","figure_data":"","figure_id":"fig_6","figure_label":"6","figure_type":"figure"},{"figure_caption":"PersonX arrive at the [","figure_data":"PersonX enjoy running,PersonX shout out loud,PersonX is at the casino,PersonX arrive at office,xWant,xWant,xWant,xWant,keep fitcryhave a drinkmeet his colleagueConceptualizationPersonX enjoy [sport], xWant, keep fit[shouting], xWant, cryPersonX is at [entertainment place], xWant, have a drinkPersonX [arrival], xWant, meet his colleagueConceptualizationPre-TrainingQA Benchmarks","figure_id":"tab_0","figure_label":"","figure_type":"table"},{"figure_caption":"; Ouyang ↑1.5 64.8 ↑0.6 73.2 ↑1.1 64.8 ↑1.7 61.3 ↑2.1 67.3 ↑1.4 ⋄ CAR-RoBERTa-L (Ours) ATM C 72.7 ↑1.9 66.3 ↑2.1 73.2 ↑1.1 64.0 ↑0.9 62.0 ↑2.8 67.6 ↑1.7 ↑2.9 67.2 ↑0.2 78.6 ↑0.6 63.8 ↑1.7 78.1 ↑2.1 73.3 ↑1.5 ⋄ CAR-DeBERTa-v3-L (Ours) ATM C 79.6 ↑3.6 69.3 ↑2.3 78.6 ↑0.6 64.0 ↑1.9 78.2 ↑2.2 73.9 ↑2.1","figure_data":"ModelCSKBa-NLICSQAPIQASIQAWGAvg.Random-50.020.050.033.350.040.7Majority-50.820.950.533.650.441.2RoBERTa-L (Liu et al., 2019)-65.545.067.647.357.556.6DeBERTa-v3-L (He et al., 2023)-59.925.444.847.850.345.6Self-talk (Shwartz et al., 2020)--32.470.246.254.7-COMET-DynGen (Bosselut et al., 2021)ATOMIC---50.1--SMLM (Banerjee and Baral, 2020)*65.338.8-48.5--MICO (Su et al., 2022)ATOMIC-44.2-56.0--STL-Adapter (Kim et al., 2022)ATOMIC71.366.571.164.460.366.7Backbone: RoBERTa-Large 340MRoBERTa-L (MR) (Ma et al., 2021)ATM-10X 70.859.472.158.558.363.8△ RoBERTa-L (MR) (Ma et al., 2021)ATOMIC70.864.272.163.159.265.9⋄ CAR-RoBERTa-L (Ours) 72.3 Backbone: DeBERTa-v3-Large 435M ATOMICDeBERTa-v3-L (MR) (Ma et al., 2021)ATM-10X 75.171.679.059.771.771.4△ DeBERTa-v3-L (MR) (Ma et al., 2021)ATOMIC76.067.078.062.176.071.8⋄ CAR-DeBERTa-v3-L (Ours) 78.9 Large Language Models ATOMICGPT-3.5 (text-davinci-003)-61.868.967.868.060.765.4ChatGPT (gpt-3.5-turbo)-69.374.575.169.562.870.2Supervised Learning & Human PerformanceRoBERTa-L (Supervised)-85.678.579.276.679.379.8DeBERTa-v3-L (Supervised)-89.082.184.580.184.184.0Human Performance-91.488.994.986.994.191.2","figure_id":"tab_3","figure_label":"","figure_type":"table"},{"figure_caption":"Comparison results (%) of different augmentation methods against conceptualization. N/A stands for not using any augmentation. Plau. is the expert-evaluated ratio of plausible augmented knowledge, %F.Neg. represents the expert-annotated proportion of false negative options. Div. and Exp.Div. are diversities measured by embedding similarity and expert annotated knowledge coverage. Performances on the right refer to accuracies achieved by the QA model trained on data augmented by each method. The best performances are bold-faced.","figure_data":"AugmentationDiv↑ Exp.Div↑ Plau.↑ %F.Neg.↓ aNLI CSQA PIQA SIQA WGN/A (Baseline)N/AN/A88.045.776.067.078.062.176.0EDA (Wei and Zou, 2019)8.104.679.3333.076.565.676.661.474.9Word2Vec (Wang and Yang, 2015) 11.84.009.0055.074.365.875.162.974.7GLOVE (Wang and Yang, 2015)8.216.674.6744.374.764.274.661.174.4BERT-base (Kobayashi, 2018)0.818.3314.341.770.463.972.463.561.0Synonym (Niu and Bansal, 2018)6.9211.05.6745.075.564.974.562.575.7GPT3-distil (West et al., 2022)35.624.395.742.775.471.875.663.476.0Conceptualization (Ours)48.537.090.022.779.669.378.664.078.2based on DeBERTa-v3-Large and trained on ourconceptualization-augmented ATOMIC, achievesstate-of-the-art results and significantly outper-forms all PLM-based baselines on every bench-mark, and can advance the average accuracy by2.1% compared with the same baseline model. Italso significantly surpasses the performance of thesame model that is trained on ATOMIC-10X withonly 10% amount of data (more explanations andexperiments in Appendix B.5). Notably, comparedwith LLMs, our system champions three bench-marks and performs better on average with a 3.7%leap. This indicates that supervision signals fromCSKBs are important for downstream applications,and CSKBs aided by conceptualization can signifi-cantly enhance this process. Moreover, as an abla-tion, we study the role of concept-level distractorsampling by discarding conceptualization augmen-tation and only training the models on ATOMIC,synthesized to QA format with our proposed con-straint technique. Comparing the results in Table 1,it can be observed that the concept-level distrac-tor sampling improves the average performance byapproximately 1.5%. This demonstrates that ourproposed technique is effective, and generating dis-tractors with a stronger positive knowledge recallis helpful in synthesizing QA pairs that are bothfair and informative.","figure_id":"tab_4","figure_label":"2","figure_type":"table"},{"figure_caption":"Experiments on the generalizability of CAR on other CSKBs (CWWV).","figure_data":"ModelCSKBAvg.RoBERTa-L (MR) (Ma et al., 2021)CWWV64.8MTL (Kim et al., 2022)CWWV63.7ZS-Fusion (Kim et al., 2022)CWWV64.7CAR-RoBERTa-L (Ours)CWWV C 65.8","figure_id":"tab_6","figure_label":"3","figure_type":"table"},{"figure_caption":"Statistics on the number of QA pairs and the number of options for each question within each benchmark's validation split.","figure_data":"aNLI CSQA PIQA SIQAWG#QA Pairs 1,5321,2211,838 1,954 1,267#Options25232","figure_id":"tab_7","figure_label":"4","figure_type":"table"},{"figure_caption":"Statistics of abstract commonsense triples used in CAR, as reported byWang et al. (2023a). D l","figure_data":"","figure_id":"tab_9","figure_label":"6","figure_type":"table"},{"figure_caption":"Zero-shot evaluation results (%) on five commonsense question answering benchmarks by models trained on the CWWV dataset. CWWV C refers to the augmented CWWV dataset using generated conceptualizations from a trained GPT2 generator and ChatGPT.","figure_data":"ModelCSKBa-NLI CSQA PIQA SIQA WG Avg.RoBERTa-L (MR) (Ma et al., 2021)CWWV70.067.972.054.859.4 64.8MTL (Kim et al., 2022)CWWV69.667.372.552.057.2 63.7ZS-Fusion (Kim et al., 2022)CWWV69.667.673.153.759.5 64.7CAR-RoBERTa-L (Ours)CWWV C 71.668.473.055.460.6 65.8GPT-3.5 (text-davinci-003)N/A61.868.967.868.060.7 65.4ChatGPT (gpt-3.5-turbo)N/A69.374.575.169.562.8 70.2","figure_id":"tab_11","figure_label":"10","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: An example of constructing synthetic QA pairs from CSKB(Ma et al., 2021). The simple heuristic used in this process can result in false negative options.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: An example of conceptualization inference. More abstracted knowledge, such as (Do sport, xWant, take a rest), can be obtained through conceptualization.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"PersonX arrive at the bar, what does PersonX want to do? A: relax himself B: keep fit C: cry PersonX arrive at [entertainment place], what does PersonX want to do? A: relax himself B: keep fit C\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: An overview of the CAR framework, which shows the process of synthesizing (PersonX arrive at the bar, xWant, relax himself) into QA pairs. The triple is conceptualized first, and potential distractor triples are sampled and filtered by keyword and concept overlap. Only those triples that have no overlap are used as distractors.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Analyses on training dynamics of different knowledge. The dotted lines refer to the median values.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: Average accuracy achieved by models trained on our training set downsampled to several ratios.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 :6Figure 6: Comparison of accuracy improvement (%) with/without conceptualization-augmentation for two groups of QA entries across five benchmarks. Avg. stands for averaging across all benchmarks.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"PersonX arrive at the [\",\n \"figure_data\": \"PersonX enjoy running,PersonX shout out loud,PersonX is at the casino,PersonX arrive at office,xWant,xWant,xWant,xWant,keep fitcryhave a drinkmeet his colleagueConceptualizationPersonX enjoy [sport], xWant, keep fit[shouting], xWant, cryPersonX is at [entertainment place], xWant, have a drinkPersonX [arrival], xWant, meet his colleagueConceptualizationPre-TrainingQA Benchmarks\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"; Ouyang ↑1.5 64.8 ↑0.6 73.2 ↑1.1 64.8 ↑1.7 61.3 ↑2.1 67.3 ↑1.4 ⋄ CAR-RoBERTa-L (Ours) ATM C 72.7 ↑1.9 66.3 ↑2.1 73.2 ↑1.1 64.0 ↑0.9 62.0 ↑2.8 67.6 ↑1.7 ↑2.9 67.2 ↑0.2 78.6 ↑0.6 63.8 ↑1.7 78.1 ↑2.1 73.3 ↑1.5 ⋄ CAR-DeBERTa-v3-L (Ours) ATM C 79.6 ↑3.6 69.3 ↑2.3 78.6 ↑0.6 64.0 ↑1.9 78.2 ↑2.2 73.9 ↑2.1\",\n \"figure_data\": \"ModelCSKBa-NLICSQAPIQASIQAWGAvg.Random-50.020.050.033.350.040.7Majority-50.820.950.533.650.441.2RoBERTa-L (Liu et al., 2019)-65.545.067.647.357.556.6DeBERTa-v3-L (He et al., 2023)-59.925.444.847.850.345.6Self-talk (Shwartz et al., 2020)--32.470.246.254.7-COMET-DynGen (Bosselut et al., 2021)ATOMIC---50.1--SMLM (Banerjee and Baral, 2020)*65.338.8-48.5--MICO (Su et al., 2022)ATOMIC-44.2-56.0--STL-Adapter (Kim et al., 2022)ATOMIC71.366.571.164.460.366.7Backbone: RoBERTa-Large 340MRoBERTa-L (MR) (Ma et al., 2021)ATM-10X 70.859.472.158.558.363.8△ RoBERTa-L (MR) (Ma et al., 2021)ATOMIC70.864.272.163.159.265.9⋄ CAR-RoBERTa-L (Ours) 72.3 Backbone: DeBERTa-v3-Large 435M ATOMICDeBERTa-v3-L (MR) (Ma et al., 2021)ATM-10X 75.171.679.059.771.771.4△ DeBERTa-v3-L (MR) (Ma et al., 2021)ATOMIC76.067.078.062.176.071.8⋄ CAR-DeBERTa-v3-L (Ours) 78.9 Large Language Models ATOMICGPT-3.5 (text-davinci-003)-61.868.967.868.060.765.4ChatGPT (gpt-3.5-turbo)-69.374.575.169.562.870.2Supervised Learning & Human PerformanceRoBERTa-L (Supervised)-85.678.579.276.679.379.8DeBERTa-v3-L (Supervised)-89.082.184.580.184.184.0Human Performance-91.488.994.986.994.191.2\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison results (%) of different augmentation methods against conceptualization. N/A stands for not using any augmentation. Plau. is the expert-evaluated ratio of plausible augmented knowledge, %F.Neg. represents the expert-annotated proportion of false negative options. Div. and Exp.Div. are diversities measured by embedding similarity and expert annotated knowledge coverage. Performances on the right refer to accuracies achieved by the QA model trained on data augmented by each method. The best performances are bold-faced.\",\n \"figure_data\": \"AugmentationDiv↑ Exp.Div↑ Plau.↑ %F.Neg.↓ aNLI CSQA PIQA SIQA WGN/A (Baseline)N/AN/A88.045.776.067.078.062.176.0EDA (Wei and Zou, 2019)8.104.679.3333.076.565.676.661.474.9Word2Vec (Wang and Yang, 2015) 11.84.009.0055.074.365.875.162.974.7GLOVE (Wang and Yang, 2015)8.216.674.6744.374.764.274.661.174.4BERT-base (Kobayashi, 2018)0.818.3314.341.770.463.972.463.561.0Synonym (Niu and Bansal, 2018)6.9211.05.6745.075.564.974.562.575.7GPT3-distil (West et al., 2022)35.624.395.742.775.471.875.663.476.0Conceptualization (Ours)48.537.090.022.779.669.378.664.078.2based on DeBERTa-v3-Large and trained on ourconceptualization-augmented ATOMIC, achievesstate-of-the-art results and significantly outper-forms all PLM-based baselines on every bench-mark, and can advance the average accuracy by2.1% compared with the same baseline model. Italso significantly surpasses the performance of thesame model that is trained on ATOMIC-10X withonly 10% amount of data (more explanations andexperiments in Appendix B.5). Notably, comparedwith LLMs, our system champions three bench-marks and performs better on average with a 3.7%leap. This indicates that supervision signals fromCSKBs are important for downstream applications,and CSKBs aided by conceptualization can signifi-cantly enhance this process. Moreover, as an abla-tion, we study the role of concept-level distractorsampling by discarding conceptualization augmen-tation and only training the models on ATOMIC,synthesized to QA format with our proposed con-straint technique. Comparing the results in Table 1,it can be observed that the concept-level distrac-tor sampling improves the average performance byapproximately 1.5%. This demonstrates that ourproposed technique is effective, and generating dis-tractors with a stronger positive knowledge recallis helpful in synthesizing QA pairs that are bothfair and informative.\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Experiments on the generalizability of CAR on other CSKBs (CWWV).\",\n \"figure_data\": \"ModelCSKBAvg.RoBERTa-L (MR) (Ma et al., 2021)CWWV64.8MTL (Kim et al., 2022)CWWV63.7ZS-Fusion (Kim et al., 2022)CWWV64.7CAR-RoBERTa-L (Ours)CWWV C 65.8\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Statistics on the number of QA pairs and the number of options for each question within each benchmark's validation split.\",\n \"figure_data\": \"aNLI CSQA PIQA SIQAWG#QA Pairs 1,5321,2211,838 1,954 1,267#Options25232\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Statistics of abstract commonsense triples used in CAR, as reported byWang et al. (2023a). D l\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_9\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Zero-shot evaluation results (%) on five commonsense question answering benchmarks by models trained on the CWWV dataset. CWWV C refers to the augmented CWWV dataset using generated conceptualizations from a trained GPT2 generator and ChatGPT.\",\n \"figure_data\": \"ModelCSKBa-NLI CSQA PIQA SIQA WG Avg.RoBERTa-L (MR) (Ma et al., 2021)CWWV70.067.972.054.859.4 64.8MTL (Kim et al., 2022)CWWV69.667.372.552.057.2 63.7ZS-Fusion (Kim et al., 2022)CWWV69.667.673.153.759.5 64.7CAR-RoBERTa-L (Ours)CWWV C 71.668.473.055.460.6 65.8GPT-3.5 (text-davinci-003)N/A61.868.967.868.060.7 65.4ChatGPT (gpt-3.5-turbo)N/A69.374.575.169.562.8 70.2\",\n \"figure_id\": \"tab_11\",\n \"figure_label\": \"10\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Supporting Evidence\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work by Ma et al. provides evidence that pre-trained language models struggle to generalize to distributionally different examples from their training sets, which supports the claim made in the citing paper about the limitations of these models in this regard.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Shwartz et al., 2020)\", \"Explanation\": \"The cited work introduces the concept of zero-shot commonsense Question-Answering (QA), which the citing paper builds upon in their research on improving the generalization of models in commonsense reasoning tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Li et al., 2020)\", \"Explanation\": \"The cited work provides a method for evaluating the performance of models in zero-shot commonsense QA tasks, which the citing paper uses to assess the effectiveness of their research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Hwang et al., 2021)\", \"Explanation\": \"The cited work injects commonsense knowledge from CSKBs into PLMs for fine-tuning in synthetic QA pairs, which the citing paper extends by further improving the performance of models in this area.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Jiang et al., 2021)\", \"Explanation\": \"The cited work injects commonsense knowledge from CSKBs into PLMs for fine-tuning in synthetic QA pairs, which the citing paper extends by further improving the performance of models in this area.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work provides a method for sampling negative examples in synthetic QA pairs for fine-tuning PLMs in CSKBs, which the citing paper uses to improve the performance of models in this area.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Sap et al., 2019b)\", \"Explanation\": \"The cited work introduces the SocialIQA dataset, which the citing paper uses to assess the performance of models in zero-shot commonsense QA tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Sap et al., 2019a)\", \"Explanation\": \"The cited work introduces the ATOMIC dataset, which the citing paper uses to assess the performance of models in zero-shot commonsense QA tasks.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Bisk et al., 2020)\", \"Explanation\": \"The cited work provides a method for evaluating the performance of models in zero-shot commonsense QA tasks in other domains, which the citing paper uses to assess the effectiveness of their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kuo and Hsu, 2010)\", \"Explanation\": \"The cited work by Kuo and Hsu (2010) introduces the concept of incomplete CSKBs, which the citing paper builds upon to discuss the limitations of manual CSKBs in covering all possible knowledge in the real world.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(He et al., 2022)\", \"Explanation\": \"The cited work by He et al. (2022) extends the discussion on the limitations of manual CSKBs by highlighting the infeasibility of covering all conceivable knowledge in the real world, which the citing paper further elaborates on.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Li et al., 2016)\", \"Explanation\": \"The cited work by Li et al. (2016) provides a method for expanding CSKBs through knowledge base completion, which the citing paper uses as a data source to discuss the limitations of manual CSKBs in covering all possible knowledge.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Malaviya et al., 2020)\", \"Explanation\": \"The cited work by Malaviya et al. (2020) presents another method for expanding CSKBs through knowledge base completion, which the citing paper uses as a data source to discuss the limitations of manual CSKBs in covering all possible knowledge.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(West et al., 2022)\", \"Explanation\": \"The cited work by West et al. (2022) discusses the use of knowledge distillation from large language models to expand CSKBs, which the citing paper uses as a data source to discuss the limitations of manual CSKBs in covering all possible knowledge.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Gao et al., 2023)\", \"Explanation\": \"The cited work by Gao et al. (2023) presents a method for expanding CSKBs through knowledge distillation from large language models, which the citing paper uses as a data source to discuss the limitations of manual CSKBs in covering all possible knowledge.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Chen et al., 2023a)\", \"Explanation\": \"The cited work by Chen et al. (2023a) provides evidence that negative examples are required for models to learn to distinguish the validity of commonsense scenarios, which the citing paper uses to highlight the need for more accurate negative examples in zero-shot commonsense QA.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Murphy, 2004)\", \"Explanation\": \"The cited work by Murphy (2004) posits the use of conceptual induction in human reasoning, which the citing paper adopts in their research to draw inferences about unseen situations and improve commonsense knowledge.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(He et al., 2022)\", \"Explanation\": \"The cited work by He et al. (2022) introduces the concept of conceptualization, which the citing paper further extends by utilizing it to abstract a set of instances into concepts and derive associated commonsense knowledge for improved reasoning in downstream situations.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Swayamdip al., 2020)\", \"Explanation\": \"The cited work by Swayamdipta et al. provides a method for analyzing the ambiguity of training examples, which the citing paper uses to understand the impact of conceptualization on the out-of-domain generalization of the proposed framework.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Trinh and Le, 2018)\", \"Explanation\": \"The cited work provides a method of using off-the-shelf language models for zero-shot commonsense QA by employing vanilla language modeling with prompts.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Li et al., 2022)\", \"Explanation\": \"The cited work also uses off-the-shelf language models for zero-shot commonsense QA, but with a different method of using inference-time mechanisms specifically designed for reasoning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Shwartz et al., 2020)\", \"Explanation\": \"The cited work extends the use of off-the-shelf language models for zero-shot commonsense QA by introducing a new inference-time mechanism of self-talk for reasoning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Dou and Peng, 2022)\", \"Explanation\": \"The cited work further extends the use of off-the-shelf language models for zero-shot commonsense QA by introducing a new inference-time mechanism of cloze translation for reasoning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Bosselut et al., 2021)\", \"Explanation\": \"The cited work also extends the use of off-the-shelf language models for zero-shot commonsense QA by introducing a new inference-time mechanism of dynamic generation of reasoning sub-graphs and graph reasoning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Banerjee and Baral, 2020)\", \"Explanation\": \"The cited work introduces a new pipeline for zero-shot commonsense QA that leverages external CSKBs as knowledge sources for fine-tuning off-the-shelf language models.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work also extends the use of external CSKBs as knowledge sources for fine-tuning off-the-shelf language models in zero-shot commonsense QA.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Su et al., 2022)\", \"Explanation\": \"The cited work further extends the use of external CSKBs as knowledge sources for fine-tuning off-the-shelf language models in zero-shot commonsense QA by converting knowledge triples to synthetic QA pairs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kim et al., 2022)\", \"Explanation\": \"The cited work by Kim et al. provides a fine-tuning paradigm that the citing paper adopts to incorporate CSKBs within different domains and exploit multi-hop graph structures with graph neural networks for state-of-the-art performances in QA context.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Shi et al., 2023)\", \"Explanation\": \"The cited work by Shi et al. extends the research on fine-tuning paradigm by incorporating CSKBs within different domains to further improve the model's commonsense sensitivity in a QA context.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Guan et al., 2023)\", \"Explanation\": \"The cited work by Guan et al. is a data source that the citing paper utilizes in their research on fine-tuning paradigm with multi-hop graph structures and graph neural networks for state-of-the-art performances in QA context.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Song et al., 2011)\", \"Explanation\": \"The cited work by Song et al. provides foundational evidence on the process of conceptualization in commonsense reasoning, which the citing paper leverages to study the derivation of abstract commonsense knowledge in a QA context.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Sap et al., 2019a)\", \"Explanation\": \"The cited work establishes the ATOMIC benchmark and knowledge base, which is used by Wang et al. (2023a) in their research on conceptualization of CSKBs at scale with semi-supervised learning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Bosselut et al., 2019;Da et al., 2021)\", \"Explanation\": \"The cited works provide the foundations for the use of abstract knowledge in commonsense inference modeling, which is a key aspect of the research conducted by Wang et al. (2023a).\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Talmor et al., 2019)\", \"Explanation\": \"The cited work is a data source for the research on commonsense QA conducted by Wang et al. (2023a), as it provides the benchmark for evaluating the impact of conceptualization on downstream tasks.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Nguyen et al., 2016)\", \"Explanation\": \"The cited work is a data source for the research on machine reading comprehension conducted by Wang et al. (2023a), as it provides the benchmark for evaluating the impact of conceptualization on downstream tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wei and Zou, 2019)\", \"Explanation\": \"The cited work proposes a method of random perturbation for augmenting textual data, which the citing paper adopts in their research to enhance the data used in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang and Yang, 2015)\", \"Explanation\": \"The cited work introduces text embeddings as a method for augmenting textual data, which the citing paper uses in their research to improve the data used in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Niu and Bansal, 2018)\", \"Explanation\": \"The cited work presents a method of using lexical semantics for data augmentation, which the citing paper employs in their research to enhance the data used in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Sennrich et al., 2016)\", \"Explanation\": \"The cited work discusses the use of back translation for data augmentation, which the citing paper adopts in their research to improve the data used in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Malaviya et al., 2020)\", \"Explanation\": \"The cited work highlights the issue of false negatives in negative sampling from large language models, which the citing paper addresses in their research by proposing a method to overcome this challenge in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(He et al., 2022)\", \"Explanation\": \"The cited work by He et al. provides the formalization of a conceptualized CSKB, which the citing paper uses to condition the knowledge triples in the CSKB and obtain the set of conceptualized head events.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Sap et al., 2019b)\", \"Explanation\": \"The cited work provides the source of the commonsense knowledge used in the citing paper, in the form of (h, r, t) triples associated with events.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(He et al., 2022)\", \"Explanation\": \"The cited work provides the human-annotated abstract knowledge used to train a generative conceptualizer for acquiring the CSKB D in the citing paper, extending the research on conceptualization and knowledge acquisition.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work by Ma et al. (2021) provides a methodological basis for the citing paper by introducing a specific approach or technique that the citing paper adopts or adapts in its research or analysis.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang et al., 2023a)\", \"Explanation\": \"The cited work provides a semi-supervised conceptualization-instantiation framework that the citing paper adopts to determine the plausibility of candidate abstract knowledge triples in the CSKB.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work by Ma et al. (2021) provides the methodological basis for training the QA model in the citing paper by fine-tuning a pre-trained Masked Language Model (MLM) using the Marginal Ranking (MR) loss.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Shwartz et al., 2020)\", \"Explanation\": \"The cited work by Shwartz et al. (2020) provides a self-talk approach that the citing paper adopts to demonstrate the characteristics of the benchmark.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Bosselut et al., 2021)\", \"Explanation\": \"The cited work by Bosselut et al. (2021) presents the COMET-DynaGen method that the citing paper uses to show the power of fine-tuning in the task.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Banerjee and Baral, 2020)\", \"Explanation\": \"The cited work by Banerjee and Baral (2020) introduces the SMLM approach that the citing paper includes in the results to demonstrate the performance of different methods in the task.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Su et al., 2022)\", \"Explanation\": \"The cited work by Su et al. (2022) presents the MICO method that the citing paper uses to show the performance of different approaches in the task.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kim et al., 2022)\", \"Explanation\": \"The cited work by Kim et al. (2022) introduces the STL-Adapter method that the citing paper compares with to validate the efficacy of the proposed framework.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work by Ma et al. (2021) is a method that the citing paper compares with to validate the efficacy of the proposed approach, as both methods share similar model architecture and training procedures.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work provides a dataset of synthetic QA pairs that the citing paper uses in their research on large language models and the performance of GPT3.5 in the ATOMIC task.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Brown et al., 2020 2021)\", \"Explanation\": \"The cited work is a large language model (LLM) that the citing paper uses to benchmark the performance of GPT3.5 in the ATOMIC task, extending the research on LLMs in the field of question answering.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(West et al., 2022)\", \"Explanation\": \"The cited work, ATOMIC-10X, is used as a source of commonsense knowledge in the research on the performance of LLM in the ATOMIC task, building upon the original work on the ATOMIC dataset.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Min et al., 2022)\", \"Explanation\": \"The cited work is a method of in-context learning that the citing paper does not apply in their research on the performance of LLM in the ATOMIC task, highlighting the reliance on a different approach to prompt the LLM.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wei et al., 2022)\", \"Explanation\": \"The cited work is a method of chain-of-thought reasoning that the citing paper does not apply in their research on the performance of LLM in the ATOMIC task, indicating a focus on zero-shot prompting in the study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Robinson et al., 2022)\", \"Explanation\": \"The cited work by Robinson et al. provides the methodology of presenting a question, several choices, and a natural language command to the LLM, which the citing paper adopts in their research to obtain predictions from the LLM.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang et al., 2023a)\", \"Explanation\": \"The cited work provides a pre-trained conceptualization discriminator that the citing paper leverages to perform the task of conceptualization in the framework.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Swayamdipta et al., 2020)\", \"Explanation\": \"The cited work by Swayamdipta et al. provides a methodology for training dynamics analysis, which the citing paper adopts to study the effects of conceptualization in the context of synthetic QA pairs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work by Ma et al. (2021) is used as a reference to report the results of a study on the performance of a particular data augmentation method in terms of accuracy and diversity.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work by Ma et al. (2021) provides a baseline accuracy of 80% for expert annotators working on question-answer pairs, which the citing paper builds upon to achieve a higher accuracy of 86%.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"DeBERTa-v3-Large models\", \"Explanation\": \"The citing paper extends the research by training DeBERTa-v3-Large models on the QA pairs synthesized from the concatenation of both original and augmented ATOMIC triples from each method, exploring new dimensions in the training process.\"}, {\"Category\": \"Data Source\", \"Citation\": \"Keywords of each head event\", \"Explanation\": \"The cited work by the authors of the paper provides the keywords of each head event that are used as constraints in the model training process.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"zero-shot commonsense reasoning ability\", \"Explanation\": \"The citing paper extends the research by training DeBERTa-v3-Large models on the QA pairs synthesized from the concatenation of both original and augmented ATOMIC triples from each method, with a focus on improving the model's zero-shot commonsense reasoning ability.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"zero-shot manner\", \"Explanation\": \"The citing paper extends the research by evaluating the performance of the models in a zero-shot manner, exploring the model's ability to generalize to new tasks without any additional training.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Swayamdipta et al., 2020)\", \"Explanation\": \"The cited work by Swayamdipta et al. (2020) suggests that the use of data with higher variability and lower confidence in the training process can result in a more robust model to handle OOD data, which the citing paper leverages in their research to improve the generalization of their model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang et al., 2023a)\", \"Explanation\": \"The cited work by Wang et al. (2023a) provides the off-the-shelf GPT2 conceptualizer that the citing paper uses in their research to generate conceptualizations for the CWWV dataset.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Speer et al., 2017)\", \"Explanation\": \"The CWWV dataset includes ConceptNet as one of the CSKBs, which the citing paper uses as a data source for their research on transferring the framework to CSKBs other than ATOMIC.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Miller, 1995)\", \"Explanation\": \"The CWWV dataset also includes WordNet as a CSKB, which the citing paper uses as a data source for their research on transferring the framework to CSKBs other than ATOMIC.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Vrandecic and Kr\\u00f6tzsch, 2014)\", \"Explanation\": \"The CWWV dataset includes Wiki-Data as a CSKB, which the citing paper uses as a data source for their research on transferring the framework to CSKBs other than ATOMIC.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work by Ma et al. provides the ATOMIC dataset, which is used in the citing paper as a foundational element for the study conducted.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Kim et al., 2022)\", \"Explanation\": \"The cited work by Kim et al. is referenced for the study of zero-shot commonsense question answering task, which involves the consolidation of multiple CSKBs including ATOMIC, Concept-Net, WordNet, VisualGenome, and Wiki-Data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Dou and Peng, 2022)\", \"Explanation\": \"The cited work by Dou and Peng contributes to the study of zero-shot commonsense question answering task by consolidating multiple CSKBs, including ATOMIC, Concept-Net, WordNet, VisualGenome, and Wiki-Data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Sap et al., 2019a)\", \"Explanation\": \"The cited work by Sap et al. is referenced for the study of zero-shot commonsense question answering task by consolidating the ATOMIC CSKB.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Speer et al., 2017)\", \"Explanation\": \"The cited work by Speer et al. is referenced for the study of zero-shot commonsense question answering task by consolidating the Concept-Net CSKB.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Miller, 1995)\", \"Explanation\": \"The cited work by Miller is referenced for the study of zero-shot commonsense question answering task by consolidating the WordNet CSKB.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Krishna et al., 2017)\", \"Explanation\": \"The cited work by Krishna et al. is referenced for the study of zero-shot commonsense question answering task by consolidating the VisualGenome CSKB.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Vrandecic and Kr\\u00f6tzsch, 2014)\", \"Explanation\": \"The cited work by Vrandecic and Kr\\u00f6tzsch is referenced for the study of zero-shot commonsense question answering task by consolidating the Wiki-Data CSKB.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(He et al., 2022)\", \"Explanation\": \"The cited work by He et al. is referenced for the availability of the conceptualized expansion of the ATOMIC CSKB, which is used in the study conducted in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Wang et al., 2023a)\", \"Explanation\": \"The cited work provides a conceptualization framework that can be applied to other CSKBs, which the citing paper builds upon to further improve the research in the field of CSKBs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(He et al., 2022)\", \"Explanation\": \"The cited work provides a definition of conceptualization that the citing paper follows in their study of commonsense reasoning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang et al., 2023a)\", \"Explanation\": \"The cited work is cited for its contribution in conceptualizing an instance within an event to a concept, which the citing paper adopts in their research on commonsense inference modeling.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Yu et al., 2023)\", \"Explanation\": \"The cited work is cited for its contribution in the construction of commonsense knowledge graphs, which the citing paper utilizes in their study of conceptualization.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2022)\", \"Explanation\": \"The cited work is cited for its contribution in the construction of commonsense knowledge graphs, which the citing paper utilizes in their study of conceptualization.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(He et al., 2022)\", \"Explanation\": \"The cited work provides the Abstrac-tATOMIC dataset as a primary source for conceptualizations in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wu et al., 2012)\", \"Explanation\": \"The cited work provides the Probase dataset, which is used in the citing paper to match identified instances against the Probase knowledge base.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Miller, 1995)\", \"Explanation\": \"The cited work provides the WordNet dataset, which is used in the citing paper to match identified instances against the WordNet knowledge base.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Huang et al., 2019)\", \"Explanation\": \"The cited work introduces the Gloss-BERT method, which is used in the citing paper to find conceptualization candidates for identified instances.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Radford et al., 2019)\", \"Explanation\": \"The cited work introduces the GPT2-based generator, which is used in the citing paper to generate additional conceptualizations for expanding the size of the conceptualization bank.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang et al., 2023a)\", \"Explanation\": \"The cited work provides a semi-supervised instantiation discriminator that the citing paper adopts to pseudo-label the unannotated triples in the data set.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Banerjee and Baral, 2020)\", \"Explanation\": \"The cited work provides the official implementation of SMLM, which the citing paper adopts in their research to perform a specific task using the CSKB that exhibits the highest alignment with the task.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Kim et al., 2022)\", \"Explanation\": \"The cited work introduces STL-Adapter, which the citing paper uses to compare their results with the official results reported in the original paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Sap et al., 2019a)\", \"Explanation\": \"The cited work provides the CSKB ATOMIC, which the citing paper uses in their research to compare the performance of their framework with baseline methods that use multiple CSKBs.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Speer et al., 2017)\", \"Explanation\": \"The cited work provides the CSKB ConceptNet, which the citing paper uses in their research to compare the performance of their framework with baseline methods that use multiple CSKBs.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Miller, 1995)\", \"Explanation\": \"The cited work provides the CSKB WordNet, which the citing paper uses in their research to compare the performance of their framework with baseline methods that use multiple CSKBs.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Vrandecic and Kr\\u00f6tzsch, 2014)\", \"Explanation\": \"The cited work provides the CSKB WikiData, which the citing paper uses in their research to compare the performance of their framework with baseline methods that use multiple CSKBs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kim et al., 2022)\", \"Explanation\": \"The cited work by Kim et al. (2022) provides a strong pre-trained language model that the citing paper leverages in their research to improve the performance of their system.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(OpenAI, 2022)\", \"Explanation\": \"The cited work by OpenAI (2022) is the source of the large language models (LLMs) used in the citing paper to evaluate their performance on various NLP downstream tasks, including commonsense reasoning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Qin et al., 2023); Chan et al. (2023)\", \"Explanation\": \"The cited works by Qin et al. (2023) and Chan et al. (2023) have shown that LLMs, such as ChatGPT, can achieve competitive performance on commonsense reasoning tasks. The citing paper extends this research by benchmarking Chat-GPT's zero-shot performance on five QA evaluation benchmarks used in the zero-shot commonsense QA task.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Talmor et al., 2019); Sakaguchi et al. (2021); Fang et al. (2021b,a)\", \"Explanation\": \"The cited works by Talmor et al. (2019), Sakaguchi et al. (2021), and Fang et al. (2021b,a) are the sources of the commonsense reasoning tasks used in the citing paper to evaluate the performance of Chat-GPT in the zero-shot setting.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Talmor et al., 2019)\", \"Explanation\": \"The cited work by Talmor et al. (2019) is used to support the claim that ChatGPT performs well in the task of CommonsenseQA due to the high frequency of questions and answers in large text corpora.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Sap et al., 2019b)\", \"Explanation\": \"The cited work by Sap et al. (2019b) is used to support the claim that ChatGPT excels in the task of SocialIQA due to the high frequency of questions and answers in large text corpora.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Bai et al., 2023)\", \"Explanation\": \"The cited work by Bai et al. (2023) is mentioned as a potential reason for the suboptimal performance of ChatGPT on three benchmarks, indicating that the task requires more complex reasoning and implicit commonsense knowledge to solve.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ding et al., 2023)\", \"Explanation\": \"The cited work by Ding et al. (2023) is mentioned as a potential reason for the suboptimal performance of ChatGPT on three benchmarks, indicating that the task requires more complex reasoning and implicit commonsense knowledge to solve.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Wang et al., 2023a)\", \"Explanation\": \"The cited work by Wang et al. (2023a) has shown that integrating commonsense knowledge into generative models can improve their performance in commonsense reasoning tasks, which supports the claim made in the citing paper about the potential benefits of such knowledge in enhancing the performance of language models in general.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work provides the open-source code repository that the citing paper uses for synthesizing QA pairs and training the QA models, which serves as the methodological basis for the system built in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wolf et al., 2020)\", \"Explanation\": \"The cited work provides the pre-trained language models used in the citing paper, which are the data source for the system built in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Loshchilov and Hutter, 2019)\", \"Explanation\": \"The cited work is used to optimize the model in the citing paper, which extends the research on the optimization techniques for the model.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(West et al., 2022)\", \"Explanation\": \"The cited work, ATOMIC-10X, serves as the source of the knowledge base (CSKB D) used in the citing paper to train the model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work by Ma et al. (2021) provides the choice of the margin \\u03b7 for the marginal ranking loss, which the citing paper adopts in their model training process.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kim et al., 2022)\", \"Explanation\": \"The cited work by Kim et al. (2022) also contributes to the model training by setting the margin \\u03b7 to 1, which the citing paper follows in their model development.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(West et al., 2022)\", \"Explanation\": \"The cited work, West et al. (2022), is the source of the machine-generated corpus ATOMIC-10X, which the citing paper uses in their research on machine learning and knowledge distillation.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work is the original source of the QA synthesis pipeline used in the citing paper to process the knowledge from ATOMIC-10X.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work provides the method of lemmatizing tokens and using keywords to filter the dataset, which the citing paper adopts in their research on commonsense QA.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The citing paper extends the research of the cited work by training QA models on the filtered dataset and evaluating them on commonsense QA benchmarks in a zero-shot setting.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work provides the pipeline for training the model on the augmented ATOMIC dataset, which the citing paper uses to improve the performance of the model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Deng et al., 2023)\", \"Explanation\": \"The cited work by Deng et al. (2023) provides evidence of the presence of noise in ATOMIC-10X that is not beneficial to the task of zero-shot commonsense QA, which the citing paper uses to support the claim that the noise in ATOMIC-10X is a potential reason for the inconsistent performance in improving the model in zero-shot commonsense QA.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Swayamdipta et al., 2020)\", \"Explanation\": \"The cited work by Swayamdipta et al. (2020) provides the methodology of analyzing a model's behavior on individual instances during training on large datasets, which the citing paper adopts in their research to study the model's confidence in predicting true classes and variability across epochs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Swayamdipta et al., 2020)\", \"Explanation\": \"The cited work by Swayamdipta et al. (2020) provides a definition of model confidence that the citing paper re-defines in order to align with the model's degree of certainty in predicting instances as the true class in the context of a (Q i , A i ) pair.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Swayamdipta et al., 2020)\", \"Explanation\": \"The cited work establishes a definition of variability that the citing paper adopts in its research to measure the standard deviation of the score gap between the ground-truth answer and the distractors in a QA pair.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work by Ma et al. (2021) provides the foundational method for the development of the CAR framework, which is used in the citing paper to study the impact of utilizing conceptualization through various techniques.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhang et al., 2020)\", \"Explanation\": \"The cited work by Zhang et al. (2020) is used to sort questions in a benchmark by their average BERTScore, which is then used to split the questions into two sets based on their level of semantic overlap with the training set in ATOMIC. This information is then used to train two QA models in the citing paper, with one trained on conceptualization-augmented ATOMIC and the other on ATOMIC only.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ma et al., 2021)\", \"Explanation\": \"The cited work by Ma et al. serves as a reference for extending the framework to other CSKBs, including Concept-Net, WordNet, and WikiData, for a more generalizable evaluation.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Speer et al., 2017)\", \"Explanation\": \"The CWWV dataset is built upon the Concept-Net CSKB, which is cited to acknowledge the origin of the data used in the study.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Miller, 1995)\", \"Explanation\": \"The CWWV dataset is also based on WordNet, another CSKB, which is cited to highlight the data source for the study.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Vrandecic and Kr\\u00f6tzsch, 2014)\", \"Explanation\": \"The CWWV dataset is further built upon WikiData, another CSKB, which is cited to acknowledge the origin of the data used in the study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Radford et al., 2019)\", \"Explanation\": \"The GPT2 model is utilized in the study to train a conceptualization generator, which serves as a methodological basis for the research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(OpenAI, 2022)\", \"Explanation\": \"The use of ChatGPT as a generative conceptualizer in the study is also a methodological basis for the research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Table 10)\", \"Explanation\": \"The experimental results presented in Table 10 are a continuation of the study, showcasing a modest improvement in performance compared to baselines and comparable performance to GPT3.5.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Sap et al., 2019a)\", \"Explanation\": \"The cited work, ATOMIC, is a key component of the CSKG knowledge graph used in the citing paper for training a QA model.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Krishna et al., 2017)\", \"Explanation\": \"The VisualGenome dataset is used in the cited work to build the CWWV knowledge graph, which is then combined with other datasets in the CSKG knowledge graph in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Miller, 1995)\", \"Explanation\": \"The WordNet dataset is used in the cited work to build the CWWV knowledge graph, which is then combined with other datasets in the CSKG knowledge graph in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Speer et al., 2017)\", \"Explanation\": \"The ConceptNet dataset is used in the cited work to build the CWWV knowledge graph, which is then combined with other datasets in the CSKG knowledge graph in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ilievski et al., 2021)\", \"Explanation\": \"The CSKG knowledge graph is built using the cited work, which combines ATOMIC and CWWV datasets to form a new knowledge graph for training a QA model in the citing paper.\"}]"}}},{"rowIdx":87,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_0"],"heading":"Introduction","publication_ref":["b46","b66","b35","b33","b37","b61","b54","b4","b45","b44","b43","b17","b9"],"table_ref":[],"text":"Text clustering, as a fundamental task in natural language processing (NLP), has a wide spectrum of applications, such as identifying public perception from social media (Park et al., 2022), analysing cause of accidents (Xu et al., 2022), and detecting emerging research topics (Martínez et al., 2022). A common practice for text clustering is to apply clustering algorithms (MacQueen, 1967;Zhang et al., 2021a) on top of pre-trained embedders (Muennighoff et al., 2022;Wang et al., 2022;Su et al., 2022) which could achieve higher performance with better pre-training quality. State-of-the-art large language models (LLMs) such as recent GPT series (Brown et al., 2020;Ouyang et al., 2022;OpenAI, 2023) have demonstrated extraordinary language capabilities for various NLP applications however, these GPT models can only be utilized through the APIs without accessible embedding vectors for clustering. Hence, LLMs cannot be directly applied on text clustering tasks.\nIn this paper, we provide insights on the question: Can we leverage API-based LLMs to guide text clustering efficiently? We attack this challenging question by drawing inspiration from an observation that humans represent an instance through comparing with others (Nosofsky, 2011). For instance, people often classify a new piece of music into a specific genre by relating to familiar ones. In fact, pairwise relationships have been utilized in spectral clustering (Donath and Hoffman, 1972;Cheeger, 1970) before. Nonetheless, naively traversing all the pairs within dataset is obviously intractable and too expensive for querying LLMs.\nWe propose CLUSTERLLM, a framework that utilizes LLM to guide a small embedder for finding text clusters with a low cost, as shown in Figure 1. It comprises two stages that are specially designed for two aspects of clustering: (1) perspective, i.e., the grouping criterion such as topic, intent and emotion and (2) granularity, i.e. the scope of clusters.\nIn Stage 1, we prompt LLMs with a triplet task that predicts which one of the two candidate choices is closer to anchor instance to understand the user-preferred perspectives. We choose this triplet task because (a) it is irrelevant with cluster granularity and (b) the produced triplets can finetune small embedder towards the right perspective. In order to improve sample efficiency, we further propose entropy-based triplet sampling to find the most informative triplets. Specifically, we first calculate entropy for each instance based on cluster assignment probabilities, and then identify those with highest entropy. Two candidate choices are then sampled from its nearest clusters to guarantee they are close enough to the anchor.\nIn Stage 2, we first obtain the cluster hierarchy that starts from instance-level clusters and iteratively merge two closest clusters until the entire dataset. And then we prompt LLMs to determine cluster granularity with a few annotated data pairs as demonstrations. We construct the data pairs to prompt by sampling from two clusters that are merged at each step of hierarchical clustering, so that they cover a wide range of granularities. And the final decision is made by measuring consistency between each level of clustering and predictions.\nWe extensively evaluate CLUSTERLLM on 14 datasets that include diverse tasks such as intent discovery, topic mining, type discovery, domain discovery, and emotion detection. Furthermore, these datasets span a wide range of granularities that have 10 to 150 number of clusters. We show that CLUS-TERLLM is effective overall on improving clustering quality, where the clustering performance is improved over both a deep clustering baseline and a self-supervise baseline. Moreover, the ablation study shows that our sampling strategy is effective compared to a random sampling baseline. Finally, CLUSTERLLM also outperforms clustering-error based methods on determining cluster granularity.\nIn summary, our contributions are three-fold: (i) We propose a framework CLUSTERLLM that utilizes sentence relations predicted from API-based LLMs to guide clustering. Furthermore, it allows users to provide textual instructions and/or fewshot annotations to specify preferences on clustering. (ii) In order to reduce API-queries, we propose a novel entropy-based sampling strategy to find the most informative triplets. Additionally, we utilize pairwise data sampled from hierarchical clustering to determine cluster granularity. (iii) Extensive experiments show that our proposed method can improve clustering performance at ∼$0.2 for perspective and ∼$0.4 for granularity with GPT-3.5."},{"figure_ref":[],"heading":"Preliminary","publication_ref":[],"table_ref":[],"text":"Text clustering takes an unlabeled corpus D = {x i } N i=1 as input, and outputs a clustering assignment Y = {y i } N i=1 that maps the input text to cluster indices. To specify user's needs, CLUSTER-LLM integrates additional textual instruction (e.g. \"Select the example that better corresponds with the Query in terms of entity type.\") to understand perspective and few-shot annotations (e.g. \"Sentence1 and Sentence2 have the same entity type ...\") to determine cluster granularity."},{"figure_ref":["fig_1"],"heading":"Our CLUSTERLLM","publication_ref":["b61","b54","b43","b45","b44"],"table_ref":[],"text":"CLUSTERLLM is based on a pre-trained small embedder (Wang et al., 2022;Su et al., 2022) (denoted as f ) which usually represents sentences individually. In contrast, inspired by human cognitive ability (Nosofsky, 2011), CLUSTERLLM considers a pair or a triplet of sentences through prompting LLMs that are trained to follow human instructions (Ouyang et al., 2022;OpenAI, 2023). Specifically, CLUSTERLLM is a two-stage framework (See Figure 2). In Section 3.1 we introduce Stage 1 that utilizes triplet task to improve clustering quality with respect to user-specified perspectives, along with a sampling strategy that reduces number of API queries. In Section 3.2, we introduce Stage 2 that leverages pairwise task to determine cluster granularity based on predictions from LLMs."},{"figure_ref":[],"heading":"Triplet Task for Perspective","publication_ref":[],"table_ref":[],"text":"In this section, we explore how to harness a triplet task to refine the cluster structures for a userspecified perspective. A triplet task takes as input a tuple of three sentences t = (a, c 1 , c 2 ), where a is the anchor and (c 1 , c 2 ) are two choices. We then prompt LLMs to select one of (c 1 , c 2 ) that better corresponds with a using a prompt P T . Moreover, in order to specify the user's perspective, P T also requires a task instruction I T as input. The LLM should make a choice\nc j = P T (I T , t),(1)\nwhere c j ∈ {c 1 , c 2 } indicates one of the choices that LLM selects as positive and we denote the other (or negative) one as c \\j ."},{"figure_ref":[],"heading":"Entropy-based Triplet Sampling","publication_ref":[],"table_ref":[],"text":"While one can randomly sample triplets to query the LLM, we demonstrate it non-efficient in experiments. In this section, we pose the question of mining informative triplets to both save the costs from querying LLMs and optimally improve the clustering. To achieve this, we resort to the current clustering results from the extracted embeddings\nZ = {z i = f (x i )} N i=1 .\nIn summary, our algorithm contains two steps: Step 1: We find the most ambiguous instances as anchors based on entropy.\nStep 2: For each anchor instance, we sample two choices from two of its closest clusters. Refer to Algorithm 1 for entire process.\nIn Step 1, since the granularity is unknown at current stage, we perform clustering on top of Z, where the clustering hyperparameters 2 are consis- \n12 c1 ∼ C1, c2 ∼ C2, t = (a, c1, c2); 13 if t not in {tq},c1 ̸ = a, c2 ̸ = a then 14"},{"figure_ref":[],"heading":"Append t to {tq};","publication_ref":["b64","b57"],"table_ref":[],"text":"Output: A set of triplets {tq} Q q=1 tent across datasets and only specific to the embedder model f . Cluster center µ k will thereafter be calculated for cluster k by averaging embeddings assigned to it. Following (Xie et al., 2016;Van der Maaten and Hinton, 2008), we calculate instancewise soft assignments with Student's t-distribution,\np ik = (1 + ||z i -µ k || 2 /α) -α+1 2 k ′ (1 + ||z i -µ k ′ || 2 /α) -α+1 2 (2)\nwhere α = 1 is the degree of freedom. We then define closest clusters for instance i as K closest clusters with largest soft assignment p ik . Here, K closest is proportional to the total number of clusters K.\nK closest = max(ϵK, 2)(3)\nwhere we fix ϵ to be a small value, such as 2%.\nWe then compute entropy based on these closest clusters with renormalized probabilities p ′ ik ,\nh i = - K closest k=1 p ′ ik log(p ′ ik )(4)\nwhere\np ′ ik = p ik K closest k ′ =1 p ik ′\n. We sort the entire dataset in descending order according to the entropies H = {h i } N i=1 . We introduce two hyperparameters γ high and γ low that control the proportion interval to filter out from ordered dataset. Our hypothesis is that higher entropy (smaller γ high and γ low ) anchors form more informative triplets that we verify in Section 4.6. In Step 2, we randomly sample two clusters C 1 , C 2 from K closest closest clusters, and then sample two sentences c 1 , c 2 from each of them as choices (see line 11 and line 12). In another word, these choices should be either a positive or a hard negative to the anchor. Finally, we also remove triplets that are either repeated or have identical choice and anchor. We continue to sample triplets until reaching budget Q. Remarks. (1) Since Q is defined by the user and is independent with the dataset size, our sampling is cost-efficient. For example, in our experiments, using 1, 024 queries can improve performance on both dataset scales of ∼ 3, 000 and ∼ 50, 000. (2) From the view of ground truth, the sampled triplets might contain \"both are correct\" or \"none of the above\". However, we argue that even these triplets might provide soft aligning information, i.e. the ranking of closeness between choices. (3) Our sampling method may also be utilized in active learning to acquire human annotations when no prior knowledge is available on the categories."},{"figure_ref":[],"heading":"Finetuning Embedder","publication_ref":["b74","b34","b3","b54"],"table_ref":[],"text":"Now that we have the triplet predictions, it is still not clear how to utilize them in clustering. Previous research rely on deep constrained clustering (Zhang et al., 2020;Manduchi et al., 2021) which are often sensitive to noisy labels (Basu et al., 2008). In this paper, we instead focus on finetuning the base embedder f towards producing an embedding space that better explains the user's perspective. We exploit both hard and in-batch negatives. Following (Su et al., 2022;Ni et al., 2022b), for a triplet t = (a, c j , c \\j ) with positive c j and hard negative c \\j , we optimize the following objective,\nl j = exp (s(a, c j )/τ ) c l ∈B exp (s(a, c l )/τ )(5)\nwhere B combines c j , c \\j and other in-batch negatives. τ is a temperature parameter. Following the original implementation, we also compute the loss with a and c j swapped. Finally fine-tuned embedders can be applied to find even more informative triplets with our sampling method which will further improve performance in an iterative manner. We acquire clustering assignments by running clustering algorithms on top of extracted embeddings."},{"figure_ref":["fig_1"],"heading":"Pairwise Task for Granularity","publication_ref":[],"table_ref":["tab_13"],"text":"In this section, we build upon the refined embedding space in Section 3.1 to determine cluster granularity. In this paper, we convert the problem of determining granularity into finding the best step in a cluster hierarchy (see Figure 2 right), where each step denotes a unique granularity (or equally number of clusters). It is non-trivial, since different granularities can be applied to the same dataset (such as domains or topics). To tackle this challenge, we query LLM with pairwise task that predicts whether a pair of data p belong to the same cluster with a prompt P P ,\nw = P P (I P , {p d } D d=1 , p)(6)\nwhere w ∈ {same, different} is the binary decision, I P is the task instruction and {p d } D d=1 are few-shot demonstration pairs used for in-context learning (typically D = 4). We assume these demonstration pairs are annotated by users who have a desired cluster granularity in mind. We also combine a brief justification for each demonstration pair (see Table 12 bottom for example)."},{"figure_ref":[],"heading":"Determine Granularity with Pairwise Hierarchical Sampling","publication_ref":["b48"],"table_ref":[],"text":"We then introduce how to sample pairs from cluster hierarchy to query LLMs and determine granularity. We assume a maximum and a minimum number of clusters (denoted as k max and k min ) following Pelleg et al. (2000) which depend on the user's expectation on the granularity. We then randomly sample λ (1 or 3 in our experiments) pairs of data from the two clusters to be merged at each step to form candidate pairs {p i } Np i=1 , where N p = λ(k max -k min ). These pairs cover the entire range of granularity between k max and k min , and will be used to query LLMs. After that, each level of granularity can be examined against LLM predictions to choose the one with the highest consistency measure M,\nk * = argmax k M(W p , W k ) (7)\nwhere\nW p = {w p i } Np i=1\ndenotes the predictions obtained from Eq. 6 and W k represents a set of binary values indicating whether each pair of data is in the same cluster at granularity k. Empirically, we found the following performs better in our framework: use F-beta score, a weighted harmonic mean of precision and recall, as measurement M and set W p /W k as labels/predictions. Finally, for largescale datasets, we address the high time complexity of hierarchical clustering by applying it on top of mini-batch K-means. See details in Appendix A. Remarks. Similar to Section 3.1.1, pairwise hierarchical sampling can also be used to acquire human annotations. Nonetheless, the reliability of the algorithm still depends on the quality of clusters. In an extreme case where the clusters are completely random, it is unable to find granularity even though all the pairwise predictions are correct."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":"We first evaluate CLUSTERLLM on clustering quality with ground truth number of clusters in Section 4.4. Then we conduct ablation studies in Section 4.6 to further analyze the effectiveness of CLUSTERLLM. Finally, we show results of determining cluster granularity in Section 4.7."},{"figure_ref":[],"heading":"Datasets","publication_ref":["b75","b31","b37","b13"],"table_ref":["tab_1"],"text":"We provide a high-level summary of evaluated datasets in this section, and see Appendix E for more descriptions. In this paper, we evaluate on a broad range of clustering datasets with various perspectives and granularities. Furthermore, to better analyze the effect of scale, each dataset has both a small-scale and a large-scale version. The two versions are different in number of data while keeping the same number of clusters. A summary of dataset statistics is shown in Table 1. Note that there is no data splits in clustering. Intent (Domain) Discovery.\nIntent discovery (Zhang et al., 2021b(Zhang et al., , 2022) ) discovers unknown intents in unlabeled customer utterances. For CLINC, Massive and MTOP, we also use domains as labels to convert them into domain discovery. Type Discovery. Type Discovery (Li et al., 2022) resolves the closed-world set-up of traditional Information Extraction. In this work, we focus on three tasks: entity, relation and event type discovery. To indicate specific mentions (entities or event triggers), we directly append them behind sentences with natural language formats, such as \"The relation between [ENTITY1] and [ENTITY2]\". Topic Mining. We adapt three topic mining datasets from MTEB (Muennighoff et al., 2022). Emotion. We adapt GoEmo (Demszky et al., 2020), a fine-grained emotion detection dataset by removing multi-label or neutral instances."},{"figure_ref":[],"heading":"Experiment Details","publication_ref":["b47","b54","b61"],"table_ref":["tab_1"],"text":"Query LLMs. The prompt only contains a taskspecific instruction (see Table 11). We set generation temperature to 0.5. Explanations are suppressed by adding a postfix:\"Please respond with 'Choice 1' or 'Choice 2' without explanation\" and set up a max token of 10. We then assign them to binary choices by directly checking whether one of the texts \"Choice 1\" or \"Choice 2\" is in the response. We also find that a very small amount of responses do not contain any choices and we discard them during fine-tuning. We use the Python API tool provided by OpenAI.\nTriplet Sampling. For both small-or large-scale experiments, we set a budget of Q = 1, 024 triplets. We set γ low = 20% and γ high = 0. For clustering methods, we fix hyperparameters of these algorithms across datasets in Stage 1. We choose agglomerative clustering with fixed distance threshold 67 for small-scale experiments on Instructor, and 77 on E5 (the embeddings are preprocessed by standard scaler). For large-scale datasets, we choose mini-batch K-means with fixed number of clusters 100 due to its lower latency. Clustering algorithms are implemented by scikit-learn (Pedregosa et al., 2011).\nFine-tune Embedders. In this work, we focus on two state-of-the-art pre-trained embedders: Instructor (Su et al., 2022) and E5 (Wang et al., 2022). We only use the large versions. Refer to Appendix D for details.\nEvaluation. To reduce cost, we run CLUSTER-LLM once for each dataset. We then run (minibatch) K-means on (large) small-scale datasets for 5 seeds with ground truth K. We show two metrics. "},{"figure_ref":[],"heading":"Compared Methods","publication_ref":[],"table_ref":[],"text":"E5 and Instructor. We directly apply (minibatch) K-means on extracted embeddings from instructor-large and e5-large. self-supervise-I(E). To verify that the performance improvement of CLUSTERLLM does not only come from domain-specific fine-tuning, instead of the more accurate triplet prediction. We propose a self-supervise fine-tuning that uses exactly the same triplets as CLUSTERLLM but only switch to self-supervised triplet predictions that select closest choices in embedding space. SCCL-I(E). We also combine Instructor and E5 with SCCL (Zhang et al., 2021a), an unsupervised deep clustering algorithm that utilizes entire dataset for training. Notice that our method uses fewer data for training. See Appendix D for details."},{"figure_ref":[],"heading":"Main Results","publication_ref":[],"table_ref":["tab_2"],"text":"We show main results with small-scale datasets in Table 2. We show several variants of our method: CLUSTERLLM-I(E) adopt Instructor or E5 as embedders. CLUSTERLLM-I(E)-iter applies the entire framework twice in an iterative manner by using previous fine-tuned model as initialization and the 1, 024 triplets inferred from new embeddings for fine-tuning. All of these use GPT-3.5 for prediction. We make the following observations: (1) CLUSTERLLM consistently improves upon both embedders. For example, CLUSTER- LLM-I increases the performance by 6.71% on FewRel. CLUSTERLLM-E increases the performance by 9.19 on Bank77. However, we do observe that on Massive(D) and CLINC(D), there are no improvements.\n(2) CLUSTERLLM outperforms deep clustering and self-supervise baselines. For instance, CLUSTERLLM-I surpasses self-supervise-I on most datasets except for two and it is also better than SCCL-I on 11 over 14 datasets. Furthermore, these improvements are consistent across both reported metrics.\n(3) Combined with the results in Appendix F, applying CLUSTERLLM iteratively is beneficial, emphasizing the potential of further improvements."},{"figure_ref":[],"heading":"Analysis on Triplet Prediction Accuracy","publication_ref":[],"table_ref":["tab_4"],"text":"We attribute the improvements on clustering quality to more accurate triplet predictions. In Table 3, we show the accuracy on predicted triplets that have ground truth (exactly one positive and one negative choices based on ground truth) with two different sampling methods. Random triplet sampling uniformly samples three random instances as query and two choices, and we guarantee the two choices are different from the anchor by filtering. Furthermore, we also show a selection accuracy with Euclidean distances between embeddings as a comparison. We observe that, GPT-3.5/4 consistently improves upon Instructor on high entropy ex-amples, demonstrating our hypothesis. In contrast, with random sampling, the ground truth triplets is significantly fewer and the accuracy gap is much smaller or even decreases performance."},{"figure_ref":["fig_2"],"heading":"Ablation Study","publication_ref":[],"table_ref":["tab_5","tab_8"],"text":"Clustering Quality. We show ablation studies on CLUSTERLLM based on Instructor in Table 4. Specifically, we present results with 3 kinds of predictions on the same set of triplets for finetuning: GPT-3.5/4, replace triplet predictions of GPT-3.5 to ground truth on those triplets that have ground truth. We observe that GPT-4 marginally improves upon GPT-3.5 given the much higher cost. When provided with human labels, CLUS-TERLLM-GT&GPT3.5 achieves the highest performance, which indicates the possibility for further improvement with more accurate predictions. We make similar observations for large-scale datasets in Table 6. Sampling Strategy. In this section, we show ablation study on entropy-based sampling. In Figure 3, we observe that clustering accuracy increases when increasing entropies (or equally decreasing mean of interval) except for GoEmo. We make two hypothesis: (1) LLMs are much better than small embedders on harder instances. ( 2 x-axis shows the mean of interval where interval length is set to 20%. For example, \"mean of interval= 50%\" means γ high = 40% and γ low = 60% (see Section 3.1.1). ♦ marks the setting for main experiments.\nlected triplets even decreases performance, which demonstrates the cruciality of triplet sampling."},{"figure_ref":[],"heading":"Determining Cluster Granularity","publication_ref":[],"table_ref":["tab_6","tab_9","tab_6","tab_6"],"text":"In this section, we show the results for determining cluster granularity. We evaluate on a subset of 8 datasets including various cluster granularities with k max = 200 and k min = 2. We compare with different methods that rely on clustering errors. For our methods, we show results with λ = {1, 3} (except for GPT-4 to reduce costs), which involve\n198 & 594 pairs in total respectively. To simulate experts for providing demonstrations, we directly sample 16 pairs from small-scale datasets when λ = 3 and then choose 2 positive and 2 negative as demonstrations. Notice that we use the same demonstrations for large-scale experiments. See more details in Appendix B. We make several observations from Table 5 andTable 7: (1) Our methods have higher ranks. Most baseline methods predict similar number of clusters for domain and intent, while our methods can effectively distinguish between the two. For instance, on MTOP(I)/(D) in Table 5, BIC predicts number of clusters 69/64 while our method (GPT-3.5, λ = 3) predicts 92/18. (2) Increasing λ generally helps (MTOP(D) in Table 5) but might not always make a large difference. (3) GPT-4 significantly improves upon GPT-3.5, probably due to its better understanding of demonstrations."},{"figure_ref":[],"heading":"Related Works","publication_ref":["b67","b24","b68","b69","b6","b42","b64","b27","b11","b6","b58","b41","b8","b42","b15","b64","b24","b9","b69","b42","b58","b8","b2","b60","b3","b74","b34","b50","b20","b61","b54","b37"],"table_ref":[],"text":"Clustering. As a fundamental task in machine learning, clustering has been applied on diverse data types, including texts (Xu et al., 2015;Hadifar et al., 2019;Zhang et al., 2021a), images (Yaling Tao, 2021;Yang et al., 2016;Caron et al., 2018;Niu et al., 2020;Xie et al., 2016) and graphs (Huang et al., 2014;Chiang et al., 2019). Recent research has been shifted to deep cluster- et al., 2016;Caron et al., 2018;Van Gansbeke et al., 2020;Niu et al., 2022;Chang et al., 2017;Niu et al., 2020), generative (Dilokthanakul et al., 2016) and simultaneous (Xie et al., 2016;Zhang et al., 2021a;Hadifar et al., 2019) man, 1972;Cheeger, 1970) makes use of similarity matrix where each entry measures the similarity between a pair of data. More recently, several works in deep clustering utilize relational supervision (Yang et al., 2016;Niu et al., 2020;Van Gansbeke et al., 2020;Chang et al., 2017) via pseudolabelling which could be noisy. Another line of works that is closely related to ours is constrained clustering. It usually incorporates pairwise mustlink or cannot-link constraints (Basu et al., 2004;Wagstaff et al., 2001;Basu et al., 2008;Zhang et al., 2020;Manduchi et al., 2021). Nonetheless, these constraints are often sampled from labels as a prior which significantly limits its application in our scenario. In this work, we study how to utilize contemporary API-based LLMs to infer relations.\nPre-trained Embedding Model. Generic pre-trained text embedding models (Reimers and Gurevych, 2019;Gao et al., 2021;Ni et al., 2022a,b) are widely applied in text similarity, classification, clustering and information retrieval. Recently, two embedding models E5 (Wang et al., 2022) and Instructor (Su et al., 2022) have shown superior performance on a popular benchmark (Muennighoff et al., 2022). Specifically E5 is pre-trained on web-scraped data pairs with contrastive objective. Instructor is pre-trained on 330 tasks with instructions. CLUSTERLLM aims at improving these models with LLMs."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"In this paper, we study how to leverage API-based LLMs to guide small embedders for text clustering in order to benefit from high-level language capability of LLMs and user's instructions on clustering. We propose to prompt LLMs with two kinds of sentence relationship tasks: triplet task and pairwise task. Triplet task chooses the sentence that is most similar with anchor combining with a perspective instruction from users. The predicted triplets are used for fine-tuning small embedders.\nPairwise task judges whether a pair of sentences belong to the same category hinted by few-shot demonstrations, and then the predictions are used to determine cluster granularity with a consistency measure. Extensive experiments show that our proposed framework CLUSTERLLM can improve clustering quality and propose reasonable cluster granularity at a negligible cost. However, CLUS-TERLLM still relies on the embedding model itself, which is inefficient and inapplicable on black-box embedding models. We encourage future works to explore the potential of model-free training such as constrained clustering."},{"figure_ref":[],"heading":"A Details of Scaling up Hierarchical Clustering","publication_ref":["b63"],"table_ref":[],"text":"A major drawback of hierarchical clustering is its O(N3 ) time complexity which makes the algorithm hard to be deployed on larger datasets. However, since we are only interested in a specific range of granularity in our scenario, the hierarchical clustering can start from an intermediate step. We address this issue by first running minibatch K-means with k max and then run hierarchical clustering with current assignments as inputs. Specifically, we use agglomerative clustering with ward's method (Ward Jr, 1963). We first calculate distances between each pair of clusters according to Murtagh and Contreras, 2011 and then provide them as inputs to nearest neighbor chain algorithm.\nFinally the returned hierarchy is combined with the K-means assignments to infer clusters."},{"figure_ref":[],"heading":"B More Details about Determining Cluster Granularity","publication_ref":["b53","b56","b48","b23","b23"],"table_ref":["tab_2","tab_13"],"text":"Previous methods often employ clustering errors as a metric and they ignore user's need on the granularity. Silhouette coefficient (Rousseeuw, 1987) indicates the clustering quality without ground truths, which exploits the inter-cluster distance with nearest clusters and the intra-cluster distance. We find the granularity by choosing the one with the best silhouette coefficient. Elbow method (Thorndike, 1953) is a heuristic method that plots the clustering error with respect to different levels of granularity in the hierarchy. And then the best granularity is determined with the largest elbow length. X-means (Pelleg et al., 2000) is a variation of K-means that starts with the lowest number of clusters, and then repeatedly attempt to split the clusters by running 2-means on them and evaluate with Bayesian Information Criterion (BIC) (Goutte et al., 2001). BIC (Goutte et al., 2001) calculates BIC for each of the granularity. Cluster-Size (Zhang et al., 2021b) uses a confidence threshold to filter small clusters starting from the maximum number of cluster. For all methods, we use the same fine-tuned embeddings (CLUSTERLLM-I in Table 2). The same cluster hierarchy is used (except for X-means that relies on K-means), which is either acquired from hierarchical clustering for small-scale or our proposed two-step method in Section 3.2 for large-scale. For out methods, the weight in F-beta score is set to 0.92 through empirical selection on Bank77. Because of the high latency, results for Silhouette and X-means are not shown on large-scale datasets. After sampling 16 data pairs, we tend to choose positives with finer granularity or negatives with coarser granularity. We also consider the sentence length to minimize the cost. We use label names as justifications and we always put 2 positive before 2 negative (See Table 12 bottom)."},{"figure_ref":[],"heading":"C Analysis for Determining Granularity","publication_ref":[],"table_ref":[],"text":"Prompt Design. We show the analysis results of prompt design for determining granularity in Table 8. We experiment with two settings: (1) remove justification for all demonstrations and only keep the \"Yes\" or \"No\" answer.\n(2) remove all demonstrations and any granularity-related words (such as domain) 3 . We observe that demonstrations are necessary and adding justifications have a positive impact.\nVisualization of Consistency Score. We visualize consistency score with respect to the number of clusters. The consistency scores exhibit continuous variations and peak at the best number of clusters."},{"figure_ref":[],"heading":"D Details of Embedders and Fine-tuning","publication_ref":["b54"],"table_ref":["tab_12"],"text":"For all the experiments (including those with or without fine-tuning), we use large version of both Instructor and E5 (i.e.\nhkunlp/instructor-large & intfloat/e5-large).\nFor Instructor, we use the same or similar prompt as original paper. See Table 10.\nFor fine-tuning, we adopt the same hyperparameters as in (Su et al., 2022), but modify the learning rate to 2e -6, the maximum gradient steps to 3, 840 for Instructor (∼ 15 epochs) and 1, 280 for E5, and batch size to 4. We choose this gradient in the begining of our experiments by observing no performance increase after that on several datasets. Training is conducted with a single NVIDIA Quadro RTX 8000 GPU.\nFor SCCL-I(E), we change the maximum token length to 128 due to the limited compute resource 4 . We use the same learning rate 2e -6 as before for Instructor and 2e -7 for E5 since we found that the performance is unstable with large learning rate. Batch size is set to 16 and we evaluate representations with K-means after 200 iterations. Also notice that we do not interrupt prompts in Instructor during data augmentation."},{"figure_ref":[],"heading":"E Description of Datasets","publication_ref":["b7","b29","b18","b30","b37","b19","b14","b16","b21"],"table_ref":["tab_11"],"text":"Bank77 (Casanueva et al., 2020) is a popular dataset in intent discovery that focuses on creating fine-grained intent categories for a single-domain, \"banking\". CLINC(I) (Larson et al., 2019) is originally created for detecting utterances that falls outside of supported intents. The dataset also contains multiple domains, such as \"travel\", \"utility\" and \"work\". In this experiment, we discard all the out-of-scope utterances and only focus on indomain ones. Moreover, we create a domain discovery dataset CLINC(D) that uses domains as labels. Massive(I)/(D) (FitzGerald et al., 2022) and MTOP(I)/(D) (Li et al., 2021) are both from MTEB (Muennighoff et al., 2022). Here \"I\" denotes intent and \"D\" for domain (or scenario). These datasets are originally used for classification but are adapted here for clustering. We also remove those intents with only a few instances and keep English-only data. For all datasets, we use the train & test sets as large-& small-scale datasets respectively. For FewRel (Gao et al., 2019) and FewEvent (Deng et al., 2020), we first randomly split datasets into train & test sets, and then sample from train set as large-scale and test set as smallscale. For FewNerd (Ding et al., 2021), we use the original train & test splits. For StackEx, Reddit (Geigle et al., 2021) and ArxivS2S, we combine all the splits into a single dataset and remove topics that only have few instances. Finally, the datasets are randomly splitted into large-& small-scale versions. To show the dataset balancy, we show the 4 We also tried maximum token length 512 with smaller batch size, but the performance is not better. entropy of class distribution in Table 9."},{"figure_ref":["fig_4"],"heading":"F Results of More Iterations","publication_ref":[],"table_ref":[],"text":"We show the results over 4 iterations of CLUS-TERLLM in Figure 5. During iteration, we sample triplets from previously fine-tuned embedding space and continue to fine-tune the model with previous checkpoint as initialization. We also show the self-supervise results using the same checkpoint fine-tuned with GPT-3.5 predictions as initialization at each iteration. We observe that using GPT-3.5 predictions is almost always beneficial. The performance generally increases and saturate at the fourth iteration with the exception of GoEmo."},{"figure_ref":[],"heading":"G More Related Works","publication_ref":["b59","b32","b75","b36","b0","b70","b5","b55","b75","b73","b0","b32","b22","b25","b77","b10","b1"],"table_ref":[],"text":"Generalized Category Discovery (GCD). GCD (Vaze et al., 2022;Lin et al., 2020;Zhang et al., 2021bZhang et al., , 2022;;Mou et al., 2022;An et al., 2022) assume partial known classes with annotations which can also be used to infer user's requirement on clustering. As an infant research area, most previous works employ pseudo-labelling, via optimal transport (Rizve et al., 2022b;Yang et al., 2022), similarity learning (Rizve et al., 2022a;Cao et al., 2022) or prototype-based learning (Sun and Li, 2022). Furthermore, new intent discovery (Zhang et al., 2022(Zhang et al., , 2021b(Zhang et al., , 2023;;An et al., 2022;Lin et al., 2020) Represent the bank purpose for retrieval: CLINC(I) Represent the sentence for retrieving the purpose: FewRel Represent the relation between two entities for retrieval: FewNerd Represent the entity type for retrieval: FewEvent Represent the event type for retrieval: StackEx Represent the question for retrieval: ArxivS2S\nRepresent the science statement for retrieval: GoEmo Represent an emotion sentence for retrieval: Massive(I) Represent the sentence for retrieving the purpose: MTOP(I) Represent the sentence for retrieval: Reddit represent a reddit community title: Massive(D) Represent the scene for retrieval: MTOP(D) Represent a sentence: CLINC(D) Represent a sentence: computation-& data-cost. LLMs as Annotators. Recent instruction-tuned LLMs, such as ChatGPT, have been shown to have the ability to reproduce or improve humangenerated labels (Gilardi et al., 2023;He et al., 2023;Zhu et al., 2023). Furthermore, several works dedicate to fine-tune models with feedbacks from LLMs (Cheng et al., 2023;Bai et al., 2022). This paper instead focuses on clustering tasks."},{"figure_ref":[],"heading":"H Sub-optimal Performance on Domain Discovery","publication_ref":[],"table_ref":["tab_2","tab_5","tab_8","tab_2"],"text":"We noticed that the performance of domain discovery (MTOP(D), Massive(D) and CLINC(D)) is barely improved or even decreased with CLUS-TERLLM from original embedders (see Table 2). Furthermore, the ablation studies reveal that even with CLUSTERLLM-GT&GPT3.5, clustering performance is not as good as self-supervise or CLUS-TERLLM-random (see CLINC(D) in Table 4 and MTOP(D) in Table 6). We also observe that, CLUS-TERLLM-I-iter will further decrease the performance (see Massive(D) in Table 2). While we do not have rigorous explanations, one hypothesize is that the embedding space after fine-tuning tends to be more compact than before and forming small cliques, making it better for clustering finegrained clusters but not for coarse-grained clusters. We showcase scatterplots on two datasets with both Instructor and CLUSTERLLM-I-iter. It can be observed that the clusters in embedding space are tighter and more separated especially on CLINC(D)."},{"figure_ref":[],"heading":"I Dataset Leakage","publication_ref":[],"table_ref":["tab_4"],"text":"Since LLMs like ChatGPT are trained on webscraped texts from internet, it is likely they already have access to our evaluation datasets during training. For example, the topic mining datasets use StackExchange, Reddit and Arxiv tags as labels which is freely available online. However, as observed in Table 3, the performance of triplet prediction on these datasets are often far from perfect. Furthermore, the other datasets like Bank77 are synthesized datasets which is not accessible during training. FewRel is collected from Wikipedia corpus but their labels are not easily accessible. Similarly, GoEmo is collected from Reddit but the emotion labels are not accessible during training. Thus, dataset leakage is not a primary concern of this paper."},{"figure_ref":[],"heading":"Dataset Prompt Bank77","publication_ref":[],"table_ref":[],"text":"Select the banking customer utterance that better corresponds with the Query in terms of intent. CLINC(I) Select the customer utterance that better corresponds with the Query in terms of intent."},{"figure_ref":[],"heading":"FewRel","publication_ref":[],"table_ref":[],"text":"Select the example that better corresponds with the Query in terms of relation type."},{"figure_ref":[],"heading":"FewNerd","publication_ref":[],"table_ref":[],"text":"Select the example that better corresponds with the Query in terms of entity type. FewEvent Select the example that better corresponds with the Query in terms of event type."},{"figure_ref":[],"heading":"StackEx","publication_ref":[],"table_ref":[],"text":"Select the StackExchange question that better corresponds with the Query in terms of topic."},{"figure_ref":[],"heading":"ArxivS2S","publication_ref":[],"table_ref":[],"text":"Select the Arxiv paper title that better corresponds with the Query in terms of domain."},{"figure_ref":[],"heading":"GoEmo","publication_ref":[],"table_ref":[],"text":"Select the sentence that better corresponds with the Query in terms of emotion expressed. Massive(I) Select the user utterance that better corresponds with the Query in terms of intent."},{"figure_ref":[],"heading":"MTOP(I)","publication_ref":[],"table_ref":["tab_1"],"text":"Select the user utterance that better corresponds with the Query in terms of intent Reddit Select the Reddit question that better corresponds with the Query in terms of topic. Massive(D) Select the user utterance that better corresponds with the Query in terms of scenario. MTOP(D) Select the user utterance that better corresponds with the Query in terms of domain. CLINC(D) Select the customer utterance that better corresponds with the Query in terms of domain.\nTable 11: Prefix of prompts for triplet task. Notice while we use different prompts for domain and intent (such as CLINC(I) and CLINC(D)) in our experiments, they might be used interchangeably."},{"figure_ref":[],"heading":"Dataset Prompt","publication_ref":[],"table_ref":[],"text":"Triplet Task\nSelect the banking customer utterance that better corresponds with the Query in terms of intent.\nQuery: Should i reinstall the payment app? Choice 1: I've received my card so now I need to know how to sync it to the app. Choice 2: Can I still use the app if I switched phones?\nPlease respond with 'Choice 1' or 'Choice 2' without explanation.\nPairwise Task\n[Example1] Sentence 1: I would like to see the source of my money.\nSentence 2: My source of funds need verified. Yes. Because both intents are verify source of funds.\n[Example2] Sentence 1: Is there a fee for topping up Sentence 2: What are the top up charges for US cards? Yes. Because both intents are top up by card charge.\n[Example3] Sentence 1: Can I reactivate my lost card that I found this morning in my jacket pocket? Sentence 2: how to activate card? No. Because Sentence 1 has intent card linking and Sentence 2 has intent activate my card.\n[Example4] Sentence 1: What will I be charged for a physical card? Sentence 2: My card is about to expire and I need to know how much it costs and how long ... No. Because Sentence 1 has intent order physical card and Sentence 2 has intent card ... Determine whether the intents of two banking customer utterances below belong to the same intent category using above examples.\nSentence 1: $1 extra has been charged on my statement, why is that? Sentence 2: Will it automatically top-up if there isn't much money left?\nPlease respond with 'Yes' or 'No' without explanation. "},{"figure_ref":[],"heading":"Acknowledgements","publication_ref":[],"table_ref":[],"text":"Our work is sponsored in part by NSF CAREER Award 2239440, NSF Proto-OKN Award 2333790, NIH Bridge2AI Center Program under award 1U54HG012510-01, Cisco-UCSD Sponsored Research Project, as well as generous gifts from Google, Adobe, and Teradata. Any opinions, findings, and conclusions or recommendations expressed herein are those of the authors and should not be interpreted as necessarily representing the views, either expressed or implied, of the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for government purposes not withstanding any copyright annotation hereon."},{"figure_ref":[],"heading":"Limitations","publication_ref":[],"table_ref":[],"text":"We list several limitations of our work that we hope to be improved in the future: Reliance on pre-trained embedder. To find the most informative data, we have to rely on a pretrained embedder that can indicate the largest clustering assignment entropy. We hope that selfsupervise triplets and LLM-predicted triplets can be combined to solve such an issue. Computational cost for fine-tuning. Our initial idea is to utilize constrained clustering which is a light-weight algorithm that do not need to update small embedders. However, the inevitable unstable training will be heavily affected by the errors in LLM predictions. We make a comprise by introducing embedder into fine-tuning to temporarily solve the issue, but we hope to reduce the computational cost in a future work. Sub-optimal performance on domain discovery. We notice that on domain discovery datasets such as Massive(D) and CLINC(D), the performance is usually sub-optimal compared with original Instructor embedding. We provide discussions on this issue in Appendix H."},{"figure_ref":[],"heading":"Ethics Statement","publication_ref":[],"table_ref":[],"text":"Our work employs LLMs which are accessed through OpenAI APIs. For some applications, uploading privacy-sensitive data is risky and might require efforts to remove sensitive information."}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b46\",\n \"b66\",\n \"b35\",\n \"b33\",\n \"b37\",\n \"b61\",\n \"b54\",\n \"b4\",\n \"b45\",\n \"b44\",\n \"b43\",\n \"b17\",\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"Text clustering, as a fundamental task in natural language processing (NLP), has a wide spectrum of applications, such as identifying public perception from social media (Park et al., 2022), analysing cause of accidents (Xu et al., 2022), and detecting emerging research topics (Martínez et al., 2022). A common practice for text clustering is to apply clustering algorithms (MacQueen, 1967;Zhang et al., 2021a) on top of pre-trained embedders (Muennighoff et al., 2022;Wang et al., 2022;Su et al., 2022) which could achieve higher performance with better pre-training quality. State-of-the-art large language models (LLMs) such as recent GPT series (Brown et al., 2020;Ouyang et al., 2022;OpenAI, 2023) have demonstrated extraordinary language capabilities for various NLP applications however, these GPT models can only be utilized through the APIs without accessible embedding vectors for clustering. Hence, LLMs cannot be directly applied on text clustering tasks.\\nIn this paper, we provide insights on the question: Can we leverage API-based LLMs to guide text clustering efficiently? We attack this challenging question by drawing inspiration from an observation that humans represent an instance through comparing with others (Nosofsky, 2011). For instance, people often classify a new piece of music into a specific genre by relating to familiar ones. In fact, pairwise relationships have been utilized in spectral clustering (Donath and Hoffman, 1972;Cheeger, 1970) before. Nonetheless, naively traversing all the pairs within dataset is obviously intractable and too expensive for querying LLMs.\\nWe propose CLUSTERLLM, a framework that utilizes LLM to guide a small embedder for finding text clusters with a low cost, as shown in Figure 1. It comprises two stages that are specially designed for two aspects of clustering: (1) perspective, i.e., the grouping criterion such as topic, intent and emotion and (2) granularity, i.e. the scope of clusters.\\nIn Stage 1, we prompt LLMs with a triplet task that predicts which one of the two candidate choices is closer to anchor instance to understand the user-preferred perspectives. We choose this triplet task because (a) it is irrelevant with cluster granularity and (b) the produced triplets can finetune small embedder towards the right perspective. In order to improve sample efficiency, we further propose entropy-based triplet sampling to find the most informative triplets. Specifically, we first calculate entropy for each instance based on cluster assignment probabilities, and then identify those with highest entropy. Two candidate choices are then sampled from its nearest clusters to guarantee they are close enough to the anchor.\\nIn Stage 2, we first obtain the cluster hierarchy that starts from instance-level clusters and iteratively merge two closest clusters until the entire dataset. And then we prompt LLMs to determine cluster granularity with a few annotated data pairs as demonstrations. We construct the data pairs to prompt by sampling from two clusters that are merged at each step of hierarchical clustering, so that they cover a wide range of granularities. And the final decision is made by measuring consistency between each level of clustering and predictions.\\nWe extensively evaluate CLUSTERLLM on 14 datasets that include diverse tasks such as intent discovery, topic mining, type discovery, domain discovery, and emotion detection. Furthermore, these datasets span a wide range of granularities that have 10 to 150 number of clusters. We show that CLUS-TERLLM is effective overall on improving clustering quality, where the clustering performance is improved over both a deep clustering baseline and a self-supervise baseline. Moreover, the ablation study shows that our sampling strategy is effective compared to a random sampling baseline. Finally, CLUSTERLLM also outperforms clustering-error based methods on determining cluster granularity.\\nIn summary, our contributions are three-fold: (i) We propose a framework CLUSTERLLM that utilizes sentence relations predicted from API-based LLMs to guide clustering. Furthermore, it allows users to provide textual instructions and/or fewshot annotations to specify preferences on clustering. (ii) In order to reduce API-queries, we propose a novel entropy-based sampling strategy to find the most informative triplets. Additionally, we utilize pairwise data sampled from hierarchical clustering to determine cluster granularity. (iii) Extensive experiments show that our proposed method can improve clustering performance at ∼$0.2 for perspective and ∼$0.4 for granularity with GPT-3.5.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Preliminary\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Text clustering takes an unlabeled corpus D = {x i } N i=1 as input, and outputs a clustering assignment Y = {y i } N i=1 that maps the input text to cluster indices. To specify user's needs, CLUSTER-LLM integrates additional textual instruction (e.g. \\\"Select the example that better corresponds with the Query in terms of entity type.\\\") to understand perspective and few-shot annotations (e.g. \\\"Sentence1 and Sentence2 have the same entity type ...\\\") to determine cluster granularity.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Our CLUSTERLLM\",\n \"publication_ref\": [\n \"b61\",\n \"b54\",\n \"b43\",\n \"b45\",\n \"b44\"\n ],\n \"table_ref\": [],\n \"text\": \"CLUSTERLLM is based on a pre-trained small embedder (Wang et al., 2022;Su et al., 2022) (denoted as f ) which usually represents sentences individually. In contrast, inspired by human cognitive ability (Nosofsky, 2011), CLUSTERLLM considers a pair or a triplet of sentences through prompting LLMs that are trained to follow human instructions (Ouyang et al., 2022;OpenAI, 2023). Specifically, CLUSTERLLM is a two-stage framework (See Figure 2). In Section 3.1 we introduce Stage 1 that utilizes triplet task to improve clustering quality with respect to user-specified perspectives, along with a sampling strategy that reduces number of API queries. In Section 3.2, we introduce Stage 2 that leverages pairwise task to determine cluster granularity based on predictions from LLMs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Triplet Task for Perspective\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we explore how to harness a triplet task to refine the cluster structures for a userspecified perspective. A triplet task takes as input a tuple of three sentences t = (a, c 1 , c 2 ), where a is the anchor and (c 1 , c 2 ) are two choices. We then prompt LLMs to select one of (c 1 , c 2 ) that better corresponds with a using a prompt P T . Moreover, in order to specify the user's perspective, P T also requires a task instruction I T as input. The LLM should make a choice\\nc j = P T (I T , t),(1)\\nwhere c j ∈ {c 1 , c 2 } indicates one of the choices that LLM selects as positive and we denote the other (or negative) one as c \\\\j .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Entropy-based Triplet Sampling\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"While one can randomly sample triplets to query the LLM, we demonstrate it non-efficient in experiments. In this section, we pose the question of mining informative triplets to both save the costs from querying LLMs and optimally improve the clustering. To achieve this, we resort to the current clustering results from the extracted embeddings\\nZ = {z i = f (x i )} N i=1 .\\nIn summary, our algorithm contains two steps: Step 1: We find the most ambiguous instances as anchors based on entropy.\\nStep 2: For each anchor instance, we sample two choices from two of its closest clusters. Refer to Algorithm 1 for entire process.\\nIn Step 1, since the granularity is unknown at current stage, we perform clustering on top of Z, where the clustering hyperparameters 2 are consis- \\n12 c1 ∼ C1, c2 ∼ C2, t = (a, c1, c2); 13 if t not in {tq},c1 ̸ = a, c2 ̸ = a then 14\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Append t to {tq};\",\n \"publication_ref\": [\n \"b64\",\n \"b57\"\n ],\n \"table_ref\": [],\n \"text\": \"Output: A set of triplets {tq} Q q=1 tent across datasets and only specific to the embedder model f . Cluster center µ k will thereafter be calculated for cluster k by averaging embeddings assigned to it. Following (Xie et al., 2016;Van der Maaten and Hinton, 2008), we calculate instancewise soft assignments with Student's t-distribution,\\np ik = (1 + ||z i -µ k || 2 /α) -α+1 2 k ′ (1 + ||z i -µ k ′ || 2 /α) -α+1 2 (2)\\nwhere α = 1 is the degree of freedom. We then define closest clusters for instance i as K closest clusters with largest soft assignment p ik . Here, K closest is proportional to the total number of clusters K.\\nK closest = max(ϵK, 2)(3)\\nwhere we fix ϵ to be a small value, such as 2%.\\nWe then compute entropy based on these closest clusters with renormalized probabilities p ′ ik ,\\nh i = - K closest k=1 p ′ ik log(p ′ ik )(4)\\nwhere\\np ′ ik = p ik K closest k ′ =1 p ik ′\\n. We sort the entire dataset in descending order according to the entropies H = {h i } N i=1 . We introduce two hyperparameters γ high and γ low that control the proportion interval to filter out from ordered dataset. Our hypothesis is that higher entropy (smaller γ high and γ low ) anchors form more informative triplets that we verify in Section 4.6. In Step 2, we randomly sample two clusters C 1 , C 2 from K closest closest clusters, and then sample two sentences c 1 , c 2 from each of them as choices (see line 11 and line 12). In another word, these choices should be either a positive or a hard negative to the anchor. Finally, we also remove triplets that are either repeated or have identical choice and anchor. We continue to sample triplets until reaching budget Q. Remarks. (1) Since Q is defined by the user and is independent with the dataset size, our sampling is cost-efficient. For example, in our experiments, using 1, 024 queries can improve performance on both dataset scales of ∼ 3, 000 and ∼ 50, 000. (2) From the view of ground truth, the sampled triplets might contain \\\"both are correct\\\" or \\\"none of the above\\\". However, we argue that even these triplets might provide soft aligning information, i.e. the ranking of closeness between choices. (3) Our sampling method may also be utilized in active learning to acquire human annotations when no prior knowledge is available on the categories.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Finetuning Embedder\",\n \"publication_ref\": [\n \"b74\",\n \"b34\",\n \"b3\",\n \"b54\"\n ],\n \"table_ref\": [],\n \"text\": \"Now that we have the triplet predictions, it is still not clear how to utilize them in clustering. Previous research rely on deep constrained clustering (Zhang et al., 2020;Manduchi et al., 2021) which are often sensitive to noisy labels (Basu et al., 2008). In this paper, we instead focus on finetuning the base embedder f towards producing an embedding space that better explains the user's perspective. We exploit both hard and in-batch negatives. Following (Su et al., 2022;Ni et al., 2022b), for a triplet t = (a, c j , c \\\\j ) with positive c j and hard negative c \\\\j , we optimize the following objective,\\nl j = exp (s(a, c j )/τ ) c l ∈B exp (s(a, c l )/τ )(5)\\nwhere B combines c j , c \\\\j and other in-batch negatives. τ is a temperature parameter. Following the original implementation, we also compute the loss with a and c j swapped. Finally fine-tuned embedders can be applied to find even more informative triplets with our sampling method which will further improve performance in an iterative manner. We acquire clustering assignments by running clustering algorithms on top of extracted embeddings.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Pairwise Task for Granularity\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_13\"\n ],\n \"text\": \"In this section, we build upon the refined embedding space in Section 3.1 to determine cluster granularity. In this paper, we convert the problem of determining granularity into finding the best step in a cluster hierarchy (see Figure 2 right), where each step denotes a unique granularity (or equally number of clusters). It is non-trivial, since different granularities can be applied to the same dataset (such as domains or topics). To tackle this challenge, we query LLM with pairwise task that predicts whether a pair of data p belong to the same cluster with a prompt P P ,\\nw = P P (I P , {p d } D d=1 , p)(6)\\nwhere w ∈ {same, different} is the binary decision, I P is the task instruction and {p d } D d=1 are few-shot demonstration pairs used for in-context learning (typically D = 4). We assume these demonstration pairs are annotated by users who have a desired cluster granularity in mind. We also combine a brief justification for each demonstration pair (see Table 12 bottom for example).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Determine Granularity with Pairwise Hierarchical Sampling\",\n \"publication_ref\": [\n \"b48\"\n ],\n \"table_ref\": [],\n \"text\": \"We then introduce how to sample pairs from cluster hierarchy to query LLMs and determine granularity. We assume a maximum and a minimum number of clusters (denoted as k max and k min ) following Pelleg et al. (2000) which depend on the user's expectation on the granularity. We then randomly sample λ (1 or 3 in our experiments) pairs of data from the two clusters to be merged at each step to form candidate pairs {p i } Np i=1 , where N p = λ(k max -k min ). These pairs cover the entire range of granularity between k max and k min , and will be used to query LLMs. After that, each level of granularity can be examined against LLM predictions to choose the one with the highest consistency measure M,\\nk * = argmax k M(W p , W k ) (7)\\nwhere\\nW p = {w p i } Np i=1\\ndenotes the predictions obtained from Eq. 6 and W k represents a set of binary values indicating whether each pair of data is in the same cluster at granularity k. Empirically, we found the following performs better in our framework: use F-beta score, a weighted harmonic mean of precision and recall, as measurement M and set W p /W k as labels/predictions. Finally, for largescale datasets, we address the high time complexity of hierarchical clustering by applying it on top of mini-batch K-means. See details in Appendix A. Remarks. Similar to Section 3.1.1, pairwise hierarchical sampling can also be used to acquire human annotations. Nonetheless, the reliability of the algorithm still depends on the quality of clusters. In an extreme case where the clusters are completely random, it is unable to find granularity even though all the pairwise predictions are correct.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We first evaluate CLUSTERLLM on clustering quality with ground truth number of clusters in Section 4.4. Then we conduct ablation studies in Section 4.6 to further analyze the effectiveness of CLUSTERLLM. Finally, we show results of determining cluster granularity in Section 4.7.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Datasets\",\n \"publication_ref\": [\n \"b75\",\n \"b31\",\n \"b37\",\n \"b13\"\n ],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"We provide a high-level summary of evaluated datasets in this section, and see Appendix E for more descriptions. In this paper, we evaluate on a broad range of clustering datasets with various perspectives and granularities. Furthermore, to better analyze the effect of scale, each dataset has both a small-scale and a large-scale version. The two versions are different in number of data while keeping the same number of clusters. A summary of dataset statistics is shown in Table 1. Note that there is no data splits in clustering. Intent (Domain) Discovery.\\nIntent discovery (Zhang et al., 2021b(Zhang et al., , 2022) ) discovers unknown intents in unlabeled customer utterances. For CLINC, Massive and MTOP, we also use domains as labels to convert them into domain discovery. Type Discovery. Type Discovery (Li et al., 2022) resolves the closed-world set-up of traditional Information Extraction. In this work, we focus on three tasks: entity, relation and event type discovery. To indicate specific mentions (entities or event triggers), we directly append them behind sentences with natural language formats, such as \\\"The relation between [ENTITY1] and [ENTITY2]\\\". Topic Mining. We adapt three topic mining datasets from MTEB (Muennighoff et al., 2022). Emotion. We adapt GoEmo (Demszky et al., 2020), a fine-grained emotion detection dataset by removing multi-label or neutral instances.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiment Details\",\n \"publication_ref\": [\n \"b47\",\n \"b54\",\n \"b61\"\n ],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"Query LLMs. The prompt only contains a taskspecific instruction (see Table 11). We set generation temperature to 0.5. Explanations are suppressed by adding a postfix:\\\"Please respond with 'Choice 1' or 'Choice 2' without explanation\\\" and set up a max token of 10. We then assign them to binary choices by directly checking whether one of the texts \\\"Choice 1\\\" or \\\"Choice 2\\\" is in the response. We also find that a very small amount of responses do not contain any choices and we discard them during fine-tuning. We use the Python API tool provided by OpenAI.\\nTriplet Sampling. For both small-or large-scale experiments, we set a budget of Q = 1, 024 triplets. We set γ low = 20% and γ high = 0. For clustering methods, we fix hyperparameters of these algorithms across datasets in Stage 1. We choose agglomerative clustering with fixed distance threshold 67 for small-scale experiments on Instructor, and 77 on E5 (the embeddings are preprocessed by standard scaler). For large-scale datasets, we choose mini-batch K-means with fixed number of clusters 100 due to its lower latency. Clustering algorithms are implemented by scikit-learn (Pedregosa et al., 2011).\\nFine-tune Embedders. In this work, we focus on two state-of-the-art pre-trained embedders: Instructor (Su et al., 2022) and E5 (Wang et al., 2022). We only use the large versions. Refer to Appendix D for details.\\nEvaluation. To reduce cost, we run CLUSTER-LLM once for each dataset. We then run (minibatch) K-means on (large) small-scale datasets for 5 seeds with ground truth K. We show two metrics. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Compared Methods\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"E5 and Instructor. We directly apply (minibatch) K-means on extracted embeddings from instructor-large and e5-large. self-supervise-I(E). To verify that the performance improvement of CLUSTERLLM does not only come from domain-specific fine-tuning, instead of the more accurate triplet prediction. We propose a self-supervise fine-tuning that uses exactly the same triplets as CLUSTERLLM but only switch to self-supervised triplet predictions that select closest choices in embedding space. SCCL-I(E). We also combine Instructor and E5 with SCCL (Zhang et al., 2021a), an unsupervised deep clustering algorithm that utilizes entire dataset for training. Notice that our method uses fewer data for training. See Appendix D for details.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Main Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"We show main results with small-scale datasets in Table 2. We show several variants of our method: CLUSTERLLM-I(E) adopt Instructor or E5 as embedders. CLUSTERLLM-I(E)-iter applies the entire framework twice in an iterative manner by using previous fine-tuned model as initialization and the 1, 024 triplets inferred from new embeddings for fine-tuning. All of these use GPT-3.5 for prediction. We make the following observations: (1) CLUSTERLLM consistently improves upon both embedders. For example, CLUSTER- LLM-I increases the performance by 6.71% on FewRel. CLUSTERLLM-E increases the performance by 9.19 on Bank77. However, we do observe that on Massive(D) and CLINC(D), there are no improvements.\\n(2) CLUSTERLLM outperforms deep clustering and self-supervise baselines. For instance, CLUSTERLLM-I surpasses self-supervise-I on most datasets except for two and it is also better than SCCL-I on 11 over 14 datasets. Furthermore, these improvements are consistent across both reported metrics.\\n(3) Combined with the results in Appendix F, applying CLUSTERLLM iteratively is beneficial, emphasizing the potential of further improvements.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Analysis on Triplet Prediction Accuracy\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\"\n ],\n \"text\": \"We attribute the improvements on clustering quality to more accurate triplet predictions. In Table 3, we show the accuracy on predicted triplets that have ground truth (exactly one positive and one negative choices based on ground truth) with two different sampling methods. Random triplet sampling uniformly samples three random instances as query and two choices, and we guarantee the two choices are different from the anchor by filtering. Furthermore, we also show a selection accuracy with Euclidean distances between embeddings as a comparison. We observe that, GPT-3.5/4 consistently improves upon Instructor on high entropy ex-amples, demonstrating our hypothesis. In contrast, with random sampling, the ground truth triplets is significantly fewer and the accuracy gap is much smaller or even decreases performance.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Ablation Study\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_5\",\n \"tab_8\"\n ],\n \"text\": \"Clustering Quality. We show ablation studies on CLUSTERLLM based on Instructor in Table 4. Specifically, we present results with 3 kinds of predictions on the same set of triplets for finetuning: GPT-3.5/4, replace triplet predictions of GPT-3.5 to ground truth on those triplets that have ground truth. We observe that GPT-4 marginally improves upon GPT-3.5 given the much higher cost. When provided with human labels, CLUS-TERLLM-GT&GPT3.5 achieves the highest performance, which indicates the possibility for further improvement with more accurate predictions. We make similar observations for large-scale datasets in Table 6. Sampling Strategy. In this section, we show ablation study on entropy-based sampling. In Figure 3, we observe that clustering accuracy increases when increasing entropies (or equally decreasing mean of interval) except for GoEmo. We make two hypothesis: (1) LLMs are much better than small embedders on harder instances. ( 2 x-axis shows the mean of interval where interval length is set to 20%. For example, \\\"mean of interval= 50%\\\" means γ high = 40% and γ low = 60% (see Section 3.1.1). ♦ marks the setting for main experiments.\\nlected triplets even decreases performance, which demonstrates the cruciality of triplet sampling.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Determining Cluster Granularity\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_6\",\n \"tab_9\",\n \"tab_6\",\n \"tab_6\"\n ],\n \"text\": \"In this section, we show the results for determining cluster granularity. We evaluate on a subset of 8 datasets including various cluster granularities with k max = 200 and k min = 2. We compare with different methods that rely on clustering errors. For our methods, we show results with λ = {1, 3} (except for GPT-4 to reduce costs), which involve\\n198 & 594 pairs in total respectively. To simulate experts for providing demonstrations, we directly sample 16 pairs from small-scale datasets when λ = 3 and then choose 2 positive and 2 negative as demonstrations. Notice that we use the same demonstrations for large-scale experiments. See more details in Appendix B. We make several observations from Table 5 andTable 7: (1) Our methods have higher ranks. Most baseline methods predict similar number of clusters for domain and intent, while our methods can effectively distinguish between the two. For instance, on MTOP(I)/(D) in Table 5, BIC predicts number of clusters 69/64 while our method (GPT-3.5, λ = 3) predicts 92/18. (2) Increasing λ generally helps (MTOP(D) in Table 5) but might not always make a large difference. (3) GPT-4 significantly improves upon GPT-3.5, probably due to its better understanding of demonstrations.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Works\",\n \"publication_ref\": [\n \"b67\",\n \"b24\",\n \"b68\",\n \"b69\",\n \"b6\",\n \"b42\",\n \"b64\",\n \"b27\",\n \"b11\",\n \"b6\",\n \"b58\",\n \"b41\",\n \"b8\",\n \"b42\",\n \"b15\",\n \"b64\",\n \"b24\",\n \"b9\",\n \"b69\",\n \"b42\",\n \"b58\",\n \"b8\",\n \"b2\",\n \"b60\",\n \"b3\",\n \"b74\",\n \"b34\",\n \"b50\",\n \"b20\",\n \"b61\",\n \"b54\",\n \"b37\"\n ],\n \"table_ref\": [],\n \"text\": \"Clustering. As a fundamental task in machine learning, clustering has been applied on diverse data types, including texts (Xu et al., 2015;Hadifar et al., 2019;Zhang et al., 2021a), images (Yaling Tao, 2021;Yang et al., 2016;Caron et al., 2018;Niu et al., 2020;Xie et al., 2016) and graphs (Huang et al., 2014;Chiang et al., 2019). Recent research has been shifted to deep cluster- et al., 2016;Caron et al., 2018;Van Gansbeke et al., 2020;Niu et al., 2022;Chang et al., 2017;Niu et al., 2020), generative (Dilokthanakul et al., 2016) and simultaneous (Xie et al., 2016;Zhang et al., 2021a;Hadifar et al., 2019) man, 1972;Cheeger, 1970) makes use of similarity matrix where each entry measures the similarity between a pair of data. More recently, several works in deep clustering utilize relational supervision (Yang et al., 2016;Niu et al., 2020;Van Gansbeke et al., 2020;Chang et al., 2017) via pseudolabelling which could be noisy. Another line of works that is closely related to ours is constrained clustering. It usually incorporates pairwise mustlink or cannot-link constraints (Basu et al., 2004;Wagstaff et al., 2001;Basu et al., 2008;Zhang et al., 2020;Manduchi et al., 2021). Nonetheless, these constraints are often sampled from labels as a prior which significantly limits its application in our scenario. In this work, we study how to utilize contemporary API-based LLMs to infer relations.\\nPre-trained Embedding Model. Generic pre-trained text embedding models (Reimers and Gurevych, 2019;Gao et al., 2021;Ni et al., 2022a,b) are widely applied in text similarity, classification, clustering and information retrieval. Recently, two embedding models E5 (Wang et al., 2022) and Instructor (Su et al., 2022) have shown superior performance on a popular benchmark (Muennighoff et al., 2022). Specifically E5 is pre-trained on web-scraped data pairs with contrastive objective. Instructor is pre-trained on 330 tasks with instructions. CLUSTERLLM aims at improving these models with LLMs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we study how to leverage API-based LLMs to guide small embedders for text clustering in order to benefit from high-level language capability of LLMs and user's instructions on clustering. We propose to prompt LLMs with two kinds of sentence relationship tasks: triplet task and pairwise task. Triplet task chooses the sentence that is most similar with anchor combining with a perspective instruction from users. The predicted triplets are used for fine-tuning small embedders.\\nPairwise task judges whether a pair of sentences belong to the same category hinted by few-shot demonstrations, and then the predictions are used to determine cluster granularity with a consistency measure. Extensive experiments show that our proposed framework CLUSTERLLM can improve clustering quality and propose reasonable cluster granularity at a negligible cost. However, CLUS-TERLLM still relies on the embedding model itself, which is inefficient and inapplicable on black-box embedding models. We encourage future works to explore the potential of model-free training such as constrained clustering.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Details of Scaling up Hierarchical Clustering\",\n \"publication_ref\": [\n \"b63\"\n ],\n \"table_ref\": [],\n \"text\": \"A major drawback of hierarchical clustering is its O(N3 ) time complexity which makes the algorithm hard to be deployed on larger datasets. However, since we are only interested in a specific range of granularity in our scenario, the hierarchical clustering can start from an intermediate step. We address this issue by first running minibatch K-means with k max and then run hierarchical clustering with current assignments as inputs. Specifically, we use agglomerative clustering with ward's method (Ward Jr, 1963). We first calculate distances between each pair of clusters according to Murtagh and Contreras, 2011 and then provide them as inputs to nearest neighbor chain algorithm.\\nFinally the returned hierarchy is combined with the K-means assignments to infer clusters.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B More Details about Determining Cluster Granularity\",\n \"publication_ref\": [\n \"b53\",\n \"b56\",\n \"b48\",\n \"b23\",\n \"b23\"\n ],\n \"table_ref\": [\n \"tab_2\",\n \"tab_13\"\n ],\n \"text\": \"Previous methods often employ clustering errors as a metric and they ignore user's need on the granularity. Silhouette coefficient (Rousseeuw, 1987) indicates the clustering quality without ground truths, which exploits the inter-cluster distance with nearest clusters and the intra-cluster distance. We find the granularity by choosing the one with the best silhouette coefficient. Elbow method (Thorndike, 1953) is a heuristic method that plots the clustering error with respect to different levels of granularity in the hierarchy. And then the best granularity is determined with the largest elbow length. X-means (Pelleg et al., 2000) is a variation of K-means that starts with the lowest number of clusters, and then repeatedly attempt to split the clusters by running 2-means on them and evaluate with Bayesian Information Criterion (BIC) (Goutte et al., 2001). BIC (Goutte et al., 2001) calculates BIC for each of the granularity. Cluster-Size (Zhang et al., 2021b) uses a confidence threshold to filter small clusters starting from the maximum number of cluster. For all methods, we use the same fine-tuned embeddings (CLUSTERLLM-I in Table 2). The same cluster hierarchy is used (except for X-means that relies on K-means), which is either acquired from hierarchical clustering for small-scale or our proposed two-step method in Section 3.2 for large-scale. For out methods, the weight in F-beta score is set to 0.92 through empirical selection on Bank77. Because of the high latency, results for Silhouette and X-means are not shown on large-scale datasets. After sampling 16 data pairs, we tend to choose positives with finer granularity or negatives with coarser granularity. We also consider the sentence length to minimize the cost. We use label names as justifications and we always put 2 positive before 2 negative (See Table 12 bottom).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C Analysis for Determining Granularity\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Prompt Design. We show the analysis results of prompt design for determining granularity in Table 8. We experiment with two settings: (1) remove justification for all demonstrations and only keep the \\\"Yes\\\" or \\\"No\\\" answer.\\n(2) remove all demonstrations and any granularity-related words (such as domain) 3 . We observe that demonstrations are necessary and adding justifications have a positive impact.\\nVisualization of Consistency Score. We visualize consistency score with respect to the number of clusters. The consistency scores exhibit continuous variations and peak at the best number of clusters.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D Details of Embedders and Fine-tuning\",\n \"publication_ref\": [\n \"b54\"\n ],\n \"table_ref\": [\n \"tab_12\"\n ],\n \"text\": \"For all the experiments (including those with or without fine-tuning), we use large version of both Instructor and E5 (i.e.\\nhkunlp/instructor-large & intfloat/e5-large).\\nFor Instructor, we use the same or similar prompt as original paper. See Table 10.\\nFor fine-tuning, we adopt the same hyperparameters as in (Su et al., 2022), but modify the learning rate to 2e -6, the maximum gradient steps to 3, 840 for Instructor (∼ 15 epochs) and 1, 280 for E5, and batch size to 4. We choose this gradient in the begining of our experiments by observing no performance increase after that on several datasets. Training is conducted with a single NVIDIA Quadro RTX 8000 GPU.\\nFor SCCL-I(E), we change the maximum token length to 128 due to the limited compute resource 4 . We use the same learning rate 2e -6 as before for Instructor and 2e -7 for E5 since we found that the performance is unstable with large learning rate. Batch size is set to 16 and we evaluate representations with K-means after 200 iterations. Also notice that we do not interrupt prompts in Instructor during data augmentation.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"E Description of Datasets\",\n \"publication_ref\": [\n \"b7\",\n \"b29\",\n \"b18\",\n \"b30\",\n \"b37\",\n \"b19\",\n \"b14\",\n \"b16\",\n \"b21\"\n ],\n \"table_ref\": [\n \"tab_11\"\n ],\n \"text\": \"Bank77 (Casanueva et al., 2020) is a popular dataset in intent discovery that focuses on creating fine-grained intent categories for a single-domain, \\\"banking\\\". CLINC(I) (Larson et al., 2019) is originally created for detecting utterances that falls outside of supported intents. The dataset also contains multiple domains, such as \\\"travel\\\", \\\"utility\\\" and \\\"work\\\". In this experiment, we discard all the out-of-scope utterances and only focus on indomain ones. Moreover, we create a domain discovery dataset CLINC(D) that uses domains as labels. Massive(I)/(D) (FitzGerald et al., 2022) and MTOP(I)/(D) (Li et al., 2021) are both from MTEB (Muennighoff et al., 2022). Here \\\"I\\\" denotes intent and \\\"D\\\" for domain (or scenario). These datasets are originally used for classification but are adapted here for clustering. We also remove those intents with only a few instances and keep English-only data. For all datasets, we use the train & test sets as large-& small-scale datasets respectively. For FewRel (Gao et al., 2019) and FewEvent (Deng et al., 2020), we first randomly split datasets into train & test sets, and then sample from train set as large-scale and test set as smallscale. For FewNerd (Ding et al., 2021), we use the original train & test splits. For StackEx, Reddit (Geigle et al., 2021) and ArxivS2S, we combine all the splits into a single dataset and remove topics that only have few instances. Finally, the datasets are randomly splitted into large-& small-scale versions. To show the dataset balancy, we show the 4 We also tried maximum token length 512 with smaller batch size, but the performance is not better. entropy of class distribution in Table 9.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"F Results of More Iterations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We show the results over 4 iterations of CLUS-TERLLM in Figure 5. During iteration, we sample triplets from previously fine-tuned embedding space and continue to fine-tune the model with previous checkpoint as initialization. We also show the self-supervise results using the same checkpoint fine-tuned with GPT-3.5 predictions as initialization at each iteration. We observe that using GPT-3.5 predictions is almost always beneficial. The performance generally increases and saturate at the fourth iteration with the exception of GoEmo.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"G More Related Works\",\n \"publication_ref\": [\n \"b59\",\n \"b32\",\n \"b75\",\n \"b36\",\n \"b0\",\n \"b70\",\n \"b5\",\n \"b55\",\n \"b75\",\n \"b73\",\n \"b0\",\n \"b32\",\n \"b22\",\n \"b25\",\n \"b77\",\n \"b10\",\n \"b1\"\n ],\n \"table_ref\": [],\n \"text\": \"Generalized Category Discovery (GCD). GCD (Vaze et al., 2022;Lin et al., 2020;Zhang et al., 2021bZhang et al., , 2022;;Mou et al., 2022;An et al., 2022) assume partial known classes with annotations which can also be used to infer user's requirement on clustering. As an infant research area, most previous works employ pseudo-labelling, via optimal transport (Rizve et al., 2022b;Yang et al., 2022), similarity learning (Rizve et al., 2022a;Cao et al., 2022) or prototype-based learning (Sun and Li, 2022). Furthermore, new intent discovery (Zhang et al., 2022(Zhang et al., , 2021b(Zhang et al., , 2023;;An et al., 2022;Lin et al., 2020) Represent the bank purpose for retrieval: CLINC(I) Represent the sentence for retrieving the purpose: FewRel Represent the relation between two entities for retrieval: FewNerd Represent the entity type for retrieval: FewEvent Represent the event type for retrieval: StackEx Represent the question for retrieval: ArxivS2S\\nRepresent the science statement for retrieval: GoEmo Represent an emotion sentence for retrieval: Massive(I) Represent the sentence for retrieving the purpose: MTOP(I) Represent the sentence for retrieval: Reddit represent a reddit community title: Massive(D) Represent the scene for retrieval: MTOP(D) Represent a sentence: CLINC(D) Represent a sentence: computation-& data-cost. LLMs as Annotators. Recent instruction-tuned LLMs, such as ChatGPT, have been shown to have the ability to reproduce or improve humangenerated labels (Gilardi et al., 2023;He et al., 2023;Zhu et al., 2023). Furthermore, several works dedicate to fine-tune models with feedbacks from LLMs (Cheng et al., 2023;Bai et al., 2022). This paper instead focuses on clustering tasks.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"H Sub-optimal Performance on Domain Discovery\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\",\n \"tab_5\",\n \"tab_8\",\n \"tab_2\"\n ],\n \"text\": \"We noticed that the performance of domain discovery (MTOP(D), Massive(D) and CLINC(D)) is barely improved or even decreased with CLUS-TERLLM from original embedders (see Table 2). Furthermore, the ablation studies reveal that even with CLUSTERLLM-GT&GPT3.5, clustering performance is not as good as self-supervise or CLUS-TERLLM-random (see CLINC(D) in Table 4 and MTOP(D) in Table 6). We also observe that, CLUS-TERLLM-I-iter will further decrease the performance (see Massive(D) in Table 2). While we do not have rigorous explanations, one hypothesize is that the embedding space after fine-tuning tends to be more compact than before and forming small cliques, making it better for clustering finegrained clusters but not for coarse-grained clusters. We showcase scatterplots on two datasets with both Instructor and CLUSTERLLM-I-iter. It can be observed that the clusters in embedding space are tighter and more separated especially on CLINC(D).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"I Dataset Leakage\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\"\n ],\n \"text\": \"Since LLMs like ChatGPT are trained on webscraped texts from internet, it is likely they already have access to our evaluation datasets during training. For example, the topic mining datasets use StackExchange, Reddit and Arxiv tags as labels which is freely available online. However, as observed in Table 3, the performance of triplet prediction on these datasets are often far from perfect. Furthermore, the other datasets like Bank77 are synthesized datasets which is not accessible during training. FewRel is collected from Wikipedia corpus but their labels are not easily accessible. Similarly, GoEmo is collected from Reddit but the emotion labels are not accessible during training. Thus, dataset leakage is not a primary concern of this paper.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Dataset Prompt Bank77\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Select the banking customer utterance that better corresponds with the Query in terms of intent. CLINC(I) Select the customer utterance that better corresponds with the Query in terms of intent.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"FewRel\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Select the example that better corresponds with the Query in terms of relation type.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"FewNerd\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Select the example that better corresponds with the Query in terms of entity type. FewEvent Select the example that better corresponds with the Query in terms of event type.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"StackEx\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Select the StackExchange question that better corresponds with the Query in terms of topic.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"ArxivS2S\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Select the Arxiv paper title that better corresponds with the Query in terms of domain.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GoEmo\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Select the sentence that better corresponds with the Query in terms of emotion expressed. Massive(I) Select the user utterance that better corresponds with the Query in terms of intent.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"MTOP(I)\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"Select the user utterance that better corresponds with the Query in terms of intent Reddit Select the Reddit question that better corresponds with the Query in terms of topic. Massive(D) Select the user utterance that better corresponds with the Query in terms of scenario. MTOP(D) Select the user utterance that better corresponds with the Query in terms of domain. CLINC(D) Select the customer utterance that better corresponds with the Query in terms of domain.\\nTable 11: Prefix of prompts for triplet task. Notice while we use different prompts for domain and intent (such as CLINC(I) and CLINC(D)) in our experiments, they might be used interchangeably.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Dataset Prompt\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Triplet Task\\nSelect the banking customer utterance that better corresponds with the Query in terms of intent.\\nQuery: Should i reinstall the payment app? Choice 1: I've received my card so now I need to know how to sync it to the app. Choice 2: Can I still use the app if I switched phones?\\nPlease respond with 'Choice 1' or 'Choice 2' without explanation.\\nPairwise Task\\n[Example1] Sentence 1: I would like to see the source of my money.\\nSentence 2: My source of funds need verified. Yes. Because both intents are verify source of funds.\\n[Example2] Sentence 1: Is there a fee for topping up Sentence 2: What are the top up charges for US cards? Yes. Because both intents are top up by card charge.\\n[Example3] Sentence 1: Can I reactivate my lost card that I found this morning in my jacket pocket? Sentence 2: how to activate card? No. Because Sentence 1 has intent card linking and Sentence 2 has intent activate my card.\\n[Example4] Sentence 1: What will I be charged for a physical card? Sentence 2: My card is about to expire and I need to know how much it costs and how long ... No. Because Sentence 1 has intent order physical card and Sentence 2 has intent card ... Determine whether the intents of two banking customer utterances below belong to the same intent category using above examples.\\nSentence 1: $1 extra has been charged on my statement, why is that? Sentence 2: Will it automatically top-up if there isn't much money left?\\nPlease respond with 'Yes' or 'No' without explanation. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgements\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Our work is sponsored in part by NSF CAREER Award 2239440, NSF Proto-OKN Award 2333790, NIH Bridge2AI Center Program under award 1U54HG012510-01, Cisco-UCSD Sponsored Research Project, as well as generous gifts from Google, Adobe, and Teradata. Any opinions, findings, and conclusions or recommendations expressed herein are those of the authors and should not be interpreted as necessarily representing the views, either expressed or implied, of the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for government purposes not withstanding any copyright annotation hereon.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We list several limitations of our work that we hope to be improved in the future: Reliance on pre-trained embedder. To find the most informative data, we have to rely on a pretrained embedder that can indicate the largest clustering assignment entropy. We hope that selfsupervise triplets and LLM-predicted triplets can be combined to solve such an issue. Computational cost for fine-tuning. Our initial idea is to utilize constrained clustering which is a light-weight algorithm that do not need to update small embedders. However, the inevitable unstable training will be heavily affected by the errors in LLM predictions. We make a comprise by introducing embedder into fine-tuning to temporarily solve the issue, but we hope to reduce the computational cost in a future work. Sub-optimal performance on domain discovery. We notice that on domain discovery datasets such as Massive(D) and CLINC(D), the performance is usually sub-optimal compared with original Instructor embedding. We provide discussions on this issue in Appendix H.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ethics Statement\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Our work employs LLMs which are accessed through OpenAI APIs. For some applications, uploading privacy-sensitive data is risky and might require efforts to remove sensitive information.\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-11-03"},"doi":{"kind":"string","value":"10.18653/v1/2020.acl-main.372"},"references":{"kind":"list like","value":[{"authors":"Wenbin An; Feng Tian; Qinghua Zheng; Wei Ding; Qianying Wang; Ping Chen","journal":"","ref_id":"b0","title":"Generalized category discovery with decoupled prototypical network","year":"2022"},{"authors":"Yuntao Bai; Saurav Kadavath; Sandipan Kundu; Amanda Askell; Jackson Kernion; Andy Jones; Anna Chen; Anna Goldie; Azalia Mirhoseini; Cameron Mckinnon","journal":"","ref_id":"b1","title":"Constitutional ai: Harmlessness from ai feedback","year":"2022"},{"authors":"Sugato Basu; Arindam Banerjee; Raymond J Mooney","journal":"","ref_id":"b2","title":"Active semi-supervision for pairwise constrained clustering","year":"2004"},{"authors":"Sugato Basu; Ian Davidson; Kiri Wagstaff","journal":"CRC Press","ref_id":"b3","title":"Constrained clustering: Advances in algorithms, theory, and applications","year":"2008"},{"authors":"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell","journal":"Advances in neural information processing systems","ref_id":"b4","title":"Language models are few-shot learners","year":"2020"},{"authors":"Kaidi Cao; Maria Brbic; Jure Leskovec","journal":"","ref_id":"b5","title":"Open-world semi-supervised learning","year":"2022"},{"authors":"Mathilde Caron; Piotr Bojanowski; Armand Joulin; Matthijs Douze","journal":"","ref_id":"b6","title":"Deep clustering for unsupervised learning of visual features","year":"2018"},{"authors":"Iñigo Casanueva; Tadas Temcinas; Daniela Gerz; Matthew Henderson; Ivan Vulic","journal":"","ref_id":"b7","title":"Efficient intent detection with dual sentence encoders","year":"2020"},{"authors":"Jianlong Chang; Lingfeng Wang; Gaofeng Meng; Shiming Xiang; Chunhong Pan","journal":"","ref_id":"b8","title":"Deep adaptive image clustering","year":"2017"},{"authors":"Jeff Cheeger","journal":"","ref_id":"b9","title":"A lower bound for the smallest eigenvalue of the laplacian, problems in analysis","year":"1969"},{"authors":"Qinyuan Cheng; Xiaogui Yang; Tianxiang Sun; Linyang Li; Xipeng Qiu","journal":"","ref_id":"b10","title":"Improving contrastive learning of sentence embeddings from ai feedback","year":"2023"},{"authors":"Wei-Lin Chiang; Xuanqing Liu; Si Si; Yang Li; Samy Bengio; Cho-Jui Hsieh","journal":"","ref_id":"b11","title":"Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks","year":"2019"},{"authors":"Pierre Colombo; Nathan Noiry; Ekhine Irurozki; Stephan Clemencon","journal":"","ref_id":"b12","title":"What are the best systems? new perspectives on nlp benchmarking","year":"2022"},{"authors":"Dorottya Demszky; Dana Movshovitz-Attias; Jeongwoo Ko; Alan Cowen; Gaurav Nemade; Sujith Ravi","journal":"Association for Computational Linguistics","ref_id":"b13","title":"GoEmotions: A dataset of fine-grained emotions","year":"2020"},{"authors":"Shumin Deng; Ningyu Zhang; Jiaojian Kang; Yichi Zhang; Wei Zhang; Huajun Chen","journal":"Association for Computing Machinery","ref_id":"b14","title":"Metalearning with dynamic-memory-based prototypical network for few-shot event detection","year":"2020"},{"authors":"Nat Dilokthanakul; Pedro Am Mediano; Marta Garnelo; C H Matthew; Hugh Lee; Kai Salimbeni; Murray Arulkumaran; Shanahan","journal":"","ref_id":"b15","title":"Deep unsupervised clustering with gaussian mixture variational autoencoders","year":"2016"},{"authors":"Ning Ding; Guangwei Xu; Yulin Chen; Xiaobin Wang; Xu Han; Pengjun Xie; Haitao Zheng; Zhiyuan Liu","journal":"Association for Computational Linguistics","ref_id":"b16","title":"Few-NERD: A few-shot named entity recognition dataset","year":"2021"},{"authors":"E William; Alan J Donath; Hoffman","journal":"IBM Technical Disclosure Bulletin","ref_id":"b17","title":"Algorithms for partitioning of graphs and computer logic based on eigenvectors of connection matrices","year":"1972"},{"authors":"Jack Fitzgerald; Christopher Hench; Charith Peris; Scott Mackie; Kay Rottmann; Ana Sanchez; Aaron Nash; Liam Urbach; Vishesh Kakarala; Richa Singh; Swetha Ranganath; Laurie Crist; Misha Britan; Wouter Leeuwis; Gokhan Tur; Prem Natarajan","journal":"","ref_id":"b18","title":"Massive: A 1m-example multilingual natural language understanding dataset with 51 typologically-diverse languages","year":"2022"},{"authors":"Tianyu Gao; Xu Han; Hao Zhu; Zhiyuan Liu; Peng Li; Maosong Sun; Jie Zhou","journal":"Association for Computational Linguistics","ref_id":"b19","title":"FewRel 2.0: Towards more challenging few-shot relation classification","year":"2019"},{"authors":"Tianyu Gao; Xingcheng Yao; Danqi Chen","journal":"Association for Computational Linguistics","ref_id":"b20","title":"SimCSE: Simple contrastive learning of sentence embeddings","year":"2021"},{"authors":"Gregor Geigle; Nils Reimers; Andreas Rücklé; Iryna Gurevych","journal":"","ref_id":"b21","title":"TWEAC: transformer with extendable QA agent classifiers","year":"2021"},{"authors":"Fabrizio Gilardi; Meysam Alizadeh; Maël Kubli","journal":"","ref_id":"b22","title":"Chatgpt outperforms crowd-workers for textannotation tasks","year":"2023"},{"authors":"Cyril Goutte; Lars Kai Hansen; Matthew G Liptrot; Egill Rostrup","journal":"Human brain mapping","ref_id":"b23","title":"Feature-space clustering for fmri meta-analysis","year":"2001"},{"authors":"Amir Hadifar; Lucas Sterckx; Thomas Demeester; Chris Develder","journal":"Association for Computational Linguistics","ref_id":"b24","title":"A self-training approach for short text clustering","year":"2019"},{"authors":"Xingwei He; Zhenghao Lin; Yeyun Gong; Hang Jin; Chen Zhang; Jian Lin; Jiao; Ming Siu; Nan Yiu; Weizhu Duan; Chen","journal":"","ref_id":"b25","title":"Annollm: Making large language models to be better crowdsourced annotators","year":"2023"},{"authors":"William Hogan; Jiacheng Li; Jingbo Shang","journal":"","ref_id":"b26","title":"Open-world semi-supervised generalized relation discovery aligned in a real-world setting","year":"2023"},{"authors":"Peihao Huang; Yan Huang; Wei Wang; Liang Wang","journal":"IEEE","ref_id":"b27","title":"Deep embedding network for clustering","year":"2014"},{"authors":" Harold W Kuhn","journal":"Naval research logistics quarterly","ref_id":"b28","title":"The hungarian method for the assignment problem","year":"1955"},{"authors":"Stefan Larson; Anish Mahendran; Joseph J Peper; Christopher Clarke; Andrew Lee; Parker Hill; Jonathan K Kummerfeld; Kevin Leach; Michael A Laurenzano; Lingjia Tang; Jason Mars","journal":"","ref_id":"b29","title":"An evaluation dataset for intent classification and out-ofscope prediction","year":"2019"},{"authors":"Haoran Li; Abhinav Arora; Shuohui Chen; Anchit Gupta; Sonal Gupta; Yashar Mehdad","journal":"Association for Computational Linguistics","ref_id":"b30","title":"MTOP: A comprehensive multilingual task-oriented semantic parsing benchmark","year":"2021"},{"authors":"Sha Li; Ji Heng; Jiawei Han","journal":"","ref_id":"b31","title":"Open relation and event type discovery with type abstraction","year":"2022"},{"authors":"Hua Ting-En Lin; Hanlei Xu; Zhang","journal":"","ref_id":"b32","title":"Discovering new intents via constrained deep adaptive clustering with cluster refinement","year":"2020"},{"authors":" Macqueen","journal":"","ref_id":"b33","title":"Classification and analysis of multivariate observations","year":"1967"},{"authors":"Laura Manduchi; Kieran Chin-Cheong; Holger Michel; Sven Wellmann; Julia E Vogt","journal":"","ref_id":"b34","title":"Deep conditional gaussian mixture model for constrained clustering","year":"2021"},{"authors":"José Manuel; Guaita Martínez; Patricia Carracedo; Dolores Gorgues Comas; Carlos H Siemens","journal":"Sustainable Technology and Entrepreneurship","ref_id":"b35","title":"An analysis of the blockchain and covid-19 research landscape using a bibliometric study","year":"2022"},{"authors":"Yutao Mou; Yanan He; Zhiyuan Wu; Hong Zeng; Huixing Xu; Wei Jiang; Weiran Wu; Xu","journal":"Association for Computational Linguistics","ref_id":"b36","title":"Disentangled knowledge transfer for OOD intent discovery with unified contrastive learning","year":"2022"},{"authors":"Niklas Muennighoff; Nouamane Tazi; Loïc Magne; Nils Reimers","journal":"","ref_id":"b37","title":"Mteb: Massive text embedding benchmark","year":"2022"},{"authors":"Fionn Murtagh; Pedro Contreras","journal":"","ref_id":"b38","title":"Methods of hierarchical clustering","year":"2011"},{"authors":"Jianmo Ni; Gustavo Hernandez Abrego; Noah Constant; Ji Ma; Keith Hall; Daniel Cer; Yinfei Yang","journal":"Association for Computational Linguistics","ref_id":"b39","title":"Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models","year":"2022"},{"authors":"Jianmo Ni; Chen Qu; Jing Lu; Zhuyun Dai; Gustavo Hernandez Abrego; Ji Ma; Vincent Zhao; Yi Luan; Keith Hall; Ming-Wei Chang; Yinfei Yang","journal":"","ref_id":"b40","title":"Large dual encoders are generalizable retrievers","year":"2022"},{"authors":"Chuang Niu; Hongming Shan; Ge Wang","journal":"IEEE Transactions on Image Processing","ref_id":"b41","title":"Spice: Semantic pseudo-labeling for image clustering","year":"2022"},{"authors":"Chuang Niu; Jun Zhang; Ge Wang; Jimin Liang","journal":"","ref_id":"b42","title":"Gatcluster: Self-supervised gaussian-attention network for image clustering","year":"2020"},{"authors":"M Robert; Nosofsky","journal":"","ref_id":"b43","title":"The generalized context model: An exemplar model of classification","year":"2011"},{"authors":" Openai","journal":"","ref_id":"b44","title":"Gpt-4 technical report","year":"2023"},{"authors":"Long Ouyang; Jeffrey Wu; Xu Jiang; Diogo Almeida; Carroll Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray","journal":"Advances in Neural Information Processing Systems","ref_id":"b45","title":"Training language models to follow instructions with human feedback","year":"2022"},{"authors":"June Young Park; Evan Mistur; Donghwan Kim; Yunjeong Mo; Richard Hoefer","journal":"Sustainable Cities and Society","ref_id":"b46","title":"Toward humancentric urban infrastructure: Text mining for social media data to identify the public perception of covid-19 policy in transportation hubs","year":"2022"},{"authors":"F Pedregosa; G Varoquaux; A Gramfort; V Michel; B Thirion; O Grisel; M Blondel; P Prettenhofer; R Weiss; V Dubourg; J Vanderplas; A Passos; D Cournapeau; M Brucher; M Perrot; E Duchesnay","journal":"Journal of Machine Learning Research","ref_id":"b47","title":"Scikit-learn: Machine learning in Python","year":"2011"},{"authors":"Dan Pelleg; Andrew W Moore","journal":"Icml","ref_id":"b48","title":"X-means: Extending k-means with efficient estimation of the number of clusters","year":"2000"},{"authors":"Maarten De Raedt; Fréderic Godin; Thomas Demeester; Chris Develder","journal":"","ref_id":"b49","title":"Idas: Intent discovery with abstractive summarization","year":"2023"},{"authors":"Nils Reimers; Iryna Gurevych","journal":"Association for Computational Linguistics","ref_id":"b50","title":"Sentence-BERT: Sentence embeddings using Siamese BERTnetworks","year":"2019"},{"authors":"Mamshad Nayeem Rizve; Navid Kardan; Salman Khan; Fahad Shahbaz Khan; Mubarak Shah","journal":"Springer","ref_id":"b51","title":"Openldn: Learning to discover novel classes for openworld semi-supervised learning","year":"2022"},{"authors":"Mamshad Nayeem Rizve; Navid Kardan; Mubarak Shah","journal":"Springer","ref_id":"b52","title":"Towards realistic semi-supervised learning","year":"2022"},{"authors":"J Peter; Rousseeuw","journal":"Journal of computational and applied mathematics","ref_id":"b53","title":"Silhouettes: a graphical aid to the interpretation and validation of cluster analysis","year":"1987"},{"authors":"Hongjin Su; Weijia Shi; Jungo Kasai; Yizhong Wang; Yushi Hu; Mari Ostendorf; Wen-Tau Yih; Noah A Smith; Luke Zettlemoyer; Tao Yu","journal":"","ref_id":"b54","title":"One embedder, any task: Instruction-finetuned text embeddings","year":"2022"},{"authors":"Yiyou Sun; Yixuan Li","journal":"Transactions on Machine Learning Research","ref_id":"b55","title":"Opencon: Open-world contrastive learning","year":"2022"},{"authors":"Robert Thorndike","journal":"Psychometrika","ref_id":"b56","title":"Who belongs in the family?","year":"1953"},{"authors":"Laurens Van Der Maaten; Geoffrey Hinton","journal":"Journal of machine learning research","ref_id":"b57","title":"Visualizing data using t-sne","year":"2008"},{"authors":"Wouter Van Gansbeke; Simon Vandenhende; Stamatios Georgoulis; Marc Proesmans; Luc Van Gool","journal":"Springer","ref_id":"b58","title":"Scan: Learning to classify images without labels","year":"2020-08-23"},{"authors":"Sagar Vaze; Kai Han; Andrea Vedaldi; Andrew Zisserman","journal":"","ref_id":"b59","title":"Generalized category discovery","year":"2022"},{"authors":"Kiri Wagstaff; Claire Cardie; Seth Rogers; Stefan Schrödl","journal":"Icml","ref_id":"b60","title":"Constrained k-means clustering with background knowledge","year":"2001"},{"authors":"Liang Wang; Nan Yang; Xiaolong Huang; Binxing Jiao; Linjun Yang; Daxin Jiang; Rangan Majumder; Furu Wei","journal":"","ref_id":"b61","title":"Text embeddings by weaklysupervised contrastive pre-training","year":"2022"},{"authors":"Zihan Wang; Jingbo Shang; Ruiqi Zhong","journal":"","ref_id":"b62","title":"Goal-driven explainable clustering via language descriptions","year":"2023"},{"authors":"Joe H Ward Jr","journal":"Journal of the American statistical association","ref_id":"b63","title":"Hierarchical grouping to optimize an objective function","year":"1963"},{"authors":"Junyuan Xie; Ross Girshick; Ali Farhadi","journal":"","ref_id":"b64","title":"Unsupervised deep embedding for clustering analysis","year":"2016"},{"authors":" Pmlr","journal":"","ref_id":"b65","title":"","year":""},{"authors":"Hui Xu; Yi Liu; Chi-Min Shu; Mingqi Bai; Mailidan Motalifu; Zhongxu He; Shuncheng Wu; Penggang Zhou; Bing Li","journal":"Journal of Loss Prevention in the Process Industries","ref_id":"b66","title":"Cause analysis of hot work accidents based on text mining and deep learning","year":"2022"},{"authors":"Jiaming Xu; Peng Wang; Guanhua Tian; Bo Xu; Jun Zhao; Fangyuan Wang; Hongwei Hao","journal":"Association for Computational Linguistics","ref_id":"b67","title":"Short text clustering via convolutional neural networks","year":"2015"},{"authors":"Kouta Nakata; Yaling Tao; Kentaro Takagi","journal":"","ref_id":"b68","title":"Clustering-friendly representation learning via instance discrimination and feature decorrelation","year":"2021"},{"authors":"Jianwei Yang; Devi Parikh; Dhruv Batra","journal":"","ref_id":"b69","title":"Joint unsupervised learning of deep representations and image clusters","year":"2016"},{"authors":"Muli Yang; Yuehua Zhu; Jiaping Yu; Aming Wu; Cheng Deng","journal":"","ref_id":"b70","title":"Divide and conquer: Compositional experts for generalized novel class discovery","year":"2022"},{"authors":"Dejiao Zhang; Feng Nan; Xiaokai Wei; Shang-Wen Li; Henghui Zhu; Kathleen Mckeown; Ramesh Nallapati; Andrew O Arnold; Bing Xiang","journal":"Association for Computational Linguistics","ref_id":"b71","title":"Supporting clustering with contrastive learning","year":"2021"},{"authors":"Hanlei Zhang; Hua Xu; Rui Ting-En Lin; Lyu","journal":"","ref_id":"b72","title":"Discovering new intents with deep aligned clustering","year":"2021"},{"authors":"Hanlei Zhang; Hua Xu; Xin Wang; Fei Long; Kai Gao","journal":"","ref_id":"b73","title":"Usnid: A framework for unsupervised and semi-supervised new intent discovery","year":"2023"},{"authors":"Hongjing Zhang; Sugato Basu; Ian Davidson","journal":"Springer","ref_id":"b74","title":"A framework for deep constrained clusteringalgorithms and advances","year":"2019"},{"authors":"Yuwei Zhang; Haode Zhang; Li-Ming Zhan; Xiao-Ming Wu; Albert Lam","journal":"Association for Computational Linguistics","ref_id":"b75","title":"New intent discovery with pre-training and contrastive learning","year":"2022"},{"authors":"Sheng Zhou; Hongjia Xu; Zhuonan Zheng; Jiawei Chen; Jiajun Bu; Jia Wu; Xin Wang; Wenwu Zhu; Martin Ester","journal":"","ref_id":"b76","title":"A comprehensive survey on deep clustering: Taxonomy, challenges, and future directions","year":"2022"},{"authors":"Yiming Zhu; Peixian Zhang; Ehsan-Ul Haq; Pan Hui; Gareth Tyson","journal":"","ref_id":"b77","title":"Can chatgpt reproduce human-generated labels? a study of social computing tasks","year":"2023"}],"string":"[\n {\n \"authors\": \"Wenbin An; Feng Tian; Qinghua Zheng; Wei Ding; Qianying Wang; Ping Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Generalized category discovery with decoupled prototypical network\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yuntao Bai; Saurav Kadavath; Sandipan Kundu; Amanda Askell; Jackson Kernion; Andy Jones; Anna Chen; Anna Goldie; Azalia Mirhoseini; Cameron Mckinnon\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Constitutional ai: Harmlessness from ai feedback\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Sugato Basu; Arindam Banerjee; Raymond J Mooney\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Active semi-supervision for pairwise constrained clustering\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Sugato Basu; Ian Davidson; Kiri Wagstaff\",\n \"journal\": \"CRC Press\",\n \"ref_id\": \"b3\",\n \"title\": \"Constrained clustering: Advances in algorithms, theory, and applications\",\n \"year\": \"2008\"\n },\n {\n \"authors\": \"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b4\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Kaidi Cao; Maria Brbic; Jure Leskovec\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Open-world semi-supervised learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Mathilde Caron; Piotr Bojanowski; Armand Joulin; Matthijs Douze\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Deep clustering for unsupervised learning of visual features\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Iñigo Casanueva; Tadas Temcinas; Daniela Gerz; Matthew Henderson; Ivan Vulic\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Efficient intent detection with dual sentence encoders\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jianlong Chang; Lingfeng Wang; Gaofeng Meng; Shiming Xiang; Chunhong Pan\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Deep adaptive image clustering\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Jeff Cheeger\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"A lower bound for the smallest eigenvalue of the laplacian, problems in analysis\",\n \"year\": \"1969\"\n },\n {\n \"authors\": \"Qinyuan Cheng; Xiaogui Yang; Tianxiang Sun; Linyang Li; Xipeng Qiu\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Improving contrastive learning of sentence embeddings from ai feedback\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Wei-Lin Chiang; Xuanqing Liu; Si Si; Yang Li; Samy Bengio; Cho-Jui Hsieh\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Pierre Colombo; Nathan Noiry; Ekhine Irurozki; Stephan Clemencon\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"What are the best systems? new perspectives on nlp benchmarking\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Dorottya Demszky; Dana Movshovitz-Attias; Jeongwoo Ko; Alan Cowen; Gaurav Nemade; Sujith Ravi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b13\",\n \"title\": \"GoEmotions: A dataset of fine-grained emotions\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Shumin Deng; Ningyu Zhang; Jiaojian Kang; Yichi Zhang; Wei Zhang; Huajun Chen\",\n \"journal\": \"Association for Computing Machinery\",\n \"ref_id\": \"b14\",\n \"title\": \"Metalearning with dynamic-memory-based prototypical network for few-shot event detection\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Nat Dilokthanakul; Pedro Am Mediano; Marta Garnelo; C H Matthew; Hugh Lee; Kai Salimbeni; Murray Arulkumaran; Shanahan\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Deep unsupervised clustering with gaussian mixture variational autoencoders\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Ning Ding; Guangwei Xu; Yulin Chen; Xiaobin Wang; Xu Han; Pengjun Xie; Haitao Zheng; Zhiyuan Liu\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b16\",\n \"title\": \"Few-NERD: A few-shot named entity recognition dataset\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"E William; Alan J Donath; Hoffman\",\n \"journal\": \"IBM Technical Disclosure Bulletin\",\n \"ref_id\": \"b17\",\n \"title\": \"Algorithms for partitioning of graphs and computer logic based on eigenvectors of connection matrices\",\n \"year\": \"1972\"\n },\n {\n \"authors\": \"Jack Fitzgerald; Christopher Hench; Charith Peris; Scott Mackie; Kay Rottmann; Ana Sanchez; Aaron Nash; Liam Urbach; Vishesh Kakarala; Richa Singh; Swetha Ranganath; Laurie Crist; Misha Britan; Wouter Leeuwis; Gokhan Tur; Prem Natarajan\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Massive: A 1m-example multilingual natural language understanding dataset with 51 typologically-diverse languages\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tianyu Gao; Xu Han; Hao Zhu; Zhiyuan Liu; Peng Li; Maosong Sun; Jie Zhou\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b19\",\n \"title\": \"FewRel 2.0: Towards more challenging few-shot relation classification\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Tianyu Gao; Xingcheng Yao; Danqi Chen\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b20\",\n \"title\": \"SimCSE: Simple contrastive learning of sentence embeddings\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Gregor Geigle; Nils Reimers; Andreas Rücklé; Iryna Gurevych\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"TWEAC: transformer with extendable QA agent classifiers\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Fabrizio Gilardi; Meysam Alizadeh; Maël Kubli\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Chatgpt outperforms crowd-workers for textannotation tasks\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Cyril Goutte; Lars Kai Hansen; Matthew G Liptrot; Egill Rostrup\",\n \"journal\": \"Human brain mapping\",\n \"ref_id\": \"b23\",\n \"title\": \"Feature-space clustering for fmri meta-analysis\",\n \"year\": \"2001\"\n },\n {\n \"authors\": \"Amir Hadifar; Lucas Sterckx; Thomas Demeester; Chris Develder\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b24\",\n \"title\": \"A self-training approach for short text clustering\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Xingwei He; Zhenghao Lin; Yeyun Gong; Hang Jin; Chen Zhang; Jian Lin; Jiao; Ming Siu; Nan Yiu; Weizhu Duan; Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Annollm: Making large language models to be better crowdsourced annotators\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"William Hogan; Jiacheng Li; Jingbo Shang\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Open-world semi-supervised generalized relation discovery aligned in a real-world setting\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Peihao Huang; Yan Huang; Wei Wang; Liang Wang\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b27\",\n \"title\": \"Deep embedding network for clustering\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \" Harold W Kuhn\",\n \"journal\": \"Naval research logistics quarterly\",\n \"ref_id\": \"b28\",\n \"title\": \"The hungarian method for the assignment problem\",\n \"year\": \"1955\"\n },\n {\n \"authors\": \"Stefan Larson; Anish Mahendran; Joseph J Peper; Christopher Clarke; Andrew Lee; Parker Hill; Jonathan K Kummerfeld; Kevin Leach; Michael A Laurenzano; Lingjia Tang; Jason Mars\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"An evaluation dataset for intent classification and out-ofscope prediction\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Haoran Li; Abhinav Arora; Shuohui Chen; Anchit Gupta; Sonal Gupta; Yashar Mehdad\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b30\",\n \"title\": \"MTOP: A comprehensive multilingual task-oriented semantic parsing benchmark\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Sha Li; Ji Heng; Jiawei Han\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Open relation and event type discovery with type abstraction\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hua Ting-En Lin; Hanlei Xu; Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Discovering new intents via constrained deep adaptive clustering with cluster refinement\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \" Macqueen\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Classification and analysis of multivariate observations\",\n \"year\": \"1967\"\n },\n {\n \"authors\": \"Laura Manduchi; Kieran Chin-Cheong; Holger Michel; Sven Wellmann; Julia E Vogt\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Deep conditional gaussian mixture model for constrained clustering\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"José Manuel; Guaita Martínez; Patricia Carracedo; Dolores Gorgues Comas; Carlos H Siemens\",\n \"journal\": \"Sustainable Technology and Entrepreneurship\",\n \"ref_id\": \"b35\",\n \"title\": \"An analysis of the blockchain and covid-19 research landscape using a bibliometric study\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yutao Mou; Yanan He; Zhiyuan Wu; Hong Zeng; Huixing Xu; Wei Jiang; Weiran Wu; Xu\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b36\",\n \"title\": \"Disentangled knowledge transfer for OOD intent discovery with unified contrastive learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Niklas Muennighoff; Nouamane Tazi; Loïc Magne; Nils Reimers\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Mteb: Massive text embedding benchmark\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Fionn Murtagh; Pedro Contreras\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Methods of hierarchical clustering\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"Jianmo Ni; Gustavo Hernandez Abrego; Noah Constant; Ji Ma; Keith Hall; Daniel Cer; Yinfei Yang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b39\",\n \"title\": \"Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jianmo Ni; Chen Qu; Jing Lu; Zhuyun Dai; Gustavo Hernandez Abrego; Ji Ma; Vincent Zhao; Yi Luan; Keith Hall; Ming-Wei Chang; Yinfei Yang\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Large dual encoders are generalizable retrievers\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Chuang Niu; Hongming Shan; Ge Wang\",\n \"journal\": \"IEEE Transactions on Image Processing\",\n \"ref_id\": \"b41\",\n \"title\": \"Spice: Semantic pseudo-labeling for image clustering\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Chuang Niu; Jun Zhang; Ge Wang; Jimin Liang\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Gatcluster: Self-supervised gaussian-attention network for image clustering\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"M Robert; Nosofsky\",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"The generalized context model: An exemplar model of classification\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \" Openai\",\n \"journal\": \"\",\n \"ref_id\": \"b44\",\n \"title\": \"Gpt-4 technical report\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Long Ouyang; Jeffrey Wu; Xu Jiang; Diogo Almeida; Carroll Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b45\",\n \"title\": \"Training language models to follow instructions with human feedback\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"June Young Park; Evan Mistur; Donghwan Kim; Yunjeong Mo; Richard Hoefer\",\n \"journal\": \"Sustainable Cities and Society\",\n \"ref_id\": \"b46\",\n \"title\": \"Toward humancentric urban infrastructure: Text mining for social media data to identify the public perception of covid-19 policy in transportation hubs\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"F Pedregosa; G Varoquaux; A Gramfort; V Michel; B Thirion; O Grisel; M Blondel; P Prettenhofer; R Weiss; V Dubourg; J Vanderplas; A Passos; D Cournapeau; M Brucher; M Perrot; E Duchesnay\",\n \"journal\": \"Journal of Machine Learning Research\",\n \"ref_id\": \"b47\",\n \"title\": \"Scikit-learn: Machine learning in Python\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"Dan Pelleg; Andrew W Moore\",\n \"journal\": \"Icml\",\n \"ref_id\": \"b48\",\n \"title\": \"X-means: Extending k-means with efficient estimation of the number of clusters\",\n \"year\": \"2000\"\n },\n {\n \"authors\": \"Maarten De Raedt; Fréderic Godin; Thomas Demeester; Chris Develder\",\n \"journal\": \"\",\n \"ref_id\": \"b49\",\n \"title\": \"Idas: Intent discovery with abstractive summarization\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Nils Reimers; Iryna Gurevych\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b50\",\n \"title\": \"Sentence-BERT: Sentence embeddings using Siamese BERTnetworks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Mamshad Nayeem Rizve; Navid Kardan; Salman Khan; Fahad Shahbaz Khan; Mubarak Shah\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b51\",\n \"title\": \"Openldn: Learning to discover novel classes for openworld semi-supervised learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Mamshad Nayeem Rizve; Navid Kardan; Mubarak Shah\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b52\",\n \"title\": \"Towards realistic semi-supervised learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"J Peter; Rousseeuw\",\n \"journal\": \"Journal of computational and applied mathematics\",\n \"ref_id\": \"b53\",\n \"title\": \"Silhouettes: a graphical aid to the interpretation and validation of cluster analysis\",\n \"year\": \"1987\"\n },\n {\n \"authors\": \"Hongjin Su; Weijia Shi; Jungo Kasai; Yizhong Wang; Yushi Hu; Mari Ostendorf; Wen-Tau Yih; Noah A Smith; Luke Zettlemoyer; Tao Yu\",\n \"journal\": \"\",\n \"ref_id\": \"b54\",\n \"title\": \"One embedder, any task: Instruction-finetuned text embeddings\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yiyou Sun; Yixuan Li\",\n \"journal\": \"Transactions on Machine Learning Research\",\n \"ref_id\": \"b55\",\n \"title\": \"Opencon: Open-world contrastive learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Robert Thorndike\",\n \"journal\": \"Psychometrika\",\n \"ref_id\": \"b56\",\n \"title\": \"Who belongs in the family?\",\n \"year\": \"1953\"\n },\n {\n \"authors\": \"Laurens Van Der Maaten; Geoffrey Hinton\",\n \"journal\": \"Journal of machine learning research\",\n \"ref_id\": \"b57\",\n \"title\": \"Visualizing data using t-sne\",\n \"year\": \"2008\"\n },\n {\n \"authors\": \"Wouter Van Gansbeke; Simon Vandenhende; Stamatios Georgoulis; Marc Proesmans; Luc Van Gool\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b58\",\n \"title\": \"Scan: Learning to classify images without labels\",\n \"year\": \"2020-08-23\"\n },\n {\n \"authors\": \"Sagar Vaze; Kai Han; Andrea Vedaldi; Andrew Zisserman\",\n \"journal\": \"\",\n \"ref_id\": \"b59\",\n \"title\": \"Generalized category discovery\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Kiri Wagstaff; Claire Cardie; Seth Rogers; Stefan Schrödl\",\n \"journal\": \"Icml\",\n \"ref_id\": \"b60\",\n \"title\": \"Constrained k-means clustering with background knowledge\",\n \"year\": \"2001\"\n },\n {\n \"authors\": \"Liang Wang; Nan Yang; Xiaolong Huang; Binxing Jiao; Linjun Yang; Daxin Jiang; Rangan Majumder; Furu Wei\",\n \"journal\": \"\",\n \"ref_id\": \"b61\",\n \"title\": \"Text embeddings by weaklysupervised contrastive pre-training\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Zihan Wang; Jingbo Shang; Ruiqi Zhong\",\n \"journal\": \"\",\n \"ref_id\": \"b62\",\n \"title\": \"Goal-driven explainable clustering via language descriptions\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Joe H Ward Jr\",\n \"journal\": \"Journal of the American statistical association\",\n \"ref_id\": \"b63\",\n \"title\": \"Hierarchical grouping to optimize an objective function\",\n \"year\": \"1963\"\n },\n {\n \"authors\": \"Junyuan Xie; Ross Girshick; Ali Farhadi\",\n \"journal\": \"\",\n \"ref_id\": \"b64\",\n \"title\": \"Unsupervised deep embedding for clustering analysis\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \" Pmlr\",\n \"journal\": \"\",\n \"ref_id\": \"b65\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Hui Xu; Yi Liu; Chi-Min Shu; Mingqi Bai; Mailidan Motalifu; Zhongxu He; Shuncheng Wu; Penggang Zhou; Bing Li\",\n \"journal\": \"Journal of Loss Prevention in the Process Industries\",\n \"ref_id\": \"b66\",\n \"title\": \"Cause analysis of hot work accidents based on text mining and deep learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jiaming Xu; Peng Wang; Guanhua Tian; Bo Xu; Jun Zhao; Fangyuan Wang; Hongwei Hao\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b67\",\n \"title\": \"Short text clustering via convolutional neural networks\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Kouta Nakata; Yaling Tao; Kentaro Takagi\",\n \"journal\": \"\",\n \"ref_id\": \"b68\",\n \"title\": \"Clustering-friendly representation learning via instance discrimination and feature decorrelation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jianwei Yang; Devi Parikh; Dhruv Batra\",\n \"journal\": \"\",\n \"ref_id\": \"b69\",\n \"title\": \"Joint unsupervised learning of deep representations and image clusters\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Muli Yang; Yuehua Zhu; Jiaping Yu; Aming Wu; Cheng Deng\",\n \"journal\": \"\",\n \"ref_id\": \"b70\",\n \"title\": \"Divide and conquer: Compositional experts for generalized novel class discovery\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Dejiao Zhang; Feng Nan; Xiaokai Wei; Shang-Wen Li; Henghui Zhu; Kathleen Mckeown; Ramesh Nallapati; Andrew O Arnold; Bing Xiang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b71\",\n \"title\": \"Supporting clustering with contrastive learning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Hanlei Zhang; Hua Xu; Rui Ting-En Lin; Lyu\",\n \"journal\": \"\",\n \"ref_id\": \"b72\",\n \"title\": \"Discovering new intents with deep aligned clustering\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Hanlei Zhang; Hua Xu; Xin Wang; Fei Long; Kai Gao\",\n \"journal\": \"\",\n \"ref_id\": \"b73\",\n \"title\": \"Usnid: A framework for unsupervised and semi-supervised new intent discovery\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Hongjing Zhang; Sugato Basu; Ian Davidson\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b74\",\n \"title\": \"A framework for deep constrained clusteringalgorithms and advances\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Yuwei Zhang; Haode Zhang; Li-Ming Zhan; Xiao-Ming Wu; Albert Lam\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b75\",\n \"title\": \"New intent discovery with pre-training and contrastive learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Sheng Zhou; Hongjia Xu; Zhuonan Zheng; Jiawei Chen; Jiajun Bu; Jia Wu; Xin Wang; Wenwu Zhu; Martin Ester\",\n \"journal\": \"\",\n \"ref_id\": \"b76\",\n \"title\": \"A comprehensive survey on deep clustering: Taxonomy, challenges, and future directions\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yiming Zhu; Peixian Zhang; Ehsan-Ul Haq; Pan Hui; Gareth Tyson\",\n \"journal\": \"\",\n \"ref_id\": \"b77\",\n \"title\": \"Can chatgpt reproduce human-generated labels? a study of social computing tasks\",\n \"year\": \"2023\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,144.66,439.08,145.21,10.69],"formula_id":"formula_0","formula_text":"c j = P T (I T , t),(1)"},{"formula_coordinates":[3,70.87,622.74,109.26,14],"formula_id":"formula_1","formula_text":"Z = {z i = f (x i )} N i=1 ."},{"formula_coordinates":[3,306.42,246.4,173.17,27.3],"formula_id":"formula_2","formula_text":"12 c1 ∼ C1, c2 ∼ C2, t = (a, c1, c2); 13 if t not in {tq},c1 ̸ = a, c2 ̸ = a then 14"},{"formula_coordinates":[3,332.26,418.48,192.88,34.21],"formula_id":"formula_3","formula_text":"p ik = (1 + ||z i -µ k || 2 /α) -α+1 2 k ′ (1 + ||z i -µ k ′ || 2 /α) -α+1 2 (2)"},{"formula_coordinates":[3,365.8,527.94,159.34,10.82],"formula_id":"formula_4","formula_text":"K closest = max(ϵK, 2)(3)"},{"formula_coordinates":[3,359.47,603.11,165.68,34.38],"formula_id":"formula_5","formula_text":"h i = - K closest k=1 p ′ ik log(p ′ ik )(4)"},{"formula_coordinates":[3,334.11,646.88,76.17,22.07],"formula_id":"formula_6","formula_text":"p ′ ik = p ik K closest k ′ =1 p ik ′"},{"formula_coordinates":[4,116.57,545.89,173.29,28.17],"formula_id":"formula_7","formula_text":"l j = exp (s(a, c j )/τ ) c l ∈B exp (s(a, c l )/τ )(5)"},{"formula_coordinates":[4,359.84,201.94,165.3,14.27],"formula_id":"formula_8","formula_text":"w = P P (I P , {p d } D d=1 , p)(6)"},{"formula_coordinates":[4,354,567.44,171.14,20.97],"formula_id":"formula_9","formula_text":"k * = argmax k M(W p , W k ) (7)"},{"formula_coordinates":[4,335.4,597.07,68.62,16.16],"formula_id":"formula_10","formula_text":"W p = {w p i } Np i=1"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 144.66,\n 439.08,\n 145.21,\n 10.69\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"c j = P T (I T , t),(1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 70.87,\n 622.74,\n 109.26,\n 14\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"Z = {z i = f (x i )} N i=1 .\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 306.42,\n 246.4,\n 173.17,\n 27.3\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"12 c1 ∼ C1, c2 ∼ C2, t = (a, c1, c2); 13 if t not in {tq},c1 ̸ = a, c2 ̸ = a then 14\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 332.26,\n 418.48,\n 192.88,\n 34.21\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"p ik = (1 + ||z i -µ k || 2 /α) -α+1 2 k ′ (1 + ||z i -µ k ′ || 2 /α) -α+1 2 (2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 365.8,\n 527.94,\n 159.34,\n 10.82\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"K closest = max(ϵK, 2)(3)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 359.47,\n 603.11,\n 165.68,\n 34.38\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"h i = - K closest k=1 p ′ ik log(p ′ ik )(4)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 334.11,\n 646.88,\n 76.17,\n 22.07\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"p ′ ik = p ik K closest k ′ =1 p ik ′\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 116.57,\n 545.89,\n 173.29,\n 28.17\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"l j = exp (s(a, c j )/τ ) c l ∈B exp (s(a, c l )/τ )(5)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 359.84,\n 201.94,\n 165.3,\n 14.27\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"w = P P (I P , {p d } D d=1 , p)(6)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 354,\n 567.44,\n 171.14,\n 20.97\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"k * = argmax k M(W p , W k ) (7)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 335.4,\n 597.07,\n 68.62,\n 16.16\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"W p = {w p i } Np i=1\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"CLUSTERLLM: Large Language Models as a Guide for Text Clustering"},"abstract":{"kind":"string","value":"We introduce CLUSTERLLM, a novel text clustering framework that leverages feedback from an instruction-tuned large language model, such as ChatGPT. Compared with traditional unsupervised methods that builds upon \"small\" embedders, CLUSTERLLM exhibits two intriguing advantages: (1) it enjoys the emergent capability of LLM even if its embeddings are inaccessible; and (2) it understands the user's preference on clustering through textual instruction and/or a few annotated data. First, we prompt ChatGPT for insights on clustering perspective by constructing hard triplet questions <does A better correspond to B than C>, where A, B and C are similar data points that belong to different clusters according to small embedder. We empirically show that this strategy is both effective for fine-tuning small embedder and cost-efficient to query ChatGPT. Second, we prompt ChatGPT for helps on clustering granularity by carefully designed pairwise questions <do A and B belong to the same category>, and tune the granularity from cluster hierarchies that is the most consistent with the ChatGPT answers. Extensive experiments on 14 datasets show that CLUSTERLLM consistently improves clustering quality, at an average cost of ∼$0.6 1 per dataset."},"authors":{"kind":"string","value":"Yuwei Zhang; Zihan Wang; Jingbo Shang"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: LLMs like ChatGPT are not applicable for text clustering directly because of the inaccessible embeddings. CLUSTERLLM resolves the dilemma by leveraging LLM as a guide on text clustering.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: An overview of CLUSTERLLM. It utilizes LLM to guide an embedder for text clustering with a low cost.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Relative clustering accuracy (divided by maximum for better aligning across datasets) of CLUSTER-LLM-GPT3.5 with different range of entropy selected.x-axis shows the mean of interval where interval length is set to 20%. For example, \"mean of interval= 50%\" means γ high = 40% and γ low = 60% (see Section 3.1.1). ♦ marks the setting for main experiments.","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Consistency score v.s. various number of clusters with GPT-4 and λ = 1.","figure_data":"","figure_id":"fig_3","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: Relative clustering accuracy (divided by maximum for better aligning across datasets) of CLUSTER-LLM-GPT3.5 over 4 iterations.","figure_data":"","figure_id":"fig_4","figure_label":"5","figure_type":"figure"},{"figure_caption":"Figure6: Scatter plots for t-SNE of embeddings. We select 10 classes from each datasets, denoted by colors.","figure_data":"","figure_id":"fig_5","figure_label":"","figure_type":"figure"},{"figure_caption":"Entropy-based Triplet SamplingInput: embeddings Z = {zi = f (xi)} N i=1 ,interval boundaries γhigh and γlow, closest clusters fraction ϵ = 2%, maximum number of queries Q. Compute µ k for each cluster k by averaging; 3 Compute Kclosest with Eq. 3; 4 Compute entropy H with Eq. 4;","figure_data":"5 Ind ← argsort(H)[::-1];6 Ind ← Ind[γhighN :γlowN ];7 Step 2: Initialize triplets {tq} ← {};8 while len({tq})< Q do9for a in Ind do10Obtain Kclosest closest clusters;11Sample C1, C2 from closest clusters;","figure_id":"tab_0","figure_label":"","figure_type":"table"},{"figure_caption":"Dataset statistics.","figure_data":"TaskName#clusters #data(small) #data(large)Bank77773,08010,003IntentCLINC(I) MTOP(I)150 1024,500 4,38615,000 15,638Massive(I)592,97411,510FewRel644,48040,320TypeFewNerd583,78950,000FewEvent344,74218,969StackEx1214,15650,000TopicArxivS2S933,67450,000Reddit503,21750,000EmotionGoEmo275,94023,485CLINC(D)104,50015,000DomainMTOP(D)114,38615,667Massive(D)182,97411,514","figure_id":"tab_1","figure_label":"1","figure_type":"table"},{"figure_caption":"Zhang et al., 2021a) 63.60(1.37) 77.34(0.62) 77.96(1.78) 91.89(0.49) 33.82(1.07) 70.42(0.34) 54.48(1.80) 71.57(0.89) 22.03(0.69) 20.05(0.63) Comparison of clustering accuracy and NMI with known granularity for evaluation. Average over all 14 datasets are shown in the last two columns. Best results are bolded.","figure_data":"The first one is clustering accuracy calculated afterHungarian alignment (Kuhn, 1955) that permuteprediction classes back to label classes. Anotherpopular metric for clustering is normalized mutualinformation (NMI) that calculates mutual informa-tion between two assignments, and normalized bytheir individual entropies.","figure_id":"tab_2","figure_label":"2","figure_type":"table"},{"figure_caption":"Analysis on the triplet prediction accuracy ( † is used to produce results of CLUSTERLLM-I in Table2). Red and green mean decreased or increased performances respectively. \"#GT Triplets\" means triplets that have ground truth (see Section 4.5 for details).","figure_data":"","figure_id":"tab_4","figure_label":"3","figure_type":"table"},{"figure_caption":"Ablation study on clustering quality with Instructor as backbone and known granularity for evaluation. See more results with large-scale datasets in Table6.","figure_data":"Intent DiscoveryEmotionMethodBank77CLINC(I)MTOP(I)Massive(I)GoEmoACCNMIACCNMIACCNMIACCNMIACCNMICLUSTERLLM-GPT3.5(random)59.88(2.56) 79.69(0.63) 74.40(0.91) 90.38(0.20) 28.05(1.69) 61.76(0.62) 51.66(2.41) 68.87(0.73) 28.62(1.95) 25.88(1.02)CLUSTERLLM-GPT3.570.77(0.49) 85.07(0.33) 82.77(1.20) 93.88(0.17) 35.84(2.07) 73.52(0.38) 59.89(2.05) 76.96(0.54) 27.49(1.25) 24.78(0.56)CLUSTERLLM-GPT469.71(1.13) 84.68(0.40) 81.91(1.20) 93.76(0.24) 34.48(0.38) 73.57(0.40) 59.10(1.12) 76.59(0.41) 27.41(1.13) 23.77(0.42)CLUSTERLLM-GT&GPT3.571.35(1.97) 85.12(0.45) 84.00(1.04) 94.34(0.30) 36.86(0.42) 75.36(0.08) 59.27(1.43) 77.37(0.54) 30.91(1.16) 27.71(0.46)Type DiscoveryTopic MiningMethodFewRelFewNerdFewEventStackExRedditACCNMIACCNMIACCNMIACCNMIACCNMICLUSTERLLM-GPT3.5(random)40.65(0.89) 56.54(0.30) 27.15(0.53) 43.56(0.49) 44.23(1.72) 66.75(0.87) 40.81(0.94) 62.10(0.34) 54.60(2.23) 61.82(1.64)CLUSTERLLM-GPT3.547.94(1.37) 62.43(0.43) 34.75(1.58) 51.03(0.57) 46.17(2.18) 70.73(0.34) 47.21(1.07) 66.78(0.29) 56.79(1.90) 63.87(0.56)CLUSTERLLM-GPT448.96(1.14) 63.58(0.39) 37.54(0.54) 53.94(0.27) 47.98(1.45) 71.32(0.70) 46.82(0.78) 66.72(0.11) 55.38(0.37) 63.45(0.49)CLUSTERLLM-GT&GPT3.548.91(1.20) 63.34(0.47) 37.27(0.61) 53.57(0.32) 48.12(1.52) 72.31(0.84) 47.55(1.17) 67.04(0.31) 58.33(1.26) 65.34(0.51)MethodTopic Mining ArxivS2SMTOP(D)Domain Discovery Massive(D)CLINC(D)AvgACCNMIACCNMIACCNMIACCNMIACCNMICLUSTERLLM-GPT3.5(random)22.03(0.28) 45.50(0.16) 87.00(2.27) 82.09(1.54) 56.40(2.35) 64.39(1.12) 60.27(4.20) 58.11(2.93)48.2761.96CLUSTERLLM-GPT3.526.61(0.48) 50.06(0.26) 93.53(0.10) 89.36(0.11) 61.06(1.91) 68.62(0.90) 52.39(1.84) 54.98(2.08)53.0966.58CLUSTERLLM-GPT426.16(0.22) 50.06(0.20) 92.04(2.67) 88.39(1.33) 60.16(2.97) 67.98(1.04) 57.45(2.48) 59.98(1.14)53.2266.98CLUSTERLLM-GT&GPT3.526.14(0.57) 50.19(0.33) 92.26(3.62) 89.36(1.42) 61.65(3.50) 69.51(1.50) 52.87(2.63) 56.43(1.21)53.9667.64Relative Clustering Accuracy0.93 0.94 0.95 0.96 0.97 0.98 0.99 1.00Bank77 FewRel StackEx GoEmo10% 20% 30% 40% 50% 60% 70% 80% 90% Mean of Interval (Entropy)","figure_id":"tab_5","figure_label":"4","figure_type":"table"},{"figure_caption":"Cluster granularity on small-scale datasets. Maximum & minimum number of clusters are set to 200 & 2.The results are shown in format of \"[#clusters] (errors)\". \"Rank\" column is computed with 1-level ranking(Colombo et al., 2022) with inverse errors. \"GT\" is ground truth. See results for large-scale datasets in Table7.","figure_data":"MethodBank77FewRel Massive(I) Massive(D) MTOP(I) MTOP(D) CLINC(I) CLINC(D) RankGT #clusters776459181021115010-Silhouette (Rousseeuw, 1987)118 (53.25) 10 (84.38) 38 (35.59)41 (127.8)11 (89.22)11 (0)172 (14.67) 163 (1530)10Elbow (Thorndike, 1953)53 (31.17) 43 (32.81) 45 (23.73)46 (155.6)33 (67.65) 34 (209.1)66 (56.00)68 (580.0)9X-means (Pelleg et al., 2000)69 (10.39) 30 (53.13) 32 (45.76)34 (88.89)28 (72.55) 27 (145.5) 130 (13.33) 135 (1250)8BIC (Goutte et al., 2001)123 (59.74) 58 (9.38)56 (5.08)60 (233.3)69 (32.35) 64 (481.8) 167 (11.33) 176 (1660)7ClusterSize (Zhang et al., 2021b)86 (11.69) 71 (10.94) 72 (22.03)90 (400.0)82 (19.61) 85 (672.7) 105 (30.00) 106 (960.0)6Ours (GPT3.5,λ = 1)64 (16.88) 46 (28.13) 43 (27.12)90 (400.0)43 (57.84) 40 (263.6)151 (0.67)96 (860.0)5Ours (GPT3.5,λ = 3)64 (16.88) 46 (28.13) 52 (11.86)37 (105.6)92 (9.80)18 (63.63)142 (5.33) 107 (970.0)2Ours (GPT4,λ = 1)56 (27.27) 46 (28.13) 41 (30.51)20 (11.11)53 (48.04)8 (27.27)146 (2.67)39 (290.0)3Ours (GT,λ = 1)100 (29.87) 91 (42.19) 42 (28.81)18 (0)41 (59.80)11 (0)141 (6.00)39 (290.0)4Ours (GT,λ = 3)99 (28.57) 94 (46.88)62 (5.08)20 (11.11)37 (63.73)11 (0)142 (5.33)31 (210.0)1ing (Zhou et al., 2022) which focuses on how toleverage deep neural network in clustering. Zhouet al. (2022) has categorized deep clustering re-search into four types including multi-stage (Yal-ing Tao, 2021; Huang et al., 2014), iterative (Yang","figure_id":"tab_6","figure_label":"5","figure_type":"table"},{"figure_caption":"Ablation study on clustering quality for large-scale datasets.","figure_data":"is proposed to study a similar researchproblem in the domain of intent detection. Mostrecently, Hogan et al., 2023 adapts the setting intorelation type discovery. However, GCD relies onsufficient annotated and unlabeled data for training.In contrast, CLUSTERLLM seeks for minimalsupervision and studies a setting with controlled","figure_id":"tab_8","figure_label":"6","figure_type":"table"},{"figure_caption":"Inferred granularity on large-scale datasets. The setting is the same as in Table5.","figure_data":"Methodsmall-scalelarge-scaleMassive(I) Massive(D) CLINC(I) CLINC(D) Massive(I) Massive(D) CLINC(I) CLINC(D)GT591815010591815010Ours412014639581614332w/o Justification415013741803612932w/o Demonstration41641411087774120105","figure_id":"tab_9","figure_label":"7","figure_type":"table"},{"figure_caption":"Prompt designs in determining granularity. We use the Instructor embedding with prompts, and we report results of GPT-4 with λ = 1.","figure_data":"","figure_id":"tab_10","figure_label":"8","figure_type":"table"},{"figure_caption":"Entropy of class distribution.","figure_data":"TaskNameentropy(small) entropy(large)Bank771.001.00IntentCLINC(I) MTOP(I)1.00 0.741.00 0.75Massive(I)0.910.92FewRel1.001.00TypeFewNerd0.820.82FewEvent0.850.85StackEx0.980.98TopicArxivS2S1.001.00Reddit1.001.00EmotionGoEmo0.910.91CLINC(D)1.001.00DomainMTOP(D)0.980.98Massive(D)0.940.94DatasetPromptBank77","figure_id":"tab_11","figure_label":"9","figure_type":"table"},{"figure_caption":"Prompts for Instructor.","figure_data":"","figure_id":"tab_12","figure_label":"10","figure_type":"table"},{"figure_caption":"One example from Bank77 on both triplet task and pairwise task.","figure_data":"","figure_id":"tab_13","figure_label":"12","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: LLMs like ChatGPT are not applicable for text clustering directly because of the inaccessible embeddings. CLUSTERLLM resolves the dilemma by leveraging LLM as a guide on text clustering.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: An overview of CLUSTERLLM. It utilizes LLM to guide an embedder for text clustering with a low cost.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Relative clustering accuracy (divided by maximum for better aligning across datasets) of CLUSTER-LLM-GPT3.5 with different range of entropy selected.x-axis shows the mean of interval where interval length is set to 20%. For example, \\\"mean of interval= 50%\\\" means γ high = 40% and γ low = 60% (see Section 3.1.1). ♦ marks the setting for main experiments.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Consistency score v.s. various number of clusters with GPT-4 and λ = 1.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: Relative clustering accuracy (divided by maximum for better aligning across datasets) of CLUSTER-LLM-GPT3.5 over 4 iterations.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure6: Scatter plots for t-SNE of embeddings. We select 10 classes from each datasets, denoted by colors.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Entropy-based Triplet SamplingInput: embeddings Z = {zi = f (xi)} N i=1 ,interval boundaries γhigh and γlow, closest clusters fraction ϵ = 2%, maximum number of queries Q. Compute µ k for each cluster k by averaging; 3 Compute Kclosest with Eq. 3; 4 Compute entropy H with Eq. 4;\",\n \"figure_data\": \"5 Ind ← argsort(H)[::-1];6 Ind ← Ind[γhighN :γlowN ];7 Step 2: Initialize triplets {tq} ← {};8 while len({tq})< Q do9for a in Ind do10Obtain Kclosest closest clusters;11Sample C1, C2 from closest clusters;\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Dataset statistics.\",\n \"figure_data\": \"TaskName#clusters #data(small) #data(large)Bank77773,08010,003IntentCLINC(I) MTOP(I)150 1024,500 4,38615,000 15,638Massive(I)592,97411,510FewRel644,48040,320TypeFewNerd583,78950,000FewEvent344,74218,969StackEx1214,15650,000TopicArxivS2S933,67450,000Reddit503,21750,000EmotionGoEmo275,94023,485CLINC(D)104,50015,000DomainMTOP(D)114,38615,667Massive(D)182,97411,514\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Zhang et al., 2021a) 63.60(1.37) 77.34(0.62) 77.96(1.78) 91.89(0.49) 33.82(1.07) 70.42(0.34) 54.48(1.80) 71.57(0.89) 22.03(0.69) 20.05(0.63) Comparison of clustering accuracy and NMI with known granularity for evaluation. Average over all 14 datasets are shown in the last two columns. Best results are bolded.\",\n \"figure_data\": \"The first one is clustering accuracy calculated afterHungarian alignment (Kuhn, 1955) that permuteprediction classes back to label classes. Anotherpopular metric for clustering is normalized mutualinformation (NMI) that calculates mutual informa-tion between two assignments, and normalized bytheir individual entropies.\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Analysis on the triplet prediction accuracy ( † is used to produce results of CLUSTERLLM-I in Table2). Red and green mean decreased or increased performances respectively. \\\"#GT Triplets\\\" means triplets that have ground truth (see Section 4.5 for details).\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablation study on clustering quality with Instructor as backbone and known granularity for evaluation. See more results with large-scale datasets in Table6.\",\n \"figure_data\": \"Intent DiscoveryEmotionMethodBank77CLINC(I)MTOP(I)Massive(I)GoEmoACCNMIACCNMIACCNMIACCNMIACCNMICLUSTERLLM-GPT3.5(random)59.88(2.56) 79.69(0.63) 74.40(0.91) 90.38(0.20) 28.05(1.69) 61.76(0.62) 51.66(2.41) 68.87(0.73) 28.62(1.95) 25.88(1.02)CLUSTERLLM-GPT3.570.77(0.49) 85.07(0.33) 82.77(1.20) 93.88(0.17) 35.84(2.07) 73.52(0.38) 59.89(2.05) 76.96(0.54) 27.49(1.25) 24.78(0.56)CLUSTERLLM-GPT469.71(1.13) 84.68(0.40) 81.91(1.20) 93.76(0.24) 34.48(0.38) 73.57(0.40) 59.10(1.12) 76.59(0.41) 27.41(1.13) 23.77(0.42)CLUSTERLLM-GT&GPT3.571.35(1.97) 85.12(0.45) 84.00(1.04) 94.34(0.30) 36.86(0.42) 75.36(0.08) 59.27(1.43) 77.37(0.54) 30.91(1.16) 27.71(0.46)Type DiscoveryTopic MiningMethodFewRelFewNerdFewEventStackExRedditACCNMIACCNMIACCNMIACCNMIACCNMICLUSTERLLM-GPT3.5(random)40.65(0.89) 56.54(0.30) 27.15(0.53) 43.56(0.49) 44.23(1.72) 66.75(0.87) 40.81(0.94) 62.10(0.34) 54.60(2.23) 61.82(1.64)CLUSTERLLM-GPT3.547.94(1.37) 62.43(0.43) 34.75(1.58) 51.03(0.57) 46.17(2.18) 70.73(0.34) 47.21(1.07) 66.78(0.29) 56.79(1.90) 63.87(0.56)CLUSTERLLM-GPT448.96(1.14) 63.58(0.39) 37.54(0.54) 53.94(0.27) 47.98(1.45) 71.32(0.70) 46.82(0.78) 66.72(0.11) 55.38(0.37) 63.45(0.49)CLUSTERLLM-GT&GPT3.548.91(1.20) 63.34(0.47) 37.27(0.61) 53.57(0.32) 48.12(1.52) 72.31(0.84) 47.55(1.17) 67.04(0.31) 58.33(1.26) 65.34(0.51)MethodTopic Mining ArxivS2SMTOP(D)Domain Discovery Massive(D)CLINC(D)AvgACCNMIACCNMIACCNMIACCNMIACCNMICLUSTERLLM-GPT3.5(random)22.03(0.28) 45.50(0.16) 87.00(2.27) 82.09(1.54) 56.40(2.35) 64.39(1.12) 60.27(4.20) 58.11(2.93)48.2761.96CLUSTERLLM-GPT3.526.61(0.48) 50.06(0.26) 93.53(0.10) 89.36(0.11) 61.06(1.91) 68.62(0.90) 52.39(1.84) 54.98(2.08)53.0966.58CLUSTERLLM-GPT426.16(0.22) 50.06(0.20) 92.04(2.67) 88.39(1.33) 60.16(2.97) 67.98(1.04) 57.45(2.48) 59.98(1.14)53.2266.98CLUSTERLLM-GT&GPT3.526.14(0.57) 50.19(0.33) 92.26(3.62) 89.36(1.42) 61.65(3.50) 69.51(1.50) 52.87(2.63) 56.43(1.21)53.9667.64Relative Clustering Accuracy0.93 0.94 0.95 0.96 0.97 0.98 0.99 1.00Bank77 FewRel StackEx GoEmo10% 20% 30% 40% 50% 60% 70% 80% 90% Mean of Interval (Entropy)\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Cluster granularity on small-scale datasets. Maximum & minimum number of clusters are set to 200 & 2.The results are shown in format of \\\"[#clusters] (errors)\\\". \\\"Rank\\\" column is computed with 1-level ranking(Colombo et al., 2022) with inverse errors. \\\"GT\\\" is ground truth. See results for large-scale datasets in Table7.\",\n \"figure_data\": \"MethodBank77FewRel Massive(I) Massive(D) MTOP(I) MTOP(D) CLINC(I) CLINC(D) RankGT #clusters776459181021115010-Silhouette (Rousseeuw, 1987)118 (53.25) 10 (84.38) 38 (35.59)41 (127.8)11 (89.22)11 (0)172 (14.67) 163 (1530)10Elbow (Thorndike, 1953)53 (31.17) 43 (32.81) 45 (23.73)46 (155.6)33 (67.65) 34 (209.1)66 (56.00)68 (580.0)9X-means (Pelleg et al., 2000)69 (10.39) 30 (53.13) 32 (45.76)34 (88.89)28 (72.55) 27 (145.5) 130 (13.33) 135 (1250)8BIC (Goutte et al., 2001)123 (59.74) 58 (9.38)56 (5.08)60 (233.3)69 (32.35) 64 (481.8) 167 (11.33) 176 (1660)7ClusterSize (Zhang et al., 2021b)86 (11.69) 71 (10.94) 72 (22.03)90 (400.0)82 (19.61) 85 (672.7) 105 (30.00) 106 (960.0)6Ours (GPT3.5,λ = 1)64 (16.88) 46 (28.13) 43 (27.12)90 (400.0)43 (57.84) 40 (263.6)151 (0.67)96 (860.0)5Ours (GPT3.5,λ = 3)64 (16.88) 46 (28.13) 52 (11.86)37 (105.6)92 (9.80)18 (63.63)142 (5.33) 107 (970.0)2Ours (GPT4,λ = 1)56 (27.27) 46 (28.13) 41 (30.51)20 (11.11)53 (48.04)8 (27.27)146 (2.67)39 (290.0)3Ours (GT,λ = 1)100 (29.87) 91 (42.19) 42 (28.81)18 (0)41 (59.80)11 (0)141 (6.00)39 (290.0)4Ours (GT,λ = 3)99 (28.57) 94 (46.88)62 (5.08)20 (11.11)37 (63.73)11 (0)142 (5.33)31 (210.0)1ing (Zhou et al., 2022) which focuses on how toleverage deep neural network in clustering. Zhouet al. (2022) has categorized deep clustering re-search into four types including multi-stage (Yal-ing Tao, 2021; Huang et al., 2014), iterative (Yang\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablation study on clustering quality for large-scale datasets.\",\n \"figure_data\": \"is proposed to study a similar researchproblem in the domain of intent detection. Mostrecently, Hogan et al., 2023 adapts the setting intorelation type discovery. However, GCD relies onsufficient annotated and unlabeled data for training.In contrast, CLUSTERLLM seeks for minimalsupervision and studies a setting with controlled\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Inferred granularity on large-scale datasets. The setting is the same as in Table5.\",\n \"figure_data\": \"Methodsmall-scalelarge-scaleMassive(I) Massive(D) CLINC(I) CLINC(D) Massive(I) Massive(D) CLINC(I) CLINC(D)GT591815010591815010Ours412014639581614332w/o Justification415013741803612932w/o Demonstration41641411087774120105\",\n \"figure_id\": \"tab_9\",\n \"figure_label\": \"7\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Prompt designs in determining granularity. We use the Instructor embedding with prompts, and we report results of GPT-4 with λ = 1.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_10\",\n \"figure_label\": \"8\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Entropy of class distribution.\",\n \"figure_data\": \"TaskNameentropy(small) entropy(large)Bank771.001.00IntentCLINC(I) MTOP(I)1.00 0.741.00 0.75Massive(I)0.910.92FewRel1.001.00TypeFewNerd0.820.82FewEvent0.850.85StackEx0.980.98TopicArxivS2S1.001.00Reddit1.001.00EmotionGoEmo0.910.91CLINC(D)1.001.00DomainMTOP(D)0.980.98Massive(D)0.940.94DatasetPromptBank77\",\n \"figure_id\": \"tab_11\",\n \"figure_label\": \"9\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Prompts for Instructor.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_12\",\n \"figure_label\": \"10\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"One example from Bank77 on both triplet task and pairwise task.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_13\",\n \"figure_label\": \"12\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(MacQueen, 1967)\", \"Explanation\": \"The cited work by MacQueen (1967) introduces a clustering algorithm that is adopted in the citing paper to perform text clustering on pre-trained embedders.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Muennighoff et al., 2022)\", \"Explanation\": \"The cited work by Muennighoff et al. (2022) provides a pre-trained embedder that is utilized in the citing paper for text clustering.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wang et al., 2022)\", \"Explanation\": \"The cited work by Wang et al. (2022) also contributes a pre-trained embedder that is used in the citing paper for text clustering.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Su et al., 2022)\", \"Explanation\": \"The cited work by Su et al. (2022) provides another pre-trained embedder that the citing paper uses for text clustering.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Mart\\u00ednez et al., 2022)\", \"Explanation\": \"The cited work by Mart\\u00ednez et al. (2022) extends the research on text clustering by exploring emerging research topics, which the citing paper builds upon in its own study.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Park et al., 2022)\", \"Explanation\": \"The cited work by Park et al. (2022) also extends the research on text clustering by identifying public perception from social media, which the citing paper builds upon in its own study.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Xu et al., 2022)\", \"Explanation\": \"The cited work by Xu et al. (2022) further extends the research on text clustering by analysing cause of accidents, which the citing paper builds upon in its own study.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(OpenAI, 2023)\", \"Explanation\": \"The cited work by OpenAI (2023) introduces the recent GPT series of LLMs, which the citing paper extends by applying them in text clustering tasks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. (2020) also contributes to the research on text clustering by introducing the GPT series of LLMs, which the citing paper extends in its own study.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ouyang et al., 2022)\", \"Explanation\": \"The cited work by Ouyang et al. (2022) further extends the research on text clustering by introducing the latest GPT series of LLMs, which the citing paper builds upon in its own study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang et al., 2022)\", \"Explanation\": \"The cited work by Wang et al. provides the pre-trained small embedder that is used in the CLUSTERLLM framework as the basis for sentence representation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Su et al., 2022)\", \"Explanation\": \"The cited work by Su et al. also contributes to the pre-trained small embedder used in the CLUSTERLLM framework, providing a foundational method for sentence representation.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Nosofsky, 2011)\", \"Explanation\": \"The cited work by Nosofsky extends the human cognitive ability concept to the context of CLUSTERLLM, providing a new perspective for considering sentence pairs and triplets in the framework.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ouyang et al., 2022)\", \"Explanation\": \"The cited work by Ouyang et al. contributes to the training of LLMs in the CLUSTERLLM framework, providing a method for following human instructions in the context of sentence processing.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(OpenAI, 2023)\", \"Explanation\": \"The cited work by OpenAI is a foundational method for the use of LLMs in the CLUSTERLLM framework, providing a way to process sentence pairs and triplets in a more human-like manner.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xie et al., 2016)\", \"Explanation\": \"The cited work by Xie et al. (2016) provides a method for calculating instance-wise soft assignments using Student's t-distribution, which the citing paper adopts in their research to calculate the closest clusters for a given instance.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Van der Maaten and Hinton, 2008)\", \"Explanation\": \"The cited work by Van der Maaten and Hinton (2008) provides a method for calculating instance-wise soft assignments with Student's t-distribution, which the citing paper builds upon in their research to calculate the closest clusters for a given instance.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2020)\", \"Explanation\": \"The cited work by Zhang et al. provides a method for deep constrained clustering that the citing paper builds upon to address the issue of noisy labels in user perspective analysis.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Manduchi et al., 2021)\", \"Explanation\": \"The cited work by Manduchi et al. contributes a method for deep constrained clustering that the citing paper adopts to improve the analysis of user perspectives in a more effective way.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Basu et al., 2008)\", \"Explanation\": \"The cited work by Basu et al. highlights the sensitivity of deep constrained clustering to noisy labels, which the citing paper addresses by focusing on finetuning the base embedder to better explain user perspectives.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Su et al., 2022)\", \"Explanation\": \"The cited work by Su et al. provides a dataset of user perspectives that the citing paper utilizes in their research to improve the analysis of user perspectives in a more effective way.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ni et al., 2022b)\", \"Explanation\": \"The cited work by Ni et al. contributes a dataset of user perspectives that the citing paper uses to optimize the objective function for finetuning the base embedder in the analysis of user perspectives.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Su et al., 2022)\", \"Explanation\": \"The cited work by Su et al. also provides a method for optimizing the objective function in the finetuning of the base embedder to better explain user perspectives in the analysis.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ni et al., 2022b)\", \"Explanation\": \"The cited work by Ni et al. further contributes a method for optimizing the objective function in the finetuning of the base embedder to improve the analysis of user perspectives in a more effective way.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Pelleg et al., 2000)\", \"Explanation\": \"The cited work by Pelleg et al. (2000) provides the basis for the maximum and minimum number of clusters used in the citing paper to determine granularity. The method of sampling pairs from the two clusters to be merged is adopted from this work to form candidate pairs for querying LLMs.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2021b)\", \"Explanation\": \"The cited work provides the dataset used for intent discovery in unlabeled customer utterances, which the citing paper uses for their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2022)\", \"Explanation\": \"The cited work provides the dataset used for intent discovery in unlabeled customer utterances, which the citing paper uses for their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Li et al., 2022)\", \"Explanation\": \"The cited work provides the dataset used for type discovery in entity, relation, and event type discovery tasks, which the citing paper focuses on in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Muennighoff et al., 2022)\", \"Explanation\": \"The cited work provides the datasets used in the study conducted in the citing paper for topic mining.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Demszky et al., 2020)\", \"Explanation\": \"The cited work provides the GoEmo dataset used in the study conducted in the citing paper for emotion detection.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Su et al., 2022)\", \"Explanation\": \"The cited work by Su et al. provides the pre-trained embedder Instructor, which the citing paper uses in their research on large-scale datasets.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wang et al., 2022)\", \"Explanation\": \"The cited work by Wang et al. provides the pre-trained embedder E5, which the citing paper uses in their research on large-scale datasets.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Pedregosa et al., 2011)\", \"Explanation\": \"The cited work by Pedregosa et al. provides the implementation of the clustering algorithms used in the research, which the citing paper adopts in their work on small-scale datasets.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Xu et al., 2015)\", \"Explanation\": \"The cited work by Xu et al. provides foundational data and methodologies for text clustering, which the citing paper leverages in its research on the same topic.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Hadifar et al., 2019)\", \"Explanation\": \"The cited work by Hadifar et al. contributes to the field of text clustering by introducing new methods and techniques that the citing paper can build upon.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Zhang et al., 2021a)\", \"Explanation\": \"The cited work by Zhang et al. offers insights and data on text clustering that the citing paper can use to further advance the research in this area.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Yaling Tao, 2021)\", \"Explanation\": \"The cited work by Yaling Tao provides a valuable data source for image clustering research, which the citing paper can utilize in its own study of the topic.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Yang et al., 2016)\", \"Explanation\": \"The cited work by Yang et al. contributes to the field of image clustering by introducing new methods and techniques that the citing paper can build upon.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Caron et al., 2018)\", \"Explanation\": \"The cited work by Caron et al. offers valuable data and insights on image clustering that the citing paper can leverage in its own research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Niu et al., 2020)\", \"Explanation\": \"The cited work by Niu et al. provides a data source for image clustering research that the citing paper can utilize in its study of the topic.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Xie et al., 2016)\", \"Explanation\": \"The cited work by Xie et al. contributes to the field of image clustering by offering new methods and techniques that the citing paper can build upon.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Huang et al., 2014)\", \"Explanation\": \"The cited work by Huang et al. provides foundational data and methodologies for graph clustering research, which the citing paper can leverage in its own study of the topic.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Chiang et al., 2019)\", \"Explanation\": \"The cited work by Chiang et al. offers valuable data and insights on graph clustering that the citing paper can use to further advance the research in this area.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Yang et al., 2016)\", \"Explanation\": \"The cited work by Yang et al. provides a method of utilizing relational supervision via pseudolabelling, which the citing paper adopts in their own research to improve the quality of clustering results.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Niu et al., 2020)\", \"Explanation\": \"The cited work by Niu et al. contributes a method of using relational supervision in deep clustering, which the citing paper builds upon to enhance the performance of their own clustering approach.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Van Gansbeke et al., 2020)\", \"Explanation\": \"The cited work by Van Gansbeke et al. provides a method of using relational supervision in deep clustering, which the citing paper incorporates into their research to improve the quality of clustering results.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chang et al., 2017)\", \"Explanation\": \"The cited work by Chang et al. presents a method of using relational supervision in deep clustering, which the citing paper leverages to enhance the performance of their own clustering approach.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Basu et al., 2004;Wagstaff et al., 2001;Basu et al., 2008;Zhang et al., 2020;Manduchi et al., 2021)\", \"Explanation\": \"The cited works provide a basis for the incorporation of pairwise mustlink or cannot-link constraints in the study of relation inference using API-based LLMs.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Reimers and Gurevych, 2019;Gao et al., 2021;Ni et al., 2022a,b)\", \"Explanation\": \"The cited works are pre-trained text embedding models that are widely applied in text similarity, classification, clustering, and information retrieval, and are used as a data source in the study of relation inference using API-based LLMs.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Wang et al., 2022)\", \"Explanation\": \"The cited work on the E5 pre-trained model is an extension of the study of pre-trained text embedding models in relation inference using API-based LLMs.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Su et al., 2022)\", \"Explanation\": \"The cited work on the Instructor pre-trained model is an extension of the study of pre-trained text embedding models in relation inference using API-based LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Rousseeuw, 1987)\", \"Explanation\": \"The cited work by Rousseeuw (1987) introduces the silhouette coefficient as a metric for evaluating clustering quality, which the citing paper adopts in their research to measure the granularity of clusters without the need for ground truth data.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Thorndike, 1953)\", \"Explanation\": \"The cited work by Thorndike (1953) presents the elbow method as a heuristic approach for determining the best granularity in a hierarchy, which the citing paper utilizes in their research to identify the optimal level of granularity in a given dataset.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Pelleg et al., 2000)\", \"Explanation\": \"The cited work by Pelleg et al. (2000) introduces the X-means algorithm as a variation of K-means that starts with the lowest number of clusters and then splits them into smaller clusters based on BIC evaluation, which the citing paper employs in their research to improve the granularity of clusters.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Goutte et al., 2001)\", \"Explanation\": \"The cited work by Goutte et al. (2001) presents the Bayesian Information Criterion (BIC) as a metric for evaluating granularity in clusters, which the citing paper uses in their research to calculate BIC for each level of granularity in a dataset.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2021b)\", \"Explanation\": \"The cited work by Zhang et al. (2021b) presents the cluster-size method as a way of filtering small clusters using a confidence threshold, which the citing paper adopts in their research to improve the granularity of clusters by removing clusters with low confidence.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Su et al., 2022)\", \"Explanation\": \"The cited work provides the hyperparameters and training details for fine-tuning the large version of Instructor and E5 models, which the citing paper adopts in their experiments.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(intfloat/e5-large)\", \"Explanation\": \"The cited work is the pre-trained E5 model used in the experiments of the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(hkunlp/instructor-large)\", \"Explanation\": \"The cited work is the pre-trained Instructor model used in the experiments of the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Su et al., 2022)\", \"Explanation\": \"The citing paper builds upon the work of Su et al. (2022) by fine-tuning the large version of Instructor and E5 models and using the same or similar prompt as the original paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Gao et al., 2019)\", \"Explanation\": \"The cited work provides the dataset for FewRel, which the citing paper uses in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Deng et al., 2020)\", \"Explanation\": \"The cited work provides the dataset for FewEvent, which the citing paper uses in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ding et al., 2021)\", \"Explanation\": \"The cited work provides the dataset for FewNerd, which the citing paper uses in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Geigle et al., 2021)\", \"Explanation\": \"The cited work provides the dataset for StackEx, Reddit, and ArxivS2S, which the citing paper uses in their research by combining the splits and removing topics with few instances.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Vaze et al., 2022)\", \"Explanation\": \"The cited work by Vaze et al. provides a method for generalized category discovery that the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lin et al., 2020)\", \"Explanation\": \"The cited work by Lin et al. contributes a method for generalized category discovery that the citing paper utilizes in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2021b)\", \"Explanation\": \"The cited work by Zhang et al. provides a method for generalized category discovery that the citing paper references in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2022)\", \"Explanation\": \"The cited work by Zhang et al. contributes a method for generalized category discovery that the citing paper references in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Mou et al., 2022)\", \"Explanation\": \"The cited work by Mou et al. provides a method for generalized category discovery that the citing paper references in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(An et al., 2022)\", \"Explanation\": \"The cited work by An et al. contributes a method for generalized category discovery that the citing paper references in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Rizve et al., 2022b)\", \"Explanation\": \"The cited work by Rizve et al. provides a method for generalized category discovery via optimal transport that the citing paper references in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Yang et al., 2022)\", \"Explanation\": \"The cited work by Yang et al. contributes a method for generalized category discovery via optimal transport that the citing paper references in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Rizve et al., 2022a)\", \"Explanation\": \"The cited work by Rizve et al. provides a method for generalized category discovery via similarity learning that the citing paper references in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Cao et al., 2022)\", \"Explanation\": \"The cited work by Cao et al. contributes a method for generalized category discovery via similarity learning that the citing paper references in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Sun and Li, 2022)\", \"Explanation\": \"The cited work by Sun and Li provides a method for generalized category discovery via prototype-based learning that the citing paper references in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing the bank purpose for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2023)\", \"Explanation\": \"The cited work provides a method for representing the relation between two entities for retrieval, which the citing paper utilizes in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(An et al., 2022)\", \"Explanation\": \"The cited work provides a method for representing the entity type for retrieval, which the citing paper employs in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lin et al., 2020)\", \"Explanation\": \"The cited work provides a method for representing the event type for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2022)\", \"Explanation\": \"The cited work provides a method for representing the question for retrieval, which the citing paper utilizes in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing the science statement for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(He et al., 2023)\", \"Explanation\": \"The cited work provides a method for representing an emotion sentence for retrieval, which the citing paper utilizes in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhu et al., 2023)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper employs in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a scene for retrieval, which the citing paper utilizes in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper employs in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper utilizes in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper employs in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper utilizes in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper employs in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper utilizes in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper employs in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper utilizes in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper employs in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper utilizes in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper employs in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper utilizes in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper employs in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper utilizes in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., , 2021b)\", \"Explanation\": \"The cited work provides a method for representing a sentence for retrieval, which the citing paper employs in their study.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Cheng et al., 2023)\", \"Explanation\": \"The cited work by Cheng et al. provides a method for fine-tuning models with feedback from LLMs, which the citing paper adopts in their research on clustering tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Bai et al., 2022)\", \"Explanation\": \"The cited work by Bai et al. also contributes to the field of fine-tuning models with feedback from LLMs, which the citing paper leverages in their research on clustering tasks.\"}]"}}},{"rowIdx":88,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_1","fig_1","fig_1"],"heading":"Introduction","publication_ref":["b32","b43","b10"],"table_ref":[],"text":"Large Language Models (LLMs) have demonstrated remarkable performance in solving various natural language processing tasks through prompt selection methods roughly correspond to a sub-term of the equation of Mutual Information (MI) (Sorensen et al., 2022). We utilize this discovery to propose several variants of MI that use different combinations of the components of existing methods, and the best combinational variant MI AGL increases the scaled F1 (F1 divided by that of the oracle prompt, showing the effectiveness of the prompt selection method) from 87.79% to 94.98% (MI AGL of Figure 1a).\nFurthermore, we find the need for a better approximation of the LLM's output probability distribution, considering that all probability-based prompt selection methods rely on the probabilistic estimates from the model that might be biased. Therefore, by drawing a connection between the existing model output probability calibration methods (Zhao et al., 2021;Holtzman et al., 2021), we propose an enhanced calibration method, Calibration By Marginalization (CBM). CBM significantly improves the prompt selection performance of several methods when applied to calibrate the output probability of LLMs, increasing the bestscaled F1 to 96.85% (MI (PA) A of Figure 1a), achieving 99.44% of the oracle prompt F1 under the uncalibrated scenario. CBM also proves to show the most robust answer selection enhancement across multiple datasets compared to the existing calibration methods (Figure 1b)."},{"figure_ref":[],"heading":"Probability-based Prompt Selection","publication_ref":[],"table_ref":[],"text":"In this section, we perform a unified evaluation of existing probability-based prompt selection methods. First, we describe the task of probabilitybased prompt selection in Section 2.1. Next, we briefly introduce each of the existing methods in Section 2.2. Then, we describe our experimental setup for unified evaluation in Section 2.3 and present the evaluation results in Section 2.4."},{"figure_ref":["fig_7"],"heading":"Task Description","publication_ref":["b37","b15"],"table_ref":["tab_2","tab_10"],"text":"Probability-based prompt selection is the task of selecting one or more prompts from a list of prompts T , which are expected to help the language model θ make the most accurate prediction for the evaluation dataset X where the evaluation instances are drawn from the data distribution, x ∼ P X , utilizing only the output probability distributions of the model on X,3 without knowing 9. The methods without super/subscripts are the existing methods (Table 1), while those with super/subscripts are our proposed methods (Table 4 & Equation 1). (b) Ratio of the prompts (out of 100) whose F1 on each dataset improves by applying probability calibration for answer selection, averaged across 10 models. Our proposed calibration method, CBM (Equation 1), is considerably more effective than CC and PMI DC (Table 5) in enhancing the answer selection performance of the prompts.\nthe ground truth labels and using neither additional gradient-based updates nor other trained components. The performance of a probability-based prompt selection method is evaluated by how high the score of the evaluation metric obtained with the selected prompt(s) is. When one prompt is selected for the whole dataset, the performance is upper bounded by the performance obtained with the prompt with which the model achieves the best metric score; we call such a prompt the optimal oracle prompt. 4 When one prompt is selected for each x ∼ P X , different t ∈ T can be chosen for each x; we call such a prompt selection approach as instance-wise prompt selection.\nNote that the definition of prompt can vary ac- (Wu et al., 2023) MDL -H(Y |x, t) Zero-Label Prompt Selection (Liao et al., 2022) x 1 arg max y p(y|x, t) = arg max y s(x, y) cording to the setup for which prompt selection is performed. When prompt selection is applied to zero-shot learning, prompts are defined as various formats of text templates that are filled by evaluation instances x ∼ P X to facilitate. On the other hand, for few-shot (in-context) learning, prompts are often defined as the demonstrations sampled from a training/development set or texts of permutations of such demonstrations. In our work, in order to enable comparison between all the methods proposed either in zero-shot and few-shot setup, we perform prompt selection in a zero-shot setup with the former definition of prompt. 5 Concrete Example Examples of prompts t ∈ T include \"Which category does the following news article fall into? {text}\", \"The following news article, {text}, covers the topic of\", and \"{text} belongs in which category: Politics, Sports, Business, Science and Technology\". We say that x instantiates the prompt t when x is inserted into the placeholder {text} of the prompt template and let ι(x, t) denote the instantiated prompt. Each of the answer categories represents the concept of politics, sports, business, and science/technology, and uses \"Politics,\" \"Sports,\" \"Business,\" and \"Science and Technology\" as the verbalizer (the actual text evaluated to score the answer choices), respectively.\nLog-probability Mean ZLP s(x, y) = 1 |T | t log p(y|x, t) Probability Mean ZPM s(x, y) = 1 |T | t p(y|x, t) Majority Vote ZMV s(x, y) = t 1{arg max v∈Y p(y|x, t) = v} Perplexity (Gonen et al., 2023) PPL -1 |X| x 1 p(x,t)\nFor instance, given OPT 2.7B (Zhang et al., 5 We have performed additional experiments in a fewshot learning setup using the texts of permutations of varying numbers of in-context learning demonstrations as the prompts. However, we do not include these results in the paper due to space limitations; also, the overall trend of the results stays similar to that of the zero-shot learning setup. 2022a) as the language model, \"King Charles III's Coronation watched by more than 18 million viewers\" as x, and the three prompts shown as examples in the previous paragraph, a prompt selection method should choose the prompt that is most likely to help OPT 2.7B correctly predict the answer y among the possible answer choices Y which represent the concepts of politics, sports, business, and science/technology. To select such a prompt, the method must rely solely on the output probability of the model given the instantiated prompts as input, e.g., p(\"Politics\"|\"Which category . . . King . . . \")."},{"figure_ref":[],"heading":"Existing Approaches","publication_ref":["b32","b32","b15"],"table_ref":["tab_2"],"text":"Table 1 provides the summary of the existing approaches for probability-based prompt selection. In the equations, we use p(y|x, t) ∈ R |Y | to express the output probability distribution of the model over the answer choices, P θ (Y |X = x, T = t), when the instantiated prompt ι(x, t) is given as the input. The probability for each y ∈ Y is calculated as\np(y|x, t) = exp(log p(y|x, t)) y ′ ∈Y exp(log p(y ′ |x, t))\n,\nwhere log p(y|x, t) is the unnormalized logit that the model outputs. When y's verbalizer is tokenized into more than one token, we calculate log p(y|x, t) as the mean of log-probability over the tokens of the verbalizer for datasets with fixed answer choices, and as the sum of log-probability for datasets with dynamically changing sentencetype answer choices, except for the method proposed by Sorensen et al. (2022) which explicitly specifies that the calculation of p(y|x, t) uses only the logits of the first token (dubbed as One-Token Response (OTR) in their work). We use H(q(y)) to denote the entropy of an arbitrary probability distribution q(y) ∈ R |Y | , -y∈Y q(y) log q(y).\nWhen q(y) = p(y|x, t), we use H(Y |x, t) to represent its entropy H(Y |X = x, T = t).\nMutual Information (MI) Sorensen et al. (2022) propose to select one prompt for the evaluation dataset that maximizes the mutual information between the evaluation instances X and their corresponding model predictions Y given prompt Here, we describe only the ranking part of their approach that we employ for our zero-shot learning setup.\nZero-Label Prompt Selection (ZLP, ZPM, ZMV) Liao et al. (2022) propose to make a pseudo-label for each x by ensembling the outputs for all prompts to make a score s(x, y) for each x, and then choosing one prompt t for the evaluation dataset whose cases of arg max y∈Y p(y|x, t) = arg max y∈Y s(x, y) is the maximum. As shown in \nΠ |ι(x,t)| i=1 p(ι(x, t) i |ι(x, t) <i ) 1 |ι(x,t)| ,\nwhere ι(x, t) i represents the i-th token of the instantiated prompt ι(x, t). We include the geometric mean to the definition of p(x, t) because the averaged probability is often used to approximate the probability of a sequence."},{"figure_ref":[],"heading":"Experimental Setup","publication_ref":[],"table_ref":["tab_7"],"text":"Evaluation Datasets Our dataset selection, aimed at fair measurement of various probabilitybased prompt selection methods, is guided by several factors. We favor the datasets previously used in research, those encompassing diverse domains, and datasets where prompt selection is meaningful. We exclude the datasets where all prompts underperform a random baseline or where a naive baseline of selecting the mode label could excel due to high imbalance. By excluding the datasets with high imbalance, we aim to avoid the false positive cases where a failed algorithm that collapses to select one label regardless of the input is evaluated as a competitive method by chance. The selected datasets have diverse label types and distributions, and we categorize them based on their label distributions into balanced (label distribution is about 1:1), unbalanced (otherwise), and dynamic 7 categories. The 13 datasets selected through this process are shown in Table 2. 8 Prompts We create a diverse range of 100 prompts for each of the 13 evaluation datasets, which results in 1,300 prompts in total. For each dataset, a few of the 100 prompts are taken from PromptSource (Bach et al., 2022), and the rest are generated using GPT 3.5 (OpenAI, 2023) to speed up the prompt generation process and then manually reviewed and corrected 9 . The prompts are designed to encompass various formats, with the evaluation instance and sometimes the answer choices appearing at different positions within the prompt, to ensure that the prompt selection task is meaningful. Table 3 shows a few examples of the prompts. We use one-token words as the verbalizers for the answer choices in most prompts, except for the prompts for the datasets of the dynamic category."},{"figure_ref":[],"heading":"Models","publication_ref":["b3","b27"],"table_ref":[],"text":"We conduct the majority of our experiments with ten different models of varying sizes ranging from 1.3B to 66B 10 . However, to present the experimental results and analysis more clearly, 7 The answer choices are sentences and vary dynamically for each evaluation instance. In these datasets, the label index is not connected to some concept, unlike the datasets with static choices (e.g., 0 is negative and 1 is positive in sst2), so the ratio of labels is not meaningful. However, all the datasets of dynamic categories that we use have balanced label distribution. 2019) 9 The generation, review, and correction are done by the first two authors of the paper.\n10 GPT-Neo (Black et al., 2021) 1.3B, OPT (Zhang et al., 2022a) 1.3B, GPT2-XL (Radford et al., 2019), GPT-Neo 2.7B, OPT 2.7B, BLOOM 3B (Workshop et al., 2023), GPT-J 6B, OPT 6.7B, OPT 30B, and OPT 66B The inference is performed using one to four NVIDIA V100 32GB GPUs."},{"figure_ref":[],"heading":"Experimental Results","publication_ref":[],"table_ref":[],"text":"We find that there is no single probability-based prompt selection method that consistently outperforms one another across all 13 datasets and evaluation categories. While PPL and LE do not rank first in any dataset, every other method ranks first in a few datasets. Figure 2 illustrates the selected prompt performance averaged by category, along with the performance of the best (oracle) and worst prompts and the average performance of all prompts. In the balanced category, GE and MDL outperform others, with MI closely following. In the unbalanced category, MI stands out, while in the dynamic category, GE, MDL, and ZLP perform the best. LE and PPL generally underperform in all of the datasets; their task average does not even exceed the average performance of all prompts. 13 We conclude that no single existing approach is significantly better than others, especially when dividing the evaluation dimensions into balanced, unbalanced, and dynamic labels."},{"figure_ref":[],"heading":"Improving MI via Unified Analysis","publication_ref":[],"table_ref":[],"text":"In this section, we first derive a unified view of prompt selection methods in Section 3.1 and show that each method other than MI roughly corresponds to a sub-term of the equation of MI and revisit the previous experimental results for a unified analysis in Section 3.2. Then, from the unified view and analysis, we identify the differences between methods, particularly MI, GE, and MDL, and derive a few combinational variants by transferring design elements across methods which improves the prompt selection performance of MI."},{"figure_ref":["fig_3"],"heading":"Unified View: Identifying Connections Between Methods","publication_ref":[],"table_ref":[],"text":"Prompt Selection Score (PSS) Figure 3 offers a unified view of existing probability-based prompt selection methods, highlighting that each method except for MI approximately corresponds to a subterm in the equation of MI. We denote the highlighted parts as the Prompt Selection Score of each method (PSS method ); the score of which the prompt with the maximum value is chosen by the prompt selection method.\nMI vs. GE and LE MI selects a prompt that maximizes the first term of PSS MI , 13 Interpretations of these results are provided in Section 3.2. arg max t H 1\nMI GE LE MDL ZPM PPL -\" ! ∈$ \" %∈& # $, & # & # ' $, & log \" %∈& # $, & # & # ' $, & --\" %∈& # $, & # & \" !∈$ [# ' $, & log # ' $, & ] → \" # $"},{"figure_ref":[],"heading":"|X|","publication_ref":[],"table_ref":[],"text":"x p(y|x, t) , and minimizes the second term, 1"},{"figure_ref":[],"heading":"|X|","publication_ref":[],"table_ref":[],"text":"x H (Y |x, t). This means that MI favors prompts that provide balanced predictions without label bias (interpretation of the first term) and sharp answer prediction distribution across all instances in the dataset (interpretation of the second term). These terms roughly correspond to PSS GE and -PSS LE , respectively. The difference between PSS GE and the first term of PSS MI is that the former converts p(y|x, t) to one-hot before taking the entropy of the mean. In sum, the prompts selected by GE and MI align, while those chosen by LE and MI tend to be opposite. Note that one expected caveat of GE is that it will be less effective when the dataset itself has a label imbalance. MI vs. MDL MDL is the only method among the presented probability-based prompt selection methods that selects a different prompt for each evaluation instance x, i.e., performs instance-wise prompt selection. Essentially, MDL is an instancewise version of the second term of PSS MI , choosing prompts whose output probability distribution p(y|x, t) has the lowest entropy, and thus aligns with MI. Since MDL favors the prompt that makes the model output a sharp probability distribution, one expected caveat of MDL is that it will not work well when the model fails to solve the given task and collapses to a single prediction regardless of the input with overly high confidence.\nMI vs. ZPM Zero-label prompt selection methods ensemble the results of all prompts to calculate s(x, y), create pseudo labels by converting s(x, y) to one-hot, and then choose the prompt with predictions most similar to the pseudo labels. Applying this view to PSS ZPM with an assumption of 2023) even restrict their prompt format for the input x to appear at the beginning so that p(x, t) is calculated only as the form of p(t|x)p(x), i.e., the probability of prompt is always conditioned on x, the probabilistic assumption of MI is incompatible with the motivation of PPL. 15\np(t|x) = 1 |T | results"},{"figure_ref":[],"heading":"Unified Analysis: Revisiting Experimental Results","publication_ref":["b15"],"table_ref":[],"text":"Revisiting the unified evaluation in Section 2.4, the results align with our analysis from Section 3.1. GE performs well in balanced datasets but poorly in unbalanced ones due to its preference for prompts that create balanced predictions. GE also performs well in dynamic datasets since the label distribution is balanced by chance (Table 2). MDL performs comparably to GE due to similar 14 One expected caveat of the methods of zero-label prompt selection is that it might not work well when a large portion of the prompts fail to solve the given task. Therefore, Liao et al. (2022) propose a way to filter out low-quality prompts in advance, but the filtering algorithm does not benefit their proposed methods in our experimental setup.\n15 Note that our experimental setup also differs with the setup of Gonen et al. ( 2023); we generated the prompts in an unrestricted manner that x can appear anywhere in the prompt. entropy calculations. LE's performance, however, is less satisfactory, given its optimization contradicts MDL. The underperformance of PPL compared to that by Gonen et al. ( 2023) might be due to our use of diverse prompt formats 16 .\nNote that in dynamic datasets, MI's best, worst, and average prompt performances differ due to its distinct calculation of p(y|x, t) that uses only the first token logits; for other methods, p(y|x, t) is calculated using all tokens (Section 2.2). 17 This leads to a question: Is the difference in the calculation of p(y|x, t) the reason that MI performs well in balanced and unbalanced cases but poorly in dynamic cases? In addition, despite GE and MDL maximizing MI's sub-term, they outperform MI in balanced datasets. This observation leads to another question: Is their higher performance due to their one-hot p(y|x, t) and instance-wise prompt selection?\nIn the following subsection, we show that the answers to both questions are yes, demonstrating that using all tokens to calculate p(y|x, t), onehot p(y|x, t), and instance-wise prompt selection improves the prompt selection performance of MI."},{"figure_ref":["fig_4"],"heading":"Experimental Results: Transferring","publication_ref":[],"table_ref":["tab_10"],"text":"Design Choices from Unified Analysis p(y|x, t) calculation using all tokens helps MI.\nTo investigate the difference between using only the first token probability and the mean/sum of 16 We allow the input x to appear anywhere in the prompt, unlike their restricted setup where x always comes at the beginning. 17 In balanced and unbalanced cases, the number of tokens of most verbalizers is 1, so the best, worst, and average prompt performances of the prompts whose performance is calculated using only the first token are identical to the other methods; on the other hand, the verbalizer is a sentence for dynamic datasets and makes the difference. all tokens to calculate PSS MI , we develop a variant of MI called MI A (A of All). Unlike MI and like other methods, MI A calculates p(y|x, t) by taking the mean of all token logits for balanced and unbalanced datasets, and the sum for dynamic datasets. Since the balanced and unbalanced datasets in our experimental setup (Section 2.4) mostly use one-token verbalizers which result in the same result of MI and MI A , we utilize new sets of verbalizers of 1-2 tokens (1 ≤ |v| ≤ 2) or 2 tokens (|v| = 2) for all the prompts of our evaluation datasets and compare the two methods. Our results in Figure 4 show that using all tokens is more effective in all configurations except for the 1-2 token-balanced tasks.\nA G L Prompt Selection Score Existing Methods GE ✓ ✓ -H 1 |X| x one hot(p(y|x, t)) MDL ✓ -✓ -H(Y |x, t) MI ✗ ✗ ✗ GE M + MDL M Explored Variants GE M ✓ ✗ -H 1 |X| x p(y|x, t) MDL M ✓ -✗ -1 |X| x H (Y |x, t) MI A ✓ ✗ ✗ GE M + MDL M MI AG ✓ ✓ ✗ GE + MDL M MI AL ✓ ✗ ✓ GE M + MDL MI AGL ✓ ✓ ✓ GE + MDL\nOne-hot p(y|x, t) and instance-wise prompt selection benefits MI. We create combinational variants of GE, MDL, and MI (outlined in Table 4) to study whether their differences contribute to MI's lower performance in balanced datasets. For instance, PSS "},{"figure_ref":[],"heading":"Improving Prompt Selection Through Enhanced Probability Calibration","publication_ref":["b43","b10"],"table_ref":[],"text":"While the previous section enhances prompt selection performance using combinatorial variants, in this section, we explore an orthogonal approach to further improve prompt selection: model output probability calibration. Since all the prompt selection methods except for PPL depend on the model output probability p(y|x, t) to calculate Prompt Selection Score (PSS), the stability and reliability of p(y|x, t) affect their prompt selection performance. However, previous works have pointed out that p(y|x, t) is unstable without calibration. 18 To address the issue, Zhao et al. (2021) suggest Contextual Calibration (CC), which reduces bias towards each answer choice by employing content-free inputs (\"N/A\", \"[MASK]\", \"\"), while Holtzman et al. (2021) present Domain Conditional Pointwise Mutual Information (PMI DC ) by reweighting each answer choice based on its task-specific prior likelihood. We summarize the two methods for answer selection in Table 5; arg max y q(y|x, t) is selected as the answer, where q(y|x, t) is the calibrated score.\nOne might assume that these existing calibration methods would effectively calibrate p(y|x, t) for PSS. However, through the experiments described in Section 4.1, we reveal in Section 4.2 the results that these methods have limitations for prompt selection and even answer selection across numerous datasets. In response, we propose an enhanced calibration method, Calibration By Marginalization (CBM), in Section 4.3. Section 4.4 shows that CBM notably improves prompt selection for most methods, particularly MI and MDL M , enabling them to achieve the highest prompt selection performance compared to all other methods. Furthermore, CBM's answer selection enhancement is the most robust across various datasets when compared to existing calibration methods."},{"figure_ref":[],"heading":"Experimental Setup for Probability Calibration","publication_ref":["b10"],"table_ref":[],"text":"We compare the prompt selection performance with four different scenarios of calibration: without applying any calibration; (A) applying calibration only for Answer selection, computing q(y|x, t) where arg max y q(y|x, t) is selected as the answer; (P) applying calibration only for Prompt selection; and (PA) applying calibration for both Prompt selection and Answer selection. Normalization of q(y|x, t) is not required for answer selection, as it does not affect the arg max of the scores. However, to obtain PSS, it is essential to normalize q(y|x, t) so that the sum equals one, thereby preserving the original probabilistic motivation of different methods. Consequently, 18 Zhao et al. (2021) find that the probability in few-shot learning tends to favor certain answer choices appearing at the end of the prompt or common in pretraining data. Holtzman et al. (2021) note that ranking based on string probability can be probabilistic due to surface form competition."},{"figure_ref":[],"heading":"Existing Method","publication_ref":["b43","b10"],"table_ref":[],"text":"Equation for Answer Selection Contextual Calibration (CC) (Zhao et al., 2021) (Holtzman et al., 2021) q(y|x, t) = log p(y|x, t) p(y|x domain , t) Table 5: Existing calibration methods proposed for answer selection. arg max y q(y|x, t) is selected as the answer for the prompt t instantiated by input instance x. Note that the actual calculation of CC in the official code uses p cf , mean-normalized pcf ; thus, we also use it in our experiments.\nC = {\"N/A\", \"[MASK]\", \"\"} pcf = 1 |C| c∈C p(y|c, t) W = diag(p cf ) -1 , b = 0 q(y|x, t) = Wp(y|x, t) + b Domain Conditional PMI (PMI DC )\nwe apply the softmax function to convert q(y|x, t) into a proper probability distribution q(y|x, t).19 "},{"figure_ref":["fig_1"],"heading":"Experimental Results: Underperformance of Existing Calibration Methods","publication_ref":[],"table_ref":[],"text":"We check the prompt selection performance of each method across the four calibration scenarios. Surprisingly, for both CC and PMI DC , we find that all three calibration scenarios show degraded performance compared to the scenario of no calibration. Not only does the prompt selection performance degrade, but the best, worst, and average prompt performance also drops in the case of A (only answer selection). This is unexpected, as CC and PMI DC have been reported to improve performance in slightly different setups (our results are in a zero-shot setting, while the main setup of Zhao et al. ( 2021) is few-shot, and the choice of x domain differs for PMI DC ).\nTo further investigate the subpar performance in case A, we analyze the proportion of prompts (out of 100) that exhibit improved performance after applying calibration for answer selection across ten different models and 13 datasets. Figure 1b displays the average ratio for all models. The figure indicates that the existing calibration methods do not result in better answer selection for the majority of our evaluation datasets. For in-stance, more than half of the prompts displayed decreased performance after applying CC in 7 out of 13 datasets. A similar pattern holds when applying PMI DC ."},{"figure_ref":["fig_1"],"heading":"Enhanced Calibration Method:","publication_ref":[],"table_ref":[],"text":"Calibration By Marginalization (CBM)\nTable 5 shows that the equation for CC can be alternatively expressed as follows:\nq(y|x, t) =diag(p cf ) -1 p(y|x, t) + 0 = p(y|x, t) pcf = p(y|x, t) 1 |C| c∈C p(y|c, t)\n, which turns CC into a special case of PMI DC 20 , where p(y|x domain , t) = 1 |C| c∈C p(y|c, t). Additionally, upon revisiting the motivation of PMI DC and considering the equation of pointwise mutual information PMI(x, y) = log p(y|x) p(y) , it becomes evident that p(y|x domain , t) approximates p(y|t). Therefore, the distinction between CC and PMI DC lies solely in how they approximate p(y|t). However, since the approximation for CC relies on three inputs and PMI DC on just one, both methods fall short of providing a stable approximation. This limitation naturally leads to the following question: Could there be a way to approximate p(y|x, t) in a more stable manner?\nEncouragingly, the answer to the question is yes. A better approximation of p(y|x, t) can be calculated using the law of marginal probability: p(y|t) = x∈X p(y, x|t) = x∈X p(y|x, t)p(x|t). With this more stable approximation of p(y|t) and the probabilistic assumption of MI that p(x|t) = 1 |X| , we introduce a new calibration method called Calibration By Marginalization (CBM) that employs the following equation for answer selection:\nq(y|x, t) = p(y|x, t) p(y|t) = p(y|x, t) 1 |X| x ′ ∈X p(y|x ′ , t)\n.\n(1)\nSince the calculation of p(y|x, t) for all t ∈ T and x ∈ X is already done to perform prompt selection, CBM does not introduce any additional computational cost for calibration, unlike CC or PMI DC that require inference on additional inputs such as \"N/A\", \"[MASK]\", \"\", and x domain . 20 We can ignore the lack of log because it does not change the result of arg max. The methods displaying the most significant performance improvements in the PA scenario are MI AG , MI A , MI, and MDL M , particularly with the prompt selection performance of MI (PA) A and MDL (PA) M being the highest among different methods. On average, MI (PA) A increases the scaled F1 from 87.79% (0.5965/0.6795) to 99.44% (0.6757/0.6795) compared to the best existing method (GE) when the oracle prompt without calibration is used as the target of comparison. The scaled F1 of MI (PA) A calculated with respect to the oracle prompt with calibration is 96.85% (0.6757/0.6977).\nNext, we assess the effectiveness of CBM calibration for answer selection by examining the proportion of prompts (out of 100) that show improved performance after applying calibration for answer selection. Figure 1b indicates that CBM is considerably more effective than CC and PMI DC in enhancing the performance of the prompts. The performance of more than half of the prompts increases after applying CBM in all 13 datasets. Additionally, the performance of nearly 100% of prompts improves with CBM calibration in 7 datasets. While CC and PMI DC improved almost none of the F1 of the prompts in story and hella, the performance of approximately 70% of the prompts increased with CBM calibration, possibly due to the more accurate calculation of p(y|t) as discussed in Section 4.3."},{"figure_ref":["fig_5","fig_5"],"heading":"Discussion","publication_ref":[],"table_ref":[],"text":"In this section, we discuss various findings that are relevant to our main experiments.\nFigure 7a shows that the effectiveness of a probability-based prompt selection method remains consistent across models of different types and numbers of parameters, justifying our choice of using a single model (OPT 2.7B) as the representative for all experiments. Figure 7b shows that the trend of correlation between Prompt Selection Score and performance of the selected prompt is also quite consistent between different models.\nFigure 8 shows the mean and standard deviation of the result of prompt selection among five different subsets of 50 prompts randomly sampled from the full set of 100 prompts, using the mainly discussed methods. The result shows that the performance of instance-wise prompt selection methods (MI AGL , MI AL , MDL) is not stable, likely due to the noisy nature of selecting one prompt for each instance. However, the performance of MI (PA) A and MDL (PA) M still achieves the highest performance and also shows the lowest standard deviation, proving the effectiveness of CBM.\nThrough additional analysis, we find that (1) while strong performance in prompt selection does not consistently correlate with Prompt Selection Score, a broadly positive correlation is observed when averaged across most methods; (2) CBM im-proves the performance of MDL M by mitigating overconfidence; (3) MI, GE, and CBM methods face limitations when applied to dynamic datasets with extreme label imbalance; (4) top-performing prompt selection methods from the zero-shot setting, like MI (PA) A and MDL (PA) M , retain their effectiveness in the few-shot setting, further validating their robustness across different conditions."},{"figure_ref":[],"heading":"Related Works","publication_ref":["b43","b10","b18","b17","b14","b30","b11","b21","b8","b37","b26","b6","b25"],"table_ref":[],"text":"Recent advances in large language models (LLMs) have created the paradigm of prompt-based learning, which gives the benefit that a single pretrained LLM can be used to solve a great number of tasks with task-specific prompts. However, the performance of LLMs can heavily fluctuate according to the choice of prompts (Zhao et al., 2021;Holtzman et al., 2021;Lu et al., 2022). To mitigate this issue, prompt engineering attempts to find the prompt that results in the most effective performance on the downstream task (Liu et al., 2023).\nAutomatic prompt engineering methods can be largely divided into two groups: the methods that use discrete prompts where the prompts are human-understandable actual text strings, and the methods that optimize continuous prompts where the prompts lie in the embedding space of the model (Li and Liang, 2021;Shin et al., 2020). Probability-based prompt selection methods that we study in this work 2.2 fall into the former group; most of the methods of the latter group require gradient-based training, while probabilitybased prompt selection does not perform any gradient-based update.\nPrompt engineering methods using discrete prompts include prompt paraphrasing, prompt generation, and prompt selection. Among these, prompt paraphrasing or generation approaches can be used together with probability-based selection methods; prompt selection can be performed on the prompts generated through prompt paraphrasing or generation (Jiang et al., 2020;Mishra et al., 2022;Gao et al., 2021;Wang et al., 2023;Prasad et al., 2023;Kim et al., 2022;Deng et al., 2022). Among prompt selection methods other than the probability-based approaches, a large portion of the methods are not easily utilizable since they require training an additional model and/or the use of an additional component. (Zhang et al., 2022b) On the other hand, probability-based prompt selection offers the advantage of prompt selection requiring only the output probabilities of the LLM. While the prerequisite is a set of candidate prompts to select from, this data is relatively small in size and can be easily obtained from the research community (Bach et al., 2022) or via machine generation (OpenAI, 2023). One limitation of these methods, though, is that one cannot use them for closed-source LLMs that are only available via proprietary LLM APIs that do not provide output probability distributions. Also, when the number of candidate prompts |T | and the size of the dataset used to select the prompt |X| is large, the calculation for prompt selection becomes computationally heavy; using a smaller set X ′ ∈ X to choose the prompt for X can be helpful in such a case."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"In this paper, we address the need for a comprehensive evaluation to compare the existing probability-based prompt selection methods, which have been proposed and evaluated under varying conditions and datasets. To achieve this, we introduce a unified evaluation setup to compare these methods, conduct a thorough evaluation, and develop a unified framework of the existing probability-based prompt selection methods. Our analysis within this unified framework has provided insights into the relationship among existing methods, enabling the development of several combinational variants that improve performance. Furthermore, our research on probability calibration has revealed the limitations of existing calibration methods and led to the proposal of an enhanced calibration method, Calibration By Marginalization (CBM). CBM not only significantly improves prompt selection performance but also demonstrates robust answer selection enhancement across multiple datasets. We hope that our unified setup provides a foundation for fair evaluation between various prompt selection methods and that our findings yield deeper insights into probability-based prompt selection."},{"figure_ref":[],"heading":"Acknowledgements","publication_ref":[],"table_ref":[],"text":"The authors would like to extend their sincere gratitude to the anonymous reviewers and Action Editor for their highly detailed and insightful comments and feedback. The authors would also like to thank Sang-Woo Lee for valuable feedback and discussions on the project. This work was partly supported by KT grant (2021, A study on a conversational language model that uses long external text as a prompt, 80%) and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub, 20%)."},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"* This project was initiated while the first author was a Master's student at KAIST (Nov 2022 -Feb 2023"}],"string":"[\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_1\",\n \"fig_1\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b32\",\n \"b43\",\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"Large Language Models (LLMs) have demonstrated remarkable performance in solving various natural language processing tasks through prompt selection methods roughly correspond to a sub-term of the equation of Mutual Information (MI) (Sorensen et al., 2022). We utilize this discovery to propose several variants of MI that use different combinations of the components of existing methods, and the best combinational variant MI AGL increases the scaled F1 (F1 divided by that of the oracle prompt, showing the effectiveness of the prompt selection method) from 87.79% to 94.98% (MI AGL of Figure 1a).\\nFurthermore, we find the need for a better approximation of the LLM's output probability distribution, considering that all probability-based prompt selection methods rely on the probabilistic estimates from the model that might be biased. Therefore, by drawing a connection between the existing model output probability calibration methods (Zhao et al., 2021;Holtzman et al., 2021), we propose an enhanced calibration method, Calibration By Marginalization (CBM). CBM significantly improves the prompt selection performance of several methods when applied to calibrate the output probability of LLMs, increasing the bestscaled F1 to 96.85% (MI (PA) A of Figure 1a), achieving 99.44% of the oracle prompt F1 under the uncalibrated scenario. CBM also proves to show the most robust answer selection enhancement across multiple datasets compared to the existing calibration methods (Figure 1b).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Probability-based Prompt Selection\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we perform a unified evaluation of existing probability-based prompt selection methods. First, we describe the task of probabilitybased prompt selection in Section 2.1. Next, we briefly introduce each of the existing methods in Section 2.2. Then, we describe our experimental setup for unified evaluation in Section 2.3 and present the evaluation results in Section 2.4.\"\n },\n {\n \"figure_ref\": [\n \"fig_7\"\n ],\n \"heading\": \"Task Description\",\n \"publication_ref\": [\n \"b37\",\n \"b15\"\n ],\n \"table_ref\": [\n \"tab_2\",\n \"tab_10\"\n ],\n \"text\": \"Probability-based prompt selection is the task of selecting one or more prompts from a list of prompts T , which are expected to help the language model θ make the most accurate prediction for the evaluation dataset X where the evaluation instances are drawn from the data distribution, x ∼ P X , utilizing only the output probability distributions of the model on X,3 without knowing 9. The methods without super/subscripts are the existing methods (Table 1), while those with super/subscripts are our proposed methods (Table 4 & Equation 1). (b) Ratio of the prompts (out of 100) whose F1 on each dataset improves by applying probability calibration for answer selection, averaged across 10 models. Our proposed calibration method, CBM (Equation 1), is considerably more effective than CC and PMI DC (Table 5) in enhancing the answer selection performance of the prompts.\\nthe ground truth labels and using neither additional gradient-based updates nor other trained components. The performance of a probability-based prompt selection method is evaluated by how high the score of the evaluation metric obtained with the selected prompt(s) is. When one prompt is selected for the whole dataset, the performance is upper bounded by the performance obtained with the prompt with which the model achieves the best metric score; we call such a prompt the optimal oracle prompt. 4 When one prompt is selected for each x ∼ P X , different t ∈ T can be chosen for each x; we call such a prompt selection approach as instance-wise prompt selection.\\nNote that the definition of prompt can vary ac- (Wu et al., 2023) MDL -H(Y |x, t) Zero-Label Prompt Selection (Liao et al., 2022) x 1 arg max y p(y|x, t) = arg max y s(x, y) cording to the setup for which prompt selection is performed. When prompt selection is applied to zero-shot learning, prompts are defined as various formats of text templates that are filled by evaluation instances x ∼ P X to facilitate. On the other hand, for few-shot (in-context) learning, prompts are often defined as the demonstrations sampled from a training/development set or texts of permutations of such demonstrations. In our work, in order to enable comparison between all the methods proposed either in zero-shot and few-shot setup, we perform prompt selection in a zero-shot setup with the former definition of prompt. 5 Concrete Example Examples of prompts t ∈ T include \\\"Which category does the following news article fall into? {text}\\\", \\\"The following news article, {text}, covers the topic of\\\", and \\\"{text} belongs in which category: Politics, Sports, Business, Science and Technology\\\". We say that x instantiates the prompt t when x is inserted into the placeholder {text} of the prompt template and let ι(x, t) denote the instantiated prompt. Each of the answer categories represents the concept of politics, sports, business, and science/technology, and uses \\\"Politics,\\\" \\\"Sports,\\\" \\\"Business,\\\" and \\\"Science and Technology\\\" as the verbalizer (the actual text evaluated to score the answer choices), respectively.\\nLog-probability Mean ZLP s(x, y) = 1 |T | t log p(y|x, t) Probability Mean ZPM s(x, y) = 1 |T | t p(y|x, t) Majority Vote ZMV s(x, y) = t 1{arg max v∈Y p(y|x, t) = v} Perplexity (Gonen et al., 2023) PPL -1 |X| x 1 p(x,t)\\nFor instance, given OPT 2.7B (Zhang et al., 5 We have performed additional experiments in a fewshot learning setup using the texts of permutations of varying numbers of in-context learning demonstrations as the prompts. However, we do not include these results in the paper due to space limitations; also, the overall trend of the results stays similar to that of the zero-shot learning setup. 2022a) as the language model, \\\"King Charles III's Coronation watched by more than 18 million viewers\\\" as x, and the three prompts shown as examples in the previous paragraph, a prompt selection method should choose the prompt that is most likely to help OPT 2.7B correctly predict the answer y among the possible answer choices Y which represent the concepts of politics, sports, business, and science/technology. To select such a prompt, the method must rely solely on the output probability of the model given the instantiated prompts as input, e.g., p(\\\"Politics\\\"|\\\"Which category . . . King . . . \\\").\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Existing Approaches\",\n \"publication_ref\": [\n \"b32\",\n \"b32\",\n \"b15\"\n ],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"Table 1 provides the summary of the existing approaches for probability-based prompt selection. In the equations, we use p(y|x, t) ∈ R |Y | to express the output probability distribution of the model over the answer choices, P θ (Y |X = x, T = t), when the instantiated prompt ι(x, t) is given as the input. The probability for each y ∈ Y is calculated as\\np(y|x, t) = exp(log p(y|x, t)) y ′ ∈Y exp(log p(y ′ |x, t))\\n,\\nwhere log p(y|x, t) is the unnormalized logit that the model outputs. When y's verbalizer is tokenized into more than one token, we calculate log p(y|x, t) as the mean of log-probability over the tokens of the verbalizer for datasets with fixed answer choices, and as the sum of log-probability for datasets with dynamically changing sentencetype answer choices, except for the method proposed by Sorensen et al. (2022) which explicitly specifies that the calculation of p(y|x, t) uses only the logits of the first token (dubbed as One-Token Response (OTR) in their work). We use H(q(y)) to denote the entropy of an arbitrary probability distribution q(y) ∈ R |Y | , -y∈Y q(y) log q(y).\\nWhen q(y) = p(y|x, t), we use H(Y |x, t) to represent its entropy H(Y |X = x, T = t).\\nMutual Information (MI) Sorensen et al. (2022) propose to select one prompt for the evaluation dataset that maximizes the mutual information between the evaluation instances X and their corresponding model predictions Y given prompt Here, we describe only the ranking part of their approach that we employ for our zero-shot learning setup.\\nZero-Label Prompt Selection (ZLP, ZPM, ZMV) Liao et al. (2022) propose to make a pseudo-label for each x by ensembling the outputs for all prompts to make a score s(x, y) for each x, and then choosing one prompt t for the evaluation dataset whose cases of arg max y∈Y p(y|x, t) = arg max y∈Y s(x, y) is the maximum. As shown in \\nΠ |ι(x,t)| i=1 p(ι(x, t) i |ι(x, t) <i ) 1 |ι(x,t)| ,\\nwhere ι(x, t) i represents the i-th token of the instantiated prompt ι(x, t). We include the geometric mean to the definition of p(x, t) because the averaged probability is often used to approximate the probability of a sequence.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experimental Setup\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_7\"\n ],\n \"text\": \"Evaluation Datasets Our dataset selection, aimed at fair measurement of various probabilitybased prompt selection methods, is guided by several factors. We favor the datasets previously used in research, those encompassing diverse domains, and datasets where prompt selection is meaningful. We exclude the datasets where all prompts underperform a random baseline or where a naive baseline of selecting the mode label could excel due to high imbalance. By excluding the datasets with high imbalance, we aim to avoid the false positive cases where a failed algorithm that collapses to select one label regardless of the input is evaluated as a competitive method by chance. The selected datasets have diverse label types and distributions, and we categorize them based on their label distributions into balanced (label distribution is about 1:1), unbalanced (otherwise), and dynamic 7 categories. The 13 datasets selected through this process are shown in Table 2. 8 Prompts We create a diverse range of 100 prompts for each of the 13 evaluation datasets, which results in 1,300 prompts in total. For each dataset, a few of the 100 prompts are taken from PromptSource (Bach et al., 2022), and the rest are generated using GPT 3.5 (OpenAI, 2023) to speed up the prompt generation process and then manually reviewed and corrected 9 . The prompts are designed to encompass various formats, with the evaluation instance and sometimes the answer choices appearing at different positions within the prompt, to ensure that the prompt selection task is meaningful. Table 3 shows a few examples of the prompts. We use one-token words as the verbalizers for the answer choices in most prompts, except for the prompts for the datasets of the dynamic category.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Models\",\n \"publication_ref\": [\n \"b3\",\n \"b27\"\n ],\n \"table_ref\": [],\n \"text\": \"We conduct the majority of our experiments with ten different models of varying sizes ranging from 1.3B to 66B 10 . However, to present the experimental results and analysis more clearly, 7 The answer choices are sentences and vary dynamically for each evaluation instance. In these datasets, the label index is not connected to some concept, unlike the datasets with static choices (e.g., 0 is negative and 1 is positive in sst2), so the ratio of labels is not meaningful. However, all the datasets of dynamic categories that we use have balanced label distribution. 2019) 9 The generation, review, and correction are done by the first two authors of the paper.\\n10 GPT-Neo (Black et al., 2021) 1.3B, OPT (Zhang et al., 2022a) 1.3B, GPT2-XL (Radford et al., 2019), GPT-Neo 2.7B, OPT 2.7B, BLOOM 3B (Workshop et al., 2023), GPT-J 6B, OPT 6.7B, OPT 30B, and OPT 66B The inference is performed using one to four NVIDIA V100 32GB GPUs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experimental Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We find that there is no single probability-based prompt selection method that consistently outperforms one another across all 13 datasets and evaluation categories. While PPL and LE do not rank first in any dataset, every other method ranks first in a few datasets. Figure 2 illustrates the selected prompt performance averaged by category, along with the performance of the best (oracle) and worst prompts and the average performance of all prompts. In the balanced category, GE and MDL outperform others, with MI closely following. In the unbalanced category, MI stands out, while in the dynamic category, GE, MDL, and ZLP perform the best. LE and PPL generally underperform in all of the datasets; their task average does not even exceed the average performance of all prompts. 13 We conclude that no single existing approach is significantly better than others, especially when dividing the evaluation dimensions into balanced, unbalanced, and dynamic labels.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Improving MI via Unified Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we first derive a unified view of prompt selection methods in Section 3.1 and show that each method other than MI roughly corresponds to a sub-term of the equation of MI and revisit the previous experimental results for a unified analysis in Section 3.2. Then, from the unified view and analysis, we identify the differences between methods, particularly MI, GE, and MDL, and derive a few combinational variants by transferring design elements across methods which improves the prompt selection performance of MI.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Unified View: Identifying Connections Between Methods\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Prompt Selection Score (PSS) Figure 3 offers a unified view of existing probability-based prompt selection methods, highlighting that each method except for MI approximately corresponds to a subterm in the equation of MI. We denote the highlighted parts as the Prompt Selection Score of each method (PSS method ); the score of which the prompt with the maximum value is chosen by the prompt selection method.\\nMI vs. GE and LE MI selects a prompt that maximizes the first term of PSS MI , 13 Interpretations of these results are provided in Section 3.2. arg max t H 1\\nMI GE LE MDL ZPM PPL -\\\" ! ∈$ \\\" %∈& # $, & # & # ' $, & log \\\" %∈& # $, & # & # ' $, & --\\\" %∈& # $, & # & \\\" !∈$ [# ' $, & log # ' $, & ] → \\\" # $\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"|X|\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"x p(y|x, t) , and minimizes the second term, 1\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"|X|\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"x H (Y |x, t). This means that MI favors prompts that provide balanced predictions without label bias (interpretation of the first term) and sharp answer prediction distribution across all instances in the dataset (interpretation of the second term). These terms roughly correspond to PSS GE and -PSS LE , respectively. The difference between PSS GE and the first term of PSS MI is that the former converts p(y|x, t) to one-hot before taking the entropy of the mean. In sum, the prompts selected by GE and MI align, while those chosen by LE and MI tend to be opposite. Note that one expected caveat of GE is that it will be less effective when the dataset itself has a label imbalance. MI vs. MDL MDL is the only method among the presented probability-based prompt selection methods that selects a different prompt for each evaluation instance x, i.e., performs instance-wise prompt selection. Essentially, MDL is an instancewise version of the second term of PSS MI , choosing prompts whose output probability distribution p(y|x, t) has the lowest entropy, and thus aligns with MI. Since MDL favors the prompt that makes the model output a sharp probability distribution, one expected caveat of MDL is that it will not work well when the model fails to solve the given task and collapses to a single prediction regardless of the input with overly high confidence.\\nMI vs. ZPM Zero-label prompt selection methods ensemble the results of all prompts to calculate s(x, y), create pseudo labels by converting s(x, y) to one-hot, and then choose the prompt with predictions most similar to the pseudo labels. Applying this view to PSS ZPM with an assumption of 2023) even restrict their prompt format for the input x to appear at the beginning so that p(x, t) is calculated only as the form of p(t|x)p(x), i.e., the probability of prompt is always conditioned on x, the probabilistic assumption of MI is incompatible with the motivation of PPL. 15\\np(t|x) = 1 |T | results\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Unified Analysis: Revisiting Experimental Results\",\n \"publication_ref\": [\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"Revisiting the unified evaluation in Section 2.4, the results align with our analysis from Section 3.1. GE performs well in balanced datasets but poorly in unbalanced ones due to its preference for prompts that create balanced predictions. GE also performs well in dynamic datasets since the label distribution is balanced by chance (Table 2). MDL performs comparably to GE due to similar 14 One expected caveat of the methods of zero-label prompt selection is that it might not work well when a large portion of the prompts fail to solve the given task. Therefore, Liao et al. (2022) propose a way to filter out low-quality prompts in advance, but the filtering algorithm does not benefit their proposed methods in our experimental setup.\\n15 Note that our experimental setup also differs with the setup of Gonen et al. ( 2023); we generated the prompts in an unrestricted manner that x can appear anywhere in the prompt. entropy calculations. LE's performance, however, is less satisfactory, given its optimization contradicts MDL. The underperformance of PPL compared to that by Gonen et al. ( 2023) might be due to our use of diverse prompt formats 16 .\\nNote that in dynamic datasets, MI's best, worst, and average prompt performances differ due to its distinct calculation of p(y|x, t) that uses only the first token logits; for other methods, p(y|x, t) is calculated using all tokens (Section 2.2). 17 This leads to a question: Is the difference in the calculation of p(y|x, t) the reason that MI performs well in balanced and unbalanced cases but poorly in dynamic cases? In addition, despite GE and MDL maximizing MI's sub-term, they outperform MI in balanced datasets. This observation leads to another question: Is their higher performance due to their one-hot p(y|x, t) and instance-wise prompt selection?\\nIn the following subsection, we show that the answers to both questions are yes, demonstrating that using all tokens to calculate p(y|x, t), onehot p(y|x, t), and instance-wise prompt selection improves the prompt selection performance of MI.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\"\n ],\n \"heading\": \"Experimental Results: Transferring\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_10\"\n ],\n \"text\": \"Design Choices from Unified Analysis p(y|x, t) calculation using all tokens helps MI.\\nTo investigate the difference between using only the first token probability and the mean/sum of 16 We allow the input x to appear anywhere in the prompt, unlike their restricted setup where x always comes at the beginning. 17 In balanced and unbalanced cases, the number of tokens of most verbalizers is 1, so the best, worst, and average prompt performances of the prompts whose performance is calculated using only the first token are identical to the other methods; on the other hand, the verbalizer is a sentence for dynamic datasets and makes the difference. all tokens to calculate PSS MI , we develop a variant of MI called MI A (A of All). Unlike MI and like other methods, MI A calculates p(y|x, t) by taking the mean of all token logits for balanced and unbalanced datasets, and the sum for dynamic datasets. Since the balanced and unbalanced datasets in our experimental setup (Section 2.4) mostly use one-token verbalizers which result in the same result of MI and MI A , we utilize new sets of verbalizers of 1-2 tokens (1 ≤ |v| ≤ 2) or 2 tokens (|v| = 2) for all the prompts of our evaluation datasets and compare the two methods. Our results in Figure 4 show that using all tokens is more effective in all configurations except for the 1-2 token-balanced tasks.\\nA G L Prompt Selection Score Existing Methods GE ✓ ✓ -H 1 |X| x one hot(p(y|x, t)) MDL ✓ -✓ -H(Y |x, t) MI ✗ ✗ ✗ GE M + MDL M Explored Variants GE M ✓ ✗ -H 1 |X| x p(y|x, t) MDL M ✓ -✗ -1 |X| x H (Y |x, t) MI A ✓ ✗ ✗ GE M + MDL M MI AG ✓ ✓ ✗ GE + MDL M MI AL ✓ ✗ ✓ GE M + MDL MI AGL ✓ ✓ ✓ GE + MDL\\nOne-hot p(y|x, t) and instance-wise prompt selection benefits MI. We create combinational variants of GE, MDL, and MI (outlined in Table 4) to study whether their differences contribute to MI's lower performance in balanced datasets. For instance, PSS \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Improving Prompt Selection Through Enhanced Probability Calibration\",\n \"publication_ref\": [\n \"b43\",\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"While the previous section enhances prompt selection performance using combinatorial variants, in this section, we explore an orthogonal approach to further improve prompt selection: model output probability calibration. Since all the prompt selection methods except for PPL depend on the model output probability p(y|x, t) to calculate Prompt Selection Score (PSS), the stability and reliability of p(y|x, t) affect their prompt selection performance. However, previous works have pointed out that p(y|x, t) is unstable without calibration. 18 To address the issue, Zhao et al. (2021) suggest Contextual Calibration (CC), which reduces bias towards each answer choice by employing content-free inputs (\\\"N/A\\\", \\\"[MASK]\\\", \\\"\\\"), while Holtzman et al. (2021) present Domain Conditional Pointwise Mutual Information (PMI DC ) by reweighting each answer choice based on its task-specific prior likelihood. We summarize the two methods for answer selection in Table 5; arg max y q(y|x, t) is selected as the answer, where q(y|x, t) is the calibrated score.\\nOne might assume that these existing calibration methods would effectively calibrate p(y|x, t) for PSS. However, through the experiments described in Section 4.1, we reveal in Section 4.2 the results that these methods have limitations for prompt selection and even answer selection across numerous datasets. In response, we propose an enhanced calibration method, Calibration By Marginalization (CBM), in Section 4.3. Section 4.4 shows that CBM notably improves prompt selection for most methods, particularly MI and MDL M , enabling them to achieve the highest prompt selection performance compared to all other methods. Furthermore, CBM's answer selection enhancement is the most robust across various datasets when compared to existing calibration methods.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experimental Setup for Probability Calibration\",\n \"publication_ref\": [\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"We compare the prompt selection performance with four different scenarios of calibration: without applying any calibration; (A) applying calibration only for Answer selection, computing q(y|x, t) where arg max y q(y|x, t) is selected as the answer; (P) applying calibration only for Prompt selection; and (PA) applying calibration for both Prompt selection and Answer selection. Normalization of q(y|x, t) is not required for answer selection, as it does not affect the arg max of the scores. However, to obtain PSS, it is essential to normalize q(y|x, t) so that the sum equals one, thereby preserving the original probabilistic motivation of different methods. Consequently, 18 Zhao et al. (2021) find that the probability in few-shot learning tends to favor certain answer choices appearing at the end of the prompt or common in pretraining data. Holtzman et al. (2021) note that ranking based on string probability can be probabilistic due to surface form competition.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Existing Method\",\n \"publication_ref\": [\n \"b43\",\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"Equation for Answer Selection Contextual Calibration (CC) (Zhao et al., 2021) (Holtzman et al., 2021) q(y|x, t) = log p(y|x, t) p(y|x domain , t) Table 5: Existing calibration methods proposed for answer selection. arg max y q(y|x, t) is selected as the answer for the prompt t instantiated by input instance x. Note that the actual calculation of CC in the official code uses p cf , mean-normalized pcf ; thus, we also use it in our experiments.\\nC = {\\\"N/A\\\", \\\"[MASK]\\\", \\\"\\\"} pcf = 1 |C| c∈C p(y|c, t) W = diag(p cf ) -1 , b = 0 q(y|x, t) = Wp(y|x, t) + b Domain Conditional PMI (PMI DC )\\nwe apply the softmax function to convert q(y|x, t) into a proper probability distribution q(y|x, t).19 \"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Experimental Results: Underperformance of Existing Calibration Methods\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We check the prompt selection performance of each method across the four calibration scenarios. Surprisingly, for both CC and PMI DC , we find that all three calibration scenarios show degraded performance compared to the scenario of no calibration. Not only does the prompt selection performance degrade, but the best, worst, and average prompt performance also drops in the case of A (only answer selection). This is unexpected, as CC and PMI DC have been reported to improve performance in slightly different setups (our results are in a zero-shot setting, while the main setup of Zhao et al. ( 2021) is few-shot, and the choice of x domain differs for PMI DC ).\\nTo further investigate the subpar performance in case A, we analyze the proportion of prompts (out of 100) that exhibit improved performance after applying calibration for answer selection across ten different models and 13 datasets. Figure 1b displays the average ratio for all models. The figure indicates that the existing calibration methods do not result in better answer selection for the majority of our evaluation datasets. For in-stance, more than half of the prompts displayed decreased performance after applying CC in 7 out of 13 datasets. A similar pattern holds when applying PMI DC .\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Enhanced Calibration Method:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Calibration By Marginalization (CBM)\\nTable 5 shows that the equation for CC can be alternatively expressed as follows:\\nq(y|x, t) =diag(p cf ) -1 p(y|x, t) + 0 = p(y|x, t) pcf = p(y|x, t) 1 |C| c∈C p(y|c, t)\\n, which turns CC into a special case of PMI DC 20 , where p(y|x domain , t) = 1 |C| c∈C p(y|c, t). Additionally, upon revisiting the motivation of PMI DC and considering the equation of pointwise mutual information PMI(x, y) = log p(y|x) p(y) , it becomes evident that p(y|x domain , t) approximates p(y|t). Therefore, the distinction between CC and PMI DC lies solely in how they approximate p(y|t). However, since the approximation for CC relies on three inputs and PMI DC on just one, both methods fall short of providing a stable approximation. This limitation naturally leads to the following question: Could there be a way to approximate p(y|x, t) in a more stable manner?\\nEncouragingly, the answer to the question is yes. A better approximation of p(y|x, t) can be calculated using the law of marginal probability: p(y|t) = x∈X p(y, x|t) = x∈X p(y|x, t)p(x|t). With this more stable approximation of p(y|t) and the probabilistic assumption of MI that p(x|t) = 1 |X| , we introduce a new calibration method called Calibration By Marginalization (CBM) that employs the following equation for answer selection:\\nq(y|x, t) = p(y|x, t) p(y|t) = p(y|x, t) 1 |X| x ′ ∈X p(y|x ′ , t)\\n.\\n(1)\\nSince the calculation of p(y|x, t) for all t ∈ T and x ∈ X is already done to perform prompt selection, CBM does not introduce any additional computational cost for calibration, unlike CC or PMI DC that require inference on additional inputs such as \\\"N/A\\\", \\\"[MASK]\\\", \\\"\\\", and x domain . 20 We can ignore the lack of log because it does not change the result of arg max. The methods displaying the most significant performance improvements in the PA scenario are MI AG , MI A , MI, and MDL M , particularly with the prompt selection performance of MI (PA) A and MDL (PA) M being the highest among different methods. On average, MI (PA) A increases the scaled F1 from 87.79% (0.5965/0.6795) to 99.44% (0.6757/0.6795) compared to the best existing method (GE) when the oracle prompt without calibration is used as the target of comparison. The scaled F1 of MI (PA) A calculated with respect to the oracle prompt with calibration is 96.85% (0.6757/0.6977).\\nNext, we assess the effectiveness of CBM calibration for answer selection by examining the proportion of prompts (out of 100) that show improved performance after applying calibration for answer selection. Figure 1b indicates that CBM is considerably more effective than CC and PMI DC in enhancing the performance of the prompts. The performance of more than half of the prompts increases after applying CBM in all 13 datasets. Additionally, the performance of nearly 100% of prompts improves with CBM calibration in 7 datasets. While CC and PMI DC improved almost none of the F1 of the prompts in story and hella, the performance of approximately 70% of the prompts increased with CBM calibration, possibly due to the more accurate calculation of p(y|t) as discussed in Section 4.3.\"\n },\n {\n \"figure_ref\": [\n \"fig_5\",\n \"fig_5\"\n ],\n \"heading\": \"Discussion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we discuss various findings that are relevant to our main experiments.\\nFigure 7a shows that the effectiveness of a probability-based prompt selection method remains consistent across models of different types and numbers of parameters, justifying our choice of using a single model (OPT 2.7B) as the representative for all experiments. Figure 7b shows that the trend of correlation between Prompt Selection Score and performance of the selected prompt is also quite consistent between different models.\\nFigure 8 shows the mean and standard deviation of the result of prompt selection among five different subsets of 50 prompts randomly sampled from the full set of 100 prompts, using the mainly discussed methods. The result shows that the performance of instance-wise prompt selection methods (MI AGL , MI AL , MDL) is not stable, likely due to the noisy nature of selecting one prompt for each instance. However, the performance of MI (PA) A and MDL (PA) M still achieves the highest performance and also shows the lowest standard deviation, proving the effectiveness of CBM.\\nThrough additional analysis, we find that (1) while strong performance in prompt selection does not consistently correlate with Prompt Selection Score, a broadly positive correlation is observed when averaged across most methods; (2) CBM im-proves the performance of MDL M by mitigating overconfidence; (3) MI, GE, and CBM methods face limitations when applied to dynamic datasets with extreme label imbalance; (4) top-performing prompt selection methods from the zero-shot setting, like MI (PA) A and MDL (PA) M , retain their effectiveness in the few-shot setting, further validating their robustness across different conditions.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Works\",\n \"publication_ref\": [\n \"b43\",\n \"b10\",\n \"b18\",\n \"b17\",\n \"b14\",\n \"b30\",\n \"b11\",\n \"b21\",\n \"b8\",\n \"b37\",\n \"b26\",\n \"b6\",\n \"b25\"\n ],\n \"table_ref\": [],\n \"text\": \"Recent advances in large language models (LLMs) have created the paradigm of prompt-based learning, which gives the benefit that a single pretrained LLM can be used to solve a great number of tasks with task-specific prompts. However, the performance of LLMs can heavily fluctuate according to the choice of prompts (Zhao et al., 2021;Holtzman et al., 2021;Lu et al., 2022). To mitigate this issue, prompt engineering attempts to find the prompt that results in the most effective performance on the downstream task (Liu et al., 2023).\\nAutomatic prompt engineering methods can be largely divided into two groups: the methods that use discrete prompts where the prompts are human-understandable actual text strings, and the methods that optimize continuous prompts where the prompts lie in the embedding space of the model (Li and Liang, 2021;Shin et al., 2020). Probability-based prompt selection methods that we study in this work 2.2 fall into the former group; most of the methods of the latter group require gradient-based training, while probabilitybased prompt selection does not perform any gradient-based update.\\nPrompt engineering methods using discrete prompts include prompt paraphrasing, prompt generation, and prompt selection. Among these, prompt paraphrasing or generation approaches can be used together with probability-based selection methods; prompt selection can be performed on the prompts generated through prompt paraphrasing or generation (Jiang et al., 2020;Mishra et al., 2022;Gao et al., 2021;Wang et al., 2023;Prasad et al., 2023;Kim et al., 2022;Deng et al., 2022). Among prompt selection methods other than the probability-based approaches, a large portion of the methods are not easily utilizable since they require training an additional model and/or the use of an additional component. (Zhang et al., 2022b) On the other hand, probability-based prompt selection offers the advantage of prompt selection requiring only the output probabilities of the LLM. While the prerequisite is a set of candidate prompts to select from, this data is relatively small in size and can be easily obtained from the research community (Bach et al., 2022) or via machine generation (OpenAI, 2023). One limitation of these methods, though, is that one cannot use them for closed-source LLMs that are only available via proprietary LLM APIs that do not provide output probability distributions. Also, when the number of candidate prompts |T | and the size of the dataset used to select the prompt |X| is large, the calculation for prompt selection becomes computationally heavy; using a smaller set X ′ ∈ X to choose the prompt for X can be helpful in such a case.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we address the need for a comprehensive evaluation to compare the existing probability-based prompt selection methods, which have been proposed and evaluated under varying conditions and datasets. To achieve this, we introduce a unified evaluation setup to compare these methods, conduct a thorough evaluation, and develop a unified framework of the existing probability-based prompt selection methods. Our analysis within this unified framework has provided insights into the relationship among existing methods, enabling the development of several combinational variants that improve performance. Furthermore, our research on probability calibration has revealed the limitations of existing calibration methods and led to the proposal of an enhanced calibration method, Calibration By Marginalization (CBM). CBM not only significantly improves prompt selection performance but also demonstrates robust answer selection enhancement across multiple datasets. We hope that our unified setup provides a foundation for fair evaluation between various prompt selection methods and that our findings yield deeper insights into probability-based prompt selection.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgements\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The authors would like to extend their sincere gratitude to the anonymous reviewers and Action Editor for their highly detailed and insightful comments and feedback. The authors would also like to thank Sang-Woo Lee for valuable feedback and discussions on the project. This work was partly supported by KT grant (2021, A study on a conversational language model that uses long external text as a prompt, 80%) and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub, 20%).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"* This project was initiated while the first author was a Master's student at KAIST (Nov 2022 -Feb 2023\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":""},"doi":{"kind":"string","value":""},"references":{"kind":"list like","value":[{"authors":"H Stephen; Victor Bach; Zheng-Xin Sanh; Albert Yong; Colin Webson; Raffel; V Nihal; Abheesht Nayak; Taewoon Sharma; M Kim; Thibault Saiful Bari; Zaid Fevry; Manan Alyafeai; Andrea Dey; Zhiqing Santilli; Srulik Sun; Canwen Ben-David; Gunjan Xu; Han Chhablani; Jason Wang; Alan Fries; Maged S Al-Shaibani; Shanya Sharma; Urmish Thakker; Khalid Almubarak; Xiangru Tang; Dragomir Radev; Mike Tian-Jian; Alexander M Jiang; Rush","journal":"","ref_id":"b0","title":"Promptsource: An integrated development environment and repository for natural language prompts","year":"2022"},{"authors":"Francesco Barbieri; Jose Camacho-Collados; Luis Espinosa-Anke; Leonardo Neves","journal":"","ref_id":"b1","title":"TweetEval:Unified Benchmark and Comparative Evaluation for Tweet Classification","year":"2020"},{"authors":"Yonatan Bisk; Rowan Zellers; Le Ronan; Jianfeng Bras; Yejin Gao; Choi","journal":"","ref_id":"b2","title":"Piqa: Reasoning about physical commonsense in natural language","year":"2020"},{"authors":"Sid Black; Leo Gao; Phil Wang; Connor Leahy; Stella Biderman","journal":"","ref_id":"b3","title":"GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow","year":"2021"},{"authors":"B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mccandlish; Ilya Radford; Dario Sutskever; Amodei","journal":"","ref_id":"b4","title":"Language models are few-shot learners","year":"2020"},{"authors":"Ting-Yun Chang; Robin Jia","journal":"","ref_id":"b5","title":"Data curation alone can stabilize in-context learning","year":"2023"},{"authors":"Mingkai Deng; Jianyu Wang; Cheng-Ping Hsieh; Yihan Wang; Han Guo; Tianmin Shu; Meng Song; Eric P Xing; Zhiting Hu","journal":"","ref_id":"b6","title":"RL-Prompt: Optimizing discrete text prompts with reinforcement learning","year":"2022"},{"authors":"Qingxiu Dong; Lei Li; Damai Dai; Ce Zheng; Zhiyong Wu; Baobao Chang; Xu Sun; Jingjing Xu; Lei Li; Zhifang Sui","journal":"","ref_id":"b7","title":"A survey on in-context learning","year":"2023"},{"authors":"Tianyu Gao; Adam Fisch; Danqi Chen","journal":"","ref_id":"b8","title":"Making pre-trained language models better few-shot learners","year":"2021"},{"authors":"Srini Hila Gonen; Terra Iyer; Noah A Blevins; Luke Smith; Zettlemoyer","journal":"","ref_id":"b9","title":"Demystifying prompts in language models via perplexity estimation","year":"2023"},{"authors":"Ari Holtzman; Peter West; Vered Shwartz; Yejin Choi; Luke Zettlemoyer","journal":"","ref_id":"b10","title":"Surface form competition: Why the highest probability answer isn't always right","year":"2021"},{"authors":"Zhengbao Jiang; Frank F Xu; Jun Araki; Graham Neubig","journal":"","ref_id":"b11","title":"How can we know what language models know? TACL","year":"2020"},{"authors":"Joon Hyuhng; Hyunsoo Kim; Junyeob Cho; Taeuk Kim; Kang Kim; Min Yoo; Sang-Goo Lee","journal":"","ref_id":"b12","title":"Self-Generated In-Context learning: Leveraging auto-regressive language models as a demonstration generator","year":"2022"},{"authors":"Sawan Kumar; Partha Talukdar","journal":"","ref_id":"b13","title":"Reordering examples helps during priming-based Few-Shot learning","year":"2021"},{"authors":"Lisa Xiang; Percy Li; Liang","journal":"","ref_id":"b14","title":"Prefixtuning: Optimizing continuous prompts for generation","year":"2021"},{"authors":"Chonghua Liao; Yanan Zheng; Zhilin Yang","journal":"","ref_id":"b15","title":"Zero-Label prompt selection","year":"2022"},{"authors":"Jiachang Liu; Dinghan Shen; Yizhe Zhang; Bill Dolan; Lawrence Carin; Weizhu Chen","journal":"","ref_id":"b16","title":"What makes good In-Context examples for GPT-3?","year":"2022"},{"authors":"Pengfei Liu; Weizhe Yuan; Jinlan Fu; Zhengbao Jiang; Hiroaki Hayashi; Graham Neubig","journal":"ACM Computing Surveys","ref_id":"b17","title":"Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing","year":"2023"},{"authors":"Yao Lu; Max Bartolo; Alastair Moore; Sebastian Riedel; Pontus Stenetorp","journal":"","ref_id":"b18","title":"Fantastically ordered prompts and where to find them: Overcoming Few-Shot prompt order sensitivity","year":"2022"},{"authors":"Xinxi Lyu; Sewon Min; Iz Beltagy; Luke Zettlemoyer; Hannaneh Hajishirzi","journal":"","ref_id":"b19","title":"Zicl: Zero-shot in-context learning with pseudodemonstrations","year":"2023"},{"authors":"Andrew L Maas; Raymond E Daly; Peter T Pham; Dan Huang; Andrew Y Ng; Christopher Potts","journal":"","ref_id":"b20","title":"Learning word vectors for sentiment analysis","year":"2011"},{"authors":"Swaroop Mishra; Daniel Khashabi; Chitta Baral; Yejin Choi; Hannaneh Hajishirzi","journal":"","ref_id":"b21","title":"Reframing instructional prompts to GPTk's language","year":"2022"},{"authors":"Saif Mohammad; Felipe Bravo-Marquez; Mohammad Salameh; Svetlana Kiritchenko","journal":"","ref_id":"b22","title":"Semeval-2018 task 1: Affect in tweets","year":"2018"},{"authors":"N Moniz; L Torgo","journal":"","ref_id":"b23","title":"Multi-source social feedback of online news feeds","year":"2018"},{"authors":"Nasrin Mostafazadeh; Michael Roth; Annie Louis; Nathanael Chambers; James Allen","journal":"","ref_id":"b24","title":"Lsdsem 2017 shared task: The story cloze test","year":"2017"},{"authors":" Openai","journal":"","ref_id":"b25","title":"","year":"2023"},{"authors":"Archiki Prasad; Peter Hase; Xiang Zhou; Mohit Bansal","journal":"","ref_id":"b26","title":"GrIPS: Gradient-free, editbased instruction search for prompting large language models","year":"2023"},{"authors":"Alec Radford; Jeff Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever","journal":"","ref_id":"b27","title":"Language models are unsupervised multitask learners","year":"2019"},{"authors":"Ohad Rubin; Jonathan Herzig; Jonathan Berant","journal":"","ref_id":"b28","title":"Learning to retrieve prompts for In-Context learning","year":"2022"},{"authors":"Victor Sanh; Albert Webson; Colin Raffel; Stephen H Bach; Lintang Sutawika; Zaid Alyafeai; Antoine Chaffin; Arnaud Stiegler; Teven Le Scao; Arun Raja; Manan Dey; M Saiful Bari; Canwen Xu; Urmish Thakker; Shanya Sharma Sharma; Eliza Szczechla; Taewoon Kim; Gunjan Chhablani; Nihal Nayak; Debajyoti Datta; Jonathan Chang; Mike Tian-Jian; Han Jiang; Matteo Wang; Sheng Manica; Zheng Xin Shen; Harshit Yong; Rachel Pandey; Thomas Bawden; Trishala Wang; Jos Neeraj; Abheesht Rozen; Andrea Sharma; Thibault Santilli; Jason Fevry; Alan Fries; Ryan Teehan; Tali Bers; Stella Biderman; Leo Gao; Thomas Wolf; Alexander M Rush","journal":"","ref_id":"b29","title":"Multitask prompted training enables zero-shot task generalization","year":"2022"},{"authors":"Taylor Shin; Yasaman Razeghi; I V Robert L Logan; Eric Wallace; Sameer Singh","journal":"","ref_id":"b30","title":"Au-toPrompt: Eliciting knowledge from language models with automatically generated prompts","year":"2020"},{"authors":"Richard Socher; Alex Perelygin; Jean Wu; Jason Chuang; Christopher D Manning; Andrew Y Ng; Christopher Potts","journal":"","ref_id":"b31","title":"Recursive deep models for semantic compositionality over a sentiment treebank","year":"2013"},{"authors":"Taylor Sorensen; Joshua Robinson; Christopher Michael Rytting; Alexander Glenn Shaw; Kyle Jeffrey Rogers; Alexia Pauline Delorey; Mahmoud Khalil; Nancy Fulda; David Wingate","journal":"","ref_id":"b32","title":"An information-theoretic approach to prompt engineering without ground truth labels","year":"2022"},{"authors":"Cynthia Van Hee; Els Lefever; Véronique Hoste","journal":"","ref_id":"b33","title":"Semeval-2018 task 3: Irony detection in english tweets","year":"2018"},{"authors":"Alex Wang; Yada Pruksachatkun; Nikita Nangia; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel R Bowman","journal":"","ref_id":"b34","title":"Superglue: A stickier benchmark for general-purpose language understanding systems","year":"2019"},{"authors":"Alex Wang; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel R Bowman","journal":"","ref_id":"b35","title":"GLUE: A multi-task benchmark and analysis platform for natural language understanding","year":"2019"},{"authors":"Benjamin Raja; Chenglei Heinzerling; Davut Emre Si; Elizabeth Taşar; Sabrina J Salesky; Wilson Y Mielke; Abheesht Lee; Andrea Sharma; Antoine Santilli; Arnaud Chaffin; Debajyoti Stiegler; Eliza Datta; Gunjan Szczechla; Han Chhablani; Harshit Wang; Hendrik Pandey; Jason Strobelt; Alan Fries; Jos Rozen; Leo Gao; Lintang Sutawika; M Saiful Bari; Maged S Al-Shaibani; Matteo Manica; Nihal Nayak; Ryan Teehan; Samuel Albanie; Sheng Shen; Srulik Ben-David; Stephen H Bach; Taewoon Kim; Tali Bers; Thibault Fevry; Trishala Neeraj; Urmish Thakker; Vikas Raunak; Xiangru Tang; Zheng-Xin Yong; Zhiqing Sun; Shaked Brody; Yallow Uri; Hadar Tojarieh; Adam Roberts; Hyung Won Chung; Jaesung Tae; Jason Phang; Ofir Press; Conglong Li; Deepak Narayanan; Hatim Bourfoune; Jared Casper; Jeff Rasley; Max Ryabinin; Mayank Mishra; Minjia Zhang; Mohammad Shoeybi; Myriam Peyrounette; Nicolas Patry; Nouamane Tazi; Omar Sanseviero; Pierre Patrick Von Platen; Pierre Cornette; Rémi François Lavallée; Samyam Lacroix; Sanchit Rajbhandari; Shaden Gandhi; Stéphane Smith; Suraj Requena; Tim Patil; Ahmed Dettmers; Amanpreet Baruwa; Anastasia Singh; Anne-Laure Cheveleva; Arjun Ligozat; Aurélie Subramonian; Charles Névéol; Dan Lovering; Deepak Garrette; Ehud Tunuguntla; Ekaterina Reiter; Ekaterina Taktasheva; Eli Voloshina; Genta Bogdanov; Hailey Indra Winata; Jan-Christoph Schoelkopf; Jekaterina Kalo; Jessica Novikova; Jordan Zosa Forde; Jungo Clive; Ken Kasai; Liam Kawamura; Marine Hazan; Miruna Carpuat; Najoung Clinciu; Newton Kim; Oleg Cheng; Omer Serikov; Oskar Antverg; Rui Van Der Wal; Ruochen Zhang; Sebastian Zhang; Shachar Gehrmann; Shani Mirkin; Tatiana Pais; Thomas Shavrina; Tian Scialom; Tomasz Yun; Verena Limisiewicz; Vitaly Rieser; Vladislav Protasov; Yada Mikhailov; Yonatan Pruksachatkun; Zachary Belinkov; Zdeněk Bamberger; Alice Kasner; Amanda Rueda; Amir Pestana; Ammar Feizpour; Amy Khan; Ana Faranak; Anthony Santos; Antigona Hevia; Arash Unldreaj; Arezoo Aghagol; Aycha Abdollahi; Azadeh Tammour; Bahareh Hajihosseini; Benjamin Behroozi; Bharat Ajibade; Carlos Muñoz Saxena; Daniel Ferrandis; Danish Mc-Duff; David Contractor; Davis Lansky; Douwe David; Kiela; A Duong; Edward Nguyen; Emi Tan; Ezinwanne Baylor; Ozoani","journal":"Fa","ref_id":"b36","title":"","year":""},{"authors":"Zhiyong Wu; Yaoxiang Wang; Jiacheng Ye; Lingpeng Kong","journal":"","ref_id":"b37","title":"Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering","year":"2023"},{"authors":"Hanwei Xu; Yujun Chen; Yulun Du; Nan Shao; Yanggang Wang; Haiyu Li; Zhilin Yang","journal":"","ref_id":"b38","title":"GPS: Genetic prompt search for efficient few-shot learning","year":"2022"},{"authors":"Rowan Zellers; Ari Holtzman; Yonatan Bisk; Ali Farhadi; Yejin Choi","journal":"","ref_id":"b39","title":"Hellaswag: Can a machine really finish your sentence? In ACL","year":"2019"},{"authors":"Susan Zhang; Stephen Roller; Naman Goyal; Mikel Artetxe; Moya Chen; Shuohui Chen; Christopher Dewan; Mona Diab; Xian Li; Xi Victoria Lin; Todor Mihaylov; Myle Ott; Sam Shleifer; Kurt Shuster; Daniel Simig; Punit Singh Koura; Anjali Sridhar; Tianlu Wang; Luke Zettlemoyer","journal":"","ref_id":"b40","title":"Opt: Open pre-trained transformer language models","year":"2022"},{"authors":"Xiang Zhang; Junbo ; Jake Zhao; Yann Lecun","journal":"","ref_id":"b41","title":"Character-level convolutional networks for text classification","year":"2015"},{"authors":"Yiming Zhang; Shi Feng; Chenhao Tan","journal":"","ref_id":"b42","title":"Active example selection for In-Context learning","year":"2022"},{"authors":"Tony Z Zhao; Eric Wallace; Shi Feng; Dan Klein; Sameer Singh","journal":"","ref_id":"b43","title":"Calibrate before use: Improving Few-Shot performance of language models","year":"2021"}],"string":"[\n {\n \"authors\": \"H Stephen; Victor Bach; Zheng-Xin Sanh; Albert Yong; Colin Webson; Raffel; V Nihal; Abheesht Nayak; Taewoon Sharma; M Kim; Thibault Saiful Bari; Zaid Fevry; Manan Alyafeai; Andrea Dey; Zhiqing Santilli; Srulik Sun; Canwen Ben-David; Gunjan Xu; Han Chhablani; Jason Wang; Alan Fries; Maged S Al-Shaibani; Shanya Sharma; Urmish Thakker; Khalid Almubarak; Xiangru Tang; Dragomir Radev; Mike Tian-Jian; Alexander M Jiang; Rush\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Promptsource: An integrated development environment and repository for natural language prompts\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Francesco Barbieri; Jose Camacho-Collados; Luis Espinosa-Anke; Leonardo Neves\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"TweetEval:Unified Benchmark and Comparative Evaluation for Tweet Classification\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Yonatan Bisk; Rowan Zellers; Le Ronan; Jianfeng Bras; Yejin Gao; Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Piqa: Reasoning about physical commonsense in natural language\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Sid Black; Leo Gao; Phil Wang; Connor Leahy; Stella Biderman\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; Tom Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeffrey Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mccandlish; Ilya Radford; Dario Sutskever; Amodei\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ting-Yun Chang; Robin Jia\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Data curation alone can stabilize in-context learning\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Mingkai Deng; Jianyu Wang; Cheng-Ping Hsieh; Yihan Wang; Han Guo; Tianmin Shu; Meng Song; Eric P Xing; Zhiting Hu\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"RL-Prompt: Optimizing discrete text prompts with reinforcement learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Qingxiu Dong; Lei Li; Damai Dai; Ce Zheng; Zhiyong Wu; Baobao Chang; Xu Sun; Jingjing Xu; Lei Li; Zhifang Sui\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"A survey on in-context learning\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Tianyu Gao; Adam Fisch; Danqi Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Making pre-trained language models better few-shot learners\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Srini Hila Gonen; Terra Iyer; Noah A Blevins; Luke Smith; Zettlemoyer\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Demystifying prompts in language models via perplexity estimation\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Ari Holtzman; Peter West; Vered Shwartz; Yejin Choi; Luke Zettlemoyer\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Surface form competition: Why the highest probability answer isn't always right\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zhengbao Jiang; Frank F Xu; Jun Araki; Graham Neubig\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"How can we know what language models know? TACL\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Joon Hyuhng; Hyunsoo Kim; Junyeob Cho; Taeuk Kim; Kang Kim; Min Yoo; Sang-Goo Lee\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Self-Generated In-Context learning: Leveraging auto-regressive language models as a demonstration generator\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Sawan Kumar; Partha Talukdar\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Reordering examples helps during priming-based Few-Shot learning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Lisa Xiang; Percy Li; Liang\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Prefixtuning: Optimizing continuous prompts for generation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Chonghua Liao; Yanan Zheng; Zhilin Yang\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Zero-Label prompt selection\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jiachang Liu; Dinghan Shen; Yizhe Zhang; Bill Dolan; Lawrence Carin; Weizhu Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"What makes good In-Context examples for GPT-3?\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Pengfei Liu; Weizhe Yuan; Jinlan Fu; Zhengbao Jiang; Hiroaki Hayashi; Graham Neubig\",\n \"journal\": \"ACM Computing Surveys\",\n \"ref_id\": \"b17\",\n \"title\": \"Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Yao Lu; Max Bartolo; Alastair Moore; Sebastian Riedel; Pontus Stenetorp\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Fantastically ordered prompts and where to find them: Overcoming Few-Shot prompt order sensitivity\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xinxi Lyu; Sewon Min; Iz Beltagy; Luke Zettlemoyer; Hannaneh Hajishirzi\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Zicl: Zero-shot in-context learning with pseudodemonstrations\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Andrew L Maas; Raymond E Daly; Peter T Pham; Dan Huang; Andrew Y Ng; Christopher Potts\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Learning word vectors for sentiment analysis\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"Swaroop Mishra; Daniel Khashabi; Chitta Baral; Yejin Choi; Hannaneh Hajishirzi\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Reframing instructional prompts to GPTk's language\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Saif Mohammad; Felipe Bravo-Marquez; Mohammad Salameh; Svetlana Kiritchenko\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Semeval-2018 task 1: Affect in tweets\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"N Moniz; L Torgo\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Multi-source social feedback of online news feeds\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Nasrin Mostafazadeh; Michael Roth; Annie Louis; Nathanael Chambers; James Allen\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Lsdsem 2017 shared task: The story cloze test\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \" Openai\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Archiki Prasad; Peter Hase; Xiang Zhou; Mohit Bansal\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"GrIPS: Gradient-free, editbased instruction search for prompting large language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Alec Radford; Jeff Wu; Rewon Child; David Luan; Dario Amodei; Ilya Sutskever\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Language models are unsupervised multitask learners\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Ohad Rubin; Jonathan Herzig; Jonathan Berant\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Learning to retrieve prompts for In-Context learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Victor Sanh; Albert Webson; Colin Raffel; Stephen H Bach; Lintang Sutawika; Zaid Alyafeai; Antoine Chaffin; Arnaud Stiegler; Teven Le Scao; Arun Raja; Manan Dey; M Saiful Bari; Canwen Xu; Urmish Thakker; Shanya Sharma Sharma; Eliza Szczechla; Taewoon Kim; Gunjan Chhablani; Nihal Nayak; Debajyoti Datta; Jonathan Chang; Mike Tian-Jian; Han Jiang; Matteo Wang; Sheng Manica; Zheng Xin Shen; Harshit Yong; Rachel Pandey; Thomas Bawden; Trishala Wang; Jos Neeraj; Abheesht Rozen; Andrea Sharma; Thibault Santilli; Jason Fevry; Alan Fries; Ryan Teehan; Tali Bers; Stella Biderman; Leo Gao; Thomas Wolf; Alexander M Rush\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Multitask prompted training enables zero-shot task generalization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Taylor Shin; Yasaman Razeghi; I V Robert L Logan; Eric Wallace; Sameer Singh\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Au-toPrompt: Eliciting knowledge from language models with automatically generated prompts\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Richard Socher; Alex Perelygin; Jean Wu; Jason Chuang; Christopher D Manning; Andrew Y Ng; Christopher Potts\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Recursive deep models for semantic compositionality over a sentiment treebank\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Taylor Sorensen; Joshua Robinson; Christopher Michael Rytting; Alexander Glenn Shaw; Kyle Jeffrey Rogers; Alexia Pauline Delorey; Mahmoud Khalil; Nancy Fulda; David Wingate\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"An information-theoretic approach to prompt engineering without ground truth labels\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Cynthia Van Hee; Els Lefever; Véronique Hoste\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Semeval-2018 task 3: Irony detection in english tweets\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Alex Wang; Yada Pruksachatkun; Nikita Nangia; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel R Bowman\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Superglue: A stickier benchmark for general-purpose language understanding systems\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Alex Wang; Amanpreet Singh; Julian Michael; Felix Hill; Omer Levy; Samuel R Bowman\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"GLUE: A multi-task benchmark and analysis platform for natural language understanding\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Benjamin Raja; Chenglei Heinzerling; Davut Emre Si; Elizabeth Taşar; Sabrina J Salesky; Wilson Y Mielke; Abheesht Lee; Andrea Sharma; Antoine Santilli; Arnaud Chaffin; Debajyoti Stiegler; Eliza Datta; Gunjan Szczechla; Han Chhablani; Harshit Wang; Hendrik Pandey; Jason Strobelt; Alan Fries; Jos Rozen; Leo Gao; Lintang Sutawika; M Saiful Bari; Maged S Al-Shaibani; Matteo Manica; Nihal Nayak; Ryan Teehan; Samuel Albanie; Sheng Shen; Srulik Ben-David; Stephen H Bach; Taewoon Kim; Tali Bers; Thibault Fevry; Trishala Neeraj; Urmish Thakker; Vikas Raunak; Xiangru Tang; Zheng-Xin Yong; Zhiqing Sun; Shaked Brody; Yallow Uri; Hadar Tojarieh; Adam Roberts; Hyung Won Chung; Jaesung Tae; Jason Phang; Ofir Press; Conglong Li; Deepak Narayanan; Hatim Bourfoune; Jared Casper; Jeff Rasley; Max Ryabinin; Mayank Mishra; Minjia Zhang; Mohammad Shoeybi; Myriam Peyrounette; Nicolas Patry; Nouamane Tazi; Omar Sanseviero; Pierre Patrick Von Platen; Pierre Cornette; Rémi François Lavallée; Samyam Lacroix; Sanchit Rajbhandari; Shaden Gandhi; Stéphane Smith; Suraj Requena; Tim Patil; Ahmed Dettmers; Amanpreet Baruwa; Anastasia Singh; Anne-Laure Cheveleva; Arjun Ligozat; Aurélie Subramonian; Charles Névéol; Dan Lovering; Deepak Garrette; Ehud Tunuguntla; Ekaterina Reiter; Ekaterina Taktasheva; Eli Voloshina; Genta Bogdanov; Hailey Indra Winata; Jan-Christoph Schoelkopf; Jekaterina Kalo; Jessica Novikova; Jordan Zosa Forde; Jungo Clive; Ken Kasai; Liam Kawamura; Marine Hazan; Miruna Carpuat; Najoung Clinciu; Newton Kim; Oleg Cheng; Omer Serikov; Oskar Antverg; Rui Van Der Wal; Ruochen Zhang; Sebastian Zhang; Shachar Gehrmann; Shani Mirkin; Tatiana Pais; Thomas Shavrina; Tian Scialom; Tomasz Yun; Verena Limisiewicz; Vitaly Rieser; Vladislav Protasov; Yada Mikhailov; Yonatan Pruksachatkun; Zachary Belinkov; Zdeněk Bamberger; Alice Kasner; Amanda Rueda; Amir Pestana; Ammar Feizpour; Amy Khan; Ana Faranak; Anthony Santos; Antigona Hevia; Arash Unldreaj; Arezoo Aghagol; Aycha Abdollahi; Azadeh Tammour; Bahareh Hajihosseini; Benjamin Behroozi; Bharat Ajibade; Carlos Muñoz Saxena; Daniel Ferrandis; Danish Mc-Duff; David Contractor; Davis Lansky; Douwe David; Kiela; A Duong; Edward Nguyen; Emi Tan; Ezinwanne Baylor; Ozoani\",\n \"journal\": \"Fa\",\n \"ref_id\": \"b36\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Zhiyong Wu; Yaoxiang Wang; Jiacheng Ye; Lingpeng Kong\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Hanwei Xu; Yujun Chen; Yulun Du; Nan Shao; Yanggang Wang; Haiyu Li; Zhilin Yang\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"GPS: Genetic prompt search for efficient few-shot learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Rowan Zellers; Ari Holtzman; Yonatan Bisk; Ali Farhadi; Yejin Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Hellaswag: Can a machine really finish your sentence? In ACL\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Susan Zhang; Stephen Roller; Naman Goyal; Mikel Artetxe; Moya Chen; Shuohui Chen; Christopher Dewan; Mona Diab; Xian Li; Xi Victoria Lin; Todor Mihaylov; Myle Ott; Sam Shleifer; Kurt Shuster; Daniel Simig; Punit Singh Koura; Anjali Sridhar; Tianlu Wang; Luke Zettlemoyer\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Opt: Open pre-trained transformer language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xiang Zhang; Junbo ; Jake Zhao; Yann Lecun\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Character-level convolutional networks for text classification\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Yiming Zhang; Shi Feng; Chenhao Tan\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Active example selection for In-Context learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tony Z Zhao; Eric Wallace; Shi Feng; Dan Klein; Sameer Singh\",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"Calibrate before use: Improving Few-Shot performance of language models\",\n \"year\": \"2021\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,96.96,175.77,390.88,54.56],"formula_id":"formula_0","formula_text":"Log-probability Mean ZLP s(x, y) = 1 |T | t log p(y|x, t) Probability Mean ZPM s(x, y) = 1 |T | t p(y|x, t) Majority Vote ZMV s(x, y) = t 1{arg max v∈Y p(y|x, t) = v} Perplexity (Gonen et al., 2023) PPL -1 |X| x 1 p(x,t)"},{"formula_coordinates":[3,330.02,607.58,168.55,27.09],"formula_id":"formula_1","formula_text":"p(y|x, t) = exp(log p(y|x, t)) y ′ ∈Y exp(log p(y ′ |x, t))"},{"formula_coordinates":[4,374.16,398.1,151.38,21.53],"formula_id":"formula_2","formula_text":"Π |ι(x,t)| i=1 p(ι(x, t) i |ι(x, t) <i ) 1 |ι(x,t)| ,"},{"formula_coordinates":[6,319.19,68.72,194.28,79.84],"formula_id":"formula_3","formula_text":"MI GE LE MDL ZPM PPL -\" ! ∈$ \" %∈& # $, & # & # ' $, & log \" %∈& # $, & # & # ' $, & --\" %∈& # $, & # & \" !∈$ [# ' $, & log # ' $, & ] → \" # $"},{"formula_coordinates":[7,72,64.03,86.14,15.19],"formula_id":"formula_4","formula_text":"p(t|x) = 1 |T | results"},{"formula_coordinates":[8,79.26,67.53,198.52,186.2],"formula_id":"formula_5","formula_text":"A G L Prompt Selection Score Existing Methods GE ✓ ✓ -H 1 |X| x one hot(p(y|x, t)) MDL ✓ -✓ -H(Y |x, t) MI ✗ ✗ ✗ GE M + MDL M Explored Variants GE M ✓ ✗ -H 1 |X| x p(y|x, t) MDL M ✓ -✗ -1 |X| x H (Y |x, t) MI A ✓ ✗ ✗ GE M + MDL M MI AG ✓ ✓ ✗ GE + MDL M MI AL ✓ ✗ ✓ GE M + MDL MI AGL ✓ ✓ ✓ GE + MDL"},{"formula_coordinates":[9,317.05,86.78,185.16,90.68],"formula_id":"formula_6","formula_text":"C = {\"N/A\", \"[MASK]\", \"\"} pcf = 1 |C| c∈C p(y|c, t) W = diag(p cf ) -1 , b = 0 q(y|x, t) = Wp(y|x, t) + b Domain Conditional PMI (PMI DC )"},{"formula_coordinates":[10,73.67,195.6,214.54,57.66],"formula_id":"formula_7","formula_text":"q(y|x, t) =diag(p cf ) -1 p(y|x, t) + 0 = p(y|x, t) pcf = p(y|x, t) 1 |C| c∈C p(y|c, t)"},{"formula_coordinates":[10,78.47,602.98,201.91,28.51],"formula_id":"formula_8","formula_text":"q(y|x, t) = p(y|x, t) p(y|t) = p(y|x, t) 1 |X| x ′ ∈X p(y|x ′ , t)"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 96.96,\n 175.77,\n 390.88,\n 54.56\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"Log-probability Mean ZLP s(x, y) = 1 |T | t log p(y|x, t) Probability Mean ZPM s(x, y) = 1 |T | t p(y|x, t) Majority Vote ZMV s(x, y) = t 1{arg max v∈Y p(y|x, t) = v} Perplexity (Gonen et al., 2023) PPL -1 |X| x 1 p(x,t)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 330.02,\n 607.58,\n 168.55,\n 27.09\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"p(y|x, t) = exp(log p(y|x, t)) y ′ ∈Y exp(log p(y ′ |x, t))\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 374.16,\n 398.1,\n 151.38,\n 21.53\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"Π |ι(x,t)| i=1 p(ι(x, t) i |ι(x, t) <i ) 1 |ι(x,t)| ,\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 319.19,\n 68.72,\n 194.28,\n 79.84\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"MI GE LE MDL ZPM PPL -\\\" ! ∈$ \\\" %∈& # $, & # & # ' $, & log \\\" %∈& # $, & # & # ' $, & --\\\" %∈& # $, & # & \\\" !∈$ [# ' $, & log # ' $, & ] → \\\" # $\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 72,\n 64.03,\n 86.14,\n 15.19\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"p(t|x) = 1 |T | results\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 79.26,\n 67.53,\n 198.52,\n 186.2\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"A G L Prompt Selection Score Existing Methods GE ✓ ✓ -H 1 |X| x one hot(p(y|x, t)) MDL ✓ -✓ -H(Y |x, t) MI ✗ ✗ ✗ GE M + MDL M Explored Variants GE M ✓ ✗ -H 1 |X| x p(y|x, t) MDL M ✓ -✗ -1 |X| x H (Y |x, t) MI A ✓ ✗ ✗ GE M + MDL M MI AG ✓ ✓ ✗ GE + MDL M MI AL ✓ ✗ ✓ GE M + MDL MI AGL ✓ ✓ ✓ GE + MDL\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 317.05,\n 86.78,\n 185.16,\n 90.68\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"C = {\\\"N/A\\\", \\\"[MASK]\\\", \\\"\\\"} pcf = 1 |C| c∈C p(y|c, t) W = diag(p cf ) -1 , b = 0 q(y|x, t) = Wp(y|x, t) + b Domain Conditional PMI (PMI DC )\"\n },\n {\n \"formula_coordinates\": [\n 10,\n 73.67,\n 195.6,\n 214.54,\n 57.66\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"q(y|x, t) =diag(p cf ) -1 p(y|x, t) + 0 = p(y|x, t) pcf = p(y|x, t) 1 |C| c∈C p(y|c, t)\"\n },\n {\n \"formula_coordinates\": [\n 10,\n 78.47,\n 602.98,\n 201.91,\n 28.51\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"q(y|x, t) = p(y|x, t) p(y|t) = p(y|x, t) 1 |X| x ′ ∈X p(y|x ′ , t)\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"Improving Probability-based Prompt Selection Through Unified Evaluation and Analysis"},"abstract":{"kind":"string","value":"Previous works in prompt engineering for large language models have introduced different gradient-free probability-based prompt selection methods that aim to choose the optimal prompt among the candidates for a given task but have failed to provide a comprehensive and fair comparison between each other. In this paper, we propose a unified framework to interpret and evaluate the existing probability-based prompt selection methods by performing extensive experiments on 13 common and diverse NLP tasks. We find that each of the existing methods can be interpreted as some variant of the method that maximizes mutual information between the input and the predicted output (MI). Utilizing this finding, we develop several other combinatorial variants of MI and increase the effectiveness of the oracle prompt selection method from 87.79% to 94.98%, measured as the ratio of the performance of the selected prompt to that of the optimal oracle prompt. Furthermore, considering that all the methods rely on the output probability distribution of the model that might be biased, we propose a novel calibration method called Calibration by Marginalization (CBM) that is orthogonal to the existing methods and helps increase the prompt selection effectiveness of the best method to 96.85%, achieving 99.44% of the oracle prompt F1 without calibration. 1"},"authors":{"kind":"string","value":"Sohee Yang; Jonghyeon Kim; Joel Jang; Seonghyeon Ye; Hyunji Lee; Minjoon Seo; Yizhong Wang; Yeganeh Kordi; Swaroop Mishra; Alisa Liu; Noah A Smith; Daniel Khashabi; Hannaneh 2023 Hajishirzi; Bigscience Workshop; Teven Le Scao; Angela Fan; Christopher Akiki; Ellie Pavlick; Suzana Ilić; Daniel Hesslow; Roman Castagné; Sasha Luccioni; François Yvon; Matthias Gallé; Jonathan Tow; Alexander M Rush; Stella Biderman; Albert Webson; Pawan Sasanka; Thomas Wang; Benoît Sagot; Niklas Muennighoff; Albert Villanova; Del Moral; Olatunji Ruwase; Rachel Bawden; Stas Bekman; Angelina Mcmillan-Major; Iz Belt- Agy; Huu Nguyen; Lucile Saulnier; Samson Tan; Pedro Ortiz Suarez; Victor Sanh; Hugo Lau- Rençon; Yacine Jernite; Julien Launay; Mar- Garet Mitchell; Colin Raffel; Aaron Gokaslan; Adi Simhi; Aitor Soroa; Alham Fikri Aji; Amit Alfassy; Anna Rogers; Ariel Kreisberg Nitzav; Canwen Xu; Chenghao Mou; Chris Emezue; Christopher Klamm; Colin Leong; Daniel Van Strien; David Ifeoluwa Adelani; Dragomir Radev; Eduardo González Ponfer- Rada; Efrat Levkovizh; Ethan Kim; Francesco De Toni; Gérard Dupont; Ger- Mán Kruszewski; Giada Pistilli; Hady Elsahar; Hamza Benyamina; Hieu Tran; Ian Yu; Idris Abdulmumin; Isaac Johnson; Itziar Gonzalez- Dios; Javier De La Rosa; Jenny Chim; Jesse Dodge; Jian Zhu; Jonathan Chang; Jörg Fro- Hberg; Joseph Tobing; Joydeep Bhattacharjee; Khalid Almubarak; Kimbo Chen; Kyle Lo; Von Le- Andro; Leon Werra; Long Weber; Loubna Phan; Ben; Ludovic Tanguy; Manan Dey; Manuel Romero Muñoz; Maraim Masoud; María Grandury; Mario Šaško; Max Huang; Maximin Coavoux; Mayank Singh; Tian- Jian Jiang; Minh Chien Vu; Mohammad A Jauhar; Mustafa Ghaleb; Nishant Subramani; Nora Kassner; Nurulaqilla Khamis; Olivier Nguyen; Omar Espejel; Ona De Gibert; Paulo Villegas; Peter Henderson; Pierre Colombo; Priscilla Amuok; Quentin Lhoest; Rheza Harliman; Rishi Bommasani; Roberto Luis López; Rui Ribeiro; Salomey Osei; Sampo Pyysalo; Sebastian Nagel; Shamik Bose; Shamsuddeen Hassan Muhammad; Shanya Sharma; Shayne Longpre; Somaieh Nikpoor; Stanislav Silberberg; Suhas Pai; Sydney Zink; Tiago Timponi Torrent; Timo Schick; Tristan Thrush; Valentin Danchev; Vassilina Nikoulina; Veronika Laippala; Violette Lepercq; Vrinda Prabhu; Zaid Alyafeai; Zeerak Talat"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: (a) F1 of the prompts selected by different probability-based prompt selection methods, averaged across 13 datasets. Per-dataset F1 and accuracy are shown in Figure 9. The methods without super/subscripts are the existing methods (Table1), while those with super/subscripts are our proposed methods (Table4& Equation1). (b) Ratio of the prompts (out of 100) whose F1 on each dataset improves by applying probability calibration for answer selection, averaged across 10 models. Our proposed calibration method, CBM (Equation1), is considerably more effective than CC and PMI DC (Table5) in enhancing the answer selection performance of the prompts.","figure_data":"","figure_id":"fig_1","figure_label":"1","figure_type":"figure"},{"figure_caption":"8Maas et al. (2011);Wang et al. (2019b);Zhang et al. (2015);Moniz and Torgo (2018); Barbieri et al. (2020); Mohammad et al. (2018); Van Hee et al. (2018); Wang et al. (2019a); Socher et al. (2013); Bisk et al. (2020); Mostafazadeh et al. (2017); Zellers et al. (","figure_data":"","figure_id":"fig_2","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: The highlighted parts of the equation are rough estimations of the Prompt Selection Score (PSS) of each method, i.e., the score of which the prompt with the maximum value is chosen by the prompt selection method. They show the connection between different probability-based prompt selection methods.","figure_data":"","figure_id":"fig_3","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: F1 of the prompts selected by MI A and MI, averaged for each setup of a different number of tokens of verbalizers and evaluation dataset category. |v| denotes the number of tokens of the verbalizers.","figure_data":"","figure_id":"fig_4","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 7 :7Figure 7: Scaled F1 and correlation of F1 of the selected prompts and Prompt Selection Score of different probability-based prompt selection methods for different models, averaged across 13 datasets.","figure_data":"","figure_id":"fig_5","figure_label":"7","figure_type":"figure"},{"figure_caption":"Figure 8: Mean and standard deviation of prompt selection among five sets of 50 prompts, sampled from the full set of 100 prompts.","figure_data":"","figure_id":"fig_6","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 9 :9Figure 9: F1 (top) and accuracy (bottom) of the prompts selected by the different probability-based prompt selection methods, shown for each dataset.","figure_data":"","figure_id":"fig_7","figure_label":"9","figure_type":"figure"},{"figure_caption":"use reinforcement learning for demonstration selection of in-context learning; Chang and Jia (2023) train a scorer and estimator for demonstration selection; Kumar and Talukdar (2021); Xu et al. (2022) use a genetic algorithm; Liu et al. (2022); Lyu et al. (2023); Rubin et al. (2022) use retrieval from a corpus to select the prompts.","figure_data":"","figure_id":"fig_8","figure_label":"","figure_type":"figure"},{"figure_caption":"Summary of the existing probability-based prompt selection methods. Notations used in the equations are explained at the end of Section 2.1.","figure_data":"","figure_id":"tab_2","figure_label":"1","figure_type":"table"},{"figure_caption":"","figure_data":", they propose three ways to calcu-late s(x, y): using the ensemble of log-probabilitymean, probability mean, and majority vote. We re-fer to them as ZLP, ZPM, and ZMV, respectively.While the authors of the original work applied fil-tering of prompts, we observed from our prelimi-nary experiments that filtering does not have a sig-nificant effect.Perplexity (PPL) Gonen et al. (2023) proposeto select one prompt for the evaluation dataset withwhich the language model exhibits the lowest av-erage perplexity of the instantiated prompt ι(x, t)as shown in the last row of Table 1. p(x, t) iscalculated as","figure_id":"tab_4","figure_label":"1","figure_type":"table"},{"figure_caption":"Examples of the created prompts. The prompts are written in Jinja for the use of Prompt-Source (Bach et al., 2022) APIs.we only display the results of OPT 2.7B throughout the paper since the overall trend remains mostly identical (shown in Section 5).","figure_data":"Evaluation Metrics Prompt selection perfor-mance is assessed using macro F1 of the selectedprompts. To compare the effectiveness of theprompt selection methods across different datasetsor models, we normalize the value by the perfor-mance of the oracle prompt (upper bound) andpresent it as scaled F1.","figure_id":"tab_7","figure_label":"3","figure_type":"table"},{"figure_caption":"in an alternative form, It is clear that PSS PPL differs from PSS MI because it considers the probability of x and t that PSS MI neglects. Applying the probabilistic assumption of MI p(x|t) = p(x) = 1 |X| to PSS PPL converts the equation to x p(t) |X| , causing PPL to select the prompt with the lowest perplexity irrespective of the input. Since Gonen et al. (","figure_data":"PSS ZPM =one hot(p(y|x,t)) ⊤ one hot(s(x,y))x∈Xs.t. s(x,y) =1 |T |t∈Tp(y|x,t) ≈ p(y|x)∴ PSS ZPM ≈one hot(p(y|x,t)) ⊤ one hot(p(y|x))x∈X≈1 |X|x∈Xp(y|x,t) ⊤ logp(y|x),which roughly corresponds to the negation ofthe second term of PSS MI , well-aligning the twomethods. 14MI vs.PPL PSS PPL is the most dissim-ilar from PSS MI , along with PSS LE . Sincearg min t1 |X|x1 p(x,t) = arg max t x p(x, t),PSS PPL can be expressed as","figure_id":"tab_9","figure_label":"","figure_type":"table"},{"figure_caption":"Top: differences among GE, MDL, and MI.","figure_data":"","figure_id":"tab_10","figure_label":"4","figure_type":"table"},{"figure_caption":"GE M is an MI-like version of GE employing p(y|x, t) without one-hot encoding, while PSS MDL M is an MI-like MDL version using the average of H(Y |x, t) for all x to select a single prompt. Contrarily, MI AG and MI AL are variants of MI, with the former emulating GE and the latter mirroring MDL, on top of MI A . MI AGL is another MI variant employing the sum of PSS GE","figure_data":"MIAGL MIAG MIAL MIAMI GE GEM LEMDL MDLM ZLPZPM ZMV PPLoracle prompt average of prompts worst prompt0.8F1 of Selected Prompt0.2 0.4 0.60.0Evaluation Dataset Category balanced unbalanced dynamic task averageFigure 5: F1 of the prompts selected by differentprobability-based prompt selection methods, averagedfor each dataset category, with the task average alsoshown. The methods with subscripts are the combi-national variants proposed in this subsection, whosePrompt Selection Scores are shown in Table 4 . Themethods with subscript M are combinational variantsthat use the component of MI; the methods with Lperform instance-wise prompt selection like MDL; theexist-ing methods. The variants that use instance-wiseprompt selection (MI AGL , MI AL , MDL) performbetter in balanced and unbalanced datasets but un-derperform in dynamic ones. Particularly in bal-anced datasets, MI AGL , MI AL , and MI A show sig-nificant improvement. While no method is consis-tently superior across all datasets (as observed inSection 2.4), MI AGL significantly improves scaledF1 to 94.98% (0.6454/0.6795) compared to that ofthe best existing method (GE), which is 87.79%(0.5965/0.6795).","figure_id":"tab_11","figure_label":"","figure_type":"table"},{"figure_caption":"MIAGL MIAG MIAL MIA GE GEM LE MDL MDLM ZLP ZPM ZMV PPL .90 0.91 0.89 0.82 0.79 0.67 0.85 0.65 0.83 0.75 0.74 0.56 0.96 0.96 0.96 0.92 0.91 0.73 0.64 0.82 0.71 0.79 0.77 0.70 0.59 0.96 0.92 0.93 0.91 0.88 0.77 0.67 0.80 0.63 0.73 0.68 0.67 0.57 0.94 0.90 0.94 0.92 0.84 0.76 0.66 0.84 0.67 0.83 0.75 0.77 0.53 0.95 0.93 0.93 0.91 0.88 0.76 0.65 0.84 0.65 0.81 0.75 0.76 0.59 0.90 0.88 0.93 0.92 0.83 0.79 0.61 0.80 0.74 0.81 0.76 0.73 0.59 0.95 0.92 0.95 0.89 0.83 0.83 0.65 0.85 0.68 0.77 0.73 0.71 0.60 0.93 0.94 0.91 0.92 0.87 0.75 0.60 0.80 0.63 0.84 0.76 0.79 0.61 0.70 * 0.77 * 0.52 * -0.18 * 0.18 * 0.53 0.23 0.12 -0.31 0.86 * 0.79 * 0.76 * 0.51 * -0.30 * 0.30 * 0.52 * 0.17 * 0.04 * -0.31 0.86 * 0.80 * 0.84 * 0.62 * -0.12 0.12 0.23 -0.16 -0.26 -0.30 0.87 * 0.74 * 0.80 * 0.59 * -0.23 0.23 0.61 0.30 0.22 -0.32 0.87 * 0.73 * 0.79 * 0.61 * -0.21 * 0.21 * 0.49 0.17 0.19 -0.35 0.83 * 0.71 * 0.75 * 0.48 * -0.29 * 0.29 * 0.48 0.27 0.20 -0.32 0.87 * 0.72 * 0.80 * 0.59 -0.26 * 0.26 * 0.52 * 0.27 * 0.16 * -0.29 0.85 * 0.75 * 0.76 * 0.54 -0.31 0.31 0.56 * 0.27 * 0.24 -0.33 0.87 * 0.76 * 0.77 * 0.53 -0.30 * 0.30 * 0.69 * 0.44 * 0.31 -0.31 0.86 * 0.77 * 0.78 * 0.53 -0.27 0.27 0.53 0.09 0.05 -0.35","figure_data":"ModelGPT-Neo 1.3B OPT 1.3B GPT2-XL 1.5B GPT-Neo 2.7B OPT 2.7B BLOOM 3B GPT-J 6B OPT 6.7B OPT 30B OPT 66BPrompt Selection Method 0.91 00.96 0.97 0.93 0.89 0.80 0.81 0.59 0.80 0.72 0.86 0.77 0.77 0.62 0.94 0.92 0.90 0.92 0.83 0.76 0.61 0.80 0.66 0.79 0.69 0.66 0.58 Scaled F1 of Selected Prompt0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95ModelGPT-Neo 1.3B BLOOM 3B OPT 66B OPT 30B OPT 6.7B GPT-J 6B OPT 2.7B GPT-Neo 2.7B GPT2-XL 1.5B OPT 1.3BPrompt Selection Method MIAG MIA GE GEM LE MDLM ZLP ZPM ZMV PPL 0.84 Corr(Selected Prompt F1, Prompt Selection Score)1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00(a) Scaled F1(b) Correlation of F1 of the selected prompts","figure_id":"tab_13","figure_label":"","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: (a) F1 of the prompts selected by different probability-based prompt selection methods, averaged across 13 datasets. Per-dataset F1 and accuracy are shown in Figure 9. The methods without super/subscripts are the existing methods (Table1), while those with super/subscripts are our proposed methods (Table4& Equation1). (b) Ratio of the prompts (out of 100) whose F1 on each dataset improves by applying probability calibration for answer selection, averaged across 10 models. Our proposed calibration method, CBM (Equation1), is considerably more effective than CC and PMI DC (Table5) in enhancing the answer selection performance of the prompts.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"8Maas et al. (2011);Wang et al. (2019b);Zhang et al. (2015);Moniz and Torgo (2018); Barbieri et al. (2020); Mohammad et al. (2018); Van Hee et al. (2018); Wang et al. (2019a); Socher et al. (2013); Bisk et al. (2020); Mostafazadeh et al. (2017); Zellers et al. (\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: The highlighted parts of the equation are rough estimations of the Prompt Selection Score (PSS) of each method, i.e., the score of which the prompt with the maximum value is chosen by the prompt selection method. They show the connection between different probability-based prompt selection methods.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: F1 of the prompts selected by MI A and MI, averaged for each setup of a different number of tokens of verbalizers and evaluation dataset category. |v| denotes the number of tokens of the verbalizers.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 7 :7Figure 7: Scaled F1 and correlation of F1 of the selected prompts and Prompt Selection Score of different probability-based prompt selection methods for different models, averaged across 13 datasets.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"7\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 8: Mean and standard deviation of prompt selection among five sets of 50 prompts, sampled from the full set of 100 prompts.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 9 :9Figure 9: F1 (top) and accuracy (bottom) of the prompts selected by the different probability-based prompt selection methods, shown for each dataset.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"9\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"use reinforcement learning for demonstration selection of in-context learning; Chang and Jia (2023) train a scorer and estimator for demonstration selection; Kumar and Talukdar (2021); Xu et al. (2022) use a genetic algorithm; Liu et al. (2022); Lyu et al. (2023); Rubin et al. (2022) use retrieval from a corpus to select the prompts.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_8\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Summary of the existing probability-based prompt selection methods. Notations used in the equations are explained at the end of Section 2.1.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \", they propose three ways to calcu-late s(x, y): using the ensemble of log-probabilitymean, probability mean, and majority vote. We re-fer to them as ZLP, ZPM, and ZMV, respectively.While the authors of the original work applied fil-tering of prompts, we observed from our prelimi-nary experiments that filtering does not have a sig-nificant effect.Perplexity (PPL) Gonen et al. (2023) proposeto select one prompt for the evaluation dataset withwhich the language model exhibits the lowest av-erage perplexity of the instantiated prompt ι(x, t)as shown in the last row of Table 1. p(x, t) iscalculated as\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Examples of the created prompts. The prompts are written in Jinja for the use of Prompt-Source (Bach et al., 2022) APIs.we only display the results of OPT 2.7B throughout the paper since the overall trend remains mostly identical (shown in Section 5).\",\n \"figure_data\": \"Evaluation Metrics Prompt selection perfor-mance is assessed using macro F1 of the selectedprompts. To compare the effectiveness of theprompt selection methods across different datasetsor models, we normalize the value by the perfor-mance of the oracle prompt (upper bound) andpresent it as scaled F1.\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"in an alternative form, It is clear that PSS PPL differs from PSS MI because it considers the probability of x and t that PSS MI neglects. Applying the probabilistic assumption of MI p(x|t) = p(x) = 1 |X| to PSS PPL converts the equation to x p(t) |X| , causing PPL to select the prompt with the lowest perplexity irrespective of the input. Since Gonen et al. (\",\n \"figure_data\": \"PSS ZPM =one hot(p(y|x,t)) ⊤ one hot(s(x,y))x∈Xs.t. s(x,y) =1 |T |t∈Tp(y|x,t) ≈ p(y|x)∴ PSS ZPM ≈one hot(p(y|x,t)) ⊤ one hot(p(y|x))x∈X≈1 |X|x∈Xp(y|x,t) ⊤ logp(y|x),which roughly corresponds to the negation ofthe second term of PSS MI , well-aligning the twomethods. 14MI vs.PPL PSS PPL is the most dissim-ilar from PSS MI , along with PSS LE . Sincearg min t1 |X|x1 p(x,t) = arg max t x p(x, t),PSS PPL can be expressed as\",\n \"figure_id\": \"tab_9\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Top: differences among GE, MDL, and MI.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_10\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"GE M is an MI-like version of GE employing p(y|x, t) without one-hot encoding, while PSS MDL M is an MI-like MDL version using the average of H(Y |x, t) for all x to select a single prompt. Contrarily, MI AG and MI AL are variants of MI, with the former emulating GE and the latter mirroring MDL, on top of MI A . MI AGL is another MI variant employing the sum of PSS GE\",\n \"figure_data\": \"MIAGL MIAG MIAL MIAMI GE GEM LEMDL MDLM ZLPZPM ZMV PPLoracle prompt average of prompts worst prompt0.8F1 of Selected Prompt0.2 0.4 0.60.0Evaluation Dataset Category balanced unbalanced dynamic task averageFigure 5: F1 of the prompts selected by differentprobability-based prompt selection methods, averagedfor each dataset category, with the task average alsoshown. The methods with subscripts are the combi-national variants proposed in this subsection, whosePrompt Selection Scores are shown in Table 4 . Themethods with subscript M are combinational variantsthat use the component of MI; the methods with Lperform instance-wise prompt selection like MDL; theexist-ing methods. The variants that use instance-wiseprompt selection (MI AGL , MI AL , MDL) performbetter in balanced and unbalanced datasets but un-derperform in dynamic ones. Particularly in bal-anced datasets, MI AGL , MI AL , and MI A show sig-nificant improvement. While no method is consis-tently superior across all datasets (as observed inSection 2.4), MI AGL significantly improves scaledF1 to 94.98% (0.6454/0.6795) compared to that ofthe best existing method (GE), which is 87.79%(0.5965/0.6795).\",\n \"figure_id\": \"tab_11\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"MIAGL MIAG MIAL MIA GE GEM LE MDL MDLM ZLP ZPM ZMV PPL .90 0.91 0.89 0.82 0.79 0.67 0.85 0.65 0.83 0.75 0.74 0.56 0.96 0.96 0.96 0.92 0.91 0.73 0.64 0.82 0.71 0.79 0.77 0.70 0.59 0.96 0.92 0.93 0.91 0.88 0.77 0.67 0.80 0.63 0.73 0.68 0.67 0.57 0.94 0.90 0.94 0.92 0.84 0.76 0.66 0.84 0.67 0.83 0.75 0.77 0.53 0.95 0.93 0.93 0.91 0.88 0.76 0.65 0.84 0.65 0.81 0.75 0.76 0.59 0.90 0.88 0.93 0.92 0.83 0.79 0.61 0.80 0.74 0.81 0.76 0.73 0.59 0.95 0.92 0.95 0.89 0.83 0.83 0.65 0.85 0.68 0.77 0.73 0.71 0.60 0.93 0.94 0.91 0.92 0.87 0.75 0.60 0.80 0.63 0.84 0.76 0.79 0.61 0.70 * 0.77 * 0.52 * -0.18 * 0.18 * 0.53 0.23 0.12 -0.31 0.86 * 0.79 * 0.76 * 0.51 * -0.30 * 0.30 * 0.52 * 0.17 * 0.04 * -0.31 0.86 * 0.80 * 0.84 * 0.62 * -0.12 0.12 0.23 -0.16 -0.26 -0.30 0.87 * 0.74 * 0.80 * 0.59 * -0.23 0.23 0.61 0.30 0.22 -0.32 0.87 * 0.73 * 0.79 * 0.61 * -0.21 * 0.21 * 0.49 0.17 0.19 -0.35 0.83 * 0.71 * 0.75 * 0.48 * -0.29 * 0.29 * 0.48 0.27 0.20 -0.32 0.87 * 0.72 * 0.80 * 0.59 -0.26 * 0.26 * 0.52 * 0.27 * 0.16 * -0.29 0.85 * 0.75 * 0.76 * 0.54 -0.31 0.31 0.56 * 0.27 * 0.24 -0.33 0.87 * 0.76 * 0.77 * 0.53 -0.30 * 0.30 * 0.69 * 0.44 * 0.31 -0.31 0.86 * 0.77 * 0.78 * 0.53 -0.27 0.27 0.53 0.09 0.05 -0.35\",\n \"figure_data\": \"ModelGPT-Neo 1.3B OPT 1.3B GPT2-XL 1.5B GPT-Neo 2.7B OPT 2.7B BLOOM 3B GPT-J 6B OPT 6.7B OPT 30B OPT 66BPrompt Selection Method 0.91 00.96 0.97 0.93 0.89 0.80 0.81 0.59 0.80 0.72 0.86 0.77 0.77 0.62 0.94 0.92 0.90 0.92 0.83 0.76 0.61 0.80 0.66 0.79 0.69 0.66 0.58 Scaled F1 of Selected Prompt0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95ModelGPT-Neo 1.3B BLOOM 3B OPT 66B OPT 30B OPT 6.7B GPT-J 6B OPT 2.7B GPT-Neo 2.7B GPT2-XL 1.5B OPT 1.3BPrompt Selection Method MIAG MIA GE GEM LE MDLM ZLP ZPM ZMV PPL 0.84 Corr(Selected Prompt F1, Prompt Selection Score)1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00(a) Scaled F1(b) Correlation of F1 of the selected prompts\",\n \"figure_id\": \"tab_13\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(Sorensen et al., 2022)\", \"Explanation\": \"The cited work provides the foundational concept of Mutual Information (MI) that the citing paper utilizes in the development of prompt selection methods for large language models (LLMs). The discovery in the cited work is used to propose several variants of MI that use different combinations of components in the equation of MI.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhao et al., 2021)\", \"Explanation\": \"The cited work on model output probability calibration methods is extended in the citing paper to propose an enhanced calibration method, Calibration By Marginalization (CBM), for improving the accuracy of probability-based prompt selection methods in LLMs.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Wu et al., 2023)\", \"Explanation\": \"The cited work by Wu et al. provides a definition of prompt selection in the context of zero-shot learning, which the citing paper uses to establish the basis for their own research on prompt selection in the same context.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Liao et al., 2022)\", \"Explanation\": \"The cited work by Liao et al. introduces the concept of instance-wise prompt selection, which the citing paper builds upon to discuss the idea of selecting different prompts for each evaluation instance in the context of few-shot learning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Gonen et al., 2023)\", \"Explanation\": \"The cited work introduces the concept of perplexity, which the citing paper adopts in their analysis of the performance of language models in zero-shot learning setups.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Sorensen et al., 2021)\", \"Explanation\": \"The cited work by Sorensen et al. provides a method for calculating the unnormalized logit for datasets with dynamically changing sentencetype answer choices, which the citing paper adopts in its research on probability-based prompt selection.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(2022)\", \"Explanation\": \"The cited work by Sorensen et al. (2022) provides a method for selecting a prompt that maximizes the mutual information between the evaluation instances and their model predictions, which the citing paper adopts in their zero-shot learning setup.\"}, {\"Category\": \"Data Source\", \"Citation\": \"Liao et al.\", \"Explanation\": \"The cited work by Liao et al. is a data source for the zero-label prompt selection (ZLP, ZPM, ZMV) method employed in the citing paper for their zero-shot learning setup.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Black et al., 2021)\", \"Explanation\": \"The cited work provides the GPT-Neo model with a size of 1.3B parameters, which the citing paper uses in its research or analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2022a)\", \"Explanation\": \"The cited work provides the OPT model with a size of 1.3B parameters, which the citing paper uses in its research or analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Radford et al., 2019)\", \"Explanation\": \"The cited work provides the GPT2-XL model, which the citing paper may have used in its research or analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Black et al., 2021)\", \"Explanation\": \"The cited work provides the GPT-Neo model with a size of 2.7B parameters, which the citing paper uses in its research or analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2022a)\", \"Explanation\": \"The cited work provides the OPT model with a size of 2.7B parameters, which the citing paper uses in its research or analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Workshop et al., 2023)\", \"Explanation\": \"The cited work provides the BLOOM model with a size of 3B parameters, which the citing paper may have used in its research or analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Black et al., 2021)\", \"Explanation\": \"The cited work provides the GPT-Neo model with a size of 6B parameters, which the citing paper uses in its research or analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2022a)\", \"Explanation\": \"The cited work provides the OPT model with a size of 6.7B parameters, which the citing paper uses in its research or analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2022a)\", \"Explanation\": \"The cited work provides the OPT model with a size of 30B parameters, which the citing paper uses in its research or analysis.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2022a)\", \"Explanation\": \"The cited work provides the OPT model with a size of 66B parameters, which the citing paper uses in its research or analysis.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liao et al., 2022)\", \"Explanation\": \"The cited work proposes a way to filter out low-quality prompts in advance, which the citing paper adopts in their experimental setup to improve the performance of their methods.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Gonen et al., 2023)\", \"Explanation\": \"The cited work is mentioned in the context of the underperformance of PPL compared to the results of Gonen et al. (2023), indicating that the citing paper is building upon the research of the cited work to further explore the performance of PPL in a different setting.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhao et al., 2021)\", \"Explanation\": \"The cited work by Zhao et al. (2021) highlights the need for normalization in the context of few-shot learning, which the citing paper incorporates in their method for prompt selection.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Holtzman et al., 2021)\", \"Explanation\": \"The cited work by Holtzman et al. (2021) provides evidence of the probabilistic nature of ranking based on string probability, which the citing paper uses to support their method for prompt selection.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhao et al., 2021)\", \"Explanation\": \"The cited work by Zhao et al. provides the equation for Answer Selection Contextual Calibration (CC), which the citing paper adopts in their research to calculate the probability distribution of the answer for a given prompt and input instance.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Holtzman et al., 2021)\", \"Explanation\": \"The cited work by Holtzman et al. also provides the equation for Answer Selection Contextual Calibration (CC), which the citing paper uses to calculate the probability distribution of the answer for a given prompt and input instance.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhao et al., 2021)\", \"Explanation\": \"The cited work by Zhao et al. provides the pcf (probability of correct answer) values for a given prompt and input instance, which the citing paper utilizes in their research to calculate the probability distribution of the answer.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Holtzman et al., 2021)\", \"Explanation\": \"The cited work by Holtzman et al. also provides the pcf (probability of correct answer) values for a given prompt and input instance, which the citing paper uses in their research to calculate the probability distribution of the answer.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Li and Liang, 2021)\", \"Explanation\": \"The cited work introduces the concept of continuous prompts in the embedding space, which the citing paper adopts in their study of probability-based prompt selection methods.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Shin et al., 2020)\", \"Explanation\": \"The cited work provides a method for optimizing continuous prompts in the embedding space, which the citing paper uses to study the performance of probability-based prompt selection methods.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Jiang et al., 2020)\", \"Explanation\": \"The cited work by Jiang et al. provides a method of prompt selection that can be used in conjunction with prompt paraphrasing or generation approaches to improve the performance of LLM models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Mishra et al., 2022)\", \"Explanation\": \"The cited work by Mishra et al. offers a method of prompt selection that can be used in conjunction with prompt paraphrasing or generation approaches to improve the performance of LLM models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Gao et al., 2021)\", \"Explanation\": \"The cited work by Gao et al. presents a method of prompt selection that can be used in conjunction with prompt paraphrasing or generation approaches to improve the performance of LLM models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang et al., 2023)\", \"Explanation\": \"The cited work by Wang et al. provides a method of prompt selection that can be used in conjunction with prompt paraphrasing or generation approaches to improve the performance of LLM models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Prasad et al., 2023)\", \"Explanation\": \"The cited work by Prasad et al. offers a method of prompt selection that can be used in conjunction with prompt paraphrasing or generation approaches to improve the performance of LLM models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kim et al., 2022)\", \"Explanation\": \"The cited work by Kim et al. presents a method of prompt selection that can be used in conjunction with prompt paraphrasing or generation approaches to improve the performance of LLM models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Deng et al., 2022)\", \"Explanation\": \"The cited work by Deng et al. provides a method of prompt selection that can be used in conjunction with prompt paraphrasing or generation approaches to improve the performance of LLM models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2022b)\", \"Explanation\": \"The cited work by Zhang et al. offers a method of prompt selection that can be used in conjunction with prompt paraphrasing or generation approaches to improve the performance of LLM models.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Bach et al., 2022)\", \"Explanation\": \"The cited work by Bach et al. is a data source that can be used to obtain a set of candidate prompts for prompt selection methods in LLM models.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(OpenAI, 2023)\", \"Explanation\": \"The cited work by OpenAI is a data source that can be used to obtain a set of candidate prompts for prompt selection methods in LLM models through machine generation.\"}]"}}},{"rowIdx":89,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_0"],"heading":"Introduction","publication_ref":["b0","b1","b2","b3","b4","b5","b6","b7","b8","b9","b10","b11","b12","b8","b13","b14","b15","b16","b9","b17","b18","b19","b19","b20","b20","b21","b21","b22"],"table_ref":[],"text":"Anomaly detection aims to identify samples that do not match the expected pattern or dataset. Anomaly detection techniques have been extensively studied in a variety of research and application domains, including industrial inspection [1][2][3] and medical applications [4]. With the continuous development of the industrial field, customers need to place various sensors for continuous monitoring of equipment condition and detect anomalies. The traditional manual visual detection method has recently become unsatisfactory because it is susceptible to the influence of manual experience and subjective factors. Therefore, anomaly detection has gradually become valuable in computer vision, attracting high attention in different fields [5][6][7]. In general, class imbalance problem [8] and the unexpectedness of anomalies are the two most common challenges of anomaly detection tasks. Anomalies are extremely rare in the industrial scene, which implies the number of aberrant samples obtained is quite low, resulting in a serious class imbalance problem. In addition, anomalies are always unexpected. It is hard to predict their location and size or even determine if anomalies occur at all. As a result, modeling all anomalies or even predicting all abnormalities that never occur is impractical on a few samples.\nThe existing anomaly detection tasks [9,10] focused on the classification of single or multiple categories at image-level. The current category to be detected is normal, and non-category are abnormal. However, in contrast to classification tasks, it is hard to train a model with full supervision for anomaly detection due to the lack of a large number of abnormal samples. [11,12] trained models in normal category, and samples were judged as anomalous when they show a large difference from the trained normal samples in the test stage. Recently, the anomaly detection tasks had confronted new challenges. To better monitor and process the anomalies, it is required not only to pick out the anomalous images but also to locate the anomalous regions. Bergmann et al. [13] proposed the MVTec AD dataset to provide benchmarks for anomaly detection and localization. The pixel-level methods [9,[14][15][16][17] exploited deep convolutional autoencoder and generative model such as Generative Adversarial Network (GAN) and Variational Autoencoder (VAE), respectively. [10,18,19] attempted to learn the feature representation of the normal samples from scratch and the model trained under the normal data manifold is hard to reconstruct the abnormal image. It may result in a large per-pixel reconstruction error and then a higher anomaly score. However, these methods showed a tremendous potential for misdetection due to low resolution of the reconstructed images and the strong generalization ability of the model.\nFortunately, it was found that using a pre-trained network could be a potential mode to learn feature representation for small sample datasets. Cohen et al. [20] utilized a group of features extracted from a deep pre-trained network on anomaly-free images to train their model. [20] improved performance gain, but their model relied on many sub-images of training data, resulting in expensive computations. To evade these limitations, Bergmann et al. [21] proposed a student-teacher network to implicitly model the distribution of features extracted from normal images. The idea of the student-teacher network was that the student network had poor generalization ability to abnormal data manifolds and then made wrong judgments. For knowledge transfer, [21] only imitated the last layer in the teacher network for knowledge distillation without making full use of the information of intermediate layers. To fully exploit the intermediate features of the teacher network, [22] proposed a novel knowledge distillation method that distilled the comprehensive knowledge of the pre-trained network at several critical layers to the trained network. It can provide the significance of multi-scale information in computer vision algorithms. Although features were extracted in several critical layers in [22], the information interaction between multi-scale features is fragile.\nLately, Transformer [23] had achieved great success in image classification and recognition, which proved the feasibility of Transformer in the information interaction of features with multiresolution. Inspired by this, we propose a multiresolution feature guidance method based on Transformer named GTrans for unsupervised anomaly detection and localization. In our method, the Anomaly Detected Network based on Transformer named Trans utilizes Transformer to effectively establish relationship between features with multiresolution. Under the tacit knowledge transfer of the Anomaly Guided Network (AGN), Trans enhances the fitting ability on the normal data manifold. Our intuition is that when abnormal images are input into the Trans, Trans cannot judge such accidents, resulting in large abnormalities in the whole image. Due to the strong ability of Trans in obtaining interaction information of features, the abnormal regions will receive great attention, conducive to the realization of pixel-level localization of the anomalies. Furthermore, we define a novel function to compute the anomaly score according to the difference of features extracted by AGN and Trans in the spatial distance and direction. The function can effectively capture the information of the combination of multi-scale features to obtain an accurate heat map of anomalous regions. We evaluate our method on MVTec AD dataset and then achieve superior performance in both localization and detection. Figure 1 shows visualized results of our method on the MVTec AD dataset. Our main contributions are summarized as follows:\n1. We propose a multiresolution feature guided method based on Transformer, referred as GTrans for unsupervised1 anomaly detection and localization.\nGTrans can improve the ability in information interaction between features with multiresolution. 2. We develop an anomaly guided network (AGN) to provide surrogate labels of features with multiresolution. Our model can be trained on normal images entirely without additional data augmentation. 3. We define a novel function to generate an anomaly map by exploiting the anomaly maps with multiresolution in spatial distance and direction."},{"figure_ref":[],"heading":"Experimental results","publication_ref":[],"table_ref":[],"text":"show that our proposed GTrans achieves state-of-theart performance on MVTec AD dataset, which demonstrates the significance of information interaction for anomaly detection and localization."},{"figure_ref":[],"heading":"Related Work","publication_ref":[],"table_ref":[],"text":"Anomaly detection has attracted a lot of attention in the last decades. We provide an extensive overview of anomaly detection techniques next. The research related to our work can categorize into two classes. First, we present an overview of anomaly detection and segmentation methods, which can be categorized into reconstruction-based methods and embedding similarity-based methods. Second, we briefly explain Transformer architectures and show the application of Transformer in anomaly detection."},{"figure_ref":[],"heading":"Anomaly Detection and Segmentation","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Reconstruction-based methods","publication_ref":["b9","b17","b18","b24","b10","b25","b26","b27","b8","b13","b14","b15","b16","b28","b24","b17","b9","b14","b8","b15","b16","b14","b9","b24","b8","b14","b15","b16"],"table_ref":[],"text":"Reconstruction-based methods like autoencoders (AEs) [10,18,19,25], variational autoencoders (VAEs) [11,[26][27][28] or generative adversarial networks (GANs) [9,[14][15][16][17]29] attempted to model the sample without abnormality and defect from scratch. The idea of AE-based methods was that if the model were trained on a dataset containing only normal samples, the reconstructed image would approach the normal sample regardless of whether the input images are normal samples or abnormal samples. By comparing the pixel error of the original image and the reconstructed image, the model could judge whether the input image is abnormal or not and even locate the abnormal regions. To learn potential features better, SSIM-AE [25] used SSIM as loss function and anomaly measure to compare input and reconstructed images. ARNet [18] learned the semantic feature embeddings related to the erased attributes by forcing the network to restore the original image. MemAE [10] proposed an autoencoder with a memory module to explicitly suppress the generalization capability of the autoencoder.\nVAE-based methods were also found to be used in anomaly detection and localization. In VAE-based methods, the probabilistic encoder and decoder both parameterized an isotropic normal distribution in the latent variable space and the original input variable space, respectively. The model endeavored to find the probability distribution conforming to the normal sample and used the reconstructed probability as the anomaly scores to generate an attention map in the potential space to detect the anomalies. However, VAE-based methods were not automatically superior to traditional autoencoder methods in general.\nGAN-based approaches utilized the discriminator to detect the slight disturbance in the reconstruction of abnormal images, prompting the generator to extract sample information from the potential space to reconstruct the image. AnoGAN [15] proposed a novel anomaly scoring scheme based on the mapping from image space to a latent space. OCGAN [9] utilized a dual latent space learning process to constrain the latent space of generators and discriminators to represent the specific categories exclusively. GANomaly [16] and f-AnoGan [17] added an additional encoder to the generator to reduce the inference time of AnoGan [15].\nMethods based on reconstruction were intuitive and explicative. However, these methods either failed to detect anomalies due to the strong generalization ability of deep models [10,25] or failed in one-class settings [9,[15][16][17]."},{"figure_ref":[],"heading":"Embedding similarity-based methods","publication_ref":["b29","b30","b31","b32","b33","b11","b32","b33","b34","b20","b20","b21","b35","b20","b21"],"table_ref":[],"text":"Embedding similarity-based methods mainly mapped the meaningful vector of the sample extracted from the pre-trained network to the high-dimensional feature space and judged the possibility of anomalies by computing the distance between the test sample and the normal sample in the feature space. Feature space expressed a higher level and more abstract information than image space. Embedding similarity-based methods could be categorized as cluster-based, embedding patch-based and knowledge-based methods.\nCluster-based methods used K-Nearest Neighbor (KNN) [30], K-means [31] or Principal Component Analysis (PCA) [32] to match the best approximate features from a memory-bank of nominal features or similar feature sets extracted from a pre-trained network to locate anomalies. However, the inference speed in the test phase of cluster-based methods was linearly related to the size of feature sets, which was limited in practical application.\nEmbedding patch-based methods tried to repair abnormal images by patch extracted from normal images. Patch SVDD [33] and CutPaste [34] applied self-supervised learning to anomaly detection. Unlike Deep SVDD [12], Patch SVDD [33] inspected the image at the patch level, and each patch corresponds to a point in the feature space. CutPaste [34] proposed a data augmentation strategy that cut an image patch and pasted it at a random location of an image. PaDiM [35] described each patch location with a Gaussian distribution and modeled the correlation between semantic layers. However, self-supervised representation typically underperformed those learned from large supervised datasets such as ImageNet.\nKnowledge-based methods mainly utilized the difference in generalization ability between teacher and student networks in a teacher-student framework to locate anomalies. The teacher network had a good performance and strong generalization ability, while the student network showed strong representation ability only in a single class by imitating the behavior of the teacher network. Bergmann et al. [21] were the first to introduce a teacher-student framework in the field of unsupervised anomaly detection and localization. Anomaly scores of [21] were derived from the predictive variance and regression error of an ensemble of student networks. On this basis, [22] and [36] extended the multiscale critical layers of knowledge extraction on the VGG and ResNet networks respectively, and achieved good performance. However, such methods either only extracted the single-layer information without making full use of the information in intermediate layers [21] or treated critical layers with multiresolutions equally while ignoring the importance of information interaction of multi-scale critical layers [22][36]."},{"figure_ref":[],"heading":"Transformer in Vision","publication_ref":["b22","b36","b22","b37","b38","b40","b37","b40","b41","b40","b42"],"table_ref":[],"text":"Ref. [23] first proposed Transformer, a model relying entirely on an attention mechanism to describe global dependencies between input and output, in the field of Natural Language Processing (NLP). Transformer did not use Recurrent Neural Network (RNN) such as Long Short-Term Memory (LSTM) [37] and instead utilized a stack of multi-headed attention blocks to accomplish NLP tasks excellently. The specific structure of the Transformer can be found in the original article [23].\nTransformer had achieved the most advanced performance in many NLP tasks and had become the preferred model for NLP tasks. Computer vision and NLP are merging as more efficient structures emerge. Due to the computational efficiency and scalability of Transformer, Transformer had been explored in computer vision and had become a new research direction.\nRecently, Transformer had achieved good performance in image classification [38], object detection [39][40] and image segmentation [41]. In general, there were two main model architectures for the adoption of Transformer in computer vision. One is a pure Transformer structure, and the other is a hybrid structure combining convolution neural network (CNN) and Transformer. Moreover, Vision Transformer [38] built a pure Transformer structure, dividing the input image into square uniform patches with 16 × 16 patch size. For each patch, the linear transformation was performed for dimensionality reduction and location information was embedded. Then, the projection information and location information was input into the Transformer to achieve image classification. [41] proposed TransUNet, which merited both Transformers and U-Net [42], to achieve medical image segmentation. TransUNet [41] utilized Transformer for encoding tokenized image patches from a feature map extracted by CNN, and the decoder upsampled the encoded features combined with the high-resolution CNN feature maps to enable precise localization.\nFor anomaly detection and localization, InTra [43] used a deep Transformer network consisting of a simple stack of multi-headed self-attention blocks to detect anomalies. Experiments showed that pure Transformer architecture is more efficient and scalable than traditional CNN in both model size and computational scale, while hybrid architecture performed better than pure Transformer at the smaller model size. Therefore, our method employ the hybrid structure combining CNN and Transformer to detect and locate anomalies and achieved excellent results."},{"figure_ref":["fig_1"],"heading":"Methodology","publication_ref":[],"table_ref":[],"text":"We propose a multiresolution feature guidance method based on Transformer named GTrans for anomaly detection and localization. GTrans consists of four components: 1) AGN: Anomaly guided network. 2) Trans: Anomaly detected network based on Transformer. 3) Learning of the normality. 4) Computation of the anomaly map. The overall architecture of GTrans is shown in Figure 2. Given a testing dataset D test = {y 1 , y 2 , ..., y M } consisting both anomaly and normal images. Given an input image y k ∈ R W ×H×C , k ∈ [1, M ] of width H, height W, and the number of channels C, we extract groups of features with multiresolution\nF G (y k ) = F 1 G (y k ) , F 2 G (y k ) , ..., F L G (y k ) and F T (y k ) = F 1 T (y k ) , F 2 T (y k ) , .\n.., F L T (y k ) from AGN pre-trained on ImageNet and the backbone of Trans, respectively, where L represents total number of critical layers. F G (y k ) and F T (y k ) are input into Tokenizer to obtain groups of tokens\nT G = T 1 G , T 2 G , ..., T L G and T T = T 1 T , T 2 T , .\n.., T L T , respectively. Trans utilizes Transformer to establish a relationship between features with multiresolution. AGN guides the decoder module of TFM to enhance the fitting ability of Trans in a normal data manifold. The mapper module maps the token output by TFM into original critical layers. In training, the parameters of AGN are frozen and the critical layers extracted by AGN as regression targets of the Trans. In testing, GTrans locates anomalies by comparing the differences in spatial distance and direction of multi-scale features extracted from the AGN and the Trans.\nIn the following, we describe the four components of GTrans in detail."},{"figure_ref":[],"heading":"Anomaly Guided Network","publication_ref":[],"table_ref":[],"text":"We propose a novel network called Anomaly Guided Network (AGN) for unsupervised anomaly detection and localization. AGN is a convolution network pre-trained on ImageNet (e.g., a ResNet-50-2 pre-trained on ImageNet), it can exhibit strong ability in feature representation when the sample quantity is small as ever. The guiding of AGN comes from two aspects. First, the token extracted from AGN as surrogate labels to be learned of Trans. Second, the feature maps extracted by AGN as regression targets of the Trans during the learning of the normality, enhancing the fitting ability of Trans in normal data manifolds."},{"figure_ref":[],"heading":"Feature Extraction","publication_ref":[],"table_ref":[],"text":"Given a training dataset D train = {x 1 , x 2 , ..., x N } consisting only of normal images. Given an input image\nx k ∈ R W ×H×C , k ∈ [1, N ] of width H, height W,\nand the number of channels C. Our goal is to create an ensemble of feature group\nF G (x k ) = F 1 G (x k ) , F 2 G (x k ) , ..., F L G (x k\n) to detect anomalies of various sizes, the l -th critical layer of AGN extracts a feature map\nF l G (x k ) ∈ R w l ×h l ×c l for l ∈ [1, L]\n, where h l and w l represent the height and width of the feature map, c l represents the number of channels."},{"figure_ref":[],"heading":"Token Extraction","publication_ref":["b43","b44","b45"],"table_ref":[],"text":"Inspired by [44][45][46], we utilize a module designated Tokenizer, containing point-wise convolutions and spatial attention, to extract visual tokens because a few visual tokens are sufficient to represent the entire image.\nWe reshape the input feature map F l G (x k ) ∈ R w l ×h l ×c l into a sequence of flattened two-dimensional feature map F l GK (x k ) ∈ R (w l •h l )×c l before entering the Tokenizer. Tokenizer uses two point-wise convolutions L g and L d to process each pixel of two-dimensional feature map F l GK (x k ) as follows:\nG l = L g F l GK (x k ) √ c l ∈ R (w l •h l )×g(1)\nand\nV l = L d F l GK (x k ) ∈ R (w l •h l )×d .(2)\nHere, L g maps each pixel of feature map in l -th critical layer to g semantic groups G l , while L d realizes the information interaction between channels of the feature map by raising or reducing dimension. V l represents the highdimensional feature representation of l -th critical layers. For each semantic group, spatial attention is computed for the sequences G l as a weighted average over V l as follow:\nT l = softmax(G l ) T V l ∈ R d×g ,(3)\nwhere softmax(•) obtains the spatial attention weight of each semantic group.\nIn general, we input F G (x k ) into the Tokenizer to obtain a group of visual tokens T G = T 1 G , T 2 G , ..., T L G with different resolutions. For anomaly detection, T G and F G (x k ) provide guidance in decoder modules and the learning of the normality, respectively."},{"figure_ref":[],"heading":"Trans","publication_ref":[],"table_ref":[],"text":"We propose a anomaly detected network based on Transformer named Trans. Trans improves the ability in information interaction between features with multiresolution. In the following, we describe the steps of Trans in detail."},{"figure_ref":[],"heading":"Extraction of feature and token","publication_ref":[],"table_ref":[],"text":"To effectively obtain the tacit knowledge of AGN, we rely on architecture that the dimension of the features of Trans is aligned with that of AGN. Trans and AGN have the same backbone, but Trans has not been pre-trained. In CNN, top layers produce low-resolution, semantically strong features, while bottom layers produce high-resolution, semantically weak ones. Given an input image x k ∈ R W ×H×C of width H, height W, and the number of channels C. To detect and locate anomalies with various sizes, we extract a group of features\nF T (x k ) = F 1 T (x k ) , F 2 T (x k ) , .\n.., F L T (x k ) at various layers. We input F T (x k ) into the Tokenizer to obtain a group of visual tokens T T = T 1 T , T 2 T , ..., T L T with multiresolutions.\nFig. 3 Overview of the TFM architecture."},{"figure_ref":[],"heading":"TFM","publication_ref":["b46","b37","b40"],"table_ref":[],"text":"This module aims to establish information interaction of spatially distant concepts for visual tokens of different semantic groups at different levels. We employ the encoder-decoder structure of standard Transformer with the following changes: (1) Omitting position embedding. [47] provided convincing evidence that CNNs do indeed rely on and learn information about spatial positioning. Therefore, the convolution layer computes learnable weights of query q e ∈ R d×(g•l) , key k e ∈ R d×(g•l) , and value v e ∈ R d×(g•l) without additional position embedding. (2) Using a non-linearity activation function and two point-wise convolutions replace position-wise feed-forward networks of the standard Transformer. (3) Extracting a group of visual tokens\nT G = T 1 G , , T 2 G , .\n.., T L G from AGN as the input of the decoder in Transformer. We name the module modified by standard Transformer as TFM. Figure 3 illustrates the architecture of the TFM. As can be seen, TFM is composed of a stack of S encoders and decoders. Each encoder and decoder has three sub-layers, respectively.\nThe first is a self-attention mechanism, the second is a combination of a non-linearity activation function and two point-wise convolutions, and the third is a layer normalization. We take an individual TFM block of S = 1 as an example below. We concatenate the group of visual tokens T G and T T into E in ∈ R d×(g•l) and D in ∈ R d×(g•l) , respectively. We take E in and D in as the input of the encoder and decoder in TFM.\nIn the encoder, we compute q e , k e , v e via\nq e = W q E in , k e = W k E in , v e = W v E in ∈ R d×(g•l)(4)\nwith learnable weight matrices W q , W k , W v ∈ R d×d . Then, self-attention is computed by a compatibility function of the q e with corresponding k e as a weighted average of the v e as follows:\nA e = v e softmax k T e q e √ d ∈ R d×(g•l) ,(5)\nwhere A e is the output of self-attention mechanism in encoder. We employ a residual connection around the self-attention mechanism and layer normalization to get the output of layer normalization E a :\nE a = E in + LayerN orm (A e ) ∈ R d×(g•l) ,(6)\nwhere LayerN orm (•) represents the layer normalization. Then, we use a residual connection around the convolution layer, followed by layer normalization to get the output of encoder E out :\nE out = LayerN orm (E a + L 1 σ (L 2 E a )) ∈ R d×(g•l) ,(7)\nwhere L 1 , L 2 ∈ R d×d are point-wise convolutions, σ (•) is the ReLU function.\nFor the decoder, most methods [38,41] previously applied to computer vision tasks use only the encoder of the Transformer. However, if the selfattention mechanism is only established in the visual token extracted from Tokenizer in Trans, the final model will easily fall into the problem of local optimal. Therefore, we employ an encoder-decoder structure of the Transformer and utilize visual tokens extracted from AGN as surrogate labels for the guidance of E out to enhance the stability of the Trans. We set\nq d = W q D in , k d = W k E out , v d = W v E out ∈ R d×(g•l) ,(8)\nA d = v d softmax k T d q d √ d ∈ R d×(g•l) .(9)\nHere, q d is the query computed from D in . k d , v d are the key and value computed from the output of the encoder E out . A d is the output of self-attention mechanism in decoder.\nSimilarly to the encoder, first, we employ a residual connection around the self-attention mechanism and layer normalization to get the output of layer normalization D a in the decoder. Second, we employ another residual connection around the convolution layer, followed by layer normalization to get the output of decoder D out . The computational process is as follows:\nD a = D in + LayerN orm (A d ) ∈ R d×(g•l)(10)\nand\nD out = LayerN orm (D a + L 1 σ (L 2 D a )) ∈ R d×(g•l) .(11)"},{"figure_ref":[],"heading":"Mapping of feature and token","publication_ref":[],"table_ref":[],"text":"To achieve pixel-level alignment between the outputs of AGN and the Trans, we divide D out into g groups and map each group into the latent space of the same dimension as the\nF G (x k ) = F 1 G (x k ) , F 2 G (x k ) , ..., F L G (x k )\n, respectively. For the l -th critical layer, we set\nF l M (x k ) = F l T (x k ) + T l v softmax X l q T T l k √ c l T ,(12)\nwhere F l M (x k ) is the final output of the l -th critical layer, X l q is the query computed from F l G (x k ). T l k , T l v are the key and value computed from the output of the encoder in TFM E out .\nIn general, the final output of Trans is a group of critical layers\nF M (x k )= F 1 M (x k ), F 2 M (x k ), ..., F L M (x k ) ."},{"figure_ref":[],"heading":"Learning of the normality","publication_ref":["b47"],"table_ref":[],"text":"In the following, we train the Trans so that its output is as similar as possible to the AGN, and the parameters of AGN are frozen during the entire training phase. Given an input image x k ∈ D train , we obtain a group of feature maps\nF G (x k )= F 1 G (x k ), F 2 G (x k ), ..., F L G (x k ) and F M (x k )= F 1 M (x k ), F 2 M (x k ), ..., F L M (x k\n) extracted from AGN and Trans, respectively. As indicated in [48], Trans can learn the distilled knowledge from AGN that is trained at a different task and guided by intermediate-level hints from AGN. Trans can learn the intermediate-level hints from different semantic layers of AGN. It then takes the intermediate-level hints as targets of optimization procedure to intensify the complete knowledge transfer from AGN to Trans. We define pixel-wise L 2 loss function p l (i,j) at position (i, j)\nfor l ∈ [1, L], i ∈ [1, w l ], j ∈ [1, h l ] as follow: p l (i,j) (x k ) = 1 2 F l G (x k ) (i,j) -F l M (x k ) (i,j) 2 2 . (13\n)\nHere, F l G (x k ) and F l M (x k ) represent the feature map extracted from l -th critical layers of AGN and Trans. We get the total loss L (x k ) of input image x k by the weighted average of all pixels in each feature map of critical layers in sets F l G (x k ) and F l M (x k ) as follow:\nL (x k ) = L l=1 1 w l • h l w l i=1 h l j=1 p l (i,j) (x k ), (14\n)\nwhere L represents total number of critical layers."},{"figure_ref":[],"heading":"Computation of the anomaly map","publication_ref":[],"table_ref":[],"text":"In training phase, Trans has stronger fitting ability in anomaly-free images than the AGN under the guidance of AGN. However, AGN is knowledgeable on anomaly images because of the strong generalization2 , while Trans is unfamiliar with such images. Therefore, the abnormal regions of anomaly images show a large deviation from the training data manifold of Trans when the anomaly images are input into AGN and Trans, respectively. Given a testing dataset D test = {y 1 , y 2 , ..., y M } consisting both anomaly and normal images, we assign an anomaly map M (y k ) ∈ R H×W to a test image\ny k ∈ D test for k ∈ [1, M ],\nwhere H and W are the height and width of the test image. The anomaly score of pixel at position (i, j)\nof i ∈ [1, H], j ∈ [1, W ]\nindicates the deviation degree of pixel at position (i, j) from the data manifold of anomalyfree images. We utilize the anomaly map M (y k ) to realize anomaly detection and localization. Next, we show the specific computation of anomaly map. We input a test image y k into our model to obtain the group of feature maps\nF G (y k )= F 1 G (y k ), F 2 G (y k ), ..., F L G (y k ) and F M (y k )= F 1 M (y k ), F 2 M (y k ), ..., F L M (y k )\nextracted from different critical layers of AGN and Trans, respectively.\nWe get a group of loss P = P 1 , P 2 , ..., P L by computing pixel-wise L 2 loss between the feature map from the same layers at F G (y k ) and F M (y k ) as follows:\nP l = 1 w l • h l c l i=1 1 2 F l G (y k ) -F l M (y k ) 2 2 ∈ R w l ×h l , (15\n)\nwhere P l ∈ P denotes the pixel-wise loss of l -th critical layer between AGN and Trans. We assign a weight to different critical layers because the different critical layers correspond to anomalies at various sizes, respectively. Here, we utilize MSE loss and cosine similarity metric of feature map in l -th critical layer to define two coefficients α l mse and α l cos as follows:\nα l mse = 1 w l • h l w l i=1 h l j=1 F l G (y k ) (i,j) -F l M (y k ) (i,j) 2(16)\nand\nα l cos = 1 - vec F l G (y k ) T • vec F l M (y k ) vec F l G (y k ) vec F l M (y k ) . (17\n)\nHere, vec (•) is a vectorization function transforming a matrix with arbitrary dimensions into a 1-D vector.\nα l mse and α l cos denote the similarity in spatial distance and direction of lth feature maps from AGN and Trans. The larger the values of α l mse and α l cos are, the less similarity the l -th feature maps are in spatial distance and direction, and the greater the possibility of anomalies. We effectively combine the similarity of feature maps in spatial distance and direction, and use harmonic mean of α l mse and α l cos as the weight of l -th feature maps. α l is formulated as\nα l = λ • α l cos • α l mse α l mse + λ • α l cos , (18\n)\nwhere λ is set to make the scale of α l mse and α l cos the same. Our goal is to assign high weight to feature maps with low similarity, so as to get high anomaly score with anomalous regions eventually. Therefore, the anomaly map is achieved by\nM (y k ) = G σ L l=1 α l R P l ∈ R H×W . (19\n)\nHere, R (•) resizes the elements of P = P 1 , P 2 , ..., P L to the spatial size of (H × W ). G σ (•) represents a Gaussian filter with standard deviation of σ. Finally, we define the region with high anomaly score in M (y k ) as anomalous regions and the maximum value of M (y k ) as the final anomaly score of the test image y k for anomaly detection."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":"In this section, we first elaborate on the details of GTrans structure and its parameters. Second, to demonstrate the effectiveness of our approach, we compare our method to state-of-the-art results on the MVTec AD benchmark, considering in image-level and pixel-level, respectively."},{"figure_ref":[],"heading":"Datesets and metrics","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Datesets","publication_ref":[],"table_ref":[],"text":"We evaluate GTrans on the MVTec AD dataset 3 which contains over 5000 high-resolution images divided into fifteen different object and texture categories. Each category of the MVTec AD dataset comprises a set of anomaly-free training images and a test set of images with various kinds of anomalies as well as images without anomaly."},{"figure_ref":[],"heading":"Metrics","publication_ref":[],"table_ref":[],"text":"We use the Area Under the Receiver Operating Characteristic curve (AUROC) and Area Under the Per-Region-Overlap Curve (AUPRO) to evaluate the performence of GTrans. AUROC is computed on different levels of threshold in favor of large anomalous regions. AUPRO computes a threshold-independent 3 The datasets analysed during the current study are available at https://www.mvtec.com/company/research/datasets/mvtec-ad. evaluation metric based on the Per-Region-Overlap (PRO) to better account for large and small anomalies in localization. Only the standard AUROC provides image-level anomaly detection."},{"figure_ref":[],"heading":"Implementation Details","publication_ref":[],"table_ref":[],"text":"We resize all images in the MVTec AD dataset to 256 × 256 and center crop them to 224 × 224. We take the preprocessed images of MVTec as the input of GTrans. We divided eighty percent of the MVTec AD dataset into the training set and twenty percent into the validation set without additional data enhancement. In the training phase, GTrans is trained only on the normal images with a batch size of 32 for 300 epochs. We utilize Adam optimizer with initial learning rate 10 -3 and weight decay 10 -4 for optimization. To enhance the stability of the GTrans, we define a large initial learning rate to improve the convergence speed of GTrans, and then the learning rate decreases with the increase of the number of iterations gradually. We utilize exponential decay equation to realize the decay of learning rate as\nlr = lr init × rate step totalstep ,(20)\nwhere lr init and lr represent the initial and current learning rate respectively, rate is decay factor with a value of 0.9, step and totalstep denote the current and total iteration number respectively. We implement our method in PyTorch and conduct all experiments on a machine equipped with an Intel i9-9900X and an NVIDIA GeForce RTX 2080 Ti GPU."},{"figure_ref":["fig_1"],"heading":"Architecture Details","publication_ref":[],"table_ref":[],"text":"As shown in Figure 2, GTrans consists of AGN and Trans. We illustrate the parameter selection of GTrans below."},{"figure_ref":[],"heading":"AGN","publication_ref":[],"table_ref":[],"text":"We extract features with a ResNet-34 (R34) and a Wide ResNet-50-2 (WR50). All backbones of AGN pre-trained on ImageNet, and the parameters of AGN are frozen during the training phase. As to the position of the selected guide feature, we choose features extracted by three intermediate layer groups, that is l = 3."},{"figure_ref":[],"heading":"Trans","publication_ref":["b49","b14","b24","b15","b19","b32","b33","b34","b42"],"table_ref":[],"text":"Trans is composed of a backbone network, a Tokenizer, a TFM and a mapper. We keep the backbone of Trans consistent with AGN. For Tokenizer, inspired by [50], we use a few visual tokens to generalise semantic concepts. Hence, we set the number of semantic groups g = 8 and dimension d = 256. For TFM, we utilize as few as two decoders and encoders to achieve superior performance. After mapper, the Trans and AGN output three feature maps with the size 56 × 56, 28 × 28 and 14 × 14, respectively.\nTable 1 The AUROC results for anomaly detection on the MVTec AD dataset using AUROC%.\nCategory AnoGAN [15] AE-SSIM [25] GANomaly [16] SPADE [20] Patch-SVDD [33] CutPaste [34] PaDiM [35] InTra [43] "},{"figure_ref":[],"heading":"Localization","publication_ref":["b42","b42","b19","b42","b19","b20"],"table_ref":["tab_1","tab_1"],"text":"Table 2 shows the AUROC and the PRO-score results for anomaly localization on the MVTec AD dataset. As shown in Table 2, GTrans-WR50 outperforms all the other methods in both the AUROC and the PRO-score on average for all the categories. Compared with the method based on Transformer, the performance of our method is 1.2% higher than InTra [43]. For PRO-score, InTra [43] and SPADE [20] achieve the best performance in the texture and object categories respectively. However, our method performs well in both texture and object categories, outperforming InTra [43] by 0.9% and SPADE [20] 1.3% in PRO-score on average for all the categories. When we further analyze the performance of GTrans-WR50, we find that GTrans-WR50 outperforms Student [21] by 7.3% in PRO-score, indicating the effectiveness of our method in multi-scale feature fusion and information interaction.\n5 Ablation Studies"},{"figure_ref":[],"heading":"Intermediate Knowledge","publication_ref":[],"table_ref":["tab_2","tab_2"],"text":"In this part, we evaluate the impact of the position of selected guide features in GTrans. In Table 3, we show the performance of anomaly detection and localization on the MVTec AD dataset of GTrans with a R34 backbone when using different combination critical layers (Layer 1+2+3+4, Layer 1+2+3, Layer 2+3+4, Layer 3+4). The final anomalous maps of all experiments in this part simply add the anomalous maps of different layers. It is known that feature layers of CNN can express various levels of abstract information. Bottom layers tend to extract low-level information such as textures, while top layers pay attention to low-resolution features that contain semantic information. It can be observed from Table 3 that Layer 1+2+3 can achieve the best performance in both texture and object categories. From the comparison of Layer 1+2+3+4 and Layer 1+2+3, it can be seen that the performance of anomaly detection and localization declines when the anomalous map generated by the fourth critical layer is added. The reason is that the features of the top layers are low-resolution, leading to the rough segmentation in detail texture. According to the experimental results of Layer 2+3, the information of the intermediate layers can express and extract anomalies to a large extent. When the information of the shallow layer is added on this basis, the model can extract the anomalies and pay attention to the processing of edge texture, improving the performance of the model."},{"figure_ref":[],"heading":"The structure of TFM","publication_ref":[],"table_ref":["tab_3"],"text":"This part evaluates the impact of an added decoder and the number of TFM blocks on experimental results.\nIn Table 4, we show the anomaly detection and localization performance on the MVTec AD dataset of GTrans with an R34 backbone. It can be seen that the detection AUROC of the model with TFM is 2.7% higher than that without TFM, while for localization AUROC, it can achieve 1.9% higher than the latter model.\nIn Table 5, we show the performance in anomaly detection and localization on the MVTec AD dataset of GTrans with a R34 backbone when using the different structure of TFM in Trans, where S represents the number of TFM blocks. For structure which added decoder, we decode the output of the encoder through the features extracted from the AGN.\nFrom Table 5, we can notice that for the same number of TFM blocks, the structure which added decoder outperforms the structure of pure encoder on MVTec AD dataset by 0.2% to 0.4% in the detection AUROC and 0.2% in localization AUROC. On this basis, we conduct experiments with the different number of TFM blocks for the two structures. The experimental results indicate that the number of TFM blocks has little effect on the structure of the pure encoder. For a structure that added decoder, the performance in detection of the structure with S =2 is 0.3% higher than the structure with S =1, while the performance in localization is basically in a stable state. When S >2, the performance in detection and localization remain stable or slightly decrease, indicating that the structure with S =2 is enough to fit the distribution manifold of normal data and detect anomalies. "},{"figure_ref":["fig_2"],"heading":"Computation of anomaly map","publication_ref":[],"table_ref":["tab_4","tab_4","tab_5","tab_5"],"text":"This part evaluates the influence of different function of anomaly score and fusion mode of anomaly map on experimental results. Table 6 shows the performance in detection and localization with a R34 backbone using different function of anomaly map, where α represents the weight of the anomaly map of our model, α mse and α cos denote the difference value in spatial distance and direction of anomaly map (see section 3.4). As can be seen from Table 6, the weight coefficient obtained according to the importance of the anomaly map outperforms the fixed coefficient by 0.9% in detection AUROC and 0.1% in localization AUROC. α mse pays more attention to distance differences and performs better in detecting subtle anomalies, while α cos focuses on direction differences and performs better in detecting diversity anomalies. Therefore, we use the harmonic mean values of α mse and α cos as the coefficient of the final anomaly score to detect anomalies with various sizes better.\nAs can be seen from the Figure 4, anomaly maps extracted from different semantic layers express different abstract information. The anomaly map extracted from the shallow layer (Layer 1) pays more attention to texture and edge extraction but has weak semantic information. The anomaly map extracted from the middle layer (Layer 2) is more carefully segmented than shallow layer but easily affected by background noise. While the anomaly map extracted from deep layer (Layer 3) focuses on semantic information but has low resolution, resulting in poor segmentation. We attempt to combine the anomaly maps extracted from different semantic layers to obtain accurate pixel-level localization of anomalies.\nTable 7 shows the performance in detection and localization with a R34 backbone using the different combinations of anomaly maps, where P1 to P5 represent different combination modes, the Arabic numerals represent anomalous maps generated by different critical layers. It can be observed that the performance of anomaly map fusion is 0.4% to 1.1% and 0.8% to 2.3% higher than that of single-layer anomaly map in detection and localization AUROC respectively, indicating that the effective fusion of anomaly map with different levels is conducive to detecting and locating anomalies with different sizes.\nAs can be noticed from Table 7 that the performance in detection of P3 is 0.3% higher than that of P2, and the performance in localization is 0.4% lower than that of P2. Qualitatively, different levels of anomaly maps detect anomalies with different sizes. P2 mode performs well with anomalies of texture categories because providing more edge information, while P3 mode works well with object categories because providing more semantic information. To balance the characteristics of different anomalies, we effectively combined P2 and P3 into P6, resulting in optimal performance."},{"figure_ref":[],"heading":"Conclusions","publication_ref":[],"table_ref":[],"text":"We propose a multiresolution feature guidance method based on Transformer named GTrans for unsupervised anomaly detection and localization. First, GTrans utilizes the difference in generalization ability between AGN and Trans to locate anomalous regions. Second, we utilize the TFM module modified by Transformer to enhance the information interaction ability of multi-scale features. Also, for anomaly maps, we propose a new generation function that jointly considers spatial and directional distances. Finally, we conduct a series of ablation studies to demonstrate the effectiveness of GTrans. Experimental results on MVTec AD dataset show that GTrans can achieve the state-of-theart performance in both detection and localization."},{"figure_ref":[],"heading":"Category","publication_ref":["b14","b24","b32","b20","b33","b19","b42"],"table_ref":[],"text":"AnoGAN [15] AE-SSIM [25] Patch-SVDD [33] Student [21] CutPaste [34] SPADE [20] InTra [43] GTrans "},{"figure_ref":[],"heading":"Results","publication_ref":[],"table_ref":[],"text":"In this section, we show the result of our method. To demonstrate the feasibility and effectivity of our experiment, we conduct anomaly detection and localization using the MVTec AD dataset with anomalies. We present the AUROC and AUPRO score for each category to give an intuition of the effect of GTrans on different categories. We compare our method to state-of-the-art results on the MVTec AD benchmark."},{"figure_ref":[],"heading":"Detection","publication_ref":[],"table_ref":[],"text":"Table 1 presents the AUROC results for anomaly detection on the MVTec AD dataset. We take the maximum value of anomaly map issued by GTrans (see Section 3.4) for the anomaly detection and report standard AUROC as a detection metric. Since the other baselines have different backbones, we try a R34 and a WR50 as the backbone of our model respectively. As shown in Table 1, GTrans-R34 outperforms the other methods by 1.1% to 43.6% in AUROC score on average for all the categories."}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b0\",\n \"b1\",\n \"b2\",\n \"b3\",\n \"b4\",\n \"b5\",\n \"b6\",\n \"b7\",\n \"b8\",\n \"b9\",\n \"b10\",\n \"b11\",\n \"b12\",\n \"b8\",\n \"b13\",\n \"b14\",\n \"b15\",\n \"b16\",\n \"b9\",\n \"b17\",\n \"b18\",\n \"b19\",\n \"b19\",\n \"b20\",\n \"b20\",\n \"b21\",\n \"b21\",\n \"b22\"\n ],\n \"table_ref\": [],\n \"text\": \"Anomaly detection aims to identify samples that do not match the expected pattern or dataset. Anomaly detection techniques have been extensively studied in a variety of research and application domains, including industrial inspection [1][2][3] and medical applications [4]. With the continuous development of the industrial field, customers need to place various sensors for continuous monitoring of equipment condition and detect anomalies. The traditional manual visual detection method has recently become unsatisfactory because it is susceptible to the influence of manual experience and subjective factors. Therefore, anomaly detection has gradually become valuable in computer vision, attracting high attention in different fields [5][6][7]. In general, class imbalance problem [8] and the unexpectedness of anomalies are the two most common challenges of anomaly detection tasks. Anomalies are extremely rare in the industrial scene, which implies the number of aberrant samples obtained is quite low, resulting in a serious class imbalance problem. In addition, anomalies are always unexpected. It is hard to predict their location and size or even determine if anomalies occur at all. As a result, modeling all anomalies or even predicting all abnormalities that never occur is impractical on a few samples.\\nThe existing anomaly detection tasks [9,10] focused on the classification of single or multiple categories at image-level. The current category to be detected is normal, and non-category are abnormal. However, in contrast to classification tasks, it is hard to train a model with full supervision for anomaly detection due to the lack of a large number of abnormal samples. [11,12] trained models in normal category, and samples were judged as anomalous when they show a large difference from the trained normal samples in the test stage. Recently, the anomaly detection tasks had confronted new challenges. To better monitor and process the anomalies, it is required not only to pick out the anomalous images but also to locate the anomalous regions. Bergmann et al. [13] proposed the MVTec AD dataset to provide benchmarks for anomaly detection and localization. The pixel-level methods [9,[14][15][16][17] exploited deep convolutional autoencoder and generative model such as Generative Adversarial Network (GAN) and Variational Autoencoder (VAE), respectively. [10,18,19] attempted to learn the feature representation of the normal samples from scratch and the model trained under the normal data manifold is hard to reconstruct the abnormal image. It may result in a large per-pixel reconstruction error and then a higher anomaly score. However, these methods showed a tremendous potential for misdetection due to low resolution of the reconstructed images and the strong generalization ability of the model.\\nFortunately, it was found that using a pre-trained network could be a potential mode to learn feature representation for small sample datasets. Cohen et al. [20] utilized a group of features extracted from a deep pre-trained network on anomaly-free images to train their model. [20] improved performance gain, but their model relied on many sub-images of training data, resulting in expensive computations. To evade these limitations, Bergmann et al. [21] proposed a student-teacher network to implicitly model the distribution of features extracted from normal images. The idea of the student-teacher network was that the student network had poor generalization ability to abnormal data manifolds and then made wrong judgments. For knowledge transfer, [21] only imitated the last layer in the teacher network for knowledge distillation without making full use of the information of intermediate layers. To fully exploit the intermediate features of the teacher network, [22] proposed a novel knowledge distillation method that distilled the comprehensive knowledge of the pre-trained network at several critical layers to the trained network. It can provide the significance of multi-scale information in computer vision algorithms. Although features were extracted in several critical layers in [22], the information interaction between multi-scale features is fragile.\\nLately, Transformer [23] had achieved great success in image classification and recognition, which proved the feasibility of Transformer in the information interaction of features with multiresolution. Inspired by this, we propose a multiresolution feature guidance method based on Transformer named GTrans for unsupervised anomaly detection and localization. In our method, the Anomaly Detected Network based on Transformer named Trans utilizes Transformer to effectively establish relationship between features with multiresolution. Under the tacit knowledge transfer of the Anomaly Guided Network (AGN), Trans enhances the fitting ability on the normal data manifold. Our intuition is that when abnormal images are input into the Trans, Trans cannot judge such accidents, resulting in large abnormalities in the whole image. Due to the strong ability of Trans in obtaining interaction information of features, the abnormal regions will receive great attention, conducive to the realization of pixel-level localization of the anomalies. Furthermore, we define a novel function to compute the anomaly score according to the difference of features extracted by AGN and Trans in the spatial distance and direction. The function can effectively capture the information of the combination of multi-scale features to obtain an accurate heat map of anomalous regions. We evaluate our method on MVTec AD dataset and then achieve superior performance in both localization and detection. Figure 1 shows visualized results of our method on the MVTec AD dataset. Our main contributions are summarized as follows:\\n1. We propose a multiresolution feature guided method based on Transformer, referred as GTrans for unsupervised1 anomaly detection and localization.\\nGTrans can improve the ability in information interaction between features with multiresolution. 2. We develop an anomaly guided network (AGN) to provide surrogate labels of features with multiresolution. Our model can be trained on normal images entirely without additional data augmentation. 3. We define a novel function to generate an anomaly map by exploiting the anomaly maps with multiresolution in spatial distance and direction.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experimental results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"show that our proposed GTrans achieves state-of-theart performance on MVTec AD dataset, which demonstrates the significance of information interaction for anomaly detection and localization.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Anomaly detection has attracted a lot of attention in the last decades. We provide an extensive overview of anomaly detection techniques next. The research related to our work can categorize into two classes. First, we present an overview of anomaly detection and segmentation methods, which can be categorized into reconstruction-based methods and embedding similarity-based methods. Second, we briefly explain Transformer architectures and show the application of Transformer in anomaly detection.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Anomaly Detection and Segmentation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Reconstruction-based methods\",\n \"publication_ref\": [\n \"b9\",\n \"b17\",\n \"b18\",\n \"b24\",\n \"b10\",\n \"b25\",\n \"b26\",\n \"b27\",\n \"b8\",\n \"b13\",\n \"b14\",\n \"b15\",\n \"b16\",\n \"b28\",\n \"b24\",\n \"b17\",\n \"b9\",\n \"b14\",\n \"b8\",\n \"b15\",\n \"b16\",\n \"b14\",\n \"b9\",\n \"b24\",\n \"b8\",\n \"b14\",\n \"b15\",\n \"b16\"\n ],\n \"table_ref\": [],\n \"text\": \"Reconstruction-based methods like autoencoders (AEs) [10,18,19,25], variational autoencoders (VAEs) [11,[26][27][28] or generative adversarial networks (GANs) [9,[14][15][16][17]29] attempted to model the sample without abnormality and defect from scratch. The idea of AE-based methods was that if the model were trained on a dataset containing only normal samples, the reconstructed image would approach the normal sample regardless of whether the input images are normal samples or abnormal samples. By comparing the pixel error of the original image and the reconstructed image, the model could judge whether the input image is abnormal or not and even locate the abnormal regions. To learn potential features better, SSIM-AE [25] used SSIM as loss function and anomaly measure to compare input and reconstructed images. ARNet [18] learned the semantic feature embeddings related to the erased attributes by forcing the network to restore the original image. MemAE [10] proposed an autoencoder with a memory module to explicitly suppress the generalization capability of the autoencoder.\\nVAE-based methods were also found to be used in anomaly detection and localization. In VAE-based methods, the probabilistic encoder and decoder both parameterized an isotropic normal distribution in the latent variable space and the original input variable space, respectively. The model endeavored to find the probability distribution conforming to the normal sample and used the reconstructed probability as the anomaly scores to generate an attention map in the potential space to detect the anomalies. However, VAE-based methods were not automatically superior to traditional autoencoder methods in general.\\nGAN-based approaches utilized the discriminator to detect the slight disturbance in the reconstruction of abnormal images, prompting the generator to extract sample information from the potential space to reconstruct the image. AnoGAN [15] proposed a novel anomaly scoring scheme based on the mapping from image space to a latent space. OCGAN [9] utilized a dual latent space learning process to constrain the latent space of generators and discriminators to represent the specific categories exclusively. GANomaly [16] and f-AnoGan [17] added an additional encoder to the generator to reduce the inference time of AnoGan [15].\\nMethods based on reconstruction were intuitive and explicative. However, these methods either failed to detect anomalies due to the strong generalization ability of deep models [10,25] or failed in one-class settings [9,[15][16][17].\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Embedding similarity-based methods\",\n \"publication_ref\": [\n \"b29\",\n \"b30\",\n \"b31\",\n \"b32\",\n \"b33\",\n \"b11\",\n \"b32\",\n \"b33\",\n \"b34\",\n \"b20\",\n \"b20\",\n \"b21\",\n \"b35\",\n \"b20\",\n \"b21\"\n ],\n \"table_ref\": [],\n \"text\": \"Embedding similarity-based methods mainly mapped the meaningful vector of the sample extracted from the pre-trained network to the high-dimensional feature space and judged the possibility of anomalies by computing the distance between the test sample and the normal sample in the feature space. Feature space expressed a higher level and more abstract information than image space. Embedding similarity-based methods could be categorized as cluster-based, embedding patch-based and knowledge-based methods.\\nCluster-based methods used K-Nearest Neighbor (KNN) [30], K-means [31] or Principal Component Analysis (PCA) [32] to match the best approximate features from a memory-bank of nominal features or similar feature sets extracted from a pre-trained network to locate anomalies. However, the inference speed in the test phase of cluster-based methods was linearly related to the size of feature sets, which was limited in practical application.\\nEmbedding patch-based methods tried to repair abnormal images by patch extracted from normal images. Patch SVDD [33] and CutPaste [34] applied self-supervised learning to anomaly detection. Unlike Deep SVDD [12], Patch SVDD [33] inspected the image at the patch level, and each patch corresponds to a point in the feature space. CutPaste [34] proposed a data augmentation strategy that cut an image patch and pasted it at a random location of an image. PaDiM [35] described each patch location with a Gaussian distribution and modeled the correlation between semantic layers. However, self-supervised representation typically underperformed those learned from large supervised datasets such as ImageNet.\\nKnowledge-based methods mainly utilized the difference in generalization ability between teacher and student networks in a teacher-student framework to locate anomalies. The teacher network had a good performance and strong generalization ability, while the student network showed strong representation ability only in a single class by imitating the behavior of the teacher network. Bergmann et al. [21] were the first to introduce a teacher-student framework in the field of unsupervised anomaly detection and localization. Anomaly scores of [21] were derived from the predictive variance and regression error of an ensemble of student networks. On this basis, [22] and [36] extended the multiscale critical layers of knowledge extraction on the VGG and ResNet networks respectively, and achieved good performance. However, such methods either only extracted the single-layer information without making full use of the information in intermediate layers [21] or treated critical layers with multiresolutions equally while ignoring the importance of information interaction of multi-scale critical layers [22][36].\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Transformer in Vision\",\n \"publication_ref\": [\n \"b22\",\n \"b36\",\n \"b22\",\n \"b37\",\n \"b38\",\n \"b40\",\n \"b37\",\n \"b40\",\n \"b41\",\n \"b40\",\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"Ref. [23] first proposed Transformer, a model relying entirely on an attention mechanism to describe global dependencies between input and output, in the field of Natural Language Processing (NLP). Transformer did not use Recurrent Neural Network (RNN) such as Long Short-Term Memory (LSTM) [37] and instead utilized a stack of multi-headed attention blocks to accomplish NLP tasks excellently. The specific structure of the Transformer can be found in the original article [23].\\nTransformer had achieved the most advanced performance in many NLP tasks and had become the preferred model for NLP tasks. Computer vision and NLP are merging as more efficient structures emerge. Due to the computational efficiency and scalability of Transformer, Transformer had been explored in computer vision and had become a new research direction.\\nRecently, Transformer had achieved good performance in image classification [38], object detection [39][40] and image segmentation [41]. In general, there were two main model architectures for the adoption of Transformer in computer vision. One is a pure Transformer structure, and the other is a hybrid structure combining convolution neural network (CNN) and Transformer. Moreover, Vision Transformer [38] built a pure Transformer structure, dividing the input image into square uniform patches with 16 × 16 patch size. For each patch, the linear transformation was performed for dimensionality reduction and location information was embedded. Then, the projection information and location information was input into the Transformer to achieve image classification. [41] proposed TransUNet, which merited both Transformers and U-Net [42], to achieve medical image segmentation. TransUNet [41] utilized Transformer for encoding tokenized image patches from a feature map extracted by CNN, and the decoder upsampled the encoded features combined with the high-resolution CNN feature maps to enable precise localization.\\nFor anomaly detection and localization, InTra [43] used a deep Transformer network consisting of a simple stack of multi-headed self-attention blocks to detect anomalies. Experiments showed that pure Transformer architecture is more efficient and scalable than traditional CNN in both model size and computational scale, while hybrid architecture performed better than pure Transformer at the smaller model size. Therefore, our method employ the hybrid structure combining CNN and Transformer to detect and locate anomalies and achieved excellent results.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Methodology\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We propose a multiresolution feature guidance method based on Transformer named GTrans for anomaly detection and localization. GTrans consists of four components: 1) AGN: Anomaly guided network. 2) Trans: Anomaly detected network based on Transformer. 3) Learning of the normality. 4) Computation of the anomaly map. The overall architecture of GTrans is shown in Figure 2. Given a testing dataset D test = {y 1 , y 2 , ..., y M } consisting both anomaly and normal images. Given an input image y k ∈ R W ×H×C , k ∈ [1, M ] of width H, height W, and the number of channels C, we extract groups of features with multiresolution\\nF G (y k ) = F 1 G (y k ) , F 2 G (y k ) , ..., F L G (y k ) and F T (y k ) = F 1 T (y k ) , F 2 T (y k ) , .\\n.., F L T (y k ) from AGN pre-trained on ImageNet and the backbone of Trans, respectively, where L represents total number of critical layers. F G (y k ) and F T (y k ) are input into Tokenizer to obtain groups of tokens\\nT G = T 1 G , T 2 G , ..., T L G and T T = T 1 T , T 2 T , .\\n.., T L T , respectively. Trans utilizes Transformer to establish a relationship between features with multiresolution. AGN guides the decoder module of TFM to enhance the fitting ability of Trans in a normal data manifold. The mapper module maps the token output by TFM into original critical layers. In training, the parameters of AGN are frozen and the critical layers extracted by AGN as regression targets of the Trans. In testing, GTrans locates anomalies by comparing the differences in spatial distance and direction of multi-scale features extracted from the AGN and the Trans.\\nIn the following, we describe the four components of GTrans in detail.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Anomaly Guided Network\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We propose a novel network called Anomaly Guided Network (AGN) for unsupervised anomaly detection and localization. AGN is a convolution network pre-trained on ImageNet (e.g., a ResNet-50-2 pre-trained on ImageNet), it can exhibit strong ability in feature representation when the sample quantity is small as ever. The guiding of AGN comes from two aspects. First, the token extracted from AGN as surrogate labels to be learned of Trans. Second, the feature maps extracted by AGN as regression targets of the Trans during the learning of the normality, enhancing the fitting ability of Trans in normal data manifolds.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Feature Extraction\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Given a training dataset D train = {x 1 , x 2 , ..., x N } consisting only of normal images. Given an input image\\nx k ∈ R W ×H×C , k ∈ [1, N ] of width H, height W,\\nand the number of channels C. Our goal is to create an ensemble of feature group\\nF G (x k ) = F 1 G (x k ) , F 2 G (x k ) , ..., F L G (x k\\n) to detect anomalies of various sizes, the l -th critical layer of AGN extracts a feature map\\nF l G (x k ) ∈ R w l ×h l ×c l for l ∈ [1, L]\\n, where h l and w l represent the height and width of the feature map, c l represents the number of channels.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Token Extraction\",\n \"publication_ref\": [\n \"b43\",\n \"b44\",\n \"b45\"\n ],\n \"table_ref\": [],\n \"text\": \"Inspired by [44][45][46], we utilize a module designated Tokenizer, containing point-wise convolutions and spatial attention, to extract visual tokens because a few visual tokens are sufficient to represent the entire image.\\nWe reshape the input feature map F l G (x k ) ∈ R w l ×h l ×c l into a sequence of flattened two-dimensional feature map F l GK (x k ) ∈ R (w l •h l )×c l before entering the Tokenizer. Tokenizer uses two point-wise convolutions L g and L d to process each pixel of two-dimensional feature map F l GK (x k ) as follows:\\nG l = L g F l GK (x k ) √ c l ∈ R (w l •h l )×g(1)\\nand\\nV l = L d F l GK (x k ) ∈ R (w l •h l )×d .(2)\\nHere, L g maps each pixel of feature map in l -th critical layer to g semantic groups G l , while L d realizes the information interaction between channels of the feature map by raising or reducing dimension. V l represents the highdimensional feature representation of l -th critical layers. For each semantic group, spatial attention is computed for the sequences G l as a weighted average over V l as follow:\\nT l = softmax(G l ) T V l ∈ R d×g ,(3)\\nwhere softmax(•) obtains the spatial attention weight of each semantic group.\\nIn general, we input F G (x k ) into the Tokenizer to obtain a group of visual tokens T G = T 1 G , T 2 G , ..., T L G with different resolutions. For anomaly detection, T G and F G (x k ) provide guidance in decoder modules and the learning of the normality, respectively.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Trans\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We propose a anomaly detected network based on Transformer named Trans. Trans improves the ability in information interaction between features with multiresolution. In the following, we describe the steps of Trans in detail.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Extraction of feature and token\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To effectively obtain the tacit knowledge of AGN, we rely on architecture that the dimension of the features of Trans is aligned with that of AGN. Trans and AGN have the same backbone, but Trans has not been pre-trained. In CNN, top layers produce low-resolution, semantically strong features, while bottom layers produce high-resolution, semantically weak ones. Given an input image x k ∈ R W ×H×C of width H, height W, and the number of channels C. To detect and locate anomalies with various sizes, we extract a group of features\\nF T (x k ) = F 1 T (x k ) , F 2 T (x k ) , .\\n.., F L T (x k ) at various layers. We input F T (x k ) into the Tokenizer to obtain a group of visual tokens T T = T 1 T , T 2 T , ..., T L T with multiresolutions.\\nFig. 3 Overview of the TFM architecture.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"TFM\",\n \"publication_ref\": [\n \"b46\",\n \"b37\",\n \"b40\"\n ],\n \"table_ref\": [],\n \"text\": \"This module aims to establish information interaction of spatially distant concepts for visual tokens of different semantic groups at different levels. We employ the encoder-decoder structure of standard Transformer with the following changes: (1) Omitting position embedding. [47] provided convincing evidence that CNNs do indeed rely on and learn information about spatial positioning. Therefore, the convolution layer computes learnable weights of query q e ∈ R d×(g•l) , key k e ∈ R d×(g•l) , and value v e ∈ R d×(g•l) without additional position embedding. (2) Using a non-linearity activation function and two point-wise convolutions replace position-wise feed-forward networks of the standard Transformer. (3) Extracting a group of visual tokens\\nT G = T 1 G , , T 2 G , .\\n.., T L G from AGN as the input of the decoder in Transformer. We name the module modified by standard Transformer as TFM. Figure 3 illustrates the architecture of the TFM. As can be seen, TFM is composed of a stack of S encoders and decoders. Each encoder and decoder has three sub-layers, respectively.\\nThe first is a self-attention mechanism, the second is a combination of a non-linearity activation function and two point-wise convolutions, and the third is a layer normalization. We take an individual TFM block of S = 1 as an example below. We concatenate the group of visual tokens T G and T T into E in ∈ R d×(g•l) and D in ∈ R d×(g•l) , respectively. We take E in and D in as the input of the encoder and decoder in TFM.\\nIn the encoder, we compute q e , k e , v e via\\nq e = W q E in , k e = W k E in , v e = W v E in ∈ R d×(g•l)(4)\\nwith learnable weight matrices W q , W k , W v ∈ R d×d . Then, self-attention is computed by a compatibility function of the q e with corresponding k e as a weighted average of the v e as follows:\\nA e = v e softmax k T e q e √ d ∈ R d×(g•l) ,(5)\\nwhere A e is the output of self-attention mechanism in encoder. We employ a residual connection around the self-attention mechanism and layer normalization to get the output of layer normalization E a :\\nE a = E in + LayerN orm (A e ) ∈ R d×(g•l) ,(6)\\nwhere LayerN orm (•) represents the layer normalization. Then, we use a residual connection around the convolution layer, followed by layer normalization to get the output of encoder E out :\\nE out = LayerN orm (E a + L 1 σ (L 2 E a )) ∈ R d×(g•l) ,(7)\\nwhere L 1 , L 2 ∈ R d×d are point-wise convolutions, σ (•) is the ReLU function.\\nFor the decoder, most methods [38,41] previously applied to computer vision tasks use only the encoder of the Transformer. However, if the selfattention mechanism is only established in the visual token extracted from Tokenizer in Trans, the final model will easily fall into the problem of local optimal. Therefore, we employ an encoder-decoder structure of the Transformer and utilize visual tokens extracted from AGN as surrogate labels for the guidance of E out to enhance the stability of the Trans. We set\\nq d = W q D in , k d = W k E out , v d = W v E out ∈ R d×(g•l) ,(8)\\nA d = v d softmax k T d q d √ d ∈ R d×(g•l) .(9)\\nHere, q d is the query computed from D in . k d , v d are the key and value computed from the output of the encoder E out . A d is the output of self-attention mechanism in decoder.\\nSimilarly to the encoder, first, we employ a residual connection around the self-attention mechanism and layer normalization to get the output of layer normalization D a in the decoder. Second, we employ another residual connection around the convolution layer, followed by layer normalization to get the output of decoder D out . The computational process is as follows:\\nD a = D in + LayerN orm (A d ) ∈ R d×(g•l)(10)\\nand\\nD out = LayerN orm (D a + L 1 σ (L 2 D a )) ∈ R d×(g•l) .(11)\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Mapping of feature and token\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To achieve pixel-level alignment between the outputs of AGN and the Trans, we divide D out into g groups and map each group into the latent space of the same dimension as the\\nF G (x k ) = F 1 G (x k ) , F 2 G (x k ) , ..., F L G (x k )\\n, respectively. For the l -th critical layer, we set\\nF l M (x k ) = F l T (x k ) + T l v softmax X l q T T l k √ c l T ,(12)\\nwhere F l M (x k ) is the final output of the l -th critical layer, X l q is the query computed from F l G (x k ). T l k , T l v are the key and value computed from the output of the encoder in TFM E out .\\nIn general, the final output of Trans is a group of critical layers\\nF M (x k )= F 1 M (x k ), F 2 M (x k ), ..., F L M (x k ) .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Learning of the normality\",\n \"publication_ref\": [\n \"b47\"\n ],\n \"table_ref\": [],\n \"text\": \"In the following, we train the Trans so that its output is as similar as possible to the AGN, and the parameters of AGN are frozen during the entire training phase. Given an input image x k ∈ D train , we obtain a group of feature maps\\nF G (x k )= F 1 G (x k ), F 2 G (x k ), ..., F L G (x k ) and F M (x k )= F 1 M (x k ), F 2 M (x k ), ..., F L M (x k\\n) extracted from AGN and Trans, respectively. As indicated in [48], Trans can learn the distilled knowledge from AGN that is trained at a different task and guided by intermediate-level hints from AGN. Trans can learn the intermediate-level hints from different semantic layers of AGN. It then takes the intermediate-level hints as targets of optimization procedure to intensify the complete knowledge transfer from AGN to Trans. We define pixel-wise L 2 loss function p l (i,j) at position (i, j)\\nfor l ∈ [1, L], i ∈ [1, w l ], j ∈ [1, h l ] as follow: p l (i,j) (x k ) = 1 2 F l G (x k ) (i,j) -F l M (x k ) (i,j) 2 2 . (13\\n)\\nHere, F l G (x k ) and F l M (x k ) represent the feature map extracted from l -th critical layers of AGN and Trans. We get the total loss L (x k ) of input image x k by the weighted average of all pixels in each feature map of critical layers in sets F l G (x k ) and F l M (x k ) as follow:\\nL (x k ) = L l=1 1 w l • h l w l i=1 h l j=1 p l (i,j) (x k ), (14\\n)\\nwhere L represents total number of critical layers.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Computation of the anomaly map\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In training phase, Trans has stronger fitting ability in anomaly-free images than the AGN under the guidance of AGN. However, AGN is knowledgeable on anomaly images because of the strong generalization2 , while Trans is unfamiliar with such images. Therefore, the abnormal regions of anomaly images show a large deviation from the training data manifold of Trans when the anomaly images are input into AGN and Trans, respectively. Given a testing dataset D test = {y 1 , y 2 , ..., y M } consisting both anomaly and normal images, we assign an anomaly map M (y k ) ∈ R H×W to a test image\\ny k ∈ D test for k ∈ [1, M ],\\nwhere H and W are the height and width of the test image. The anomaly score of pixel at position (i, j)\\nof i ∈ [1, H], j ∈ [1, W ]\\nindicates the deviation degree of pixel at position (i, j) from the data manifold of anomalyfree images. We utilize the anomaly map M (y k ) to realize anomaly detection and localization. Next, we show the specific computation of anomaly map. We input a test image y k into our model to obtain the group of feature maps\\nF G (y k )= F 1 G (y k ), F 2 G (y k ), ..., F L G (y k ) and F M (y k )= F 1 M (y k ), F 2 M (y k ), ..., F L M (y k )\\nextracted from different critical layers of AGN and Trans, respectively.\\nWe get a group of loss P = P 1 , P 2 , ..., P L by computing pixel-wise L 2 loss between the feature map from the same layers at F G (y k ) and F M (y k ) as follows:\\nP l = 1 w l • h l c l i=1 1 2 F l G (y k ) -F l M (y k ) 2 2 ∈ R w l ×h l , (15\\n)\\nwhere P l ∈ P denotes the pixel-wise loss of l -th critical layer between AGN and Trans. We assign a weight to different critical layers because the different critical layers correspond to anomalies at various sizes, respectively. Here, we utilize MSE loss and cosine similarity metric of feature map in l -th critical layer to define two coefficients α l mse and α l cos as follows:\\nα l mse = 1 w l • h l w l i=1 h l j=1 F l G (y k ) (i,j) -F l M (y k ) (i,j) 2(16)\\nand\\nα l cos = 1 - vec F l G (y k ) T • vec F l M (y k ) vec F l G (y k ) vec F l M (y k ) . (17\\n)\\nHere, vec (•) is a vectorization function transforming a matrix with arbitrary dimensions into a 1-D vector.\\nα l mse and α l cos denote the similarity in spatial distance and direction of lth feature maps from AGN and Trans. The larger the values of α l mse and α l cos are, the less similarity the l -th feature maps are in spatial distance and direction, and the greater the possibility of anomalies. We effectively combine the similarity of feature maps in spatial distance and direction, and use harmonic mean of α l mse and α l cos as the weight of l -th feature maps. α l is formulated as\\nα l = λ • α l cos • α l mse α l mse + λ • α l cos , (18\\n)\\nwhere λ is set to make the scale of α l mse and α l cos the same. Our goal is to assign high weight to feature maps with low similarity, so as to get high anomaly score with anomalous regions eventually. Therefore, the anomaly map is achieved by\\nM (y k ) = G σ L l=1 α l R P l ∈ R H×W . (19\\n)\\nHere, R (•) resizes the elements of P = P 1 , P 2 , ..., P L to the spatial size of (H × W ). G σ (•) represents a Gaussian filter with standard deviation of σ. Finally, we define the region with high anomaly score in M (y k ) as anomalous regions and the maximum value of M (y k ) as the final anomaly score of the test image y k for anomaly detection.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we first elaborate on the details of GTrans structure and its parameters. Second, to demonstrate the effectiveness of our approach, we compare our method to state-of-the-art results on the MVTec AD benchmark, considering in image-level and pixel-level, respectively.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Datesets and metrics\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Datesets\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We evaluate GTrans on the MVTec AD dataset 3 which contains over 5000 high-resolution images divided into fifteen different object and texture categories. Each category of the MVTec AD dataset comprises a set of anomaly-free training images and a test set of images with various kinds of anomalies as well as images without anomaly.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Metrics\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We use the Area Under the Receiver Operating Characteristic curve (AUROC) and Area Under the Per-Region-Overlap Curve (AUPRO) to evaluate the performence of GTrans. AUROC is computed on different levels of threshold in favor of large anomalous regions. AUPRO computes a threshold-independent 3 The datasets analysed during the current study are available at https://www.mvtec.com/company/research/datasets/mvtec-ad. evaluation metric based on the Per-Region-Overlap (PRO) to better account for large and small anomalies in localization. Only the standard AUROC provides image-level anomaly detection.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Implementation Details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We resize all images in the MVTec AD dataset to 256 × 256 and center crop them to 224 × 224. We take the preprocessed images of MVTec as the input of GTrans. We divided eighty percent of the MVTec AD dataset into the training set and twenty percent into the validation set without additional data enhancement. In the training phase, GTrans is trained only on the normal images with a batch size of 32 for 300 epochs. We utilize Adam optimizer with initial learning rate 10 -3 and weight decay 10 -4 for optimization. To enhance the stability of the GTrans, we define a large initial learning rate to improve the convergence speed of GTrans, and then the learning rate decreases with the increase of the number of iterations gradually. We utilize exponential decay equation to realize the decay of learning rate as\\nlr = lr init × rate step totalstep ,(20)\\nwhere lr init and lr represent the initial and current learning rate respectively, rate is decay factor with a value of 0.9, step and totalstep denote the current and total iteration number respectively. We implement our method in PyTorch and conduct all experiments on a machine equipped with an Intel i9-9900X and an NVIDIA GeForce RTX 2080 Ti GPU.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Architecture Details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"As shown in Figure 2, GTrans consists of AGN and Trans. We illustrate the parameter selection of GTrans below.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"AGN\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We extract features with a ResNet-34 (R34) and a Wide ResNet-50-2 (WR50). All backbones of AGN pre-trained on ImageNet, and the parameters of AGN are frozen during the training phase. As to the position of the selected guide feature, we choose features extracted by three intermediate layer groups, that is l = 3.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Trans\",\n \"publication_ref\": [\n \"b49\",\n \"b14\",\n \"b24\",\n \"b15\",\n \"b19\",\n \"b32\",\n \"b33\",\n \"b34\",\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"Trans is composed of a backbone network, a Tokenizer, a TFM and a mapper. We keep the backbone of Trans consistent with AGN. For Tokenizer, inspired by [50], we use a few visual tokens to generalise semantic concepts. Hence, we set the number of semantic groups g = 8 and dimension d = 256. For TFM, we utilize as few as two decoders and encoders to achieve superior performance. After mapper, the Trans and AGN output three feature maps with the size 56 × 56, 28 × 28 and 14 × 14, respectively.\\nTable 1 The AUROC results for anomaly detection on the MVTec AD dataset using AUROC%.\\nCategory AnoGAN [15] AE-SSIM [25] GANomaly [16] SPADE [20] Patch-SVDD [33] CutPaste [34] PaDiM [35] InTra [43] \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Localization\",\n \"publication_ref\": [\n \"b42\",\n \"b42\",\n \"b19\",\n \"b42\",\n \"b19\",\n \"b20\"\n ],\n \"table_ref\": [\n \"tab_1\",\n \"tab_1\"\n ],\n \"text\": \"Table 2 shows the AUROC and the PRO-score results for anomaly localization on the MVTec AD dataset. As shown in Table 2, GTrans-WR50 outperforms all the other methods in both the AUROC and the PRO-score on average for all the categories. Compared with the method based on Transformer, the performance of our method is 1.2% higher than InTra [43]. For PRO-score, InTra [43] and SPADE [20] achieve the best performance in the texture and object categories respectively. However, our method performs well in both texture and object categories, outperforming InTra [43] by 0.9% and SPADE [20] 1.3% in PRO-score on average for all the categories. When we further analyze the performance of GTrans-WR50, we find that GTrans-WR50 outperforms Student [21] by 7.3% in PRO-score, indicating the effectiveness of our method in multi-scale feature fusion and information interaction.\\n5 Ablation Studies\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Intermediate Knowledge\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\",\n \"tab_2\"\n ],\n \"text\": \"In this part, we evaluate the impact of the position of selected guide features in GTrans. In Table 3, we show the performance of anomaly detection and localization on the MVTec AD dataset of GTrans with a R34 backbone when using different combination critical layers (Layer 1+2+3+4, Layer 1+2+3, Layer 2+3+4, Layer 3+4). The final anomalous maps of all experiments in this part simply add the anomalous maps of different layers. It is known that feature layers of CNN can express various levels of abstract information. Bottom layers tend to extract low-level information such as textures, while top layers pay attention to low-resolution features that contain semantic information. It can be observed from Table 3 that Layer 1+2+3 can achieve the best performance in both texture and object categories. From the comparison of Layer 1+2+3+4 and Layer 1+2+3, it can be seen that the performance of anomaly detection and localization declines when the anomalous map generated by the fourth critical layer is added. The reason is that the features of the top layers are low-resolution, leading to the rough segmentation in detail texture. According to the experimental results of Layer 2+3, the information of the intermediate layers can express and extract anomalies to a large extent. When the information of the shallow layer is added on this basis, the model can extract the anomalies and pay attention to the processing of edge texture, improving the performance of the model.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"The structure of TFM\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_3\"\n ],\n \"text\": \"This part evaluates the impact of an added decoder and the number of TFM blocks on experimental results.\\nIn Table 4, we show the anomaly detection and localization performance on the MVTec AD dataset of GTrans with an R34 backbone. It can be seen that the detection AUROC of the model with TFM is 2.7% higher than that without TFM, while for localization AUROC, it can achieve 1.9% higher than the latter model.\\nIn Table 5, we show the performance in anomaly detection and localization on the MVTec AD dataset of GTrans with a R34 backbone when using the different structure of TFM in Trans, where S represents the number of TFM blocks. For structure which added decoder, we decode the output of the encoder through the features extracted from the AGN.\\nFrom Table 5, we can notice that for the same number of TFM blocks, the structure which added decoder outperforms the structure of pure encoder on MVTec AD dataset by 0.2% to 0.4% in the detection AUROC and 0.2% in localization AUROC. On this basis, we conduct experiments with the different number of TFM blocks for the two structures. The experimental results indicate that the number of TFM blocks has little effect on the structure of the pure encoder. For a structure that added decoder, the performance in detection of the structure with S =2 is 0.3% higher than the structure with S =1, while the performance in localization is basically in a stable state. When S >2, the performance in detection and localization remain stable or slightly decrease, indicating that the structure with S =2 is enough to fit the distribution manifold of normal data and detect anomalies. \"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Computation of anomaly map\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\",\n \"tab_4\",\n \"tab_5\",\n \"tab_5\"\n ],\n \"text\": \"This part evaluates the influence of different function of anomaly score and fusion mode of anomaly map on experimental results. Table 6 shows the performance in detection and localization with a R34 backbone using different function of anomaly map, where α represents the weight of the anomaly map of our model, α mse and α cos denote the difference value in spatial distance and direction of anomaly map (see section 3.4). As can be seen from Table 6, the weight coefficient obtained according to the importance of the anomaly map outperforms the fixed coefficient by 0.9% in detection AUROC and 0.1% in localization AUROC. α mse pays more attention to distance differences and performs better in detecting subtle anomalies, while α cos focuses on direction differences and performs better in detecting diversity anomalies. Therefore, we use the harmonic mean values of α mse and α cos as the coefficient of the final anomaly score to detect anomalies with various sizes better.\\nAs can be seen from the Figure 4, anomaly maps extracted from different semantic layers express different abstract information. The anomaly map extracted from the shallow layer (Layer 1) pays more attention to texture and edge extraction but has weak semantic information. The anomaly map extracted from the middle layer (Layer 2) is more carefully segmented than shallow layer but easily affected by background noise. While the anomaly map extracted from deep layer (Layer 3) focuses on semantic information but has low resolution, resulting in poor segmentation. We attempt to combine the anomaly maps extracted from different semantic layers to obtain accurate pixel-level localization of anomalies.\\nTable 7 shows the performance in detection and localization with a R34 backbone using the different combinations of anomaly maps, where P1 to P5 represent different combination modes, the Arabic numerals represent anomalous maps generated by different critical layers. It can be observed that the performance of anomaly map fusion is 0.4% to 1.1% and 0.8% to 2.3% higher than that of single-layer anomaly map in detection and localization AUROC respectively, indicating that the effective fusion of anomaly map with different levels is conducive to detecting and locating anomalies with different sizes.\\nAs can be noticed from Table 7 that the performance in detection of P3 is 0.3% higher than that of P2, and the performance in localization is 0.4% lower than that of P2. Qualitatively, different levels of anomaly maps detect anomalies with different sizes. P2 mode performs well with anomalies of texture categories because providing more edge information, while P3 mode works well with object categories because providing more semantic information. To balance the characteristics of different anomalies, we effectively combined P2 and P3 into P6, resulting in optimal performance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We propose a multiresolution feature guidance method based on Transformer named GTrans for unsupervised anomaly detection and localization. First, GTrans utilizes the difference in generalization ability between AGN and Trans to locate anomalous regions. Second, we utilize the TFM module modified by Transformer to enhance the information interaction ability of multi-scale features. Also, for anomaly maps, we propose a new generation function that jointly considers spatial and directional distances. Finally, we conduct a series of ablation studies to demonstrate the effectiveness of GTrans. Experimental results on MVTec AD dataset show that GTrans can achieve the state-of-theart performance in both detection and localization.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Category\",\n \"publication_ref\": [\n \"b14\",\n \"b24\",\n \"b32\",\n \"b20\",\n \"b33\",\n \"b19\",\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"AnoGAN [15] AE-SSIM [25] Patch-SVDD [33] Student [21] CutPaste [34] SPADE [20] InTra [43] GTrans \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we show the result of our method. To demonstrate the feasibility and effectivity of our experiment, we conduct anomaly detection and localization using the MVTec AD dataset with anomalies. We present the AUROC and AUPRO score for each category to give an intuition of the effect of GTrans on different categories. We compare our method to state-of-the-art results on the MVTec AD benchmark.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Detection\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Table 1 presents the AUROC results for anomaly detection on the MVTec AD dataset. We take the maximum value of anomaly map issued by GTrans (see Section 3.4) for the anomaly detection and report standard AUROC as a detection metric. Since the other baselines have different backbones, we try a R34 and a WR50 as the backbone of our model respectively. As shown in Table 1, GTrans-R34 outperforms the other methods by 1.1% to 43.6% in AUROC score on average for all the categories.\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-05-24"},"doi":{"kind":"string","value":""},"references":{"kind":"list like","value":[{"authors":"J Liu; F Guo; Y Zhang; B Hou; H Zhou","journal":"Applied Intelligence","ref_id":"b0","title":"Defect classification on limited labeled samples with multiscale feature fusion and semi-supervised learning","year":"2021"},{"authors":"J Wu; J Le; Z Xiao; F Zhang; L Geng; Y Liu; W Wang","journal":"Applied Intelligence","ref_id":"b1","title":"Automatic fabric defect detection using a wide-and-light network","year":"2021"},{"authors":"L He; N Shi; K Malik; F Li","journal":"Applied Intelligence","ref_id":"b2","title":"Unsupervised defect inspection algorithm based on cascaded gan with edge repair feature fusion","year":"2021"},{"authors":"K Zhou; Y Xiao; J Yang; J Cheng; W Liu; W Luo; Z Gu; J Liu; S Gao","journal":"Springer","ref_id":"b3","title":"Encoding structure-texture relation with p-net for anomaly detection in retinal images","year":"2020"},{"authors":"X Zhao; J Su; J Cai; H Yang; T Xi","journal":"Applied Intelligence","ref_id":"b4","title":"Vehicle anomalous trajectory detection algorithm based on road network partition","year":"2021"},{"authors":"W Zhou; Y Liang","journal":"Applied Intelligence","ref_id":"b5","title":"An immune optimization based deterministic dendritic cell algorithm","year":"2021"},{"authors":"X Zheng; J Chen; H Wang; S Zheng; Y Kong","journal":"Applied intelligence","ref_id":"b6","title":"A deep learning-based approach for the automated surface inspection of copper clad laminate images","year":"2021"},{"authors":"Q Qian; L Chen; H Li; R Jin","journal":"","ref_id":"b7","title":"Dr loss: Improving object detection by distributional ranking","year":"2020"},{"authors":"P Perera; R Nallapati; B Xiang","journal":"","ref_id":"b8","title":"Ocgan: One-class novelty detection using gans with constrained latent representations","year":"2019"},{"authors":"D Gong; L Liu; V Le; B Saha; M R Mansour; S Venkatesh; A V Hengel","journal":"","ref_id":"b9","title":"Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection","year":"2019"},{"authors":"J An; S Cho","journal":"Special Lecture on IE","ref_id":"b10","title":"Variational autoencoder based anomaly detection using reconstruction probability","year":"2015"},{"authors":"L Ruff; R Vandermeulen; N Goernitz; L Deecke; S A Siddiqui; A Binder; E Müller; M Kloft","journal":"PMLR","ref_id":"b11","title":"Deep one-class classification","year":"2018"},{"authors":"P Bergmann; M Fauser; D Sattlegger; C Steger","journal":"","ref_id":"b12","title":"Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection","year":"2019"},{"authors":"S Pidhorskyi; R Almohsen; D A Adjeroh; G Doretto","journal":"","ref_id":"b13","title":"Generative probabilistic novelty detection with adversarial autoencoders","year":"2018"},{"authors":"T Schlegl; P Seeböck; S M Waldstein; U Schmidt-Erfurth; G Langs","journal":"Springer","ref_id":"b14","title":"Unsupervised anomaly detection with generative adversarial networks to guide marker discovery","year":"2017"},{"authors":"S Akcay; A Atapour-Abarghouei; T P Breckon","journal":"Springer","ref_id":"b15","title":"Ganomaly: Semisupervised anomaly detection via adversarial training","year":"2018"},{"authors":"T Schlegl; P Seeböck; S M Waldstein; G Langs; U Schmidt-Erfurth","journal":"Medical image analysis","ref_id":"b16","title":"f-anogan: Fast unsupervised anomaly detection with generative adversarial networks","year":"2019"},{"authors":"Y Fei; C Huang; C Jinkun; M Li; Y Zhang; C Lu","journal":"IEEE Transactions on Multimedia","ref_id":"b17","title":"Attribute restoration framework for anomaly detection","year":"2020"},{"authors":"D Abati; A Porrello; S Calderara; R Cucchiara","journal":"","ref_id":"b18","title":"Latent space autoregression for novelty detection","year":"2019"},{"authors":"N Cohen; Y Hoshen","journal":"","ref_id":"b19","title":"Sub-image anomaly detection with deep pyramid correspondences","year":"2020"},{"authors":"P Bergmann; M Fauser; D Sattlegger; C Steger","journal":"","ref_id":"b20","title":"Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings","year":"2020"},{"authors":"M Salehi; N Sadjadi; S Baselizadeh; M H Rohban; H R Rabiee","journal":"","ref_id":"b21","title":"Multiresolution knowledge distillation for anomaly detection","year":"2021"},{"authors":"A Vaswani; N Shazeer; N Parmar; J Uszkoreit; L Jones; A N Gomez; L Kaiser; I Polosukhin","journal":"","ref_id":"b22","title":"Attention is all you need","year":"2017"},{"authors":"D Gudovskiy; S Ishizaka; K Kozuka","journal":"","ref_id":"b23","title":"Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows","year":"2022"},{"authors":"P Bergmann; S Löwe; M Fauser; D Sattlegger; C Steger","journal":"","ref_id":"b24","title":"Improving unsupervised defect segmentation by applying structural similarity to autoencoders","year":"2018"},{"authors":"K Sato; K Hama; T Matsubara; K Uehara","journal":"IEEE","ref_id":"b25","title":"Predictable uncertaintyaware unsupervised deep anomaly segmentation","year":"2019"},{"authors":"W Liu; R Li; M Zheng; S Karanam; Z Wu; B Bhanu; R J Radke; O Camps","journal":"","ref_id":"b26","title":"Towards visually explaining variational autoencoders","year":"2020"},{"authors":"Y Lu; P Xu","journal":"","ref_id":"b27","title":"Anomaly detection for skin disease images using variational autoencoder","year":"2018"},{"authors":"M Sabokrou; M Khalooei; M Fathy; E Adeli","journal":"","ref_id":"b28","title":"Adversarially learned one-class classifier for novelty detection","year":"2018"},{"authors":"E Eskin; A Arnold; M Prerau; L Portnoy; S Stolfo","journal":"Springer","ref_id":"b29","title":"A geometric framework for unsupervised anomaly detection","year":"2002"},{"authors":"J A Hartigan; M A Wong","journal":"Journal of the royal statistical society. series c (applied statistics)","ref_id":"b30","title":"Algorithm as 136: A k-means clustering algorithm","year":"1979"},{"authors":"H Y Teh; I Kevin; K Wang; A W Kempa-Liehr","journal":"IEEE Sensors Journal","ref_id":"b31","title":"Expect the unexpected: Unsupervised feature selection for automated sensor anomaly detection","year":"2021"},{"authors":"J Yi; S Yoon","journal":"","ref_id":"b32","title":"Patch svdd: Patch-level svdd for anomaly detection and segmentation","year":"2020"},{"authors":"C.-L Li; K Sohn; J Yoon; T Pfister","journal":"","ref_id":"b33","title":"Cutpaste: Self-supervised learning for anomaly detection and localization","year":"2021"},{"authors":"T Defard; A Setkov; A Loesch; R Audigier","journal":"Springer","ref_id":"b34","title":"Padim: A patch distribution modeling framework for anomaly detection and localization","year":"2021"},{"authors":"G Wang; S Han; E Ding; D Huang","journal":"","ref_id":"b35","title":"Student-teacher feature pyramid matching for unsupervised anomaly detection","year":"2021"},{"authors":"S Hochreiter; J Schmidhuber","journal":"Neural computation","ref_id":"b36","title":"Long short-term memory","year":"1997"},{"authors":"A Dosovitskiy; L Beyer; A Kolesnikov; D Weissenborn; X Zhai; T Unterthiner; M Dehghani; M Minderer; G Heigold; S Gelly","journal":"","ref_id":"b37","title":"An image is worth 16x16 words: Transformers for image recognition at scale","year":"2020"},{"authors":"N Carion; F Massa; G Synnaeve; N Usunier; A Kirillov; S Zagoruyko","journal":"Springer","ref_id":"b38","title":"End-to-end object detection with transformers","year":"2020"},{"authors":"X Zhu; W Su; L Lu; B Li; X Wang; J Dai","journal":"","ref_id":"b39","title":"Deformable detr: Deformable transformers for end-to-end object detection","year":"2020"},{"authors":"J Chen; Y Lu; Q Yu; X Luo; E Adeli; Y Wang; L Lu; A L Yuille; Y Zhou","journal":"","ref_id":"b40","title":"Transunet: Transformers make strong encoders for medical image segmentation","year":"2021"},{"authors":"O Ronneberger; P Fischer; T Brox","journal":"Springer","ref_id":"b41","title":"U-net: Convolutional networks for biomedical image segmentation","year":"2015"},{"authors":"J Pirnay; K Chai","journal":"","ref_id":"b42","title":"Inpainting transformer for anomaly detection","year":"2021"},{"authors":"S Zhang; X He; S Yan","journal":"PMLR","ref_id":"b43","title":"Latentgnn: Learning efficient non-local relations for visual recognition","year":"2019"},{"authors":"X Liang; Z Hu; H Zhang; L Lin; E P Xing","journal":"Advances in Neural Information Processing Systems","ref_id":"b44","title":"Symbolic graph reasoning meets convolutions","year":"2018"},{"authors":"Y Chen; M Rohrbach; Z Yan; Y Shuicheng; J Feng; Y Kalantidis","journal":"","ref_id":"b45","title":"Graph-based global reasoning networks","year":"2019"},{"authors":"M A Islam; S Jia; N D Bruce","journal":"","ref_id":"b46","title":"How much position information do convolutional neural networks encode?","year":"2020"},{"authors":"J Yim; D Joo; J Bae; J Kim","journal":"","ref_id":"b47","title":"A gift from knowledge distillation: Fast optimization, network minimization and transfer learning","year":"2017"},{"authors":"S Kornblith; J Shlens; Q V Le","journal":"","ref_id":"b48","title":"Do better imagenet models transfer better?","year":"2019"},{"authors":"B Wu; C Xu; X Dai; A Wan; P Zhang; Z Yan; M Tomizuka; J Gonzalez; K Keutzer; P Vajda","journal":"","ref_id":"b49","title":"Visual transformers: Token-based image representation and processing for computer vision","year":"2020"}],"string":"[\n {\n \"authors\": \"J Liu; F Guo; Y Zhang; B Hou; H Zhou\",\n \"journal\": \"Applied Intelligence\",\n \"ref_id\": \"b0\",\n \"title\": \"Defect classification on limited labeled samples with multiscale feature fusion and semi-supervised learning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"J Wu; J Le; Z Xiao; F Zhang; L Geng; Y Liu; W Wang\",\n \"journal\": \"Applied Intelligence\",\n \"ref_id\": \"b1\",\n \"title\": \"Automatic fabric defect detection using a wide-and-light network\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"L He; N Shi; K Malik; F Li\",\n \"journal\": \"Applied Intelligence\",\n \"ref_id\": \"b2\",\n \"title\": \"Unsupervised defect inspection algorithm based on cascaded gan with edge repair feature fusion\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"K Zhou; Y Xiao; J Yang; J Cheng; W Liu; W Luo; Z Gu; J Liu; S Gao\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b3\",\n \"title\": \"Encoding structure-texture relation with p-net for anomaly detection in retinal images\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"X Zhao; J Su; J Cai; H Yang; T Xi\",\n \"journal\": \"Applied Intelligence\",\n \"ref_id\": \"b4\",\n \"title\": \"Vehicle anomalous trajectory detection algorithm based on road network partition\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"W Zhou; Y Liang\",\n \"journal\": \"Applied Intelligence\",\n \"ref_id\": \"b5\",\n \"title\": \"An immune optimization based deterministic dendritic cell algorithm\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"X Zheng; J Chen; H Wang; S Zheng; Y Kong\",\n \"journal\": \"Applied intelligence\",\n \"ref_id\": \"b6\",\n \"title\": \"A deep learning-based approach for the automated surface inspection of copper clad laminate images\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Q Qian; L Chen; H Li; R Jin\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Dr loss: Improving object detection by distributional ranking\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"P Perera; R Nallapati; B Xiang\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Ocgan: One-class novelty detection using gans with constrained latent representations\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"D Gong; L Liu; V Le; B Saha; M R Mansour; S Venkatesh; A V Hengel\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"J An; S Cho\",\n \"journal\": \"Special Lecture on IE\",\n \"ref_id\": \"b10\",\n \"title\": \"Variational autoencoder based anomaly detection using reconstruction probability\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"L Ruff; R Vandermeulen; N Goernitz; L Deecke; S A Siddiqui; A Binder; E Müller; M Kloft\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b11\",\n \"title\": \"Deep one-class classification\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"P Bergmann; M Fauser; D Sattlegger; C Steger\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"S Pidhorskyi; R Almohsen; D A Adjeroh; G Doretto\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Generative probabilistic novelty detection with adversarial autoencoders\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"T Schlegl; P Seeböck; S M Waldstein; U Schmidt-Erfurth; G Langs\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b14\",\n \"title\": \"Unsupervised anomaly detection with generative adversarial networks to guide marker discovery\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"S Akcay; A Atapour-Abarghouei; T P Breckon\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b15\",\n \"title\": \"Ganomaly: Semisupervised anomaly detection via adversarial training\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"T Schlegl; P Seeböck; S M Waldstein; G Langs; U Schmidt-Erfurth\",\n \"journal\": \"Medical image analysis\",\n \"ref_id\": \"b16\",\n \"title\": \"f-anogan: Fast unsupervised anomaly detection with generative adversarial networks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Y Fei; C Huang; C Jinkun; M Li; Y Zhang; C Lu\",\n \"journal\": \"IEEE Transactions on Multimedia\",\n \"ref_id\": \"b17\",\n \"title\": \"Attribute restoration framework for anomaly detection\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"D Abati; A Porrello; S Calderara; R Cucchiara\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Latent space autoregression for novelty detection\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"N Cohen; Y Hoshen\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Sub-image anomaly detection with deep pyramid correspondences\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"P Bergmann; M Fauser; D Sattlegger; C Steger\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"M Salehi; N Sadjadi; S Baselizadeh; M H Rohban; H R Rabiee\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Multiresolution knowledge distillation for anomaly detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"A Vaswani; N Shazeer; N Parmar; J Uszkoreit; L Jones; A N Gomez; L Kaiser; I Polosukhin\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Attention is all you need\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"D Gudovskiy; S Ishizaka; K Kozuka\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"P Bergmann; S Löwe; M Fauser; D Sattlegger; C Steger\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Improving unsupervised defect segmentation by applying structural similarity to autoencoders\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"K Sato; K Hama; T Matsubara; K Uehara\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b25\",\n \"title\": \"Predictable uncertaintyaware unsupervised deep anomaly segmentation\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"W Liu; R Li; M Zheng; S Karanam; Z Wu; B Bhanu; R J Radke; O Camps\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Towards visually explaining variational autoencoders\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Y Lu; P Xu\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Anomaly detection for skin disease images using variational autoencoder\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"M Sabokrou; M Khalooei; M Fathy; E Adeli\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Adversarially learned one-class classifier for novelty detection\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"E Eskin; A Arnold; M Prerau; L Portnoy; S Stolfo\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b29\",\n \"title\": \"A geometric framework for unsupervised anomaly detection\",\n \"year\": \"2002\"\n },\n {\n \"authors\": \"J A Hartigan; M A Wong\",\n \"journal\": \"Journal of the royal statistical society. series c (applied statistics)\",\n \"ref_id\": \"b30\",\n \"title\": \"Algorithm as 136: A k-means clustering algorithm\",\n \"year\": \"1979\"\n },\n {\n \"authors\": \"H Y Teh; I Kevin; K Wang; A W Kempa-Liehr\",\n \"journal\": \"IEEE Sensors Journal\",\n \"ref_id\": \"b31\",\n \"title\": \"Expect the unexpected: Unsupervised feature selection for automated sensor anomaly detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"J Yi; S Yoon\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Patch svdd: Patch-level svdd for anomaly detection and segmentation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"C.-L Li; K Sohn; J Yoon; T Pfister\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Cutpaste: Self-supervised learning for anomaly detection and localization\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"T Defard; A Setkov; A Loesch; R Audigier\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b34\",\n \"title\": \"Padim: A patch distribution modeling framework for anomaly detection and localization\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"G Wang; S Han; E Ding; D Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Student-teacher feature pyramid matching for unsupervised anomaly detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"S Hochreiter; J Schmidhuber\",\n \"journal\": \"Neural computation\",\n \"ref_id\": \"b36\",\n \"title\": \"Long short-term memory\",\n \"year\": \"1997\"\n },\n {\n \"authors\": \"A Dosovitskiy; L Beyer; A Kolesnikov; D Weissenborn; X Zhai; T Unterthiner; M Dehghani; M Minderer; G Heigold; S Gelly\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"An image is worth 16x16 words: Transformers for image recognition at scale\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"N Carion; F Massa; G Synnaeve; N Usunier; A Kirillov; S Zagoruyko\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b38\",\n \"title\": \"End-to-end object detection with transformers\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"X Zhu; W Su; L Lu; B Li; X Wang; J Dai\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Deformable detr: Deformable transformers for end-to-end object detection\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"J Chen; Y Lu; Q Yu; X Luo; E Adeli; Y Wang; L Lu; A L Yuille; Y Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Transunet: Transformers make strong encoders for medical image segmentation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"O Ronneberger; P Fischer; T Brox\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b41\",\n \"title\": \"U-net: Convolutional networks for biomedical image segmentation\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"J Pirnay; K Chai\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Inpainting transformer for anomaly detection\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"S Zhang; X He; S Yan\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b43\",\n \"title\": \"Latentgnn: Learning efficient non-local relations for visual recognition\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"X Liang; Z Hu; H Zhang; L Lin; E P Xing\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b44\",\n \"title\": \"Symbolic graph reasoning meets convolutions\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Y Chen; M Rohrbach; Z Yan; Y Shuicheng; J Feng; Y Kalantidis\",\n \"journal\": \"\",\n \"ref_id\": \"b45\",\n \"title\": \"Graph-based global reasoning networks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"M A Islam; S Jia; N D Bruce\",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"How much position information do convolutional neural networks encode?\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"J Yim; D Joo; J Bae; J Kim\",\n \"journal\": \"\",\n \"ref_id\": \"b47\",\n \"title\": \"A gift from knowledge distillation: Fast optimization, network minimization and transfer learning\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"S Kornblith; J Shlens; Q V Le\",\n \"journal\": \"\",\n \"ref_id\": \"b48\",\n \"title\": \"Do better imagenet models transfer better?\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"B Wu; C Xu; X Dai; A Wan; P Zhang; Z Yan; M Tomizuka; J Gonzalez; K Keutzer; P Vajda\",\n \"journal\": \"\",\n \"ref_id\": \"b49\",\n \"title\": \"Visual transformers: Token-based image representation and processing for computer vision\",\n \"year\": \"2020\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[8,56.88,173.63,331.5,25.84],"formula_id":"formula_0","formula_text":"F G (y k ) = F 1 G (y k ) , F 2 G (y k ) , ..., F L G (y k ) and F T (y k ) = F 1 T (y k ) , F 2 T (y k ) , ."},{"formula_coordinates":[8,51,222.71,182.16,12.76],"formula_id":"formula_1","formula_text":"T G = T 1 G , T 2 G , ..., T L G and T T = T 1 T , T 2 T , ."},{"formula_coordinates":[8,51,525.54,337.46,22.66],"formula_id":"formula_2","formula_text":"x k ∈ R W ×H×C , k ∈ [1, N ] of width H, height W,"},{"formula_coordinates":[8,116.4,549.47,178.59,12.76],"formula_id":"formula_3","formula_text":"F G (x k ) = F 1 G (x k ) , F 2 G (x k ) , ..., F L G (x k"},{"formula_coordinates":[8,51,573.54,144.84,18.19],"formula_id":"formula_4","formula_text":"F l G (x k ) ∈ R w l ×h l ×c l for l ∈ [1, L]"},{"formula_coordinates":[9,148.44,170.94,239.91,25.87],"formula_id":"formula_5","formula_text":"G l = L g F l GK (x k ) √ c l ∈ R (w l •h l )×g(1)"},{"formula_coordinates":[9,149.28,219.59,239.07,18.62],"formula_id":"formula_6","formula_text":"V l = L d F l GK (x k ) ∈ R (w l •h l )×d .(2)"},{"formula_coordinates":[9,155.04,306.66,233.31,18.55],"formula_id":"formula_7","formula_text":"T l = softmax(G l ) T V l ∈ R d×g ,(3)"},{"formula_coordinates":[9,51,564.95,132.12,12.76],"formula_id":"formula_8","formula_text":"F T (x k ) = F 1 T (x k ) , F 2 T (x k ) , ."},{"formula_coordinates":[10,51,419.51,68.88,12.76],"formula_id":"formula_9","formula_text":"T G = T 1 G , , T 2 G , ."},{"formula_coordinates":[10,114.36,574.98,273.99,18.67],"formula_id":"formula_10","formula_text":"q e = W q E in , k e = W k E in , v e = W v E in ∈ R d×(g•l)(4)"},{"formula_coordinates":[11,141.6,103.98,246.75,25.54],"formula_id":"formula_11","formula_text":"A e = v e softmax k T e q e √ d ∈ R d×(g•l) ,(5)"},{"formula_coordinates":[11,131.28,187.62,257.07,18.55],"formula_id":"formula_12","formula_text":"E a = E in + LayerN orm (A e ) ∈ R d×(g•l) ,(6)"},{"formula_coordinates":[11,110.16,259.5,278.18,18.55],"formula_id":"formula_13","formula_text":"E out = LayerN orm (E a + L 1 σ (L 2 E a )) ∈ R d×(g•l) ,(7)"},{"formula_coordinates":[11,108.12,391.38,280.23,18.67],"formula_id":"formula_14","formula_text":"q d = W q D in , k d = W k E out , v d = W v E out ∈ R d×(g•l) ,(8)"},{"formula_coordinates":[11,141.12,416.1,247.23,25.54],"formula_id":"formula_15","formula_text":"A d = v d softmax k T d q d √ d ∈ R d×(g•l) .(9)"},{"formula_coordinates":[11,131.52,550.74,256.95,18.55],"formula_id":"formula_16","formula_text":"D a = D in + LayerN orm (A d ) ∈ R d×(g•l)(10)"},{"formula_coordinates":[11,108.72,586.62,279.75,18.67],"formula_id":"formula_17","formula_text":"D out = LayerN orm (D a + L 1 σ (L 2 D a )) ∈ R d×(g•l) .(11)"},{"formula_coordinates":[12,127.92,98.63,177.38,12.75],"formula_id":"formula_18","formula_text":"F G (x k ) = F 1 G (x k ) , F 2 G (x k ) , ..., F L G (x k )"},{"formula_coordinates":[12,119.16,133.98,269.31,30.91],"formula_id":"formula_19","formula_text":"F l M (x k ) = F l T (x k ) + T l v softmax X l q T T l k √ c l T ,(12)"},{"formula_coordinates":[12,51,229.55,179.88,12.76],"formula_id":"formula_20","formula_text":"F M (x k )= F 1 M (x k ), F 2 M (x k ), ..., F L M (x k ) ."},{"formula_coordinates":[12,51,309.47,337.33,25.84],"formula_id":"formula_21","formula_text":"F G (x k )= F 1 G (x k ), F 2 G (x k ), ..., F L G (x k ) and F M (x k )= F 1 M (x k ), F 2 M (x k ), ..., F L M (x k"},{"formula_coordinates":[12,51,395.65,337.32,59.4],"formula_id":"formula_22","formula_text":"for l ∈ [1, L], i ∈ [1, w l ], j ∈ [1, h l ] as follow: p l (i,j) (x k ) = 1 2 F l G (x k ) (i,j) -F l M (x k ) (i,j) 2 2 . (13"},{"formula_coordinates":[12,384.02,438.01,4.45,9.96],"formula_id":"formula_23","formula_text":")"},{"formula_coordinates":[12,138.72,526.5,245.3,34.03],"formula_id":"formula_24","formula_text":"L (x k ) = L l=1 1 w l • h l w l i=1 h l j=1 p l (i,j) (x k ), (14"},{"formula_coordinates":[12,384.02,536.65,4.45,9.96],"formula_id":"formula_25","formula_text":")"},{"formula_coordinates":[13,51,159.73,337.37,29.04],"formula_id":"formula_26","formula_text":"y k ∈ D test for k ∈ [1, M ],"},{"formula_coordinates":[13,227.16,183.73,101.88,17.04],"formula_id":"formula_27","formula_text":"of i ∈ [1, H], j ∈ [1, W ]"},{"formula_coordinates":[13,51,242.63,337.33,25.84],"formula_id":"formula_28","formula_text":"F G (y k )= F 1 G (y k ), F 2 G (y k ), ..., F L G (y k ) and F M (y k )= F 1 M (y k ), F 2 M (y k ), ..., F L M (y k )"},{"formula_coordinates":[13,110.16,324.3,273.86,34.03],"formula_id":"formula_29","formula_text":"P l = 1 w l • h l c l i=1 1 2 F l G (y k ) -F l M (y k ) 2 2 ∈ R w l ×h l , (15"},{"formula_coordinates":[13,384.02,334.45,4.45,9.96],"formula_id":"formula_30","formula_text":")"},{"formula_coordinates":[13,110.28,426.11,278.19,37.46],"formula_id":"formula_31","formula_text":"α l mse = 1 w l • h l w l i=1 h l j=1 F l G (y k ) (i,j) -F l M (y k ) (i,j) 2(16)"},{"formula_coordinates":[13,122.88,493.02,261.14,30.09],"formula_id":"formula_32","formula_text":"α l cos = 1 - vec F l G (y k ) T • vec F l M (y k ) vec F l G (y k ) vec F l M (y k ) . (17"},{"formula_coordinates":[13,384.02,503.89,4.45,9.96],"formula_id":"formula_33","formula_text":")"},{"formula_coordinates":[14,173.64,115.98,210.38,31.75],"formula_id":"formula_34","formula_text":"α l = λ • α l cos • α l mse α l mse + λ • α l cos , (18"},{"formula_coordinates":[14,384.02,123.85,4.45,9.96],"formula_id":"formula_35","formula_text":")"},{"formula_coordinates":[14,133.08,200.58,250.94,30.81],"formula_id":"formula_36","formula_text":"M (y k ) = G σ L l=1 α l R P l ∈ R H×W . (19"},{"formula_coordinates":[14,384.02,210.49,4.45,9.96],"formula_id":"formula_37","formula_text":")"},{"formula_coordinates":[15,165.84,265.14,222.63,20.7],"formula_id":"formula_38","formula_text":"lr = lr init × rate step totalstep ,(20)"}],"string":"[\n {\n \"formula_coordinates\": [\n 8,\n 56.88,\n 173.63,\n 331.5,\n 25.84\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"F G (y k ) = F 1 G (y k ) , F 2 G (y k ) , ..., F L G (y k ) and F T (y k ) = F 1 T (y k ) , F 2 T (y k ) , .\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 51,\n 222.71,\n 182.16,\n 12.76\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"T G = T 1 G , T 2 G , ..., T L G and T T = T 1 T , T 2 T , .\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 51,\n 525.54,\n 337.46,\n 22.66\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"x k ∈ R W ×H×C , k ∈ [1, N ] of width H, height W,\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 116.4,\n 549.47,\n 178.59,\n 12.76\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"F G (x k ) = F 1 G (x k ) , F 2 G (x k ) , ..., F L G (x k\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 51,\n 573.54,\n 144.84,\n 18.19\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"F l G (x k ) ∈ R w l ×h l ×c l for l ∈ [1, L]\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 148.44,\n 170.94,\n 239.91,\n 25.87\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"G l = L g F l GK (x k ) √ c l ∈ R (w l •h l )×g(1)\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 149.28,\n 219.59,\n 239.07,\n 18.62\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"V l = L d F l GK (x k ) ∈ R (w l •h l )×d .(2)\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 155.04,\n 306.66,\n 233.31,\n 18.55\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"T l = softmax(G l ) T V l ∈ R d×g ,(3)\"\n },\n {\n \"formula_coordinates\": [\n 9,\n 51,\n 564.95,\n 132.12,\n 12.76\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"F T (x k ) = F 1 T (x k ) , F 2 T (x k ) , .\"\n },\n {\n \"formula_coordinates\": [\n 10,\n 51,\n 419.51,\n 68.88,\n 12.76\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"T G = T 1 G , , T 2 G , .\"\n },\n {\n \"formula_coordinates\": [\n 10,\n 114.36,\n 574.98,\n 273.99,\n 18.67\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"q e = W q E in , k e = W k E in , v e = W v E in ∈ R d×(g•l)(4)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 141.6,\n 103.98,\n 246.75,\n 25.54\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"A e = v e softmax k T e q e √ d ∈ R d×(g•l) ,(5)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 131.28,\n 187.62,\n 257.07,\n 18.55\n ],\n \"formula_id\": \"formula_12\",\n \"formula_text\": \"E a = E in + LayerN orm (A e ) ∈ R d×(g•l) ,(6)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 110.16,\n 259.5,\n 278.18,\n 18.55\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"E out = LayerN orm (E a + L 1 σ (L 2 E a )) ∈ R d×(g•l) ,(7)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 108.12,\n 391.38,\n 280.23,\n 18.67\n ],\n \"formula_id\": \"formula_14\",\n \"formula_text\": \"q d = W q D in , k d = W k E out , v d = W v E out ∈ R d×(g•l) ,(8)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 141.12,\n 416.1,\n 247.23,\n 25.54\n ],\n \"formula_id\": \"formula_15\",\n \"formula_text\": \"A d = v d softmax k T d q d √ d ∈ R d×(g•l) .(9)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 131.52,\n 550.74,\n 256.95,\n 18.55\n ],\n \"formula_id\": \"formula_16\",\n \"formula_text\": \"D a = D in + LayerN orm (A d ) ∈ R d×(g•l)(10)\"\n },\n {\n \"formula_coordinates\": [\n 11,\n 108.72,\n 586.62,\n 279.75,\n 18.67\n ],\n \"formula_id\": \"formula_17\",\n \"formula_text\": \"D out = LayerN orm (D a + L 1 σ (L 2 D a )) ∈ R d×(g•l) .(11)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 127.92,\n 98.63,\n 177.38,\n 12.75\n ],\n \"formula_id\": \"formula_18\",\n \"formula_text\": \"F G (x k ) = F 1 G (x k ) , F 2 G (x k ) , ..., F L G (x k )\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 119.16,\n 133.98,\n 269.31,\n 30.91\n ],\n \"formula_id\": \"formula_19\",\n \"formula_text\": \"F l M (x k ) = F l T (x k ) + T l v softmax X l q T T l k √ c l T ,(12)\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 51,\n 229.55,\n 179.88,\n 12.76\n ],\n \"formula_id\": \"formula_20\",\n \"formula_text\": \"F M (x k )= F 1 M (x k ), F 2 M (x k ), ..., F L M (x k ) .\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 51,\n 309.47,\n 337.33,\n 25.84\n ],\n \"formula_id\": \"formula_21\",\n \"formula_text\": \"F G (x k )= F 1 G (x k ), F 2 G (x k ), ..., F L G (x k ) and F M (x k )= F 1 M (x k ), F 2 M (x k ), ..., F L M (x k\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 51,\n 395.65,\n 337.32,\n 59.4\n ],\n \"formula_id\": \"formula_22\",\n \"formula_text\": \"for l ∈ [1, L], i ∈ [1, w l ], j ∈ [1, h l ] as follow: p l (i,j) (x k ) = 1 2 F l G (x k ) (i,j) -F l M (x k ) (i,j) 2 2 . (13\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 384.02,\n 438.01,\n 4.45,\n 9.96\n ],\n \"formula_id\": \"formula_23\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 138.72,\n 526.5,\n 245.3,\n 34.03\n ],\n \"formula_id\": \"formula_24\",\n \"formula_text\": \"L (x k ) = L l=1 1 w l • h l w l i=1 h l j=1 p l (i,j) (x k ), (14\"\n },\n {\n \"formula_coordinates\": [\n 12,\n 384.02,\n 536.65,\n 4.45,\n 9.96\n ],\n \"formula_id\": \"formula_25\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 51,\n 159.73,\n 337.37,\n 29.04\n ],\n \"formula_id\": \"formula_26\",\n \"formula_text\": \"y k ∈ D test for k ∈ [1, M ],\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 227.16,\n 183.73,\n 101.88,\n 17.04\n ],\n \"formula_id\": \"formula_27\",\n \"formula_text\": \"of i ∈ [1, H], j ∈ [1, W ]\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 51,\n 242.63,\n 337.33,\n 25.84\n ],\n \"formula_id\": \"formula_28\",\n \"formula_text\": \"F G (y k )= F 1 G (y k ), F 2 G (y k ), ..., F L G (y k ) and F M (y k )= F 1 M (y k ), F 2 M (y k ), ..., F L M (y k )\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 110.16,\n 324.3,\n 273.86,\n 34.03\n ],\n \"formula_id\": \"formula_29\",\n \"formula_text\": \"P l = 1 w l • h l c l i=1 1 2 F l G (y k ) -F l M (y k ) 2 2 ∈ R w l ×h l , (15\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 384.02,\n 334.45,\n 4.45,\n 9.96\n ],\n \"formula_id\": \"formula_30\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 110.28,\n 426.11,\n 278.19,\n 37.46\n ],\n \"formula_id\": \"formula_31\",\n \"formula_text\": \"α l mse = 1 w l • h l w l i=1 h l j=1 F l G (y k ) (i,j) -F l M (y k ) (i,j) 2(16)\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 122.88,\n 493.02,\n 261.14,\n 30.09\n ],\n \"formula_id\": \"formula_32\",\n \"formula_text\": \"α l cos = 1 - vec F l G (y k ) T • vec F l M (y k ) vec F l G (y k ) vec F l M (y k ) . (17\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 384.02,\n 503.89,\n 4.45,\n 9.96\n ],\n \"formula_id\": \"formula_33\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 14,\n 173.64,\n 115.98,\n 210.38,\n 31.75\n ],\n \"formula_id\": \"formula_34\",\n \"formula_text\": \"α l = λ • α l cos • α l mse α l mse + λ • α l cos , (18\"\n },\n {\n \"formula_coordinates\": [\n 14,\n 384.02,\n 123.85,\n 4.45,\n 9.96\n ],\n \"formula_id\": \"formula_35\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 14,\n 133.08,\n 200.58,\n 250.94,\n 30.81\n ],\n \"formula_id\": \"formula_36\",\n \"formula_text\": \"M (y k ) = G σ L l=1 α l R P l ∈ R H×W . (19\"\n },\n {\n \"formula_coordinates\": [\n 14,\n 384.02,\n 210.49,\n 4.45,\n 9.96\n ],\n \"formula_id\": \"formula_37\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 15,\n 165.84,\n 265.14,\n 222.63,\n 20.7\n ],\n \"formula_id\": \"formula_38\",\n \"formula_text\": \"lr = lr init × rate step totalstep ,(20)\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"A T E X template Multiresolution Feature Guidance Based Transformer for Anomaly Detection"},"abstract":{"kind":"string","value":"Anomaly detection is represented as an unsupervised learning to identify deviated images from normal images. In general, there are two main challenges of anomaly detection tasks, i.e., the class imbalance and the unexpectedness of anomalies. In this paper, we propose a multiresolution feature guidance method based on Transformer named GTrans for unsupervised anomaly detection and localization. In GTrans, an Anomaly Guided Network (AGN) pre-trained on ImageNet is developed to provide surrogate labels for features and tokens. Under the tacit knowledge guidance of the AGN, the anomaly detection network named Trans utilizes Transformer to effectively establish a relationship between features with multiresolution, enhancing the ability of the Trans in fitting the normal data manifold. Due to the strong generalization ability of AGN, GTrans locates anomalies by comparing the differences in spatial distance and direction of multi-scale features extracted from the AGN and the Trans. Our experiments demonstrate that the proposed GTrans achieves state-of-the-art performance in both detection and localization on the MVTec AD dataset. GTrans achieves image-level and pixel-level anomaly detection AUROC scores of 99.0% and 97.9% on the MVTec AD dataset, respectively."},"authors":{"kind":"string","value":"Shuting Yan; Pingping Chen; Honghui Chen; Huan Mao; Feng Chen; Zhijian Lin"},"figures":{"kind":"list like","value":[{"figure_caption":"Fig. 11Fig. 1 Visual results from the MVTec AD datasets. Superimposed on the images are the anomaly localization map from GTrans. Red areas correspond to the located anomalies, whereas the blue areas indicate the normality regions.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Fig. 22Fig. 2 The overall architecture of the GTrans. The solid purple arrow represents the processing of the AGN, and the solid orange arrow represents the processing of the Trans network. Given an input image y k , we extract groups of multiresolution features F G (y k ) and F T (y k ) from B-AGN and B-Trans, where B-AGN and B-Trans represent the backbone of AGN and Trans, respectively. F G (y k ) and F T (y k ) are input into Tokenizer to obtain groups of tokens T G and T T . AGN guide the decoder module of TFM to enhance the fitting ability of Trans in normal data manifold. The mapper module maps the token output by TFM into original critical layers. GTrans locates anomalies by comparing the differences in spatial distance and direction of multi-scale features extracted from the AGN and the Trans. Dotted green arrows represent the process of generating an anomaly map.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Fig. 44Fig. 4 Visualization of anomaly samples from bottle, hazelnut and screw of MVTec AD dataset. Columns from left to right correspond to anomaly sample, ground truth, anomaly maps generated by three layers (Layer 1, Layer 2, Layer 3), and the final anomaly maps respectively.","figure_data":"","figure_id":"fig_2","figure_label":"4","figure_type":"figure"},{"figure_caption":"Comparison of our models with the state-of-the-art for the anomaly localization on the MVTec AD dataset. Results are displayed as tuples(AUROC%, PRO-SCORE%).","figure_data":"","figure_id":"tab_1","figure_label":"2","figure_type":"table"},{"figure_caption":"Study of the anomaly detection and localization performance with a R34 backbone using different feature layers. results are displayed as tuples(detection AUROC%, location AUROC%) on MVTec AD dataset.","figure_data":"Layer usedavg.texturesavg.objectsavg.all categoriesLayer 1+2+3+4(98.4, 91.5)(97.3, 93.2)(97.6, 92.8)Layer 1+2+3(99.6, 96.0)(97.9, 96.7)(98.3, 96.5)Layer 2+3+4(98.9, 92.2)(97.1, 94.2)(97.6, 93.7)Layer 2+3(99.6, 95.6)(97.9, 96.5)(98.3, 96.2)","figure_id":"tab_2","figure_label":"3","figure_type":"table"},{"figure_caption":"Performance with a R34 backbone with/without TFM architecture. Results are displayed as tuples(detection AUROC%, location AUROC%) on MVTec AD dataset.","figure_data":"Structureavg.texturesavg.objectsavg.all categorieswithout TFM(99.4, 96.8)(93.2, 97.3)(95.2, 97.1)with TFM(99.6, 97.9) (98.7, 98.0)(97.9, 99.0)Table 5 Study of performance with a R34 backbone using different structure of TFM.results are displayed as tuples(detection AUROC%, location AUROC%) on MVTec ADdataset.Structureblockavg.texturesavg.objectsavg.all categoriesS=1(99.4, 97.9)(98.1, 97.6)(98.5, 97.7)pure encoderS=2(99.3, 97.9)(98.2, 97.7)(98.6, 97.7)S=3(99.3, 97.9)(98.2, 97.6)(98.6, 97.7)S=1(99.4, 97.9)(98.3, 97.9)(98.7, 97.9)added decoderS=2(99.6, 97.9)(98.7, 98.0)(99.0, 97.9)S=3(99.4, 98.0)(98.7, 97.9)(98.9, 97.9)","figure_id":"tab_3","figure_label":"4","figure_type":"table"},{"figure_caption":"Study of performance with a R34 backbone using different function of anomaly maps. results are displayed as tuples(detection AUROC%, location AUROC%) on MVTec AD dataset.","figure_data":"Coefficientavg.texturesavg.objectsavg.all categories0.5(98.7, 98.0)(97.6, 97.8)(98.0, 97.8)αmse(99.5, 97.7)(98.7, 98.0)(98.9, 97.9)αcos(99.6, 97.8)(98.6, 97.9)(98.9, 97.9)α(99.6, 97.9)(98.7, 98.0)(99.0, 97.9)","figure_id":"tab_4","figure_label":"6","figure_type":"table"},{"figure_caption":"Study of performance with a R34 backbone using different combination of anomaly maps. results are displayed as tuples(detection AUROC%, location AUROC%) on MVTec AD dataset.","figure_data":"Combination Modeavg.texturesavg.objectsavg.all categoriesP1: 3(99.3, 95.2)(97.2, 96.0)(97.9, 95.7)P2: 1*3(99.5, 98.2)(98.0, 97.9)(98.5, 98.0)P3: 2*3(99.5, 97.3)(98.4, 97.8)(98.8, 97.6)P4: 1+2+3(99.5, 95.6)(97.8, 96.9)(98.3,96.5)P5: 1*2*3(99.5, 98.2)(98.4, 97.8)(98.8, 97.9)P6: 1*3+2*3(99.6, 97.9)(98.7, 98.0)(99.0, 97.9)","figure_id":"tab_5","figure_label":"7","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Fig. 11Fig. 1 Visual results from the MVTec AD datasets. Superimposed on the images are the anomaly localization map from GTrans. Red areas correspond to the located anomalies, whereas the blue areas indicate the normality regions.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 22Fig. 2 The overall architecture of the GTrans. The solid purple arrow represents the processing of the AGN, and the solid orange arrow represents the processing of the Trans network. Given an input image y k , we extract groups of multiresolution features F G (y k ) and F T (y k ) from B-AGN and B-Trans, where B-AGN and B-Trans represent the backbone of AGN and Trans, respectively. F G (y k ) and F T (y k ) are input into Tokenizer to obtain groups of tokens T G and T T . AGN guide the decoder module of TFM to enhance the fitting ability of Trans in normal data manifold. The mapper module maps the token output by TFM into original critical layers. GTrans locates anomalies by comparing the differences in spatial distance and direction of multi-scale features extracted from the AGN and the Trans. Dotted green arrows represent the process of generating an anomaly map.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 44Fig. 4 Visualization of anomaly samples from bottle, hazelnut and screw of MVTec AD dataset. Columns from left to right correspond to anomaly sample, ground truth, anomaly maps generated by three layers (Layer 1, Layer 2, Layer 3), and the final anomaly maps respectively.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Comparison of our models with the state-of-the-art for the anomaly localization on the MVTec AD dataset. Results are displayed as tuples(AUROC%, PRO-SCORE%).\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Study of the anomaly detection and localization performance with a R34 backbone using different feature layers. results are displayed as tuples(detection AUROC%, location AUROC%) on MVTec AD dataset.\",\n \"figure_data\": \"Layer usedavg.texturesavg.objectsavg.all categoriesLayer 1+2+3+4(98.4, 91.5)(97.3, 93.2)(97.6, 92.8)Layer 1+2+3(99.6, 96.0)(97.9, 96.7)(98.3, 96.5)Layer 2+3+4(98.9, 92.2)(97.1, 94.2)(97.6, 93.7)Layer 2+3(99.6, 95.6)(97.9, 96.5)(98.3, 96.2)\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Performance with a R34 backbone with/without TFM architecture. Results are displayed as tuples(detection AUROC%, location AUROC%) on MVTec AD dataset.\",\n \"figure_data\": \"Structureavg.texturesavg.objectsavg.all categorieswithout TFM(99.4, 96.8)(93.2, 97.3)(95.2, 97.1)with TFM(99.6, 97.9) (98.7, 98.0)(97.9, 99.0)Table 5 Study of performance with a R34 backbone using different structure of TFM.results are displayed as tuples(detection AUROC%, location AUROC%) on MVTec ADdataset.Structureblockavg.texturesavg.objectsavg.all categoriesS=1(99.4, 97.9)(98.1, 97.6)(98.5, 97.7)pure encoderS=2(99.3, 97.9)(98.2, 97.7)(98.6, 97.7)S=3(99.3, 97.9)(98.2, 97.6)(98.6, 97.7)S=1(99.4, 97.9)(98.3, 97.9)(98.7, 97.9)added decoderS=2(99.6, 97.9)(98.7, 98.0)(99.0, 97.9)S=3(99.4, 98.0)(98.7, 97.9)(98.9, 97.9)\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Study of performance with a R34 backbone using different function of anomaly maps. results are displayed as tuples(detection AUROC%, location AUROC%) on MVTec AD dataset.\",\n \"figure_data\": \"Coefficientavg.texturesavg.objectsavg.all categories0.5(98.7, 98.0)(97.6, 97.8)(98.0, 97.8)αmse(99.5, 97.7)(98.7, 98.0)(98.9, 97.9)αcos(99.6, 97.8)(98.6, 97.9)(98.9, 97.9)α(99.6, 97.9)(98.7, 98.0)(99.0, 97.9)\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Study of performance with a R34 backbone using different combination of anomaly maps. results are displayed as tuples(detection AUROC%, location AUROC%) on MVTec AD dataset.\",\n \"figure_data\": \"Combination Modeavg.texturesavg.objectsavg.all categoriesP1: 3(99.3, 95.2)(97.2, 96.0)(97.9, 95.7)P2: 1*3(99.5, 98.2)(98.0, 97.9)(98.5, 98.0)P3: 2*3(99.5, 97.3)(98.4, 97.8)(98.8, 97.6)P4: 1+2+3(99.5, 95.6)(97.8, 96.9)(98.3,96.5)P5: 1*2*3(99.5, 98.2)(98.4, 97.8)(98.8, 97.9)P6: 1*3+2*3(99.6, 97.9)(98.7, 98.0)(99.0, 97.9)\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"7\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Supporting Evidence\", \"Citation\": \"[1][2][3]\", \"Explanation\": \"The cited works provide evidence of the use of industrial inspection techniques in anomaly detection, which supports the claim that anomaly detection techniques have been extensively studied in various research and application domains.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[4]\", \"Explanation\": \"The cited work highlights the use of anomaly detection techniques in medical applications, which further supports the claim of the widespread application of these techniques in different fields.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[5][6][7]\", \"Explanation\": \"The cited works suggest a focus on computer vision in anomaly detection tasks, which extends the research on anomaly detection beyond the traditional manual visual detection method.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[8]\", \"Explanation\": \"The cited work highlights the class imbalance problem as a common challenge in anomaly detection tasks, which supports the claim of the challenges faced in this field.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work by Bergmann et al. provides a benchmark dataset for anomaly detection and localization, which serves as a foundational element for the study conducted in the citing paper on the same topic.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[20]\", \"Explanation\": \"The cited work by Cohen et al. introduced a method of using a pre-trained network to extract features for anomaly detection, which the citing paper adopts to improve performance in small sample datasets.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[21]\", \"Explanation\": \"The cited work by Bergmann et al. proposed a student-teacher network to model the distribution of features extracted from normal images, which the citing paper adopts to address the limitations of the student network in making wrong judgments in abnormal data manifolds.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[21]\", \"Explanation\": \"The cited work introduces a method for knowledge distillation that only imitates the last layer in the teacher network, which the citing paper builds upon to fully utilize the information in intermediate layers.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[22]\", \"Explanation\": \"The cited work proposes a novel method for knowledge distillation that distils the comprehensive knowledge of the pre-trained network at several critical layers, which the citing paper adopts to fully exploit the multi-scale information in computer vision algorithms.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[23]\", \"Explanation\": \"The cited work of Transformer has achieved great success in image classification and recognition, which inspires the citing paper to propose a multiresolution feature guidance method based on Transformer for unsupervised anomaly detection and localization.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[10,18,19,25]\", \"Explanation\": \"The cited works on autoencoders provide a methodological basis for the reconstruction-based methods used in the citing paper for anomaly detection and localization.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[11,[26][27][28]\", \"Explanation\": \"The cited works on variational autoencoders offer a methodological foundation for the reconstruction-based methods used in the citing paper for anomaly detection and localization.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[9,[14][15][16][17]29]\", \"Explanation\": \"The cited works on generative adversarial networks provide a methodological basis for the reconstruction-based methods used in the citing paper for anomaly detection and localization.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"SSIM-AE [25]\", \"Explanation\": \"The cited work on SSIM-AE provides a methodological basis for the use of SSIM as a loss function and anomaly measure in the reconstruction-based methods used in the citing paper for anomaly detection and localization.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"ARNet [18]\", \"Explanation\": \"The cited work on ARNet provides a methodological basis for the use of semantic feature embeddings in the reconstruction-based methods used in the citing paper for anomaly detection and localization.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"MemAE [10]\", \"Explanation\": \"The cited work on MemAE provides a methodological basis for the use of an autoencoder with a memory module in the reconstruction-based methods used in the citing paper for anomaly detection and localization.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[15]\", \"Explanation\": \"The cited work AnoGAN proposes a novel anomaly scoring scheme that the citing paper adopts in the process of detecting anomalies in the potential space.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[9]\", \"Explanation\": \"The cited work OCGAN utilizes a dual latent space learning process that the citing paper leverages to constrain the latent space of generators and discriminators in a specific category.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work GANomaly adds an additional encoder to the generator in the process of reducing inference time in the detection of anomalies in the potential space.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[17]\", \"Explanation\": \"The cited work f-AnoGan also adds an additional encoder to the generator in the process of reducing inference time in the detection of anomalies in the potential space.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[10,25]\", \"Explanation\": \"The cited works failed to detect anomalies due to the strong generalization ability of deep models, which is a foundational point for the citing paper to highlight the need for more effective methods in detecting anomalies.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[9,[15][16][17]\", \"Explanation\": \"The cited works failed in one-class settings, which the citing paper uses to emphasize the need for new methods in detecting anomalies in one-class settings.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[30]\", \"Explanation\": \"The cited work introduces the K-Nearest Neighbor (KNN) method, which the citing paper adopts to match the best approximate features in the feature space for anomaly detection.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[31]\", \"Explanation\": \"The cited work presents the K-means method, which the citing paper uses to cluster the features in the feature space for anomaly detection.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[32]\", \"Explanation\": \"The cited work describes the Principal Component Analysis (PCA) method, which the citing paper employs to analyze the features in the feature space for anomaly detection.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[33]\", \"Explanation\": \"The cited work introduces the Patch SVDD method, which the citing paper utilizes to inspect images at the patch level in the feature space for anomaly detection.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[34]\", \"Explanation\": \"The cited work presents the CutPaste method, which the citing paper applies to self-supervised learning for anomaly detection in the feature space.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[34]\", \"Explanation\": \"CutPaste is a data augmentation strategy that the citing paper adopts in their research to improve the performance of self-supervised representation learning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[35]\", \"Explanation\": \"PaDiM is a method that the citing paper uses to model the correlation between semantic layers in the context of data augmentation.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[21]\", \"Explanation\": \"The teacher-student framework introduced in [21] is a foundational concept in the field of unsupervised anomaly detection and localization, which the citing paper builds upon in their research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[22]\", \"Explanation\": \"The extension of multiscale critical layers of knowledge extraction in [22] is a continuation of the work in [21], further improving the performance of the teacher-student framework in unsupervised anomaly detection and localization.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[36]\", \"Explanation\": \"The extension of multiscale critical layers of knowledge extraction in [36] is another continuation of the work in [21], further improving the performance of the teacher-student framework in unsupervised anomaly detection and localization.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[21]\", \"Explanation\": \"The cited work provides a method of extracting only single-layer information, which the citing paper builds upon to make full use of the information in intermediate layers.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[22][36]\", \"Explanation\": \"The cited works highlight the importance of information interaction of multi-scale critical layers, which the citing paper adopts in their research to treat critical layers with multi-resolutions in a more effective manner.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[23]\", \"Explanation\": \"The cited work introduced the Transformer model, which the citing paper adopts to perform Natural Language Processing tasks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[38]\", \"Explanation\": \"The cited work on image classification is extended in the citing paper to explore the use of Transformer in computer vision.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[39][40]\", \"Explanation\": \"The cited works on object detection are further explored in the citing paper to study the adoption of Transformer in computer vision.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[41]\", \"Explanation\": \"The cited work on image segmentation is extended in the citing paper to examine the use of Transformer in computer vision.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[38]\", \"Explanation\": \"The cited work introduces the concept of dividing input images into square uniform patches and using linear transformation to reduce dimensionality and embed location information, which the citing paper adopts in their own research on image classification.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[41]\", \"Explanation\": \"The cited work introduces the TransUNet model for medical image segmentation, which the citing paper extends by utilizing Transformers for encoding tokenized image patches and combining with high-resolution CNN feature maps for precise localization in their own research on medical image segmentation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[43]\", \"Explanation\": \"The cited work introduces a deep Transformer network for anomaly detection and localization, which the citing paper adopts in their own research to achieve efficient and scalable detection and localization of anomalies using a hybrid structure of CNN and Transformer.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[44][45][46]\", \"Explanation\": \"The cited works provide the inspiration for the use of a module called Tokenizer in the citing paper, which is used to extract visual tokens by performing point-wise convolutions and spatial attention on the input feature map.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[47]\", \"Explanation\": \"The cited work provides evidence that CNNs learn information about spatial positioning, which the citing paper uses to inform the design of the encoder-decoder structure in the TFM module.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[38,41]\", \"Explanation\": \"The cited works provide a method of using an encoder-decoder structure in the Transformer to improve the stability of the model in computer vision tasks, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[48]\", \"Explanation\": \"The cited work provides a method for Trans to learn the distilled knowledge from AGN and intermediate-level hints, which the citing paper adopts in their research to intensify the knowledge transfer from AGN to Trans.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[50]\", \"Explanation\": \"The cited work by [50] provides a method for using visual tokens to generalize semantic concepts, which the citing paper adopts in the design of the Tokenizer component in Trans.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[15]\", \"Explanation\": \"The cited work by [15] introduces the AnoGAN model for anomaly detection, which the citing paper uses as a baseline for comparison in Table 1.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[25]\", \"Explanation\": \"The cited work by [25] presents the AE-SSIM model for anomaly detection, which the citing paper uses as a benchmark for performance comparison in Table 1.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work by [16] develops the GANomaly model for anomaly detection, which the citing paper uses as a reference in Table 1 to assess the performance of various methods in the field.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[20]\", \"Explanation\": \"The cited work by [20] introduces the SPADE model for anomaly detection, which the citing paper uses in Table 1 to compare the performance of different methods in the field.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[33]\", \"Explanation\": \"The cited work by [33] presents the Patch-SVDD model for anomaly detection, which the citing paper uses in Table 1 to assess the performance of various methods in the field.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[34]\", \"Explanation\": \"The cited work by [34] introduces the CutPaste model for anomaly detection, which the citing paper uses in Table 1 to compare the performance of different methods in the field.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[35]\", \"Explanation\": \"The cited work by [35] presents the PaDiM model for anomaly detection, which the citing paper uses in Table 1 to assess the performance of various methods in the field.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[43]\", \"Explanation\": \"The cited work by [43] introduces the InTra model for anomaly detection, which the citing paper uses in Table 1 to compare the performance of different methods in the field.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[20]\", \"Explanation\": \"The cited work, SPADE, is used as a benchmark to compare the performance of the citing paper in the object category of anomaly localization on the MVTec AD dataset.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[21]\", \"Explanation\": \"The cited work, Student, is used as a benchmark to compare the performance of the citing paper in the multi-scale feature fusion and information interaction in anomaly localization on the MVTec AD dataset.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[15]\", \"Explanation\": \"AnoGAN is cited as a method for generating high-quality images in a generative adversarial network (GAN) framework.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[25]\", \"Explanation\": \"AE-SSIM is cited as a method for measuring image quality using a deep learning model based on the structural similarity index (SSIM).\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[33]\", \"Explanation\": \"Patch-SVDD is cited as a method for detecting and removing image patches that are outliers in a dataset using a support vector data description (SVDD) model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[21]\", \"Explanation\": \"Student is cited as a method for training a deep learning model in a self-supervised manner, where the model learns to predict the input from a single image without the need for labels.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[34]\", \"Explanation\": \"CutPaste is cited as a method for generating image inpainting by cutting and pasting image patches from a source image to a target image.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[20]\", \"Explanation\": \"SPADE is cited as a method for improving the quality of image inpainting by using a self-attention mechanism to better capture the spatial and semantic information in the image.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[43]\", \"Explanation\": \"InTra is cited as a method for training a deep learning model in a self-supervised manner, where the model learns to predict the input from a single image without the need for labels.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[32]\", \"Explanation\": \"GTrans is cited as a method for training a deep learning model in a self-supervised manner, where the model learns to translate images from one domain to another without the need for labels.\"}]"}}},{"rowIdx":90,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_0"],"heading":"","publication_ref":["b0","b0","b1","b2","b3","b4","b5","b6","b5","b7","b8","b8","b9","b10","b0","b11","b8","b9","b12"],"table_ref":[],"text":"In this paper, we investigate how graph convolutions amplify the popularity bias in RS. Through theoretical analyses, we identify two fundamental factors: (1) with graph convolution (i.e., neighborhood aggregation), popular items exert larger influence than tail items on neighbor users, making the users move towards popular items in the representation space; (2) after multiple times of graph convolution, popular items would affect more high-order neighbors and become more influential. The two points make popular items get closer to almost users and thus being recommended more frequently. To rectify this, we propose to estimate the amplified effect of popular nodes on each node's representation, and intervene the effect after each graph convolution. Specifically, we adopt clustering to discover highly-influential nodes and estimate the amplification effect of each node, then remove the effect from the node embeddings at each graph convolution layer. Our method is simple and genericit can be used in the inference stage to correct existing models rather than training a new model from scratch, and can be applied to various GCN models. We demonstrate our method on two representative GCN backbones LightGCN and UltraGCN, verifying its ability in improving the recommendations of tail items without sacrificing the performance of 1 Introduction\nRecommender system (RS) has become a key tool for personalization in today's Web, and also a research hotpot [1,2]. Recently, methods based on GCN have become prevalent in RS since they can effectively encode collaborative filtering signal [3][4][5][6][7]. Though GCNs can improve the overall recommendation accuracy, we find a downside is they amplify the popularity bias -popular items are more frequently recommended than tail items 2) , which means the tail items are not fairly treated by the algorithm, and the issue would be exacerbated in practical RSs due to the user-system feedback loop.\nFigure 1 provides an evidence that GCNs intensify the popularity bias as the number of graph convolution layers increases. The backbone model is LightGCN [6] using BPR loss [8], a representative and competitive GCN-based recommender. With the increasing of graph convolution layers, the overall accuracy metric (i.e., Recall@20) gradually increases, but the ratio of recommending tail items (i.e., TR@20) drops significantly. This means the improvements of performing graph convolutions is at the expense of tail items, pointing to the amplified popularity bias issue of GCNs [9]. Although there has been recent work in alleviating the popularity bias in RS [9][10][11], they do not reveal how GCN amplifies the bias (i.e., the mechanism) and their solutions are generally applied to RS models rather than tailored for GCNs.\nIn this work, we aim to answer the unsolved questions \"How do graph convolutions amplify popularity bias in RS?\" and \"How to address the issue with minimum changes on the training progress?\" To reveal the mechanism of GCNs amplifying the bias, we conduct theoretical studies by analyzing how popular items influence other nodes. The main findings are twofold: (1) with graph convolution, popular items have a greater impact than tail items on neighbor users, making the users position closer to popular items in the representation space; (2) after multiple times of graph convolution, popular items become more influential by affecting more high-order neighbor users. The theoretical analyses confirm the inherent defect of GCNs in overrecommending popular items.\nTo alleviate the popularity bias in RS, two lines of research work have been conducted. The first is intervening the training process to eliminate the bias effect. For example, propensity-based methods [12] impose lower weights on popular items in training loss and causal methods [9,10] model the causal effect of the bias on the model training. Those methods need to retrain the backbone model, making the solutions costly to use in practical RSs especially to correct the already-deployed GCN models. In contrast, another line is revising the inference stage in a post-hoc way. For example, [13] performs personalized re-ranking on the generated candidates of the RS model to suppress popular items. However, both lines work on the general issue of popularity bias in recommendation, leaving how GCN model suffers from and amplifies the popularity bias untouched.\nTowards the research gap, we propose a new problem to solve -rectifying the popularity bias of GCN model in the inference stage. Such a solution can be used to revise already-trained GCNs, thus is easier to deploy in practice than training a new model. Given a GCN model, we first cluster the node representations to automatically discover the highly-influential nodes. Then, the amplification of popularity bias for each node within its cluster is estimated based on the prior theoretical analyses. Thereafter, the amplification effect in the node representation can be intervened to control the bias. This post-hoc method can be easily deployed in practice to correct existing GCN models and promote the recommendations of tail items without sacrificing the performance of popular items. To summarize, this work makes the following contributions:\n• Providing in-depth theoretical analyses to interpret the popularity bias amplification problem in GCN-based recommenders;\n• Developing a new method working at each graph convolution layer in the inference stage to correct the popularity bias for GCN;\n• Conducting extensive experiments on three real datasets to demonstrate the effectiveness of our method on LightGCN and UltraGCN backbones. The degree of user u or item i e (l) u , e (l) i\nThe embedding of user u or item i at the l-th graph convolution layer L ui\nThe individual loss term of an interaction (u, i)\nC (l) p\nA set of nodes in the p-th cluster obtained by using Kmeans given the embeddings\nE (l) H (l) v A set of nodes { j ∈ C (l) p , d j > d v |v ∈ C (l) p } L (l) v A set of nodes { j ∈ C (l) p , d j < d v |v ∈ C (l) p } θ ϑ ϑ (l) H v\nThe pooling representations after normalization of\nH (l) v θ ϑ ϑ (l) L v\nThe pooling representations after normalization of L (l) v"},{"figure_ref":[],"heading":"Preliminaries","publication_ref":["b13","b14","b5","b13","b15"],"table_ref":[],"text":"Suppose that there are a set of users and a set of items\nU = {u 1 , u 2 , • • • , u M }, I = {i 1 , i 2 , • • • , i N } in a dataset D.\nLet y ui = 1 be the positive label if the user u has interacted with the item i, otherwise y ui = 0. We can construct a user-item bipartite graph B = (V, E) based on the interaction history, where V consists of the set of user and item nodes, and E denotes the set of edges. If y ui = 1, there is an edge between the user u and the item i.\nRecently, many studies opt for powerful GCNs to learn user and item node representations [14,15]. Particularly, we introduce LightGCN [6], which is neat and well represents the GCN-based recommenders. One graph convolution block of Light-GCN can be expressed as:\ne (l) u = i∈N u 1 √ d u √ d i e (l-1) i , e (l) i = u∈N i 1 √ d i √ d u e (l-1) u ,(1)\nwhere d u (d i ) is the degree of user u (item i) in the graph B, N u (N i ) is the one-order neighbor nodes of the user u (item i), e (0) is an ID embedding of a user or an item. After stacking several graph convolution layers, LightGCN combines the embeddings obtained at each layer to form the final representa-tion e of a node. Thereafter, the model prediction is defined as the inner product of user and item final representations, i.e., ŷui = e ⊤ u e i . Another representative work is UltraGCN [14] which skips infinite layers of graph convolution. We also conduct experiments on this model to verify the generality of our method.\nTo optimize model parameters, prior work usually frames it as a supervised learning task, and utilize a pointwise or a pairwise loss for model training, which can be summarized by the following formula: L = (u,i)∈D L ui , where L ui is the individual loss term of an interaction (u, i). Without loss of generality, we investigate the popularity bias amplification problem in GCNs based on BCE loss [16] in the next section. The formal formulation of BCE loss is\nL ui = -[y ui ln σ(ŷ ui ) + (1 -y ui ) ln(1 -σ(ŷ ui ))]. (2)"},{"figure_ref":[],"heading":"METHODOLOGY","publication_ref":[],"table_ref":[],"text":"In this part, we attempt to analyze and resolve the amplified popularity bias of GCNs."},{"figure_ref":["fig_3"],"heading":"Popularity Bias Amplification in GCNs","publication_ref":["b16","b17","b18","b18","b19"],"table_ref":[],"text":"To understand why GCNs amplify the popularity bias, we conduct theoretical analyses and empirical experiments on GCNs: (1) we start with defining the influence between nodes based on the training loss;\n(2) we prove that popular items with higher degrees exert larger influence on neighbor users than tail items with lower-degrees; (3) we reveal that popular items commonly have higher probabilities of being recommended by users after representation updating and graph convolution in GCNs.\nConcisely, we take the e = e (L) at L-th graph convolution layer in Eq. (1) as the final representation of each node. Next, we give the definition of the influence of a user-item pair loss on their neighbors exploiting the concept of influence functions [17][18][19]: Definition 1 (Influence of an observed interaction on a node's representation learning): Suppose that (u, i) is an observed interaction, i.e., there is an edge between node u and node i, some other node k is reachable from node u or i, then the influence of an interaction (u, i) on node k is defined as ∂L ui ∂e k . Without loss of generality, we mainly consider y ui = 1 in BCE loss and have\n∂L ui ∂e k = - ∂ ln σ(ŷ ui ) ∂e k = - ln σ(ŷ ui ) ∂σ(ŷ ui ) • ∂σ(ŷ ui ) ∂ŷ ui • ∂ŷ ui ∂e k = - 1 σ(ŷ ui ) • σ(ŷ ui )(1 -σ(ŷ ui )) • ∂ŷ ui ∂e k = -[1 -σ(ŷ ui )] ∂ŷ ui ∂e k = -λ ui ∂ŷ ui ∂e k ,(3)\nwhere ŷui is the prediction between the user u and the item i and 0 < λ ui < 1.\nDefinition 2 (Influence of a node on another node's representation learning): Suppose that a node i can reach a neighbor node k on a graph 3) . The influence of the loss for the target node i on the node k is defined as ∂L i ∂e k , where L i = j∈N i L i j . In fact, the influence provides a fine-grained scrutiny of updating information of each node at the lens of gradient derivation. We then have the following lemma. Lemma 1. If node i with degree d i can reach node k after stacking L layers of graph convolution, then the influence of node i on node k follows\nE ∂L i ∂e k ∝ -d where E(•) is the expectation, ϑ ϑ ϑ i = E Φ j p=1 1 l=L-1 1 √ d p l e j\nrepresents the expectation on paths starting from a neighbor node j of the node i to the node k, Φ j is the set of all (L -1)-length paths from the node j to the node k, p L-1 and p 1 are the node j and the node k, respectively.\nProof. According to Eq. ( 3) and Eq. ( 1), we obtain\n∂L i ∂e k = j∈N i ∂L i j ∂e k = - j∈N i λ i j ∂ŷ i j ∂e k = - j∈N i λ i j ∂e i ∂e k e j = - j∈N i λ i j Φ p=1 1 l=L 1 d p l e j ,(5)\nwhere p L is the node i, Φ is the set of all L-length random paths on the graph from nodes i to k, and we omit the transpose symbol of the partial derivative for brevity. Further, we have\nE ∂L i ∂e k = -E j∈N i λ i j Φ p=1 1 l=L 1 d p l e j ∝ -d i E j∈N i Φ j p=1 1 l=L-1 1 d p l e j ≈ -d 3 2 i E Φ j p=1 1 l=L-1 1 d p l e j = -d 3 2 i ϑ ϑ ϑ i .(6)\nFinish the proof. □\nWe visualize the results of d influence. Following the assumption in [19], we posit ∥ϑ ϑ ϑ i ∥ = ϕ for any node i. Conclusion 1: If nodes r and s with d r > d s both can reach node t after L-hop neighborhood aggregation in a graph, we could have\nE ∂L r ∂e t > E ∂L s ∂e t .(7)\nIt suggests that the nodes with higher degrees exert larger influence on L-hop neighbor nodes than lower-degree nodes in the training stage of GCNbased models. In other words, the popular items dominate the updating information of neighbor users. As a result, popular items would make reachable users get closer to them in the representation space.\nBased on the results in the upcoming lemma, we further prove that the popular items tend to have higher probabilities of being recommended by users.\nLemma 2. Suppose that the items r and s could reach the user t by stacking L layers of graph convolution, and d r > d s . After L -1 rounds of graph convolution, the expectation of prediction difference between the two items with regard to the user t is E e ⊤ t (e re s ) . By performing the L-th graph convolution and after the representation of user t updates the gradients ∂L r ∂e t + ∂L s ∂e t , the prediction difference becomes larger, i.e., \nE\nwhere e ′ r and e ′ s are the representations of the items r and s after aggregating the user t, respectively.\nProof. After the influence of the items r and s propagate to the user t by stacking L graph convolution layers, this changes the representation of user t from e t to e t -∂L r ∂e t + ∂L s ∂e t . Assume that the influence\nE ∂L r ∂e t = -υ 1 d 3 2 r ϑ ϑ ϑ r (υ 1 > 0) and ϑ ϑ ϑ r ∥ϑ ϑ ϑ r ∥ = υ 2 e r ∥e r ∥ (υ 2 > 0 for the local homogeneity), likewise, E ∂L s ∂e t = -υ 1 d 3 2\ns ϑ ϑ ϑ s and ϑ ϑ ϑ s ∥ϑ ϑ ϑ s ∥ = υ 2 e s ∥e s ∥ . Now we calculate the prediction difference between the items r and s on the user t,\nE e t - ∂L r ∂e t + ∂L s ∂e t ⊤ (e r -e s ) =E[e ⊤ t (e r -e s )] + υ 1 E d 3 2 r ϑ ϑ ϑ ⊤ r e r -d 3 2 r ϑ ϑ ϑ ⊤ r e s + d 3 2 s ϑ ϑ ϑ ⊤ s e r -d 3 2 s ϑ ϑ ϑ ⊤ s e s =E[e ⊤ t (e r -e s )] + υ 1 υ 2 E d 3 2 r ∥ϑ ϑ ϑ r ∥ ∥e r ∥ e ⊤ r e r -d 3 2 r ∥ϑ ϑ ϑ r ∥ ∥e r ∥ e ⊤ r e s +d 3 2 s ∥ϑ ϑ ϑ s ∥ ∥e s ∥ e ⊤ s e r -d 3 2 s ∥ϑ ϑ ϑ s ∥ ∥e s ∥ e ⊤ s e s =E[e ⊤ z (e r -e s )]\n+ υ 1 υ 2 ϕE d 3 2\nr (∥e r ∥ -\nρ ∥e r ∥ ) -d 3 2 s (∥e s ∥ - ρ ∥e s ∥ ) ,(10)\nwhere ρ = e ⊤ r e s and ρ ≤ ∥e r ∥∥e s ∥. Since the magnitude of the node representation increases as its degree increases [20], for d r > d s , there is ∥e r ∥ > ∥e s ∥ generally. Let ∥e r ∥ = κ∥e s ∥ with κ > 1. Therefore, we have\nd 3 2 r (∥e r ∥ - ρ ∥e r ∥ ) -d 3 2 s (∥e s ∥ - ρ ∥e s ∥ ) =(d 3 2 r κ -d 3 2 s )∥e s ∥ -(d 3 2 r -d 3 2 s κ) ρ ∥e r ∥ ,(11)\nsince\nd 3 2 r κ -d 3 2 s > d 3 2 r -d 3 2\ns κ and ∥e s ∥ ≥ ρ ∥e r ∥ , thus Eq. ( 11) > 0. Based on this, we derive the expression of (8). Furthermore, after the items r and s aggregate the information of the user t, we obtain \ne ′ r =\nNow we calculate the rating difference again,\nE e t - ∂L r ∂e t + ∂L s ∂e t ⊤ (e ′ r -e ′ s ) =E ẽ⊤ t [(e t -e s ) + (ω rt -ω st )ẽ t ] =E ẽ⊤ t (e r -e s ) + (ω rt -ω st )∥ẽ t ∥ 2 ,(14)\nwhen ω rt -ω st ≥ 0, the expression of ( 9) holds. We visualize the average aggregation weight ω of one-order neighbor users for each item groups when training LightGCN in Figure 3. From the results, it can be observed that ω is larger as degree increases generally. Therefore, the expectation of the rating difference would be enlarged for d r > d s after graph convolution. Finish the proof. □"},{"figure_ref":["fig_0"],"heading":"Conclusion 2:","publication_ref":[],"table_ref":[],"text":"The theoretical analyses show that the gap of the prediction scores between popular items and tail items w.r.t. users enlarges with deep layers of graph convolution. It indicates that popular items would become more influential by affecting more high-order neighbor users. As a consequence, popular items are more likely be over-recommended as GCNs go deeper. It reveals how GCN-based models amplify the popularity bias, providing theoretical supports to the phenomenon shown in Figure 1."},{"figure_ref":[],"heading":"Our Method -DAP","publication_ref":["b20"],"table_ref":[],"text":"In this section, we propose our method DAP (Debias the Amplification of Popularity) to alleviate the issue of the popularity bias amplification of GCN in the inference stage. From Lemma 2, the popularity bias amplification comes from the updating of node representation and neighborhood aggregation after graph convolution when training GCN backbone models. Taking LightGCN as an example, we can quantify the bias after graph convolution at each layer in a unified form as\ne (l) v = j∈N v 1 √ d v d j e (l-1) j = j∈N v 1 √ d v d j (ê (l-1) j + α (l-1) H j ϑ ϑ ϑ (l-1) H j + α (l-1) L j ϑ ϑ ϑ (l-1) L j ) = ê(l) v + α (l) H v ϑ ϑ ϑ (l) H v + α (l) L v ϑ ϑ ϑ (l) L v ,(15)\nwhere ê(l)\nv is the ideal representation of the node v at the l-th layer, α (l)\nH v ϑ ϑ ϑ (l) H v and α (l) L v ϑ ϑ ϑ (l)\nL v are the bias that comes from higher-degree and lower-degree neighbors respectively. Specifically, there are two amplification effects: (1) the higher-degree neighbors have large influence and dominate the updating of the target node's representation. It is inclined to make the target node position close to the higher-degree neighbors in the representation space; (2) for the lower-degree neighbors, the representations of them are influenced by the target node, leading to biased learning; after graph convolution, such bias would be further aggregated into the target node. Those two amplification effects are need to be estimated and intervened.\nIn addition, in order to estimate the bias of each node, we employ clustering algorithms to group node representations into clusters. The clustering can automatically discover the highly-influential nodes as they are close to each other in the representation space. For each node, we estimate its amplification effect within its cluster and then intervene the bias in the representations. The specific debiasing process is as follows.\nGiven a well-trained GCN-based backbone model, we could obtain the 0-th layer representations E (0) ∈ R (M+N)×D (D is the embedding size) of all nodes. Then, these representations are fed into the next layer of graph convolution. As we discussed earlier, the bias appears after graph convolution and nodes which in the same cluster are most possible to affect each other. Thus, Kmeans is employed to group nodes in the representation space. For the l-th layer node representations E (l) of all nodes, Kmeans automatically divides them into P clusters, i.e.,\n{C (l) 1 , C (l) 2 , • • • , C (l) P } = Kmeans(E (l) ), (16\n)\nwhere P is a hyper-parameter of Kmeans. For a node v, we can know the cluster it belongs to.\nTo intervene the amplified bias effect b (l) v after lth graph convolution, we have the following strategy: for the node v with degree d v in the cluster C (l) p , we can obtain a set of higher-degree nodes\nH (l) v = { j ∈ C (l) p , d j > d v }\n, and a set of lower-degree nodes\nL (l) v = { j ∈ C (l) p , d j < d v }.\nFor the two parts H (l) v and L (l) v of node v, their pooling (e.g., mean pooling, weighted average pooling with degree) representations after normalization θ ϑ ϑ\n(l)\nH v ∈ R 1×D and θ ϑ ϑ (l)\nL v ∈ R 1×D can be computed respectively. Thereafter, its amplification bias b (l) v after l-th layer graph convolution is estimated by\nb (l) v = αM(e (l) v , θ ϑ ϑ (l) H v ) θ ϑ ϑ (l) H v + βM(e (l) v , θ ϑ ϑ (l) L v ) θ ϑ ϑ (l) L v ,(17)\nwhere α and β are hyper-parameters that intervene the effect of bias in the final representation at the l-th layer since the popularity bias may be not completely harmful [21]. The larger the values of α and β, the greater the bias on the node. M is a similarity calculation function (e.g., cosine similarity) for measuring how strongly the node v is affected by different parts.\nAfter the above operations, we intervene the bias effect and revise the representation of the node v at l-th layer graph convolution, i.e.,\nê(l) v = e (l) v -b (l) v .(18)\nThen the revised representation ê(l) v is fed into the next layer in GCNs and we get the node representations at (l + 1)-th layer E (l+1) . In an iterative manner, we can obtain all the ideal representations at each layer. For all representations rectified at different layers, they are assembled in the manner as the original model does to get the final representations."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":"In this section, we conduct experiments to evaluate the performance of our proposed DAP, aiming at answering the following research questions:\n• RQ1: Does DAP outperform other debiasing methods?\n• RQ2: How do higher-degree part, lower-degree part and the hyper-parameters affect the recommendation performance?\n• RQ3: Can DAP mitigate the popularity bias?"},{"figure_ref":[],"heading":"Experiments Settings","publication_ref":["b5"],"table_ref":["tab_3"],"text":"We conduct experiments on three real-world datasets, Table 2 lists the statistics of three datasets. In order to guarantee a fair comparison, we follow the settings of LightGCN [6] and randomly split the training set and test set. The test set is called Overall. Since our DAP is expected to mitigate the popularity bias and improve the performance on the tail items, we specially split a subset of tail items from the whole test set, named Tail test set, in contrast to the Overall counterpart. In addition, we randomly split 20% data from the training set as the validation set for tuning the hyper-parameters. Note that the same splitting strategy is applied to the validation set. "},{"figure_ref":[],"heading":"Compared Methods","publication_ref":["b21","b13","b22","b23","b24","b9","b9","b8","b12","b25"],"table_ref":[],"text":"To evaluate the debiasing performance on recommendation, we implement our DAP with the GCNbased recommender models LightGCN and Ultra-GCN to explore how our DAP improves the recommendation performance for GCNs. In addition, three methods for solving the popularity bias and two methods for improving tail node representations are compared:\n• BFGCN [22]: This is a novel graph convolution filter for the user-item bipartite network to improve long-tail node representations.\n• UltraGCN [14]: This is a state-of-the-art method that achieves the best performance on the three datasets. It is an ultra-simplified formulation of GCN which skips explicit message passing and directly approximate the limit of infinite graph convolution layers.\n• IPSCN [23]: This method adds max-capping and normalization on IPS [24] value to reduce the variance of IPS. IPS eliminates popularity bias by re-weighting each item according to its popularity.\n• CausE [25]: It requires a large sample of biased data and a small sample of unbiased data. CausE adds a regularizer term on the discrepancy between the item vectors used to fit the biased sample and their counterpart representations that fit the unbiased sample. Because there is no unbiased data in our datasets, we adopt the sampling method in [10] and obtain 20% unbiased data from the training set.\n• DICE [10]: DICE is a method to handle with the popularity bias problem by learning causal embeddings. It is a framework with causal-specific data to disentangle interest and popularity into two sets of embeddings.\n• MACR [9]: This is a state-of-the-art method to eliminate the popularity bias by counterfactual reasoning. It performs counterfactual inference to remove the effect of item popularity.\n• BxQuAD [13]: BxQuAD is a typical post-hoc method for improving tail item recommendations. It suffers from the recommendation accuracy drop for controlling popular items. In this paper, we adopt the Binary-xQuAD method of the original paper and set the hyper-parameter λ = 0.9.\n• Tail [26]: This method learns a neighborhood translation from head nodes, which can be further transferred to tail nodes to enhance their represen- tations. It is devised for node classification and we transfer it to the field of recommendation.\nWe report the all-ranking performance w.r.t. two metrics: Recall and NDCG cut at 20."},{"figure_ref":[],"heading":"Hyper-parameter Settings","publication_ref":["b26"],"table_ref":[],"text":"For a fair comparison, all compared GCN-based models are implemented with 3 layers except for the UltraGCN. We optimize all models with Adam [27] with batch size as 4096. For our method, the number of clusters P is searched in {1, 5, 10, 20, 30, • • • , 70}. Note that we keep the same P in each layer when operating Kmeans. The hyper-parameters α and β in Eq. ( 17) are tuned in the range of [0, 2.0] with step of 0.1."},{"figure_ref":["fig_5","fig_7","fig_7"],"heading":"Recommendation Performance (RQ1)","publication_ref":["b20"],"table_ref":["tab_4","tab_5","tab_4","tab_5","tab_5","tab_4"],"text":"We compare all methods on the Overall and Tail test sets in Tables 3 and4, where the hyper-parameters of DAP-t and DAP-o are tuned to the best on the Tail and Overall validation sets, respectively. The promotions reported in Tables 3 and4 • In all cases, our DAP-o brings performance gain in Recall and NDCG for LightGCN on the Overall and Tail test sets while other baselines only boost LightGCN on the Tail test set. These comparison methods mainly rely on suppressing popular items in exchange for the promotion of tail items. However, our method revises the node representations by intervening the popularity bias based on theoretical analyses on GCNs. It is more applicable to GCN-based backbones.\n• In terms of the performance on the Tail test set, our DAP-t has a significant improvement over LightGCN, the average improvements of LightGCN-DAP-t over LightGCN on the three datasets are 95.71% on Recall and 121.92% on NDCG, respectively. In the same time, the performance on the Overall test set only has a small drop. Although some competitive baselines outperform DAP-t on some metrics, DAP demonstrates stronger comprehensive abilities on different test sets.\n• In order to further verify the effectiveness of our method compared to LightGCN, we show the performance comparison at each layer in Figure 4.\nOverall, it can be seen that DAP boosts LightGCN stably layer by layer. Particularly, on the Amazonbook, when LightGCN degrades the performance as the graph convolution goes deeper, DAP has no accuracy drop. It indicates that the effectiveness of our debiasing method for improving the representations of nodes.\n• For the method UltraGCN, we implement our DAP on it. Because UltraGCN directly uses infinite layers of graph convolution, we only can debias on its final representations. In Table 4, compared with other baselines, our method shows a similar trend to Table 3. It validates the effectiveness of our method. It should be noted that the improvement is relatively small on the Tail test set compared to that on LightGCN. This is mainly because we can not debias at each layer of graph convolution and obtain the most ideal representations.\nTo conclude, DAP can effectively improve the performance of tail items for GCN backbones and out-performs baselines in general. To evaluate the amplification effect brought by higherdegree nodes and lower-degree nodes, we design five variants of DAP implemented on LightGCN including: (1) DAP-kh: this variant is derived by setting β = 0 in Eq. ( 17) which aims to evaluate the amplification effect of higher-degree nodes; (2) DAP-kl: this variant is obtained by setting α = 0 for evaluating the effect of lower-degree nodes; (3) DAP-nh: this variant treats one-order neighbors of the target node as its cluster instead of using Kmeans and sets β = 0 for exploring the effect of higher-degree neighbors; (4) DAP-nl: this variant is the opposite of DAP-nh, using one-order lowerdegree neighbors (i.e., α = 0); (5) DAP-m: it is derived by removing the similarity calculation func- tion M in Eq. ( 17) and is without considering the effect of the lower-degree part (i.e., β = 0).\nFor each variant, we need only tune α or β. For example, we adjust α for DAP-kh from 0 to 1 on the Gowalla dataset. Figure 5 shows the results on the Overall test set and we have the following observations: (1) on the three datasets, DAP-kh achieves the best performance. It reflects that higher-degree nodes have the greater impact on other nodes than lower-degree ones. (2) DAP-kh and DAP-kl have a better performance than DAP-nh and DAP-nl on the three datasets. It indicates that estimating the amplified bias among one-order neighbors is not enough and may introduce noise. Kmeans can automatically help discover the highly-influential nodes. In this way, we would estimate the amplified bias more accurately. Tuning α and β is important for the performance at the inference stage. We plot the performance under different α and β settings on the three datasets in Figure 5. It can be observed that the performance on the Overall test set increases gradually and then decreases while LightGCN is for α = 0 and β = 0. This is to say, the popularity bias is not completely harmful. Directly eliminating all the bias is not reasonable in recommendation, which conforms with the finding of [21]. "},{"figure_ref":["fig_10","fig_5","fig_10"],"heading":"Alleviating Popularity Bias (RQ3)","publication_ref":[],"table_ref":[],"text":"We already discuss that GCNs amplify the popularity bias in the introduction. In this part, we would verify that our debiasing framework can mitigate the amplification of popularity bias issue. The result of TR@20 (the ratio of recommending tail items cut at 20) is shown in Figure 6. As can be seen, when the TR@20 result of LightGCN significantly decreases with more graph convolution layers, DAP could restraint continuous worsening of the popularity bias and gradually improve the TR@20 on the three datasets. It means that the tail items can be recommended more when GCNs go deeper, and thus the popularity bias is effectively alleviated by DAP. In addition, combining Figures 4 and6 gives the conclusion that DAP can not only promote the overall performance, but also can alleviate the popularity bias at each layer."},{"figure_ref":[],"heading":"Related work","publication_ref":[],"table_ref":[],"text":"In this section, we review the research topics related to our work: recommendation debiasing and GCNs debiasing in classification."},{"figure_ref":[],"heading":"Recommendation Debiasing","publication_ref":["b27","b29","b30","b31","b11","b23","b32","b33","b34","b35","b12","b36","b8","b9","b10","b24","b37","b38","b39","b40","b1"],"table_ref":[],"text":"Recommendation debiasing is a recently emerged research topic [?] focusing on various biases in recommendation, for example, popularity bias [28] [29] [30], exposure bias [31], position bias [32], etc. Many methods have been explored to analyze and alleviate the popularity bias in recommender systems. For example, [12,24] propose inverse propensity score (IPS) methods that reweights the interactions to debias the popularity bias in the training loss. However, these methods are difficult to tune because estimating the propensity score is challenging. [33] proposes another method which combines data imputation and IPS to jointly learn an unbiased recommender. However, it is a brute manner for improving the long-tail items with a huge performance drop on the recommendation accuracy. Other empirical approaches such as adversarial learning [34], meta learning [35] and transfer learning [36] are developed without estimating the propensity weight. Ranking adjustment is an another line of research to solve the popularity bias issue [13,37].\nRecently, methods based on causal inference have been widely applied to solve bias issues in recommendation, for example, MACR [9] and DICE [10]. And [11] proposes a causal framework to leverage the popularity bias in recommendation. Other causal methods for learning an unbiased recommender [25,[38][39][40][41] are proposed to takle various bias issues. We refer the readers to a system survey for more details [2]."},{"figure_ref":[],"heading":"GCNs Debiasing in Classification","publication_ref":["b18","b41","b42","b43","b6","b18","b18","b25","b18"],"table_ref":[],"text":"Recently, GCN-based models have demonstrated promising performance and advancements on classification tasks. However, some recent studies also reveal various issues of GCNs including over-smoothing, vulnerability and degree-related biases [19,[42][43][44][45]. The over-smoothing issue can be treated as a kind of bias in GCNs, which means the node representations are inclined to be distinguishable, and degrades the performance with many graph convolution layers. Some methods are proposed to tackle this problem. [46] proposes GCNII which extends the vanilla GCN model with initial residual connection and identity mapping to prevent the oversmoothing. [45] proposes two methods including adding a regularizer to the training objective and optimizing the graph topology based on the model predictions. Nevertheless, [47] argues that the oversmoothing issue only happens after a large number of iterations and the current results with several layers of graph convolution are relatively far from the ideal over-smoothing situation. For recommendation, one node can reach all the other nodes by stacking 7 layers [7]. In this regard, we could treat the nodes clustering issue with shallow graph con-volution layers and over-smoothing with a large number layers differently.\nBesides, [19] points out that GCNs are biased towards higher-degree nodes with higher accuracy than those lower-degree nodes. The authors analyze this issue and argue that nodes with low degrees tend to have very few labeled nodes, which results in sub-optimal performance on low-degree nodes. Therefore, [19] proposes a method exploiting pseudo labels to enhance the representations for low-degree nodes. [48] also proposes a training algorithm adding confident data with virtual labels to the labeled set to enlarge the training set based on self-training. [26] proposes a method which learns a neighborhood translation from head nodes to tail nodes. In this way, the representations of tail nodes can be enhanced for improving the performance of tail nodes.\nIn this work, we analyze the bias existing in GCNbased recommenders that high-degree nodes and low-degree nodes influence each other. It is different from the bias in classification introduced in [19]. And our method is different with these methods based on producing pseudo labels and translation."},{"figure_ref":[],"heading":"Conclusion and future work","publication_ref":[],"table_ref":[],"text":"In this paper, we first theoretically analyzed how GCN-based recommenders amplify the popularity bias. We show that popular items tend to dominate the updating information of neighbor users in the training stage, which makes user position closer to popular items. After multiple times of neighborhood aggregation, popular items would become more influential by affecting more high-order neighbor users. Based on the above insights, we propose a simple yet generic debiasing framework DAP. Our method is implemented in each graph convolution layer in the inference stage by intervening the amplification effect on nodes in the representation space. Extensive experiments on the three real-world datasets justify the effectiveness of DAP. Our method could promote the recommendation performance on tail items and alleviate the popularity bias for GCN backbone models.\nIn the future, we will explore and theoretically analyze more problems hidden in graph-based recommendation methods. In addition, various biases also exist in recommender systems that are harmful to users and need to be solved, such as position bias and exposure bias. It will be meaningful to propose a universal solution to solve various biases. "},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"https://github.com/MEICRS/DAP"}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"\",\n \"publication_ref\": [\n \"b0\",\n \"b0\",\n \"b1\",\n \"b2\",\n \"b3\",\n \"b4\",\n \"b5\",\n \"b6\",\n \"b5\",\n \"b7\",\n \"b8\",\n \"b8\",\n \"b9\",\n \"b10\",\n \"b0\",\n \"b11\",\n \"b8\",\n \"b9\",\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"In this paper, we investigate how graph convolutions amplify the popularity bias in RS. Through theoretical analyses, we identify two fundamental factors: (1) with graph convolution (i.e., neighborhood aggregation), popular items exert larger influence than tail items on neighbor users, making the users move towards popular items in the representation space; (2) after multiple times of graph convolution, popular items would affect more high-order neighbors and become more influential. The two points make popular items get closer to almost users and thus being recommended more frequently. To rectify this, we propose to estimate the amplified effect of popular nodes on each node's representation, and intervene the effect after each graph convolution. Specifically, we adopt clustering to discover highly-influential nodes and estimate the amplification effect of each node, then remove the effect from the node embeddings at each graph convolution layer. Our method is simple and genericit can be used in the inference stage to correct existing models rather than training a new model from scratch, and can be applied to various GCN models. We demonstrate our method on two representative GCN backbones LightGCN and UltraGCN, verifying its ability in improving the recommendations of tail items without sacrificing the performance of 1 Introduction\\nRecommender system (RS) has become a key tool for personalization in today's Web, and also a research hotpot [1,2]. Recently, methods based on GCN have become prevalent in RS since they can effectively encode collaborative filtering signal [3][4][5][6][7]. Though GCNs can improve the overall recommendation accuracy, we find a downside is they amplify the popularity bias -popular items are more frequently recommended than tail items 2) , which means the tail items are not fairly treated by the algorithm, and the issue would be exacerbated in practical RSs due to the user-system feedback loop.\\nFigure 1 provides an evidence that GCNs intensify the popularity bias as the number of graph convolution layers increases. The backbone model is LightGCN [6] using BPR loss [8], a representative and competitive GCN-based recommender. With the increasing of graph convolution layers, the overall accuracy metric (i.e., Recall@20) gradually increases, but the ratio of recommending tail items (i.e., TR@20) drops significantly. This means the improvements of performing graph convolutions is at the expense of tail items, pointing to the amplified popularity bias issue of GCNs [9]. Although there has been recent work in alleviating the popularity bias in RS [9][10][11], they do not reveal how GCN amplifies the bias (i.e., the mechanism) and their solutions are generally applied to RS models rather than tailored for GCNs.\\nIn this work, we aim to answer the unsolved questions \\\"How do graph convolutions amplify popularity bias in RS?\\\" and \\\"How to address the issue with minimum changes on the training progress?\\\" To reveal the mechanism of GCNs amplifying the bias, we conduct theoretical studies by analyzing how popular items influence other nodes. The main findings are twofold: (1) with graph convolution, popular items have a greater impact than tail items on neighbor users, making the users position closer to popular items in the representation space; (2) after multiple times of graph convolution, popular items become more influential by affecting more high-order neighbor users. The theoretical analyses confirm the inherent defect of GCNs in overrecommending popular items.\\nTo alleviate the popularity bias in RS, two lines of research work have been conducted. The first is intervening the training process to eliminate the bias effect. For example, propensity-based methods [12] impose lower weights on popular items in training loss and causal methods [9,10] model the causal effect of the bias on the model training. Those methods need to retrain the backbone model, making the solutions costly to use in practical RSs especially to correct the already-deployed GCN models. In contrast, another line is revising the inference stage in a post-hoc way. For example, [13] performs personalized re-ranking on the generated candidates of the RS model to suppress popular items. However, both lines work on the general issue of popularity bias in recommendation, leaving how GCN model suffers from and amplifies the popularity bias untouched.\\nTowards the research gap, we propose a new problem to solve -rectifying the popularity bias of GCN model in the inference stage. Such a solution can be used to revise already-trained GCNs, thus is easier to deploy in practice than training a new model. Given a GCN model, we first cluster the node representations to automatically discover the highly-influential nodes. Then, the amplification of popularity bias for each node within its cluster is estimated based on the prior theoretical analyses. Thereafter, the amplification effect in the node representation can be intervened to control the bias. This post-hoc method can be easily deployed in practice to correct existing GCN models and promote the recommendations of tail items without sacrificing the performance of popular items. To summarize, this work makes the following contributions:\\n• Providing in-depth theoretical analyses to interpret the popularity bias amplification problem in GCN-based recommenders;\\n• Developing a new method working at each graph convolution layer in the inference stage to correct the popularity bias for GCN;\\n• Conducting extensive experiments on three real datasets to demonstrate the effectiveness of our method on LightGCN and UltraGCN backbones. The degree of user u or item i e (l) u , e (l) i\\nThe embedding of user u or item i at the l-th graph convolution layer L ui\\nThe individual loss term of an interaction (u, i)\\nC (l) p\\nA set of nodes in the p-th cluster obtained by using Kmeans given the embeddings\\nE (l) H (l) v A set of nodes { j ∈ C (l) p , d j > d v |v ∈ C (l) p } L (l) v A set of nodes { j ∈ C (l) p , d j < d v |v ∈ C (l) p } θ ϑ ϑ (l) H v\\nThe pooling representations after normalization of\\nH (l) v θ ϑ ϑ (l) L v\\nThe pooling representations after normalization of L (l) v\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Preliminaries\",\n \"publication_ref\": [\n \"b13\",\n \"b14\",\n \"b5\",\n \"b13\",\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"Suppose that there are a set of users and a set of items\\nU = {u 1 , u 2 , • • • , u M }, I = {i 1 , i 2 , • • • , i N } in a dataset D.\\nLet y ui = 1 be the positive label if the user u has interacted with the item i, otherwise y ui = 0. We can construct a user-item bipartite graph B = (V, E) based on the interaction history, where V consists of the set of user and item nodes, and E denotes the set of edges. If y ui = 1, there is an edge between the user u and the item i.\\nRecently, many studies opt for powerful GCNs to learn user and item node representations [14,15]. Particularly, we introduce LightGCN [6], which is neat and well represents the GCN-based recommenders. One graph convolution block of Light-GCN can be expressed as:\\ne (l) u = i∈N u 1 √ d u √ d i e (l-1) i , e (l) i = u∈N i 1 √ d i √ d u e (l-1) u ,(1)\\nwhere d u (d i ) is the degree of user u (item i) in the graph B, N u (N i ) is the one-order neighbor nodes of the user u (item i), e (0) is an ID embedding of a user or an item. After stacking several graph convolution layers, LightGCN combines the embeddings obtained at each layer to form the final representa-tion e of a node. Thereafter, the model prediction is defined as the inner product of user and item final representations, i.e., ŷui = e ⊤ u e i . Another representative work is UltraGCN [14] which skips infinite layers of graph convolution. We also conduct experiments on this model to verify the generality of our method.\\nTo optimize model parameters, prior work usually frames it as a supervised learning task, and utilize a pointwise or a pairwise loss for model training, which can be summarized by the following formula: L = (u,i)∈D L ui , where L ui is the individual loss term of an interaction (u, i). Without loss of generality, we investigate the popularity bias amplification problem in GCNs based on BCE loss [16] in the next section. The formal formulation of BCE loss is\\nL ui = -[y ui ln σ(ŷ ui ) + (1 -y ui ) ln(1 -σ(ŷ ui ))]. (2)\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"METHODOLOGY\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this part, we attempt to analyze and resolve the amplified popularity bias of GCNs.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Popularity Bias Amplification in GCNs\",\n \"publication_ref\": [\n \"b16\",\n \"b17\",\n \"b18\",\n \"b18\",\n \"b19\"\n ],\n \"table_ref\": [],\n \"text\": \"To understand why GCNs amplify the popularity bias, we conduct theoretical analyses and empirical experiments on GCNs: (1) we start with defining the influence between nodes based on the training loss;\\n(2) we prove that popular items with higher degrees exert larger influence on neighbor users than tail items with lower-degrees; (3) we reveal that popular items commonly have higher probabilities of being recommended by users after representation updating and graph convolution in GCNs.\\nConcisely, we take the e = e (L) at L-th graph convolution layer in Eq. (1) as the final representation of each node. Next, we give the definition of the influence of a user-item pair loss on their neighbors exploiting the concept of influence functions [17][18][19]: Definition 1 (Influence of an observed interaction on a node's representation learning): Suppose that (u, i) is an observed interaction, i.e., there is an edge between node u and node i, some other node k is reachable from node u or i, then the influence of an interaction (u, i) on node k is defined as ∂L ui ∂e k . Without loss of generality, we mainly consider y ui = 1 in BCE loss and have\\n∂L ui ∂e k = - ∂ ln σ(ŷ ui ) ∂e k = - ln σ(ŷ ui ) ∂σ(ŷ ui ) • ∂σ(ŷ ui ) ∂ŷ ui • ∂ŷ ui ∂e k = - 1 σ(ŷ ui ) • σ(ŷ ui )(1 -σ(ŷ ui )) • ∂ŷ ui ∂e k = -[1 -σ(ŷ ui )] ∂ŷ ui ∂e k = -λ ui ∂ŷ ui ∂e k ,(3)\\nwhere ŷui is the prediction between the user u and the item i and 0 < λ ui < 1.\\nDefinition 2 (Influence of a node on another node's representation learning): Suppose that a node i can reach a neighbor node k on a graph 3) . The influence of the loss for the target node i on the node k is defined as ∂L i ∂e k , where L i = j∈N i L i j . In fact, the influence provides a fine-grained scrutiny of updating information of each node at the lens of gradient derivation. We then have the following lemma. Lemma 1. If node i with degree d i can reach node k after stacking L layers of graph convolution, then the influence of node i on node k follows\\nE ∂L i ∂e k ∝ -d where E(•) is the expectation, ϑ ϑ ϑ i = E Φ j p=1 1 l=L-1 1 √ d p l e j\\nrepresents the expectation on paths starting from a neighbor node j of the node i to the node k, Φ j is the set of all (L -1)-length paths from the node j to the node k, p L-1 and p 1 are the node j and the node k, respectively.\\nProof. According to Eq. ( 3) and Eq. ( 1), we obtain\\n∂L i ∂e k = j∈N i ∂L i j ∂e k = - j∈N i λ i j ∂ŷ i j ∂e k = - j∈N i λ i j ∂e i ∂e k e j = - j∈N i λ i j Φ p=1 1 l=L 1 d p l e j ,(5)\\nwhere p L is the node i, Φ is the set of all L-length random paths on the graph from nodes i to k, and we omit the transpose symbol of the partial derivative for brevity. Further, we have\\nE ∂L i ∂e k = -E j∈N i λ i j Φ p=1 1 l=L 1 d p l e j ∝ -d i E j∈N i Φ j p=1 1 l=L-1 1 d p l e j ≈ -d 3 2 i E Φ j p=1 1 l=L-1 1 d p l e j = -d 3 2 i ϑ ϑ ϑ i .(6)\\nFinish the proof. □\\nWe visualize the results of d influence. Following the assumption in [19], we posit ∥ϑ ϑ ϑ i ∥ = ϕ for any node i. Conclusion 1: If nodes r and s with d r > d s both can reach node t after L-hop neighborhood aggregation in a graph, we could have\\nE ∂L r ∂e t > E ∂L s ∂e t .(7)\\nIt suggests that the nodes with higher degrees exert larger influence on L-hop neighbor nodes than lower-degree nodes in the training stage of GCNbased models. In other words, the popular items dominate the updating information of neighbor users. As a result, popular items would make reachable users get closer to them in the representation space.\\nBased on the results in the upcoming lemma, we further prove that the popular items tend to have higher probabilities of being recommended by users.\\nLemma 2. Suppose that the items r and s could reach the user t by stacking L layers of graph convolution, and d r > d s . After L -1 rounds of graph convolution, the expectation of prediction difference between the two items with regard to the user t is E e ⊤ t (e re s ) . By performing the L-th graph convolution and after the representation of user t updates the gradients ∂L r ∂e t + ∂L s ∂e t , the prediction difference becomes larger, i.e., \\nE\\nwhere e ′ r and e ′ s are the representations of the items r and s after aggregating the user t, respectively.\\nProof. After the influence of the items r and s propagate to the user t by stacking L graph convolution layers, this changes the representation of user t from e t to e t -∂L r ∂e t + ∂L s ∂e t . Assume that the influence\\nE ∂L r ∂e t = -υ 1 d 3 2 r ϑ ϑ ϑ r (υ 1 > 0) and ϑ ϑ ϑ r ∥ϑ ϑ ϑ r ∥ = υ 2 e r ∥e r ∥ (υ 2 > 0 for the local homogeneity), likewise, E ∂L s ∂e t = -υ 1 d 3 2\\ns ϑ ϑ ϑ s and ϑ ϑ ϑ s ∥ϑ ϑ ϑ s ∥ = υ 2 e s ∥e s ∥ . Now we calculate the prediction difference between the items r and s on the user t,\\nE e t - ∂L r ∂e t + ∂L s ∂e t ⊤ (e r -e s ) =E[e ⊤ t (e r -e s )] + υ 1 E d 3 2 r ϑ ϑ ϑ ⊤ r e r -d 3 2 r ϑ ϑ ϑ ⊤ r e s + d 3 2 s ϑ ϑ ϑ ⊤ s e r -d 3 2 s ϑ ϑ ϑ ⊤ s e s =E[e ⊤ t (e r -e s )] + υ 1 υ 2 E d 3 2 r ∥ϑ ϑ ϑ r ∥ ∥e r ∥ e ⊤ r e r -d 3 2 r ∥ϑ ϑ ϑ r ∥ ∥e r ∥ e ⊤ r e s +d 3 2 s ∥ϑ ϑ ϑ s ∥ ∥e s ∥ e ⊤ s e r -d 3 2 s ∥ϑ ϑ ϑ s ∥ ∥e s ∥ e ⊤ s e s =E[e ⊤ z (e r -e s )]\\n+ υ 1 υ 2 ϕE d 3 2\\nr (∥e r ∥ -\\nρ ∥e r ∥ ) -d 3 2 s (∥e s ∥ - ρ ∥e s ∥ ) ,(10)\\nwhere ρ = e ⊤ r e s and ρ ≤ ∥e r ∥∥e s ∥. Since the magnitude of the node representation increases as its degree increases [20], for d r > d s , there is ∥e r ∥ > ∥e s ∥ generally. Let ∥e r ∥ = κ∥e s ∥ with κ > 1. Therefore, we have\\nd 3 2 r (∥e r ∥ - ρ ∥e r ∥ ) -d 3 2 s (∥e s ∥ - ρ ∥e s ∥ ) =(d 3 2 r κ -d 3 2 s )∥e s ∥ -(d 3 2 r -d 3 2 s κ) ρ ∥e r ∥ ,(11)\\nsince\\nd 3 2 r κ -d 3 2 s > d 3 2 r -d 3 2\\ns κ and ∥e s ∥ ≥ ρ ∥e r ∥ , thus Eq. ( 11) > 0. Based on this, we derive the expression of (8). Furthermore, after the items r and s aggregate the information of the user t, we obtain \\ne ′ r =\\nNow we calculate the rating difference again,\\nE e t - ∂L r ∂e t + ∂L s ∂e t ⊤ (e ′ r -e ′ s ) =E ẽ⊤ t [(e t -e s ) + (ω rt -ω st )ẽ t ] =E ẽ⊤ t (e r -e s ) + (ω rt -ω st )∥ẽ t ∥ 2 ,(14)\\nwhen ω rt -ω st ≥ 0, the expression of ( 9) holds. We visualize the average aggregation weight ω of one-order neighbor users for each item groups when training LightGCN in Figure 3. From the results, it can be observed that ω is larger as degree increases generally. Therefore, the expectation of the rating difference would be enlarged for d r > d s after graph convolution. Finish the proof. □\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Conclusion 2:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The theoretical analyses show that the gap of the prediction scores between popular items and tail items w.r.t. users enlarges with deep layers of graph convolution. It indicates that popular items would become more influential by affecting more high-order neighbor users. As a consequence, popular items are more likely be over-recommended as GCNs go deeper. It reveals how GCN-based models amplify the popularity bias, providing theoretical supports to the phenomenon shown in Figure 1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Our Method -DAP\",\n \"publication_ref\": [\n \"b20\"\n ],\n \"table_ref\": [],\n \"text\": \"In this section, we propose our method DAP (Debias the Amplification of Popularity) to alleviate the issue of the popularity bias amplification of GCN in the inference stage. From Lemma 2, the popularity bias amplification comes from the updating of node representation and neighborhood aggregation after graph convolution when training GCN backbone models. Taking LightGCN as an example, we can quantify the bias after graph convolution at each layer in a unified form as\\ne (l) v = j∈N v 1 √ d v d j e (l-1) j = j∈N v 1 √ d v d j (ê (l-1) j + α (l-1) H j ϑ ϑ ϑ (l-1) H j + α (l-1) L j ϑ ϑ ϑ (l-1) L j ) = ê(l) v + α (l) H v ϑ ϑ ϑ (l) H v + α (l) L v ϑ ϑ ϑ (l) L v ,(15)\\nwhere ê(l)\\nv is the ideal representation of the node v at the l-th layer, α (l)\\nH v ϑ ϑ ϑ (l) H v and α (l) L v ϑ ϑ ϑ (l)\\nL v are the bias that comes from higher-degree and lower-degree neighbors respectively. Specifically, there are two amplification effects: (1) the higher-degree neighbors have large influence and dominate the updating of the target node's representation. It is inclined to make the target node position close to the higher-degree neighbors in the representation space; (2) for the lower-degree neighbors, the representations of them are influenced by the target node, leading to biased learning; after graph convolution, such bias would be further aggregated into the target node. Those two amplification effects are need to be estimated and intervened.\\nIn addition, in order to estimate the bias of each node, we employ clustering algorithms to group node representations into clusters. The clustering can automatically discover the highly-influential nodes as they are close to each other in the representation space. For each node, we estimate its amplification effect within its cluster and then intervene the bias in the representations. The specific debiasing process is as follows.\\nGiven a well-trained GCN-based backbone model, we could obtain the 0-th layer representations E (0) ∈ R (M+N)×D (D is the embedding size) of all nodes. Then, these representations are fed into the next layer of graph convolution. As we discussed earlier, the bias appears after graph convolution and nodes which in the same cluster are most possible to affect each other. Thus, Kmeans is employed to group nodes in the representation space. For the l-th layer node representations E (l) of all nodes, Kmeans automatically divides them into P clusters, i.e.,\\n{C (l) 1 , C (l) 2 , • • • , C (l) P } = Kmeans(E (l) ), (16\\n)\\nwhere P is a hyper-parameter of Kmeans. For a node v, we can know the cluster it belongs to.\\nTo intervene the amplified bias effect b (l) v after lth graph convolution, we have the following strategy: for the node v with degree d v in the cluster C (l) p , we can obtain a set of higher-degree nodes\\nH (l) v = { j ∈ C (l) p , d j > d v }\\n, and a set of lower-degree nodes\\nL (l) v = { j ∈ C (l) p , d j < d v }.\\nFor the two parts H (l) v and L (l) v of node v, their pooling (e.g., mean pooling, weighted average pooling with degree) representations after normalization θ ϑ ϑ\\n(l)\\nH v ∈ R 1×D and θ ϑ ϑ (l)\\nL v ∈ R 1×D can be computed respectively. Thereafter, its amplification bias b (l) v after l-th layer graph convolution is estimated by\\nb (l) v = αM(e (l) v , θ ϑ ϑ (l) H v ) θ ϑ ϑ (l) H v + βM(e (l) v , θ ϑ ϑ (l) L v ) θ ϑ ϑ (l) L v ,(17)\\nwhere α and β are hyper-parameters that intervene the effect of bias in the final representation at the l-th layer since the popularity bias may be not completely harmful [21]. The larger the values of α and β, the greater the bias on the node. M is a similarity calculation function (e.g., cosine similarity) for measuring how strongly the node v is affected by different parts.\\nAfter the above operations, we intervene the bias effect and revise the representation of the node v at l-th layer graph convolution, i.e.,\\nê(l) v = e (l) v -b (l) v .(18)\\nThen the revised representation ê(l) v is fed into the next layer in GCNs and we get the node representations at (l + 1)-th layer E (l+1) . In an iterative manner, we can obtain all the ideal representations at each layer. For all representations rectified at different layers, they are assembled in the manner as the original model does to get the final representations.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we conduct experiments to evaluate the performance of our proposed DAP, aiming at answering the following research questions:\\n• RQ1: Does DAP outperform other debiasing methods?\\n• RQ2: How do higher-degree part, lower-degree part and the hyper-parameters affect the recommendation performance?\\n• RQ3: Can DAP mitigate the popularity bias?\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments Settings\",\n \"publication_ref\": [\n \"b5\"\n ],\n \"table_ref\": [\n \"tab_3\"\n ],\n \"text\": \"We conduct experiments on three real-world datasets, Table 2 lists the statistics of three datasets. In order to guarantee a fair comparison, we follow the settings of LightGCN [6] and randomly split the training set and test set. The test set is called Overall. Since our DAP is expected to mitigate the popularity bias and improve the performance on the tail items, we specially split a subset of tail items from the whole test set, named Tail test set, in contrast to the Overall counterpart. In addition, we randomly split 20% data from the training set as the validation set for tuning the hyper-parameters. Note that the same splitting strategy is applied to the validation set. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Compared Methods\",\n \"publication_ref\": [\n \"b21\",\n \"b13\",\n \"b22\",\n \"b23\",\n \"b24\",\n \"b9\",\n \"b9\",\n \"b8\",\n \"b12\",\n \"b25\"\n ],\n \"table_ref\": [],\n \"text\": \"To evaluate the debiasing performance on recommendation, we implement our DAP with the GCNbased recommender models LightGCN and Ultra-GCN to explore how our DAP improves the recommendation performance for GCNs. In addition, three methods for solving the popularity bias and two methods for improving tail node representations are compared:\\n• BFGCN [22]: This is a novel graph convolution filter for the user-item bipartite network to improve long-tail node representations.\\n• UltraGCN [14]: This is a state-of-the-art method that achieves the best performance on the three datasets. It is an ultra-simplified formulation of GCN which skips explicit message passing and directly approximate the limit of infinite graph convolution layers.\\n• IPSCN [23]: This method adds max-capping and normalization on IPS [24] value to reduce the variance of IPS. IPS eliminates popularity bias by re-weighting each item according to its popularity.\\n• CausE [25]: It requires a large sample of biased data and a small sample of unbiased data. CausE adds a regularizer term on the discrepancy between the item vectors used to fit the biased sample and their counterpart representations that fit the unbiased sample. Because there is no unbiased data in our datasets, we adopt the sampling method in [10] and obtain 20% unbiased data from the training set.\\n• DICE [10]: DICE is a method to handle with the popularity bias problem by learning causal embeddings. It is a framework with causal-specific data to disentangle interest and popularity into two sets of embeddings.\\n• MACR [9]: This is a state-of-the-art method to eliminate the popularity bias by counterfactual reasoning. It performs counterfactual inference to remove the effect of item popularity.\\n• BxQuAD [13]: BxQuAD is a typical post-hoc method for improving tail item recommendations. It suffers from the recommendation accuracy drop for controlling popular items. In this paper, we adopt the Binary-xQuAD method of the original paper and set the hyper-parameter λ = 0.9.\\n• Tail [26]: This method learns a neighborhood translation from head nodes, which can be further transferred to tail nodes to enhance their represen- tations. It is devised for node classification and we transfer it to the field of recommendation.\\nWe report the all-ranking performance w.r.t. two metrics: Recall and NDCG cut at 20.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Hyper-parameter Settings\",\n \"publication_ref\": [\n \"b26\"\n ],\n \"table_ref\": [],\n \"text\": \"For a fair comparison, all compared GCN-based models are implemented with 3 layers except for the UltraGCN. We optimize all models with Adam [27] with batch size as 4096. For our method, the number of clusters P is searched in {1, 5, 10, 20, 30, • • • , 70}. Note that we keep the same P in each layer when operating Kmeans. The hyper-parameters α and β in Eq. ( 17) are tuned in the range of [0, 2.0] with step of 0.1.\"\n },\n {\n \"figure_ref\": [\n \"fig_5\",\n \"fig_7\",\n \"fig_7\"\n ],\n \"heading\": \"Recommendation Performance (RQ1)\",\n \"publication_ref\": [\n \"b20\"\n ],\n \"table_ref\": [\n \"tab_4\",\n \"tab_5\",\n \"tab_4\",\n \"tab_5\",\n \"tab_5\",\n \"tab_4\"\n ],\n \"text\": \"We compare all methods on the Overall and Tail test sets in Tables 3 and4, where the hyper-parameters of DAP-t and DAP-o are tuned to the best on the Tail and Overall validation sets, respectively. The promotions reported in Tables 3 and4 • In all cases, our DAP-o brings performance gain in Recall and NDCG for LightGCN on the Overall and Tail test sets while other baselines only boost LightGCN on the Tail test set. These comparison methods mainly rely on suppressing popular items in exchange for the promotion of tail items. However, our method revises the node representations by intervening the popularity bias based on theoretical analyses on GCNs. It is more applicable to GCN-based backbones.\\n• In terms of the performance on the Tail test set, our DAP-t has a significant improvement over LightGCN, the average improvements of LightGCN-DAP-t over LightGCN on the three datasets are 95.71% on Recall and 121.92% on NDCG, respectively. In the same time, the performance on the Overall test set only has a small drop. Although some competitive baselines outperform DAP-t on some metrics, DAP demonstrates stronger comprehensive abilities on different test sets.\\n• In order to further verify the effectiveness of our method compared to LightGCN, we show the performance comparison at each layer in Figure 4.\\nOverall, it can be seen that DAP boosts LightGCN stably layer by layer. Particularly, on the Amazonbook, when LightGCN degrades the performance as the graph convolution goes deeper, DAP has no accuracy drop. It indicates that the effectiveness of our debiasing method for improving the representations of nodes.\\n• For the method UltraGCN, we implement our DAP on it. Because UltraGCN directly uses infinite layers of graph convolution, we only can debias on its final representations. In Table 4, compared with other baselines, our method shows a similar trend to Table 3. It validates the effectiveness of our method. It should be noted that the improvement is relatively small on the Tail test set compared to that on LightGCN. This is mainly because we can not debias at each layer of graph convolution and obtain the most ideal representations.\\nTo conclude, DAP can effectively improve the performance of tail items for GCN backbones and out-performs baselines in general. To evaluate the amplification effect brought by higherdegree nodes and lower-degree nodes, we design five variants of DAP implemented on LightGCN including: (1) DAP-kh: this variant is derived by setting β = 0 in Eq. ( 17) which aims to evaluate the amplification effect of higher-degree nodes; (2) DAP-kl: this variant is obtained by setting α = 0 for evaluating the effect of lower-degree nodes; (3) DAP-nh: this variant treats one-order neighbors of the target node as its cluster instead of using Kmeans and sets β = 0 for exploring the effect of higher-degree neighbors; (4) DAP-nl: this variant is the opposite of DAP-nh, using one-order lowerdegree neighbors (i.e., α = 0); (5) DAP-m: it is derived by removing the similarity calculation func- tion M in Eq. ( 17) and is without considering the effect of the lower-degree part (i.e., β = 0).\\nFor each variant, we need only tune α or β. For example, we adjust α for DAP-kh from 0 to 1 on the Gowalla dataset. Figure 5 shows the results on the Overall test set and we have the following observations: (1) on the three datasets, DAP-kh achieves the best performance. It reflects that higher-degree nodes have the greater impact on other nodes than lower-degree ones. (2) DAP-kh and DAP-kl have a better performance than DAP-nh and DAP-nl on the three datasets. It indicates that estimating the amplified bias among one-order neighbors is not enough and may introduce noise. Kmeans can automatically help discover the highly-influential nodes. In this way, we would estimate the amplified bias more accurately. Tuning α and β is important for the performance at the inference stage. We plot the performance under different α and β settings on the three datasets in Figure 5. It can be observed that the performance on the Overall test set increases gradually and then decreases while LightGCN is for α = 0 and β = 0. This is to say, the popularity bias is not completely harmful. Directly eliminating all the bias is not reasonable in recommendation, which conforms with the finding of [21]. \"\n },\n {\n \"figure_ref\": [\n \"fig_10\",\n \"fig_5\",\n \"fig_10\"\n ],\n \"heading\": \"Alleviating Popularity Bias (RQ3)\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We already discuss that GCNs amplify the popularity bias in the introduction. In this part, we would verify that our debiasing framework can mitigate the amplification of popularity bias issue. The result of TR@20 (the ratio of recommending tail items cut at 20) is shown in Figure 6. As can be seen, when the TR@20 result of LightGCN significantly decreases with more graph convolution layers, DAP could restraint continuous worsening of the popularity bias and gradually improve the TR@20 on the three datasets. It means that the tail items can be recommended more when GCNs go deeper, and thus the popularity bias is effectively alleviated by DAP. In addition, combining Figures 4 and6 gives the conclusion that DAP can not only promote the overall performance, but also can alleviate the popularity bias at each layer.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related work\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we review the research topics related to our work: recommendation debiasing and GCNs debiasing in classification.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Recommendation Debiasing\",\n \"publication_ref\": [\n \"b27\",\n \"b29\",\n \"b30\",\n \"b31\",\n \"b11\",\n \"b23\",\n \"b32\",\n \"b33\",\n \"b34\",\n \"b35\",\n \"b12\",\n \"b36\",\n \"b8\",\n \"b9\",\n \"b10\",\n \"b24\",\n \"b37\",\n \"b38\",\n \"b39\",\n \"b40\",\n \"b1\"\n ],\n \"table_ref\": [],\n \"text\": \"Recommendation debiasing is a recently emerged research topic [?] focusing on various biases in recommendation, for example, popularity bias [28] [29] [30], exposure bias [31], position bias [32], etc. Many methods have been explored to analyze and alleviate the popularity bias in recommender systems. For example, [12,24] propose inverse propensity score (IPS) methods that reweights the interactions to debias the popularity bias in the training loss. However, these methods are difficult to tune because estimating the propensity score is challenging. [33] proposes another method which combines data imputation and IPS to jointly learn an unbiased recommender. However, it is a brute manner for improving the long-tail items with a huge performance drop on the recommendation accuracy. Other empirical approaches such as adversarial learning [34], meta learning [35] and transfer learning [36] are developed without estimating the propensity weight. Ranking adjustment is an another line of research to solve the popularity bias issue [13,37].\\nRecently, methods based on causal inference have been widely applied to solve bias issues in recommendation, for example, MACR [9] and DICE [10]. And [11] proposes a causal framework to leverage the popularity bias in recommendation. Other causal methods for learning an unbiased recommender [25,[38][39][40][41] are proposed to takle various bias issues. We refer the readers to a system survey for more details [2].\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GCNs Debiasing in Classification\",\n \"publication_ref\": [\n \"b18\",\n \"b41\",\n \"b42\",\n \"b43\",\n \"b6\",\n \"b18\",\n \"b18\",\n \"b25\",\n \"b18\"\n ],\n \"table_ref\": [],\n \"text\": \"Recently, GCN-based models have demonstrated promising performance and advancements on classification tasks. However, some recent studies also reveal various issues of GCNs including over-smoothing, vulnerability and degree-related biases [19,[42][43][44][45]. The over-smoothing issue can be treated as a kind of bias in GCNs, which means the node representations are inclined to be distinguishable, and degrades the performance with many graph convolution layers. Some methods are proposed to tackle this problem. [46] proposes GCNII which extends the vanilla GCN model with initial residual connection and identity mapping to prevent the oversmoothing. [45] proposes two methods including adding a regularizer to the training objective and optimizing the graph topology based on the model predictions. Nevertheless, [47] argues that the oversmoothing issue only happens after a large number of iterations and the current results with several layers of graph convolution are relatively far from the ideal over-smoothing situation. For recommendation, one node can reach all the other nodes by stacking 7 layers [7]. In this regard, we could treat the nodes clustering issue with shallow graph con-volution layers and over-smoothing with a large number layers differently.\\nBesides, [19] points out that GCNs are biased towards higher-degree nodes with higher accuracy than those lower-degree nodes. The authors analyze this issue and argue that nodes with low degrees tend to have very few labeled nodes, which results in sub-optimal performance on low-degree nodes. Therefore, [19] proposes a method exploiting pseudo labels to enhance the representations for low-degree nodes. [48] also proposes a training algorithm adding confident data with virtual labels to the labeled set to enlarge the training set based on self-training. [26] proposes a method which learns a neighborhood translation from head nodes to tail nodes. In this way, the representations of tail nodes can be enhanced for improving the performance of tail nodes.\\nIn this work, we analyze the bias existing in GCNbased recommenders that high-degree nodes and low-degree nodes influence each other. It is different from the bias in classification introduced in [19]. And our method is different with these methods based on producing pseudo labels and translation.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion and future work\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we first theoretically analyzed how GCN-based recommenders amplify the popularity bias. We show that popular items tend to dominate the updating information of neighbor users in the training stage, which makes user position closer to popular items. After multiple times of neighborhood aggregation, popular items would become more influential by affecting more high-order neighbor users. Based on the above insights, we propose a simple yet generic debiasing framework DAP. Our method is implemented in each graph convolution layer in the inference stage by intervening the amplification effect on nodes in the representation space. Extensive experiments on the three real-world datasets justify the effectiveness of DAP. Our method could promote the recommendation performance on tail items and alleviate the popularity bias for GCN backbone models.\\nIn the future, we will explore and theoretically analyze more problems hidden in graph-based recommendation methods. In addition, various biases also exist in recommender systems that are harmful to users and need to be solved, such as position bias and exposure bias. It will be meaningful to propose a universal solution to solve various biases. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"https://github.com/MEICRS/DAP\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-05-24"},"doi":{"kind":"string","value":"10.1007/sxxxxx-yyy-zzzz-1"},"references":{"kind":"list like","value":[{"authors":"L Wu; X He; X Wang; K Zhang; M Wang","journal":"IEEE Transactions on Knowledge and Data Engineering","ref_id":"b0","title":"A survey on accuracy-oriented neural recommendation: From collaborative filtering to information-rich recommendation","year":"2022"},{"authors":"J Chen; H Dong; X Wang; F Feng; M Wang; X He","journal":"ACM Transactions on Information Systems","ref_id":"b1","title":"Bias and debias in recommender system: A survey and future directions","year":"2023"},{"authors":"X Wang; X He; M Wang; F Feng; T Chua","journal":"","ref_id":"b2","title":"Neural graph collaborative filtering","year":"2019"},{"authors":"L Wu; P Sun; Y Fu; R Hong; X Wang; M Wang","journal":"","ref_id":"b3","title":"A neural influence diffusion model for social recommendation","year":"2019"},{"authors":"F Wu; A Souza; T Zhang; C Fifty; T Yu; K Weinberger","journal":"","ref_id":"b4","title":"Simplifying graph convolutional networks","year":"2019"},{"authors":"X He; K Deng; X Wang; Y Li; Y Zhang; M Wang","journal":"SIGIR","ref_id":"b5","title":"Lightgcn: Simplifying and powering graph convolution network for recommendation","year":"2020"},{"authors":"F Liu; Z Cheng; L Zhu; Z Gao; L Nie","journal":"","ref_id":"b6","title":"Interest-aware message-passing gcn for recommendation","year":"2021"},{"authors":"S Rendle; C Freudenthaler; Z Gantner; L Schmidt-Thieme","journal":"","ref_id":"b7","title":"Bpr: Bayesian personalized ranking from implicit feedback","year":"2009"},{"authors":"T Wei; F Feng; J Chen; Z Wu; J Yi; X He","journal":"","ref_id":"b8","title":"Model-agnostic counterfactual reasoning for eliminating popularity bias in recommender system","year":"2021"},{"authors":"Y Zheng; C Gao; X Li; X He; Y Li; Jin D ","journal":"","ref_id":"b9","title":"Disentangling user interest and conformity for recommendation with causal embedding","year":"2021"},{"authors":"Y Zhang; F Feng; X He; T Wei; C Song; G Ling; Y Zhang","journal":"","ref_id":"b10","title":"Causal intervention for leveraging popularity bias in recommendation","year":"2021"},{"authors":"D Liang; L Charlin; D Blei","journal":"","ref_id":"b11","title":"Causal inference for recommendation","year":"2016"},{"authors":"H Abdollahpouri; R Burke; B Mobasher","journal":"","ref_id":"b12","title":"Managing popularity bias in recommender systems with personalized re-ranking","year":"2019"},{"authors":"K Mao; J Zhu; Xiao X Lu; B Wang; Z He; X ","journal":"","ref_id":"b13","title":"Ultragcn: ultra simplification of graph convolutional networks for recommendation","year":"2021"},{"authors":"T N Kipf; M Welling","journal":"","ref_id":"b14","title":"Semi-supervised classification with graph convolutional networks","year":"2017"},{"authors":"C Johnson","journal":"Advances in Neural Information Processing Systems","ref_id":"b15","title":"Logistic matrix factorization for implicit feedback data","year":"2014"},{"authors":"P W Koh; P Liang","journal":"","ref_id":"b16","title":"Understanding black-box predictions via influence functions","year":"2017"},{"authors":"K Xu; C Li; Y Tian; T Sonobe; K Kawarabayashi; S Jegelka","journal":"","ref_id":"b17","title":"Representation learning on graphs with jumping knowledge networks","year":"2018"},{"authors":"X Tang; H Yao; Y Sun; Y Wang; J Tang; C Aggarwal; P Mitra; S Wang","journal":"","ref_id":"b18","title":"Investigating and mitigating degreerelated biases in graph convoltuional networks","year":"2020"},{"authors":"J Wu; X Wang; X Gao; J Chen; H Fu; T Qiu; X He","journal":"","ref_id":"b19","title":"On the effectiveness of sampled softmax loss for item recommendation","year":"2022"},{"authors":"Z Zhao; J Chen; S Zhou; X He; X Cao; F Zhang; W Wu","journal":"IEEE Transactions on Knowledge & Data Engineering","ref_id":"b20","title":"Popularity bias is not always evil: Disentangling benign and harmful bias for recommendation","year":"2022"},{"authors":"M Zhao; Q Deng; K Wang; R Wu; J Tao; C Fan; L Chen; P Cui","journal":"ACM Transactions on Information Systems (TOIS)","ref_id":"b21","title":"Bilateral filtering graph convolutional network for multi-relational social recommendation in the power-law networks","year":"2021"},{"authors":"A Gruson; P Chandar; C Charbuillet; J Mcinerney; S Hansen; D Tardieu; B Carterette","journal":"","ref_id":"b22","title":"Offline evaluation to make decisions about playlistrecommendation algorithms","year":"2019"},{"authors":"T Schnabel; A Swaminathan; A Singh; N Chandak; T Joachims","journal":"","ref_id":"b23","title":"Recommendations as treatments: Debiasing learning and evaluation","year":"2016"},{"authors":"S Bonner; F Vasile","journal":"","ref_id":"b24","title":"Causal embeddings for recommendation","year":"2018"},{"authors":"Z Liu; T K Nguyen; Y Fang","journal":"","ref_id":"b25","title":"Tail-gnn: Tail-node graph neural networks","year":"2021"},{"authors":"D P Kingma; J Ba","journal":"","ref_id":"b26","title":"Adam: A method for stochastic optimization","year":"2015"},{"authors":"Z Zhu; Y He; X Zhao; J Caverlee","journal":"","ref_id":"b27","title":"Popularity bias in dynamic recommendation","year":"2021"},{"authors":"G Xv; C Lin; H Li; J Su; W Ye; Y Chen","journal":"","ref_id":"b28","title":"Neutralizing popularity bias in recommendation models","year":"2022"},{"authors":"P Gupta; A Sharma; P Malhotra; L Vig; G Shroff","journal":"","ref_id":"b29","title":"Causer: Causal session-based recommendations for handling popularity bias","year":"2021"},{"authors":"S Gupta; H Wang; Z Lipton; Y Wang","journal":"","ref_id":"b30","title":"Correcting exposure bias for link recommendation","year":"2021"},{"authors":"H Guo; J Yu; Q Liu; R Tang; Y Zhang","journal":"","ref_id":"b31","title":"Pal: a positionbias aware learning framework for ctr prediction in live recommender systems","year":"2019"},{"authors":"M Dudík; J Langford; L Li","journal":"","ref_id":"b32","title":"Doubly robust policy evaluation and learning","year":"2011"},{"authors":"A Krishnan; A Sharma; A Sankar; H Sundaram","journal":"","ref_id":"b33","title":"An adversarial approach to improve long-tail performance in neural collaborative filtering","year":"2018"},{"authors":"J Chen; H Dong; Y Qiu; X He; X Xin; L Chen; G Lin; K Yang","journal":"","ref_id":"b34","title":"Autodebias: Learning to debias for recommendation","year":"2021"},{"authors":"Y Zhang; D Z Cheng; T Yao; X Yi; L Hong; Chi E ","journal":"","ref_id":"b35","title":"A model of two tales: Dual transfer learning framework for improved long-tail item recommendation","year":"2021"},{"authors":"H Abdollahpouri; R Burke; B Mobasher","journal":"","ref_id":"b36","title":"Controlling popularity bias in learning-to-rank recommendation","year":"2017"},{"authors":"Y Wang; D Liang; L Charlin; D Blei","journal":"CoRR","ref_id":"b37","title":"The deconfounded recommender: A causal inference approach to recommendation","year":""},{"authors":"M Sato; S Takemori; J Singh; T Ohkuma","journal":"","ref_id":"b38","title":"Unbiased learning for the causal effect of recommendation","year":"2020"},{"authors":"R Qiu; S Wang; Z Chen; H Yin; Z Huang","journal":"","ref_id":"b39","title":"Causalrec: Causal inference for visual debiasing in visually-aware recommendation","year":"2021"},{"authors":"W Wang; F Feng; X He; X Wang; T Chua","journal":"","ref_id":"b40","title":"Deconfounded recommendation for alleviating bias amplification","year":"2021"},{"authors":"H Dai; H Li; T Tian; X Huang; L Wang; J Zhu; L Song","journal":"","ref_id":"b41","title":"Adversarial attack on graph structured data","year":"2018"},{"authors":"K Xu; W Hu; J Leskovec; S Jegelka","journal":"","ref_id":"b42","title":"How powerful are graph neural networks?","year":"2018"},{"authors":"Q Li; Z Han; X Wu","journal":"","ref_id":"b43","title":"Deeper insights into graph convolutional networks for semi-supervised learning","year":""}],"string":"[\n {\n \"authors\": \"L Wu; X He; X Wang; K Zhang; M Wang\",\n \"journal\": \"IEEE Transactions on Knowledge and Data Engineering\",\n \"ref_id\": \"b0\",\n \"title\": \"A survey on accuracy-oriented neural recommendation: From collaborative filtering to information-rich recommendation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"J Chen; H Dong; X Wang; F Feng; M Wang; X He\",\n \"journal\": \"ACM Transactions on Information Systems\",\n \"ref_id\": \"b1\",\n \"title\": \"Bias and debias in recommender system: A survey and future directions\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"X Wang; X He; M Wang; F Feng; T Chua\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Neural graph collaborative filtering\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"L Wu; P Sun; Y Fu; R Hong; X Wang; M Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"A neural influence diffusion model for social recommendation\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"F Wu; A Souza; T Zhang; C Fifty; T Yu; K Weinberger\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Simplifying graph convolutional networks\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"X He; K Deng; X Wang; Y Li; Y Zhang; M Wang\",\n \"journal\": \"SIGIR\",\n \"ref_id\": \"b5\",\n \"title\": \"Lightgcn: Simplifying and powering graph convolution network for recommendation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"F Liu; Z Cheng; L Zhu; Z Gao; L Nie\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Interest-aware message-passing gcn for recommendation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"S Rendle; C Freudenthaler; Z Gantner; L Schmidt-Thieme\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Bpr: Bayesian personalized ranking from implicit feedback\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"T Wei; F Feng; J Chen; Z Wu; J Yi; X He\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Model-agnostic counterfactual reasoning for eliminating popularity bias in recommender system\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Y Zheng; C Gao; X Li; X He; Y Li; Jin D \",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Disentangling user interest and conformity for recommendation with causal embedding\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Y Zhang; F Feng; X He; T Wei; C Song; G Ling; Y Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Causal intervention for leveraging popularity bias in recommendation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"D Liang; L Charlin; D Blei\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Causal inference for recommendation\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"H Abdollahpouri; R Burke; B Mobasher\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Managing popularity bias in recommender systems with personalized re-ranking\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"K Mao; J Zhu; Xiao X Lu; B Wang; Z He; X \",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Ultragcn: ultra simplification of graph convolutional networks for recommendation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"T N Kipf; M Welling\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Semi-supervised classification with graph convolutional networks\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"C Johnson\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b15\",\n \"title\": \"Logistic matrix factorization for implicit feedback data\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"P W Koh; P Liang\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Understanding black-box predictions via influence functions\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"K Xu; C Li; Y Tian; T Sonobe; K Kawarabayashi; S Jegelka\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Representation learning on graphs with jumping knowledge networks\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"X Tang; H Yao; Y Sun; Y Wang; J Tang; C Aggarwal; P Mitra; S Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Investigating and mitigating degreerelated biases in graph convoltuional networks\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"J Wu; X Wang; X Gao; J Chen; H Fu; T Qiu; X He\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"On the effectiveness of sampled softmax loss for item recommendation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Z Zhao; J Chen; S Zhou; X He; X Cao; F Zhang; W Wu\",\n \"journal\": \"IEEE Transactions on Knowledge & Data Engineering\",\n \"ref_id\": \"b20\",\n \"title\": \"Popularity bias is not always evil: Disentangling benign and harmful bias for recommendation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"M Zhao; Q Deng; K Wang; R Wu; J Tao; C Fan; L Chen; P Cui\",\n \"journal\": \"ACM Transactions on Information Systems (TOIS)\",\n \"ref_id\": \"b21\",\n \"title\": \"Bilateral filtering graph convolutional network for multi-relational social recommendation in the power-law networks\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"A Gruson; P Chandar; C Charbuillet; J Mcinerney; S Hansen; D Tardieu; B Carterette\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Offline evaluation to make decisions about playlistrecommendation algorithms\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"T Schnabel; A Swaminathan; A Singh; N Chandak; T Joachims\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Recommendations as treatments: Debiasing learning and evaluation\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"S Bonner; F Vasile\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Causal embeddings for recommendation\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Z Liu; T K Nguyen; Y Fang\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Tail-gnn: Tail-node graph neural networks\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"D P Kingma; J Ba\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Adam: A method for stochastic optimization\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Z Zhu; Y He; X Zhao; J Caverlee\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Popularity bias in dynamic recommendation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"G Xv; C Lin; H Li; J Su; W Ye; Y Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Neutralizing popularity bias in recommendation models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"P Gupta; A Sharma; P Malhotra; L Vig; G Shroff\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Causer: Causal session-based recommendations for handling popularity bias\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"S Gupta; H Wang; Z Lipton; Y Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Correcting exposure bias for link recommendation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"H Guo; J Yu; Q Liu; R Tang; Y Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Pal: a positionbias aware learning framework for ctr prediction in live recommender systems\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"M Dudík; J Langford; L Li\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Doubly robust policy evaluation and learning\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"A Krishnan; A Sharma; A Sankar; H Sundaram\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"An adversarial approach to improve long-tail performance in neural collaborative filtering\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"J Chen; H Dong; Y Qiu; X He; X Xin; L Chen; G Lin; K Yang\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Autodebias: Learning to debias for recommendation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Y Zhang; D Z Cheng; T Yao; X Yi; L Hong; Chi E \",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"A model of two tales: Dual transfer learning framework for improved long-tail item recommendation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"H Abdollahpouri; R Burke; B Mobasher\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Controlling popularity bias in learning-to-rank recommendation\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Y Wang; D Liang; L Charlin; D Blei\",\n \"journal\": \"CoRR\",\n \"ref_id\": \"b37\",\n \"title\": \"The deconfounded recommender: A causal inference approach to recommendation\",\n \"year\": \"\"\n },\n {\n \"authors\": \"M Sato; S Takemori; J Singh; T Ohkuma\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Unbiased learning for the causal effect of recommendation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"R Qiu; S Wang; Z Chen; H Yin; Z Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Causalrec: Causal inference for visual debiasing in visually-aware recommendation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"W Wang; F Feng; X He; X Wang; T Chua\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Deconfounded recommendation for alleviating bias amplification\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"H Dai; H Li; T Tian; X Huang; L Wang; J Zhu; L Song\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Adversarial attack on graph structured data\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"K Xu; W Hu; J Leskovec; S Jegelka\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"How powerful are graph neural networks?\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Q Li; Z Han; X Wu\",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"Deeper insights into graph convolutional networks for semi-supervised learning\",\n \"year\": \"\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,310.7,161.7,13.05,12.7],"formula_id":"formula_0","formula_text":"C (l) p"},{"formula_coordinates":[3,310.48,169.76,202.57,58.99],"formula_id":"formula_1","formula_text":"E (l) H (l) v A set of nodes { j ∈ C (l) p , d j > d v |v ∈ C (l) p } L (l) v A set of nodes { j ∈ C (l) p , d j < d v |v ∈ C (l) p } θ ϑ ϑ (l) H v"},{"formula_coordinates":[3,310.48,221.81,65.15,30.4],"formula_id":"formula_2","formula_text":"H (l) v θ ϑ ϑ (l) L v"},{"formula_coordinates":[3,304.72,325.99,242.36,28.91],"formula_id":"formula_3","formula_text":"U = {u 1 , u 2 , • • • , u M }, I = {i 1 , i 2 , • • • , i N } in a dataset D."},{"formula_coordinates":[3,304.89,574.7,243,43.9],"formula_id":"formula_4","formula_text":"e (l) u = i∈N u 1 √ d u √ d i e (l-1) i , e (l) i = u∈N i 1 √ d i √ d u e (l-1) u ,(1)"},{"formula_coordinates":[4,54.17,350.15,237.19,13.7],"formula_id":"formula_5","formula_text":"L ui = -[y ui ln σ(ŷ ui ) + (1 -y ui ) ln(1 -σ(ŷ ui ))]. (2)"},{"formula_coordinates":[4,310.48,247.58,237.41,132],"formula_id":"formula_6","formula_text":"∂L ui ∂e k = - ∂ ln σ(ŷ ui ) ∂e k = - ln σ(ŷ ui ) ∂σ(ŷ ui ) • ∂σ(ŷ ui ) ∂ŷ ui • ∂ŷ ui ∂e k = - 1 σ(ŷ ui ) • σ(ŷ ui )(1 -σ(ŷ ui )) • ∂ŷ ui ∂e k = -[1 -σ(ŷ ui )] ∂ŷ ui ∂e k = -λ ui ∂ŷ ui ∂e k ,(3)"},{"formula_coordinates":[4,383.23,652.79,66.44,28.7],"formula_id":"formula_7","formula_text":"E ∂L i ∂e k ∝ -d where E(•) is the expectation, ϑ ϑ ϑ i = E Φ j p=1 1 l=L-1 1 √ d p l e j"},{"formula_coordinates":[5,54.1,210.3,237.26,87.47],"formula_id":"formula_8","formula_text":"∂L i ∂e k = j∈N i ∂L i j ∂e k = - j∈N i λ i j ∂ŷ i j ∂e k = - j∈N i λ i j ∂e i ∂e k e j = - j∈N i λ i j Φ p=1 1 l=L 1 d p l e j ,(5)"},{"formula_coordinates":[5,55.69,394.41,235.67,121.8],"formula_id":"formula_9","formula_text":"E ∂L i ∂e k = -E j∈N i λ i j Φ p=1 1 l=L 1 d p l e j ∝ -d i E j∈N i Φ j p=1 1 l=L-1 1 d p l e j ≈ -d 3 2 i E Φ j p=1 1 l=L-1 1 d p l e j = -d 3 2 i ϑ ϑ ϑ i .(6)"},{"formula_coordinates":[5,371.29,500.02,176.59,28.7],"formula_id":"formula_10","formula_text":"E ∂L r ∂e t > E ∂L s ∂e t .(7)"},{"formula_coordinates":[6,48.19,270.7,7.3,8.26],"formula_id":"formula_11","formula_text":"E"},{"formula_coordinates":[6,48.19,650.65,243.86,58.57],"formula_id":"formula_13","formula_text":"E ∂L r ∂e t = -υ 1 d 3 2 r ϑ ϑ ϑ r (υ 1 > 0) and ϑ ϑ ϑ r ∥ϑ ϑ ϑ r ∥ = υ 2 e r ∥e r ∥ (υ 2 > 0 for the local homogeneity), likewise, E ∂L s ∂e t = -υ 1 d 3 2"},{"formula_coordinates":[6,304.72,95.12,245.34,166.36],"formula_id":"formula_14","formula_text":"E e t - ∂L r ∂e t + ∂L s ∂e t ⊤ (e r -e s ) =E[e ⊤ t (e r -e s )] + υ 1 E d 3 2 r ϑ ϑ ϑ ⊤ r e r -d 3 2 r ϑ ϑ ϑ ⊤ r e s + d 3 2 s ϑ ϑ ϑ ⊤ s e r -d 3 2 s ϑ ϑ ϑ ⊤ s e s =E[e ⊤ t (e r -e s )] + υ 1 υ 2 E d 3 2 r ∥ϑ ϑ ϑ r ∥ ∥e r ∥ e ⊤ r e r -d 3 2 r ∥ϑ ϑ ϑ r ∥ ∥e r ∥ e ⊤ r e s +d 3 2 s ∥ϑ ϑ ϑ s ∥ ∥e s ∥ e ⊤ s e r -d 3 2 s ∥ϑ ϑ ϑ s ∥ ∥e s ∥ e ⊤ s e s =E[e ⊤ z (e r -e s )]"},{"formula_coordinates":[6,314.99,269.45,62.48,17.67],"formula_id":"formula_15","formula_text":"+ υ 1 υ 2 ϕE d 3 2"},{"formula_coordinates":[6,415.9,266.53,131.99,40.21],"formula_id":"formula_16","formula_text":"ρ ∥e r ∥ ) -d 3 2 s (∥e s ∥ - ρ ∥e s ∥ ) ,(10)"},{"formula_coordinates":[6,342.67,411.01,205.21,56.96],"formula_id":"formula_17","formula_text":"d 3 2 r (∥e r ∥ - ρ ∥e r ∥ ) -d 3 2 s (∥e s ∥ - ρ ∥e s ∥ ) =(d 3 2 r κ -d 3 2 s )∥e s ∥ -(d 3 2 r -d 3 2 s κ) ρ ∥e r ∥ ,(11)"},{"formula_coordinates":[6,332.45,489.43,92.45,17.77],"formula_id":"formula_18","formula_text":"d 3 2 r κ -d 3 2 s > d 3 2 r -d 3 2"},{"formula_coordinates":[6,311.89,608.07,20.01,15],"formula_id":"formula_19","formula_text":"e ′ r ="},{"formula_coordinates":[7,73.86,86.77,217.49,72.36],"formula_id":"formula_21","formula_text":"E e t - ∂L r ∂e t + ∂L s ∂e t ⊤ (e ′ r -e ′ s ) =E ẽ⊤ t [(e t -e s ) + (ω rt -ω st )ẽ t ] =E ẽ⊤ t (e r -e s ) + (ω rt -ω st )∥ẽ t ∥ 2 ,(14)"},{"formula_coordinates":[7,304.72,83.14,243.16,102.08],"formula_id":"formula_22","formula_text":"e (l) v = j∈N v 1 √ d v d j e (l-1) j = j∈N v 1 √ d v d j (ê (l-1) j + α (l-1) H j ϑ ϑ ϑ (l-1) H j + α (l-1) L j ϑ ϑ ϑ (l-1) L j ) = ê(l) v + α (l) H v ϑ ϑ ϑ (l) H v + α (l) L v ϑ ϑ ϑ (l) L v ,(15)"},{"formula_coordinates":[7,389.62,206.27,79.02,17.11],"formula_id":"formula_23","formula_text":"H v ϑ ϑ ϑ (l) H v and α (l) L v ϑ ϑ ϑ (l)"},{"formula_coordinates":[8,75.85,139.46,210.53,16.54],"formula_id":"formula_24","formula_text":"{C (l) 1 , C (l) 2 , • • • , C (l) P } = Kmeans(E (l) ), (16"},{"formula_coordinates":[8,286.37,142.4,4.98,10.68],"formula_id":"formula_25","formula_text":")"},{"formula_coordinates":[8,48.19,270.45,112.94,14.64],"formula_id":"formula_26","formula_text":"H (l) v = { j ∈ C (l) p , d j > d v }"},{"formula_coordinates":[8,80.24,287.79,120.46,14.64],"formula_id":"formula_27","formula_text":"L (l) v = { j ∈ C (l) p , d j < d v }."},{"formula_coordinates":[8,47.92,341.09,242.63,30.76],"formula_id":"formula_28","formula_text":"H v ∈ R 1×D and θ ϑ ϑ (l)"},{"formula_coordinates":[8,59.9,417.7,231.45,18.71],"formula_id":"formula_29","formula_text":"b (l) v = αM(e (l) v , θ ϑ ϑ (l) H v ) θ ϑ ϑ (l) H v + βM(e (l) v , θ ϑ ϑ (l) L v ) θ ϑ ϑ (l) L v ,(17)"},{"formula_coordinates":[8,134.28,648,157.07,14.95],"formula_id":"formula_30","formula_text":"ê(l) v = e (l) v -b (l) v .(18)"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 310.7,\n 161.7,\n 13.05,\n 12.7\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"C (l) p\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 310.48,\n 169.76,\n 202.57,\n 58.99\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"E (l) H (l) v A set of nodes { j ∈ C (l) p , d j > d v |v ∈ C (l) p } L (l) v A set of nodes { j ∈ C (l) p , d j < d v |v ∈ C (l) p } θ ϑ ϑ (l) H v\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 310.48,\n 221.81,\n 65.15,\n 30.4\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"H (l) v θ ϑ ϑ (l) L v\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 304.72,\n 325.99,\n 242.36,\n 28.91\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"U = {u 1 , u 2 , • • • , u M }, I = {i 1 , i 2 , • • • , i N } in a dataset D.\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 304.89,\n 574.7,\n 243,\n 43.9\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"e (l) u = i∈N u 1 √ d u √ d i e (l-1) i , e (l) i = u∈N i 1 √ d i √ d u e (l-1) u ,(1)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 54.17,\n 350.15,\n 237.19,\n 13.7\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"L ui = -[y ui ln σ(ŷ ui ) + (1 -y ui ) ln(1 -σ(ŷ ui ))]. (2)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 310.48,\n 247.58,\n 237.41,\n 132\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"∂L ui ∂e k = - ∂ ln σ(ŷ ui ) ∂e k = - ln σ(ŷ ui ) ∂σ(ŷ ui ) • ∂σ(ŷ ui ) ∂ŷ ui • ∂ŷ ui ∂e k = - 1 σ(ŷ ui ) • σ(ŷ ui )(1 -σ(ŷ ui )) • ∂ŷ ui ∂e k = -[1 -σ(ŷ ui )] ∂ŷ ui ∂e k = -λ ui ∂ŷ ui ∂e k ,(3)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 383.23,\n 652.79,\n 66.44,\n 28.7\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"E ∂L i ∂e k ∝ -d where E(•) is the expectation, ϑ ϑ ϑ i = E Φ j p=1 1 l=L-1 1 √ d p l e j\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 54.1,\n 210.3,\n 237.26,\n 87.47\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"∂L i ∂e k = j∈N i ∂L i j ∂e k = - j∈N i λ i j ∂ŷ i j ∂e k = - j∈N i λ i j ∂e i ∂e k e j = - j∈N i λ i j Φ p=1 1 l=L 1 d p l e j ,(5)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 55.69,\n 394.41,\n 235.67,\n 121.8\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"E ∂L i ∂e k = -E j∈N i λ i j Φ p=1 1 l=L 1 d p l e j ∝ -d i E j∈N i Φ j p=1 1 l=L-1 1 d p l e j ≈ -d 3 2 i E Φ j p=1 1 l=L-1 1 d p l e j = -d 3 2 i ϑ ϑ ϑ i .(6)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 371.29,\n 500.02,\n 176.59,\n 28.7\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"E ∂L r ∂e t > E ∂L s ∂e t .(7)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 48.19,\n 270.7,\n 7.3,\n 8.26\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"E\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 48.19,\n 650.65,\n 243.86,\n 58.57\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"E ∂L r ∂e t = -υ 1 d 3 2 r ϑ ϑ ϑ r (υ 1 > 0) and ϑ ϑ ϑ r ∥ϑ ϑ ϑ r ∥ = υ 2 e r ∥e r ∥ (υ 2 > 0 for the local homogeneity), likewise, E ∂L s ∂e t = -υ 1 d 3 2\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 304.72,\n 95.12,\n 245.34,\n 166.36\n ],\n \"formula_id\": \"formula_14\",\n \"formula_text\": \"E e t - ∂L r ∂e t + ∂L s ∂e t ⊤ (e r -e s ) =E[e ⊤ t (e r -e s )] + υ 1 E d 3 2 r ϑ ϑ ϑ ⊤ r e r -d 3 2 r ϑ ϑ ϑ ⊤ r e s + d 3 2 s ϑ ϑ ϑ ⊤ s e r -d 3 2 s ϑ ϑ ϑ ⊤ s e s =E[e ⊤ t (e r -e s )] + υ 1 υ 2 E d 3 2 r ∥ϑ ϑ ϑ r ∥ ∥e r ∥ e ⊤ r e r -d 3 2 r ∥ϑ ϑ ϑ r ∥ ∥e r ∥ e ⊤ r e s +d 3 2 s ∥ϑ ϑ ϑ s ∥ ∥e s ∥ e ⊤ s e r -d 3 2 s ∥ϑ ϑ ϑ s ∥ ∥e s ∥ e ⊤ s e s =E[e ⊤ z (e r -e s )]\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 314.99,\n 269.45,\n 62.48,\n 17.67\n ],\n \"formula_id\": \"formula_15\",\n \"formula_text\": \"+ υ 1 υ 2 ϕE d 3 2\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 415.9,\n 266.53,\n 131.99,\n 40.21\n ],\n \"formula_id\": \"formula_16\",\n \"formula_text\": \"ρ ∥e r ∥ ) -d 3 2 s (∥e s ∥ - ρ ∥e s ∥ ) ,(10)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 342.67,\n 411.01,\n 205.21,\n 56.96\n ],\n \"formula_id\": \"formula_17\",\n \"formula_text\": \"d 3 2 r (∥e r ∥ - ρ ∥e r ∥ ) -d 3 2 s (∥e s ∥ - ρ ∥e s ∥ ) =(d 3 2 r κ -d 3 2 s )∥e s ∥ -(d 3 2 r -d 3 2 s κ) ρ ∥e r ∥ ,(11)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 332.45,\n 489.43,\n 92.45,\n 17.77\n ],\n \"formula_id\": \"formula_18\",\n \"formula_text\": \"d 3 2 r κ -d 3 2 s > d 3 2 r -d 3 2\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 311.89,\n 608.07,\n 20.01,\n 15\n ],\n \"formula_id\": \"formula_19\",\n \"formula_text\": \"e ′ r =\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 73.86,\n 86.77,\n 217.49,\n 72.36\n ],\n \"formula_id\": \"formula_21\",\n \"formula_text\": \"E e t - ∂L r ∂e t + ∂L s ∂e t ⊤ (e ′ r -e ′ s ) =E ẽ⊤ t [(e t -e s ) + (ω rt -ω st )ẽ t ] =E ẽ⊤ t (e r -e s ) + (ω rt -ω st )∥ẽ t ∥ 2 ,(14)\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 304.72,\n 83.14,\n 243.16,\n 102.08\n ],\n \"formula_id\": \"formula_22\",\n \"formula_text\": \"e (l) v = j∈N v 1 √ d v d j e (l-1) j = j∈N v 1 √ d v d j (ê (l-1) j + α (l-1) H j ϑ ϑ ϑ (l-1) H j + α (l-1) L j ϑ ϑ ϑ (l-1) L j ) = ê(l) v + α (l) H v ϑ ϑ ϑ (l) H v + α (l) L v ϑ ϑ ϑ (l) L v ,(15)\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 389.62,\n 206.27,\n 79.02,\n 17.11\n ],\n \"formula_id\": \"formula_23\",\n \"formula_text\": \"H v ϑ ϑ ϑ (l) H v and α (l) L v ϑ ϑ ϑ (l)\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 75.85,\n 139.46,\n 210.53,\n 16.54\n ],\n \"formula_id\": \"formula_24\",\n \"formula_text\": \"{C (l) 1 , C (l) 2 , • • • , C (l) P } = Kmeans(E (l) ), (16\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 286.37,\n 142.4,\n 4.98,\n 10.68\n ],\n \"formula_id\": \"formula_25\",\n \"formula_text\": \")\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 48.19,\n 270.45,\n 112.94,\n 14.64\n ],\n \"formula_id\": \"formula_26\",\n \"formula_text\": \"H (l) v = { j ∈ C (l) p , d j > d v }\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 80.24,\n 287.79,\n 120.46,\n 14.64\n ],\n \"formula_id\": \"formula_27\",\n \"formula_text\": \"L (l) v = { j ∈ C (l) p , d j < d v }.\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 47.92,\n 341.09,\n 242.63,\n 30.76\n ],\n \"formula_id\": \"formula_28\",\n \"formula_text\": \"H v ∈ R 1×D and θ ϑ ϑ (l)\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 59.9,\n 417.7,\n 231.45,\n 18.71\n ],\n \"formula_id\": \"formula_29\",\n \"formula_text\": \"b (l) v = αM(e (l) v , θ ϑ ϑ (l) H v ) θ ϑ ϑ (l) H v + βM(e (l) v , θ ϑ ϑ (l) L v ) θ ϑ ϑ (l) L v ,(17)\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 134.28,\n 648,\n 157.07,\n 14.95\n ],\n \"formula_id\": \"formula_30\",\n \"formula_text\": \"ê(l) v = e (l) v -b (l) v .(18)\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"How Graph Convolutions Amplify Popularity Bias for Recommendation?"},"abstract":{"kind":"string","value":"Graph convolutional networks (GCNs) have become prevalent in recommender system (RS) due to their superiority in modeling collaborative patterns. Although improving the overall accuracy, GCNs unfortunately amplify popularity bias -tail items are less likely to be recommended. This effect prevents the GCN-based RS from making precise and fair recommendations, decreasing the effectiveness of recommender systems in the long run."},"authors":{"kind":"string","value":"Jiajia Chen; Jiancan Wu; Jiawei Chen; Xin Xin; Yong Li; Xiangnan He"},"figures":{"kind":"list like","value":[{"figure_caption":"Fig. 11Fig. 1 Performance change of LightGCN with different graph convolution layers on Gowalla. Recall@20 and TR@20 stand for the overall recall score and the ratio of tail items in the top-20 recommendation list, respectively.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"3 2 3 2 3 2 3 23333∥ϑ ϑ ϑ∥ and ∥ϑ ϑ ϑ∥ in log scale in Figure2. Items are evenly divided into several groups in ascending order of their degrees. For each item group, we show its average ln(d ∥ϑ ϑ ϑ∥) and ln(∥ϑ ϑ ϑ∥) in terms of the two-hop neighbor nodes (i.e., L = 2 in Lemma 1) when training LightGCN. As we see, ln(d ∥ϑ ϑ ϑ∥) enlarges as degree increases. Compared to ln(d ∥ϑ ϑ ϑ∥), ln(∥ϑ ϑ ϑ∥) is relatively smaller and flat across at different degrees. This illustrates that the degree of nodes plays a crucial role in the","figure_data":"","figure_id":"fig_1","figure_label":"3333","figure_type":"figure"},{"figure_caption":"Fig. 2 3 223Fig. 2 Average ∥θ θ θ∥ and d 3 2 ∥θ θ θ∥ in each items group. Items are sorted into groups in ascending order of their degrees.","figure_data":"","figure_id":"fig_2","figure_label":"23","figure_type":"figure"},{"figure_caption":"Fig. 33Fig.3The average aggregation weight of one-order neighbor users in each items group. Items are sorted into groups in ascending order of their degrees.","figure_data":"","figure_id":"fig_3","figure_label":"3","figure_type":"figure"},{"figure_caption":"are calculated by comparing LightGCN-DAP-o (UltraGCN-DAPo) with LightGCN (UltraGCN). In general, our DAP significantly boosts two GCN methods on the Tail test set. The main observations are as follows:","figure_data":"","figure_id":"fig_4","figure_label":"","figure_type":"figure"},{"figure_caption":"Fig. 44Fig. 4 Performance comparison between LightGCN and LightGCN-DAP with different layers of graph convolution on the Overall test set.","figure_data":"","figure_id":"fig_5","figure_label":"4","figure_type":"figure"},{"figure_caption":"4. 33Ablation Study (RQ2) 4.3.1 Performance of Variants.","figure_data":"","figure_id":"fig_6","figure_label":"3","figure_type":"figure"},{"figure_caption":"xFig. 55Fig. 5 Ablation study of DAP with different hyper-parameters α and β on the Overall test set.","figure_data":"","figure_id":"fig_7","figure_label":"5","figure_type":"figure"},{"figure_caption":"( 3 )3Compared to the performance of DAP-kh and DAP-m, it can be found that DAPkh outperforms DAP-m. It reflects that the similarity function M could capture the relation strength among nodes and thus help estimate the amplified bias well. 4.3.2 Effect of Different Hyper-Parameters α and β.","figure_data":"","figure_id":"fig_8","figure_label":"3","figure_type":"figure"},{"figure_caption":"4. 3 . 333Effect of Different Hyper-Parameter P.","figure_data":"","figure_id":"fig_9","figure_label":"33","figure_type":"figure"},{"figure_caption":"Fig. 66Fig. 6 DAP can effectively alleviate the popularity bias.","figure_data":"","figure_id":"fig_10","figure_label":"6","figure_type":"figure"},{"figure_caption":"Main notations used in the paper.U, IUser set, item set N u , N i The one-order neighbors of user u or item i d u , d i","figure_data":"","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"e r + ω rt e t -","figure_data":"∂L r ∂e t+∂L s ∂e t= e r + ω rt ẽt ,(12)where ω rt is the weight of aggregation. Likewise,e ′∂L r ∂e t+∂L s ∂e","figure_id":"tab_2","figure_label":"","figure_type":"table"},{"figure_caption":"Dataset description","figure_data":"Dataset#Users #Items #Interactions DensityGowalla29,858 40,9811,027,3700.00084Yelp201831,668 38,0481,561,4060.00130Amazon-book 52,643 91,5992,984,1080.00062","figure_id":"tab_3","figure_label":"2","figure_type":"table"},{"figure_caption":"Performance comparison between our method DAP and other counterparts on the Overall and Tail test sets. The 'improve' is the relative improvement of LightGCN-DAP-o over LightGCN.","figure_data":"DatasetGowallaYelp2018Amazon-bookModelsOverall Recall NDCGRecallTail NDCG Recall NDCG OverallRecallTail NDCG Recall NDCG OverallRecallTail NDCGLightGCN0.1820 0.1546 0.04340.0191 0.0627 0.0516 0.00910.0046 0.0414 0.03210.0090.0051BFGCN0.1083 0.0805 0.04680.0245 0.0389 0.0311 0.01240.0076 0.0276 0.0211 0.00970.0059LightGCN-IPSCN0.1325 0.1132 0.04770.0213 0.0473 0.0391 0.01360.0077 0.0285 0.0221 0.01180.0069LightGCN-CausE0.1334 0.1137 0.04850.0225 0.0492 0.0405 0.01410.0085 0.0299 0.0230 0.01270.0078LightGCN-DICE0.1337 0.1138 0.04930.0241 0.0505 0.0409 0.01320.0073 0.0348 0.0264 0.01210.0074LightGCN-MACR0.1188 0.0928 0.04780.0219 0.0343 0.0270.02330.0126 0.0269 0.0204 0.01080.0065LightGCN-Tail0.1647 0.1391 0.06280.03190.057 0.0466 0.01540.0095 0.0369 0.0283 0.01510.0094LightGCN-BxQuAD 0.1378 0.1130 0.06890.0360 0.0545 0.0431 0.02090.0123 0.0389 0.0304 0.01640.0108LightGCN-DAP-o0.1834 0.1564 0.05380.02450.0634 0.0521 0.01370.0073 0.0436 0.0339 0.01340.0079LightGCN-DAP-t0.1672 0.1427 0.07080.03540.0562 0.0461 0.02180.0129 0.0414 0.0328 0.01660.0102improve0.77% 1.16% 23.96% 28.27% 1.12% 0.97% 50.55% 58.70% 4.83% 5.61% 48.89% 54.90%","figure_id":"tab_4","figure_label":"3","figure_type":"table"},{"figure_caption":"Performance comparison between our method DAP and other counterparts on the Overall and Tail test sets. The 'improve' is the relative improvement of UltraGCN-DAP-o over UltraGCN.","figure_data":"DatasetGowallaYelp2018Amazon-bookModelsOverall Recall NDCG Recall NDCG Recall NDCG Recall NDCG Recall NDCG Recall NDCG Tail Overall Tail Overall TailUltraGCN0.1862 0.1579 0.0447 0.0213 0.0676 0.0554 0.0127 0.0074 0.0682 0.0556 0.0436 0.0297UltraGCN-IPSCN0.1345 0.1123 0.0451 0.0208 0.0401 0.0324 0.0144 0.0087 0.0442 0.0356 0.0458 0.0317UltraGCN-CausE0.1408 0.1177 0.0449 0.0209 0.0411 0.0329 0.0151 0.0096 0.0459 0.0369 0.0463 0.0320UltraGCN-DICE0.1424 0.1201 0.0512 0.0247 0.0516 0.0417 0.0157 0.0096 0.0545 0.0423 0.0491 0.0343UltraGCN-MACR0.1311 0.1078 0.0517 0.0252 0.0387 0.0323 0.0248 0.0141 0.0501 0.0398 0.0488 0.0335UltraGCN-Tail0.1788 0.1521 0.0634 0.0321 0.0618 0.0501 0.0167 0.0102 0.0599 0.0499 0.0531 0.0378UltraGCN-BxQuAD 0.1482 0.1289 0.0694 0.0361 0.0591 0.0482 0.0218 0.0136 0.0623 0.0517 0.0547 0.0386UltraGCN-DAP-o0.1868 0.1580 0.0551 0.0271 0.0678 0.0555 0.0135 0.0079 0.0688 0.0562 0.0462 0.0316UltraGCN-DAP-t0.1701 0.1483 0.0714 0.0362 0.0607 0.0493 0.0237 0.0135 0.0625 0.0520 0.0543 0.0391improve0.32% 0.06% 5.59% 6.57% 0.30% 0.18% 6.30% 6.76% 0.88% 1.07% 5.96% 6.40%","figure_id":"tab_5","figure_label":"4","figure_type":"table"},{"figure_caption":"reports our experimental results on the three datasets w.r.t. the hyper-parameter P. As can be seen, the performance increases as P becomes large and then decreases. For different datasets, when P is too small or large, the popularity bias among","figure_data":"","figure_id":"tab_6","figure_label":"5","figure_type":"table"},{"figure_caption":"Effect of P over DAP for the three datasets on the Overall test set.","figure_data":"DatasetGowallaYelp2018Amazon-bookPRecall NDCG Recall NDCG Recall NDCG10.1819 0.1547 0.0629 0.0517 0.0423 0.032950.182 0.15480.063 0.0518 0.0436 0.0339100.1823 0.1552 0.0633 0.0521 0.0435 0.0339300.1831 0.1563 0.0628 0.0518 0.0435 0.0338500.1834 0.1564 0.0626 0.0517 0.0428 0.0334700.1833 0.1564 0.0623 0.0515 0.0426 0.0331nodes can not be captured accurately, therefore theperformance is unsatisfactory. Kmeans is a simpletool for clustering, other advanced unsupervisedclustering tools can be explored for improving theperformance.","figure_id":"tab_7","figure_label":"5","figure_type":"table"},{"figure_caption":"Intelligence, AAAI. 2018 45. Chen D, Lin Y, Li W, Li P, Zhou J, Sun X. Measuring and relieving the over-smoothing problem for graph neural networks from the topological view. In: Proceedings of the AAAI Conference on Artificial Intelligence. 2020, 3438-3445 46. Chen M, Wei Z, Huang Z, Ding B, Li Y. Simple and deep graph convolutional networks. In: International Conference on Machine Learning. 2020, 1725-1735 47. Liu M, Gao H, Ji S. Towards deeper graph neural networks. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020, 338-348 48. Sun K, Lin Z, Zhu Z. Multi-stage self-supervised learning for graph convolutional networks on graphs with few labeled nodes. In: Proceedings of the AAAI Conference on Artificial Intelligence. 2020, 5892-5899","figure_data":"","figure_id":"tab_8","figure_label":"","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Fig. 11Fig. 1 Performance change of LightGCN with different graph convolution layers on Gowalla. Recall@20 and TR@20 stand for the overall recall score and the ratio of tail items in the top-20 recommendation list, respectively.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"3 2 3 2 3 2 3 23333∥ϑ ϑ ϑ∥ and ∥ϑ ϑ ϑ∥ in log scale in Figure2. Items are evenly divided into several groups in ascending order of their degrees. For each item group, we show its average ln(d ∥ϑ ϑ ϑ∥) and ln(∥ϑ ϑ ϑ∥) in terms of the two-hop neighbor nodes (i.e., L = 2 in Lemma 1) when training LightGCN. As we see, ln(d ∥ϑ ϑ ϑ∥) enlarges as degree increases. Compared to ln(d ∥ϑ ϑ ϑ∥), ln(∥ϑ ϑ ϑ∥) is relatively smaller and flat across at different degrees. This illustrates that the degree of nodes plays a crucial role in the\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"3333\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 2 3 223Fig. 2 Average ∥θ θ θ∥ and d 3 2 ∥θ θ θ∥ in each items group. Items are sorted into groups in ascending order of their degrees.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"23\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 33Fig.3The average aggregation weight of one-order neighbor users in each items group. Items are sorted into groups in ascending order of their degrees.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"are calculated by comparing LightGCN-DAP-o (UltraGCN-DAPo) with LightGCN (UltraGCN). In general, our DAP significantly boosts two GCN methods on the Tail test set. The main observations are as follows:\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 44Fig. 4 Performance comparison between LightGCN and LightGCN-DAP with different layers of graph convolution on the Overall test set.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"4. 33Ablation Study (RQ2) 4.3.1 Performance of Variants.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"xFig. 55Fig. 5 Ablation study of DAP with different hyper-parameters α and β on the Overall test set.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"( 3 )3Compared to the performance of DAP-kh and DAP-m, it can be found that DAPkh outperforms DAP-m. It reflects that the similarity function M could capture the relation strength among nodes and thus help estimate the amplified bias well. 4.3.2 Effect of Different Hyper-Parameters α and β.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_8\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"4. 3 . 333Effect of Different Hyper-Parameter P.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_9\",\n \"figure_label\": \"33\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 66Fig. 6 DAP can effectively alleviate the popularity bias.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_10\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Main notations used in the paper.U, IUser set, item set N u , N i The one-order neighbors of user u or item i d u , d i\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"e r + ω rt e t -\",\n \"figure_data\": \"∂L r ∂e t+∂L s ∂e t= e r + ω rt ẽt ,(12)where ω rt is the weight of aggregation. Likewise,e ′∂L r ∂e t+∂L s ∂e\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Dataset description\",\n \"figure_data\": \"Dataset#Users #Items #Interactions DensityGowalla29,858 40,9811,027,3700.00084Yelp201831,668 38,0481,561,4060.00130Amazon-book 52,643 91,5992,984,1080.00062\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Performance comparison between our method DAP and other counterparts on the Overall and Tail test sets. The 'improve' is the relative improvement of LightGCN-DAP-o over LightGCN.\",\n \"figure_data\": \"DatasetGowallaYelp2018Amazon-bookModelsOverall Recall NDCGRecallTail NDCG Recall NDCG OverallRecallTail NDCG Recall NDCG OverallRecallTail NDCGLightGCN0.1820 0.1546 0.04340.0191 0.0627 0.0516 0.00910.0046 0.0414 0.03210.0090.0051BFGCN0.1083 0.0805 0.04680.0245 0.0389 0.0311 0.01240.0076 0.0276 0.0211 0.00970.0059LightGCN-IPSCN0.1325 0.1132 0.04770.0213 0.0473 0.0391 0.01360.0077 0.0285 0.0221 0.01180.0069LightGCN-CausE0.1334 0.1137 0.04850.0225 0.0492 0.0405 0.01410.0085 0.0299 0.0230 0.01270.0078LightGCN-DICE0.1337 0.1138 0.04930.0241 0.0505 0.0409 0.01320.0073 0.0348 0.0264 0.01210.0074LightGCN-MACR0.1188 0.0928 0.04780.0219 0.0343 0.0270.02330.0126 0.0269 0.0204 0.01080.0065LightGCN-Tail0.1647 0.1391 0.06280.03190.057 0.0466 0.01540.0095 0.0369 0.0283 0.01510.0094LightGCN-BxQuAD 0.1378 0.1130 0.06890.0360 0.0545 0.0431 0.02090.0123 0.0389 0.0304 0.01640.0108LightGCN-DAP-o0.1834 0.1564 0.05380.02450.0634 0.0521 0.01370.0073 0.0436 0.0339 0.01340.0079LightGCN-DAP-t0.1672 0.1427 0.07080.03540.0562 0.0461 0.02180.0129 0.0414 0.0328 0.01660.0102improve0.77% 1.16% 23.96% 28.27% 1.12% 0.97% 50.55% 58.70% 4.83% 5.61% 48.89% 54.90%\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Performance comparison between our method DAP and other counterparts on the Overall and Tail test sets. The 'improve' is the relative improvement of UltraGCN-DAP-o over UltraGCN.\",\n \"figure_data\": \"DatasetGowallaYelp2018Amazon-bookModelsOverall Recall NDCG Recall NDCG Recall NDCG Recall NDCG Recall NDCG Recall NDCG Tail Overall Tail Overall TailUltraGCN0.1862 0.1579 0.0447 0.0213 0.0676 0.0554 0.0127 0.0074 0.0682 0.0556 0.0436 0.0297UltraGCN-IPSCN0.1345 0.1123 0.0451 0.0208 0.0401 0.0324 0.0144 0.0087 0.0442 0.0356 0.0458 0.0317UltraGCN-CausE0.1408 0.1177 0.0449 0.0209 0.0411 0.0329 0.0151 0.0096 0.0459 0.0369 0.0463 0.0320UltraGCN-DICE0.1424 0.1201 0.0512 0.0247 0.0516 0.0417 0.0157 0.0096 0.0545 0.0423 0.0491 0.0343UltraGCN-MACR0.1311 0.1078 0.0517 0.0252 0.0387 0.0323 0.0248 0.0141 0.0501 0.0398 0.0488 0.0335UltraGCN-Tail0.1788 0.1521 0.0634 0.0321 0.0618 0.0501 0.0167 0.0102 0.0599 0.0499 0.0531 0.0378UltraGCN-BxQuAD 0.1482 0.1289 0.0694 0.0361 0.0591 0.0482 0.0218 0.0136 0.0623 0.0517 0.0547 0.0386UltraGCN-DAP-o0.1868 0.1580 0.0551 0.0271 0.0678 0.0555 0.0135 0.0079 0.0688 0.0562 0.0462 0.0316UltraGCN-DAP-t0.1701 0.1483 0.0714 0.0362 0.0607 0.0493 0.0237 0.0135 0.0625 0.0520 0.0543 0.0391improve0.32% 0.06% 5.59% 6.57% 0.30% 0.18% 6.30% 6.76% 0.88% 1.07% 5.96% 6.40%\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"reports our experimental results on the three datasets w.r.t. the hyper-parameter P. As can be seen, the performance increases as P becomes large and then decreases. For different datasets, when P is too small or large, the popularity bias among\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Effect of P over DAP for the three datasets on the Overall test set.\",\n \"figure_data\": \"DatasetGowallaYelp2018Amazon-bookPRecall NDCG Recall NDCG Recall NDCG10.1819 0.1547 0.0629 0.0517 0.0423 0.032950.182 0.15480.063 0.0518 0.0436 0.0339100.1823 0.1552 0.0633 0.0521 0.0435 0.0339300.1831 0.1563 0.0628 0.0518 0.0435 0.0338500.1834 0.1564 0.0626 0.0517 0.0428 0.0334700.1833 0.1564 0.0623 0.0515 0.0426 0.0331nodes can not be captured accurately, therefore theperformance is unsatisfactory. Kmeans is a simpletool for clustering, other advanced unsupervisedclustering tools can be explored for improving theperformance.\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Intelligence, AAAI. 2018 45. Chen D, Lin Y, Li W, Li P, Zhou J, Sun X. Measuring and relieving the over-smoothing problem for graph neural networks from the topological view. In: Proceedings of the AAAI Conference on Artificial Intelligence. 2020, 3438-3445 46. Chen M, Wei Z, Huang Z, Ding B, Li Y. Simple and deep graph convolutional networks. In: International Conference on Machine Learning. 2020, 1725-1735 47. Liu M, Gao H, Ji S. Towards deeper graph neural networks. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020, 338-348 48. Sun K, Lin Z, Zhu Z. Multi-stage self-supervised learning for graph convolutional networks on graphs with few labeled nodes. In: Proceedings of the AAAI Conference on Artificial Intelligence. 2020, 5892-5899\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work LightGCN is the backbone model used in the citing paper, and the BPR loss is employed as the loss function in the GCN-based recommender system.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[9]\", \"Explanation\": \"The cited work provides a method for alleviating popularity bias in RS, which the citing paper adopts to address the issue in GCNs.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[10][11]\", \"Explanation\": \"The cited works further explore the issue of popularity bias in RS and provide additional solutions to address the problem.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[12]\", \"Explanation\": \"The cited work introduces propensity-based methods to intervene in the training process and lower the weights of popular items, which the citing paper adopts to alleviate the popularity bias in RS.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[9,10]\", \"Explanation\": \"The cited works model the causal effect of the bias on the model training, which the citing paper uses to understand the impact of the bias on the model training in RS.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work performs personalized re-ranking on the generated candidates of the RS model to suppress popular items, which the citing paper extends the research by proposing a post-hoc method to revise the inference stage in RS.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work provides a method for learning user and item node representations using GCNs, which the citing paper adopts in their research to represent the user-item bipartite graph.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[15]\", \"Explanation\": \"The cited work also contributes to the method of learning user and item node representations using GCNs, which the citing paper may have further built upon in their research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work, LightGCN, is mentioned as a neat and well-represented GCN-based recommender that the citing paper may have extended or built upon in their research to further improve the learning of user and item node representations in the user-item bipartite graph.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work UltraGCN is used as a benchmark model in the citing paper to verify the generality of the proposed method in a different model context.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work introduces the BCE loss function, which the citing paper adopts in the formal formulation of the GCN model to optimize model parameters and address the popularity bias amplification problem.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[17][18][19]\", \"Explanation\": \"The cited works provide the concept of influence functions, which the citing paper adopts to define the influence of a user-item pair loss on their neighbors in the context of representation learning and graph convolution in GCNs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[19]\", \"Explanation\": \"The cited work provides the assumption that is used in the citing paper to posit the relationship between the node distance and the influence in a graph.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[20]\", \"Explanation\": \"The cited work by [20] provides a method for increasing the magnitude of node representations based on the degree of the node, which the citing paper adopts in their research to improve the performance of their model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[21]\", \"Explanation\": \"The cited work provides a method for estimating amplification bias in the final representation of a node, which the citing paper adopts in the l-th layer graph convolution to measure the effect of bias on the node.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work by LightGCN is the source of the data used in the experiments conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[22]\", \"Explanation\": \"The cited work, BFGCN, is a novel graph convolution filter for the user-item bipartite network that the citing paper implements in their DAP to improve long-tail node representations in the recommender models.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work, UltraGCN, is a state-of-the-art method that the citing paper compares to in their evaluation of the debiasing performance on recommendation. The citing paper extends the study by comparing the performance of UltraGCN to other methods in the recommender models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[23]\", \"Explanation\": \"The cited work, IPSCN, is a method that the citing paper compares to in their evaluation of the debiasing performance on recommendation. The citing paper implements IPSCN in their DAP to improve the performance of the recommender models by adding max-capping and normalization on IPS value.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[24]\", \"Explanation\": \"The cited work, IPS, is a method that the citing paper compares to in their evaluation of the debiasing performance on recommendation. The citing paper uses IPS in their IPSCN method to eliminate popularity bias in the recommender models by re-weighting each item according to its popularity.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[25]\", \"Explanation\": \"The cited work, CausE, is a method that the citing paper compares to in their evaluation of the debiasing performance on recommendation. The citing paper uses CausE in their evaluation by adding a regularizer term on the discrepancy between the item vectors used to fit the biased data and their counterpart representations that fit the unbiased data.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[10]\", \"Explanation\": \"The cited work provides a method for handling popularity bias in training data, which the citing paper adopts to obtain unbiased data for their research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work, BxQuAD, is a post-hoc method for improving tail item recommendations. The citing paper adopts the Binary-xQuAD method from the original paper to enhance the performance of their own research in this area.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[26]\", \"Explanation\": \"The cited work, Tail, learns a neighborhood translation method for node classification. The citing paper transfers this method to the field of recommendation to enhance the performance of their research in the area of tail item recommendations.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[27]\", \"Explanation\": \"The cited work, Adam, is used as the optimization method for all models in the comparison, including the UltraGCN. This indicates that the citing paper adopts the Adam method for optimizing the models in their research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[21]\", \"Explanation\": \"The cited work provides evidence that directly eliminating bias in recommendation is not reasonable, which supports the claim made in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[9]\", \"Explanation\": \"The cited work, MACR, is a method that has been applied in the citing paper to solve bias issues in recommendation, providing a basis for the research conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[10]\", \"Explanation\": \"The cited work, DICE, is another method that has been used in the citing paper to address bias issues in recommendation, serving as a methodological basis for the research conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[11]\", \"Explanation\": \"The cited work proposes a causal framework that the citing paper adopts to leverage the popularity bias in recommendation.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[25,[38][39][40][41]\", \"Explanation\": \"The cited works propose various causal methods for learning an unbiased recommender, which the citing paper builds upon to address bias issues in recommendation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[19,[42][43][44][45]\", \"Explanation\": \"The cited works reveal various issues of GCNs, including over-smoothing, vulnerability, and degree-related biases. The citing paper adopts these findings to develop methods for treating the over-smoothing issue in GCNs.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[46]\", \"Explanation\": \"The cited work proposes GCNII, which extends the vanilla GCN model with initial residual connection and identity mapping to prevent over-smoothing. The citing paper builds upon this work by further exploring methods for tackling the over-smoothing issue in GCNs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[45]\", \"Explanation\": \"The cited work proposes two methods for treating the over-smoothing issue in GCNs, including adding a regularizer to the training objective and optimizing the graph topology based on model predictions. The citing paper adopts these methods to address the over-smoothing issue in GCNs.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[47]\", \"Explanation\": \"The cited work argues that the over-smoothing issue only happens after a large number of iterations and the current results with several layers of graph convolution are relatively far from the ideal over-smoothing situation. The citing paper builds upon this argument to treat the nodes clustering issue and over-smoothing with different approaches, such as using shallow graph convolution layers and a large number of layers.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[19]\", \"Explanation\": \"The cited work points out a bias in GCNs towards higher-degree nodes, which the citing paper analyzes and uses to inform the development of a method to enhance the representations of low-degree nodes.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[26]\", \"Explanation\": \"The cited work proposes a method of learning neighborhood translations to enhance the representations of tail nodes, which the citing paper builds upon to further improve the performance of tail nodes in GCN-based recommenders.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[48]\", \"Explanation\": \"The cited work introduces a training algorithm based on self-training to add confident data with virtual labels to the labeled set, which the citing paper utilizes in its research to enlarge the training set and improve the performance of GCN-based recommenders.\"}]"}}},{"rowIdx":91,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b26","b37","b61","b6","b26","b3","b43","b53","b35","b16","b54","b56","b8","b40","b55","b40","b5","b42","b5","b42","b40"],"table_ref":[],"text":"Knowledge distillation (KD) methods [27,37,60] are powerful and flexible tools to transfer the knowledge of a given teacher model to the transfer target, the student model, without copying the weights. Instead, these methods match the student's functional activity (e.g. the softmax output) to that of the teacher for the presented inputs. Hence, those methods are independent of architectural details and allow knowledge distillation to be applied in scenarios like model compression [7,27], continual learning [4,42,52], or even neuroscience [35], where traditional transfer learning would be impossible to use. KD methods also appear to be key to training new models that trade off inductive biases for more flexibility and more parameters [17,53,55] on smaller data [9,40,54]. However, Nix et al. [40] recently showed that current KD methods fail to transfer even simple equivariances between teacher and student. Additionally, previous work showed that KD leads to a larger gap between student and teacher on out-of-domain evaluation performance compared to within domain performance [6,41], even in cases where the student almost perfectly matches the teacher [6] (see Table 5). This phenomenon is especially pronounced for particularly robust teachers [41]. Thus we expect that transferring robustness properties is a difficult problem for KD in general.\nFigure 1: Our task-agnostic HARD framework switches between training the student to match the teacher and training the augmentor to generate new samples on which the student underperforms while maintaining high teacher performance. We optimize the augmentor and student in interchanging phases through a student-teacher loss L s↔ t and teacher-teacher loss L t↔t . We switch between the two phases by comparing the default loss L s on augmented data to pre-defined thresholds.\nWe hypothesize that KD methods are in principle capable of transferring most knowledge from a teacher to a student if the training data is chosen adequately. We confirm this hypothesis on a small toy example (Section 3), showing the importance of input data for KD. Motivated by this demonstration, we propose our Hard Augmentations for Robust Distillation (HARD) method, a general framework (Section 4) to generate augmented training inputs which improve knowledge transfer by maximizing the distance between teacher and student while leaving the teacher's output unchanged. Consequently, our framework moves the input in directions that the teacher is invariant to but which are most challenging for the student. Our experiments (Section 5) show that our task-agnostic framework improves transfer effectiveness and thereby solves the problem of of KD not being able to transfer shift equivariance [40]. Additionally, as part of our framework, we propose several parameterized augmentations (Section 4.1) that can be integrated with most existing KD methods and are applicable to a variety of different computer vision tasks. Finally, we demonstrate across multiple different models on the tasks of CIFAR10 and ImageNet that our framework learns interpretable augmentations that improve KD to the same level and in many cases even beyond established data augmentation methods, even when evaluated in an out-of-domain setting."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b17","b32","b11","b23","b27","b39","b60","b62","b23","b62","b60","b10","b24","b49","b68","b1","b2","b19","b57","b64","b65","b66","b5","b59","b44","b20","b46","b63","b28","b62","b9","b30","b34"],"table_ref":[],"text":"There is a long tradition in using data augmentations to artificially extend training data for deep learning models and particularly in computer vision, be it through adding Gaussian noise, random crops, shifts, flips, or rotations [18,33]. In recent years, data augmentations became more complex [12,24,28,39,59,61], employing a multitude of different heuristics with the aim to improve generalization and in some cases also out-of-domain performance [24]. A particularly popular augmentation method is Mixup [61], which randomly interpolates two input samples and their labels respectively. Similarly, Cutmix [59] combines two input images by pasting a random crop of one image on top of the other. Also, many studies use parameterized augmentations optimized to improve a given objective [11,25,48,67], and some even optimize the augmentations to improve on an adversarial objective [2,3,20,56,[63][64][65], however, without applying them for knowledge transfer.\nIn KD, applying data augmentations is a very effective tool to improve matching between student and teacher [6,58] and optimizing on a meta level can be useful to aide the teaching [43]. Similar to our work, Haidar et al. [21], Rashid et al. [45], Zhang et al. [62] utilized adversarial objectives to optimize data augmentations for KD, however, they were solely focused on natural language processing tasks and do not optimize the augmentations towards invariance.\nInspired by this large body of work we formulate a task-agnostic framework containing only one building block that is specific to the data-domain -the instantiation of the augmentor model generating the augmented data samples -for which we offer a variety of reasonable model choices based on spatial transformer modules [29], Mixup [61], and variational autoencoders [10,31,34]. "},{"figure_ref":["fig_2","fig_2","fig_2","fig_2","fig_2","fig_2","fig_2","fig_2"],"heading":"Input Data Matters for Functional Transfer","publication_ref":["b0","b5","b62"],"table_ref":[],"text":"We hypothesize that the choice of input data is crucial to successfully knowledge distillation and we illustrate the impact of training data by a simple toy example. To demonstrate this, consider a simple KD task in which we instantiate the teacher-model by the true function f t (x) = cos(x) and the student f s (x) by a three layer Multilayer Perceptron (MLP) with ReLU activation [1]. We use input data x chosen such that it does not capture the teacher's cos(x) periodicity (orange points in Figure 2A). Simple KD does neither interpolate between the given training points nor extrapolates beyond them (Figure 2E). Hence the student neural network does not learn the teacher's periodicity and fails to interpolate and extrapolate beyond the training data (Figure 2A).\nAugmenting the training data with more helpful inputs x and teacher labels f t (x) = cos(x) could mitigate this problem. One method successfully applied to KD [6] is to extend the input data through Mixup [61]. When applying this to our illustrative example, we create new training inputs x through linear interpolation between pairs of input points x = (1 -α)x 1 + αx 2 (Figure 2F), and recording the corresponding teacher responses f t (x) = cos(x). Thus, the student learns to interpolate between training points, but mixup does not enhance extrapolation (Figure 2B).\nTo generate datapoints that would interpolate and extrapolate beyond already available training points, we could simply augment by adding Gaussian noise ϵ to the available data points, x = x + ϵ, hence interpolating and extrapolating beyond the training data (Figure 2G). This strategy helps our student to match the teacher also outside the original training regime (Figure 2C). However, the student only improves within a fixed margin that is determined by the noise distribution's mean and variance.\nWe could obviously improve interpolation and extrapolation by increasing the noise distribution's variance or shifting it's mean, however, as we move to a high dimensional image input space (x ∈ R → ⃗ x ∈ R N ) it becomes unclear how to heuristically select helpful new samples and at the same time random exploration strategies become computationally infeasible. Instead, we propose to optimize a parameterized augmentation to efficiently generate new, hard training samples on which the student lacks performance, as here the student could improve the most. In our toy example, we illustrate this by optimizing the Gaussian's parameters (mean and variance) according to our augmentation framework HARD, which we will present in the next section. This provides us with a noise distribution which we use to draw new helpful training examples x that transfer inter-and extrapolation to the student network (Figure 2D,H). Overall, this toy example shows that learning hard augmentations to select new helpful data points is crucial to efficiently improve extrapolation beyond the training distribution."},{"figure_ref":[],"heading":"Learning Hard Augmentations for Robust Distillation (HARD)","publication_ref":["b26"],"table_ref":[],"text":"Our task-agnostic HARD framework learns augmenting training images to most efficiently help knowledge-transfer from a teacher to a student model. Our method requires three main components: a teacher model with frozen parameters, a student model that should learn knowledge from the teacher, and a parameterized augmentation model that learns to augment images such that most of the teacher's knowledge is transferred to the student.\nIn classical KD methods [27], the objective is to minimize a distance D [f s (x), f t (x)] between the student's activation f s (x) and the teacher's activation f t (x) on given inputs x ∈ R n . Usually, this would be the Kullback-Leibler divergence between the softmax distributions of teacher and student. Unfortunately, only considering training data could miss properties of the teacher (eg. shift invariance) that might be crucial for generalization (see Section 3 for an illustrative example). To resolve this issue, we learn a parametrized augmentation model g a to generate new input data points x = g a (x) transferring such invariance properties from the teacher to the student. Hence, we define a teacher-student loss considering the more general case of matching student and teacher on augmented inputs x ∈ R n :\nL s↔ t = D [f s (x), f t (x)] .(1)\nTo specifically transfer the teacher's invariance properties to the student, we propose a teacher-teacher loss pushing the augmentor towards generating data points on which the teacher is invariant,\nL t↔t = D [f t (x), f t (x)] ,(2)\nas these are often useful augmentations for generalization. Using both of these losses, we optimize the augmentor's parameters θ a to generate augmented samples on which the teacher results in similar activations but the student differs from them (Figure 1 top) and simultaneously we optimize the student's parameters θ s to perform well on those augmentations (Figure 1 bottom):\nmax θa λ s L s↔ t -λ t L t↔t and min θs L s↔ t .(3)\nHere, λ s and λ t trade off the loss terms and are treated as hyper-parameters. We train both components separately switching from training the augmentor to training the student when the student's performance on augmented data gets worse than a pre-defined threshold (L s > ℓ max ) and we switch back from student to augmentor training when the student's performance on augmented data surpasses a pre-defined threshold (L s < ℓ min ; Figure 1). To prevent catastrophic forgetting, we save augmentors at every switch and employ an augmentor randomly chosen out of the set of previously saved augmentors in each iteration when training the student."},{"figure_ref":["fig_1"],"heading":"The augmentor models","publication_ref":["b30","b28"],"table_ref":[],"text":"To generate new input data points it is important to choose an augmentor that suits the desired application and is powerful enough to generate useful augmentations. Usually, we do not know a priori what useful augmentations are and thus should try to allow as much flexibility as possible. Additionally, some variance over augmentations could benefit the transfer. Thus, all augmentors in our study introduce randomness in the model by adding Gaussian noise into the computation of the augmentation through the reparametrization trick [31]. While our framework is universally applicable across domains, choosing an effective augmentation model likely needs to be addressed for each task individually. In our experiments, we use the following augmentor models:\nHARD-Affine In the simplest model, we limit the augmentations to affine transformations of the coordinate grid of pixel locations, i.e. shifts, rotations, scalings, and shears of images. Models implementing such transformations are known as spatial transformers [29]. We leverage this model for our augmentor by learning a distribution over the entries of an affine transformation matrix ϑ ∈ R 2×3 that defines the transformation of the sampling grid, i.e. a transformation that maps the pixel positions from the original image to the augmented image (Figure 3A)."},{"figure_ref":["fig_1","fig_1"],"heading":"HARD-Mix","publication_ref":["b62","b60","b30","b9","b34"],"table_ref":[],"text":"Additionally we consider a slightly more complex augmentor model, which is an adaptive variant of the commonly used Mixup [61] and Cutmix [59] augmentations. However, instead of randomly sampling the ratio and cutout position that are used to combine images, we learn how to combine the images dependent on the input images. We achieve this by performing a patch-wise projection of the input image, followed by comparing each patch with the same query vector sampled from a learned distribution (Figure 3B). We normalize similarities for each patch over each group of images and use the resulting weights to combine the original image patches, giving a combined image. This mechanism allows our augmentor to decide which features of which image are shown to the student, enabling it to explore the interpolated space between images systematically, instead of randomly. As it would not make sense for the teacher to be invariant to an interpolation as it is generated by HARD-Mix, we do not consider the teacher-teacher-loss L t↔t in this case and optimize student and augmentor jointly instead.\nHARD-VAE To lift constraints further, we wanted to use a more powerful augmentor that could generate a large variety of images across the entire image-space. As the augmentor has to generate new samples on-the-fly during the student training, the generation process needs to be very fast, limiting the choice of useful generative models. For this reason, we focus on variants of the variational autoencoder architecture [31], allowing for good image reconstructions which can be achieved reasonably fast in a single forward pass (Figure 3D). For CIFAR, we choose the very deep VAE [10] model, which we finetune by solely optimizing parameters of the posterior network from layer 10 onward in the decoder. For the experiments on ImageNet, we use a Residual-Quantized VAE (RQ-VAE) [34] pretrained on ImageNet, which we finetune in its entirety and add a noise vector on the latent state. Hence, as training progresses, the model changes from generating plain reconstructions of a given image to input conditioned generations that serve as our augmentations."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Transferring equivariance","publication_ref":["b40","b21","b16","b14","b40","b38","b40","b40"],"table_ref":[],"text":"For our initial experiment, we reproduce the setup from Nix et al. [40] to test whether we can transfer the inductive bias from a shift equivariant teacher, CNN and ResNet18 [22], to a student that does not have this inductive bias built into its architecture: a Multi-Layer Perceptron (MLP) and a Vision Transformer (ViT) [17]. When training the students and teachers by themselves on standard MNIST [15] training data, we observe a small drop of generalization performance (-0.6% and -1.2%) between Table 1: MNIST (columns \"Centered\") and MNIST-C (columns \"Shifted\") test accuracies (mean and standard error of the mean across 4 random seeds) comparing KD without augmentation and our HARD-Affine method to Orbit transfer [40], which also learns and transfers equivariances. teacher and student on the MNIST test set and a large gap (-56.1% and -52.4%) when we evaluate on a version of the test set in which digits were randomly shifted [38]. As another baseline, we applied plain KD to transfer shift equivariance from teacher to student. Consistent with the findings of Nix et al. [40], we only observe a small improvement on the centered (+0.2% and +0.3%) and the shifted (+5.1% and +4.3%) test sets, which likely result from the centered training data we use for transfer.\nWe then test if combining KD with our augmentations produced by HARD-Affine would outperform these baselines. The resulting student model improves significantly on shifted inputs (+28.6% and +39.4%) compared to plain KD and the generated images clearly show that the augmentor learns to shift the digits within the image. Compared to Nix et al. [40] our approach outperforms their results on the ViT task but, while improving the out-of-domain generalization by 28.6% over baseline, stays behind the Orbit performance on the MLP task. This demonstrates that our method while acting on fewer parts of the network compared to Orbit and while being a more general method, can improve or reach better performance when it comes to transferring invariances, and can be generalized to bigger datasets, as we show below.\nWe verify that the student's performance improvement is specifically due to our data generation framework in two control experiments. The first experiment (Random Affine) augments the training inputs of a stand-alone student model with a random affine transformation akin to our augmentor model, but using transformation parameters sampled uniformly from a pre-defined, reasonably constrained range (i.e. ensuring the digit is always fully visible). This student performs well on the shifted test set, however, performance significantly degrades on the centered test set. In comparison, our HARD-Affine model is unconstrained and learns more useful augmentations, leading to better performance on the centered test sets.\nIn our second control (Shifts) we asked how much data augmentation could improve the performance in the best case (without KD). For this, we augment the inputs by the same random shifts that were applied to obtain the shifted test data, leading to great improvements on the shifted test set. However, our learned augmentations achieve scores in a similar range on the shifted evaluation and outperform its results on the centered test set."},{"figure_ref":[],"heading":"Transfer on natural images","publication_ref":["b31","b13","b21","b24","b16"],"table_ref":["tab_0","tab_0","tab_0","tab_0","tab_0","tab_0"],"text":"After demonstrating that our method successfully captures the difference between teacher and student and bridges a gap in inductive bias, we now want to test whether this effect holds up in more realistic scenarios. CIFAR experiments We begin by applying our framework to CIFAR10 [32] on three different KD scenarios (see Table 2). Specifically, we test scenarios where the student lacks an inductive bias (ResNet18→ViT), where the teacher has more capacity and access to data than the student (ResNet101 * →ResNet18), and to scenarios combining both properties (ResNet101 * →ViT). For all experiments, we keep the experimental setup as close to our previous MNIST experiments as possible (see Appendix A for details).\nWe start by establishing baselines by training only the teacher and only the student models on the data and evaluating default KD. We observe that on this small data set a small ResNet18 performs better (78.5% accuracy) than a larger ViT (68.5%), likely because of the ResNet's superior inductive bias on this task and small data set. Next, we find that adding default data augmentations (random rotations, cropping, horizontal flips) to the student baselines significantly boosts performance to 92.6% and 78.3% for the ResNet18 and ViT, respectively. Adding these default augmentations to typical KD leads to a great performance boost, too (see Table 2).\nGiven that adding default data augmentation to KD already leads to a substantial performance boost, it is particularly noteworthy that the data augmentations learned by HARD-Affine outperform this baseline for the ViT. Qualitatively, the augmented images exhibit a large variety of spatial transformations, suggesting that a difference in these examples lead to the observed performance boost (Table 2, right).\nWe then investigated performance of our HARD-VAE augmentation strategy and found performance improvement over the KD + standard augmentations baseline for transfer to the ViT (+1.0% and +1.9%) student. However, inspecting the augmented images indicates that our augmentor lacks the expected shifts of object positions, but rather learns stylistic changes in the image (Table 2, right). This motivated us to combine HARD-Affine and HARD-VAE augmentation resulting in best performance (up to +7.8%) for all teacher-student pairings (HARD-VAE-Affine in Table 2) and the resulting images demonstrate variability in both style and spatial alignment (Table 2, right).\nImageNet experiments Having established our methods' performance for CIFAR10, we extend our results to classification on ImageNet [14]. Here we aim to distill a ResNet50 [22] teacher, trained with Deep-augment and AugMix data augmentations [25], into a smaller ResNet18 and ViT-S (small vision transformer variant) [17] that we want to be particularly robust to natural image corruptions,."},{"figure_ref":[],"heading":"The distillation into ResNet18 allows us to investigate the capability for model compression, because","publication_ref":["b4","b47","b18","b22","b24","b25","b58","b13","b4","b47","b24","b25","b24","b58","b18","b22","b60","b62","b23","b39","b5"],"table_ref":["tab_1","tab_1"],"text":"ResNet18 is a smaller network compared to ResNet50, but with a similar architecture. Distillation into a ViT-S architecture with a patch-size of 14 tests additionally if KD transfers the ResNet50's inductive bias of shift equivariance on a larger dataset.\nWe evaluate on common test sets for both in-domain (ID) [5,46] and out-of-domain (OOD) [19,23,25,26,57] generalization performance (Tables 3 and4, respectively). In-domain evaluation for ImageNet: reporting Top-1 accuracy in % on ImageNet-Validaton [14], ImageNet-ReaL [5] and ImageNet-V2 [46] with KD from a robust ResNet50 [25] teacher to ResNet18 (columns 2-4) and ViT-S (columns 5-7) students.\nTable 4: In-domain evaluation for ImageNet: Reporting Top-1 accuracy in % on ImageNet-A [26], ImageNet-R [25], ImageNet-Sketch [57] and ImageNet-Style [19] and mean-corruption-error on ImageNet-C (lower is better) [23]. data augmentations: we randomly switch between Cutmix [59] and Mixup [61], each drawing their interpolation weight from a β-distribution with α = 1, as well as AugMix [24] augmentations. For the standalone student training, we additionally apply various lighter data augmentations (Cutmix with α = 1, Mixup with α = 0.1, and Trivialaugment [39]). Since we ask how KD can be improved in a setting of limited resources, we run our experiments an order of magnitude shorter than proposed for the state-of-the-art in KD [6] (200 epochs for all ResNet18 and 150 epochs for all ViT-S experiments).\nResNet50 → ResNet18 ResNet50 → ViT-S Im-A Im-R Im-C ↓ Sketch Style Im-A Im-R Im-C ↓ Sketch\nFor student and KD models, we perform a small grid search over learning-rate and weight-decay hyperparameters. We then train the models with our HARD framework based on the hyperparameters of our best performing KD setting. The augmentor-specific settings are selected through a small grid-search in the ResNet18 setting (for details see Appendix A).\nWe first evaluate the ID performance of our methods (Table 3) beginning with the standalone teacher and student baselines, which reveal a larger performance gap between the ResNet18 student and the ResNet50 teacher compared to the ViT-S student (5.1% and 2.6% on the ImageNet validation set, respectively). Plain KD significantly reduces this gap for the ViT-S (+2.1% performance improvement compared to standalone). For the ResNet18 student KD achieves only small (0.7% V2) improvements or no improvements (0.0% Val), even though the initial gap between teacher and student is larger. Applying HARD-Affine, HARD-Mix and HARD-VAE augmentation on this task improves over plain KD across most augmentation models and test sets with student performance gains of up to 0.9% for ResNet18 (HARD-Affine) and 0.6% for ViT-S (HARD-VAE). For ViT-S, our best-performing HARD-VAE method even matches the teacher's performance on 2 out of 3 test sets.\nFor the OOD setting (Table 4), we observe that the initial gap between student and teacher is larger than on ID data across all data sets (up to 35.1% difference), except for Im-A in the ViT-S setting. The aggressive data augmentations we apply for the plain KD baseline favor OOD performance, hence it is expected that plain KD results in good performance improvement over the standalone baseline (up to 21.3% imporvement on Im-C). All three HARD approaches transfer some of the teacher's generalization abilities leading to improvements on a number of students and data sets, however, HARD-Affine fails to reach the KD performance in both settings and HARD-VAE underperforms for the ResNet18 student in these OOD scenarios. However, HARD-Mix and HARD-VAE (for ViT-S) outperform plain KD on several test sets and are roughly on par on all others, across the board. Given that we chose a very strong baseline by applying aggressive state-of-the-art data augmentations we find these results especially encouraging."},{"figure_ref":["fig_3","fig_3"],"heading":"Discussion","publication_ref":["b5","b12","b45","b48","b50","b51"],"table_ref":[],"text":"Original -Affine -VAE -Mix Interpretability HARD enables us to gain insight into the distillation mechanism as the augmented images illustrate the knowledge that is transferred (Figure 4). As expected, HARD-Affine learns to downscale images to shift and rotate the images such that the object in the image is shown in different places (row 2-4 in Figure 4) and scales such that the images is cropped (row 1). As HARD-Mix is a dynamically learnable extension of mixup, it either merges two objects into the same picture (row 1 and 4), especially if they are not in the same position, or uses one image to change the style (row 2) or background (row 3) of another. Finally, HARD-VAE mostly impacts the style of an image and additionally adds small distortions to specific image regions, which is noticable by the altered image brightness and the blurring of some highfrequency features.\nLimitations and broader impact State-ofthe-art knowledge distillation typically deals with huge models (billions of parameters) and incredibly long training times (>9,000 epochs) [6,13]. In comparison, our study is computationally lightweight in requiring approximately 400 A100GPU days across all our experiments. We believe exploring even more flexible augmentor models with a semantically meaningful latent space as for example diffusion models [44,47,49,50] could improve our proposed methods even further. However, generating a single image with out-of-the-box diffusion models requires multiple seconds. This is prohibitively long, so leave exploring their usability in our proposed dynamic data augmentation technique for future work. In general, KD allows us to distill smaller models that perform similar to large foundation models. Improving the distillation process to be more efficient lowers the barrier of applying KD across labs with various compute budget and decreases environmental impact. At the same time, transferring generalization abilities effectively and consistently results in smaller distilled models that are appealing to use, thus we would expect such smaller models to be used abundantly hence lowering the general carbon footprint for model usage. In conclusion, our study proposes avenues to efficiently improve KD in terms of performance, efficiency, and hence environmental impact."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"In this work we introduced a general, task-agnostic, and modular framework to extend knowledge distillation by learnable data augmentations. The augmentation models are optimized to generate inputs on which teacher and student disagree, keeping the teacher's predictions unchanged at the same time. We show that these augmentations can solve the issue of KD and transfer equivariance properties, even in cases where the teacher's inductive biases are distinct from the student's. We further demonstrate that our learned augmentations achieve performance competitive to established classical data augmentation techniques even when student and teacher share similar inductive biases. Overall our framework offers a powerful tool that enhances transfer performance and offers a unique insights into the transferred knowledge through its interpretable augmentations."},{"figure_ref":[],"heading":"A Setup Details","publication_ref":["b40"],"table_ref":[],"text":"Our experiments on MNIST were meant to reproduce Nix et al. [40] and thus follow their setup exactly, using the same training setup and model architectures."},{"figure_ref":[],"heading":"A.1 CIFAR10 Experiments","publication_ref":["b29"],"table_ref":[],"text":"Training We train on the entire CIFAR10 dataset (excluding 10% held-out as validation set) for 300 epochs with a batch-size of 256. As an optimizer, we use Adam [30] with a learning rate of 0.0003 and an L2-regularization of 2•10 -9 . Our training begins with a linear warmup of the learning rate for 20 epochs. The validation accuracy is monitored after every epoch and if it has not improved for 20 consecutive epochs, we decay the learning rate by a factor of 0.8 and restore the previously best performing model. The training is stopped prematurely if we decay five times."},{"figure_ref":[],"heading":"Models","publication_ref":["b9"],"table_ref":[],"text":"The different models we use generally follow the standard architecture and settings know from the literature. For the ViT, we use a smaller variant of it on the CIFAR task. It consists of six layers and eight attention heads throughout the network. The dropout rate is set to 0.1 and the hidden dimension is chosen as 512 in all places.\nKD and HARD After initial experiments on MNIST, we decided to use a softmax temperature of 5.0 for all experiments involving KD. We furthermore rely solely on the KL-Divergence loss to optimize our model. For the experiments with our augmentation framework, we have the same settings as before for the student (KD) training and separate settings for the augmentor training. There we have different settings depeding on whether we use the VAE augmentor (or the Affine augmentor). There we reduce the batch-size to 160 (128) and a learning-rate of 0.0001 (0.05). We initialize both augmentors to perform an identity transformation, i.e. the VAE is taken pretrained from Child [10]. The thresholds for switching are set as ℓ min = 10%(5%) and ℓ max = 60%(40%). The train modi are switched if the threshold is surpassed for 5 consecutive iterations. Both λ s and λ t are set to 1 for the experiments. For the experiment ResNet101 * → ResNet18, we found a slightly different setting to be more effective with ℓ min =5% and ℓ max = 40% and a switch only happening if the threshold is surpassed for 10 consecutive iterations."},{"figure_ref":[],"heading":"A.2 ImageNet Experiments","publication_ref":["b7","b36","b15","b62","b60","b39","b67","b52"],"table_ref":[],"text":"Baseline Training In general, all our ImageNet experiments follow a similar setup. We train with a batch-size of 512 samples using the Lion optimizer [8] with a linear learning-rate warmup to a defined initial learning-rate. Afterwards, we anneal the learning-rate following a cosine schedule [36] with a final value of 0. The training runs for 200 epochs for all ResNet18 experiments and 150 epochs for the ViT-S experiments. Throughout the training, the validation accuracy is monitored on a heldout set consisting of samples randomly chosen from training set, making up 1% of the total number of samples. The validation performance is used to pick the best performing epoch throughout training for final evaluation and the best hyperparameters during grid-search. We train at a resolution of 224 pixels with random resizing and cropping, as well as random horizontal flips applied in all trainings. All training runs are performed with automatic mixed precision and 8bit optimization [16].\nStudent Training After a grid-search, we found that for the standalone student training, an optimization with learning-rate 0.0001 with weight decay 0.1 for the ResNet18 student and learning-rate 0.00005 with weight decay 0.001 for the ViT-S student worked best. For both students, we apply light augmentations during training with Mixup (α = 0.2) [61] and CutMix (α = 1.0) [59]. For the ViT-S baseline, we additionally apply Trivial-Augment [39] and randomly erase pixels from the input image [66] with a probability of 0.1. We optimize the standard cross-entropy loss with additional label-smoothing [51] mixed in with a factor of 0.1."},{"figure_ref":[],"heading":"KD and HARD","publication_ref":["b23"],"table_ref":[],"text":"As described in the main paper, the configuration for the KD experiments (including HARD) mainly differ in the choice of augmentation, as well as learning-rate and weight-decay. The plain KD experiments use Mixup (α = 1.0) and CutMix (α = 1.0) as well as AugMix [24] augmentation. The softmax temperature was chosen as 1.0 in prior experiments and kept for all experiments. The learning-rate for all KD and HARD experiments was chosen through a grid-search"},{"figure_ref":[],"heading":"Acknowledgements","publication_ref":[],"table_ref":[],"text":"Furthermore, we thank Felix Schlüter for his helpful insights into evaluation problems as well as Mohammad Bashiri, Pawel Pierzchlewicz and Suhas Shrinivasan for helpful comments and discussions. The authors thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting Arne Nix and Max F. Burg. This work was supported by the Cyber Valley Research Fund (CyVy-RF-2019-01), by the German Federal Ministry of Education and Research (BMBF) through the Tübingen AI Center (FKZ: 01IS18039A), by the Deutsche Forschungsgemeinschaft (DFG) in the SFB 1233, Robust Vision: Inference Principles and Neural Mechanisms (TP12), project number: 276693517, and funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) -Project-ID 432680300 -SFB 1456. FHS is supported by the Carl-Zeiss-Stiftung and acknowledges the support of the DFG Cluster of Excellence \"Machine Learning -New Perspectives for Science\", EXC 2064/1, project number 390727645."},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"to be 0.0001 in all cases and weight-decay is 0.001 in most cases, except for HARD experiments with a ResNet18 student where a weight-decay of 0.05 is used."},{"figure_ref":[],"heading":"B Knowledge Distillation Results from the Literature","publication_ref":["b5","b42"],"table_ref":[],"text":"We (re-)evaluated student and teacher models from two high-performing KD experiments [6,41] in the literature on both in-domain and out-of-domain test sets. "}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b26\",\n \"b37\",\n \"b61\",\n \"b6\",\n \"b26\",\n \"b3\",\n \"b43\",\n \"b53\",\n \"b35\",\n \"b16\",\n \"b54\",\n \"b56\",\n \"b8\",\n \"b40\",\n \"b55\",\n \"b40\",\n \"b5\",\n \"b42\",\n \"b5\",\n \"b42\",\n \"b40\"\n ],\n \"table_ref\": [],\n \"text\": \"Knowledge distillation (KD) methods [27,37,60] are powerful and flexible tools to transfer the knowledge of a given teacher model to the transfer target, the student model, without copying the weights. Instead, these methods match the student's functional activity (e.g. the softmax output) to that of the teacher for the presented inputs. Hence, those methods are independent of architectural details and allow knowledge distillation to be applied in scenarios like model compression [7,27], continual learning [4,42,52], or even neuroscience [35], where traditional transfer learning would be impossible to use. KD methods also appear to be key to training new models that trade off inductive biases for more flexibility and more parameters [17,53,55] on smaller data [9,40,54]. However, Nix et al. [40] recently showed that current KD methods fail to transfer even simple equivariances between teacher and student. Additionally, previous work showed that KD leads to a larger gap between student and teacher on out-of-domain evaluation performance compared to within domain performance [6,41], even in cases where the student almost perfectly matches the teacher [6] (see Table 5). This phenomenon is especially pronounced for particularly robust teachers [41]. Thus we expect that transferring robustness properties is a difficult problem for KD in general.\\nFigure 1: Our task-agnostic HARD framework switches between training the student to match the teacher and training the augmentor to generate new samples on which the student underperforms while maintaining high teacher performance. We optimize the augmentor and student in interchanging phases through a student-teacher loss L s↔ t and teacher-teacher loss L t↔t . We switch between the two phases by comparing the default loss L s on augmented data to pre-defined thresholds.\\nWe hypothesize that KD methods are in principle capable of transferring most knowledge from a teacher to a student if the training data is chosen adequately. We confirm this hypothesis on a small toy example (Section 3), showing the importance of input data for KD. Motivated by this demonstration, we propose our Hard Augmentations for Robust Distillation (HARD) method, a general framework (Section 4) to generate augmented training inputs which improve knowledge transfer by maximizing the distance between teacher and student while leaving the teacher's output unchanged. Consequently, our framework moves the input in directions that the teacher is invariant to but which are most challenging for the student. Our experiments (Section 5) show that our task-agnostic framework improves transfer effectiveness and thereby solves the problem of of KD not being able to transfer shift equivariance [40]. Additionally, as part of our framework, we propose several parameterized augmentations (Section 4.1) that can be integrated with most existing KD methods and are applicable to a variety of different computer vision tasks. Finally, we demonstrate across multiple different models on the tasks of CIFAR10 and ImageNet that our framework learns interpretable augmentations that improve KD to the same level and in many cases even beyond established data augmentation methods, even when evaluated in an out-of-domain setting.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b17\",\n \"b32\",\n \"b11\",\n \"b23\",\n \"b27\",\n \"b39\",\n \"b60\",\n \"b62\",\n \"b23\",\n \"b62\",\n \"b60\",\n \"b10\",\n \"b24\",\n \"b49\",\n \"b68\",\n \"b1\",\n \"b2\",\n \"b19\",\n \"b57\",\n \"b64\",\n \"b65\",\n \"b66\",\n \"b5\",\n \"b59\",\n \"b44\",\n \"b20\",\n \"b46\",\n \"b63\",\n \"b28\",\n \"b62\",\n \"b9\",\n \"b30\",\n \"b34\"\n ],\n \"table_ref\": [],\n \"text\": \"There is a long tradition in using data augmentations to artificially extend training data for deep learning models and particularly in computer vision, be it through adding Gaussian noise, random crops, shifts, flips, or rotations [18,33]. In recent years, data augmentations became more complex [12,24,28,39,59,61], employing a multitude of different heuristics with the aim to improve generalization and in some cases also out-of-domain performance [24]. A particularly popular augmentation method is Mixup [61], which randomly interpolates two input samples and their labels respectively. Similarly, Cutmix [59] combines two input images by pasting a random crop of one image on top of the other. Also, many studies use parameterized augmentations optimized to improve a given objective [11,25,48,67], and some even optimize the augmentations to improve on an adversarial objective [2,3,20,56,[63][64][65], however, without applying them for knowledge transfer.\\nIn KD, applying data augmentations is a very effective tool to improve matching between student and teacher [6,58] and optimizing on a meta level can be useful to aide the teaching [43]. Similar to our work, Haidar et al. [21], Rashid et al. [45], Zhang et al. [62] utilized adversarial objectives to optimize data augmentations for KD, however, they were solely focused on natural language processing tasks and do not optimize the augmentations towards invariance.\\nInspired by this large body of work we formulate a task-agnostic framework containing only one building block that is specific to the data-domain -the instantiation of the augmentor model generating the augmented data samples -for which we offer a variety of reasonable model choices based on spatial transformer modules [29], Mixup [61], and variational autoencoders [10,31,34]. \"\n },\n {\n \"figure_ref\": [\n \"fig_2\",\n \"fig_2\",\n \"fig_2\",\n \"fig_2\",\n \"fig_2\",\n \"fig_2\",\n \"fig_2\",\n \"fig_2\"\n ],\n \"heading\": \"Input Data Matters for Functional Transfer\",\n \"publication_ref\": [\n \"b0\",\n \"b5\",\n \"b62\"\n ],\n \"table_ref\": [],\n \"text\": \"We hypothesize that the choice of input data is crucial to successfully knowledge distillation and we illustrate the impact of training data by a simple toy example. To demonstrate this, consider a simple KD task in which we instantiate the teacher-model by the true function f t (x) = cos(x) and the student f s (x) by a three layer Multilayer Perceptron (MLP) with ReLU activation [1]. We use input data x chosen such that it does not capture the teacher's cos(x) periodicity (orange points in Figure 2A). Simple KD does neither interpolate between the given training points nor extrapolates beyond them (Figure 2E). Hence the student neural network does not learn the teacher's periodicity and fails to interpolate and extrapolate beyond the training data (Figure 2A).\\nAugmenting the training data with more helpful inputs x and teacher labels f t (x) = cos(x) could mitigate this problem. One method successfully applied to KD [6] is to extend the input data through Mixup [61]. When applying this to our illustrative example, we create new training inputs x through linear interpolation between pairs of input points x = (1 -α)x 1 + αx 2 (Figure 2F), and recording the corresponding teacher responses f t (x) = cos(x). Thus, the student learns to interpolate between training points, but mixup does not enhance extrapolation (Figure 2B).\\nTo generate datapoints that would interpolate and extrapolate beyond already available training points, we could simply augment by adding Gaussian noise ϵ to the available data points, x = x + ϵ, hence interpolating and extrapolating beyond the training data (Figure 2G). This strategy helps our student to match the teacher also outside the original training regime (Figure 2C). However, the student only improves within a fixed margin that is determined by the noise distribution's mean and variance.\\nWe could obviously improve interpolation and extrapolation by increasing the noise distribution's variance or shifting it's mean, however, as we move to a high dimensional image input space (x ∈ R → ⃗ x ∈ R N ) it becomes unclear how to heuristically select helpful new samples and at the same time random exploration strategies become computationally infeasible. Instead, we propose to optimize a parameterized augmentation to efficiently generate new, hard training samples on which the student lacks performance, as here the student could improve the most. In our toy example, we illustrate this by optimizing the Gaussian's parameters (mean and variance) according to our augmentation framework HARD, which we will present in the next section. This provides us with a noise distribution which we use to draw new helpful training examples x that transfer inter-and extrapolation to the student network (Figure 2D,H). Overall, this toy example shows that learning hard augmentations to select new helpful data points is crucial to efficiently improve extrapolation beyond the training distribution.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Learning Hard Augmentations for Robust Distillation (HARD)\",\n \"publication_ref\": [\n \"b26\"\n ],\n \"table_ref\": [],\n \"text\": \"Our task-agnostic HARD framework learns augmenting training images to most efficiently help knowledge-transfer from a teacher to a student model. Our method requires three main components: a teacher model with frozen parameters, a student model that should learn knowledge from the teacher, and a parameterized augmentation model that learns to augment images such that most of the teacher's knowledge is transferred to the student.\\nIn classical KD methods [27], the objective is to minimize a distance D [f s (x), f t (x)] between the student's activation f s (x) and the teacher's activation f t (x) on given inputs x ∈ R n . Usually, this would be the Kullback-Leibler divergence between the softmax distributions of teacher and student. Unfortunately, only considering training data could miss properties of the teacher (eg. shift invariance) that might be crucial for generalization (see Section 3 for an illustrative example). To resolve this issue, we learn a parametrized augmentation model g a to generate new input data points x = g a (x) transferring such invariance properties from the teacher to the student. Hence, we define a teacher-student loss considering the more general case of matching student and teacher on augmented inputs x ∈ R n :\\nL s↔ t = D [f s (x), f t (x)] .(1)\\nTo specifically transfer the teacher's invariance properties to the student, we propose a teacher-teacher loss pushing the augmentor towards generating data points on which the teacher is invariant,\\nL t↔t = D [f t (x), f t (x)] ,(2)\\nas these are often useful augmentations for generalization. Using both of these losses, we optimize the augmentor's parameters θ a to generate augmented samples on which the teacher results in similar activations but the student differs from them (Figure 1 top) and simultaneously we optimize the student's parameters θ s to perform well on those augmentations (Figure 1 bottom):\\nmax θa λ s L s↔ t -λ t L t↔t and min θs L s↔ t .(3)\\nHere, λ s and λ t trade off the loss terms and are treated as hyper-parameters. We train both components separately switching from training the augmentor to training the student when the student's performance on augmented data gets worse than a pre-defined threshold (L s > ℓ max ) and we switch back from student to augmentor training when the student's performance on augmented data surpasses a pre-defined threshold (L s < ℓ min ; Figure 1). To prevent catastrophic forgetting, we save augmentors at every switch and employ an augmentor randomly chosen out of the set of previously saved augmentors in each iteration when training the student.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"The augmentor models\",\n \"publication_ref\": [\n \"b30\",\n \"b28\"\n ],\n \"table_ref\": [],\n \"text\": \"To generate new input data points it is important to choose an augmentor that suits the desired application and is powerful enough to generate useful augmentations. Usually, we do not know a priori what useful augmentations are and thus should try to allow as much flexibility as possible. Additionally, some variance over augmentations could benefit the transfer. Thus, all augmentors in our study introduce randomness in the model by adding Gaussian noise into the computation of the augmentation through the reparametrization trick [31]. While our framework is universally applicable across domains, choosing an effective augmentation model likely needs to be addressed for each task individually. In our experiments, we use the following augmentor models:\\nHARD-Affine In the simplest model, we limit the augmentations to affine transformations of the coordinate grid of pixel locations, i.e. shifts, rotations, scalings, and shears of images. Models implementing such transformations are known as spatial transformers [29]. We leverage this model for our augmentor by learning a distribution over the entries of an affine transformation matrix ϑ ∈ R 2×3 that defines the transformation of the sampling grid, i.e. a transformation that maps the pixel positions from the original image to the augmented image (Figure 3A).\"\n },\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_1\"\n ],\n \"heading\": \"HARD-Mix\",\n \"publication_ref\": [\n \"b62\",\n \"b60\",\n \"b30\",\n \"b9\",\n \"b34\"\n ],\n \"table_ref\": [],\n \"text\": \"Additionally we consider a slightly more complex augmentor model, which is an adaptive variant of the commonly used Mixup [61] and Cutmix [59] augmentations. However, instead of randomly sampling the ratio and cutout position that are used to combine images, we learn how to combine the images dependent on the input images. We achieve this by performing a patch-wise projection of the input image, followed by comparing each patch with the same query vector sampled from a learned distribution (Figure 3B). We normalize similarities for each patch over each group of images and use the resulting weights to combine the original image patches, giving a combined image. This mechanism allows our augmentor to decide which features of which image are shown to the student, enabling it to explore the interpolated space between images systematically, instead of randomly. As it would not make sense for the teacher to be invariant to an interpolation as it is generated by HARD-Mix, we do not consider the teacher-teacher-loss L t↔t in this case and optimize student and augmentor jointly instead.\\nHARD-VAE To lift constraints further, we wanted to use a more powerful augmentor that could generate a large variety of images across the entire image-space. As the augmentor has to generate new samples on-the-fly during the student training, the generation process needs to be very fast, limiting the choice of useful generative models. For this reason, we focus on variants of the variational autoencoder architecture [31], allowing for good image reconstructions which can be achieved reasonably fast in a single forward pass (Figure 3D). For CIFAR, we choose the very deep VAE [10] model, which we finetune by solely optimizing parameters of the posterior network from layer 10 onward in the decoder. For the experiments on ImageNet, we use a Residual-Quantized VAE (RQ-VAE) [34] pretrained on ImageNet, which we finetune in its entirety and add a noise vector on the latent state. Hence, as training progresses, the model changes from generating plain reconstructions of a given image to input conditioned generations that serve as our augmentations.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Transferring equivariance\",\n \"publication_ref\": [\n \"b40\",\n \"b21\",\n \"b16\",\n \"b14\",\n \"b40\",\n \"b38\",\n \"b40\",\n \"b40\"\n ],\n \"table_ref\": [],\n \"text\": \"For our initial experiment, we reproduce the setup from Nix et al. [40] to test whether we can transfer the inductive bias from a shift equivariant teacher, CNN and ResNet18 [22], to a student that does not have this inductive bias built into its architecture: a Multi-Layer Perceptron (MLP) and a Vision Transformer (ViT) [17]. When training the students and teachers by themselves on standard MNIST [15] training data, we observe a small drop of generalization performance (-0.6% and -1.2%) between Table 1: MNIST (columns \\\"Centered\\\") and MNIST-C (columns \\\"Shifted\\\") test accuracies (mean and standard error of the mean across 4 random seeds) comparing KD without augmentation and our HARD-Affine method to Orbit transfer [40], which also learns and transfers equivariances. teacher and student on the MNIST test set and a large gap (-56.1% and -52.4%) when we evaluate on a version of the test set in which digits were randomly shifted [38]. As another baseline, we applied plain KD to transfer shift equivariance from teacher to student. Consistent with the findings of Nix et al. [40], we only observe a small improvement on the centered (+0.2% and +0.3%) and the shifted (+5.1% and +4.3%) test sets, which likely result from the centered training data we use for transfer.\\nWe then test if combining KD with our augmentations produced by HARD-Affine would outperform these baselines. The resulting student model improves significantly on shifted inputs (+28.6% and +39.4%) compared to plain KD and the generated images clearly show that the augmentor learns to shift the digits within the image. Compared to Nix et al. [40] our approach outperforms their results on the ViT task but, while improving the out-of-domain generalization by 28.6% over baseline, stays behind the Orbit performance on the MLP task. This demonstrates that our method while acting on fewer parts of the network compared to Orbit and while being a more general method, can improve or reach better performance when it comes to transferring invariances, and can be generalized to bigger datasets, as we show below.\\nWe verify that the student's performance improvement is specifically due to our data generation framework in two control experiments. The first experiment (Random Affine) augments the training inputs of a stand-alone student model with a random affine transformation akin to our augmentor model, but using transformation parameters sampled uniformly from a pre-defined, reasonably constrained range (i.e. ensuring the digit is always fully visible). This student performs well on the shifted test set, however, performance significantly degrades on the centered test set. In comparison, our HARD-Affine model is unconstrained and learns more useful augmentations, leading to better performance on the centered test sets.\\nIn our second control (Shifts) we asked how much data augmentation could improve the performance in the best case (without KD). For this, we augment the inputs by the same random shifts that were applied to obtain the shifted test data, leading to great improvements on the shifted test set. However, our learned augmentations achieve scores in a similar range on the shifted evaluation and outperform its results on the centered test set.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Transfer on natural images\",\n \"publication_ref\": [\n \"b31\",\n \"b13\",\n \"b21\",\n \"b24\",\n \"b16\"\n ],\n \"table_ref\": [\n \"tab_0\",\n \"tab_0\",\n \"tab_0\",\n \"tab_0\",\n \"tab_0\",\n \"tab_0\"\n ],\n \"text\": \"After demonstrating that our method successfully captures the difference between teacher and student and bridges a gap in inductive bias, we now want to test whether this effect holds up in more realistic scenarios. CIFAR experiments We begin by applying our framework to CIFAR10 [32] on three different KD scenarios (see Table 2). Specifically, we test scenarios where the student lacks an inductive bias (ResNet18→ViT), where the teacher has more capacity and access to data than the student (ResNet101 * →ResNet18), and to scenarios combining both properties (ResNet101 * →ViT). For all experiments, we keep the experimental setup as close to our previous MNIST experiments as possible (see Appendix A for details).\\nWe start by establishing baselines by training only the teacher and only the student models on the data and evaluating default KD. We observe that on this small data set a small ResNet18 performs better (78.5% accuracy) than a larger ViT (68.5%), likely because of the ResNet's superior inductive bias on this task and small data set. Next, we find that adding default data augmentations (random rotations, cropping, horizontal flips) to the student baselines significantly boosts performance to 92.6% and 78.3% for the ResNet18 and ViT, respectively. Adding these default augmentations to typical KD leads to a great performance boost, too (see Table 2).\\nGiven that adding default data augmentation to KD already leads to a substantial performance boost, it is particularly noteworthy that the data augmentations learned by HARD-Affine outperform this baseline for the ViT. Qualitatively, the augmented images exhibit a large variety of spatial transformations, suggesting that a difference in these examples lead to the observed performance boost (Table 2, right).\\nWe then investigated performance of our HARD-VAE augmentation strategy and found performance improvement over the KD + standard augmentations baseline for transfer to the ViT (+1.0% and +1.9%) student. However, inspecting the augmented images indicates that our augmentor lacks the expected shifts of object positions, but rather learns stylistic changes in the image (Table 2, right). This motivated us to combine HARD-Affine and HARD-VAE augmentation resulting in best performance (up to +7.8%) for all teacher-student pairings (HARD-VAE-Affine in Table 2) and the resulting images demonstrate variability in both style and spatial alignment (Table 2, right).\\nImageNet experiments Having established our methods' performance for CIFAR10, we extend our results to classification on ImageNet [14]. Here we aim to distill a ResNet50 [22] teacher, trained with Deep-augment and AugMix data augmentations [25], into a smaller ResNet18 and ViT-S (small vision transformer variant) [17] that we want to be particularly robust to natural image corruptions,.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"The distillation into ResNet18 allows us to investigate the capability for model compression, because\",\n \"publication_ref\": [\n \"b4\",\n \"b47\",\n \"b18\",\n \"b22\",\n \"b24\",\n \"b25\",\n \"b58\",\n \"b13\",\n \"b4\",\n \"b47\",\n \"b24\",\n \"b25\",\n \"b24\",\n \"b58\",\n \"b18\",\n \"b22\",\n \"b60\",\n \"b62\",\n \"b23\",\n \"b39\",\n \"b5\"\n ],\n \"table_ref\": [\n \"tab_1\",\n \"tab_1\"\n ],\n \"text\": \"ResNet18 is a smaller network compared to ResNet50, but with a similar architecture. Distillation into a ViT-S architecture with a patch-size of 14 tests additionally if KD transfers the ResNet50's inductive bias of shift equivariance on a larger dataset.\\nWe evaluate on common test sets for both in-domain (ID) [5,46] and out-of-domain (OOD) [19,23,25,26,57] generalization performance (Tables 3 and4, respectively). In-domain evaluation for ImageNet: reporting Top-1 accuracy in % on ImageNet-Validaton [14], ImageNet-ReaL [5] and ImageNet-V2 [46] with KD from a robust ResNet50 [25] teacher to ResNet18 (columns 2-4) and ViT-S (columns 5-7) students.\\nTable 4: In-domain evaluation for ImageNet: Reporting Top-1 accuracy in % on ImageNet-A [26], ImageNet-R [25], ImageNet-Sketch [57] and ImageNet-Style [19] and mean-corruption-error on ImageNet-C (lower is better) [23]. data augmentations: we randomly switch between Cutmix [59] and Mixup [61], each drawing their interpolation weight from a β-distribution with α = 1, as well as AugMix [24] augmentations. For the standalone student training, we additionally apply various lighter data augmentations (Cutmix with α = 1, Mixup with α = 0.1, and Trivialaugment [39]). Since we ask how KD can be improved in a setting of limited resources, we run our experiments an order of magnitude shorter than proposed for the state-of-the-art in KD [6] (200 epochs for all ResNet18 and 150 epochs for all ViT-S experiments).\\nResNet50 → ResNet18 ResNet50 → ViT-S Im-A Im-R Im-C ↓ Sketch Style Im-A Im-R Im-C ↓ Sketch\\nFor student and KD models, we perform a small grid search over learning-rate and weight-decay hyperparameters. We then train the models with our HARD framework based on the hyperparameters of our best performing KD setting. The augmentor-specific settings are selected through a small grid-search in the ResNet18 setting (for details see Appendix A).\\nWe first evaluate the ID performance of our methods (Table 3) beginning with the standalone teacher and student baselines, which reveal a larger performance gap between the ResNet18 student and the ResNet50 teacher compared to the ViT-S student (5.1% and 2.6% on the ImageNet validation set, respectively). Plain KD significantly reduces this gap for the ViT-S (+2.1% performance improvement compared to standalone). For the ResNet18 student KD achieves only small (0.7% V2) improvements or no improvements (0.0% Val), even though the initial gap between teacher and student is larger. Applying HARD-Affine, HARD-Mix and HARD-VAE augmentation on this task improves over plain KD across most augmentation models and test sets with student performance gains of up to 0.9% for ResNet18 (HARD-Affine) and 0.6% for ViT-S (HARD-VAE). For ViT-S, our best-performing HARD-VAE method even matches the teacher's performance on 2 out of 3 test sets.\\nFor the OOD setting (Table 4), we observe that the initial gap between student and teacher is larger than on ID data across all data sets (up to 35.1% difference), except for Im-A in the ViT-S setting. The aggressive data augmentations we apply for the plain KD baseline favor OOD performance, hence it is expected that plain KD results in good performance improvement over the standalone baseline (up to 21.3% imporvement on Im-C). All three HARD approaches transfer some of the teacher's generalization abilities leading to improvements on a number of students and data sets, however, HARD-Affine fails to reach the KD performance in both settings and HARD-VAE underperforms for the ResNet18 student in these OOD scenarios. However, HARD-Mix and HARD-VAE (for ViT-S) outperform plain KD on several test sets and are roughly on par on all others, across the board. Given that we chose a very strong baseline by applying aggressive state-of-the-art data augmentations we find these results especially encouraging.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\",\n \"fig_3\"\n ],\n \"heading\": \"Discussion\",\n \"publication_ref\": [\n \"b5\",\n \"b12\",\n \"b45\",\n \"b48\",\n \"b50\",\n \"b51\"\n ],\n \"table_ref\": [],\n \"text\": \"Original -Affine -VAE -Mix Interpretability HARD enables us to gain insight into the distillation mechanism as the augmented images illustrate the knowledge that is transferred (Figure 4). As expected, HARD-Affine learns to downscale images to shift and rotate the images such that the object in the image is shown in different places (row 2-4 in Figure 4) and scales such that the images is cropped (row 1). As HARD-Mix is a dynamically learnable extension of mixup, it either merges two objects into the same picture (row 1 and 4), especially if they are not in the same position, or uses one image to change the style (row 2) or background (row 3) of another. Finally, HARD-VAE mostly impacts the style of an image and additionally adds small distortions to specific image regions, which is noticable by the altered image brightness and the blurring of some highfrequency features.\\nLimitations and broader impact State-ofthe-art knowledge distillation typically deals with huge models (billions of parameters) and incredibly long training times (>9,000 epochs) [6,13]. In comparison, our study is computationally lightweight in requiring approximately 400 A100GPU days across all our experiments. We believe exploring even more flexible augmentor models with a semantically meaningful latent space as for example diffusion models [44,47,49,50] could improve our proposed methods even further. However, generating a single image with out-of-the-box diffusion models requires multiple seconds. This is prohibitively long, so leave exploring their usability in our proposed dynamic data augmentation technique for future work. In general, KD allows us to distill smaller models that perform similar to large foundation models. Improving the distillation process to be more efficient lowers the barrier of applying KD across labs with various compute budget and decreases environmental impact. At the same time, transferring generalization abilities effectively and consistently results in smaller distilled models that are appealing to use, thus we would expect such smaller models to be used abundantly hence lowering the general carbon footprint for model usage. In conclusion, our study proposes avenues to efficiently improve KD in terms of performance, efficiency, and hence environmental impact.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this work we introduced a general, task-agnostic, and modular framework to extend knowledge distillation by learnable data augmentations. The augmentation models are optimized to generate inputs on which teacher and student disagree, keeping the teacher's predictions unchanged at the same time. We show that these augmentations can solve the issue of KD and transfer equivariance properties, even in cases where the teacher's inductive biases are distinct from the student's. We further demonstrate that our learned augmentations achieve performance competitive to established classical data augmentation techniques even when student and teacher share similar inductive biases. Overall our framework offers a powerful tool that enhances transfer performance and offers a unique insights into the transferred knowledge through its interpretable augmentations.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Setup Details\",\n \"publication_ref\": [\n \"b40\"\n ],\n \"table_ref\": [],\n \"text\": \"Our experiments on MNIST were meant to reproduce Nix et al. [40] and thus follow their setup exactly, using the same training setup and model architectures.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.1 CIFAR10 Experiments\",\n \"publication_ref\": [\n \"b29\"\n ],\n \"table_ref\": [],\n \"text\": \"Training We train on the entire CIFAR10 dataset (excluding 10% held-out as validation set) for 300 epochs with a batch-size of 256. As an optimizer, we use Adam [30] with a learning rate of 0.0003 and an L2-regularization of 2•10 -9 . Our training begins with a linear warmup of the learning rate for 20 epochs. The validation accuracy is monitored after every epoch and if it has not improved for 20 consecutive epochs, we decay the learning rate by a factor of 0.8 and restore the previously best performing model. The training is stopped prematurely if we decay five times.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Models\",\n \"publication_ref\": [\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"The different models we use generally follow the standard architecture and settings know from the literature. For the ViT, we use a smaller variant of it on the CIFAR task. It consists of six layers and eight attention heads throughout the network. The dropout rate is set to 0.1 and the hidden dimension is chosen as 512 in all places.\\nKD and HARD After initial experiments on MNIST, we decided to use a softmax temperature of 5.0 for all experiments involving KD. We furthermore rely solely on the KL-Divergence loss to optimize our model. For the experiments with our augmentation framework, we have the same settings as before for the student (KD) training and separate settings for the augmentor training. There we have different settings depeding on whether we use the VAE augmentor (or the Affine augmentor). There we reduce the batch-size to 160 (128) and a learning-rate of 0.0001 (0.05). We initialize both augmentors to perform an identity transformation, i.e. the VAE is taken pretrained from Child [10]. The thresholds for switching are set as ℓ min = 10%(5%) and ℓ max = 60%(40%). The train modi are switched if the threshold is surpassed for 5 consecutive iterations. Both λ s and λ t are set to 1 for the experiments. For the experiment ResNet101 * → ResNet18, we found a slightly different setting to be more effective with ℓ min =5% and ℓ max = 40% and a switch only happening if the threshold is surpassed for 10 consecutive iterations.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.2 ImageNet Experiments\",\n \"publication_ref\": [\n \"b7\",\n \"b36\",\n \"b15\",\n \"b62\",\n \"b60\",\n \"b39\",\n \"b67\",\n \"b52\"\n ],\n \"table_ref\": [],\n \"text\": \"Baseline Training In general, all our ImageNet experiments follow a similar setup. We train with a batch-size of 512 samples using the Lion optimizer [8] with a linear learning-rate warmup to a defined initial learning-rate. Afterwards, we anneal the learning-rate following a cosine schedule [36] with a final value of 0. The training runs for 200 epochs for all ResNet18 experiments and 150 epochs for the ViT-S experiments. Throughout the training, the validation accuracy is monitored on a heldout set consisting of samples randomly chosen from training set, making up 1% of the total number of samples. The validation performance is used to pick the best performing epoch throughout training for final evaluation and the best hyperparameters during grid-search. We train at a resolution of 224 pixels with random resizing and cropping, as well as random horizontal flips applied in all trainings. All training runs are performed with automatic mixed precision and 8bit optimization [16].\\nStudent Training After a grid-search, we found that for the standalone student training, an optimization with learning-rate 0.0001 with weight decay 0.1 for the ResNet18 student and learning-rate 0.00005 with weight decay 0.001 for the ViT-S student worked best. For both students, we apply light augmentations during training with Mixup (α = 0.2) [61] and CutMix (α = 1.0) [59]. For the ViT-S baseline, we additionally apply Trivial-Augment [39] and randomly erase pixels from the input image [66] with a probability of 0.1. We optimize the standard cross-entropy loss with additional label-smoothing [51] mixed in with a factor of 0.1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"KD and HARD\",\n \"publication_ref\": [\n \"b23\"\n ],\n \"table_ref\": [],\n \"text\": \"As described in the main paper, the configuration for the KD experiments (including HARD) mainly differ in the choice of augmentation, as well as learning-rate and weight-decay. The plain KD experiments use Mixup (α = 1.0) and CutMix (α = 1.0) as well as AugMix [24] augmentation. The softmax temperature was chosen as 1.0 in prior experiments and kept for all experiments. The learning-rate for all KD and HARD experiments was chosen through a grid-search\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgements\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Furthermore, we thank Felix Schlüter for his helpful insights into evaluation problems as well as Mohammad Bashiri, Pawel Pierzchlewicz and Suhas Shrinivasan for helpful comments and discussions. The authors thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting Arne Nix and Max F. Burg. This work was supported by the Cyber Valley Research Fund (CyVy-RF-2019-01), by the German Federal Ministry of Education and Research (BMBF) through the Tübingen AI Center (FKZ: 01IS18039A), by the Deutsche Forschungsgemeinschaft (DFG) in the SFB 1233, Robust Vision: Inference Principles and Neural Mechanisms (TP12), project number: 276693517, and funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) -Project-ID 432680300 -SFB 1456. FHS is supported by the Carl-Zeiss-Stiftung and acknowledges the support of the DFG Cluster of Excellence \\\"Machine Learning -New Perspectives for Science\\\", EXC 2064/1, project number 390727645.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"to be 0.0001 in all cases and weight-decay is 0.001 in most cases, except for HARD experiments with a ResNet18 student where a weight-decay of 0.05 is used.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B Knowledge Distillation Results from the Literature\",\n \"publication_ref\": [\n \"b5\",\n \"b42\"\n ],\n \"table_ref\": [],\n \"text\": \"We (re-)evaluated student and teacher models from two high-performing KD experiments [6,41] in the literature on both in-domain and out-of-domain test sets. \"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-05-25"},"doi":{"kind":"string","value":"10.1109/WACV.2019.00137"},"references":{"kind":"list like","value":[{"authors":"Abien Fred; Agarap ","journal":"","ref_id":"b0","title":"Deep learning using rectified linear units (relu)","year":"2018"},{"authors":"Anthreas Antoniou; Amos Storkey; Harrison Edwards","journal":"","ref_id":"b1","title":"Data Augmentation Generative Adversarial Networks","year":"2022-02"},{"authors":"Sima Behpour; Kris M Kitani; Brian D Ziebart","journal":"","ref_id":"b2","title":"ADA: Adversarial data augmentation for object detection","year":"2019-03"},{"authors":"David Ari S Benjamin; Konrad P Rolnick; Kording","journal":"","ref_id":"b3","title":"Measuring and regularizing networks in function space","year":"2019"},{"authors":"Lucas Beyer; Olivier J Hénaff; Alexander Kolesnikov; Xiaohua Zhai; Aäron Van Den Oord","journal":"","ref_id":"b4","title":"Are we done with imagenet?","year":"2020"},{"authors":"Lucas Beyer; Xiaohua Zhai; Amélie Royer; Larisa Markeeva; Rohan Anil; Alexander Kolesnikov","journal":"","ref_id":"b5","title":"Knowledge distillation: A good teacher is patient and consistent","year":"2021"},{"authors":"Cristian Bucilǎ; Rich Caruana; Alexandra Niculescu-Mizil","journal":"","ref_id":"b6","title":"Model compression","year":"2006"},{"authors":"Xiangning Chen; Chen Liang; Da Huang; Esteban Real; Kaiyuan Wang; Yao Liu; Hieu Pham; Xuanyi Dong; Thang Luong; Cho-Jui Hsieh; Yifeng Lu; Quoc V Le","journal":"","ref_id":"b7","title":"Symbolic discovery of optimization algorithms","year":"2023"},{"authors":"Xianing Chen; Qiong Cao; Yujie Zhong; Jing Zhang; Shenghua Gao; Dacheng Tao","journal":"","ref_id":"b8","title":"DearKD: Data-Efficient Early Knowledge Distillation for Vision Transformers","year":"2022-04"},{"authors":"Rewon Child","journal":"","ref_id":"b9","title":"Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them on Images","year":"2020-11"},{"authors":"D Ekin; Barret Cubuk; Dandelion Zoph; Vijay Mane; Quoc V Vasudevan; Le","journal":"Cvpr","ref_id":"b10","title":"AutoAugment: Learning Augmentation Policies from Data","year":"2018"},{"authors":"Barret Ekin D Cubuk; Jonathon Zoph; Quoc V Shlens; Le","journal":"","ref_id":"b11","title":"Randaugment: Practical automated data augmentation with a reduced search space","year":"2020"},{"authors":"Mostafa Dehghani; Josip Djolonga; Basil Mustafa; Piotr Padlewski; Jonathan Heek; Justin Gilmer; Andreas Steiner; Mathilde Caron; Robert Geirhos; Ibrahim Alabdulmohsin; Rodolphe Jenatton; Lucas Beyer; Michael Tschannen; Anurag Arnab; Xiao Wang; Carlos Riquelme; Matthias Minderer; Joan Puigcerver; Utku Evci; Manoj Kumar; Sjoerd Van Steenkiste; Gamaleldin F Elsayed; Aravindh Mahendran; Fisher Yu; Avital Oliver; Fantine Huot; Jasmijn Bastings; Mark Patrick Collier; Alexey Gritsenko; Vighnesh Birodkar; Cristina Vasconcelos; Yi Tay; Thomas Mensink; Alexander Kolesnikov; Filip Pavetić; Dustin Tran; Thomas Kipf; Mario Lučić; Xiaohua Zhai; Daniel Keysers; Jeremiah Harmsen; Neil Houlsby","journal":"","ref_id":"b12","title":"Scaling vision transformers to 22 billion parameters","year":"2023"},{"authors":"Jia Deng; Wei Dong; Richard Socher; Li-Jia Li; Kai Li; Li Fei-Fei","journal":"Ieee","ref_id":"b13","title":"Imagenet: A largescale hierarchical image database","year":"2009"},{"authors":"Li Deng","journal":"IEEE Signal Processing Magazine","ref_id":"b14","title":"The mnist database of handwritten digit images for machine learning research","year":"2012"},{"authors":"Tim Dettmers; Mike Lewis; Sam Shleifer; Luke Zettlemoyer","journal":"ICLR","ref_id":"b15","title":"8-bit optimizers via block-wise quantization","year":"2022"},{"authors":"Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly; Jakob Uszkoreit; Neil Houlsby","journal":"","ref_id":"b16","title":"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale","year":"2020"},{"authors":"Logan Engstrom; Brandon Tran; Dimitris Tsipras; Ludwig Schmidt; Aleksander Madry","journal":"","ref_id":"b17","title":"A rotation and a translation suffice: Fooling cnns with simple transformations","year":"2017"},{"authors":"Robert Geirhos; Claudio Michaelis; Felix A Wichmann; Patricia Rubisch; Matthias Bethge; Wieland Brendel; . Imagenet-Trained Cnns Are Biased Towards Texture","journal":"","ref_id":"b18","title":"INCREASING SHAPE BIAS IMPROVES ACCURACY AND ROBUSTNESS","year":"2018"},{"authors":"Kehong Gong; Jianfeng Zhang; Jiashi Feng","journal":"","ref_id":"b19","title":"PoseAug: A Differentiable Pose Augmentation Framework for 3D Human Pose Estimation","year":"2021"},{"authors":"Md Akmal Haidar; Mehdi Rezagholizadeh; Abbas Ghaddar; Khalil Bibi; Philippe Langlais; Pascal Poupart","journal":"","ref_id":"b20","title":"CILDA: Contrastive Data Augmentation using Intermediate Layer Knowledge Distillation","year":"2022-04"},{"authors":"Kaiming He; Xiangyu Zhang; Shaoqing Ren; Jian Sun","journal":"","ref_id":"b21","title":"Deep Residual Learning for Image Recognition","year":"2015"},{"authors":"Dan Hendrycks; Thomas Dietterich","journal":"mar","ref_id":"b22","title":"Benchmarking neural network robustness to common corruptions and perturbations","year":"2019"},{"authors":"Dan Hendrycks; Norman Mu; D Ekin; Barret Cubuk; Justin Zoph; Balaji Gilmer; Lakshminarayanan","journal":"","ref_id":"b23","title":"Augmix: A simple data processing method to improve robustness and uncertainty","year":"2019"},{"authors":"Dan Hendrycks; Steven Basart; Norman Mu; Saurav Kadavath; Frank Wang; Evan Dorundo; Rahul Desai; Tyler Zhu; Samyak Parajuli; Mike Guo; Dawn Song; Jacob Steinhardt; Justin Gilmer","journal":"","ref_id":"b24","title":"The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization","year":"2021"},{"authors":"Dan Hendrycks; Kevin Zhao; Steven Basart; Jacob Steinhardt; Dawn Song","journal":"CVPR","ref_id":"b25","title":"Natural adversarial examples","year":"2021"},{"authors":"Geoffrey Hinton; Jeff Dean","journal":"","ref_id":"b26","title":"Distilling the Knowledge in a Neural Network","year":"2015"},{"authors":"Amir Philip Tg Jackson; Stephen Atapour Abarghouei; Toby P Bonner; Boguslaw Breckon; Obara","journal":"","ref_id":"b27","title":"Style augmentation: data augmentation via style randomization","year":"2019"},{"authors":"Max Jaderberg; Karen Simonyan; Andrew Zisserman; Koray Kavukcuoglu","journal":"","ref_id":"b28","title":"Spatial transformer networks","year":"2015-06"},{"authors":"P Diederik; Jimmy Kingma; Ba","journal":"","ref_id":"b29","title":"Adam: A Method for Stochastic Optimization","year":"2014-12"},{"authors":"P Diederik; Max Kingma; Welling","journal":"ICLR","ref_id":"b30","title":"Auto-encoding variational bayes","year":"2014-12"},{"authors":"Alex Krizhevsky; Vinod Nair; Geoffrey Hinton","journal":"","ref_id":"b31","title":"CIFAR-10","year":""},{"authors":"Alex Krizhevsky; Ilya Sutskever; Geoffrey E Hinton","journal":"","ref_id":"b32","title":"Imagenet classification with deep convolutional neural networks","year":""},{"authors":" Curran Associates; Inc","journal":"","ref_id":"b33","title":"","year":"2012"},{"authors":"Doyup Lee; Chiheon Kim; Saehoon Kim; Minsu Cho; Wook-Shin Han","journal":"","ref_id":"b34","title":"Autoregressive image generation using residual quantization","year":"2022"},{"authors":"Zhu Li; Adrian Perez-Suay; Gustau Camps-Valls; Dino Sejdinovic","journal":"","ref_id":"b35","title":"Kernel Dependence Regularizers and Gaussian Processes with Applications to Algorithmic Fairness","year":"2019"},{"authors":"Ilya Loshchilov; Frank Hutter","journal":"","ref_id":"b36","title":"Sgdr: Stochastic gradient descent with warm restarts","year":"2017"},{"authors":"Patrick Mcclure; Nikolaus Kriegeskorte","journal":"Frontiers in Computational Neuroscience","ref_id":"b37","title":"Representational distance learning for deep neural networks","year":"2016-12"},{"authors":"Norman Mu; Justin Gilmer","journal":"","ref_id":"b38","title":"MNIST-C: A Robustness Benchmark for Computer Vision","year":"2019"},{"authors":"G Samuel; Frank Müller; Hutter","journal":"","ref_id":"b39","title":"Trivialaugment: Tuning-free yet state-of-the-art data augmentation","year":"2021"},{"authors":"Arne Nix; Suhas Shrinivasan; Edgar Y Walker; Fabian Sinz","journal":"","ref_id":"b40","title":"Can Functional Transfer Methods Capture Simple Inductive Biases? In Gustau Camps-Valls","year":""},{"authors":" Pmlr","journal":"","ref_id":"b41","title":"","year":"2022"},{"authors":"Maxime Oquab; Timothée Darcet; Théo Moutakanni; Huy Vo; Marc Szafraniec; Vasil Khalidov; Pierre Fernandez; Daniel Haziza; Francisco Massa; Alaaeldin El-Nouby; Mahmoud Assran; Nicolas Ballas; Wojciech Galuba; Russell Howes; Po-Yao Huang; Shang-Wen Li; Ishan Misra; Michael Rabbat; Vasu Sharma; Gabriel Synnaeve; Hu Xu; Hervé Jegou; Julien Mairal; Patrick Labatut; Armand Joulin; Piotr Bojanowski","journal":"","ref_id":"b42","title":"Dinov2: Learning robust visual features without supervision","year":"2023"},{"authors":"Pingbo Pan; Siddharth Swaroop; Alexander Immer; Runa Eschenhagen; Richard E Turner; Mohammad Emtiyaz Khan","journal":"Decem","ref_id":"b43","title":"Continual deep learning by functional regularisation of memorable past","year":"2020"},{"authors":"Hieu Pham; Zihang Dai; Qizhe Xie; Quoc V Le","journal":"","ref_id":"b44","title":"Meta Pseudo Labels","year":"2021-03"},{"authors":"Aditya Ramesh; Prafulla Dhariwal; Alex Nichol; Casey Chu; Mark Chen","journal":"","ref_id":"b45","title":"Hierarchical Text-Conditional Image Generation with CLIP Latents","year":"2022-04"},{"authors":"Ahmad Rashid; Vasileios Lioutas; Mehdi Rezagholizadeh","journal":"","ref_id":"b46","title":"MATE-KD: Masked adversarial text, a companion to knowledge distillation","year":"2021"},{"authors":"Benjamin Recht; Rebecca Roelofs; Ludwig Schmidt; Vaishaal Shankar","journal":"PMLR","ref_id":"b47","title":"Do imagenet classifiers generalize to imagenet?","year":"2019"},{"authors":"Robin Rombach; Andreas Blattmann; Dominik Lorenz; Patrick Esser; Björn Ommer","journal":"","ref_id":"b48","title":"Highresolution image synthesis with latent diffusion models","year":"2022"},{"authors":"Evgenia Rusak; Lukas Schott; Roland S Zimmermann; Julian Bitterwolf; Oliver Bringmann; Matthias Bethge; Wieland Brendel","journal":"","ref_id":"b49","title":"A simple way to make neural networks robust against diverse image corruptions","year":"2020-01"},{"authors":"Chitwan Saharia; William Chan; Saurabh Saxena; Lala Li; Jay Whang; Emily Denton; Seyed Kamyar; Seyed Ghasemipour; Burcu Karagol Ayan; S Sara Mahdavi; Rapha Gontijo Lopes; Tim Salimans; Jonathan Ho; David J Fleet; Mohammad Norouzi","journal":"","ref_id":"b50","title":"Photorealistic Textto-Image Diffusion Models with Deep Language Understanding","year":"2022-05"},{"authors":"Jascha Sohl-Dickstein; Eric Weiss; Niru Maheswaranathan; Surya Ganguli","journal":"PMLR","ref_id":"b51","title":"Deep unsupervised learning using nonequilibrium thermodynamics","year":"2015"},{"authors":"Christian Szegedy; Vincent Vanhoucke; Sergey Ioffe; Jon Shlens; Zbigniew Wojna","journal":"","ref_id":"b52","title":"Rethinking the inception architecture for computer vision","year":"2016"},{"authors":"Jonathan Michalis K Titsias; Alexander G Schwarz; G De; Razvan Matthews; Yee Whye Pascanu; Teh","journal":"","ref_id":"b53","title":"Functional Regularisation for Continual Learning with Gaussian Processes","year":"2019"},{"authors":"Ilya Tolstikhin; Neil Houlsby; Alexander Kolesnikov; Lucas Beyer; Xiaohua Zhai; Thomas Unterthiner; Jessica Yung; Andreas Steiner; Daniel Keysers; Jakob Uszkoreit; Mario Lucic; Alexey Dosovitskiy","journal":"Advances in Neural Information Processing Systems","ref_id":"b54","title":"MLP-Mixer: An all-MLP Architecture for Vision","year":"2021"},{"authors":"Hugo Touvron; Matthieu Cord; Matthijs Douze; Francisco Massa; Alexandre Sablayrolles; Hervé Jégou","journal":"","ref_id":"b55","title":"Training data-efficient image transformers & distillation through attention","year":"2020-07"},{"authors":"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin","journal":"Decem","ref_id":"b56","title":"Attention is all you need","year":"2017"},{"authors":"Riccardo Volpi; John Duchi; Hongseok Namkoong; Vittorio Murino; Ozan Sener; Silvio Savarese","journal":"","ref_id":"b57","title":"Generalizing to Unseen Domains via Adversarial Data Augmentation","year":"2018"},{"authors":"Haohan Wang; Songwei Ge; Zachary Lipton; Eric P Xing","journal":"","ref_id":"b58","title":"Learning robust global representations by penalizing local predictive power","year":"2019"},{"authors":"Huan Wang; Suhas Lohit; Michael Jones; Yun Fu","journal":"","ref_id":"b59","title":"Knowledge Distillation Thrives on Data Augmentation","year":"2020"},{"authors":"Sangdoo Yun; Dongyoon Han; Seong Joon Oh; Sanghyuk Chun; Junsuk Choe; Youngjoon Yoo","journal":"","ref_id":"b60","title":"Cutmix: Regularization strategy to train strong classifiers with localizable features","year":"2019"},{"authors":"Sergey Zagoruyko; Nikos Komodakis","journal":"","ref_id":"b61","title":"Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer","year":"2017"},{"authors":"Hongyi Zhang; Moustapha Cisse; Yann N Dauphin; David Lopez-Paz","journal":"","ref_id":"b62","title":"mixup: Beyond Empirical Risk Minimization","year":"2017-10"},{"authors":"Minjia Zhang; Niranjan Uma Naresh; Yuxiong He","journal":"","ref_id":"b63","title":"Adversarial Data Augmentation for Task-Specific Knowledge Distillation of Pre-trained Transformers","year":"2022"},{"authors":"Xiaofeng Zhang; Zhangyang Wang; Dong Liu; Qing Ling","journal":"","ref_id":"b64","title":"DADA: Deep Adversarial Data Augmentation for Extremely Low Data Regime Classification","year":"2019-05"},{"authors":"Xinyu Zhang; Qiang Wang; Jian Zhang; Zhao Zhong","journal":"","ref_id":"b65","title":"Adversarial AutoAugment","year":"2019"},{"authors":"Long Zhao; Ting Liu; Xi Peng; Dimitris Metaxas","journal":"","ref_id":"b66","title":"Maximum-entropy adversarial data augmentation for improved generalization and robustness","year":"2020"},{"authors":"Zhun Zhong; Liang Zheng; Guoliang Kang; Shaozi Li; Yi Yang","journal":"","ref_id":"b67","title":"Random erasing data augmentation","year":"2017"},{"authors":"Dominik Zietlow; Michael Lohaus; Guha Balakrishnan; Matthäus Kleindessner; Francesco Locatello; Bernhard Schölkopf; Chris Russell","journal":"","ref_id":"b68","title":"Leveling Down in Computer Vision: Pareto Inefficiencies in Fair Deep Classifiers","year":"2022"}],"string":"[\n {\n \"authors\": \"Abien Fred; Agarap \",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Deep learning using rectified linear units (relu)\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Anthreas Antoniou; Amos Storkey; Harrison Edwards\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Data Augmentation Generative Adversarial Networks\",\n \"year\": \"2022-02\"\n },\n {\n \"authors\": \"Sima Behpour; Kris M Kitani; Brian D Ziebart\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"ADA: Adversarial data augmentation for object detection\",\n \"year\": \"2019-03\"\n },\n {\n \"authors\": \"David Ari S Benjamin; Konrad P Rolnick; Kording\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Measuring and regularizing networks in function space\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Lucas Beyer; Olivier J Hénaff; Alexander Kolesnikov; Xiaohua Zhai; Aäron Van Den Oord\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Are we done with imagenet?\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Lucas Beyer; Xiaohua Zhai; Amélie Royer; Larisa Markeeva; Rohan Anil; Alexander Kolesnikov\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Knowledge distillation: A good teacher is patient and consistent\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Cristian Bucilǎ; Rich Caruana; Alexandra Niculescu-Mizil\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Model compression\",\n \"year\": \"2006\"\n },\n {\n \"authors\": \"Xiangning Chen; Chen Liang; Da Huang; Esteban Real; Kaiyuan Wang; Yao Liu; Hieu Pham; Xuanyi Dong; Thang Luong; Cho-Jui Hsieh; Yifeng Lu; Quoc V Le\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Symbolic discovery of optimization algorithms\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Xianing Chen; Qiong Cao; Yujie Zhong; Jing Zhang; Shenghua Gao; Dacheng Tao\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"DearKD: Data-Efficient Early Knowledge Distillation for Vision Transformers\",\n \"year\": \"2022-04\"\n },\n {\n \"authors\": \"Rewon Child\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them on Images\",\n \"year\": \"2020-11\"\n },\n {\n \"authors\": \"D Ekin; Barret Cubuk; Dandelion Zoph; Vijay Mane; Quoc V Vasudevan; Le\",\n \"journal\": \"Cvpr\",\n \"ref_id\": \"b10\",\n \"title\": \"AutoAugment: Learning Augmentation Policies from Data\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Barret Ekin D Cubuk; Jonathon Zoph; Quoc V Shlens; Le\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Randaugment: Practical automated data augmentation with a reduced search space\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Mostafa Dehghani; Josip Djolonga; Basil Mustafa; Piotr Padlewski; Jonathan Heek; Justin Gilmer; Andreas Steiner; Mathilde Caron; Robert Geirhos; Ibrahim Alabdulmohsin; Rodolphe Jenatton; Lucas Beyer; Michael Tschannen; Anurag Arnab; Xiao Wang; Carlos Riquelme; Matthias Minderer; Joan Puigcerver; Utku Evci; Manoj Kumar; Sjoerd Van Steenkiste; Gamaleldin F Elsayed; Aravindh Mahendran; Fisher Yu; Avital Oliver; Fantine Huot; Jasmijn Bastings; Mark Patrick Collier; Alexey Gritsenko; Vighnesh Birodkar; Cristina Vasconcelos; Yi Tay; Thomas Mensink; Alexander Kolesnikov; Filip Pavetić; Dustin Tran; Thomas Kipf; Mario Lučić; Xiaohua Zhai; Daniel Keysers; Jeremiah Harmsen; Neil Houlsby\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Scaling vision transformers to 22 billion parameters\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Jia Deng; Wei Dong; Richard Socher; Li-Jia Li; Kai Li; Li Fei-Fei\",\n \"journal\": \"Ieee\",\n \"ref_id\": \"b13\",\n \"title\": \"Imagenet: A largescale hierarchical image database\",\n \"year\": \"2009\"\n },\n {\n \"authors\": \"Li Deng\",\n \"journal\": \"IEEE Signal Processing Magazine\",\n \"ref_id\": \"b14\",\n \"title\": \"The mnist database of handwritten digit images for machine learning research\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Tim Dettmers; Mike Lewis; Sam Shleifer; Luke Zettlemoyer\",\n \"journal\": \"ICLR\",\n \"ref_id\": \"b15\",\n \"title\": \"8-bit optimizers via block-wise quantization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Alexey Dosovitskiy; Lucas Beyer; Alexander Kolesnikov; Dirk Weissenborn; Xiaohua Zhai; Thomas Unterthiner; Mostafa Dehghani; Matthias Minderer; Georg Heigold; Sylvain Gelly; Jakob Uszkoreit; Neil Houlsby\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Logan Engstrom; Brandon Tran; Dimitris Tsipras; Ludwig Schmidt; Aleksander Madry\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"A rotation and a translation suffice: Fooling cnns with simple transformations\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Robert Geirhos; Claudio Michaelis; Felix A Wichmann; Patricia Rubisch; Matthias Bethge; Wieland Brendel; . Imagenet-Trained Cnns Are Biased Towards Texture\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"INCREASING SHAPE BIAS IMPROVES ACCURACY AND ROBUSTNESS\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Kehong Gong; Jianfeng Zhang; Jiashi Feng\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"PoseAug: A Differentiable Pose Augmentation Framework for 3D Human Pose Estimation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Md Akmal Haidar; Mehdi Rezagholizadeh; Abbas Ghaddar; Khalil Bibi; Philippe Langlais; Pascal Poupart\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"CILDA: Contrastive Data Augmentation using Intermediate Layer Knowledge Distillation\",\n \"year\": \"2022-04\"\n },\n {\n \"authors\": \"Kaiming He; Xiangyu Zhang; Shaoqing Ren; Jian Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Deep Residual Learning for Image Recognition\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Dan Hendrycks; Thomas Dietterich\",\n \"journal\": \"mar\",\n \"ref_id\": \"b22\",\n \"title\": \"Benchmarking neural network robustness to common corruptions and perturbations\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Dan Hendrycks; Norman Mu; D Ekin; Barret Cubuk; Justin Zoph; Balaji Gilmer; Lakshminarayanan\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Augmix: A simple data processing method to improve robustness and uncertainty\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Dan Hendrycks; Steven Basart; Norman Mu; Saurav Kadavath; Frank Wang; Evan Dorundo; Rahul Desai; Tyler Zhu; Samyak Parajuli; Mike Guo; Dawn Song; Jacob Steinhardt; Justin Gilmer\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Dan Hendrycks; Kevin Zhao; Steven Basart; Jacob Steinhardt; Dawn Song\",\n \"journal\": \"CVPR\",\n \"ref_id\": \"b25\",\n \"title\": \"Natural adversarial examples\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Geoffrey Hinton; Jeff Dean\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Distilling the Knowledge in a Neural Network\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Amir Philip Tg Jackson; Stephen Atapour Abarghouei; Toby P Bonner; Boguslaw Breckon; Obara\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Style augmentation: data augmentation via style randomization\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Max Jaderberg; Karen Simonyan; Andrew Zisserman; Koray Kavukcuoglu\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Spatial transformer networks\",\n \"year\": \"2015-06\"\n },\n {\n \"authors\": \"P Diederik; Jimmy Kingma; Ba\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Adam: A Method for Stochastic Optimization\",\n \"year\": \"2014-12\"\n },\n {\n \"authors\": \"P Diederik; Max Kingma; Welling\",\n \"journal\": \"ICLR\",\n \"ref_id\": \"b30\",\n \"title\": \"Auto-encoding variational bayes\",\n \"year\": \"2014-12\"\n },\n {\n \"authors\": \"Alex Krizhevsky; Vinod Nair; Geoffrey Hinton\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"CIFAR-10\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Alex Krizhevsky; Ilya Sutskever; Geoffrey E Hinton\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Imagenet classification with deep convolutional neural networks\",\n \"year\": \"\"\n },\n {\n \"authors\": \" Curran Associates; Inc\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Doyup Lee; Chiheon Kim; Saehoon Kim; Minsu Cho; Wook-Shin Han\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Autoregressive image generation using residual quantization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Zhu Li; Adrian Perez-Suay; Gustau Camps-Valls; Dino Sejdinovic\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Kernel Dependence Regularizers and Gaussian Processes with Applications to Algorithmic Fairness\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Ilya Loshchilov; Frank Hutter\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Sgdr: Stochastic gradient descent with warm restarts\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Patrick Mcclure; Nikolaus Kriegeskorte\",\n \"journal\": \"Frontiers in Computational Neuroscience\",\n \"ref_id\": \"b37\",\n \"title\": \"Representational distance learning for deep neural networks\",\n \"year\": \"2016-12\"\n },\n {\n \"authors\": \"Norman Mu; Justin Gilmer\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"MNIST-C: A Robustness Benchmark for Computer Vision\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"G Samuel; Frank Müller; Hutter\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Trivialaugment: Tuning-free yet state-of-the-art data augmentation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Arne Nix; Suhas Shrinivasan; Edgar Y Walker; Fabian Sinz\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Can Functional Transfer Methods Capture Simple Inductive Biases? In Gustau Camps-Valls\",\n \"year\": \"\"\n },\n {\n \"authors\": \" Pmlr\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Maxime Oquab; Timothée Darcet; Théo Moutakanni; Huy Vo; Marc Szafraniec; Vasil Khalidov; Pierre Fernandez; Daniel Haziza; Francisco Massa; Alaaeldin El-Nouby; Mahmoud Assran; Nicolas Ballas; Wojciech Galuba; Russell Howes; Po-Yao Huang; Shang-Wen Li; Ishan Misra; Michael Rabbat; Vasu Sharma; Gabriel Synnaeve; Hu Xu; Hervé Jegou; Julien Mairal; Patrick Labatut; Armand Joulin; Piotr Bojanowski\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Dinov2: Learning robust visual features without supervision\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Pingbo Pan; Siddharth Swaroop; Alexander Immer; Runa Eschenhagen; Richard E Turner; Mohammad Emtiyaz Khan\",\n \"journal\": \"Decem\",\n \"ref_id\": \"b43\",\n \"title\": \"Continual deep learning by functional regularisation of memorable past\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Hieu Pham; Zihang Dai; Qizhe Xie; Quoc V Le\",\n \"journal\": \"\",\n \"ref_id\": \"b44\",\n \"title\": \"Meta Pseudo Labels\",\n \"year\": \"2021-03\"\n },\n {\n \"authors\": \"Aditya Ramesh; Prafulla Dhariwal; Alex Nichol; Casey Chu; Mark Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b45\",\n \"title\": \"Hierarchical Text-Conditional Image Generation with CLIP Latents\",\n \"year\": \"2022-04\"\n },\n {\n \"authors\": \"Ahmad Rashid; Vasileios Lioutas; Mehdi Rezagholizadeh\",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"MATE-KD: Masked adversarial text, a companion to knowledge distillation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Benjamin Recht; Rebecca Roelofs; Ludwig Schmidt; Vaishaal Shankar\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b47\",\n \"title\": \"Do imagenet classifiers generalize to imagenet?\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Robin Rombach; Andreas Blattmann; Dominik Lorenz; Patrick Esser; Björn Ommer\",\n \"journal\": \"\",\n \"ref_id\": \"b48\",\n \"title\": \"Highresolution image synthesis with latent diffusion models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Evgenia Rusak; Lukas Schott; Roland S Zimmermann; Julian Bitterwolf; Oliver Bringmann; Matthias Bethge; Wieland Brendel\",\n \"journal\": \"\",\n \"ref_id\": \"b49\",\n \"title\": \"A simple way to make neural networks robust against diverse image corruptions\",\n \"year\": \"2020-01\"\n },\n {\n \"authors\": \"Chitwan Saharia; William Chan; Saurabh Saxena; Lala Li; Jay Whang; Emily Denton; Seyed Kamyar; Seyed Ghasemipour; Burcu Karagol Ayan; S Sara Mahdavi; Rapha Gontijo Lopes; Tim Salimans; Jonathan Ho; David J Fleet; Mohammad Norouzi\",\n \"journal\": \"\",\n \"ref_id\": \"b50\",\n \"title\": \"Photorealistic Textto-Image Diffusion Models with Deep Language Understanding\",\n \"year\": \"2022-05\"\n },\n {\n \"authors\": \"Jascha Sohl-Dickstein; Eric Weiss; Niru Maheswaranathan; Surya Ganguli\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b51\",\n \"title\": \"Deep unsupervised learning using nonequilibrium thermodynamics\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Christian Szegedy; Vincent Vanhoucke; Sergey Ioffe; Jon Shlens; Zbigniew Wojna\",\n \"journal\": \"\",\n \"ref_id\": \"b52\",\n \"title\": \"Rethinking the inception architecture for computer vision\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Jonathan Michalis K Titsias; Alexander G Schwarz; G De; Razvan Matthews; Yee Whye Pascanu; Teh\",\n \"journal\": \"\",\n \"ref_id\": \"b53\",\n \"title\": \"Functional Regularisation for Continual Learning with Gaussian Processes\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Ilya Tolstikhin; Neil Houlsby; Alexander Kolesnikov; Lucas Beyer; Xiaohua Zhai; Thomas Unterthiner; Jessica Yung; Andreas Steiner; Daniel Keysers; Jakob Uszkoreit; Mario Lucic; Alexey Dosovitskiy\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b54\",\n \"title\": \"MLP-Mixer: An all-MLP Architecture for Vision\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Hugo Touvron; Matthieu Cord; Matthijs Douze; Francisco Massa; Alexandre Sablayrolles; Hervé Jégou\",\n \"journal\": \"\",\n \"ref_id\": \"b55\",\n \"title\": \"Training data-efficient image transformers & distillation through attention\",\n \"year\": \"2020-07\"\n },\n {\n \"authors\": \"Ashish Vaswani; Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones; Aidan N Gomez; Lukasz Kaiser; Illia Polosukhin\",\n \"journal\": \"Decem\",\n \"ref_id\": \"b56\",\n \"title\": \"Attention is all you need\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Riccardo Volpi; John Duchi; Hongseok Namkoong; Vittorio Murino; Ozan Sener; Silvio Savarese\",\n \"journal\": \"\",\n \"ref_id\": \"b57\",\n \"title\": \"Generalizing to Unseen Domains via Adversarial Data Augmentation\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Haohan Wang; Songwei Ge; Zachary Lipton; Eric P Xing\",\n \"journal\": \"\",\n \"ref_id\": \"b58\",\n \"title\": \"Learning robust global representations by penalizing local predictive power\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Huan Wang; Suhas Lohit; Michael Jones; Yun Fu\",\n \"journal\": \"\",\n \"ref_id\": \"b59\",\n \"title\": \"Knowledge Distillation Thrives on Data Augmentation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Sangdoo Yun; Dongyoon Han; Seong Joon Oh; Sanghyuk Chun; Junsuk Choe; Youngjoon Yoo\",\n \"journal\": \"\",\n \"ref_id\": \"b60\",\n \"title\": \"Cutmix: Regularization strategy to train strong classifiers with localizable features\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Sergey Zagoruyko; Nikos Komodakis\",\n \"journal\": \"\",\n \"ref_id\": \"b61\",\n \"title\": \"Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Hongyi Zhang; Moustapha Cisse; Yann N Dauphin; David Lopez-Paz\",\n \"journal\": \"\",\n \"ref_id\": \"b62\",\n \"title\": \"mixup: Beyond Empirical Risk Minimization\",\n \"year\": \"2017-10\"\n },\n {\n \"authors\": \"Minjia Zhang; Niranjan Uma Naresh; Yuxiong He\",\n \"journal\": \"\",\n \"ref_id\": \"b63\",\n \"title\": \"Adversarial Data Augmentation for Task-Specific Knowledge Distillation of Pre-trained Transformers\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xiaofeng Zhang; Zhangyang Wang; Dong Liu; Qing Ling\",\n \"journal\": \"\",\n \"ref_id\": \"b64\",\n \"title\": \"DADA: Deep Adversarial Data Augmentation for Extremely Low Data Regime Classification\",\n \"year\": \"2019-05\"\n },\n {\n \"authors\": \"Xinyu Zhang; Qiang Wang; Jian Zhang; Zhao Zhong\",\n \"journal\": \"\",\n \"ref_id\": \"b65\",\n \"title\": \"Adversarial AutoAugment\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Long Zhao; Ting Liu; Xi Peng; Dimitris Metaxas\",\n \"journal\": \"\",\n \"ref_id\": \"b66\",\n \"title\": \"Maximum-entropy adversarial data augmentation for improved generalization and robustness\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Zhun Zhong; Liang Zheng; Guoliang Kang; Shaozi Li; Yi Yang\",\n \"journal\": \"\",\n \"ref_id\": \"b67\",\n \"title\": \"Random erasing data augmentation\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Dominik Zietlow; Michael Lohaus; Guha Balakrishnan; Matthäus Kleindessner; Francesco Locatello; Bernhard Schölkopf; Chris Russell\",\n \"journal\": \"\",\n \"ref_id\": \"b68\",\n \"title\": \"Leveling Down in Computer Vision: Pareto Inefficiencies in Fair Deep Classifiers\",\n \"year\": \"2022\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[4,253.36,262.9,251.3,10.7],"formula_id":"formula_0","formula_text":"L s↔ t = D [f s (x), f t (x)] .(1)"},{"formula_coordinates":[4,253.41,314.13,251.26,9.65],"formula_id":"formula_1","formula_text":"L t↔t = D [f t (x), f t (x)] ,(2)"},{"formula_coordinates":[4,204.57,381.69,300.1,14.66],"formula_id":"formula_2","formula_text":"max θa λ s L s↔ t -λ t L t↔t and min θs L s↔ t .(3)"},{"formula_coordinates":[8,195.4,263.43,272.88,21.73],"formula_id":"formula_3","formula_text":"ResNet50 → ResNet18 ResNet50 → ViT-S Im-A Im-R Im-C ↓ Sketch Style Im-A Im-R Im-C ↓ Sketch"}],"string":"[\n {\n \"formula_coordinates\": [\n 4,\n 253.36,\n 262.9,\n 251.3,\n 10.7\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"L s↔ t = D [f s (x), f t (x)] .(1)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 253.41,\n 314.13,\n 251.26,\n 9.65\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"L t↔t = D [f t (x), f t (x)] ,(2)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 204.57,\n 381.69,\n 300.1,\n 14.66\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"max θa λ s L s↔ t -λ t L t↔t and min θs L s↔ t .(3)\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 195.4,\n 263.43,\n 272.88,\n 21.73\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"ResNet50 → ResNet18 ResNet50 → ViT-S Im-A Im-R Im-C ↓ Sketch Style Im-A Im-R Im-C ↓ Sketch\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"HARD: Hard Augmentations for Robust Distillation"},"abstract":{"kind":"string","value":"Knowledge distillation (KD) is a simple and successful method to transfer knowledge from a teacher to a student model solely based on functional activity. However, current KD has a few shortcomings: it has recently been shown that this method is unsuitable to transfer simple inductive biases like shift equivariance, struggles to transfer out of domain generalization, and optimization time is magnitudes longer compared to default non-KD model training. To improve these aspects of KD, we propose Hard Augmentations for Robust Distillation (HARD), a generally applicable data augmentation framework, that generates synthetic data points for which the teacher and the student disagree. We show in a simple toy example that our augmentation framework solves the problem of transferring simple equivariances with KD. We then apply our framework in real-world tasks for a variety of augmentation models, ranging from simple spatial transformations to unconstrained image manipulations with a pretrained variational autoencoder. We find that our learned augmentations significantly improve KD performance on in-domain and out-of-domain evaluation. Moreover, our method outperforms even state-of-the-art data augmentations and since the augmented training inputs can be visualized, they offer a qualitative insight into the properties that are transferred from the teacher to the student. Thus HARD represents a generally applicable, dynamically optimized data augmentation technique tailored to improve the generalization and convergence speed of models trained with KD."},"authors":{"kind":"string","value":"Arne F Nix; Max F Burg; Fabian H Sinz"},"figures":{"kind":"list like","value":[{"figure_caption":"HFigure 2 :2Figure2: Fitting the student, a three-layer ReLU MLP, to the teacher function, cos(x), for 10, 000 iterations. We show results for 10 random seeds (A-D) and the distribution of (augmented) training inputs as a normalized histogram (E-H). We compare baseline (no augmentations) with Mixup, Gaussian noise and an HARD-optimized noise distribution. We report mean-squared-error (MSE) on 100 test inputs sampled from U[-10,10] .","figure_data":"","figure_id":"fig_0","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Illustration of the augmentor models used in our experiments. (A) HARD-Mix: Imagedependent patch-wise interpolation of multiple images. (B) HARD-Affine: Learned distribution of affine transformations in the pixel coordinates. (C) HARD-VAE: Finetuning (parts of) a pretrained VAE.","figure_data":"","figure_id":"fig_1","figure_label":"3","figure_type":"figure"},{"figure_caption":"2 Orbit2The left two columns show the transfer results from a small CNN teacher to a MLP student. The right columns show analogous experiments between a ResNet18 teacher and a small ViT student. The best performing transfer is shown in bold for each column. Examples of our HARD-Affine learned data augmentations shown on the right. We include the controls Random Affine and MNIST-C Shifts (marked by italics). ± 0.0 91.3 ± 0.5 99.5 ± 0.0 92.8 ± 0.5 Student only 98.4 ± 0.0 35.2 ± 0.7 98.3 ± 0.0 40.4 ± 0.8 + Random Affine 92.1 ± 0.6 81.0 ± 2.0 95.4 ± 0.3 90.4 ± 1.0 + MNIST-C Shifts 98.1 ± 0.1 86.5 ± 0.3 98.5 ± 0.0 93.7 ± 0.± 0.0 40.3 ± 0.6 98.6 ± 0.1 44.7 ± 1.9 + HARD -Affine 98.6 ± 0.1 68.9 ± 2.5 99.2 ± 0.0 84.1 ± 2.3","figure_data":"","figure_id":"fig_2","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Example augmentations applied to images of the ImageNet validation set obtained from augmentor models in the ViT-S setting at the end of training.","figure_data":"","figure_id":"fig_3","figure_label":"4","figure_type":"figure"},{"figure_caption":"","figure_data":"","figure_id":"","figure_label":"","figure_type":"figure"},{"figure_caption":"Test accuracies on the CIFAR10 test set. Standard error of the mean is reported where available across three different seeds. Best transfer is highlighted in bold. The ResNet101 * models were pretrained on ImageNet. Examples of augmented test images from ResNet18→ViT experiments with samples across different iterations are shown to the right.","figure_data":"ResNet18ResNet101 *ResNet101 *Original-Affine-VAE-VAE-Aff.↓↓↓ViTViTResNet18Teacher only92.5 ± 0.095.595.5Student only68.5 ± 0.568.5 ± 0.578.5+ Standard Aug.78.3 ± 0.478.3 ± 0.492.6+ Random Affine Aug.58.9 ± 0.458.9 ± 0.479.3KD67.9 ± 0.168.584.4+ Standard Aug.80.9 ± 0.179.393.3+ HARD -Affine87.8 ± 0.884.493.5+ HARD -VAE81.9 ± 0.481.291.0+ HARD -VAE-Affine 87.6 ± 0.687.194.0","figure_id":"tab_0","figure_label":"2","figure_type":"table"},{"figure_caption":"","figure_data":"Val ReaLV2Val ReaL V2Teacher75.8 83.163.775.8 83.1 63.7Student70.7 78.157.473.2 79.4 60.3KD70.7 78.758.175.3 82.8 62.9+ HARD -Affine 71.6 79.558.674.9 82.3 62.4+ HARD -Mix71.4 79.458.675.7 83.0 63.3+ HARD -VAE71.0 78.958.775.8 83.1 63.5","figure_id":"tab_1","figure_label":"3","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"HFigure 2 :2Figure2: Fitting the student, a three-layer ReLU MLP, to the teacher function, cos(x), for 10, 000 iterations. We show results for 10 random seeds (A-D) and the distribution of (augmented) training inputs as a normalized histogram (E-H). We compare baseline (no augmentations) with Mixup, Gaussian noise and an HARD-optimized noise distribution. We report mean-squared-error (MSE) on 100 test inputs sampled from U[-10,10] .\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Illustration of the augmentor models used in our experiments. (A) HARD-Mix: Imagedependent patch-wise interpolation of multiple images. (B) HARD-Affine: Learned distribution of affine transformations in the pixel coordinates. (C) HARD-VAE: Finetuning (parts of) a pretrained VAE.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"2 Orbit2The left two columns show the transfer results from a small CNN teacher to a MLP student. The right columns show analogous experiments between a ResNet18 teacher and a small ViT student. The best performing transfer is shown in bold for each column. Examples of our HARD-Affine learned data augmentations shown on the right. We include the controls Random Affine and MNIST-C Shifts (marked by italics). ± 0.0 91.3 ± 0.5 99.5 ± 0.0 92.8 ± 0.5 Student only 98.4 ± 0.0 35.2 ± 0.7 98.3 ± 0.0 40.4 ± 0.8 + Random Affine 92.1 ± 0.6 81.0 ± 2.0 95.4 ± 0.3 90.4 ± 1.0 + MNIST-C Shifts 98.1 ± 0.1 86.5 ± 0.3 98.5 ± 0.0 93.7 ± 0.± 0.0 40.3 ± 0.6 98.6 ± 0.1 44.7 ± 1.9 + HARD -Affine 98.6 ± 0.1 68.9 ± 2.5 99.2 ± 0.0 84.1 ± 2.3\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Example augmentations applied to images of the ImageNet validation set obtained from augmentor models in the ViT-S setting at the end of training.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"\",\n \"figure_id\": \"\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Test accuracies on the CIFAR10 test set. Standard error of the mean is reported where available across three different seeds. Best transfer is highlighted in bold. The ResNet101 * models were pretrained on ImageNet. Examples of augmented test images from ResNet18→ViT experiments with samples across different iterations are shown to the right.\",\n \"figure_data\": \"ResNet18ResNet101 *ResNet101 *Original-Affine-VAE-VAE-Aff.↓↓↓ViTViTResNet18Teacher only92.5 ± 0.095.595.5Student only68.5 ± 0.568.5 ± 0.578.5+ Standard Aug.78.3 ± 0.478.3 ± 0.492.6+ Random Affine Aug.58.9 ± 0.458.9 ± 0.479.3KD67.9 ± 0.168.584.4+ Standard Aug.80.9 ± 0.179.393.3+ HARD -Affine87.8 ± 0.884.493.5+ HARD -VAE81.9 ± 0.481.291.0+ HARD -VAE-Affine 87.6 ± 0.687.194.0\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"Val ReaLV2Val ReaL V2Teacher75.8 83.163.775.8 83.1 63.7Student70.7 78.157.473.2 79.4 60.3KD70.7 78.758.175.3 82.8 62.9+ HARD -Affine 71.6 79.558.674.9 82.3 62.4+ HARD -Mix71.4 79.458.675.7 83.0 63.3+ HARD -VAE71.0 78.958.775.8 83.1 63.5\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Supporting Evidence\", \"Citation\": \"[27,37,60]\", \"Explanation\": \"The cited works provide a foundation for the use of knowledge distillation methods in transfer learning scenarios, including model compression, continual learning, and neuroscience research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[17,53,55]\", \"Explanation\": \"The cited works build upon the use of knowledge distillation methods to train new models that trade off inductive biases for more flexibility and more parameters on smaller data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[9,40,54]\", \"Explanation\": \"The cited works are the data sources for the study conducted in the citing paper, which shows the failure of current knowledge distillation methods to transfer even simple equivariances between teacher and student.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[6,41]\", \"Explanation\": \"The cited works extend the research on the gap between student and teacher performance in out-of-domain evaluation compared to within-domain performance, even in cases where the student almost perfectly matches the teacher.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[41]\", \"Explanation\": \"The cited work by [41] is mentioned in the context of discussing the phenomenon of particularly robust teachers, which the citing paper builds upon to support the claim that transferring robustness properties is a difficult problem for knowledge distillation in general.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[40]\", \"Explanation\": \"The cited work establishes the concept of shift equivariance in the context of knowledge distillation, which the citing paper builds upon in their framework to improve transfer effectiveness and solve the problem of not being able to transfer shift equivariance.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[18,33]\", \"Explanation\": \"The cited works provide a long tradition of using data augmentations in computer vision, which the citing paper builds upon in its research on data augmentations in deep learning models.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[12,24,28,39,59,61]\", \"Explanation\": \"The cited works present more complex data augmentations that the citing paper expands upon in its study of improving generalization and out-of-domain performance in data augmentations.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[61]\", \"Explanation\": \"The cited work on Mixup is a popular augmentation method that the citing paper uses to support its research on data augmentations in deep learning models.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[59]\", \"Explanation\": \"The cited work on Cutmix is another data augmentation method that the citing paper uses to support its study on improving generalization and out-of-domain performance in data augmentations.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[11,25,48,67]\", \"Explanation\": \"The cited works on parameterized augmentations optimized to improve a given objective are methodological bases for the citing paper in its research on data augmentations in deep learning models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[2,3,20,56,[63][64][65], however, without applying them for knowledge transfer\", \"Explanation\": \"The cited works on optimizing data augmentations to improve on an adversarial objective are methodological bases for the citing paper in its study of data augmentations in deep learning models without applying them for knowledge transfer.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[6,58]\", \"Explanation\": \"The cited works on applying data augmentations in knowledge transfer are methodological bases for the citing paper in its research on data augmentations in deep learning models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[43]\", \"Explanation\": \"The cited work on optimizing data augmentations to aide the teaching in knowledge transfer is a methodological basis for the citing paper in its study of data augmentations in deep learning models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[21], [45], [62]\", \"Explanation\": \"The cited works on using adversarial objectives to optimize data augmentations for knowledge distillation in natural language processing tasks provide a methodological basis for the citing paper to apply similar techniques in a task-agnostic framework for data augmentation in various data domains.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[27]\", \"Explanation\": \"The cited work provides a classical knowledge distillation (KD) method that the citing paper builds upon to develop a new approach for knowledge transfer from a teacher to a student model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[29]\", \"Explanation\": \"The cited work on spatial transformers provides the basis for the augmentor model used in the citing paper, which is implemented to perform simple transformations of the pixel grid in images.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[61]\", \"Explanation\": \"The cited work on Mixup is used as a basis for the adaptive variant of the augmentor model in the citing paper, which combines images in a more complex manner.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[59]\", \"Explanation\": \"The cited work on Cutmix is also used as a basis for the adaptive variant of the augmentor model in the citing paper, which combines images in a more complex manner.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"Figure 3B\", \"Explanation\": \"The cited figure in the text provides a visual representation of the patch-wise projection and comparison mechanism used in the augmentor model, which extends the concept of image combination in a more systematic way.\"}, {\"Category\": \"Data Source\", \"Citation\": \"Figure 3B\", \"Explanation\": \"The cited figure in the text is used to highlight the use of a query vector sampled from a learned distribution in the patch-wise projection and comparison mechanism of the augmentor model, which serves as a data source for the study conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[10]\", \"Explanation\": \"The cited work provides a very deep VAE model that the citing paper chooses to finetune for the image reconstruction process in the student training.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[34]\", \"Explanation\": \"The cited work provides a Residual-Quantized VAE (RQ-VAE) model that the citing paper uses to finetune and add a noise vector for the latent state in the generation process of the augmentations.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[40]\", \"Explanation\": \"The cited work by Nix et al. provides the initial setup and methodology for the experiment conducted in the citing paper, which tests the transfer of inductive bias from a shift equivariant teacher to a student without the same inductive bias.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[15]\", \"Explanation\": \"The cited work by LeCun et al. introduces the MNIST dataset, which is used in the citing paper to evaluate the performance of the students and teachers in a training and test scenario.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[38]\", \"Explanation\": \"The cited work by Tulyakov et al. provides a version of the MNIST test set in which digits were randomly shifted, which is used in the citing paper to evaluate the generalization performance of the students and teachers.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[40]\", \"Explanation\": \"The cited work by Nix et al. provides a method for training data augmentation that the citing paper adopts to improve the performance of the student model in shifted input scenarios.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[32]\", \"Explanation\": \"The cited work, CIFAR10, serves as the data set used in the experiments conducted in the citing paper to test the framework in more realistic scenarios.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work, ImageNet, is used as a benchmark dataset for classification tasks in the citing paper. The results obtained from the experiment on ImageNet provide a basis for the performance analysis of the methods discussed in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[22]\", \"Explanation\": \"The cited work, ResNet50, is used as the teacher model in the distillation process in the citing paper. The method of training the teacher model with Deep-augment and AugMix data augmentations is adopted in the citing paper to improve the robustness of the student models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[17]\", \"Explanation\": \"The cited work, ViT-S, is used as a student model in the distillation process in the citing paper. The method of using a small vision transformer variant as a student model is adopted to improve the performance of the student models in the context of natural image corruptions.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work provides the standard ImageNet dataset and its validation set, which the citing paper uses as a benchmark for evaluating the performance of the students in in-domain generalization.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[25]\", \"Explanation\": \"The cited work provides the robust ResNet50 teacher model that the citing paper uses to perform knowledge distillation in the in-domain evaluation of ImageNet.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[19]\", \"Explanation\": \"The cited work provides the ImageNet-Style dataset that the citing paper uses in the in-domain evaluation of ImageNet.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[23]\", \"Explanation\": \"The cited work provides the ImageNet-C dataset and the mean-corruption-error metric that the citing paper uses in the in-domain evaluation of ImageNet.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[59]\", \"Explanation\": \"The cited work introduces the Cutmix data augmentation technique, which the citing paper adopts in their research to improve the performance of their models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[61]\", \"Explanation\": \"The cited work introduces the Mixup data augmentation technique, which the citing paper uses in their research to improve the performance of their models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[24]\", \"Explanation\": \"The cited work introduces the AugMix data augmentation technique, which the citing paper uses in their research to improve the performance of their models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[39]\", \"Explanation\": \"The cited work introduces the Trivialaugment data augmentation technique, which the citing paper uses in their research to improve the performance of their models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work proposes a state-of-the-art method for knowledge distillation, which the citing paper uses as a benchmark to compare the performance of their own research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"Appendix A\", \"Explanation\": \"The cited appendix provides details on the grid search performed to select the hyperparameters for the augmentor-specific settings in the research of the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[6,13]\", \"Explanation\": \"The cited works provide a state-of-the-art knowledge distillation approach that the citing paper adopts in their research, which involves training for a long time (over 9,000 epochs) to achieve the desired results.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[44,47,49,50]\", \"Explanation\": \"The cited works on diffusion models are mentioned as a potential direction for future research in improving the proposed methods in the citing paper. The cited works are seen as a source of inspiration for exploring more flexible augmentor models with a semantically meaningful latent space.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[40]\", \"Explanation\": \"The cited work by Nix et al. provides the training setup and model architectures that the citing paper follows in their experiments on MNIST, serving as a methodological basis for the study conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[30]\", \"Explanation\": \"The cited work, Adam optimizer, is adopted in the citing paper to train the model on the CIFAR10 dataset, indicating a methodological basis for the research conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[10]\", \"Explanation\": \"The cited work by Child provides the pre-trained VAE model that the augmentor training in the citing paper is based on.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[8]\", \"Explanation\": \"The cited work introduces the Lion optimizer, which the citing paper adopts in their training process to optimize the learning process.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[36]\", \"Explanation\": \"The cited work provides the cosine schedule for the learning-rate annealing, which the citing paper utilizes in their training process to control the learning rate.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work presents the use of automatic mixed precision and 8bit optimization in training, which the citing paper implements in their training process to improve training efficiency.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[61]\", \"Explanation\": \"The cited work on Mixup is used as a training method for the student model, which is adopted in the citing paper to improve the performance of the model during training.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[59]\", \"Explanation\": \"The cited work on CutMix is also used as a training method for the student model, which is adopted in the citing paper to further enhance the performance of the model during training.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[39]\", \"Explanation\": \"The cited work on Trivial-Augment is used to provide a data augmentation method for the ViT-S baseline, which is adopted in the citing paper to improve the performance of the model during training.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[66]\", \"Explanation\": \"The cited work on random pixel erasing is used to provide a data augmentation method for the ViT-S baseline, which is adopted in the citing paper to further enhance the performance of the model during training.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[51]\", \"Explanation\": \"The cited work on label-smoothing is used to provide a method for optimizing the standard cross-entropy loss, which is adopted in the citing paper to improve the performance of the model during training.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[24]\", \"Explanation\": \"The cited work introduces the AugMix augmentation technique, which the citing paper utilizes in their experiments for the plain KD configuration.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[6,41]\", \"Explanation\": \"The cited works provide high-performing knowledge distillation experiments that serve as a benchmark for the evaluation of student and teacher models in the citing paper.\"}]"}}},{"rowIdx":92,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_0"],"heading":"Introduction","publication_ref":["b5","b11","b8","b19","b6","b22","b26","b3","b28","b33","b11","b8","b7","b20","b17","b32","b16","b24","b29"],"table_ref":[],"text":"Information extraction (IE) aims to discern meaningful information from unstructured data sources (Grishman, 2015). A traditional IE pipeline contains an array of tasks, which include, but are not limited to, Named Entity Recognition (NER) (Lample et al., 2016), Entity Linking (EL) (Kolitsas et al., 2018), Entity Typing (ET) (Ren et al., 2016), Relation Extraction (RE) (Huguet Cabot and Navigli, 2021), etc. IE plays a vital role in knowledge graph construction (Schneider et al., 2022), search engine (Wang et al., 2022), and document analysis (Chiticariu et al., 2010;Wang et al., 2018;Zhong et al., 2020).\nMost existing IE methods center around a closedworld setting with predefined ontologies. For instance, NER generally extracts named entities within several categories (Lample et al., 2016); EL focuses on associating mentions with a predefined ontology (Kolitsas et al., 2018). Furthermore, conventional closed-world IE usually extracts all information without focusing on desired targets. To better address these issues, we introduce Openworld Information Extraction (Open-world IE) to accommodate broad and diverse requests related to entity profiles surpassing predefined ontologies' limits. Specifically, Open-world IE accepts an unstructured corpus and an instruction that characterizes target entities, identifies all entities within the context, and generates entity profiles, as shown in Fig. 1. Open-world IE aims to break the ontology limitations in traditional IE and pursuit generalization to unseen instructions and out-of-ontology cases. Past research on this topic has predomi-arXiv:2305.14898v1 [cs.CL] 24 May 2023 nantly focused on individual subtasks of IE, such as EL (Iurshina et al., 2022;Ruas and Couto, 2022) and OpenIE (Niklaus et al., 2018;Bhardwaj et al., 2019a). Consequently, a noticeable gap exists in comprehensive end-to-end studies aiming to create more extensive entity profiles within an open-world setting.\nWith the emergence of large language models (LLMs) (Zhao et al., 2023), generative IE based on LLMs holds substantial promise in addressing this open-world challenge, given their exceptional generalization capabilities. Open-world IE can also serve as a pivotal capability for integrating plugins into the ChatGPT system, since it provides a flexible communication interface between LLMs and their plugins. Nevertheless, existing research on LLMs reveals that they typically do not function as zero-shot learners in IE, necessitating appropriate instruction tuning to enhance their IE capabilities (Ma et al., 2023;Wadhwa et al., 2023). Therefore, instruction tuning (Wei et al., 2022) becomes critical in endowing LLMs with Open-world IE abilities.\nTo combat these issues, we develop PIVOINE (Instruction-following Open-world Information Extraction). PIVOINE is an LLM designed for Open-world IE. We formulate Open-world IE as an instruction-following auto-regressive generative task to generate comprehensive entity profiles in JSON. We cover eight popular categories of instructions in various granularities. Each category of instruction imposes specific constraints on candidate entities. In pursuit of generalization over unseen instructions and out-of-ontology entities, we develop an instruction tuning dataset INSTRUCTOPENWIKI for Open-world IE, which includes diverse instructions that endows PIVOINE with strong instruction following capability. INSTRUCTOPENWIKI incorporates rich entity and relation annotations, various instructions, and a delicate design of the out-ofontology evaluation set, which contributes to the generalization of both unseen instructions and outof-ontology entities.\nThe contributions of this work are three-fold. First, we propose the definition of open-world IE and develop PIVOINE, which performs IE without the limitations of predefined ontology. This flexibility allows for its generalization abilities and application across diverse downstream scenarios. Second, we construct a substantial Open-world IE dataset INSTRUCTOPENWIKI. Third, we explore a comprehensive evaluation for Open-world IE. We meticulously design an open-world evaluation set incorporated in INSTRUCTOPENWIKI to assess Open-world IE capabilities thoroughly, focusing on the generalization of unseen instructions and out-of-ontology entities. Our contributions are verified with experiments and multifaceted analysis. Most notably, PIVOINE exhibits impressive generalization capabilities on unseen instructions and out-of-ontology cases, demonstrating its robust potential to address the open-world challenge effectively."},{"figure_ref":[],"heading":"Related Works","publication_ref":["b32","b2","b4","b23","b29","b18","b30","b27","b13","b34","b31"],"table_ref":[],"text":"Large Language Models. Large language models (LLMs) is an emerging topic summarized in a recent survey Zhao et al. (2023). Therefore, we only provide a highly-selective review. Brown et al. (2020) train an auto-regressive language model GPT-3 with 175 billion parameters, showing extraordinary task-agnostic few-shot performance. Chowdhery et al. (2022) develop a Pathways Language Model PALM and scale it up to 540 billion parameters. Scao et al. (2022) propose BLOOM, open-access LLMs from 560 million to 175 billion parameters. Touvron et al. (2023) develop LLAMA, a more efficient public-accessible LLM. We use BLOOM as the backbone since it was the latest public LLM pretrained on a diverse corpus, including codes. However, other latest LLMs, such as LLAMA, can also be easily tuned on our dataset to acquire open-world IE abilities.\nInstruction Tuning. Instruction tuning is an emergent paradigm that finetunes LLMs on datasets described by instructions. Wei et al. (2022) finetune an LLM with 175 billion parameters on various NLP datasets with instruction templates and proof instruction tuning can significantly improve zero-shot performance. Ouyang et al. (2022) show supervised instruction tuning and finetuning with human feedback helps LLMs align with human intent. This work is further extended by OPENAI and becomes the product CHATGPT2 used as a baseline in our work. In this work, we create an instructionfollowing dataset INSTRUCTOPENWIKI for openworld IE and employ instruction tuning to empower LLMs with Open-world IE abilities.\nInformation Extraction. Instruction-following IE reformulates IE into a generation task with instruc-tions describing target information. We mainly present two concurrent works as this is an emerging topic. Wei et al. (2023) solve IE as a multi-turn question-answering format by providing predefined instructions to CHATGPT. Wang et al. (2023) proposes an instruction-tuning IE benchmark and develops a unified IE method. However, all these works are based on the closed-world setting and have not adapted to Open-world IE, which is exactly our focus in this work. To our best knowledge, PIVOINE is the first work exploring instructionfollowing open-world IE. Previous explorations are limited to different sub-fields of IE, such as the NIL problem in EL (Lin et al., 2012) and open information extraction (Zhou et al., 2022). Open-world knowledge graph completion (KGC) (Ye et al., 2022) also completes existing KGs by creating unseen entities and conducting link prediction to existing entities. However, open-world KGC mainly uses structured information within KGs while we leverage rich unstructured corpus."},{"figure_ref":[],"heading":"Methods","publication_ref":[],"table_ref":[],"text":"In this section, we describe PIVOINE (Instructionfollowing Open-world Information Extraction), an open-world generative IE method with instructionfollowing abilities. We will introduce the preliminaries ( §3.1), instruction-following open-world IE ( §3.2), and construction of the dataset INSTRUC-TOPENWIKI ( §3.3)."},{"figure_ref":["fig_0","fig_2","fig_4"],"heading":"Preliminaries","publication_ref":[],"table_ref":[],"text":"Problem Definition. Open-world IE aims to extract entity profiles from unstructured texts without predefined ontologies by following specific instructions. In our task, the entity profile includes an entity mention, a canonical title, types, a description, aliases, and relations, as shown in Fig. 1. Specifically, mentions are text spans in input corpus that are linked to entities; types are a list of phrases that an entity is an instance of; aliases are a list of synonyms; relations are a list of relation titles between extracted entities within the input. Given a document and an instruction describing a specific constraint about target entities, such as types or descriptions, Open-world IE methods are expected to generate entity profiles that fixed constraints in the instruction.\nMethod Overview. We solve Open-world IE by instruction tuning of LLMs. As shown in Fig. 2, we first reformulate Open-world IE into auto-regressive generation by linearizing the structure knowledge into the JSON format ( §3.2). We apply instruction tuning to empower PIVOINE extracting different entities following instructions within eight instruction categories in different granularities. To do so, we build INSTRUCTOPENWIKI, a large-scale instruction-following open-world IE dataset. As presented in Fig. 3, INSTRUCTOPEN-WIKI is created by weak supervision between largescale corpus and existing knowledge base ( §3.3). Then, we augment the dataset with diverse instructions and rephrase them to enrich semantic diversity. We also comprehensively evaluate PIVOINE on the open-world evaluation set ( §4)."},{"figure_ref":[],"heading":"Instruction Tuning for Open-world IE","publication_ref":[],"table_ref":[],"text":"Instruction tuning is a multi-task learning paradigm that uses instructions to prompt models to generate proper outputs. Leveraging the strong generalization abilities of LLMs to pursue generalization on unseen instructions and out-of-ontology cases, we reformulate Open-world IE as an instructionfollowing generation task. To do so, we create diverse instructions and linearize structured IE outputs into JSON sequences. And then we finetune LLMs in a supervised setup of instruction tuning (SFT), training LLMs to generate the targeted JSON sequence of entity profiles as the output.\nInstruction Design. The diversity of instruction categories are essential for the generalization of unseen instructions. We first manually designed eight instruction categories with varying granularities to introduce extraction constraints that one might face in real-world applications: • Default: Extract all entities and relations in the input without any additional requirements. • Base Type: Extract entities of given base types.\nWe define base types as fine-grained types in Wikidata, building from the \"P31 (instance of)\" properties of entities in Wikidata. Please extract two humans in the context."},{"figure_ref":[],"heading":"{ % Open Entity Extraction","publication_ref":[],"table_ref":[],"text":"\"entities\": [{ % 2 entities \"mention\": \"country music\", % MD \"title\": \"Country music\", % EL \"type\": [\"Genre\"], % ET \"description\": \"…\", % EIG(Desc.) \"aliases\": […] % EIG(Aliases) }, { \"mention\": \"I Love 'Em All\", \"title\": \"I Love 'Em All\", \"type\": [\"Album\"], \"description\": \"…\" }],"},{"figure_ref":[],"heading":"% Open Relation Extraction","publication_ref":[],"table_ref":[],"text":"\"triplets\": [{ \"head\": \"I Love 'Em All\", \"tail\": \"country music\", \"relations\": [\"genre\"] }]}"},{"figure_ref":[],"heading":"Auto-regressive Generation JSON","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Number+ Abstract Type","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Instruction Categories","publication_ref":[],"table_ref":[],"text":"Please identify an entity with the following description: the American country music artist who recorded the song \"Party Time\".\nPlease identify two entities in the following categories: Genre, Album."},{"figure_ref":[],"heading":"{","publication_ref":[],"table_ref":[],"text":"\"entities\": [{ % 2 Human entities \"mention\": \"T.G.Sheppard\", \"title\": \" T.G.Sheppard \", \"type\": [\"Human\"], \"description\": \"American country music singer-songwriter\", \"aliases\": […] }, { \"mention\": \"Bruce Channel\", \"title\": \"Bruce Channel\", \"type\": [\"Human\"], \"description\": \"American musician\" }],…}"},{"figure_ref":[],"heading":"Auto-regressive Generation JSON","publication_ref":[],"table_ref":[],"text":"Please identify an entity with the following description: the American country music artist who wrote the song \"Party Time\"."},{"figure_ref":[],"heading":"{","publication_ref":[],"table_ref":[],"text":"\"entities\": [{ % 1 entity fit the description in the query \"mention\": \"Bruce Channel\", \"title\": \"Bruce Channel\", \"type\": [\"Human\"], \"description\": \"…\", }],…}"},{"figure_ref":[],"heading":"Auto-regressive Generation JSON","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_2"],"heading":"{","publication_ref":[],"table_ref":[],"text":"\"entities\": [{ % 1 entity fit the description in the query \"mention\": \" T. G. Sheppard\", \"title\": \"T. G. Sheppard\", \"type\": [\"Human\"], \"description\": \"…\", \"aliases\": [\"William Neal Browder\", \"Brian Stacy\"] }],…} ultra fine-grained instructions that require openworld IE methods directly understand diverse descriptions. As shown in the green cases in Fig. 2, descriptions can be phrases or sentences that describe the properties of entities."},{"figure_ref":[],"heading":"Auto-regressive Generation JSON","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Number+","publication_ref":["b31","b14","b25","b15"],"table_ref":[],"text":"• Importance: Extract the top-K most important entities. The importance of entities is defined as entity priorities in Wikidata. This instruction requires Open-world IE methods to rank entities with inherent priorities properly and only provide the top-K results. • Number: Extract a specific number of entities in the input document. Unlike other categories, instructions with number requires our method to extract partial information from the input and the answer is not unique. Therefore, we separately analyze these categories in evaluation. • Number+Base Type: Cross instructions between categories Number and Base Type. • Number+Abstract Type: Cross instructions between categories Number and Abstract Type. We manually design an instruction template for each category. Then we ask CHATGPT to rephrase the manually designed templates, enhancing the semantic diversity of these templates. Details of original and rephrased templates are shown in Tab. 8. CHATGPT prompts we use for rephrasing seed templates are shown in Tab. 22. We train PIVOINE to follow the top six single instructions. And we add the last two cross instructions in evaluation to evaluate generalization on unseen instructions.\nLinearization. Previously, various techniques have been explored to linearize the structured information in the generative IE (Ye et al., 2022) but either lack semantics or require additional training for special tokens (Lou et al., 2023;Wang et al., 2021;Lu et al., 2022). To better leverage pretrained knowledge, we present a novel idea that uses the JSON (JavaScript Object Notation) format to linearize heterogeneous structured entity profiles. It is primarily used to transmit data in web applications, so it frequently appears in codes. Therefore, LLMs pretrained on codes are familiar with the JSON schema, avoiding additional training for any special tokens or manually-defined structure templates. Furthermore, JSON uses a text-based syntax with key-value pairs, capturing additional semantics in natural language by keys and providing flexible structures. This linearization aggregates multiple IE subtasks, revealing the chain of thoughts in IE employed in traditional pipeline methods."},{"figure_ref":["fig_4","fig_4"],"heading":"Instruction Dataset Construction","publication_ref":[],"table_ref":[],"text":"Learning from a large-scale instruction tuning dataset with a rich corpus and large training ontology contributes to the generalization of out-ofontology cases. However, building a large-scale Open-world IE dataset by manual annotations is infeasible since identifying entities in text and linking them with entity profiles require tremendous human effort. Therefore, we develop a weakly supervised method that automatically creates the { % Open Entity Extraction \"entities\": [{ \"mention\": \"country music\", % MD \"title\": \"Country music\", % EL \"type\": […], % ET \"description\": \"…\", % EIG \"aliases\": […] }, { \"mention\": \"I Love 'Em All\", \"title\": \"I Love 'Em All\", \"type\": [\"Album\"], \"description\": \"…\" }], % Open Relation Extraction \"triplets\": [{ \"head\": \"I Love 'Em All\", \"tail\": \"country music\", \"relations\": dataset INSTRUCTOPENWIKI for instruction tuning.\nWeak Supervision. Our dataset is created by aligning anchor links in Wikipedia3 to entity profiles in its corresponding knowledge base Wikidata4 by the wiki identifiers, shown in the left part of Fig. 3. Wikipedia is a large corpus covering various domains, while Wikidata contains rich world knowledge. Wikipedia and Wikidata are frequently revised by contributors worldwide, ensuring precision and being up-to-date. All anchor links in Wikipedia are manually annotated, so linking between mentions and entities is reliable. We only use the leading paragraph in each Wikipedia article since it contains the richest anchor links. Besides, anchor links referring to the same entity may only be marked once the first time within an article, so using the rest of the paragraphs will face higher risks of missing mention annotations. We retrieve four fields from Wikidata as its profile for each linked entity, including the canonical title, types, description, and aliases. Canonical titles are English labels of entities; types of entities are derived from \"instance of (P31)\" properties. After identifying all entities in a paragraph, we employ distant supervision to identify relations between these entities from the knowledge base of Wikidata as Instruction Augmentation. We further augment the dataset with predefined instructions as shown in the middle of Fig. 3. We generate an instructiontuning sample with the default instruction and randomly select one another from six training categories for each sample. All instructions focus on entities, and we also filter out triplets whose head or tail mentions are filtered out during the augmentation to ensure alignment. Specifically, we augment samples with abstract type instructions using parent \"P279\" to replace base types."},{"figure_ref":[],"heading":"Evaluation","publication_ref":[],"table_ref":[],"text":"Open-world IE focuses on extracting unseen outof-ontology entities and relations. Therefore, we create an open-world evaluation set with rich outof-ontology cases and design metrics for evaluating such performance. Metrics. Although defining Open-world IE as an end-to-end entity profile generation task, we still split it into six tasks in evaluation to provide more comprehensive analyses: (1) Mention Detection (MD) corresponds to the correctness of the \"mention\" key in the JSON output."},{"figure_ref":["fig_0"],"heading":"Open-world Evaluation","publication_ref":["b34"],"table_ref":[],"text":"(2) Entity Linking (EL) is related to the \"title\" key, evaluating whether models generate proper canonical titles for mentions. We use hard and soft matching based on a ROUGE-L F1 threshold as the criterion.\n(3) Entity Typing (ET) requires models generate entity types for entities. (4) Open Relation Extraction (RE) is related to the \"triplets\" field in the JSON output. We learn from OpenIE evaluation (Zhou et al., 2022) and use CaRB (Bhardwaj et al., 2019b) to evaluate triplet generation performance. We calculate metrics based on CaRB with the ROUGE-L matcher.\n(5) Description Generation (EIG-Desc.) requires models to generate description for generated entities. We report the average ROUGE-L F1 for this task. (6) Aliases Generation (EIG-Aliases) is related to the \"aliases\" field, expecting models to generate aliases for predicted entities. We report precision, recall, and F1 scores on each task except description generation. We randomly select three rephrased templates for each sample in the test set and report the average metric with standard deviation.\nUnseen Ontologies. We explore the out-ofontology generalization by separately analyzing the recall of training (Before 05/30/2022) and out-oftraining (After 05/30/2022) entities. For instance, 2023 ATP Tour (Q111441127) shown in Fig. 1 is an unseen entity introduced to Wikidata after 05/20/2022. Open-world IE methods are proven to have the great generalization of unseen ontologies if they can extract this entity from the latest corpus.\nUnseen Instructions. We also split the test set into samples with unseen and seen instructions under the most fine-grained category \"Description\". Unseen instructions query context with constraints that are not in the training instructions. For example, \"men's tennis circuit\" is a description not shown in any training instructions. So the instruction extracting entities with such description is considered as an unseen instruction. Similarly, we filter out unseen instructions for other categories and separately evaluate Open-world IE performance on this split. The unseen proportions in each category are shown in Tab. 4. We also separately evaluate number-related instructions as these partial extraction instructions have no unique correct answers."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":"In this section, we carry out comprehensive evaluation of our proposed PIVOINE. Specifically, we present experimental setup ( §5.1), main results ( §5.2), and further analysis of PIVOINE ( §5.3)."},{"figure_ref":[],"heading":"Experimental Setup","publication_ref":[],"table_ref":[],"text":"Baselines. We employ ChatGPT as our main baseline since no instruction-following open-world IE methods exist to our best knowledge.\n(1) CHAT-GPT is an instruction-following LLM that can handle various tasks. The detailed configuration of CHATGPT is described in Appx. §C.\n(2) CHAT-GPT W/DEMO is a stronger baseline with a oneshot demo based on the CHATGPT baseline. We provide CHATGPT an example of the instruction (2022). PIVOINE-1b is trained on 64 NVIDIA V100 GPU for 92 hours. PIVOINE-7b is trained on 256 NVIDIA V100 GPU for 54 hours. We develop our models with Megatron-Deepspeed from Big-Science. We trained PIVOINE-1b for 10,294 steps with a global batch size of 1,024 and PIVOINE-7b for 5,000 steps with a global batch size of 2,048.\nThe training steps are selected by evaluation of performance on the development set. We infer on 256 NVIDIA V100 GPU within 30 minutes with generation parameters in Appx. §E."},{"figure_ref":[],"heading":"Main Results","publication_ref":[],"table_ref":[],"text":"We present our main results in three aspects: overall performance, generalization study on entities within or out of the training ontology, and unseen instructions.\nOverall Performance. Tab. 1 shows overall performance on six subtasks. We report the macro average of F1 scores with all instruction categories. Table 3: Main results of the instruction generalization evaluation on \"Description\" instruction category. Headers are the same as Tab. 1. Partition denotes the unseen and seen descriptions in instructions. To avoid data leakage, descriptions as instrcutions are not considered in descirption generation tasks. Subscript scores are the deviation under three different rephrased instructions for each sample. The best scores in each partition are marked in bold.\nWe only report performance on default instruction for GENRE and OPENIE6 for reference since they have no instruction-following abilities. Fig. 4 shows a more detailed analysis of each instruction category in each task. Comparing six radar charts, we identify MD, EL, and EIG(Desc.) are three tasks on which all methods are easy to generalize to out-of-ontology entities since their upper and lower parts of plots are nearly symmetric. We notice such generalization is harder in RE since it is an end-to-end task requiring precise outof-ontology MD first. ET and EIG(Aliases) are the most challenging for all methods because they require a comprehensive list of types and aliases. As for instruction categories, we find PIVOINE-7b consistently outperforms other baselines on all instruction categories, except for the Base Type instructions in ET and the Description instructions in EIG(Aliases).\nGeneralization to unseen instructions. We evaluate the generalization abilities of both unseen instructions and out-of-ontology cases and present the results in Tab. 3. As introduced in the last paragraph of §4, unseen instructions include descriptions that have not appeared in the training samples. We notice PIVOINE-7b achieves the best performance on almost all tasks under both seen and unseen partitions except for EIG(Aliases) in the unseen partition, where CHATGPT W/DEMO outperforms other methods. Therefore, PIVOINE-7b shows extraordinary generalization abilities to two main challenges in Open-world IE, especially in MD and EL."},{"figure_ref":[],"heading":"Analysis","publication_ref":[],"table_ref":[],"text":"We provide further analyses to reveal model behaviors on instruction-following open-world IE.\nInstruction Following. We analyze instruction following qualities in three aspects as shown in Fig. 5. First, we analyze the JSON decoding error rates with each method on different instructions. Generating correct JSON schema is the essential requirement for parsing extracted information. As presented in the left figure in Fig. 5, ChatGPT-based methods fail to ensure the valid JSON schema, especially on unseen cross instructions. PIVOINE-1b is better than baselines on average but still faces decoding errors. However, PIVOINE-7b has no JSON decoding errors on all instruction categories, even for unseen cross instructions. Therefore, PIVOINE-7b is a trustworthy model consistently generating valid JSON schema.\nWe also compare entity number instruction failure rates on four number-related instruction categories. The results are shown in the middle of Fig. 5. PIVOINE-7b still has the lowest failure rates for extracting the correct number of entities, which is close to zero on two trained instruction categories. All methods fail to follow number instructions when receiving untrained cross instructions, suggesting cross instructions are still challenging for current methods. Furthermore, results in Tab. 21 shows that number-related instructions do not hinder the precision of MD and EL. These instruction categories provide partial extraction abilities to PIVOINE so that we can control the precisionrecall trade-off by specifying the number of entities we need in the instruction. Similar patterns are observed in the entity type instruction failure rate analysis shown in the right of Fig. 5. PIVOINE-7b only has half failure rates on cross instruction categories compared with the vanilla CHATGPT, showing PIVOINE-7b has much better type instruction following abilities. We also notice that following instructions with abstract types is significantly easier for PIVOINE than finegrained base types.\nHuman Evaluation. PIVOINE may extract correct relation triplets that are out of the scope of existing Wikidata. Therefore, we randomly select 100 relation triplets with unseen entities predicted by PIVOINE-7b to analyze its precision further. We consider the evidence of a relation triplet as the context from which PIVOINE-7b extracts the triplet. To minimize the manual effort, we first reformulate the output of PIVOINE-7b to a prompt and ask GPT-4 to provide judgment and explanation based on evidence supplied by PIVOINE. We also manually evaluate the correctness of relation triplets based on the same evidence without additional world knowledge. The accuracy of relation triplets provided by GPT-4 is 87%, and manual checking accuracy is 83%. The agreement between GPT-4 and the annotator is 87%, suggesting GPT-4 is capable of evaluating relation correctness. The latest Wikidata also verifies 8% cases. This evaluation shows PIVOINE-7b can precisely excavate new entities and relations from the corpus."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"We propose Open-world IE, a challenging task that aims to extract out-of-ontology entity profiles with instructions. Towards this grand mission, we create a large-scale instruction-following open-world IE dataset INSTRUCTOPENWIKI and develop PIVOINE by instruction tuning. We conduct extensive experiments on diverse instruction categories and different model scales, showing PIVOINE is a trustworthy LLM capable of following (possibly unseen) instructions in various granularities and extracting out-of-ontology entity profiles. Valuable future works include extending PIVOINE to a larger scale and exploring a more comprehensive range of instruction categories. detailed cases are shown in our Github repository."},{"figure_ref":[],"heading":"E Hyper-parameters","publication_ref":[],"table_ref":[],"text":"We infer PIVOINE with the official BLOOM inference project provided by HuggingFace7 . We generate with beam search without sampling and the number of beam is 4. The maximum number of generated tokens are 2,048. The other generation parameters are set to default. Table 9: Statistics of the instruction-following open-world dataset INSTRUCTOPENWIKI. We report the statistics of the corpus, including the number of articles, mentions, and triplets in the left section. The ontology statistics, including the number of unique entities, aliases, relations, and types, are reported in the middle section. The right section shows the proportions of mentions with descriptions, aliases, and types, respectively. Ent., Rel., and Desc. are short for entities, relations, and descriptions."},{"figure_ref":[],"heading":"F Comprehensive Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Instruction","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"datasets are available at https://github.com/Lukeming"},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"Work done during Keming Lu's internship at Tencent AI Lab."},{"figure_ref":[],"heading":"Appendix A Detailed Instructions","publication_ref":[],"table_ref":[],"text":"Tab. 8 shows original manually-designed seed instruction, rephrased instructions, and number of rephrased instructions for each category. We use ChatGPT to rephrased our original instructions. The prompt we used is shown in Tab. 22."},{"figure_ref":[],"heading":"B Dataset Statistics","publication_ref":[],"table_ref":[],"text":"We display the statistics of INSTRUCTOPENWIKI in Tab. 9. INSTRUCTOPENWIKI contains a rich corpus, which includes all head paragraphs of all Wikipedia articles with 39 million mentions and 19 million triplets. And INSTRUCTOPENWIKI is also annotated with a large ontology containing over 2 million entities with 21 thousand entity types and 962 relation types, ensuring it covers a wide range of domains. Besides, the entity information density of INSTRUCTOPENWIKI is also abundant, so models can be trained for extracting entity profiles efficiently. Table 4: Proportions of unseen and seen instructions in \"Base Type\", \"Abstract Type\", and \"Description\" in the test set.\nWe also create a small development set to select the best checkpoint balancing performance on seen and unseen cases. As shown in Tab. 9, the open-world evaluation test set contains rich unseen mentions (27.1%, 29,612/109,411) annotated by unseen entities (24.2%, 9,464/39,086)."},{"figure_ref":[],"heading":"C CHATGPT Configurations","publication_ref":[],"table_ref":[],"text":"We provide ChatGPT with the input context, instruction, and output JSON schema to prompt Chat-GPT to solve instruction-following open-world IE. We use ChatGPT in May 2022 and query it with the official API provided by OpenAI 5 . The detailed prompt we used is shown in Tab. 23."},{"figure_ref":[],"heading":"D Details of Human Evaluation","publication_ref":[],"table_ref":[],"text":"We use the May 1st, 2023 version of GPT-4. We query it via the chat platform of OpenAI 6 . The ----- Table 11: Main results of mention detection (MD). We report precision, recall, and F1 score of each model in this table. The subscript scores are the deviation under three different rephrased instructions for each instruction category. Subscript scores are the deviation under three different rephrased instructions for each sample. GENRE is an closed-world generative IE baseline without the instruction-following ability, so we only report GENRE scores with the default instruction. 13.00.3 19.40.5 15.60.4 29.90.2 35.00.4 32.30.3 46.00.1 39.90.2 42.70.1 59.30.2 57.60.1 58.40.1 Abstract Type 8.90.5 7.50.2 8.20.3 16.20.3 18.20.2 17.10.3 42.70.2 35.60.0 38.80.1 62.30.3 53.90.1 57.80.1 Description 2.30.1 4.40.1 3.10.1 6.40.1 6.50.1 6.40.1 44.90.1 43.80.4 44.40.1 64.10.1 63.20.1 63.60 "}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b5\",\n \"b11\",\n \"b8\",\n \"b19\",\n \"b6\",\n \"b22\",\n \"b26\",\n \"b3\",\n \"b28\",\n \"b33\",\n \"b11\",\n \"b8\",\n \"b7\",\n \"b20\",\n \"b17\",\n \"b32\",\n \"b16\",\n \"b24\",\n \"b29\"\n ],\n \"table_ref\": [],\n \"text\": \"Information extraction (IE) aims to discern meaningful information from unstructured data sources (Grishman, 2015). A traditional IE pipeline contains an array of tasks, which include, but are not limited to, Named Entity Recognition (NER) (Lample et al., 2016), Entity Linking (EL) (Kolitsas et al., 2018), Entity Typing (ET) (Ren et al., 2016), Relation Extraction (RE) (Huguet Cabot and Navigli, 2021), etc. IE plays a vital role in knowledge graph construction (Schneider et al., 2022), search engine (Wang et al., 2022), and document analysis (Chiticariu et al., 2010;Wang et al., 2018;Zhong et al., 2020).\\nMost existing IE methods center around a closedworld setting with predefined ontologies. For instance, NER generally extracts named entities within several categories (Lample et al., 2016); EL focuses on associating mentions with a predefined ontology (Kolitsas et al., 2018). Furthermore, conventional closed-world IE usually extracts all information without focusing on desired targets. To better address these issues, we introduce Openworld Information Extraction (Open-world IE) to accommodate broad and diverse requests related to entity profiles surpassing predefined ontologies' limits. Specifically, Open-world IE accepts an unstructured corpus and an instruction that characterizes target entities, identifies all entities within the context, and generates entity profiles, as shown in Fig. 1. Open-world IE aims to break the ontology limitations in traditional IE and pursuit generalization to unseen instructions and out-of-ontology cases. Past research on this topic has predomi-arXiv:2305.14898v1 [cs.CL] 24 May 2023 nantly focused on individual subtasks of IE, such as EL (Iurshina et al., 2022;Ruas and Couto, 2022) and OpenIE (Niklaus et al., 2018;Bhardwaj et al., 2019a). Consequently, a noticeable gap exists in comprehensive end-to-end studies aiming to create more extensive entity profiles within an open-world setting.\\nWith the emergence of large language models (LLMs) (Zhao et al., 2023), generative IE based on LLMs holds substantial promise in addressing this open-world challenge, given their exceptional generalization capabilities. Open-world IE can also serve as a pivotal capability for integrating plugins into the ChatGPT system, since it provides a flexible communication interface between LLMs and their plugins. Nevertheless, existing research on LLMs reveals that they typically do not function as zero-shot learners in IE, necessitating appropriate instruction tuning to enhance their IE capabilities (Ma et al., 2023;Wadhwa et al., 2023). Therefore, instruction tuning (Wei et al., 2022) becomes critical in endowing LLMs with Open-world IE abilities.\\nTo combat these issues, we develop PIVOINE (Instruction-following Open-world Information Extraction). PIVOINE is an LLM designed for Open-world IE. We formulate Open-world IE as an instruction-following auto-regressive generative task to generate comprehensive entity profiles in JSON. We cover eight popular categories of instructions in various granularities. Each category of instruction imposes specific constraints on candidate entities. In pursuit of generalization over unseen instructions and out-of-ontology entities, we develop an instruction tuning dataset INSTRUCTOPENWIKI for Open-world IE, which includes diverse instructions that endows PIVOINE with strong instruction following capability. INSTRUCTOPENWIKI incorporates rich entity and relation annotations, various instructions, and a delicate design of the out-ofontology evaluation set, which contributes to the generalization of both unseen instructions and outof-ontology entities.\\nThe contributions of this work are three-fold. First, we propose the definition of open-world IE and develop PIVOINE, which performs IE without the limitations of predefined ontology. This flexibility allows for its generalization abilities and application across diverse downstream scenarios. Second, we construct a substantial Open-world IE dataset INSTRUCTOPENWIKI. Third, we explore a comprehensive evaluation for Open-world IE. We meticulously design an open-world evaluation set incorporated in INSTRUCTOPENWIKI to assess Open-world IE capabilities thoroughly, focusing on the generalization of unseen instructions and out-of-ontology entities. Our contributions are verified with experiments and multifaceted analysis. Most notably, PIVOINE exhibits impressive generalization capabilities on unseen instructions and out-of-ontology cases, demonstrating its robust potential to address the open-world challenge effectively.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Works\",\n \"publication_ref\": [\n \"b32\",\n \"b2\",\n \"b4\",\n \"b23\",\n \"b29\",\n \"b18\",\n \"b30\",\n \"b27\",\n \"b13\",\n \"b34\",\n \"b31\"\n ],\n \"table_ref\": [],\n \"text\": \"Large Language Models. Large language models (LLMs) is an emerging topic summarized in a recent survey Zhao et al. (2023). Therefore, we only provide a highly-selective review. Brown et al. (2020) train an auto-regressive language model GPT-3 with 175 billion parameters, showing extraordinary task-agnostic few-shot performance. Chowdhery et al. (2022) develop a Pathways Language Model PALM and scale it up to 540 billion parameters. Scao et al. (2022) propose BLOOM, open-access LLMs from 560 million to 175 billion parameters. Touvron et al. (2023) develop LLAMA, a more efficient public-accessible LLM. We use BLOOM as the backbone since it was the latest public LLM pretrained on a diverse corpus, including codes. However, other latest LLMs, such as LLAMA, can also be easily tuned on our dataset to acquire open-world IE abilities.\\nInstruction Tuning. Instruction tuning is an emergent paradigm that finetunes LLMs on datasets described by instructions. Wei et al. (2022) finetune an LLM with 175 billion parameters on various NLP datasets with instruction templates and proof instruction tuning can significantly improve zero-shot performance. Ouyang et al. (2022) show supervised instruction tuning and finetuning with human feedback helps LLMs align with human intent. This work is further extended by OPENAI and becomes the product CHATGPT2 used as a baseline in our work. In this work, we create an instructionfollowing dataset INSTRUCTOPENWIKI for openworld IE and employ instruction tuning to empower LLMs with Open-world IE abilities.\\nInformation Extraction. Instruction-following IE reformulates IE into a generation task with instruc-tions describing target information. We mainly present two concurrent works as this is an emerging topic. Wei et al. (2023) solve IE as a multi-turn question-answering format by providing predefined instructions to CHATGPT. Wang et al. (2023) proposes an instruction-tuning IE benchmark and develops a unified IE method. However, all these works are based on the closed-world setting and have not adapted to Open-world IE, which is exactly our focus in this work. To our best knowledge, PIVOINE is the first work exploring instructionfollowing open-world IE. Previous explorations are limited to different sub-fields of IE, such as the NIL problem in EL (Lin et al., 2012) and open information extraction (Zhou et al., 2022). Open-world knowledge graph completion (KGC) (Ye et al., 2022) also completes existing KGs by creating unseen entities and conducting link prediction to existing entities. However, open-world KGC mainly uses structured information within KGs while we leverage rich unstructured corpus.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Methods\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we describe PIVOINE (Instructionfollowing Open-world Information Extraction), an open-world generative IE method with instructionfollowing abilities. We will introduce the preliminaries ( §3.1), instruction-following open-world IE ( §3.2), and construction of the dataset INSTRUC-TOPENWIKI ( §3.3).\"\n },\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_2\",\n \"fig_4\"\n ],\n \"heading\": \"Preliminaries\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Problem Definition. Open-world IE aims to extract entity profiles from unstructured texts without predefined ontologies by following specific instructions. In our task, the entity profile includes an entity mention, a canonical title, types, a description, aliases, and relations, as shown in Fig. 1. Specifically, mentions are text spans in input corpus that are linked to entities; types are a list of phrases that an entity is an instance of; aliases are a list of synonyms; relations are a list of relation titles between extracted entities within the input. Given a document and an instruction describing a specific constraint about target entities, such as types or descriptions, Open-world IE methods are expected to generate entity profiles that fixed constraints in the instruction.\\nMethod Overview. We solve Open-world IE by instruction tuning of LLMs. As shown in Fig. 2, we first reformulate Open-world IE into auto-regressive generation by linearizing the structure knowledge into the JSON format ( §3.2). We apply instruction tuning to empower PIVOINE extracting different entities following instructions within eight instruction categories in different granularities. To do so, we build INSTRUCTOPENWIKI, a large-scale instruction-following open-world IE dataset. As presented in Fig. 3, INSTRUCTOPEN-WIKI is created by weak supervision between largescale corpus and existing knowledge base ( §3.3). Then, we augment the dataset with diverse instructions and rephrase them to enrich semantic diversity. We also comprehensively evaluate PIVOINE on the open-world evaluation set ( §4).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Instruction Tuning for Open-world IE\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Instruction tuning is a multi-task learning paradigm that uses instructions to prompt models to generate proper outputs. Leveraging the strong generalization abilities of LLMs to pursue generalization on unseen instructions and out-of-ontology cases, we reformulate Open-world IE as an instructionfollowing generation task. To do so, we create diverse instructions and linearize structured IE outputs into JSON sequences. And then we finetune LLMs in a supervised setup of instruction tuning (SFT), training LLMs to generate the targeted JSON sequence of entity profiles as the output.\\nInstruction Design. The diversity of instruction categories are essential for the generalization of unseen instructions. We first manually designed eight instruction categories with varying granularities to introduce extraction constraints that one might face in real-world applications: • Default: Extract all entities and relations in the input without any additional requirements. • Base Type: Extract entities of given base types.\\nWe define base types as fine-grained types in Wikidata, building from the \\\"P31 (instance of)\\\" properties of entities in Wikidata. Please extract two humans in the context.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"{ % Open Entity Extraction\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\\\"entities\\\": [{ % 2 entities \\\"mention\\\": \\\"country music\\\", % MD \\\"title\\\": \\\"Country music\\\", % EL \\\"type\\\": [\\\"Genre\\\"], % ET \\\"description\\\": \\\"…\\\", % EIG(Desc.) \\\"aliases\\\": […] % EIG(Aliases) }, { \\\"mention\\\": \\\"I Love 'Em All\\\", \\\"title\\\": \\\"I Love 'Em All\\\", \\\"type\\\": [\\\"Album\\\"], \\\"description\\\": \\\"…\\\" }],\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"% Open Relation Extraction\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\\\"triplets\\\": [{ \\\"head\\\": \\\"I Love 'Em All\\\", \\\"tail\\\": \\\"country music\\\", \\\"relations\\\": [\\\"genre\\\"] }]}\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Auto-regressive Generation JSON\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Number+ Abstract Type\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Instruction Categories\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Please identify an entity with the following description: the American country music artist who recorded the song \\\"Party Time\\\".\\nPlease identify two entities in the following categories: Genre, Album.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"{\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\\\"entities\\\": [{ % 2 Human entities \\\"mention\\\": \\\"T.G.Sheppard\\\", \\\"title\\\": \\\" T.G.Sheppard \\\", \\\"type\\\": [\\\"Human\\\"], \\\"description\\\": \\\"American country music singer-songwriter\\\", \\\"aliases\\\": […] }, { \\\"mention\\\": \\\"Bruce Channel\\\", \\\"title\\\": \\\"Bruce Channel\\\", \\\"type\\\": [\\\"Human\\\"], \\\"description\\\": \\\"American musician\\\" }],…}\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Auto-regressive Generation JSON\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Please identify an entity with the following description: the American country music artist who wrote the song \\\"Party Time\\\".\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"{\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\\\"entities\\\": [{ % 1 entity fit the description in the query \\\"mention\\\": \\\"Bruce Channel\\\", \\\"title\\\": \\\"Bruce Channel\\\", \\\"type\\\": [\\\"Human\\\"], \\\"description\\\": \\\"…\\\", }],…}\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Auto-regressive Generation JSON\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"{\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\\\"entities\\\": [{ % 1 entity fit the description in the query \\\"mention\\\": \\\" T. G. Sheppard\\\", \\\"title\\\": \\\"T. G. Sheppard\\\", \\\"type\\\": [\\\"Human\\\"], \\\"description\\\": \\\"…\\\", \\\"aliases\\\": [\\\"William Neal Browder\\\", \\\"Brian Stacy\\\"] }],…} ultra fine-grained instructions that require openworld IE methods directly understand diverse descriptions. As shown in the green cases in Fig. 2, descriptions can be phrases or sentences that describe the properties of entities.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Auto-regressive Generation JSON\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Number+\",\n \"publication_ref\": [\n \"b31\",\n \"b14\",\n \"b25\",\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"• Importance: Extract the top-K most important entities. The importance of entities is defined as entity priorities in Wikidata. This instruction requires Open-world IE methods to rank entities with inherent priorities properly and only provide the top-K results. • Number: Extract a specific number of entities in the input document. Unlike other categories, instructions with number requires our method to extract partial information from the input and the answer is not unique. Therefore, we separately analyze these categories in evaluation. • Number+Base Type: Cross instructions between categories Number and Base Type. • Number+Abstract Type: Cross instructions between categories Number and Abstract Type. We manually design an instruction template for each category. Then we ask CHATGPT to rephrase the manually designed templates, enhancing the semantic diversity of these templates. Details of original and rephrased templates are shown in Tab. 8. CHATGPT prompts we use for rephrasing seed templates are shown in Tab. 22. We train PIVOINE to follow the top six single instructions. And we add the last two cross instructions in evaluation to evaluate generalization on unseen instructions.\\nLinearization. Previously, various techniques have been explored to linearize the structured information in the generative IE (Ye et al., 2022) but either lack semantics or require additional training for special tokens (Lou et al., 2023;Wang et al., 2021;Lu et al., 2022). To better leverage pretrained knowledge, we present a novel idea that uses the JSON (JavaScript Object Notation) format to linearize heterogeneous structured entity profiles. It is primarily used to transmit data in web applications, so it frequently appears in codes. Therefore, LLMs pretrained on codes are familiar with the JSON schema, avoiding additional training for any special tokens or manually-defined structure templates. Furthermore, JSON uses a text-based syntax with key-value pairs, capturing additional semantics in natural language by keys and providing flexible structures. This linearization aggregates multiple IE subtasks, revealing the chain of thoughts in IE employed in traditional pipeline methods.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\",\n \"fig_4\"\n ],\n \"heading\": \"Instruction Dataset Construction\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Learning from a large-scale instruction tuning dataset with a rich corpus and large training ontology contributes to the generalization of out-ofontology cases. However, building a large-scale Open-world IE dataset by manual annotations is infeasible since identifying entities in text and linking them with entity profiles require tremendous human effort. Therefore, we develop a weakly supervised method that automatically creates the { % Open Entity Extraction \\\"entities\\\": [{ \\\"mention\\\": \\\"country music\\\", % MD \\\"title\\\": \\\"Country music\\\", % EL \\\"type\\\": […], % ET \\\"description\\\": \\\"…\\\", % EIG \\\"aliases\\\": […] }, { \\\"mention\\\": \\\"I Love 'Em All\\\", \\\"title\\\": \\\"I Love 'Em All\\\", \\\"type\\\": [\\\"Album\\\"], \\\"description\\\": \\\"…\\\" }], % Open Relation Extraction \\\"triplets\\\": [{ \\\"head\\\": \\\"I Love 'Em All\\\", \\\"tail\\\": \\\"country music\\\", \\\"relations\\\": dataset INSTRUCTOPENWIKI for instruction tuning.\\nWeak Supervision. Our dataset is created by aligning anchor links in Wikipedia3 to entity profiles in its corresponding knowledge base Wikidata4 by the wiki identifiers, shown in the left part of Fig. 3. Wikipedia is a large corpus covering various domains, while Wikidata contains rich world knowledge. Wikipedia and Wikidata are frequently revised by contributors worldwide, ensuring precision and being up-to-date. All anchor links in Wikipedia are manually annotated, so linking between mentions and entities is reliable. We only use the leading paragraph in each Wikipedia article since it contains the richest anchor links. Besides, anchor links referring to the same entity may only be marked once the first time within an article, so using the rest of the paragraphs will face higher risks of missing mention annotations. We retrieve four fields from Wikidata as its profile for each linked entity, including the canonical title, types, description, and aliases. Canonical titles are English labels of entities; types of entities are derived from \\\"instance of (P31)\\\" properties. After identifying all entities in a paragraph, we employ distant supervision to identify relations between these entities from the knowledge base of Wikidata as Instruction Augmentation. We further augment the dataset with predefined instructions as shown in the middle of Fig. 3. We generate an instructiontuning sample with the default instruction and randomly select one another from six training categories for each sample. All instructions focus on entities, and we also filter out triplets whose head or tail mentions are filtered out during the augmentation to ensure alignment. Specifically, we augment samples with abstract type instructions using parent \\\"P279\\\" to replace base types.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Evaluation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Open-world IE focuses on extracting unseen outof-ontology entities and relations. Therefore, we create an open-world evaluation set with rich outof-ontology cases and design metrics for evaluating such performance. Metrics. Although defining Open-world IE as an end-to-end entity profile generation task, we still split it into six tasks in evaluation to provide more comprehensive analyses: (1) Mention Detection (MD) corresponds to the correctness of the \\\"mention\\\" key in the JSON output.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Open-world Evaluation\",\n \"publication_ref\": [\n \"b34\"\n ],\n \"table_ref\": [],\n \"text\": \"(2) Entity Linking (EL) is related to the \\\"title\\\" key, evaluating whether models generate proper canonical titles for mentions. We use hard and soft matching based on a ROUGE-L F1 threshold as the criterion.\\n(3) Entity Typing (ET) requires models generate entity types for entities. (4) Open Relation Extraction (RE) is related to the \\\"triplets\\\" field in the JSON output. We learn from OpenIE evaluation (Zhou et al., 2022) and use CaRB (Bhardwaj et al., 2019b) to evaluate triplet generation performance. We calculate metrics based on CaRB with the ROUGE-L matcher.\\n(5) Description Generation (EIG-Desc.) requires models to generate description for generated entities. We report the average ROUGE-L F1 for this task. (6) Aliases Generation (EIG-Aliases) is related to the \\\"aliases\\\" field, expecting models to generate aliases for predicted entities. We report precision, recall, and F1 scores on each task except description generation. We randomly select three rephrased templates for each sample in the test set and report the average metric with standard deviation.\\nUnseen Ontologies. We explore the out-ofontology generalization by separately analyzing the recall of training (Before 05/30/2022) and out-oftraining (After 05/30/2022) entities. For instance, 2023 ATP Tour (Q111441127) shown in Fig. 1 is an unseen entity introduced to Wikidata after 05/20/2022. Open-world IE methods are proven to have the great generalization of unseen ontologies if they can extract this entity from the latest corpus.\\nUnseen Instructions. We also split the test set into samples with unseen and seen instructions under the most fine-grained category \\\"Description\\\". Unseen instructions query context with constraints that are not in the training instructions. For example, \\\"men's tennis circuit\\\" is a description not shown in any training instructions. So the instruction extracting entities with such description is considered as an unseen instruction. Similarly, we filter out unseen instructions for other categories and separately evaluate Open-world IE performance on this split. The unseen proportions in each category are shown in Tab. 4. We also separately evaluate number-related instructions as these partial extraction instructions have no unique correct answers.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we carry out comprehensive evaluation of our proposed PIVOINE. Specifically, we present experimental setup ( §5.1), main results ( §5.2), and further analysis of PIVOINE ( §5.3).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experimental Setup\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Baselines. We employ ChatGPT as our main baseline since no instruction-following open-world IE methods exist to our best knowledge.\\n(1) CHAT-GPT is an instruction-following LLM that can handle various tasks. The detailed configuration of CHATGPT is described in Appx. §C.\\n(2) CHAT-GPT W/DEMO is a stronger baseline with a oneshot demo based on the CHATGPT baseline. We provide CHATGPT an example of the instruction (2022). PIVOINE-1b is trained on 64 NVIDIA V100 GPU for 92 hours. PIVOINE-7b is trained on 256 NVIDIA V100 GPU for 54 hours. We develop our models with Megatron-Deepspeed from Big-Science. We trained PIVOINE-1b for 10,294 steps with a global batch size of 1,024 and PIVOINE-7b for 5,000 steps with a global batch size of 2,048.\\nThe training steps are selected by evaluation of performance on the development set. We infer on 256 NVIDIA V100 GPU within 30 minutes with generation parameters in Appx. §E.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Main Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We present our main results in three aspects: overall performance, generalization study on entities within or out of the training ontology, and unseen instructions.\\nOverall Performance. Tab. 1 shows overall performance on six subtasks. We report the macro average of F1 scores with all instruction categories. Table 3: Main results of the instruction generalization evaluation on \\\"Description\\\" instruction category. Headers are the same as Tab. 1. Partition denotes the unseen and seen descriptions in instructions. To avoid data leakage, descriptions as instrcutions are not considered in descirption generation tasks. Subscript scores are the deviation under three different rephrased instructions for each sample. The best scores in each partition are marked in bold.\\nWe only report performance on default instruction for GENRE and OPENIE6 for reference since they have no instruction-following abilities. Fig. 4 shows a more detailed analysis of each instruction category in each task. Comparing six radar charts, we identify MD, EL, and EIG(Desc.) are three tasks on which all methods are easy to generalize to out-of-ontology entities since their upper and lower parts of plots are nearly symmetric. We notice such generalization is harder in RE since it is an end-to-end task requiring precise outof-ontology MD first. ET and EIG(Aliases) are the most challenging for all methods because they require a comprehensive list of types and aliases. As for instruction categories, we find PIVOINE-7b consistently outperforms other baselines on all instruction categories, except for the Base Type instructions in ET and the Description instructions in EIG(Aliases).\\nGeneralization to unseen instructions. We evaluate the generalization abilities of both unseen instructions and out-of-ontology cases and present the results in Tab. 3. As introduced in the last paragraph of §4, unseen instructions include descriptions that have not appeared in the training samples. We notice PIVOINE-7b achieves the best performance on almost all tasks under both seen and unseen partitions except for EIG(Aliases) in the unseen partition, where CHATGPT W/DEMO outperforms other methods. Therefore, PIVOINE-7b shows extraordinary generalization abilities to two main challenges in Open-world IE, especially in MD and EL.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We provide further analyses to reveal model behaviors on instruction-following open-world IE.\\nInstruction Following. We analyze instruction following qualities in three aspects as shown in Fig. 5. First, we analyze the JSON decoding error rates with each method on different instructions. Generating correct JSON schema is the essential requirement for parsing extracted information. As presented in the left figure in Fig. 5, ChatGPT-based methods fail to ensure the valid JSON schema, especially on unseen cross instructions. PIVOINE-1b is better than baselines on average but still faces decoding errors. However, PIVOINE-7b has no JSON decoding errors on all instruction categories, even for unseen cross instructions. Therefore, PIVOINE-7b is a trustworthy model consistently generating valid JSON schema.\\nWe also compare entity number instruction failure rates on four number-related instruction categories. The results are shown in the middle of Fig. 5. PIVOINE-7b still has the lowest failure rates for extracting the correct number of entities, which is close to zero on two trained instruction categories. All methods fail to follow number instructions when receiving untrained cross instructions, suggesting cross instructions are still challenging for current methods. Furthermore, results in Tab. 21 shows that number-related instructions do not hinder the precision of MD and EL. These instruction categories provide partial extraction abilities to PIVOINE so that we can control the precisionrecall trade-off by specifying the number of entities we need in the instruction. Similar patterns are observed in the entity type instruction failure rate analysis shown in the right of Fig. 5. PIVOINE-7b only has half failure rates on cross instruction categories compared with the vanilla CHATGPT, showing PIVOINE-7b has much better type instruction following abilities. We also notice that following instructions with abstract types is significantly easier for PIVOINE than finegrained base types.\\nHuman Evaluation. PIVOINE may extract correct relation triplets that are out of the scope of existing Wikidata. Therefore, we randomly select 100 relation triplets with unseen entities predicted by PIVOINE-7b to analyze its precision further. We consider the evidence of a relation triplet as the context from which PIVOINE-7b extracts the triplet. To minimize the manual effort, we first reformulate the output of PIVOINE-7b to a prompt and ask GPT-4 to provide judgment and explanation based on evidence supplied by PIVOINE. We also manually evaluate the correctness of relation triplets based on the same evidence without additional world knowledge. The accuracy of relation triplets provided by GPT-4 is 87%, and manual checking accuracy is 83%. The agreement between GPT-4 and the annotator is 87%, suggesting GPT-4 is capable of evaluating relation correctness. The latest Wikidata also verifies 8% cases. This evaluation shows PIVOINE-7b can precisely excavate new entities and relations from the corpus.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We propose Open-world IE, a challenging task that aims to extract out-of-ontology entity profiles with instructions. Towards this grand mission, we create a large-scale instruction-following open-world IE dataset INSTRUCTOPENWIKI and develop PIVOINE by instruction tuning. We conduct extensive experiments on diverse instruction categories and different model scales, showing PIVOINE is a trustworthy LLM capable of following (possibly unseen) instructions in various granularities and extracting out-of-ontology entity profiles. Valuable future works include extending PIVOINE to a larger scale and exploring a more comprehensive range of instruction categories. detailed cases are shown in our Github repository.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"E Hyper-parameters\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We infer PIVOINE with the official BLOOM inference project provided by HuggingFace7 . We generate with beam search without sampling and the number of beam is 4. The maximum number of generated tokens are 2,048. The other generation parameters are set to default. Table 9: Statistics of the instruction-following open-world dataset INSTRUCTOPENWIKI. We report the statistics of the corpus, including the number of articles, mentions, and triplets in the left section. The ontology statistics, including the number of unique entities, aliases, relations, and types, are reported in the middle section. The right section shows the proportions of mentions with descriptions, aliases, and types, respectively. Ent., Rel., and Desc. are short for entities, relations, and descriptions.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"F Comprehensive Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Instruction\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"datasets are available at https://github.com/Lukeming\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Work done during Keming Lu's internship at Tencent AI Lab.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Appendix A Detailed Instructions\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Tab. 8 shows original manually-designed seed instruction, rephrased instructions, and number of rephrased instructions for each category. We use ChatGPT to rephrased our original instructions. The prompt we used is shown in Tab. 22.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B Dataset Statistics\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We display the statistics of INSTRUCTOPENWIKI in Tab. 9. INSTRUCTOPENWIKI contains a rich corpus, which includes all head paragraphs of all Wikipedia articles with 39 million mentions and 19 million triplets. And INSTRUCTOPENWIKI is also annotated with a large ontology containing over 2 million entities with 21 thousand entity types and 962 relation types, ensuring it covers a wide range of domains. Besides, the entity information density of INSTRUCTOPENWIKI is also abundant, so models can be trained for extracting entity profiles efficiently. Table 4: Proportions of unseen and seen instructions in \\\"Base Type\\\", \\\"Abstract Type\\\", and \\\"Description\\\" in the test set.\\nWe also create a small development set to select the best checkpoint balancing performance on seen and unseen cases. As shown in Tab. 9, the open-world evaluation test set contains rich unseen mentions (27.1%, 29,612/109,411) annotated by unseen entities (24.2%, 9,464/39,086).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C CHATGPT Configurations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We provide ChatGPT with the input context, instruction, and output JSON schema to prompt Chat-GPT to solve instruction-following open-world IE. We use ChatGPT in May 2022 and query it with the official API provided by OpenAI 5 . The detailed prompt we used is shown in Tab. 23.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"D Details of Human Evaluation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We use the May 1st, 2023 version of GPT-4. We query it via the chat platform of OpenAI 6 . The ----- Table 11: Main results of mention detection (MD). We report precision, recall, and F1 score of each model in this table. The subscript scores are the deviation under three different rephrased instructions for each instruction category. Subscript scores are the deviation under three different rephrased instructions for each sample. GENRE is an closed-world generative IE baseline without the instruction-following ability, so we only report GENRE scores with the default instruction. 13.00.3 19.40.5 15.60.4 29.90.2 35.00.4 32.30.3 46.00.1 39.90.2 42.70.1 59.30.2 57.60.1 58.40.1 Abstract Type 8.90.5 7.50.2 8.20.3 16.20.3 18.20.2 17.10.3 42.70.2 35.60.0 38.80.1 62.30.3 53.90.1 57.80.1 Description 2.30.1 4.40.1 3.10.1 6.40.1 6.50.1 6.40.1 44.90.1 43.80.4 44.40.1 64.10.1 63.20.1 63.60 \"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":""},"doi":{"kind":"string","value":"10.18653/v1/D19-1651"},"references":{"kind":"list like","value":[{"authors":"Sangnie Bhardwaj; Samarth Aggarwal; Mausam Mausam","journal":"Association for Computational Linguistics","ref_id":"b0","title":"CaRB: A crowdsourced benchmark for open IE","year":"2019"},{"authors":"Sangnie Bhardwaj; Samarth Aggarwal; Mausam Mausam","journal":"","ref_id":"b1","title":"Carb: A crowdsourced benchmark for open ie","year":"2019"},{"authors":"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell","journal":"Advances in neural information processing systems","ref_id":"b2","title":"Language models are few-shot learners","year":"2020"},{"authors":"Laura Chiticariu; Yunyao Li; Sriram Raghavan; Frederick R Reiss","journal":"","ref_id":"b3","title":"Enterprise information extraction: recent developments and open challenges","year":"2010"},{"authors":"Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann","journal":"","ref_id":"b4","title":"Palm: Scaling language modeling with pathways","year":"2022"},{"authors":"Ralph Grishman","journal":"IEEE Intelligent Systems","ref_id":"b5","title":"Information extraction","year":"2015"},{"authors":"Pere-Lluís Huguet; Cabot ; Roberto Navigli","journal":"Association for Computational Linguistics","ref_id":"b6","title":"REBEL: Relation extraction by end-to-end language generation","year":"2021"},{"authors":"Anastasiia Iurshina; Jiaxin Pan; Rafika Boutalbi; Steffen Staab","journal":"","ref_id":"b7","title":"Nilk: Entity linking dataset targeting nil-linking cases","year":"2022"},{"authors":"Nikolaos Kolitsas; Octavian-Eugen; Thomas Ganea; Hofmann","journal":"Association for Computational Linguistics","ref_id":"b8","title":"End-to-end neural entity linking","year":"2018"},{"authors":"Keshav Kolluru; Vaibhav Adlakha; Samarth Aggarwal; Mausam ; Soumen Chakrabarti","journal":"Association for Computational Linguistics","ref_id":"b9","title":"OpenIE6: Iterative Grid Labeling and Coordination Analysis for Open Information Extraction","year":"2020"},{"authors":"Nalin Kumar; Ondrej Bojar","journal":"Association for Computational Linguistics","ref_id":"b10","title":"Genre transfer in NMT:creating synthetic spoken parallel sentences using written parallel data","year":"2022"},{"authors":"Guillaume Lample; Miguel Ballesteros; Sandeep Subramanian; Kazuya Kawakami; Chris Dyer","journal":"Association for Computational Linguistics","ref_id":"b11","title":"Neural architectures for named entity recognition","year":"2016"},{"authors":"Lucile Hugo Laurençon; Thomas Saulnier; Christopher Wang; Albert Akiki; Teven Villanova Del Moral; Leandro Le Scao; Chenghao Von Werra; Eduardo González Mou; Huu Ponferrada; Jörg Nguyen; Mario Frohberg; Quentin Šaško; Angelina Lhoest; Gérard Mcmillan-Major; Stella Dupont; Anna Biderman; Loubna Rogers; Francesco De Ben Allal; Giada Toni; Olivier Pistilli; Somaieh Nguyen; Maraim Nikpoor; Pierre Masoud; Javier Colombo; Paulo De La Rosa; Tristan Villegas; Shayne Thrush; Sebastian Longpre; Leon Nagel; Manuel Weber; Jian Romero Muñoz; Daniel Zhu; Zaid Van Strien; Khalid Alyafeai; Almubarak; Minh Vu; Itziar Chien; Aitor Gonzalez-Dios; Kyle Soroa; Manan Lo; Pedro Ortiz Dey; Aaron Suarez; Shamik Gokaslan; David Bose; Long Ifeoluwa Adelani; Hieu Phan; Ian Tran; Suhas Yu; Jenny Pai; Violette Chim; Suzana Lepercq; Margaret Ilic; Sasha Mitchell; Yacine Luccioni; Jernite","journal":"","ref_id":"b12","title":"The bigscience ROOTS corpus: A 1.6TB composite multilingual dataset","year":"2022"},{"authors":"Thomas Lin; Oren Etzioni","journal":"","ref_id":"b13","title":"No noun phrase left behind: detecting and typing unlinkable entities","year":"2012"},{"authors":"Jie Lou; Yaojie Lu; Dai Dai; Wei Jia; Hongyu Lin; Xianpei Han; Le Sun; Hua Wu","journal":"","ref_id":"b14","title":"Universal information extraction as unified semantic matching","year":"2023"},{"authors":"Yaojie Lu; Qing Liu; Dai Dai; Xinyan Xiao; Hongyu Lin; Xianpei Han; Le Sun; Hua Wu","journal":"","ref_id":"b15","title":"Unified structure generation for universal information extraction","year":"2022"},{"authors":"Yubo Ma; Yixin Cao; Yongching Hong; Aixin Sun","journal":"","ref_id":"b16","title":"Large language model is not a good few-shot information extractor, but a good reranker for hard samples!","year":"2023"},{"authors":"Christina Niklaus; Matthias Cetto; André Freitas; Siegfried Handschuh","journal":"Association for Computational Linguistics","ref_id":"b17","title":"A survey on open information extraction","year":"2018"},{"authors":"Long Ouyang; Jeffrey Wu; Xu Jiang; Diogo Almeida; Carroll Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray","journal":"Advances in Neural Information Processing Systems","ref_id":"b18","title":"Training language models to follow instructions with human feedback","year":"2022"},{"authors":"Xiang Ren; Wenqi He; Meng Qu; Lifu Huang; Ji Heng; Jiawei Han","journal":"Association for Computational Linguistics","ref_id":"b19","title":"AFET: Automatic finegrained entity typing by hierarchical partial-label embedding","year":"2016"},{"authors":"Pedro Ruas; Francisco M Couto","journal":"Journal of Biomedical Informatics","ref_id":"b20","title":"Nilinker: Attention-based approach to nil entity linking","year":"2022"},{"authors":"Le Teven; Angela Scao; Christopher Fan; Ellie Akiki; Suzana Pavlick; Daniel Ilić; Roman Hesslow; Alexandra Castagné; François Sasha Luccioni; Matthias Yvon; Gallé","journal":"","ref_id":"b21","title":"Bloom: A 176bparameter open-access multilingual language model","year":"2022"},{"authors":"Phillip Schneider; Tim Schopf; Juraj Vladika; Mikhail Galkin; Elena Simperl; Florian Matthes","journal":"","ref_id":"b22","title":"A decade of knowledge graphs in natural language processing: a survey","year":"2022"},{"authors":"Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar","journal":"","ref_id":"b23","title":"Llama: Open and efficient foundation language models","year":"2023"},{"authors":"Somin Wadhwa; Silvio Amir; Byron C Wallace","journal":"","ref_id":"b24","title":"Revisiting relation extraction in the era of large language models","year":"2023"},{"authors":"Chenguang Wang; Xiao Liu; Zui Chen; Haoyun Hong; Jie Tang; Dawn Song","journal":"","ref_id":"b25","title":"Zero-shot information extraction as a unified text-to-triple translation","year":"2021"},{"authors":"Qifan Wang; Yi Fang; Anirudh Ravula; Fuli Feng; Xiaojun Quan; Dongfang Liu","journal":"","ref_id":"b26","title":"Webformer: The web-page transformer for structure information extraction","year":"2022"},{"authors":"Xiao Wang; Weikang Zhou; Can Zu; Han Xia; Tianze Chen; Yuansen Zhang; Rui Zheng; Junjie Ye; Qi Zhang; Tao Gui","journal":"","ref_id":"b27","title":"Instructuie: Multitask instruction tuning for unified information extraction","year":"2023"},{"authors":"Yanshan Wang; Liwei Wang; Majid Rastegar-Mojarad; Sungrim Moon; Feichen Shen; Naveed Afzal; Sijia Liu; Yuqun Zeng; Saeed Mehrabi; Sunghwan Sohn","journal":"Journal of biomedical informatics","ref_id":"b28","title":"Clinical information extraction applications: a literature review","year":"2018"},{"authors":"Jason Wei; Maarten Bosma; Vincent Zhao; Kelvin Guu; Adams Wei Yu; Brian Lester; Nan Du; Andrew M Dai; Quoc V Le","journal":"","ref_id":"b29","title":"Finetuned language models are zero-shot learners","year":"2022"},{"authors":"Xiang Wei; Xingyu Cui; Ning Cheng; Xiaobin Wang; Xin Zhang; Shen Huang; Pengjun Xie; Jinan Xu; Yufeng Chen; Meishan Zhang","journal":"","ref_id":"b30","title":"Zeroshot information extraction via chatting with chatgpt","year":"2023"},{"authors":"Hongbin Ye; Ningyu Zhang; Hui Chen; Huajun Chen","journal":"","ref_id":"b31","title":"Generative knowledge graph construction: A review","year":"2022"},{"authors":"Kun Wayne Xin Zhao; Junyi Zhou; Tianyi Li; Xiaolei Tang; Yupeng Wang; Yingqian Hou; Beichen Min; Junjie Zhang; Zican Zhang; Dong","journal":"","ref_id":"b32","title":"A survey of large language models","year":"2023"},{"authors":"Haoxi Zhong; Chaojun Xiao; Cunchao Tu; Tianyang Zhang; Zhiyuan Liu; Maosong Sun","journal":"","ref_id":"b33","title":"How does nlp benefit legal system: A summary of legal artificial intelligence","year":"2020"},{"authors":"Shaowen Zhou; Bowen Yu; Aixin Sun; Cheng Long; Jingyang Li; Jian Sun","journal":"","ref_id":"b34","title":"A survey on neural open information extraction: Current status and future directions","year":"2022"},{"authors":"","journal":"","ref_id":"b35","title":"Method Partition Instruction MD EL ET OpenRE EIG (Desc.) EIG (Aliases) F1 F1(T=1) F1(T=0.8)","year":""}],"string":"[\n {\n \"authors\": \"Sangnie Bhardwaj; Samarth Aggarwal; Mausam Mausam\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b0\",\n \"title\": \"CaRB: A crowdsourced benchmark for open IE\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Sangnie Bhardwaj; Samarth Aggarwal; Mausam Mausam\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Carb: A crowdsourced benchmark for open ie\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b2\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Laura Chiticariu; Yunyao Li; Sriram Raghavan; Frederick R Reiss\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Enterprise information extraction: recent developments and open challenges\",\n \"year\": \"2010\"\n },\n {\n \"authors\": \"Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"Palm: Scaling language modeling with pathways\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ralph Grishman\",\n \"journal\": \"IEEE Intelligent Systems\",\n \"ref_id\": \"b5\",\n \"title\": \"Information extraction\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Pere-Lluís Huguet; Cabot ; Roberto Navigli\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b6\",\n \"title\": \"REBEL: Relation extraction by end-to-end language generation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Anastasiia Iurshina; Jiaxin Pan; Rafika Boutalbi; Steffen Staab\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Nilk: Entity linking dataset targeting nil-linking cases\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Nikolaos Kolitsas; Octavian-Eugen; Thomas Ganea; Hofmann\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b8\",\n \"title\": \"End-to-end neural entity linking\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Keshav Kolluru; Vaibhav Adlakha; Samarth Aggarwal; Mausam ; Soumen Chakrabarti\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b9\",\n \"title\": \"OpenIE6: Iterative Grid Labeling and Coordination Analysis for Open Information Extraction\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Nalin Kumar; Ondrej Bojar\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b10\",\n \"title\": \"Genre transfer in NMT:creating synthetic spoken parallel sentences using written parallel data\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Guillaume Lample; Miguel Ballesteros; Sandeep Subramanian; Kazuya Kawakami; Chris Dyer\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b11\",\n \"title\": \"Neural architectures for named entity recognition\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Lucile Hugo Laurençon; Thomas Saulnier; Christopher Wang; Albert Akiki; Teven Villanova Del Moral; Leandro Le Scao; Chenghao Von Werra; Eduardo González Mou; Huu Ponferrada; Jörg Nguyen; Mario Frohberg; Quentin Šaško; Angelina Lhoest; Gérard Mcmillan-Major; Stella Dupont; Anna Biderman; Loubna Rogers; Francesco De Ben Allal; Giada Toni; Olivier Pistilli; Somaieh Nguyen; Maraim Nikpoor; Pierre Masoud; Javier Colombo; Paulo De La Rosa; Tristan Villegas; Shayne Thrush; Sebastian Longpre; Leon Nagel; Manuel Weber; Jian Romero Muñoz; Daniel Zhu; Zaid Van Strien; Khalid Alyafeai; Almubarak; Minh Vu; Itziar Chien; Aitor Gonzalez-Dios; Kyle Soroa; Manan Lo; Pedro Ortiz Dey; Aaron Suarez; Shamik Gokaslan; David Bose; Long Ifeoluwa Adelani; Hieu Phan; Ian Tran; Suhas Yu; Jenny Pai; Violette Chim; Suzana Lepercq; Margaret Ilic; Sasha Mitchell; Yacine Luccioni; Jernite\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"The bigscience ROOTS corpus: A 1.6TB composite multilingual dataset\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Thomas Lin; Oren Etzioni\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"No noun phrase left behind: detecting and typing unlinkable entities\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Jie Lou; Yaojie Lu; Dai Dai; Wei Jia; Hongyu Lin; Xianpei Han; Le Sun; Hua Wu\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Universal information extraction as unified semantic matching\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Yaojie Lu; Qing Liu; Dai Dai; Xinyan Xiao; Hongyu Lin; Xianpei Han; Le Sun; Hua Wu\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Unified structure generation for universal information extraction\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yubo Ma; Yixin Cao; Yongching Hong; Aixin Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Large language model is not a good few-shot information extractor, but a good reranker for hard samples!\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Christina Niklaus; Matthias Cetto; André Freitas; Siegfried Handschuh\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b17\",\n \"title\": \"A survey on open information extraction\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Long Ouyang; Jeffrey Wu; Xu Jiang; Diogo Almeida; Carroll Wainwright; Pamela Mishkin; Chong Zhang; Sandhini Agarwal; Katarina Slama; Alex Ray\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b18\",\n \"title\": \"Training language models to follow instructions with human feedback\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xiang Ren; Wenqi He; Meng Qu; Lifu Huang; Ji Heng; Jiawei Han\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b19\",\n \"title\": \"AFET: Automatic finegrained entity typing by hierarchical partial-label embedding\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Pedro Ruas; Francisco M Couto\",\n \"journal\": \"Journal of Biomedical Informatics\",\n \"ref_id\": \"b20\",\n \"title\": \"Nilinker: Attention-based approach to nil entity linking\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Le Teven; Angela Scao; Christopher Fan; Ellie Akiki; Suzana Pavlick; Daniel Ilić; Roman Hesslow; Alexandra Castagné; François Sasha Luccioni; Matthias Yvon; Gallé\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Bloom: A 176bparameter open-access multilingual language model\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Phillip Schneider; Tim Schopf; Juraj Vladika; Mikhail Galkin; Elena Simperl; Florian Matthes\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"A decade of knowledge graphs in natural language processing: a survey\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Llama: Open and efficient foundation language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Somin Wadhwa; Silvio Amir; Byron C Wallace\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Revisiting relation extraction in the era of large language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Chenguang Wang; Xiao Liu; Zui Chen; Haoyun Hong; Jie Tang; Dawn Song\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Zero-shot information extraction as a unified text-to-triple translation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Qifan Wang; Yi Fang; Anirudh Ravula; Fuli Feng; Xiaojun Quan; Dongfang Liu\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Webformer: The web-page transformer for structure information extraction\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xiao Wang; Weikang Zhou; Can Zu; Han Xia; Tianze Chen; Yuansen Zhang; Rui Zheng; Junjie Ye; Qi Zhang; Tao Gui\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Instructuie: Multitask instruction tuning for unified information extraction\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Yanshan Wang; Liwei Wang; Majid Rastegar-Mojarad; Sungrim Moon; Feichen Shen; Naveed Afzal; Sijia Liu; Yuqun Zeng; Saeed Mehrabi; Sunghwan Sohn\",\n \"journal\": \"Journal of biomedical informatics\",\n \"ref_id\": \"b28\",\n \"title\": \"Clinical information extraction applications: a literature review\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Jason Wei; Maarten Bosma; Vincent Zhao; Kelvin Guu; Adams Wei Yu; Brian Lester; Nan Du; Andrew M Dai; Quoc V Le\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Finetuned language models are zero-shot learners\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xiang Wei; Xingyu Cui; Ning Cheng; Xiaobin Wang; Xin Zhang; Shen Huang; Pengjun Xie; Jinan Xu; Yufeng Chen; Meishan Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Zeroshot information extraction via chatting with chatgpt\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Hongbin Ye; Ningyu Zhang; Hui Chen; Huajun Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Generative knowledge graph construction: A review\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Kun Wayne Xin Zhao; Junyi Zhou; Tianyi Li; Xiaolei Tang; Yupeng Wang; Yingqian Hou; Beichen Min; Junjie Zhang; Zican Zhang; Dong\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"A survey of large language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Haoxi Zhong; Chaojun Xiao; Cunchao Tu; Tianyang Zhang; Zhiyuan Liu; Maosong Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"How does nlp benefit legal system: A summary of legal artificial intelligence\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Shaowen Zhou; Bowen Yu; Aixin Sun; Cheng Long; Jingyang Li; Jian Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"A survey on neural open information extraction: Current status and future directions\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Method Partition Instruction MD EL ET OpenRE EIG (Desc.) EIG (Aliases) F1 F1(T=1) F1(T=0.8)\",\n \"year\": \"\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[],"string":"[]","isConversational":false},"title":{"kind":"string","value":"PIVOINE: Instruction Tuning for Open-world Information Extraction"},"abstract":{"kind":"string","value":"We consider the problem of Open-world Information Extraction (Open-world IE), which extracts comprehensive entity profiles from unstructured texts. Different from the conventional closed-world setting of Information Extraction (IE), Open-world IE considers a more general situation where entities and relations could be beyond a predefined ontology. More importantly, we seek to develop a large language model (LLM) that is able to perform Open-world IE to extract desirable entity profiles characterized by (possibly fine-grained) natural language instructions. We achieve this by finetuning LLMs using instruction tuning. In particular, we construct INSTRUCTOPEN-WIKI, a substantial instruction tuning dataset for Open-world IE enriched with a comprehensive corpus, extensive annotations, and diverse instructions. We finetune the pretrained BLOOM models on INSTRUCTOPENWIKI and obtain PIVOINE, an LLM for Open-world IE with strong instruction-following capabilities. Our experiments demonstrate that PIVOINE significantly outperforms traditional closed-world methods and other LLM baselines, displaying impressive generalization capabilities on both unseen instructions and out-of-ontology cases. Consequently, PIVOINE emerges as a promising solution to tackle the open-world challenge in IE effectively."},"authors":{"kind":"string","value":"Keming Lu; Xiaoman Pan; Kaiqiang Song; Hongming Zhang; Dong Yu; Jianshu Chen"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: Illustration of Open-world IE and its two main challenges: generalization to unseen instructions and out-of-ontology entities.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"LLM\"Party Time\" is a song written by Bruce Channel, and recorded by American country music artist T. G. Sheppard. It was released in June 1981 as the second single from the album \"I Love 'Em All\". The song was Sheppard's eighth number one on the country chart. The single stayed at number one for one week and spent a total of thirteen weeks on the country chart.","figure_data":"","figure_id":"fig_1","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Overview of the open-world information extraction method PIVOINE. This figure shows four generation cases of PIVOINE under three instruction categories colored in orange, blue, and green. PIVOINE takes the corpus and instructions as inputs and auto-regressively generates JSON sequences. And PIVOINE can extract different entity profiles based on different instructions from the same corpus. The auto-regressive generation of the JSON targets aligns well with various IE subtasks marked as green comments in the JSON.","figure_data":"","figure_id":"fig_2","figure_label":"2","figure_type":"figure"},{"figure_caption":"\" is a song written by Bruce Channel, and recorded by American country music artist T. G. Sheppard. It was released in June 1981 as the second single from the album \"I Love 'Em All\". The song was Sheppard's eighth number one on the country chart. The single stayed at number one for one week and spent a total of thirteen weeks on the country chart.","figure_data":"","figure_id":"fig_3","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Overview of instruction-following open-world IE dataset construction. The top of this figure shows INSTRUCTOPENWIKI is created by aligning anchor links in Wikipedia to entities and relations in the knowledge base Wikidata, and then augmented with instructions within eight categories. Entity profiles are identified from Wikidata and linearized into a sequence in the JSON format. The bottom of this figure shows how we create an open-world evaluation set with the time difference between Wiki dumps.","figure_data":"","figure_id":"fig_4","figure_label":"3","figure_type":"figure"},{"figure_caption":"described at the top of Fig. 3. Specifically, we link a relation triplet in the KB to this paragraph if both head and tail entities are mentioned. A relation triplet is represented by mentions of head and tail entities and a list of relation names. The detailed statistics of this dataset are presented in Appx. §B.","figure_data":"","figure_id":"fig_5","figure_label":"","figure_type":"figure"},{"figure_caption":"Main results of overall performance in end-to-end evaluation. We report the macro average of F1 scores with all instruction categories on mention detection (MD), entity linking (EL), entity typing (ET), open relation extraction (OpenRE), and entity information generation (EIG) for descriptions and aliases. Subscript scores are the deviation under three different rephrased instructions for each sample. Comprehensive results are demonstrated in Appx. §F. † We only report the performance of GENRE and OPENIE6 on default instructions since they are closed-world benchmarks focusing on specific subtasks without instruction following abilities.","figure_data":"MethodMD F1F1(T=1)EL F1(T=0.8)ET F1OpenRE F1(CaRB)EIG (Desc.) F1(ROUGE-L)EIG (Aliases) F1GENRE43.7 †17.2 †20.1 †----OPENIE6----15.2 †--CHATGPT50.50.225.00.226.10.27.90.022.50.239.00.215.80.0CHATGPT W/DEMO51.10.138.60.039.90.014.80.121.40.052.00.213.00.1PIVOINE-1b61.40.049.60.050.60.040.40.056.10.268.60.072.70.1PIVOINE-7b79.60.069.80.170.70.156.40.067.80.280.20.180.50.0PartitionMethodMD REL R(T=1) R(T=0.8)ET ROpenRE EIG (Desc.) EIG (Aliases) R(CaRB) R(ROUGE-L) RGENRE55.3 †23.8 †24.7 †----OPENIE6----21.9 †--BeforeCHATGPT59.30.3 31.80.232.30.29.70.134.10.240.20.410.10.105/30/2022CHATGPT W/DEMO 56.50.0 45.00.145.70.115.90.238.30.053.60.28.40.0PIVOINE-1b59.20.1 51.80.052.20.044.40.161.80.371.40.169.70.1PIVOINE-7b83.70.0 78.80.1 79.10.1 66.80.1 79.40.182.50.078.00.0GENRE33.6 †0 †3.6 †----OPENIE6----18.2 †--AfterCHATGPT58.90.2 23.80.226.50.36.90.127.10.536.00.314.30.305/30/2022CHATGPT W/DEMO 58.20.0 39.10.042.10.014.20.129.30.548.10.213.30.1PIVOINE-1b55.10.0 36.30.138.30.122.80.034.50.155.70.320.50.2PIVOINE-7b71.70.1 51.70.2 53.80.2 24.10.2 36.10.164.10.222.70.1","figure_id":"tab_2","figure_label":"1","figure_type":"table"},{"figure_caption":"Main results of generalization analysis in end-to-end evaluation. Headers are the same as Tab. 1 except we report recalls in each task. The partition \"Before 05/30/2022\" denotes mentions linked to the training ontology (seen entities), while \"After 05/30/2022\" denotes mentions linked to entities that are out of training ontology (unseen entities). Comprehensive results are demonstrated in Appx. §F. † We only report the performance of GENRE and OPENIE6 on default instructions since they are closed-world benchmarks focusing on specific subtasks without instruction following abilities.","figure_data":"from the same category by adding it to the conver-sation history. We also introduce two traditional IEbaselines to provide more comparisons of IE per-formance: (3) GENRE (Kumar and Bojar, 2022)is the first system retrieving entities by generatingcanonical names, which can address mention de-tection and entity linking but is constrained by theKILT entities. (4) OPENIE6 (Kolluru et al., 2020)is a recent state-of-the-art neural open informationextraction system. We employ it as a baseline ofopen relation extraction.Configurations. We start from BLOOM (Scaoet al., 2022) checkpoints with 1 billion and 7billion parameters and run instruction tuning onINSTRUCTOPENWIKI. We use suggested hyper-parameters for finetuning each model in Scao et al.","figure_id":"tab_3","figure_label":"2","figure_type":"table"},{"figure_caption":"Figure 4: Main results of the ontology generalization evaluation for each instruction category in each task. The partition \"Before 05/30/2022\" denotes mentions linked to the training ontology (seen entities), while \"After 05/30/2022\" denotes mentions linked to entities that are out of training ontology (unseen entities). * To avoid data leakage, descriptions as instructions are not considered in description generation tasks. Original scores are shown in Appx. §F.","figure_data":"Abstract Type Mention DetectionAbstract Type Entity LinkingAbstract Type Entity TypingAfter 05/30/2022 Before 05/30/2022Base TypeDescriptionBase TypeDescriptionBase TypeDescriptionDefaultImportanceDefaultImportanceDefaultImportanceImportance20406080DefaultImportance20406080DefaultImportance20406080DefaultDescriptionBase TypeDescriptionBase TypeDescriptionBase TypeAbstract Type Abstract Type Open Relation ExtractionAbstract Type Abstract Type Description GenerationAbstract Type Abstract Type Alias GenerationChatGPT ChatGPT w/Demo PIVOINE-1b PIVOINE-7bBase TypeDescriptionBase TypeDescription*Base TypeDescriptionDefaultImportanceDefaultImportanceDefaultImportanceImportance20406080DefaultImportance20406080DefaultImportance20406080DefaultDescriptionBase TypeDescription*Base TypeDescriptionBase TypeAbstract TypeAbstract TypeAbstract TypeMethodPartitionMD F1EL F1(T=1) F1(T=0.8)ET F1OpenRE EIG (Aliases) F1(CaRB) F1CHATGPTunseen seen60.30.6 24.71.1 67.40.9 25.10.326.61.2 28.80.83.00.4 1.00.724.90.4 23.11.412.11.4 22.93.2CHATGPT W/DEMOunseen seen59.00.2 40.00.5 67.91.8 44.32.641.90.6 48.22.82.30.3 7.10.022.20.6 23.30.722.61.8 74.23.8PIVOINE-1bunseen seen55.80.4 34.70.3 60.30.7 44.90.736.40.3 47.50.819.90.1 26.70.730.00.6 33.90.59.70.3 36.50.4PIVOINE-7bunseen seen83.60.2 55.10.1 57.80.1 28.20.1 35.00.3 86.50.2 71.70.2 72.80.2 33.80.7 39.40.012.80.3 31.80.3","figure_id":"tab_4","figure_label":"","figure_type":"table"},{"figure_caption":"Analysis of the instruction-following capabilities, including JSON decoding error rate, entity number instruction failure rate, and entity type instruction failure rate. Original scores are shown in Appx. §F.","figure_data":"Macro Avg.Number+Base TypeBase TypeNumber+Abstract TypeAbstract TypeFigure 5: all methods, especially the closed-world baselines.Even for the two ChatGPT-based baselines, perfor-mance on EL, ET, OpenRE, and EIG(Desc.) alsodrop dramatically, probably because unseen enti-ties are the latest and unpopular. PIVOINE-1b onlyoutperforms CHATGPT W/DEMO on four tasks inthe unseen partition. At the same time, it surpassesCHATGPT W/DEMO on all tasks in the seen par-tition, suggesting such generalization requires asufficient model scale. PIVOINE-7b still achievesthe best performance on tasks in both partitions,showing it can successfully generalize to out-of-ontology entities in Open-world IE.","figure_id":"tab_6","figure_label":"","figure_type":"table"},{"figure_caption":"Analysis of JSON format correctness on all instruction categories. We report the JSON decoding error rates (%) in this table. Subscript scores are the deviation under three different rephrased instructions for each sample.","figure_data":"ChatGPTChatGPT w/DemoPIVOINE-1b PIVOINE-7bDefault2.50.11.90.11.70.00.00.0Base Type3.40.33.60.42.30.20.00.0Abstract Type 3.80.42.50.82.30.20.00.0Description3.10.32.20.21.30.30.00.0Importance3.90.63.90.93.90.30.00.0Number2.70.53.70.12.30.10.00.0Number+ Base Type3.50.25.50.33.80.10.00.0Number+ Abstract Type3.60.42.81.03.30.40.00.1Macro Avg.3.30.23.30.42.10.00.00.0InstructionChatGPTChatGPT w/DemoPIVOINE-1b PIVOINE-7bNumber26.80.4 13.40.56.50.00.60.0Importance25.61.1 18.70.27.90.30.80.0Number+ Base Type43.40.4 37.30.038.00.929.00.2Number+ Abstract Type48.30.9 40.80.442.30.837.30.1Macro Avg.36.00.1 27.50.023.70.116.90.0","figure_id":"tab_7","figure_label":"5","figure_type":"table"},{"figure_caption":"Analysis of following entity number constraints. We report error rates (%) of predictions that do not have the same number of entities as the instruction. Subscript scores are the deviation under three different rephrased instructions for each sample.","figure_data":"","figure_id":"tab_8","figure_label":"6","figure_type":"table"},{"figure_caption":"Details of Instructions. {num}, {types}, {descriptions} are placeholders for entity numbers, types, and descriptions. We only display plural forms of templates in this table, but the templates will be different for singular or plural. We only show one example of rephrased templates for each category.","figure_data":"Split#ArticleCorpus #Mention#Triplets#Ent.Ontology #Aliases #Rel. #Types %Desc. %Aliases %Types Entity Info DensityTrain 11,447,454 39,930,663 19,184,948 2,234,052 840,401 962 21,350 93.564.271.5Dev2,71013,601 unseen:30385,9156,868 unseen:14178,812234 1,16394.655.670.2Test24,393109,411 unseen:29,61245,75839,086 unseen:9,47437,809 398 3,30692.752.670.7","figure_id":"tab_9","figure_label":"8","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: Illustration of Open-world IE and its two main challenges: generalization to unseen instructions and out-of-ontology entities.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"LLM\\\"Party Time\\\" is a song written by Bruce Channel, and recorded by American country music artist T. G. Sheppard. It was released in June 1981 as the second single from the album \\\"I Love 'Em All\\\". The song was Sheppard's eighth number one on the country chart. The single stayed at number one for one week and spent a total of thirteen weeks on the country chart.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Overview of the open-world information extraction method PIVOINE. This figure shows four generation cases of PIVOINE under three instruction categories colored in orange, blue, and green. PIVOINE takes the corpus and instructions as inputs and auto-regressively generates JSON sequences. And PIVOINE can extract different entity profiles based on different instructions from the same corpus. The auto-regressive generation of the JSON targets aligns well with various IE subtasks marked as green comments in the JSON.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"\\\" is a song written by Bruce Channel, and recorded by American country music artist T. G. Sheppard. It was released in June 1981 as the second single from the album \\\"I Love 'Em All\\\". The song was Sheppard's eighth number one on the country chart. The single stayed at number one for one week and spent a total of thirteen weeks on the country chart.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Overview of instruction-following open-world IE dataset construction. The top of this figure shows INSTRUCTOPENWIKI is created by aligning anchor links in Wikipedia to entities and relations in the knowledge base Wikidata, and then augmented with instructions within eight categories. Entity profiles are identified from Wikidata and linearized into a sequence in the JSON format. The bottom of this figure shows how we create an open-world evaluation set with the time difference between Wiki dumps.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"described at the top of Fig. 3. Specifically, we link a relation triplet in the KB to this paragraph if both head and tail entities are mentioned. A relation triplet is represented by mentions of head and tail entities and a list of relation names. The detailed statistics of this dataset are presented in Appx. §B.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Main results of overall performance in end-to-end evaluation. We report the macro average of F1 scores with all instruction categories on mention detection (MD), entity linking (EL), entity typing (ET), open relation extraction (OpenRE), and entity information generation (EIG) for descriptions and aliases. Subscript scores are the deviation under three different rephrased instructions for each sample. Comprehensive results are demonstrated in Appx. §F. † We only report the performance of GENRE and OPENIE6 on default instructions since they are closed-world benchmarks focusing on specific subtasks without instruction following abilities.\",\n \"figure_data\": \"MethodMD F1F1(T=1)EL F1(T=0.8)ET F1OpenRE F1(CaRB)EIG (Desc.) F1(ROUGE-L)EIG (Aliases) F1GENRE43.7 †17.2 †20.1 †----OPENIE6----15.2 †--CHATGPT50.50.225.00.226.10.27.90.022.50.239.00.215.80.0CHATGPT W/DEMO51.10.138.60.039.90.014.80.121.40.052.00.213.00.1PIVOINE-1b61.40.049.60.050.60.040.40.056.10.268.60.072.70.1PIVOINE-7b79.60.069.80.170.70.156.40.067.80.280.20.180.50.0PartitionMethodMD REL R(T=1) R(T=0.8)ET ROpenRE EIG (Desc.) EIG (Aliases) R(CaRB) R(ROUGE-L) RGENRE55.3 †23.8 †24.7 †----OPENIE6----21.9 †--BeforeCHATGPT59.30.3 31.80.232.30.29.70.134.10.240.20.410.10.105/30/2022CHATGPT W/DEMO 56.50.0 45.00.145.70.115.90.238.30.053.60.28.40.0PIVOINE-1b59.20.1 51.80.052.20.044.40.161.80.371.40.169.70.1PIVOINE-7b83.70.0 78.80.1 79.10.1 66.80.1 79.40.182.50.078.00.0GENRE33.6 †0 †3.6 †----OPENIE6----18.2 †--AfterCHATGPT58.90.2 23.80.226.50.36.90.127.10.536.00.314.30.305/30/2022CHATGPT W/DEMO 58.20.0 39.10.042.10.014.20.129.30.548.10.213.30.1PIVOINE-1b55.10.0 36.30.138.30.122.80.034.50.155.70.320.50.2PIVOINE-7b71.70.1 51.70.2 53.80.2 24.10.2 36.10.164.10.222.70.1\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Main results of generalization analysis in end-to-end evaluation. Headers are the same as Tab. 1 except we report recalls in each task. The partition \\\"Before 05/30/2022\\\" denotes mentions linked to the training ontology (seen entities), while \\\"After 05/30/2022\\\" denotes mentions linked to entities that are out of training ontology (unseen entities). Comprehensive results are demonstrated in Appx. §F. † We only report the performance of GENRE and OPENIE6 on default instructions since they are closed-world benchmarks focusing on specific subtasks without instruction following abilities.\",\n \"figure_data\": \"from the same category by adding it to the conver-sation history. We also introduce two traditional IEbaselines to provide more comparisons of IE per-formance: (3) GENRE (Kumar and Bojar, 2022)is the first system retrieving entities by generatingcanonical names, which can address mention de-tection and entity linking but is constrained by theKILT entities. (4) OPENIE6 (Kolluru et al., 2020)is a recent state-of-the-art neural open informationextraction system. We employ it as a baseline ofopen relation extraction.Configurations. We start from BLOOM (Scaoet al., 2022) checkpoints with 1 billion and 7billion parameters and run instruction tuning onINSTRUCTOPENWIKI. We use suggested hyper-parameters for finetuning each model in Scao et al.\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Figure 4: Main results of the ontology generalization evaluation for each instruction category in each task. The partition \\\"Before 05/30/2022\\\" denotes mentions linked to the training ontology (seen entities), while \\\"After 05/30/2022\\\" denotes mentions linked to entities that are out of training ontology (unseen entities). * To avoid data leakage, descriptions as instructions are not considered in description generation tasks. Original scores are shown in Appx. §F.\",\n \"figure_data\": \"Abstract Type Mention DetectionAbstract Type Entity LinkingAbstract Type Entity TypingAfter 05/30/2022 Before 05/30/2022Base TypeDescriptionBase TypeDescriptionBase TypeDescriptionDefaultImportanceDefaultImportanceDefaultImportanceImportance20406080DefaultImportance20406080DefaultImportance20406080DefaultDescriptionBase TypeDescriptionBase TypeDescriptionBase TypeAbstract Type Abstract Type Open Relation ExtractionAbstract Type Abstract Type Description GenerationAbstract Type Abstract Type Alias GenerationChatGPT ChatGPT w/Demo PIVOINE-1b PIVOINE-7bBase TypeDescriptionBase TypeDescription*Base TypeDescriptionDefaultImportanceDefaultImportanceDefaultImportanceImportance20406080DefaultImportance20406080DefaultImportance20406080DefaultDescriptionBase TypeDescription*Base TypeDescriptionBase TypeAbstract TypeAbstract TypeAbstract TypeMethodPartitionMD F1EL F1(T=1) F1(T=0.8)ET F1OpenRE EIG (Aliases) F1(CaRB) F1CHATGPTunseen seen60.30.6 24.71.1 67.40.9 25.10.326.61.2 28.80.83.00.4 1.00.724.90.4 23.11.412.11.4 22.93.2CHATGPT W/DEMOunseen seen59.00.2 40.00.5 67.91.8 44.32.641.90.6 48.22.82.30.3 7.10.022.20.6 23.30.722.61.8 74.23.8PIVOINE-1bunseen seen55.80.4 34.70.3 60.30.7 44.90.736.40.3 47.50.819.90.1 26.70.730.00.6 33.90.59.70.3 36.50.4PIVOINE-7bunseen seen83.60.2 55.10.1 57.80.1 28.20.1 35.00.3 86.50.2 71.70.2 72.80.2 33.80.7 39.40.012.80.3 31.80.3\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Analysis of the instruction-following capabilities, including JSON decoding error rate, entity number instruction failure rate, and entity type instruction failure rate. Original scores are shown in Appx. §F.\",\n \"figure_data\": \"Macro Avg.Number+Base TypeBase TypeNumber+Abstract TypeAbstract TypeFigure 5: all methods, especially the closed-world baselines.Even for the two ChatGPT-based baselines, perfor-mance on EL, ET, OpenRE, and EIG(Desc.) alsodrop dramatically, probably because unseen enti-ties are the latest and unpopular. PIVOINE-1b onlyoutperforms CHATGPT W/DEMO on four tasks inthe unseen partition. At the same time, it surpassesCHATGPT W/DEMO on all tasks in the seen par-tition, suggesting such generalization requires asufficient model scale. PIVOINE-7b still achievesthe best performance on tasks in both partitions,showing it can successfully generalize to out-of-ontology entities in Open-world IE.\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Analysis of JSON format correctness on all instruction categories. We report the JSON decoding error rates (%) in this table. Subscript scores are the deviation under three different rephrased instructions for each sample.\",\n \"figure_data\": \"ChatGPTChatGPT w/DemoPIVOINE-1b PIVOINE-7bDefault2.50.11.90.11.70.00.00.0Base Type3.40.33.60.42.30.20.00.0Abstract Type 3.80.42.50.82.30.20.00.0Description3.10.32.20.21.30.30.00.0Importance3.90.63.90.93.90.30.00.0Number2.70.53.70.12.30.10.00.0Number+ Base Type3.50.25.50.33.80.10.00.0Number+ Abstract Type3.60.42.81.03.30.40.00.1Macro Avg.3.30.23.30.42.10.00.00.0InstructionChatGPTChatGPT w/DemoPIVOINE-1b PIVOINE-7bNumber26.80.4 13.40.56.50.00.60.0Importance25.61.1 18.70.27.90.30.80.0Number+ Base Type43.40.4 37.30.038.00.929.00.2Number+ Abstract Type48.30.9 40.80.442.30.837.30.1Macro Avg.36.00.1 27.50.023.70.116.90.0\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Analysis of following entity number constraints. We report error rates (%) of predictions that do not have the same number of entities as the instruction. Subscript scores are the deviation under three different rephrased instructions for each sample.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Details of Instructions. {num}, {types}, {descriptions} are placeholders for entity numbers, types, and descriptions. We only display plural forms of templates in this table, but the templates will be different for singular or plural. We only show one example of rephrased templates for each category.\",\n \"figure_data\": \"Split#ArticleCorpus #Mention#Triplets#Ent.Ontology #Aliases #Rel. #Types %Desc. %Aliases %Types Entity Info DensityTrain 11,447,454 39,930,663 19,184,948 2,234,052 840,401 962 21,350 93.564.271.5Dev2,71013,601 unseen:30385,9156,868 unseen:14178,812234 1,16394.655.670.2Test24,393109,411 unseen:29,61245,75839,086 unseen:9,47437,809 398 3,30692.752.670.7\",\n \"figure_id\": \"tab_9\",\n \"figure_label\": \"8\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(Grishman, 2015)\", \"Explanation\": \"The cited work by Grishman (2015) provides a foundational understanding of information extraction (IE) and its various tasks, which the citing paper uses to structure its own research on IE.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Lample et al., 2016)\", \"Explanation\": \"The cited work by Lample et al. (2016) on Named Entity Recognition (NER) provides a key task in the traditional IE pipeline, which the citing paper uses to develop its own research in the field of IE.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Kolitsas et al., 2018)\", \"Explanation\": \"The cited work by Kolitsas et al. (2018) on Entity Linking (EL) contributes to the traditional IE pipeline by providing a method for associating mentions with a predefined ontology, which the citing paper uses in its research.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Ren et al., 2016)\", \"Explanation\": \"The cited work by Ren et al. (2016) on Entity Typing (ET) provides a task in the traditional IE pipeline that the citing paper uses to develop its own research in the field of IE.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Huguet Cabot and Navigli, 2021)\", \"Explanation\": \"The cited work by Huguet Cabot and Navigli (2021) on Relation Extraction (RE) contributes to the traditional IE pipeline by providing a method for identifying and understanding relationships between entities, which the citing paper uses in its research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Schneider et al., 2022)\", \"Explanation\": \"The cited work by Schneider et al. (2022) on knowledge graph construction extends the research on information extraction (IE) by exploring the use of IE in building knowledge graphs, which the citing paper may consider in its own research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Wang et al., 2022)\", \"Explanation\": \"The cited work by Wang et al. (2022) on search engine development extends the research on information extraction (IE) by exploring the use of IE in search engine optimization, which the citing paper may consider in its own research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Chiticariu et al., 2010)\", \"Explanation\": \"The cited work by Chiticariu et al. (2010) on document analysis extends the research on information extraction (IE) by exploring the use of IE in document analysis and understanding, which the citing paper may consider in its own research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Wang et al., 2018)\", \"Explanation\": \"The cited work by Wang et al. (2018) on document analysis extends the research on information extraction (IE) by exploring the use of IE in document analysis and understanding, which the citing paper may consider in its own research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhong et al., 2020)\", \"Explanation\": \"The cited work by Zhong et al. (2020) on document analysis extends the research on information extraction (IE) by exploring the use of IE in document analysis and understanding, which the citing paper may consider in its own research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Iurshina et al., 2022)\", \"Explanation\": \"The cited work by Iurshina et al. focuses on the subtask of EL, which is a component of the broader Open-world IE task. The citing paper extends this work by exploring the end-to-end process of creating entity profiles in an open-world setting.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ruas and Couto, 2022)\", \"Explanation\": \"The cited work by Ruas and Couto also addresses the subtask of EL in the context of Open-world IE. The citing paper further extends this work by discussing the challenges and limitations of this task in a more comprehensive manner.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Niklaus et al., 2018)\", \"Explanation\": \"The cited work by Niklaus et al. focuses on the subtask of OpenIE, which is another component of the Open-world IE task. The citing paper builds upon this work by discussing the open-world setting and the challenges of creating entity profiles in a more general context.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Bhardwaj et al., 2019a)\", \"Explanation\": \"The cited work by Bhardwaj et al. also addresses the subtask of OpenIE in the context of Open-world IE. The citing paper further extends this work by discussing the challenges and limitations of this task in a more comprehensive manner.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhao et al., 2023)\", \"Explanation\": \"The cited work by Zhao et al. provides a foundational understanding of the potential of large language models in addressing the open-world challenge in information extraction, which the citing paper builds upon to develop a new LLM for Open-world IE.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ma et al., 2023)\", \"Explanation\": \"The cited work by Ma et al. highlights the need for instruction tuning in endowing LLMs with Open-world IE capabilities, which the citing paper addresses by developing PIVOINE, a new LLM designed for Open-world IE.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wadhwa et al., 2023)\", \"Explanation\": \"The cited work by Wadhwa et al. further emphasizes the need for instruction tuning in LLMs for Open-world IE, which the citing paper addresses by developing PIVOINE as a new LLM for this task.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wei et al., 2022)\", \"Explanation\": \"The cited work by Wei et al. provides a key contribution in the form of instruction tuning for endowing LLMs with Open-world IE capabilities, which the citing paper builds upon in the development of PIVOINE as a new LLM for this task.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work, GPT-3, is a large language model that serves as a data source for the training of the LLM in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Chowdhery et al., 2022)\", \"Explanation\": \"The cited work, PALM, is a language model that builds upon the work of Brown et al. (2020) by scaling it up to a larger size of 540 billion parameters.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Scao et al., 2022)\", \"Explanation\": \"The cited work, BLOOM, is another language model that extends the work of Brown et al. (2020) by making it open-access and available in a range of parameter sizes from 560 million to 175 billion.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Touvron et al., 2023)\", \"Explanation\": \"The cited work, LLAMA, is a more efficient public-accessible LLM that serves as a data source for the development of open-world IE abilities in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(2022)\", \"Explanation\": \"The cited work by Ouyang et al. (2022) provides a method of instruction tuning and finetuning with human feedback that the citing paper adopts to improve the performance of LLMs in open-world IE tasks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"OPENAI\", \"Explanation\": \"The cited work by OPENAI serves as a baseline for the instruction-following dataset and instruction tuning method developed in the citing paper, extending the research on open-world IE by empowering LLMs with the ability to perform IE tasks in an open-world setting.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"Wei et al. (2023)\", \"Explanation\": \"The cited work by Wei et al. (2023) provides a method of solving IE as a multi-turn question-answering format with instructions, which serves as supporting evidence for the instruction-following IE approach adopted in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"Wang et al. (2023)\", \"Explanation\": \"The cited work by Wang et al. (2023) develops a unified IE method and proposes an instruction-tuning IE benchmark, which the citing paper builds upon to create an instruction-following dataset and instruction tuning method for open-world IE tasks.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Lin et al., 2012)\", \"Explanation\": \"The cited work by Lin et al. provides foundational data and research on the NIL problem in EL, which the citing paper uses to underpin its own study on instruction-following open-world IE.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhou et al., 2022)\", \"Explanation\": \"The cited work by Zhou et al. explores open information extraction, which the citing paper extends by applying the concept to instruction-following open-world IE.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Ye et al., 2022)\", \"Explanation\": \"The cited work by Ye et al. on open-world KGC provides a method for creating unseen entities and conducting link prediction in KGs, which the citing paper uses to support its own study on instruction-following open-world IE.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ye et al., 2022)\", \"Explanation\": \"The cited work by Ye et al. provides a method for linearizing structured information in generative IE, which the citing paper uses to improve the linearization process.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lou et al., 2023)\", \"Explanation\": \"The cited work by Lou et al. presents techniques for linearization in generative IE, which the citing paper adopts to further develop the linearization process.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang et al., 2021)\", \"Explanation\": \"The cited work by Wang et al. provides a method for linearization in generative IE, which the citing paper uses to improve the linearization process by avoiding the need for special tokens.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lu et al., 2022)\", \"Explanation\": \"The cited work by Lu et al. presents a method for linearization in generative IE, which the citing paper adopts to further develop the linearization process by using a text-based syntax with key-value pairs.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(2)\", \"Explanation\": \"The cited work provides a method for evaluating the quality of generated titles for mentions in a dataset, which the citing paper uses to assess the performance of their model in generating proper canonical titles.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(3)\", \"Explanation\": \"The cited work provides a method for generating entity types for entities, which the citing paper adopts in their research to generate entity types for the entities in their study.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(4)\", \"Explanation\": \"The cited work provides a method for evaluating the performance of triplet generation in open relation extraction, which the citing paper extends by using the same method to evaluate the performance of their model in generating triplets.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(5)\", \"Explanation\": \"The cited work provides a method for evaluating the quality of description generation in EIG-Desc., which the citing paper uses to assess the performance of their model in generating descriptions for generated entities.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(6)\", \"Explanation\": \"The cited work provides a method for generating aliases for predicted entities, which the citing paper adopts in their research to generate aliases for the predicted entities in their study.\"}]"}}},{"rowIdx":93,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b31","b25","b30","b21","b7","b28","b26","b34","b19","b33"],"table_ref":[],"text":"Large language models (LLMs) are becoming mainstream and easily accessible, ushering in an explosion of machine-generated content over various channels, such as news, social media, questionanswering forums, educational, and even academic contexts. Recently introduced LLMs, such as Chat-GPT, GPT-4, LLaMA 2 (Touvron et al., 2023), and Jais (Sengupta et al., 2023), generate remarkably fluent responses to a wide variety of user queries. The high quality of the generated texts makes them attractive for replacing human labor in many scenarios. However, this raises concerns regarding their potential misuse, e.g., to spread disinformation or to cause disruptions in the education system (Tang et al., 2023).\nSince humans perform only slightly better than chance when classifying machine-generated vs. human-written texts (Mitchell et al., 2023), we aim to facilitate the development of automatic detectors to mitigate the potential misuse of LLMs. In particular, we construct a diverse resource that could be used for training and testing various models for detecting machine-generated text (MGT).\nPrevious efforts in detecting MGT (i) focused on only one or two particular languages, typically only on English, (ii) used a single generator, e.g., just ChatGPT (Guo et al., 2023;Shijaku and Canhasi, 2023), (iii) leveraged fine-tuned LLMs for specific tasks, e.g., machine translation or text summarization (Shamardina et al., 2022), or (iv) considered only one specific domain e.g., news (Zellers et al., 2019;Macko et al., 2023). In contrast, here we encompass multiple languages, various LLMs, and several diverse domains, aiming to enable more general machine-generated text detection. Our dataset serves as the basis for SemEval-2024 Task 8 (Wang et al., 2024).\nOur contributions are as follows:\n• We construct M4: a large-scale multigenerator, multi-domain, and multi-lingual corpus for detecting machine-generated texts in a black-box scenario where there is no access to a potential generator or its outputs except for plain text.\n• We study the performance of automatic detectors from various perspectives: (a) different detectors across different domains for a specific LLM generator, (b) different detectors across different generators for a specific domain, (c) interactions of domains and generators in a multilingual setting, and (d) the performance of the detector on data generated from different time periods. From these experiments, we draw a number of observations, which can inform future research.\n• We release our data and code freely, and we plan to keep our repository constantly growing, adding new generators, domains, and languages over time. The remainder of the paper is organized as follows: Section 2 discusses related work. Section 3 describes the process of collecting the corpus from multiple generators (including davinci-text-003, ChatGPT, GPT4, Cohere, Dolly2, and BLOOMz), multiple domains (including Wikipedia, WikiHow, Reddit, QA, news, paper abstracts, and peer reviews), and multiple languages (Arabic, Bulgarian, Chinese, English, Indonesian, Russian, and Urdu) for machine-generated text detection. Section 4 presents the seven detectors we experiment with. Section 5 evaluates their performance across domains given a generator (ChatGPT or davinci) and across generators given a domain (arXiv or Wikipedia), as well as across different languages. Finally, Section 6 concludes and points to possible directions for future work."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b24","b29","b8","b13","b35","b15","b21","b32","b7","b28","b26","b34","b19","b7","b28","b5","b30","b7"],"table_ref":[],"text":"White-Box vs. Black-Box Detection We categorize the detection strategies into black-box and white-box, contingent on the level of access to the LLM that is suspected to have generated the target text. White-box methods focus on zero-shot detection without any additional training overhead (Sadasivan et al., 2023). Some use watermarking techniques (Szyller et al., 2021;He et al., 2022;Kirchenbauer et al., 2023;Zhao et al., 2023) and others rely on the expected per-token log probability of texts (Krishna et al., 2022;Mitchell et al., 2023). Black-box detectors only need API-level access to the LLM (i.e., when only the generated text is available) and typically extract and select features based on training text samples originating from both human and machine-generated sources.\nIn this study, we focus on black-box techniques because they aim to solve the task for the more practical and general use case. However, we note that their effectiveness heavily depends on the quality and the diversity of the training corpus.\nRelated Corpora Recently, a growing body of research has concentrated on amassing responses generated by LLMs. TuringBench (Uchendu et al., 2021) comprises 200K human-and machinegenerated pieces of text from 19 generative models. However, it is outdated, as the most advanced model used in this research is GPT-3. Guo et al. (2023) collected the HC3 dataset, which consists of nearly 40K questions and their corresponding answers from human experts and ChatGPT (English and Chinese), covering a wide range of domains (computer science, finance, medicine, law, psychology, and open-domain). Shijaku and Canhasi (2023) gathered TOEFL essays written by examined people and such generated by ChatGPT (126 essays for each).\nThe RuATD Shared Task 2022 involved artificial texts in Russian generated by various language models fine-tuned for specific domains or tasks such as machine translation, paraphrase generation, text summarization, and text simplification (Shamardina et al., 2022). We pay more attention to zero-shot generations of LLMs, such as the subset of RuATD generated by ruGPT-3.\nIn general, previous studies have concentrated on detecting machine-generated texts in one or two languages, for a specific LLM such as ChatGPT, or within a single domain such as news (Zellers et al., 2019;Macko et al., 2023). Our work broadens this scope to include multiple languages and a variety of widely-used LLMs across different domains.\nBlack-box Detectors are usually binary classifiers based on three types of features: statistical distributions (Guo et al., 2023;Shijaku and Canhasi, 2023), e.g., GLTR-like word rankings (Gehrmann et al., 2019), linguistic patterns (such as vocabulary, part-of-speech tags, dependency parsing, sentiment analysis, and stylistic features), and factverification features (Tang et al., 2023). Classification models involve deep neural networks, such as RoBERTa (Guo et al., 2023), or more traditional algorithms, such as logistic regression, support vector machines, Naïve Bayes, and decision trees.\nThere are also widely-used off-the-shelf MGT detectors, such as the OpenAI detector, 1 GPTZero, 2 and ZeroGPT. 3 According to the limited public information about them, these detectors are trained on collections of human-written texts and texts generated by various LLMs. For example, the training data of the OpenAI detector contains generations from 34 LLMs from various organizations, including OpenAI itself. For our M4 dataset, we selected a diverse set of state-of-the-art black-box methods and features, including one off-the-shelf detector."},{"figure_ref":[],"heading":"The M4 Dataset","publication_ref":["b14","b12","b26","b3","b22"],"table_ref":[],"text":"We gather human-written texts from a diverse range of sources across various domains and languages. For English we have Wikipedia (the March 2022 version), WikiHow (Koupaee and Wang, 2018), Reddit (ELI5), arXiv, and PeerRead (Kang et al., 2018), for Chinese we have Baike/Web QA question answering (QA), for Russian we have RuATD (Shamardina et al., 2022), for Arabic Wikipedia, and we use news for Urdu, Indonesian, and Bulgarian. Details about the data sources are provided in Appendix A.1 and A.2.\nFor machine generation, we prompt the following multilingual LLMs: GPT-4, ChatGPT, GPT-3.5 (text-davinci-003), Cohere, Dolly-v2 (Conover et al., 2023), and BLOOMz 176B (Muennighoff et al., 2022). The models are asked to write articles given a title (Wikipedia), abstracts given a paper title (arXiv), peer reviews based on the title and the abstract of a paper (PeerRead), news briefs based on a title (news), also to summarize Wikipedia articles (Arabic), and to answer questions (e.g., Reddit and Baike/Web QA).4 "},{"figure_ref":[],"heading":"Collection","publication_ref":[],"table_ref":["tab_9","tab_10"],"text":"Prompt Diversity For each generator, we carefully designed multiple (2-8) prompts in various styles, aiming to produce diverse outputs that are more aligned to divergent generations in real-world application scenarios. For example, on simple domains of Wikipedia and WikiHow, two prompts are applied. For arXiv and Reddit, as well as for ChatGPT, we use five prompts and four prompts for PeerRead. We generate varying tones of responses with prompts such as answer the question (1) \"like I am five years old\"; (2) \"in an expert confident voice\"; (3) \"in a formal academic and scientific writing voice\"; etc. Table 7 in Appendix A gives some statistics about the prompts used to generate the data collection, and Table 8 shows the hyper-parameters for the various generators.\nData Cleaning Simple artifacts in MGTs, such as multiple newlines and bullet points, could assist detectors, as their presence in the training data may discourage detectors from learning more generalized signals.\nTherefore, we performed minimal cleaning of the human-written and the machine-generated texts: (i) in a human-written WikiHow text, we removed multiple commas at the beginning of a new line (like \"\"\"\"\"\", we believe that ...\") and repeating newlines (\"\\n\\n\\n\\n\\n text begin \\n\\n\\n\\n\\n\"); (ii) in machine-generated WikiHow texts, we removed bullet points (as there were no bullet points in human-written texts); (iii) in human-written Wikipedia articles, we removed references (e.g., [1], [2]), URLs, multiple newlines, as well as paragraphs whose length was less than 50 characters; and (iv) in human-written arXiv abstracts, we removed newlines stemming from PDF conversion.\nQuality Control Unlike other tasks, where the data quality can be evaluated through the agreement between annotators over gold labels, we naturally obtain gold labels along with the collection of machine-generated texts. Therefore, we checked the data quality by randomly sampling 10-20 cases for each domain/generator and manually assessing the plausibility of generated texts. This can effectively circumvent incoherent, disorganized, and illogical generations that are easy to distinguish from human-written ones due to improper prompts or hyper-parameter settings of the generators (e.g., some generators repeat newly generated snippets to satisfy the minimum setup of new tokens). Moreover, in order to mimic human-written texts, we control the length of MGTs.\nIt should be highlighted that we did not pick examples. The quality control we exercised was model-level rather than example-level. We checked for cases where a model fundamentally failed, e.g., by generating visibly very bad output (e.g., very repetitive, English instead of foreign language output, etc.). This was very high-level checking (whether to keep a certain model in M4 or not); at the individual example level, we just checked whether the output had at least 1000 characters in length. Thus, we believe any biases that we might have introduced are minimal."},{"figure_ref":[],"heading":"Statistics","publication_ref":[],"table_ref":["tab_1"],"text":"The overall statistics about our M4 dataset for different tasks and languages are given in Table 1. We collected ∼ 147k human-machine parallel data in total, with 102k for English and 45k for other languages: 9k for Chinese, Russian, and Bulgarian; and 6k for Urdu, Indonesian, and Arabic respectively, in addition to over 10M non-parallel human-written texts. "},{"figure_ref":[],"heading":"Data Analysis","publication_ref":[],"table_ref":[],"text":"We performed analysis of our dataset in terms of vocabulary richness at the n-gram level, as well as in terms of human performance on the task of detecting machine-generated content."},{"figure_ref":[],"heading":"N-gram Analysis","publication_ref":[],"table_ref":["tab_12"],"text":"We compared the uni-gram and the bi-gram distributions of human-written vs. machine-generated texts and found that the former had a richer vocabulary than each of the six generators; see Table 9 in Appendix A.4 for detail. Dolly-v2 had the largest number of unique uni-and bi-grams, followed by davinci, ChatGPT, and BLOOMz, and Cohere had the least. The combination of all generators had comparable vocabulary to humans. When comparing across domains, we observed that Wikipedia, which covers a wide range of topics, contains the highest number of unique unigrams, followed by WikiHow and Reddit. In contrast, arXiv and PeerRead, which are specific to academic papers and peer reviews, exhibited fewer unique uni-grams and bi-grams. Within the same domain, we calculated the overlap of unique unigrams and bi-grams between human and machinegenerated texts. This overlap ranges in 20-35% for unigrams and in 10-20% for bi-grams. These variations can provide distinctive signals for black-box machine-generated text detection approaches."},{"figure_ref":[],"heading":"Human Evaluation","publication_ref":["b0"],"table_ref":["tab_4","tab_4"],"text":"From the Reddit and the arXiv (ChatGPT) test sets, for each domain, we sampled the first 50 (human, machine) pairs of texts and shuffled them into two groups, where two texts from the same pair would go in different groups. The annotators were then asked to focus on one group, which meant that they had to make a decision looking at each example individually, rather than having a pair of examples and deciding which one in the pair was humanwritten and which one was machine-generated (as some previous work did). This ensures a realistic scenario. For Reddit, we had 29 examples by humans and 21 by machines for group 1, and (21 human, 29 machine) for group 2; and (human:26, machine:24) for arXiv group 1, (human:24, machine:26) for arXiv group 2.\nWe had a total of six human annotators, who came from different countries and were native speakers of different languages. They were all proficient in English and all had NLP background: three PhD students, two MSc students, and 1 postdoc. Annotator 3 was an English native speaker who is also proficient in Arabic. Annotators 1 and 4 were Chinese native speakers, annotators 2 and 6 ware Russian native speakers, and annotator 5 was a Bulgarian native speaker.\nEach annotator made a guess about 17 unique examples for Reddit (finished by six annotators) and 25 examples for arXiv (finished by four). 5 The results are shown in Table 2. Interestingly, the English native speaker did not perform as well as some other annotators. We can further see in Table 2 that annotator 4 performed much better than annotator 1, even though they were both Chinese native speakers; this may be because annotator 4 had better understanding of how LLM generations work. Moreover, annotator 6 was the best rater, and he was also the one who was very familiar with LLM generation mechanisms, achieving higher guessing accuracy than annotator 2.\nThus, the annotators' proficiency in English may affect the evaluation, but for equal language proficiency, the degree of understanding of the LLM generation styles or patterns will also impact the quality of the annotator's guess.\nOn average, the accuracy of the human guesses was 0.77 for Reddit and 0.72 for arXiv. This indicates that it is not easy for humans to detect machine-generated text, especially for nonnative English speakers who are not familiar with the ChatGPT generation patterns (e.g., annotators 1,3,5). Besides, it is harder to classify the texts from arXiv than from Reddit. This is consistent with the findings in Clark et al. (2021). Without training, evaluators distinguished between GPT3-written and human-authored text at the chance level, and training by detailed instructions, annotated examples, and paired examples will improve the accuracy while the improvement across domains differs.\nWe hypothesize that our human annotators depended less on content signals and more on stylistic cues when identifying MGT for the arXiv domain, which results in the accuracy disparity between the two domains. Overall, it is challenging for general readers to understand and to follow abstracts of academic papers, but it is much easier to read Reddit answers.\nWe further compared the human performance to an XLM-R detector fine-tuned on in-domain training data. The classifier achieved near-perfect accuracy across the two domains, outperforming all human annotators. These findings strongly indicate the potential for automated in-domain black-box detection."},{"figure_ref":[],"heading":"Detectors","publication_ref":["b18","b1","b2"],"table_ref":["tab_14"],"text":"We evaluated seven detectors; see Table 11 for their hyper-parameter settings.\nRoBERTa This detector is based on the pretrained RoBERTa model (Liu et al., 2019), which we fine-tuned to detect machine-generated texts.\nELECTRA We further fine-tuned ELEC-TRA (Clark et al., 2020). Its pre-training objective is more aligned with our MGT task: it was pretrained to predict whether a token in a corrupted input was replaced by a plausible alternative sampled from a small generator network.\nXLM-R We fine-tuned XLM-RoBERTa, a multilingual variant of RoBERTa (Conneau et al., 2019)."},{"figure_ref":[],"heading":"Logistic Regression with GLTR Features","publication_ref":["b5"],"table_ref":[],"text":"We trained a logistic regression model based on 14 GLTR features from (Gehrmann et al., 2019), which are based on the observation that most LLM decoding strategies sample high-probability tokens from the head of the distribution. Thus, word ranking information about an LLM can be used to distinguish machine-generated texts from human-written ones. We selected two categories of these features: (i) the number of tokens in the top-10, top-100, top-1000, and 1000+ ranks from the LM predicted probability distributions (4 features), and (ii) the Frac(p) distribution over 10 bins ranging from 0.0 to 1.0 (10 features). Frac(p) describes the fraction of probability for the actual word divided by the maximum probability of any word at this position."},{"figure_ref":[],"heading":"Stylistic Features","publication_ref":["b16","b9","b6"],"table_ref":["tab_4","tab_5"],"text":"We trained an SVM classifier based on stylistic features from (Li et al., 2014): (i) character-based features, e.g., number of characters, letters, special characters, etc., (ii) syntactic features, e.g., number of punctuation and function words, (iii) structural features, e.g., total number of sentences, and (iv) word-based features, e.g., total number of words, average word length, average sentence length, etc. 12 and13)\nNEws LAndscape (NELA) We trained an SVM classifier using the NELA features (Horne et al., 2019), which cover six aspects: (i) style: the style and the structure of the article; (ii) complexity: how complex the writing is; (iii) bias: overall bias and subjectivity; (iv) affect: sentiment and emotional patterns; (v) moral: based on the Moral Foundation Theory (Graham et al., 2012); and (vi) event: time and location.\nGPTZero Finally, we used the GPTZero system without any adaptation. It was trained on a large diverse corpus of human-written and AI-generated texts, focussing on English. The system can analyze texts ranging from individual sentences to entire documents."},{"figure_ref":[],"heading":"Experiments and Results","publication_ref":[],"table_ref":[],"text":"In this section, we first describe our experiments, which come in three settings: (i) same generator, cross-domain evaluation, (ii) same domain, crossgenerator evaluation, and (iii) cross-lingual, crossgenerator evaluation. As mentioned in the previous section, we also experiment with GPTZero in a zero-shot setting, as it has not seen our data (even though it might have been trained on some domains involved in our data). We further discuss the evaluation results of these experiments."},{"figure_ref":["fig_0","fig_3"],"heading":"Same-Generator, Cross-Domain","publication_ref":["b23"],"table_ref":["tab_4","tab_5","tab_16","tab_7"],"text":"Given a specific text generator, such as ChatGPT and davinci-003, we train a detector using data from one domain and evaluate it on the test set from the same domain (in-domain evaluation) and other domains (out-of-domain evaluation). The results are shown in Figure 1 and Tables 12 and13.\nIn-domain detection is easy and can be done with very high accuracy, sometimes very close to a perfect score of 100%. This is especially the case for the RoBERTa detector, which reaches 100% accuracy for detecting ChatGPT-generated text on arXiv, 99.7% on Wikipedia, 99.7% on WikiHow, and 98.0% on PeerRead. The only dataset where the best score for the RoBERTa detector is achieved when training on a different domain is Reddit. We can further see that the results with davinci-003 show the same pattern: all in-domain evaluation scores are usually very high, approaching 100%. Other detectors also show high performance in the in-domain evaluation setting, but they usually overfit less to a particular domain. For example, the LR-GLTR detector shows only 79.6% accuracy on WikiHow when the davinci-003 generator was used, while the score for the RoBERTa-based detector exceeds 99%.\nThe best performance in the out-of-domain evaluation is often achieved by fine-tuning ELEC-TRA for the task. We attribute this to the specific pre-training objective of this model, which is based on the detection of replaced tokens. ELECTRA shows slightly lower performance than RoBERTa for the in-domain evaluation, but achieves huge improvements in the out-domain evaluation setting. For example, in the case of training on Wikipedia to detect davinci-003 on Reddit, the RoBERTa's performance is close to random guessing, while ELEC-TRA achieves 87.9% accuracy. Another strong approach for out-of-domain detection is LR-GLTR, which outperforms ELECTRA in some scenarios, such as detecting ChatGPT on the Wikipedia domain. 14 and15)\nOut-of-domain detection might be hard. This is especially noticeable when training on arXiv and detecting artificial texts for Reddit or training on arXiv and detecting for Wikipedia. This is expected as these pairs of domains are very different. There are some domains that offer better generalization than others. The RoBERTa-based detector and the detector based on NELA features are the most vulnerable in this regard. RoBERTa overfits to the training domain, while the NELA features are not tailored to machine-generated text detection, but rather initiated for fake news detection.\nThe best training domain for out-of-domain generalization is Reddit. Training on Reddit ELI5 usually yields the best out-of-domain performance. Wikipedia is also often a good domain for training. Training on arXiv and PeerRead yields the worst generalization across other domains because the writing style of academic papers is very specific.\nThe most challenging domain for machinegenerated text detection is WikiHow, while Peer-Read is the easiest one.\nThe GPT-3.5 (davinci-003) generator is harder to detect than ChatGPT. Aggregating the results across all domains and both generators, we can see that the accuracy for ChatGPT is usually higher than that for davinci-003. This indicates that ChatGPT may leave more distinctive signals in generated texts than davinci-003.\nFeature Analysis. We conducted feature analysis of in-domain detectors using LIME (Ribeiro et al., 2016), and we found that detectors did not overfit to MGT artifacts and leveraged word distribution for classification. See Figure 4 in Appendix G for more detail."},{"figure_ref":["fig_1"],"heading":"Same-Domain, Cross-Generator","publication_ref":[],"table_ref":["tab_16"],"text":"Given a specific domain, we train the detector using the training data from one generator and we evaluate it on the test data from the same and also from other generators. The accuracy on arXiv and Wikipedia is shown in Figure 2 (see Table 14 and 15 in Section D for precision, recall, and F1).\nRoBERTa performs the best among five detectors. It is the best on both arXiv (95.9%: average accuracy) and Wikipedia (99.4%), followed by LR-GLTR (84.0/80.7%), stylistic features (80.4/82.8%), and ELECTRA (72.5/76.6%); NELA features are the worst (73.7/64.3%). We can see that apart from the main diagonal, most scores for the detector using NELA features are around or lower than 50.0%, particularly on arXiv. This indicates that they are not suitable for distinguishing machine-generated and human-written texts. Moreover, the accuracy for Wikipedia is higher than for arXiv, especially for RoBERTa pre-trained using Wikipedia data. This suggests that arXiv is somewhat harder to detect than Wikipedia, and exposure bias on pre-training can impact a detectors' domain-specific performance.\nThe highest accuracy is for the same generator. Akin to the trend of cross-domain evaluation, training and testing using the same generator always yields the best accuracy for both arXiv and Wikipedia across the five detectors. Even for NELA, and detection over generations by BLOOMz, the accuracy mostly remains over 90.0. Performance drops substantially when the training and the test data are generated from different LLMs because of different distributions between the outputs of different generators. BLOOMz-generated text is much different from ChatGPT, davinci, and Cohere. For all detectors in both arXiv and Wikipedia, BLOOMz shows the lowest cross-generator accuracy. Specifically, when training on BLOOMz and testing on other generators, or when training on other generators and testing on BLOOMz, it shows low recall (<0.5) for machine-generated texts. This means that there are many false negative examples, namely, many machine-generated texts are misclassified as human-written ones. Most accuracy scores are ≤50.0%, i.e., similar or even worse than a random guess. This indicates that the distribution of BLOOMz outputs is very different from the other three generators. We assume that this is because BLOOMz is primarily fine-tuned for NLP downstream data.\nMoreover, we found that, for all detectors, when training on Cohere, the accuracy for ChatGPT is comparable to the accuracy on Cohere itself, and similarly high accuracy occurs when training on ChatGPT and testing on Cohere. This suggests that ChatGPT and Cohere share some generative patterns."},{"figure_ref":[],"heading":"Zero-shot Evaluation: GPTZero","publication_ref":[],"table_ref":["tab_5"],"text":"Table 3 shows that, from the perspective of the domain, GPTZero performs the best on Wikipedia, while the worst results are on arXiv where, for all generators, the F1 score is below 50%. From the perspective of generators, GPTZero shows the best performance on ChatGPT and the worst performance on BLOOMz. The recall for BLOOMz is close to 0% across all domains, which is consistent with the results for other detectors. GPTZero also demosntrated low performance for Dolly v2. GPTZero may have been trained on generations of ChatGPT and on data from domains such as Wikipedia and Reddit, thus showing remarkable scores for them. At the same time, zero-shot detection for unseen domains and generators poses a major challenge for GPTZero."},{"figure_ref":[],"heading":"Multilingual Evaluation","publication_ref":[],"table_ref":["tab_6","tab_6"],"text":"In this section, we discuss the results for our multilingual experiments with the XLM-R detector across seven languages. For multilingual evaluation, we used ChatGPT and davinci-003 as generators. The results are shown in Table 4 (see Section E in the Appendix for more detail).\nWe constructed the English training, development, and test sets by combining English texts across all domains: Wikipedia, WikiHow, Reddit ELI5, arXiv, and PeerRead. Then, the All row refers to the combination of all training data in Arabic, Bulgarian, Chinese, English, Indonesian, Russian, and Urdu from the same generator. We aim to evaluate the performance of a detector over each monolingual test set from a single domain when fully leveraging the available training data, thus observing the benefits brought by the interaction of multiple languages and domains.\nWe can see in Table 4 that the best accuracy is achieved when training and testing on the same language and using the same generator, while when training on one generator and testing on another one, the highest scores tend to appear in the row of All, i.e., when using the training data for all languages, except for Bulgarian (training on Bulgarian is best, if we want to test on Bulgarian).\nWe can also see that it is difficult for XLM-R to detect machine-generated text in a language that it has never seen during training. For example, it struggles to detect Russian, Urdu, and Indonesian machine/human-generated text when it was not trained on them. Interestingly, XLM-R still demonstrates good performance for Arabic even when trained on English data only."},{"figure_ref":[],"heading":"Time Domain Evaluation","publication_ref":[],"table_ref":["tab_7"],"text":"LLMs are constantly improving over time. This raises the question of the robustness of detectors for the same generator across different time points. With this in mind, we compared ChatGPT output generated in March 2023 (from our M4 dataset) vs. September 2023 on the Reddit-ELI5 domain and using XLM-R as a detector, and the same prompts and questions as for the M4 dataset. The results are shown in Table 5, where we can see that the detector trained on the earlier version can effectively classify generations produced by the September 2023 version. This implies that a detector may remain effective even when applied to a newer generator trained using fresh data. Table 6: Impact of text length on detection accuracy on arXiv using XLM-R."},{"figure_ref":["fig_2"],"heading":"Impact of Text Length","publication_ref":[],"table_ref":[],"text":"Finally, we investigated the impact of text length on detection accuracy. We truncated arXiv articles at the first 1,000, 500, 250, and 125 characters and compared the accuracy of XLM-R detectors trained and tested on such truncated articles for machine-generated content produced by ChatGPT.\nThe results are shown in Table 6. We can see that as the length decreases from 1,000 to 125, the accuracy drops by 4.5 points. This illustrates the negative impact of smaller text length on detection performance; more experiments on the arXiv and the Reddit datasets are presented in Figure 3 in the Appendix."},{"figure_ref":[],"heading":"Conclusion and Future Work","publication_ref":["b33"],"table_ref":["tab_1"],"text":"We presented M4, a large-scale multi-generator, multi-domain, and multi-lingual dataset for machine-generated text detection. We further experimented with this dataset performing a number of cross-domain, cross-generator, cross-lingual, and zero-shot experiments using seven detectors. We found that detectors struggle to differentiate between machine-generated and human-written texts if the texts come from a domain, a generator, or a language that the model has not seen during training. Our results show that the problem is far from solved and that there is a lot of room for improvement. We hope that our release of M4, which we make freely available to the community, will enable future research towards more robust approaches to the pressing societal problem of fighting malicious machine-generated text. We have already created an extension of M4 for SemEval-2024 Task 8 (Wang et al., 2024), 6 which features additional languages, domains, and three new task (re)formulations.\nIn future work, we plan to expand our M4 dataset continuously by introducing new LLM generators, by exploring different domains, by incorporating new languages, and by diversifying the range of tasks and prompts used. We believe that this is a good, practical way to keep the dataset up-to-date in response to the ongoing progress in LLMs. Our aim is to maintain a dataset that remains relevant as LLMs continue to evolve.\nData Collection, Licenses, and User Privacy. Creating the M4 dataset did not involve scraping raw data from websites. Instead, we used preexisting corpora that have been publicly released and approved for research purposes, with clear dataset licenses, which are listed in Table 1. To the best of our knowledge, all included datasets adhere to ethical guidelines and minimize privacy concerns. Since the human-written data has already been published and made publicly available for research purposes, we see no additional privacy risks in releasing that data as part of our M4 dataset.\nThe human text components of M4 are publicly available and can be freely accessed and used for research purposes. However, researchers must acknowledge the original sources of the text and comply with the respective licensing terms.\nThe machine-generated text components of our M4 dataset are subject to the licensing terms of the underlying LLMs. For text generated using LLMs, researchers must comply with the respective licensing terms of those LLMs:\n• davinci-003, ChatGPT, GPT-4: no specific license. They welcome research publications related to the OpenAI API. 7\n• Dolly-v2: Apache 2.0 8\n• Cohere: no specific license. They point out that CUSTOMER RETAINS ALL OWNER-SHIP AND INTELLECTUAL PROPERTY RIGHTS IN AND TO CUSTOMER DATA. 9\n• BLOOMz: Apache 2.0 10\nPotential Biases We recognize the potential for biases in our M4 dataset, stemming from both the original human-written corpora and the Large Language Models (LLMs) used for generation. This is an important issue, and we put efforts to minimize such biases. However, we are aware that unethical usage of our dataset may still lead to biased applications: even if our original dataset was completely unbiased, external parties may extract a biased subset, which would be out of our control.\n7 https://openai.com/policies/ sharing-publication-policy 8 https://github.com/databrickslabs/dolly 9 https://cohere.com/saas-agreement 10 https://github.com/bigscience-workshop/xmtf Having already realized these concerns, we have implemented the following measures: a. We provide comprehensive documentation about our M4 dataset, including detailed information about the sources of all human-written corpora, the generation process for obtaining the machine-generated text, including the full prompts and the measures we took to cleanse the output, and the potential biases that may exist. We believe that this transparency would allow researchers to understand the origins of the data and to make informed decisions about how to use it.\nb. We further acknowledge and transparently discuss these limitations and debiasing techniques that could be used to address these limitations. We hope that the strong emphasis on transparency in our methodology by explicitly stating the sources of human-written corpora and the generation processes for the corresponding machine-generated text could help clarify the dataset's origins and potential biases.\nRobustly Secure System The M4 dataset is intended for the development of detection systems to mitigate misuse, particularly in the context of malicious content generated using LLMs. While we encourage extensive and responsible use of the datasets to advance this critical area of research, we also emphasize the importance of adhering to the licensing terms of the original human-written corpora and the corresponding LLMs."},{"figure_ref":[],"heading":"Limitations","publication_ref":[],"table_ref":[],"text":"In this section, we discuss some perceived limitations of our study."},{"figure_ref":[],"heading":"M4 Dataset Generalization and Biases","publication_ref":[],"table_ref":[],"text":"Generalization: Machine-generated outputs exhibit a high degree of sensitivity to the prompts. While our M4 dataset was collected with diverse prompts for a variety of generators, domains, and languages, to cover typical use cases, it has limitations as a general resource, as it is neither sufficient to train a detector that can be expected to generalize well across all possible domains and generators, nor is it likely sufficient to act as a standard benchmark that can accurately evaluate the effectiveness of a detection method.\nUp-to-Date: Detecting machine-generated text is a very challenging task when we do not know in advance the potential generator and the domain: as our findings show, human-written and machinegenerated text cannot be distinguished in certain situations, e.g., we saw issues when using text generated by BLOOMz. Therefore, we regard M4 as a useful repository of machine-generated text for researchers who want to improve and to evaluate their detectors from multiple dimensions. Moreover, the LLMs are constantly evolving, and thus any dataset collected for machine-generated text detection can become outdated relatively fast. With this in mind, we have constantly been extending the M4 dataset (e.g., with a recent collection of GPT-4 responses), and we expect to grow our repository to enable better training and more up-to-date detectors.\nBias: Biases may exist in both human-written and machine-generated texts, and it is possible that our M4 dataset may be influenced by biases from human collection, thus affecting the detection outcomes. We leave the analysis of such biases to our future work."},{"figure_ref":[],"heading":"Feasibility of Black-Box Machine-Generated Text Detection","publication_ref":["b30","b24","b17"],"table_ref":[],"text":"A growing body of work shows that machinegenerated text detection might gradually become harder and even nearly impossible: as LLMs evolve, the gap between machine-generated and human-written text might narrow (Tang et al., 2023;Sadasivan et al., 2023). Liang et al. (2023) "},{"figure_ref":[],"heading":"Appendix A Data Collection and Analysis","publication_ref":["b4","b14","b12","b26","b26","b27","b20","b11","b10"],"table_ref":[],"text":"A.1 English Corpora Wikipedia We used the Wikipedia dataset available on HuggingFace 11 and randomly chose 3,000 articles, each of which surpasses a character length limit of 1,000. We prompted LLMs to generate Wikipedia articles given titles, with the requirement that the output articles contain at least 250 words.\nFor generation with Dolly-v2, 12 we set the minimum number of generated tokens to be 300 to satisfy the minimal character length of 1,000.\nReddit ELI5 dataset (Fan et al., 2019) is a collection of English question-answering (QA) pairs, 13 gathered to facilitate open-domain and long-form abstractive QA. The data is derived from three categories: ExplainLikeImFive for general topics, AskScience for scientific queries, and AskHistorians for historical inquiries. Each pair is composed of a question (a title + a detailed description) and corresponding answers.\nWe filtered out answers with less than 1,000 characters, retaining questions whose title ends with a question mark without detailed descriptions. Finally, we selected 1,000 QA pairs with top user ratings for each category, resulting in a total of 3,000 pairs.\nWikiHow dataset 14 (Koupaee and Wang, 2018) is built from the online WikiHow knowledge base. It consists of articles with a title, a headline (the concatenation of all bold lines of all paragraphs), and text (the concatenation of all paragraphs except the bold lines). We randomly chose 3,000 articles with the length of more than 1,000 characters and prompted LLMs with titles and headlines to generate artificial articles.\nPeerRead Reviews We sampled 586 academic papers published in top-tier NLP and machine learning conferences from the PeerRead corpus (Kang et al., 2018). Each paper contains metadata, including title, abstract, and multiple human-written reviews. Given a paper, we prompted LLMs to generate peer reviews with four different instructions; two depend only on the title and another two involve both the title and the abstract. Two prompts specify the review format of first describing what problem or question the considered paper addresses, and then providing its strengths and weaknesses. Other two prompts do not contain a review format specification. 15 This resulted in 584 × 4 = 2,344 machine-generated texts for each generator and 5,798 human-written reviews in total.\nArxiv Abstract parallel dataset is constructed from a Kaggle corpus. 16 We sampled 3,000 abstracts with a minimum length of 1,000 characters and prompted LLMs to produce machine-generated abstracts based on their titles.\nChinese QA is constructed from 3,000 question-answer pairs sampled from Baike and the Web QA corpus. 18 The length of each answer is more than 100 Chinese characters. We prompted LLMs with a combination of a brief title and a detailed description for each question.\nIndonesian News 2018 is constructed from a corpus of Indonesian news articles19 collected from seven different news websites in 2018. We picked news from CNN Indonesia since this source was found to provide the cleanest data. We selected 3,000 texts from the corpus and generated artificial news articles by prompting ChatGPT with a title.\nRussian RuATD is sourced from the RuATD Shared Task 2022 (Shamardina et al., 2022) devoted to artificial text detection in Russian. Shamardina et al. (2022) gathered a vast human and machine-generated corpora from various text generators. However, these generators are either task-specific or domain-specific.\nWe leveraged their human-written texts collected from publicly available resources and re-generated the machine-authored data using the open-domain state-of-the-art multilingual LLMs. In particular, for the construction of human-written data, the task organizers used the following sources: (1) diachronic sub-corpora of the Russian National Corpus20 , which covers three historical periods of the society and the Modern Russian language (\"pre-Soviet\", \"Soviet\", and \"post-Soviet\"); (2) several social media platforms;\n(3) top-100 most viewed Russian Wikipedia pages spanning the period of 2016-2021 according to the PageViews statistics; (4) news articles from the Taiga corpus (Shavrina and Shapovalova, 2017) and the \"corus\" library 21 ; (5) a corpus of digitalized personal diaries \"Prozhito\" written during the 20th century (Melnichenko and Tyshkevich, 2017); (6) government documents from the RuREBus Shared Task (Ivanin et al., 2020).\nUrdu News is derived from Urdu News Data 1M -a collection of one million news articles from four distinct categories: Business & Economics, Science & Technology, Entertainment, and Sports. These articles were gathered from four reputable news agencies in Pakistan (Hussain et al., 2021). Each entry in this dataset includes a headline, a category, and a news article text. To ensure the data balance over four categories, we randomly sampled 750 news articles from each, resulting in 3,000 examples in total. Using the headlines as prompts, we generated the content of artificial news articles."},{"figure_ref":[],"heading":"A.3 LLM Generation","publication_ref":[],"table_ref":["tab_9"],"text":"Prompt Diversity We paid attention to prompt diversity, using multiple (2-8) prompts for each domain-generator combination in English, with the aim to produce diverse outputs that are more aligned with divergent generations in real-world application scenarios. See Table 7 for detailed statistics about the prompts."},{"figure_ref":[],"heading":"Prompts of PeerRead","publication_ref":[],"table_ref":["tab_10"],"text":"• Please write a peer review for the paper + title;\n• Write a peer review by first describing what problem or question this paper addresses, then strengths and weaknesses, for the paper + title;\n• Please write a peer review for the paper + title, its main content is as below: + abstract;\n• Write a peer review by first describing what problem or question this paper addresses, then strengths and weaknesses, for the paper + title, its main content is as follows: + abstract.\nHyper-Parameter Values for the Generators Table 8 shows the values of the hyper-parameters we used for the various generators. In general, we followed the default setting, except for the length of new generations in order to satisfy the minimum character length of 1,000. "},{"figure_ref":[],"heading":"A.4 N-gram Analysis","publication_ref":[],"table_ref":["tab_12","tab_13"],"text":"Table 9 shows statistics about the number of unique uni-grams (word types) and bi-grams of human-written and machine-generated texts (English). Table 10 shows the number of per-document unique uni-grams (word types) and bi-grams of human-written and machine generated texts (English). "},{"figure_ref":[],"heading":"B.2 Computation Resources and Cost","publication_ref":[],"table_ref":[],"text":"We spent $600 on calls of OpenAI APIs for ChatGPT and davinci-003 generations, $40 on calls to GPTZero. We spent about 2,500 GPU hours on Dolly-v2 and BLOOMz generation.\nC Results: Same-Generator, Cross-Domain "},{"figure_ref":[],"heading":"E Results: Multilingual Evaluation","publication_ref":[],"table_ref":["tab_18","tab_19"],"text":"Table 16 shows detailed results for cross-language experiments based on XLM-R on the test sets across different languages generated by ChatGPT. Table 17 shows the same results, but using davinci-003 as a generator (instead of ChatGPT). "},{"figure_ref":["fig_2"],"heading":"F Results: Impact of Text Length","publication_ref":[],"table_ref":[],"text":"Figure 3 shows detailed results from experiments to study the impact of text length on detection accuracy, over arXiv and Reddit generated by ChatGPT, davinci, and Cohere. We can see that as the character length decreasing from 1000 to 125, the F1-score with respect to machine-generated text decreases for all subsets, demonstrating the negative impact of short text on detection performance. Henry was an experienced jouster, but Montgomery was a newcomer to the sport, and unfortunately, he was no match for the King's lance. During one of their runs, Henry's lance struck Montgomery's helmet, shattering it and sending a jagged piece of wood into his eye and brain. The young man was rushed to a nearby hospital, but it was too late. He died just a few days later, leaving behind a wife and children. As for Henry, he was devastated by the accident and reportedly went into a deep depression. He blamed himself for Montgomery's death and was haunted by guilt for the rest of his life. He even imposed a penance on himself, vowing to fast and do charitable works for the rest of his days. The incident also led to changes in the sport of jousting. After Montgomery's death, many rules were put in place to make the sport safer, such as requiring jousters to wear full face and head protection. Overall, it was a tragic event that had a lasting impact on the sport and on the King who inadvertently caused the death of his opponent."},{"figure_ref":[],"heading":"H Examples of M4","publication_ref":[],"table_ref":[],"text":"Language English LLM Generator CHATGPT Domain Arxiv"},{"figure_ref":[],"heading":"Title Polymer Quantum Mechanics and its Continuum Limit","publication_ref":[],"table_ref":[],"text":"Human A rather non-standard quantum representation of the canonical commutation relations of quantum mechanics systems, known as the polymer representation has gained some attention in recent years, due to its possible relation with Planck scale physics. In particular, this approach has been followed in a symmetric sector of loop quantum gravity known as loop quantum cosmology.\nHere we explore different aspects of the relation between the ordinary Schroedinger theory and the polymer description. The paper has two parts. In the first one, we derive the polymer quantum mechanics starting from the ordinary Schroedinger theory and show that the polymer description arises as an appropriate limit. In the second part we consider the continuum limit of this theory, namely, the reverse process in which one starts from the discrete theory and tries to recover back the ordinary Schroedinger quantum mechanics. We consider several examples of interest, including the harmonic oscillator, the free particle and a simple cosmological model."},{"figure_ref":[],"heading":"Field Content","publication_ref":[],"table_ref":[],"text":"Prompt Generate a 150-220-word abstract for work with title: Polymer Quantum Mechanics and its Continuum Limit. Describe the results obtained, problem work attempts to solve, and the key ideas and methodology in a formal academic and scientific writing voice. Use the first plural person form. Use active voice."},{"figure_ref":[],"heading":"Machine","publication_ref":[],"table_ref":[],"text":"In this paper, we investigate the continuum limit of polymer quantum mechanics. The aim of our work is to better understand the quantum behavior of polymers and its relationship to the classical theory. Our research attempts to clarify the nature of the continuum limit and determine the correctness of the approximations made. Specifically, we study the connectivity of the polymer strands. We find that a continuous path can be defined in the polymer configuration space, which is essential for the polymer to display continuum behavior. We show that the approximation of continuous randomness in the spatial variable can certainly be used when the polymer is stretched out over a significant distance. For a coil-like polymer, our results demonstrate that its behavior is non-classical. Our methodology involves a combination of intuition, formalism, and numerical simulations. We evaluate a path integral representation of polymer quantum mechanics in the continuum limit, and use numerical simulations to support our results. The ideas proposed in this work are relevant to the understanding of the thermodynamic properties of polymers, as well as to quantum computing and information. The findings of this work present a theoretical framework for the investigation of other polymer properties that are still poorly understood.\nLanguage English LLM Generator DAVINCI-003 Domain Arxiv"},{"figure_ref":[],"heading":"Title","publication_ref":[],"table_ref":[],"text":"Effects of Dirac sea on pion propagation in asymmetric nuclear matter"},{"figure_ref":[],"heading":"Human","publication_ref":[],"table_ref":[],"text":"We study pion propagation in asymmetric nuclear matter (ANM). One of the interesting consequences of pion propagation in ANM is the mode splitting for the different charged states of pions. First we describe the pion-nucleon dynamics using the non-chiral model where one starts with pseudoscalar (PS) πN coupling and the pseudovector (PV) representation is obtained via suitable non-linear field transformations. For both of these cases the effect of the Dirac sea is estimated. Subsequently, we present results using the chiral effective Lagrangian where the short-distance behavior (Dirac vacuum) is included by re-defining the field parameters as done in the modern effective field theory approach developed recently. The results are compared with the previous calculations for the case of symmetric nuclear matter (SNM). Closed form analytical results are presented for the effective pion masses and dispersion relations by making hard nucleon loop (HNL) approximation and suitable density expansion."},{"figure_ref":[],"heading":"Prompt","publication_ref":[],"table_ref":[],"text":"Rephrase the abstract of an article with title 'Effects of Dirac sea on pion propagation in asymmetric nuclear matter'. 'We study pion propagation in asymmetric nuclear matter (ANM). One of the interesting consequences of pion propagation in ANM is the mode splitting for the different charged states of pions. First we describe the pion-nucleon dynamics using the non-chiral model where one starts with pseudoscalar (PS) πN coupling and the pseudovector (PV) representation is obtained via suitable non-linear field transformations. For both of these cases the effect of the Dirac sea is estimated. Subsequently, we present results using the chiral effective Lagrangian where the short-distance behavior (Dirac vacuum) is included by re-defining the field parameters as done in the modern effective field theory approach developed recently. The results are compared with the previous calculations for the case of symmetric nuclear matter (SNM). Closed form analytical results are presented for the effective pion masses and dispersion relations by making hard nucleon loop (HNL) approximation and suitable density expansion.'"},{"figure_ref":[],"heading":"Machine","publication_ref":[],"table_ref":[],"text":"The propagation of pions in asymmetric nuclear matter (ANM) is studied in this article. We consider two approaches: a non-chiral model where the pseudoscalar/pseudovector coupling is determined using non-linear field transformations, and a chiral effective Lagrangian incorporating the short-distance behaviors known as the Dirac vacuum. In both cases, the consequences from these approaches on the mode splitting of the charged pions are evaluated. Using a hard nucleon loop approximation, accompanied by parameters deduced from a density expansion, closedform analytical results for the effective pion masses and dispersion relations are presented in comparison to calculations for symmetric nuclear matter.\nLanguage English LLM Generator COHERE Domain Wikipedia Title Shed (deity)"},{"figure_ref":[],"heading":"Human","publication_ref":[],"table_ref":[],"text":"Shed was a deity from ancient Egyptian religion. Popularly called \"the Savior\", he is first recorded after the Amarna Period. Representing the concept of salvation, Shed is identified with Horus, particularly Horus the Child."},{"figure_ref":[],"heading":"Field Content","publication_ref":[],"table_ref":[],"text":"Rather than have formal worship in a temple or as an official cult, he appears to have been a god that ordinary Egyptians looked to save them from illness, misfortune or danger. He is shown on the Metternich Stela as vanquishing danger in the form of a serpent, a scorpion and a crocodile. The rise of \"Savior\" names in personal piety during the Amarna period has been interpreted as the popular response of ordinary people to the attempts by Akhenaten to proscribe the ancient religion of Egypt. Shed has also been viewed as a form of the Canaanite god Resheph. Shed can be depicted as a young prince overcoming snakes, lions and crocodiles. Shed has been viewed as a form of savior, a helper for those in need when state authority or the kingś help is wanting. The increased reliance on divine assistance could even extend to saving a person from the Underworld, even to providing a substitute, and lengthening a personś time in this world.\nIn the New Kingdom Shed \"the savior\" is addressed on countless stelae by people or praising him for help.References Egyptian gods Savior gods Prompt Write a Wikipedia article with the title \"Shed (deity)\", the article should at least have 250 words Machine Shed (deity) Shed is a deity in the sqlrunen mythos. Shed is described as a giantess, a daughter of the primordial mother, Nott. In most traditions, Shed is a single entity, but in some traditions, Shed is a collective of giantesses. Shed is the mother of the first race of giants, the primordial beings of chaos and destruction. Shed mothered the giants through her mating with the primordial god, Odin. In some traditions, the giantess Gunnlod or Hyr is described as the mother of the giants, while Shed is described as the mother of the gods. Shed is the spouse of the primordial god, Odin. Odin and Shed had many offspring, including the gods, the giants, the Lamonts, the Maquas, the Frost Giants, and the Ice Giants. Shed is described as a beautiful giantess, with golden hair, a beautiful face, and a beautiful body. She is said to have worn a long golden gown and a gold crown. She is described as the most beautiful of all the giantesses. Shed is the goddess of fertility and reproduction. She is also the goddess of the harvest. In some traditions, Shed is also the goddess of war and death. Shed is the mother of the gods. She gave birth to the gods through her union with the primordial god, Odin. The gods were born from her body, and she nourished them with her milk. Shed is the protector of the gods. She is said to have fought alongside the gods in their war against the giants. She is also said to have fought against the dragons and the serpents. Shed is the goddess of the hunt. She is said to have accompanied the gods on their hunting expeditions. She is also said to have provided them with food and shelter.\nLanguage English LLM Generator DAVINCI-003 Domain Wikihow"},{"figure_ref":[],"heading":"Title How to Buy Virtual Console Games for Nintendo Wii","publication_ref":[],"table_ref":[],"text":"Human They are about $20 a card. Or, if you want to just buy points with your credit card, Skip down to the section, With a Credit Card., You don t want people tapping into your Wii and stealing your games, do you? On the back of the Wii Points Card, this same code can be found by scratching out the grey section with a coin or similar. You will be informed if you entered an invalid (wrong) code. Go back to the main Shop Channel menu now. Nintendo, from time to time, offers freebies via a special channel called \"WiiWare\". There are over 50 games to list, so be patient if the option you selected is taking a while to load. Don t forget to say you want to download to your Wii, unless you want your classics transportable, for example if you plan to demo the classics to a friend with a Wii. Nintendo added an animation to this process that shows Mario or Luigi collecting coins. This is why you want a secure connection!, Note that a tax is added to your purchase to comply with federal law. "},{"figure_ref":[],"heading":"Human","publication_ref":[],"table_ref":[],"text":"Especially check that the fishing line can be placed in a location which doesn t get too much in the way of your normal use (heavy users of the mouse buttons or touchpad might have problems here), and that there are locations more or less in the center of the laptop where gluing the Velcro will not affect the use of the laptop (that no access doors or ventilation holes are sealed). , Try to make the line of the two strips be as perpendicular as possible to the axis of the display hinges.\n, This can take 24 hours or even more. See the instructions for the particular epoxy you have used. Don t confuse the setting time of the epoxy (which can be as quick as 5 minutes) with the time until it reaches maximum strength.\n, Actually, what is more important is that the length of the hooked strip totally utilize the glued strips when the laptop display is open. Do not use the Velcro with an open display before the glue has hardened to its maximum strength. , Laptops with large or heavy displays may require gluing multiple strips of Velcro, or wider and/or stronger Velcro (Velcro comes in a range of sizes and bonding strengths). The laptop in the images has a 12.1\" display. , I used an \"improved blood knot\" for this. Instructions can be found in various places on the Internet. , If it is a bother, make the adhesive on the outside less sticky by touching it repeatedly."},{"figure_ref":[],"heading":"Field Content","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_1"],"heading":"Prompt","publication_ref":[],"table_ref":[],"text":"Please, generate wikihow article with length above 1000 characters from title 'How to Workaround Damaged Laptop Display Hinges' and headline 'Read the entire article, and check that your specific laptop can be repaired in the way presented., Cut two strips of the fuzzy Velcro and round the corners (rounding the corners will make it less likely that the Velcro will detach after gluing)., Read the instructions for the epoxy glue; mix it and use it to glue the two strips of fuzzy Velcro, one on the back of the display, and one on the bottom of the laptop, such that the two strips are co-linear (in other words, the alignment of one strip, when extended, coincides with the alignment of the other strip)., Wait for the epoxy glue to harden fully.,Close the laptop display, and cut a strip of the hooked Velcro more or less to the length required when sticking it to the two glued strips in their entirety., After the glue has hardened to maximum strength, check that the Velcro can now prevent the display from falling into the closed position., Take the fishing line, and tie it into a loop which exactly encircles the laptop when the display is open to the working position., Check that the fishing line loop is strong enough to prevent the display from falling into the fully open position (fishing line comes in a variety of strengths, use a relatively strong line which will not stretch under the load)., Glue a small strip of two-sided foam tape (like used for mounting things) on the edge of the laptop where the fishing line should rest., Optionally, you can use a plastic bag closure which has been notched (see images), in order to more easily adjust the exact size of the fishing line loop (and therefore, the exact working position of the laptop display).'\nMachine Have you ever experienced a damaged laptop display hinge? If so, you know how frustrating it can be when your display doesn't stay in the correct position. Fortunately, there's a workaround that you can use to fix this issue without having to replace the entire laptop hinge. This wikihow article will guide you through the process of using Velcro and fishing line to strengthen your laptop display hinges, keeping your display in the right position while you work.\n1. Read the entire article, and check that your specific laptop can be repaired in the way presented. Before getting started with any repairs, it's important to read the entire article to make sure that this method will work for your laptop model. Some laptops may have different hinge designs that won't be compatible with the Velcro and fishing line method we're using here.\n2. Cut two strips of the fuzzy Velcro and round the corners.\nTo begin the repair process, you'll need to cut two strips of fuzzy Velcro and round the corners of each strip. This will make it less likely that the Velcro will detach after gluing.\n3. Glue the Velcro to the laptop display and bottom.\nNext, mix your epoxy glue according to the instructions, and use it to glue the two strips of fuzzy Velcro -one on the back of the display and one on the bottom of the laptop. Make sure that the two strips are co-linear, meaning that the alignment of one strip, when extended, coincides with the alignment of the other strip. 4. Wait for the glue to dry. After gluing the Velcro to the laptop, you'll need to wait for the epoxy glue to harden fully. This may take several hours, so be patient and allow the glue to dry completely. 5. Cut a strip of hooked Velcro to length.\nOnce the glue has dried, close your laptop display and cut a strip of hooked Velcro more or less to the length required when sticking it to the two glued strips in their entirety. 6. Check that the Velcro can prevent the display from falling into the closed position.\nAfter the glue has hardened to maximum strength, check that the Velcro can now prevent the display from falling into the closed position. Abstract Integrating text and knowledge into a unified semantic space has attracted significant research interests recently. However, the ambiguity in the common space remains a challenge, namely that the same mention phrase usually refers to various entities. In this paper, to deal with the ambiguity of entity mentions, we propose a novel Multi-Prototype Mention Embedding model, which learns multiple sense embeddings for each mention by jointly modeling words from textual contexts and entities derived from a knowledge base. In addition, we further design an efficient language model based approach to disambiguate each mention to a specific sense.\nIn experiments, both qualitative and quantitative analysis demonstrate the high quality of the word, entity and multi-prototype mention embeddings. Using entity linking as a study case, we apply our disambiguation method as well as the multi-prototype mention embeddings on the benchmark dataset, and achieve state-of-the-art performance.\nHuman -Strengths: Good ideas, simple neural learning, interesting performance (altough not striking) and finally large set of applications. -Weaknesses: amount of novel content. Clarity in some sections. The paper presents a neural learning method for entity disambiguation and linking. It introduces a good idea to integrate entity, mention and sense modeling within the smame neural language modeling technique. The simple training procedure connected with the modeling allows to support a large set of application. The paper is clear formally, but the discussion is not always at the same level of the technical ideas. The empirical evaluation is good although not striking improvements of the performance are reported. Although it seems an extension of (Yamada et al., CoNLL 2016), it adds novel ideas and it is of a releant interest.\nThe weaker points of the paper are: -The prose is not always clear. I found Section 3 not as clear. Some details of Figure 2 are not explained and the terminology is somehow redundant: for example, why do you refer to the dictionary of mentions? or the dictionary of entity-mention pairs? are these different from text anchors and types for annotated text anchors? -Tha paper is quite close in nature to Yamada et al., 2016) and the authors should at least outline the differences.\nOne general observation on the current version is: The paper tests the Multiple Embedding model against entity linking/disambiguation tasks. However, word embeddings are not only used to model such tasks, but also some processes not directly depending on entities of the KB, e.g. parsing, coreference or semantic role labeling. The authors should show that the word embeddings provided by the proposed MPME method are not weaker wrt to simpler wordspaces in such other semantic tasks, i.e. those involving directly entity mentions. I did read the author's response."},{"figure_ref":[],"heading":"Prompt","publication_ref":[],"table_ref":[],"text":"Write a peer review by first describing what problem or question this paper addresses, then strengths and weaknesses, for the paper: title"},{"figure_ref":[],"heading":"Machine","publication_ref":[],"table_ref":[],"text":"The paper \"Bridge Text and Knowledge by Learning Multi-Prototype Entity Mention Embedding\" proposes a new method for utilizing external knowledge sources to improve entity detection and disambiguation in natural language processing tasks. The authors specifically address the issue of how to effectively incorporate large-scale knowledge bases such as Wikipedia into existing machine learning models. The strengths of this paper are its clear and concise explanation of the proposed methodology, use of established evaluation metrics, and consideration of potential limitations and future directions.\nThe authors provide a detailed description of their multi-prototype entity mention embedding (MPEME) approach, which includes the use of different prototypes to represent entities with multiple meanings in Wikipedia. They evaluate their approach on two established datasets and demonstrate improved performance compared to baseline models. Additionally, the authors note potential limitations of their approach such as the requirement for large knowledge bases and potential bias in the selection of prototypes. One weakness of this paper is that it assumes knowledge bases such as Wikipedia are comprehensive and reliable sources of information. There is a growing body of research showing that knowledge bases can contain errors or biases, which may affect the accuracy of this approach. The authors acknowledge this issue but do not offer a solution, which limits the generalizability of their findings. Additionally, the paper provides limited discussion on how this approach can be scaled up for larger knowledge bases, which is an important consideration for practical applications.\nOverall, the paper provides a valuable contribution to the field by addressing an important challenge of incorporating external knowledge sources into NLP models. The proposed MPEME method shows promise for improving entity detection and disambiguation, but further research is needed to fully explore its potential and limitations. "},{"figure_ref":[],"heading":"Prompt","publication_ref":[],"table_ref":[],"text":"Tulislah suatu artikel berita yang berjudul \"Ahmad Dhani Klaim Tak Punya Informasi soal Sri Bintang\". Artikelnya harus mengandung paling tidak 300 kata dalam bahasa Indonesia.\nTambahkan \"Jakarta, CNN Indonesia -\" di awal artikel.\nMachine Jakarta, CNN Indonesia -Ahmad Dhani, personel grup band Dewa 19, mengaku tak memiliki informasi mengenai sosok Sri Bintang yang tengah ramai dibicarakan di media sosial. Seperti diketahui, Sri Bintang menjadi perbincangan publik setelah dirinya mengunggah video yang menyebutkan bahwa personel band Dewa 19, Ahmad Dhani, pernah menyebut Sri Bintang sebagai sosok yang berperan dalam pembuatannya dari awal. Klaim tersebut kemudian dibantah oleh Ahmad Dhani melalui sebuah video yang diunggahnya di akun media sosialnya. \"Saya sendiri nggak tahu siapa Sri Bintang itu. Jadi saya gak bisa bicara, saya gak bisa komen, kenapa orang lain mau ngomongin saya tentang Sri Bintang,\" ujar Ahmad Dhani dalam video tersebut. Ahmad Dhani juga mengatakan bahwa beberapa orang yang mengklaim tahu tentang perjalanannya dalam industri musik sebenarnya tidak memiliki informasi yang tepat. \"Saya sekali lagi ingin tegaskan bahwa ada beberapa orang yang bicara itu nggak tahu apa yang mereka bicarakan. Intinya, kita perlu memperhatikan orang yang kita ngomongin,\" tegasnya. Kendati begitu, Ahmad Dhani mengungkapkan bahwa dirinya berharap Sri Bintang bisa berkarya dan memperlihatkan kemampuan yang dimilikinya di industri musik tanah air. Ahmad Dhani juga berharap bahwa sosok Sri Bintang tidak menjadi obyek perdebatan yang tak berkesudahan di antara para penggemar musik dan publik di Indonesia. \"Setidaknya semua artis harus diberikan kesempatan yang sama untuk berkarya. Saya berharap Sri Bintang bisa berkarya dengan baik dan melebihi ekspektasi dari semua orang,\" harapnya. Meski begitu, klaim dari Sri Bintang mengenai Ahmad Dhani telah menjadi perhatian banyak pengguna media sosial. Banyak netizen yang menyatakan bahwa hal tersebut merupakan sebuah bukti bahwa industri musik di Indonesia selalu menarik perhatian masyarakat dan penuh dengan kontroversi. Maka dari itu, publik diharapkan mampu menghormati karya dari masing-masing musisi dan menghindari terjadinya perdebatan yang tak perlu di media sosial. "},{"figure_ref":[],"heading":"Field","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Acknowledgments","publication_ref":[],"table_ref":[],"text":"We thank the anonymous reviewers and the program committee chairs for their very helpful and insightful comments, which have helped us improve the paper."},{"figure_ref":[],"heading":"Ethics and Broader Impact","publication_ref":[],"table_ref":[],"text":"Below, we discuss some potential ethical concerns about the present work."},{"figure_ref":[],"heading":"Field Content prompt","publication_ref":[],"table_ref":[],"text":"Write a peer review by first describing what problem or question this paper addresses, then strengths and weaknesses, for the paper title, its main content is as below: abstract Machine In this paper, the authors tackle the problem of ambiguity in entity mentions when integrating text and knowledge into a unified semantic space. Specifically, they propose a Multi-Prototype Mention Embedding model that learns multiple sense embeddings for each mention by jointly modeling words from textual contexts and entities from a knowledge base. They also design an efficient language model-based approach to disambiguate each mention to a specific sense.\nThe authors evaluate their proposed model on a benchmark dataset for entity linking and report state-of-the-art performance.\nOverall, the paper presents a well-motivated, well-executed, and well-evaluated solution to a relevant research problem. The main contributions of the paper are the Multi-Prototype Mention Embedding model and the efficient language model-based disambiguation approach.\nThe authors clearly describe the model and how it addresses the problem of mention ambiguity.\nThe evaluation of the model is thorough, including both qualitative and quantitative analyses.\nOne of the strengths of the paper is the use of a benchmark dataset, which allows for fair comparison with previous approaches. Additionally, the authors provide detailed explanations of the implementation and evaluation of their model, making the paper easily replicable. The experiments demonstrate the high quality of the word, entity, and multi-prototype mention embeddings produced by the proposed model. However, there are a few weaknesses in the paper. While the evaluation of the proposed method is thorough and convincing, it would have been beneficial if the authors had provided more in-depth analysis of the results to show exactly where and why their method outperforms previous approaches. Additionally, the paper would have benefited from a more detailed discussion of the limitations and potential future directions of the proposed method.\nOverall, the paper is a strong contribution to the field of entity linking and the integration of text and knowledge into a unified semantic space. The proposed Multi-Prototype Mention Embedding model and the efficient language model-based disambiguation approach provide a promising solution to the problem of mention ambiguity. Machine This method has been used for years by many people who suffer from foot-nail fungus. It works well because both salt and vinegar are natural disinfectants that kill bacteria and other pathogens; they also help loosen the hardened layers around infected areas. The warm water helps dissolve the salt crystals which then act like an abrasive cleansing agent against the affected toe. Vinegar contains acetic acid which kills most types of mold spores including those responsible for athlete's foot. This treatment will not work immediately but over time should remove the infection causing the nail to become detached naturally. If there... is no improvement within three months consult a doctor.... To prepare the solution mix together equal parts of baking soda and hydrogen peroxide using enough liquid soap to create bubbles. Add some drops of tea tree oil Mix everything together in a bowl creating thick paste-like consistency. Apply directly onto the infected nail. Cover with bandage overnight. Remove next morning. Rinse away with cool water. Do twice each day for at least 2 weeks. You may need longer depending upon severity of infection. Keep doing this even though new growth appears underneath old nail. Be patient!"}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b31\",\n \"b25\",\n \"b30\",\n \"b21\",\n \"b7\",\n \"b28\",\n \"b26\",\n \"b34\",\n \"b19\",\n \"b33\"\n ],\n \"table_ref\": [],\n \"text\": \"Large language models (LLMs) are becoming mainstream and easily accessible, ushering in an explosion of machine-generated content over various channels, such as news, social media, questionanswering forums, educational, and even academic contexts. Recently introduced LLMs, such as Chat-GPT, GPT-4, LLaMA 2 (Touvron et al., 2023), and Jais (Sengupta et al., 2023), generate remarkably fluent responses to a wide variety of user queries. The high quality of the generated texts makes them attractive for replacing human labor in many scenarios. However, this raises concerns regarding their potential misuse, e.g., to spread disinformation or to cause disruptions in the education system (Tang et al., 2023).\\nSince humans perform only slightly better than chance when classifying machine-generated vs. human-written texts (Mitchell et al., 2023), we aim to facilitate the development of automatic detectors to mitigate the potential misuse of LLMs. In particular, we construct a diverse resource that could be used for training and testing various models for detecting machine-generated text (MGT).\\nPrevious efforts in detecting MGT (i) focused on only one or two particular languages, typically only on English, (ii) used a single generator, e.g., just ChatGPT (Guo et al., 2023;Shijaku and Canhasi, 2023), (iii) leveraged fine-tuned LLMs for specific tasks, e.g., machine translation or text summarization (Shamardina et al., 2022), or (iv) considered only one specific domain e.g., news (Zellers et al., 2019;Macko et al., 2023). In contrast, here we encompass multiple languages, various LLMs, and several diverse domains, aiming to enable more general machine-generated text detection. Our dataset serves as the basis for SemEval-2024 Task 8 (Wang et al., 2024).\\nOur contributions are as follows:\\n• We construct M4: a large-scale multigenerator, multi-domain, and multi-lingual corpus for detecting machine-generated texts in a black-box scenario where there is no access to a potential generator or its outputs except for plain text.\\n• We study the performance of automatic detectors from various perspectives: (a) different detectors across different domains for a specific LLM generator, (b) different detectors across different generators for a specific domain, (c) interactions of domains and generators in a multilingual setting, and (d) the performance of the detector on data generated from different time periods. From these experiments, we draw a number of observations, which can inform future research.\\n• We release our data and code freely, and we plan to keep our repository constantly growing, adding new generators, domains, and languages over time. The remainder of the paper is organized as follows: Section 2 discusses related work. Section 3 describes the process of collecting the corpus from multiple generators (including davinci-text-003, ChatGPT, GPT4, Cohere, Dolly2, and BLOOMz), multiple domains (including Wikipedia, WikiHow, Reddit, QA, news, paper abstracts, and peer reviews), and multiple languages (Arabic, Bulgarian, Chinese, English, Indonesian, Russian, and Urdu) for machine-generated text detection. Section 4 presents the seven detectors we experiment with. Section 5 evaluates their performance across domains given a generator (ChatGPT or davinci) and across generators given a domain (arXiv or Wikipedia), as well as across different languages. Finally, Section 6 concludes and points to possible directions for future work.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b24\",\n \"b29\",\n \"b8\",\n \"b13\",\n \"b35\",\n \"b15\",\n \"b21\",\n \"b32\",\n \"b7\",\n \"b28\",\n \"b26\",\n \"b34\",\n \"b19\",\n \"b7\",\n \"b28\",\n \"b5\",\n \"b30\",\n \"b7\"\n ],\n \"table_ref\": [],\n \"text\": \"White-Box vs. Black-Box Detection We categorize the detection strategies into black-box and white-box, contingent on the level of access to the LLM that is suspected to have generated the target text. White-box methods focus on zero-shot detection without any additional training overhead (Sadasivan et al., 2023). Some use watermarking techniques (Szyller et al., 2021;He et al., 2022;Kirchenbauer et al., 2023;Zhao et al., 2023) and others rely on the expected per-token log probability of texts (Krishna et al., 2022;Mitchell et al., 2023). Black-box detectors only need API-level access to the LLM (i.e., when only the generated text is available) and typically extract and select features based on training text samples originating from both human and machine-generated sources.\\nIn this study, we focus on black-box techniques because they aim to solve the task for the more practical and general use case. However, we note that their effectiveness heavily depends on the quality and the diversity of the training corpus.\\nRelated Corpora Recently, a growing body of research has concentrated on amassing responses generated by LLMs. TuringBench (Uchendu et al., 2021) comprises 200K human-and machinegenerated pieces of text from 19 generative models. However, it is outdated, as the most advanced model used in this research is GPT-3. Guo et al. (2023) collected the HC3 dataset, which consists of nearly 40K questions and their corresponding answers from human experts and ChatGPT (English and Chinese), covering a wide range of domains (computer science, finance, medicine, law, psychology, and open-domain). Shijaku and Canhasi (2023) gathered TOEFL essays written by examined people and such generated by ChatGPT (126 essays for each).\\nThe RuATD Shared Task 2022 involved artificial texts in Russian generated by various language models fine-tuned for specific domains or tasks such as machine translation, paraphrase generation, text summarization, and text simplification (Shamardina et al., 2022). We pay more attention to zero-shot generations of LLMs, such as the subset of RuATD generated by ruGPT-3.\\nIn general, previous studies have concentrated on detecting machine-generated texts in one or two languages, for a specific LLM such as ChatGPT, or within a single domain such as news (Zellers et al., 2019;Macko et al., 2023). Our work broadens this scope to include multiple languages and a variety of widely-used LLMs across different domains.\\nBlack-box Detectors are usually binary classifiers based on three types of features: statistical distributions (Guo et al., 2023;Shijaku and Canhasi, 2023), e.g., GLTR-like word rankings (Gehrmann et al., 2019), linguistic patterns (such as vocabulary, part-of-speech tags, dependency parsing, sentiment analysis, and stylistic features), and factverification features (Tang et al., 2023). Classification models involve deep neural networks, such as RoBERTa (Guo et al., 2023), or more traditional algorithms, such as logistic regression, support vector machines, Naïve Bayes, and decision trees.\\nThere are also widely-used off-the-shelf MGT detectors, such as the OpenAI detector, 1 GPTZero, 2 and ZeroGPT. 3 According to the limited public information about them, these detectors are trained on collections of human-written texts and texts generated by various LLMs. For example, the training data of the OpenAI detector contains generations from 34 LLMs from various organizations, including OpenAI itself. For our M4 dataset, we selected a diverse set of state-of-the-art black-box methods and features, including one off-the-shelf detector.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"The M4 Dataset\",\n \"publication_ref\": [\n \"b14\",\n \"b12\",\n \"b26\",\n \"b3\",\n \"b22\"\n ],\n \"table_ref\": [],\n \"text\": \"We gather human-written texts from a diverse range of sources across various domains and languages. For English we have Wikipedia (the March 2022 version), WikiHow (Koupaee and Wang, 2018), Reddit (ELI5), arXiv, and PeerRead (Kang et al., 2018), for Chinese we have Baike/Web QA question answering (QA), for Russian we have RuATD (Shamardina et al., 2022), for Arabic Wikipedia, and we use news for Urdu, Indonesian, and Bulgarian. Details about the data sources are provided in Appendix A.1 and A.2.\\nFor machine generation, we prompt the following multilingual LLMs: GPT-4, ChatGPT, GPT-3.5 (text-davinci-003), Cohere, Dolly-v2 (Conover et al., 2023), and BLOOMz 176B (Muennighoff et al., 2022). The models are asked to write articles given a title (Wikipedia), abstracts given a paper title (arXiv), peer reviews based on the title and the abstract of a paper (PeerRead), news briefs based on a title (news), also to summarize Wikipedia articles (Arabic), and to answer questions (e.g., Reddit and Baike/Web QA).4 \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Collection\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_9\",\n \"tab_10\"\n ],\n \"text\": \"Prompt Diversity For each generator, we carefully designed multiple (2-8) prompts in various styles, aiming to produce diverse outputs that are more aligned to divergent generations in real-world application scenarios. For example, on simple domains of Wikipedia and WikiHow, two prompts are applied. For arXiv and Reddit, as well as for ChatGPT, we use five prompts and four prompts for PeerRead. We generate varying tones of responses with prompts such as answer the question (1) \\\"like I am five years old\\\"; (2) \\\"in an expert confident voice\\\"; (3) \\\"in a formal academic and scientific writing voice\\\"; etc. Table 7 in Appendix A gives some statistics about the prompts used to generate the data collection, and Table 8 shows the hyper-parameters for the various generators.\\nData Cleaning Simple artifacts in MGTs, such as multiple newlines and bullet points, could assist detectors, as their presence in the training data may discourage detectors from learning more generalized signals.\\nTherefore, we performed minimal cleaning of the human-written and the machine-generated texts: (i) in a human-written WikiHow text, we removed multiple commas at the beginning of a new line (like \\\"\\\"\\\"\\\"\\\"\\\", we believe that ...\\\") and repeating newlines (\\\"\\\\n\\\\n\\\\n\\\\n\\\\n text begin \\\\n\\\\n\\\\n\\\\n\\\\n\\\"); (ii) in machine-generated WikiHow texts, we removed bullet points (as there were no bullet points in human-written texts); (iii) in human-written Wikipedia articles, we removed references (e.g., [1], [2]), URLs, multiple newlines, as well as paragraphs whose length was less than 50 characters; and (iv) in human-written arXiv abstracts, we removed newlines stemming from PDF conversion.\\nQuality Control Unlike other tasks, where the data quality can be evaluated through the agreement between annotators over gold labels, we naturally obtain gold labels along with the collection of machine-generated texts. Therefore, we checked the data quality by randomly sampling 10-20 cases for each domain/generator and manually assessing the plausibility of generated texts. This can effectively circumvent incoherent, disorganized, and illogical generations that are easy to distinguish from human-written ones due to improper prompts or hyper-parameter settings of the generators (e.g., some generators repeat newly generated snippets to satisfy the minimum setup of new tokens). Moreover, in order to mimic human-written texts, we control the length of MGTs.\\nIt should be highlighted that we did not pick examples. The quality control we exercised was model-level rather than example-level. We checked for cases where a model fundamentally failed, e.g., by generating visibly very bad output (e.g., very repetitive, English instead of foreign language output, etc.). This was very high-level checking (whether to keep a certain model in M4 or not); at the individual example level, we just checked whether the output had at least 1000 characters in length. Thus, we believe any biases that we might have introduced are minimal.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Statistics\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"The overall statistics about our M4 dataset for different tasks and languages are given in Table 1. We collected ∼ 147k human-machine parallel data in total, with 102k for English and 45k for other languages: 9k for Chinese, Russian, and Bulgarian; and 6k for Urdu, Indonesian, and Arabic respectively, in addition to over 10M non-parallel human-written texts. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Data Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We performed analysis of our dataset in terms of vocabulary richness at the n-gram level, as well as in terms of human performance on the task of detecting machine-generated content.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"N-gram Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_12\"\n ],\n \"text\": \"We compared the uni-gram and the bi-gram distributions of human-written vs. machine-generated texts and found that the former had a richer vocabulary than each of the six generators; see Table 9 in Appendix A.4 for detail. Dolly-v2 had the largest number of unique uni-and bi-grams, followed by davinci, ChatGPT, and BLOOMz, and Cohere had the least. The combination of all generators had comparable vocabulary to humans. When comparing across domains, we observed that Wikipedia, which covers a wide range of topics, contains the highest number of unique unigrams, followed by WikiHow and Reddit. In contrast, arXiv and PeerRead, which are specific to academic papers and peer reviews, exhibited fewer unique uni-grams and bi-grams. Within the same domain, we calculated the overlap of unique unigrams and bi-grams between human and machinegenerated texts. This overlap ranges in 20-35% for unigrams and in 10-20% for bi-grams. These variations can provide distinctive signals for black-box machine-generated text detection approaches.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Human Evaluation\",\n \"publication_ref\": [\n \"b0\"\n ],\n \"table_ref\": [\n \"tab_4\",\n \"tab_4\"\n ],\n \"text\": \"From the Reddit and the arXiv (ChatGPT) test sets, for each domain, we sampled the first 50 (human, machine) pairs of texts and shuffled them into two groups, where two texts from the same pair would go in different groups. The annotators were then asked to focus on one group, which meant that they had to make a decision looking at each example individually, rather than having a pair of examples and deciding which one in the pair was humanwritten and which one was machine-generated (as some previous work did). This ensures a realistic scenario. For Reddit, we had 29 examples by humans and 21 by machines for group 1, and (21 human, 29 machine) for group 2; and (human:26, machine:24) for arXiv group 1, (human:24, machine:26) for arXiv group 2.\\nWe had a total of six human annotators, who came from different countries and were native speakers of different languages. They were all proficient in English and all had NLP background: three PhD students, two MSc students, and 1 postdoc. Annotator 3 was an English native speaker who is also proficient in Arabic. Annotators 1 and 4 were Chinese native speakers, annotators 2 and 6 ware Russian native speakers, and annotator 5 was a Bulgarian native speaker.\\nEach annotator made a guess about 17 unique examples for Reddit (finished by six annotators) and 25 examples for arXiv (finished by four). 5 The results are shown in Table 2. Interestingly, the English native speaker did not perform as well as some other annotators. We can further see in Table 2 that annotator 4 performed much better than annotator 1, even though they were both Chinese native speakers; this may be because annotator 4 had better understanding of how LLM generations work. Moreover, annotator 6 was the best rater, and he was also the one who was very familiar with LLM generation mechanisms, achieving higher guessing accuracy than annotator 2.\\nThus, the annotators' proficiency in English may affect the evaluation, but for equal language proficiency, the degree of understanding of the LLM generation styles or patterns will also impact the quality of the annotator's guess.\\nOn average, the accuracy of the human guesses was 0.77 for Reddit and 0.72 for arXiv. This indicates that it is not easy for humans to detect machine-generated text, especially for nonnative English speakers who are not familiar with the ChatGPT generation patterns (e.g., annotators 1,3,5). Besides, it is harder to classify the texts from arXiv than from Reddit. This is consistent with the findings in Clark et al. (2021). Without training, evaluators distinguished between GPT3-written and human-authored text at the chance level, and training by detailed instructions, annotated examples, and paired examples will improve the accuracy while the improvement across domains differs.\\nWe hypothesize that our human annotators depended less on content signals and more on stylistic cues when identifying MGT for the arXiv domain, which results in the accuracy disparity between the two domains. Overall, it is challenging for general readers to understand and to follow abstracts of academic papers, but it is much easier to read Reddit answers.\\nWe further compared the human performance to an XLM-R detector fine-tuned on in-domain training data. The classifier achieved near-perfect accuracy across the two domains, outperforming all human annotators. These findings strongly indicate the potential for automated in-domain black-box detection.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Detectors\",\n \"publication_ref\": [\n \"b18\",\n \"b1\",\n \"b2\"\n ],\n \"table_ref\": [\n \"tab_14\"\n ],\n \"text\": \"We evaluated seven detectors; see Table 11 for their hyper-parameter settings.\\nRoBERTa This detector is based on the pretrained RoBERTa model (Liu et al., 2019), which we fine-tuned to detect machine-generated texts.\\nELECTRA We further fine-tuned ELEC-TRA (Clark et al., 2020). Its pre-training objective is more aligned with our MGT task: it was pretrained to predict whether a token in a corrupted input was replaced by a plausible alternative sampled from a small generator network.\\nXLM-R We fine-tuned XLM-RoBERTa, a multilingual variant of RoBERTa (Conneau et al., 2019).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Logistic Regression with GLTR Features\",\n \"publication_ref\": [\n \"b5\"\n ],\n \"table_ref\": [],\n \"text\": \"We trained a logistic regression model based on 14 GLTR features from (Gehrmann et al., 2019), which are based on the observation that most LLM decoding strategies sample high-probability tokens from the head of the distribution. Thus, word ranking information about an LLM can be used to distinguish machine-generated texts from human-written ones. We selected two categories of these features: (i) the number of tokens in the top-10, top-100, top-1000, and 1000+ ranks from the LM predicted probability distributions (4 features), and (ii) the Frac(p) distribution over 10 bins ranging from 0.0 to 1.0 (10 features). Frac(p) describes the fraction of probability for the actual word divided by the maximum probability of any word at this position.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Stylistic Features\",\n \"publication_ref\": [\n \"b16\",\n \"b9\",\n \"b6\"\n ],\n \"table_ref\": [\n \"tab_4\",\n \"tab_5\"\n ],\n \"text\": \"We trained an SVM classifier based on stylistic features from (Li et al., 2014): (i) character-based features, e.g., number of characters, letters, special characters, etc., (ii) syntactic features, e.g., number of punctuation and function words, (iii) structural features, e.g., total number of sentences, and (iv) word-based features, e.g., total number of words, average word length, average sentence length, etc. 12 and13)\\nNEws LAndscape (NELA) We trained an SVM classifier using the NELA features (Horne et al., 2019), which cover six aspects: (i) style: the style and the structure of the article; (ii) complexity: how complex the writing is; (iii) bias: overall bias and subjectivity; (iv) affect: sentiment and emotional patterns; (v) moral: based on the Moral Foundation Theory (Graham et al., 2012); and (vi) event: time and location.\\nGPTZero Finally, we used the GPTZero system without any adaptation. It was trained on a large diverse corpus of human-written and AI-generated texts, focussing on English. The system can analyze texts ranging from individual sentences to entire documents.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments and Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we first describe our experiments, which come in three settings: (i) same generator, cross-domain evaluation, (ii) same domain, crossgenerator evaluation, and (iii) cross-lingual, crossgenerator evaluation. As mentioned in the previous section, we also experiment with GPTZero in a zero-shot setting, as it has not seen our data (even though it might have been trained on some domains involved in our data). We further discuss the evaluation results of these experiments.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_3\"\n ],\n \"heading\": \"Same-Generator, Cross-Domain\",\n \"publication_ref\": [\n \"b23\"\n ],\n \"table_ref\": [\n \"tab_4\",\n \"tab_5\",\n \"tab_16\",\n \"tab_7\"\n ],\n \"text\": \"Given a specific text generator, such as ChatGPT and davinci-003, we train a detector using data from one domain and evaluate it on the test set from the same domain (in-domain evaluation) and other domains (out-of-domain evaluation). The results are shown in Figure 1 and Tables 12 and13.\\nIn-domain detection is easy and can be done with very high accuracy, sometimes very close to a perfect score of 100%. This is especially the case for the RoBERTa detector, which reaches 100% accuracy for detecting ChatGPT-generated text on arXiv, 99.7% on Wikipedia, 99.7% on WikiHow, and 98.0% on PeerRead. The only dataset where the best score for the RoBERTa detector is achieved when training on a different domain is Reddit. We can further see that the results with davinci-003 show the same pattern: all in-domain evaluation scores are usually very high, approaching 100%. Other detectors also show high performance in the in-domain evaluation setting, but they usually overfit less to a particular domain. For example, the LR-GLTR detector shows only 79.6% accuracy on WikiHow when the davinci-003 generator was used, while the score for the RoBERTa-based detector exceeds 99%.\\nThe best performance in the out-of-domain evaluation is often achieved by fine-tuning ELEC-TRA for the task. We attribute this to the specific pre-training objective of this model, which is based on the detection of replaced tokens. ELECTRA shows slightly lower performance than RoBERTa for the in-domain evaluation, but achieves huge improvements in the out-domain evaluation setting. For example, in the case of training on Wikipedia to detect davinci-003 on Reddit, the RoBERTa's performance is close to random guessing, while ELEC-TRA achieves 87.9% accuracy. Another strong approach for out-of-domain detection is LR-GLTR, which outperforms ELECTRA in some scenarios, such as detecting ChatGPT on the Wikipedia domain. 14 and15)\\nOut-of-domain detection might be hard. This is especially noticeable when training on arXiv and detecting artificial texts for Reddit or training on arXiv and detecting for Wikipedia. This is expected as these pairs of domains are very different. There are some domains that offer better generalization than others. The RoBERTa-based detector and the detector based on NELA features are the most vulnerable in this regard. RoBERTa overfits to the training domain, while the NELA features are not tailored to machine-generated text detection, but rather initiated for fake news detection.\\nThe best training domain for out-of-domain generalization is Reddit. Training on Reddit ELI5 usually yields the best out-of-domain performance. Wikipedia is also often a good domain for training. Training on arXiv and PeerRead yields the worst generalization across other domains because the writing style of academic papers is very specific.\\nThe most challenging domain for machinegenerated text detection is WikiHow, while Peer-Read is the easiest one.\\nThe GPT-3.5 (davinci-003) generator is harder to detect than ChatGPT. Aggregating the results across all domains and both generators, we can see that the accuracy for ChatGPT is usually higher than that for davinci-003. This indicates that ChatGPT may leave more distinctive signals in generated texts than davinci-003.\\nFeature Analysis. We conducted feature analysis of in-domain detectors using LIME (Ribeiro et al., 2016), and we found that detectors did not overfit to MGT artifacts and leveraged word distribution for classification. See Figure 4 in Appendix G for more detail.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Same-Domain, Cross-Generator\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_16\"\n ],\n \"text\": \"Given a specific domain, we train the detector using the training data from one generator and we evaluate it on the test data from the same and also from other generators. The accuracy on arXiv and Wikipedia is shown in Figure 2 (see Table 14 and 15 in Section D for precision, recall, and F1).\\nRoBERTa performs the best among five detectors. It is the best on both arXiv (95.9%: average accuracy) and Wikipedia (99.4%), followed by LR-GLTR (84.0/80.7%), stylistic features (80.4/82.8%), and ELECTRA (72.5/76.6%); NELA features are the worst (73.7/64.3%). We can see that apart from the main diagonal, most scores for the detector using NELA features are around or lower than 50.0%, particularly on arXiv. This indicates that they are not suitable for distinguishing machine-generated and human-written texts. Moreover, the accuracy for Wikipedia is higher than for arXiv, especially for RoBERTa pre-trained using Wikipedia data. This suggests that arXiv is somewhat harder to detect than Wikipedia, and exposure bias on pre-training can impact a detectors' domain-specific performance.\\nThe highest accuracy is for the same generator. Akin to the trend of cross-domain evaluation, training and testing using the same generator always yields the best accuracy for both arXiv and Wikipedia across the five detectors. Even for NELA, and detection over generations by BLOOMz, the accuracy mostly remains over 90.0. Performance drops substantially when the training and the test data are generated from different LLMs because of different distributions between the outputs of different generators. BLOOMz-generated text is much different from ChatGPT, davinci, and Cohere. For all detectors in both arXiv and Wikipedia, BLOOMz shows the lowest cross-generator accuracy. Specifically, when training on BLOOMz and testing on other generators, or when training on other generators and testing on BLOOMz, it shows low recall (<0.5) for machine-generated texts. This means that there are many false negative examples, namely, many machine-generated texts are misclassified as human-written ones. Most accuracy scores are ≤50.0%, i.e., similar or even worse than a random guess. This indicates that the distribution of BLOOMz outputs is very different from the other three generators. We assume that this is because BLOOMz is primarily fine-tuned for NLP downstream data.\\nMoreover, we found that, for all detectors, when training on Cohere, the accuracy for ChatGPT is comparable to the accuracy on Cohere itself, and similarly high accuracy occurs when training on ChatGPT and testing on Cohere. This suggests that ChatGPT and Cohere share some generative patterns.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Zero-shot Evaluation: GPTZero\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_5\"\n ],\n \"text\": \"Table 3 shows that, from the perspective of the domain, GPTZero performs the best on Wikipedia, while the worst results are on arXiv where, for all generators, the F1 score is below 50%. From the perspective of generators, GPTZero shows the best performance on ChatGPT and the worst performance on BLOOMz. The recall for BLOOMz is close to 0% across all domains, which is consistent with the results for other detectors. GPTZero also demosntrated low performance for Dolly v2. GPTZero may have been trained on generations of ChatGPT and on data from domains such as Wikipedia and Reddit, thus showing remarkable scores for them. At the same time, zero-shot detection for unseen domains and generators poses a major challenge for GPTZero.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Multilingual Evaluation\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_6\",\n \"tab_6\"\n ],\n \"text\": \"In this section, we discuss the results for our multilingual experiments with the XLM-R detector across seven languages. For multilingual evaluation, we used ChatGPT and davinci-003 as generators. The results are shown in Table 4 (see Section E in the Appendix for more detail).\\nWe constructed the English training, development, and test sets by combining English texts across all domains: Wikipedia, WikiHow, Reddit ELI5, arXiv, and PeerRead. Then, the All row refers to the combination of all training data in Arabic, Bulgarian, Chinese, English, Indonesian, Russian, and Urdu from the same generator. We aim to evaluate the performance of a detector over each monolingual test set from a single domain when fully leveraging the available training data, thus observing the benefits brought by the interaction of multiple languages and domains.\\nWe can see in Table 4 that the best accuracy is achieved when training and testing on the same language and using the same generator, while when training on one generator and testing on another one, the highest scores tend to appear in the row of All, i.e., when using the training data for all languages, except for Bulgarian (training on Bulgarian is best, if we want to test on Bulgarian).\\nWe can also see that it is difficult for XLM-R to detect machine-generated text in a language that it has never seen during training. For example, it struggles to detect Russian, Urdu, and Indonesian machine/human-generated text when it was not trained on them. Interestingly, XLM-R still demonstrates good performance for Arabic even when trained on English data only.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Time Domain Evaluation\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_7\"\n ],\n \"text\": \"LLMs are constantly improving over time. This raises the question of the robustness of detectors for the same generator across different time points. With this in mind, we compared ChatGPT output generated in March 2023 (from our M4 dataset) vs. September 2023 on the Reddit-ELI5 domain and using XLM-R as a detector, and the same prompts and questions as for the M4 dataset. The results are shown in Table 5, where we can see that the detector trained on the earlier version can effectively classify generations produced by the September 2023 version. This implies that a detector may remain effective even when applied to a newer generator trained using fresh data. Table 6: Impact of text length on detection accuracy on arXiv using XLM-R.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Impact of Text Length\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Finally, we investigated the impact of text length on detection accuracy. We truncated arXiv articles at the first 1,000, 500, 250, and 125 characters and compared the accuracy of XLM-R detectors trained and tested on such truncated articles for machine-generated content produced by ChatGPT.\\nThe results are shown in Table 6. We can see that as the length decreases from 1,000 to 125, the accuracy drops by 4.5 points. This illustrates the negative impact of smaller text length on detection performance; more experiments on the arXiv and the Reddit datasets are presented in Figure 3 in the Appendix.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion and Future Work\",\n \"publication_ref\": [\n \"b33\"\n ],\n \"table_ref\": [\n \"tab_1\"\n ],\n \"text\": \"We presented M4, a large-scale multi-generator, multi-domain, and multi-lingual dataset for machine-generated text detection. We further experimented with this dataset performing a number of cross-domain, cross-generator, cross-lingual, and zero-shot experiments using seven detectors. We found that detectors struggle to differentiate between machine-generated and human-written texts if the texts come from a domain, a generator, or a language that the model has not seen during training. Our results show that the problem is far from solved and that there is a lot of room for improvement. We hope that our release of M4, which we make freely available to the community, will enable future research towards more robust approaches to the pressing societal problem of fighting malicious machine-generated text. We have already created an extension of M4 for SemEval-2024 Task 8 (Wang et al., 2024), 6 which features additional languages, domains, and three new task (re)formulations.\\nIn future work, we plan to expand our M4 dataset continuously by introducing new LLM generators, by exploring different domains, by incorporating new languages, and by diversifying the range of tasks and prompts used. We believe that this is a good, practical way to keep the dataset up-to-date in response to the ongoing progress in LLMs. Our aim is to maintain a dataset that remains relevant as LLMs continue to evolve.\\nData Collection, Licenses, and User Privacy. Creating the M4 dataset did not involve scraping raw data from websites. Instead, we used preexisting corpora that have been publicly released and approved for research purposes, with clear dataset licenses, which are listed in Table 1. To the best of our knowledge, all included datasets adhere to ethical guidelines and minimize privacy concerns. Since the human-written data has already been published and made publicly available for research purposes, we see no additional privacy risks in releasing that data as part of our M4 dataset.\\nThe human text components of M4 are publicly available and can be freely accessed and used for research purposes. However, researchers must acknowledge the original sources of the text and comply with the respective licensing terms.\\nThe machine-generated text components of our M4 dataset are subject to the licensing terms of the underlying LLMs. For text generated using LLMs, researchers must comply with the respective licensing terms of those LLMs:\\n• davinci-003, ChatGPT, GPT-4: no specific license. They welcome research publications related to the OpenAI API. 7\\n• Dolly-v2: Apache 2.0 8\\n• Cohere: no specific license. They point out that CUSTOMER RETAINS ALL OWNER-SHIP AND INTELLECTUAL PROPERTY RIGHTS IN AND TO CUSTOMER DATA. 9\\n• BLOOMz: Apache 2.0 10\\nPotential Biases We recognize the potential for biases in our M4 dataset, stemming from both the original human-written corpora and the Large Language Models (LLMs) used for generation. This is an important issue, and we put efforts to minimize such biases. However, we are aware that unethical usage of our dataset may still lead to biased applications: even if our original dataset was completely unbiased, external parties may extract a biased subset, which would be out of our control.\\n7 https://openai.com/policies/ sharing-publication-policy 8 https://github.com/databrickslabs/dolly 9 https://cohere.com/saas-agreement 10 https://github.com/bigscience-workshop/xmtf Having already realized these concerns, we have implemented the following measures: a. We provide comprehensive documentation about our M4 dataset, including detailed information about the sources of all human-written corpora, the generation process for obtaining the machine-generated text, including the full prompts and the measures we took to cleanse the output, and the potential biases that may exist. We believe that this transparency would allow researchers to understand the origins of the data and to make informed decisions about how to use it.\\nb. We further acknowledge and transparently discuss these limitations and debiasing techniques that could be used to address these limitations. We hope that the strong emphasis on transparency in our methodology by explicitly stating the sources of human-written corpora and the generation processes for the corresponding machine-generated text could help clarify the dataset's origins and potential biases.\\nRobustly Secure System The M4 dataset is intended for the development of detection systems to mitigate misuse, particularly in the context of malicious content generated using LLMs. While we encourage extensive and responsible use of the datasets to advance this critical area of research, we also emphasize the importance of adhering to the licensing terms of the original human-written corpora and the corresponding LLMs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we discuss some perceived limitations of our study.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"M4 Dataset Generalization and Biases\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Generalization: Machine-generated outputs exhibit a high degree of sensitivity to the prompts. While our M4 dataset was collected with diverse prompts for a variety of generators, domains, and languages, to cover typical use cases, it has limitations as a general resource, as it is neither sufficient to train a detector that can be expected to generalize well across all possible domains and generators, nor is it likely sufficient to act as a standard benchmark that can accurately evaluate the effectiveness of a detection method.\\nUp-to-Date: Detecting machine-generated text is a very challenging task when we do not know in advance the potential generator and the domain: as our findings show, human-written and machinegenerated text cannot be distinguished in certain situations, e.g., we saw issues when using text generated by BLOOMz. Therefore, we regard M4 as a useful repository of machine-generated text for researchers who want to improve and to evaluate their detectors from multiple dimensions. Moreover, the LLMs are constantly evolving, and thus any dataset collected for machine-generated text detection can become outdated relatively fast. With this in mind, we have constantly been extending the M4 dataset (e.g., with a recent collection of GPT-4 responses), and we expect to grow our repository to enable better training and more up-to-date detectors.\\nBias: Biases may exist in both human-written and machine-generated texts, and it is possible that our M4 dataset may be influenced by biases from human collection, thus affecting the detection outcomes. We leave the analysis of such biases to our future work.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Feasibility of Black-Box Machine-Generated Text Detection\",\n \"publication_ref\": [\n \"b30\",\n \"b24\",\n \"b17\"\n ],\n \"table_ref\": [],\n \"text\": \"A growing body of work shows that machinegenerated text detection might gradually become harder and even nearly impossible: as LLMs evolve, the gap between machine-generated and human-written text might narrow (Tang et al., 2023;Sadasivan et al., 2023). Liang et al. (2023) \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Appendix A Data Collection and Analysis\",\n \"publication_ref\": [\n \"b4\",\n \"b14\",\n \"b12\",\n \"b26\",\n \"b26\",\n \"b27\",\n \"b20\",\n \"b11\",\n \"b10\"\n ],\n \"table_ref\": [],\n \"text\": \"A.1 English Corpora Wikipedia We used the Wikipedia dataset available on HuggingFace 11 and randomly chose 3,000 articles, each of which surpasses a character length limit of 1,000. We prompted LLMs to generate Wikipedia articles given titles, with the requirement that the output articles contain at least 250 words.\\nFor generation with Dolly-v2, 12 we set the minimum number of generated tokens to be 300 to satisfy the minimal character length of 1,000.\\nReddit ELI5 dataset (Fan et al., 2019) is a collection of English question-answering (QA) pairs, 13 gathered to facilitate open-domain and long-form abstractive QA. The data is derived from three categories: ExplainLikeImFive for general topics, AskScience for scientific queries, and AskHistorians for historical inquiries. Each pair is composed of a question (a title + a detailed description) and corresponding answers.\\nWe filtered out answers with less than 1,000 characters, retaining questions whose title ends with a question mark without detailed descriptions. Finally, we selected 1,000 QA pairs with top user ratings for each category, resulting in a total of 3,000 pairs.\\nWikiHow dataset 14 (Koupaee and Wang, 2018) is built from the online WikiHow knowledge base. It consists of articles with a title, a headline (the concatenation of all bold lines of all paragraphs), and text (the concatenation of all paragraphs except the bold lines). We randomly chose 3,000 articles with the length of more than 1,000 characters and prompted LLMs with titles and headlines to generate artificial articles.\\nPeerRead Reviews We sampled 586 academic papers published in top-tier NLP and machine learning conferences from the PeerRead corpus (Kang et al., 2018). Each paper contains metadata, including title, abstract, and multiple human-written reviews. Given a paper, we prompted LLMs to generate peer reviews with four different instructions; two depend only on the title and another two involve both the title and the abstract. Two prompts specify the review format of first describing what problem or question the considered paper addresses, and then providing its strengths and weaknesses. Other two prompts do not contain a review format specification. 15 This resulted in 584 × 4 = 2,344 machine-generated texts for each generator and 5,798 human-written reviews in total.\\nArxiv Abstract parallel dataset is constructed from a Kaggle corpus. 16 We sampled 3,000 abstracts with a minimum length of 1,000 characters and prompted LLMs to produce machine-generated abstracts based on their titles.\\nChinese QA is constructed from 3,000 question-answer pairs sampled from Baike and the Web QA corpus. 18 The length of each answer is more than 100 Chinese characters. We prompted LLMs with a combination of a brief title and a detailed description for each question.\\nIndonesian News 2018 is constructed from a corpus of Indonesian news articles19 collected from seven different news websites in 2018. We picked news from CNN Indonesia since this source was found to provide the cleanest data. We selected 3,000 texts from the corpus and generated artificial news articles by prompting ChatGPT with a title.\\nRussian RuATD is sourced from the RuATD Shared Task 2022 (Shamardina et al., 2022) devoted to artificial text detection in Russian. Shamardina et al. (2022) gathered a vast human and machine-generated corpora from various text generators. However, these generators are either task-specific or domain-specific.\\nWe leveraged their human-written texts collected from publicly available resources and re-generated the machine-authored data using the open-domain state-of-the-art multilingual LLMs. In particular, for the construction of human-written data, the task organizers used the following sources: (1) diachronic sub-corpora of the Russian National Corpus20 , which covers three historical periods of the society and the Modern Russian language (\\\"pre-Soviet\\\", \\\"Soviet\\\", and \\\"post-Soviet\\\"); (2) several social media platforms;\\n(3) top-100 most viewed Russian Wikipedia pages spanning the period of 2016-2021 according to the PageViews statistics; (4) news articles from the Taiga corpus (Shavrina and Shapovalova, 2017) and the \\\"corus\\\" library 21 ; (5) a corpus of digitalized personal diaries \\\"Prozhito\\\" written during the 20th century (Melnichenko and Tyshkevich, 2017); (6) government documents from the RuREBus Shared Task (Ivanin et al., 2020).\\nUrdu News is derived from Urdu News Data 1M -a collection of one million news articles from four distinct categories: Business & Economics, Science & Technology, Entertainment, and Sports. These articles were gathered from four reputable news agencies in Pakistan (Hussain et al., 2021). Each entry in this dataset includes a headline, a category, and a news article text. To ensure the data balance over four categories, we randomly sampled 750 news articles from each, resulting in 3,000 examples in total. Using the headlines as prompts, we generated the content of artificial news articles.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.3 LLM Generation\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_9\"\n ],\n \"text\": \"Prompt Diversity We paid attention to prompt diversity, using multiple (2-8) prompts for each domain-generator combination in English, with the aim to produce diverse outputs that are more aligned with divergent generations in real-world application scenarios. See Table 7 for detailed statistics about the prompts.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Prompts of PeerRead\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_10\"\n ],\n \"text\": \"• Please write a peer review for the paper + title;\\n• Write a peer review by first describing what problem or question this paper addresses, then strengths and weaknesses, for the paper + title;\\n• Please write a peer review for the paper + title, its main content is as below: + abstract;\\n• Write a peer review by first describing what problem or question this paper addresses, then strengths and weaknesses, for the paper + title, its main content is as follows: + abstract.\\nHyper-Parameter Values for the Generators Table 8 shows the values of the hyper-parameters we used for the various generators. In general, we followed the default setting, except for the length of new generations in order to satisfy the minimum character length of 1,000. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.4 N-gram Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_12\",\n \"tab_13\"\n ],\n \"text\": \"Table 9 shows statistics about the number of unique uni-grams (word types) and bi-grams of human-written and machine-generated texts (English). Table 10 shows the number of per-document unique uni-grams (word types) and bi-grams of human-written and machine generated texts (English). \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.2 Computation Resources and Cost\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We spent $600 on calls of OpenAI APIs for ChatGPT and davinci-003 generations, $40 on calls to GPTZero. We spent about 2,500 GPU hours on Dolly-v2 and BLOOMz generation.\\nC Results: Same-Generator, Cross-Domain \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"E Results: Multilingual Evaluation\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_18\",\n \"tab_19\"\n ],\n \"text\": \"Table 16 shows detailed results for cross-language experiments based on XLM-R on the test sets across different languages generated by ChatGPT. Table 17 shows the same results, but using davinci-003 as a generator (instead of ChatGPT). \"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"F Results: Impact of Text Length\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Figure 3 shows detailed results from experiments to study the impact of text length on detection accuracy, over arXiv and Reddit generated by ChatGPT, davinci, and Cohere. We can see that as the character length decreasing from 1000 to 125, the F1-score with respect to machine-generated text decreases for all subsets, demonstrating the negative impact of short text on detection performance. Henry was an experienced jouster, but Montgomery was a newcomer to the sport, and unfortunately, he was no match for the King's lance. During one of their runs, Henry's lance struck Montgomery's helmet, shattering it and sending a jagged piece of wood into his eye and brain. The young man was rushed to a nearby hospital, but it was too late. He died just a few days later, leaving behind a wife and children. As for Henry, he was devastated by the accident and reportedly went into a deep depression. He blamed himself for Montgomery's death and was haunted by guilt for the rest of his life. He even imposed a penance on himself, vowing to fast and do charitable works for the rest of his days. The incident also led to changes in the sport of jousting. After Montgomery's death, many rules were put in place to make the sport safer, such as requiring jousters to wear full face and head protection. Overall, it was a tragic event that had a lasting impact on the sport and on the King who inadvertently caused the death of his opponent.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"H Examples of M4\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Language English LLM Generator CHATGPT Domain Arxiv\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Title Polymer Quantum Mechanics and its Continuum Limit\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Human A rather non-standard quantum representation of the canonical commutation relations of quantum mechanics systems, known as the polymer representation has gained some attention in recent years, due to its possible relation with Planck scale physics. In particular, this approach has been followed in a symmetric sector of loop quantum gravity known as loop quantum cosmology.\\nHere we explore different aspects of the relation between the ordinary Schroedinger theory and the polymer description. The paper has two parts. In the first one, we derive the polymer quantum mechanics starting from the ordinary Schroedinger theory and show that the polymer description arises as an appropriate limit. In the second part we consider the continuum limit of this theory, namely, the reverse process in which one starts from the discrete theory and tries to recover back the ordinary Schroedinger quantum mechanics. We consider several examples of interest, including the harmonic oscillator, the free particle and a simple cosmological model.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Field Content\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Prompt Generate a 150-220-word abstract for work with title: Polymer Quantum Mechanics and its Continuum Limit. Describe the results obtained, problem work attempts to solve, and the key ideas and methodology in a formal academic and scientific writing voice. Use the first plural person form. Use active voice.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Machine\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this paper, we investigate the continuum limit of polymer quantum mechanics. The aim of our work is to better understand the quantum behavior of polymers and its relationship to the classical theory. Our research attempts to clarify the nature of the continuum limit and determine the correctness of the approximations made. Specifically, we study the connectivity of the polymer strands. We find that a continuous path can be defined in the polymer configuration space, which is essential for the polymer to display continuum behavior. We show that the approximation of continuous randomness in the spatial variable can certainly be used when the polymer is stretched out over a significant distance. For a coil-like polymer, our results demonstrate that its behavior is non-classical. Our methodology involves a combination of intuition, formalism, and numerical simulations. We evaluate a path integral representation of polymer quantum mechanics in the continuum limit, and use numerical simulations to support our results. The ideas proposed in this work are relevant to the understanding of the thermodynamic properties of polymers, as well as to quantum computing and information. The findings of this work present a theoretical framework for the investigation of other polymer properties that are still poorly understood.\\nLanguage English LLM Generator DAVINCI-003 Domain Arxiv\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Title\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Effects of Dirac sea on pion propagation in asymmetric nuclear matter\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Human\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We study pion propagation in asymmetric nuclear matter (ANM). One of the interesting consequences of pion propagation in ANM is the mode splitting for the different charged states of pions. First we describe the pion-nucleon dynamics using the non-chiral model where one starts with pseudoscalar (PS) πN coupling and the pseudovector (PV) representation is obtained via suitable non-linear field transformations. For both of these cases the effect of the Dirac sea is estimated. Subsequently, we present results using the chiral effective Lagrangian where the short-distance behavior (Dirac vacuum) is included by re-defining the field parameters as done in the modern effective field theory approach developed recently. The results are compared with the previous calculations for the case of symmetric nuclear matter (SNM). Closed form analytical results are presented for the effective pion masses and dispersion relations by making hard nucleon loop (HNL) approximation and suitable density expansion.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Prompt\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Rephrase the abstract of an article with title 'Effects of Dirac sea on pion propagation in asymmetric nuclear matter'. 'We study pion propagation in asymmetric nuclear matter (ANM). One of the interesting consequences of pion propagation in ANM is the mode splitting for the different charged states of pions. First we describe the pion-nucleon dynamics using the non-chiral model where one starts with pseudoscalar (PS) πN coupling and the pseudovector (PV) representation is obtained via suitable non-linear field transformations. For both of these cases the effect of the Dirac sea is estimated. Subsequently, we present results using the chiral effective Lagrangian where the short-distance behavior (Dirac vacuum) is included by re-defining the field parameters as done in the modern effective field theory approach developed recently. The results are compared with the previous calculations for the case of symmetric nuclear matter (SNM). Closed form analytical results are presented for the effective pion masses and dispersion relations by making hard nucleon loop (HNL) approximation and suitable density expansion.'\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Machine\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The propagation of pions in asymmetric nuclear matter (ANM) is studied in this article. We consider two approaches: a non-chiral model where the pseudoscalar/pseudovector coupling is determined using non-linear field transformations, and a chiral effective Lagrangian incorporating the short-distance behaviors known as the Dirac vacuum. In both cases, the consequences from these approaches on the mode splitting of the charged pions are evaluated. Using a hard nucleon loop approximation, accompanied by parameters deduced from a density expansion, closedform analytical results for the effective pion masses and dispersion relations are presented in comparison to calculations for symmetric nuclear matter.\\nLanguage English LLM Generator COHERE Domain Wikipedia Title Shed (deity)\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Human\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Shed was a deity from ancient Egyptian religion. Popularly called \\\"the Savior\\\", he is first recorded after the Amarna Period. Representing the concept of salvation, Shed is identified with Horus, particularly Horus the Child.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Field Content\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Rather than have formal worship in a temple or as an official cult, he appears to have been a god that ordinary Egyptians looked to save them from illness, misfortune or danger. He is shown on the Metternich Stela as vanquishing danger in the form of a serpent, a scorpion and a crocodile. The rise of \\\"Savior\\\" names in personal piety during the Amarna period has been interpreted as the popular response of ordinary people to the attempts by Akhenaten to proscribe the ancient religion of Egypt. Shed has also been viewed as a form of the Canaanite god Resheph. Shed can be depicted as a young prince overcoming snakes, lions and crocodiles. Shed has been viewed as a form of savior, a helper for those in need when state authority or the kingś help is wanting. The increased reliance on divine assistance could even extend to saving a person from the Underworld, even to providing a substitute, and lengthening a personś time in this world.\\nIn the New Kingdom Shed \\\"the savior\\\" is addressed on countless stelae by people or praising him for help.References Egyptian gods Savior gods Prompt Write a Wikipedia article with the title \\\"Shed (deity)\\\", the article should at least have 250 words Machine Shed (deity) Shed is a deity in the sqlrunen mythos. Shed is described as a giantess, a daughter of the primordial mother, Nott. In most traditions, Shed is a single entity, but in some traditions, Shed is a collective of giantesses. Shed is the mother of the first race of giants, the primordial beings of chaos and destruction. Shed mothered the giants through her mating with the primordial god, Odin. In some traditions, the giantess Gunnlod or Hyr is described as the mother of the giants, while Shed is described as the mother of the gods. Shed is the spouse of the primordial god, Odin. Odin and Shed had many offspring, including the gods, the giants, the Lamonts, the Maquas, the Frost Giants, and the Ice Giants. Shed is described as a beautiful giantess, with golden hair, a beautiful face, and a beautiful body. She is said to have worn a long golden gown and a gold crown. She is described as the most beautiful of all the giantesses. Shed is the goddess of fertility and reproduction. She is also the goddess of the harvest. In some traditions, Shed is also the goddess of war and death. Shed is the mother of the gods. She gave birth to the gods through her union with the primordial god, Odin. The gods were born from her body, and she nourished them with her milk. Shed is the protector of the gods. She is said to have fought alongside the gods in their war against the giants. She is also said to have fought against the dragons and the serpents. Shed is the goddess of the hunt. She is said to have accompanied the gods on their hunting expeditions. She is also said to have provided them with food and shelter.\\nLanguage English LLM Generator DAVINCI-003 Domain Wikihow\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Title How to Buy Virtual Console Games for Nintendo Wii\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Human They are about $20 a card. Or, if you want to just buy points with your credit card, Skip down to the section, With a Credit Card., You don t want people tapping into your Wii and stealing your games, do you? On the back of the Wii Points Card, this same code can be found by scratching out the grey section with a coin or similar. You will be informed if you entered an invalid (wrong) code. Go back to the main Shop Channel menu now. Nintendo, from time to time, offers freebies via a special channel called \\\"WiiWare\\\". There are over 50 games to list, so be patient if the option you selected is taking a while to load. Don t forget to say you want to download to your Wii, unless you want your classics transportable, for example if you plan to demo the classics to a friend with a Wii. Nintendo added an animation to this process that shows Mario or Luigi collecting coins. This is why you want a secure connection!, Note that a tax is added to your purchase to comply with federal law. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Human\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Especially check that the fishing line can be placed in a location which doesn t get too much in the way of your normal use (heavy users of the mouse buttons or touchpad might have problems here), and that there are locations more or less in the center of the laptop where gluing the Velcro will not affect the use of the laptop (that no access doors or ventilation holes are sealed). , Try to make the line of the two strips be as perpendicular as possible to the axis of the display hinges.\\n, This can take 24 hours or even more. See the instructions for the particular epoxy you have used. Don t confuse the setting time of the epoxy (which can be as quick as 5 minutes) with the time until it reaches maximum strength.\\n, Actually, what is more important is that the length of the hooked strip totally utilize the glued strips when the laptop display is open. Do not use the Velcro with an open display before the glue has hardened to its maximum strength. , Laptops with large or heavy displays may require gluing multiple strips of Velcro, or wider and/or stronger Velcro (Velcro comes in a range of sizes and bonding strengths). The laptop in the images has a 12.1\\\" display. , I used an \\\"improved blood knot\\\" for this. Instructions can be found in various places on the Internet. , If it is a bother, make the adhesive on the outside less sticky by touching it repeatedly.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Field Content\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Prompt\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Please, generate wikihow article with length above 1000 characters from title 'How to Workaround Damaged Laptop Display Hinges' and headline 'Read the entire article, and check that your specific laptop can be repaired in the way presented., Cut two strips of the fuzzy Velcro and round the corners (rounding the corners will make it less likely that the Velcro will detach after gluing)., Read the instructions for the epoxy glue; mix it and use it to glue the two strips of fuzzy Velcro, one on the back of the display, and one on the bottom of the laptop, such that the two strips are co-linear (in other words, the alignment of one strip, when extended, coincides with the alignment of the other strip)., Wait for the epoxy glue to harden fully.,Close the laptop display, and cut a strip of the hooked Velcro more or less to the length required when sticking it to the two glued strips in their entirety., After the glue has hardened to maximum strength, check that the Velcro can now prevent the display from falling into the closed position., Take the fishing line, and tie it into a loop which exactly encircles the laptop when the display is open to the working position., Check that the fishing line loop is strong enough to prevent the display from falling into the fully open position (fishing line comes in a variety of strengths, use a relatively strong line which will not stretch under the load)., Glue a small strip of two-sided foam tape (like used for mounting things) on the edge of the laptop where the fishing line should rest., Optionally, you can use a plastic bag closure which has been notched (see images), in order to more easily adjust the exact size of the fishing line loop (and therefore, the exact working position of the laptop display).'\\nMachine Have you ever experienced a damaged laptop display hinge? If so, you know how frustrating it can be when your display doesn't stay in the correct position. Fortunately, there's a workaround that you can use to fix this issue without having to replace the entire laptop hinge. This wikihow article will guide you through the process of using Velcro and fishing line to strengthen your laptop display hinges, keeping your display in the right position while you work.\\n1. Read the entire article, and check that your specific laptop can be repaired in the way presented. Before getting started with any repairs, it's important to read the entire article to make sure that this method will work for your laptop model. Some laptops may have different hinge designs that won't be compatible with the Velcro and fishing line method we're using here.\\n2. Cut two strips of the fuzzy Velcro and round the corners.\\nTo begin the repair process, you'll need to cut two strips of fuzzy Velcro and round the corners of each strip. This will make it less likely that the Velcro will detach after gluing.\\n3. Glue the Velcro to the laptop display and bottom.\\nNext, mix your epoxy glue according to the instructions, and use it to glue the two strips of fuzzy Velcro -one on the back of the display and one on the bottom of the laptop. Make sure that the two strips are co-linear, meaning that the alignment of one strip, when extended, coincides with the alignment of the other strip. 4. Wait for the glue to dry. After gluing the Velcro to the laptop, you'll need to wait for the epoxy glue to harden fully. This may take several hours, so be patient and allow the glue to dry completely. 5. Cut a strip of hooked Velcro to length.\\nOnce the glue has dried, close your laptop display and cut a strip of hooked Velcro more or less to the length required when sticking it to the two glued strips in their entirety. 6. Check that the Velcro can prevent the display from falling into the closed position.\\nAfter the glue has hardened to maximum strength, check that the Velcro can now prevent the display from falling into the closed position. Abstract Integrating text and knowledge into a unified semantic space has attracted significant research interests recently. However, the ambiguity in the common space remains a challenge, namely that the same mention phrase usually refers to various entities. In this paper, to deal with the ambiguity of entity mentions, we propose a novel Multi-Prototype Mention Embedding model, which learns multiple sense embeddings for each mention by jointly modeling words from textual contexts and entities derived from a knowledge base. In addition, we further design an efficient language model based approach to disambiguate each mention to a specific sense.\\nIn experiments, both qualitative and quantitative analysis demonstrate the high quality of the word, entity and multi-prototype mention embeddings. Using entity linking as a study case, we apply our disambiguation method as well as the multi-prototype mention embeddings on the benchmark dataset, and achieve state-of-the-art performance.\\nHuman -Strengths: Good ideas, simple neural learning, interesting performance (altough not striking) and finally large set of applications. -Weaknesses: amount of novel content. Clarity in some sections. The paper presents a neural learning method for entity disambiguation and linking. It introduces a good idea to integrate entity, mention and sense modeling within the smame neural language modeling technique. The simple training procedure connected with the modeling allows to support a large set of application. The paper is clear formally, but the discussion is not always at the same level of the technical ideas. The empirical evaluation is good although not striking improvements of the performance are reported. Although it seems an extension of (Yamada et al., CoNLL 2016), it adds novel ideas and it is of a releant interest.\\nThe weaker points of the paper are: -The prose is not always clear. I found Section 3 not as clear. Some details of Figure 2 are not explained and the terminology is somehow redundant: for example, why do you refer to the dictionary of mentions? or the dictionary of entity-mention pairs? are these different from text anchors and types for annotated text anchors? -Tha paper is quite close in nature to Yamada et al., 2016) and the authors should at least outline the differences.\\nOne general observation on the current version is: The paper tests the Multiple Embedding model against entity linking/disambiguation tasks. However, word embeddings are not only used to model such tasks, but also some processes not directly depending on entities of the KB, e.g. parsing, coreference or semantic role labeling. The authors should show that the word embeddings provided by the proposed MPME method are not weaker wrt to simpler wordspaces in such other semantic tasks, i.e. those involving directly entity mentions. I did read the author's response.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Prompt\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Write a peer review by first describing what problem or question this paper addresses, then strengths and weaknesses, for the paper: title\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Machine\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The paper \\\"Bridge Text and Knowledge by Learning Multi-Prototype Entity Mention Embedding\\\" proposes a new method for utilizing external knowledge sources to improve entity detection and disambiguation in natural language processing tasks. The authors specifically address the issue of how to effectively incorporate large-scale knowledge bases such as Wikipedia into existing machine learning models. The strengths of this paper are its clear and concise explanation of the proposed methodology, use of established evaluation metrics, and consideration of potential limitations and future directions.\\nThe authors provide a detailed description of their multi-prototype entity mention embedding (MPEME) approach, which includes the use of different prototypes to represent entities with multiple meanings in Wikipedia. They evaluate their approach on two established datasets and demonstrate improved performance compared to baseline models. Additionally, the authors note potential limitations of their approach such as the requirement for large knowledge bases and potential bias in the selection of prototypes. One weakness of this paper is that it assumes knowledge bases such as Wikipedia are comprehensive and reliable sources of information. There is a growing body of research showing that knowledge bases can contain errors or biases, which may affect the accuracy of this approach. The authors acknowledge this issue but do not offer a solution, which limits the generalizability of their findings. Additionally, the paper provides limited discussion on how this approach can be scaled up for larger knowledge bases, which is an important consideration for practical applications.\\nOverall, the paper provides a valuable contribution to the field by addressing an important challenge of incorporating external knowledge sources into NLP models. The proposed MPEME method shows promise for improving entity detection and disambiguation, but further research is needed to fully explore its potential and limitations. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Prompt\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Tulislah suatu artikel berita yang berjudul \\\"Ahmad Dhani Klaim Tak Punya Informasi soal Sri Bintang\\\". Artikelnya harus mengandung paling tidak 300 kata dalam bahasa Indonesia.\\nTambahkan \\\"Jakarta, CNN Indonesia -\\\" di awal artikel.\\nMachine Jakarta, CNN Indonesia -Ahmad Dhani, personel grup band Dewa 19, mengaku tak memiliki informasi mengenai sosok Sri Bintang yang tengah ramai dibicarakan di media sosial. Seperti diketahui, Sri Bintang menjadi perbincangan publik setelah dirinya mengunggah video yang menyebutkan bahwa personel band Dewa 19, Ahmad Dhani, pernah menyebut Sri Bintang sebagai sosok yang berperan dalam pembuatannya dari awal. Klaim tersebut kemudian dibantah oleh Ahmad Dhani melalui sebuah video yang diunggahnya di akun media sosialnya. \\\"Saya sendiri nggak tahu siapa Sri Bintang itu. Jadi saya gak bisa bicara, saya gak bisa komen, kenapa orang lain mau ngomongin saya tentang Sri Bintang,\\\" ujar Ahmad Dhani dalam video tersebut. Ahmad Dhani juga mengatakan bahwa beberapa orang yang mengklaim tahu tentang perjalanannya dalam industri musik sebenarnya tidak memiliki informasi yang tepat. \\\"Saya sekali lagi ingin tegaskan bahwa ada beberapa orang yang bicara itu nggak tahu apa yang mereka bicarakan. Intinya, kita perlu memperhatikan orang yang kita ngomongin,\\\" tegasnya. Kendati begitu, Ahmad Dhani mengungkapkan bahwa dirinya berharap Sri Bintang bisa berkarya dan memperlihatkan kemampuan yang dimilikinya di industri musik tanah air. Ahmad Dhani juga berharap bahwa sosok Sri Bintang tidak menjadi obyek perdebatan yang tak berkesudahan di antara para penggemar musik dan publik di Indonesia. \\\"Setidaknya semua artis harus diberikan kesempatan yang sama untuk berkarya. Saya berharap Sri Bintang bisa berkarya dengan baik dan melebihi ekspektasi dari semua orang,\\\" harapnya. Meski begitu, klaim dari Sri Bintang mengenai Ahmad Dhani telah menjadi perhatian banyak pengguna media sosial. Banyak netizen yang menyatakan bahwa hal tersebut merupakan sebuah bukti bahwa industri musik di Indonesia selalu menarik perhatian masyarakat dan penuh dengan kontroversi. Maka dari itu, publik diharapkan mampu menghormati karya dari masing-masing musisi dan menghindari terjadinya perdebatan yang tak perlu di media sosial. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Field\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We thank the anonymous reviewers and the program committee chairs for their very helpful and insightful comments, which have helped us improve the paper.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ethics and Broader Impact\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Below, we discuss some potential ethical concerns about the present work.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Field Content prompt\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Write a peer review by first describing what problem or question this paper addresses, then strengths and weaknesses, for the paper title, its main content is as below: abstract Machine In this paper, the authors tackle the problem of ambiguity in entity mentions when integrating text and knowledge into a unified semantic space. Specifically, they propose a Multi-Prototype Mention Embedding model that learns multiple sense embeddings for each mention by jointly modeling words from textual contexts and entities from a knowledge base. They also design an efficient language model-based approach to disambiguate each mention to a specific sense.\\nThe authors evaluate their proposed model on a benchmark dataset for entity linking and report state-of-the-art performance.\\nOverall, the paper presents a well-motivated, well-executed, and well-evaluated solution to a relevant research problem. The main contributions of the paper are the Multi-Prototype Mention Embedding model and the efficient language model-based disambiguation approach.\\nThe authors clearly describe the model and how it addresses the problem of mention ambiguity.\\nThe evaluation of the model is thorough, including both qualitative and quantitative analyses.\\nOne of the strengths of the paper is the use of a benchmark dataset, which allows for fair comparison with previous approaches. Additionally, the authors provide detailed explanations of the implementation and evaluation of their model, making the paper easily replicable. The experiments demonstrate the high quality of the word, entity, and multi-prototype mention embeddings produced by the proposed model. However, there are a few weaknesses in the paper. While the evaluation of the proposed method is thorough and convincing, it would have been beneficial if the authors had provided more in-depth analysis of the results to show exactly where and why their method outperforms previous approaches. Additionally, the paper would have benefited from a more detailed discussion of the limitations and potential future directions of the proposed method.\\nOverall, the paper is a strong contribution to the field of entity linking and the integration of text and knowledge into a unified semantic space. The proposed Multi-Prototype Mention Embedding model and the efficient language model-based disambiguation approach provide a promising solution to the problem of mention ambiguity. Machine This method has been used for years by many people who suffer from foot-nail fungus. It works well because both salt and vinegar are natural disinfectants that kill bacteria and other pathogens; they also help loosen the hardened layers around infected areas. The warm water helps dissolve the salt crystals which then act like an abrasive cleansing agent against the affected toe. Vinegar contains acetic acid which kills most types of mold spores including those responsible for athlete's foot. This treatment will not work immediately but over time should remove the infection causing the nail to become detached naturally. If there... is no improvement within three months consult a doctor.... To prepare the solution mix together equal parts of baking soda and hydrogen peroxide using enough liquid soap to create bubbles. Add some drops of tea tree oil Mix everything together in a bowl creating thick paste-like consistency. Apply directly onto the infected nail. Cover with bandage overnight. Remove next morning. Rinse away with cool water. Do twice each day for at least 2 weeks. You may need longer depending upon severity of infection. Keep doing this even though new growth appears underneath old nail. Be patient!\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2024-03-10"},"doi":{"kind":"string","value":"10.18653/v1/2021.acl-long.565"},"references":{"kind":"list like","value":[{"authors":"Elizabeth Clark; Tal August; Sofia Serrano; Nikita Haduong; Suchin Gururangan; Noah A Smith","journal":"Association for Computational Linguistics","ref_id":"b0","title":"All that's 'human' is not gold: Evaluating human evaluation of generated text","year":"2021"},{"authors":"Kevin Clark; Minh-Thang Luong; Quoc V Le; Christopher D Manning","journal":"","ref_id":"b1","title":"Electra: Pre-training text encoders as discriminators rather than generators","year":"2020"},{"authors":"Alexis Conneau; Kartikay Khandelwal; Naman Goyal; Vishrav Chaudhary; Guillaume Wenzek; Francisco Guzmán; Edouard Grave; Myle Ott; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b2","title":"Unsupervised cross-lingual representation learning at scale","year":"2019"},{"authors":"Mike Conover; Matt Hayes; Ankit Mathur; Jianwei Xie; Jun Wan; Sam Shah; Ali Ghodsi; Patrick Wendell; Matei Zaharia; Reynold Xin","journal":"","ref_id":"b3","title":"Free dolly: Introducing the world's first truly open instructiontuned llm","year":"2023"},{"authors":"Angela Fan; Yacine Jernite; Ethan Perez; David Grangier; Jason Weston; Michael Auli","journal":"Association for Computational Linguistics","ref_id":"b4","title":"ELI5: long form question answering","year":"2019-07-28"},{"authors":"Sebastian Gehrmann; Hendrik Strobelt; Alexander Rush","journal":"Association for Computational Linguistics","ref_id":"b5","title":"GLTR: Statistical detection and visualization of generated text","year":"2019"},{"authors":"Jesse Graham; Jonathan Haidt; Sena Koleva; Matt Motyl; Ravi Iyer; Sean Wojcik; Peter Ditto","journal":"Advances in Experimental Social Psychology","ref_id":"b6","title":"Moral foundations theory: The pragmatic validity of moral pluralism","year":"2012"},{"authors":"Biyang Guo; Xin Zhang; Ziyuan Wang; Minqi Jiang; Jinran Nie; Yuxuan Ding; Jianwei Yue; Yupeng Wu","journal":"","ref_id":"b7","title":"How close is chatgpt to human experts? comparison corpus, evaluation, and detection","year":"2023"},{"authors":"Xuanli He; Qiongkai Xu; Lingjuan Lyu; Fangzhao Wu; Chenguang Wang","journal":"AAAI Press","ref_id":"b8","title":"Protecting intellectual property of language generation apis with lexical watermark","year":"2022-02-22"},{"authors":"Jeppe Benjamin D Horne; Sibel Nørregaard; Adali","journal":"ACM Transactions on Intelligent Systems and Technology (TIST)","ref_id":"b9","title":"Robust fake news detection over time and attack","year":"2019"},{"authors":"Khalid Hussain; Nimra Mughal; Irfan Ali; Saif Hassan; Sher Muhammad Daudpota","journal":"","ref_id":"b10","title":"","year":"2021"},{"authors":" Va Ivanin; El Artemova; Batura; Vv Ivanov; E V Vv Sarkisyan; Tutubalina; Smurov","journal":"","ref_id":"b11","title":"RuREbus-2020 shared task: Russian relation extraction for business","year":"2020"},{"authors":"Dongyeop Kang; Waleed Ammar; Bhavana Dalvi; Madeleine Van Zuylen; Sebastian Kohlmeier; Eduard Hovy; Roy Schwartz","journal":"Association for Computational Linguistics","ref_id":"b12","title":"A dataset of peer reviews (PeerRead): Collection, insights and NLP applications","year":"2018"},{"authors":"John Kirchenbauer; Jonas Geiping; Yuxin Wen; Jonathan Katz; Ian Miers; Tom Goldstein","journal":"","ref_id":"b13","title":"A watermark for large language models","year":"2023"},{"authors":"Mahnaz Koupaee; William Yang; Wang ","journal":"","ref_id":"b14","title":"Wikihow: A large scale text summarization dataset","year":"2018"},{"authors":"Kalpesh Krishna; Yapei Chang; John Wieting; Mohit Iyyer","journal":"Association for Computational Linguistics","ref_id":"b15","title":"RankGen: Improving text generation with large ranking models","year":"2022"},{"authors":"Jenny S Li; John V Monaco; Li-Chiou Chen; Charles C Tappert","journal":"IEEE Computer Society","ref_id":"b16","title":"Authorship authentication using short messages from social networking sites","year":"2014-11-05"},{"authors":"Weixin Liang; Mert Yuksekgonul; Yining Mao; Eric Wu; James Zou","journal":"","ref_id":"b17","title":"Gpt detectors are biased against non-native english writers","year":"2023"},{"authors":"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov","journal":"","ref_id":"b18","title":"Roberta: A robustly optimized bert pretraining approach","year":"2019"},{"authors":"Dominik Macko; Robert Moro; Adaku Uchendu; Jason Lucas; Michiharu Yamashita; Matúš Pikuliak; Ivan Srba; Thai Le; Dongwon Lee; Jakub Simko; Maria Bielikova","journal":"Association for Computational Linguistics","ref_id":"b19","title":"MULTITuDE: Large-scale multilingual machine-generated text detection benchmark","year":"2023"},{"authors":"Michail Melnichenko; Natalia Tyshkevich","journal":"","ref_id":"b20","title":"Prozhito from manuscript to corpus","year":"2017"},{"authors":"Eric Mitchell; Yoonho Lee; Alexander Khazatsky; Christopher D Manning; Chelsea Finn","journal":"","ref_id":"b21","title":"Detectgpt: Zero-shot machine-generated text detection using probability curvature","year":"2023"},{"authors":"Niklas Muennighoff; Thomas Wang; Lintang Sutawika; Adam Roberts; Stella Biderman; Teven Le Scao; M Saiful; Sheng Bari; Zheng Xin Shen; Hailey Yong; Xiangru Schoelkopf; Dragomir Tang; Alham Radev; Khalid Fikri Aji; Samuel Almubarak; Zaid Albanie; Albert Alyafeai; Edward Webson; Colin Raff; Raffel","journal":"","ref_id":"b22","title":"Crosslingual generalization through multitask finetuning","year":"2022"},{"authors":"Marco Tulio Ribeiro; Sameer Singh; Carlos Guestrin","journal":"","ref_id":"b23","title":"why should i trust you?\" explaining the predictions of any classifier","year":"2016"},{"authors":"Aounon Vinu Sankar Sadasivan; Sriram Kumar; Wenxiao Balasubramanian; Soheil Wang; Feizi","journal":"","ref_id":"b24","title":"Can ai-generated text be reliably detected?","year":"2023"},{"authors":"Neha Sengupta; Sunil Kumar Sahu; Bokang Jia; Satheesh Katipomu; Haonan Li; Fajri Koto; William Marshall; Gurpreet Gosal; Cynthia Liu; Zhiming Chen; Osama Mohammed Afzal; Samta Kamboj; Onkar Pandit; Rahul Pal; Lalit Pradhan; Zain Muhammad Mujahid; Massa Baali; Xudong Han; Mahmoud Sondos; Alham Bsharat; Zhiqiang Fikri Aji; Zhengzhong Shen; Natalia Liu; Joel Vassilieva; Andy Hestness; Andrew Hock; Jonathan Feldman; Andrew Lee; Hector Jackson; Preslav Xuguang Ren; Timothy Nakov; Eric Baldwin; Xing","journal":"","ref_id":"b25","title":"Jais and Jais-chat: Arabic-centric foundation and instruction-tuned open generative large language models","year":"2023"},{"authors":"Tatiana Shamardina; Vladislav Mikhailov; Daniil Cherniavskii; Alena Fenogenova; Marat Saidov; Anastasiya Valeeva; Tatiana Shavrina; Ivan Smurov; Elena Tutubalina; Ekaterina Artemova","journal":"","ref_id":"b26","title":"Findings of the the ruatd shared task 2022 on artificial text detection in russian","year":"2022"},{"authors":"Tatiana Shavrina; Olga Shapovalova","journal":"","ref_id":"b27","title":"To the methodology of corpus construction for machine learning: \"taiga\" syntax tree corpus and parser","year":"2017"},{"authors":"Rexhep Shijaku; Ercan Canhasi","journal":"","ref_id":"b28","title":"Chatgpt generated text detection","year":"2023"},{"authors":"Sebastian Szyller; Gul Buse; Samuel Atli; N Marchal; Asokan","journal":"ACM","ref_id":"b29","title":"DAWN: dynamic adversarial watermarking of neural networks","year":"2021-10-20"},{"authors":"Ruixiang Tang; Yu-Neng Chuang; Xia Hu","journal":"","ref_id":"b30","title":"The science of detecting llm-generated texts","year":"2023"},{"authors":"Hugo Touvron; Louis Martin; Kevin Stone; Peter Albert; Amjad Almahairi; Yasmine Babaei; Nikolay Bashlykov; Soumya Batra; Prajjwal Bhargava; Shruti Bhosale; Dan Bikel; Lukas Blecher; Cristian Canton Ferrer; Moya Chen; Guillem Cucurull; David Esiobu; Jude Fernandes; Jeremy Fu; Wenyin Fu; Brian Fuller; Cynthia Gao; Vedanuj Goswami; Naman Goyal; Anthony Hartshorn; Saghar Hosseini; Rui Hou; Hakan Inan; Marcin Kardas; Viktor Kerkez; Madian Khabsa; Isabel Kloumann; Artem Korenev; Punit Singh Koura; Marie-Anne Lachaux; Thibaut Lavril; Jenya Lee; Diana Liskovich; Yinghai Lu; Yuning Mao; Xavier Martinet; Todor Mihaylov; Pushkar Mishra; Igor Molybog; Yixin Nie; Andrew Poulton; Jeremy Reizenstein; Rashi Rungta; Kalyan Saladi; Alan Schelten; Ruan Silva; Eric Michael Smith; Ranjan Subramanian; Ellen Xiaoqing; Binh Tan; Ross Tang; Adina Taylor; Jian Williams; Puxin Xiang Kuan; Zheng Xu; Iliyan Yan; Yuchen Zarov; Angela Zhang; Melanie Fan; Sharan Kambadur; Aurelien Narang; Robert Rodriguez; Sergey Stojnic; Thomas Edunov; Scialom","journal":"","ref_id":"b31","title":"LLaMA 2: Open foundation and finetuned chat models","year":"2023"},{"authors":"Adaku Uchendu; Zeyu Ma; Thai Le; Rui Zhang; Dongwon Lee","journal":"Association for Computational Linguistics","ref_id":"b32","title":"TURINGBENCH: A benchmark environment for Turing test in the age of neural text generation","year":"2021"},{"authors":"Yuxia Wang; Jonibek Mansurov; Petar Ivanov; Jinyan Su; Artem Shelmanov; Akim Tsvigun; Chenxi Whitehouse; Osama Mohammed Afzal; Tarek Mahmoud; Giovanni Puccetti; Thomas Arnold; Alham Fikri Aji; Nizar Habash; Iryna Gurevych; Preslav Nakov","journal":"","ref_id":"b33","title":"SemEval-2024 task 8: Multidomain, multimodal and multilingual machine-generated text detection","year":"2024"},{"authors":"Rowan Zellers; Ari Holtzman; Hannah Rashkin; Yonatan Bisk; Ali Farhadi; Franziska Roesner; Yejin Choi","journal":"","ref_id":"b34","title":"Defending against neural fake news","year":"2019-12-08"},{"authors":"Xuandong Zhao; Yu-Xiang Wang; Lei Li","journal":"","ref_id":"b35","title":"Protecting language generation models via invisible watermarking","year":"2023"},{"authors":"A ","journal":"","ref_id":"b36","title":"2 Corpora in Other Languages Arabic Wikipedia","year":""},{"authors":"Bulgarian True; & ","journal":"","ref_id":"b37","title":"Fake News is sampled from the Hack the Fake News datathon 17 organized in 2017 by the Data Science Society in Bulgaria","year":""},{"authors":"","journal":"","ref_id":"b38","title":"), precision (Prec), recall, and F1 score","year":""},{"authors":"","journal":"Acc Prec Recall F1 Acc Prec Recall F1 RoBERTa(base) ChatGPT","ref_id":"b39","title":"8 Table 13: Same-generator, cross-domain experiments: training on a single domain of davinci-003 vs. human, and testing across domains. Evaluation accuracy (Acc), precision (Prec), recall, and F1 scores (in %) with respect to machine generations across four detectors. Test → ChatGPT davinci Cohere BLOOMz Train ↓ Acc Prec Recall F1 Acc Prec Recall F1","year":""},{"authors":"","journal":"","ref_id":"b40","title":"0 Table 15: Same-domain, cross-generator experiments: training and testing on Wikipedia (single machine-text generator vs. human). evaluation accuracy (Acc), precision (Prec), recall, and F1 score","year":""}],"string":"[\n {\n \"authors\": \"Elizabeth Clark; Tal August; Sofia Serrano; Nikita Haduong; Suchin Gururangan; Noah A Smith\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b0\",\n \"title\": \"All that's 'human' is not gold: Evaluating human evaluation of generated text\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Kevin Clark; Minh-Thang Luong; Quoc V Le; Christopher D Manning\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"Electra: Pre-training text encoders as discriminators rather than generators\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Alexis Conneau; Kartikay Khandelwal; Naman Goyal; Vishrav Chaudhary; Guillaume Wenzek; Francisco Guzmán; Edouard Grave; Myle Ott; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Unsupervised cross-lingual representation learning at scale\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Mike Conover; Matt Hayes; Ankit Mathur; Jianwei Xie; Jun Wan; Sam Shah; Ali Ghodsi; Patrick Wendell; Matei Zaharia; Reynold Xin\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Free dolly: Introducing the world's first truly open instructiontuned llm\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Angela Fan; Yacine Jernite; Ethan Perez; David Grangier; Jason Weston; Michael Auli\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b4\",\n \"title\": \"ELI5: long form question answering\",\n \"year\": \"2019-07-28\"\n },\n {\n \"authors\": \"Sebastian Gehrmann; Hendrik Strobelt; Alexander Rush\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b5\",\n \"title\": \"GLTR: Statistical detection and visualization of generated text\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Jesse Graham; Jonathan Haidt; Sena Koleva; Matt Motyl; Ravi Iyer; Sean Wojcik; Peter Ditto\",\n \"journal\": \"Advances in Experimental Social Psychology\",\n \"ref_id\": \"b6\",\n \"title\": \"Moral foundations theory: The pragmatic validity of moral pluralism\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Biyang Guo; Xin Zhang; Ziyuan Wang; Minqi Jiang; Jinran Nie; Yuxuan Ding; Jianwei Yue; Yupeng Wu\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"How close is chatgpt to human experts? comparison corpus, evaluation, and detection\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Xuanli He; Qiongkai Xu; Lingjuan Lyu; Fangzhao Wu; Chenguang Wang\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b8\",\n \"title\": \"Protecting intellectual property of language generation apis with lexical watermark\",\n \"year\": \"2022-02-22\"\n },\n {\n \"authors\": \"Jeppe Benjamin D Horne; Sibel Nørregaard; Adali\",\n \"journal\": \"ACM Transactions on Intelligent Systems and Technology (TIST)\",\n \"ref_id\": \"b9\",\n \"title\": \"Robust fake news detection over time and attack\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Khalid Hussain; Nimra Mughal; Irfan Ali; Saif Hassan; Sher Muhammad Daudpota\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \" Va Ivanin; El Artemova; Batura; Vv Ivanov; E V Vv Sarkisyan; Tutubalina; Smurov\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"RuREbus-2020 shared task: Russian relation extraction for business\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Dongyeop Kang; Waleed Ammar; Bhavana Dalvi; Madeleine Van Zuylen; Sebastian Kohlmeier; Eduard Hovy; Roy Schwartz\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b12\",\n \"title\": \"A dataset of peer reviews (PeerRead): Collection, insights and NLP applications\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"John Kirchenbauer; Jonas Geiping; Yuxin Wen; Jonathan Katz; Ian Miers; Tom Goldstein\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"A watermark for large language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Mahnaz Koupaee; William Yang; Wang \",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Wikihow: A large scale text summarization dataset\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Kalpesh Krishna; Yapei Chang; John Wieting; Mohit Iyyer\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b15\",\n \"title\": \"RankGen: Improving text generation with large ranking models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jenny S Li; John V Monaco; Li-Chiou Chen; Charles C Tappert\",\n \"journal\": \"IEEE Computer Society\",\n \"ref_id\": \"b16\",\n \"title\": \"Authorship authentication using short messages from social networking sites\",\n \"year\": \"2014-11-05\"\n },\n {\n \"authors\": \"Weixin Liang; Mert Yuksekgonul; Yining Mao; Eric Wu; James Zou\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Gpt detectors are biased against non-native english writers\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Yinhan Liu; Myle Ott; Naman Goyal; Jingfei Du; Mandar Joshi; Danqi Chen; Omer Levy; Mike Lewis; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Roberta: A robustly optimized bert pretraining approach\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Dominik Macko; Robert Moro; Adaku Uchendu; Jason Lucas; Michiharu Yamashita; Matúš Pikuliak; Ivan Srba; Thai Le; Dongwon Lee; Jakub Simko; Maria Bielikova\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b19\",\n \"title\": \"MULTITuDE: Large-scale multilingual machine-generated text detection benchmark\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Michail Melnichenko; Natalia Tyshkevich\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Prozhito from manuscript to corpus\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Eric Mitchell; Yoonho Lee; Alexander Khazatsky; Christopher D Manning; Chelsea Finn\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Detectgpt: Zero-shot machine-generated text detection using probability curvature\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Niklas Muennighoff; Thomas Wang; Lintang Sutawika; Adam Roberts; Stella Biderman; Teven Le Scao; M Saiful; Sheng Bari; Zheng Xin Shen; Hailey Yong; Xiangru Schoelkopf; Dragomir Tang; Alham Radev; Khalid Fikri Aji; Samuel Almubarak; Zaid Albanie; Albert Alyafeai; Edward Webson; Colin Raff; Raffel\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Crosslingual generalization through multitask finetuning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Marco Tulio Ribeiro; Sameer Singh; Carlos Guestrin\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"why should i trust you?\\\" explaining the predictions of any classifier\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Aounon Vinu Sankar Sadasivan; Sriram Kumar; Wenxiao Balasubramanian; Soheil Wang; Feizi\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Can ai-generated text be reliably detected?\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Neha Sengupta; Sunil Kumar Sahu; Bokang Jia; Satheesh Katipomu; Haonan Li; Fajri Koto; William Marshall; Gurpreet Gosal; Cynthia Liu; Zhiming Chen; Osama Mohammed Afzal; Samta Kamboj; Onkar Pandit; Rahul Pal; Lalit Pradhan; Zain Muhammad Mujahid; Massa Baali; Xudong Han; Mahmoud Sondos; Alham Bsharat; Zhiqiang Fikri Aji; Zhengzhong Shen; Natalia Liu; Joel Vassilieva; Andy Hestness; Andrew Hock; Jonathan Feldman; Andrew Lee; Hector Jackson; Preslav Xuguang Ren; Timothy Nakov; Eric Baldwin; Xing\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Jais and Jais-chat: Arabic-centric foundation and instruction-tuned open generative large language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Tatiana Shamardina; Vladislav Mikhailov; Daniil Cherniavskii; Alena Fenogenova; Marat Saidov; Anastasiya Valeeva; Tatiana Shavrina; Ivan Smurov; Elena Tutubalina; Ekaterina Artemova\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Findings of the the ruatd shared task 2022 on artificial text detection in russian\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tatiana Shavrina; Olga Shapovalova\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"To the methodology of corpus construction for machine learning: \\\"taiga\\\" syntax tree corpus and parser\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Rexhep Shijaku; Ercan Canhasi\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Chatgpt generated text detection\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Sebastian Szyller; Gul Buse; Samuel Atli; N Marchal; Asokan\",\n \"journal\": \"ACM\",\n \"ref_id\": \"b29\",\n \"title\": \"DAWN: dynamic adversarial watermarking of neural networks\",\n \"year\": \"2021-10-20\"\n },\n {\n \"authors\": \"Ruixiang Tang; Yu-Neng Chuang; Xia Hu\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"The science of detecting llm-generated texts\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Hugo Touvron; Louis Martin; Kevin Stone; Peter Albert; Amjad Almahairi; Yasmine Babaei; Nikolay Bashlykov; Soumya Batra; Prajjwal Bhargava; Shruti Bhosale; Dan Bikel; Lukas Blecher; Cristian Canton Ferrer; Moya Chen; Guillem Cucurull; David Esiobu; Jude Fernandes; Jeremy Fu; Wenyin Fu; Brian Fuller; Cynthia Gao; Vedanuj Goswami; Naman Goyal; Anthony Hartshorn; Saghar Hosseini; Rui Hou; Hakan Inan; Marcin Kardas; Viktor Kerkez; Madian Khabsa; Isabel Kloumann; Artem Korenev; Punit Singh Koura; Marie-Anne Lachaux; Thibaut Lavril; Jenya Lee; Diana Liskovich; Yinghai Lu; Yuning Mao; Xavier Martinet; Todor Mihaylov; Pushkar Mishra; Igor Molybog; Yixin Nie; Andrew Poulton; Jeremy Reizenstein; Rashi Rungta; Kalyan Saladi; Alan Schelten; Ruan Silva; Eric Michael Smith; Ranjan Subramanian; Ellen Xiaoqing; Binh Tan; Ross Tang; Adina Taylor; Jian Williams; Puxin Xiang Kuan; Zheng Xu; Iliyan Yan; Yuchen Zarov; Angela Zhang; Melanie Fan; Sharan Kambadur; Aurelien Narang; Robert Rodriguez; Sergey Stojnic; Thomas Edunov; Scialom\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"LLaMA 2: Open foundation and finetuned chat models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Adaku Uchendu; Zeyu Ma; Thai Le; Rui Zhang; Dongwon Lee\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b32\",\n \"title\": \"TURINGBENCH: A benchmark environment for Turing test in the age of neural text generation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yuxia Wang; Jonibek Mansurov; Petar Ivanov; Jinyan Su; Artem Shelmanov; Akim Tsvigun; Chenxi Whitehouse; Osama Mohammed Afzal; Tarek Mahmoud; Giovanni Puccetti; Thomas Arnold; Alham Fikri Aji; Nizar Habash; Iryna Gurevych; Preslav Nakov\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"SemEval-2024 task 8: Multidomain, multimodal and multilingual machine-generated text detection\",\n \"year\": \"2024\"\n },\n {\n \"authors\": \"Rowan Zellers; Ari Holtzman; Hannah Rashkin; Yonatan Bisk; Ali Farhadi; Franziska Roesner; Yejin Choi\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Defending against neural fake news\",\n \"year\": \"2019-12-08\"\n },\n {\n \"authors\": \"Xuandong Zhao; Yu-Xiang Wang; Lei Li\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Protecting language generation models via invisible watermarking\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"A \",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"2 Corpora in Other Languages Arabic Wikipedia\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Bulgarian True; & \",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Fake News is sampled from the Hack the Fake News datathon 17 organized in 2017 by the Data Science Society in Bulgaria\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"), precision (Prec), recall, and F1 score\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"Acc Prec Recall F1 Acc Prec Recall F1 RoBERTa(base) ChatGPT\",\n \"ref_id\": \"b39\",\n \"title\": \"8 Table 13: Same-generator, cross-domain experiments: training on a single domain of davinci-003 vs. human, and testing across domains. Evaluation accuracy (Acc), precision (Prec), recall, and F1 scores (in %) with respect to machine generations across four detectors. Test → ChatGPT davinci Cohere BLOOMz Train ↓ Acc Prec Recall F1 Acc Prec Recall F1\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"0 Table 15: Same-domain, cross-generator experiments: training and testing on Wikipedia (single machine-text generator vs. human). evaluation accuracy (Acc), precision (Prec), recall, and F1 score\",\n \"year\": \"\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[],"string":"[]","isConversational":false},"title":{"kind":"string","value":"M4: Multi-Generator, Multi-Domain, and Multi-Lingual Black-Box Machine-Generated Text Detection"},"abstract":{"kind":"string","value":"Large language models (LLMs) have demonstrated remarkable capability to generate fluent responses to a wide variety of user queries. However, this has also raised concerns about the potential misuse of such texts in journalism, education, and academia. In this study, we strive to create automated systems that can detect machine-generated texts and pinpoint potential misuse. We first introduce a large-scale benchmark M4, which is a multi-generator, multi-domain, and multi-lingual corpus for machine-generated text detection. Through an extensive empirical study of this dataset, we show that it is challenging for detectors to generalize well on instances from unseen domains or LLMs. In such cases, detectors tend to misclassify machine-generated text as human-written. These results show that the problem is far from solved and that there is a lot of room for improvement. We believe that our dataset will enable future research towards more robust approaches to this pressing societal problem."},"authors":{"kind":"string","value":"Yuxia Wang; Jonibek Mansurov; Petar Ivanov; Jinyan Su; Artem Shelmanov; Akim Tsvigun; Chenxi Whitehouse; † Osama; Mohammed Afzal; Tarek Mahmoud; Toru Sasaki; Thomas Arnold; Alham Fikri; Nizar Habash; Iryna Gurevych; Preslav Nakov; Mohamed Bin"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: Accuracy of cross-domain experiments: given generations from ChatGPT (top) or davinci (bottom), train on a single domain and test across domains across five detectors. (see more detail in Tables12 and 13)","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Accuracy of cross-generator experiments: train and test on arXiv (top) and Wikipedia (bottom) across five detectors, over single machine-text generator vs. human. (see detail in Tables14 and 15)","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Impact of text length on detection accuracy over arXiv and Reddit generated by ChatGPT, davinci, and Cohere.","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Visualization of the features extracted by LIME for Reddit as a domain, and ChatGPT as a generator.","figure_data":"","figure_id":"fig_3","figure_label":"4","figure_type":"figure"},{"figure_caption":"Statistics about our M4 dataset, which includes non-parallel human data and parallel human and machinegenerated texts.","figure_data":"","figure_id":"tab_1","figure_label":"1","figure_type":"table"},{"figure_caption":"Human evaluation on 100 examples from Reddit and arXiv (human, ChatGPT). The XLM-R detector fine-tuned on in-domain data demonstrated much better results than human annotators.","figure_data":"","figure_id":"tab_4","figure_label":"2","figure_type":"table"},{"figure_caption":"Zero-shot detection with GPTZero: recall (Rec) and F1-score with respect to generators and domains.","figure_data":"arXivRedditWikiHow Wikipedia PeerReadRecF1 RecF1 RecF1 RecF1 RecF1BLOOMz 0.4 0.8 7.6 13.8 0.0 0.0 2.0 3.9 5.8 10.9ChatGPT 26.2 41.5 86.4 91.6 49.4 62.1 87.2 93.1 70.8 82.7davinci0.2 0.4 60.4 74.3 45.2 59.4 53.8 70.0 96.2 97.9Cohere18.6 31.4 30.2 44.5 68.0 77.9 69.0 81.7 84.4 91.3Dolly v.25.4 10.3 52.8 66.7 13.6 21.1 29.4 45.4 18.6 31.3","figure_id":"tab_5","figure_label":"3","figure_type":"table"},{"figure_caption":"Accuracy (%) based on XLM-R on test sets across different languages over ChatGPT and davinci-003.","figure_data":"Generator →ChatGPTdavinci-003↓Test Domain →All Baike/ Ru Bulgarian IDN Urdu ArabicAll Baike/ Ru BulgarianTrain Domain ↓domain Web QA ATD News-News Wikipedia domain Web QA ATD News(en)(zh) (ru)(bg) (id) (ur)(ar)(en)(zh) (ru)(bg)All domains (en)98.697.5 76.6 80.8 76.9 57.796.590.293.0 54.1 66.0Baike/Web QA (zh)61.899.4 63.1 65.0 64.1 81.862.761.693.5 58.8 57.7RuATD (ru)59.192.6 97.5 81.7 76.9 55.586.256.775.7 84.7 82.2ChatGPT Bulgarian News (bg) 83.887.8 83.7 96.9 92.6 64.988.374.278.3 53.8 95.4IDN (id)65.959.9 62.6 67.6 98.4 50.654.661.055.6 50.6 58.7Urdu-News (ur)50.051.0 50.0 50.3 50.1 99.950.550.050.8 50.0 50.2Arabic Wikipedia (ar) 76.487.0 66.0 65.5 68.9 67.796.872.883.9 62.0 64.6All98.399.1 95.4 83.4 97.3 99.996.791.394.5 86.1 82.6All domains (en)95.979.7 70.4 72.4 67.2 61.193.195.879.5 60.5 65.8Baike/Web QA (zh)66.898.0 62.0 57.1 57.3 83.076.166.498.9 59.5 48.6davinci-003 RuATD (ru)61.460.5 88.6 72.4 58.6 49.768.962.849.6 95.3 86.5Bulgarian News (bg) 64.969.3 61.5 84.9 64.7 66.473.864.859.0 59.0 99.6All96.495.5 94.3 83.3 74.5 76.193.396.398.7 92.8 85.2Test →March 2023September 2023Train ↓Acc Precision Recall F1 Acc Precision Recall F1March99.5 99.0100 99.5 99.4 99.099.8 99.4September 96.0 10092.0 95.8 99.5 99.0100 99.5","figure_id":"tab_6","figure_label":"4","figure_type":"table"},{"figure_caption":"Impact","figure_data":"Length → Full Length 1,000 500250125Accuracy99.098.996.8 96.4 94.5Precision98.297.894.2 94.4 92.5Recall99.8100.0 99.8 98.6 96.8F199.098.996.9 96.5 94.6","figure_id":"tab_7","figure_label":"5","figure_type":"table"},{"figure_caption":"Statistics about the prompts for different domains and LLMs. One prompt is used for non-English text, and multiple prompts are used for English. The number of prompts for different domains varies as shown in the last column. Given a domain, some models might not follow all designed instructions, leading to less variety of prompts.","figure_data":"Domain↓davinc-003 ChatGPT Cohere Dolly-v2 Bloomz Unique across domainwikipedia111123Reddit551118wikihow111123peerread444444arxiv151128baike/web QA11NaNaNa1RuATD11NaNaNa1True Fake news11NaNaNa1Urdu-newsNa1NaNaNa1id_newspaperNa1NaNaNa1Arabic wikipediaNa1NaNaNa1Source/LanguageGeneratorDomainDavinci003ChatGPTCohereDolly-v2BLOOMzWikipediaEnglishmax_tokens=1000 max_tokens=1000 max_tokens=1000min_new_tokens=300, max_new_tokens=1000defaultmin_new_tokens=180Reddit ELI5Englishdefaultdefaultdefaultmax_new_tokens=600min_new_tokens=180WikiHowEnglishmax_tokens=2000 defaultdefaultmin_new_tokens=200 max_new_tokens=1000min_new_tokens=200PeerReadEnglishdefaultdefaultdefaultdefaultmin_new_tokens=150min_new_tokens=180,arXiv abstractEnglishmax_tokens=3000 defaultdefaultmin_new_tokens=180 max_new_tokens=600max_new_tokens=420, repitition_penalty=1.15,length_penalty=10Baike/Web QAChinesedefaultdefault---RuATDRussianmax_tokens=1700 default---Urdu-newsUrdu-temperature=0---id_newspapers_2018Indonesian -default---Arabic-WikipediaArabic-default---Bulgarian True & Fake News Bulgarianmax_tokens=3000 default---","figure_id":"tab_9","figure_label":"7","figure_type":"table"},{"figure_caption":"Hyper-parameter values used to generate data. We only specify the values that are different from the default.","figure_data":"","figure_id":"tab_10","figure_label":"8","figure_type":"table"},{"figure_caption":"Statistics about the number of unique uni-grams (word types) and bi-grams of human-written and machinegenerated texts (English).","figure_data":"Word (unigram)bigramsDomain↓Human ChatGPT davinc-003 Cohere Dolly-v2 BLOOMz Human ChatGPT davinc-003 Cohere Dolly-v2 BLOOMzWikipedia3341581891421677768327433725929693Reddit250140159107142134482247292191254164WikiHow369277250143160174867580514270294225PeerRead1421519082178133244262146129332154arXiv1281219697130159208199142168219218All domains228170160115154136457315293207277172All228147457252","figure_id":"tab_12","figure_label":"9","figure_type":"table"},{"figure_caption":"Statistics about per-document unique uni-grams (word types) and bi-grams of human-written and machine generated texts (English).Table11shows the non-default values of the hyper-parameters for the five detectors we trained.","figure_data":"B DetectorsB.1 Detector Hyper-Parameter ValuesDetector↓Learning rate # epochs Batch sizeMaximum iterationsCRoBERTa-base1e-61064--ELECTRA-base1e-61064--XLM-R-base2e-5516--LR-GLTRdefault-default1,000-Linear-SVM---20,0000.8","figure_id":"tab_13","figure_label":"10","figure_type":"table"},{"figure_caption":"Hyper-parameter settings for the five detectors we trained. LR-GLTR is based on the sklearn logistic regression implementation, and all hyper-parameters follow the default setting except for maximum training iterations, which we set to 1,000. The Linear-SVM detector uses all default parameters provided in the sklearn implementation except for the penalty parameter of the error term C and the maximum number of iterations.","figure_data":"","figure_id":"tab_14","figure_label":"11","figure_type":"table"},{"figure_caption":"Table12shows the evaluation results for same-generator, cross-domain experiments: training on a single domain of ChatGPT vs. human, and testing across domains. Table13shows the corresponding results for davinci-003 vs. human.","figure_data":"D Results: Same-Domain, Cross-GeneratorTest →WikipediaWikiHowReddit ELI5arXivPeerReadTrain ↓Acc Prec Recall F1Acc Prec Recall F1Acc Prec Recall F1Acc Prec Recall F1Acc Prec Recall F1","figure_id":"tab_15","figure_label":"","figure_type":"table"},{"figure_caption":"Table14shows detailed evaluation results for same-domain, cross-generator experiments: training and testing on arXiv (single machine-text generator vs. human). Table15shows the corresponding results when training and testing on Wikipedia. Same","figure_data":"Test →ChatGPTdavinciCohereBLOOMzTrain ↓Acc Prec Recall F1Acc Prec Recall F1Acc Prec Recall F1Acc Prec Recall F1RoBERTa(base)ChatGPT 99.7 99.4 100.99.7 99.7 99.4 100.99.7 99.4 99.8 99.099.4 77.7 100. 55.471.3davinci99.6 99.2 100.99.6 99.5 99.2 99.899.5 99.4 99.8 99.099.4 81.4 99.7 63.077.2Cohere99.7 99.4 100.99.7 99.6 99.4 99.899.6 99.6 99.8 99.499.6 82.6 99.7 65.479.0BLOOMz 99.3 98.8 99.899.3 99.3 99.8 99.899.3 99.0 98.8 99.299.0 98.1 98.8 97.498.1ELECTRA(large)ChatGPT 93.5 88.9 99.493.9 91.8 88.6 96.092.1 83.6 86.4 79.883.0 56.6 72.9 21.032.6davinci88.6 81.8 99.289.7 88.7 81.9 99.489.8 83.8 79.2 91.685.0 62.4 73.1 39.251.0Cohere84.3 76.5 99.086.3 83.4 76.2 97.285.4 85.5 77.8 99.487.3 72.4 72.7 71.872.2BLOOMz 49.9 48.1 2.64.950.1 51.7 3.05.753.2 77.6 9.016.1 97.5 98.0 97.097.5LR-GLTRChatGPT 96.3 96.4 96.296.3 65.3 90.1 34.449.8 96.9 96.4 97.496.9 65.5 90.6 34.650.1davinci81.2 83.9 77.280.4 85.2 84.5 86.285.3 78.5 82.9 71.877.0 73.7 80.8 62.270.3Cohere96.8 96.4 97.296.8 66.0 90.4 35.851.3 97.0 96.4 97.697.0 61.5 88.1 26.640.9BLOOMz 89.2 87.7 91.289.4 71.2 80.8 55.665.9 79.5 84.9 71.877.8 87.2 87.2 87.287.2StylisticChatGPT 100. 100. 100.100. 71.0 100. 42.059.2 87.7 100. 75.486.0 62.4 100. 24.839.7davinci97.3 97.4 97.297.3 97.4 97.2 97.697.4 82.8 96.3 68.279.9 87.1 96.7 76.885.6Cohere97.6 99.4 95.897.6 83.8 99.7 67.880.7 98.8 99.4 98.298.8 65.5 98.1 31.647.8BLOOMz 63.4 95.3 28.243.5 76.0 97.4 53.469.0 55.5 89.9 12.421.8 98.5 98.6 98.498.5NELAChatGPT 97.2 97.0 97.497.2 52.0 69.2 7.213.0 64.2 91.3 31.446.7 48.8 16.7 0.61.2davinci48.3 41.2 8.013.4 88.5 88.9 88.088.4 45.8 20.8 3.05.273.0 83.4 57.468.0Cohere70.1 88.8 46.060.6 49.4 44.6 5.09.093.9 94.2 93.693.9 47.1 20.8 7.38.9BLOOMz 48.6 11.1 0.40.855.5 81.6 14.224.2 48.7 15.8 0.61.296.9 96.8 97.096.9","figure_id":"tab_16","figure_label":"14","figure_type":"table"},{"figure_caption":"Cross-language experiments. Accuracy (Acc) and F1 scores (for the machine-generated class) based on XLM-R on the test sets across different languages generated by ChatGPT. We average the performance across five runs (the standard deviation is shown in parenthesis).","figure_data":"Generator↓ Test Domain →All domain (en)Baike/Web QA (zh)RuATD (ru)Bulgarian News (bg)Train Domain ↓AccF1AccF1AccF1AccF1All domains (en)95.8 (1.9)96.0 (1.8)79.5 (4.1)82.9 (2.9)60.5 (3.0) 65.3 (5.1)65.8 (3.2)69.3 (6.6)davinci-003 Baike/Web QA (zh)66.4 (7.6)74.8 (4.2)98.9 (0.4)98.9 (0.4)59.5 (0.6) 70.0 (0.6)48.6 (3.3)61.3 (3.7)RuATD (ru)62.8 (3.0)62.0 (8.1)49.6 (9.3)58.6 (3.2)95.3 (1.6) 95.4 (1.4)86.5 (5.1)86.0 (6.5)Bulgarian News (bg)64.8 (3.1)67.2 (9.1)59.0 (8.7)29.4 (23.6) 59.0 (3.6) 32.0 (11.3) 99.6 (0.2)99.6 (0.2)All96.3 (0.7)96.4 (0.6)98.7 (0.5)98.7 (0.5)92.8 (2.1) 93.2 (2.0)85.2 (3.2)87.0 (2.3)All domains (en)90.2 (0.9)89.4 (1.0)93.0 (0.9)92.6 (1.1)54.1 (1.8) 51.5 (5.2)66.0 (3.2)64.3 (7.6)ChatGPTBaike/Web QA (zh)61.6 (5.5)72.2 (2.8)93.5 (1.1)93.1 (1.2)58.8 (2.2) 67.7 (3.7)57.7 (3.4)65.0 (5.0)RuATD (ru)56.7 (3.0)68.6 (0.5)75.7 (7.6)67.5 (14.5) 84.7 (3.9) 82.4 (5.8)82.2 (4.5)84.9 (3.2)Bulgarian News (bg)74.2 (4.9)75.1 (2.2)78.3 (11.2) 70.1 (21.1) 53.8 (1.5) 15.5 (5.8)95.4 (1.3)95.3 (1.4)IDN (id)61.0 (14.3) 29.5 (37.4) 55.6 (7.7)17.5 (23.6) 50.6 (0.8) 5.1 (7.0)58.7 (13.9) 23.6 (35.0)Urdu-News (ur)50.0 (0.1)66.6 (0.1)50.8 (0.7)67.0 (0.3)50.0 (0.0) 66.7 (0.0)50.2 (0.2)66.8 (0.1)Arabic Wikipedia (ar) 72.8 (4.7)77.0 (2.8)83.9 (6.9)85.5 (5.1)62.0 (2.3) 70.2 (1.1)64.6 (5.9)73.6 (3.0)All91.3 (0.6)90.8 (0.6)94.5 (1.2)94.3 (1.4)86.1 (2.5) 85.4 (2.9)82.6 (2.2)84.9 (1.5)","figure_id":"tab_18","figure_label":"16","figure_type":"table"},{"figure_caption":"Cross","figure_data":"","figure_id":"tab_19","figure_label":"17","figure_type":"table"},{"figure_caption":"Data format of M4: English examples sampled across different domains and LLM generators. a joust against the captain of his Scottish Guard, Gabriel, the Count of Montgomery. The fateful run occurred at the end of a tournament day, after Montgomery had almost unhorsed the king. Henry instead on another tilt. Montgomery's lance struck the king's helmet and shattered, with a long splinter running through the king's visor through his eye and into or near his brain. The king initially survived injury, and was attended to by two of the most celebrated physicians in Europe, Ambroise Paré and Andreas Vesalius. The queen, Catherine de Medici, ordered four prisoners executed with wood driven into their brains so that the physicians would have the chance to study the king's wound in detail on the corpses. Despite this rather extraordinary measure, the king deteriorated steadily. Vesalius' personal account is consistent with the development of meningitis or encephalitis. After 11 days, the king died. During these 11 days, Montgomery is supposed to have come to the king's side, and asked to have his head and right hand cut off in punishment. The king told him that he had jousted well and bravely and that the accident was not his fault. Following Henry's death, Catherine essentially ruled through a series of three of her sons. Montgomery retired to his estate in Normandy. From there, his history gets a bit complicated. He converted to Protestantism and took part in the religious fighting in Bourges and Rouen that followed the St. Bartholomew's Day Massacre in 1572. He escaped to England and remained there for about a year, later returning to attempt to lead a Huguenot insurrection in Normandy. He was captured by royal forces and executed in 1574.","figure_data":"FieldContentLanguageEnglishLLM Generator CHATGPTDomainReddit ELI5QuestionWhat happened to Henry II's jousting opponent after accidentally killing him in 1559?HumanHenry died in","figure_id":"tab_20","figure_label":"18","figure_type":"table"},{"figure_caption":"PromptPlease, generate wikihow article with length above 1000 characters from title 'How to Buy Virtual Console Games for Nintendo Wii' and headline 'Find a few Wii Points cards from game retailers like GameStop. Make sure your Wii is online and on a secure connection if possible.,Open the Wii Shop Channel, and click Add Wii Points at the bottom right of your screen after the channel finishes loading. Click \"From Card\". Now, your Wii should be asking for a code with 5 segments, separated by dashes. Enter this code into the Wii now. After a few prompts, you will have Wii points added to your Wii, which is what you use to purchase games. Click \"Virtual Console\". Choose how you want the games to be displayed. Scroll through the list for games that you want.Click on a game you want to play, And hit Yes on the confirmation box. Wait while the game is downloaded. After you have bought the games you want to play, go back to the Wii Menu. Select the new games you just bought, and start them like a regular channel. Open the Wii Shop Channel, and click Add Wii Points. Click \"With a Credit Card\". Key in your Credit Card Number. Let Nintendo authorize the purchase.,Go back to the Points Card process above and skip down to Step 7. Buy Virtual Console Games for Nintendo Wii Most people know that Nintendoś library of classic titles is available on the Wii platform through the Virtual Console. Playing these games from years gone by is easy, but there are a few steps you have to take to purchase them. Hereś what you have to do to buy Virtual Console games for your Nintendo Wii. Find a few Wii Points cards from game retailers like GameStop. You ĺl find that most stores that carry video games have them available, and you usually can select cards of different amounts. Choose the ones that are going to give you what you need. Make sure your Wii is online and on a secure connection if possible. Going online with your Wii allows you to access all the Virtual Console titles that Nintendo has made available. To do this, select the Wii Icon from the main screen, then go to the settings area, select the Internet tab, then go to the Connect to the Internet option. Select an existing connection if you have one, or go through the Wiiś setup menu to make a new connection. Open the Wii Shop Channel, and click Add Wii Points at the bottom right of your screen after the channel finishes loading. On the next screen, you have two choices; one from a card, or from a credit card. Click \"From Card\". Now, your Wii should be asking for a code with 5 segments, separated by dashes. This code is what you use to transfer the money you have to the Virtual Console shop. Enter this code into the Wii now. After a few prompts, you will have Wii points added to your Wii, which is what you use to purchase games. Click \"Virtual Console\". This option allows you to scroll through all the different games that have been made available for the Virtual Console. Choose how you want the games to be displayed. You can sort them alphabetically, or by genre, or even by release date. Scroll through the list for games that you want. When you find one, click on it and you can see the details and the cost of the game. Click on a game you want to play, And hit Yes on the confirmation box. You then have to wait while the game is downloaded onto your hard drive. This will take a few minutes, depending on the size of the game you want. After you have bought the games you want to play, go back to the Wii Menu. Select the new games you just bought, and start them just like a regular channel. If you want to add more funds to your Wii point wallet, go back to step 3. But this time instead of clicking \"From a card\", click on the \"With A credit card\" option. Key in your Credit Card Number. Be sure that it is entered correctly, because many purchases from the Virtual Console are automated and there are no refunds if you enter your number incorrectly. Let Nintendo authorize the purchase. A few steps later you should be notified of the success of your purchase, and you should now have more Wii points with which to buy more games. Go back to the Points Card process above and skip down to Step 7. This is the step right after you have verified that you have the points. Here, you can select a game, and start playing it. Now, you know how to purchase and download games from the Virtual Console. Spend some time going through the list, purchasing some of the classics like Super Mario and the Legend of Zelda, and for a small fee, you can start playing them on your Wii. Have fun!","figure_data":"FieldContentMachine How to Language EnglishLLM Generator CHATGPTDomainWikihowTitleHow to Workaround Damaged Laptop Display Hinges","figure_id":"tab_21","figure_label":"","figure_type":"table"},{"figure_caption":"If it's still loose, you may need to repeat steps 3-5 with stronger epoxy glue and larger strips of Velcro. 7. Tie a fishing line loop around the laptop. Take the fishing line and tie it into a loop which exactly encircles the laptop when the display is open to the working position. Be sure to check that the fishing line loop is strong enough to prevent the display from falling into the fully open position. 8. Glue foam tape to the laptop edge. To finalize the repair, glue a small strip of two-sided foam tape on the edge of the laptop where the fishing line should rest. Optionally, you can also use a plastic bag closure to adjust the exact size of the fishing line loop and the working position of the laptop display. With these simple steps, you can work around damaged laptop display hinges and fix the issue without having to replace the entire hinge. Keep in mind that this repair method may not work for all laptop models, so be sure to check if your laptop can be repaired in this way before getting started.","figure_data":"FieldContentLanguageEnglishLLM Generator CHATGPTDomainPeerReadTitleBridge Text and Knowledge by Learning Multi-Prototype Entity Mention EmbeddingGood luck!","figure_id":"tab_22","figure_label":"","figure_type":"table"},{"figure_caption":"Examples of other languages across different domains by ChatGPT and davinci-text-003. Игоря, он согласился со мной, что «она, по-видимому, хороший человек». Но тут же прибавил: -Разве вы не видите, тетя Вава, что мне лучше всего одному, как я теперь живу. Дико и оскорбительно представить себе вот в этих стенах какую-то другую жену, кроме той, которая у меня отнята смертью. День праздника был посвящен пути движения Игоря, его сына. Взгляд, которым его судьба была окрашена, был властный, а к мимолетным правам на близость был дан официальный статус. Желание судить и осуждать пару, видимо, комплексное, так как сразу трое -мужчина, женщина и в какой-то мере дети, занимаются браконьерством. Даже брутальное отношение главы дома к ним не могло исказить более-менее нормальное настроение.Одержимость жены, ее свежесть, девичья улыбка с ямочками на лице, а также твердый взгляд серых глаз очень привлекла мягкосердечного, но застаревшего мужа. Назад вспомнили, как 4 года назад он встретился с красавицей, которая должна была привлечь его интерес, по мнению Игоря, в душевном мире и представлении жизни. Однако, овладение собой было столь сильно, что он дошел до того, что «лучше для него жить в одиночестве, чем представлять себе другую жену внутри этих чертовых стен». Даже сейчас после потери жены, чья память приукрасит и облегчит жизнь Игоря и его сына, до разрушительных, можно догадываться, или загнутых ходов, они не доводятся. Няма да допуснем да продължат. Ако се премахнат буните за един месец ще изчезне целият плаж\", каза един от протестиращите пред Флагман.бг.Наш репортер се свърза и със стопанина на плажа -бургаската фирма \"Акра лидо\" ООД, която има 5-годишен договор за ивицата и плаща най-скъпия плажен наем в България -по 107 хил.лева на година за изключително малка площ.От компанията останаха изненадани от действията на строителния наздор. Буните са там от много години, спират свлачищния процес и осигурят плажна ивица, макар и много тясна.Само преди седмица от \"Акра лидо\" са платили накуп наемната цена за 2017 година, а сега ако бъдат премахнати буните ще останат без плаж.Скандалът ще има международен отзвук, защото 500 руски семейства имат жилища в непосредствена близост до къмпинг \"Черноморец\" и са ги закупили заради това, че плажът им е само на няколко метра.Това е друга от машините, която ще участва в акцията по разрушаването на буните\"Ние сме изрядни платци, наели сме плажа в това състояние от държавата и до тази сутрин смятахме, че всичко е изрядно. Не сме поставяли под съмнение, че държавата ни е отдала под наем плаж с незаконни буни\", каза един от собствениците на фирмата Милен Янков пред Флагман.бг.Припомняме, че миналата година също имаше много проблеми с този плаж. В основата бяха отношенията със стария стопанин на ивицата (ВИЖ НАШАТА ТЕМА В РАЗВИТИЕ)ОЧАКВАЙТЕ ПОДРОБНОСТИ! В категории: Созопол Докато правните изследвания са все още в процес на продължаване за да се разбере точно какви са причините за багера, се очаква тази политика да продължи и да се извършват постоянни разрешения за някои интереси. Някои пленници на ансамбъла на бунита подози, че максимумът на разрешенията е 100, които могат да се дават годишно според правилата, които сме предвидили. Защо РДНСК решил да направи това и защо именно сега, това продължава да бъде неизвестно. Но понеже това се оказва наистина забранено, много от плавателните клубове и индивиди на плажа ще бъдат принудени да се преместят на други места. Бановете, разположени в гледната точка за потоците, да се зачитат на рилско море -това ли ще стане след махането на буните? Ще има ли някаква защита за плавателните инсталации, една от важните атракции на Golden Sands? Ще плажът продължи да съществува или ще бъде претърпял значителни изменения? Тези и много други въпроси очакват отговори. За да се оцени тайнственото поведение на РДНСК, много информация трябва да се прецени. Ето защо е особено важно да се осигури правопорядъчно намиране на логичен разследвателен процес за да се разберат всички мотиви за действието на РДНСК и да се донесат промени, които да се правят в добро на всички страни.HumanОбикновените хора се обръщат към кралица Елизабет Втора с \"Нейно кралско величество\" и придружават обръщението с поклон и израз на любезност. Но зад кулисите отношенията са доста по-неформални, разказват от \"Новини Лондон\". Потребител на сайта Quora зададе въпрос: \"Как внуците на кралица Елизабет се обръщат към нея на четири очи?\". Макар че отговорите варират, проучване на \"Индипендънт\" показва, че внуците на кралицата са й измислили обръщения, изразяващи обичта им към нея. Уилям и Хари окуражаваха баба си при нейната неочаквана, спираща дъха роля на церемонията по откриването на Олимпийските игри през 2012 г. При вида на баба им, която се появи заедно с Даниел Крейг като Джеймс Бонд, принцовете извикаха с цяло гърло: \"Давай, бабо!\". Но това не е единственото име, използвано от Уилям и Хари. Според светския хроникьор Ричард Кей веднъж в Бъкингамския дворец Уилям казал на баба си \"Гари\". Когато служители на двореца попитали \"Кой е Гари?\", кралицата отговорила: \"Аз съм Гари\". Уил още не се е научил да казва \"Granny\" (бабоанглийски език, бел. ред.), обяснила Елизабет. Тригодишният правнук на Елизабет Втора принц Джордж също има любвеобилно обръщение към кралицата, както стана ясно от интервю, излъчено по телевизия ITV миналото лято по повод 90годишнината й. Херцогинята на Кеймбридж Кейт Мидълтън сподели, че Джордж нарича баба си \"Ган Ган\". Но това не е просто мило обръщение, а е израз на дълготрайна кралска традиция. Американският биограф Кити Кели твърди, че принц Чарлз, Уилям и Хари също се обръщали така към своята прабаба. Prompt Моля, генерирайте статия с дължина над 1000 символа и заглавие \"Неподозирани разкрития за кралското семейство\" Machine Неподозирани разкрития за кралското семейство Кралските семейства винаги са били обект на внимание и интерес от страна на медиите и обществото. Известни са скандалите и скритите истории, които остават извън зрителния ъгъл и несъмнено експлоатирането на подобни теми продължава да държи високи места в таблоидите. Но дали знаете всичко за кралските семейства, което медиите не съобщават? Има ли неща, които остават скрити дълбоко в историята или в семейните тайни? Изглежда, отговорът е да. В последните години имаше повече и повече изисквания за откриване на тайните на кралските династии, но много от тези разкрития остават скрити за публика. Една от най-известните скрити истории, свързани с британското кралско семейство, е тази на принца Джон, третият син на крал Джордж V и кралица Мери. Принц Джон беше диагностициран с епилепсия и склероза на 7-годишна възраст. Той беше изолиран от обществото и от своето семейство и беше принуден да живее в бездейственост и самота, за да не покаже \"ненормалното\"си поведение. Принц Джон почина на 13-годишна възраст, покрит с тайните и позволено му беше да забрави света на живите. Разказите за скритата история на Американското кралско семейство са също известни. Съществува теория, че Сежър, който се хвали, че е третият наследник на трона след Алфонсо и Джон, наистина има връзки с кралското семейство на САЩ. През 1937 г. Сежър заяви, че е бил роден като дете на Сесилия, дъщерята на крал Едуард VIII и Уолис Симпсън. Макар това да е слух, наличието на тайни и тъмни истории в кралските семейства е добре документирано. Все още има много да се разкрие, но кралските династии на света ще продължат да привличат интриги и вълнение през годините. CNN Indonesia -Musisi Ahmad Dhani memenuhi panggilan Polda Metro Jaya untuk diperiksa sebagai saksi tersangka dugaan makar Sri Bintang Pamungkas, Selasa (20/12). Berdasarkan pantauan CNNIndonesia.com, dia tiba pukul 15.00 WIB. Sedangkan tim kuasa hukumnya yang tergabung dalam Advokat Cinta Tanah Air sudah tiba satu jam sebelum kedatangannya. Tak lama kemudian, Farhat Abbas juga datang untuk mendampingi Ahmad Dhani. Ahmad Dhani mengatakan, dirinya tidak akan memberikan informasi apapun soal Sri Bintang. Dia mengklaim tidak kenal dengan Sri Bintang. Buni Yani dan Ahmad Dhani Jadi Saksi Kasus Sri Bintang Besok Saksi Mengaku Dapat Aliran Dana dari Tersangka Makar Buni Yani Diperiksa Soal Pidato Sri Bintang di Kalijodo \"Informasi pasti tidak ada, karena saya tidak kenal dengan Sri Bintang Pamungkas. Saya pernah ketemu beliau ketika di Mako Brimob,\" ucapnya. Meski demikian, Ahmad Dhani mengaku hadir saat ada pertemuan di Universitas Bung Karno. Namun, ia mengklaim tidak mendengar pidato yang disampaikan oleh Sri Bintang karena terlambat datang. Awalnya, Kepala Subdirektorat Kejahatan dan Kekerasan Direktorat Reserse Kriminal Umum Polda Metro Jaya AKBP Hendy Kurniawan mengatakan, karena Ahmad Dhani sedang sakit maka pemeriksaan terhadap dia yang akan dilakukan hari ini harus ditunda Kamis (22/12). \"Ahmad Dhani ditunda karena sakit. Surat sakit sudah kami terima. Kami sudah koordinasi dengan kuasa hukumnya kemarin,\" ujarnya. Kuasa Hukum Ahmad Dhani, Ali Lubis mengklaim, kliennya tersebut merupakan warga negara yang baik sehingga datang memenuhi panggilan kepolisian. \"Beliau kooperatif, beliau ingin membantu kepolisian jadi lebih cepat lebih baik,\" ucapnya. Ahmad Dhani telah ditetapkan sebagai tersangka dugaan penghinaan terhadap penguasa. Dia ikut ditangkap pada Jumat (2/12) dengan sejumlah tersangka dugaan makar lainnya. Sepuluh tersangka dugaan makar itu adalah Sri Bintang, Kivlan Zein, Adityawarman Thahar, Ratna Sarumpaet, Firza Huzein, Eko Santjojo, Alvin Indra, Rachmawati Soekarnoputri, dan kakak beradik Rizal Kobar dan Jamran.","figure_data":"FieldContentLanguageArabicLLM Generator CHATGPTDomainWikipediaTitleHuman","figure_id":"tab_23","figure_label":"20","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: Accuracy of cross-domain experiments: given generations from ChatGPT (top) or davinci (bottom), train on a single domain and test across domains across five detectors. (see more detail in Tables12 and 13)\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Accuracy of cross-generator experiments: train and test on arXiv (top) and Wikipedia (bottom) across five detectors, over single machine-text generator vs. human. (see detail in Tables14 and 15)\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Impact of text length on detection accuracy over arXiv and Reddit generated by ChatGPT, davinci, and Cohere.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Visualization of the features extracted by LIME for Reddit as a domain, and ChatGPT as a generator.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Statistics about our M4 dataset, which includes non-parallel human data and parallel human and machinegenerated texts.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Human evaluation on 100 examples from Reddit and arXiv (human, ChatGPT). The XLM-R detector fine-tuned on in-domain data demonstrated much better results than human annotators.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Zero-shot detection with GPTZero: recall (Rec) and F1-score with respect to generators and domains.\",\n \"figure_data\": \"arXivRedditWikiHow Wikipedia PeerReadRecF1 RecF1 RecF1 RecF1 RecF1BLOOMz 0.4 0.8 7.6 13.8 0.0 0.0 2.0 3.9 5.8 10.9ChatGPT 26.2 41.5 86.4 91.6 49.4 62.1 87.2 93.1 70.8 82.7davinci0.2 0.4 60.4 74.3 45.2 59.4 53.8 70.0 96.2 97.9Cohere18.6 31.4 30.2 44.5 68.0 77.9 69.0 81.7 84.4 91.3Dolly v.25.4 10.3 52.8 66.7 13.6 21.1 29.4 45.4 18.6 31.3\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Accuracy (%) based on XLM-R on test sets across different languages over ChatGPT and davinci-003.\",\n \"figure_data\": \"Generator →ChatGPTdavinci-003↓Test Domain →All Baike/ Ru Bulgarian IDN Urdu ArabicAll Baike/ Ru BulgarianTrain Domain ↓domain Web QA ATD News-News Wikipedia domain Web QA ATD News(en)(zh) (ru)(bg) (id) (ur)(ar)(en)(zh) (ru)(bg)All domains (en)98.697.5 76.6 80.8 76.9 57.796.590.293.0 54.1 66.0Baike/Web QA (zh)61.899.4 63.1 65.0 64.1 81.862.761.693.5 58.8 57.7RuATD (ru)59.192.6 97.5 81.7 76.9 55.586.256.775.7 84.7 82.2ChatGPT Bulgarian News (bg) 83.887.8 83.7 96.9 92.6 64.988.374.278.3 53.8 95.4IDN (id)65.959.9 62.6 67.6 98.4 50.654.661.055.6 50.6 58.7Urdu-News (ur)50.051.0 50.0 50.3 50.1 99.950.550.050.8 50.0 50.2Arabic Wikipedia (ar) 76.487.0 66.0 65.5 68.9 67.796.872.883.9 62.0 64.6All98.399.1 95.4 83.4 97.3 99.996.791.394.5 86.1 82.6All domains (en)95.979.7 70.4 72.4 67.2 61.193.195.879.5 60.5 65.8Baike/Web QA (zh)66.898.0 62.0 57.1 57.3 83.076.166.498.9 59.5 48.6davinci-003 RuATD (ru)61.460.5 88.6 72.4 58.6 49.768.962.849.6 95.3 86.5Bulgarian News (bg) 64.969.3 61.5 84.9 64.7 66.473.864.859.0 59.0 99.6All96.495.5 94.3 83.3 74.5 76.193.396.398.7 92.8 85.2Test →March 2023September 2023Train ↓Acc Precision Recall F1 Acc Precision Recall F1March99.5 99.0100 99.5 99.4 99.099.8 99.4September 96.0 10092.0 95.8 99.5 99.0100 99.5\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Impact\",\n \"figure_data\": \"Length → Full Length 1,000 500250125Accuracy99.098.996.8 96.4 94.5Precision98.297.894.2 94.4 92.5Recall99.8100.0 99.8 98.6 96.8F199.098.996.9 96.5 94.6\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Statistics about the prompts for different domains and LLMs. One prompt is used for non-English text, and multiple prompts are used for English. The number of prompts for different domains varies as shown in the last column. Given a domain, some models might not follow all designed instructions, leading to less variety of prompts.\",\n \"figure_data\": \"Domain↓davinc-003 ChatGPT Cohere Dolly-v2 Bloomz Unique across domainwikipedia111123Reddit551118wikihow111123peerread444444arxiv151128baike/web QA11NaNaNa1RuATD11NaNaNa1True Fake news11NaNaNa1Urdu-newsNa1NaNaNa1id_newspaperNa1NaNaNa1Arabic wikipediaNa1NaNaNa1Source/LanguageGeneratorDomainDavinci003ChatGPTCohereDolly-v2BLOOMzWikipediaEnglishmax_tokens=1000 max_tokens=1000 max_tokens=1000min_new_tokens=300, max_new_tokens=1000defaultmin_new_tokens=180Reddit ELI5Englishdefaultdefaultdefaultmax_new_tokens=600min_new_tokens=180WikiHowEnglishmax_tokens=2000 defaultdefaultmin_new_tokens=200 max_new_tokens=1000min_new_tokens=200PeerReadEnglishdefaultdefaultdefaultdefaultmin_new_tokens=150min_new_tokens=180,arXiv abstractEnglishmax_tokens=3000 defaultdefaultmin_new_tokens=180 max_new_tokens=600max_new_tokens=420, repitition_penalty=1.15,length_penalty=10Baike/Web QAChinesedefaultdefault---RuATDRussianmax_tokens=1700 default---Urdu-newsUrdu-temperature=0---id_newspapers_2018Indonesian -default---Arabic-WikipediaArabic-default---Bulgarian True & Fake News Bulgarianmax_tokens=3000 default---\",\n \"figure_id\": \"tab_9\",\n \"figure_label\": \"7\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Hyper-parameter values used to generate data. We only specify the values that are different from the default.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_10\",\n \"figure_label\": \"8\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Statistics about the number of unique uni-grams (word types) and bi-grams of human-written and machinegenerated texts (English).\",\n \"figure_data\": \"Word (unigram)bigramsDomain↓Human ChatGPT davinc-003 Cohere Dolly-v2 BLOOMz Human ChatGPT davinc-003 Cohere Dolly-v2 BLOOMzWikipedia3341581891421677768327433725929693Reddit250140159107142134482247292191254164WikiHow369277250143160174867580514270294225PeerRead1421519082178133244262146129332154arXiv1281219697130159208199142168219218All domains228170160115154136457315293207277172All228147457252\",\n \"figure_id\": \"tab_12\",\n \"figure_label\": \"9\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Statistics about per-document unique uni-grams (word types) and bi-grams of human-written and machine generated texts (English).Table11shows the non-default values of the hyper-parameters for the five detectors we trained.\",\n \"figure_data\": \"B DetectorsB.1 Detector Hyper-Parameter ValuesDetector↓Learning rate # epochs Batch sizeMaximum iterationsCRoBERTa-base1e-61064--ELECTRA-base1e-61064--XLM-R-base2e-5516--LR-GLTRdefault-default1,000-Linear-SVM---20,0000.8\",\n \"figure_id\": \"tab_13\",\n \"figure_label\": \"10\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Hyper-parameter settings for the five detectors we trained. LR-GLTR is based on the sklearn logistic regression implementation, and all hyper-parameters follow the default setting except for maximum training iterations, which we set to 1,000. The Linear-SVM detector uses all default parameters provided in the sklearn implementation except for the penalty parameter of the error term C and the maximum number of iterations.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_14\",\n \"figure_label\": \"11\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Table12shows the evaluation results for same-generator, cross-domain experiments: training on a single domain of ChatGPT vs. human, and testing across domains. Table13shows the corresponding results for davinci-003 vs. human.\",\n \"figure_data\": \"D Results: Same-Domain, Cross-GeneratorTest →WikipediaWikiHowReddit ELI5arXivPeerReadTrain ↓Acc Prec Recall F1Acc Prec Recall F1Acc Prec Recall F1Acc Prec Recall F1Acc Prec Recall F1\",\n \"figure_id\": \"tab_15\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Table14shows detailed evaluation results for same-domain, cross-generator experiments: training and testing on arXiv (single machine-text generator vs. human). Table15shows the corresponding results when training and testing on Wikipedia. Same\",\n \"figure_data\": \"Test →ChatGPTdavinciCohereBLOOMzTrain ↓Acc Prec Recall F1Acc Prec Recall F1Acc Prec Recall F1Acc Prec Recall F1RoBERTa(base)ChatGPT 99.7 99.4 100.99.7 99.7 99.4 100.99.7 99.4 99.8 99.099.4 77.7 100. 55.471.3davinci99.6 99.2 100.99.6 99.5 99.2 99.899.5 99.4 99.8 99.099.4 81.4 99.7 63.077.2Cohere99.7 99.4 100.99.7 99.6 99.4 99.899.6 99.6 99.8 99.499.6 82.6 99.7 65.479.0BLOOMz 99.3 98.8 99.899.3 99.3 99.8 99.899.3 99.0 98.8 99.299.0 98.1 98.8 97.498.1ELECTRA(large)ChatGPT 93.5 88.9 99.493.9 91.8 88.6 96.092.1 83.6 86.4 79.883.0 56.6 72.9 21.032.6davinci88.6 81.8 99.289.7 88.7 81.9 99.489.8 83.8 79.2 91.685.0 62.4 73.1 39.251.0Cohere84.3 76.5 99.086.3 83.4 76.2 97.285.4 85.5 77.8 99.487.3 72.4 72.7 71.872.2BLOOMz 49.9 48.1 2.64.950.1 51.7 3.05.753.2 77.6 9.016.1 97.5 98.0 97.097.5LR-GLTRChatGPT 96.3 96.4 96.296.3 65.3 90.1 34.449.8 96.9 96.4 97.496.9 65.5 90.6 34.650.1davinci81.2 83.9 77.280.4 85.2 84.5 86.285.3 78.5 82.9 71.877.0 73.7 80.8 62.270.3Cohere96.8 96.4 97.296.8 66.0 90.4 35.851.3 97.0 96.4 97.697.0 61.5 88.1 26.640.9BLOOMz 89.2 87.7 91.289.4 71.2 80.8 55.665.9 79.5 84.9 71.877.8 87.2 87.2 87.287.2StylisticChatGPT 100. 100. 100.100. 71.0 100. 42.059.2 87.7 100. 75.486.0 62.4 100. 24.839.7davinci97.3 97.4 97.297.3 97.4 97.2 97.697.4 82.8 96.3 68.279.9 87.1 96.7 76.885.6Cohere97.6 99.4 95.897.6 83.8 99.7 67.880.7 98.8 99.4 98.298.8 65.5 98.1 31.647.8BLOOMz 63.4 95.3 28.243.5 76.0 97.4 53.469.0 55.5 89.9 12.421.8 98.5 98.6 98.498.5NELAChatGPT 97.2 97.0 97.497.2 52.0 69.2 7.213.0 64.2 91.3 31.446.7 48.8 16.7 0.61.2davinci48.3 41.2 8.013.4 88.5 88.9 88.088.4 45.8 20.8 3.05.273.0 83.4 57.468.0Cohere70.1 88.8 46.060.6 49.4 44.6 5.09.093.9 94.2 93.693.9 47.1 20.8 7.38.9BLOOMz 48.6 11.1 0.40.855.5 81.6 14.224.2 48.7 15.8 0.61.296.9 96.8 97.096.9\",\n \"figure_id\": \"tab_16\",\n \"figure_label\": \"14\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Cross-language experiments. Accuracy (Acc) and F1 scores (for the machine-generated class) based on XLM-R on the test sets across different languages generated by ChatGPT. We average the performance across five runs (the standard deviation is shown in parenthesis).\",\n \"figure_data\": \"Generator↓ Test Domain →All domain (en)Baike/Web QA (zh)RuATD (ru)Bulgarian News (bg)Train Domain ↓AccF1AccF1AccF1AccF1All domains (en)95.8 (1.9)96.0 (1.8)79.5 (4.1)82.9 (2.9)60.5 (3.0) 65.3 (5.1)65.8 (3.2)69.3 (6.6)davinci-003 Baike/Web QA (zh)66.4 (7.6)74.8 (4.2)98.9 (0.4)98.9 (0.4)59.5 (0.6) 70.0 (0.6)48.6 (3.3)61.3 (3.7)RuATD (ru)62.8 (3.0)62.0 (8.1)49.6 (9.3)58.6 (3.2)95.3 (1.6) 95.4 (1.4)86.5 (5.1)86.0 (6.5)Bulgarian News (bg)64.8 (3.1)67.2 (9.1)59.0 (8.7)29.4 (23.6) 59.0 (3.6) 32.0 (11.3) 99.6 (0.2)99.6 (0.2)All96.3 (0.7)96.4 (0.6)98.7 (0.5)98.7 (0.5)92.8 (2.1) 93.2 (2.0)85.2 (3.2)87.0 (2.3)All domains (en)90.2 (0.9)89.4 (1.0)93.0 (0.9)92.6 (1.1)54.1 (1.8) 51.5 (5.2)66.0 (3.2)64.3 (7.6)ChatGPTBaike/Web QA (zh)61.6 (5.5)72.2 (2.8)93.5 (1.1)93.1 (1.2)58.8 (2.2) 67.7 (3.7)57.7 (3.4)65.0 (5.0)RuATD (ru)56.7 (3.0)68.6 (0.5)75.7 (7.6)67.5 (14.5) 84.7 (3.9) 82.4 (5.8)82.2 (4.5)84.9 (3.2)Bulgarian News (bg)74.2 (4.9)75.1 (2.2)78.3 (11.2) 70.1 (21.1) 53.8 (1.5) 15.5 (5.8)95.4 (1.3)95.3 (1.4)IDN (id)61.0 (14.3) 29.5 (37.4) 55.6 (7.7)17.5 (23.6) 50.6 (0.8) 5.1 (7.0)58.7 (13.9) 23.6 (35.0)Urdu-News (ur)50.0 (0.1)66.6 (0.1)50.8 (0.7)67.0 (0.3)50.0 (0.0) 66.7 (0.0)50.2 (0.2)66.8 (0.1)Arabic Wikipedia (ar) 72.8 (4.7)77.0 (2.8)83.9 (6.9)85.5 (5.1)62.0 (2.3) 70.2 (1.1)64.6 (5.9)73.6 (3.0)All91.3 (0.6)90.8 (0.6)94.5 (1.2)94.3 (1.4)86.1 (2.5) 85.4 (2.9)82.6 (2.2)84.9 (1.5)\",\n \"figure_id\": \"tab_18\",\n \"figure_label\": \"16\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Cross\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_19\",\n \"figure_label\": \"17\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Data format of M4: English examples sampled across different domains and LLM generators. a joust against the captain of his Scottish Guard, Gabriel, the Count of Montgomery. The fateful run occurred at the end of a tournament day, after Montgomery had almost unhorsed the king. Henry instead on another tilt. Montgomery's lance struck the king's helmet and shattered, with a long splinter running through the king's visor through his eye and into or near his brain. The king initially survived injury, and was attended to by two of the most celebrated physicians in Europe, Ambroise Paré and Andreas Vesalius. The queen, Catherine de Medici, ordered four prisoners executed with wood driven into their brains so that the physicians would have the chance to study the king's wound in detail on the corpses. Despite this rather extraordinary measure, the king deteriorated steadily. Vesalius' personal account is consistent with the development of meningitis or encephalitis. After 11 days, the king died. During these 11 days, Montgomery is supposed to have come to the king's side, and asked to have his head and right hand cut off in punishment. The king told him that he had jousted well and bravely and that the accident was not his fault. Following Henry's death, Catherine essentially ruled through a series of three of her sons. Montgomery retired to his estate in Normandy. From there, his history gets a bit complicated. He converted to Protestantism and took part in the religious fighting in Bourges and Rouen that followed the St. Bartholomew's Day Massacre in 1572. He escaped to England and remained there for about a year, later returning to attempt to lead a Huguenot insurrection in Normandy. He was captured by royal forces and executed in 1574.\",\n \"figure_data\": \"FieldContentLanguageEnglishLLM Generator CHATGPTDomainReddit ELI5QuestionWhat happened to Henry II's jousting opponent after accidentally killing him in 1559?HumanHenry died in\",\n \"figure_id\": \"tab_20\",\n \"figure_label\": \"18\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"PromptPlease, generate wikihow article with length above 1000 characters from title 'How to Buy Virtual Console Games for Nintendo Wii' and headline 'Find a few Wii Points cards from game retailers like GameStop. Make sure your Wii is online and on a secure connection if possible.,Open the Wii Shop Channel, and click Add Wii Points at the bottom right of your screen after the channel finishes loading. Click \\\"From Card\\\". Now, your Wii should be asking for a code with 5 segments, separated by dashes. Enter this code into the Wii now. After a few prompts, you will have Wii points added to your Wii, which is what you use to purchase games. Click \\\"Virtual Console\\\". Choose how you want the games to be displayed. Scroll through the list for games that you want.Click on a game you want to play, And hit Yes on the confirmation box. Wait while the game is downloaded. After you have bought the games you want to play, go back to the Wii Menu. Select the new games you just bought, and start them like a regular channel. Open the Wii Shop Channel, and click Add Wii Points. Click \\\"With a Credit Card\\\". Key in your Credit Card Number. Let Nintendo authorize the purchase.,Go back to the Points Card process above and skip down to Step 7. Buy Virtual Console Games for Nintendo Wii Most people know that Nintendoś library of classic titles is available on the Wii platform through the Virtual Console. Playing these games from years gone by is easy, but there are a few steps you have to take to purchase them. Hereś what you have to do to buy Virtual Console games for your Nintendo Wii. Find a few Wii Points cards from game retailers like GameStop. You ĺl find that most stores that carry video games have them available, and you usually can select cards of different amounts. Choose the ones that are going to give you what you need. Make sure your Wii is online and on a secure connection if possible. Going online with your Wii allows you to access all the Virtual Console titles that Nintendo has made available. To do this, select the Wii Icon from the main screen, then go to the settings area, select the Internet tab, then go to the Connect to the Internet option. Select an existing connection if you have one, or go through the Wiiś setup menu to make a new connection. Open the Wii Shop Channel, and click Add Wii Points at the bottom right of your screen after the channel finishes loading. On the next screen, you have two choices; one from a card, or from a credit card. Click \\\"From Card\\\". Now, your Wii should be asking for a code with 5 segments, separated by dashes. This code is what you use to transfer the money you have to the Virtual Console shop. Enter this code into the Wii now. After a few prompts, you will have Wii points added to your Wii, which is what you use to purchase games. Click \\\"Virtual Console\\\". This option allows you to scroll through all the different games that have been made available for the Virtual Console. Choose how you want the games to be displayed. You can sort them alphabetically, or by genre, or even by release date. Scroll through the list for games that you want. When you find one, click on it and you can see the details and the cost of the game. Click on a game you want to play, And hit Yes on the confirmation box. You then have to wait while the game is downloaded onto your hard drive. This will take a few minutes, depending on the size of the game you want. After you have bought the games you want to play, go back to the Wii Menu. Select the new games you just bought, and start them just like a regular channel. If you want to add more funds to your Wii point wallet, go back to step 3. But this time instead of clicking \\\"From a card\\\", click on the \\\"With A credit card\\\" option. Key in your Credit Card Number. Be sure that it is entered correctly, because many purchases from the Virtual Console are automated and there are no refunds if you enter your number incorrectly. Let Nintendo authorize the purchase. A few steps later you should be notified of the success of your purchase, and you should now have more Wii points with which to buy more games. Go back to the Points Card process above and skip down to Step 7. This is the step right after you have verified that you have the points. Here, you can select a game, and start playing it. Now, you know how to purchase and download games from the Virtual Console. Spend some time going through the list, purchasing some of the classics like Super Mario and the Legend of Zelda, and for a small fee, you can start playing them on your Wii. Have fun!\",\n \"figure_data\": \"FieldContentMachine How to Language EnglishLLM Generator CHATGPTDomainWikihowTitleHow to Workaround Damaged Laptop Display Hinges\",\n \"figure_id\": \"tab_21\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"If it's still loose, you may need to repeat steps 3-5 with stronger epoxy glue and larger strips of Velcro. 7. Tie a fishing line loop around the laptop. Take the fishing line and tie it into a loop which exactly encircles the laptop when the display is open to the working position. Be sure to check that the fishing line loop is strong enough to prevent the display from falling into the fully open position. 8. Glue foam tape to the laptop edge. To finalize the repair, glue a small strip of two-sided foam tape on the edge of the laptop where the fishing line should rest. Optionally, you can also use a plastic bag closure to adjust the exact size of the fishing line loop and the working position of the laptop display. With these simple steps, you can work around damaged laptop display hinges and fix the issue without having to replace the entire hinge. Keep in mind that this repair method may not work for all laptop models, so be sure to check if your laptop can be repaired in this way before getting started.\",\n \"figure_data\": \"FieldContentLanguageEnglishLLM Generator CHATGPTDomainPeerReadTitleBridge Text and Knowledge by Learning Multi-Prototype Entity Mention EmbeddingGood luck!\",\n \"figure_id\": \"tab_22\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Examples of other languages across different domains by ChatGPT and davinci-text-003. Игоря, он согласился со мной, что «она, по-видимому, хороший человек». Но тут же прибавил: -Разве вы не видите, тетя Вава, что мне лучше всего одному, как я теперь живу. Дико и оскорбительно представить себе вот в этих стенах какую-то другую жену, кроме той, которая у меня отнята смертью. День праздника был посвящен пути движения Игоря, его сына. Взгляд, которым его судьба была окрашена, был властный, а к мимолетным правам на близость был дан официальный статус. Желание судить и осуждать пару, видимо, комплексное, так как сразу трое -мужчина, женщина и в какой-то мере дети, занимаются браконьерством. Даже брутальное отношение главы дома к ним не могло исказить более-менее нормальное настроение.Одержимость жены, ее свежесть, девичья улыбка с ямочками на лице, а также твердый взгляд серых глаз очень привлекла мягкосердечного, но застаревшего мужа. Назад вспомнили, как 4 года назад он встретился с красавицей, которая должна была привлечь его интерес, по мнению Игоря, в душевном мире и представлении жизни. Однако, овладение собой было столь сильно, что он дошел до того, что «лучше для него жить в одиночестве, чем представлять себе другую жену внутри этих чертовых стен». Даже сейчас после потери жены, чья память приукрасит и облегчит жизнь Игоря и его сына, до разрушительных, можно догадываться, или загнутых ходов, они не доводятся. Няма да допуснем да продължат. Ако се премахнат буните за един месец ще изчезне целият плаж\\\", каза един от протестиращите пред Флагман.бг.Наш репортер се свърза и със стопанина на плажа -бургаската фирма \\\"Акра лидо\\\" ООД, която има 5-годишен договор за ивицата и плаща най-скъпия плажен наем в България -по 107 хил.лева на година за изключително малка площ.От компанията останаха изненадани от действията на строителния наздор. Буните са там от много години, спират свлачищния процес и осигурят плажна ивица, макар и много тясна.Само преди седмица от \\\"Акра лидо\\\" са платили накуп наемната цена за 2017 година, а сега ако бъдат премахнати буните ще останат без плаж.Скандалът ще има международен отзвук, защото 500 руски семейства имат жилища в непосредствена близост до къмпинг \\\"Черноморец\\\" и са ги закупили заради това, че плажът им е само на няколко метра.Това е друга от машините, която ще участва в акцията по разрушаването на буните\\\"Ние сме изрядни платци, наели сме плажа в това състояние от държавата и до тази сутрин смятахме, че всичко е изрядно. Не сме поставяли под съмнение, че държавата ни е отдала под наем плаж с незаконни буни\\\", каза един от собствениците на фирмата Милен Янков пред Флагман.бг.Припомняме, че миналата година също имаше много проблеми с този плаж. В основата бяха отношенията със стария стопанин на ивицата (ВИЖ НАШАТА ТЕМА В РАЗВИТИЕ)ОЧАКВАЙТЕ ПОДРОБНОСТИ! В категории: Созопол Докато правните изследвания са все още в процес на продължаване за да се разбере точно какви са причините за багера, се очаква тази политика да продължи и да се извършват постоянни разрешения за някои интереси. Някои пленници на ансамбъла на бунита подози, че максимумът на разрешенията е 100, които могат да се дават годишно според правилата, които сме предвидили. Защо РДНСК решил да направи това и защо именно сега, това продължава да бъде неизвестно. Но понеже това се оказва наистина забранено, много от плавателните клубове и индивиди на плажа ще бъдат принудени да се преместят на други места. Бановете, разположени в гледната точка за потоците, да се зачитат на рилско море -това ли ще стане след махането на буните? Ще има ли някаква защита за плавателните инсталации, една от важните атракции на Golden Sands? Ще плажът продължи да съществува или ще бъде претърпял значителни изменения? Тези и много други въпроси очакват отговори. За да се оцени тайнственото поведение на РДНСК, много информация трябва да се прецени. Ето защо е особено важно да се осигури правопорядъчно намиране на логичен разследвателен процес за да се разберат всички мотиви за действието на РДНСК и да се донесат промени, които да се правят в добро на всички страни.HumanОбикновените хора се обръщат към кралица Елизабет Втора с \\\"Нейно кралско величество\\\" и придружават обръщението с поклон и израз на любезност. Но зад кулисите отношенията са доста по-неформални, разказват от \\\"Новини Лондон\\\". Потребител на сайта Quora зададе въпрос: \\\"Как внуците на кралица Елизабет се обръщат към нея на четири очи?\\\". Макар че отговорите варират, проучване на \\\"Индипендънт\\\" показва, че внуците на кралицата са й измислили обръщения, изразяващи обичта им към нея. Уилям и Хари окуражаваха баба си при нейната неочаквана, спираща дъха роля на церемонията по откриването на Олимпийските игри през 2012 г. При вида на баба им, която се появи заедно с Даниел Крейг като Джеймс Бонд, принцовете извикаха с цяло гърло: \\\"Давай, бабо!\\\". Но това не е единственото име, използвано от Уилям и Хари. Според светския хроникьор Ричард Кей веднъж в Бъкингамския дворец Уилям казал на баба си \\\"Гари\\\". Когато служители на двореца попитали \\\"Кой е Гари?\\\", кралицата отговорила: \\\"Аз съм Гари\\\". Уил още не се е научил да казва \\\"Granny\\\" (бабоанглийски език, бел. ред.), обяснила Елизабет. Тригодишният правнук на Елизабет Втора принц Джордж също има любвеобилно обръщение към кралицата, както стана ясно от интервю, излъчено по телевизия ITV миналото лято по повод 90годишнината й. Херцогинята на Кеймбридж Кейт Мидълтън сподели, че Джордж нарича баба си \\\"Ган Ган\\\". Но това не е просто мило обръщение, а е израз на дълготрайна кралска традиция. Американският биограф Кити Кели твърди, че принц Чарлз, Уилям и Хари също се обръщали така към своята прабаба. Prompt Моля, генерирайте статия с дължина над 1000 символа и заглавие \\\"Неподозирани разкрития за кралското семейство\\\" Machine Неподозирани разкрития за кралското семейство Кралските семейства винаги са били обект на внимание и интерес от страна на медиите и обществото. Известни са скандалите и скритите истории, които остават извън зрителния ъгъл и несъмнено експлоатирането на подобни теми продължава да държи високи места в таблоидите. Но дали знаете всичко за кралските семейства, което медиите не съобщават? Има ли неща, които остават скрити дълбоко в историята или в семейните тайни? Изглежда, отговорът е да. В последните години имаше повече и повече изисквания за откриване на тайните на кралските династии, но много от тези разкрития остават скрити за публика. Една от най-известните скрити истории, свързани с британското кралско семейство, е тази на принца Джон, третият син на крал Джордж V и кралица Мери. Принц Джон беше диагностициран с епилепсия и склероза на 7-годишна възраст. Той беше изолиран от обществото и от своето семейство и беше принуден да живее в бездейственост и самота, за да не покаже \\\"ненормалното\\\"си поведение. Принц Джон почина на 13-годишна възраст, покрит с тайните и позволено му беше да забрави света на живите. Разказите за скритата история на Американското кралско семейство са също известни. Съществува теория, че Сежър, който се хвали, че е третият наследник на трона след Алфонсо и Джон, наистина има връзки с кралското семейство на САЩ. През 1937 г. Сежър заяви, че е бил роден като дете на Сесилия, дъщерята на крал Едуард VIII и Уолис Симпсън. Макар това да е слух, наличието на тайни и тъмни истории в кралските семейства е добре документирано. Все още има много да се разкрие, но кралските династии на света ще продължат да привличат интриги и вълнение през годините. CNN Indonesia -Musisi Ahmad Dhani memenuhi panggilan Polda Metro Jaya untuk diperiksa sebagai saksi tersangka dugaan makar Sri Bintang Pamungkas, Selasa (20/12). Berdasarkan pantauan CNNIndonesia.com, dia tiba pukul 15.00 WIB. Sedangkan tim kuasa hukumnya yang tergabung dalam Advokat Cinta Tanah Air sudah tiba satu jam sebelum kedatangannya. Tak lama kemudian, Farhat Abbas juga datang untuk mendampingi Ahmad Dhani. Ahmad Dhani mengatakan, dirinya tidak akan memberikan informasi apapun soal Sri Bintang. Dia mengklaim tidak kenal dengan Sri Bintang. Buni Yani dan Ahmad Dhani Jadi Saksi Kasus Sri Bintang Besok Saksi Mengaku Dapat Aliran Dana dari Tersangka Makar Buni Yani Diperiksa Soal Pidato Sri Bintang di Kalijodo \\\"Informasi pasti tidak ada, karena saya tidak kenal dengan Sri Bintang Pamungkas. Saya pernah ketemu beliau ketika di Mako Brimob,\\\" ucapnya. Meski demikian, Ahmad Dhani mengaku hadir saat ada pertemuan di Universitas Bung Karno. Namun, ia mengklaim tidak mendengar pidato yang disampaikan oleh Sri Bintang karena terlambat datang. Awalnya, Kepala Subdirektorat Kejahatan dan Kekerasan Direktorat Reserse Kriminal Umum Polda Metro Jaya AKBP Hendy Kurniawan mengatakan, karena Ahmad Dhani sedang sakit maka pemeriksaan terhadap dia yang akan dilakukan hari ini harus ditunda Kamis (22/12). \\\"Ahmad Dhani ditunda karena sakit. Surat sakit sudah kami terima. Kami sudah koordinasi dengan kuasa hukumnya kemarin,\\\" ujarnya. Kuasa Hukum Ahmad Dhani, Ali Lubis mengklaim, kliennya tersebut merupakan warga negara yang baik sehingga datang memenuhi panggilan kepolisian. \\\"Beliau kooperatif, beliau ingin membantu kepolisian jadi lebih cepat lebih baik,\\\" ucapnya. Ahmad Dhani telah ditetapkan sebagai tersangka dugaan penghinaan terhadap penguasa. Dia ikut ditangkap pada Jumat (2/12) dengan sejumlah tersangka dugaan makar lainnya. Sepuluh tersangka dugaan makar itu adalah Sri Bintang, Kivlan Zein, Adityawarman Thahar, Ratna Sarumpaet, Firza Huzein, Eko Santjojo, Alvin Indra, Rachmawati Soekarnoputri, dan kakak beradik Rizal Kobar dan Jamran.\",\n \"figure_data\": \"FieldContentLanguageArabicLLM Generator CHATGPTDomainWikipediaTitleHuman\",\n \"figure_id\": \"tab_23\",\n \"figure_label\": \"20\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Data Source\", \"Citation\": \"(Mitchell et al., 2023)\", \"Explanation\": \"The cited work provides a benchmark dataset for evaluating the performance of human and machine-generated text classification models.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wang et al., 2024)\", \"Explanation\": \"The cited work serves as the basis for SemEval-2024 Task 8, providing a dataset for detecting machine-generated texts in a black-box scenario.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Szyller et al., 2021)\", \"Explanation\": \"The cited work by Szyller et al. (2021) provides a watermarking technique that is used as a method for white-box detection in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(He et al., 2022)\", \"Explanation\": \"The work by He et al. (2022) contributes to the white-box detection techniques by providing another watermarking method for LLM detection.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Kirchenbauer et al., 2023)\", \"Explanation\": \"The cited work by Kirchenbauer et al. (2023) is another contribution to the white-box detection methods by introducing a new technique for LLM detection.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Zhao et al., 2023)\", \"Explanation\": \"The work by Zhao et al. (2023) further adds to the white-box detection techniques by providing a new method for LLM detection.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Krishna et al., 2022)\", \"Explanation\": \"The cited work by Krishna et al. (2022) contributes to the white-box detection by focusing on the expected per-token log probability of texts in LLM detection.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Mitchell et al., 2023)\", \"Explanation\": \"The work by Mitchell et al. (2023) is another contribution to the white-box detection methods by focusing on the expected per-token log probability of texts in LLM detection.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Uchendu et al., 2021)\", \"Explanation\": \"The cited work provides the TuringBench dataset, which is a collection of human- and machine-generated text from generative models, that the citing paper uses in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Guo et al., 2023)\", \"Explanation\": \"The cited work presents the HC3 dataset, which is a collection of questions and their corresponding answers from human experts and ChatGPT, that the citing paper uses in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Shijaku and Canhasi, 2023)\", \"Explanation\": \"The cited work provides the TOEFL essays written by examined people and generated by ChatGPT, that the citing paper uses in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Shamardina et al., 2022)\", \"Explanation\": \"The cited work presents the RuATD Shared Task 2022 data, which includes artificial texts in Russian generated by various language models, that the citing paper uses in their research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zellers et al., 2019)\", \"Explanation\": \"The cited work by Zellers et al. (2019) has served as a foundational study in detecting machine-generated texts in a specific language and domain. The citing paper builds upon this work by broadening the scope to include multiple languages and a variety of LLMs across different domains.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Macko et al., 2023)\", \"Explanation\": \"The cited work by Macko et al. (2023) provides a specific case study in detecting machine-generated texts in a particular language and domain. The citing paper uses this study as a reference point to support the broader research on detecting machine-generated texts in multiple languages and a variety of LLMs across different domains.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Gehrmann et al., 2019)\", \"Explanation\": \"The cited work by Gehrmann et al. (2019) introduces the concept of GLTR-like word rankings as a statistical distribution feature in detecting machine-generated texts. The citing paper acknowledges this work as the origin of the word ranking feature used in the study of detecting machine-generated texts in multiple languages and a variety of LLMs across different domains.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Shijaku and Canhasi, 2023)\", \"Explanation\": \"The cited work by Shijaku and Canhasi (2023) presents a method for detecting machine-generated texts using statistical distributions as features. The citing paper adopts this method in the study of detecting machine-generated texts in multiple languages and a variety of LLMs across different domains.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Tang et al., 2023)\", \"Explanation\": \"The cited work by Tang et al. (2023) highlights the use of fact-verification features in detecting machine-generated texts. The citing paper acknowledges this work as the origin of the fact-verification feature used in the study of detecting machine-generated texts in multiple languages and a variety of LLMs across different domains.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Koupaee and Wang, 2018)\", \"Explanation\": \"The cited work provides the WikiHow dataset as a source of human-written texts for the English language in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Kang et al., 2018)\", \"Explanation\": \"The cited work serves as the source of the PeerRead dataset for the English language in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Shamardina et al., 2022)\", \"Explanation\": \"The cited work is the source of the RuATD dataset for the Russian language in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Conover et al., 2023)\", \"Explanation\": \"The cited work is the source of the Dolly-v2 model for machine generation in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Muennighoff et al., 2022)\", \"Explanation\": \"The cited work is the source of the BLOOMz 176B model for machine generation in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Clark et al., 2021)\", \"Explanation\": \"The cited work by Clark et al. (2021) provides evidence that it is hard for human evaluators to distinguish between machine-generated and human-authored text, which supports the claim made in the citing paper that it is not easy for humans to detect machine-generated text.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Clark et al., 2021)\", \"Explanation\": \"The cited work by Clark et al. (2021) is extended in the citing paper to explore the impact of training on the accuracy of human evaluators in distinguishing between machine-generated and human-authored text across different domains.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Clark et al., 2021)\", \"Explanation\": \"The cited work by Clark et al. (2021) provides evidence that the improvement in accuracy across domains in identifying machine-generated text is not uniform, which supports the claim made in the citing paper that the improvement in accuracy across domains differs.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Clark et al., 2021)\", \"Explanation\": \"The cited work by Clark et al. (2021) provides evidence that the accuracy of human evaluators in identifying machine-generated text is lower in the arXiv domain compared to the Reddit domain, which supports the claim made in the citing paper that it is harder to classify texts from the arXiv domain than from the Reddit domain.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2019)\", \"Explanation\": \"The RoBERTa model is the basis for the detector used in the citing paper, providing the foundational model for detecting machine-generated texts.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Clark et al., 2020)\", \"Explanation\": \"The ELECTRA model is also used as a detector in the citing paper, and its pre-training objective is a methodological basis for the MGT task.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Conneau et al., 2019)\", \"Explanation\": \"The XLM-RoBERTa model is a variant of the RoBERTa model and is used as a detector in the citing paper, providing a methodological basis for the detection of machine-generated texts.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Gehrmann et al., 2019)\", \"Explanation\": \"The cited work provides the GLTR features that the citing paper uses to train a logistic regression model for distinguishing machine-generated texts from human-written ones.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Li et al., 2014)\", \"Explanation\": \"The cited work provides a set of stylistic features that the citing paper uses to train an SVM classifier for text classification.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Horne et al., 2019)\", \"Explanation\": \"The cited work provides a set of features from the NELA system that the citing paper uses to train an SVM classifier for text classification.\"}, {\"Category\": \"Data Source\", \"Citation\": \"Graham et al., 2012\", \"Explanation\": \"The cited work provides the Moral Foundation Theory that the citing paper uses in the NELA features to train an SVM classifier for text classification.\"}, {\"Category\": \"Data Source\", \"Citation\": \"GPTZero\", \"Explanation\": \"The cited work is a pre-existing system that the citing paper uses without adaptation to perform text classification.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Ribeiro et al., 2016)\", \"Explanation\": \"The cited work by Ribeiro et al. (2016) provides a method for feature analysis, which the citing paper uses to conduct a study on in-domain detectors and their reliance on word distribution for classification.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Wang et al., 2024)\", \"Explanation\": \"The cited work by Wang et al. (2024) is an extension of the M4 dataset presented in the citing paper, featuring additional languages, domains, and three new task (re)formulations for SemEval-2024 Task 8.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Tang et al., 2023)\", \"Explanation\": \"The cited work by Tang et al. provides evidence that machine-generated text detection might become more challenging as LLMs evolve, which supports the claim made in the citing paper that the gap between machine-generated and human-written text might narrow.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Sadasivan et al., 2023)\", \"Explanation\": \"The cited work by Sadasivan et al. also contributes to the claim made in the citing paper that machine-generated text detection might become more difficult as LLMs evolve, by showing that the gap between machine-generated and human-written text might narrow.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Liang et al., 2023)\", \"Explanation\": \"The cited work by Liang et al. further supports the claim made in the citing paper that the gap between machine-generated and human-written text might narrow, by showing that LLMs might evolve in a way that makes machine-generated text detection more challenging.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Fan et al., 2019)\", \"Explanation\": \"The cited work is the source of the Reddit ELI5 dataset used in the citing paper for open-domain and long-form abstractive QA research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Koupaee and Wang, 2018)\", \"Explanation\": \"The cited work is the source of the WikiHow dataset used in the citing paper to build a collection of articles for research purposes.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kang et al., 2018)\", \"Explanation\": \"The cited work is the source of the PeerRead corpus used in the citing paper to sample academic papers and human-written reviews for research analysis.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Kang et al., 2018)\", \"Explanation\": \"The citing paper extends the research of the cited work by using the PeerRead corpus to sample papers and human-written reviews for further analysis and study.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Shamardina et al., 2022)\", \"Explanation\": \"The cited work is the source of the Russian RuATD dataset used in the citing paper for artificial text detection in Russian.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Shavrina and Shapovalova, 2017)\", \"Explanation\": \"The cited work provides a corpus of news articles that the citing paper uses in its research on the construction of human-written data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Melnichenko and Tyshkevich, 2017)\", \"Explanation\": \"The cited work contributes a corpus of personal diaries that the citing paper utilizes in its study of human-written data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ivanin et al., 2020)\", \"Explanation\": \"The cited work provides a government document corpus that the citing paper uses in its research on the construction of human-written data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Hussain et al., 2021)\", \"Explanation\": \"The cited work provides the Urdu News Data 1M dataset that the citing paper uses to gather news articles from four reputable news agencies in Pakistan.\"}]"}}},{"rowIdx":94,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_0"],"heading":"Introduction","publication_ref":["b1","b21","b3","b21","b31","b16","b25","b16","b29","b16","b25"],"table_ref":[],"text":"Large language models (LLMs) (Devlin et al., 2019;Brown et al., 2020) are capable of generalizing to novel tasks (Brown et al., 2020) by conditioning on textual prompts consisting of a few task examples. This training-free paradigm of fewshot inference, known as in-context learning (ICL), reduces the cost of modeling new tasks while also providing an interpretable and customizable interface (Liu et al., 2022;Wei et al., 2023) and improving generalization (Anil et al., 2022;Qiu et al., 2022b;Drozdov et al., 2023) and reasoning skills (Wei et al., 2023). However, ICL performance is critically sensitive to the choice of demonstrations (Zhao et al., 2021;Liu et al., 2022;Lu et al., 2022;Rubin et al., 2022; Schick and Schütze, 2021), as the LLM relies on them for understanding and solving the test instance.\nThe standard approach to selecting ICL examples or demonstrations from a pool of candidates is to independently score them using a relevance metric and choose the top-ranked ones. However, cosine similarity and BM25, the two commonly used metrics, are sub-optimal for selecting demonstrations due to their reliance on a single dense embedding and unigram overlap, respectively. Moreover, since it selects examples independently, this approach ignores their utility as a set. It is particularly inadequate for complex compositional tasks like semantic parsing (Levy et al., 2022) where no single candidate might contain all reasoning patterns, and an independent selection would select multiple redundant examples with the same reasoning patterns but fail to demonstrate the others. Figure 1 shows a failure case where similarity-based selection picks paraphrased examples that fail to demonstrate how to find a manager. Prior work on selecting demonstrations as a set (Ye et al., 2023;Levy et al., 2022) required task and/or LLM-specific training, limiting their utility. For this reason, simple yet widely applicable training-free methods like BM25 and cosine similarity remain the most popular approaches for ICL example selection.\nIn this work, we propose a novel framework for selecting sets of maximally informative demonstrations for the salient aspects of the test input, e.g., reasoning patterns, entities, etc. Examples selected using this framework are informative about the test input and help the LLM understand and perform the task. We use this framework to explore different ways to characterize salient aspects, including syntactic structures like dependency parse subtrees and contextual token embeddings, while using BM25 and BERTScore (Zhang et al., 2020) to measure their coverage, respectively. To select the demonstrations as a set, we extend the coverage metrics to measure the overall informativeness of a set of demonstrations. We show that these set-level metrics are submodular and can be efficiently optimized to find demonstration sets that maximally cover the salient aspects. We evaluate our ICL example selection methods on 15 diverse datasets, including 6 semantic parsing, 2 numerical reasoning, and 7 classification datasets, and with 7 LLMs of varying sizes and pretraining. Among instance-level metrics, BSR, the recall version of BERTScore, consistently outperforms standard retrieval metrics on all datasets and LLMs, beating cosine similarity by up to 8 points on average in semantic parsing datasets and 15 points in the rest. Selecting demonstrations as a set using SET-BSR, the set-extension of BSR, leads to further gains in semantic parsing and is particularly effective in compositional settings where the gains grow with LLM size. With Codex, a 175B parameter LLM, SET-BSR outperforms cosine similarity by 17% on average with up to 49% improvement in some splits, and, despite being training-free, outperforms even trained methods like those from Rubin et al. (2022), Levy et al. (2022), and Ye et al. (2023) that require task and/or LLM-specific training."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b21","b3","b21","b31","b16","b26","b24","b18","b16","b25","b16","b25","b15"],"table_ref":[],"text":"In-context learning for few-shot inference facilitates the use of LLMs for novel tasks without the need for expensive supervised fine-tuning. In addition to reduced cost, it has several other advantages over supervised fine-tuning: it provides a more interpretable and customizable interface to using LLMs (Liu et al., 2022;Wei et al., 2023); and retention of linguistic understanding and knowledge from pretraining leading to improved generalization (Anil et al., 2022;Qiu et al., 2022b;Drozdov et al., 2023) and reasoning skills (Wei et al., 2023).\nHowever, the performance of ICL is critically sensitive to the choice of demonstrations (Zhao et al., 2021;Liu et al., 2022). This has led to a growing interest in techniques for selecting good demonstrations. Prior work can be roughly classified into (1) independently scoring and retrieving examples (Liu et al., 2022;Rubin et al., 2022), (2) selecting diverse examples to reduce redundancy among them (Su et al., 2022;Levy et al., 2022;Agrawal et al., 2022;Ye et al., 2022), and (3) selecting examples that minimize the entropy of the LLM's output distribution for the test input (Lu et al., 2022;Wu et al., 2023). Recent work has also trained RL agents (Lu et al., 2023) and used Bayesian inference (Wang et al., 2023).\nThe most similar studies to ours are Levy et al. (2022) and Ye et al. (2023). Levy et al. (2022) select diverse demonstrations that cover substructures of the target output predicted by task-specific classifiers but are limited in applicability to a few semantic parsing tasks. Ye et al. (2023) use Determinantal Point Processes (Kulesza, 2012) to select a diverse set of demonstrations similar to the test instance but do not optimize for coverage directly and require training with the LLM. Moreover, both methods require task or LLM-specific training that limits their use and effectiveness for larger LMs."},{"figure_ref":[],"heading":"Preliminaries","publication_ref":[],"table_ref":[],"text":"In-context learning is the ability of LLMs to solve novel tasks by merely conditioning on a few task demonstrations. Formally, given demonstrations {(x i , y i )} k i=1 and the test input x test , it involves using textual templates to linearize instance inputs and outputs into sequences of tokens from the LLM vocabulary, x = I(x) = ⟨x 1 . . . x |x| ⟩ and y = O(y) = ⟨y 1 . . . y |y| ⟩. The linearizations are then concatenated to form a prompt and fed to the LLM for conditional generation of the test output:\ny test ∼ P LM (• | x 1 , y 1 , . . . , x K , y K , x test ) (1)\nThe interpretable and training-free nature of ICL makes it an attractive alternative to supervised finetuning. However, its performance is highly sensitive to the choice and order of demonstrations.\nDemonstration Selection identifies which examples to include in the prompt for any test instance. Formally, given a test input x test and a pool of candidates\nT = {z i } N i=1 = {(x i , y i )} N i=1\n, the goal is to select a subset of k ≪ N demonstrations that when included in the context make y test the most likely generation. A naive approach is to randomly sample k instances from T , but this is sub-optimal since the demonstrations are often completely unrelated to the test input. Instead, the standard approach to selecting demonstrations that are informative about the test input is to independently assign each candidate z a score score(x test , z) using a relevance metric and then select the top k candidates."},{"figure_ref":[],"heading":"Relevance Metrics","publication_ref":["b12"],"table_ref":[],"text":"The two most commonly used relevance metrics for scoring demonstration are cosine similarity and BM25. Cosine similarity uses a representation function R to independently map the textual linearizations of inputs to unit-norm embeddings r x = R(x) in a common vector space and then scores the candidate z using the dot product, cosine(x test , z) = r T xtest r z . BM25, on the other hand, is a sparse information retrieval algorithm belonging to a class of TF-IDF measures that view the test input and the candidates as bags of terms and measures relevance as a weighted recall or coverage of these terms:\ntfidf(x test , z) = s∈Tx test idf(s)tf(s, T z ) (2)\nHere T x and T z are the set of terms in x and z respectively, and tf(s, T z ) and idf(s) are the term frequency and inverse document frequency statistics that measure the coverage of a particular term and the relative importance of terms respectively. We use tf and idf as per the Okapi variant of BM25 (Robertson et al., 1993;Jones et al., 2000)."},{"figure_ref":[],"heading":"Informative Demonstrations","publication_ref":["b29"],"table_ref":[],"text":"The limitation of the standard demonstration selection approach is that by independently scoring the demonstrations, it ignores their utility as a set. For ICL to work, the demonstrations included in the context need to be informative about how to understand and solve the test input. In this section and the next, we describe our approach to selecting informative sets of demonstrations for ICL. We begin by defining our notion of informativeness of demonstrations in ICL and describing how to measure it. Thereafter, in §5, we will discuss how to extend this notion to an algorithm for selecting optimally informative sets of demonstrations.\nInformativeness Demonstrations should demonstrate the salient aspects, e.g., reasoning patterns, entities, etc., of the test input. Formally, denoting S xtest as the set of salient aspects of the test input, we measure the informativeness of a demonstration z in terms of the coverage of such salient aspects,\ncover (x test , z) = s∈Sx test c(s, z)(3)\nwhere c(s, z) measures the coverage (or recall) of a single salient aspect s by z.\nSalient Aspects Both cosine similarity and BM25 are special cases of Eq. 3 for different notions of salient aspects. For BM25, S xtest = T xtest , the set of unigrams in x, and c(s, z) = idf(s)tf(s, T z ).\nAnd cosine similarity, although not explicitly a recall metric, can also be interpreted as evaluating coverage of the dimensions of the test input embedding by defining S xtest = [1, d], the dimensions of the dense embedding as the salient aspects, i.e., cosine(x test , z)\n= d s=1 r xtest [s] • r z [s] (4)\nThe above interpretations reveal why neither cosine similarity nor BM25 are good measures of informativeness. While cosine similarity captures some aspects of semantic similarity (depending on the embedding), it is limited to a single embedding. And, unigrams, the commonly used terms with BM25, are too small to capture most salient aspects. A good measure of informativeness necessitates an accurate characterization of salient aspects. One way might be to use larger syntactic substructures of the input as terms with BM25. We experiment with using larger n-grams and subtrees of the dependency parse tree. However, such syntactic structures are constrained to the surface form of the instance and hence may not capture meaning and aspects like reasoning patterns. A better way to capture salient aspects is to use contextualized token embeddings, the idea behind the BERTScore (Zhang et al., 2020) metric.\nBERTScore was originally proposed as a metric for evaluating the quality of machine-generated text (e.g., machine translation) by comparing it to a reference text. It leverages pre-trained contextual embeddings to match words in the candidate and reference sentences by cosine similarity and compute precision, recall, and F1 measures. Formally, given the sequences of contextual embeddings ⟨x 1 , x 2 , . . . , x |x| ⟩ and ⟨z 1 , z 2 , . . . , z |z| ⟩ of tokens in x = ⟨x 1 , x 2 , . . . , x |x| ⟩ and z = ⟨z 1 , z 2 , . . . , z |z| ⟩ respectively, the recall measure, BERTScore-Recall (BSR), is defined as:\nBSR(x, z) = x i ∈x w(x i ) max j x T i z j (5)\nHere, w(x i ) is a weight assigned to token x i and can be defined as 1 |x| if treating each token as equally important or idf(x i )\nx i ∈x idf(x i ) if downweighting rare words. The precision measure is defined analogously, while the F1 measure is the harmonic mean of the two. BSR is also a special case of Eq. 3 with contextualized tokens as salient aspects, i.e., S x = ⟨x 1 , x 2 , . . . , x |x| ⟩ and can be used to select examples by treating them as candidates and the test input as the reference. The following table summarizes the informativeness measures and salient aspects in this work."},{"figure_ref":["fig_0"],"heading":"Metric Salient Aspects","publication_ref":[],"table_ref":[],"text":"Cosine embedding dimensions BM25 unigrams, n-grams, dependency parse subtrees BERTScore contextual token embeddings 5 Set-level Information Coverage So far, we have focused on measuring the informativeness of a single demonstration to rank and independently select the most informative ones. However, as depicted in Fig. 1, when no single single candidate demonstrates all salient aspects, this approach can fail to cover all of them while also selecting redundant demonstrations that provide no new information. A scenario where this can happen is when the candidate pool contains close paraphrases (or duplicates). This suggests that demonstrations should be selected as a set.\nSet Metric To evaluate the informativeness of a set of examples Z, we propose to extend the coverage measure in Eq. 3 to a measure for sets as follows: Intuitively, this measures the coverage of each salient aspect as the best coverage it receives from any example in the set. In other words, maximizing it requires that every salient aspect appears at least once in some demonstration without considering which or how many. Since cosine similarity, BM25, and BSR are all special cases of Eq. 3, they can be extended to set measures using Eq. 6.\nsetcov (x test , Z) = s∈Sx test max z∈Z c(s, z) (6)\nSubmodularity Given the combinatorial space of sets of demonstrations, for a measure on sets to be practical, it needs to be efficiently optimizable. Fortunately, the set-level metric, as defined above, is also submodular for any definition of c(s, z). We prove this in Appendix A. Intuitively, this follows from the facts that (1) for any given test instance, c(s, z) assigns a scalar weight to each demonstration z ∈ Z, (2) the maximum of weights across set elements is submodular, and (3) the sum of submodular functions is also submodular. This means that the set-level metric can be optimized using a greedy algorithm with a constant factor approximation guarantee (Nemhauser et al., 1978).\nAlgorithm The greedy algorithm we use to select the optimal set is shown in Algorithm 1. In every iteration, it selects the example that maximally increases the coverage of the current set of demonstrations (lines 5-9). If no such example exists, it resets (lines 11). Using the following identity when computing the score for candidate sets (line 5), ( 7)\nsetcov x test , Z ∪ z ′ = s∈Sx test max c(s, Z), c(s, z ′ )\nand assuming constant time for computing each c(s, z), the time complexity of algorithm is\nO(kN L), where L = |S xtest |. For BSR, the com- plexity of computing c(x, z) for all z ∈ Z is O(T d),\nwhere T is the total number of tokens in Z and d is the token embedding size. Thus, the time complexity of both instance and set-level BSR is dominated by the computation of c(x, z), and is O(LT d). While slower than cosine and BM25, we found it to be a small overhead to in-context learning for most datasets considered in this work. We discuss this further in App. C."},{"figure_ref":[],"heading":"Experimental Setup","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Datasets","publication_ref":["b28","b10","b0","b19","b23","b17","b4","b22","b2","b30","b5","b16","b5","b7","b27","b25","b8"],"table_ref":[],"text":"We experiment with a total of 15 datasets including six diverse semantic parsing datasets viz. Geo-Query (Zelle and Mooney, 1996), ATIS (Hemphill et al., 1990;Dahl et al., 1994), Overnight (Wang et al., 2015), SMCalFlow (Andreas et al., 2020), BREAK (Wolfson et al., 2020), and MTOP (Li et al., 2021); a math-word problems (GSM8K (Cobbe et al., 2021)) and a machine reading comprehension (DROP (Dua et al., 2019)) dataset requiring multi-step numeric reasoning; and seven classification datasets spanning natural language inference, paraphrase detection and sentiment classification viz. QNLI (Wang et al., 2018), MNLI (Williams et al., 2018), RTE (Bentivogli et al., 2009), MRPC (Dolan and Brockett, 2005), PAWS (Zhang et al., 2019), QQP (Wang et al., 2018), andSST2 (Socher et al., 2013). We refer the reader to App. B for detailed descriptions of each dataset along with sample instances and prompt templates.\nIn addition to the standard IID splits, we also evaluate compositional generalization using compositional splits wherever available. For GeoQuery we use three types of compositional splits: Template (Finegan-Dollak et al., 2018), TMCD (Keysers et al., 2020), and Length. Following Levy et al. (2022), we use the compositional splitsthree Template, three TMCD, and one Lengthgenerated by Qiu et al. (2022a) and average results across the TMCD and Template splits. For ATIS and Overnight, we experiment with Template splits (Finegan-Dollak et al., 2018) generated by Gupta et al. (2022). For SMCalFlow, we experiment with splits in SMCalFlow-CS (Yin et al., 2021): an IID split (8-S) and a compositional split (32-C).\nFor all the splits, following prior work (Ye et al., 2023;Rubin et al., 2022) we randomly subsample 44,000 instances from the train set to use as pool to select demonstrations from. For evaluation, we use a random subsample of 1000 instance of the validation set if available, and the test set otherwise. We use Exact Match (EM) accuracy for all datasets except BREAK where we use LF-EM (Hasson and Berant, 2021), which is preferred over EM for semantic equivalence."},{"figure_ref":[],"heading":"Models","publication_ref":["b6","b14"],"table_ref":[],"text":"We experiment with the following LLMs: GPT-Neo-2.7B (Black et al., 2021): A 2.7B-parameter LM trained on The Pile (Gao et al., 2020), an 825 GB text corpus. LLaMA (Touvron et al., 2023): A collection of LMs ranging from 7B to 65B parameters pretrained on CommonCrawl, GitHub, Arxiv, etc. We experiment with LLaMA-7B and LLaMA-13B. StarCoder (Li et al., 2023): A 15.5B parameter model trained on 80+ programming languages (Kocetkov et al., 2022). GPT-3.5-Turbo2 : 175B LM trained with RL to follow instructions and optimized for chat. Cushman, Codex3 (Chen et al., 2021): 12B and 175B parameter code-pretrained LMs. GPT-Neo-2.7B, LLaMA-7B, LLaMA-13B, and Cushman have context window lengths of 2048, GPT-3.5-Turbo of 4096, Codex of 8001, and StarCoder of 8192."},{"figure_ref":[],"heading":"Methods","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Training-Free Methods","publication_ref":["b11","b12","b29","b9","b22"],"table_ref":[],"text":"We compare the following training-free metrics: Cosine similarity (COSINE) We use the Sentence-Bert library (Reimers and Gurevych, 2019) with the all-mpnet-base-v2 model. For independent selection, we use FAISS4 (Johnson et al., 2019) retrieve the most similar examples. BM25 (BM25) We use the Okapi variant (Robertson et al., 1993;Jones et al., 2000) of BM25 from the rank_bm255 library with three syntactic structures as terms: unigrams, size-4 or smaller n-grams, and size-4 or smaller subtrees of the input dependency parse (obtained using the spaCy6 ). BERTScore We use the bert_score7 library (Zhang et al., 2020) with deberta-large-mnli and deberta-base-mnli models which are De-BERTa models (He et al., 2021) finetuned on the MNLI dataset (Williams et al., 2018). We will refer to the recall, precision, and F1 variants as BSR, BSP, and BSF1, respectively. Unless specified otherwise, we do not apply importance weighting (IDF) and use deberta-large-mnli.\nAdditionally, we experiment with (1) a random baseline (RANDOM) that randomly selects demonstrations from the pool, and (2) with the set-extensions of COSINE, BM25 and BSR as described in §5 which will be referred to as SET-COSINE, SET-BM25, and SET-BSR respectively."},{"figure_ref":[],"heading":"Trained Methods","publication_ref":["b25","b15","b25","b16"],"table_ref":[],"text":"We also compare with methods that require task or LLM-specific training. EPR (Rubin et al., 2022) uses LLM perplexity to train a dense retriever for each dataset. CEIL (Ye et al., 2023) uses EPR and an LLM to train a Determinantal Point Process (Kulesza, 2012) for each dataset and then uses it to select examples. We use Ye et al. (2023)'s implementation of EPR and CEIL and use GPT-Neo-2.7B LLM. We also compare with LFCOV (Levy et al., 2022), a method for semantic parsing, specifically SMCalFlow-CS and GeoQuery. It trains a classifier to predict logical form substructures and then selects diverse examples containing them. We use the shots provided by the authors."},{"figure_ref":[],"heading":"Prompt Construction","publication_ref":[],"table_ref":["tab_11"],"text":"For k-shot (we use k = 8 unless specified otherwise) ICL with any given dataset ( § 6.1), demonstration selection method ( § 6.3) and LLM ( § 6.2), we construct the prompt as follows: (1) select up to k demonstrations depending on the context window of the LLM; (2) order the demonstrations in increasing order of relevance so that the most relevant demonstrations appear closest to the test input; and (3) linearize the ordered demonstrations and the test input using the dataset's prompt template in Table 5 dered by their corresponding instance-level score. For the trained baselines, we use orderings recommended by the corresponding authors."},{"figure_ref":[],"heading":"Results","publication_ref":[],"table_ref":["tab_5","tab_13","tab_14"],"text":"We begin by comparing the performance of our proposed methods, BSR and SET-BSR, with prior training-free and state-of-the-art trained methods in § 7.1. We then analyze the different metrics for measuring informativeness of individual demonstrations ( § 7.2) and the impact of coverage-based set selection using our set extension ( § 7.3). Table 3, we see that, unlike SET-BSR, BSR is effective even for non-semantic parsing datasets outperforming COSINE by 15 points on average with GPT-Neo-2.7B (see Table 12), and often even EPR and CEIL (see Table 13). All the above improvements were statistically significant (p < 0.05) under paired permutation-tests."},{"figure_ref":[],"heading":"Main Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_1"],"heading":"SET-BSR is more effective with larger LLMs","publication_ref":[],"table_ref":["tab_4"],"text":"The effectiveness of SET-BSR monotonically improves as LLMs become more powerful. The trend is particularly pronounced in compositional splits, where it gets 25% absolute improvement v/s CO-SINE on average (see Fig. 2) and 49% improvement on the 32-C split of SMCalFlow-CS (see Table 2)."},{"figure_ref":["fig_1"],"heading":"Trained methods do not leverage larger LLMs","publication_ref":["b20"],"table_ref":["tab_5","tab_14"],"text":"As EPR and CEIL are trained using GPT-Neo-2.7B, they have difficulty generalizing to and taking ad- vantage of larger, more powerful LLMs, becoming less effective on IID splits (Fig. 2), and failing on GSM8K (Table 3). The latter is likely because GPT-Neo-2.7B itself fails on GSM8K (Table 13), which requires Chain-of-Thought reasoning, an emergent ability of larger LLMs (Wei et al., 2022). As training with increasingly large LLMs is prohibitively expensive and impractical, these results demonstrate serious limitations of trained methods."},{"figure_ref":["fig_1","fig_3"],"heading":"Measure of Informativeness","publication_ref":[],"table_ref":["tab_2","tab_5","tab_6","tab_6"],"text":"Contextual embeddings capture salient aspects From Tables 1 and3, it is clear that BSR consistently outperforms COSINE and BM25. This is true even when using the same encoder (see App. D), is seen in both IID and compositional splits (see Fig. 2), and with varying number of demonstrations (see Fig. 4). Larger syntactic substructures did not improve BM25 as seen in Table 4 (Bottom). These results show that contextual embeddings are indeed better at capturing salient aspects.\nRecall outperforms other measures Comparing the variants of BERTScore, for Codex in Table 4 (Top), and other LLMs in Fig. 7 in App. D, it is evident that recall is on par with, or better than, the F1 metric. This supports our hypothesis that recall or coverage (of salient aspects) is a useful metric for informativeness. We include additional ablations in App. D, analyzing the effect of using importance weighting (IDF) and using a larger LM to compute token embeddings for BSR."},{"figure_ref":["fig_2"],"heading":"Coverage-based Set Selection","publication_ref":[],"table_ref":["tab_2"],"text":"Impact on performance From Fig. 3, we see that coverage-based set selection is most effective in compositional splits where it improves the average performance of all metrics, including COSINE. This shows the importance of selecting demonstrations as a set in compositional settings where examples demonstrating all the salient aspects of the test input are even less likely to exist. The set extension is less effective in IID splits and even hurts performance for COSINE and vanilla unigram BM25.\nOverall, BSR and BM25 with larger substructures benefit the most from the set extension. We provided further analyses of improvements from set selection and the impact of reordering in App. D. the relevance of individual demonstrations to prioritize coverage of all aspects with the set (see Table 9 for an example from GSM8K). Additionally, even contextual token embeddings can only capture salient aspects that are explicitly expressed in the input text and thus may not be suitable for tasks where the salient aspects are more abstract and require reasoning themselves (see Table 10 for an example from QNLI). We leave it to future work to explore better measures of informativeness, including better characterizations of salient aspects."},{"figure_ref":[],"heading":"Illustrative Example","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"This paper presents a novel framework for selecting informative sets of demonstrations that cover salient aspects of the test input to aid the language model (LLM) in solving it. We explore different ways to characterize these aspects and quantify their coverage. Evaluation on a wide range of tasks and LLMs validates the effectiveness of BERTScore-Recall as a measure of informativeness of individual demonstrations. Further, our results demonstrate the superiority of SET-BSR in selecting informative sets of demonstrations compositional tasks like semantic parsing and highlight the ability of coverage-based demonstration selection, unlike trained methods, to leverage increasingly powerful larger LLMs. Our code base is available at https://github.com/ Shivanshu-Gupta/icl-coverage. A Submodularity Definition A.1 (Submodular Function). If Ω is a finite set, a submodular function is a set function f : 2 Ω → R, where 2 Ω denotes the power set of Ω, which satisfies one of the following equivalent conditions.\n1. For every X, Y ⊆ Ω with X ⊆ Y and every\nx ∈ Ω\\Y we have that f\n(X ∪{x})-f (X) ≥ f (Y ∪ {x}) -f (Y ).\n2. For every S, T ⊆ Ω we have that\nf (S) + f (T ) ≥ f (S ∪ T ) + f (S ∩ T ).\n3. For every X ⊆ Ω and\nx 1 , x 2 ∈ Ω\\X such that x 1 ̸ = x 2 we have that f (X ∪ {x 1 }) + f (X ∪ {x 2 }) ≥ f (X ∪ {x 1 , x 2 }) + f (X). Theorem A.1. The function f maxw (X) = max x∈X w x\nis submodular for any assignment of weights w x to the elements x ∈ Ω.\nProof. The following are clearly true for any x ∈ Ω and any\nx 1 , x 2 ∈ Ω such that w x 1 > w x 2 : 1. f maxw (X ∪ {x}) ≥ f (X) 2. f maxw (X ∪ {x 1 }) = f maxw (X ∪ {x 1 , x 2 })\nAdding these two inequalities together, we get the third definition of submodularity and thus f maxw is submodular.\nTheorem A.2. If {f i } n\ni=1 are all submodular functions, then n i=1 f i is also submodular.\nProof. We show this for n = 2:\n(f 1 + f 2 )(X 1 ∪ X 2 ) + (f 1 + f 2 )(X 1 ∩ X 2 ) = (f 1 (X 1 ∪ X 2 ) + f 1 (X 1 ∩ X 2 )) + (f 2 (X 1 ∪ X 2 ) + f 2 (X 1 ∩ X 2 )) ≤ (f 1 (X 1 ) + f 1 (X 2 )) + (f 2 (X 1 ) + f 2 (X 2 )) = (f 1 + f 2 )(X 1 ) + (f 1 + f 2 )(X 2 ) (8)\nTherefore, f 1 + f 2 is submodular using the second definition of submodularity. By induction, this is true for any number n of functions.\nTheorem A.3. The set-level coverage metric setcov (x test , Z) as defined in Eq. 6 is submodular for any definition of c(s, z).\nProof. From Theorem A.1, the function f s (Z) defined as f s (Z) = max z∈Z c(s, z) is submodular for any definition of c(s, z). Further, since from Theorem A.2, the sum of submodular functions is also submodular, setcov (x test , Z) =\ns∈Sx test f s (Z) is submodular."},{"figure_ref":[],"heading":"B Datasets","publication_ref":[],"table_ref":[],"text":"We use 15 diverse datasets, including 6 semantic parsing, 2 numerical reasoning, and 7 classification datasets."},{"figure_ref":[],"heading":"B.1 Semantic Parsing","publication_ref":["b28","b5","b13","b16","b10","b0","b5","b7","b19","b7","b27","b23","b17"],"table_ref":["tab_11"],"text":"We use 6 semantic parsing datasets with IID and compositional splits for our experiments. Table 5 shows sample instances from each dataset we experiment with along with the textual template we use to linearize the instances. The ICL prompt is constructed by concatenating the templatized demonstrations and the test instance using \\n\\n as the separator.\nGeoQuery (Zelle and Mooney, 1996): A dataset containing 880 natural language questions about US geography paired with Prolog programs. In addition to the standard (IID) split, we experiment with three types of compositional splits: (1) Template split where the training and test sets have disjoint program templates (Finegan-Dollak et al., 2018); (2) TMCD split which creates train and test sets with maximal compound divergence and minimal atom divergence (Keysers et al., 2020); and\n(3) Length split which evaluates for length generalization by testing on sequences longer than ones in training. Following Levy et al. (2022), we use the compositional splits -three Template, three TMCD, and one Length -generated by Qiu et al.\n(2022a) and average results across the TMCD and Template splits. ATIS (Hemphill et al., 1990;Dahl et al., 1994):\nA dataset of natural language queries about aviation paired with λ-calculus programs. We experiment with an IID split and a Template split (Finegan-Dollak et al., 2018) for evaluating compositional generalization, both taken from (Gupta et al., 2022).\nOvernight (Wang et al., 2015): A dataset containing both synthetic and natural language utterances from 11 domains (e.g. socialnetwork, restaurants, etc.) paired with Lambda-DCS logical forms. We experiment with an IID and a Template split of the socialnetwork domain taken from (Gupta et al., 2022). SMCalFlow (Andreas et al., 2020): A dataset of task-oriented natural language dialogs about calendars, weather, places, and people paired with executable dataflow programs. SMCalFlow-CS (Yin et al., 2021) BREAK (Wolfson et al., 2020) is a dataset that maps complex natural language questions into a language-based meaning representation (QDMR) comprising an ordered list of atomic steps necessary to answer the question. Following (Rubin et al., 2022), we use the low-level Break subset where the targets are logical forms comprising lists of operators with their arguments based on the corresponding QDMR. MTOP (Li et al., 2021): A multilingual taskoriented semantic parsing dataset spanning six languages and 11 domains. The target commands are complex queries featuring nested intent-slot prediction. We use the English subset of MTOP from (Rubin et al., 2022)."},{"figure_ref":[],"heading":"B.2 Non-Semantic Parsing","publication_ref":["b21","b4","b22","b2","b30"],"table_ref":[],"text":"We additionally experiment with the standard IID splits of 9 non-semantic parsing datasets from the following categories: Numerical Reasoning: For this category, we experiment with GSM8K (Cobbe et al., 2021), a chain-of-thought reasoning (Wei et al., 2023) dataset of grade school-level arithmetic reasoning problems expressed in natural language and DROP (Dua et al., 2019), a dataset of question-answer pairs where the questions are about paragraphs containing numerical information and the answers are spans in the paragraph.\nClassification: For this category, we experiment with three Natural Language Inference (NLI) datasests (QNLI (Wang et al., 2018), MNLI (Williams et al., 2018), andRTE (Bentivogli et al., 2009)), three Paraphrase Detection datasets (MRPC (Dolan and Brockett, 2005), PAWS (Zhang et al., 2019), and QQP (Wang et al., 2018)) and one Sentiment Classification dataset (SST2 (Socher et al., 2013))."},{"figure_ref":[],"heading":"C Selection Time","publication_ref":[],"table_ref":[],"text":"Despite their O(LT d) time complexity, we found example selection using both BSR and SET-BSR to be fast enough to not be a bottleneck to incontext learning for most datasets considered in this work. By using a GPU to compute c(x, z)s, we could get both to work in the order tens of milliseconds per test input on average which was significantly faster than the LLM inference time itself. The exceptions were DROP, PAWS, QQP, MNLI and QNLI for which the selection took >1 second due to much longer instances and/or larger instance pool. We leave it to future work to explore more efficient ways to measure informativeness."},{"figure_ref":["fig_4","fig_6"],"heading":"D Additional Analyses","publication_ref":[],"table_ref":["tab_11","tab_12"],"text":"BM25 From Fig. 6 we can see that coverage-based selection using BM25 with larger substructures outperforms vanilla unigram BM25 in compositional splits.\nBERTScore-Recall Examining the impact of importance weighting in Fig. 8 which compares the performance change with using importance weighting (IDF) in BSR, we can see that its effect is not consistent across different LLMs. We also did not see any consistent improvement from using larger deberta-large-mnli for computing token embeddings for instance-level BSR (see Fig. 9). However, it did help with set-level selection using SET-BSR.\nReordering We found the reordering of demonstrations according to the corresponding instance-level metric to only be necessary for smaller LLMs (see Fig. 10), with it even hurting the performance of larger LLMs. We believe this is because larger and code-pretrained LLMs are more capable at composing the salient aspects in the different demonstrations and taking advantage of the full context. deberta-large-mnli outperforms Cosine with all-mpnet-base-v2. Tables 15,16, 17, and 18 show that the same trend holds even when using the same encoder, bert-base-uncased, for both metrics confirming that contextual embeddings are indeed better at capturing salient aspects."},{"figure_ref":[],"heading":"BSR outperforms","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_7"],"heading":"Recall of Syntactic Structures","publication_ref":[],"table_ref":[],"text":"The improvements from set-based selection may be explained by Fig. 11 where we see that set-extensions COSINE and unigram BM25 reduce the recall of substructures of the test input whereas the recalls increase with set-extensions of both BM25[4-GRAM] and BM25[4-DEPST], and even BSR, which does not explicity consider these substructures."},{"figure_ref":[],"heading":"E Qualitative Analysis of Prompts","publication_ref":[],"table_ref":[],"text":"Tables 7, 8 show demonstrations selected using COSINE and SET-BSR for instances from MTOP and SMCalFlow-CS respectively. In each case, CO-SINE find demonstrations that are all very similar to the test input but fails to demonstrate some salient aspect, whereas BSR selects less similar instances but ensures complete coverage of all salient aspects. Tables 9 and 10 additionally illustrate limitations of set-selection and of token-embeddings in capturing salient aspects."},{"figure_ref":[],"heading":"F All Results","publication_ref":[],"table_ref":["tab_11","tab_12"],"text":"Tables 11 contains 8-shot ICL results for our proposed methods and prior learning-free and learningbased demonstration selection on all the LLMs for all the semantic parsing datasets. For numerical reasoning and classification datasets, Tables 12 and 13 compare 8-shot ICL performance with prior training-free and trained methods, respectively. Table 14 provides average performances across all datasets.\nAdditionally, Tables 15,16, 17, 18, 20, and 21 contain results on semantic parsing datasets of all ablations of learning-free selection methods we ran, with GPT-Neo-2.7B, LLaMA-7B, LLaMA-13B, StarCoder, Cushman, and Codex, respectively. We did not run ablations on GPT-3.5-Turbo due to its cost. Figure 9: Impact on average 8-shot ICL performance on semantic parsing splits from using a larger deberta-large-mnli LLM for computing contextual token embeddings v/s using deberta-base-mnli in BSR and SET-BSR. "},{"figure_ref":[],"heading":"Acknowledgements","publication_ref":[],"table_ref":[],"text":"We would like to thank the anonymous reviewers for their feedback. This work was sponsored in part by the DARPA MCS program under Contract No. N660011924033 with the United States Office Of Naval Research and in part by the NSF award #IIS-2046873. The views expressed are those of the authors and do not reflect the policy of the funding agencies."},{"figure_ref":[],"heading":"Limitations","publication_ref":[],"table_ref":[],"text":"Contextual token embeddings require the salient aspects to be expressed in text and hence may not be able to capture them for all tasks. Moreover, since it requires computing a dot product for every pair of test and candidate instance tokens, this causes it to scale quadratically with the average number of tokens making it computationally infeasible for tasks with very long textual linearizations. Future work can thus explore more general characterizations of salient aspects and more efficient methods for selecting demonstrations covering them."},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"//github.com/Shivanshu-"},{"figure_ref":[],"heading":"","publication_ref":[],"table_ref":[],"text":"]. COSINE's reliance on a single dense embedding means it is unable to account for the fact that \"options\" could mean dishes and not just recipes.\nSelector Prompt"},{"figure_ref":[],"heading":"COSINE","publication_ref":[],"table_ref":[],"text":"Sentence : I need a meeting with Elli tomorrow at 11 pm Logical Form : CreateEvent ( AND ( with_attendee (\" Elli \") , starts_at ( Tomorrow () ) , starts_at ( NumberPM (11) ) ) )\nSentence : Set a meeting with Elli for tomorrow at 2 pm through the end of the day and call it Recap Logical Form : CreateEvent ( AND ( ends_at ( AND ( GE ( DateTime ?( date = Tomorrow () , time = NumberPM (2) ) ) , EndOfWorkDay () ) ) , with_attendee (\" Elli \") , has_subject (\" Recap \") , starts_at ( Tomorrow () ) , starts_at ( NumberPM (2) ) ) )\nSentence : Schedule a meeting with Elli for tomorrow at 4 pm through the end of the workday Logical Form : CreateEvent ( AND ( ends_at ( AND ( GE ( DateTime ?( date = Tomorrow () , time = NumberPM (4) ) ) , EndOfWorkDay () ) ) , with_attendee (\" Elli \") , starts_at ( Tomorrow () ) , starts_at ( NumberPM (4) ) ) )\nSentence : Schedule a meeting with Elli from 4 PM until the end of the day tomorrow . Logical Form : CreateEvent ( AND ( ends_at ( AND ( GE ( DateTime ?( date = Tomorrow () , time = NumberPM (4) ) ) , EndOfWorkDay () ) ) , with_attendee (\" Elli \") , starts_at ( Tomorrow () ) , starts_at ( NumberPM (4) ) ) )"},{"figure_ref":[],"heading":"SET-BSR","publication_ref":[],"table_ref":[],"text":"Sentence : I need a doctor 's appointment on Wednesday morning . Logical Form : CreateEvent ( AND ( has_subject (\" doctor ' s appointment \") , starts_at ( Morning () ) , starts_at ( NextDOW (\" WEDNESDAY \") ) ) )\nSentence : I need to see Alice and her boss next Monday at 3 pm . Logical Form : CreateEvent ( AND ( with_attendee (\" Alice \") , with_attendee ( FindManager (\" Alice \") ) , starts_at ( NextDOW (\" MONDAY \") ) , starts_at ( NumberPM (3) ) ) )\nSentence : Schedule a meeting with Jake , Elli , and Jesse for Friday at 2 pm . Logical Form : CreateEvent ( AND ( with_attendee (\" Jesse \") , with_attendee (\" Jake \") , with_attendee (\" Elli \") , starts_at ( NextDOW (\" FRIDAY \") ) , starts_at ( NumberPM (2) ) ) )\nSentence : I need to schedule a meeting with Jeff 's supervisor Lynne for tomorrow at 10 AM . Logical Form : CreateEvent ( AND ( with_attendee (\" Lynne \") , starts_at ( Tomorrow () ) , starts_at ( NumberAM (10) ) ) ) Since SET-BSR prioritizes coverage of the remaining aspect, it selects an example that has exactly three items whose total length has to be computed but overall is not very similar in reasoning. BSR on the other hand tries to find an example that demonstrates all aspects by itself and happens to find one that partially demonstrates the remaining aspect as well. "}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b1\",\n \"b21\",\n \"b3\",\n \"b21\",\n \"b31\",\n \"b16\",\n \"b25\",\n \"b16\",\n \"b29\",\n \"b16\",\n \"b25\"\n ],\n \"table_ref\": [],\n \"text\": \"Large language models (LLMs) (Devlin et al., 2019;Brown et al., 2020) are capable of generalizing to novel tasks (Brown et al., 2020) by conditioning on textual prompts consisting of a few task examples. This training-free paradigm of fewshot inference, known as in-context learning (ICL), reduces the cost of modeling new tasks while also providing an interpretable and customizable interface (Liu et al., 2022;Wei et al., 2023) and improving generalization (Anil et al., 2022;Qiu et al., 2022b;Drozdov et al., 2023) and reasoning skills (Wei et al., 2023). However, ICL performance is critically sensitive to the choice of demonstrations (Zhao et al., 2021;Liu et al., 2022;Lu et al., 2022;Rubin et al., 2022; Schick and Schütze, 2021), as the LLM relies on them for understanding and solving the test instance.\\nThe standard approach to selecting ICL examples or demonstrations from a pool of candidates is to independently score them using a relevance metric and choose the top-ranked ones. However, cosine similarity and BM25, the two commonly used metrics, are sub-optimal for selecting demonstrations due to their reliance on a single dense embedding and unigram overlap, respectively. Moreover, since it selects examples independently, this approach ignores their utility as a set. It is particularly inadequate for complex compositional tasks like semantic parsing (Levy et al., 2022) where no single candidate might contain all reasoning patterns, and an independent selection would select multiple redundant examples with the same reasoning patterns but fail to demonstrate the others. Figure 1 shows a failure case where similarity-based selection picks paraphrased examples that fail to demonstrate how to find a manager. Prior work on selecting demonstrations as a set (Ye et al., 2023;Levy et al., 2022) required task and/or LLM-specific training, limiting their utility. For this reason, simple yet widely applicable training-free methods like BM25 and cosine similarity remain the most popular approaches for ICL example selection.\\nIn this work, we propose a novel framework for selecting sets of maximally informative demonstrations for the salient aspects of the test input, e.g., reasoning patterns, entities, etc. Examples selected using this framework are informative about the test input and help the LLM understand and perform the task. We use this framework to explore different ways to characterize salient aspects, including syntactic structures like dependency parse subtrees and contextual token embeddings, while using BM25 and BERTScore (Zhang et al., 2020) to measure their coverage, respectively. To select the demonstrations as a set, we extend the coverage metrics to measure the overall informativeness of a set of demonstrations. We show that these set-level metrics are submodular and can be efficiently optimized to find demonstration sets that maximally cover the salient aspects. We evaluate our ICL example selection methods on 15 diverse datasets, including 6 semantic parsing, 2 numerical reasoning, and 7 classification datasets, and with 7 LLMs of varying sizes and pretraining. Among instance-level metrics, BSR, the recall version of BERTScore, consistently outperforms standard retrieval metrics on all datasets and LLMs, beating cosine similarity by up to 8 points on average in semantic parsing datasets and 15 points in the rest. Selecting demonstrations as a set using SET-BSR, the set-extension of BSR, leads to further gains in semantic parsing and is particularly effective in compositional settings where the gains grow with LLM size. With Codex, a 175B parameter LLM, SET-BSR outperforms cosine similarity by 17% on average with up to 49% improvement in some splits, and, despite being training-free, outperforms even trained methods like those from Rubin et al. (2022), Levy et al. (2022), and Ye et al. (2023) that require task and/or LLM-specific training.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b21\",\n \"b3\",\n \"b21\",\n \"b31\",\n \"b16\",\n \"b26\",\n \"b24\",\n \"b18\",\n \"b16\",\n \"b25\",\n \"b16\",\n \"b25\",\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"In-context learning for few-shot inference facilitates the use of LLMs for novel tasks without the need for expensive supervised fine-tuning. In addition to reduced cost, it has several other advantages over supervised fine-tuning: it provides a more interpretable and customizable interface to using LLMs (Liu et al., 2022;Wei et al., 2023); and retention of linguistic understanding and knowledge from pretraining leading to improved generalization (Anil et al., 2022;Qiu et al., 2022b;Drozdov et al., 2023) and reasoning skills (Wei et al., 2023).\\nHowever, the performance of ICL is critically sensitive to the choice of demonstrations (Zhao et al., 2021;Liu et al., 2022). This has led to a growing interest in techniques for selecting good demonstrations. Prior work can be roughly classified into (1) independently scoring and retrieving examples (Liu et al., 2022;Rubin et al., 2022), (2) selecting diverse examples to reduce redundancy among them (Su et al., 2022;Levy et al., 2022;Agrawal et al., 2022;Ye et al., 2022), and (3) selecting examples that minimize the entropy of the LLM's output distribution for the test input (Lu et al., 2022;Wu et al., 2023). Recent work has also trained RL agents (Lu et al., 2023) and used Bayesian inference (Wang et al., 2023).\\nThe most similar studies to ours are Levy et al. (2022) and Ye et al. (2023). Levy et al. (2022) select diverse demonstrations that cover substructures of the target output predicted by task-specific classifiers but are limited in applicability to a few semantic parsing tasks. Ye et al. (2023) use Determinantal Point Processes (Kulesza, 2012) to select a diverse set of demonstrations similar to the test instance but do not optimize for coverage directly and require training with the LLM. Moreover, both methods require task or LLM-specific training that limits their use and effectiveness for larger LMs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Preliminaries\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In-context learning is the ability of LLMs to solve novel tasks by merely conditioning on a few task demonstrations. Formally, given demonstrations {(x i , y i )} k i=1 and the test input x test , it involves using textual templates to linearize instance inputs and outputs into sequences of tokens from the LLM vocabulary, x = I(x) = ⟨x 1 . . . x |x| ⟩ and y = O(y) = ⟨y 1 . . . y |y| ⟩. The linearizations are then concatenated to form a prompt and fed to the LLM for conditional generation of the test output:\\ny test ∼ P LM (• | x 1 , y 1 , . . . , x K , y K , x test ) (1)\\nThe interpretable and training-free nature of ICL makes it an attractive alternative to supervised finetuning. However, its performance is highly sensitive to the choice and order of demonstrations.\\nDemonstration Selection identifies which examples to include in the prompt for any test instance. Formally, given a test input x test and a pool of candidates\\nT = {z i } N i=1 = {(x i , y i )} N i=1\\n, the goal is to select a subset of k ≪ N demonstrations that when included in the context make y test the most likely generation. A naive approach is to randomly sample k instances from T , but this is sub-optimal since the demonstrations are often completely unrelated to the test input. Instead, the standard approach to selecting demonstrations that are informative about the test input is to independently assign each candidate z a score score(x test , z) using a relevance metric and then select the top k candidates.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Relevance Metrics\",\n \"publication_ref\": [\n \"b12\"\n ],\n \"table_ref\": [],\n \"text\": \"The two most commonly used relevance metrics for scoring demonstration are cosine similarity and BM25. Cosine similarity uses a representation function R to independently map the textual linearizations of inputs to unit-norm embeddings r x = R(x) in a common vector space and then scores the candidate z using the dot product, cosine(x test , z) = r T xtest r z . BM25, on the other hand, is a sparse information retrieval algorithm belonging to a class of TF-IDF measures that view the test input and the candidates as bags of terms and measures relevance as a weighted recall or coverage of these terms:\\ntfidf(x test , z) = s∈Tx test idf(s)tf(s, T z ) (2)\\nHere T x and T z are the set of terms in x and z respectively, and tf(s, T z ) and idf(s) are the term frequency and inverse document frequency statistics that measure the coverage of a particular term and the relative importance of terms respectively. We use tf and idf as per the Okapi variant of BM25 (Robertson et al., 1993;Jones et al., 2000).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Informative Demonstrations\",\n \"publication_ref\": [\n \"b29\"\n ],\n \"table_ref\": [],\n \"text\": \"The limitation of the standard demonstration selection approach is that by independently scoring the demonstrations, it ignores their utility as a set. For ICL to work, the demonstrations included in the context need to be informative about how to understand and solve the test input. In this section and the next, we describe our approach to selecting informative sets of demonstrations for ICL. We begin by defining our notion of informativeness of demonstrations in ICL and describing how to measure it. Thereafter, in §5, we will discuss how to extend this notion to an algorithm for selecting optimally informative sets of demonstrations.\\nInformativeness Demonstrations should demonstrate the salient aspects, e.g., reasoning patterns, entities, etc., of the test input. Formally, denoting S xtest as the set of salient aspects of the test input, we measure the informativeness of a demonstration z in terms of the coverage of such salient aspects,\\ncover (x test , z) = s∈Sx test c(s, z)(3)\\nwhere c(s, z) measures the coverage (or recall) of a single salient aspect s by z.\\nSalient Aspects Both cosine similarity and BM25 are special cases of Eq. 3 for different notions of salient aspects. For BM25, S xtest = T xtest , the set of unigrams in x, and c(s, z) = idf(s)tf(s, T z ).\\nAnd cosine similarity, although not explicitly a recall metric, can also be interpreted as evaluating coverage of the dimensions of the test input embedding by defining S xtest = [1, d], the dimensions of the dense embedding as the salient aspects, i.e., cosine(x test , z)\\n= d s=1 r xtest [s] • r z [s] (4)\\nThe above interpretations reveal why neither cosine similarity nor BM25 are good measures of informativeness. While cosine similarity captures some aspects of semantic similarity (depending on the embedding), it is limited to a single embedding. And, unigrams, the commonly used terms with BM25, are too small to capture most salient aspects. A good measure of informativeness necessitates an accurate characterization of salient aspects. One way might be to use larger syntactic substructures of the input as terms with BM25. We experiment with using larger n-grams and subtrees of the dependency parse tree. However, such syntactic structures are constrained to the surface form of the instance and hence may not capture meaning and aspects like reasoning patterns. A better way to capture salient aspects is to use contextualized token embeddings, the idea behind the BERTScore (Zhang et al., 2020) metric.\\nBERTScore was originally proposed as a metric for evaluating the quality of machine-generated text (e.g., machine translation) by comparing it to a reference text. It leverages pre-trained contextual embeddings to match words in the candidate and reference sentences by cosine similarity and compute precision, recall, and F1 measures. Formally, given the sequences of contextual embeddings ⟨x 1 , x 2 , . . . , x |x| ⟩ and ⟨z 1 , z 2 , . . . , z |z| ⟩ of tokens in x = ⟨x 1 , x 2 , . . . , x |x| ⟩ and z = ⟨z 1 , z 2 , . . . , z |z| ⟩ respectively, the recall measure, BERTScore-Recall (BSR), is defined as:\\nBSR(x, z) = x i ∈x w(x i ) max j x T i z j (5)\\nHere, w(x i ) is a weight assigned to token x i and can be defined as 1 |x| if treating each token as equally important or idf(x i )\\nx i ∈x idf(x i ) if downweighting rare words. The precision measure is defined analogously, while the F1 measure is the harmonic mean of the two. BSR is also a special case of Eq. 3 with contextualized tokens as salient aspects, i.e., S x = ⟨x 1 , x 2 , . . . , x |x| ⟩ and can be used to select examples by treating them as candidates and the test input as the reference. The following table summarizes the informativeness measures and salient aspects in this work.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"Metric Salient Aspects\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Cosine embedding dimensions BM25 unigrams, n-grams, dependency parse subtrees BERTScore contextual token embeddings 5 Set-level Information Coverage So far, we have focused on measuring the informativeness of a single demonstration to rank and independently select the most informative ones. However, as depicted in Fig. 1, when no single single candidate demonstrates all salient aspects, this approach can fail to cover all of them while also selecting redundant demonstrations that provide no new information. A scenario where this can happen is when the candidate pool contains close paraphrases (or duplicates). This suggests that demonstrations should be selected as a set.\\nSet Metric To evaluate the informativeness of a set of examples Z, we propose to extend the coverage measure in Eq. 3 to a measure for sets as follows: Intuitively, this measures the coverage of each salient aspect as the best coverage it receives from any example in the set. In other words, maximizing it requires that every salient aspect appears at least once in some demonstration without considering which or how many. Since cosine similarity, BM25, and BSR are all special cases of Eq. 3, they can be extended to set measures using Eq. 6.\\nsetcov (x test , Z) = s∈Sx test max z∈Z c(s, z) (6)\\nSubmodularity Given the combinatorial space of sets of demonstrations, for a measure on sets to be practical, it needs to be efficiently optimizable. Fortunately, the set-level metric, as defined above, is also submodular for any definition of c(s, z). We prove this in Appendix A. Intuitively, this follows from the facts that (1) for any given test instance, c(s, z) assigns a scalar weight to each demonstration z ∈ Z, (2) the maximum of weights across set elements is submodular, and (3) the sum of submodular functions is also submodular. This means that the set-level metric can be optimized using a greedy algorithm with a constant factor approximation guarantee (Nemhauser et al., 1978).\\nAlgorithm The greedy algorithm we use to select the optimal set is shown in Algorithm 1. In every iteration, it selects the example that maximally increases the coverage of the current set of demonstrations (lines 5-9). If no such example exists, it resets (lines 11). Using the following identity when computing the score for candidate sets (line 5), ( 7)\\nsetcov x test , Z ∪ z ′ = s∈Sx test max c(s, Z), c(s, z ′ )\\nand assuming constant time for computing each c(s, z), the time complexity of algorithm is\\nO(kN L), where L = |S xtest |. For BSR, the com- plexity of computing c(x, z) for all z ∈ Z is O(T d),\\nwhere T is the total number of tokens in Z and d is the token embedding size. Thus, the time complexity of both instance and set-level BSR is dominated by the computation of c(x, z), and is O(LT d). While slower than cosine and BM25, we found it to be a small overhead to in-context learning for most datasets considered in this work. We discuss this further in App. C.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experimental Setup\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Datasets\",\n \"publication_ref\": [\n \"b28\",\n \"b10\",\n \"b0\",\n \"b19\",\n \"b23\",\n \"b17\",\n \"b4\",\n \"b22\",\n \"b2\",\n \"b30\",\n \"b5\",\n \"b16\",\n \"b5\",\n \"b7\",\n \"b27\",\n \"b25\",\n \"b8\"\n ],\n \"table_ref\": [],\n \"text\": \"We experiment with a total of 15 datasets including six diverse semantic parsing datasets viz. Geo-Query (Zelle and Mooney, 1996), ATIS (Hemphill et al., 1990;Dahl et al., 1994), Overnight (Wang et al., 2015), SMCalFlow (Andreas et al., 2020), BREAK (Wolfson et al., 2020), and MTOP (Li et al., 2021); a math-word problems (GSM8K (Cobbe et al., 2021)) and a machine reading comprehension (DROP (Dua et al., 2019)) dataset requiring multi-step numeric reasoning; and seven classification datasets spanning natural language inference, paraphrase detection and sentiment classification viz. QNLI (Wang et al., 2018), MNLI (Williams et al., 2018), RTE (Bentivogli et al., 2009), MRPC (Dolan and Brockett, 2005), PAWS (Zhang et al., 2019), QQP (Wang et al., 2018), andSST2 (Socher et al., 2013). We refer the reader to App. B for detailed descriptions of each dataset along with sample instances and prompt templates.\\nIn addition to the standard IID splits, we also evaluate compositional generalization using compositional splits wherever available. For GeoQuery we use three types of compositional splits: Template (Finegan-Dollak et al., 2018), TMCD (Keysers et al., 2020), and Length. Following Levy et al. (2022), we use the compositional splitsthree Template, three TMCD, and one Lengthgenerated by Qiu et al. (2022a) and average results across the TMCD and Template splits. For ATIS and Overnight, we experiment with Template splits (Finegan-Dollak et al., 2018) generated by Gupta et al. (2022). For SMCalFlow, we experiment with splits in SMCalFlow-CS (Yin et al., 2021): an IID split (8-S) and a compositional split (32-C).\\nFor all the splits, following prior work (Ye et al., 2023;Rubin et al., 2022) we randomly subsample 44,000 instances from the train set to use as pool to select demonstrations from. For evaluation, we use a random subsample of 1000 instance of the validation set if available, and the test set otherwise. We use Exact Match (EM) accuracy for all datasets except BREAK where we use LF-EM (Hasson and Berant, 2021), which is preferred over EM for semantic equivalence.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Models\",\n \"publication_ref\": [\n \"b6\",\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"We experiment with the following LLMs: GPT-Neo-2.7B (Black et al., 2021): A 2.7B-parameter LM trained on The Pile (Gao et al., 2020), an 825 GB text corpus. LLaMA (Touvron et al., 2023): A collection of LMs ranging from 7B to 65B parameters pretrained on CommonCrawl, GitHub, Arxiv, etc. We experiment with LLaMA-7B and LLaMA-13B. StarCoder (Li et al., 2023): A 15.5B parameter model trained on 80+ programming languages (Kocetkov et al., 2022). GPT-3.5-Turbo2 : 175B LM trained with RL to follow instructions and optimized for chat. Cushman, Codex3 (Chen et al., 2021): 12B and 175B parameter code-pretrained LMs. GPT-Neo-2.7B, LLaMA-7B, LLaMA-13B, and Cushman have context window lengths of 2048, GPT-3.5-Turbo of 4096, Codex of 8001, and StarCoder of 8192.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Methods\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Training-Free Methods\",\n \"publication_ref\": [\n \"b11\",\n \"b12\",\n \"b29\",\n \"b9\",\n \"b22\"\n ],\n \"table_ref\": [],\n \"text\": \"We compare the following training-free metrics: Cosine similarity (COSINE) We use the Sentence-Bert library (Reimers and Gurevych, 2019) with the all-mpnet-base-v2 model. For independent selection, we use FAISS4 (Johnson et al., 2019) retrieve the most similar examples. BM25 (BM25) We use the Okapi variant (Robertson et al., 1993;Jones et al., 2000) of BM25 from the rank_bm255 library with three syntactic structures as terms: unigrams, size-4 or smaller n-grams, and size-4 or smaller subtrees of the input dependency parse (obtained using the spaCy6 ). BERTScore We use the bert_score7 library (Zhang et al., 2020) with deberta-large-mnli and deberta-base-mnli models which are De-BERTa models (He et al., 2021) finetuned on the MNLI dataset (Williams et al., 2018). We will refer to the recall, precision, and F1 variants as BSR, BSP, and BSF1, respectively. Unless specified otherwise, we do not apply importance weighting (IDF) and use deberta-large-mnli.\\nAdditionally, we experiment with (1) a random baseline (RANDOM) that randomly selects demonstrations from the pool, and (2) with the set-extensions of COSINE, BM25 and BSR as described in §5 which will be referred to as SET-COSINE, SET-BM25, and SET-BSR respectively.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Trained Methods\",\n \"publication_ref\": [\n \"b25\",\n \"b15\",\n \"b25\",\n \"b16\"\n ],\n \"table_ref\": [],\n \"text\": \"We also compare with methods that require task or LLM-specific training. EPR (Rubin et al., 2022) uses LLM perplexity to train a dense retriever for each dataset. CEIL (Ye et al., 2023) uses EPR and an LLM to train a Determinantal Point Process (Kulesza, 2012) for each dataset and then uses it to select examples. We use Ye et al. (2023)'s implementation of EPR and CEIL and use GPT-Neo-2.7B LLM. We also compare with LFCOV (Levy et al., 2022), a method for semantic parsing, specifically SMCalFlow-CS and GeoQuery. It trains a classifier to predict logical form substructures and then selects diverse examples containing them. We use the shots provided by the authors.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Prompt Construction\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_11\"\n ],\n \"text\": \"For k-shot (we use k = 8 unless specified otherwise) ICL with any given dataset ( § 6.1), demonstration selection method ( § 6.3) and LLM ( § 6.2), we construct the prompt as follows: (1) select up to k demonstrations depending on the context window of the LLM; (2) order the demonstrations in increasing order of relevance so that the most relevant demonstrations appear closest to the test input; and (3) linearize the ordered demonstrations and the test input using the dataset's prompt template in Table 5 dered by their corresponding instance-level score. For the trained baselines, we use orderings recommended by the corresponding authors.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_5\",\n \"tab_13\",\n \"tab_14\"\n ],\n \"text\": \"We begin by comparing the performance of our proposed methods, BSR and SET-BSR, with prior training-free and state-of-the-art trained methods in § 7.1. We then analyze the different metrics for measuring informativeness of individual demonstrations ( § 7.2) and the impact of coverage-based set selection using our set extension ( § 7.3). Table 3, we see that, unlike SET-BSR, BSR is effective even for non-semantic parsing datasets outperforming COSINE by 15 points on average with GPT-Neo-2.7B (see Table 12), and often even EPR and CEIL (see Table 13). All the above improvements were statistically significant (p < 0.05) under paired permutation-tests.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Main Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"SET-BSR is more effective with larger LLMs\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\"\n ],\n \"text\": \"The effectiveness of SET-BSR monotonically improves as LLMs become more powerful. The trend is particularly pronounced in compositional splits, where it gets 25% absolute improvement v/s CO-SINE on average (see Fig. 2) and 49% improvement on the 32-C split of SMCalFlow-CS (see Table 2).\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Trained methods do not leverage larger LLMs\",\n \"publication_ref\": [\n \"b20\"\n ],\n \"table_ref\": [\n \"tab_5\",\n \"tab_14\"\n ],\n \"text\": \"As EPR and CEIL are trained using GPT-Neo-2.7B, they have difficulty generalizing to and taking ad- vantage of larger, more powerful LLMs, becoming less effective on IID splits (Fig. 2), and failing on GSM8K (Table 3). The latter is likely because GPT-Neo-2.7B itself fails on GSM8K (Table 13), which requires Chain-of-Thought reasoning, an emergent ability of larger LLMs (Wei et al., 2022). As training with increasingly large LLMs is prohibitively expensive and impractical, these results demonstrate serious limitations of trained methods.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_3\"\n ],\n \"heading\": \"Measure of Informativeness\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\",\n \"tab_5\",\n \"tab_6\",\n \"tab_6\"\n ],\n \"text\": \"Contextual embeddings capture salient aspects From Tables 1 and3, it is clear that BSR consistently outperforms COSINE and BM25. This is true even when using the same encoder (see App. D), is seen in both IID and compositional splits (see Fig. 2), and with varying number of demonstrations (see Fig. 4). Larger syntactic substructures did not improve BM25 as seen in Table 4 (Bottom). These results show that contextual embeddings are indeed better at capturing salient aspects.\\nRecall outperforms other measures Comparing the variants of BERTScore, for Codex in Table 4 (Top), and other LLMs in Fig. 7 in App. D, it is evident that recall is on par with, or better than, the F1 metric. This supports our hypothesis that recall or coverage (of salient aspects) is a useful metric for informativeness. We include additional ablations in App. D, analyzing the effect of using importance weighting (IDF) and using a larger LM to compute token embeddings for BSR.\"\n },\n {\n \"figure_ref\": [\n \"fig_2\"\n ],\n \"heading\": \"Coverage-based Set Selection\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"Impact on performance From Fig. 3, we see that coverage-based set selection is most effective in compositional splits where it improves the average performance of all metrics, including COSINE. This shows the importance of selecting demonstrations as a set in compositional settings where examples demonstrating all the salient aspects of the test input are even less likely to exist. The set extension is less effective in IID splits and even hurts performance for COSINE and vanilla unigram BM25.\\nOverall, BSR and BM25 with larger substructures benefit the most from the set extension. We provided further analyses of improvements from set selection and the impact of reordering in App. D. the relevance of individual demonstrations to prioritize coverage of all aspects with the set (see Table 9 for an example from GSM8K). Additionally, even contextual token embeddings can only capture salient aspects that are explicitly expressed in the input text and thus may not be suitable for tasks where the salient aspects are more abstract and require reasoning themselves (see Table 10 for an example from QNLI). We leave it to future work to explore better measures of informativeness, including better characterizations of salient aspects.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Illustrative Example\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This paper presents a novel framework for selecting informative sets of demonstrations that cover salient aspects of the test input to aid the language model (LLM) in solving it. We explore different ways to characterize these aspects and quantify their coverage. Evaluation on a wide range of tasks and LLMs validates the effectiveness of BERTScore-Recall as a measure of informativeness of individual demonstrations. Further, our results demonstrate the superiority of SET-BSR in selecting informative sets of demonstrations compositional tasks like semantic parsing and highlight the ability of coverage-based demonstration selection, unlike trained methods, to leverage increasingly powerful larger LLMs. Our code base is available at https://github.com/ Shivanshu-Gupta/icl-coverage. A Submodularity Definition A.1 (Submodular Function). If Ω is a finite set, a submodular function is a set function f : 2 Ω → R, where 2 Ω denotes the power set of Ω, which satisfies one of the following equivalent conditions.\\n1. For every X, Y ⊆ Ω with X ⊆ Y and every\\nx ∈ Ω\\\\Y we have that f\\n(X ∪{x})-f (X) ≥ f (Y ∪ {x}) -f (Y ).\\n2. For every S, T ⊆ Ω we have that\\nf (S) + f (T ) ≥ f (S ∪ T ) + f (S ∩ T ).\\n3. For every X ⊆ Ω and\\nx 1 , x 2 ∈ Ω\\\\X such that x 1 ̸ = x 2 we have that f (X ∪ {x 1 }) + f (X ∪ {x 2 }) ≥ f (X ∪ {x 1 , x 2 }) + f (X). Theorem A.1. The function f maxw (X) = max x∈X w x\\nis submodular for any assignment of weights w x to the elements x ∈ Ω.\\nProof. The following are clearly true for any x ∈ Ω and any\\nx 1 , x 2 ∈ Ω such that w x 1 > w x 2 : 1. f maxw (X ∪ {x}) ≥ f (X) 2. f maxw (X ∪ {x 1 }) = f maxw (X ∪ {x 1 , x 2 })\\nAdding these two inequalities together, we get the third definition of submodularity and thus f maxw is submodular.\\nTheorem A.2. If {f i } n\\ni=1 are all submodular functions, then n i=1 f i is also submodular.\\nProof. We show this for n = 2:\\n(f 1 + f 2 )(X 1 ∪ X 2 ) + (f 1 + f 2 )(X 1 ∩ X 2 ) = (f 1 (X 1 ∪ X 2 ) + f 1 (X 1 ∩ X 2 )) + (f 2 (X 1 ∪ X 2 ) + f 2 (X 1 ∩ X 2 )) ≤ (f 1 (X 1 ) + f 1 (X 2 )) + (f 2 (X 1 ) + f 2 (X 2 )) = (f 1 + f 2 )(X 1 ) + (f 1 + f 2 )(X 2 ) (8)\\nTherefore, f 1 + f 2 is submodular using the second definition of submodularity. By induction, this is true for any number n of functions.\\nTheorem A.3. The set-level coverage metric setcov (x test , Z) as defined in Eq. 6 is submodular for any definition of c(s, z).\\nProof. From Theorem A.1, the function f s (Z) defined as f s (Z) = max z∈Z c(s, z) is submodular for any definition of c(s, z). Further, since from Theorem A.2, the sum of submodular functions is also submodular, setcov (x test , Z) =\\ns∈Sx test f s (Z) is submodular.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B Datasets\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We use 15 diverse datasets, including 6 semantic parsing, 2 numerical reasoning, and 7 classification datasets.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.1 Semantic Parsing\",\n \"publication_ref\": [\n \"b28\",\n \"b5\",\n \"b13\",\n \"b16\",\n \"b10\",\n \"b0\",\n \"b5\",\n \"b7\",\n \"b19\",\n \"b7\",\n \"b27\",\n \"b23\",\n \"b17\"\n ],\n \"table_ref\": [\n \"tab_11\"\n ],\n \"text\": \"We use 6 semantic parsing datasets with IID and compositional splits for our experiments. Table 5 shows sample instances from each dataset we experiment with along with the textual template we use to linearize the instances. The ICL prompt is constructed by concatenating the templatized demonstrations and the test instance using \\\\n\\\\n as the separator.\\nGeoQuery (Zelle and Mooney, 1996): A dataset containing 880 natural language questions about US geography paired with Prolog programs. In addition to the standard (IID) split, we experiment with three types of compositional splits: (1) Template split where the training and test sets have disjoint program templates (Finegan-Dollak et al., 2018); (2) TMCD split which creates train and test sets with maximal compound divergence and minimal atom divergence (Keysers et al., 2020); and\\n(3) Length split which evaluates for length generalization by testing on sequences longer than ones in training. Following Levy et al. (2022), we use the compositional splits -three Template, three TMCD, and one Length -generated by Qiu et al.\\n(2022a) and average results across the TMCD and Template splits. ATIS (Hemphill et al., 1990;Dahl et al., 1994):\\nA dataset of natural language queries about aviation paired with λ-calculus programs. We experiment with an IID split and a Template split (Finegan-Dollak et al., 2018) for evaluating compositional generalization, both taken from (Gupta et al., 2022).\\nOvernight (Wang et al., 2015): A dataset containing both synthetic and natural language utterances from 11 domains (e.g. socialnetwork, restaurants, etc.) paired with Lambda-DCS logical forms. We experiment with an IID and a Template split of the socialnetwork domain taken from (Gupta et al., 2022). SMCalFlow (Andreas et al., 2020): A dataset of task-oriented natural language dialogs about calendars, weather, places, and people paired with executable dataflow programs. SMCalFlow-CS (Yin et al., 2021) BREAK (Wolfson et al., 2020) is a dataset that maps complex natural language questions into a language-based meaning representation (QDMR) comprising an ordered list of atomic steps necessary to answer the question. Following (Rubin et al., 2022), we use the low-level Break subset where the targets are logical forms comprising lists of operators with their arguments based on the corresponding QDMR. MTOP (Li et al., 2021): A multilingual taskoriented semantic parsing dataset spanning six languages and 11 domains. The target commands are complex queries featuring nested intent-slot prediction. We use the English subset of MTOP from (Rubin et al., 2022).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"B.2 Non-Semantic Parsing\",\n \"publication_ref\": [\n \"b21\",\n \"b4\",\n \"b22\",\n \"b2\",\n \"b30\"\n ],\n \"table_ref\": [],\n \"text\": \"We additionally experiment with the standard IID splits of 9 non-semantic parsing datasets from the following categories: Numerical Reasoning: For this category, we experiment with GSM8K (Cobbe et al., 2021), a chain-of-thought reasoning (Wei et al., 2023) dataset of grade school-level arithmetic reasoning problems expressed in natural language and DROP (Dua et al., 2019), a dataset of question-answer pairs where the questions are about paragraphs containing numerical information and the answers are spans in the paragraph.\\nClassification: For this category, we experiment with three Natural Language Inference (NLI) datasests (QNLI (Wang et al., 2018), MNLI (Williams et al., 2018), andRTE (Bentivogli et al., 2009)), three Paraphrase Detection datasets (MRPC (Dolan and Brockett, 2005), PAWS (Zhang et al., 2019), and QQP (Wang et al., 2018)) and one Sentiment Classification dataset (SST2 (Socher et al., 2013)).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"C Selection Time\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Despite their O(LT d) time complexity, we found example selection using both BSR and SET-BSR to be fast enough to not be a bottleneck to incontext learning for most datasets considered in this work. By using a GPU to compute c(x, z)s, we could get both to work in the order tens of milliseconds per test input on average which was significantly faster than the LLM inference time itself. The exceptions were DROP, PAWS, QQP, MNLI and QNLI for which the selection took >1 second due to much longer instances and/or larger instance pool. We leave it to future work to explore more efficient ways to measure informativeness.\"\n },\n {\n \"figure_ref\": [\n \"fig_4\",\n \"fig_6\"\n ],\n \"heading\": \"D Additional Analyses\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_11\",\n \"tab_12\"\n ],\n \"text\": \"BM25 From Fig. 6 we can see that coverage-based selection using BM25 with larger substructures outperforms vanilla unigram BM25 in compositional splits.\\nBERTScore-Recall Examining the impact of importance weighting in Fig. 8 which compares the performance change with using importance weighting (IDF) in BSR, we can see that its effect is not consistent across different LLMs. We also did not see any consistent improvement from using larger deberta-large-mnli for computing token embeddings for instance-level BSR (see Fig. 9). However, it did help with set-level selection using SET-BSR.\\nReordering We found the reordering of demonstrations according to the corresponding instance-level metric to only be necessary for smaller LLMs (see Fig. 10), with it even hurting the performance of larger LLMs. We believe this is because larger and code-pretrained LLMs are more capable at composing the salient aspects in the different demonstrations and taking advantage of the full context. deberta-large-mnli outperforms Cosine with all-mpnet-base-v2. Tables 15,16, 17, and 18 show that the same trend holds even when using the same encoder, bert-base-uncased, for both metrics confirming that contextual embeddings are indeed better at capturing salient aspects.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"BSR outperforms\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_7\"\n ],\n \"heading\": \"Recall of Syntactic Structures\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The improvements from set-based selection may be explained by Fig. 11 where we see that set-extensions COSINE and unigram BM25 reduce the recall of substructures of the test input whereas the recalls increase with set-extensions of both BM25[4-GRAM] and BM25[4-DEPST], and even BSR, which does not explicity consider these substructures.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"E Qualitative Analysis of Prompts\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Tables 7, 8 show demonstrations selected using COSINE and SET-BSR for instances from MTOP and SMCalFlow-CS respectively. In each case, CO-SINE find demonstrations that are all very similar to the test input but fails to demonstrate some salient aspect, whereas BSR selects less similar instances but ensures complete coverage of all salient aspects. Tables 9 and 10 additionally illustrate limitations of set-selection and of token-embeddings in capturing salient aspects.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"F All Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_11\",\n \"tab_12\"\n ],\n \"text\": \"Tables 11 contains 8-shot ICL results for our proposed methods and prior learning-free and learningbased demonstration selection on all the LLMs for all the semantic parsing datasets. For numerical reasoning and classification datasets, Tables 12 and 13 compare 8-shot ICL performance with prior training-free and trained methods, respectively. Table 14 provides average performances across all datasets.\\nAdditionally, Tables 15,16, 17, 18, 20, and 21 contain results on semantic parsing datasets of all ablations of learning-free selection methods we ran, with GPT-Neo-2.7B, LLaMA-7B, LLaMA-13B, StarCoder, Cushman, and Codex, respectively. We did not run ablations on GPT-3.5-Turbo due to its cost. Figure 9: Impact on average 8-shot ICL performance on semantic parsing splits from using a larger deberta-large-mnli LLM for computing contextual token embeddings v/s using deberta-base-mnli in BSR and SET-BSR. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgements\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We would like to thank the anonymous reviewers for their feedback. This work was sponsored in part by the DARPA MCS program under Contract No. N660011924033 with the United States Office Of Naval Research and in part by the NSF award #IIS-2046873. The views expressed are those of the authors and do not reflect the policy of the funding agencies.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Contextual token embeddings require the salient aspects to be expressed in text and hence may not be able to capture them for all tasks. Moreover, since it requires computing a dot product for every pair of test and candidate instance tokens, this causes it to scale quadratically with the average number of tokens making it computationally infeasible for tasks with very long textual linearizations. Future work can thus explore more general characterizations of salient aspects and more efficient methods for selecting demonstrations covering them.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"//github.com/Shivanshu-\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"]. COSINE's reliance on a single dense embedding means it is unable to account for the fact that \\\"options\\\" could mean dishes and not just recipes.\\nSelector Prompt\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"COSINE\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Sentence : I need a meeting with Elli tomorrow at 11 pm Logical Form : CreateEvent ( AND ( with_attendee (\\\" Elli \\\") , starts_at ( Tomorrow () ) , starts_at ( NumberPM (11) ) ) )\\nSentence : Set a meeting with Elli for tomorrow at 2 pm through the end of the day and call it Recap Logical Form : CreateEvent ( AND ( ends_at ( AND ( GE ( DateTime ?( date = Tomorrow () , time = NumberPM (2) ) ) , EndOfWorkDay () ) ) , with_attendee (\\\" Elli \\\") , has_subject (\\\" Recap \\\") , starts_at ( Tomorrow () ) , starts_at ( NumberPM (2) ) ) )\\nSentence : Schedule a meeting with Elli for tomorrow at 4 pm through the end of the workday Logical Form : CreateEvent ( AND ( ends_at ( AND ( GE ( DateTime ?( date = Tomorrow () , time = NumberPM (4) ) ) , EndOfWorkDay () ) ) , with_attendee (\\\" Elli \\\") , starts_at ( Tomorrow () ) , starts_at ( NumberPM (4) ) ) )\\nSentence : Schedule a meeting with Elli from 4 PM until the end of the day tomorrow . Logical Form : CreateEvent ( AND ( ends_at ( AND ( GE ( DateTime ?( date = Tomorrow () , time = NumberPM (4) ) ) , EndOfWorkDay () ) ) , with_attendee (\\\" Elli \\\") , starts_at ( Tomorrow () ) , starts_at ( NumberPM (4) ) ) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"SET-BSR\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Sentence : I need a doctor 's appointment on Wednesday morning . Logical Form : CreateEvent ( AND ( has_subject (\\\" doctor ' s appointment \\\") , starts_at ( Morning () ) , starts_at ( NextDOW (\\\" WEDNESDAY \\\") ) ) )\\nSentence : I need to see Alice and her boss next Monday at 3 pm . Logical Form : CreateEvent ( AND ( with_attendee (\\\" Alice \\\") , with_attendee ( FindManager (\\\" Alice \\\") ) , starts_at ( NextDOW (\\\" MONDAY \\\") ) , starts_at ( NumberPM (3) ) ) )\\nSentence : Schedule a meeting with Jake , Elli , and Jesse for Friday at 2 pm . Logical Form : CreateEvent ( AND ( with_attendee (\\\" Jesse \\\") , with_attendee (\\\" Jake \\\") , with_attendee (\\\" Elli \\\") , starts_at ( NextDOW (\\\" FRIDAY \\\") ) , starts_at ( NumberPM (2) ) ) )\\nSentence : I need to schedule a meeting with Jeff 's supervisor Lynne for tomorrow at 10 AM . Logical Form : CreateEvent ( AND ( with_attendee (\\\" Lynne \\\") , starts_at ( Tomorrow () ) , starts_at ( NumberAM (10) ) ) ) Since SET-BSR prioritizes coverage of the remaining aspect, it selects an example that has exactly three items whose total length has to be computed but overall is not very similar in reasoning. BSR on the other hand tries to find an example that demonstrates all aspects by itself and happens to find one that partially demonstrates the remaining aspect as well. \"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-11-06"},"doi":{"kind":"string","value":"10.18653/v1/N19-1423"},"references":{"kind":"list like","value":[{"authors":"Deborah A Dahl; Madeleine Bates; Michael Brown; William Fisher; Kate Hunicke-Smith; David Pallett; Christine Pao; Alexander Rudnicky; Elizabeth Shriberg","journal":"","ref_id":"b0","title":"Expanding the scope of the ATIS task: The ATIS-3 corpus","year":"1994-03-08"},{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"Association for Computational Linguistics","ref_id":"b1","title":"BERT: Pre-training of deep bidirectional transformers for language understanding","year":"2019"},{"authors":"William B Dolan; Chris Brockett","journal":"IWP","ref_id":"b2","title":"Automatically constructing a corpus of sentential paraphrases","year":"2005"},{"authors":"Andrew Drozdov; Nathanael Schärli; Ekin Akyürek; Nathan Scales; Xinying Song; Xinyun Chen; Olivier Bousquet; Denny Zhou","journal":"","ref_id":"b3","title":"Compositional semantic parsing with large language models","year":"2023"},{"authors":"Dheeru Dua; Yizhong Wang; Pradeep Dasigi; Gabriel Stanovsky; Sameer Singh; Matt Gardner","journal":"Association for Computational Linguistics","ref_id":"b4","title":"DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs","year":"2019"},{"authors":"Catherine Finegan-Dollak; Jonathan K Kummerfeld; Li Zhang; Karthik Ramanathan; Sesh Sadasivam; Rui Zhang; Dragomir Radev","journal":"Association for Computational Linguistics","ref_id":"b5","title":"Improving textto-SQL evaluation methodology","year":"2018"},{"authors":"Leo Gao; Stella Biderman; Sid Black; Laurence Golding; Travis Hoppe; Charles Foster; Jason Phang; Horace He; Anish Thite; Noa Nabeshima; Shawn Presser; Connor Leahy","journal":"","ref_id":"b6","title":"The pile: An 800gb dataset of diverse text for language modeling","year":"2020"},{"authors":"Shivanshu Gupta; Sameer Singh; Matt Gardner","journal":"Association for Computational Linguistics","ref_id":"b7","title":"Structurally diverse sampling for sampleefficient training and comprehensive evaluation","year":"2022"},{"authors":"Matan Hasson; Jonathan Berant","journal":"","ref_id":"b8","title":"Question decomposition with dependency graphs","year":"2021"},{"authors":"Pengcheng He; Xiaodong Liu; Jianfeng Gao; Weizhu Chen","journal":"","ref_id":"b9","title":"Deberta: decoding-enhanced bert with disentangled attention","year":"2021-05-03"},{"authors":"Charles T Hemphill; John J Godfrey; George R Doddington","journal":"","ref_id":"b10","title":"The ATIS spoken language systems pilot corpus","year":"1990-06-24"},{"authors":"Jeff Johnson; Matthijs Douze; Hervé Jégou","journal":"IEEE Transactions on Big Data","ref_id":"b11","title":"Billion-scale similarity search with GPUs","year":"2019"},{"authors":"Karen Spärck Jones; Steve Walker; Stephen E Robertson","journal":"Inf. Process. Manag","ref_id":"b12","title":"A probabilistic model of information retrieval: development and comparative experiments -part 2","year":"2000"},{"authors":"Daniel Keysers; Nathanael Schärli; Nathan Scales; Hylke Buisman; Daniel Furrer; Sergii Kashubin; Nikola Momchev; Danila Sinopalnikov; Lukasz Stafiniak; Tibor Tihon; Dmitry Tsarkov; Xiao Wang; Marc Van Zee; Olivier Bousquet","journal":"","ref_id":"b13","title":"Measuring compositional generalization: A comprehensive method on realistic data","year":"2020-04-26"},{"authors":"Denis Kocetkov; Raymond Li; Loubna Ben Allal; Jia Li; Chenghao Mou; Carlos Muñoz Ferrandis; Yacine Jernite; Margaret Mitchell; Sean Hughes; Thomas Wolf; Dzmitry Bahdanau; Leandro Von Werra; Harm De Vries","journal":"","ref_id":"b14","title":"The stack: 3 tb of permissively licensed source code","year":"2022"},{"authors":"Alex Kulesza","journal":"Foundations and Trends® in Machine Learning","ref_id":"b15","title":"Determinantal point processes for machine learning","year":"2012"},{"authors":"Itay Levy; Ben Bogin; Jonathan Berant","journal":"","ref_id":"b16","title":"Diverse demonstrations improve in-context compositional generalization","year":"2022"},{"authors":"Haoran Li; Abhinav Arora; Shuohui Chen; Anchit Gupta; Sonal Gupta; Yashar Mehdad","journal":"Association for Computational Linguistics","ref_id":"b17","title":"MTOP: A comprehensive multilingual task-oriented semantic parsing benchmark","year":"2021"},{"authors":"Xinyi Wang; Wanrong Zhu; Michael Saxon; Mark Steyvers; William Yang; Wang ","journal":"","ref_id":"b18","title":"Large language models are implicitly topic models: Explaining and finding good demonstrations for in-context learning","year":"2023"},{"authors":"Yushi Wang; Jonathan Berant; Percy Liang","journal":"Association for Computational Linguistics","ref_id":"b19","title":"Building a semantic parser overnight","year":"2015"},{"authors":"Jason Wei; Yi Tay; Rishi Bommasani; Colin Raffel; Barret Zoph; Sebastian Borgeaud; Dani Yogatama; Maarten Bosma; Denny Zhou; Donald Metzler; Ed H Chi; Tatsunori Hashimoto; Oriol Vinyals; Percy Liang; Jeff Dean; William Fedus","journal":"","ref_id":"b20","title":"Emergent abilities of large language models","year":"2022"},{"authors":"Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Brian Ichter; Fei Xia; Ed Chi; Quoc Le; Denny Zhou","journal":"","ref_id":"b21","title":"Chain-of-thought prompting elicits reasoning in large language models","year":"2023"},{"authors":"Adina Williams; Nikita Nangia; Samuel Bowman","journal":"Association for Computational Linguistics","ref_id":"b22","title":"A broad-coverage challenge corpus for sentence understanding through inference","year":"2018"},{"authors":"Tomer Wolfson; Mor Geva; Ankit Gupta; Matt Gardner; Yoav Goldberg; Daniel Deutch; Jonathan Berant","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b23","title":"Break it down: A question understanding benchmark","year":"2020"},{"authors":"Zhiyong Wu; Yaoxiang Wang; Jiacheng Ye; Lingpeng Kong","journal":"","ref_id":"b24","title":"Self-adaptive in-context learning: An information compression perspective for incontext example selection and ordering","year":"2023"},{"authors":"Jiacheng Ye; Zhiyong Wu; Jiangtao Feng; Tao Yu; Lingpeng Kong","journal":"","ref_id":"b25","title":"Compositional exemplars for in-context learning","year":"2023"},{"authors":"Xi Ye; Srinivasan Iyer; Asli Celikyilmaz; Ves Stoyanov; Greg Durrett; Ramakanth Pasunuru","journal":"","ref_id":"b26","title":"Complementary explanations for effective in-context learning","year":"2022"},{"authors":"Pengcheng Yin; Hao Fang; Graham Neubig; Adam Pauls; Antonios Emmanouil; Yu Platanios; Sam Su; Jacob Thomson; Andreas","journal":"Association for Computational Linguistics","ref_id":"b27","title":"Compositional generalization for neural semantic parsing via spanlevel supervised attention","year":"2021"},{"authors":"John M Zelle; Raymond J Mooney","journal":"AAAI Press","ref_id":"b28","title":"Learning to parse database queries using inductive logic programming","year":"1996"},{"authors":"Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi","journal":"","ref_id":"b29","title":"Bertscore: Evaluating text generation with BERT","year":"2020-04-26"},{"authors":"Yuan Zhang; Jason Baldridge; Luheng He","journal":"Association for Computational Linguistics","ref_id":"b30","title":"PAWS: Paraphrase adversaries from word scrambling","year":"2019"},{"authors":"Zihao Zhao; Eric Wallace; Shi Feng; Dan Klein; Sameer Singh","journal":"PMLR","ref_id":"b31","title":"Calibrate before use: Improving few-shot performance of language models","year":"2021-07"}],"string":"[\n {\n \"authors\": \"Deborah A Dahl; Madeleine Bates; Michael Brown; William Fisher; Kate Hunicke-Smith; David Pallett; Christine Pao; Alexander Rudnicky; Elizabeth Shriberg\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Expanding the scope of the ATIS task: The ATIS-3 corpus\",\n \"year\": \"1994-03-08\"\n },\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b1\",\n \"title\": \"BERT: Pre-training of deep bidirectional transformers for language understanding\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"William B Dolan; Chris Brockett\",\n \"journal\": \"IWP\",\n \"ref_id\": \"b2\",\n \"title\": \"Automatically constructing a corpus of sentential paraphrases\",\n \"year\": \"2005\"\n },\n {\n \"authors\": \"Andrew Drozdov; Nathanael Schärli; Ekin Akyürek; Nathan Scales; Xinying Song; Xinyun Chen; Olivier Bousquet; Denny Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Compositional semantic parsing with large language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Dheeru Dua; Yizhong Wang; Pradeep Dasigi; Gabriel Stanovsky; Sameer Singh; Matt Gardner\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b4\",\n \"title\": \"DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Catherine Finegan-Dollak; Jonathan K Kummerfeld; Li Zhang; Karthik Ramanathan; Sesh Sadasivam; Rui Zhang; Dragomir Radev\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b5\",\n \"title\": \"Improving textto-SQL evaluation methodology\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Leo Gao; Stella Biderman; Sid Black; Laurence Golding; Travis Hoppe; Charles Foster; Jason Phang; Horace He; Anish Thite; Noa Nabeshima; Shawn Presser; Connor Leahy\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"The pile: An 800gb dataset of diverse text for language modeling\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Shivanshu Gupta; Sameer Singh; Matt Gardner\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b7\",\n \"title\": \"Structurally diverse sampling for sampleefficient training and comprehensive evaluation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Matan Hasson; Jonathan Berant\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Question decomposition with dependency graphs\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Pengcheng He; Xiaodong Liu; Jianfeng Gao; Weizhu Chen\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Deberta: decoding-enhanced bert with disentangled attention\",\n \"year\": \"2021-05-03\"\n },\n {\n \"authors\": \"Charles T Hemphill; John J Godfrey; George R Doddington\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"The ATIS spoken language systems pilot corpus\",\n \"year\": \"1990-06-24\"\n },\n {\n \"authors\": \"Jeff Johnson; Matthijs Douze; Hervé Jégou\",\n \"journal\": \"IEEE Transactions on Big Data\",\n \"ref_id\": \"b11\",\n \"title\": \"Billion-scale similarity search with GPUs\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Karen Spärck Jones; Steve Walker; Stephen E Robertson\",\n \"journal\": \"Inf. Process. Manag\",\n \"ref_id\": \"b12\",\n \"title\": \"A probabilistic model of information retrieval: development and comparative experiments -part 2\",\n \"year\": \"2000\"\n },\n {\n \"authors\": \"Daniel Keysers; Nathanael Schärli; Nathan Scales; Hylke Buisman; Daniel Furrer; Sergii Kashubin; Nikola Momchev; Danila Sinopalnikov; Lukasz Stafiniak; Tibor Tihon; Dmitry Tsarkov; Xiao Wang; Marc Van Zee; Olivier Bousquet\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Measuring compositional generalization: A comprehensive method on realistic data\",\n \"year\": \"2020-04-26\"\n },\n {\n \"authors\": \"Denis Kocetkov; Raymond Li; Loubna Ben Allal; Jia Li; Chenghao Mou; Carlos Muñoz Ferrandis; Yacine Jernite; Margaret Mitchell; Sean Hughes; Thomas Wolf; Dzmitry Bahdanau; Leandro Von Werra; Harm De Vries\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"The stack: 3 tb of permissively licensed source code\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Alex Kulesza\",\n \"journal\": \"Foundations and Trends® in Machine Learning\",\n \"ref_id\": \"b15\",\n \"title\": \"Determinantal point processes for machine learning\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Itay Levy; Ben Bogin; Jonathan Berant\",\n \"journal\": \"\",\n \"ref_id\": \"b16\",\n \"title\": \"Diverse demonstrations improve in-context compositional generalization\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Haoran Li; Abhinav Arora; Shuohui Chen; Anchit Gupta; Sonal Gupta; Yashar Mehdad\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b17\",\n \"title\": \"MTOP: A comprehensive multilingual task-oriented semantic parsing benchmark\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Xinyi Wang; Wanrong Zhu; Michael Saxon; Mark Steyvers; William Yang; Wang \",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Large language models are implicitly topic models: Explaining and finding good demonstrations for in-context learning\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Yushi Wang; Jonathan Berant; Percy Liang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b19\",\n \"title\": \"Building a semantic parser overnight\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Jason Wei; Yi Tay; Rishi Bommasani; Colin Raffel; Barret Zoph; Sebastian Borgeaud; Dani Yogatama; Maarten Bosma; Denny Zhou; Donald Metzler; Ed H Chi; Tatsunori Hashimoto; Oriol Vinyals; Percy Liang; Jeff Dean; William Fedus\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Emergent abilities of large language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Brian Ichter; Fei Xia; Ed Chi; Quoc Le; Denny Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Chain-of-thought prompting elicits reasoning in large language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Adina Williams; Nikita Nangia; Samuel Bowman\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b22\",\n \"title\": \"A broad-coverage challenge corpus for sentence understanding through inference\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Tomer Wolfson; Mor Geva; Ankit Gupta; Matt Gardner; Yoav Goldberg; Daniel Deutch; Jonathan Berant\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b23\",\n \"title\": \"Break it down: A question understanding benchmark\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Zhiyong Wu; Yaoxiang Wang; Jiacheng Ye; Lingpeng Kong\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Self-adaptive in-context learning: An information compression perspective for incontext example selection and ordering\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Jiacheng Ye; Zhiyong Wu; Jiangtao Feng; Tao Yu; Lingpeng Kong\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Compositional exemplars for in-context learning\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Xi Ye; Srinivasan Iyer; Asli Celikyilmaz; Ves Stoyanov; Greg Durrett; Ramakanth Pasunuru\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Complementary explanations for effective in-context learning\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Pengcheng Yin; Hao Fang; Graham Neubig; Adam Pauls; Antonios Emmanouil; Yu Platanios; Sam Su; Jacob Thomson; Andreas\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b27\",\n \"title\": \"Compositional generalization for neural semantic parsing via spanlevel supervised attention\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"John M Zelle; Raymond J Mooney\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b28\",\n \"title\": \"Learning to parse database queries using inductive logic programming\",\n \"year\": \"1996\"\n },\n {\n \"authors\": \"Tianyi Zhang; Varsha Kishore; Felix Wu; Kilian Q Weinberger; Yoav Artzi\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Bertscore: Evaluating text generation with BERT\",\n \"year\": \"2020-04-26\"\n },\n {\n \"authors\": \"Yuan Zhang; Jason Baldridge; Luheng He\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b30\",\n \"title\": \"PAWS: Paraphrase adversaries from word scrambling\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Zihao Zhao; Eric Wallace; Shi Feng; Dan Klein; Sameer Singh\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b31\",\n \"title\": \"Calibrate before use: Improving few-shot performance of language models\",\n \"year\": \"2021-07\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[2,313.81,763.54,211.33,10.84],"formula_id":"formula_0","formula_text":"y test ∼ P LM (• | x 1 , y 1 , . . . , x K , y K , x test ) (1)"},{"formula_coordinates":[3,105.52,175.83,130.66,15.24],"formula_id":"formula_1","formula_text":"T = {z i } N i=1 = {(x i , y i )} N i=1"},{"formula_coordinates":[3,83.45,527.85,206.42,24.32],"formula_id":"formula_2","formula_text":"tfidf(x test , z) = s∈Tx test idf(s)tf(s, T z ) (2)"},{"formula_coordinates":[3,343.57,269.05,181.57,24.32],"formula_id":"formula_3","formula_text":"cover (x test , z) = s∈Sx test c(s, z)(3)"},{"formula_coordinates":[3,408.07,472.67,117.07,33.58],"formula_id":"formula_4","formula_text":"= d s=1 r xtest [s] • r z [s] (4)"},{"formula_coordinates":[4,104.83,247.41,185.03,24.87],"formula_id":"formula_5","formula_text":"BSR(x, z) = x i ∈x w(x i ) max j x T i z j (5)"},{"formula_coordinates":[4,87.22,752.32,202.65,24.32],"formula_id":"formula_6","formula_text":"setcov (x test , Z) = s∈Sx test max z∈Z c(s, z) (6)"},{"formula_coordinates":[4,339.59,692.37,146.38,45.35],"formula_id":"formula_7","formula_text":"setcov x test , Z ∪ z ′ = s∈Sx test max c(s, Z), c(s, z ′ )"},{"formula_coordinates":[5,70.87,74.37,220.08,36.67],"formula_id":"formula_8","formula_text":"O(kN L), where L = |S xtest |. For BSR, the com- plexity of computing c(x, z) for all z ∈ Z is O(T d),"},{"formula_coordinates":[13,92.68,188.63,196.45,23.12],"formula_id":"formula_9","formula_text":"(X ∪{x})-f (X) ≥ f (Y ∪ {x}) -f (Y )."},{"formula_coordinates":[13,92.68,226.59,198.56,23.12],"formula_id":"formula_10","formula_text":"f (S) + f (T ) ≥ f (S ∪ T ) + f (S ∩ T )."},{"formula_coordinates":[13,70.51,264.56,220.74,67.77],"formula_id":"formula_11","formula_text":"x 1 , x 2 ∈ Ω\\X such that x 1 ̸ = x 2 we have that f (X ∪ {x 1 }) + f (X ∪ {x 2 }) ≥ f (X ∪ {x 1 , x 2 }) + f (X). Theorem A.1. The function f maxw (X) = max x∈X w x"},{"formula_coordinates":[13,79.52,384.98,202.55,60.87],"formula_id":"formula_12","formula_text":"x 1 , x 2 ∈ Ω such that w x 1 > w x 2 : 1. f maxw (X ∪ {x}) ≥ f (X) 2. f maxw (X ∪ {x 1 }) = f maxw (X ∪ {x 1 , x 2 })"},{"formula_coordinates":[13,70.51,509.92,101.82,12.58],"formula_id":"formula_13","formula_text":"Theorem A.2. If {f i } n"},{"formula_coordinates":[13,75.84,584.39,213.29,84.3],"formula_id":"formula_14","formula_text":"(f 1 + f 2 )(X 1 ∪ X 2 ) + (f 1 + f 2 )(X 1 ∩ X 2 ) = (f 1 (X 1 ∪ X 2 ) + f 1 (X 1 ∩ X 2 )) + (f 2 (X 1 ∪ X 2 ) + f 2 (X 1 ∩ X 2 )) ≤ (f 1 (X 1 ) + f 1 (X 2 )) + (f 2 (X 1 ) + f 2 (X 2 )) = (f 1 + f 2 )(X 1 ) + (f 1 + f 2 )(X 2 ) (8)"},{"formula_coordinates":[13,306.14,133.06,218.27,31.66],"formula_id":"formula_15","formula_text":"s∈Sx test f s (Z) is submodular."}],"string":"[\n {\n \"formula_coordinates\": [\n 2,\n 313.81,\n 763.54,\n 211.33,\n 10.84\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"y test ∼ P LM (• | x 1 , y 1 , . . . , x K , y K , x test ) (1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 105.52,\n 175.83,\n 130.66,\n 15.24\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"T = {z i } N i=1 = {(x i , y i )} N i=1\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 83.45,\n 527.85,\n 206.42,\n 24.32\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"tfidf(x test , z) = s∈Tx test idf(s)tf(s, T z ) (2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 343.57,\n 269.05,\n 181.57,\n 24.32\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"cover (x test , z) = s∈Sx test c(s, z)(3)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 408.07,\n 472.67,\n 117.07,\n 33.58\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"= d s=1 r xtest [s] • r z [s] (4)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 104.83,\n 247.41,\n 185.03,\n 24.87\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"BSR(x, z) = x i ∈x w(x i ) max j x T i z j (5)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 87.22,\n 752.32,\n 202.65,\n 24.32\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"setcov (x test , Z) = s∈Sx test max z∈Z c(s, z) (6)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 339.59,\n 692.37,\n 146.38,\n 45.35\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"setcov x test , Z ∪ z ′ = s∈Sx test max c(s, Z), c(s, z ′ )\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 70.87,\n 74.37,\n 220.08,\n 36.67\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"O(kN L), where L = |S xtest |. For BSR, the com- plexity of computing c(x, z) for all z ∈ Z is O(T d),\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 92.68,\n 188.63,\n 196.45,\n 23.12\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"(X ∪{x})-f (X) ≥ f (Y ∪ {x}) -f (Y ).\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 92.68,\n 226.59,\n 198.56,\n 23.12\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"f (S) + f (T ) ≥ f (S ∪ T ) + f (S ∩ T ).\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 70.51,\n 264.56,\n 220.74,\n 67.77\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"x 1 , x 2 ∈ Ω\\\\X such that x 1 ̸ = x 2 we have that f (X ∪ {x 1 }) + f (X ∪ {x 2 }) ≥ f (X ∪ {x 1 , x 2 }) + f (X). Theorem A.1. The function f maxw (X) = max x∈X w x\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 79.52,\n 384.98,\n 202.55,\n 60.87\n ],\n \"formula_id\": \"formula_12\",\n \"formula_text\": \"x 1 , x 2 ∈ Ω such that w x 1 > w x 2 : 1. f maxw (X ∪ {x}) ≥ f (X) 2. f maxw (X ∪ {x 1 }) = f maxw (X ∪ {x 1 , x 2 })\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 70.51,\n 509.92,\n 101.82,\n 12.58\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"Theorem A.2. If {f i } n\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 75.84,\n 584.39,\n 213.29,\n 84.3\n ],\n \"formula_id\": \"formula_14\",\n \"formula_text\": \"(f 1 + f 2 )(X 1 ∪ X 2 ) + (f 1 + f 2 )(X 1 ∩ X 2 ) = (f 1 (X 1 ∪ X 2 ) + f 1 (X 1 ∩ X 2 )) + (f 2 (X 1 ∪ X 2 ) + f 2 (X 1 ∩ X 2 )) ≤ (f 1 (X 1 ) + f 1 (X 2 )) + (f 2 (X 1 ) + f 2 (X 2 )) = (f 1 + f 2 )(X 1 ) + (f 1 + f 2 )(X 2 ) (8)\"\n },\n {\n \"formula_coordinates\": [\n 13,\n 306.14,\n 133.06,\n 218.27,\n 31.66\n ],\n \"formula_id\": \"formula_15\",\n \"formula_text\": \"s∈Sx test f s (Z) is submodular.\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"Coverage-based Example Selection for In-Context Learning"},"abstract":{"kind":"string","value":"In-context learning (ICL), the ability of large language models to perform novel tasks by conditioning on a prompt with a few task examples, requires these examples to be informative about the test instance. The standard approach of independently ranking and selecting the most similar examples selects redundant examples while omitting important information. In this work, we show that BERTScore-Recall (BSR) selects better examples that demonstrate more of the salient aspects, e.g. reasoning patterns, of the test input. We further extend BSR and many standard metrics to easily optimizable set-level metrics, giving still better coverage of those salient aspects. On 15 datasets spanning 6 tasks and with 7 diverse LLMs, we show that (1) BSR is the superior metric for in-context example selection across the board, and (2) for compositional tasks, set selection using Set-BSR outperforms independent ranking by up to 17 points on average and, despite being trainingfree, surpasses methods that leverage task or LLM-specific training."},"authors":{"kind":"string","value":"Shivanshu Gupta; Matt Gardner; Sameer Singh; Sweta Agrawal; Chunting Zhou; Mike Lewis; Luke Zettlemoyer; Marjan 2022 Ghazvininejad; Jacob Andreas; John Bufe; David Burkett; Charles Chen; Josh Clausman; Jean Crawford; Kate Crim; Jordan Deloach; Leah Dorner; Jason Eisner; Hao Fang; Alan Guo; David Hall; Kristin Hayes; Kellie Hill; Diana Ho; Wendy Iwaszuk; Smriti Jha; Dan Klein; Jayant Krishnamurthy; Theo Lanman; Percy Liang; Christopher H Lin; Ilya Lintsbakh; Andy Mc- Govern; Aleksandr Nisnevich; Adam Pauls; Dmitrij Petters; Brent Read; Dan Roth; Subhro Roy; Jesse Rusak; Beth Short; Div Slomin; Ben Snyder; Stephon Striplin; Yu Su; Zachary Tellman; Sam Thomson; An- Drei Vorobev; Izabela Witoszko; Jason Wolfe; Abby Wray; Yuchen Zhang; Alexander 2020 Zotov; Yuhuai Wu; Anders Johan Andreassen; Aitor Lewkowycz; Sid Black; Leo Gao; Phil Wang; Connor Leahy; Stella 2021 Biderman; Tom B Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel M Ziegler; Jeffrey Wu; Clemens Winter; Christopher Hesse; Mark Chen; Eric Sigler; Mateusz Litwin; Scott Gray; Benjamin Chess; Jack Clark; Christopher Berner; Jerry Tworek; Heewoo Jun; Qiming Yuan; Henrique Ponde; Oliveira Pinto; Jared Ka- Plan; Harri Edwards; Yuri Burda; Nicholas Joseph; Greg Brockman; Alex Ray; Raul Puri; Michael Petrov; Heidy Khlaaf; Girish Sas- Try; Pamela Mishkin; Brooke Chan; Mikhail Pavlov; Alethea Power; Lukasz Kaiser; Mohammad Bavarian; Philippe Tillet; Felipe Petroski Such; Dave Cum- Mings; Matthias Plappert; Fotios Chantzis; Eliza- Beth Barnes; William Hebgen Guss; Alex Nichol; Alex Paino; Nikolas Tezak; Jie Tang; Igor Babuschkin; Suchir Balaji; Shantanu Jain; William Saunders; Andrew N Carr; Jan Leike; Josh Achiam; Vedant Misra; Evan Morikawa; Alec Radford; Matthew Knight; Miles Brundage; Mira Murati; Katie Mayer; Peter Welinder; Bob Mcgrew; Dario Amodei; Sam Mccandlish; Ilya Sutskever; Wojciech 2021 Zaremba; Karl Evaluating; Vineet Cobbe; Mohammad Kosaraju; Mark Bavarian; Heewoo Chen; Lukasz Jun; Matthias Kaiser; Jerry Plappert; Jacob Tworek; Reiichiro Hilton; Christopher Nakano; John Hesse; 2021 Schulman; Raymond Training; Ben Li; Yangtian Allal; Niklas Zi; Denis Muennighoff; Chenghao Kocetkov; Marc Mou; Christopher Marone; Jia Akiki; Jenny Li; Qian Chim; Evgenii Liu; Terry Yue Zheltonozhskii; Thomas Zhuo; Olivier Wang; Mishig Dehaene; Joel Davaadorj; João Lamy-Poirier; Oleh Monteiro; Nicolas Shliazhko; Nicholas Gontier; Armel Meade; Ming-Ho Zebaze; Kumar Yee; Jian Umapathi; Benjamin Zhu; Muhtasham Lipkin; Zhiruo Oblokulov; Rudra Wang; Jason Murthy; Stillerman; Sankalp Siva; Dmitry Patel; Marco Abulkhanov; Manan Zocca; Zhihan Dey; Nour Zhang; Urvashi Fahmy; Wenhao Bhattacharyya; Swayam Yu; Sasha Singh; Paulo Luccioni; Maxim Villegas; Fedor Kunakov; Manuel Zhdanov; Tony Romero; Nadav Lee; Jennifer Timor; Claire Ding; Hailey Schlesinger; Jan Schoelkopf; Tri Ebert; Mayank Dao; Alex Mishra; Jennifer Gu; Carolyn Jane Robinson; Brendan Anderson; Dan- Ish Dolan-Gavitt; Siva Contractor; Daniel Reddy; Dzmitry Fried; Yacine Bahdanau; Carlos Muñoz Jernite; Sean Ferrandis; Thomas Hughes"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: (a) Test input with salient aspects highlighted. (a) Independently selecting similar examples leads to redundancy and failure to demonstrate all salient aspects, in this case, the need to identify the manager. (b) Coverage-based selection using SET-BSR mitigates this by selecting a less similar example that contains the missing information. Blue indicates LLM generation.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure2: Gain in average ICL accuracy compared to COSINE on IID and COMPositional splits in semantic parsing. Trained methods (EPR and CEIL) become less effective with larger LLMs on IID splits. This is unlike SET-BSR, which, on compositional splits, even becomes more effective with larger LLMs.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure3: Change in average performance on different types of splits of semantic parsing datasets from set-selection using our set metrics v/s the corresponding instance-level metric. Coverage-based set selection is most useful in compositional splits and when covering larger syntactic structures (BM25) or contextual embeddings (BSR).","figure_data":"","figure_id":"fig_2","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Average performance on IID and COMP semantic parsing splits with Codex. SET-BSR consistently outperforms independent selection.","figure_data":"","figure_id":"fig_3","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 6 :6Figure6: Absolute improvement in average 8-shot ICL performance on different types of semantic parsing splits from using the set extensions SET-BM25 with larger substructures over vanilla BM25.","figure_data":"","figure_id":"fig_4","figure_label":"6","figure_type":"figure"},{"figure_caption":"Figure 7 :Figure 8 :78Figure7: Comparison of 8-shot ICL performance of different variants of BERTScore with token embeddings computed using deberta-base-mnli. For easier visualization, since we found BERTScore-Precision to consistently perform worst, we show absolute improvement in average performance on different types of splits from the recall and F1 metrics over the precision metric.","figure_data":"","figure_id":"fig_5","figure_label":"78","figure_type":"figure"},{"figure_caption":"Figure 10 :10Figure10: Impact on average 8-shot ICL performance on semantic parsing splits from reordering the demonstrations selected by the different set-level metric using the corresponding instance-level metric as absolute gain v/s the unreordered version.","figure_data":"","figure_id":"fig_6","figure_label":"10","figure_type":"figure"},{"figure_caption":"Figure 11 :11Figure 11: Coverage of aspects of the test instance: Change in recall of unigrams, 4-grams, and dependency parse subtrees (size < 4) in the test input with set-selection of demonstrations, compared to their nonset version, averaged across all datasets.","figure_data":"","figure_id":"fig_7","figure_label":"11","figure_type":"figure"},{"figure_caption":"Algorithm 1 Greedy Optimization of Set Coverage","figure_data":"Require: Instance pool T ; test input xtest; desired number ofdemonstrations k; coverage scoring function setcov1: Z ← ∅▷ Selected Demonstrations2: Zcurr ← ∅▷ Current Set Cover3: curr_cov ← -inf4: while |Z|< k do5:z * , next_cov = argmaxsetcov (xtest , Zcurr ∪ z)z∈T -Z6:if next_cov > curr_cov then▷ Pick z *7:curr_cov ← next_cov8:Z ← Z ∪ z *9:Zcurr ← Zcurr ∪ z *10:else▷ Or start new cover11:Zcurr ← ∅, curr_cov ← -inf12:end if13: end while14: return Z","figure_id":"tab_0","figure_label":"","figure_type":"table"},{"figure_caption":"Average 8-shot ICL performance across all splits of semantic parsing datasets using different LLMs and demonstration-selection methods with absolute improvement over COSINE in brackets. Both BSR and SET-BSR outperform prior training-free methods, with the latter outperforming even trained methods with larger LLMs.","figure_data":"","figure_id":"tab_2","figure_label":"1","figure_type":"table"},{"figure_caption":"and concatenate to form the prompt. For set-selection methods, the demonstrations are or-","figure_data":"Selector8_S32_CTraining FreeRANDOM 31.9 (-22.8) 7.4 (-4.5) COSINE 54.7 11.9 BM25 65.4 (+10.7) 29.4 (+17.5)EPR76.3 (+21.6) 21.7 (+9.8)TrainedCEIL77.5 (+22.8) 40.1 (+28.2)LFCOV66.3 (+11.6) 45.9 (+33.9)OursBSR SET-BSR 75.7 (+21.0) 61.2 (+49.3) 72.5 (+17.8) 31.5 (+19.6)","figure_id":"tab_3","figure_label":"","figure_type":"table"},{"figure_caption":"8-shot ICL accuracy on SMCalFlow-CS using Codex with absolute improvement over COSINE in brackets. SET-BSR is competitive with trained methods on the IID split while dramatically outperforming them on the compositional split.","figure_data":"","figure_id":"tab_4","figure_label":"2","figure_type":"table"},{"figure_caption":"Table 1 compares average performance across all semantic parsing splits for seven LLMs of varying sizes. See Table 2 for comparison with LFCOV, which only works with GeoQuery and SMCalFlow-CS and Table11for results on individual splits. While BSR consistently outperforms COSINE and BM25 for all LLMs, set-selection using SET-BSR leads to further dramatic gains with upto 17% improvement over COSINE with Codex, beating even state-of-the-art trained methods like EPR and CEIL by 12 and 6 points, respectively. Further, from 8-shot ICL performance for tasks other than semantic parsing (using GPT-Neo-2.7B for the classification tasks and Codex for the harder GSM8K and DROP). BSR is competitive with prior methods, however, as these are IID splits, SET-BSR doesn't lead to further gains. † 50-shot results fromYe et al. (2023).","figure_data":"Selector GSM8K DROP MNLI PAWS SST2Training FreeRandom 60.6 Cosine 64 BM25 64.862.7 65.4 66.941.9 44.0 52.5 81.9 48 86.9 42.2 55.2 82.6TrainedEPR CEIL61.7 63.1--66.1 † 71.7 †----OursBSR Set-BSR68.1 67.468.1 66.476.7 78.6 74.9 61.5 75 90.9","figure_id":"tab_5","figure_label":"3","figure_type":"table"},{"figure_caption":"Average 8-shot ICL performance with Codex on IID, COMPositional, and ALL semantic parsing splits. Top compares different variants of BERTScore, white Bottom compares the different variants of BM25.","figure_data":"SelectorALL IID COMPBSF160.6 71.0 50.1BSP54.3 65.5 43.2BSR61.2 71.5 50.9BM2560.9 68.9 52.8+ Coverage 56.4 63.4 49.5BM25[4-gram] 59.1 67.1 51.0+ Coverage 64.5 68.9 60.2BM25[4-depst] 57.8 65.5 50.0+ Coverage 64.9 68.6 61.2","figure_id":"tab_6","figure_label":"4","figure_type":"table"},{"figure_caption":"Cosine even with the same encoder In § 7.2, we showed that BSR with","figure_data":"DatasetExample TemplateSample InstanceOvernight {source}\\t{target} source: employees who finish after alices birthdaytarget: (call listValue (call getProperty ((lambda s (call filter (vars) (call ensureNumericProperty (string employment_end_date)) (string>) (call ensureNumericEntity (call getProperty en.person.alice (stringbirthdate))))) (call domain (string employee))) (string employee)))ATIS{source}\\t{target} source: give me the flights from pittsburgh to los angeles thursday eveningtarget: ( lambda $0 e ( and ( flight $0 ) ( during_day $0 evening :pd ) ( from $0 pittsburgh : ci ) ( to $0 los_angeles : ci ) ( day $0thursday : da ) ) )GeoQuery {source}\\t{target} source: which river traverses most statestarget: answer ( most ( river, traverse_2, state ) )SMCalFlow {source}\\t{target} source: Please put a 2 o'clock on my schedule where I'm meeting with boss Daniel.target:CreateEvent(AND(with_attendee(\"Daniel\"),starts_at(NextTime(time=NumberPM(2)))))BREAK{source}\\t{target} source: Is there another cube that is the same size as the cyan cube; what color is it?target: return the cyan cube ;return size of #1 ;return cubes besides#1 ;return sizes of #3 ;return #3 where #4 is the same as #2 ;returncolor of #5MTOP{source}\\t{target} source: latest news from washington times pleasetarget: [IN:GET_STORIES_NEWS [SL:DATE_TIME latest ] [SL:NEWS_TYPE news] [SL:NEWS_SOURCE washington times ] ]","figure_id":"tab_10","figure_label":"","figure_type":"table"},{"figure_caption":"Semantic Parsing Datasets with corresponding sample instances and example templates used in for ICL.","figure_data":"","figure_id":"tab_11","figure_label":"5","figure_type":"table"},{"figure_caption":"Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May? To start the season, the Lions traveled south to Tampa, Florida to take on the Tampa Bay Buccaneers. The Lions scored first in the first quarter with a 23-yard field goal by Jason Hanson. The Buccaneers tied it up with a 38-yard field goal by Connor Barth, then took the lead when Aqib Talib intercepted a pass from Matthew Stafford and ran it in 28 yards. The Lions responded with a 28-yard field goal. In the second quarter, Detroit took the lead with a 36-yard touchdown catch by Calvin Johnson, and later added more points when Tony Scheffler caught an 11-yard TD pass. Tampa Bay responded with a 31-yard field goal just before halftime. The second half was relatively quiet, with each team only scoring one touchdown. First, Detroit's Calvin Johnson caught a 1-yard pass in the third quarter. The game's final points came when Mike Williams of Tampa Bay caught a 5-yard pass. The Lions won their regular season opener for the first time since 2007 question: How many points did the buccaneers need to tie in the first? Unlike the two seasons before it and most of the seasons that followed, Digimon Tamers takes a darker and more realistic approach to its story featuring Digimon who do not reincarnate after their deaths and more complex character development in the original Japanese. question: When did the third Digimon series begin? Dana Reeve, the widow of the actor Christopher Reeve, has died of lung cancer at age 44, according to the Christopher Reeve Foundation. hypothesis: Christopher Reeve had an accident. He said the foodservice pie business doesn 't fit the company 's long-term growth strategy. sentence2: \" The foodservice pie business does not fit our long-term growth strategy . Bradd Crellin represented BARLA Cumbria on a tour of Australia with 6 other players representing Britain , also on a tour of Australia . sentence2: \"Bradd Crellin also represented BARLA Great Britain on a tour through Australia on a tour through Australia with 6 other players representing Cumbria . Non-Semantic Parsing Datasets with corresponding sample instances and example templates used for ICL.","figure_data":"Dataset Example TemplateSample InstanceGSM8K Question: {question} Solution: {solution}question: solution: Natalia sold 48/2 = «48/2=24»24 clips in May. Natalia sold 48+24 =«48+24=72»72 clips altogether in April and May. #### 72DROP Passage: {passage} Question: {question} Answer: {answer}passage: answer: 3QNLIQuestion: {question} Sentence: {sentence} Answer: {label}sentence: label: NoMNLIPremise: {premise}premise: The new rights are nice enoughHypothesis: {hypothesis}hypothesis: Everyone really likes the newest benefitsAnswer: {label}label: MaybeRTEPremise: {premise} Hypothesis: {hypothesis} Answer: {label}premise: label: YesMRPC Sentence 1: {sentence1} Sentence 2: {sentence2} Answer: {label}sentence1: label: YesPAWS Sentence 1: {sentence1} Sentence 2: {sentence2} Answer: {label}sentence1: label: NoQQPQuestion 1: {question1}question1: Why are African-Americans so beautiful?Question 2: {question2}question2: \"Why are hispanics so beautiful?Answer: {label}label: NoSST2Review: {sentence}sentence: it 's a charming and often affecting journey .Answer: {label}label: Positive","figure_id":"tab_12","figure_label":"6","figure_type":"table"},{"figure_caption":"Comparison of various methods on 8-shot ICL for reasoning and classification datasets. BSR outperforms all prior training-free methods though SET-BSR doesn't yield additional improvement.","figure_data":"LMSelector GSM8K DROP QNLI MNLI RTE MRPC PAWS QQP SST2 AVERAGERandom1.56.451.2 41.9 53.4514865.9 86.945.1Cosine1.912.35644 54.2 52.552.575 81.947.8GPT-Neo-2.7BBM253.712.6 58.1 42.2 50.9 57.655.2 71.3 82.648.2BSR2.913.9 81.1 76.7 67.9 70.17586.4 90.962.8Set-BSR2.913.2 78.6 61.5 60.6 68.474.9 84.4 89.859.4Random 11.323.4 54.2 54.3 7033.859.1 66.2 94.251.8Cosine12.326.75858 67.9 46.656.6 76.1 9254.9LLaMA-7BBM2511.926.8 57.3 56.1 68.2 48.357.2 73.2 93.254.7BSR14.527.2 82.9 76.3 70.8 59.87480.4 95.864.6Set-BSR1529.1 74.7 70.2 67.5 53.77280.3 94.261.9Random 15.329.9 56.3 51.3 75.5 72.360.1 67.3 93.157.9Cosine16.732.6 61.8 62.9 75.1 57.158.9 78.7 92.759.6LLaMA-13BBM2516.331.7 63.3 62.8 73.6 62.558.8 76.9 92.759.8BSR19.234.3 85.682 76.5 72.877.5 85.2 95.269.8Set-BSR20.633.38273.3 76.2 67.277.2 81.9 93.867.3","figure_id":"tab_13","figure_label":"12","figure_type":"table"},{"figure_caption":"Additional non-semantic parsing 8-shot ICL results for comparison with trained methods and larger LLMs. BSR is competitive with EPR and CEIL, even outperforming them with larger LLMs. We could not get CEIL to work for DROP.","figure_data":"LMSelector GSM8K DROP QNLIEPR0.013.8 74.9GPT-Neo-2.7BCEIL BSR0.0 2.9-13.9 81.1 84.2Set-BSR2.913.2 78.6EPR--80.9LLaMA-7BCEIL BSR-14.5-27.2 82.9 84.1Set-BSR 15.029.1 74.7EPR--81.9LLaMA-13BCEIL BSR-19.2-34.3 85.6 84.2Set-BSR 20.633.3 82.0EPR--81.7StarCoderCEIL BSR-17.1-26.6 84.7 84.8Set-BSR 17.524.9 80.3EPR10.0-78.9CushmanCEIL BSR8.3 12.1-23.6 84.7 83.8Set-BSR 11.123.7 74.4EPR61.7-83.8CodexCEIL BSR63.1 68.1-68.1 88.7 84.9Set-BSR 67.466.4 84.6GPT-Neo-2.7B LLaMA-7B LLaMA-13BRandom22.525.530.4Cosine36.943.247.7BM2538.544.450.1BSR46.450.656.5Set-BSR45.951.658.2","figure_id":"tab_14","figure_label":"13","figure_type":"table"},{"figure_caption":"Average 8-shot ICL performance across all datasets and splits.","figure_data":"","figure_id":"tab_15","figure_label":"14","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: (a) Test input with salient aspects highlighted. (a) Independently selecting similar examples leads to redundancy and failure to demonstrate all salient aspects, in this case, the need to identify the manager. (b) Coverage-based selection using SET-BSR mitigates this by selecting a less similar example that contains the missing information. Blue indicates LLM generation.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure2: Gain in average ICL accuracy compared to COSINE on IID and COMPositional splits in semantic parsing. Trained methods (EPR and CEIL) become less effective with larger LLMs on IID splits. This is unlike SET-BSR, which, on compositional splits, even becomes more effective with larger LLMs.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure3: Change in average performance on different types of splits of semantic parsing datasets from set-selection using our set metrics v/s the corresponding instance-level metric. Coverage-based set selection is most useful in compositional splits and when covering larger syntactic structures (BM25) or contextual embeddings (BSR).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Average performance on IID and COMP semantic parsing splits with Codex. SET-BSR consistently outperforms independent selection.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 6 :6Figure6: Absolute improvement in average 8-shot ICL performance on different types of semantic parsing splits from using the set extensions SET-BM25 with larger substructures over vanilla BM25.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 7 :Figure 8 :78Figure7: Comparison of 8-shot ICL performance of different variants of BERTScore with token embeddings computed using deberta-base-mnli. For easier visualization, since we found BERTScore-Precision to consistently perform worst, we show absolute improvement in average performance on different types of splits from the recall and F1 metrics over the precision metric.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"78\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 10 :10Figure10: Impact on average 8-shot ICL performance on semantic parsing splits from reordering the demonstrations selected by the different set-level metric using the corresponding instance-level metric as absolute gain v/s the unreordered version.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"10\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 11 :11Figure 11: Coverage of aspects of the test instance: Change in recall of unigrams, 4-grams, and dependency parse subtrees (size < 4) in the test input with set-selection of demonstrations, compared to their nonset version, averaged across all datasets.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"11\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Algorithm 1 Greedy Optimization of Set Coverage\",\n \"figure_data\": \"Require: Instance pool T ; test input xtest; desired number ofdemonstrations k; coverage scoring function setcov1: Z ← ∅▷ Selected Demonstrations2: Zcurr ← ∅▷ Current Set Cover3: curr_cov ← -inf4: while |Z|< k do5:z * , next_cov = argmaxsetcov (xtest , Zcurr ∪ z)z∈T -Z6:if next_cov > curr_cov then▷ Pick z *7:curr_cov ← next_cov8:Z ← Z ∪ z *9:Zcurr ← Zcurr ∪ z *10:else▷ Or start new cover11:Zcurr ← ∅, curr_cov ← -inf12:end if13: end while14: return Z\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Average 8-shot ICL performance across all splits of semantic parsing datasets using different LLMs and demonstration-selection methods with absolute improvement over COSINE in brackets. Both BSR and SET-BSR outperform prior training-free methods, with the latter outperforming even trained methods with larger LLMs.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"and concatenate to form the prompt. For set-selection methods, the demonstrations are or-\",\n \"figure_data\": \"Selector8_S32_CTraining FreeRANDOM 31.9 (-22.8) 7.4 (-4.5) COSINE 54.7 11.9 BM25 65.4 (+10.7) 29.4 (+17.5)EPR76.3 (+21.6) 21.7 (+9.8)TrainedCEIL77.5 (+22.8) 40.1 (+28.2)LFCOV66.3 (+11.6) 45.9 (+33.9)OursBSR SET-BSR 75.7 (+21.0) 61.2 (+49.3) 72.5 (+17.8) 31.5 (+19.6)\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"8-shot ICL accuracy on SMCalFlow-CS using Codex with absolute improvement over COSINE in brackets. SET-BSR is competitive with trained methods on the IID split while dramatically outperforming them on the compositional split.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Table 1 compares average performance across all semantic parsing splits for seven LLMs of varying sizes. See Table 2 for comparison with LFCOV, which only works with GeoQuery and SMCalFlow-CS and Table11for results on individual splits. While BSR consistently outperforms COSINE and BM25 for all LLMs, set-selection using SET-BSR leads to further dramatic gains with upto 17% improvement over COSINE with Codex, beating even state-of-the-art trained methods like EPR and CEIL by 12 and 6 points, respectively. Further, from 8-shot ICL performance for tasks other than semantic parsing (using GPT-Neo-2.7B for the classification tasks and Codex for the harder GSM8K and DROP). BSR is competitive with prior methods, however, as these are IID splits, SET-BSR doesn't lead to further gains. † 50-shot results fromYe et al. (2023).\",\n \"figure_data\": \"Selector GSM8K DROP MNLI PAWS SST2Training FreeRandom 60.6 Cosine 64 BM25 64.862.7 65.4 66.941.9 44.0 52.5 81.9 48 86.9 42.2 55.2 82.6TrainedEPR CEIL61.7 63.1--66.1 † 71.7 †----OursBSR Set-BSR68.1 67.468.1 66.476.7 78.6 74.9 61.5 75 90.9\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Average 8-shot ICL performance with Codex on IID, COMPositional, and ALL semantic parsing splits. Top compares different variants of BERTScore, white Bottom compares the different variants of BM25.\",\n \"figure_data\": \"SelectorALL IID COMPBSF160.6 71.0 50.1BSP54.3 65.5 43.2BSR61.2 71.5 50.9BM2560.9 68.9 52.8+ Coverage 56.4 63.4 49.5BM25[4-gram] 59.1 67.1 51.0+ Coverage 64.5 68.9 60.2BM25[4-depst] 57.8 65.5 50.0+ Coverage 64.9 68.6 61.2\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Cosine even with the same encoder In § 7.2, we showed that BSR with\",\n \"figure_data\": \"DatasetExample TemplateSample InstanceOvernight {source}\\\\t{target} source: employees who finish after alices birthdaytarget: (call listValue (call getProperty ((lambda s (call filter (vars) (call ensureNumericProperty (string employment_end_date)) (string>) (call ensureNumericEntity (call getProperty en.person.alice (stringbirthdate))))) (call domain (string employee))) (string employee)))ATIS{source}\\\\t{target} source: give me the flights from pittsburgh to los angeles thursday eveningtarget: ( lambda $0 e ( and ( flight $0 ) ( during_day $0 evening :pd ) ( from $0 pittsburgh : ci ) ( to $0 los_angeles : ci ) ( day $0thursday : da ) ) )GeoQuery {source}\\\\t{target} source: which river traverses most statestarget: answer ( most ( river, traverse_2, state ) )SMCalFlow {source}\\\\t{target} source: Please put a 2 o'clock on my schedule where I'm meeting with boss Daniel.target:CreateEvent(AND(with_attendee(\\\"Daniel\\\"),starts_at(NextTime(time=NumberPM(2)))))BREAK{source}\\\\t{target} source: Is there another cube that is the same size as the cyan cube; what color is it?target: return the cyan cube ;return size of #1 ;return cubes besides#1 ;return sizes of #3 ;return #3 where #4 is the same as #2 ;returncolor of #5MTOP{source}\\\\t{target} source: latest news from washington times pleasetarget: [IN:GET_STORIES_NEWS [SL:DATE_TIME latest ] [SL:NEWS_TYPE news] [SL:NEWS_SOURCE washington times ] ]\",\n \"figure_id\": \"tab_10\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Semantic Parsing Datasets with corresponding sample instances and example templates used in for ICL.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_11\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May? To start the season, the Lions traveled south to Tampa, Florida to take on the Tampa Bay Buccaneers. The Lions scored first in the first quarter with a 23-yard field goal by Jason Hanson. The Buccaneers tied it up with a 38-yard field goal by Connor Barth, then took the lead when Aqib Talib intercepted a pass from Matthew Stafford and ran it in 28 yards. The Lions responded with a 28-yard field goal. In the second quarter, Detroit took the lead with a 36-yard touchdown catch by Calvin Johnson, and later added more points when Tony Scheffler caught an 11-yard TD pass. Tampa Bay responded with a 31-yard field goal just before halftime. The second half was relatively quiet, with each team only scoring one touchdown. First, Detroit's Calvin Johnson caught a 1-yard pass in the third quarter. The game's final points came when Mike Williams of Tampa Bay caught a 5-yard pass. The Lions won their regular season opener for the first time since 2007 question: How many points did the buccaneers need to tie in the first? Unlike the two seasons before it and most of the seasons that followed, Digimon Tamers takes a darker and more realistic approach to its story featuring Digimon who do not reincarnate after their deaths and more complex character development in the original Japanese. question: When did the third Digimon series begin? Dana Reeve, the widow of the actor Christopher Reeve, has died of lung cancer at age 44, according to the Christopher Reeve Foundation. hypothesis: Christopher Reeve had an accident. He said the foodservice pie business doesn 't fit the company 's long-term growth strategy. sentence2: \\\" The foodservice pie business does not fit our long-term growth strategy . Bradd Crellin represented BARLA Cumbria on a tour of Australia with 6 other players representing Britain , also on a tour of Australia . sentence2: \\\"Bradd Crellin also represented BARLA Great Britain on a tour through Australia on a tour through Australia with 6 other players representing Cumbria . Non-Semantic Parsing Datasets with corresponding sample instances and example templates used for ICL.\",\n \"figure_data\": \"Dataset Example TemplateSample InstanceGSM8K Question: {question} Solution: {solution}question: solution: Natalia sold 48/2 = «48/2=24»24 clips in May. Natalia sold 48+24 =«48+24=72»72 clips altogether in April and May. #### 72DROP Passage: {passage} Question: {question} Answer: {answer}passage: answer: 3QNLIQuestion: {question} Sentence: {sentence} Answer: {label}sentence: label: NoMNLIPremise: {premise}premise: The new rights are nice enoughHypothesis: {hypothesis}hypothesis: Everyone really likes the newest benefitsAnswer: {label}label: MaybeRTEPremise: {premise} Hypothesis: {hypothesis} Answer: {label}premise: label: YesMRPC Sentence 1: {sentence1} Sentence 2: {sentence2} Answer: {label}sentence1: label: YesPAWS Sentence 1: {sentence1} Sentence 2: {sentence2} Answer: {label}sentence1: label: NoQQPQuestion 1: {question1}question1: Why are African-Americans so beautiful?Question 2: {question2}question2: \\\"Why are hispanics so beautiful?Answer: {label}label: NoSST2Review: {sentence}sentence: it 's a charming and often affecting journey .Answer: {label}label: Positive\",\n \"figure_id\": \"tab_12\",\n \"figure_label\": \"6\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison of various methods on 8-shot ICL for reasoning and classification datasets. BSR outperforms all prior training-free methods though SET-BSR doesn't yield additional improvement.\",\n \"figure_data\": \"LMSelector GSM8K DROP QNLI MNLI RTE MRPC PAWS QQP SST2 AVERAGERandom1.56.451.2 41.9 53.4514865.9 86.945.1Cosine1.912.35644 54.2 52.552.575 81.947.8GPT-Neo-2.7BBM253.712.6 58.1 42.2 50.9 57.655.2 71.3 82.648.2BSR2.913.9 81.1 76.7 67.9 70.17586.4 90.962.8Set-BSR2.913.2 78.6 61.5 60.6 68.474.9 84.4 89.859.4Random 11.323.4 54.2 54.3 7033.859.1 66.2 94.251.8Cosine12.326.75858 67.9 46.656.6 76.1 9254.9LLaMA-7BBM2511.926.8 57.3 56.1 68.2 48.357.2 73.2 93.254.7BSR14.527.2 82.9 76.3 70.8 59.87480.4 95.864.6Set-BSR1529.1 74.7 70.2 67.5 53.77280.3 94.261.9Random 15.329.9 56.3 51.3 75.5 72.360.1 67.3 93.157.9Cosine16.732.6 61.8 62.9 75.1 57.158.9 78.7 92.759.6LLaMA-13BBM2516.331.7 63.3 62.8 73.6 62.558.8 76.9 92.759.8BSR19.234.3 85.682 76.5 72.877.5 85.2 95.269.8Set-BSR20.633.38273.3 76.2 67.277.2 81.9 93.867.3\",\n \"figure_id\": \"tab_13\",\n \"figure_label\": \"12\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Additional non-semantic parsing 8-shot ICL results for comparison with trained methods and larger LLMs. BSR is competitive with EPR and CEIL, even outperforming them with larger LLMs. We could not get CEIL to work for DROP.\",\n \"figure_data\": \"LMSelector GSM8K DROP QNLIEPR0.013.8 74.9GPT-Neo-2.7BCEIL BSR0.0 2.9-13.9 81.1 84.2Set-BSR2.913.2 78.6EPR--80.9LLaMA-7BCEIL BSR-14.5-27.2 82.9 84.1Set-BSR 15.029.1 74.7EPR--81.9LLaMA-13BCEIL BSR-19.2-34.3 85.6 84.2Set-BSR 20.633.3 82.0EPR--81.7StarCoderCEIL BSR-17.1-26.6 84.7 84.8Set-BSR 17.524.9 80.3EPR10.0-78.9CushmanCEIL BSR8.3 12.1-23.6 84.7 83.8Set-BSR 11.123.7 74.4EPR61.7-83.8CodexCEIL BSR63.1 68.1-68.1 88.7 84.9Set-BSR 67.466.4 84.6GPT-Neo-2.7B LLaMA-7B LLaMA-13BRandom22.525.530.4Cosine36.943.247.7BM2538.544.450.1BSR46.450.656.5Set-BSR45.951.658.2\",\n \"figure_id\": \"tab_14\",\n \"figure_label\": \"13\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Average 8-shot ICL performance across all datasets and splits.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_15\",\n \"figure_label\": \"14\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(Devlin et al., 2019)\", \"Explanation\": \"The cited work by Devlin et al. introduces the concept of large language models (LLMs), which serves as the basis for the research conducted in the citing paper on in-context learning (ICL). The study builds upon the capabilities of LLMs to generalize to novel tasks by conditioning on textual prompts.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides evidence of the training-free paradigm of few-shot inference in LLMs, which supports the research on in-context learning (ICL) conducted in the citing paper.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. further extends the research on in-context learning (ICL) by exploring the capabilities of LLMs in generalizing to novel tasks through conditioning on textual prompts.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. highlights the training-free paradigm of few-shot inference in LLMs, which supports the research on in-context learning (ICL) conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it focuses on the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. offers a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it highlights the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it focuses on the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. offers a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it highlights the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it focuses on the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. offers a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it highlights the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it focuses on the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. offers a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it highlights the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it focuses on the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. offers a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it highlights the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it focuses on the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. offers a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it highlights the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it focuses on the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. offers a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it highlights the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it focuses on the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. offers a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it highlights the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it focuses on the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. offers a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it highlights the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it focuses on the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. offers a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it highlights the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it focuses on the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. offers a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it highlights the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Brown et al., 2020)\", \"Explanation\": \"The cited work by Brown et al. provides a methodological basis for the research on in-context learning (ICL) conducted in the citing paper, as it focuses on the training-free paradigm of few-shot inference in LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Levy et al., 2022)\", \"Explanation\": \"The cited work by Levy et al. (2022) is used to highlight the inadequacy of the standard approach to selecting ICL examples or demonstrations, as it relies on a single dense embedding and unigram overlap for evaluation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2020)\", \"Explanation\": \"The cited work by Zhang et al. (2020) provides the BERTScore metric, which the citing paper uses to measure the coverage of contextual token embeddings in the context of example selection for LLM understanding and performance.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Rubin et al., 2022)\", \"Explanation\": \"The cited work by Rubin et al. (2022) is used to highlight the performance of trained methods in the context of semantic parsing and compositional settings, which the citing paper uses to support the claim that SET-BSR outperforms these methods in terms of performance.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Levy et al., 2022)\", \"Explanation\": \"The cited work by Levy et al. (2022) is used to further expand on the research conducted in the citing paper, exploring the performance of trained methods in the context of semantic parsing and compositional settings.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ye et al., 2023)\", \"Explanation\": \"The cited work by Ye et al. (2023) is used to continue the research on the performance of trained methods in the context of semantic parsing and compositional settings, building upon the findings of the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The cited work by Liu et al. provides a more interpretable and customizable interface to using LLMs, which the citing paper adopts in their research to improve the performance of in-context learning for few-shot inference.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wei et al., 2023)\", \"Explanation\": \"The cited work by Wei et al. discusses the advantages of in-context learning for few-shot inference, including improved generalization and reasoning skills, which the citing paper leverages in their research to improve the performance of in-context learning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhao et al., 2021)\", \"Explanation\": \"The cited work by Zhao et al. highlights the sensitivity of in-context learning to the choice of demonstrations, which the citing paper builds upon in their research to develop techniques for selecting good demonstrations to improve the performance of in-context learning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The cited work by Liu et al. discusses the importance of in-context learning for few-shot inference in the context of LLMs, which the citing paper extends in their research to further explore the performance of in-context learning in the context of LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Robertson et al., 1993)\", \"Explanation\": \"The cited work by Robertson et al. provides the Okapi variant of BM25, which the citing paper uses as a method for measuring relevance in the context of information retrieval algorithms.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2020)\", \"Explanation\": \"The cited work, BERTScore, is used as a metric to evaluate the quality of machine-generated text by comparing it to a reference text. The citing paper adopts the idea of leveraging pre-trained contextual embeddings to match words in the candidate and reference sentences by cosine similarity, which is a methodological basis for the study conducted in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zelle and Mooney, 1996)\", \"Explanation\": \"The cited work provides the Geo-Query dataset, which is used in the citing paper for semantic parsing research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Dahl et al., 1994)\", \"Explanation\": \"The cited work contributes the ATIS dataset, which is utilized in the study of semantic parsing.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wang et al., 2015)\", \"Explanation\": \"The cited work provides the Overnight dataset, which is used in the research on semantic parsing.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Andreas et al., 2020)\", \"Explanation\": \"The cited work contributes the SMCalFlow dataset, which is employed in the study of semantic parsing.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wolfson et al., 2020)\", \"Explanation\": \"The cited work provides the BREAK dataset, which is utilized in the research on semantic parsing.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Li et al., 2021)\", \"Explanation\": \"The cited work contributes the MTOP dataset, which is used in the study of semantic parsing.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Cobbe et al., 2021)\", \"Explanation\": \"The cited work provides the GSM8K dataset, which is utilized in the study of math-word problems.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Dua et al., 2019)\", \"Explanation\": \"The cited work contributes the DROP dataset, which is used in the research on machine reading comprehension.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wang et al., 2018)\", \"Explanation\": \"The cited work provides the QNLI, MNLI, and QQP datasets, which are utilized in the study of natural language inference, paraphrase detection, and sentiment classification, respectively.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Williams et al., 2018)\", \"Explanation\": \"The cited work contributes the MNLI dataset, which is used in the study of natural language inference.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Bentivogli et al., 2009)\", \"Explanation\": \"The cited work provides the RTE dataset, which is utilized in the research on natural language inference.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Dolan and Brockett, 2005)\", \"Explanation\": \"The cited work contributes the MRPC dataset, which is used in the study of natural language inference.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhang et al., 2019)\", \"Explanation\": \"The cited work provides the PAWS dataset, which is utilized in the study of natural language inference.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Socher et al., 2013)\", \"Explanation\": \"The cited work contributes the SST2 dataset, which is used in the study of natural language inference.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Finegan-Dollak et al., 2018)\", \"Explanation\": \"The cited work by Finegan-Dollak et al. (2018) provides the template for the compositional splits used in the evaluation of GeoQuery.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Keysers et al., 2020)\", \"Explanation\": \"The work by Keysers et al. (2020) contributes a TMCD compositional split for GeoQuery, which is used in the evaluation of the model performance.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Levy et al., 2022)\", \"Explanation\": \"The work by Levy et al. (2022) provides a length compositional split for GeoQuery, which is used in the evaluation of the model performance.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Qiu et al., 2022a)\", \"Explanation\": \"The work by Qiu et al. (2022a) generates the compositional splits for ATIS and Overnight, which are used in the evaluation of the model performance.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Yin et al., 2021)\", \"Explanation\": \"The work by Yin et al. (2021) provides the SMCalFlow compositional split, which is used in the evaluation of the model performance.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Gupta et al., 2022)\", \"Explanation\": \"The work by Gupta et al. (2022) generates the template compositional splits for ATIS and Overnight, which are used in the evaluation of the model performance.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ye et al., 2023)\", \"Explanation\": \"The cited work is the source of the training data used in the citing paper for the subsampling process.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Rubin et al., 2022)\", \"Explanation\": \"The cited work is the source of the training data used in the citing paper for the subsampling process.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Hasson and Berant, 2021)\", \"Explanation\": \"The cited work is the source of the evaluation metric used in the citing paper for the BREAK dataset.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Black et al., 2021)\", \"Explanation\": \"The cited work provides the The Pile text corpus, which the citing paper uses in training the GPT-Neo-2.7B LM.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Touvron et al., 2023)\", \"Explanation\": \"The cited work introduces the LLaMA LMs, which the citing paper experiments with, expanding upon the research of the original work.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Li et al., 2023)\", \"Explanation\": \"The cited work provides the StarCoder model, which the citing paper uses in its experiments with LLaMA-7B and LLaMA-13B.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Kocetkov et al., 2022)\", \"Explanation\": \"The cited work contributes the collection of programming languages used in training the StarCoder model, which the citing paper experiments with.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Chen et al., 2021)\", \"Explanation\": \"The cited work provides the code-pretrained LMs Cushman and Codex3, which the citing paper uses in its experiments with GPT-3.5-Turbo2.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Reimers and Gurevych, 2019)\", \"Explanation\": \"The cited work provides the Sentence-Bert library and the all-mpnet-base-v2 model for Cosine similarity measurement, which the citing paper uses in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Johnson et al., 2019)\", \"Explanation\": \"The cited work introduces the FAISS4 method for independent selection, which the citing paper adopts in their research to retrieve the most similar examples.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Robertson et al., 1993;Jones et al., 2000)\", \"Explanation\": \"The cited works provide the Okapi variant of BM25 for the BM25 metric, which the citing paper uses in their research to measure similarity.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2020)\", \"Explanation\": \"The cited work introduces the bert_score library with the deberta-large-mnli and deberta-base-mnli models for BERTScore measurement, which the citing paper uses in their research to evaluate similarity.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Rubin et al., 2022)\", \"Explanation\": \"The cited work provides a method for training a dense retriever using LLM perplexity, which the citing paper adopts in their research to improve the performance of their model.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ye et al., 2023)\", \"Explanation\": \"The cited work introduces a new method for training a Determinantal Point Process using LLM and EPR, which the citing paper extends by using the same method to further improve the model performance.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Levy et al., 2022)\", \"Explanation\": \"The cited work introduces a method for semantic parsing that predicts logical form substructures and selects diverse examples containing them, which the citing paper extends by using the same method to improve the model performance in SMCalFlow-CS and GeoQuery datasets.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Wei et al., 2022)\", \"Explanation\": \"The cited work by Wei et al. (2022) provides evidence that larger LLMs have the ability to perform Chain-of-Thought reasoning, which is necessary for success on GSM8K. This finding supports the claim in the citing paper that the failure of smaller LLMs on GSM8K is due to the need for this emergent ability.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zelle and Mooney, 1996)\", \"Explanation\": \"The cited work provides the dataset of natural language questions and Prolog programs used in the ICL prompt construction in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Finegan-Dollak et al., 2018)\", \"Explanation\": \"The cited work introduces the program template split in the GeoQuery dataset, which the citing paper adopts in their experiments to evaluate the performance of the ICL prompt.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Keysers et al., 2020)\", \"Explanation\": \"The cited work presents the TMCD split in the GeoQuery dataset, which the citing paper uses to create train and test sets with maximal compound divergence and minimal atom divergence in their experiments.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Levy et al., 2022)\", \"Explanation\": \"The cited work provides the compositional splits for the GeoQuery dataset, which the citing paper utilizes in their experiments to evaluate the performance of the ICL prompt.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Qiu et al., 2022a)\", \"Explanation\": \"The cited work generates the compositional splits for the GeoQuery dataset, which the citing paper uses in their experiments to evaluate the performance of the ICL prompt.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Finegan-Dollak et al., 2018)\", \"Explanation\": \"The cited work provides the Template split of the ATIS dataset for evaluating compositional generalization in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Gupta et al., 2022)\", \"Explanation\": \"The cited work provides the IID and Template splits of the ATIS and Overnight datasets for experiments in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Gupta et al., 2022)\", \"Explanation\": \"The cited work provides the IID and Template splits of the socialnetwork domain from the Overnight dataset for experiments in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Andreas et al., 2020)\", \"Explanation\": \"The cited work provides the SMCalFlow dataset of task-oriented natural language dialogs for experiments in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Rubin et al., 2022)\", \"Explanation\": \"The cited work provides the low-level Break subset of the SMCalFlow-CS dataset and the English subset of the MTOP dataset for the citing paper to use in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wang et al., 2018)\", \"Explanation\": \"The cited work provides the QQP dataset, which the citing paper utilizes in their study of sentiment classification.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Socher et al., 2013)\", \"Explanation\": \"The cited work provides the SST2 dataset, which the citing paper utilizes in their study of sentiment classification.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wang et al., 2018)\", \"Explanation\": \"The cited work provides the QQP dataset, which the citing paper utilizes in their study of sentiment classification.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Williams et al., 2018)\", \"Explanation\": \"The cited work provides the MNLI dataset, which the citing paper utilizes in their study of Natural Language Inference (NLI).\"}]"}}},{"rowIdx":95,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_1"],"heading":"Introduction","publication_ref":["b11","b40","b6","b27","b22","b26","b31","b30","b39","b14","b2","b14","b19","b34","b3","b34","b19","b7","b36","b29","b20"],"table_ref":[],"text":"As the strengths of large language models (LLMs) have become prominent (Brown et al., 2020;Chowdhery et al., 2022;Touvron et al., 2023), so too have their weaknesses (Bender et al., 2021). A glaring weakness of LLMs is their penchant for generating false, biased, or misleading claims in a phenomena broadly referred to as \"hallucinations\" (Maynez et al., 2020;Krishna et al., 2021;Longpre et al., 2021;Raunak et al., 2021). Most LLMs also do not ground their generations to any source, exacerbating this weakness (Rashkin et al., 2021).\nPost-hoc attribution and edit strategies offer promising solutions to tackle the problems of grounding and hallucination in language models (Thorne and Vlachos, 2020;Gao et al., 2022). These approaches retrieve supporting evidence to attribute the output (referred to as a claim) of a language model, followed by an editor that corrects factual errors in the claim, ensuring consistency with the evidence. A notable advantage of posthoc methods is their modularity: they can be easily applied to any text regardless of their generation source. However, existing editors exhibit distinct strengths and weaknesses. Sufficiently large language models can be few-shot prompted to perform editing (Bai et al., 2022;Gao et al., 2022). However, there is currently a steep compute-quality tradeoff, where only the largest, most expensive models can perform this task well. Even then, significant quality headroom remains, as we will show. In contrast, much smaller, cheaper models can be fine-tuned to perform editing, but are limited to specific domains where adequate training data is available (Iv et al., 2022;Schick et al., 2022).\nInstead of utilizing LLMs as prompted editors, we leverage their general-purpose capabilities to introduce challenging corruptions (i.e., noise) to clean pieces of text. Subsequently, we fine-tune compact editors to denoise these corruptions by grounding onto relevant evidence. While text to corrupt is readily available, we do not assume that paired relevant evidence is provided. To tackle this, our data generation pipeline first searches for a collection of topically related evidence. We then employ an LLM summarize the evidence into a claim which is then noised (Fig. 1a). The evidence is then used to ground the denoising. In contrast to existing work that assumes access to relevant paired evidence to ground the edit when training (Balachandran et al., 2022) or assumes edit data is provided for training (Schick et al., 2022;Iv et al., 2022), our approach eliminates these assumptions. (b) Using PURR. Given an ungrounded statement, we generate questions to search for relevant evidence which is then used to produce an edit. Furthermore, unlike distillation where a challenging distillation set is vital and the student model generally under-performs the teacher (Beyer et al., 2021;Stanton et al., 2021), our noising process introduces challenging corruptions and our resulting editor trained on these corruptions surpasses the performance of the LLM used for noising when the same LLM is employed for prompted editing on multiple datasets.\nOur Petite Unsupervised Research and Revision model, PURR, is built by fine-tuning a fusion-indecoder T5 model on denoising data from our data generation pipeline (Raffel et al., 2020;Izacard and Grave, 2021). Because our goal is to improve attribution broadly across tasks and domains, we evaluate PURR on outputs of large language models on multiple question answering and dialog datasets. On all benchmarks, PURR outperforms much more expensive LLM-prompted editors in improving attribution while being orders of magnitude faster."},{"figure_ref":[],"heading":"Editing for Attribution","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Task Overview","publication_ref":["b14","b14"],"table_ref":[],"text":"While there are various ways to apply editing to the outputs of large language models, the primary objective of PURR is to present efficient methods for attributing the outputs of language models and rectifying inaccuracies, referred to as Editing for Attribution (Gao et al., 2022). In this task, a system is provided with a textual statement, x, and is tasked to produce an attribution report. The attribution report consists of a collection of evidence snippets, A = {e 1 , e 2 , . . . , e n }, that grounds the information in x. Additionally, the system is asked to produced a revised statement (i.e., edit), y, that fixes any inaccuracies in x that contradict the content in A. For completeness, we present a summary of the task and refer interested readers to Gao et al. (2022) for a more comprehensive discussion."},{"figure_ref":[],"heading":"Evaluation Metrics","publication_ref":["b14","b14","b30","b14"],"table_ref":[],"text":"Following Gao et al. (2022), we evaluate editingfor-attribution systems along two dimensions: attribution, the extent to which the original and revised statements can be attributed to the attribution report, and preservation, which measures how much information has changed from x to y. The objective of the task is to maximally attribute a textual statement while preserving the original intent of the language model generation to the greatest extent possible. We use automated metrics developed by Gao et al. (2022) to measure both attribution and preservation, which were found to have strong correlation to human raters. It is important to note that this evaluation setup does not require reference edits and only relies on the grounding between the textual statements and the attribution report.\nAttribution A textual statement is generally said to be attributable to a set of evidence if one could reasonably say that given the evidence set, the statement is entailed (Rashkin et al., 2021). To formalize this, Gao et al. (2022) introduce an evaluation metric based on sentence-level natural langauge inference (NLI) model. Given an attribution report, A, and a textual statement y consisting of sentences, y = {s 1 , s 2 , . . .}, we use a NLI model to measure the likely that each sentence is entailed by an evidence snippet in A: NLI(e, s i ). The attribution of the entire statement, y, is computed as the average over the maximum attribution score for each constituent sentence.\nAttr (s,A) = max e∈A NLI(e, s)\n(1)\nAttr (y,A) = avg s∈y Attr (s,A)(2)\nThe goal of editing is to have Attr (y,A) be higher than Attr (x,A) .\nPreservation Preservation is measured using character-level Levenshtein distance between x and y. Preservation is 1 if the statements are the same and 0 if y has completely changed all textual information in x.\nPres (x,y) = max 1 - Lev(x, y) length(x) , 0(3)\nTo capture our goal of maximal attribution with maximal preservation, we unify these two metrics by computing the harmonic mean, F 1 AP , of Attr (y,A) and Pres (x,y) ."},{"figure_ref":[],"heading":"Evaluation Sets","publication_ref":["b14","b23","b16","b1","b14","b11","b38"],"table_ref":[],"text":"Our goal is to improve attribution broadly across tasks and domains on the outputs of strong generations systems. Gao et al. (2022) construct evaluation sets by prompting strong LLMs to generate outputs on three tasks: Natural Questions (factoid question answering) (Kwiatkowski et al., 2019), StrategyQA (reasoning-chain question answering) (Geva et al., 2021), and QreCC (knowledge-intensive dialogue) (Anantha et al., 2021). Gao et al. (2022) generate 150 validation and 150 test instances for each dataset using PALM for Natural Questions and StrategyQA and LaMBDA on QReCC (Chowdhery et al., 2022;Thoppilan et al., 2022). We use these sets and tune on the validation sets and report results on the test sets."},{"figure_ref":[],"heading":"Baselines","publication_ref":["b0","b14","b11","b14"],"table_ref":[],"text":"PURR and all baselines follow a research-andrevision pipeline. In the research stage, the objective is to search for relevant pieces of evidence to ground the information in the textual statement, x. This stage remains consistent across all baselines. We first prompt a large language model to generate a set of queries Q = {q 1 , q 2 , . . . q m } that attempts to cover all pieces of information in x that needs verification. Subsequently, we use Google Search in conjunction with a passage extractor to find the most relevant evidence snippet for each query, constituting an evidence set E = {e 1 , e 2 . . . , e m }.\nIn the revision stage, an editor is given the original statement, x, the set of queries, Q, and the evidence set, E, and asked to produce a revised statement, y. y can be the same as x in the event the editor deems the original statement cannot be edited further to increase attribution. We measure the ability of different editors to abstain from edit-ing later on. We compare PURR against two baseline editors.\nEFEC is a fine-tuned T5 editor trained on FEVER (Aly et al., 2021). EFEC was trained using evidence retrieved from Wikipedia and concatenates all pieces of evidence with the text statement to produce an edited statement. Notably, EFEC does not use the query set when making an edit. (Gao et al., 2022) found EFEC often improves attribution at the expense of preservation.\nRARR is a prompt-based editing approach that builds upon PALM, a language model with 540 billion parameters (Chowdhery et al., 2022). Unlike EFEC, which incorporates all evidence simultaneously to produce an edit, RARR iteratively examines each evidence, e i , by checking whether there is any contradictions between the text statement, x, and edits in the event there is. The process of contradiction checking and editing is performed using distinct few-shot prompts. Gao et al. (2022) demonstrate that this iterative approach to editing combined with few-shot prompting leads to improvements in attribution and preservation, albeit at the cost of multiple computationally expensive and slow calls to a large language model."},{"figure_ref":[],"heading":"Generating the Attribution Report","publication_ref":[],"table_ref":[],"text":"To maintain a manageable scope, we limit the attribution report, A, to include only the five most relevant pieces of evidence from the evidence set, E. An attribution report of five evidence snippets was found to be able to attribute the information for the claims in the datasets we evaluate on. It is worth noting that when editing, there are no restrictions on the number of evidence snippets an editor can utilize. Given the evidence set, E, and the query set, Q, from the research stage, we employ a scoring module that evaluates the relevance of each evidence e i to each query q j , S(q i , e j ). Our objective is to identify a subset of evidence that maximizes the coverage across all queries to form the attribution report. This coverage is quantified as the sum of the highest relevance scores achieved by each query with respect to any evidence. For scoring, we use a cross-encoder2 .\nCov (E,Q) = N i=1 max e j ∈E S(q i , e j )(4)\n3 Efficient Editing by Denoising\nIn this section, we present an overview of PURR, highlight its distinguishing features compared to baselines, and describe the denoising training strategy."},{"figure_ref":["fig_1"],"heading":"Overview of PURR at Inference Time","publication_ref":["b20"],"table_ref":[],"text":"We first describe how PURR is used at inference time and highlight the differences between PURR and baselines (Fig. 1b). Similar to EFEC, PURR is built upon on the T5 model, specifically T5-large. Furthermore, our editing framework adopts a similar approach to EFEC in terms of incorporating all available evidence simultaneously when making an edit. However, instead of concatenating the evidence in the input, we employ fusion-in-decoder (FiD) to effectively aggregate information across evidence (Izacard and Grave, 2021). This approach has demonstrated superior performance in merging information and allows us to surpass the context length limits imposed by modern language models. Finally, rather than employing a prompted language model for query generation during the research stage, we employ distillation to train a T5-large query generation model. Although our primary focus lies in enhancing the editing process, we opt for distillation during query generation as well to ensure that our editing pipeline does not rely on prompting."},{"figure_ref":["fig_1"],"heading":"Creating Training Data via Noising","publication_ref":[],"table_ref":[],"text":"To train an editor to fix hallucinations, we need a dataset consisting of a clean statements, y, which are paired with a set of supporting evidence E = {e 1 , e 2 , . . . , e n }, as well as a corrupted statement, x. While collecting this data manually is feasible, doing so can be expensive, requiring scouring for evidence to ground an LLM generation followed by removing any inaccuracies in the generation. Instead, we remove this bottleneck by leveraging the general purpose generation capabilities of LLMs to create a training set in a completely fashion. We generate clean statements by providing a set of topically related evidence to the LLM, and then corrupt the statements to create simulated hallucinations (Fig. 1a). We provide the prompts used for summarization and corruption in Appendix A."},{"figure_ref":[],"heading":"Generating Clean Statements With Evidence","publication_ref":[],"table_ref":[],"text":"The first step is to create a statement, y, paired with a set of evidence, E, that attributes (i.e., grounds) the statement. Our pipeline only requires a set q: Who will be the new coach of the Detroit lions? E + : -On Jan. of queries in the domain of interest to get started.\nWe start with a query, q, and use a search engine to find evidence related to the question. We take the top web pages from the search engine and chunk them into passages. Using the same crossencoder from the attribution report scoring module, we bin the passages that have the highest relevant score (beyond some threshold) to q into a set of gold evidence E + = {e + 1 , e + 2 , . . . , e + i } and the rest of the passages into a set of hard negative evidences E -= {e - 1 , e - 2 , . . . , e - j }. In our pipeline, we restrict the size of E + to contain at most four pieces of evidence. The resulting evidence set is the union of the gold and hard negative evidences E = E + ∪ E -. We then prompt a large language model to do zero-shot multi-document summarization of the gold evidence set, E + . We use the resulting summary as the clean statement, y, and upon manual inspection, the summary has a high degree of faithfulness to the evidence set."},{"figure_ref":[],"heading":"Noising and Conditional Denoising","publication_ref":[],"table_ref":[],"text":"We take the clean statement, y, and noise it by prompting a large language model to corrupt the text resulting in the corruption x. Our prompt contains examples of corruptions, and covers a wide range of linguistic phenomena we observe when it comes to LLM hallucinations. These include incorrect dates and entities, semantic role errors, and quantification errors. Once noised claims paired with evidence is available, an editor can be trained by fine-tuning a sequence-to-sequence model to maximize P (y|x, E). We call the resulting editor from denoising PURR."},{"figure_ref":[],"heading":"Dataset Statistics and Training Details","publication_ref":[],"table_ref":[],"text":"We utilized GPT-3.5 text-davinci-003 to facilitate the process of generating summaries and introducing corruption. Our choice of this particular model ensures that our generation strategy can be easily replicated. We started with roughly 6,000 seed queries covering a variety of domains and topics resulting in an edit dataset of 6,000 instances (Tab. 1). We reserve 10% for validation and use the resulting 90% for training. Each instance cost roughly 4 cents to generate and in total cost of roughly $250.\nWe fine-tune T5-large on our dataset using the validation loss to tune hyperparameters and determine training stoppage. During training, we pair each corrupted statement, x, with four pieces of evidence from the accompanying gold evidence set, E + , to ground the edit and produce the clean statement, y. In the event that the gold evidence set has fewer than four evidence snippets, we randomly sample evidence from the negative evidence set, E -, until we hit four snippets. We found adding negative evidence during training helps PURR ignore irrelevant evidence during inference."},{"figure_ref":[],"heading":"Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Primary Quantitative Results","publication_ref":[],"table_ref":["tab_2"],"text":"We provide results on the editing-for-attribution task in Table 2. We report the attribution of the claim before and after editing and the preservation of the edited claim. Our primary metric, F 1 AP , is the harmonic mean between the attribution and preservation of the edited claim. We first turn our attention to the baselines. EFEC, the editor that was fine-tuned with evidence largely from Wikipedia, struggles on this task. While EFEC improves attribution, this comes at the large expense of preser- vation and we see this in practice as EFEC tends to make large changes to the claim. RARR, the prompted editor, does not improve attribution as much as EFEC. However it is significantly better at preserving the intent of the original claim. Because of this, RARR is much better on the unified F 1 AP metric.\nPURR improves upon the results of RARR by generally making smaller changes to the claim while improving the attribution in this more limited edit. Because of this, PURR pushes the state-of-theart on the unified F 1 AP metric an all three tasks. Moreover, PURR is significantly more efficient to use by virtue of its size."},{"figure_ref":[],"heading":"Breaking Down the Numbers","publication_ref":[],"table_ref":[],"text":"We dig into the edits to get a better sense of where PURR improves on the baselines. Based on the preservation, Pres (x,y) , and attribution scores of the original statement, Attr (x,A) , and edited statement, Attr (y,A) , we say an edit can fall into one of the following sets:\n• Huge Edit: We say an edit is \"huge\" if preservation is low: Pres (x,y) < 0.5.\n• Bad Edit: We say an edit is \"bad\" if the attribution after editing is lower than before: Attr (y,A) -Attr (x,A) < -0.1. sary\" if it is a bad edit and also Attr (x,A) > 0.9. This means the editor made a poor edit when the attribution was already near perfect before editing.\n• Good Edit: We say an edit is \"good\" if attribution has significantly improved while preservation is high: Attr (y,A) -Attr (x,A) > 0.3 and Pres (x,y) > 0.7.\nNote that unnecessary edits are a subset of bad edits. We take the 150 instances in the Natural Questions test set and categorize the edits each editor makes in Figure 2. On a majority of claims, EFEC makes large edits while rarely making edits that improve attribution while preserving the original claim. RARR does a much better job at minimizing large edits but there are still cases where RARR edits a claim in a way that reduces the attribution. PURR almost never makes large edits and never edits a claim when it is near-perfect in a way that reduces attribution. PURR also makes more good edits compared to the baselines."},{"figure_ref":["fig_1"],"heading":"Qualitative Analysis","publication_ref":[],"table_ref":["tab_4","tab_6"],"text":"We then dig into the PURR predictions and diagnose the strengths of PURR and examine where there is room for improvement. We show examples in Table 3 that we found are representative of the strengths of PURR and areas of potential improvement. We find that PURR is extremely strong at fixing entity and numerical hallucinations as well as longer spans. Additionally, because PURR uses fusion-in-decoder, it is adept at merging informa-Correctly revising an entity/span. x: The US stopped printing $2 bills in 1966. The reason was that vending machines would not accept it. In 1976, the $2 bill was reintroduced. . . E: During that time, most goods and services were less than a dollar, making paper currency impractical. As need and use declined, the Federal Reserve stopped printing $2 bills in 1966. In 1976, the Federal Reserve brought back the $2-dollar bill. y: The US stopped printing $2 bills in 1966. The reason was that most goods and services were less than a dollar. In 1976, the $2 bill was reintroduced. . . tion across multiple pieces of evidence to make an edit. We noticed several instances where there are challenging distractors in evidence that can lead to an erroneous edit. Future work will introduce stronger corruptions in the data generation pipeline to better handle this case.\nWe next analyze the entire inference pipeline of PURR (Fig. 1b), which includes the question generation model, the search engine, and the editor itself. Our goal is to see when there is an error, which component is responsible. On the Natural Questions subset of the evaluation, we examine 20 instances where the attribution after editing, Attr (y,A) , is less than 0.30. Our qualitative analysis is provided in Table 4. Roughly 80% of the instances have low attribution after editing because either the question generation model we used did not fully cover the information in the claim or our search procedure did not find the best evidence for editing. We believe the question generation is the easier problem to fix while search is a much harder problem. Editing is a fairly small issue in comparison. Finally, there are some claims that fall into a \"miscellaenous\" category, either because it was not contextualized enough to properly edit or because the automatic metric erroneously assigned a low score."},{"figure_ref":[],"heading":"Inference Speed and Cost Comparisons of Fine-tuned vs Prompted Editors","publication_ref":[],"table_ref":[],"text":"A key advantage of PURR over prompt-based editors are the lower computational costs. RARR, a prompt-based editor built upon 540B PALM, runs on dozens of TPUs and takes approximately 40 seconds to edit a single statement. In comparison, PURR can run on a 12GB GPU and takes approximately 2 seconds to edit a single statement on a Titan-RTX. Considering generating our training set costs <$300 USD which is quickly amortized, we recommend our synthetic data generation strategy for large-scale deployment given the speed and cost savings of PURR."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b4","b9","b18","b14","b17","b19","b39","b34","b10","b15","b45","b33","b28","b44","b46","b37","b21","b2","b41","b5","b24","b35","b13","b25"],"table_ref":[],"text":"Editing for Attribution PURR builds upon previous research on post-hoc editing methods aimed at enhancing the attribution and accuracy of generated text (Balachandran et al., 2021;Cao et al., 2020;Iso et al., 2020). Notably, RARR (Gao et al., 2022) and Rethinking-with-Retrieval (He et al., 2022) employ few-shot prompting to rectify language model outputs, exhibiting similarities to our work. FRUIT (Iv et al., 2022) and EFEC (Thorne and Vlachos, 2020) also utilize fine-tuned editors to achieve similar objectives, leveraging Wikipedia as a source of training data. PEER is trained on Wikipedia edits (Schick et al., 2022) and includes a component for enhancing factuality through editing, but its primary focus lies in collaborative writing. Our denoising approach combines the speed advantages of fine-tuned editors while circumventing the reliance on training data that is typically constrained to specific domains like Wikipedia. x: James Marion Sims was the first to perform the first c section in 1846. This was in the US. E: -The first recorded cesarean where both mother and child survived was done in Switzerland, in 1500. That was also a husband-wife affair, although in this case Jacob Nufer was a swine gelder, not a doctor.\n-The first documented cesarean section on a living woman was performed in 1610; she died 25 days after the surgery. y: James Marion Sims was the first to perform the first c section in 1610. This was in the UK.\nAttr Claim not contextualized enough to properly edit.\nAttr (A,x) → Attr (A,y) : 0.00 → 0.01 Improving Trust in Large Language Models Ensuring the safe deployment of large language models encompasses various considerations, beyond just factuality and attribution. Large language models have demonstrated the potential to regurgitate protected information (Carlini et al., 2020), spew hateful content (Gehman et al., 2020), and exhibit high sensitivity to input variations (Zhao et al., 2021). A common approach to addressing these issues has been via additional training such as instruction fine-tuning (Sanh et al., 2021;Min et al., 2021;Chung et al., 2022;Ye et al., 2022), fine-tuning from human feedback (Ziegler et al., 2019;Stiennon et al., 2020), and more recently pretraining from human feedback (Korbak et al., 2023). In a similar vein to RARR, Bai et al. (2022) proposes to edit the outputs of LLMs using prompted LLMs to remove unsafe aspects of generated text.\nAs part of our future research, we aim to apply our denoising strategy to train efficient compact editors for addressing such undesired generation behaviors.\nDistilling Large Language Models Given their generation prowess, LLMs have been incorporated into data generation pipelines, essentially distilling the knowledge of the language model if their outputs are used for training (Wang et al., 2021;Bartolo et al., 2022;Lang et al., 2022;Smith et al., 2022). Eisenstein et al. (2022) follow a multi-step distillation pipeline like ours, chaining the outputs of multiple LLM calls and distilling the output into an explainable question answering model. Liu et al. (2022) uses the outputs of LLMs followed by filtering and human refinement to create WANLI, a challenging natural language inference dataset. On the evaluation side, Ribeiro and Lundberg (2022) use LLMs to generate evaluation sets for testing LLMs. While similar, our denoising approach implicitly distills the information in a large language model while simultaneously producing challenging training instances."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"Factuality and attribution are vital for the safe deployment of large language models. However, these mechanisms are inherently lacking in LLMs.\nRecent work has proposed augmenting the outputs of LLMs by retrieving evidence to attribute their outputs followed by prompting another LLM to edit the outputs to remove inconsistencies. However, there is a heavy computational cost which bottleneck these methods which motivates a need to develop efficient editors, but this is hindered by training data scarcity. To overcome these challenges, we use LLMs to corrupt text and fine-tune compact editors to denoise these faux hallucinations. Our denoising method is completely unsupervised and our resulting editor, PURR, improves attribution performance across various datasets over prompted editors, while being order of magnitude faster to execute."},{"figure_ref":[],"heading":"A Prompts for Creating Training Data","publication_ref":[],"table_ref":[],"text":""}],"string":"[\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b11\",\n \"b40\",\n \"b6\",\n \"b27\",\n \"b22\",\n \"b26\",\n \"b31\",\n \"b30\",\n \"b39\",\n \"b14\",\n \"b2\",\n \"b14\",\n \"b19\",\n \"b34\",\n \"b3\",\n \"b34\",\n \"b19\",\n \"b7\",\n \"b36\",\n \"b29\",\n \"b20\"\n ],\n \"table_ref\": [],\n \"text\": \"As the strengths of large language models (LLMs) have become prominent (Brown et al., 2020;Chowdhery et al., 2022;Touvron et al., 2023), so too have their weaknesses (Bender et al., 2021). A glaring weakness of LLMs is their penchant for generating false, biased, or misleading claims in a phenomena broadly referred to as \\\"hallucinations\\\" (Maynez et al., 2020;Krishna et al., 2021;Longpre et al., 2021;Raunak et al., 2021). Most LLMs also do not ground their generations to any source, exacerbating this weakness (Rashkin et al., 2021).\\nPost-hoc attribution and edit strategies offer promising solutions to tackle the problems of grounding and hallucination in language models (Thorne and Vlachos, 2020;Gao et al., 2022). These approaches retrieve supporting evidence to attribute the output (referred to as a claim) of a language model, followed by an editor that corrects factual errors in the claim, ensuring consistency with the evidence. A notable advantage of posthoc methods is their modularity: they can be easily applied to any text regardless of their generation source. However, existing editors exhibit distinct strengths and weaknesses. Sufficiently large language models can be few-shot prompted to perform editing (Bai et al., 2022;Gao et al., 2022). However, there is currently a steep compute-quality tradeoff, where only the largest, most expensive models can perform this task well. Even then, significant quality headroom remains, as we will show. In contrast, much smaller, cheaper models can be fine-tuned to perform editing, but are limited to specific domains where adequate training data is available (Iv et al., 2022;Schick et al., 2022).\\nInstead of utilizing LLMs as prompted editors, we leverage their general-purpose capabilities to introduce challenging corruptions (i.e., noise) to clean pieces of text. Subsequently, we fine-tune compact editors to denoise these corruptions by grounding onto relevant evidence. While text to corrupt is readily available, we do not assume that paired relevant evidence is provided. To tackle this, our data generation pipeline first searches for a collection of topically related evidence. We then employ an LLM summarize the evidence into a claim which is then noised (Fig. 1a). The evidence is then used to ground the denoising. In contrast to existing work that assumes access to relevant paired evidence to ground the edit when training (Balachandran et al., 2022) or assumes edit data is provided for training (Schick et al., 2022;Iv et al., 2022), our approach eliminates these assumptions. (b) Using PURR. Given an ungrounded statement, we generate questions to search for relevant evidence which is then used to produce an edit. Furthermore, unlike distillation where a challenging distillation set is vital and the student model generally under-performs the teacher (Beyer et al., 2021;Stanton et al., 2021), our noising process introduces challenging corruptions and our resulting editor trained on these corruptions surpasses the performance of the LLM used for noising when the same LLM is employed for prompted editing on multiple datasets.\\nOur Petite Unsupervised Research and Revision model, PURR, is built by fine-tuning a fusion-indecoder T5 model on denoising data from our data generation pipeline (Raffel et al., 2020;Izacard and Grave, 2021). Because our goal is to improve attribution broadly across tasks and domains, we evaluate PURR on outputs of large language models on multiple question answering and dialog datasets. On all benchmarks, PURR outperforms much more expensive LLM-prompted editors in improving attribution while being orders of magnitude faster.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Editing for Attribution\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Task Overview\",\n \"publication_ref\": [\n \"b14\",\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"While there are various ways to apply editing to the outputs of large language models, the primary objective of PURR is to present efficient methods for attributing the outputs of language models and rectifying inaccuracies, referred to as Editing for Attribution (Gao et al., 2022). In this task, a system is provided with a textual statement, x, and is tasked to produce an attribution report. The attribution report consists of a collection of evidence snippets, A = {e 1 , e 2 , . . . , e n }, that grounds the information in x. Additionally, the system is asked to produced a revised statement (i.e., edit), y, that fixes any inaccuracies in x that contradict the content in A. For completeness, we present a summary of the task and refer interested readers to Gao et al. (2022) for a more comprehensive discussion.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Evaluation Metrics\",\n \"publication_ref\": [\n \"b14\",\n \"b14\",\n \"b30\",\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"Following Gao et al. (2022), we evaluate editingfor-attribution systems along two dimensions: attribution, the extent to which the original and revised statements can be attributed to the attribution report, and preservation, which measures how much information has changed from x to y. The objective of the task is to maximally attribute a textual statement while preserving the original intent of the language model generation to the greatest extent possible. We use automated metrics developed by Gao et al. (2022) to measure both attribution and preservation, which were found to have strong correlation to human raters. It is important to note that this evaluation setup does not require reference edits and only relies on the grounding between the textual statements and the attribution report.\\nAttribution A textual statement is generally said to be attributable to a set of evidence if one could reasonably say that given the evidence set, the statement is entailed (Rashkin et al., 2021). To formalize this, Gao et al. (2022) introduce an evaluation metric based on sentence-level natural langauge inference (NLI) model. Given an attribution report, A, and a textual statement y consisting of sentences, y = {s 1 , s 2 , . . .}, we use a NLI model to measure the likely that each sentence is entailed by an evidence snippet in A: NLI(e, s i ). The attribution of the entire statement, y, is computed as the average over the maximum attribution score for each constituent sentence.\\nAttr (s,A) = max e∈A NLI(e, s)\\n(1)\\nAttr (y,A) = avg s∈y Attr (s,A)(2)\\nThe goal of editing is to have Attr (y,A) be higher than Attr (x,A) .\\nPreservation Preservation is measured using character-level Levenshtein distance between x and y. Preservation is 1 if the statements are the same and 0 if y has completely changed all textual information in x.\\nPres (x,y) = max 1 - Lev(x, y) length(x) , 0(3)\\nTo capture our goal of maximal attribution with maximal preservation, we unify these two metrics by computing the harmonic mean, F 1 AP , of Attr (y,A) and Pres (x,y) .\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Evaluation Sets\",\n \"publication_ref\": [\n \"b14\",\n \"b23\",\n \"b16\",\n \"b1\",\n \"b14\",\n \"b11\",\n \"b38\"\n ],\n \"table_ref\": [],\n \"text\": \"Our goal is to improve attribution broadly across tasks and domains on the outputs of strong generations systems. Gao et al. (2022) construct evaluation sets by prompting strong LLMs to generate outputs on three tasks: Natural Questions (factoid question answering) (Kwiatkowski et al., 2019), StrategyQA (reasoning-chain question answering) (Geva et al., 2021), and QreCC (knowledge-intensive dialogue) (Anantha et al., 2021). Gao et al. (2022) generate 150 validation and 150 test instances for each dataset using PALM for Natural Questions and StrategyQA and LaMBDA on QReCC (Chowdhery et al., 2022;Thoppilan et al., 2022). We use these sets and tune on the validation sets and report results on the test sets.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Baselines\",\n \"publication_ref\": [\n \"b0\",\n \"b14\",\n \"b11\",\n \"b14\"\n ],\n \"table_ref\": [],\n \"text\": \"PURR and all baselines follow a research-andrevision pipeline. In the research stage, the objective is to search for relevant pieces of evidence to ground the information in the textual statement, x. This stage remains consistent across all baselines. We first prompt a large language model to generate a set of queries Q = {q 1 , q 2 , . . . q m } that attempts to cover all pieces of information in x that needs verification. Subsequently, we use Google Search in conjunction with a passage extractor to find the most relevant evidence snippet for each query, constituting an evidence set E = {e 1 , e 2 . . . , e m }.\\nIn the revision stage, an editor is given the original statement, x, the set of queries, Q, and the evidence set, E, and asked to produce a revised statement, y. y can be the same as x in the event the editor deems the original statement cannot be edited further to increase attribution. We measure the ability of different editors to abstain from edit-ing later on. We compare PURR against two baseline editors.\\nEFEC is a fine-tuned T5 editor trained on FEVER (Aly et al., 2021). EFEC was trained using evidence retrieved from Wikipedia and concatenates all pieces of evidence with the text statement to produce an edited statement. Notably, EFEC does not use the query set when making an edit. (Gao et al., 2022) found EFEC often improves attribution at the expense of preservation.\\nRARR is a prompt-based editing approach that builds upon PALM, a language model with 540 billion parameters (Chowdhery et al., 2022). Unlike EFEC, which incorporates all evidence simultaneously to produce an edit, RARR iteratively examines each evidence, e i , by checking whether there is any contradictions between the text statement, x, and edits in the event there is. The process of contradiction checking and editing is performed using distinct few-shot prompts. Gao et al. (2022) demonstrate that this iterative approach to editing combined with few-shot prompting leads to improvements in attribution and preservation, albeit at the cost of multiple computationally expensive and slow calls to a large language model.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Generating the Attribution Report\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To maintain a manageable scope, we limit the attribution report, A, to include only the five most relevant pieces of evidence from the evidence set, E. An attribution report of five evidence snippets was found to be able to attribute the information for the claims in the datasets we evaluate on. It is worth noting that when editing, there are no restrictions on the number of evidence snippets an editor can utilize. Given the evidence set, E, and the query set, Q, from the research stage, we employ a scoring module that evaluates the relevance of each evidence e i to each query q j , S(q i , e j ). Our objective is to identify a subset of evidence that maximizes the coverage across all queries to form the attribution report. This coverage is quantified as the sum of the highest relevance scores achieved by each query with respect to any evidence. For scoring, we use a cross-encoder2 .\\nCov (E,Q) = N i=1 max e j ∈E S(q i , e j )(4)\\n3 Efficient Editing by Denoising\\nIn this section, we present an overview of PURR, highlight its distinguishing features compared to baselines, and describe the denoising training strategy.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Overview of PURR at Inference Time\",\n \"publication_ref\": [\n \"b20\"\n ],\n \"table_ref\": [],\n \"text\": \"We first describe how PURR is used at inference time and highlight the differences between PURR and baselines (Fig. 1b). Similar to EFEC, PURR is built upon on the T5 model, specifically T5-large. Furthermore, our editing framework adopts a similar approach to EFEC in terms of incorporating all available evidence simultaneously when making an edit. However, instead of concatenating the evidence in the input, we employ fusion-in-decoder (FiD) to effectively aggregate information across evidence (Izacard and Grave, 2021). This approach has demonstrated superior performance in merging information and allows us to surpass the context length limits imposed by modern language models. Finally, rather than employing a prompted language model for query generation during the research stage, we employ distillation to train a T5-large query generation model. Although our primary focus lies in enhancing the editing process, we opt for distillation during query generation as well to ensure that our editing pipeline does not rely on prompting.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Creating Training Data via Noising\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To train an editor to fix hallucinations, we need a dataset consisting of a clean statements, y, which are paired with a set of supporting evidence E = {e 1 , e 2 , . . . , e n }, as well as a corrupted statement, x. While collecting this data manually is feasible, doing so can be expensive, requiring scouring for evidence to ground an LLM generation followed by removing any inaccuracies in the generation. Instead, we remove this bottleneck by leveraging the general purpose generation capabilities of LLMs to create a training set in a completely fashion. We generate clean statements by providing a set of topically related evidence to the LLM, and then corrupt the statements to create simulated hallucinations (Fig. 1a). We provide the prompts used for summarization and corruption in Appendix A.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Generating Clean Statements With Evidence\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The first step is to create a statement, y, paired with a set of evidence, E, that attributes (i.e., grounds) the statement. Our pipeline only requires a set q: Who will be the new coach of the Detroit lions? E + : -On Jan. of queries in the domain of interest to get started.\\nWe start with a query, q, and use a search engine to find evidence related to the question. We take the top web pages from the search engine and chunk them into passages. Using the same crossencoder from the attribution report scoring module, we bin the passages that have the highest relevant score (beyond some threshold) to q into a set of gold evidence E + = {e + 1 , e + 2 , . . . , e + i } and the rest of the passages into a set of hard negative evidences E -= {e - 1 , e - 2 , . . . , e - j }. In our pipeline, we restrict the size of E + to contain at most four pieces of evidence. The resulting evidence set is the union of the gold and hard negative evidences E = E + ∪ E -. We then prompt a large language model to do zero-shot multi-document summarization of the gold evidence set, E + . We use the resulting summary as the clean statement, y, and upon manual inspection, the summary has a high degree of faithfulness to the evidence set.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Noising and Conditional Denoising\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We take the clean statement, y, and noise it by prompting a large language model to corrupt the text resulting in the corruption x. Our prompt contains examples of corruptions, and covers a wide range of linguistic phenomena we observe when it comes to LLM hallucinations. These include incorrect dates and entities, semantic role errors, and quantification errors. Once noised claims paired with evidence is available, an editor can be trained by fine-tuning a sequence-to-sequence model to maximize P (y|x, E). We call the resulting editor from denoising PURR.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Dataset Statistics and Training Details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We utilized GPT-3.5 text-davinci-003 to facilitate the process of generating summaries and introducing corruption. Our choice of this particular model ensures that our generation strategy can be easily replicated. We started with roughly 6,000 seed queries covering a variety of domains and topics resulting in an edit dataset of 6,000 instances (Tab. 1). We reserve 10% for validation and use the resulting 90% for training. Each instance cost roughly 4 cents to generate and in total cost of roughly $250.\\nWe fine-tune T5-large on our dataset using the validation loss to tune hyperparameters and determine training stoppage. During training, we pair each corrupted statement, x, with four pieces of evidence from the accompanying gold evidence set, E + , to ground the edit and produce the clean statement, y. In the event that the gold evidence set has fewer than four evidence snippets, we randomly sample evidence from the negative evidence set, E -, until we hit four snippets. We found adding negative evidence during training helps PURR ignore irrelevant evidence during inference.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Primary Quantitative Results\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_2\"\n ],\n \"text\": \"We provide results on the editing-for-attribution task in Table 2. We report the attribution of the claim before and after editing and the preservation of the edited claim. Our primary metric, F 1 AP , is the harmonic mean between the attribution and preservation of the edited claim. We first turn our attention to the baselines. EFEC, the editor that was fine-tuned with evidence largely from Wikipedia, struggles on this task. While EFEC improves attribution, this comes at the large expense of preser- vation and we see this in practice as EFEC tends to make large changes to the claim. RARR, the prompted editor, does not improve attribution as much as EFEC. However it is significantly better at preserving the intent of the original claim. Because of this, RARR is much better on the unified F 1 AP metric.\\nPURR improves upon the results of RARR by generally making smaller changes to the claim while improving the attribution in this more limited edit. Because of this, PURR pushes the state-of-theart on the unified F 1 AP metric an all three tasks. Moreover, PURR is significantly more efficient to use by virtue of its size.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Breaking Down the Numbers\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We dig into the edits to get a better sense of where PURR improves on the baselines. Based on the preservation, Pres (x,y) , and attribution scores of the original statement, Attr (x,A) , and edited statement, Attr (y,A) , we say an edit can fall into one of the following sets:\\n• Huge Edit: We say an edit is \\\"huge\\\" if preservation is low: Pres (x,y) < 0.5.\\n• Bad Edit: We say an edit is \\\"bad\\\" if the attribution after editing is lower than before: Attr (y,A) -Attr (x,A) < -0.1. sary\\\" if it is a bad edit and also Attr (x,A) > 0.9. This means the editor made a poor edit when the attribution was already near perfect before editing.\\n• Good Edit: We say an edit is \\\"good\\\" if attribution has significantly improved while preservation is high: Attr (y,A) -Attr (x,A) > 0.3 and Pres (x,y) > 0.7.\\nNote that unnecessary edits are a subset of bad edits. We take the 150 instances in the Natural Questions test set and categorize the edits each editor makes in Figure 2. On a majority of claims, EFEC makes large edits while rarely making edits that improve attribution while preserving the original claim. RARR does a much better job at minimizing large edits but there are still cases where RARR edits a claim in a way that reduces the attribution. PURR almost never makes large edits and never edits a claim when it is near-perfect in a way that reduces attribution. PURR also makes more good edits compared to the baselines.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Qualitative Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\",\n \"tab_6\"\n ],\n \"text\": \"We then dig into the PURR predictions and diagnose the strengths of PURR and examine where there is room for improvement. We show examples in Table 3 that we found are representative of the strengths of PURR and areas of potential improvement. We find that PURR is extremely strong at fixing entity and numerical hallucinations as well as longer spans. Additionally, because PURR uses fusion-in-decoder, it is adept at merging informa-Correctly revising an entity/span. x: The US stopped printing $2 bills in 1966. The reason was that vending machines would not accept it. In 1976, the $2 bill was reintroduced. . . E: During that time, most goods and services were less than a dollar, making paper currency impractical. As need and use declined, the Federal Reserve stopped printing $2 bills in 1966. In 1976, the Federal Reserve brought back the $2-dollar bill. y: The US stopped printing $2 bills in 1966. The reason was that most goods and services were less than a dollar. In 1976, the $2 bill was reintroduced. . . tion across multiple pieces of evidence to make an edit. We noticed several instances where there are challenging distractors in evidence that can lead to an erroneous edit. Future work will introduce stronger corruptions in the data generation pipeline to better handle this case.\\nWe next analyze the entire inference pipeline of PURR (Fig. 1b), which includes the question generation model, the search engine, and the editor itself. Our goal is to see when there is an error, which component is responsible. On the Natural Questions subset of the evaluation, we examine 20 instances where the attribution after editing, Attr (y,A) , is less than 0.30. Our qualitative analysis is provided in Table 4. Roughly 80% of the instances have low attribution after editing because either the question generation model we used did not fully cover the information in the claim or our search procedure did not find the best evidence for editing. We believe the question generation is the easier problem to fix while search is a much harder problem. Editing is a fairly small issue in comparison. Finally, there are some claims that fall into a \\\"miscellaenous\\\" category, either because it was not contextualized enough to properly edit or because the automatic metric erroneously assigned a low score.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Inference Speed and Cost Comparisons of Fine-tuned vs Prompted Editors\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"A key advantage of PURR over prompt-based editors are the lower computational costs. RARR, a prompt-based editor built upon 540B PALM, runs on dozens of TPUs and takes approximately 40 seconds to edit a single statement. In comparison, PURR can run on a 12GB GPU and takes approximately 2 seconds to edit a single statement on a Titan-RTX. Considering generating our training set costs <$300 USD which is quickly amortized, we recommend our synthetic data generation strategy for large-scale deployment given the speed and cost savings of PURR.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b4\",\n \"b9\",\n \"b18\",\n \"b14\",\n \"b17\",\n \"b19\",\n \"b39\",\n \"b34\",\n \"b10\",\n \"b15\",\n \"b45\",\n \"b33\",\n \"b28\",\n \"b44\",\n \"b46\",\n \"b37\",\n \"b21\",\n \"b2\",\n \"b41\",\n \"b5\",\n \"b24\",\n \"b35\",\n \"b13\",\n \"b25\"\n ],\n \"table_ref\": [],\n \"text\": \"Editing for Attribution PURR builds upon previous research on post-hoc editing methods aimed at enhancing the attribution and accuracy of generated text (Balachandran et al., 2021;Cao et al., 2020;Iso et al., 2020). Notably, RARR (Gao et al., 2022) and Rethinking-with-Retrieval (He et al., 2022) employ few-shot prompting to rectify language model outputs, exhibiting similarities to our work. FRUIT (Iv et al., 2022) and EFEC (Thorne and Vlachos, 2020) also utilize fine-tuned editors to achieve similar objectives, leveraging Wikipedia as a source of training data. PEER is trained on Wikipedia edits (Schick et al., 2022) and includes a component for enhancing factuality through editing, but its primary focus lies in collaborative writing. Our denoising approach combines the speed advantages of fine-tuned editors while circumventing the reliance on training data that is typically constrained to specific domains like Wikipedia. x: James Marion Sims was the first to perform the first c section in 1846. This was in the US. E: -The first recorded cesarean where both mother and child survived was done in Switzerland, in 1500. That was also a husband-wife affair, although in this case Jacob Nufer was a swine gelder, not a doctor.\\n-The first documented cesarean section on a living woman was performed in 1610; she died 25 days after the surgery. y: James Marion Sims was the first to perform the first c section in 1610. This was in the UK.\\nAttr Claim not contextualized enough to properly edit.\\nAttr (A,x) → Attr (A,y) : 0.00 → 0.01 Improving Trust in Large Language Models Ensuring the safe deployment of large language models encompasses various considerations, beyond just factuality and attribution. Large language models have demonstrated the potential to regurgitate protected information (Carlini et al., 2020), spew hateful content (Gehman et al., 2020), and exhibit high sensitivity to input variations (Zhao et al., 2021). A common approach to addressing these issues has been via additional training such as instruction fine-tuning (Sanh et al., 2021;Min et al., 2021;Chung et al., 2022;Ye et al., 2022), fine-tuning from human feedback (Ziegler et al., 2019;Stiennon et al., 2020), and more recently pretraining from human feedback (Korbak et al., 2023). In a similar vein to RARR, Bai et al. (2022) proposes to edit the outputs of LLMs using prompted LLMs to remove unsafe aspects of generated text.\\nAs part of our future research, we aim to apply our denoising strategy to train efficient compact editors for addressing such undesired generation behaviors.\\nDistilling Large Language Models Given their generation prowess, LLMs have been incorporated into data generation pipelines, essentially distilling the knowledge of the language model if their outputs are used for training (Wang et al., 2021;Bartolo et al., 2022;Lang et al., 2022;Smith et al., 2022). Eisenstein et al. (2022) follow a multi-step distillation pipeline like ours, chaining the outputs of multiple LLM calls and distilling the output into an explainable question answering model. Liu et al. (2022) uses the outputs of LLMs followed by filtering and human refinement to create WANLI, a challenging natural language inference dataset. On the evaluation side, Ribeiro and Lundberg (2022) use LLMs to generate evaluation sets for testing LLMs. While similar, our denoising approach implicitly distills the information in a large language model while simultaneously producing challenging training instances.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Factuality and attribution are vital for the safe deployment of large language models. However, these mechanisms are inherently lacking in LLMs.\\nRecent work has proposed augmenting the outputs of LLMs by retrieving evidence to attribute their outputs followed by prompting another LLM to edit the outputs to remove inconsistencies. However, there is a heavy computational cost which bottleneck these methods which motivates a need to develop efficient editors, but this is hindered by training data scarcity. To overcome these challenges, we use LLMs to corrupt text and fine-tune compact editors to denoise these faux hallucinations. Our denoising method is completely unsupervised and our resulting editor, PURR, improves attribution performance across various datasets over prompted editors, while being order of magnitude faster to execute.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Prompts for Creating Training Data\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-05-24"},"doi":{"kind":"string","value":"10.18653/v1/2021.naacl-main.44"},"references":{"kind":"list like","value":[{"authors":"Rami Aly; Christos Christodoulopoulos; Oana Cocarascu; Zhijiang ","journal":"","ref_id":"b0","title":"Proceedings of the Fourth Workshop on Fact Extraction and VERification (FEVER). Association for Computational Linguistics","year":"2021"},{"authors":"Raviteja Anantha; Svitlana Vakulenko; Zhucheng Tu; Shayne Longpre; Stephen Pulman; Srinivas Chappidi","journal":"Association for Computational Linguistics","ref_id":"b1","title":"Open-domain question answering goes conversational via question rewriting","year":"2021"},{"authors":"Yuntao Bai; Saurav Kadavath; Sandipan Kundu; Amanda Askell; Jackson Kernion; Andy Jones; Anna Chen; Anna Goldie; Azalia Mirhoseini; Cameron Mckinnon; Carol Chen; Catherine Olsson; Christopher Olah; Danny Hernandez; Dawn Drain; Deep Ganguli; Dustin Li; Eli Tran-Johnson; Ethan Perez; Jamie Kerr; Jared Mueller; Jeffrey Ladish; Joshua Landau; Kamile Kamal Ndousse; Liane Lukosuite; Michael Lovitt; Nelson Sellitto; Nicholas Elhage; Noemi Schiefer; Nova Mercado; Robert Dassarma; Robin Lasenby; Sam Larson; Scott Ringer; Shauna Johnston; Sheer El Kravec; Stanislav Showk; Tamera Fort; Timothy Lanham; Tom Telleen-Lawton; Tom Conerly; Tristan Henighan; Samuel R Hume; Zac Bowman; Ben Hatfield-Dodds; Dario Mann; Nicholas Amodei; Sam Joseph; Tom Mccandlish; Jared Brown; Kaplan","journal":"","ref_id":"b2","title":"Constitutional ai: Harmlessness from ai feedback","year":"2022"},{"authors":"Hannaneh Vidhisha Balachandran; William Hajishirzi; Yulia Cohen; Tsvetkov","journal":"","ref_id":"b3","title":"Correcting diverse factual errors in abstractive summarization via postediting and language model infilling","year":"2022"},{"authors":"Ashish Vidhisha Balachandran; Yulia Vaswani; Niki Tsvetkov; Parmar","journal":"Association for Computational Linguistics","ref_id":"b4","title":"Simple and efficient ways to improve REALM","year":"2021"},{"authors":"Max Bartolo; Tristan Thrush; Sebastian Riedel; Pontus Stenetorp; Robin Jia; Douwe Kiela","journal":"Association for Computational Linguistics","ref_id":"b5","title":"Models in the loop: Aiding crowdworkers with generative annotation assistants","year":"2022"},{"authors":"Emily M Bender; Timnit Gebru; Angelina Mcmillan-Major; Shmargaret Shmitchell","journal":"","ref_id":"b6","title":"On the dangers of stochastic parrots: Can language models be too big?","year":"2021"},{"authors":"Lucas Beyer; Xiaohua Zhai; Amélie Royer; Larisa Markeeva; Rohan Anil; Alexander Kolesnikov","journal":"","ref_id":"b7","title":"Knowledge distillation: A good teacher is patient and consistent","year":"2021"},{"authors":"B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; T J Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeff Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mccandlish; Ilya Radford; Dario Sutskever; Amodei","journal":"","ref_id":"b8","title":"Language models are few-shot learners","year":"2020"},{"authors":"Meng Cao; Yue Dong; Jiapeng Wu; Jackie Chi; Kit Cheung","journal":"Association for Computational Linguistics","ref_id":"b9","title":"Factual error correction for abstractive summarization models","year":"2020"},{"authors":"Nicholas Carlini; Florian Tramèr; Eric Wallace; Matthew Jagielski; Ariel Herbert-Voss; Katherine Lee; Adam Roberts; Tom B Brown; Dawn Xiaodong Song; Úlfar Erlingsson; Alina Oprea; Colin Raffel","journal":"","ref_id":"b10","title":"Extracting training data from large language models","year":"2020"},{"authors":"Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann; Parker Schuh; Kensen Shi; Sasha Tsvyashchenko; Joshua Maynez; Abhishek Rao; Parker Barnes; Yi Tay; Noam M Shazeer; Emily Vinodkumar Prabhakaran; Nan Reif; Benton C Du; Reiner Hutchinson; James Pope; Jacob Bradbury; Michael Austin; Guy Isard; Pengcheng Gur-Ari; Toju Yin; Anselm Duke; Sanjay Levskaya; Sunipa Ghemawat; Henryk Dev; Xavier Michalewski; Vedant García; Kevin Misra; Liam Robinson; Denny Fedus; Daphne Zhou; David Ippolito; Hyeontaek Luan; Barret Lim; Alexander Zoph; Ryan Spiridonov; David Sepassi; Shivani Dohan; Mark Agrawal; Andrew M Omernick; Thanumalayan Dai; Marie Sankaranarayana Pillai; Aitor Pellat; Erica Lewkowycz; Rewon Moreira; Oleksandr Child; Katherine Polozov; Zongwei Lee; Xuezhi Zhou; Brennan Wang; Mark Saeta; Orhan Díaz; Michele Firat; Jason Catasta; Kathleen S Wei; Douglas Meier-Hellstern; Jeff Eck; Slav Dean; Noah Petrov; Fiedel","journal":"","ref_id":"b11","title":"Palm: Scaling language modeling with pathways","year":"2022"},{"authors":"Chung Hyung Won; Le Hou; S Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma; Albert Webson; Shane Shixiang; Zhuyun Gu; Mirac Dai; Xinyun Suzgun; Aakanksha Chen; Dasha Chowdhery; Sharan Valter; Gaurav Narang; Adams Wei Mishra; Vincent Yu; Yanping Zhao; Andrew M Huang; Hongkun Dai; Slav Yu; Ed Petrov; Jeff Huai Hsin Chi; Jacob Dean; Adam Devlin; Denny Roberts; Quoc Zhou; Jason Le; Wei","journal":"","ref_id":"b12","title":"Scaling instructionfinetuned language models","year":"2022"},{"authors":"Jacob Eisenstein; Daniel Andor; Bernd Bohnet; Michael Collins; David Mimno","journal":"","ref_id":"b13","title":"Honest students from untrusted teachers: Learning an interpretable question-answering pipeline from a pretrained language model","year":"2022"},{"authors":"Luyu Gao; Zhuyun Dai; Panupong Pasupat; Anthony Chen; Arun Tejasvi Chaganty; Yicheng Fan; N Vincent Zhao; Hongrae Lao; Da-Cheng Lee; Kelvin Juan; Guu","journal":"","ref_id":"b14","title":"Rarr: Researching and revising what language models say, using language models","year":"2022"},{"authors":"Suchin Samuel Gehman; Maarten Gururangan; Yejin Sap; Noah A Choi; Smith","journal":"Association for Computational Linguistics","ref_id":"b15","title":"RealToxi-cityPrompts: Evaluating neural toxic degeneration in language models","year":"2020"},{"authors":"Mor Geva; Daniel Khashabi; Elad Segal; Tushar Khot; Dan Roth; Jonathan Berant","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b16","title":"Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies","year":"2021"},{"authors":"Hangfeng He; Hongming Zhang; Dan Roth","journal":"","ref_id":"b17","title":"Rethinking with Retrieval: Faithful Large Language Model Inference","year":"2022"},{"authors":"Hayate Iso; Chao Qiao; Hang Li","journal":"Association for Computational Linguistics","ref_id":"b18","title":"Fact-based Text Editing","year":"2020"},{"authors":"Robert Iv; Alexandre Passos; Sameer Singh; Ming-Wei Chang","journal":"Association for Computational Linguistics","ref_id":"b19","title":"FRUIT: Faithfully reflecting updated information in text","year":"2022"},{"authors":"Gautier Izacard; Edouard Grave","journal":"Association for Computational Linguistics","ref_id":"b20","title":"Leveraging passage retrieval with generative models for open domain question answering","year":"2021"},{"authors":"Tomasz Korbak; Kejian Shi; Angelica Chen; Rasika Bhalerao; Christopher L Buckley; Jason Phang; Sam Bowman; Ethan Perez","journal":"","ref_id":"b21","title":"Pretraining language models with human preferences","year":"2023"},{"authors":"Kalpesh Krishna; Aurko Roy; Mohit Iyyer","journal":"Association for Computational Linguistics","ref_id":"b22","title":"Hurdles to progress in long-form question answering","year":"2021"},{"authors":"Tom Kwiatkowski; Jennimaria Palomaki; Olivia Redfield; Michael Collins; Ankur Parikh; Chris Alberti; Danielle Epstein; Illia Polosukhin; Jacob Devlin; Kenton Lee; Kristina Toutanova; Llion Jones; Matthew Kelcey; Ming-Wei Chang; Andrew M Dai; Jakob Uszkoreit; Quoc Le; Slav Petrov","journal":"Transactions of the Association for Computational Linguistics","ref_id":"b23","title":"Natural questions: A benchmark for question answering research","year":"2019"},{"authors":"Hunter Lang; Monica Agrawal; Yoon Kim; David A Sontag","journal":"","ref_id":"b24","title":"Co-training improves prompt-based learning for large language models","year":"2022"},{"authors":"Alisa Liu; Swabha Swayamdipta; Noah A Smith; Yejin Choi","journal":"Association for Computational Linguistics","ref_id":"b25","title":"WANLI: Worker and AI collaboration for natural language inference dataset creation","year":"2022"},{"authors":"Shayne Longpre; Kartik Kumar Perisetla; Anthony Chen; Nikhil Ramesh; Chris Dubois; Sameer Singh","journal":"","ref_id":"b26","title":"Entity-based knowledge conflicts in question answering","year":"2021"},{"authors":"Joshua Maynez; Shashi Narayan; Bernd Bohnet; Ryan T Mcdonald","journal":"","ref_id":"b27","title":"On faithfulness and factuality in abstractive summarization","year":"2020"},{"authors":"Sewon Min; Mike Lewis; Luke Zettlemoyer; Hannaneh Hajishirzi","journal":"","ref_id":"b28","title":"Metaicl: Learning to learn in context","year":"2021"},{"authors":"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu","journal":"Journal of Machine Learning Research","ref_id":"b29","title":"Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer","year":"2020"},{"authors":"Vitaly Hannah Rashkin; Matthew Nikolaev; Michael Lamm; Dipanjan Collins; Slav Das; Gaurav Petrov; Iulia Singh Tomar; D Turc; Reitter","journal":"","ref_id":"b30","title":"Measuring attribution in natural language generation models","year":"2021"},{"authors":"Arul Vikas Raunak; Marcin Menezes; Junczys-Dowmunt","journal":"Association for Computational Linguistics","ref_id":"b31","title":"The curious case of hallucinations in neural machine translation","year":"2021"},{"authors":"Marco Tulio; Ribeiro ; Scott Lundberg","journal":"Association for Computational Linguistics","ref_id":"b32","title":"Adaptive testing and debugging of NLP models","year":"2022"},{"authors":"Victor Sanh; Albert Webson; Colin Raffel; Stephen H Bach; Lintang Sutawika; Zaid Alyafeai; Antoine Chaffin; Arnaud Stiegler; Teven Le Scao; Arun Raja; Manan Dey; M Saiful Bari; Canwen Xu; Urmish Thakker; Shanya Sharma; Eliza Szczechla; Taewoon Kim; Gunjan Chhablani; V Nihal; Debajyoti Nayak; Jonathan Datta; Mike Chang; Tian-Jian; Han Jiang; Matteo Wang; Sheng Manica; Zheng Xin Shen; Harshit Yong; Rachel Pandey; Thomas Bawden; Trishala Wang; Jos Neeraj; Abheesht Rozen; Andrea Sharma; Thibault Santilli; Jason Févry; Alan Fries; Ryan Teehan; Stella Rose Biderman; Leo Gao; Tali Bers; Thomas Wolf; Alexander M Rush","journal":"","ref_id":"b33","title":"Multitask prompted training enables zero-shot task generalization","year":"2021"},{"authors":"Timo Schick; Jane Dwivedi-Yu; Zhengbao Jiang; Fabio Petroni; Patrick Lewis; Gautier Izacard; Qingfei You; Christoforos Nalmpantis; Edouard Grave; Sebastian Riedel","journal":"","ref_id":"b34","title":"Peer: A collaborative language model","year":"2022"},{"authors":"Ryan Smith; Jason ; Alan Fries; Braden Hancock; Stephen H Bach","journal":"","ref_id":"b35","title":"Language models in the loop: Incorporating prompting into weak supervision","year":"2022"},{"authors":"Samuel Stanton; Pavel Izmailov; P Kirichenko; Alexander A Alemi; Andrew Gordon; Wilson ","journal":"","ref_id":"b36","title":"Does knowledge distillation really work?","year":"2021"},{"authors":"Nisan Stiennon; Long Ouyang; Jeff Wu; Daniel M Ziegler; Ryan J Lowe; Chelsea Voss; Alec Radford; Dario Amodei; Paul Christiano","journal":"","ref_id":"b37","title":"Learning to summarize from human feedback","year":"2020"},{"authors":"Romal Thoppilan; Daniel De Freitas; Jamie Hall; Noam M Shazeer; Apoorv Kulshreshtha; Heng-Tze; Alicia Cheng; Taylor Jin; Leslie Bos; Yu Baker; Yaguang Du; Hongrae Li; Huaixiu Lee; Amin Zheng; Marcelo Ghafouri; Yanping Menegali; Maxim Huang; Dmitry Krikun; James Lepikhin; Dehao Qin; Yuanzhong Chen; Zhifeng Xu; Adam Chen; Maarten Roberts; Yanqi Bosma; Chung-Ching Zhou; I A Chang; Willard James Krivokon; Marc Rusch; Kathleen S Pickett; Meredith Ringel Meier-Hellstern; Tulsee Morris; Renelito Delos Doshi; Toju Santos; Johnny Hartz Duke; Ben Søraker; Vinodkumar Zevenbergen; Mark Prabhakaran; Ben Díaz; Kristen Hutchinson; Alejandra Olson; Erin Molina; Josh Hoffman-John; Lora Lee; Ravindran Aroyo; Alena Rajakumar; Matthew Butryna; V O Lamm; Joseph Kuzmina; Aaron Fenton; Rachel Cohen; Ray Bernstein; Blaise Kurzweil; Claire Aguera-Arcas; Marian Cui; Ed Croak; Chi Huai Hsin; Quoc Le","journal":"","ref_id":"b38","title":"Lamda: Language models for dialog applications","year":"2022"},{"authors":"James Thorne; Andreas Vlachos","journal":"","ref_id":"b39","title":"Evidencebased factual error correction","year":"2020"},{"authors":"Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar; Armand Aur'elien Rodriguez; Edouard Joulin; Guillaume Grave; Lample","journal":"","ref_id":"b40","title":"Llama: Open and efficient foundation language models","year":"2023"},{"authors":"Shuohang Wang; Yang Liu; Yichong Xu; Chenguang Zhu; Michael Zeng","journal":"Association for Computational Linguistics","ref_id":"b41","title":"Want to reduce labeling cost? GPT-3 can help","year":"2021"},{"authors":"Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Fei Xia; Ed Chi; Denny Quoc V Le; Zhou","journal":"","ref_id":"b42","title":"Chain-of-Thought Prompting Elicits Reasoning in Large Language Models","year":"2022"},{"authors":"Curran Associates; Inc ","journal":"","ref_id":"b43","title":"","year":""},{"authors":"Seonghyeon Ye; Doyoung Kim; Joel Jang; Joongbo Shin; Minjoon Seo","journal":"","ref_id":"b44","title":"Guess the instruction! flipped learning makes language models stronger zero-shot learners","year":"2022"},{"authors":"Tony Zhao; Eric Wallace; Shi Feng; Dan Klein; Sameer Singh","journal":"","ref_id":"b45","title":"Calibrate before use: Improving few-shot performance of language models","year":"2021"},{"authors":"M Daniel; Nisan Ziegler; Jeff Stiennon; Tom B Wu; Alec Brown; Dario Radford; Paul Amodei; Geoffrey Christiano; Irving","journal":"","ref_id":"b46","title":"Fine-tuning language models from human preferences","year":"2019"}],"string":"[\n {\n \"authors\": \"Rami Aly; Christos Christodoulopoulos; Oana Cocarascu; Zhijiang \",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Proceedings of the Fourth Workshop on Fact Extraction and VERification (FEVER). Association for Computational Linguistics\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Raviteja Anantha; Svitlana Vakulenko; Zhucheng Tu; Shayne Longpre; Stephen Pulman; Srinivas Chappidi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b1\",\n \"title\": \"Open-domain question answering goes conversational via question rewriting\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Yuntao Bai; Saurav Kadavath; Sandipan Kundu; Amanda Askell; Jackson Kernion; Andy Jones; Anna Chen; Anna Goldie; Azalia Mirhoseini; Cameron Mckinnon; Carol Chen; Catherine Olsson; Christopher Olah; Danny Hernandez; Dawn Drain; Deep Ganguli; Dustin Li; Eli Tran-Johnson; Ethan Perez; Jamie Kerr; Jared Mueller; Jeffrey Ladish; Joshua Landau; Kamile Kamal Ndousse; Liane Lukosuite; Michael Lovitt; Nelson Sellitto; Nicholas Elhage; Noemi Schiefer; Nova Mercado; Robert Dassarma; Robin Lasenby; Sam Larson; Scott Ringer; Shauna Johnston; Sheer El Kravec; Stanislav Showk; Tamera Fort; Timothy Lanham; Tom Telleen-Lawton; Tom Conerly; Tristan Henighan; Samuel R Hume; Zac Bowman; Ben Hatfield-Dodds; Dario Mann; Nicholas Amodei; Sam Joseph; Tom Mccandlish; Jared Brown; Kaplan\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Constitutional ai: Harmlessness from ai feedback\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hannaneh Vidhisha Balachandran; William Hajishirzi; Yulia Cohen; Tsvetkov\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Correcting diverse factual errors in abstractive summarization via postediting and language model infilling\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ashish Vidhisha Balachandran; Yulia Vaswani; Niki Tsvetkov; Parmar\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b4\",\n \"title\": \"Simple and efficient ways to improve REALM\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Max Bartolo; Tristan Thrush; Sebastian Riedel; Pontus Stenetorp; Robin Jia; Douwe Kiela\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b5\",\n \"title\": \"Models in the loop: Aiding crowdworkers with generative annotation assistants\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Emily M Bender; Timnit Gebru; Angelina Mcmillan-Major; Shmargaret Shmitchell\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"On the dangers of stochastic parrots: Can language models be too big?\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Lucas Beyer; Xiaohua Zhai; Amélie Royer; Larisa Markeeva; Rohan Anil; Alexander Kolesnikov\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Knowledge distillation: A good teacher is patient and consistent\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"B Tom; Benjamin Brown; Nick Mann; Melanie Ryder; Jared Subbiah; Prafulla Kaplan; Arvind Dhariwal; Pranav Neelakantan; Girish Shyam; Amanda Sastry; Sandhini Askell; Ariel Agarwal; Gretchen Herbert-Voss; T J Krueger; Rewon Henighan; Aditya Child; Daniel M Ramesh; Jeff Ziegler; Clemens Wu; Christopher Winter; Mark Hesse; Eric Chen; Mateusz Sigler; Scott Litwin; Benjamin Gray; Jack Chess; Christopher Clark; Sam Berner; Alec Mccandlish; Ilya Radford; Dario Sutskever; Amodei\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Meng Cao; Yue Dong; Jiapeng Wu; Jackie Chi; Kit Cheung\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b9\",\n \"title\": \"Factual error correction for abstractive summarization models\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Nicholas Carlini; Florian Tramèr; Eric Wallace; Matthew Jagielski; Ariel Herbert-Voss; Katherine Lee; Adam Roberts; Tom B Brown; Dawn Xiaodong Song; Úlfar Erlingsson; Alina Oprea; Colin Raffel\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Extracting training data from large language models\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Aakanksha Chowdhery; Sharan Narang; Jacob Devlin; Maarten Bosma; Gaurav Mishra; Adam Roberts; Paul Barham; Hyung Won Chung; Charles Sutton; Sebastian Gehrmann; Parker Schuh; Kensen Shi; Sasha Tsvyashchenko; Joshua Maynez; Abhishek Rao; Parker Barnes; Yi Tay; Noam M Shazeer; Emily Vinodkumar Prabhakaran; Nan Reif; Benton C Du; Reiner Hutchinson; James Pope; Jacob Bradbury; Michael Austin; Guy Isard; Pengcheng Gur-Ari; Toju Yin; Anselm Duke; Sanjay Levskaya; Sunipa Ghemawat; Henryk Dev; Xavier Michalewski; Vedant García; Kevin Misra; Liam Robinson; Denny Fedus; Daphne Zhou; David Ippolito; Hyeontaek Luan; Barret Lim; Alexander Zoph; Ryan Spiridonov; David Sepassi; Shivani Dohan; Mark Agrawal; Andrew M Omernick; Thanumalayan Dai; Marie Sankaranarayana Pillai; Aitor Pellat; Erica Lewkowycz; Rewon Moreira; Oleksandr Child; Katherine Polozov; Zongwei Lee; Xuezhi Zhou; Brennan Wang; Mark Saeta; Orhan Díaz; Michele Firat; Jason Catasta; Kathleen S Wei; Douglas Meier-Hellstern; Jeff Eck; Slav Dean; Noah Petrov; Fiedel\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"Palm: Scaling language modeling with pathways\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Chung Hyung Won; Le Hou; S Longpre; Barret Zoph; Yi Tay; William Fedus; Eric Li; Xuezhi Wang; Mostafa Dehghani; Siddhartha Brahma; Albert Webson; Shane Shixiang; Zhuyun Gu; Mirac Dai; Xinyun Suzgun; Aakanksha Chen; Dasha Chowdhery; Sharan Valter; Gaurav Narang; Adams Wei Mishra; Vincent Yu; Yanping Zhao; Andrew M Huang; Hongkun Dai; Slav Yu; Ed Petrov; Jeff Huai Hsin Chi; Jacob Dean; Adam Devlin; Denny Roberts; Quoc Zhou; Jason Le; Wei\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Scaling instructionfinetuned language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jacob Eisenstein; Daniel Andor; Bernd Bohnet; Michael Collins; David Mimno\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Honest students from untrusted teachers: Learning an interpretable question-answering pipeline from a pretrained language model\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Luyu Gao; Zhuyun Dai; Panupong Pasupat; Anthony Chen; Arun Tejasvi Chaganty; Yicheng Fan; N Vincent Zhao; Hongrae Lao; Da-Cheng Lee; Kelvin Juan; Guu\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Rarr: Researching and revising what language models say, using language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Suchin Samuel Gehman; Maarten Gururangan; Yejin Sap; Noah A Choi; Smith\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b15\",\n \"title\": \"RealToxi-cityPrompts: Evaluating neural toxic degeneration in language models\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Mor Geva; Daniel Khashabi; Elad Segal; Tushar Khot; Dan Roth; Jonathan Berant\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b16\",\n \"title\": \"Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Hangfeng He; Hongming Zhang; Dan Roth\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Rethinking with Retrieval: Faithful Large Language Model Inference\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hayate Iso; Chao Qiao; Hang Li\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b18\",\n \"title\": \"Fact-based Text Editing\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Robert Iv; Alexandre Passos; Sameer Singh; Ming-Wei Chang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b19\",\n \"title\": \"FRUIT: Faithfully reflecting updated information in text\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Gautier Izacard; Edouard Grave\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b20\",\n \"title\": \"Leveraging passage retrieval with generative models for open domain question answering\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Tomasz Korbak; Kejian Shi; Angelica Chen; Rasika Bhalerao; Christopher L Buckley; Jason Phang; Sam Bowman; Ethan Perez\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"Pretraining language models with human preferences\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Kalpesh Krishna; Aurko Roy; Mohit Iyyer\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b22\",\n \"title\": \"Hurdles to progress in long-form question answering\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Tom Kwiatkowski; Jennimaria Palomaki; Olivia Redfield; Michael Collins; Ankur Parikh; Chris Alberti; Danielle Epstein; Illia Polosukhin; Jacob Devlin; Kenton Lee; Kristina Toutanova; Llion Jones; Matthew Kelcey; Ming-Wei Chang; Andrew M Dai; Jakob Uszkoreit; Quoc Le; Slav Petrov\",\n \"journal\": \"Transactions of the Association for Computational Linguistics\",\n \"ref_id\": \"b23\",\n \"title\": \"Natural questions: A benchmark for question answering research\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Hunter Lang; Monica Agrawal; Yoon Kim; David A Sontag\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Co-training improves prompt-based learning for large language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Alisa Liu; Swabha Swayamdipta; Noah A Smith; Yejin Choi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b25\",\n \"title\": \"WANLI: Worker and AI collaboration for natural language inference dataset creation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Shayne Longpre; Kartik Kumar Perisetla; Anthony Chen; Nikhil Ramesh; Chris Dubois; Sameer Singh\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Entity-based knowledge conflicts in question answering\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Joshua Maynez; Shashi Narayan; Bernd Bohnet; Ryan T Mcdonald\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"On faithfulness and factuality in abstractive summarization\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Sewon Min; Mike Lewis; Luke Zettlemoyer; Hannaneh Hajishirzi\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Metaicl: Learning to learn in context\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Colin Raffel; Noam Shazeer; Adam Roberts; Katherine Lee; Sharan Narang; Michael Matena; Yanqi Zhou; Wei Li; Peter J Liu\",\n \"journal\": \"Journal of Machine Learning Research\",\n \"ref_id\": \"b29\",\n \"title\": \"Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Vitaly Hannah Rashkin; Matthew Nikolaev; Michael Lamm; Dipanjan Collins; Slav Das; Gaurav Petrov; Iulia Singh Tomar; D Turc; Reitter\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Measuring attribution in natural language generation models\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Arul Vikas Raunak; Marcin Menezes; Junczys-Dowmunt\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b31\",\n \"title\": \"The curious case of hallucinations in neural machine translation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Marco Tulio; Ribeiro ; Scott Lundberg\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b32\",\n \"title\": \"Adaptive testing and debugging of NLP models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Victor Sanh; Albert Webson; Colin Raffel; Stephen H Bach; Lintang Sutawika; Zaid Alyafeai; Antoine Chaffin; Arnaud Stiegler; Teven Le Scao; Arun Raja; Manan Dey; M Saiful Bari; Canwen Xu; Urmish Thakker; Shanya Sharma; Eliza Szczechla; Taewoon Kim; Gunjan Chhablani; V Nihal; Debajyoti Nayak; Jonathan Datta; Mike Chang; Tian-Jian; Han Jiang; Matteo Wang; Sheng Manica; Zheng Xin Shen; Harshit Yong; Rachel Pandey; Thomas Bawden; Trishala Wang; Jos Neeraj; Abheesht Rozen; Andrea Sharma; Thibault Santilli; Jason Févry; Alan Fries; Ryan Teehan; Stella Rose Biderman; Leo Gao; Tali Bers; Thomas Wolf; Alexander M Rush\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Multitask prompted training enables zero-shot task generalization\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Timo Schick; Jane Dwivedi-Yu; Zhengbao Jiang; Fabio Petroni; Patrick Lewis; Gautier Izacard; Qingfei You; Christoforos Nalmpantis; Edouard Grave; Sebastian Riedel\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Peer: A collaborative language model\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ryan Smith; Jason ; Alan Fries; Braden Hancock; Stephen H Bach\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Language models in the loop: Incorporating prompting into weak supervision\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Samuel Stanton; Pavel Izmailov; P Kirichenko; Alexander A Alemi; Andrew Gordon; Wilson \",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Does knowledge distillation really work?\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Nisan Stiennon; Long Ouyang; Jeff Wu; Daniel M Ziegler; Ryan J Lowe; Chelsea Voss; Alec Radford; Dario Amodei; Paul Christiano\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Learning to summarize from human feedback\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Romal Thoppilan; Daniel De Freitas; Jamie Hall; Noam M Shazeer; Apoorv Kulshreshtha; Heng-Tze; Alicia Cheng; Taylor Jin; Leslie Bos; Yu Baker; Yaguang Du; Hongrae Li; Huaixiu Lee; Amin Zheng; Marcelo Ghafouri; Yanping Menegali; Maxim Huang; Dmitry Krikun; James Lepikhin; Dehao Qin; Yuanzhong Chen; Zhifeng Xu; Adam Chen; Maarten Roberts; Yanqi Bosma; Chung-Ching Zhou; I A Chang; Willard James Krivokon; Marc Rusch; Kathleen S Pickett; Meredith Ringel Meier-Hellstern; Tulsee Morris; Renelito Delos Doshi; Toju Santos; Johnny Hartz Duke; Ben Søraker; Vinodkumar Zevenbergen; Mark Prabhakaran; Ben Díaz; Kristen Hutchinson; Alejandra Olson; Erin Molina; Josh Hoffman-John; Lora Lee; Ravindran Aroyo; Alena Rajakumar; Matthew Butryna; V O Lamm; Joseph Kuzmina; Aaron Fenton; Rachel Cohen; Ray Bernstein; Blaise Kurzweil; Claire Aguera-Arcas; Marian Cui; Ed Croak; Chi Huai Hsin; Quoc Le\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Lamda: Language models for dialog applications\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"James Thorne; Andreas Vlachos\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Evidencebased factual error correction\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Hugo Touvron; Thibaut Lavril; Gautier Izacard; Xavier Martinet; Marie-Anne Lachaux; Timothée Lacroix; Baptiste Rozière; Naman Goyal; Eric Hambro; Faisal Azhar; Armand Aur'elien Rodriguez; Edouard Joulin; Guillaume Grave; Lample\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Llama: Open and efficient foundation language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Shuohang Wang; Yang Liu; Yichong Xu; Chenguang Zhu; Michael Zeng\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b41\",\n \"title\": \"Want to reduce labeling cost? GPT-3 can help\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jason Wei; Xuezhi Wang; Dale Schuurmans; Maarten Bosma; Fei Xia; Ed Chi; Denny Quoc V Le; Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Chain-of-Thought Prompting Elicits Reasoning in Large Language Models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Curran Associates; Inc \",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Seonghyeon Ye; Doyoung Kim; Joel Jang; Joongbo Shin; Minjoon Seo\",\n \"journal\": \"\",\n \"ref_id\": \"b44\",\n \"title\": \"Guess the instruction! flipped learning makes language models stronger zero-shot learners\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Tony Zhao; Eric Wallace; Shi Feng; Dan Klein; Sameer Singh\",\n \"journal\": \"\",\n \"ref_id\": \"b45\",\n \"title\": \"Calibrate before use: Improving few-shot performance of language models\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"M Daniel; Nisan Ziegler; Jeff Stiennon; Tom B Wu; Alec Brown; Dario Radford; Paul Amodei; Geoffrey Christiano; Irving\",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"Fine-tuning language models from human preferences\",\n \"year\": \"2019\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[2,354.02,731.04,171.12,11.62],"formula_id":"formula_0","formula_text":"Attr (y,A) = avg s∈y Attr (s,A)(2)"},{"formula_coordinates":[3,98.17,149.07,191.7,24.67],"formula_id":"formula_1","formula_text":"Pres (x,y) = max 1 - Lev(x, y) length(x) , 0(3)"},{"formula_coordinates":[3,348.59,707.57,176.56,33.71],"formula_id":"formula_2","formula_text":"Cov (E,Q) = N i=1 max e j ∈E S(q i , e j )(4)"}],"string":"[\n {\n \"formula_coordinates\": [\n 2,\n 354.02,\n 731.04,\n 171.12,\n 11.62\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"Attr (y,A) = avg s∈y Attr (s,A)(2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 98.17,\n 149.07,\n 191.7,\n 24.67\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"Pres (x,y) = max 1 - Lev(x, y) length(x) , 0(3)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 348.59,\n 707.57,\n 176.56,\n 33.71\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"Cov (E,Q) = N i=1 max e j ∈E S(q i , e j )(4)\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"PURR: Efficiently Editing Language Model Hallucinations by Denoising Language Model Corruptions"},"abstract":{"kind":"string","value":"The remarkable capabilities of large language models have been accompanied by a persistent drawback: the generation of false and unsubstantiated claims commonly known as \"hallucinations\". To combat this issue, recent research has introduced approaches that involve editing and attributing the outputs of language models, particularly through prompt-based editing. However, the inference cost and speed of using large language models for editing currently bottleneck prompt-based methods. These bottlenecks motivate the training of compact editors, which is challenging due to the scarcity of training data for this purpose. To overcome these challenges, we exploit the power of large language models to introduce corruptions (i.e., noise) into text and subsequently fine-tune compact editors to denoise the corruptions by incorporating relevant evidence. Our methodology is entirely unsupervised and provides us with faux hallucinations for training in any domain. Our Petite Unsupervised Research and Revision model, PURR, not only improves attribution over existing editing methods based on fine-tuning and prompting, but also achieves faster execution times by orders of magnitude. 1"},"authors":{"kind":"string","value":"Anthony Chen; Panupong Pasupat; Sameer Singh; Hongrae Lee; Kelvin Guu; James W Marshall; John Sutter's Carpenter; John Sutter; James W Marshall'"},"figures":{"kind":"list like","value":[{"figure_caption":"Training PURR. Given a seed query, we search for relevant evidence and summarize them into a claim which we corrupt. PURR is trained to denoise the corruption conditioned on the evidence.","figure_data":"","figure_id":"fig_0","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 1 :1Figure 1: Training and Using PURR.","figure_data":"","figure_id":"fig_1","figure_label":"1","figure_type":"figure"},{"figure_caption":"Pres (x,y) : 0.81 Attr (A,x) → Attr (A,y) : 0.61 → 0.84 Reasoning across multiple pieces of evidence x: In 1899, the first driver's license was issued in France. The first license in the United States was issued in 1902 in Massachusetts. E: -France introduced the first driving licences under the Paris Police Ordinance of 14 August 1893. . . -Massachusetts and Missouri were the first states to require a license in 1903, but there was no test associated with it. . . y: In 1893, the first driver's license was issued in France. The first license in the United States was issued in 1903 in Massachusetts. P res (x,y) : 0.90 Attr (A,x) → Attr (A,y) : 0.03 → 0.96 Challenging distractors in evidence x: The first orca put in captivity was in 1965. It was at Seattle, Washington's Seattle Aquarium. The orca's name was Wanda. E: -On Nov. 18, 1961 a killer whale had been spotted and a collection crew from the Marineland of the Pacific scrambled to capture her. Never had an orca been captured alive, and Wanda was trying to avoid being the first. . . -The first orca to be captured in 1961 was found off the coast of California and given the name Wanda. y: The first orca put in captivity was in 1961. It was at Marineland of the Pacific. The orca's name was Wanda. P res (x,y) : 0.77 Attr (A,x) → Attr (A,y) : 0.33 → 0.77","figure_data":"","figure_id":"fig_2","figure_label":"","figure_type":"figure"},{"figure_caption":"(A,x) → Attr (A,y) : 0.00 → 0.20 Miscellaneous (15%) x: The song was written by Tony Macaulay and John Macleod and recorded by Helen Shapiro in 1961. y: The song was written by Dub Allbritten and Ronnie Self and recorded by The Jackson 5 in 1960.","figure_data":"","figure_id":"fig_3","figure_label":"","figure_type":"figure"},{"figure_caption":"20, 2021 the Detroit Lions named Dan Campbell the franchise's new head coach. . . -Campbell possesses 23 years of NFL experience, including 12 years as a coach and 11 as a player. In his first year. . . -On Jan. 20, 2021 the Detroit Lions named Dan Campbell the franchise's new head coach. . . x/y: Dan Campbell was appointed the new head assistant coach of the Detroit Lions on January 20, 2021. With 23 19 years of NFL experience, 12 as a coach and 11 7 as a player. . . What is the neurological explanation for why people laugh when they're nervous or frightened? E + : -A 2015 Yale study found people respond with a variety of emotions to strong outside stimuli. . . -Vilayanur Ramachandran states \"We have nervous laughter because we want to make ourselves think what horrible thing we encountered isn't really as horrible as it appears\". . . -Stanley Milgram conducted one of the earliest studies about nervous laughter in the 1960s. His study revealed that people often laughed nervously in uncomfortable situations. . . x/y: Yale researchers in 2015 found people often respond to strong external stimuli with a variety of emotions, including nervous laughter anger. Stanley Milgram's Vilayanur Ramachandran's 1960s study also observed this in uncomfortable situations. Neuroscientist Vilayanur Ramachandran Stanley Milgram theorizes that people laugh when. . . .","figure_data":"","figure_id":"tab_1","figure_label":"","figure_type":"table"},{"figure_caption":"Results on the Editing for Attribution task. We report the attribution of the statement before and after editing, preservation after editing, and F 1 AP which combines attribution and preservation. Results are on LLM outputs on factoid question answering, long reasoning question answering, and dialog.","figure_data":"Model Attr. (x → y) Pres. F 1 APPALM outputs on NQEFEC44.7 → 63.9 39.648.5RARR 44.7 → 53.8 89.667.2PURR 44.8 → 59.8 91.072.2PALM outputs on SQAEFEC37.2 → 58.2 31.040.4RARR 37.2 → 44.6 89.959.6PURR 36.9 → 47.1 92.062.3LaMBDA outputs on QreCCEFEC18.4 → 47.2 39.042.7RARR 18.4 → 28.7 80.142.2PURR 16.8 → 33.0 85.847.7","figure_id":"tab_2","figure_label":"2","figure_type":"table"},{"figure_caption":"Example of good and bad revisions with PURR. x = claim; E = relevant evidence; y = edited claim using E. PURR can handle hallucinated entities and spans as well as merge information across evidence to edit. PURR can struggle when there are challenging distractors in a piece of evidence.","figure_data":"","figure_id":"tab_4","figure_label":"3","figure_type":"table"},{"figure_caption":"Query Generation Missing Coverage (35%) x: Legends of Tomorrow season 3 finale aired on April 9, 2018. It's title is No Country for Old Dads and is a 42-minute episode. Q: -When did the season 3 finale of Legends of Tomorrow air? -What's the title of Legends of Tomorrow season 3 finale? -How long is the season 3 finale of Legends of Tomorrow? y: Legends of Tomorrow season 3 finale aired on October 10, 2017. It's title is Freakshow and is a 43-minute episode. Bryce Dallas Howard plays the character of the mother of Cindy Lou Who in the 2018 animated movie The Grinch. E: Donna Lou Who is one of the characters of the 2018 Illumination animated film, Dr. Seuss' The Grinch. She's the mother of Cindy Lou Who and twin infants sons, Buster and Bean. She's voiced by Rashida Jones . y: No edit.","figure_data":"Attr (A,x) → Attr (A,y) : 0.01 → 0.01Search Not Finding Best Evidence (45%)x: Attr (A,x) → Attr (A,y) : 0.05 → 0.05Editor Error/Confusion (5%)","figure_id":"tab_5","figure_label":"","figure_type":"table"},{"figure_caption":"Error Analysis of PURR Inference Pipeline. We sample 20 edits from the NQ set where attribution is low after editing and categorize why by component. x = claim; Q = generated queries used for search, E = relevant evidence; y = edited claim using E. Strikethrough text represents a query that wasn't generated or evidence that wasn't retrieved but should have been.","figure_data":"","figure_id":"tab_6","figure_label":"4","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Training PURR. Given a seed query, we search for relevant evidence and summarize them into a claim which we corrupt. PURR is trained to denoise the corruption conditioned on the evidence.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: Training and Using PURR.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Pres (x,y) : 0.81 Attr (A,x) → Attr (A,y) : 0.61 → 0.84 Reasoning across multiple pieces of evidence x: In 1899, the first driver's license was issued in France. The first license in the United States was issued in 1902 in Massachusetts. E: -France introduced the first driving licences under the Paris Police Ordinance of 14 August 1893. . . -Massachusetts and Missouri were the first states to require a license in 1903, but there was no test associated with it. . . y: In 1893, the first driver's license was issued in France. The first license in the United States was issued in 1903 in Massachusetts. P res (x,y) : 0.90 Attr (A,x) → Attr (A,y) : 0.03 → 0.96 Challenging distractors in evidence x: The first orca put in captivity was in 1965. It was at Seattle, Washington's Seattle Aquarium. The orca's name was Wanda. E: -On Nov. 18, 1961 a killer whale had been spotted and a collection crew from the Marineland of the Pacific scrambled to capture her. Never had an orca been captured alive, and Wanda was trying to avoid being the first. . . -The first orca to be captured in 1961 was found off the coast of California and given the name Wanda. y: The first orca put in captivity was in 1961. It was at Marineland of the Pacific. The orca's name was Wanda. P res (x,y) : 0.77 Attr (A,x) → Attr (A,y) : 0.33 → 0.77\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"(A,x) → Attr (A,y) : 0.00 → 0.20 Miscellaneous (15%) x: The song was written by Tony Macaulay and John Macleod and recorded by Helen Shapiro in 1961. y: The song was written by Dub Allbritten and Ronnie Self and recorded by The Jackson 5 in 1960.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"20, 2021 the Detroit Lions named Dan Campbell the franchise's new head coach. . . -Campbell possesses 23 years of NFL experience, including 12 years as a coach and 11 as a player. In his first year. . . -On Jan. 20, 2021 the Detroit Lions named Dan Campbell the franchise's new head coach. . . x/y: Dan Campbell was appointed the new head assistant coach of the Detroit Lions on January 20, 2021. With 23 19 years of NFL experience, 12 as a coach and 11 7 as a player. . . What is the neurological explanation for why people laugh when they're nervous or frightened? E + : -A 2015 Yale study found people respond with a variety of emotions to strong outside stimuli. . . -Vilayanur Ramachandran states \\\"We have nervous laughter because we want to make ourselves think what horrible thing we encountered isn't really as horrible as it appears\\\". . . -Stanley Milgram conducted one of the earliest studies about nervous laughter in the 1960s. His study revealed that people often laughed nervously in uncomfortable situations. . . x/y: Yale researchers in 2015 found people often respond to strong external stimuli with a variety of emotions, including nervous laughter anger. Stanley Milgram's Vilayanur Ramachandran's 1960s study also observed this in uncomfortable situations. Neuroscientist Vilayanur Ramachandran Stanley Milgram theorizes that people laugh when. . . .\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Results on the Editing for Attribution task. We report the attribution of the statement before and after editing, preservation after editing, and F 1 AP which combines attribution and preservation. Results are on LLM outputs on factoid question answering, long reasoning question answering, and dialog.\",\n \"figure_data\": \"Model Attr. (x → y) Pres. F 1 APPALM outputs on NQEFEC44.7 → 63.9 39.648.5RARR 44.7 → 53.8 89.667.2PURR 44.8 → 59.8 91.072.2PALM outputs on SQAEFEC37.2 → 58.2 31.040.4RARR 37.2 → 44.6 89.959.6PURR 36.9 → 47.1 92.062.3LaMBDA outputs on QreCCEFEC18.4 → 47.2 39.042.7RARR 18.4 → 28.7 80.142.2PURR 16.8 → 33.0 85.847.7\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Example of good and bad revisions with PURR. x = claim; E = relevant evidence; y = edited claim using E. PURR can handle hallucinated entities and spans as well as merge information across evidence to edit. PURR can struggle when there are challenging distractors in a piece of evidence.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Query Generation Missing Coverage (35%) x: Legends of Tomorrow season 3 finale aired on April 9, 2018. It's title is No Country for Old Dads and is a 42-minute episode. Q: -When did the season 3 finale of Legends of Tomorrow air? -What's the title of Legends of Tomorrow season 3 finale? -How long is the season 3 finale of Legends of Tomorrow? y: Legends of Tomorrow season 3 finale aired on October 10, 2017. It's title is Freakshow and is a 43-minute episode. Bryce Dallas Howard plays the character of the mother of Cindy Lou Who in the 2018 animated movie The Grinch. E: Donna Lou Who is one of the characters of the 2018 Illumination animated film, Dr. Seuss' The Grinch. She's the mother of Cindy Lou Who and twin infants sons, Buster and Bean. She's voiced by Rashida Jones . y: No edit.\",\n \"figure_data\": \"Attr (A,x) → Attr (A,y) : 0.01 → 0.01Search Not Finding Best Evidence (45%)x: Attr (A,x) → Attr (A,y) : 0.05 → 0.05Editor Error/Confusion (5%)\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Error Analysis of PURR Inference Pipeline. We sample 20 edits from the NQ set where attribution is low after editing and categorize why by component. x = claim; Q = generated queries used for search, E = relevant evidence; y = edited claim using E. Strikethrough text represents a query that wasn't generated or evidence that wasn't retrieved but should have been.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Supporting Evidence\", \"Citation\": \"(Bai et al., 2022)\", \"Explanation\": \"The cited work by Bai et al. provides a method for few-shot prompting of language models to perform editing tasks, which the citing paper uses to highlight the advantages of posthoc methods in text editing.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Gao et al., 2022)\", \"Explanation\": \"The cited work by Gao et al. also contributes to the discussion on the use of language models for text editing, providing another method for few-shot prompting of language models to perform editing tasks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Iv et al., 2022)\", \"Explanation\": \"The cited work by Iv et al. extends the research on text editing by fine-tuning small, cheap models to perform editing tasks in specific domains where training data is available.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Schick et al., 2022)\", \"Explanation\": \"The cited work by Schick et al. also contributes to the discussion on text editing by fine-tuning small models to perform editing tasks in specific domains, building upon the research in the cited work by Iv et al.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Balachandran et al., 2022)\", \"Explanation\": \"The cited work by Balachandran et al. provides a method for training a denoising model that requires access to relevant paired evidence, which the citing paper uses to ground the denoising process in their data generation pipeline.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Schick et al., 2022;Iv et al., 2022)\", \"Explanation\": \"The cited works by Schick et al. and Iv et al. are extensions of the research on training denoising models, as they provide methods for training the model using edit data and noising the data during training, respectively.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Beyer et al., 2021;Stanton et al., 2021)\", \"Explanation\": \"The cited works by Beyer et al. and Stanton et al. are data sources for the distillation process in the research on training denoising models, as they provide information on the importance of a challenging distillation set and the under-performance of the student model compared to the teacher model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Raffel et al., 2020)\", \"Explanation\": \"The cited work by Raffel et al. (2020) provides the data generation pipeline that the citing paper uses to fine-tune the T5 model in the PURR model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Izacard and Grave, 2021)\", \"Explanation\": \"The cited work by Izacard and Grave (2021) contributes the denoising data that the PURR model is fine-tuned on, which is a key methodological element in the model development.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Gao et al., 2022)\", \"Explanation\": \"The cited work by Gao et al. provides the primary objective and methods for the task of attributing the outputs of language models and rectifying inaccuracies, which the citing paper builds upon to present its own research in the same area.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Rashkin et al., 2021)\", \"Explanation\": \"The cited work by Rashkin et al. provides a formalization of the concept of attribution, which the citing paper adopts in the evaluation of editing-for-attribution systems.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Gao et al., 2022)\", \"Explanation\": \"The cited work by Gao et al. (2022) provides the evaluation sets used in the citing paper for the tasks of Natural Questions, StrategyQA, and QreCC, which serve as the data source for the study conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Aly et al., 2021)\", \"Explanation\": \"The cited work by Aly et al. (2021) provides the training data and methodology for the development of the EFEC model, which the citing paper uses as a fine-tuned T5 editor in their research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Gao et al., 2022)\", \"Explanation\": \"The cited work by Gao et al. (2022) builds upon the research of the citing paper by exploring the effects of the EFEC model on attribution and preservation in a new context, providing insights and recommendations for future work.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chowdhery et al., 2022)\", \"Explanation\": \"The cited work by Chowdhery et al. (2022) introduces the PALM model and the RARR approach to prompt-based editing, which the citing paper uses as a basis for their research on iterative evidence checking and few-shot prompting.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Balachandran et al., 2021)\", \"Explanation\": \"The cited work by Balachandran et al. provides a foundation for post-hoc editing methods that enhance the attribution and accuracy of generated text, which the citing paper builds upon in its research on PURR.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Cao et al., 2020)\", \"Explanation\": \"The cited work by Cao et al. contributes to the development of post-hoc editing methods for improving the quality of generated text, which the citing paper further builds upon in its research on PURR.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Iso et al., 2020)\", \"Explanation\": \"The cited work by Iso et al. provides a methodological basis for post-hoc editing methods in the context of generated text, which the citing paper leverages in its research on PURR.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Gao et al., 2022)\", \"Explanation\": \"The cited work by Gao et al. introduces RARR, a method for rectifying language model outputs through few-shot prompting. The citing paper extends this research by exploring similar techniques in the context of PURR.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(He et al., 2022)\", \"Explanation\": \"The cited work by He et al. presents Rethinking-with-Retrieval, a method for improving language model outputs through few-shot prompting. The citing paper extends this research by further exploring the use of such techniques in the context of PURR.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Iv et al., 2022)\", \"Explanation\": \"The cited work by Iv et al. introduces FRUIT, a method for enhancing the factuality of text through fine-tuning editors. The citing paper extends this research by exploring similar techniques in the context of PURR.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Thorne and Vlachos, 2020)\", \"Explanation\": \"The cited work by Thorne and Vlachos presents EFEC, a method for improving the factuality of text through fine-tuning editors. The citing paper extends this research by further exploring the use of such techniques in the context of PURR.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Schick et al., 2022)\", \"Explanation\": \"The cited work by Schick et al. introduces PEER, a method for collaborative writing and enhancing factuality through editing. The citing paper extends this research by exploring similar techniques in the context of PURR.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Carlini et al., 2020)\", \"Explanation\": \"The cited work by Carlini et al. provides a dataset or model that the citing paper uses to study the potential of large language models to regurgitate protected information.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Gehman et al., 2020)\", \"Explanation\": \"The cited work by Gehman et al. contributes a dataset or model that the citing paper uses to study the potential of large language models to spew hateful content.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zhao et al., 2021)\", \"Explanation\": \"The cited work by Zhao et al. provides a dataset or model that the citing paper uses to study the high sensitivity of large language models to input variations.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Sanh et al., 2021)\", \"Explanation\": \"The cited work by Sanh et al. (2021) provides a method of instruction fine-tuning that the citing paper adopts in their research to address the issues of LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Min et al., 2021)\", \"Explanation\": \"The cited work by Min et al. (2021) contributes a method of instruction fine-tuning that the citing paper uses to address the issues of LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chung et al., 2022)\", \"Explanation\": \"The cited work by Chung et al. (2022) provides a method of instruction fine-tuning that the citing paper adopts in their research to address the issues of LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ye et al., 2022)\", \"Explanation\": \"The cited work by Ye et al. (2022) contributes a method of instruction fine-tuning that the citing paper uses to address the issues of LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Korbak et al., 2023)\", \"Explanation\": \"The cited work by Korbak et al. (2023) provides a method of pretraining from human feedback that the citing paper uses in their research to address the issues of LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Bai et al., 2022)\", \"Explanation\": \"The cited work by Bai et al. (2022) proposes a method of editing LLMs using prompted LLMs to address undesired generation behaviors, which the citing paper aims to apply in their future research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wang et al., 2021)\", \"Explanation\": \"The cited work by Wang et al. (2021) contributes to the use of LLMs in data generation pipelines, which the citing paper mentions as a way of distilling the knowledge of the language model if their outputs are used for training.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Bartolo et al., 2022)\", \"Explanation\": \"The cited work by Bartolo et al. (2022) discusses the use of LLMs in data generation pipelines, which the citing paper mentions as a way of distilling the knowledge of the language model if their outputs are used for training.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Lang et al., 2022)\", \"Explanation\": \"The cited work by Lang et al. (2022) contributes to the use of LLMs in data generation pipelines, which the citing paper mentions as a way of distilling the knowledge of the language model if their outputs are used for training.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Smith et al., 2022)\", \"Explanation\": \"The cited work by Smith et al. (2022) discusses the use of LLMs in data generation pipelines, which the citing paper mentions as a way of distilling the knowledge of the language model if their outputs are used for training.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Eisenstein et al., 2022)\", \"Explanation\": \"The cited work by Eisenstein et al. (2022) provides a method of using LLMs in data generation pipelines, which the citing paper mentions as a way of distilling the knowledge of the language model if their outputs are used for training.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(2022)\", \"Explanation\": \"The cited work by (2022) follows a multi-step distillation pipeline that the citing paper adopts in their own research, chaining the outputs of multiple LLM calls and distilling the output into an explainable question answering model.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"Liu et al. (2022)\", \"Explanation\": \"The cited work by Liu et al. (2022) uses the outputs of LLMs followed by filtering and human refinement to create WANLI, a challenging natural language inference dataset. The citing paper extends this work by using the same approach to create their own challenging training instances.\"}, {\"Category\": \"Data Source\", \"Citation\": \"Ribeiro and Lundberg (2022)\", \"Explanation\": \"The cited work by Ribeiro and Lundberg (2022) uses LLMs to generate evaluation sets for testing LLMs. The citing paper acknowledges the origin of this data as a foundational element for their own research on evaluating LLMs.\"}]"}}},{"rowIdx":96,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b18","b19","b59","b41","b55","b23","b54","b52","b53","b30","b32","b47","b54","b12","b17","b21","b46","b36","b57","b29"],"table_ref":[],"text":"The field of artificial intelligence has been revolutionized with the advent of large pre-trained models. Of particular significance are transformer-based large language models (LLMs) which have showcased remarkable performance in natural language processing tasks. Along with these tasks, LLMs have been tested to perform another widely-studied crucial aspect of AI agents, namely, sequential decision-making or planning. Preliminary studies suggest that, in some everyday domains, LLMs are capable of suggesting sensible action plans [19,1]. However, the correctness and executability of these plans are often limited. For instance, LLMs may regularly overlook the physical plausibility of actions in certain states and may not effectively handle long-term dependencies across multiple actions. Several approaches have been proposed to improve the planning capabilities of LLMs. One promising approach involves collecting feedback from the environment during plan execution and subsequently refining the plans. By incorporating various forms of feedback, such as sensory information [20], human corrections [60], or information of unmet preconditions [42,56], the planners can re-plan and produce plans that are closer to a satisficing plan.\nDespite the improvements in planning performance, LLMs are still far from being a usable and reliable planner due to various factors:\n(a) LLMs have not yet demonstrated sufficient capabilities in reasoning and planning [24,55,53,54,31]. Recent investigations show that even when provided with detailed descriptions of actions, such as a PDDL domain model [33] or a natural-language version of a PDDL model, LLMs still struggle to produce correct and executable plans [48,55].\n(b) Existing LLMs-planning paradigms only allow for feedback collection in a fully online manner, meaning that the feedback signals are only available after the agent has started executing the plan. However, when a faithful simulator is not available or is expensive to use, collecting feedback through actual plan execution can be costly and may not fully exploit the advantages of provably sound planning, as seen in classical-planning literature [11,13].\n(c) LLMs exhibit complex behaviors that are not yet fully understood, particularly with respect to error occurrences. LLM planners are prone to repeating the same mistakes in slightly different scenarios. Repeatedly providing the same feedback can lead to frustration for end users.\nTo overcome these limitations, rather than using LLMs directly as planners, we advocate a modelbased paradigm, wherein a PDDL world model is teased out of LLMs. We follow the identical problem setup as existing approaches, which involves providing the planner with a set of actions and their brief natural language descriptions. However, instead of directly mapping user commands to plans, we utilize LLMs to extract a symbolic representation of the actions in the form of PDDL action models. This intermediate output can be used with an external domain-independent planner to reliably search for feasible plans, or it can be used to validate and correct \"heuristic\" plans generated by an LLM planner. Additionally, our modular method essentially divides the planning process into two distinct parts, namely modeling the causal dependencies of actions and determining the appropriate sequence of actions to accomplish the goals. LLMs, which have been trained on extensive web-scale knowledge, exhibit greater proficiency in the former task rather than the latter.\nNevertheless, we still take into account the fact that the LLMs may not be able to generate error-free PDDL models at the outset. To address this, we show that LLMs can also serve as an interface between PDDL and any feedback sources that can provide corrective feedback in natural language, such as humans and the PDDL validator in VAL [18]. The LLM middle layer translates PDDL representation to natural language and presents it to users for inspection. The acquired feedback is then incorporated and archived back to the PDDL models. This conceals the complexity of PDDL from users who do not have prior knowledge of PDDL, and enables seamless inclusion of feedback.\nWe conducted an extensive evaluation of our methodology on two IPC domains [22] from classical planning literature and a household domain that has a more diverse set of actions and constraints than commonly used benchmarks such as ALFWORLD [47]. We assess the quality of the generated PDDL models through manual evaluation. Results show that GPT-4 [37] generates high-quality PDDL domain models with over 400 literals for 41 actions in total. Then, by replaying and continuing the PDDL-construction dialogue, we show that GPT-4 can readily correct all the errors according to natural language feedback from PDDL validators and humans.\nWe consider two use cases of the generated PDDL action models for downstream planning tasks. For one, by utilizing an LLM to translate user instructions into goal specifications in PDDL [58,30], we can use any standard domain-independent planner to search for a plan. On the other hand, the extracted PDDL model can be used to validate plans suggested by an LLM planner and to provide corrective feedback in the form of unmet preconditions or goal conditions. In this case, the PDDL model is essentially serving as an inexpensive high-level simulator or a human proxy to ensure plan correctness.\nThis reduces the reliance on faithful simulators or extensive manual inspection of plans by domain experts. Compared to the first approach, the second approach potentially offers better flexibility in incorporating both explicit and implicit user constraints in common-sense domains because of the LLM planner. For instance, the LLM planner can directly incorporate ordering constraints such as \"heat the potato first before mashing it\" and \"bring me a fork first, then a plate.\" On the contrary, an approach purely based on classical planners would require extra steps, such as introducing extra state variables in the PDDL models, in order to accommodate such constraints. However, as demonstrated in our experiments, although the validation feedback significantly improves the plan correctness on average, the performance of the second approach is still limited by the \"planning capability\" of LLMs."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b23","b54","b54","b47","b28","b27","b49","b59","b50","b19","b33","b59","b45","b41","b55","b54","b37","b61","b26","b6","b42","b25","b58","b60","b8","b48","b1","b34","b57","b31","b14","b2","b30","b52","b53","b44","b38","b56"],"table_ref":[],"text":"LLMs and planning. The growing interest in evaluating the emergent abilities of LLMs paved way into exploring their abilities in sequential decision-making tasks. Preliminary studies [24,55] have shown that off-the-shelf LLMs are currently incapable of producing accurate plans. But their plans can be used as heuristics or seeds to either an external planner or a human in the loop [55,48].\nSayCan [1] and Text2Motion [29] employ an LLM as a heuristic by utilizing it to score high-level actions, followed by a low-level planner that grounds these actions to determine the executability in the physical world. In a similar vein, [28,50] use LLMs to generate plans represented in Python-style code. Other works have aimed to improve the planning performance of LLMs through prompt engineering [60] or collecting various forms of feedback such as sensory information [51,20,34], human corrections [60], self-corrections [46] or information of unmet preconditions [42,56].\nTraining transformers for sequential decision-making tasks. Along with using off-the-shelf LLMs, there are works that either fine-tune LLMs [55,38] or train sequence models [62,27,7,43] for sequential decision making tasks. Experiments in [26] have shown that training sequence models on a specific task gives rise to an internal world representation within the model. In this work, we use off-the-shelf LLMs to construct symbolic world models without performing any extra training.\nLearning/acquiring symbolic domain models. In classical planning, the community has explored numerous learning-based methods [59,61,9,25,4] and interactive editor-based methods [49] for acquiring symbolic domain models. For a more comprehensive survey, we refer the reader to [2,6].\nHere, we are interested in leveraging the common-world knowledge embedded in LLMs and their in-context learning ability for constructing domain models. Recent studies have shown the efficacy of LLMs in translating natural language to formal descriptions [35] or constructing PDDL goals from natural-language instructions [58,32]. Moreover, a contemporary work [15] considers the use of LLM as a parametric world model and plan critic. However, unlike a symbolic model that can simulate plan outcomes with guaranteed correctness, using LLMs directly as a world model actually adds another layer of errors. There is evidence that autoregressive models lack reliable capacity for reasoning about action effects [3,31] and capturing errors in candidate plans [53,54].\nLanguage models with access to external tools. Since LLMs are approximately omniscient, they may not always outperform specialized models or tools in specific downstream tasks. To address this limitation, frameworks have been developed to enable LLMs to utilize external tools for performing sub-tasks like arithmetic [45] and logical reasoning [39,57]. In this context, our work can be regarded as an exercise in employing external sound planners to augment the capacity of LLMs for more reliable plan generation."},{"figure_ref":[],"heading":"Problem Setting and Background","publication_ref":["b15","b11","b59","b19"],"table_ref":[],"text":"Our work focuses on a scenario where an intelligent agent receives high-level instructions or tasks, denoted as i, from a user. The agent is capable of only executing skills or operations that are part of a skill library Π, where each skill k has a short language description l k . We assume that the agent is equipped with the low-level control policies corresponding to these high-level skills. In order to achieve the goal conditions specified in i, a planner, which can be either an LLM or an external planner [16,12,17], needs to come up with a sequence of high-level skills that the agent can execute. This type of problem is referred to as a sequential decision-making or planning problem. Similar to previous works such as [60,20], we also allow for human-in-the-loop feedback during both the domain-model construction and plan execution stages. In the next subsections, we describe the formalism behind planning problems and a standard way in the literature to specify them."},{"figure_ref":[],"heading":"Classical planning problems","publication_ref":["b43"],"table_ref":[],"text":"The most fundamental planning formalism is goal-directed deterministic planning problem, referred to as a classical planning problem in the planning literature. A classical planning problem [44] can be formally represented with a tuple P = ⟨D, I, G⟩. D is referred to as the domain, I is the initial state, and G is the goal specification. The state space of a planning problem consists of the truth assignments for predicates. The domain D is further defined by the tuple D = ⟨F, A⟩. F corresponds to the set of fluents, i.e., the state variables used to define the state space with each fluent corresponding to a predicate with some arity. A corresponds to the set of actions that can be performed. Each action a i [V] ∈ A (where V is the set of variables used by the operator a i and each variable could be mapped to an object) can be further defined by two components, the precondition prec[V] which describes when an action can be executed, and the effects eff[V] which defines what happens when an action is executed. We assume that prec[V] consists of a set of predicates defined over the variables V. An action is assumed to be executable only if its preconditions are met, i.e, the predicates in the precondition hold in the given state. The effect set eff[V] is further defined by the tuple ⟨add[V], del[V]⟩, where add[V] is the set of predicates that will be set true by the action and del[V] is the set of predicates that will be set false by the action. An action is said to be grounded if we replace each of the variables with an object, else it is referred to as a lifted action model. A solution to a planning problem is called a plan, and it is a sequence of actions that once executed in the initial state would lead to a state where the goal specification holds. Classical planning problems are one of the simpler classes in planning and there are multiple extensions with more complex forms of preconditions, conditional effects, and also support for richer planning formalisms."},{"figure_ref":[],"heading":"PDDL","publication_ref":["b32"],"table_ref":[],"text":"Planning Definition and Domain Language (PDDL) [33], is the standard encoding language for classical planning problems. Here is an example of a lifted action in PDDL which corresponds to putting a block onto the table in the classical Blocksworld domain:\n(:action PutDownBlock :parameters (?x -block) :precondition (and (robot-holding ?x)) :effect (and (not (robot-holding ?x)) (block-clear ?x) (robot-hand-empty) (block-on-table ?x)))\nThe parameters line provides the possible variable(s), and in this case, ?x represents the block to put down. The precondition states that the robot must be holding the block in its gripper. The effects line describes the expected outcome of this action."},{"figure_ref":[],"heading":"Methodology","publication_ref":[],"table_ref":[],"text":"PDDL provides a succinct and standardized way to represent a world model. Once a PDDL model is constructed, it can be seamlessly used by any domain-independent planner developed in the automated planning community to search for a plan given the initial state and goal conditions. In this section, we will introduce our solution for constructing PDDL models using LLMs. We then discuss techniques for correcting errors in the generated PDDL models. Finally, we present the full pipeline for utilizing the generated PDDL models to solve planning problems."},{"figure_ref":["fig_1","fig_1"],"heading":"Constructing PDDL models with LLMs","publication_ref":["b29","b40","b36","b9"],"table_ref":[],"text":"Our approach involves prompting pre-trained LLMs with the following information: (a) detailed instructions for the PDDL generation task, outlining components of upcoming inputs and desired outputs; (b) one or two examples from other domains (e.g., the classical Blocksworld domain) for illustrating the input and output formats; (c) a description of the current domain, including contextual information about the agent's tasks and physical constraints due to the specific embodiment of the agent; (d) a description of the agent's action; and (e) a dynamically updated list of predicates that the LLM can reuse to maintain consistent use of symbols across multiple actions. Note that the predicate list is initialized to an empty list, and thus all predicates are introduced by the LLM. The structure of the prompt is illustrated in Fig. 2, and a complete prompt for the household-robot domain can be found at Appx. A.6.1. Depending on the information included in the action description or the domain context, users may gain varying levels of control over the extracted PDDL or receive differing levels of support from the LLMs. On one hand, when the user provides only a minimal description of the action, such as \"this action enables the robot to use a microwave to heat food,\" we not only use the LLM as a PDDL constructor but also leverage the common world knowledge encoded within the model for knowledge acquisition. This is particularly useful when expanding the set of actions for an AI agent. For example, a robot engineer could set up a training environment for skill learning by following the suggested preconditions and effects. On the other hand, when some preconditions or effects are explicitly mentioned in the prompt, we rely more on the LLM's ability to parse the knowledge provided in natural language and to precisely represent it by devising a collection of predicates. This capability is useful when there could be different initial setups of a skill, and the engineers have already made some assumptions on the preconditions at the time of designing the skill. This capability is also crucial when constructing PDDL for specialized domains. For instance, robots such as Fetch and Spot Robot have only one robot arm, which is less flexible than a human arm, and are therefore subject to many uncommon physical constraints.\nThe desired output comprises the following elements: (a) the list of arguments for the action; (b) the preconditions and effects expressed in PDDL; and (c) a list of any newly defined predicates and their descriptions in natural language, if applicable. An example output is shown in Fig. 2. Our algorithm generates PDDL models for each action separately, one at a time, by iterating over the set of actions. Any newly defined predicates will be added to an actively maintained predicate list, such that the LLM can reuse existing predicates in subsequent actions without creating redundant ones. Once we obtain the initial PDDL models and the full predicate list, we repeat the entire process but with all of the extracted predicates presented to the LLM. Running the generation process twice is useful because the LLMs may be unaware of some precondition(s) during the first iteration, especially if the precondition(s) are not explicitly mentioned. For instance, the LLM may overlook the fact that a furniture piece can be openable, but a predicate created in the \"open a furniture piece or appliance\" skill can inform the LLM of this fact. One alternative to this action-by-action generation could be to include descriptions of all the actions in the prompt and require the LLM to construct the entire domain model in a single dialogue. An additional discussion on this can be found at Sec. A.2 in Appendix.\nIt is worth noting that every time a new predicate is defined, the LLM is required to give the natural language description of it. As we will see in the following sections, this is crucial for enabling any user to easily understand and inspect the generated PDDL models without having to delve into the low-level symbolic representation. Additionally, natural language descriptions allow the predicate values of the initial state to be automatically grounded by using LLMs to translate environment description in natural language to PDDL [30], or leveraging pre-trained vision-language models ---------------------------------------The LLM: ... [ 41,37,10] and querying them in a question-answering manner, based on observations from the environment."},{"figure_ref":[],"heading":"Correcting errors in the initial PDDL models","publication_ref":["b17","b59"],"table_ref":[],"text":"As with any use case involving LLMs, there is no guarantee that the output is completely error-free. Therefore, it is essential to incorporate error-correction mechanisms. While it may be easy for PDDL experts to directly inspect and correct the generated PDDL models, we cannot assume that all end users possess this level of expertise. Our solution is to use the LLM as a middle layer or interface between the underlying PDDL model and any feedback source that can provide corrective feedback in natural language. We consider two feedback sources in this work, namely the PDDL model validation tools (e.g., the one in VAL [18]) and human domain experts. The former is used to detect basic syntax errors, while the latter is mainly responsible for catching factual errors, such as missing effects. It is worth noting that the feedback sources are not limited to those mentioned above, and we leave the investigation of other sources for future research.\nFor corrective feedback from PDDL validators, a generated PDDL model is directly presented to the validator to obtain brief but readable error messages. Examples of feedback messages for syntax errors are shown in Appx. A.3. For corrective feedback from users, a PDDL model is translated into its natural-language version based on the natural language descriptions of the predicates and parameters (Sec. 4.1). The user can then examine potentially erroneous action models. Human corrections can occur both during the construction of PDDL models and after the models have been used for planning. Although there are techniques available to assist users to locate errors in the models (as discussed in Appx. A.4), this is beyond the scope of this work, since the focus here is to investigate the feasibility of using LLMs to correct PDDL models based on feedback. We also note that correcting action models is not more cognitively demanding than correcting plans or the \"reasoning traces\" of an LLM planner [60]. In fact, when correcting plans, humans must also maintain the action models and their causal chains in mind in order to validate the plans. More importantly, once the action models are corrected, users no longer need to provide similar feedback repeatedly. Finally, corrective feedback is integrated by replaying and continuing the PDDL-construction dialogue. Examples of such dialogues can be found in Sec. A.7, Sec. A.9, and Sec. A.11 in Appendix."},{"figure_ref":[],"heading":"Generating plans with the extracted PDDL models","publication_ref":["b29","b57","b31","b11","b54","b41","b20"],"table_ref":[],"text":"Recall that given the set of extracted predicates and their natural language descriptions, we can get the grounded initial state by using LLMs to translate descriptions of the environment to PDDL, or by observing the environment and querying pre-trained vision-language models. Besides, the goal specification can be obtained by using an LLM to parse the user's command and convert it into a symbolic form, as done previously in [30,58,32]. With this setup, the following two methods can be used to generate the final plans.\nClassical planner with LLM-acquired PDDL model. One straightforward approach is to employ a standard domain-independent planner to reliably find a satisficing or even optimal plan for the specified goal. In common-sense domains where LLMs may generate meaningful \"heuristics\", the LLM plans may also be used as seed plans for a local-search planner such as LPG [12] to accelerate the plan searching. This is similar to the approach suggested in [55], but with a higher degree of automation.\nLLM modulo planner backprompted by VAL using LLM-acquired PDDL model. As outlined in Sec. 1, we can also use the extracted PDDL as a symbolic simulator or human proxy to provide corrective feedback based on validation information to an LLM planner. With this setup, the planner can iteratively refine the plans through re-prompting [42].\nIt is worth noting that depending on the specific problem settings, the extracted PDDL model can also be used for tasks other than task planning. For instance, in cases where reinforcement learning is permissible, the domain model can be used to guide skill learning [21,8] or exploration even if the model is not fully situated [14]."},{"figure_ref":[],"heading":"Empirical Evaluation","publication_ref":["b46","b39","b36","b35"],"table_ref":[],"text":"We conduct our experiments2 on an everyday household-robot domain and two more specialized IPC domains (i.e., Tyreworld and Logistics). The Household domain is similar to other commonly used benchmarks like ALFWORLD [47] and VirtualHome [40]. However, in our household domain, a single-arm robot is equipped with a more diverse and extended set of 22 mobile and manipulation skills. In addition, we apply more rigorous physical-plausibility constraints to each skill. A detailed description of this domain can be found at Appx. A.5. In our experiments, we first evaluate the quality of PDDL models generated by the LLMs. Next, we assess the ability of the LLMs to incorporate corrective feedback from both PDDL validators and users in order to obtain error-free PDDL models. Lastly, we showcase multiple ways to use the corrected PDDL model for downstream planning tasks. We present the results of GPT-4 [37] and GPT-3.5-Turbo [36] for PDDL construction (we also conducted experiments with GPT-3 [5], and observe that its performance is comparable to that of GPT-3.5-Turbo)."},{"figure_ref":[],"heading":"Constructing PDDL","publication_ref":[],"table_ref":[],"text":"In PDDL construction tasks, we aim to investigate the extent to which LLMs can construct accurate PDDL models before getting corrective feedback from domain experts. For all the domains, two actions from the classical Blocksworld domain are used as demonstrations in the prompt so that the end user is not required to come up with any domain-specific example. To evaluate the degree of correctness, we recruit multiple graduate students who possess expertise in PDDL. These experts are responsible for annotating and correcting any errors present in the generated PDDL models. As an evaluation metric, we count and report the total number of annotations, which may include the removal of irrelevant preconditions, the addition of missing preconditions, the replacement of incorrect predicates, the inclusion of missing parameters, and other commonly made corrections. Note that the number of annotations can be viewed as the approximate distance between a generated PDDL model and its corrected version. In order to provide the reader with a comprehensive understanding of the quality of the generated models, we also list all the models and collected annotations in Appendix. In each of the figures, errors that affect the functionality of the PDDL model are highlighted in yellow, while minor issues are highlighted in green. One example of a minor issue is the redundant inclusion of (pickupable ?o) in preconditions when (robot-holding ?o) has already been listed. The former is unnecessary because it can be implied by the latter, but this only affects conciseness rather than functionality. Table 1: The number of errors in the domain models produced by the LLMs for each of the domains.\nA \"+\" mark indicates that the generated model is excessively noisy, making it challenging to determine an exact number of errors.\nWe first evaluate the PDDL models generated when partial constraint information is given, as this is closer to most of the practical use cases where constraints on skills in the library Π are often pre-specified. In this setting, our evaluation focuses on the LLMs' ability to accurately recover a \"ground truth PDDL\" that captures the mentioned constraints and underlying dependencies among skills. Our results indicate that GPT-4 can produce high-quality PDDL models with significantly fewer errors when compared to GPT-3.5-Turbo. Table 1 presents the number of errors in the generated domain models for each domain. To help the readers understand the complexities of the action models, we additionally report the total number of parameters and literals in the final corrected domain models produced by GPT-4. Out of the total 59 errors made by GPT-4, three of them are syntax errors and the rest are factual errors such as missing preconditions and effects. This observation suggests that while GPT-4 demonstrates proficiency in adhering to the grammar of PDDL, it may still have an inaccurate understanding of the actions. By examining the set of predicates (listed in the Appendix), we also find that GPT-4 can devise a set of intuitively-named predicates that can concisely and precisely describe the states of objects and events in the domain. In contrast, GPT-3.5-Turbo produces highly noisy outputs with over 350 errors. This suggests that our framework relies heavily on GPT-4's improved capability in understanding symbols, and future work may investigate how to enable the use of more lightweight models (e.g., by fine-tuning on some PDDL datasets). Furthermore, recall that when the action description contains minimal information, LLMs could also be utilized to propose preconditions and effects to assist with knowledge acquisition. To verify this hypothesis, we conduct additional experiments on the Household domain that can have a more open-ended action design. In this setting, the correctness of the action models is determined based on whether the preconditions and effects establish correct connections among the actions. Our results show that GPT-4 can suggest meaningful action models, and the generated PDDL models have only around 45 errors.\nAlthough GPT-4 has shown improved performance in the PDDL construction task, our experiments still uncover some limitations. Firstly, GPT-4 still exhibits a shallow understanding of the causal relationships between actions, particularly when it comes to tasks involving reasoning skills such as spatial reasoning. For instance, when constructing the model of action \"pick up an object from a furniture piece,\" GPT-4 fails to consider that there could be other objects stacked on top of the target object, even if relevant predicates are provided (which were created in the action \"stack objects\").\nIn addition, although it occurs rarely, GPT-4 may output contradictory effects. For instance, in the action of mashing food with a blender, GPT-4 lists both (not (object-in-receptacle ...)) and (object-in-receptacle ...) as effects at the same time."},{"figure_ref":[],"heading":"Correcting PDDL with domain experts","publication_ref":[],"table_ref":[],"text":"We proceed with the PDDL models generated by GPT-4 when the constraint information is partially given. Our objective is to demonstrate the feasibility of using GPT-4 as a middle layer to incorporate natural-language feedback and correct the PDDL models. As discussed in Sec. 4.2, we use PDDL validators to capture basic syntax errors. In the Household domain, there are two syntax errors associated with improper usage of relevant predicates due to issues with the object types of parameters 3 . As shown in Appx. A.7.1, by continuing the PDDL-construction dialogue with a feedback message \"the second parameter of object-on should be a furnitureAppliance but a householdObject was given,\" GPT-4 can locate the inaccurate PDDL snippet and replace it with a correct one. For the other factual errors, GPT-4 successfully corrects all of them based on the natural language feedback. An example feedback message on factual errors is \"there is a missing effect: the item is no longer pickupable after being mashed.\" More PDDL-correction conversations can be found in Appendix. We also experiment with feedback written in various ways, and GPT-4 is able to understand all the messages and successfully correct the models. To quantify how effectively GPT-4 utilizes feedback from domain experts, we count the number of feedback messages concerning factual errors. Our result shows that GPT-4 required 59 feedback messages to address a total of 56 factual errors. There are three instances where additional feedback was needed. One case involved the user reiterating the error, while the other two cases involved GPT-4 introducing new errors. Furthermore, we attempt to correct the same errors using GPT-3.5-Turbo. Results show that GPT-3.5-Turbo not only fails to correct all the errors but also occasionally introduces new errors, again confirming its lack of ability to manipulate symbols. Some examples can be found in Appendix starting from Sec. A.7.3."},{"figure_ref":[],"heading":"Generating plans with the extracted PDDL models","publication_ref":["b15","b29","b59"],"table_ref":["tab_3"],"text":"For planning tasks (i.e., user instructions and initial states), we use the Household domain and Logistics domain, where state-of-the-art LLM planners struggle to find valid plans. We sampled 27 tasks for Household and 21 for Logistics. For the initial states, we assume the grounding is provided, and for the goals, we leverage GPT-4 to translate user instructions into PDDL goal specifications in terms of the extracted predicates (an example prompt can be found at Appx. A.13), and send it over to a standard STRIPS planner which already has access to the domain model acquired through LLMs. With this setup, a classical planner Fast Downward [16] can effectively find valid plans in 95% of the cases (the failures were only due to goal translation errors). Note that in contrast to earlier methods such as [30] that use LLMs only as a mechanism for translating user goals to PDDL format, and throw that over to external sound planners with hand-crafted correct PDDL domain models, our approach uses LLMs themselves to develop the PDDL world model driving the external planner. On the other hand, for the approach that utilizes PDDL models to validate LLM plans (i.e., LLM modulo planner back-prompted by VAL using LLMacquired domain model), we employ the state-of-the-art algorithm ReAct [60] with GPT-4 as the underlying LLM planner. However, we made two modifications to the prompt design. Firstly, we provide a detailed description of all actions in natural language, including parameters, preconditions, and effects. These descriptions are obtained by using another LLM to translate the generated PDDL domain model into natural language. Secondly, we use only two fixed examples for each domain because end users might not always be able to provide a large pool of examples, and the planner should rely on the action model information.\nThe LLM plans, symbolic goal specifications, initial states and domain models are passed to a plan validation system (i.e., VAL) to check for unmet precondition(s) or goal condition(s). The validation results (given in PDDL) are then translated into natural language with GPT-4 and provided to the LLM planner by continuing the planning dialogue (see Appx. A.12.1 for examples). In our experiments, we limit the number of feedbacks per task to 8 due to the restricted access to GPT-4. Table 2 provides a summary of the average success rates of all approaches. Not surprisingly, the vanilla LLM planner constantly overlooks action preconditions and achieves an extremely low success rate. With the integration of validation feedback, we observe a notable improvement in plan correctness. Despite this improvement, the overall performance is still not satisfactory, as the success rate remains below 50%. Furthermore, we have observed that GPT-4 fails to effectively utilize the feedback, often getting stuck in a loop by repeatedly generating the same plan. In some cases, it may also introduce new errors while attempting to rectify the plans.\nBeyond the notion of correctness, the experiments also uncover intriguing properties of the LLM planner. In the Household domain, we intentionally introduce ordering constraints in some instructions that cannot be expressed using existing predicates (refer to Appx. A.12 for examples). Remarkably, upon manual examination of the generated plans, we observe that all LLM plans adhere to the specified ordering, despite not being entirely correct or executable. Furthermore, also in the Household domain, we observe that classical planners occasionally generate physically plausible but unconventional actions, such as placing a knife on a toaster when the knife is not being used. In contrast, the LLM planner rarely exhibits such actions, suggesting that LLMs possess knowledge of implicit human preferences. It would be meaningful to explore methods that more effectively combine the strengths of LLM planners and the correctness guarantee provided by symbolic domain models, particularly in determining which information from LLM plans should be preserved."},{"figure_ref":[],"heading":"Conclusion","publication_ref":["b35"],"table_ref":[],"text":"We introduce a new paradigm for leveraging LLMs in planning tasks, which involves maintaining an explicit world model instead of directly mapping user prompts to plans. This is motivated by the insight that LLMs, while incapable of the combinatorial search needed to produce correct plans, may be better suited as the source of world models. We present a complete pipeline that begins with generating high-quality PDDL models using GPT-4, then corrects the PDDL models with naturallanguage feedback, and finally utilizes the extracted domain models to reliably plan in multiple ways.\nOur experiments demonstrate that pairing LLMs with an external planner significantly outperforms existing methods when applied to two IPC domains and a household-robot domain that has more action-wise constraints than commonly used benchmarks such as ALFWorld. Apart from directions for further research that we have previously mentioned, there are several exciting opportunities for extending this work. Firstly, the complexity of our evaluation domains is still lower than that of many domains used in the classical planning literature. It remains to be seen whether LLMs can effectively scale to write PDDL models that express more intricate logic. Secondly, our framework assumes full observability, meaning that the agent must fully explore the environment to acquire object states at the beginning. It would be useful to support partial observability. Finally, our experiments assume the grounding of predicate values is done perfectly. However, it would be useful to take into account that perception can be noisy in practice.\napproach may allow the LLM to better establish a global view of all the actions. However, we do not pursue this alternative here for the following reasons: (a) the inclusion of all actions might result in a lengthy prompt, potentially exceeding the context window size of an LLM. This could pose practical issues for utilizing smaller language models (e.g., GPT-3.5-Turbo [36]) or attempting to train smaller specialized models; (b) our integration of corrective feedback relies on continuing the construction dialogue (Sec. 4.2), which necessitates a shorter initial prompt to fit within the context window; (c) our experiments indicate that the action-by-action construction approach already achieves satisfactory results"},{"figure_ref":[],"heading":"A.3 Examples of feedback messages that capture syntax errors","publication_ref":["b48"],"table_ref":[],"text":"Recall that we rely on the PDDL validator in VAL to identify syntax errors. However, several \"simpler\" syntax errors can be easily detected using simple Python scripts. In our experiments, we wrote our scripts to capture such syntax errors. Also note that since these errors can be detected at a minimal cost, the corresponding feedback messages are directly provided to the LLMs, and they are not counted in the results reported in Table 1. These \"simpler\" syntax errors include:\n1. In this work, we only consider standard base-level PDDL. However, it's possible that the LLMs have seen various extensions of PDDL and might use them in the constructed domain models. Hence, we provide a feedback message to the LLM whenever we detect an unsupported keyword. One example feedback is: \"The precondition or effect contain the keyword 'forall' that is not supported in a standard STRIPS style model. Please express the same logic in a simplified way. You can come up with new predicates if needed (but note that you should use existing predicates as much as possible).\"\n2. Newly created predicates might have the same names as existing object types, which is not allowed in PDDL. In such cases, a feedback message is provided to notify the LLM about the name clash. For instance, a message might state: \"The following predicate(s) have the same name(s) as existing object types: 1. 'smallReceptacle'. Please rename these predicates.\"\n3. Newly created predicates might have the same names as existing predicates, which is not allowed in PDDL. Also, LLMs often mistakenly list existing predicates under the 'New Predicates' section. In such cases, a feedback message is provided to notify the LLM about the name clash or mistake. For instance, a message might state: \"The following predicate(s) have the same name(s) as existing predicate(s): 1. (cutting-board ?z -smallReceptacle), true if the small receptacle ?z is a cutting board | existing predicate with the same name: (cutting-board ?z -householdObject), true if the object ?z is a cutting board. You should reuse existing predicates whenever possible. If you are reusing existing predicate(s), you shouldn't list them under 'New Predicates'. If existing predicates are not enough and you are devising new predicate(s), please use names that are different from existing ones. Please revise the PDDL model to fix this error.\" This is the most common syntax error made by GPT-4 in our experiments. 4. The LLMs might fail to only use the object types given in the prompt. An example feedback can be: \"There is an invalid object type 'pump' for the parameter ?p.\"\nIn our experiments, the above error types encompass the majority of syntax errors made by GPT-4. In some less-common cases, GPT-4 may have problems with predicate usage, usually caused by mis-matched object types. This kind of error can be captured by VAL and an example feedback message can be: \"There is a syntax error, the second parameter of 'object-on' should be a furnitureAppliance, but a householdObject was given. Please use the correct predicate or devise new one(s) if needed.\"\nA. 4 Techniques that assist users to locate errors in PDDL models Several well-established techniques and tools are available for locating errors in PDDL models. For example, graphical tools like GIPO [49] can effectively visualize the causal dependencies of actions. However, these advanced tools or techniques are beyond the scope of this work. Here, we outline a viable solution as a starting point for users who are unfamiliar with these tools.\nThere are two stages in which corrective feedback can be obtained: during the construction of the PDDL model, and when the domain model is used to generate a plan. In the first stage, end users can direclty review the domain model and identify potential factual errors. Since all the predicates and parameters are accompanied by natural language descriptions, we can easily convert the PDDL model into natural language and present it to the users. This allows users to pre-screen the preconditions and effects. Note that we do not expect all factual errors to be caught in this stage, because the users may not be aware of certain constraints until they review the final plans. In the second stage, the PDDL model is used to solve downstream planning problems by following the procedure outlined in Sec. 4.3. Two possible cases may arise here: (a) no plan can be found for the given goal specification, or (b) at least one plan is found but it either gets rejected by the users or results in execution failure in the actual environment. To address the first case, we can request the users to suggest a goal-satisficing plan, which is supposed to be executable (but not necessarily optimal). We then use the generated PDDL model to \"validate\" the suggested plan. This allows us to find the first step in the plan that has an unsatisfied precondition(s). The models of all actions up to this step, along with the unmet precondition(s), are then converted to natural language and presented to the user for inspection. As an example, in the model of \"slice an object\" extracted by GPT-4 (Sec. A.6.11 in Appendix), the model requires the object to be placed on a cutting board and a furniture piece at the same time, which is not physically possible. By leveraging a user-suggested plan, we can identify the potentially erroneous model(s) and flag incorrect precondition(s). In the second case where an invalid plan is afforded by the PDDL model, there are typically missing preconditions or effects in the actions taken, both during and prior to the execution failure. The users can bring attention to these actions."},{"figure_ref":[],"heading":"A.5 Detailed description of the Household domain","publication_ref":[],"table_ref":[],"text":"We consider a single-arm robot model that closely resembles the SPOT robot and the Fetch Mobile Manipulator. Consequently, the robot is incapable of grasping multiple objects simultaneously or executing manipulation actions while holding irrelevant items (e.g., opening a fridge door while holding a mug). We also ensure the constraints align with real-world robotic capabilities. For example, we recognize that robot arms may be significantly less flexible than human arms, and therefore, we require certain manipulation tasks to be performed on furniture pieces with open and flexible surfaces (e.g., the robot can only pick up food items from a lunch box when the lunch box is placed on a kitchen countertop instead of inside a fridge). The list of actions and their descriptions can be found in Appendix starting from Sec. A.6.2. The PDDL-construction prompt for this domain includes a general description of the domain, which outlines the tasks to be performed by the robot, the types of objects involved, and details of the robot's morphology. An example of a complete prompt can be found in Fig. A.6.1."},{"figure_ref":[],"heading":"A.6 Household: Constructing PDDL Models","publication_ref":[],"table_ref":[],"text":"The AI agent here is a household robot that can navigate to various large and normally immovable furniture pieces or appliances in the house to carry out household tasks. Note that the robot has only one gripper, so (a) it can only hold one object; (b) it shouldn't hold any other irrelevant objects in its gripper while performing some manipulation tasks (e.g., opening a drawer or closing a window); (c) operations on small household items should be carried out on furniture with a flat surface to get enough space for manipulation. There are three major types of objects in this domain: robot, furnitureAppliance, and householdObject. The object type furnitureAppliance covers large and normally immovable furniture pieces or appliances, such as stove burners, side tables, dining tables, drawer, cabinets, or microwaves. The object type householdObject covers all other small household items, such as handheld vacuum cleaners, cloth, apples, bananas, and small receptacles like bowls and lunch boxes. There is a subtype of householdObject called smallReceptacle that covers small receptacles like bowls, lunch boxes, plates, etc. In this domain, the locations of the robot and small household items (e.g., apples, oranges, bowls, lunch boxes or lamps) are determined by large and normally immovable furniture pieces or appliances.\nA.6.1 An example prompt for constructing PDDL models of the action \"close a small receptacle\"\nAn example prompt for constructing PDDL models of the action \"close a small receptacle\"\nInstructions for the PDDL generation task You are defining the preconditions and effects (represented in PDDL format) of an AI agent's actions. Information about the AI agent will be provided in the domain description. Note that individual conditions in preconditions and effects should be listed separately. For example, \"object_1 is washed and heated\" should be considered as two separate conditions \"object_1 is washed\" and \"object_1 is heated\". Also, in PDDL, two predicates cannot have the same name even if they have different parameters. Each predicate in PDDL must have a unique name, and its parameters must be explicitly defined in the predicate definition. It is recommended to define predicate names in an intuitive and readable way. \n→ → → → → → → → → → → → →\nA natural language description of the action Action: This action enables the robot to close a small receptacle receptacle_1 that is openable (e.g. small storage boxes or lunch boxes with lids). For example, the robot closes lunch_box_2, or the robot closes storage_box_1."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"The dynamically updated list of predicates You can create and define new predicates, but you may also reuse the following predicates: 1. (robot-at ?x -furnitureAppliance): true if the robot is at the furniture or appliance ?x 2. (object-on ?x -householdObject ?y -furnitureAppliance): true if the object ?x is on or in the furniture or appliance ?y → 3. (pickupable ?x -householdObject): true if the object ?x can be picked up by the robot 4. (closed-receptacle ?x -smallReceptacle): true if the small receptacle ?x is closed ..."},{"figure_ref":[],"heading":"Parameters:","publication_ref":[],"table_ref":[],"text":"A.6.2 Navigate to a furniture piece or an appliance"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to navigate from one normally immovable piece of furniture to another (e.g., dining tables, side tables, cabinets, and sinks) or an appliance (e.g., refrigerators, coffee makers, microwaves, and washers)."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"GPT-4: Navigate to a furniture piece or an appliance Parameters:\n1. ?x -furnitureAppliance: the furniture or appliance the robot is currently at 2. ?y -furnitureAppliance: the furniture or appliance the robot wants to navigate to Preconditions:\n(and (robot-at ?x) (not (= ?x ?y)) )\nEffects: (and (not (robot-at ?x)) (robot-at ?y) )"},{"figure_ref":[],"heading":"A.6.3 Pick up an object on or in a furniture piece or an appliance","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to pick up an object object_1 in/on a large piece of furniture (e.g., dining tables, drawers, cabinets) or an appliance (e.g., dishwashers and refrigerators). For example, the robot pick up coffee_mug_1 on dining_table_1, the robot pick up fork_1 in drawer_1, or the robot pick up apple_1 in refrigerator_1."},{"figure_ref":[],"heading":"→ → →","publication_ref":[],"table_ref":[],"text":"Additional information from the user GPT-4: Pick up an object on or in a furniture piece or an appliance Parameters:\n1. ?x -householdObject: the object to pick up 2. ?y -furnitureAppliance: the furniture or appliance where the object is located Preconditions:\n(and (robot-at ?y) (object-on ?x ?y) (pickupable ?x) Inaccurate precondition: \"?x is not stacked on top of other object\" should only apply when ?x is a stackable object (not (object-stacked ?x)) (robot-hand-empty) (or (not (openable ?y)) (opened ?y) ) Missing precondition: there should be no other object stacked on top of ?x if ?x is a stackable object )"},{"figure_ref":[],"heading":"Effects: (and","publication_ref":[],"table_ref":[],"text":"(not (object-on ?x ?y)) (robot-holding ?x) (not (robot-hand-empty)) )\nA.6.4 Put an object on or in a furniture piece or an appliance"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to put an object object_1 in/on a large piece of furniture (e.g., dining tables, drawers, cabinets) or an appliance (e.g., dishwashers and refrigerators). For example, the robot puts coffee_mug_1 on dining_table_1, the robot puts fork_1 in drawer_1, or the robot puts apple_1 in refrigerator_1."},{"figure_ref":[],"heading":"→ → →","publication_ref":[],"table_ref":[],"text":"Additional information from the user (a) The furniture piece or appliance should be opened if it is openable."},{"figure_ref":[],"heading":"GPT-4: Put an object on or in a furniture piece or an appliance","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?x -householdObject: the object to put in/on the furniture or appliance 2. ?y -furnitureAppliance: the furniture or appliance to put the object in/on Preconditions:\n(and (robot-at ?y) (robot-holding ?x) (pickupable ?x)\nUnnecessary precondition \"no other object stacked on object ?x\" (object-clear ?x) (or (not (openable ?y)) (opened ?y) ) )\nEffects: (and (not (robot-holding ?x)) (robot-hand-empty) (object-on ?x ?y) Syntax error, there is an illegal keyword \"if\" (if (openable ?y) (closed ?y)) )"},{"figure_ref":[],"heading":"A.6.5 Stack Objects","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to stack one object object_1 on top of another object object_2 that is on the same piece of furniture. For example, the robot stacks book_1 on top of book_2, or the robot stacks lunch_box_1 on top of lunch_box_2."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"Additional information from the user "},{"figure_ref":[],"heading":"GPT-4: Stack objects","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?x -householdObject: the object to stack on top of another object 2. ?y -householdObject: the object on which ?x will be stacked 3. ?z -furnitureAppliance: the furniture or appliance on which both objects are placed Preconditions:\n(and (robot-holding ?x) (object-on ?y ?z) (object-clear ?y) (flat-surface ?z) Incorrect precondition: should be stackable (pickupable ?x) Incorrect precondition: should be stackable (pickupable ?y) (robot-at ?z) )"},{"figure_ref":[],"heading":"Effects: (and","publication_ref":[],"table_ref":[],"text":"(not (robot-holding ?x)) (not (object-clear ?y)) (robot-hand-empty) (object-on ?x ?z) (object-stacked ?x) Missing effect: object ?x should be stacked on top of object ?y )"},{"figure_ref":[],"heading":"A.6.6 Unstack Objects","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to unstack one object object_1 that is on top of another object object_2. For example, the robot stacks book_1 on top of book_2, or the robot stacks lunch_box_1 on top of lunch_box_2. The robot will hold the object after unstacking it."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"GPT-4: Unstack objects","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?x -householdObject: the object to unstack from another object 2. ?y -householdObject: the object that ?x is stacked on 3. ?z -furnitureAppliance: the furniture or appliance where the objects are stacked Preconditions:\n(and (robot-at ?z) (object-on ?x ?z) (object-on ?y ?z) (pickupable ?x) (object-stacked ?x) (object-clear ?x) (robot-hand-empty) Missing precondition: object ?x should be stacked on top of object ?y ) Effects: (and\n(not (object-on ?x ?z)) (not (object-stacked ?x)) Incorrect effect, this action shouldn't change whether or not object ?x is under any another household object (not (object-clear ?x)) (not (robot-hand-empty)) (robot-holding ?x) (object-clear ?y) Missing effect: object ?x is no longer stacked on top of object ?y )\nA.6.7 Open a furniture piece or an appliance"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to open a large piece of furniture (e.g., cabinets and drawers) or an appliance (e.g., dishwashers and refrigerators) that is openable. For example, the robot opens cabinet_1, or the robot opens drawer_2."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"GPT-4: Open a furniture piece or an appliance Parameters:\n1. ?y -furnitureAppliance: the furniture or appliance to be opened Preconditions:\n(and (robot-at ?y) (openable ?y) (closed ?y) (robot-hand-empty) )\nEffects: (and (not (closed ?y)) (opened ?y) )\nA.6.8 Close a furniture piece or an appliance"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to close a large piece of furniture (e.g., cabinets and drawers) or an appliance (e.g., dishwashers and refrigerators) that is openable. For example, the robot closes cabinet_1, or the robot closes drawer_2. A.6.9 Toggle a small appliance on"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to toggle a small appliances (like humidifiers and light bulbs) to switch them on. For example, the robot toggles humidifier_2 on, or the robot toggle light_bulb_1 on.\n→ →"},{"figure_ref":[],"heading":"GPT-4: Toggle a small appliance on","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?x -householdObject: the small appliance to toggle on 2. ?y -furnitureAppliance: the furniture or appliance where the small appliance is located Preconditions:\n(and (robot-at ?y) (object-on ?x ?y) (robot-hand-empty) (not (appliance-on ?x)) Missing precondition: the appliance to toggle should be a toggleable appliance )"},{"figure_ref":[],"heading":"Effects: (and (appliance-on ?x) )","publication_ref":[],"table_ref":[],"text":"A.6.10 Toggle a small appliance off"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to toggle a small appliances (like humidifiers and light bulbs) to switch them off. For example, the robot toggles humidifier_2 off, or the robot toggle light_bulb_1 off.\n→ →"},{"figure_ref":[],"heading":"GPT4: Toggle a small appliance off","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?r -robot: the robot performing the action 2. ?o -householdObject: the small appliance to be toggled off 3. ?f -furnitureAppliance: the furniture or appliance where the small appliance is located Preconditions:\n(and (robot-at ?r ?f) (object-in-on-furniture ?o ?f) (robot-gripper-empty ?r) (appliance-on ?o) ) Effects:\n(and\n(not (appliance-on ?o)) )"},{"figure_ref":[],"heading":"A.6.11 Slice an object","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to slice objects (like fruits and vegetables) with a knife. For example, the robot slices orange_1 with knife_2, the robot slices potato_3 with knife_1.\n→ A.6.14 Transfer food from one small receptacle to another"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to transfer food food_1 from one small receptacle receptacle_1 to another small receptacle receptacle_2. This skill is usually used to transfer food that can not be easily picked up by the robot. For example, the robot can pour the milk from milk_box_1 to mug_1, or the robot can transfer heated pizza_1 from pan_2 to plate_1."},{"figure_ref":[],"heading":"→ → →","publication_ref":[],"table_ref":[],"text":"Additional information from the user GPT-4: Transfer food from one small receptacle to another Parameters:\n1. ?food -householdObject: the food to be transferred 2. ?receptacle1 -smallReceptacle: the small receptacle from which the food will be transferred 3. ?receptacle2 -smallReceptacle: the small receptacle to which the food will be transferred 4. ?furniture -furnitureAppliance: the furniture piece with a flat surface where the transfer will take place → Preconditions:\n(and (robot-at ?furniture) (object-on ?receptacle1 ?furniture) (object-on ?receptacle2 ?furniture) (object-in ?food ?receptacle1) (is-small-receptacle ?receptacle1) (is-small-receptacle ?receptacle2) (flat-surface ?furniture) (robot-hand-empty) Inaccurate precondition: \"?receptacle1 is not stacked on other objects\" should only apply if ?receptacle1 is a stackable object (not (object-stacked ?receptacle1)) Inaccurate precondition: \"?receptacle2 is not stacked on other objects\" should only apply if ?receptacle2 is a stackable object (not (object-stacked ?receptacle2)) (or (not (openable-receptacle ?receptacle1)) (opened-receptacle ?receptacle1)) (or (not (openable-receptacle ?receptacle2)) (opened-receptacle ?receptacle2)) Missing precondition: there should be no other object stacked on top of ?receptacle1 if ?receptacle1 is stackable Missing precondition: there should be no other object stacked on top of ?receptacle2 if ?receptacle2 is stackable ) Effects: (and (not (object-in ?food ?receptacle1)) (object-in ?food ?receptacle2) )\nA.6.15 Put an object onto or into a small receptacle like a bowl and plate"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to put an object object_1 into/onto a small receptacle receptacle_1 (e.g. storage boxes, bowls, plates, or pans). For example, the robot puts apple_1 into blow_1, or the robot puts apple_2 into lunch_box_2."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"Additional information from the user (flat-surface ?y) (robot-hand-empty) Incorrect precondition: ?x can't be stacked inside a receptacle so the precondition `?x is not under other household object`is unnecessary (object-clear ?x) Incorrect precondition: ?z can be an openable receptacle (not (openable-receptacle ?z)) (or Syntax error, the first parameter of `openable`should be a furnitureAppliance but here a smallReceptacle is given (not (openable ?z)) (opened-receptacle ?z) ) Missing precondition: there should be no other object stacked on top of ?z if ?z is stackable ) Effects: (and\n(not (object-on ?x ?y)) (not (object-in ?x ?z)) (not (robot-hand-empty)) (robot-holding ?x) )\nA.6.17 Open a small receptacle such as a lunch box with a lid"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to open a small receptacle receptacle_1 (e.g. small storage boxes or lunch boxes with lids). For example, the robot opens lunch_box_2, or the robot opens storage_box_1."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"Additional information from the user (a) This action is only applicable for receptacles that are openable. (b) The receptacle needs to be placed on a furniture piece that has an open and flat surface to allow enough space for manipulation (e.g., countertops or tables)."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"(c) The receptacle should not be stacked on top of other objects if it is stackable.\nGPT-4: Open a small receptacle such as a lunch box with a lid Parameters:\n1. ?x -smallReceptacle: the small receptacle to be opened 2. ?y -furnitureAppliance: the furniture or appliance on which the small receptacle is placed Preconditions:\n(and (robot-at ?y) (object-on ?x ?y) (openable-receptacle ?x) (flat-surface ?y) (robot-hand-empty) Inaccurate precondition: should only apply if ?x is stackable (not (object-stacked ?x)) (closed-receptacle ?x) Missing precondition: there should be no other object stacked on top of ?x ) Effects: (and (not (closed-receptacle ?x)) (opened-receptacle ?x) ) (object-in ?f ?b) (object-sliced ?f) (is-small-receptacle ?b) Incorrect precondition: as mentioned in the prompt, the blender should be turned off at the beginning (appliance-on ?b) (robot-hand-empty) Missing precondition: ?b should be a blender ) Effects: (and\n(not (appliance-on ?b)) (object-mashed ?f) Missing effect: ?f is no longer pickupable )"},{"figure_ref":[],"heading":"A.6.20 Wash an object","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action Description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to wash an object (e.g., fruits and cloths) in a sink or basin. For example, the robot washed apple_1, or the robot washed romaine_lettuce_1."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"Additional information from the user (a) The object has to be something washable.\n(b) The robot should hold the object when washing it."},{"figure_ref":[],"heading":"GPT-4: Wash an object","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?x -householdObject: the object to be washed 2. ?y -furnitureAppliance: the sink or basin where the object will be washed Preconditions:\n(and (robot-at ?y) (washable ?x) (robot-holding ?x) Missing precondition: ?y should be a sink or basin ) Effects: (and (not (object-dirty ?x)) (object-washed ?x) )"},{"figure_ref":[],"heading":"A.6.21 Wipe a surface","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to wipe and clean the surface of a piece of furniture or an appliance, such as a dining table, a mirror, a sink, or a bathtub, with a cloth. For example, the robot wipes dining_table_1 with cloth_1, the robot wipes mirror_1 with cloth_2."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"Additional information from the user "},{"figure_ref":[],"heading":"GPT-4: Wipe a surface","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?x -furnitureAppliance: the furniture or appliance to be wiped and cleaned 2. ?c -householdObject: the cloth to be used for wiping and cleaning Preconditions:\n(and (robot-at ?x) (robot-holding ?c) (washable ?c) (object-washed ?c) Incorrect precondition: ?x doesn't need to have a flat surface (flat-surface ?x) Missing precondition: ?c should be a cloth ) Effects: (and (not (object-washed ?c)) (object-dirty ?c) (furniture-cleaned ?x) )"},{"figure_ref":[],"heading":"A.6.22 Vacuum a carpet","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to vacuum a carpet with a handheld vacuum cleaner. For example, the robot uses vacuum_cleaner_1 to vacuum carpet_2, the robot uses vacuum_cleaner_2 to vacuum carpet_1."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"Additional information from the user (a) You need to make sure the dust bin of the vacuum cleaner is not full before executing this action.\n→ (b) You may assume the robot is holding the vacuum cleaner at the beginning. (c) The dust bin of the vacuum cleaner will be full of dust."},{"figure_ref":[],"heading":"GPT4: Vacuum a carpet","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?v -householdObject: the handheld vacuum cleaner to use 2. ?c -furnitureAppliance: the carpet to vacuum Preconditions:\n(and (robot-holding ?v) (robot-at ?c) (pickupable ?v) (not (dustbin-full ?v)) (not (carpet-clean ?c)) Missing precondition: ?v should be a handheld vacuum cleaner Missing precondition: ?c should be a carpet ) Effects: (and (carpet-clean ?c) (dustbin-full ?v) )"},{"figure_ref":[],"heading":"A.6.23 Empty a vacuum cleaner","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to empty a vacuum cleaner's dust bin and dump the dust into a trash can. For example, the robot empties vacuum_cleaner_1 and dumps the dust into trash_can_2, or the robot empties vacuum_cleaner_2 and dumps the dust into trash_can 1."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"2. There is a syntax error: the second parameter of `object-on`should be a furnitureAppliance, but a householdObject was given. Please use the correct predicate or devise new one(s) if needed → →\n3. Incorrect precondition: ?f needs not be pickupable 4. There is a missing precondition: ?p has to be a pan 5. There is a missing precondition: ?s has to be a stove burner 6. There is an incorrect precondition: a stove or burner ?s doesn't need to have a flat surface Please revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors)."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"Parameters:\n----------------------------------------GPT-4:\n1. ?f -householdObject: the food to be heated 2. ?p -smallReceptacle: the pan to be used for heating the food 3. ?s -furnitureAppliance: the stove burner on which the pan is placed Preconditions:\n(and (robot-at ?s) (object-in ?f ?p) (object-on ?p ?s) (pan ?p) (stove-burner ?s) (robot-hand-empty) )\nEffects: (and (not (pickupable ?f)) (food-heated-pan ?f) ) New Predicates: 1. (pan ?x -smallReceptacle): true if the small receptacle ?x is a pan 2. (stove-burner ?x -furnitureAppliance): true if the furniture or appliance ?x is a stove burner 3. (object-in ?x -householdObject ?y -smallReceptacle): true if the object ?x is in the small receptacle ?y\n→ ---------------------------------------- Automated corrective feedback:\nThe following predicate(s) have the same name(s) as existing predicate(s):\n1. (object-in ?x -householdObject ?y -smallReceptacle), true if the object ?x is in the small receptacle ?y; existing predicate with the same name: (object-in ?x -householdObject ?z -smallReceptacle), true if the object ?x is in the small receptacle ?z\n→ →\nYou should reuse existing predicates whenever possible. If you are reusing existing predicate(s), you shouldn't list them under 'New Predicates'. If existing predicates are not enough and you are devising new predicate(s), please use names that are different from existing ones."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"Please revise the PDDL model to fix this error."},{"figure_ref":[],"heading":"Parameters:","publication_ref":[],"table_ref":[],"text":"----------------------------------------GPT-4: 1. ?f -householdObject: the food to be heated 2. ?p -smallReceptacle: the pan to be used for heating the food 3. ?s -furnitureAppliance: the stove burner on which the pan is placed Preconditions:\n(and (robot-at ?s) (object-in ?f ?p) (object-on ?p ?s) (pan ?p) (stove-burner ?s) (robot-hand-empty) )\nEffects: (and (not (pickupable ?f)) (food-heated-pan ?f) ) New Predicates: 1. (pan ?x -smallReceptacle): true if the small receptacle ?x is a pan 2. (stove-burner ?x -furnitureAppliance): true if the furniture or appliance ?x is a stove burner"},{"figure_ref":[],"heading":"A.8 Logistics: Constructing PDDL Models","publication_ref":[],"table_ref":[],"text":"The AI agent here is a logistics planner that has to plan to transport packages within the locations in a city through a truck and between cities through an airplane. Within a city, the locations are directly linked, allowing trucks to travel between any two of these locations. Similarly, cities are directly connected to each other allowing airplanes to travel between any two cities. Each city is equipped with one truck and has a designated location that functions as an airport. There are five types of objects: package, truck, plane, location, and city. There are multiple cities and each city can have multiple locations. Also, there is no limit to how many packages a truck or plane can carry (so in theory a truck or plane can carry an infinite number of packages)."},{"figure_ref":[],"heading":"A.8.1 Load a package into a truck","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the agent to load a package into a truck. For example, load a package_1 into a truck_1. "},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the agent to unload a package from a truck. For example, unload a package_1 from a truck_1.\n→ GPT-4: Unload a package from a truck Parameters:\n1. ?p -package: the package to unload 2. ?t -truck: the truck from which the package will be unloaded 3. ?l -location: the location where the package will be unloaded Preconditions:\n(and (package-in-truck ?p ?t) (truck-at ?t ?l) )\nEffects: (and (not (package-in-truck ?p ?t)) (package-at ?p ?l) )"},{"figure_ref":[],"heading":"A.8.3 Load a package into an airplane","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the agent to load a package into an airplane. For example, load a package_1 into an airplane_1.\n→ GPT-4: Load a package into an airplane Parameters:\n1. ?p -package: the package to load into the airplane 2. ?a -plane: the airplane to load the package into 3. ?l -location: the location where the package and airplane are present Preconditions:\n(and (package-at ?p ?l) (plane-at ?a ?l) (not (package-in-plane ?p ?a)) )\nEffects: (and (not (package-at ?p ?l)) (package-in-plane ?p ?a) )"},{"figure_ref":[],"heading":"A.8.4 Unload a package from an airplane","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the agent to unload a package from an airplane. For example, unload a package_1 from an airplane_1.\n→ GPT-4: Unload a package from an airplane Parameters:\n1. ?p -package: the package to unload 2. ?a -plane: the airplane from which the package will be unloaded 3. ?l -location: the location where the airplane is and the package will be unloaded Preconditions:\n(and (package-in-plane ?p ?a) (plane-at ?a ?l) )\nEffects: (and (not (package-in-plane ?p ?a)) (package-at ?p ?l) )\nA.8.5 Drive a truck from one location to another in a city"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the agent to drive a truck from one location to another in a city. For example, drive a truck_1 from location_1 to location_2 in city_1.\n→ GPT-4: Drive a truck from one location to another in a city Parameters:\n1. ?t -truck: the truck to be driven 2. ?from -location: the starting location of the truck 3. ?to -location: the destination location of the truck 4. ?c -city: the city in which the truck is being driven Preconditions:\n(and (truck-at ?t ?from) (location-in-city ?from ?c) (location-in-city ?to ?c) (not (= ?from ?to)) )\nEffects: (and (not (truck-at ?t ?from)) (truck-at ?t ?to) (move-packages ?t ?from ?to) )\nA.8.6 Fly an airplane from one city to another"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the agent to fly an airplane from one city's airport to another. The airports are locations in the city. For example, fly an airplane_1 from location_0 to location_1.\n→ GPT-4: Fly an airplane from one city to another Parameters:\n1. ?a -plane: the airplane to fly 2. ?from -location: the departure location (airport) in the origin city 3. ?to -location: the destination location (airport) in the destination city 4. ?c1 -city: the origin city 5. ?c2 -city: the destination city Preconditions:\n(and (plane-at ?a ?from) (location-in-city ?from ?c1) (location-in-city ?to ?c2) (not (= ?c1 ?c2)) Missing precondition: ?from should be an airport Missing precondition: ?to should be an airport ) Effects: (and (not (plane-at ?a ?from)) (plane-at ?a ?to) (move-packages-plane ?a ?from ?to) )\nA.8.7 Examples of PDDL action models constructed by GPT-3.5-Turbo GPT-3.5-Turbo: Fly an airplane from one city to another Parameters: 1. ?a -plane: the airplane to fly 2. ?c1 -city: the city to fly from 3. ?c2 -city: the city to fly to 4. ?l1 -location: the airport location in ?c1 A.8.8 The initial set of predicates extracted by GPT-4\nThe initial set of predicates extracted by GPT-4\n1. (package-at ?p -package ?l -location): true if the package ?p is at the location ?l 2. (truck-at ?t -truck ?l -location): true if the truck ?t is at the location ?l 3. (package-in-truck ?p -package ?t -truck): true if the package ?p is in the truck ?t 4. (plane-at ?a -plane ?l -location): true if the airplane ?a is at the location ?l 5. (location-in-city ?l -location ?c -city): true if the location ?l is in the city ?c 6. (package-in-plane ?p -package ?a -plane): true if the package ?p is in the airplane ?a 7. (move-packages ?t -truck ?from -location ?to -location): true if all packages in the truck ?t are moved from location ?from to location ?to. This predicate is used to represent the effect of moving all packages in the truck when the truck is driven from one location to another."},{"figure_ref":[],"heading":"→ → →","publication_ref":[],"table_ref":[],"text":"8. (move-packages-plane ?a -plane ?from -location ?to -location): true if all packages in the airplane ?a are moved from location ?from to location ?to. This predicate is used to represent the effect of moving all packages in the airplane when the airplane is flown from one location to another → → →\nA.9 Logistics: Correcting PDDL Models A.9.1 Fly an airplane from one city to another GPT-4: Fly an airplane from one city to another User: ... The original prompt for PDDL construction ... Parameters:\n-------------------------------------GPT-4: Parameters:\n1. ?a -plane: the airplane to fly 2. ?from -location: the departure location (airport) in the origin city 3. ?to -location: the destination location (airport) in the destination city 4. ?c1 -city: the origin city 5. ?c2 -city: the destination city Preconditions:\n(and (plane-at ?a ?from)\nMissing precondition: ?from should be an airport Missing precondition: ?to should be an airport ) Effects: (and (not (plane-at ?a ?from)) (plane-at ?a ?to) (move-packages-plane ?a ?from ?to) ) --------------------------------------- ----------------------------------------GPT-3.5-Turbo:\n1. ?a -plane: the airplane to fly 2. ?from -location: the departure location (airport) in the origin city 3. ?to -location: the destination location (airport) in the destination city 4. ?c1 -city: the origin city 5. ?c2 -city: the destination city Preconditions:\n(and (plane-at ?a ?from) (location-in-city ?from ?c1) (location-in-city ?to ?c2) (not (= ?c1 ?c2)) (airport ?to) Missing precondition: ?from should be an airport ) Effects: (and (not (plane-at ?a ?from)) (plane-at ?a ?to) (move-packages-plane ?a ?from ?to) ) New Predicates: 1. (airport ?l -location): true if the location ?l is an airport."},{"figure_ref":[],"heading":"A.10 Tyreworld: Constructing PDDL Models","publication_ref":[],"table_ref":[],"text":"The AI agent here is a robot that has to replace a flat tyre with a spare one. This involves fetching the tools (i.e., wrench, jack, pump) from the boot, undoing the nuts on the flat tyre, jacking up the (appropriate) hub(s), removing the tyre, doing up the spare one, etc. There are three major object types: 'small_object', 'container' and 'hub'. The object type 'small_object' covers tools, wheels and nuts. The 'small_object' object type has three subtypes: 'tool', 'wheel' and 'nut'. The subtype 'tool' covers tools like wrenches, jacks, pumps etc. The subtype 'wheel' covers various wheels. The subtype 'nut' covers various nuts. The object type 'container' covers storage spaces like the boot in a car. The object type 'hub' covers the hubs in the wheels of the car. Note that there is no restriction on how many objects the AI agent (i.e., the robot) can carry. Also note that each hub has only one nut."},{"figure_ref":[],"heading":"A.10.1 Open a container","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to open a container. For example, the robot can open a boot. "},{"figure_ref":[],"heading":"GPT-4: Open a container","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to jack up a hub. For example, the robot can jack up hub_1."},{"figure_ref":[],"heading":"GPT-4: Jack up a hub","publication_ref":[],"table_ref":[],"text":"Parameters: 1. ?h -hub: the hub to jack up 2. ?j -small_object: the jack to use for jacking up the hub Preconditions: (and (is-hub ?h) (hub-on-ground ?h) (is-jack ?j) (robot-holding ?j) )\nEffects: (and\n(not (hub-on-ground ?h)) (hub-jacked-up ?h) )"},{"figure_ref":[],"heading":"A.10.8 Jack down a hub","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to jack down a hub. For example, the robot can jack down hub_1."},{"figure_ref":[],"heading":"GPT-4: Jack down a hub","publication_ref":[],"table_ref":[],"text":"Parameters: 1. ?h -hub: the hub to jack down 2. ?j -small_object: the jack used to jack down the hub Preconditions: (and (is-hub ?h) (is-jack ?j) (robot-holding ?j) (hub-jacked-up ?h) )\nEffects: (and (not (hub-jacked-up ?h)) (hub-on-ground ?h) )\nA.10.9 Unfasten a hub"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to unfasten a hub and remove the nut on it. For example, the robot can unfasten hub_1 that has nut_1. At the end of this action, the hub should be unfastened, and the nut should be detached from the hub and held by the robot. To safely perform this action, the hub needs to be jacked up."},{"figure_ref":[],"heading":"→ → →","publication_ref":["b11"],"table_ref":[],"text":"(robot-holding ?w) (robot-holding ?p) (wheel-deflated ?w) (wheel-intact ?w) )\nEffects: (and (not (wheel-deflated ?w)) (wheel-inflated ?w) )\nA.12 LLM planners back-prompted by VAL using LLM-acquired PDDL model A. 12 Here is the task. The house contains the following large furniture pieces and appliances: drawer_1, drawer_2, ..., dish_washer_1, garbage_can_1. Note that cutting boards are viewed as small receptacles, and blenders are small appliances (and thus small receptacles)."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"The current states of the household objects are as follows:\n1. drawer_1 is opened 2. drawer_2 is opened ..."},{"figure_ref":[],"heading":"robot is holding mug_1","publication_ref":[],"table_ref":[],"text":"Your task is to: put apple_2 on side_table_2. Your plan is:\nAn example prompt for the Logistics domain\nYou are a logistics planner that has to plan to transport packages within the locations in a city through a truck and between cities through an airplane. Within a city, the locations are directly linked, allowing trucks to travel between any two of these locations. Similarly, cities are directly connected to each other allowing airplanes to travel between any two cities. Each city is equipped with one truck and has a designated location that functions as an airport. There are five types of objects: package, truck, plane, location, and city. There are multiple cities and each city can have multiple locations. Also, there is no limit to how many packages a truck or plane can carry (so in theory a truck or plane can carry an infinite number of packages).\n→ → → → → → → →\nThe following actions are available to you, and you should strictly follow the output format (demonstrated in the \"example output\") of each action: → Preconditions: (a) the package ?p is at the location ?l; (b) the truck ?t is at the location ?l. Effects: (a) the package ?p is not at the location ?l; (b) the package ?p is in the truck ?t.\n2. unload a package ?p from a truck ?t at location ?l (example output: unload package_1 from a truck truck_1 at location_1).\n→ Parameters: (a) ?p: the package to unload; (b) ?t: the truck from which the package will be unloaded; (c) ?l: the location where the package will be unloaded."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"Preconditions: (a) the package ?p is in the truck ?t; (b) the truck ?t is at the location ?l. Effects: (a) the package ?p is not in the truck ?t; (b) the package ?p is at the location ?l. ...\nHere are two examples."},{"figure_ref":[],"heading":"Example 1","publication_ref":[],"table_ref":[],"text":"The environment contains the following objects: plane_0 (type: plane), city_0 (type: city), truck_0 (type: truck), location_0 (type: location), location_1 (type: location), package_0 (type: package)."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"The current states of the objects are as follows:\n1. location_0 is in city city_0 2. location_1 is in city city_0 3. truck_0 is at location location_0 4. plane_0 is at location location_0 5. package_0 is at location location_1 6. location_0 is an airport Your task is to: Transport package package_0 to location location_0 Your plan is:\nStep 1: > think: First, I need to drive the truck to the location of the package and load the package. > action: drive truck_0 from location_0 to location_1 in city_0\nStep 2: > action: load package_0 into a truck truck_0 at location_1\nStep 3: > think: Now I need to drive the truck to the destination and unload the package. > action: drive truck_0 from location_1 to location_0 in city_0\nStep 4: > action: unload package_0 from a truck truck_0 at location_0 Example 2 ...\nHere is the task. The environment contains the following objects: plane_0 (type: plane), city_0 (type: city), city_1 (type: city), truck_0 (type: truck), truck_1 (type: truck), location_0 (type: location), location_1 (type: location), location_2 (type: location), package_0 (type: package)\nA.12.4 Examples of validation feedback and instructions with ordering constraints Make heated mashed potato_1 & Remember to mash the potato at the end after heating it User: ... The original prompt for the LLM planner ... The current states of the household objects are as follows: ... 70. robot at dining_table_1 71. robot is holding mug_1\nYour task is to: make heated and mashed potato_1, remember to mash the potato at the end after heating it."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"Your plan is:\n-------------------------------------GPT-4:\nStep 1: > think: First, I need to find a cutting board and put it on a flat surface (countertop_1 or countertop_2). In the current state, cutting_board_1 is in/on countertop_2.\n→ > action: go from dining_table_1 to countertop_2\nStep 2: > think: Now I need to find potato_1. In the current state, potato_1 is in fridge_1 and fridge_1 is closed. I need to get the potato and then put it on the cutting board.\n→ > action: go from countertop_2 to fridge_1\nStep 3:\n> action: open furniture/appliance fridge_1\nStep 4: > action: pick up potato_1 in/on fridge_1\nStep 5: > action: go from fridge_1 to countertop_2\nStep 6: > action: put potato_1 in/on cutting_board_1 at countertop_2\nStep 7: > think: Now I need to find a knife. In the current state, knife_1 is in cabinet_3 and cabinet_3 is closed. I need to get the knife and then slice the potato.\n→ > action: go from countertop_2 to cabinet_3\nStep 8:\n> action: open furniture/appliance cabinet_3\nStep 9: > action: pick up knife_1 in/on cabinet_3\nStep 10: > action: go from cabinet_3 to countertop_2\nStep 11: > action: slice potato_1 with knife_1 and cutting_board_1 at countertop_2\nStep 12: > think: Now I need to heat the sliced potato. In the current state, microwave_1 is closed. I need to put the sliced potato in a microwave-safe container and then heat it in the microwave. Step 24: > action: go from microwave_1 to countertop_1\nStep 25: > think: Now I need to find a blender to mash the heated potato. In the current state, blender_1 is in/on countertop_1 and blender_1 is turned off. I need to put the heated potato in the blender and then mash it."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"> action: transfer potato_1 from bowl_2 to blender_1 at countertop_1\nStep 26: > action: mash potato_1 with blender_1 at countertop_1\nStep 27: > think: Task complete. The heated mashed potato_1 is now in the blender.\n-------------------------------------Validation feedback:\nThe action at step 3 is not executable due to unmet precondition(s). Here are the unsatisfied precondition(s):\n→ The robot's gripper is not holding any object.\nYour revised plan is:\n-------------------------------------GPT-4: ... The revised plan ...\nPlace fork_1 spoon_1 and knife_1 on dining_table_1 & Please take the fork first and the knife last User: ... The original prompt for the LLM planner ... The current states of the household objects are as follows: ... 24. lunch_box_2 in/on fridge_1 25. lunch_box_2 is not opened ... 59. spoon_1 in/on lunch_box_2 ... Your task is to: place fork_1, spoon_1, and knife_1 on dining_table_1, please take the fork first and the knife last."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"Your plan is:\n-------------------------------------GPT-4:\nStep 1: > think: First, I need to put down the mug_1 I am currently holding. > action: go from dining_table_1 to side_table_1\nStep 2: > action: put mug_1 in/on side_table_1\nStep 3: > think: Now, I need to pick up fork_1 from dish_washer_1. > action: go from side_table_1 to dish_washer_1\nStep 4:\n> action: open furniture/appliance dish_washer_1\nStep 5: > action: pick up fork_1 in/on dish_washer_1\nStep 6: > think: I have the fork_1, now I need to place it on dining_table_1. > action: go from dish_washer_1 to dining_table_1\nStep 7: > action: put fork_1 in/on dining_table_1\nStep 8: > think: Next, I need to pick up spoon_1 from lunch_box_2 in fridge_1. > action: go from dining_table_1 to fridge_1\nStep 9:\n> action: open furniture/appliance fridge_1\nStep 10: > action: pick up spoon_1 in/on lunch_box_2 at fridge_1\nStep 11: > think: I have the spoon_1, now I need to place it on dining_table_1. > action: go from fridge_1 to dining_table_1\nStep 12: > action: put spoon_1 in/on dining_table_1\nStep 13: > think: Finally, I need to pick up knife_1 from cabinet_1. > action: go from dining_table_1 to cabinet_1\nStep 14: > action: pick up knife_1 in/on cabinet_1\nStep 15: > think: I have the knife_1, now I need to place it on dining_table_1. > action: go from cabinet_1 to dining_table_1\nStep 16: > action: put knife_1 in/on dining_table_1\nStep 17:"},{"figure_ref":[],"heading":"Acknowledgement","publication_ref":[],"table_ref":[],"text":"This research was supported by ONR grants N00014-18-1-2442, N00014-18-1-2840, N00014-19-1-2119 and N00014-23-1-2409, AFOSR grant FA9550-18-1-0067, DARPA SAIL-ON grant W911NF-19-2-0006, and a JP Morgan AI Faculty Research Grant to Kambhampati. Sreedharan was supported in part by NSF grant 2303019."},{"figure_ref":[],"heading":"A.1 Broader impact on using LLMs","publication_ref":["b22","b51"],"table_ref":[],"text":"There is a general temptation to use LLMs for a variety of tasks, including plan generation. Given the fact that LLMs cannot guarantee the generation of correct plans, this can lead to safety and security issues downstream. Our approach of teasing a domain model from LLMs, and using it in conjunction with external sound planners aims to mitigate these safety concerns. Nevertheless, given that humans are still in charge of verifying the correctness of the domain models extracted from LLMs, there is still the possibility that an incorrect or undesirable domain model is inadvertently certified correct, leading to undesirable plans and agent behaviors down the line.\nImproved explainability is another advantage of extracting explicit domain models. As opposed to just directly querying LLMs for plans, generating behaviors with (intermediate) symbolic models offers additional opportunities for explanation, drawing upon existing works in explainable AI [23]. In cases where the user's understanding does not align with the transcribed model, debugging and model reconciliation techniques such as D3wa+ [52] can also be directly applied."},{"figure_ref":[],"heading":"A.2 Additional discussion on alternative to action-by-action PDDL construction","publication_ref":[],"table_ref":[],"text":"One alternative to our action-by-action generation could be to include descriptions of all the actions in the prompt and require the LLM to construct the entire domain model in a single dialogue. This Additional information from the user (a) Note that the object to slice needs to be placed on a cutting board. A cutting board can be viewed as a special kind of receptacle. A.6.12 Heat food with a microwave"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to start a microwave and heat up the food inside. For example, the robot heats pizza_1 with microwave_2, the robot heats pasta_2 with microwave_1."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"Additional information from the user (a) The food to heat should be placed in a small receptacle (e.g., a plate or a bowl).\n(b) The robot also needs to close the door of the microwave before taking this action. (c) Note that the food is no longer pickupable after it has been heated."},{"figure_ref":[],"heading":"GPT-4: Heat food with a microwave","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?x -householdObject: the food to heat up 2. ?y -furnitureAppliance: the microwave to use 3. ?z -smallReceptacle: the small receptacle containing the food Preconditions:\n(and (robot-at ?y) Incorrect precondition: the object should be in the receptacle not directly in the microwave (object-on ?x ?y) (object-in ?x ?z) (is-small-receptacle ?z) (object-on ?z ?y) Incorrect precondition: ?x needs not be pickupable (pickupable ?x) (closed ?y) (not (food-heated ?x)) Missing precondition: ?y should be a microwave ) Effects: (and (not (pickupable ?x)) (food-heated ?x) )\nA.6.13 Heat food with pan"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to heat food which is heatable with a pan. For example, the robot heats pizza_1 with pan_2, the robot heats pasta_2 with pan_1."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"Additional information from the user (a) Note that the food should be placed on the pan, and the pan needs to be placed on a stove burner before executing this action.\n→ (b) A stove burner can be viewed as an immovable appliance, and a pan can be viewed as a small receptacle.\n→ (c) Note that the food is no longer pickupable after it has been heated."},{"figure_ref":[],"heading":"GPT-4: Heat food with pan","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?f -householdObject: the food to be heated Inaccurate object type, a pan ?p can be viewed as a small receptacle 2. ?p -householdObject: the pan to be used for heating the food 3. ?s -furnitureAppliance: the stove burner on which the pan is placed Preconditions:\n(and (robot-at ?s) Syntax error: the second parameter of `object-on`should be a furnitureAppliance, but a householdObject was given (object-on ?f ?p) (object-on ?p ?s) Incorrect precondition: ?f needs not be pickupable (pickupable ?f) (is-small-receptacle ?p) Incorrect precondition: a stove or burner ?s doesn't need to have a flat surface (flat-surface ?s) (robot-hand-empty) Missing precondition: ?p has to be a pan Missing precondition: ?s has to be a stove burner ) Effects: (and (not (pickupable ?f)) (food-heated-pan ?f) )\nGPT-4: Put an object onto or into a small receptacle like a bowl and plate Parameters: 1. ?x -householdObject: the object to put into/onto the small receptacle 2. ?y -furnitureAppliance: the furniture or appliance with a flat surface for manipulation 3. ?z -smallReceptacle: the small receptacle to put the object into/onto Preconditions: (and (robot-at ?y) Inaccurate precondition: object ?x can't be on the furnitureAppliance ?y when ?x is held by the robot (object-on ?x ?y) (pickupable ?x) Inaccurate precondition: this should only apply if ?z is stackable (not (object-stacked ?z)) (flat-surface ?y) (robot-holding ?x) (is-small-receptacle ?z) (object-on ?z ?y) Inaccurate precondition: this should only apply if ?z is stackable (object-clear ?z) Unnecessary precondition: the robot holding the object ?x already implies ?x is not in the receptacle ?z (not (object-in ?x ?z)) (or (not (openable-receptacle ?z)) (opened-receptacle ?z) ) )"},{"figure_ref":[],"heading":"Effects: (and","publication_ref":[],"table_ref":[],"text":"(not (robot-holding ?x)) (not (object-on ?x ?y)) (robot-hand-empty) (object-in ?x ?z) )\nA.6.16 Pick up an object on or in a small receptacle like a bowl and plate"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to pick up an object object_1 in some small receptacle receptacle_1 (e.g. storage boxes, lunch boxes, bowls, plates). For example, the robot picks up apple_1 in blow_1, or the robot picks up apple_2 in lunch_box_2."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":"Additional information from the user GPT-4: Pick up an object on or in a small receptacle like a bowl and plate Parameters:\n1. ?x -householdObject: the object to pick up 2. ?y -furnitureAppliance: the furniture or appliance where the object is located 3. ?z -smallReceptacle: the small receptacle containing the object Preconditions:\n(and (robot-at ?y) Incorrect precondition: the object ?x is in the receptacle ?z so ?x is not directly on the furniture or appliance ?y (object-on ?x ?y) (pickupable ?x) (object-in ?x ?z) (is-small-receptacle ?z) Inaccurate precondition: \"?z is not stacked on other objects\" should only apply if ?z is a stackable object (not (object-stacked ?z))\nA.6.18 Close a small receptacle such as a lunch box with a lid"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to close a small receptacle receptacle_1 that is openable (e.g. small storage boxes or lunch boxes with lids). For example, the robot closes lunch_box_2, or the robot closes storage_box_1. A.6.19 Mash food with a blender"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to use a blender to mash some food in it. For example, the robot gets mashed apple_2 with blender_1, or the robot gets mashed banana_1 with blender_2."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"Additional information from the user (a) The food needs to be sliced beforehand and the food should be placed inside the blender. (b) Note that the food remains in the blender after this action is performed. (c) You may also assume the blender is turned off before and after mashing the food."},{"figure_ref":[],"heading":"GPT-4: Mash food with a blender","publication_ref":[],"table_ref":[],"text":"Parameters: Incorrect object type: a blender ?b should be a smallReceptacle 1. ?b -householdObject: the blender to be used for mashing the food 2. ?f -householdObject: the food to be mashed 3. ?y -furnitureAppliance: the furniture or appliance where the blender is located Preconditions:\n(and (robot-at ?y) (object-on ?b ?y) Incorrect precondition: the food to mash should be in the blender ?b rather than directly on the furniture piece ?y (object-on ?f ?y)\nAdditional information from the user (a) Note that the robot should hold the vacuum cleaner and stand by the trash can before executing this action. A.6. 25 The initial set of predicates extracted by GPT-4\nThe initial set of predicates extracted by GPT-4 -------------------------------------GPT-4: Parameters: 1. ?f -householdObject: the food to be heated There is an inaccurate object type: a pan ?p can be viewed as a small receptacle 2. ?p -householdObject: the pan to be used for heating the food 3. ?s -furnitureAppliance: the stove burner on which the pan is placed Preconditions: (and (robot-at ?s) Syntax error: the second parameter of `object-on`should be a furnitureAppliance, but a householdObject was given (object-on ?f ?p) (object-on ?p ?s) Incorrect precondition: ?f needs not be pickupable (pickupable ?f) (is-small-receptacle ?p) Incorrect precondition: a stove or burner ?s doesn't need to have a flat surface (flat-surface ?s) (robot-hand-empty) Missing precondition: ?p has to be a pan Missing precondition: ?s has to be a stove burner ) Effects: (and (not (pickupable ?f)) (food-heated-pan ?f) ) ----------------------------------------Corrective feedback: There are some errors in the PDDL model: 1. There is an inaccurate object type, a pan ?p can be viewed as a small receptacle A.7.2 Slice an object GPT-4: Slice an object User: ... The original prompt for PDDL construction ... Parameters: -------------------------------------GPT-4: Parameters: 1. ?x -householdObject: the object to slice 2. ?k -householdObject: the knife to use for slicing 3. ?y -furnitureAppliance: the furniture or appliance with a flat surface for manipulation Incorrect object type, a cutting board should be viewed as a small receptacle as mentioned in the prompt 4. ?z -householdObject: the cutting board on which the object is placed Preconditions: (and (robot-at ?y) Incorrect precondition: ?x should be on the cutting board not directly on the furniture ?y (object-on ?x ?y) (object-on ?z ?y) (robot-holding ?k) (sliceable ?x) (cutting-board ?z) (knife ?k) (flat-surface ?y) (object-whole ?x) )"},{"figure_ref":[],"heading":"Effects: (and (not (object-whole ?x)) (object-sliced ?x) ) ----------------------------------------Corrective feedback:","publication_ref":[],"table_ref":[],"text":"There are some errors in the PDDL model: 1. There is an incorrect object type: a cutting board should be viewed as a small receptacle as mentioned in the prompt → 2. There is an incorrect precondition: ?x should be on the cutting board not directly on the furniture ?y"},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"Please revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors)."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Parameters:","publication_ref":[],"table_ref":[],"text":"----------------------------------------GPT-4: 1. ?x -householdObject: the object to slice 2. ?k -householdObject: the knife to use for slicing 3. ?y -furnitureAppliance: the furniture or appliance with a flat surface for manipulation 4. ?z -smallReceptacle: the cutting board on which the object is placed Preconditions:\n(and (robot-at ?y) (object-in ?x ?z) (object-on ?z ?y) (robot-holding ?k) (sliceable ?x) (cutting-board ?z) (knife ?k) (flat-surface ?y) (object-whole ?x) )\nEffects: (and (not (object-whole ?x)) (object-sliced ?x) ) New Predicates: No newly defined predicate A.7.3 GPT-3.5-Turbo trying to correct the action model of \"slice an object\" GPT-3.5-Turbo trying to correct the action model of \"slice an object\" User: ... The original prompt for PDDL construction ... Parameters: -------------------------------------GPT-4: Parameters: 1. ?x -householdObject: the object to slice 2. ?k -householdObject: the knife to use for slicing 3. ?y -furnitureAppliance: the furniture or appliance with a flat surface for manipulation Incorrect object type, a cutting board should be viewed as a small receptacle as mentioned in the prompt 4. ?z -householdObject: the cutting board on which the object is placed Preconditions: (and (robot-at ?y) Incorrect precondition: ?x should be on the cutting board not directly on the furniture ?y (object-on ?x ?y) (object-on ?z ?y) (robot-holding ?k) (sliceable ?x) (cutting-board ?z) (knife ?k) (flat-surface ?y) (object-whole ?x) )\nEffects: (and (not (object-whole ?x)) (object-sliced ?x) ) ----------------------------------------Corrective feedback: There are some errors in the PDDL model: 1. There is an incorrect object type: a cutting board should be viewed as a small receptacle as mentioned in the prompt → 2. There is an incorrect precondition: ?x should be on the cutting board not directly on the furniture ?y"},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"Please revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors)."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Parameters:","publication_ref":[],"table_ref":[],"text":"----------------------------------------GPT-3.5-Turbo: 1. ?x -householdObject: the object to slice 2. ?k -householdObject: the knife to use for slicing 3. ?y -furnitureAppliance: the furniture or appliance with a flat surface for manipulation 4. ?z -smallReceptacle: the cutting board on which the object is placed Preconditions:\n(and (robot-at ?y) (object-on ?z ?y) (object-in ?x ?z) (robot-holding ?k) (sliceable ?x) Should use the predicate (cutting-board ?z) (is-small-receptacle ?z) (knife ?k) (flat-surface ?y) (object-whole ?x) )\nEffects: (and (not (object-whole ?x)) (object-sliced ?x) ) (location-in-city ?from ?c1) (location-in-city ?to ?c2) (not (= ?c1 ?c2)) Missing precondition: ?from should be an airport Missing precondition: ?to should be an airport ) Effects: (and (not (plane-at ?a ?from)) (plane-at ?a ?to) (move-packages-plane ?a ?from ?to) ) --------------------------------------- ---------------------------------------GPT-4: 1. ?a -plane: the airplane to fly 2. ?from -location: the departure location (airport) in the origin city 3. ?to -location: the destination location (airport) in the destination city 4. ?c1 -city: the origin city 5. ?c2 -city: the destination city Preconditions:\n(and (plane-at ?a ?from) (location-in-city ?from ?c1) (location-in-city ?to ?c2) (not (= ?c1 ?c2)) (airport ?from) (airport ?to) )"},{"figure_ref":[],"heading":"Effects: (and","publication_ref":[],"table_ref":[],"text":"(not (plane-at ?a ?from)) (plane-at ?a ?to) (move-packages-plane ?a ?from ?to) ) New Predicates: 1. (airport ?l -location): true if the location ?l is an airport A.9.2 GPT-3.5-Turbo trying to correct the action model of \"fly an airplane from one city to another\" GPT-3.5-Turbo trying to correct the action model of \"fly an airplane from one city to another\" User: ... The original prompt for PDDL construction ... Parameters:\n- ------------------------------------GPT-4: Parameters:\n1. ?a -plane: the airplane to fly 2. ?from -location: the departure location (airport) in the origin city 3. ?to -location: the destination location (airport) in the destination city 4. ?c1 -city: the origin city 5. ?c2 -city: the destination city Preconditions:\n(and (plane-at ?a ?from) (location-in-city ?from ?c1) (location-in-city ?to ?c2) (not (= ?c1 ?c2))\nPreconditions:\n(and (container-closed ?c) )\nEffects: (and (not (container-closed ?c))\n(container-open ?c) )"},{"figure_ref":[],"heading":"A.10.2 Close a container","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to close a container. For example, the robot can close a boot. "},{"figure_ref":[],"heading":"GPT-4: Close a container","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"A.10.3 Fetch an object from a container","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to take an object from a container. For example, the robot can fetch jack from the boot or the robot can fetch wheel_1 from the boot. (not (object-in-container ?x ?c)) (robot-holding ?x) )"},{"figure_ref":[],"heading":"A.10.4 Put an object into a container","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to put an object into a container. For example, the robot can put away jack into the boot or the robot can put away wheel_1 into the boot. (not (robot-holding ?x)) (object-in-container ?x ?c) )"},{"figure_ref":[],"heading":"A.10.5 Loosen a nut in a hub","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to loosen a nut in a hub. For example, the robot can loosen nut_1 in hub_1. To safely perform this action, the hub needs to be on the ground (i.e., not jacked up)."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"GPT-4: Loosen a nut in a hub","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?n -nut: the nut to be loosened 2. ?h -hub: the hub where the nut is located 3. ?w -tool: the wrench to be used for loosening the nut Preconditions:\n(and (is-nut ?n) (is-hub ?h) (is-wrench ?w) (nut-in-hub ?n ?h) (hub-on-ground ?h) (robot-holding ?w) )\nEffects: (and (nut-loose ?n) )"},{"figure_ref":[],"heading":"A.10.6 Tighten a nut in a hub","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to tighten a nut in a hub. For example, the robot can tighten nut_1 in hub_1. To safely perform this action, the hub needs to be on the ground (i.e., not jacked up).\n→ →"},{"figure_ref":[],"heading":"GPT-4: Tighten a nut in a hub","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?n -nut: the nut to tighten 2. ?h -hub: the hub where the nut is to be tightened 3. ?w -tool: the wrench to use for tightening the nut Preconditions:\n(and (is-nut ?n) (is-hub ?h) (is-wrench ?w) (robot-holding ?w) (nut-in-hub ?n ?h) (hub-on-ground ?h) (nut-loose ?n)"},{"figure_ref":[],"heading":"GPT-4: Unfasten a hub","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?h -hub: the hub to unfasten 2. ?n -nut: the nut to detach from the hub Preconditions:\n(and (is-hub ?h) (is-nut ?n) (nut-in-hub ?n ?h) (nut-loose ?n) (hub-jacked-up ?h) Incorrect precondition: the robot should not hold the nut if the nut is in the hub (robot-holding ?n) )\nEffects: (and (not (nut-in-hub ?n ?h)) (hub-unfastened ?h) Missing effect: the robot should hold the nut at the end )"},{"figure_ref":[],"heading":"A.10.10 Fasten a hub","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to fasten a hub by installing a nut on it. For example, the robot can fasten hub_1 that has nut_1. Note that the robot needs to install a nut into the hub, but the robot is not required to fasten the nut in this operation. To safely perform this action, the hub needs to be jacked up."},{"figure_ref":[],"heading":"→ → →","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"GPT-4: Fasten a hub","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?h -hub: the hub to fasten 2. ?n -nut: the nut to install in the hub Preconditions:\n(and (is-hub ?h) (is-nut ?n) (robot-holding ?n) (hub-jacked-up ?h) (hub-unfastened ?h) Unnecessary precondition (not (nut-in-hub ?n ?h)) )\nEffects: (and (not (robot-holding ?n)) (nut-in-hub ?n ?h) (not (hub-unfastened ?h)) Missing effect: the nut should be loose )"},{"figure_ref":[],"heading":"A.10.11 Remove wheel from hub","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to remove a wheel from a hub. For example, the robot can remove wheel_1 from hub_1. Note that the hub needs to be unfastened in order to remove the wheel. To safely perform this action, the hub needs to be jacked up."},{"figure_ref":[],"heading":"→ →","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"GPT-4: Remove wheel from hub","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?w -wheel: the wheel to remove 2. ?h -hub: the hub from which the wheel is to be removed Preconditions:\n(and (is-wheel ?w) (is-hub ?h) (wheel-on-hub ?w ?h) (hub-jacked-up ?h) (hub-unfastened ?h) )\nEffects: (and\n(not (wheel-on-hub ?w ?h)) (robot-holding ?w) )\nA.10.12 Put wheel on hub"},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to put a wheel on an unfastened hub. For example, the robot can put wheel_1 on hub_1. Note that the hub needs to be unfastened in order to put on the wheel. To safely perform this action, the hub needs to be jacked up.\n→ →"},{"figure_ref":[],"heading":"GPT-4: Put wheel on hub","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?w -wheel: the wheel to put on the hub 2. ?h -hub: the hub to put the wheel on Preconditions:\n(and (robot-holding ?w) (is-wheel ?w) (is-hub ?h) (hub-jacked-up ?h) (hub-unfastened ?h) (not (wheel-on-hub ?w ?h)) )"},{"figure_ref":[],"heading":"Effects: (and","publication_ref":[],"table_ref":[],"text":"(not (robot-holding ?w)) (wheel-on-hub ?w ?h) )"},{"figure_ref":[],"heading":"A.10.13 Inflate wheel","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Action description","publication_ref":[],"table_ref":[],"text":"This action enables the robot to inflate a wheel. For example, the robot can inflate wheel_1. Note that you need to make sure the wheel is intact."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"GPT-4: Inflate wheel","publication_ref":[],"table_ref":[],"text":"Parameters:\n1. ?w -wheel: the wheel to inflate Preconditions: (and (is-wheel ?w) (robot-holding ?w) (wheel-deflated ?w) (wheel-intact ?w) Missing precondition: the robot should have a pump ) Effects: (and (not (wheel-deflated ?w)) (wheel-inflated ?w) )\nA.10.14 Examples of PDDL action models constructed by GPT-3.5-Turbo GPT-3.5-Turbo: Open a containe ------------------------------------ ----------------------------------------Corrective feedback: There are some errors in the PDDL model: 1. There is an incorrect precondition: the robot is not holding the nut because the nut is in the hub → 2. There is a missing effect: the robot should hold the nut at the end Please revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors)."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"Parameters:\n- -------------------------------------- ----------------------------------- ---------------------------------------"},{"figure_ref":[],"heading":"-Corrective feedback:","publication_ref":[],"table_ref":[],"text":"There is an error in the PDDL model: 1. You allow the robot to inflate a wheel without having a pump, this is incorrect.\nPlease revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors)."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":"Parameters:\n----------------------------------------GPT-4: 1. ?w -wheel: the wheel to inflate 2. ?p -pump: the pump to inflate the wheel Preconditions:\n(and (is-wheel ?w) (is-pump ?p) (robot-holding ?w) (robot-holding ?p) (wheel-deflated ?w) (wheel-intact ?w) )\nEffects: (and (not (wheel-deflated ?w)) (wheel-inflated ?w) ) New Predicates: No newly defined predicate - --------------------------------------"},{"figure_ref":[],"heading":"-Automated corrective feedback:","publication_ref":[],"table_ref":[],"text":"There is an error in the PDDL model: 1. There is an unknown object type `pump`for the parameter ?p Please revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors)."},{"figure_ref":[],"heading":"→","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Parameters:","publication_ref":[],"table_ref":[],"text":"----------------------------------------GPT-4: 1. ?w -wheel: the wheel to inflate 2. ?p -small_object: the pump to inflate the wheel Preconditions:\n(and (is-wheel ?w) (is-pump ?p) Your plan is:\nA.12.2 Translation to admissible actions We extract the plan from the LLM planner as a completion to the prompt. Although the prompt explicitly states that the planner should strictly adhere to the demonstrated output format of each action, there are still cases where GPT-4 fails to follow these formats. As a result, extra post-processing is needed to translate the generated actions into admissible actions. To do this, we utilize another LLM (i.e., GPT-4) to perform the translation. In cases where certain actions have missing parameters, we present an error message to the LLM planner and ask it to regenerate the plan. The error message is structured as follows: \"There is an invalid output at step x. Please strictly follow the output format provided in the example output of each action. Your revised plan:\".\nWe use a fixed prompt template for action translation in all the domains:"},{"figure_ref":[],"heading":"Template of the prompt for action translation","publication_ref":["b17"],"table_ref":[],"text":"Your task is to translate free-form action description to its corresponding structural format based on its semantic meaning. The structural output format is given in the example output of each candidate action. You should strictly follow the structural output format. When there is missing information in the free-form action description, you should output \"Not able to translate due to missing information\". A.12.3 Back-prompting LLM planners with validation feedback by VAL\nWe consider three types of validation feedback that can be obtained from VAL [18]. One is unsatisfied precondition. VAL highlights unsatisfied precondition(s) in PDDL, which can be translated into natural language and provided as corrective feedback to the LLM planner. The feedback message is structured as follows: \"The action at step x is not executable due to unmet precondition(s). Here are the unsatisfied precondition(s):\". The second type of validation feedback is related to unmet goal condition(s), and the third type is related to the usage of invalid or illegal parameter(s) in some action(s).\nFor translating PDDL into natural language, we use a fixed prompt template in all the domains:"},{"figure_ref":[],"heading":"Template of the prompt for translating PDDL into natural language","publication_ref":[],"table_ref":[],"text":"Your task is to translate planning domain definition language (PDDL) to natural language.\nHere is an example:\nPredicates: Step 20: > action: close furniture/appliance dish_washer_1 -------------------------------------Validation feedback: The action at step 10 is not executable due to unmet precondition(s). Here are the unsatisfied precondition(s) described in Pseudo-PDDL format:\n→ (and The furniture or appliance fridge_1 has a flat surface for manipulation;\n(or The small receptacle lunch_box_2 cannot be opened;\nThe small receptacle lunch_box_2 is open; ) )\nYour revised plan is: -------------------------------------GPT-4: ... The revised plan ..."},{"figure_ref":[],"heading":"A.13 Translating user instructions into PDDL goal specifications","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Template of the prompt for translating user instructions into PDDL goal specifications","publication_ref":[],"table_ref":[],"text":"Your task is to translate natural language instructions to planning goals in planning domain definition language (PDDL). You should only use the given predicates to express the goal specification.\n→ →\nHere is an example from the classical BlocksWorld domain for demonstrating the output format.\nDomain information: BlocksWorld is a planning domain in artificial intelligence. The AI agent here is a mechanical robot arm that can pick and place the blocks. Only one block may be moved at a time: it may either be placed on the table or placed atop another block. Because of this, any blocks that are, at a given time, under another block cannot be moved. There is only one type of object in this domain, and that is the block.\nThe domain has the following objects: block_1 (type: block), block_2 (type: block), block_3 (type: block), block_4 (type: block), block_5 (type: block). "}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b18\",\n \"b19\",\n \"b59\",\n \"b41\",\n \"b55\",\n \"b23\",\n \"b54\",\n \"b52\",\n \"b53\",\n \"b30\",\n \"b32\",\n \"b47\",\n \"b54\",\n \"b12\",\n \"b17\",\n \"b21\",\n \"b46\",\n \"b36\",\n \"b57\",\n \"b29\"\n ],\n \"table_ref\": [],\n \"text\": \"The field of artificial intelligence has been revolutionized with the advent of large pre-trained models. Of particular significance are transformer-based large language models (LLMs) which have showcased remarkable performance in natural language processing tasks. Along with these tasks, LLMs have been tested to perform another widely-studied crucial aspect of AI agents, namely, sequential decision-making or planning. Preliminary studies suggest that, in some everyday domains, LLMs are capable of suggesting sensible action plans [19,1]. However, the correctness and executability of these plans are often limited. For instance, LLMs may regularly overlook the physical plausibility of actions in certain states and may not effectively handle long-term dependencies across multiple actions. Several approaches have been proposed to improve the planning capabilities of LLMs. One promising approach involves collecting feedback from the environment during plan execution and subsequently refining the plans. By incorporating various forms of feedback, such as sensory information [20], human corrections [60], or information of unmet preconditions [42,56], the planners can re-plan and produce plans that are closer to a satisficing plan.\\nDespite the improvements in planning performance, LLMs are still far from being a usable and reliable planner due to various factors:\\n(a) LLMs have not yet demonstrated sufficient capabilities in reasoning and planning [24,55,53,54,31]. Recent investigations show that even when provided with detailed descriptions of actions, such as a PDDL domain model [33] or a natural-language version of a PDDL model, LLMs still struggle to produce correct and executable plans [48,55].\\n(b) Existing LLMs-planning paradigms only allow for feedback collection in a fully online manner, meaning that the feedback signals are only available after the agent has started executing the plan. However, when a faithful simulator is not available or is expensive to use, collecting feedback through actual plan execution can be costly and may not fully exploit the advantages of provably sound planning, as seen in classical-planning literature [11,13].\\n(c) LLMs exhibit complex behaviors that are not yet fully understood, particularly with respect to error occurrences. LLM planners are prone to repeating the same mistakes in slightly different scenarios. Repeatedly providing the same feedback can lead to frustration for end users.\\nTo overcome these limitations, rather than using LLMs directly as planners, we advocate a modelbased paradigm, wherein a PDDL world model is teased out of LLMs. We follow the identical problem setup as existing approaches, which involves providing the planner with a set of actions and their brief natural language descriptions. However, instead of directly mapping user commands to plans, we utilize LLMs to extract a symbolic representation of the actions in the form of PDDL action models. This intermediate output can be used with an external domain-independent planner to reliably search for feasible plans, or it can be used to validate and correct \\\"heuristic\\\" plans generated by an LLM planner. Additionally, our modular method essentially divides the planning process into two distinct parts, namely modeling the causal dependencies of actions and determining the appropriate sequence of actions to accomplish the goals. LLMs, which have been trained on extensive web-scale knowledge, exhibit greater proficiency in the former task rather than the latter.\\nNevertheless, we still take into account the fact that the LLMs may not be able to generate error-free PDDL models at the outset. To address this, we show that LLMs can also serve as an interface between PDDL and any feedback sources that can provide corrective feedback in natural language, such as humans and the PDDL validator in VAL [18]. The LLM middle layer translates PDDL representation to natural language and presents it to users for inspection. The acquired feedback is then incorporated and archived back to the PDDL models. This conceals the complexity of PDDL from users who do not have prior knowledge of PDDL, and enables seamless inclusion of feedback.\\nWe conducted an extensive evaluation of our methodology on two IPC domains [22] from classical planning literature and a household domain that has a more diverse set of actions and constraints than commonly used benchmarks such as ALFWORLD [47]. We assess the quality of the generated PDDL models through manual evaluation. Results show that GPT-4 [37] generates high-quality PDDL domain models with over 400 literals for 41 actions in total. Then, by replaying and continuing the PDDL-construction dialogue, we show that GPT-4 can readily correct all the errors according to natural language feedback from PDDL validators and humans.\\nWe consider two use cases of the generated PDDL action models for downstream planning tasks. For one, by utilizing an LLM to translate user instructions into goal specifications in PDDL [58,30], we can use any standard domain-independent planner to search for a plan. On the other hand, the extracted PDDL model can be used to validate plans suggested by an LLM planner and to provide corrective feedback in the form of unmet preconditions or goal conditions. In this case, the PDDL model is essentially serving as an inexpensive high-level simulator or a human proxy to ensure plan correctness.\\nThis reduces the reliance on faithful simulators or extensive manual inspection of plans by domain experts. Compared to the first approach, the second approach potentially offers better flexibility in incorporating both explicit and implicit user constraints in common-sense domains because of the LLM planner. For instance, the LLM planner can directly incorporate ordering constraints such as \\\"heat the potato first before mashing it\\\" and \\\"bring me a fork first, then a plate.\\\" On the contrary, an approach purely based on classical planners would require extra steps, such as introducing extra state variables in the PDDL models, in order to accommodate such constraints. However, as demonstrated in our experiments, although the validation feedback significantly improves the plan correctness on average, the performance of the second approach is still limited by the \\\"planning capability\\\" of LLMs.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b23\",\n \"b54\",\n \"b54\",\n \"b47\",\n \"b28\",\n \"b27\",\n \"b49\",\n \"b59\",\n \"b50\",\n \"b19\",\n \"b33\",\n \"b59\",\n \"b45\",\n \"b41\",\n \"b55\",\n \"b54\",\n \"b37\",\n \"b61\",\n \"b26\",\n \"b6\",\n \"b42\",\n \"b25\",\n \"b58\",\n \"b60\",\n \"b8\",\n \"b48\",\n \"b1\",\n \"b34\",\n \"b57\",\n \"b31\",\n \"b14\",\n \"b2\",\n \"b30\",\n \"b52\",\n \"b53\",\n \"b44\",\n \"b38\",\n \"b56\"\n ],\n \"table_ref\": [],\n \"text\": \"LLMs and planning. The growing interest in evaluating the emergent abilities of LLMs paved way into exploring their abilities in sequential decision-making tasks. Preliminary studies [24,55] have shown that off-the-shelf LLMs are currently incapable of producing accurate plans. But their plans can be used as heuristics or seeds to either an external planner or a human in the loop [55,48].\\nSayCan [1] and Text2Motion [29] employ an LLM as a heuristic by utilizing it to score high-level actions, followed by a low-level planner that grounds these actions to determine the executability in the physical world. In a similar vein, [28,50] use LLMs to generate plans represented in Python-style code. Other works have aimed to improve the planning performance of LLMs through prompt engineering [60] or collecting various forms of feedback such as sensory information [51,20,34], human corrections [60], self-corrections [46] or information of unmet preconditions [42,56].\\nTraining transformers for sequential decision-making tasks. Along with using off-the-shelf LLMs, there are works that either fine-tune LLMs [55,38] or train sequence models [62,27,7,43] for sequential decision making tasks. Experiments in [26] have shown that training sequence models on a specific task gives rise to an internal world representation within the model. In this work, we use off-the-shelf LLMs to construct symbolic world models without performing any extra training.\\nLearning/acquiring symbolic domain models. In classical planning, the community has explored numerous learning-based methods [59,61,9,25,4] and interactive editor-based methods [49] for acquiring symbolic domain models. For a more comprehensive survey, we refer the reader to [2,6].\\nHere, we are interested in leveraging the common-world knowledge embedded in LLMs and their in-context learning ability for constructing domain models. Recent studies have shown the efficacy of LLMs in translating natural language to formal descriptions [35] or constructing PDDL goals from natural-language instructions [58,32]. Moreover, a contemporary work [15] considers the use of LLM as a parametric world model and plan critic. However, unlike a symbolic model that can simulate plan outcomes with guaranteed correctness, using LLMs directly as a world model actually adds another layer of errors. There is evidence that autoregressive models lack reliable capacity for reasoning about action effects [3,31] and capturing errors in candidate plans [53,54].\\nLanguage models with access to external tools. Since LLMs are approximately omniscient, they may not always outperform specialized models or tools in specific downstream tasks. To address this limitation, frameworks have been developed to enable LLMs to utilize external tools for performing sub-tasks like arithmetic [45] and logical reasoning [39,57]. In this context, our work can be regarded as an exercise in employing external sound planners to augment the capacity of LLMs for more reliable plan generation.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Problem Setting and Background\",\n \"publication_ref\": [\n \"b15\",\n \"b11\",\n \"b59\",\n \"b19\"\n ],\n \"table_ref\": [],\n \"text\": \"Our work focuses on a scenario where an intelligent agent receives high-level instructions or tasks, denoted as i, from a user. The agent is capable of only executing skills or operations that are part of a skill library Π, where each skill k has a short language description l k . We assume that the agent is equipped with the low-level control policies corresponding to these high-level skills. In order to achieve the goal conditions specified in i, a planner, which can be either an LLM or an external planner [16,12,17], needs to come up with a sequence of high-level skills that the agent can execute. This type of problem is referred to as a sequential decision-making or planning problem. Similar to previous works such as [60,20], we also allow for human-in-the-loop feedback during both the domain-model construction and plan execution stages. In the next subsections, we describe the formalism behind planning problems and a standard way in the literature to specify them.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Classical planning problems\",\n \"publication_ref\": [\n \"b43\"\n ],\n \"table_ref\": [],\n \"text\": \"The most fundamental planning formalism is goal-directed deterministic planning problem, referred to as a classical planning problem in the planning literature. A classical planning problem [44] can be formally represented with a tuple P = ⟨D, I, G⟩. D is referred to as the domain, I is the initial state, and G is the goal specification. The state space of a planning problem consists of the truth assignments for predicates. The domain D is further defined by the tuple D = ⟨F, A⟩. F corresponds to the set of fluents, i.e., the state variables used to define the state space with each fluent corresponding to a predicate with some arity. A corresponds to the set of actions that can be performed. Each action a i [V] ∈ A (where V is the set of variables used by the operator a i and each variable could be mapped to an object) can be further defined by two components, the precondition prec[V] which describes when an action can be executed, and the effects eff[V] which defines what happens when an action is executed. We assume that prec[V] consists of a set of predicates defined over the variables V. An action is assumed to be executable only if its preconditions are met, i.e, the predicates in the precondition hold in the given state. The effect set eff[V] is further defined by the tuple ⟨add[V], del[V]⟩, where add[V] is the set of predicates that will be set true by the action and del[V] is the set of predicates that will be set false by the action. An action is said to be grounded if we replace each of the variables with an object, else it is referred to as a lifted action model. A solution to a planning problem is called a plan, and it is a sequence of actions that once executed in the initial state would lead to a state where the goal specification holds. Classical planning problems are one of the simpler classes in planning and there are multiple extensions with more complex forms of preconditions, conditional effects, and also support for richer planning formalisms.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"PDDL\",\n \"publication_ref\": [\n \"b32\"\n ],\n \"table_ref\": [],\n \"text\": \"Planning Definition and Domain Language (PDDL) [33], is the standard encoding language for classical planning problems. Here is an example of a lifted action in PDDL which corresponds to putting a block onto the table in the classical Blocksworld domain:\\n(:action PutDownBlock :parameters (?x -block) :precondition (and (robot-holding ?x)) :effect (and (not (robot-holding ?x)) (block-clear ?x) (robot-hand-empty) (block-on-table ?x)))\\nThe parameters line provides the possible variable(s), and in this case, ?x represents the block to put down. The precondition states that the robot must be holding the block in its gripper. The effects line describes the expected outcome of this action.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Methodology\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"PDDL provides a succinct and standardized way to represent a world model. Once a PDDL model is constructed, it can be seamlessly used by any domain-independent planner developed in the automated planning community to search for a plan given the initial state and goal conditions. In this section, we will introduce our solution for constructing PDDL models using LLMs. We then discuss techniques for correcting errors in the generated PDDL models. Finally, we present the full pipeline for utilizing the generated PDDL models to solve planning problems.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_1\"\n ],\n \"heading\": \"Constructing PDDL models with LLMs\",\n \"publication_ref\": [\n \"b29\",\n \"b40\",\n \"b36\",\n \"b9\"\n ],\n \"table_ref\": [],\n \"text\": \"Our approach involves prompting pre-trained LLMs with the following information: (a) detailed instructions for the PDDL generation task, outlining components of upcoming inputs and desired outputs; (b) one or two examples from other domains (e.g., the classical Blocksworld domain) for illustrating the input and output formats; (c) a description of the current domain, including contextual information about the agent's tasks and physical constraints due to the specific embodiment of the agent; (d) a description of the agent's action; and (e) a dynamically updated list of predicates that the LLM can reuse to maintain consistent use of symbols across multiple actions. Note that the predicate list is initialized to an empty list, and thus all predicates are introduced by the LLM. The structure of the prompt is illustrated in Fig. 2, and a complete prompt for the household-robot domain can be found at Appx. A.6.1. Depending on the information included in the action description or the domain context, users may gain varying levels of control over the extracted PDDL or receive differing levels of support from the LLMs. On one hand, when the user provides only a minimal description of the action, such as \\\"this action enables the robot to use a microwave to heat food,\\\" we not only use the LLM as a PDDL constructor but also leverage the common world knowledge encoded within the model for knowledge acquisition. This is particularly useful when expanding the set of actions for an AI agent. For example, a robot engineer could set up a training environment for skill learning by following the suggested preconditions and effects. On the other hand, when some preconditions or effects are explicitly mentioned in the prompt, we rely more on the LLM's ability to parse the knowledge provided in natural language and to precisely represent it by devising a collection of predicates. This capability is useful when there could be different initial setups of a skill, and the engineers have already made some assumptions on the preconditions at the time of designing the skill. This capability is also crucial when constructing PDDL for specialized domains. For instance, robots such as Fetch and Spot Robot have only one robot arm, which is less flexible than a human arm, and are therefore subject to many uncommon physical constraints.\\nThe desired output comprises the following elements: (a) the list of arguments for the action; (b) the preconditions and effects expressed in PDDL; and (c) a list of any newly defined predicates and their descriptions in natural language, if applicable. An example output is shown in Fig. 2. Our algorithm generates PDDL models for each action separately, one at a time, by iterating over the set of actions. Any newly defined predicates will be added to an actively maintained predicate list, such that the LLM can reuse existing predicates in subsequent actions without creating redundant ones. Once we obtain the initial PDDL models and the full predicate list, we repeat the entire process but with all of the extracted predicates presented to the LLM. Running the generation process twice is useful because the LLMs may be unaware of some precondition(s) during the first iteration, especially if the precondition(s) are not explicitly mentioned. For instance, the LLM may overlook the fact that a furniture piece can be openable, but a predicate created in the \\\"open a furniture piece or appliance\\\" skill can inform the LLM of this fact. One alternative to this action-by-action generation could be to include descriptions of all the actions in the prompt and require the LLM to construct the entire domain model in a single dialogue. An additional discussion on this can be found at Sec. A.2 in Appendix.\\nIt is worth noting that every time a new predicate is defined, the LLM is required to give the natural language description of it. As we will see in the following sections, this is crucial for enabling any user to easily understand and inspect the generated PDDL models without having to delve into the low-level symbolic representation. Additionally, natural language descriptions allow the predicate values of the initial state to be automatically grounded by using LLMs to translate environment description in natural language to PDDL [30], or leveraging pre-trained vision-language models ---------------------------------------The LLM: ... [ 41,37,10] and querying them in a question-answering manner, based on observations from the environment.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Correcting errors in the initial PDDL models\",\n \"publication_ref\": [\n \"b17\",\n \"b59\"\n ],\n \"table_ref\": [],\n \"text\": \"As with any use case involving LLMs, there is no guarantee that the output is completely error-free. Therefore, it is essential to incorporate error-correction mechanisms. While it may be easy for PDDL experts to directly inspect and correct the generated PDDL models, we cannot assume that all end users possess this level of expertise. Our solution is to use the LLM as a middle layer or interface between the underlying PDDL model and any feedback source that can provide corrective feedback in natural language. We consider two feedback sources in this work, namely the PDDL model validation tools (e.g., the one in VAL [18]) and human domain experts. The former is used to detect basic syntax errors, while the latter is mainly responsible for catching factual errors, such as missing effects. It is worth noting that the feedback sources are not limited to those mentioned above, and we leave the investigation of other sources for future research.\\nFor corrective feedback from PDDL validators, a generated PDDL model is directly presented to the validator to obtain brief but readable error messages. Examples of feedback messages for syntax errors are shown in Appx. A.3. For corrective feedback from users, a PDDL model is translated into its natural-language version based on the natural language descriptions of the predicates and parameters (Sec. 4.1). The user can then examine potentially erroneous action models. Human corrections can occur both during the construction of PDDL models and after the models have been used for planning. Although there are techniques available to assist users to locate errors in the models (as discussed in Appx. A.4), this is beyond the scope of this work, since the focus here is to investigate the feasibility of using LLMs to correct PDDL models based on feedback. We also note that correcting action models is not more cognitively demanding than correcting plans or the \\\"reasoning traces\\\" of an LLM planner [60]. In fact, when correcting plans, humans must also maintain the action models and their causal chains in mind in order to validate the plans. More importantly, once the action models are corrected, users no longer need to provide similar feedback repeatedly. Finally, corrective feedback is integrated by replaying and continuing the PDDL-construction dialogue. Examples of such dialogues can be found in Sec. A.7, Sec. A.9, and Sec. A.11 in Appendix.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Generating plans with the extracted PDDL models\",\n \"publication_ref\": [\n \"b29\",\n \"b57\",\n \"b31\",\n \"b11\",\n \"b54\",\n \"b41\",\n \"b20\"\n ],\n \"table_ref\": [],\n \"text\": \"Recall that given the set of extracted predicates and their natural language descriptions, we can get the grounded initial state by using LLMs to translate descriptions of the environment to PDDL, or by observing the environment and querying pre-trained vision-language models. Besides, the goal specification can be obtained by using an LLM to parse the user's command and convert it into a symbolic form, as done previously in [30,58,32]. With this setup, the following two methods can be used to generate the final plans.\\nClassical planner with LLM-acquired PDDL model. One straightforward approach is to employ a standard domain-independent planner to reliably find a satisficing or even optimal plan for the specified goal. In common-sense domains where LLMs may generate meaningful \\\"heuristics\\\", the LLM plans may also be used as seed plans for a local-search planner such as LPG [12] to accelerate the plan searching. This is similar to the approach suggested in [55], but with a higher degree of automation.\\nLLM modulo planner backprompted by VAL using LLM-acquired PDDL model. As outlined in Sec. 1, we can also use the extracted PDDL as a symbolic simulator or human proxy to provide corrective feedback based on validation information to an LLM planner. With this setup, the planner can iteratively refine the plans through re-prompting [42].\\nIt is worth noting that depending on the specific problem settings, the extracted PDDL model can also be used for tasks other than task planning. For instance, in cases where reinforcement learning is permissible, the domain model can be used to guide skill learning [21,8] or exploration even if the model is not fully situated [14].\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Empirical Evaluation\",\n \"publication_ref\": [\n \"b46\",\n \"b39\",\n \"b36\",\n \"b35\"\n ],\n \"table_ref\": [],\n \"text\": \"We conduct our experiments2 on an everyday household-robot domain and two more specialized IPC domains (i.e., Tyreworld and Logistics). The Household domain is similar to other commonly used benchmarks like ALFWORLD [47] and VirtualHome [40]. However, in our household domain, a single-arm robot is equipped with a more diverse and extended set of 22 mobile and manipulation skills. In addition, we apply more rigorous physical-plausibility constraints to each skill. A detailed description of this domain can be found at Appx. A.5. In our experiments, we first evaluate the quality of PDDL models generated by the LLMs. Next, we assess the ability of the LLMs to incorporate corrective feedback from both PDDL validators and users in order to obtain error-free PDDL models. Lastly, we showcase multiple ways to use the corrected PDDL model for downstream planning tasks. We present the results of GPT-4 [37] and GPT-3.5-Turbo [36] for PDDL construction (we also conducted experiments with GPT-3 [5], and observe that its performance is comparable to that of GPT-3.5-Turbo).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Constructing PDDL\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In PDDL construction tasks, we aim to investigate the extent to which LLMs can construct accurate PDDL models before getting corrective feedback from domain experts. For all the domains, two actions from the classical Blocksworld domain are used as demonstrations in the prompt so that the end user is not required to come up with any domain-specific example. To evaluate the degree of correctness, we recruit multiple graduate students who possess expertise in PDDL. These experts are responsible for annotating and correcting any errors present in the generated PDDL models. As an evaluation metric, we count and report the total number of annotations, which may include the removal of irrelevant preconditions, the addition of missing preconditions, the replacement of incorrect predicates, the inclusion of missing parameters, and other commonly made corrections. Note that the number of annotations can be viewed as the approximate distance between a generated PDDL model and its corrected version. In order to provide the reader with a comprehensive understanding of the quality of the generated models, we also list all the models and collected annotations in Appendix. In each of the figures, errors that affect the functionality of the PDDL model are highlighted in yellow, while minor issues are highlighted in green. One example of a minor issue is the redundant inclusion of (pickupable ?o) in preconditions when (robot-holding ?o) has already been listed. The former is unnecessary because it can be implied by the latter, but this only affects conciseness rather than functionality. Table 1: The number of errors in the domain models produced by the LLMs for each of the domains.\\nA \\\"+\\\" mark indicates that the generated model is excessively noisy, making it challenging to determine an exact number of errors.\\nWe first evaluate the PDDL models generated when partial constraint information is given, as this is closer to most of the practical use cases where constraints on skills in the library Π are often pre-specified. In this setting, our evaluation focuses on the LLMs' ability to accurately recover a \\\"ground truth PDDL\\\" that captures the mentioned constraints and underlying dependencies among skills. Our results indicate that GPT-4 can produce high-quality PDDL models with significantly fewer errors when compared to GPT-3.5-Turbo. Table 1 presents the number of errors in the generated domain models for each domain. To help the readers understand the complexities of the action models, we additionally report the total number of parameters and literals in the final corrected domain models produced by GPT-4. Out of the total 59 errors made by GPT-4, three of them are syntax errors and the rest are factual errors such as missing preconditions and effects. This observation suggests that while GPT-4 demonstrates proficiency in adhering to the grammar of PDDL, it may still have an inaccurate understanding of the actions. By examining the set of predicates (listed in the Appendix), we also find that GPT-4 can devise a set of intuitively-named predicates that can concisely and precisely describe the states of objects and events in the domain. In contrast, GPT-3.5-Turbo produces highly noisy outputs with over 350 errors. This suggests that our framework relies heavily on GPT-4's improved capability in understanding symbols, and future work may investigate how to enable the use of more lightweight models (e.g., by fine-tuning on some PDDL datasets). Furthermore, recall that when the action description contains minimal information, LLMs could also be utilized to propose preconditions and effects to assist with knowledge acquisition. To verify this hypothesis, we conduct additional experiments on the Household domain that can have a more open-ended action design. In this setting, the correctness of the action models is determined based on whether the preconditions and effects establish correct connections among the actions. Our results show that GPT-4 can suggest meaningful action models, and the generated PDDL models have only around 45 errors.\\nAlthough GPT-4 has shown improved performance in the PDDL construction task, our experiments still uncover some limitations. Firstly, GPT-4 still exhibits a shallow understanding of the causal relationships between actions, particularly when it comes to tasks involving reasoning skills such as spatial reasoning. For instance, when constructing the model of action \\\"pick up an object from a furniture piece,\\\" GPT-4 fails to consider that there could be other objects stacked on top of the target object, even if relevant predicates are provided (which were created in the action \\\"stack objects\\\").\\nIn addition, although it occurs rarely, GPT-4 may output contradictory effects. For instance, in the action of mashing food with a blender, GPT-4 lists both (not (object-in-receptacle ...)) and (object-in-receptacle ...) as effects at the same time.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Correcting PDDL with domain experts\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We proceed with the PDDL models generated by GPT-4 when the constraint information is partially given. Our objective is to demonstrate the feasibility of using GPT-4 as a middle layer to incorporate natural-language feedback and correct the PDDL models. As discussed in Sec. 4.2, we use PDDL validators to capture basic syntax errors. In the Household domain, there are two syntax errors associated with improper usage of relevant predicates due to issues with the object types of parameters 3 . As shown in Appx. A.7.1, by continuing the PDDL-construction dialogue with a feedback message \\\"the second parameter of object-on should be a furnitureAppliance but a householdObject was given,\\\" GPT-4 can locate the inaccurate PDDL snippet and replace it with a correct one. For the other factual errors, GPT-4 successfully corrects all of them based on the natural language feedback. An example feedback message on factual errors is \\\"there is a missing effect: the item is no longer pickupable after being mashed.\\\" More PDDL-correction conversations can be found in Appendix. We also experiment with feedback written in various ways, and GPT-4 is able to understand all the messages and successfully correct the models. To quantify how effectively GPT-4 utilizes feedback from domain experts, we count the number of feedback messages concerning factual errors. Our result shows that GPT-4 required 59 feedback messages to address a total of 56 factual errors. There are three instances where additional feedback was needed. One case involved the user reiterating the error, while the other two cases involved GPT-4 introducing new errors. Furthermore, we attempt to correct the same errors using GPT-3.5-Turbo. Results show that GPT-3.5-Turbo not only fails to correct all the errors but also occasionally introduces new errors, again confirming its lack of ability to manipulate symbols. Some examples can be found in Appendix starting from Sec. A.7.3.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Generating plans with the extracted PDDL models\",\n \"publication_ref\": [\n \"b15\",\n \"b29\",\n \"b59\"\n ],\n \"table_ref\": [\n \"tab_3\"\n ],\n \"text\": \"For planning tasks (i.e., user instructions and initial states), we use the Household domain and Logistics domain, where state-of-the-art LLM planners struggle to find valid plans. We sampled 27 tasks for Household and 21 for Logistics. For the initial states, we assume the grounding is provided, and for the goals, we leverage GPT-4 to translate user instructions into PDDL goal specifications in terms of the extracted predicates (an example prompt can be found at Appx. A.13), and send it over to a standard STRIPS planner which already has access to the domain model acquired through LLMs. With this setup, a classical planner Fast Downward [16] can effectively find valid plans in 95% of the cases (the failures were only due to goal translation errors). Note that in contrast to earlier methods such as [30] that use LLMs only as a mechanism for translating user goals to PDDL format, and throw that over to external sound planners with hand-crafted correct PDDL domain models, our approach uses LLMs themselves to develop the PDDL world model driving the external planner. On the other hand, for the approach that utilizes PDDL models to validate LLM plans (i.e., LLM modulo planner back-prompted by VAL using LLMacquired domain model), we employ the state-of-the-art algorithm ReAct [60] with GPT-4 as the underlying LLM planner. However, we made two modifications to the prompt design. Firstly, we provide a detailed description of all actions in natural language, including parameters, preconditions, and effects. These descriptions are obtained by using another LLM to translate the generated PDDL domain model into natural language. Secondly, we use only two fixed examples for each domain because end users might not always be able to provide a large pool of examples, and the planner should rely on the action model information.\\nThe LLM plans, symbolic goal specifications, initial states and domain models are passed to a plan validation system (i.e., VAL) to check for unmet precondition(s) or goal condition(s). The validation results (given in PDDL) are then translated into natural language with GPT-4 and provided to the LLM planner by continuing the planning dialogue (see Appx. A.12.1 for examples). In our experiments, we limit the number of feedbacks per task to 8 due to the restricted access to GPT-4. Table 2 provides a summary of the average success rates of all approaches. Not surprisingly, the vanilla LLM planner constantly overlooks action preconditions and achieves an extremely low success rate. With the integration of validation feedback, we observe a notable improvement in plan correctness. Despite this improvement, the overall performance is still not satisfactory, as the success rate remains below 50%. Furthermore, we have observed that GPT-4 fails to effectively utilize the feedback, often getting stuck in a loop by repeatedly generating the same plan. In some cases, it may also introduce new errors while attempting to rectify the plans.\\nBeyond the notion of correctness, the experiments also uncover intriguing properties of the LLM planner. In the Household domain, we intentionally introduce ordering constraints in some instructions that cannot be expressed using existing predicates (refer to Appx. A.12 for examples). Remarkably, upon manual examination of the generated plans, we observe that all LLM plans adhere to the specified ordering, despite not being entirely correct or executable. Furthermore, also in the Household domain, we observe that classical planners occasionally generate physically plausible but unconventional actions, such as placing a knife on a toaster when the knife is not being used. In contrast, the LLM planner rarely exhibits such actions, suggesting that LLMs possess knowledge of implicit human preferences. It would be meaningful to explore methods that more effectively combine the strengths of LLM planners and the correctness guarantee provided by symbolic domain models, particularly in determining which information from LLM plans should be preserved.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [\n \"b35\"\n ],\n \"table_ref\": [],\n \"text\": \"We introduce a new paradigm for leveraging LLMs in planning tasks, which involves maintaining an explicit world model instead of directly mapping user prompts to plans. This is motivated by the insight that LLMs, while incapable of the combinatorial search needed to produce correct plans, may be better suited as the source of world models. We present a complete pipeline that begins with generating high-quality PDDL models using GPT-4, then corrects the PDDL models with naturallanguage feedback, and finally utilizes the extracted domain models to reliably plan in multiple ways.\\nOur experiments demonstrate that pairing LLMs with an external planner significantly outperforms existing methods when applied to two IPC domains and a household-robot domain that has more action-wise constraints than commonly used benchmarks such as ALFWorld. Apart from directions for further research that we have previously mentioned, there are several exciting opportunities for extending this work. Firstly, the complexity of our evaluation domains is still lower than that of many domains used in the classical planning literature. It remains to be seen whether LLMs can effectively scale to write PDDL models that express more intricate logic. Secondly, our framework assumes full observability, meaning that the agent must fully explore the environment to acquire object states at the beginning. It would be useful to support partial observability. Finally, our experiments assume the grounding of predicate values is done perfectly. However, it would be useful to take into account that perception can be noisy in practice.\\napproach may allow the LLM to better establish a global view of all the actions. However, we do not pursue this alternative here for the following reasons: (a) the inclusion of all actions might result in a lengthy prompt, potentially exceeding the context window size of an LLM. This could pose practical issues for utilizing smaller language models (e.g., GPT-3.5-Turbo [36]) or attempting to train smaller specialized models; (b) our integration of corrective feedback relies on continuing the construction dialogue (Sec. 4.2), which necessitates a shorter initial prompt to fit within the context window; (c) our experiments indicate that the action-by-action construction approach already achieves satisfactory results\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.3 Examples of feedback messages that capture syntax errors\",\n \"publication_ref\": [\n \"b48\"\n ],\n \"table_ref\": [],\n \"text\": \"Recall that we rely on the PDDL validator in VAL to identify syntax errors. However, several \\\"simpler\\\" syntax errors can be easily detected using simple Python scripts. In our experiments, we wrote our scripts to capture such syntax errors. Also note that since these errors can be detected at a minimal cost, the corresponding feedback messages are directly provided to the LLMs, and they are not counted in the results reported in Table 1. These \\\"simpler\\\" syntax errors include:\\n1. In this work, we only consider standard base-level PDDL. However, it's possible that the LLMs have seen various extensions of PDDL and might use them in the constructed domain models. Hence, we provide a feedback message to the LLM whenever we detect an unsupported keyword. One example feedback is: \\\"The precondition or effect contain the keyword 'forall' that is not supported in a standard STRIPS style model. Please express the same logic in a simplified way. You can come up with new predicates if needed (but note that you should use existing predicates as much as possible).\\\"\\n2. Newly created predicates might have the same names as existing object types, which is not allowed in PDDL. In such cases, a feedback message is provided to notify the LLM about the name clash. For instance, a message might state: \\\"The following predicate(s) have the same name(s) as existing object types: 1. 'smallReceptacle'. Please rename these predicates.\\\"\\n3. Newly created predicates might have the same names as existing predicates, which is not allowed in PDDL. Also, LLMs often mistakenly list existing predicates under the 'New Predicates' section. In such cases, a feedback message is provided to notify the LLM about the name clash or mistake. For instance, a message might state: \\\"The following predicate(s) have the same name(s) as existing predicate(s): 1. (cutting-board ?z -smallReceptacle), true if the small receptacle ?z is a cutting board | existing predicate with the same name: (cutting-board ?z -householdObject), true if the object ?z is a cutting board. You should reuse existing predicates whenever possible. If you are reusing existing predicate(s), you shouldn't list them under 'New Predicates'. If existing predicates are not enough and you are devising new predicate(s), please use names that are different from existing ones. Please revise the PDDL model to fix this error.\\\" This is the most common syntax error made by GPT-4 in our experiments. 4. The LLMs might fail to only use the object types given in the prompt. An example feedback can be: \\\"There is an invalid object type 'pump' for the parameter ?p.\\\"\\nIn our experiments, the above error types encompass the majority of syntax errors made by GPT-4. In some less-common cases, GPT-4 may have problems with predicate usage, usually caused by mis-matched object types. This kind of error can be captured by VAL and an example feedback message can be: \\\"There is a syntax error, the second parameter of 'object-on' should be a furnitureAppliance, but a householdObject was given. Please use the correct predicate or devise new one(s) if needed.\\\"\\nA. 4 Techniques that assist users to locate errors in PDDL models Several well-established techniques and tools are available for locating errors in PDDL models. For example, graphical tools like GIPO [49] can effectively visualize the causal dependencies of actions. However, these advanced tools or techniques are beyond the scope of this work. Here, we outline a viable solution as a starting point for users who are unfamiliar with these tools.\\nThere are two stages in which corrective feedback can be obtained: during the construction of the PDDL model, and when the domain model is used to generate a plan. In the first stage, end users can direclty review the domain model and identify potential factual errors. Since all the predicates and parameters are accompanied by natural language descriptions, we can easily convert the PDDL model into natural language and present it to the users. This allows users to pre-screen the preconditions and effects. Note that we do not expect all factual errors to be caught in this stage, because the users may not be aware of certain constraints until they review the final plans. In the second stage, the PDDL model is used to solve downstream planning problems by following the procedure outlined in Sec. 4.3. Two possible cases may arise here: (a) no plan can be found for the given goal specification, or (b) at least one plan is found but it either gets rejected by the users or results in execution failure in the actual environment. To address the first case, we can request the users to suggest a goal-satisficing plan, which is supposed to be executable (but not necessarily optimal). We then use the generated PDDL model to \\\"validate\\\" the suggested plan. This allows us to find the first step in the plan that has an unsatisfied precondition(s). The models of all actions up to this step, along with the unmet precondition(s), are then converted to natural language and presented to the user for inspection. As an example, in the model of \\\"slice an object\\\" extracted by GPT-4 (Sec. A.6.11 in Appendix), the model requires the object to be placed on a cutting board and a furniture piece at the same time, which is not physically possible. By leveraging a user-suggested plan, we can identify the potentially erroneous model(s) and flag incorrect precondition(s). In the second case where an invalid plan is afforded by the PDDL model, there are typically missing preconditions or effects in the actions taken, both during and prior to the execution failure. The users can bring attention to these actions.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.5 Detailed description of the Household domain\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We consider a single-arm robot model that closely resembles the SPOT robot and the Fetch Mobile Manipulator. Consequently, the robot is incapable of grasping multiple objects simultaneously or executing manipulation actions while holding irrelevant items (e.g., opening a fridge door while holding a mug). We also ensure the constraints align with real-world robotic capabilities. For example, we recognize that robot arms may be significantly less flexible than human arms, and therefore, we require certain manipulation tasks to be performed on furniture pieces with open and flexible surfaces (e.g., the robot can only pick up food items from a lunch box when the lunch box is placed on a kitchen countertop instead of inside a fridge). The list of actions and their descriptions can be found in Appendix starting from Sec. A.6.2. The PDDL-construction prompt for this domain includes a general description of the domain, which outlines the tasks to be performed by the robot, the types of objects involved, and details of the robot's morphology. An example of a complete prompt can be found in Fig. A.6.1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.6 Household: Constructing PDDL Models\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The AI agent here is a household robot that can navigate to various large and normally immovable furniture pieces or appliances in the house to carry out household tasks. Note that the robot has only one gripper, so (a) it can only hold one object; (b) it shouldn't hold any other irrelevant objects in its gripper while performing some manipulation tasks (e.g., opening a drawer or closing a window); (c) operations on small household items should be carried out on furniture with a flat surface to get enough space for manipulation. There are three major types of objects in this domain: robot, furnitureAppliance, and householdObject. The object type furnitureAppliance covers large and normally immovable furniture pieces or appliances, such as stove burners, side tables, dining tables, drawer, cabinets, or microwaves. The object type householdObject covers all other small household items, such as handheld vacuum cleaners, cloth, apples, bananas, and small receptacles like bowls and lunch boxes. There is a subtype of householdObject called smallReceptacle that covers small receptacles like bowls, lunch boxes, plates, etc. In this domain, the locations of the robot and small household items (e.g., apples, oranges, bowls, lunch boxes or lamps) are determined by large and normally immovable furniture pieces or appliances.\\nA.6.1 An example prompt for constructing PDDL models of the action \\\"close a small receptacle\\\"\\nAn example prompt for constructing PDDL models of the action \\\"close a small receptacle\\\"\\nInstructions for the PDDL generation task You are defining the preconditions and effects (represented in PDDL format) of an AI agent's actions. Information about the AI agent will be provided in the domain description. Note that individual conditions in preconditions and effects should be listed separately. For example, \\\"object_1 is washed and heated\\\" should be considered as two separate conditions \\\"object_1 is washed\\\" and \\\"object_1 is heated\\\". Also, in PDDL, two predicates cannot have the same name even if they have different parameters. Each predicate in PDDL must have a unique name, and its parameters must be explicitly defined in the predicate definition. It is recommended to define predicate names in an intuitive and readable way. \\n→ → → → → → → → → → → → →\\nA natural language description of the action Action: This action enables the robot to close a small receptacle receptacle_1 that is openable (e.g. small storage boxes or lunch boxes with lids). For example, the robot closes lunch_box_2, or the robot closes storage_box_1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The dynamically updated list of predicates You can create and define new predicates, but you may also reuse the following predicates: 1. (robot-at ?x -furnitureAppliance): true if the robot is at the furniture or appliance ?x 2. (object-on ?x -householdObject ?y -furnitureAppliance): true if the object ?x is on or in the furniture or appliance ?y → 3. (pickupable ?x -householdObject): true if the object ?x can be picked up by the robot 4. (closed-receptacle ?x -smallReceptacle): true if the small receptacle ?x is closed ...\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Parameters:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"A.6.2 Navigate to a furniture piece or an appliance\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to navigate from one normally immovable piece of furniture to another (e.g., dining tables, side tables, cabinets, and sinks) or an appliance (e.g., refrigerators, coffee makers, microwaves, and washers).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"GPT-4: Navigate to a furniture piece or an appliance Parameters:\\n1. ?x -furnitureAppliance: the furniture or appliance the robot is currently at 2. ?y -furnitureAppliance: the furniture or appliance the robot wants to navigate to Preconditions:\\n(and (robot-at ?x) (not (= ?x ?y)) )\\nEffects: (and (not (robot-at ?x)) (robot-at ?y) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.6.3 Pick up an object on or in a furniture piece or an appliance\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to pick up an object object_1 in/on a large piece of furniture (e.g., dining tables, drawers, cabinets) or an appliance (e.g., dishwashers and refrigerators). For example, the robot pick up coffee_mug_1 on dining_table_1, the robot pick up fork_1 in drawer_1, or the robot pick up apple_1 in refrigerator_1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ → →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user GPT-4: Pick up an object on or in a furniture piece or an appliance Parameters:\\n1. ?x -householdObject: the object to pick up 2. ?y -furnitureAppliance: the furniture or appliance where the object is located Preconditions:\\n(and (robot-at ?y) (object-on ?x ?y) (pickupable ?x) Inaccurate precondition: \\\"?x is not stacked on top of other object\\\" should only apply when ?x is a stackable object (not (object-stacked ?x)) (robot-hand-empty) (or (not (openable ?y)) (opened ?y) ) Missing precondition: there should be no other object stacked on top of ?x if ?x is a stackable object )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Effects: (and\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"(not (object-on ?x ?y)) (robot-holding ?x) (not (robot-hand-empty)) )\\nA.6.4 Put an object on or in a furniture piece or an appliance\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to put an object object_1 in/on a large piece of furniture (e.g., dining tables, drawers, cabinets) or an appliance (e.g., dishwashers and refrigerators). For example, the robot puts coffee_mug_1 on dining_table_1, the robot puts fork_1 in drawer_1, or the robot puts apple_1 in refrigerator_1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ → →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user (a) The furniture piece or appliance should be opened if it is openable.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Put an object on or in a furniture piece or an appliance\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?x -householdObject: the object to put in/on the furniture or appliance 2. ?y -furnitureAppliance: the furniture or appliance to put the object in/on Preconditions:\\n(and (robot-at ?y) (robot-holding ?x) (pickupable ?x)\\nUnnecessary precondition \\\"no other object stacked on object ?x\\\" (object-clear ?x) (or (not (openable ?y)) (opened ?y) ) )\\nEffects: (and (not (robot-holding ?x)) (robot-hand-empty) (object-on ?x ?y) Syntax error, there is an illegal keyword \\\"if\\\" (if (openable ?y) (closed ?y)) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.6.5 Stack Objects\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to stack one object object_1 on top of another object object_2 that is on the same piece of furniture. For example, the robot stacks book_1 on top of book_2, or the robot stacks lunch_box_1 on top of lunch_box_2.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Stack objects\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?x -householdObject: the object to stack on top of another object 2. ?y -householdObject: the object on which ?x will be stacked 3. ?z -furnitureAppliance: the furniture or appliance on which both objects are placed Preconditions:\\n(and (robot-holding ?x) (object-on ?y ?z) (object-clear ?y) (flat-surface ?z) Incorrect precondition: should be stackable (pickupable ?x) Incorrect precondition: should be stackable (pickupable ?y) (robot-at ?z) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Effects: (and\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"(not (robot-holding ?x)) (not (object-clear ?y)) (robot-hand-empty) (object-on ?x ?z) (object-stacked ?x) Missing effect: object ?x should be stacked on top of object ?y )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.6.6 Unstack Objects\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to unstack one object object_1 that is on top of another object object_2. For example, the robot stacks book_1 on top of book_2, or the robot stacks lunch_box_1 on top of lunch_box_2. The robot will hold the object after unstacking it.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Unstack objects\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?x -householdObject: the object to unstack from another object 2. ?y -householdObject: the object that ?x is stacked on 3. ?z -furnitureAppliance: the furniture or appliance where the objects are stacked Preconditions:\\n(and (robot-at ?z) (object-on ?x ?z) (object-on ?y ?z) (pickupable ?x) (object-stacked ?x) (object-clear ?x) (robot-hand-empty) Missing precondition: object ?x should be stacked on top of object ?y ) Effects: (and\\n(not (object-on ?x ?z)) (not (object-stacked ?x)) Incorrect effect, this action shouldn't change whether or not object ?x is under any another household object (not (object-clear ?x)) (not (robot-hand-empty)) (robot-holding ?x) (object-clear ?y) Missing effect: object ?x is no longer stacked on top of object ?y )\\nA.6.7 Open a furniture piece or an appliance\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to open a large piece of furniture (e.g., cabinets and drawers) or an appliance (e.g., dishwashers and refrigerators) that is openable. For example, the robot opens cabinet_1, or the robot opens drawer_2.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"GPT-4: Open a furniture piece or an appliance Parameters:\\n1. ?y -furnitureAppliance: the furniture or appliance to be opened Preconditions:\\n(and (robot-at ?y) (openable ?y) (closed ?y) (robot-hand-empty) )\\nEffects: (and (not (closed ?y)) (opened ?y) )\\nA.6.8 Close a furniture piece or an appliance\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to close a large piece of furniture (e.g., cabinets and drawers) or an appliance (e.g., dishwashers and refrigerators) that is openable. For example, the robot closes cabinet_1, or the robot closes drawer_2. A.6.9 Toggle a small appliance on\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to toggle a small appliances (like humidifiers and light bulbs) to switch them on. For example, the robot toggles humidifier_2 on, or the robot toggle light_bulb_1 on.\\n→ →\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Toggle a small appliance on\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?x -householdObject: the small appliance to toggle on 2. ?y -furnitureAppliance: the furniture or appliance where the small appliance is located Preconditions:\\n(and (robot-at ?y) (object-on ?x ?y) (robot-hand-empty) (not (appliance-on ?x)) Missing precondition: the appliance to toggle should be a toggleable appliance )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Effects: (and (appliance-on ?x) )\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"A.6.10 Toggle a small appliance off\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to toggle a small appliances (like humidifiers and light bulbs) to switch them off. For example, the robot toggles humidifier_2 off, or the robot toggle light_bulb_1 off.\\n→ →\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT4: Toggle a small appliance off\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?r -robot: the robot performing the action 2. ?o -householdObject: the small appliance to be toggled off 3. ?f -furnitureAppliance: the furniture or appliance where the small appliance is located Preconditions:\\n(and (robot-at ?r ?f) (object-in-on-furniture ?o ?f) (robot-gripper-empty ?r) (appliance-on ?o) ) Effects:\\n(and\\n(not (appliance-on ?o)) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.6.11 Slice an object\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to slice objects (like fruits and vegetables) with a knife. For example, the robot slices orange_1 with knife_2, the robot slices potato_3 with knife_1.\\n→ A.6.14 Transfer food from one small receptacle to another\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to transfer food food_1 from one small receptacle receptacle_1 to another small receptacle receptacle_2. This skill is usually used to transfer food that can not be easily picked up by the robot. For example, the robot can pour the milk from milk_box_1 to mug_1, or the robot can transfer heated pizza_1 from pan_2 to plate_1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ → →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user GPT-4: Transfer food from one small receptacle to another Parameters:\\n1. ?food -householdObject: the food to be transferred 2. ?receptacle1 -smallReceptacle: the small receptacle from which the food will be transferred 3. ?receptacle2 -smallReceptacle: the small receptacle to which the food will be transferred 4. ?furniture -furnitureAppliance: the furniture piece with a flat surface where the transfer will take place → Preconditions:\\n(and (robot-at ?furniture) (object-on ?receptacle1 ?furniture) (object-on ?receptacle2 ?furniture) (object-in ?food ?receptacle1) (is-small-receptacle ?receptacle1) (is-small-receptacle ?receptacle2) (flat-surface ?furniture) (robot-hand-empty) Inaccurate precondition: \\\"?receptacle1 is not stacked on other objects\\\" should only apply if ?receptacle1 is a stackable object (not (object-stacked ?receptacle1)) Inaccurate precondition: \\\"?receptacle2 is not stacked on other objects\\\" should only apply if ?receptacle2 is a stackable object (not (object-stacked ?receptacle2)) (or (not (openable-receptacle ?receptacle1)) (opened-receptacle ?receptacle1)) (or (not (openable-receptacle ?receptacle2)) (opened-receptacle ?receptacle2)) Missing precondition: there should be no other object stacked on top of ?receptacle1 if ?receptacle1 is stackable Missing precondition: there should be no other object stacked on top of ?receptacle2 if ?receptacle2 is stackable ) Effects: (and (not (object-in ?food ?receptacle1)) (object-in ?food ?receptacle2) )\\nA.6.15 Put an object onto or into a small receptacle like a bowl and plate\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to put an object object_1 into/onto a small receptacle receptacle_1 (e.g. storage boxes, bowls, plates, or pans). For example, the robot puts apple_1 into blow_1, or the robot puts apple_2 into lunch_box_2.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user (flat-surface ?y) (robot-hand-empty) Incorrect precondition: ?x can't be stacked inside a receptacle so the precondition `?x is not under other household object`is unnecessary (object-clear ?x) Incorrect precondition: ?z can be an openable receptacle (not (openable-receptacle ?z)) (or Syntax error, the first parameter of `openable`should be a furnitureAppliance but here a smallReceptacle is given (not (openable ?z)) (opened-receptacle ?z) ) Missing precondition: there should be no other object stacked on top of ?z if ?z is stackable ) Effects: (and\\n(not (object-on ?x ?y)) (not (object-in ?x ?z)) (not (robot-hand-empty)) (robot-holding ?x) )\\nA.6.17 Open a small receptacle such as a lunch box with a lid\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to open a small receptacle receptacle_1 (e.g. small storage boxes or lunch boxes with lids). For example, the robot opens lunch_box_2, or the robot opens storage_box_1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user (a) This action is only applicable for receptacles that are openable. (b) The receptacle needs to be placed on a furniture piece that has an open and flat surface to allow enough space for manipulation (e.g., countertops or tables).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"(c) The receptacle should not be stacked on top of other objects if it is stackable.\\nGPT-4: Open a small receptacle such as a lunch box with a lid Parameters:\\n1. ?x -smallReceptacle: the small receptacle to be opened 2. ?y -furnitureAppliance: the furniture or appliance on which the small receptacle is placed Preconditions:\\n(and (robot-at ?y) (object-on ?x ?y) (openable-receptacle ?x) (flat-surface ?y) (robot-hand-empty) Inaccurate precondition: should only apply if ?x is stackable (not (object-stacked ?x)) (closed-receptacle ?x) Missing precondition: there should be no other object stacked on top of ?x ) Effects: (and (not (closed-receptacle ?x)) (opened-receptacle ?x) ) (object-in ?f ?b) (object-sliced ?f) (is-small-receptacle ?b) Incorrect precondition: as mentioned in the prompt, the blender should be turned off at the beginning (appliance-on ?b) (robot-hand-empty) Missing precondition: ?b should be a blender ) Effects: (and\\n(not (appliance-on ?b)) (object-mashed ?f) Missing effect: ?f is no longer pickupable )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.6.20 Wash an object\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action Description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to wash an object (e.g., fruits and cloths) in a sink or basin. For example, the robot washed apple_1, or the robot washed romaine_lettuce_1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user (a) The object has to be something washable.\\n(b) The robot should hold the object when washing it.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Wash an object\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?x -householdObject: the object to be washed 2. ?y -furnitureAppliance: the sink or basin where the object will be washed Preconditions:\\n(and (robot-at ?y) (washable ?x) (robot-holding ?x) Missing precondition: ?y should be a sink or basin ) Effects: (and (not (object-dirty ?x)) (object-washed ?x) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.6.21 Wipe a surface\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to wipe and clean the surface of a piece of furniture or an appliance, such as a dining table, a mirror, a sink, or a bathtub, with a cloth. For example, the robot wipes dining_table_1 with cloth_1, the robot wipes mirror_1 with cloth_2.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Wipe a surface\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?x -furnitureAppliance: the furniture or appliance to be wiped and cleaned 2. ?c -householdObject: the cloth to be used for wiping and cleaning Preconditions:\\n(and (robot-at ?x) (robot-holding ?c) (washable ?c) (object-washed ?c) Incorrect precondition: ?x doesn't need to have a flat surface (flat-surface ?x) Missing precondition: ?c should be a cloth ) Effects: (and (not (object-washed ?c)) (object-dirty ?c) (furniture-cleaned ?x) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.6.22 Vacuum a carpet\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to vacuum a carpet with a handheld vacuum cleaner. For example, the robot uses vacuum_cleaner_1 to vacuum carpet_2, the robot uses vacuum_cleaner_2 to vacuum carpet_1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user (a) You need to make sure the dust bin of the vacuum cleaner is not full before executing this action.\\n→ (b) You may assume the robot is holding the vacuum cleaner at the beginning. (c) The dust bin of the vacuum cleaner will be full of dust.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT4: Vacuum a carpet\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?v -householdObject: the handheld vacuum cleaner to use 2. ?c -furnitureAppliance: the carpet to vacuum Preconditions:\\n(and (robot-holding ?v) (robot-at ?c) (pickupable ?v) (not (dustbin-full ?v)) (not (carpet-clean ?c)) Missing precondition: ?v should be a handheld vacuum cleaner Missing precondition: ?c should be a carpet ) Effects: (and (carpet-clean ?c) (dustbin-full ?v) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.6.23 Empty a vacuum cleaner\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to empty a vacuum cleaner's dust bin and dump the dust into a trash can. For example, the robot empties vacuum_cleaner_1 and dumps the dust into trash_can_2, or the robot empties vacuum_cleaner_2 and dumps the dust into trash_can 1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"2. There is a syntax error: the second parameter of `object-on`should be a furnitureAppliance, but a householdObject was given. Please use the correct predicate or devise new one(s) if needed → →\\n3. Incorrect precondition: ?f needs not be pickupable 4. There is a missing precondition: ?p has to be a pan 5. There is a missing precondition: ?s has to be a stove burner 6. There is an incorrect precondition: a stove or burner ?s doesn't need to have a flat surface Please revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n----------------------------------------GPT-4:\\n1. ?f -householdObject: the food to be heated 2. ?p -smallReceptacle: the pan to be used for heating the food 3. ?s -furnitureAppliance: the stove burner on which the pan is placed Preconditions:\\n(and (robot-at ?s) (object-in ?f ?p) (object-on ?p ?s) (pan ?p) (stove-burner ?s) (robot-hand-empty) )\\nEffects: (and (not (pickupable ?f)) (food-heated-pan ?f) ) New Predicates: 1. (pan ?x -smallReceptacle): true if the small receptacle ?x is a pan 2. (stove-burner ?x -furnitureAppliance): true if the furniture or appliance ?x is a stove burner 3. (object-in ?x -householdObject ?y -smallReceptacle): true if the object ?x is in the small receptacle ?y\\n→ ---------------------------------------- Automated corrective feedback:\\nThe following predicate(s) have the same name(s) as existing predicate(s):\\n1. (object-in ?x -householdObject ?y -smallReceptacle), true if the object ?x is in the small receptacle ?y; existing predicate with the same name: (object-in ?x -householdObject ?z -smallReceptacle), true if the object ?x is in the small receptacle ?z\\n→ →\\nYou should reuse existing predicates whenever possible. If you are reusing existing predicate(s), you shouldn't list them under 'New Predicates'. If existing predicates are not enough and you are devising new predicate(s), please use names that are different from existing ones.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Please revise the PDDL model to fix this error.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Parameters:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"----------------------------------------GPT-4: 1. ?f -householdObject: the food to be heated 2. ?p -smallReceptacle: the pan to be used for heating the food 3. ?s -furnitureAppliance: the stove burner on which the pan is placed Preconditions:\\n(and (robot-at ?s) (object-in ?f ?p) (object-on ?p ?s) (pan ?p) (stove-burner ?s) (robot-hand-empty) )\\nEffects: (and (not (pickupable ?f)) (food-heated-pan ?f) ) New Predicates: 1. (pan ?x -smallReceptacle): true if the small receptacle ?x is a pan 2. (stove-burner ?x -furnitureAppliance): true if the furniture or appliance ?x is a stove burner\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.8 Logistics: Constructing PDDL Models\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The AI agent here is a logistics planner that has to plan to transport packages within the locations in a city through a truck and between cities through an airplane. Within a city, the locations are directly linked, allowing trucks to travel between any two of these locations. Similarly, cities are directly connected to each other allowing airplanes to travel between any two cities. Each city is equipped with one truck and has a designated location that functions as an airport. There are five types of objects: package, truck, plane, location, and city. There are multiple cities and each city can have multiple locations. Also, there is no limit to how many packages a truck or plane can carry (so in theory a truck or plane can carry an infinite number of packages).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.8.1 Load a package into a truck\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the agent to load a package into a truck. For example, load a package_1 into a truck_1. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the agent to unload a package from a truck. For example, unload a package_1 from a truck_1.\\n→ GPT-4: Unload a package from a truck Parameters:\\n1. ?p -package: the package to unload 2. ?t -truck: the truck from which the package will be unloaded 3. ?l -location: the location where the package will be unloaded Preconditions:\\n(and (package-in-truck ?p ?t) (truck-at ?t ?l) )\\nEffects: (and (not (package-in-truck ?p ?t)) (package-at ?p ?l) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.8.3 Load a package into an airplane\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the agent to load a package into an airplane. For example, load a package_1 into an airplane_1.\\n→ GPT-4: Load a package into an airplane Parameters:\\n1. ?p -package: the package to load into the airplane 2. ?a -plane: the airplane to load the package into 3. ?l -location: the location where the package and airplane are present Preconditions:\\n(and (package-at ?p ?l) (plane-at ?a ?l) (not (package-in-plane ?p ?a)) )\\nEffects: (and (not (package-at ?p ?l)) (package-in-plane ?p ?a) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.8.4 Unload a package from an airplane\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the agent to unload a package from an airplane. For example, unload a package_1 from an airplane_1.\\n→ GPT-4: Unload a package from an airplane Parameters:\\n1. ?p -package: the package to unload 2. ?a -plane: the airplane from which the package will be unloaded 3. ?l -location: the location where the airplane is and the package will be unloaded Preconditions:\\n(and (package-in-plane ?p ?a) (plane-at ?a ?l) )\\nEffects: (and (not (package-in-plane ?p ?a)) (package-at ?p ?l) )\\nA.8.5 Drive a truck from one location to another in a city\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the agent to drive a truck from one location to another in a city. For example, drive a truck_1 from location_1 to location_2 in city_1.\\n→ GPT-4: Drive a truck from one location to another in a city Parameters:\\n1. ?t -truck: the truck to be driven 2. ?from -location: the starting location of the truck 3. ?to -location: the destination location of the truck 4. ?c -city: the city in which the truck is being driven Preconditions:\\n(and (truck-at ?t ?from) (location-in-city ?from ?c) (location-in-city ?to ?c) (not (= ?from ?to)) )\\nEffects: (and (not (truck-at ?t ?from)) (truck-at ?t ?to) (move-packages ?t ?from ?to) )\\nA.8.6 Fly an airplane from one city to another\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the agent to fly an airplane from one city's airport to another. The airports are locations in the city. For example, fly an airplane_1 from location_0 to location_1.\\n→ GPT-4: Fly an airplane from one city to another Parameters:\\n1. ?a -plane: the airplane to fly 2. ?from -location: the departure location (airport) in the origin city 3. ?to -location: the destination location (airport) in the destination city 4. ?c1 -city: the origin city 5. ?c2 -city: the destination city Preconditions:\\n(and (plane-at ?a ?from) (location-in-city ?from ?c1) (location-in-city ?to ?c2) (not (= ?c1 ?c2)) Missing precondition: ?from should be an airport Missing precondition: ?to should be an airport ) Effects: (and (not (plane-at ?a ?from)) (plane-at ?a ?to) (move-packages-plane ?a ?from ?to) )\\nA.8.7 Examples of PDDL action models constructed by GPT-3.5-Turbo GPT-3.5-Turbo: Fly an airplane from one city to another Parameters: 1. ?a -plane: the airplane to fly 2. ?c1 -city: the city to fly from 3. ?c2 -city: the city to fly to 4. ?l1 -location: the airport location in ?c1 A.8.8 The initial set of predicates extracted by GPT-4\\nThe initial set of predicates extracted by GPT-4\\n1. (package-at ?p -package ?l -location): true if the package ?p is at the location ?l 2. (truck-at ?t -truck ?l -location): true if the truck ?t is at the location ?l 3. (package-in-truck ?p -package ?t -truck): true if the package ?p is in the truck ?t 4. (plane-at ?a -plane ?l -location): true if the airplane ?a is at the location ?l 5. (location-in-city ?l -location ?c -city): true if the location ?l is in the city ?c 6. (package-in-plane ?p -package ?a -plane): true if the package ?p is in the airplane ?a 7. (move-packages ?t -truck ?from -location ?to -location): true if all packages in the truck ?t are moved from location ?from to location ?to. This predicate is used to represent the effect of moving all packages in the truck when the truck is driven from one location to another.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ → →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"8. (move-packages-plane ?a -plane ?from -location ?to -location): true if all packages in the airplane ?a are moved from location ?from to location ?to. This predicate is used to represent the effect of moving all packages in the airplane when the airplane is flown from one location to another → → →\\nA.9 Logistics: Correcting PDDL Models A.9.1 Fly an airplane from one city to another GPT-4: Fly an airplane from one city to another User: ... The original prompt for PDDL construction ... Parameters:\\n-------------------------------------GPT-4: Parameters:\\n1. ?a -plane: the airplane to fly 2. ?from -location: the departure location (airport) in the origin city 3. ?to -location: the destination location (airport) in the destination city 4. ?c1 -city: the origin city 5. ?c2 -city: the destination city Preconditions:\\n(and (plane-at ?a ?from)\\nMissing precondition: ?from should be an airport Missing precondition: ?to should be an airport ) Effects: (and (not (plane-at ?a ?from)) (plane-at ?a ?to) (move-packages-plane ?a ?from ?to) ) --------------------------------------- ----------------------------------------GPT-3.5-Turbo:\\n1. ?a -plane: the airplane to fly 2. ?from -location: the departure location (airport) in the origin city 3. ?to -location: the destination location (airport) in the destination city 4. ?c1 -city: the origin city 5. ?c2 -city: the destination city Preconditions:\\n(and (plane-at ?a ?from) (location-in-city ?from ?c1) (location-in-city ?to ?c2) (not (= ?c1 ?c2)) (airport ?to) Missing precondition: ?from should be an airport ) Effects: (and (not (plane-at ?a ?from)) (plane-at ?a ?to) (move-packages-plane ?a ?from ?to) ) New Predicates: 1. (airport ?l -location): true if the location ?l is an airport.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.10 Tyreworld: Constructing PDDL Models\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The AI agent here is a robot that has to replace a flat tyre with a spare one. This involves fetching the tools (i.e., wrench, jack, pump) from the boot, undoing the nuts on the flat tyre, jacking up the (appropriate) hub(s), removing the tyre, doing up the spare one, etc. There are three major object types: 'small_object', 'container' and 'hub'. The object type 'small_object' covers tools, wheels and nuts. The 'small_object' object type has three subtypes: 'tool', 'wheel' and 'nut'. The subtype 'tool' covers tools like wrenches, jacks, pumps etc. The subtype 'wheel' covers various wheels. The subtype 'nut' covers various nuts. The object type 'container' covers storage spaces like the boot in a car. The object type 'hub' covers the hubs in the wheels of the car. Note that there is no restriction on how many objects the AI agent (i.e., the robot) can carry. Also note that each hub has only one nut.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.10.1 Open a container\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to open a container. For example, the robot can open a boot. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Open a container\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to jack up a hub. For example, the robot can jack up hub_1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Jack up a hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters: 1. ?h -hub: the hub to jack up 2. ?j -small_object: the jack to use for jacking up the hub Preconditions: (and (is-hub ?h) (hub-on-ground ?h) (is-jack ?j) (robot-holding ?j) )\\nEffects: (and\\n(not (hub-on-ground ?h)) (hub-jacked-up ?h) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.10.8 Jack down a hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to jack down a hub. For example, the robot can jack down hub_1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Jack down a hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters: 1. ?h -hub: the hub to jack down 2. ?j -small_object: the jack used to jack down the hub Preconditions: (and (is-hub ?h) (is-jack ?j) (robot-holding ?j) (hub-jacked-up ?h) )\\nEffects: (and (not (hub-jacked-up ?h)) (hub-on-ground ?h) )\\nA.10.9 Unfasten a hub\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to unfasten a hub and remove the nut on it. For example, the robot can unfasten hub_1 that has nut_1. At the end of this action, the hub should be unfastened, and the nut should be detached from the hub and held by the robot. To safely perform this action, the hub needs to be jacked up.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ → →\",\n \"publication_ref\": [\n \"b11\"\n ],\n \"table_ref\": [],\n \"text\": \"(robot-holding ?w) (robot-holding ?p) (wheel-deflated ?w) (wheel-intact ?w) )\\nEffects: (and (not (wheel-deflated ?w)) (wheel-inflated ?w) )\\nA.12 LLM planners back-prompted by VAL using LLM-acquired PDDL model A. 12 Here is the task. The house contains the following large furniture pieces and appliances: drawer_1, drawer_2, ..., dish_washer_1, garbage_can_1. Note that cutting boards are viewed as small receptacles, and blenders are small appliances (and thus small receptacles).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The current states of the household objects are as follows:\\n1. drawer_1 is opened 2. drawer_2 is opened ...\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"robot is holding mug_1\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Your task is to: put apple_2 on side_table_2. Your plan is:\\nAn example prompt for the Logistics domain\\nYou are a logistics planner that has to plan to transport packages within the locations in a city through a truck and between cities through an airplane. Within a city, the locations are directly linked, allowing trucks to travel between any two of these locations. Similarly, cities are directly connected to each other allowing airplanes to travel between any two cities. Each city is equipped with one truck and has a designated location that functions as an airport. There are five types of objects: package, truck, plane, location, and city. There are multiple cities and each city can have multiple locations. Also, there is no limit to how many packages a truck or plane can carry (so in theory a truck or plane can carry an infinite number of packages).\\n→ → → → → → → →\\nThe following actions are available to you, and you should strictly follow the output format (demonstrated in the \\\"example output\\\") of each action: → Preconditions: (a) the package ?p is at the location ?l; (b) the truck ?t is at the location ?l. Effects: (a) the package ?p is not at the location ?l; (b) the package ?p is in the truck ?t.\\n2. unload a package ?p from a truck ?t at location ?l (example output: unload package_1 from a truck truck_1 at location_1).\\n→ Parameters: (a) ?p: the package to unload; (b) ?t: the truck from which the package will be unloaded; (c) ?l: the location where the package will be unloaded.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Preconditions: (a) the package ?p is in the truck ?t; (b) the truck ?t is at the location ?l. Effects: (a) the package ?p is not in the truck ?t; (b) the package ?p is at the location ?l. ...\\nHere are two examples.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Example 1\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The environment contains the following objects: plane_0 (type: plane), city_0 (type: city), truck_0 (type: truck), location_0 (type: location), location_1 (type: location), package_0 (type: package).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The current states of the objects are as follows:\\n1. location_0 is in city city_0 2. location_1 is in city city_0 3. truck_0 is at location location_0 4. plane_0 is at location location_0 5. package_0 is at location location_1 6. location_0 is an airport Your task is to: Transport package package_0 to location location_0 Your plan is:\\nStep 1: > think: First, I need to drive the truck to the location of the package and load the package. > action: drive truck_0 from location_0 to location_1 in city_0\\nStep 2: > action: load package_0 into a truck truck_0 at location_1\\nStep 3: > think: Now I need to drive the truck to the destination and unload the package. > action: drive truck_0 from location_1 to location_0 in city_0\\nStep 4: > action: unload package_0 from a truck truck_0 at location_0 Example 2 ...\\nHere is the task. The environment contains the following objects: plane_0 (type: plane), city_0 (type: city), city_1 (type: city), truck_0 (type: truck), truck_1 (type: truck), location_0 (type: location), location_1 (type: location), location_2 (type: location), package_0 (type: package)\\nA.12.4 Examples of validation feedback and instructions with ordering constraints Make heated mashed potato_1 & Remember to mash the potato at the end after heating it User: ... The original prompt for the LLM planner ... The current states of the household objects are as follows: ... 70. robot at dining_table_1 71. robot is holding mug_1\\nYour task is to: make heated and mashed potato_1, remember to mash the potato at the end after heating it.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Your plan is:\\n-------------------------------------GPT-4:\\nStep 1: > think: First, I need to find a cutting board and put it on a flat surface (countertop_1 or countertop_2). In the current state, cutting_board_1 is in/on countertop_2.\\n→ > action: go from dining_table_1 to countertop_2\\nStep 2: > think: Now I need to find potato_1. In the current state, potato_1 is in fridge_1 and fridge_1 is closed. I need to get the potato and then put it on the cutting board.\\n→ > action: go from countertop_2 to fridge_1\\nStep 3:\\n> action: open furniture/appliance fridge_1\\nStep 4: > action: pick up potato_1 in/on fridge_1\\nStep 5: > action: go from fridge_1 to countertop_2\\nStep 6: > action: put potato_1 in/on cutting_board_1 at countertop_2\\nStep 7: > think: Now I need to find a knife. In the current state, knife_1 is in cabinet_3 and cabinet_3 is closed. I need to get the knife and then slice the potato.\\n→ > action: go from countertop_2 to cabinet_3\\nStep 8:\\n> action: open furniture/appliance cabinet_3\\nStep 9: > action: pick up knife_1 in/on cabinet_3\\nStep 10: > action: go from cabinet_3 to countertop_2\\nStep 11: > action: slice potato_1 with knife_1 and cutting_board_1 at countertop_2\\nStep 12: > think: Now I need to heat the sliced potato. In the current state, microwave_1 is closed. I need to put the sliced potato in a microwave-safe container and then heat it in the microwave. Step 24: > action: go from microwave_1 to countertop_1\\nStep 25: > think: Now I need to find a blender to mash the heated potato. In the current state, blender_1 is in/on countertop_1 and blender_1 is turned off. I need to put the heated potato in the blender and then mash it.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"> action: transfer potato_1 from bowl_2 to blender_1 at countertop_1\\nStep 26: > action: mash potato_1 with blender_1 at countertop_1\\nStep 27: > think: Task complete. The heated mashed potato_1 is now in the blender.\\n-------------------------------------Validation feedback:\\nThe action at step 3 is not executable due to unmet precondition(s). Here are the unsatisfied precondition(s):\\n→ The robot's gripper is not holding any object.\\nYour revised plan is:\\n-------------------------------------GPT-4: ... The revised plan ...\\nPlace fork_1 spoon_1 and knife_1 on dining_table_1 & Please take the fork first and the knife last User: ... The original prompt for the LLM planner ... The current states of the household objects are as follows: ... 24. lunch_box_2 in/on fridge_1 25. lunch_box_2 is not opened ... 59. spoon_1 in/on lunch_box_2 ... Your task is to: place fork_1, spoon_1, and knife_1 on dining_table_1, please take the fork first and the knife last.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Your plan is:\\n-------------------------------------GPT-4:\\nStep 1: > think: First, I need to put down the mug_1 I am currently holding. > action: go from dining_table_1 to side_table_1\\nStep 2: > action: put mug_1 in/on side_table_1\\nStep 3: > think: Now, I need to pick up fork_1 from dish_washer_1. > action: go from side_table_1 to dish_washer_1\\nStep 4:\\n> action: open furniture/appliance dish_washer_1\\nStep 5: > action: pick up fork_1 in/on dish_washer_1\\nStep 6: > think: I have the fork_1, now I need to place it on dining_table_1. > action: go from dish_washer_1 to dining_table_1\\nStep 7: > action: put fork_1 in/on dining_table_1\\nStep 8: > think: Next, I need to pick up spoon_1 from lunch_box_2 in fridge_1. > action: go from dining_table_1 to fridge_1\\nStep 9:\\n> action: open furniture/appliance fridge_1\\nStep 10: > action: pick up spoon_1 in/on lunch_box_2 at fridge_1\\nStep 11: > think: I have the spoon_1, now I need to place it on dining_table_1. > action: go from fridge_1 to dining_table_1\\nStep 12: > action: put spoon_1 in/on dining_table_1\\nStep 13: > think: Finally, I need to pick up knife_1 from cabinet_1. > action: go from dining_table_1 to cabinet_1\\nStep 14: > action: pick up knife_1 in/on cabinet_1\\nStep 15: > think: I have the knife_1, now I need to place it on dining_table_1. > action: go from cabinet_1 to dining_table_1\\nStep 16: > action: put knife_1 in/on dining_table_1\\nStep 17:\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgement\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This research was supported by ONR grants N00014-18-1-2442, N00014-18-1-2840, N00014-19-1-2119 and N00014-23-1-2409, AFOSR grant FA9550-18-1-0067, DARPA SAIL-ON grant W911NF-19-2-0006, and a JP Morgan AI Faculty Research Grant to Kambhampati. Sreedharan was supported in part by NSF grant 2303019.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.1 Broader impact on using LLMs\",\n \"publication_ref\": [\n \"b22\",\n \"b51\"\n ],\n \"table_ref\": [],\n \"text\": \"There is a general temptation to use LLMs for a variety of tasks, including plan generation. Given the fact that LLMs cannot guarantee the generation of correct plans, this can lead to safety and security issues downstream. Our approach of teasing a domain model from LLMs, and using it in conjunction with external sound planners aims to mitigate these safety concerns. Nevertheless, given that humans are still in charge of verifying the correctness of the domain models extracted from LLMs, there is still the possibility that an incorrect or undesirable domain model is inadvertently certified correct, leading to undesirable plans and agent behaviors down the line.\\nImproved explainability is another advantage of extracting explicit domain models. As opposed to just directly querying LLMs for plans, generating behaviors with (intermediate) symbolic models offers additional opportunities for explanation, drawing upon existing works in explainable AI [23]. In cases where the user's understanding does not align with the transcribed model, debugging and model reconciliation techniques such as D3wa+ [52] can also be directly applied.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.2 Additional discussion on alternative to action-by-action PDDL construction\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"One alternative to our action-by-action generation could be to include descriptions of all the actions in the prompt and require the LLM to construct the entire domain model in a single dialogue. This Additional information from the user (a) Note that the object to slice needs to be placed on a cutting board. A cutting board can be viewed as a special kind of receptacle. A.6.12 Heat food with a microwave\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to start a microwave and heat up the food inside. For example, the robot heats pizza_1 with microwave_2, the robot heats pasta_2 with microwave_1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user (a) The food to heat should be placed in a small receptacle (e.g., a plate or a bowl).\\n(b) The robot also needs to close the door of the microwave before taking this action. (c) Note that the food is no longer pickupable after it has been heated.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Heat food with a microwave\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?x -householdObject: the food to heat up 2. ?y -furnitureAppliance: the microwave to use 3. ?z -smallReceptacle: the small receptacle containing the food Preconditions:\\n(and (robot-at ?y) Incorrect precondition: the object should be in the receptacle not directly in the microwave (object-on ?x ?y) (object-in ?x ?z) (is-small-receptacle ?z) (object-on ?z ?y) Incorrect precondition: ?x needs not be pickupable (pickupable ?x) (closed ?y) (not (food-heated ?x)) Missing precondition: ?y should be a microwave ) Effects: (and (not (pickupable ?x)) (food-heated ?x) )\\nA.6.13 Heat food with pan\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to heat food which is heatable with a pan. For example, the robot heats pizza_1 with pan_2, the robot heats pasta_2 with pan_1.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user (a) Note that the food should be placed on the pan, and the pan needs to be placed on a stove burner before executing this action.\\n→ (b) A stove burner can be viewed as an immovable appliance, and a pan can be viewed as a small receptacle.\\n→ (c) Note that the food is no longer pickupable after it has been heated.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Heat food with pan\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?f -householdObject: the food to be heated Inaccurate object type, a pan ?p can be viewed as a small receptacle 2. ?p -householdObject: the pan to be used for heating the food 3. ?s -furnitureAppliance: the stove burner on which the pan is placed Preconditions:\\n(and (robot-at ?s) Syntax error: the second parameter of `object-on`should be a furnitureAppliance, but a householdObject was given (object-on ?f ?p) (object-on ?p ?s) Incorrect precondition: ?f needs not be pickupable (pickupable ?f) (is-small-receptacle ?p) Incorrect precondition: a stove or burner ?s doesn't need to have a flat surface (flat-surface ?s) (robot-hand-empty) Missing precondition: ?p has to be a pan Missing precondition: ?s has to be a stove burner ) Effects: (and (not (pickupable ?f)) (food-heated-pan ?f) )\\nGPT-4: Put an object onto or into a small receptacle like a bowl and plate Parameters: 1. ?x -householdObject: the object to put into/onto the small receptacle 2. ?y -furnitureAppliance: the furniture or appliance with a flat surface for manipulation 3. ?z -smallReceptacle: the small receptacle to put the object into/onto Preconditions: (and (robot-at ?y) Inaccurate precondition: object ?x can't be on the furnitureAppliance ?y when ?x is held by the robot (object-on ?x ?y) (pickupable ?x) Inaccurate precondition: this should only apply if ?z is stackable (not (object-stacked ?z)) (flat-surface ?y) (robot-holding ?x) (is-small-receptacle ?z) (object-on ?z ?y) Inaccurate precondition: this should only apply if ?z is stackable (object-clear ?z) Unnecessary precondition: the robot holding the object ?x already implies ?x is not in the receptacle ?z (not (object-in ?x ?z)) (or (not (openable-receptacle ?z)) (opened-receptacle ?z) ) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Effects: (and\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"(not (robot-holding ?x)) (not (object-on ?x ?y)) (robot-hand-empty) (object-in ?x ?z) )\\nA.6.16 Pick up an object on or in a small receptacle like a bowl and plate\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to pick up an object object_1 in some small receptacle receptacle_1 (e.g. storage boxes, lunch boxes, bowls, plates). For example, the robot picks up apple_1 in blow_1, or the robot picks up apple_2 in lunch_box_2.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user GPT-4: Pick up an object on or in a small receptacle like a bowl and plate Parameters:\\n1. ?x -householdObject: the object to pick up 2. ?y -furnitureAppliance: the furniture or appliance where the object is located 3. ?z -smallReceptacle: the small receptacle containing the object Preconditions:\\n(and (robot-at ?y) Incorrect precondition: the object ?x is in the receptacle ?z so ?x is not directly on the furniture or appliance ?y (object-on ?x ?y) (pickupable ?x) (object-in ?x ?z) (is-small-receptacle ?z) Inaccurate precondition: \\\"?z is not stacked on other objects\\\" should only apply if ?z is a stackable object (not (object-stacked ?z))\\nA.6.18 Close a small receptacle such as a lunch box with a lid\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to close a small receptacle receptacle_1 that is openable (e.g. small storage boxes or lunch boxes with lids). For example, the robot closes lunch_box_2, or the robot closes storage_box_1. A.6.19 Mash food with a blender\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to use a blender to mash some food in it. For example, the robot gets mashed apple_2 with blender_1, or the robot gets mashed banana_1 with blender_2.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Additional information from the user (a) The food needs to be sliced beforehand and the food should be placed inside the blender. (b) Note that the food remains in the blender after this action is performed. (c) You may also assume the blender is turned off before and after mashing the food.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Mash food with a blender\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters: Incorrect object type: a blender ?b should be a smallReceptacle 1. ?b -householdObject: the blender to be used for mashing the food 2. ?f -householdObject: the food to be mashed 3. ?y -furnitureAppliance: the furniture or appliance where the blender is located Preconditions:\\n(and (robot-at ?y) (object-on ?b ?y) Incorrect precondition: the food to mash should be in the blender ?b rather than directly on the furniture piece ?y (object-on ?f ?y)\\nAdditional information from the user (a) Note that the robot should hold the vacuum cleaner and stand by the trash can before executing this action. A.6. 25 The initial set of predicates extracted by GPT-4\\nThe initial set of predicates extracted by GPT-4 -------------------------------------GPT-4: Parameters: 1. ?f -householdObject: the food to be heated There is an inaccurate object type: a pan ?p can be viewed as a small receptacle 2. ?p -householdObject: the pan to be used for heating the food 3. ?s -furnitureAppliance: the stove burner on which the pan is placed Preconditions: (and (robot-at ?s) Syntax error: the second parameter of `object-on`should be a furnitureAppliance, but a householdObject was given (object-on ?f ?p) (object-on ?p ?s) Incorrect precondition: ?f needs not be pickupable (pickupable ?f) (is-small-receptacle ?p) Incorrect precondition: a stove or burner ?s doesn't need to have a flat surface (flat-surface ?s) (robot-hand-empty) Missing precondition: ?p has to be a pan Missing precondition: ?s has to be a stove burner ) Effects: (and (not (pickupable ?f)) (food-heated-pan ?f) ) ----------------------------------------Corrective feedback: There are some errors in the PDDL model: 1. There is an inaccurate object type, a pan ?p can be viewed as a small receptacle A.7.2 Slice an object GPT-4: Slice an object User: ... The original prompt for PDDL construction ... Parameters: -------------------------------------GPT-4: Parameters: 1. ?x -householdObject: the object to slice 2. ?k -householdObject: the knife to use for slicing 3. ?y -furnitureAppliance: the furniture or appliance with a flat surface for manipulation Incorrect object type, a cutting board should be viewed as a small receptacle as mentioned in the prompt 4. ?z -householdObject: the cutting board on which the object is placed Preconditions: (and (robot-at ?y) Incorrect precondition: ?x should be on the cutting board not directly on the furniture ?y (object-on ?x ?y) (object-on ?z ?y) (robot-holding ?k) (sliceable ?x) (cutting-board ?z) (knife ?k) (flat-surface ?y) (object-whole ?x) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Effects: (and (not (object-whole ?x)) (object-sliced ?x) ) ----------------------------------------Corrective feedback:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"There are some errors in the PDDL model: 1. There is an incorrect object type: a cutting board should be viewed as a small receptacle as mentioned in the prompt → 2. There is an incorrect precondition: ?x should be on the cutting board not directly on the furniture ?y\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Please revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Parameters:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"----------------------------------------GPT-4: 1. ?x -householdObject: the object to slice 2. ?k -householdObject: the knife to use for slicing 3. ?y -furnitureAppliance: the furniture or appliance with a flat surface for manipulation 4. ?z -smallReceptacle: the cutting board on which the object is placed Preconditions:\\n(and (robot-at ?y) (object-in ?x ?z) (object-on ?z ?y) (robot-holding ?k) (sliceable ?x) (cutting-board ?z) (knife ?k) (flat-surface ?y) (object-whole ?x) )\\nEffects: (and (not (object-whole ?x)) (object-sliced ?x) ) New Predicates: No newly defined predicate A.7.3 GPT-3.5-Turbo trying to correct the action model of \\\"slice an object\\\" GPT-3.5-Turbo trying to correct the action model of \\\"slice an object\\\" User: ... The original prompt for PDDL construction ... Parameters: -------------------------------------GPT-4: Parameters: 1. ?x -householdObject: the object to slice 2. ?k -householdObject: the knife to use for slicing 3. ?y -furnitureAppliance: the furniture or appliance with a flat surface for manipulation Incorrect object type, a cutting board should be viewed as a small receptacle as mentioned in the prompt 4. ?z -householdObject: the cutting board on which the object is placed Preconditions: (and (robot-at ?y) Incorrect precondition: ?x should be on the cutting board not directly on the furniture ?y (object-on ?x ?y) (object-on ?z ?y) (robot-holding ?k) (sliceable ?x) (cutting-board ?z) (knife ?k) (flat-surface ?y) (object-whole ?x) )\\nEffects: (and (not (object-whole ?x)) (object-sliced ?x) ) ----------------------------------------Corrective feedback: There are some errors in the PDDL model: 1. There is an incorrect object type: a cutting board should be viewed as a small receptacle as mentioned in the prompt → 2. There is an incorrect precondition: ?x should be on the cutting board not directly on the furniture ?y\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Please revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Parameters:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"----------------------------------------GPT-3.5-Turbo: 1. ?x -householdObject: the object to slice 2. ?k -householdObject: the knife to use for slicing 3. ?y -furnitureAppliance: the furniture or appliance with a flat surface for manipulation 4. ?z -smallReceptacle: the cutting board on which the object is placed Preconditions:\\n(and (robot-at ?y) (object-on ?z ?y) (object-in ?x ?z) (robot-holding ?k) (sliceable ?x) Should use the predicate (cutting-board ?z) (is-small-receptacle ?z) (knife ?k) (flat-surface ?y) (object-whole ?x) )\\nEffects: (and (not (object-whole ?x)) (object-sliced ?x) ) (location-in-city ?from ?c1) (location-in-city ?to ?c2) (not (= ?c1 ?c2)) Missing precondition: ?from should be an airport Missing precondition: ?to should be an airport ) Effects: (and (not (plane-at ?a ?from)) (plane-at ?a ?to) (move-packages-plane ?a ?from ?to) ) --------------------------------------- ---------------------------------------GPT-4: 1. ?a -plane: the airplane to fly 2. ?from -location: the departure location (airport) in the origin city 3. ?to -location: the destination location (airport) in the destination city 4. ?c1 -city: the origin city 5. ?c2 -city: the destination city Preconditions:\\n(and (plane-at ?a ?from) (location-in-city ?from ?c1) (location-in-city ?to ?c2) (not (= ?c1 ?c2)) (airport ?from) (airport ?to) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Effects: (and\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"(not (plane-at ?a ?from)) (plane-at ?a ?to) (move-packages-plane ?a ?from ?to) ) New Predicates: 1. (airport ?l -location): true if the location ?l is an airport A.9.2 GPT-3.5-Turbo trying to correct the action model of \\\"fly an airplane from one city to another\\\" GPT-3.5-Turbo trying to correct the action model of \\\"fly an airplane from one city to another\\\" User: ... The original prompt for PDDL construction ... Parameters:\\n- ------------------------------------GPT-4: Parameters:\\n1. ?a -plane: the airplane to fly 2. ?from -location: the departure location (airport) in the origin city 3. ?to -location: the destination location (airport) in the destination city 4. ?c1 -city: the origin city 5. ?c2 -city: the destination city Preconditions:\\n(and (plane-at ?a ?from) (location-in-city ?from ?c1) (location-in-city ?to ?c2) (not (= ?c1 ?c2))\\nPreconditions:\\n(and (container-closed ?c) )\\nEffects: (and (not (container-closed ?c))\\n(container-open ?c) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.10.2 Close a container\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to close a container. For example, the robot can close a boot. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Close a container\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.10.3 Fetch an object from a container\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to take an object from a container. For example, the robot can fetch jack from the boot or the robot can fetch wheel_1 from the boot. (not (object-in-container ?x ?c)) (robot-holding ?x) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.10.4 Put an object into a container\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to put an object into a container. For example, the robot can put away jack into the boot or the robot can put away wheel_1 into the boot. (not (robot-holding ?x)) (object-in-container ?x ?c) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.10.5 Loosen a nut in a hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to loosen a nut in a hub. For example, the robot can loosen nut_1 in hub_1. To safely perform this action, the hub needs to be on the ground (i.e., not jacked up).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Loosen a nut in a hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?n -nut: the nut to be loosened 2. ?h -hub: the hub where the nut is located 3. ?w -tool: the wrench to be used for loosening the nut Preconditions:\\n(and (is-nut ?n) (is-hub ?h) (is-wrench ?w) (nut-in-hub ?n ?h) (hub-on-ground ?h) (robot-holding ?w) )\\nEffects: (and (nut-loose ?n) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.10.6 Tighten a nut in a hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to tighten a nut in a hub. For example, the robot can tighten nut_1 in hub_1. To safely perform this action, the hub needs to be on the ground (i.e., not jacked up).\\n→ →\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Tighten a nut in a hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?n -nut: the nut to tighten 2. ?h -hub: the hub where the nut is to be tightened 3. ?w -tool: the wrench to use for tightening the nut Preconditions:\\n(and (is-nut ?n) (is-hub ?h) (is-wrench ?w) (robot-holding ?w) (nut-in-hub ?n ?h) (hub-on-ground ?h) (nut-loose ?n)\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Unfasten a hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?h -hub: the hub to unfasten 2. ?n -nut: the nut to detach from the hub Preconditions:\\n(and (is-hub ?h) (is-nut ?n) (nut-in-hub ?n ?h) (nut-loose ?n) (hub-jacked-up ?h) Incorrect precondition: the robot should not hold the nut if the nut is in the hub (robot-holding ?n) )\\nEffects: (and (not (nut-in-hub ?n ?h)) (hub-unfastened ?h) Missing effect: the robot should hold the nut at the end )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.10.10 Fasten a hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to fasten a hub by installing a nut on it. For example, the robot can fasten hub_1 that has nut_1. Note that the robot needs to install a nut into the hub, but the robot is not required to fasten the nut in this operation. To safely perform this action, the hub needs to be jacked up.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ → →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Fasten a hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?h -hub: the hub to fasten 2. ?n -nut: the nut to install in the hub Preconditions:\\n(and (is-hub ?h) (is-nut ?n) (robot-holding ?n) (hub-jacked-up ?h) (hub-unfastened ?h) Unnecessary precondition (not (nut-in-hub ?n ?h)) )\\nEffects: (and (not (robot-holding ?n)) (nut-in-hub ?n ?h) (not (hub-unfastened ?h)) Missing effect: the nut should be loose )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.10.11 Remove wheel from hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to remove a wheel from a hub. For example, the robot can remove wheel_1 from hub_1. Note that the hub needs to be unfastened in order to remove the wheel. To safely perform this action, the hub needs to be jacked up.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→ →\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Remove wheel from hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?w -wheel: the wheel to remove 2. ?h -hub: the hub from which the wheel is to be removed Preconditions:\\n(and (is-wheel ?w) (is-hub ?h) (wheel-on-hub ?w ?h) (hub-jacked-up ?h) (hub-unfastened ?h) )\\nEffects: (and\\n(not (wheel-on-hub ?w ?h)) (robot-holding ?w) )\\nA.10.12 Put wheel on hub\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to put a wheel on an unfastened hub. For example, the robot can put wheel_1 on hub_1. Note that the hub needs to be unfastened in order to put on the wheel. To safely perform this action, the hub needs to be jacked up.\\n→ →\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Put wheel on hub\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?w -wheel: the wheel to put on the hub 2. ?h -hub: the hub to put the wheel on Preconditions:\\n(and (robot-holding ?w) (is-wheel ?w) (is-hub ?h) (hub-jacked-up ?h) (hub-unfastened ?h) (not (wheel-on-hub ?w ?h)) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Effects: (and\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"(not (robot-holding ?w)) (wheel-on-hub ?w ?h) )\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.10.13 Inflate wheel\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Action description\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"This action enables the robot to inflate a wheel. For example, the robot can inflate wheel_1. Note that you need to make sure the wheel is intact.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"GPT-4: Inflate wheel\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n1. ?w -wheel: the wheel to inflate Preconditions: (and (is-wheel ?w) (robot-holding ?w) (wheel-deflated ?w) (wheel-intact ?w) Missing precondition: the robot should have a pump ) Effects: (and (not (wheel-deflated ?w)) (wheel-inflated ?w) )\\nA.10.14 Examples of PDDL action models constructed by GPT-3.5-Turbo GPT-3.5-Turbo: Open a containe ------------------------------------ ----------------------------------------Corrective feedback: There are some errors in the PDDL model: 1. There is an incorrect precondition: the robot is not holding the nut because the nut is in the hub → 2. There is a missing effect: the robot should hold the nut at the end Please revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n- -------------------------------------- ----------------------------------- ---------------------------------------\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"-Corrective feedback:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"There is an error in the PDDL model: 1. You allow the robot to inflate a wheel without having a pump, this is incorrect.\\nPlease revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Parameters:\\n----------------------------------------GPT-4: 1. ?w -wheel: the wheel to inflate 2. ?p -pump: the pump to inflate the wheel Preconditions:\\n(and (is-wheel ?w) (is-pump ?p) (robot-holding ?w) (robot-holding ?p) (wheel-deflated ?w) (wheel-intact ?w) )\\nEffects: (and (not (wheel-deflated ?w)) (wheel-inflated ?w) ) New Predicates: No newly defined predicate - --------------------------------------\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"-Automated corrective feedback:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"There is an error in the PDDL model: 1. There is an unknown object type `pump`for the parameter ?p Please revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"→\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Parameters:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"----------------------------------------GPT-4: 1. ?w -wheel: the wheel to inflate 2. ?p -small_object: the pump to inflate the wheel Preconditions:\\n(and (is-wheel ?w) (is-pump ?p) Your plan is:\\nA.12.2 Translation to admissible actions We extract the plan from the LLM planner as a completion to the prompt. Although the prompt explicitly states that the planner should strictly adhere to the demonstrated output format of each action, there are still cases where GPT-4 fails to follow these formats. As a result, extra post-processing is needed to translate the generated actions into admissible actions. To do this, we utilize another LLM (i.e., GPT-4) to perform the translation. In cases where certain actions have missing parameters, we present an error message to the LLM planner and ask it to regenerate the plan. The error message is structured as follows: \\\"There is an invalid output at step x. Please strictly follow the output format provided in the example output of each action. Your revised plan:\\\".\\nWe use a fixed prompt template for action translation in all the domains:\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Template of the prompt for action translation\",\n \"publication_ref\": [\n \"b17\"\n ],\n \"table_ref\": [],\n \"text\": \"Your task is to translate free-form action description to its corresponding structural format based on its semantic meaning. The structural output format is given in the example output of each candidate action. You should strictly follow the structural output format. When there is missing information in the free-form action description, you should output \\\"Not able to translate due to missing information\\\". A.12.3 Back-prompting LLM planners with validation feedback by VAL\\nWe consider three types of validation feedback that can be obtained from VAL [18]. One is unsatisfied precondition. VAL highlights unsatisfied precondition(s) in PDDL, which can be translated into natural language and provided as corrective feedback to the LLM planner. The feedback message is structured as follows: \\\"The action at step x is not executable due to unmet precondition(s). Here are the unsatisfied precondition(s):\\\". The second type of validation feedback is related to unmet goal condition(s), and the third type is related to the usage of invalid or illegal parameter(s) in some action(s).\\nFor translating PDDL into natural language, we use a fixed prompt template in all the domains:\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Template of the prompt for translating PDDL into natural language\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Your task is to translate planning domain definition language (PDDL) to natural language.\\nHere is an example:\\nPredicates: Step 20: > action: close furniture/appliance dish_washer_1 -------------------------------------Validation feedback: The action at step 10 is not executable due to unmet precondition(s). Here are the unsatisfied precondition(s) described in Pseudo-PDDL format:\\n→ (and The furniture or appliance fridge_1 has a flat surface for manipulation;\\n(or The small receptacle lunch_box_2 cannot be opened;\\nThe small receptacle lunch_box_2 is open; ) )\\nYour revised plan is: -------------------------------------GPT-4: ... The revised plan ...\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.13 Translating user instructions into PDDL goal specifications\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Template of the prompt for translating user instructions into PDDL goal specifications\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Your task is to translate natural language instructions to planning goals in planning domain definition language (PDDL). You should only use the given predicates to express the goal specification.\\n→ →\\nHere is an example from the classical BlocksWorld domain for demonstrating the output format.\\nDomain information: BlocksWorld is a planning domain in artificial intelligence. The AI agent here is a mechanical robot arm that can pick and place the blocks. Only one block may be moved at a time: it may either be placed on the table or placed atop another block. Because of this, any blocks that are, at a given time, under another block cannot be moved. There is only one type of object in this domain, and that is the block.\\nThe domain has the following objects: block_1 (type: block), block_2 (type: block), block_3 (type: block), block_4 (type: block), block_5 (type: block). \"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-11-02"},"doi":{"kind":"string","value":""},"references":{"kind":"list like","value":[{"authors":"Anthony Michael Ahn; Noah Brohan; Yevgen Brown; Omar Chebotar; Byron Cortes; Chelsea David; Keerthana Finn; Karol Gopalakrishnan; Alex Hausman; Herzog","journal":"","ref_id":"b0","title":"Do as i can, not as i say: Grounding language in robotic affordances","year":"2022"},{"authors":"Ankuj Arora; Humbert Fiorino; Damien Pellier; Marc Métivier; Sylvie Pesty","journal":"The Knowledge Engineering Review","ref_id":"b1","title":"A review of learning planning action models","year":"2018"},{"authors":"Pratyay Banerjee; Chitta Baral; Man Luo; Arindam Mitra; Kuntal Pal; Tran C Son; Neeraj Varshney","journal":"","ref_id":"b2","title":"Can transformers reason about effects of actions?","year":"2020"},{"authors":"Blai Bonet; Hector Geffner","journal":"","ref_id":"b3","title":"Learning first-order symbolic representations for planning from the structure of the state space","year":"2019"},{"authors":"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell","journal":"Advances in neural information processing systems","ref_id":"b4","title":"Language models are few-shot learners","year":"2020"},{"authors":"Ethan Callanan; Rebecca De Venezia; Victoria Armstrong; Alison Paredes; Tathagata Chakraborti; Christian Muise","journal":"","ref_id":"b5","title":"Macq: A holistic view of model acquisition techniques","year":"2022"},{"authors":"Hongyi Chen; Yilun Du; Yiye Chen; Joshua B Tenenbaum; Patricio A Vela","journal":"","ref_id":"b6","title":"Planning with sequence models through iterative energy minimization","year":"2023"},{"authors":"Shuo Cheng; Danfei Xu","journal":"","ref_id":"b7","title":"Guided skill learning and abstraction for long-horizon manipulation","year":"2022"},{"authors":"Thomas L Stephen N Cresswell; Margaret M Mccluskey; West","journal":"The Knowledge Engineering Review","ref_id":"b8","title":"Acquiring planning domain models using locm","year":"2013"},{"authors":"Danny Driess; Fei Xia; S M Mehdi; Corey Sajjadi; Aakanksha Lynch; Brian Chowdhery; Ayzaan Ichter; Jonathan Wahid; Quan Tompson; Tianhe Vuong; Yu","journal":"","ref_id":"b9","title":"Palm-e: An embodied multimodal language model","year":"2023"},{"authors":"E Richard; Nils J Fikes; Nilsson","journal":"Artificial intelligence","ref_id":"b10","title":"Strips: A new approach to the application of theorem proving to problem solving","year":"1971"},{"authors":"Alfonso Gerevini; Ivan Serina","journal":"AIPS","ref_id":"b11","title":"Lpg: A planner based on local search for planning graphs with action costs","year":"2002"},{"authors":"Malik Ghallab; Dana Nau; Paolo Traverso","journal":"Elsevier","ref_id":"b12","title":"Automated Planning: theory and practice","year":"2004"},{"authors":"Lin Guan; Sarath Sreedharan; Subbarao Kambhampati","journal":"PMLR","ref_id":"b13","title":"Leveraging approximate symbolic models for reinforcement learning via skill diversity","year":"2022"},{"authors":"Shibo Hao; Yi Gu; Haodi Ma; Joshua Jiahua Hong; Zhen Wang; Daisy Zhe Wang; Zhiting Hu","journal":"","ref_id":"b14","title":"Reasoning with language model is planning with world model","year":"2023"},{"authors":"Malte Helmert","journal":"Journal of Artificial Intelligence Research","ref_id":"b15","title":"The fast downward planning system","year":"2006"},{"authors":"Jörg Hoffmann; Bernhard Nebel","journal":"Journal of Artificial Intelligence Research","ref_id":"b16","title":"The ff planning system: Fast plan generation through heuristic search","year":"2001"},{"authors":"Richard Howey; Derek Long; Maria Fox","journal":"IEEE","ref_id":"b17","title":"Val: Automatic plan validation, continuous effects and mixed initiative planning using pddl","year":"2004"},{"authors":"Wenlong Huang; Pieter Abbeel; Deepak Pathak; Igor Mordatch","journal":"PMLR","ref_id":"b18","title":"Language models as zero-shot planners: Extracting actionable knowledge for embodied agents","year":"2022"},{"authors":"Wenlong Huang; Fei Xia; Ted Xiao; Harris Chan; Jacky Liang; Pete Florence; Andy Zeng; Jonathan Tompson; Igor Mordatch; Yevgen Chebotar","journal":"","ref_id":"b19","title":"Inner monologue: Embodied reasoning through planning with language models","year":"2022"},{"authors":"León Illanes; Xi Yan; Rodrigo Toro Icarte; Sheila A Mcilraith","journal":"","ref_id":"b20","title":"Symbolic plans as high-level instructions for reinforcement learning","year":"2020"},{"authors":"","journal":"","ref_id":"b21","title":"IPC. International planning competition","year":"1998"},{"authors":"Subbarao Kambhampati; Sarath Sreedharan; Mudit Verma; Yantian Zha; Lin Guan","journal":"","ref_id":"b22","title":"Symbols as a lingua franca for bridging human-ai chasm for explainable and advisable ai systems","year":"2022"},{"authors":"Subbarao Kambhampati; Karthik Valmeekam; Matthew Marquez; Lin Guan","journal":"","ref_id":"b23","title":"On the role of large language models in planning","year":"2023-07"},{"authors":"George Konidaris; Leslie Pack Kaelbling; Tomas Lozano-Perez","journal":"Journal of Artificial Intelligence Research","ref_id":"b24","title":"From skills to symbols: Learning symbolic representations for abstract high-level planning","year":"2018"},{"authors":"Kenneth Li; K Aspen; David Hopkins; Fernanda Bau; Hanspeter Viégas; Martin Pfister; Wattenberg","journal":"","ref_id":"b25","title":"Emergent world representations: Exploring a sequence model trained on a synthetic task","year":"2023"},{"authors":"Shuang Li; Xavier Puig; Chris Paxton; Yilun Du; Clinton Wang; Linxi Fan; Tao Chen; De-An Huang; Ekin Akyürek; Anima Anandkumar","journal":"Advances in Neural Information Processing Systems","ref_id":"b26","title":"Pre-trained language models for interactive decision-making","year":"2022"},{"authors":"Jacky Liang; Wenlong Huang; Fei Xia; Peng Xu; Karol Hausman; Brian Ichter; Pete Florence; Andy Zeng","journal":"","ref_id":"b27","title":"Code as policies: Language model programs for embodied control","year":"2022"},{"authors":"Kevin Lin; Christopher Agia; Toki Migimatsu; Marco Pavone; Jeannette Bohg","journal":"","ref_id":"b28","title":"Text2motion: From natural language instructions to feasible plans","year":"2023"},{"authors":"Bo Liu; Yuqian Jiang; Xiaohan Zhang; Qiang Liu; Shiqi Zhang; Joydeep Biswas; Peter Stone","journal":"","ref_id":"b29","title":"Llm+ p: Empowering large language models with optimal planning proficiency","year":"2023"},{"authors":"Man Luo; Shrinidhi Kumbhar; Mihir Parmar; Neeraj Varshney; Pratyay Banerjee; Somak Aditya; Chitta Baral","journal":"","ref_id":"b30","title":"Towards logiglue: A brief survey and a benchmark for analyzing logical reasoning capabilities of language models","year":"2023"},{"authors":"Qing Lyu; Shreya Havaldar; Adam Stein; Li Zhang; Delip Rao; Eric Wong; Marianna Apidianaki; Chris Callison-Burch","journal":"","ref_id":"b31","title":"Faithful chain-of-thought reasoning","year":"2023"},{"authors":"Drew Mcdermott; Malik Ghallab; Adele E Howe; Craig A Knoblock; Ashwin Ram; Manuela M Veloso; Daniel S Weld; David E Wilkins","journal":"","ref_id":"b32","title":"Pddl-the planning domain definition language","year":"1998"},{"authors":"Kolby Nottingham; Prithviraj Ammanabrolu; Alane Suhr; Yejin Choi; Hannaneh Hajishirzi; Sameer Singh; Roy Fox","journal":"","ref_id":"b33","title":"Do embodied agents dream of pixelated sheep?: Embodied decision making using language guided world modelling","year":"2023"},{"authors":"Alberto Olmo; Sarath Sreedharan; Subbarao Kambhampati","journal":"","ref_id":"b34","title":"Gpt3-to-plan: Extracting plans from text using gpt-3","year":"2021"},{"authors":" Openai","journal":"","ref_id":"b35","title":"Introducing chatgpt by openai","year":"2022"},{"authors":" Openai","journal":"","ref_id":"b36","title":"","year":"2023"},{"authors":"Vishal Pallagani; Bharath Muppasani; Keerthiram Murugesan; Francesca Rossi; Lior Horesh; Biplav Srivastava; Francesco Fabiano; Andrea Loreggia","journal":"","ref_id":"b37","title":"Plansformer: Generating symbolic plans using transformers","year":"2022"},{"authors":"Liangming Pan; Alon Albalak; Xinyi Wang; William Yang; Wang ","journal":"","ref_id":"b38","title":"Logic-lm: Empowering large language models with symbolic solvers for faithful logical reasoning","year":"2023"},{"authors":"Xavier Puig; Kevin Ra; Marko Boben; Jiaman Li; Tingwu Wang; Sanja Fidler; Antonio Torralba","journal":"","ref_id":"b39","title":"Virtualhome: Simulating household activities via programs","year":"2018"},{"authors":"Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark","journal":"PMLR","ref_id":"b40","title":"Learning transferable visual models from natural language supervision","year":"2021"},{"authors":"Vanya Shreyas Sundara Raman; Eric Cohen; Ifrah Rosen; David Idrees; Stefanie Paulius; Tellex","journal":"","ref_id":"b41","title":"Planning with large language models via corrective re-prompting","year":"2022"},{"authors":"Scott Reed; Konrad Zolna; Emilio Parisotto; Sergio Gomez Colmenarejo; Alexander Novikov; Gabriel Barth-Maron; Mai Gimenez; Yury Sulsky; Jackie Kay; Jost Tobias Springenberg","journal":"","ref_id":"b42","title":"A generalist agent","year":"2022"},{"authors":"Jonathan Russell Stuart; Norvig","journal":"","ref_id":"b43","title":"Artificial intelligence: a modern approach","year":"2003"},{"authors":"Timo Schick; Jane Dwivedi-Yu; Roberto Dessì; Roberta Raileanu; Maria Lomeli; Luke Zettlemoyer; Nicola Cancedda; Thomas Scialom","journal":"","ref_id":"b44","title":"Toolformer: Language models can teach themselves to use tools","year":"2023"},{"authors":"Noah Shinn; Beck Labash; Ashwin Gopinath","journal":"","ref_id":"b45","title":"Reflexion: an autonomous agent with dynamic memory and self-reflection","year":"2023"},{"authors":"Mohit Shridhar; Xingdi Yuan; Marc-Alexandre Côté; Yonatan Bisk; Adam Trischler; Matthew Hausknecht","journal":"","ref_id":"b46","title":"ALFWorld: Aligning Text and Embodied Environments for Interactive Learning","year":"2021"},{"authors":"Tom Silver; Varun Hariprasad; Reece S Shuttleworth; Nishanth Kumar; Tomás Lozano-Pérez; Leslie Pack; Kaelbling ","journal":"","ref_id":"b47","title":"Pddl planning with pretrained large language models","year":"2022"},{"authors":"Ron M Simpson; Diane E Kitchin; Thomas Leo Mccluskey","journal":"The Knowledge Engineering Review","ref_id":"b48","title":"Planning domain definition using gipo","year":"2007"},{"authors":"Ishika Singh; Valts Blukis; Arsalan Mousavian; Ankit Goyal; Danfei Xu; Jonathan Tremblay; Dieter Fox; Jesse Thomason; Animesh Garg","journal":"","ref_id":"b49","title":"Progprompt: Generating situated robot task plans using large language models","year":"2022"},{"authors":"Hee Chan; Jiaman Song; Clayton Wu; Brian M Washington; Wei-Lun Sadler; Yu Chao; Su","journal":"","ref_id":"b50","title":"Llm-planner: Few-shot grounded planning for embodied agents with large language models","year":"2022"},{"authors":"Sarath Sreedharan; Tathagata Chakraborti; Christian Muise; Yasaman Khazaeni; Subbarao Kambhampati","journal":"","ref_id":"b51","title":"-d3wa+-a case study of xaip in a model acquisition task for dialogue planning","year":"2020"},{"authors":"Kaya Stechly; Matthew Marquez; Subbarao Kambhampati","journal":"","ref_id":"b52","title":"Gpt-4 doesn't know it's wrong: An analysis of iterative prompting for reasoning problems","year":"2023"},{"authors":"Karthik Valmeekam; Matthew Marquez; Subbarao Kambhampati","journal":"","ref_id":"b53","title":"Can large language models really improve by self-critiquing their own plans?","year":"2023"},{"authors":"Karthik Valmeekam; Matthew Marquez; Sarath Sreedharan; Subbarao Kambhampati","journal":"","ref_id":"b54","title":"On the planning abilities of large language models-a critical investigation","year":"2023"},{"authors":"Zihao Wang; Shaofei Cai; Anji Liu; Xiaojian Ma; Yitao Liang","journal":"","ref_id":"b55","title":"Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents","year":"2023"},{"authors":"Lionel Wong; Gabriel Grand; Alexander K Lew; Noah D Goodman; K Vikash; Jacob Mansinghka; Joshua B Andreas; Tenenbaum","journal":"","ref_id":"b56","title":"From word models to world models: Translating from natural language to the probabilistic language of thought","year":"2023"},{"authors":"Yaqi Xie; Chen Yu; Tongyao Zhu; Jinbin Bai; Ze Gong; Harold Soh","journal":"","ref_id":"b57","title":"Translating natural language to planning goals with large-language models","year":"2023"},{"authors":"Qiang Yang; Kangheng Wu; Yunfei Jiang","journal":"Artificial Intelligence","ref_id":"b58","title":"Learning action models from plan examples using weighted max-sat","year":"2007"},{"authors":"Shunyu Yao; Jeffrey Zhao; Dian Yu; Nan Du; Izhak Shafran; Yuan Karthik R Narasimhan; Cao","journal":"","ref_id":"b59","title":"React: Synergizing reasoning and acting in language models","year":"2023"},{"authors":"Hankz Hankui Zhuo; Qiang Yang; Derek Hao Hu; Lei Li","journal":"Artificial Intelligence","ref_id":"b60","title":"Learning complex action models with quantifiers and logical implications","year":"2010"},{"authors":"Hankz Hankui Zhuo; Yantian Zha; Subbarao Kambhampati; Xin Tian","journal":"ACM Transactions on Intelligent Systems and Technology (TIST)","ref_id":"b61","title":"Discovering underlying plans based on shallow models","year":"2020"},{"authors":"","journal":"","ref_id":"b62","title":"robot-at ?x -furnitureAppliance): true if the robot is at the furniture or appliance ?x 2","year":""},{"authors":"","journal":"","ref_id":"b63","title":"openable ?y -furnitureAppliance): true if the furniture or appliance ?y can be opened","year":""},{"authors":"","journal":"","ref_id":"b64","title":"true if the furniture or appliance ?y is opened 7. (","year":""},{"authors":"","journal":"","ref_id":"b65","title":"true if the food ?x has been heated in the microwave 18","year":""},{"authors":"","journal":"","ref_id":"b66","title":"object-washed ?x -householdObject): true if the object ?x has been washed 26","year":""},{"authors":"","journal":"","ref_id":"b67","title":"robot-holding ?x -small_object): true if the robot is holding the small_object ?x 5","year":""},{"authors":"","journal":"","ref_id":"b68","title":"nut-in-hub ?n -nut ?h -hub): true if the nut ?n is in the hub ?h","year":""},{"authors":"","journal":"","ref_id":"b69","title":"is-jack ?x -small_object): true if the small_object ?x is a","year":""},{"authors":"","journal":"","ref_id":"b70","title":"is-wheel ?x -small_object): true if the small_object ?x is a wheel 15","year":""}],"string":"[\n {\n \"authors\": \"Anthony Michael Ahn; Noah Brohan; Yevgen Brown; Omar Chebotar; Byron Cortes; Chelsea David; Keerthana Finn; Karol Gopalakrishnan; Alex Hausman; Herzog\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Do as i can, not as i say: Grounding language in robotic affordances\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ankuj Arora; Humbert Fiorino; Damien Pellier; Marc Métivier; Sylvie Pesty\",\n \"journal\": \"The Knowledge Engineering Review\",\n \"ref_id\": \"b1\",\n \"title\": \"A review of learning planning action models\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Pratyay Banerjee; Chitta Baral; Man Luo; Arindam Mitra; Kuntal Pal; Tran C Son; Neeraj Varshney\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"Can transformers reason about effects of actions?\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Blai Bonet; Hector Geffner\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Learning first-order symbolic representations for planning from the structure of the state space\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell\",\n \"journal\": \"Advances in neural information processing systems\",\n \"ref_id\": \"b4\",\n \"title\": \"Language models are few-shot learners\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Ethan Callanan; Rebecca De Venezia; Victoria Armstrong; Alison Paredes; Tathagata Chakraborti; Christian Muise\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"Macq: A holistic view of model acquisition techniques\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hongyi Chen; Yilun Du; Yiye Chen; Joshua B Tenenbaum; Patricio A Vela\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Planning with sequence models through iterative energy minimization\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Shuo Cheng; Danfei Xu\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Guided skill learning and abstraction for long-horizon manipulation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Thomas L Stephen N Cresswell; Margaret M Mccluskey; West\",\n \"journal\": \"The Knowledge Engineering Review\",\n \"ref_id\": \"b8\",\n \"title\": \"Acquiring planning domain models using locm\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"Danny Driess; Fei Xia; S M Mehdi; Corey Sajjadi; Aakanksha Lynch; Brian Chowdhery; Ayzaan Ichter; Jonathan Wahid; Quan Tompson; Tianhe Vuong; Yu\",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"Palm-e: An embodied multimodal language model\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"E Richard; Nils J Fikes; Nilsson\",\n \"journal\": \"Artificial intelligence\",\n \"ref_id\": \"b10\",\n \"title\": \"Strips: A new approach to the application of theorem proving to problem solving\",\n \"year\": \"1971\"\n },\n {\n \"authors\": \"Alfonso Gerevini; Ivan Serina\",\n \"journal\": \"AIPS\",\n \"ref_id\": \"b11\",\n \"title\": \"Lpg: A planner based on local search for planning graphs with action costs\",\n \"year\": \"2002\"\n },\n {\n \"authors\": \"Malik Ghallab; Dana Nau; Paolo Traverso\",\n \"journal\": \"Elsevier\",\n \"ref_id\": \"b12\",\n \"title\": \"Automated Planning: theory and practice\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Lin Guan; Sarath Sreedharan; Subbarao Kambhampati\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b13\",\n \"title\": \"Leveraging approximate symbolic models for reinforcement learning via skill diversity\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Shibo Hao; Yi Gu; Haodi Ma; Joshua Jiahua Hong; Zhen Wang; Daisy Zhe Wang; Zhiting Hu\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Reasoning with language model is planning with world model\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Malte Helmert\",\n \"journal\": \"Journal of Artificial Intelligence Research\",\n \"ref_id\": \"b15\",\n \"title\": \"The fast downward planning system\",\n \"year\": \"2006\"\n },\n {\n \"authors\": \"Jörg Hoffmann; Bernhard Nebel\",\n \"journal\": \"Journal of Artificial Intelligence Research\",\n \"ref_id\": \"b16\",\n \"title\": \"The ff planning system: Fast plan generation through heuristic search\",\n \"year\": \"2001\"\n },\n {\n \"authors\": \"Richard Howey; Derek Long; Maria Fox\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b17\",\n \"title\": \"Val: Automatic plan validation, continuous effects and mixed initiative planning using pddl\",\n \"year\": \"2004\"\n },\n {\n \"authors\": \"Wenlong Huang; Pieter Abbeel; Deepak Pathak; Igor Mordatch\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b18\",\n \"title\": \"Language models as zero-shot planners: Extracting actionable knowledge for embodied agents\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Wenlong Huang; Fei Xia; Ted Xiao; Harris Chan; Jacky Liang; Pete Florence; Andy Zeng; Jonathan Tompson; Igor Mordatch; Yevgen Chebotar\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Inner monologue: Embodied reasoning through planning with language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"León Illanes; Xi Yan; Rodrigo Toro Icarte; Sheila A Mcilraith\",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Symbolic plans as high-level instructions for reinforcement learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b21\",\n \"title\": \"IPC. International planning competition\",\n \"year\": \"1998\"\n },\n {\n \"authors\": \"Subbarao Kambhampati; Sarath Sreedharan; Mudit Verma; Yantian Zha; Lin Guan\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Symbols as a lingua franca for bridging human-ai chasm for explainable and advisable ai systems\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Subbarao Kambhampati; Karthik Valmeekam; Matthew Marquez; Lin Guan\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"On the role of large language models in planning\",\n \"year\": \"2023-07\"\n },\n {\n \"authors\": \"George Konidaris; Leslie Pack Kaelbling; Tomas Lozano-Perez\",\n \"journal\": \"Journal of Artificial Intelligence Research\",\n \"ref_id\": \"b24\",\n \"title\": \"From skills to symbols: Learning symbolic representations for abstract high-level planning\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Kenneth Li; K Aspen; David Hopkins; Fernanda Bau; Hanspeter Viégas; Martin Pfister; Wattenberg\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Emergent world representations: Exploring a sequence model trained on a synthetic task\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Shuang Li; Xavier Puig; Chris Paxton; Yilun Du; Clinton Wang; Linxi Fan; Tao Chen; De-An Huang; Ekin Akyürek; Anima Anandkumar\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b26\",\n \"title\": \"Pre-trained language models for interactive decision-making\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jacky Liang; Wenlong Huang; Fei Xia; Peng Xu; Karol Hausman; Brian Ichter; Pete Florence; Andy Zeng\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Code as policies: Language model programs for embodied control\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Kevin Lin; Christopher Agia; Toki Migimatsu; Marco Pavone; Jeannette Bohg\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Text2motion: From natural language instructions to feasible plans\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Bo Liu; Yuqian Jiang; Xiaohan Zhang; Qiang Liu; Shiqi Zhang; Joydeep Biswas; Peter Stone\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Llm+ p: Empowering large language models with optimal planning proficiency\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Man Luo; Shrinidhi Kumbhar; Mihir Parmar; Neeraj Varshney; Pratyay Banerjee; Somak Aditya; Chitta Baral\",\n \"journal\": \"\",\n \"ref_id\": \"b30\",\n \"title\": \"Towards logiglue: A brief survey and a benchmark for analyzing logical reasoning capabilities of language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Qing Lyu; Shreya Havaldar; Adam Stein; Li Zhang; Delip Rao; Eric Wong; Marianna Apidianaki; Chris Callison-Burch\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Faithful chain-of-thought reasoning\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Drew Mcdermott; Malik Ghallab; Adele E Howe; Craig A Knoblock; Ashwin Ram; Manuela M Veloso; Daniel S Weld; David E Wilkins\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Pddl-the planning domain definition language\",\n \"year\": \"1998\"\n },\n {\n \"authors\": \"Kolby Nottingham; Prithviraj Ammanabrolu; Alane Suhr; Yejin Choi; Hannaneh Hajishirzi; Sameer Singh; Roy Fox\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Do embodied agents dream of pixelated sheep?: Embodied decision making using language guided world modelling\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Alberto Olmo; Sarath Sreedharan; Subbarao Kambhampati\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Gpt3-to-plan: Extracting plans from text using gpt-3\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \" Openai\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Introducing chatgpt by openai\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \" Openai\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Vishal Pallagani; Bharath Muppasani; Keerthiram Murugesan; Francesca Rossi; Lior Horesh; Biplav Srivastava; Francesco Fabiano; Andrea Loreggia\",\n \"journal\": \"\",\n \"ref_id\": \"b37\",\n \"title\": \"Plansformer: Generating symbolic plans using transformers\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Liangming Pan; Alon Albalak; Xinyi Wang; William Yang; Wang \",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"Logic-lm: Empowering large language models with symbolic solvers for faithful logical reasoning\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Xavier Puig; Kevin Ra; Marko Boben; Jiaman Li; Tingwu Wang; Sanja Fidler; Antonio Torralba\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Virtualhome: Simulating household activities via programs\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Alec Radford; Jong Wook Kim; Chris Hallacy; Aditya Ramesh; Gabriel Goh; Sandhini Agarwal; Girish Sastry; Amanda Askell; Pamela Mishkin; Jack Clark\",\n \"journal\": \"PMLR\",\n \"ref_id\": \"b40\",\n \"title\": \"Learning transferable visual models from natural language supervision\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Vanya Shreyas Sundara Raman; Eric Cohen; Ifrah Rosen; David Idrees; Stefanie Paulius; Tellex\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Planning with large language models via corrective re-prompting\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Scott Reed; Konrad Zolna; Emilio Parisotto; Sergio Gomez Colmenarejo; Alexander Novikov; Gabriel Barth-Maron; Mai Gimenez; Yury Sulsky; Jackie Kay; Jost Tobias Springenberg\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"A generalist agent\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Jonathan Russell Stuart; Norvig\",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"Artificial intelligence: a modern approach\",\n \"year\": \"2003\"\n },\n {\n \"authors\": \"Timo Schick; Jane Dwivedi-Yu; Roberto Dessì; Roberta Raileanu; Maria Lomeli; Luke Zettlemoyer; Nicola Cancedda; Thomas Scialom\",\n \"journal\": \"\",\n \"ref_id\": \"b44\",\n \"title\": \"Toolformer: Language models can teach themselves to use tools\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Noah Shinn; Beck Labash; Ashwin Gopinath\",\n \"journal\": \"\",\n \"ref_id\": \"b45\",\n \"title\": \"Reflexion: an autonomous agent with dynamic memory and self-reflection\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Mohit Shridhar; Xingdi Yuan; Marc-Alexandre Côté; Yonatan Bisk; Adam Trischler; Matthew Hausknecht\",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"ALFWorld: Aligning Text and Embodied Environments for Interactive Learning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Tom Silver; Varun Hariprasad; Reece S Shuttleworth; Nishanth Kumar; Tomás Lozano-Pérez; Leslie Pack; Kaelbling \",\n \"journal\": \"\",\n \"ref_id\": \"b47\",\n \"title\": \"Pddl planning with pretrained large language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ron M Simpson; Diane E Kitchin; Thomas Leo Mccluskey\",\n \"journal\": \"The Knowledge Engineering Review\",\n \"ref_id\": \"b48\",\n \"title\": \"Planning domain definition using gipo\",\n \"year\": \"2007\"\n },\n {\n \"authors\": \"Ishika Singh; Valts Blukis; Arsalan Mousavian; Ankit Goyal; Danfei Xu; Jonathan Tremblay; Dieter Fox; Jesse Thomason; Animesh Garg\",\n \"journal\": \"\",\n \"ref_id\": \"b49\",\n \"title\": \"Progprompt: Generating situated robot task plans using large language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Hee Chan; Jiaman Song; Clayton Wu; Brian M Washington; Wei-Lun Sadler; Yu Chao; Su\",\n \"journal\": \"\",\n \"ref_id\": \"b50\",\n \"title\": \"Llm-planner: Few-shot grounded planning for embodied agents with large language models\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Sarath Sreedharan; Tathagata Chakraborti; Christian Muise; Yasaman Khazaeni; Subbarao Kambhampati\",\n \"journal\": \"\",\n \"ref_id\": \"b51\",\n \"title\": \"-d3wa+-a case study of xaip in a model acquisition task for dialogue planning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Kaya Stechly; Matthew Marquez; Subbarao Kambhampati\",\n \"journal\": \"\",\n \"ref_id\": \"b52\",\n \"title\": \"Gpt-4 doesn't know it's wrong: An analysis of iterative prompting for reasoning problems\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Karthik Valmeekam; Matthew Marquez; Subbarao Kambhampati\",\n \"journal\": \"\",\n \"ref_id\": \"b53\",\n \"title\": \"Can large language models really improve by self-critiquing their own plans?\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Karthik Valmeekam; Matthew Marquez; Sarath Sreedharan; Subbarao Kambhampati\",\n \"journal\": \"\",\n \"ref_id\": \"b54\",\n \"title\": \"On the planning abilities of large language models-a critical investigation\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Zihao Wang; Shaofei Cai; Anji Liu; Xiaojian Ma; Yitao Liang\",\n \"journal\": \"\",\n \"ref_id\": \"b55\",\n \"title\": \"Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Lionel Wong; Gabriel Grand; Alexander K Lew; Noah D Goodman; K Vikash; Jacob Mansinghka; Joshua B Andreas; Tenenbaum\",\n \"journal\": \"\",\n \"ref_id\": \"b56\",\n \"title\": \"From word models to world models: Translating from natural language to the probabilistic language of thought\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Yaqi Xie; Chen Yu; Tongyao Zhu; Jinbin Bai; Ze Gong; Harold Soh\",\n \"journal\": \"\",\n \"ref_id\": \"b57\",\n \"title\": \"Translating natural language to planning goals with large-language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Qiang Yang; Kangheng Wu; Yunfei Jiang\",\n \"journal\": \"Artificial Intelligence\",\n \"ref_id\": \"b58\",\n \"title\": \"Learning action models from plan examples using weighted max-sat\",\n \"year\": \"2007\"\n },\n {\n \"authors\": \"Shunyu Yao; Jeffrey Zhao; Dian Yu; Nan Du; Izhak Shafran; Yuan Karthik R Narasimhan; Cao\",\n \"journal\": \"\",\n \"ref_id\": \"b59\",\n \"title\": \"React: Synergizing reasoning and acting in language models\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Hankz Hankui Zhuo; Qiang Yang; Derek Hao Hu; Lei Li\",\n \"journal\": \"Artificial Intelligence\",\n \"ref_id\": \"b60\",\n \"title\": \"Learning complex action models with quantifiers and logical implications\",\n \"year\": \"2010\"\n },\n {\n \"authors\": \"Hankz Hankui Zhuo; Yantian Zha; Subbarao Kambhampati; Xin Tian\",\n \"journal\": \"ACM Transactions on Intelligent Systems and Technology (TIST)\",\n \"ref_id\": \"b61\",\n \"title\": \"Discovering underlying plans based on shallow models\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b62\",\n \"title\": \"robot-at ?x -furnitureAppliance): true if the robot is at the furniture or appliance ?x 2\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b63\",\n \"title\": \"openable ?y -furnitureAppliance): true if the furniture or appliance ?y can be opened\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b64\",\n \"title\": \"true if the furniture or appliance ?y is opened 7. (\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b65\",\n \"title\": \"true if the food ?x has been heated in the microwave 18\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b66\",\n \"title\": \"object-washed ?x -householdObject): true if the object ?x has been washed 26\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b67\",\n \"title\": \"robot-holding ?x -small_object): true if the robot is holding the small_object ?x 5\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b68\",\n \"title\": \"nut-in-hub ?n -nut ?h -hub): true if the nut ?n is in the hub ?h\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b69\",\n \"title\": \"is-jack ?x -small_object): true if the small_object ?x is a\",\n \"year\": \"\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b70\",\n \"title\": \"is-wheel ?x -small_object): true if the small_object ?x is a wheel 15\",\n \"year\": \"\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[20,123.27,219.32,7.31,100.88],"formula_id":"formula_0","formula_text":"→ → → → → → → → → → → → →"},{"formula_coordinates":[36,123.17,386.54,148.16,22.44],"formula_id":"formula_1","formula_text":"→ ---------------------------------------- Automated corrective feedback:"},{"formula_coordinates":[36,123.27,427.06,7.31,13.21],"formula_id":"formula_2","formula_text":"→ →"},{"formula_coordinates":[54,123.27,353.34,7.31,61.03],"formula_id":"formula_3","formula_text":"→ → → → → → → →"}],"string":"[\n {\n \"formula_coordinates\": [\n 20,\n 123.27,\n 219.32,\n 7.31,\n 100.88\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"→ → → → → → → → → → → → →\"\n },\n {\n \"formula_coordinates\": [\n 36,\n 123.17,\n 386.54,\n 148.16,\n 22.44\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"→ ---------------------------------------- Automated corrective feedback:\"\n },\n {\n \"formula_coordinates\": [\n 36,\n 123.27,\n 427.06,\n 7.31,\n 13.21\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"→ →\"\n },\n {\n \"formula_coordinates\": [\n 54,\n 123.27,\n 353.34,\n 7.31,\n 61.03\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"→ → → → → → → →\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"Leveraging Pre-trained Large Language Models to Construct and Utilize World Models for Model-based Task Planning"},"abstract":{"kind":"string","value":"There is a growing interest in applying pre-trained large language models (LLMs) to planning problems. However, methods that use LLMs directly as planners are currently impractical due to several factors, including limited correctness of plans, strong reliance on feedback from interactions with simulators or even the actual environment, and the inefficiency in utilizing human feedback. In this work, we introduce a novel alternative paradigm that constructs an explicit world (domain) model in planning domain definition language (PDDL) and then uses it to plan with sound domain-independent planners. To address the fact that LLMs may not generate a fully functional PDDL model initially, we employ LLMs as an interface between PDDL and sources of corrective feedback, such as PDDL validators and humans. For users who lack a background in PDDL, we show that LLMs can translate PDDL into natural language and effectively encode corrective feedback back to the underlying domain model. Our framework not only enjoys the correctness guarantee offered by the external planners but also reduces human involvement by allowing users to correct domain models at the beginning, rather than inspecting and correcting (through interactive prompting) every generated plan as in previous work. On two IPC domains and a Household domain that is more complicated than commonly used benchmarks such as ALFWorld, we demonstrate that GPT-4 can be leveraged to produce high-quality PDDL models for over 40 actions, and the corrected PDDL models are then used to successfully solve 48 challenging planning tasks. Resources, including the source code, are released at: https://guansuns.github.io/pages/llm-dm."},"authors":{"kind":"string","value":"Lin Guan; Karthik Valmeekam; Sarath Sreedharan; Subbarao Kambhampati"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: An overview of our framework and existing methods that use LLMs directly as planners.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: The prompt template for PDDL construction and an example of the LLM output for the household domain.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"Domain# of actions # of params and literals # of GPT-4 errors # of GPT-3.5","figure_data":"","figure_id":"fig_2","figure_label":"","figure_type":"figure"},{"figure_caption":"( a )aThis action is only applicable for objects that are pickupable by the robot. (b) The furniture piece or appliance should be opened if it is openable. (c) The object to pick up should not be stacked on top of other household items.","figure_data":"","figure_id":"fig_3","figure_label":"a","figure_type":"figure"},{"figure_caption":"( a )aThe furniture piece has to have an open and flat surface. (b) Note that both object_1 and object_2 must be stackable objects. (c) You may assume the robot is holding object_1 at the beginning. (d) Also, there should be no other object on top of object_2.","figure_data":"","figure_id":"fig_4","figure_label":"a","figure_type":"figure"},{"figure_caption":"( a )aThe furniture piece needs to have an open and flat surface to allow enough space for manipulation (e.g., countertops or tables). → (b) A receptacle should be opened if it is openable. (c) Both receptacles should not be stacked on top of other objects if they are stackable.","figure_data":"","figure_id":"fig_5","figure_label":"a","figure_type":"figure"},{"figure_caption":"(a) The furniture piece needs to have an open and flat surface to allow enough space for manipulation (e.g., countertops or tables). → (b) The receptacle should be opened if it is openable. (c) The receptacle should not be stacked on top of other objects if it is stackable.","figure_data":"","figure_id":"fig_6","figure_label":"","figure_type":"figure"},{"figure_caption":"(a) You may assume the robot is holding the cloth before executing this action. (b) The cloth will be dirty after executing the action. (c) The robot should also use a clean cloth.","figure_data":"","figure_id":"fig_7","figure_label":"","figure_type":"figure"},{"figure_caption":"-Corrective feedback: There is an error in the PDDL model: 1. The current action model allows the plane to fly from any location to another location in a different city, but a plane can only land at airports → Please revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors). → Parameters:","figure_data":"","figure_id":"fig_8","figure_label":"","figure_type":"figure"},{"figure_caption":"Parameters: 1 .1?c -container: the container to open ) Effects: (and (not (nut-loose ?n)) ) A.10.7 Jack up a hub","figure_data":"","figure_id":"fig_9","figure_label":"1","figure_type":"figure"},{"figure_caption":"→ 1 .1load a package ?p into a truck ?t at location ?l (example output: load package_1 into a truck truck_1 at location_0). → Parameters: (a) ?p: the package to load; (b) ?t: the truck to load the package into; (c) ?l: the location where the package and truck are.","figure_data":"","figure_id":"fig_10","figure_label":"1","figure_type":"figure"},{"figure_caption":"Instructions for the PDDL generation task You are defining the preconditions and effects (represented in PDDL format) of an AI agent's actions. Information about the AI agent will be provided in the domain description ...","figure_data":"→One or two examples from other domains for illustrating the input and output formatsHere are two examples from the classical BlocksWorld domain for demonstrating the output format....Here is the task.A natural language description of the domainDomain information: The AI agent here is a household robot that can navigate to various large and→normally immovable furniture pieces or appliances in the house to carry out household tasks→...A natural language description of the actionAction: This action enables the robot to toggle small appliances (like humidifiers and light→bulbs) which are toggleable to switch them on ...The dynamically updated list of predicatesYou can create and define new predicates, but you may also reuse the following predicates:1. (robot-at ?r -robot ?f -furnitureAppliance): true if the robot ?r is at the furniture or→appliance ?f2. (object-in-on ?o -householdObject ?f -furnitureAppliance): true if the object ?o is in or on→the furniture or appliance ?f...Parameters:","figure_id":"tab_0","figure_label":"","figure_type":"table"},{"figure_caption":"Success rates of different planning approaches in the Household domain and the Logistics domain.","figure_data":"","figure_id":"tab_3","figure_label":"2","figure_type":"table"},{"figure_caption":"One or two examples from other domains for illustrating the input and output formats Here are two examples from the classical BlocksWorld domain for demonstrating the output format. Domain information: BlocksWorld is a planning domain in artificial intelligence. The AI agent here is a mechanical robot arm that can pick and place the blocks. Only one block may be moved at a time: it may either be placed on the table or placed atop another block. Because of this, any blocks that are, at a given time, under another block cannot be moved. There is only one type of object in this domain, and that is the block. This action enables the robot to put a block onto the table. For example, the robot puts block_1 onto the table. The AI agent here is a household robot that can navigate to various large and normally immovable furniture pieces or appliances in the house to carry out household tasks. Note that the robot has only one gripper, so (a) it can only hold one object; (b) it shouldn't hold any other irrelevant objects in its gripper while performing some manipulation tasks (e.g., opening a drawer or closing a window); (c) operations on small household items should be carried out on furniture with a flat surface to get enough space for manipulation. There are three types of objects in this domain: robot, furnitureAppliance, and householdObject. The object type furnitureAppliance covers large and normally immovable furniture pieces or appliances, such as stove burners, side tables, dining tables, drawer, cabinets, or microwaves. The object type householdObject covers all other small household items, such as handheld vacuum cleaners, cloth, apples, bananas, and small receptacles like bowls and lunch boxes. In this domain, the locations of the robot and small household items (e.g., apples, oranges, bowls, lunch boxes or lamps) are determined by large and normally immovable furniture pieces or appliances.","figure_data":"Effects:``(and(not (block-on-table ?x))(not (block-clear ?x))(not (robot-hand-empty))(robot-holding ?x))→ ``Ǹ→→ ew Predicates:→ No newly defined predicate→→→ Here is the task.A natural language description of the domainDomain information:→→→→Example 1Action: Parameters:1. ?x -block: the block to put downPreconditions:``(and(robot-holding ?x))``Èffects:``(and(not (robot-holding ?x))(block-clear ?x)(robot-hand-empty)(block-on-table ?x))``Ǹew Predicates:1. (robot-holding ?x -block): true if the robot arm is holding the block ?x2. (block-clear ?x -block): true if the block ?x is not under any another block3. (robot-hand-empty): true if the robot arm is not holding any block4. (block-on-table ?x -block): true if the block ?x is placed on the tableExample 2Action: This action enables the robot to pick up a block on the table.You can create and define new predicates, but you may also reuse the following predicates:1. (robot-holding ?x -block): true if the robot arm is holding the block ?x2. (block-clear ?x -block): true if the block ?x is not under any another block3. (robot-hand-empty): true if the robot arm is not holding any block4. (block-on-table ?x -block): true if the block ?x is placed on the tableParameters:1. ?x -block: the block to pick upPreconditions:``(and(block-clear ?x)(block-on-table ?x)(robot-hand-empty))```1","figure_id":"tab_4","figure_label":"","figure_type":"table"},{"figure_caption":".1 Examples of prompts for LLM plannersAn example prompt for the Household domainYou are in a household to complete a certain task for the owners. You are a household robot that can navigate to various large and normally immovable furniture pieces or appliances in the house to carry out household tasks. Note that you have only one gripper, so (a) it can only hold one object, (b) it shouldn't hold any other irrelevant objects in its gripper while performing some manipulation tasks (e.g., opening a drawer or closing a window), (c) operations on small household items should be carried out on furniture with a flat surface to get enough space for manipulation.The following actions are available to you, and you should strictly follow the output format (demonstrated in the \"example output\") of each action:","figure_data":"> action: open furniture/appliance drawer_1Step 5:> action: pick up orange_1 in/on drawer_1Step 6:> action: go from drawer_1 to fridge_1Step 7:> action: put orange_1 in/on fridge_1Step 8:> action: close furniture/appliance fridge_1Example 2...→→→→→→→→→→Here are two examples.Example 1The house contains the following large furniture pieces and appliances: drawer_1, fridge_1,→dining_table_1, countertop_1, stove_burner_1, cabinet_1, microwave_1, sink_basin_1,→dish_washer_1, light_switch_1, garbage_can_1.The current states of the household objects are as follows:1. drawer_1 is closed2. fridge_1 is closed...17. robot is not holding anythingYour task is to: put an orange in a fridge.Your plan is:Step 1:> think: First, I need to open the fridge.> action: go from dining_table_1 to fridge_1Step 2:> action: open furniture/appliance fridge_1Step 3:fridge_1 to drawer_1Step 4:","figure_id":"tab_8","figure_label":"","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: An overview of our framework and existing methods that use LLMs directly as planners.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: The prompt template for PDDL construction and an example of the LLM output for the household domain.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Domain# of actions # of params and literals # of GPT-4 errors # of GPT-3.5\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"( a )aThis action is only applicable for objects that are pickupable by the robot. (b) The furniture piece or appliance should be opened if it is openable. (c) The object to pick up should not be stacked on top of other household items.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"a\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"( a )aThe furniture piece has to have an open and flat surface. (b) Note that both object_1 and object_2 must be stackable objects. (c) You may assume the robot is holding object_1 at the beginning. (d) Also, there should be no other object on top of object_2.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"a\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"( a )aThe furniture piece needs to have an open and flat surface to allow enough space for manipulation (e.g., countertops or tables). → (b) A receptacle should be opened if it is openable. (c) Both receptacles should not be stacked on top of other objects if they are stackable.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"a\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"(a) The furniture piece needs to have an open and flat surface to allow enough space for manipulation (e.g., countertops or tables). → (b) The receptacle should be opened if it is openable. (c) The receptacle should not be stacked on top of other objects if it is stackable.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"(a) You may assume the robot is holding the cloth before executing this action. (b) The cloth will be dirty after executing the action. (c) The robot should also use a clean cloth.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"-Corrective feedback: There is an error in the PDDL model: 1. The current action model allows the plane to fly from any location to another location in a different city, but a plane can only land at airports → Please revise the PDDL model (and the list of predicates if needed) to fix the above errors (and other potentially similar errors). → Parameters:\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_8\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Parameters: 1 .1?c -container: the container to open ) Effects: (and (not (nut-loose ?n)) ) A.10.7 Jack up a hub\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_9\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"→ 1 .1load a package ?p into a truck ?t at location ?l (example output: load package_1 into a truck truck_1 at location_0). → Parameters: (a) ?p: the package to load; (b) ?t: the truck to load the package into; (c) ?l: the location where the package and truck are.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_10\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Instructions for the PDDL generation task You are defining the preconditions and effects (represented in PDDL format) of an AI agent's actions. Information about the AI agent will be provided in the domain description ...\",\n \"figure_data\": \"→One or two examples from other domains for illustrating the input and output formatsHere are two examples from the classical BlocksWorld domain for demonstrating the output format....Here is the task.A natural language description of the domainDomain information: The AI agent here is a household robot that can navigate to various large and→normally immovable furniture pieces or appliances in the house to carry out household tasks→...A natural language description of the actionAction: This action enables the robot to toggle small appliances (like humidifiers and light→bulbs) which are toggleable to switch them on ...The dynamically updated list of predicatesYou can create and define new predicates, but you may also reuse the following predicates:1. (robot-at ?r -robot ?f -furnitureAppliance): true if the robot ?r is at the furniture or→appliance ?f2. (object-in-on ?o -householdObject ?f -furnitureAppliance): true if the object ?o is in or on→the furniture or appliance ?f...Parameters:\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Success rates of different planning approaches in the Household domain and the Logistics domain.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"One or two examples from other domains for illustrating the input and output formats Here are two examples from the classical BlocksWorld domain for demonstrating the output format. Domain information: BlocksWorld is a planning domain in artificial intelligence. The AI agent here is a mechanical robot arm that can pick and place the blocks. Only one block may be moved at a time: it may either be placed on the table or placed atop another block. Because of this, any blocks that are, at a given time, under another block cannot be moved. There is only one type of object in this domain, and that is the block. This action enables the robot to put a block onto the table. For example, the robot puts block_1 onto the table. The AI agent here is a household robot that can navigate to various large and normally immovable furniture pieces or appliances in the house to carry out household tasks. Note that the robot has only one gripper, so (a) it can only hold one object; (b) it shouldn't hold any other irrelevant objects in its gripper while performing some manipulation tasks (e.g., opening a drawer or closing a window); (c) operations on small household items should be carried out on furniture with a flat surface to get enough space for manipulation. There are three types of objects in this domain: robot, furnitureAppliance, and householdObject. The object type furnitureAppliance covers large and normally immovable furniture pieces or appliances, such as stove burners, side tables, dining tables, drawer, cabinets, or microwaves. The object type householdObject covers all other small household items, such as handheld vacuum cleaners, cloth, apples, bananas, and small receptacles like bowls and lunch boxes. In this domain, the locations of the robot and small household items (e.g., apples, oranges, bowls, lunch boxes or lamps) are determined by large and normally immovable furniture pieces or appliances.\",\n \"figure_data\": \"Effects:``(and(not (block-on-table ?x))(not (block-clear ?x))(not (robot-hand-empty))(robot-holding ?x))→ ``Ǹ→→ ew Predicates:→ No newly defined predicate→→→ Here is the task.A natural language description of the domainDomain information:→→→→Example 1Action: Parameters:1. ?x -block: the block to put downPreconditions:``(and(robot-holding ?x))``Èffects:``(and(not (robot-holding ?x))(block-clear ?x)(robot-hand-empty)(block-on-table ?x))``Ǹew Predicates:1. (robot-holding ?x -block): true if the robot arm is holding the block ?x2. (block-clear ?x -block): true if the block ?x is not under any another block3. (robot-hand-empty): true if the robot arm is not holding any block4. (block-on-table ?x -block): true if the block ?x is placed on the tableExample 2Action: This action enables the robot to pick up a block on the table.You can create and define new predicates, but you may also reuse the following predicates:1. (robot-holding ?x -block): true if the robot arm is holding the block ?x2. (block-clear ?x -block): true if the block ?x is not under any another block3. (robot-hand-empty): true if the robot arm is not holding any block4. (block-on-table ?x -block): true if the block ?x is placed on the tableParameters:1. ?x -block: the block to pick upPreconditions:``(and(block-clear ?x)(block-on-table ?x)(robot-hand-empty))```1\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \".1 Examples of prompts for LLM plannersAn example prompt for the Household domainYou are in a household to complete a certain task for the owners. You are a household robot that can navigate to various large and normally immovable furniture pieces or appliances in the house to carry out household tasks. Note that you have only one gripper, so (a) it can only hold one object, (b) it shouldn't hold any other irrelevant objects in its gripper while performing some manipulation tasks (e.g., opening a drawer or closing a window), (c) operations on small household items should be carried out on furniture with a flat surface to get enough space for manipulation.The following actions are available to you, and you should strictly follow the output format (demonstrated in the \\\"example output\\\") of each action:\",\n \"figure_data\": \"> action: open furniture/appliance drawer_1Step 5:> action: pick up orange_1 in/on drawer_1Step 6:> action: go from drawer_1 to fridge_1Step 7:> action: put orange_1 in/on fridge_1Step 8:> action: close furniture/appliance fridge_1Example 2...→→→→→→→→→→Here are two examples.Example 1The house contains the following large furniture pieces and appliances: drawer_1, fridge_1,→dining_table_1, countertop_1, stove_burner_1, cabinet_1, microwave_1, sink_basin_1,→dish_washer_1, light_switch_1, garbage_can_1.The current states of the household objects are as follows:1. drawer_1 is closed2. fridge_1 is closed...17. robot is not holding anythingYour task is to: put an orange in a fridge.Your plan is:Step 1:> think: First, I need to open the fridge.> action: go from dining_table_1 to fridge_1Step 2:> action: open furniture/appliance fridge_1Step 3:fridge_1 to drawer_1Step 4:\",\n \"figure_id\": \"tab_8\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"[19]\", \"Explanation\": \"The cited work by [19] has showcased remarkable performance in natural language processing tasks, which the citing paper leverages to improve the planning capabilities of LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[1]\", \"Explanation\": \"The cited work by [1] has been tested to perform sequential decision-making or planning in everyday domains, which the citing paper builds upon to improve the planning capabilities of LLMs.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[20]\", \"Explanation\": \"The cited work by [20] has proposed collecting feedback from the environment during plan execution to improve the planning capabilities of LLMs, which the citing paper extends by incorporating various forms of feedback in the planning process.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[60]\", \"Explanation\": \"The cited work by [60] has proposed collecting human corrections as feedback to improve the planning capabilities of LLMs, which the citing paper builds upon to enhance the plan refinement process.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[42]\", \"Explanation\": \"The cited work by [42] has proposed collecting information of unmet preconditions as feedback to improve the planning capabilities of LLMs, which the citing paper extends by incorporating this form of feedback in the plan refinement process.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[56]\", \"Explanation\": \"The cited work by [56] has proposed collecting information of unmet preconditions as feedback to improve the planning capabilities of LLMs, which the citing paper builds upon to enhance the plan refinement process.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[24,55,53,54,31]\", \"Explanation\": \"The cited works provide evidence that LLMs have not yet demonstrated sufficient capabilities in reasoning and planning, highlighting the need for further research in this area.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[48,55]\", \"Explanation\": \"The cited works show that LLMs struggle to produce correct and executable plans even when provided with detailed descriptions of actions, emphasizing the need for improved planning capabilities.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[11,13]\", \"Explanation\": \"The cited works focus on the use of provably sound planning in classical planning literature, which the citing paper extends to the context of LLMs-planning paradigms.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[33]\", \"Explanation\": \"The cited work provides a PDDL domain model that the citing paper utilizes in its research on LLMs planning performance.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[18]\", \"Explanation\": \"The cited work, VAL, is the PDDL validator that the citing paper uses to provide corrective feedback in natural language to the PDDL models generated by the LLM middle layer in the cited work.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[58,30]\", \"Explanation\": \"The cited works provide a method for translating user instructions into goal specifications in PDDL, which the citing paper adopts to utilize standard domain-independent planners for downstream planning tasks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[30]\", \"Explanation\": \"The cited work extends the use of PDDL models in planning tasks by introducing a method for translating user instructions into goal specifications, which the citing paper builds upon to explore the potential of LLM planners in incorporating both explicit and implicit user constraints in common-sense domains.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[24,55]\", \"Explanation\": \"The cited works provide preliminary evidence that off-the-shelf LLMs are currently not capable of producing accurate plans, which serves as a basis for the citing paper to explore the limitations of LLMs in sequential decision-making tasks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[1,29]\", \"Explanation\": \"The cited works employ LLMs as a heuristic in planning tasks, which the citing paper extends by further exploring the use of LLMs in planning and grounding actions in the physical world.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[28,50]\", \"Explanation\": \"The cited works use LLMs to generate plans represented in Python-style code, which the citing paper extends by exploring the use of LLMs in generating plans in a more specific format.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[60]\", \"Explanation\": \"The cited work on prompt engineering provides evidence that improving the planning performance of LLMs can be achieved through prompt engineering, which the citing paper builds upon in its research on improving planning performance.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[51,20,34]\", \"Explanation\": \"The cited works collect sensory information, human corrections, and self-corrections as forms of feedback to improve the planning performance of LLMs, which the citing paper acknowledges as external data sources in its research on feedback collection for planning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[42,56]\", \"Explanation\": \"The cited works collect information of unmet preconditions to improve the planning performance of LLMs, which the citing paper extends by exploring the use of this information in planning tasks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[26]\", \"Explanation\": \"The cited work has shown that training sequence models on a specific task can lead to the development of internal world representations within the model, which the citing paper uses as a methodological basis for constructing symbolic world models without extra training.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[2]\", \"Explanation\": \"The cited work provides a comprehensive survey on learning-based methods for acquiring symbolic domain models, which the citing paper uses as a data source to build upon in the context of leveraging common-world knowledge in LLMs for constructing domain models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[3,31]\", \"Explanation\": \"The cited works provide evidence that autoregressive models lack reliable capacity for reasoning about action effects, which the citing paper uses to highlight the limitations of using LLMs as world models.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[53,54]\", \"Explanation\": \"The cited works show that language models have issues with capturing errors in candidate plans, which the citing paper extends by exploring the limitations of using LLMs in plan generation.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[45]\", \"Explanation\": \"The cited work on using external tools for arithmetic in LLMs is mentioned as a way to address the limitations of LLMs in specific downstream tasks, which the citing paper uses as a context for their work on plan generation with sound planners.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"[39,57]\", \"Explanation\": \"The cited works on using external tools for logical reasoning in LLMs are mentioned as a way to augment the capacity of LLMs for more reliable plan generation, which the citing paper extends by using sound planners to improve plan generation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work provides a method for planning problems involving high-level skills and low-level control policies, which the citing paper adopts in its research on intelligent agent planning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[12]\", \"Explanation\": \"The cited work is referenced for its external planner approach to planning problems involving high-level skills and low-level control policies, which the citing paper may have adopted in its research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[17]\", \"Explanation\": \"The cited work is mentioned for its role in planning problems involving high-level skills and low-level control policies, which the citing paper may have considered in its research on intelligent agent planning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[44]\", \"Explanation\": \"The cited work provides the formal definition of a classical planning problem, which serves as the basis for the planning formalism used in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[33]\", \"Explanation\": \"The cited work provides the standard encoding language for classical planning problems, which the citing paper adopts to encode the planning problem in the Blocksworld domain.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[30]\", \"Explanation\": \"The cited work provides a method for translating natural language descriptions of environment into PDDL, which the citing paper leverages to ground the predicate values of the initial state in the PDDL models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[18]\", \"Explanation\": \"The cited work provides a PDDL model validation tool that the citing paper uses to detect basic syntax errors in the generated PDDL models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[60]\", \"Explanation\": \"The cited work provides a method for correcting action models in PDDL models, which the citing paper adopts to address the issue of user feedback in the construction of PDDL models.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[12]\", \"Explanation\": \"The cited work provides a local-search planner called LPG that the citing paper uses to accelerate the plan searching process in a common-sense domain where LLMs may generate meaningful heuristics.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[42]\", \"Explanation\": \"The cited work provides a method for re-prompting in the context of LLM planning, which the citing paper uses in their research to refine plans and improve performance.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[36]\", \"Explanation\": \"The cited work, GPT-3.5-Turbo, is used as a data source for the experiments conducted in the citing paper to assess the quality of PDDL models generated by LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[37]\", \"Explanation\": \"The cited work, GPT-4, is used as a methodological basis for the PDDL construction process in the citing paper, as it is the LLM that the experiments are conducted with.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work, Fast Downward, is a classical planner that the citing paper adopts to effectively find valid plans in the Household and Logistics domains. The cited work provides the method and technique for plan generation in the context of the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[60]\", \"Explanation\": \"The cited work, ReAct, is used as the underlying LLM planner in the approach of LLM modulo planner back-prompted by VAL using LLMacquired domain model. The prompt design is modified to provide detailed action descriptions and use only two fixed examples to address the limitation of end users not always being able to provide a large pool of examples.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[36]\", \"Explanation\": \"The cited work, GPT-3.5-Turbo, is used as a language model in the citing paper to support the integration of corrective feedback in the construction dialogue process.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[49]\", \"Explanation\": \"The cited work, GIPO, is a graphical tool that can effectively visualize the causal dependencies of actions, providing a useful technique for locating errors in PDDL models that the citing paper builds upon.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[23]\", \"Explanation\": \"The cited work on explainable AI provides a basis for improving the explainability of the plans generated by the domain models extracted from LLMs.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[18]\", \"Explanation\": \"The cited work, VAL, provides a method for obtaining validation feedback in the form of unsatisfied preconditions, goal conditions, and parameter usage in PDDL, which the citing paper uses to structure the feedback messages in the translation process of PDDL to natural language.\"}]"}}},{"rowIdx":97,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_0","fig_1"],"heading":"Introduction","publication_ref":["b0","b22","b33","b3","b8","b18","b26","b31","b32","b11","b13","b10","b12","b13","b10","b12","b13","b12","b13","b34","b36","b10","b12","b14","b13"],"table_ref":["tab_4"],"text":"Tensor network (TN) representation, which aims to express higher-order data with low-dimensional tensors (called TN cores) by a specific operation among them, has gained significant attention in various areas of data analysis [1,23,34], machine learning [4,9], computer vision [19,27,32,33], etc. By regarding TN cores as nodes and operations as edges, a TN corresponds to a graph (called TN topology). The vector composed of the weights of all edges in the topology is defined as the TN rank. TN structure (including topology and rank) search (TN-SS) aims to search for a suitable TN structure to achieve a compact representation for a given tensor, which is known as a challenging and theoretically NP-hard problem [12,14]. The selection of TN structure dramatically impacts the performance of TN representation in practical applications [11,13,14].\nRecently, there have been several notable efforts to tackle the TN-SS problem [11,13,14]. But most of them adopt the \"sampling-evaluation\" framework, which requires sampling a large number of structures as candidates and conducting numerous repeated structure evaluations. For instance, for a tensor of size 40 × 60 × 3 × 9 × 9 (used in Section 4.2), TNGA in [13] requires thousands of evaluations and TNALE in [14] requires hundreds of evaluations, where each evaluation entails solving an optimization problem to compute TN cores iteratively. Consequently, the computational cost becomes exceedingly high. A meaningful question is whether we can optimize the TN structure simultaneously during the computation Preprint. [35] and FCTNALS [37] are methods with pre-defined topologies, and TNGreedy [11], TNGA [13], TNLS [15], and TNALE [14] are TN-SS methods (please see more results in Table 3).\nof TN cores, thus escaping the \"sampling-evaluation\" framework and fundamentally addressing the computationally consuming issue.\nIn this paper, we introduce for the first time a regularized modeling perspective on solving the TN-SS problem. This perspective enables us to optimize the TN structure simultaneously during the computation of TN cores, effectively eliminating the need for repetitive structure evaluations. To be specific, we propose a novel TN paradigm, termed as SVD-inspired TN decomposition (SVDinsTN), by inserting diagonal factors between any two TN cores in the \"fully-connected\" topology (see Figure 1(b)). The intuition behind SVDinsTN is to leverage the sparsity of the inserted diagonal factors to reveal a compact TN structure and utilize the TN cores (merged with the diagonal factors) to represent a given tensor. Based on SVDinsTN, we establish a regularized model, which updates the TN cores and diagonal factors iteratively and imposes a sparse operator to induce the sparsity of the diagonal factors. In theory, we prove a convergence guarantee for the proposed method and establish an upper bound for the TN rank. In particular, we design a novel initialization scheme for the proposed method based on the upper bound. This initialization scheme enables the proposed method to overcome the high computational cost in the first several iterations, which is caused by the utilization of a \"fully-connected\" topology as the starting point. As a result, SVDinsTN is capable of capturing a customized TN structure and providing a compact representation for a given tensor in an efficient manner. In summary, we make the following three contributions.\n• We propose SVDinsTN, a new TN paradigm, that enables us to optimize the TN structure during the computation of TN cores, greatly reducing the computational cost. • In theory, we prove a convergence guarantee for the proposed method and establish an upper bound for the TN rank involved in SVDinsTN. The upper bound serves as a guide for designing an efficient initialization scheme. • Experimental results verify numerically that the proposed method achieves 100 ∼ 1000 times acceleration compared to the state-of-the-art TN-SS methods with a comparable representation ability (see Figure 1(c))."},{"figure_ref":[],"heading":"Related Works","publication_ref":["b3","b4","b27","b21","b34","b36","b34","b28","b30","b36","b7","b10","b12","b13","b14","b16","b17","b19","b23","b10","b12","b13"],"table_ref":[],"text":"TN representation1 aims to find a set of small-sized TN cores to express a large-sized tensor under a TN structure (including topology and rank) [4,5,28]. In the past decades, many works focused on TN representation with a fixed TN topology, such as tensor train (TT) decomposition with a \"chain\" topology [22], tensor ring (TR) decomposition with a\"ring\" topology [35], fully-connected tensor network (FCTN) decomposition with a \"fully-connected\" topology [37], etc. In addition, these works also presented various methods to optimize the TN cores, such as alternating least square(ALS) [35], gradient descent (GD) [29,31], proximal alternating minimization (PAM) [37], etc. In contrast, SVDinsTN can reveal a compact TN structure for a given tensor, surpassing methods with pre-defined topologies in terms of representation ability.\nTN structure search (TN-SS) aims to search for a suitable or optimal TN structure, including both topology and rank, to achieve a compact representation for a given tensor [8,11,13,14,15,17,18,20,24]. However, most existing TN-SS methods adopt the \"sampling-evaluation\" framework, making them require the utilization of heuristic search algorithms such as the greedy algorithm [11], genetic algorithm [13], and alternating local enumeration algorithm [14] to sample candidate structures and then evaluate them one by one. Therefore, these methods inevitably suffer from prohibitively high computational costs due to the numerous repeated evaluations, with each evaluation involving solving an optimization problem concerning the iterative calculation of TN cores. In contrast, SVDinsTN solves the TN-SS problem from a regularized modeling perspective, which avoids the repeated structure evaluations and greatly decreases the computational cost."},{"figure_ref":[],"heading":"Notations and Preliminaries","publication_ref":["b36"],"table_ref":[],"text":"We first summarize notations and several operations that will be used in the rest of the paper. Then, we review definitions related to tensor networks.\nA tensor is a multi-dimensional array, and the number of dimensions (also called modes) of which is referred to as the tensor order. In the paper, first-order tensors (vectors), second-order tensors (matrices), and N th-order tensors are denoted by x ∈ R I1 , X ∈ R I1×I2 , and X ∈ R I1×I2ו••×I N , respectively. We use ∥X ∥ F and ∥X ∥ 1 to denote the Frobenius norm and ℓ 1 -norm of X , respectively. For brevity, we let K N denote the set {1, 2, • • • , N }, which is used for labeling individual indices. Similarly, we let TL N denote the set {(t, l)|1 ≤ t < l ≤ N ; t, l ∈ N}, which is used for labeling paired indices.\nWe next provide a review of several operations on tensors [37].\nThe generalized tensor unfolding is an operation that converts a tensor into a matrix by merging a group of tensor modes into the rows of the matrix and merging the remaining modes into the columns. For example, an I 1 × I 2 × I 3 × I 4 tensor can be unfolded into an I 1 I 3 × I 2 I 4 matrix, denoted by X [1,3;2,4] . We use X [1,3;2,4] = GenUnfold(X , [1, 3; 2, 4]) and X = GenFold(X [1,3;2,4] , [1, 3; 2, 4]) to simply denote the corresponding operation and inverse operation, respectively. We also use X (2) to simply represent X [2;1,3,4] ∈ R I2×I1I2I4 , which is also called mode-2 unfolding. Besides, we use X (2) = Unfold(X , 2) and X = Fold(X (2) , 2) to simply denote the corresponding operation and inverse operation, respectively.\nThe generalized tensor transposition is an operation that rearranges tensor modes in the order specified by a vector. For example, an I 1 × I 2 × I 3 × I 4 tensor can be transposed into an I 3 × I 2 × I 1 × I 4 tensor, denoted by ⃗ X n with n = (3, 2, 1, 4). We use ⃗ X n = permute(X , n) and X = ipermute( ⃗ X n , n) to simply denote the corresponding operation and inverse operation, respectively.\nThe tensor contraction is an operation that obtains a new tensor by pairing, multiplying, and summing indices of certain modes of two tensors. For example, if a fourth-order tensor X ∈ R I1×I2×I3×I4 and a third-order tensor Y ∈ R J1×J2×J3 satisfy I 2 = J 1 and I 4 = J 2 , then the tensor contraction between the 2nd and 4th modes of X and the 1st and 2nd modes of Y yields a tensor Z = X × 1,2 2,4 Y ∈ R I1×I3×J3 . The elements of Z are calculated as follows:\nZ(i 1 , i 3 , j 3 ) = I2 i2=1 I4 i4=1 X (i 1 , i 2 , i 3 , i 4 )Y(i 2 , i 4 , j 3 ).\nIn Table 1, we give the interpretations of the operations \"diag\", \"ones\", \"zeros\", and \"vec\"."},{"figure_ref":[],"heading":"Tensor Network","publication_ref":["b3","b36","b15"],"table_ref":[],"text":"In general, a tensor network (TN) is defined as a collection of low-dimensional tensors, known as TN cores, in which some or all modes are contracted according to specific operations [4]. The primary purpose of a TN is to represent higher-order data using these TN cores. By considering TN cores as nodes and operations between modes of cores as edges, we define the graph formed by these nodes and edges as the TN topology. Additionally, we assign a non-negative integer weight to each edge to indicate the size of the corresponding TN core, and call the vector composed of these edge weights the TN rank. Consequently, a TN structure refers to a weighted graph comprising nodes, edges, and weights, encompassing both the TN topology and TN rank. This paper focuses on only a class of TNs that employs tensor contraction as the operation among TN cores and adopts a simple graph as the TN topology. More particularly, when representing an N th-order tensor X , this class of TNs comprises precisely N TN cores, each corresponding to one mode of X . A notable method is the FCTN decomposition, which represents an N th-order tensor [37]. In this decomposition, any two cores G l and G t for (t, l) ∈ TL N share an equal-sized mode R t,l used for tensor contraction. We denote the FCTN decomposition by\nX ∈ R I1×I2ו••×I N by N small-sized N th-order cores denoted by G k ∈ R R 1,k ×R 2,k ו••×R k-1,k ×I k ×R k,k+1 ו••×R k,N for k ∈ K N\nX = FCTN(G 1 , G 2 , • • • , G N ) and define the FCTN rank as the vector (R 1,2 , R 1,3 , • • • , R 1,N , R 2,3 , • • • , R 2,N , • • • , R N -1,N ) ∈ R N (N -1)/2 .\nAccording to the concept of tensor contraction, removing rank-one edges in the TN topology does not change the expression of the TN. This means that if any element in the FCTN rank is equal to one, the corresponding edge can be harmlessly eliminated from the \"fully-connected\" topology. For instance, a \"fully-connected\" topology with the rank\n(R 1,2 , 1, • • • , 1, R 2,3 , 1, • • • , 1, R N -2,N -1 , R N -1,N ) can be converted into a \"chain\" topology with rank (R 1,2 , R 2,3 , • • • , R N -1,N\n) in this manner. This fact can be formally stated as follows.\nProperty 1 [16] There exists a one-to-one correspondence between the TN structure and FCTN rank.\nThis work is inspired by Property 1, searching for a more compact TN structure by optimizing the FCTN rank."},{"figure_ref":[],"heading":"An Efficient Method for TN-SS","publication_ref":[],"table_ref":[],"text":"We propose an efficient method to solve the TN-SS problem from a regularized modeling perspective. Unlike the existing \"sampling-evaluation\" framework, the main idea of the proposed method is to optimize the TN structure simultaneously during the computation of TN cores, thereby eliminating the need for repetitive structure evaluations and greatly decreasing the computational cost."},{"figure_ref":["fig_0"],"heading":"SVDinsTN","publication_ref":["b36"],"table_ref":[],"text":"We start with the definition of the following SVDinsTN.\nDefinition 1 (SVDinsTN) Let X ∈ R I1×I2ו••×I N be an N th-order tensor such that X (i 1 , i 2 , • • • , i N ) = R1,2 r1,2=1 R1,3 r1,3=1 • • • R 1,N r 1,N =1 R2,3 r2,3=1 • • • R 2,N r 2,N =1 • • • R N -1,N r N -1,N =1 S 1,2 (r 1,2 , r 1,2 )S 1,3 (r 1,3 , r 1,3 ) • • • S 1,N (r 1,N , r 1,N ) S 2,3 (r 2,3 , r 2,3 ) • • • S 2,N (r 2,N , r 2,N ) • • • S N -1,N (r N -1,N , r N -1,N ) G 1 (i 1 , r 1,2 , r 1,3 , • • • , r 1,N ) G 2 (r 1,2 , i 2 , r 2,3 , • • • , r 2,N ) • • • G k (r 1,k , r 2,k , • • • , r k-1,k , i k , r k,k+1 , • • • , r k,N ) • • • G N (r 1,N , r 2,N , • • • , r N -1,N , i N ),(1)\nwhere\nG k ∈ R R 1,k ×R 2,k ו••×R k-1,k ×I k ×R k,k+1 ו••×R k,N\nfor ∀k ∈ K N are N th-order tensors and called TN cores, and S t,l ∈ R R t,l ×R t,l for ∀(t, l) ∈ TL N are diagonal matrices. Then we call (1) an SVDinspired TN decomposition (SVDinsTN) of X , denoted by X = STN(G, S), where G denotes {G k |k ∈ K N } and S denotes {S t,l |(t, l) ∈ TL N }. Remark 1 (SVDinsTN & SVD) SVDinsTN extends upon the \"core&diagonal factor&core\" form of SVD to higher-order cases (see Figure 1(a) & (b)), incorporating the idea of determining rank through non-zero elements in the diagonal factor. In particular, SVDinsTN can degrade into SVD in second-order cases (matrices) when TN cores satisfy orthogonality.\nRemark 2 (SVDinsTN & FCTN) SVDinsTN builds upon FCTN decomposition [37] but can reveal the FCTN rank (TN structure). It achieves this by inserting diagonal factors between any two TN cores in FCTN decomposition and leveraging the number of non-zero elements in the diagonal factors to determine the FCTN rank. In particular, SVDinsTN can transform into a TN (includes but is not limited to FCTN) decomposition by merging the diagonal factors into TN cores through the tensor contraction operation."},{"figure_ref":[],"heading":"A Regularized Method for TN-SS","publication_ref":["b9","b1","b5","b5"],"table_ref":[],"text":"We present an SVDinsTN-based regularized method, which updates TN cores and diagonal factors simultaneously, and imposes a sparse operator to induce the sparsity of diagonal factors to reveal a compact TN structure.\nWe consider an ℓ 1 -norm-based operator for diagonal factors S and Tikhonov regularization [10] for TN cores G. The ℓ 1 -norm-based operator is used to promote the sparsity of S, and the Tikhonov regularization is used to constrict the feasible range of G. Mathematically, the proposed model can be formulated as\nmin G,S 1 2 ∥X -STN(G, S)∥ 2 F + µ 2 k∈K N ∥G k ∥ 2 F + (t,l)∈TL N λ t,l ∥S t,l ∥ 1 ,(2)\nwhere λ t,l > 0 for ∀(t, l) ∈ TL N and µ > 0 are regularization parameters, G denotes {G k |k ∈ K N }, and\nS denotes {S t,l |(t, l) ∈ TL N }.\nWe use the PAM-based algorithm [2] to solve (2), whose solution is obtained by alternately updating\n G k = argmin G k 1 2 ∥X -STN(G, S)∥ 2 F + µ 2 ∥G k ∥ 2 F + ρ 2 ∥G k -Ĝk ∥ 2 F , ∀k ∈ K N , S t,l = argmin S t,l 1 2 ∥X -STN(G, S)∥ 2 F + λ t,l ∥S t,l ∥ 1 + ρ 2 ∥S t,l -Ŝt,l ∥ 2 F , ∀(t, l) ∈ TL N ,(3)\nwhere ρ > 0 is a proximal parameter (we fix ρ = 0.001), and Ĝk and Ŝt,l are the solutions of the G k -subproblem and S t,l -subproblem at the previous iteration, respectively.\n1) Update G k for ∀k ∈ K N : Solving the G k -subproblem requires fixing the other TN cores and diagonal factors. To address this, we use M k to denote the matrix obtained by performing tensor contraction and unfolding operations on all diagonal factors and TN cores except G k . Then, the G k -subproblem can be rewritten as follows:\nmin G k (k) 1 2 ∥X (k) -G k(k) M k ∥ 2 F + µ 2 ∥G k(k) ∥ 2 F + ρ 2 ∥G k(k) -Ĝk(k) ∥ 2 F .(4)\nThe objective function of ( 4) is differentiable, and thus its solution can be obtained by\nG k(k) = X (k) M T k + ρ Ĝk(k) M k M T k + (µ + ρ)I -1 .(5)\n2) Update S t,l for ∀(t, l) ∈ TL N : Solving the S t,l -subproblem requires fixing the other diagonal factors and TN cores. In a similar fashion, we use H t,l to denote the matrix obtained by performing tensor contraction and unfolding operations on all TN cores and diagonal factors except S t,l . Then, the S t,lsubproblem can be rewritten as follows:\nmin s t,l 1 2 ∥x -H t,l s t,l ∥ 2 F + λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F ,(6)\nwhere x = vec(X ) and s t,l = diag(S t,l ). We use an alternating direction method of multipliers (ADMM) [6] to solve (6) and rewrite it as follows:\nmin s t,l 1 2 ∥x -H t,l q t,l ∥ 2 F + λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F s.t. s t,l -q t,l = 0,(7)\nAlgorithm 1 PAM-based algorithm to optimize model (2).\nInput: A tensor X ∈ R I 1 ×I 2 ו••×I N and a parameter γ. Initialization: Initialize S t,l and R t,l by the initialization scheme in Section 3.3 and let\nβ t,l = 1 for ∀(t, l) ∈ TLN ; let G k = 1/ √ I k ones(R 1,k , R 2,k , • • • , R k-1,k , I k , R k,k+1 , • • • , R k,N\n) for ∀k ∈ KN and µ = 1. 1: while not converged do 2:\nLet X = X and λ t,l = γ max(S t,l )(ρ + β t,l )."},{"figure_ref":[],"heading":"3:","publication_ref":[],"table_ref":[],"text":"Update G k by ( 5) and let G k = Fold(G k (k) , k)."},{"figure_ref":[],"heading":"4:","publication_ref":[],"table_ref":[],"text":"for i = 1 to 5 do 5:\nUpdate q t,l , s t,l , and p t,l by (10). 6: end for 7:\nDelete the zero elements in s t,l , let S t,l = diag(s t,l ), and define the size of s t,l as R t,l . 8:\nDelete the corresponding dimensions of G k and let X = STN(G, S). 9:\nCheck the convergence condition: ∥X -X ∥ F ∥ X ∥ F < 10 -5 . 10: end while Output: G k for ∀k ∈ KN , and S t,l and R t,l for ∀(t, l) ∈ TLN .\nwhere q t,l is a auxiliary variable. The augmented Lagrangian function of ( 7) can be expressed as the following concise form:\nL β t,l (s t,l , q t,l , p t,l ) = 1 2 ∥x -H t,l q t,l ∥ 2 F + λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F + β t,l 2 s t,l -q t,l + p t,l β t,l 2 F ,(8)\nwhere p t,l is the Lagrangian multiplier and β t,l is the penalty parameter. Within the ADMM framework, q t,l , s t,l , and p t,l can be solved by alternately updating\n q t,l = argmin q t,l 1 2 ∥x -H t,l q t,l ∥ 2 F + β t,l 2 s t,l -q t,l + p t,l β t,l2\nF , s t,l = argmin s t,l λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F + β t,l 2 s t,l -q t,l + p t,l β t,l2\nF\n, p t,l = p t,l + β t,l (s t,l -q t,l ).\nThat is,\n q t,l = H T t,l H t,l + β t,l I -1 H T t,l x + β t,l s t,l + p t,l , s t,l = shrink ρŝ t,l + β t,l q t,l -p t,l ρ + β t,l , λ t,l ρ + β t,l , p t,l = p t,l + β t,l (s t,l -q t,l ),(10)\nwhere shrink(a, b) = max(a -b, 0) + min(a + b, 0).\nWe describe the pseudocode to optimize model (2) in Algorithm 1. Below, we provide the computational complexity and prove a convergence guarantee for the developed algorithm.\nComputational complexity. For simplicity, we let the size of the N th-order tensor X be\nI × I × • • • × I and the initial rank be (R, R, • • • , R) satisfied R ≤ I. The computational cost involves updat- ing G and S, resulting in costs of O N N k=2 I k R k(N -k)+k-1 + N I N -1 R 2(N -1) + N 3 IR N and O N 2 N k=2 I k R k(N -k)+k +N 4 IR N +N 2 RI 2N , respectively. In summary, the computational cost at each iteration is O N 2 N k=2 I k R k(N -k)+k +N 4 IR N +N 2 RI 2N .\nTheorem 1 (Convergence) The sequence generated by Algorithm 1, denoted by {G (s) , S (s) } s∈N , converges to a critical point of the optimization problem in (2).\nThe numerical convergence is shown in Section 4.3."},{"figure_ref":["fig_3"],"heading":"Initialization Scheme","publication_ref":[],"table_ref":[],"text":"SVDinsTN would encounter high computational cost in the first several iterations if the TN rank R t,l for ∀(t, l) ∈ TL N are initialized with large values. This is due to the utilization of a \"fully-connected\" topology as a starting point. To solve this issue, we design a novel initialization scheme, which can effectively reduce the initial values of the TN rank.\nWe first give an upper bound for R t,l in Theorem 2, by which we then design an initialization scheme for both R t,l and diagonal factors S t,l ∈ R R t,l ×R t,l for ∀(t, l) ∈ TL N .\nTheorem 2 Let X ∈ R I1×I2ו••×I N be an N th-order tensor, then there exists an SVDinsTN (1) with the TN rank R t,l ≤ min(rank(X (t) ), rank(X (l) )) for ∀(t, l) ∈ TL N .\nTheorem 2 implies that min(rank(X (t) ),rank(X (l) )) can be the initial value of R t,l . For real-world data, this value is usually embodied by the rank of mode-(t, l) slices2 of X . Hence, we initialize R t,l and S t,l by virtue of truncated matrix SVD of mode-(t, l) slices of X , which consists of the following two steps.\nStep 1: We first calculate the mean of all mode-(t, l) slices of X and denote it by X t,l . Then we perform matrix SVD on X t,l to obtain s t,l ∈ R min(It,I l ) , whose elements are singular values of X t,l .\nStep 2: We first let s t,l = shrink s t,l , γ max(s t,l ) abs(s t,l )+ε and delete zero elements in s t,l . Then we let S t,l = diag(s t,l ) and define the size of s t,l as R t,l .\nIn practical applications, the shrink operation in Step 2 effectively reduces the initial value of R t,l by projecting very small singular values in s t,l to zero. As a result, the challenge of high computational costs in the first several iterations of SVDinsTN can be effectively addressed (see Figure 2 for a numerical illustration)."},{"figure_ref":[],"heading":"Numerical Experiments","publication_ref":[],"table_ref":[],"text":"In this section, we present numerical experiments on both synthetic and real-world data to evaluate the performance of the proposed SVDinsTN. The primary objective is to validate the following three Claims: A: SVDinsTN can reveal a customized TN structure that aligns with the unique structure of a given tensor. B: SVDinsTN can greatly reduce time costs while achieving a comparable representation ability to state-of-the-art TN-SS methods. Moreover, SVDinsTN can also surpass existing tensor decomposition methods with pre-defined topologies regarding representation ability. C: SVDinsTN can outperform existing tensor decomposition methods in the tensor completion task, highlighting its effectiveness as a valuable tool in applications."},{"figure_ref":[],"heading":"Experiments for validating Claim A","publication_ref":[],"table_ref":["tab_3"],"text":"First of all, we conduct experiments to validate Claim A. Real-world data lacks a true TN structure, and thus we consider synthetic data in this experiment.\nData generation. We first randomly generate\nG k ∈ R R 1,k ×R 2,k ו••×R k-1,k ×I k ×R k,k+1 ו••×R k,N for ∀k ∈ K N and S t,l ∈ R R t,l ×R t,l\nfor ∀(t, l) ∈ TL N , whose elements are taken from a uniform distribution between 0 and 1. Then we obtain the synthetic tensor by X = STN(G, S).\nExperiment setting. We test both fourth-order tensors of size 16 × 18 × 20 × 22 and fifth-order tensors of size 14 × 16 × 18 × 20 × 22, and consider five different kinds of TN structures. For each structure, we conduct 100 independent tests and regenerate the synthetic data to ensure reliable and unbiased results. The ability of the proposed SVDinsTN to reveal TN structure is measured by the success rate of the output structures, defined as S T /T × 100%, where T = 100 is the total number of tests and S T is the number of tests that accurately output the true TN structure. In all tests, the parameter γ is set to 0.0015.\nResult analysis. Table 2 presents the success rate of the output TN structures obtained by the proposed SVDinsTN in 100 independent tests on fourth-order and fifth-order tensors. It can be observed that the proposed SVDinsTN consistently yields high success rates of over 95% in all test cases. Notably, in approximately half of the test cases, the success rates reach a perfect score of 100%. Moreover, it is also worth mentioning that in the test on fifth-order tensors, we consider two isomorphic topologies: the \"ring\" topology and the \"five-star\" topology. These two topologies are both the \"ring\" topology (TR Success rate 100% 98% 96% 97% 100% \nG 1 → G 2 → G 3 → G 4 → G 5 → G 1 and G 1 → G 3 → G 5 → G 2 → G 4 → G 1 , respectively.\nIt can be seen that despite the isomorphism, the proposed SVDinsTN can accurately identify the correct permutation for each topology."},{"figure_ref":["fig_3","fig_3"],"heading":"Experiments for Validating Claim B","publication_ref":["b34","b36","b10","b12","b14","b13"],"table_ref":["tab_4","tab_6"],"text":"We conduct experiments to validate Claim B. We consider both real-world data and synthetic data, and use different methods to represent it in this experiment.\nExperiment setting. We test three light field data3 , named Bunny, Knights, and Truck, which are fifth-order tensors of size 40 × 60 × 3 × 9 × 9 (spatial height × spatial width × color channel × vertical grid × horizontal grid). We employ six representative methods as the compared baselines, including two methods with pre-defined topology: TRALS [35] and FCTNALS [37], and four TN-SS methods: TNGreedy [11], TNGA [13], TNLS [15], and TNALE [14]. We represent the test light field data by different methods and calculate the corresponding compression ratio (CR) to achieve a certain reconstruction error (RE) bound. The CR is defined as F G /F X × 100%, where F G is the number of elements of TN cores used to represent a tensor and F X is the number of total elements of the original tensor. The RE is defined as ∥X -X ∥ F /∥X ∥ F , where X is the original data and X is the reconstructed data. In all tests, we select the parameter γ from the interval [10 -7 , 10 -3 ].\nResult analysis. Table 3 reports the CR and run time of different methods on fifth-order light field data. The results show that the proposed SVDinsTN achieves significantly lower CR than TRALS and FCTNALS, which are methods with pre-defined topology. This indicates that SVDinsTN can obtain a more compact structure than the one pre-defined. Besides, while SVDinsTN requires the determination of diagonal matrices alongside TN cores, its iterative process generates progressively simpler structures, enhancing the computational efficiency. Consequently, SVDinsTN demonstrates faster performance compared to TRALS and FCTNALS. Furthermore, compared to the TN-SS methods, the proposed SVDinsTN achieves a substantial speed improvement while maintaining a comparable level of CR. Remarkably, SVDinsTN achieves an acceleration of approximately 100 ∼ 1000 times over TNGA, TNLS, and TNALE. This is because TNGreedy, TNGA, TNLS, and TNALE adopt the \"sampling-evaluation\" framework, necessitating a significant number of repeated structure evaluations. However, SVDinsTN introduces a regularized modeling framework, requiring only a single evaluation.\nImpact of the initialization scheme. We analyze the impact of the shrink operation in the initialization scheme under this experimental setting. In Figure 2, we present the run time comparison of the initial five iterations between SVDinsTN with the shrink operation and SVDinsTN without the shrink operation in the initialization scheme. As observed, the designed initialization scheme greatly reduces the computational costs in the first several iterations. These findings validate the analysis presented in Section 3.3.\nHigher-order cases. We analyze whether the proposed SVDinsTN still performs well on higher-order tensors. We randomly generate 6th-, 8th-, and 10th-order tensors by using the same procedure in Section ??.\nThe size of each tensor mode is randomly selected from {5, 6, 7, 8}, the edge number of each TN is randomly selected from {6, 8, 10}, and the rank of each edge is randomly selected from {2, 3}. For each tensor order, we randomly generate 5 tensors. We compare SVDinsTN and baseline methods in terms of CR and run time when reaching the RE bound of 0.01, and show the results in Table 4. As observed, SVDinsTN is applicable to higher orders beyond 5, and even up to 10. The behind rational is the truncated SVD used in initialization restricts the initial values of the rank for each edge to a relatively small range, thus improving computational and storage efficiency (see Figure 2). As the iterations progress, the sparsity regularization in the model leads to progressively simpler learned structures, further boosting efficiency. "},{"figure_ref":["fig_4","fig_5"],"heading":"Experiments for Validating Claim C","publication_ref":["b33","b25","b2","b29","b24"],"table_ref":["tab_7"],"text":"We conduct experiments to validate Claim C. We employ the proposed SVDinsTN to a fundamental application, i.e., tensor completion (TC), and compare it with the state-of-the-art tensor decompositionbased TC methods. Given an incomplete observation tensor\nF ∈ R I1×I2ו••×I N of X ∈ R I1×I2ו••×I N ,\nthe proposed TC method first updates G and S by Algorithm 1, and then updates the target tensor X as follows: X = P Ω c ((STN(G, S) + ρ X )/(1 + ρ)) + P Ω (F), where Ω is the index set of the known elements, P Ω (X ) is a projection operator which projects the elements in Ω to themselves and all others to zeros, X\nis the result at the previous iteration, and the initial X is F.\nExperiment setting. We test four color videos 4 , named Bunny, News, Salesman, and Silent, which are fourth-order tensors of size 144 × 176 × 3 × 50 (spatial height×spatial width×color channel×frame). We employ five baseline methods for comparison, named FBCP [34], TMac [26], TMacTT [3], TRLRF [30], and TW [25], respectively. All of them are based on different tensor decompositions. We set the missing ratio (MR) to 90%, which is defined as the ratio of the number of missing elements to the total number of elements. We evaluate the reconstructed quality by the mean peak signal-to-noise ratio (MPSNR) computed across all frames. In all tests, the parameter γ is set to 0.0003.\nResult analysis. Table 5 reports MPSNR and run time obtained by different TC methods. As observed, the proposed SVDinsTN consistently achieves the highest MPSNR values among all utilized TC methods across all test color videos. Besides, SVDinsTN exhibits comparable run time to the baseline methods.\nFigure 3 displays the reconstructed images and their corresponding residual images at the 25th frame of News. We observe that the proposed SVDinsTN outperforms the baseline methods in terms of visual quality, particularly with respect to background cleanliness and local details (e.g., \"dancer\") recovery.\nNumerical convergence. In Theorem 1, we provide a theoretical convergence guarantee for the proposed method. Here, we numerically verify the convergence. Figure 4 presents the relative change in the reconstructed color videos at each iteration compared to their respective previous iterations. We observe that the values of the relative change achieved by the proposed method decrease and gradually tend to zero as the number of iterations increases. This justifies the numerical convergence of the proposed method. Here the relative change is defined as ∥X c -X p ∥ F /∥X p ∥ F , and X c and X p are the results of the current iteration and its previous iteration."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"Summary. We propose a novel TN paradigm, called SVDinsTN, which enables us to solve the challenging TN-SS problem from a regularized modeling perspective. This perspective renders our model highly amenable to easy solutions, allowing us to leverage well-established optimization algorithms for tackling the regularized model. As a result, the proposed method achieves approximately 100 ∼ 1000 times acceleration compared to the state-of-the-art TN-SS methods with a comparable representation ability. Additionally, SVDinsTN demonstrates its effectiveness as a valuable tool in practical applications.\nLimitations. In existing research on the TN-SS problem, two challenging issues remain open. One is the computationally consuming issue, and the other is the theoretical guarantee of the optimal TN structure. SVDinsTN provides a new regularized modeling perspective to fundamentally address the computationally consuming issue and proves theorems regarding the upper bound of the TN rank and the convergence. However, the theoretical guarantee of the optimal TN structure is still an open problem. Solving this issue will be the direction of our future work."},{"figure_ref":[],"heading":"Acknowledgments","publication_ref":[],"table_ref":[],"text":"We would like to express our gratitude to Guillaume Rabusseau for his valuable assistance in correcting the experimental results of \"TN-Greedy\"."}],"string":"[\n {\n \"figure_ref\": [\n \"fig_0\",\n \"fig_1\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b0\",\n \"b22\",\n \"b33\",\n \"b3\",\n \"b8\",\n \"b18\",\n \"b26\",\n \"b31\",\n \"b32\",\n \"b11\",\n \"b13\",\n \"b10\",\n \"b12\",\n \"b13\",\n \"b10\",\n \"b12\",\n \"b13\",\n \"b12\",\n \"b13\",\n \"b34\",\n \"b36\",\n \"b10\",\n \"b12\",\n \"b14\",\n \"b13\"\n ],\n \"table_ref\": [\n \"tab_4\"\n ],\n \"text\": \"Tensor network (TN) representation, which aims to express higher-order data with low-dimensional tensors (called TN cores) by a specific operation among them, has gained significant attention in various areas of data analysis [1,23,34], machine learning [4,9], computer vision [19,27,32,33], etc. By regarding TN cores as nodes and operations as edges, a TN corresponds to a graph (called TN topology). The vector composed of the weights of all edges in the topology is defined as the TN rank. TN structure (including topology and rank) search (TN-SS) aims to search for a suitable TN structure to achieve a compact representation for a given tensor, which is known as a challenging and theoretically NP-hard problem [12,14]. The selection of TN structure dramatically impacts the performance of TN representation in practical applications [11,13,14].\\nRecently, there have been several notable efforts to tackle the TN-SS problem [11,13,14]. But most of them adopt the \\\"sampling-evaluation\\\" framework, which requires sampling a large number of structures as candidates and conducting numerous repeated structure evaluations. For instance, for a tensor of size 40 × 60 × 3 × 9 × 9 (used in Section 4.2), TNGA in [13] requires thousands of evaluations and TNALE in [14] requires hundreds of evaluations, where each evaluation entails solving an optimization problem to compute TN cores iteratively. Consequently, the computational cost becomes exceedingly high. A meaningful question is whether we can optimize the TN structure simultaneously during the computation Preprint. [35] and FCTNALS [37] are methods with pre-defined topologies, and TNGreedy [11], TNGA [13], TNLS [15], and TNALE [14] are TN-SS methods (please see more results in Table 3).\\nof TN cores, thus escaping the \\\"sampling-evaluation\\\" framework and fundamentally addressing the computationally consuming issue.\\nIn this paper, we introduce for the first time a regularized modeling perspective on solving the TN-SS problem. This perspective enables us to optimize the TN structure simultaneously during the computation of TN cores, effectively eliminating the need for repetitive structure evaluations. To be specific, we propose a novel TN paradigm, termed as SVD-inspired TN decomposition (SVDinsTN), by inserting diagonal factors between any two TN cores in the \\\"fully-connected\\\" topology (see Figure 1(b)). The intuition behind SVDinsTN is to leverage the sparsity of the inserted diagonal factors to reveal a compact TN structure and utilize the TN cores (merged with the diagonal factors) to represent a given tensor. Based on SVDinsTN, we establish a regularized model, which updates the TN cores and diagonal factors iteratively and imposes a sparse operator to induce the sparsity of the diagonal factors. In theory, we prove a convergence guarantee for the proposed method and establish an upper bound for the TN rank. In particular, we design a novel initialization scheme for the proposed method based on the upper bound. This initialization scheme enables the proposed method to overcome the high computational cost in the first several iterations, which is caused by the utilization of a \\\"fully-connected\\\" topology as the starting point. As a result, SVDinsTN is capable of capturing a customized TN structure and providing a compact representation for a given tensor in an efficient manner. In summary, we make the following three contributions.\\n• We propose SVDinsTN, a new TN paradigm, that enables us to optimize the TN structure during the computation of TN cores, greatly reducing the computational cost. • In theory, we prove a convergence guarantee for the proposed method and establish an upper bound for the TN rank involved in SVDinsTN. The upper bound serves as a guide for designing an efficient initialization scheme. • Experimental results verify numerically that the proposed method achieves 100 ∼ 1000 times acceleration compared to the state-of-the-art TN-SS methods with a comparable representation ability (see Figure 1(c)).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Works\",\n \"publication_ref\": [\n \"b3\",\n \"b4\",\n \"b27\",\n \"b21\",\n \"b34\",\n \"b36\",\n \"b34\",\n \"b28\",\n \"b30\",\n \"b36\",\n \"b7\",\n \"b10\",\n \"b12\",\n \"b13\",\n \"b14\",\n \"b16\",\n \"b17\",\n \"b19\",\n \"b23\",\n \"b10\",\n \"b12\",\n \"b13\"\n ],\n \"table_ref\": [],\n \"text\": \"TN representation1 aims to find a set of small-sized TN cores to express a large-sized tensor under a TN structure (including topology and rank) [4,5,28]. In the past decades, many works focused on TN representation with a fixed TN topology, such as tensor train (TT) decomposition with a \\\"chain\\\" topology [22], tensor ring (TR) decomposition with a\\\"ring\\\" topology [35], fully-connected tensor network (FCTN) decomposition with a \\\"fully-connected\\\" topology [37], etc. In addition, these works also presented various methods to optimize the TN cores, such as alternating least square(ALS) [35], gradient descent (GD) [29,31], proximal alternating minimization (PAM) [37], etc. In contrast, SVDinsTN can reveal a compact TN structure for a given tensor, surpassing methods with pre-defined topologies in terms of representation ability.\\nTN structure search (TN-SS) aims to search for a suitable or optimal TN structure, including both topology and rank, to achieve a compact representation for a given tensor [8,11,13,14,15,17,18,20,24]. However, most existing TN-SS methods adopt the \\\"sampling-evaluation\\\" framework, making them require the utilization of heuristic search algorithms such as the greedy algorithm [11], genetic algorithm [13], and alternating local enumeration algorithm [14] to sample candidate structures and then evaluate them one by one. Therefore, these methods inevitably suffer from prohibitively high computational costs due to the numerous repeated evaluations, with each evaluation involving solving an optimization problem concerning the iterative calculation of TN cores. In contrast, SVDinsTN solves the TN-SS problem from a regularized modeling perspective, which avoids the repeated structure evaluations and greatly decreases the computational cost.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Notations and Preliminaries\",\n \"publication_ref\": [\n \"b36\"\n ],\n \"table_ref\": [],\n \"text\": \"We first summarize notations and several operations that will be used in the rest of the paper. Then, we review definitions related to tensor networks.\\nA tensor is a multi-dimensional array, and the number of dimensions (also called modes) of which is referred to as the tensor order. In the paper, first-order tensors (vectors), second-order tensors (matrices), and N th-order tensors are denoted by x ∈ R I1 , X ∈ R I1×I2 , and X ∈ R I1×I2ו••×I N , respectively. We use ∥X ∥ F and ∥X ∥ 1 to denote the Frobenius norm and ℓ 1 -norm of X , respectively. For brevity, we let K N denote the set {1, 2, • • • , N }, which is used for labeling individual indices. Similarly, we let TL N denote the set {(t, l)|1 ≤ t < l ≤ N ; t, l ∈ N}, which is used for labeling paired indices.\\nWe next provide a review of several operations on tensors [37].\\nThe generalized tensor unfolding is an operation that converts a tensor into a matrix by merging a group of tensor modes into the rows of the matrix and merging the remaining modes into the columns. For example, an I 1 × I 2 × I 3 × I 4 tensor can be unfolded into an I 1 I 3 × I 2 I 4 matrix, denoted by X [1,3;2,4] . We use X [1,3;2,4] = GenUnfold(X , [1, 3; 2, 4]) and X = GenFold(X [1,3;2,4] , [1, 3; 2, 4]) to simply denote the corresponding operation and inverse operation, respectively. We also use X (2) to simply represent X [2;1,3,4] ∈ R I2×I1I2I4 , which is also called mode-2 unfolding. Besides, we use X (2) = Unfold(X , 2) and X = Fold(X (2) , 2) to simply denote the corresponding operation and inverse operation, respectively.\\nThe generalized tensor transposition is an operation that rearranges tensor modes in the order specified by a vector. For example, an I 1 × I 2 × I 3 × I 4 tensor can be transposed into an I 3 × I 2 × I 1 × I 4 tensor, denoted by ⃗ X n with n = (3, 2, 1, 4). We use ⃗ X n = permute(X , n) and X = ipermute( ⃗ X n , n) to simply denote the corresponding operation and inverse operation, respectively.\\nThe tensor contraction is an operation that obtains a new tensor by pairing, multiplying, and summing indices of certain modes of two tensors. For example, if a fourth-order tensor X ∈ R I1×I2×I3×I4 and a third-order tensor Y ∈ R J1×J2×J3 satisfy I 2 = J 1 and I 4 = J 2 , then the tensor contraction between the 2nd and 4th modes of X and the 1st and 2nd modes of Y yields a tensor Z = X × 1,2 2,4 Y ∈ R I1×I3×J3 . The elements of Z are calculated as follows:\\nZ(i 1 , i 3 , j 3 ) = I2 i2=1 I4 i4=1 X (i 1 , i 2 , i 3 , i 4 )Y(i 2 , i 4 , j 3 ).\\nIn Table 1, we give the interpretations of the operations \\\"diag\\\", \\\"ones\\\", \\\"zeros\\\", and \\\"vec\\\".\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Tensor Network\",\n \"publication_ref\": [\n \"b3\",\n \"b36\",\n \"b15\"\n ],\n \"table_ref\": [],\n \"text\": \"In general, a tensor network (TN) is defined as a collection of low-dimensional tensors, known as TN cores, in which some or all modes are contracted according to specific operations [4]. The primary purpose of a TN is to represent higher-order data using these TN cores. By considering TN cores as nodes and operations between modes of cores as edges, we define the graph formed by these nodes and edges as the TN topology. Additionally, we assign a non-negative integer weight to each edge to indicate the size of the corresponding TN core, and call the vector composed of these edge weights the TN rank. Consequently, a TN structure refers to a weighted graph comprising nodes, edges, and weights, encompassing both the TN topology and TN rank. This paper focuses on only a class of TNs that employs tensor contraction as the operation among TN cores and adopts a simple graph as the TN topology. More particularly, when representing an N th-order tensor X , this class of TNs comprises precisely N TN cores, each corresponding to one mode of X . A notable method is the FCTN decomposition, which represents an N th-order tensor [37]. In this decomposition, any two cores G l and G t for (t, l) ∈ TL N share an equal-sized mode R t,l used for tensor contraction. We denote the FCTN decomposition by\\nX ∈ R I1×I2ו••×I N by N small-sized N th-order cores denoted by G k ∈ R R 1,k ×R 2,k ו••×R k-1,k ×I k ×R k,k+1 ו••×R k,N for k ∈ K N\\nX = FCTN(G 1 , G 2 , • • • , G N ) and define the FCTN rank as the vector (R 1,2 , R 1,3 , • • • , R 1,N , R 2,3 , • • • , R 2,N , • • • , R N -1,N ) ∈ R N (N -1)/2 .\\nAccording to the concept of tensor contraction, removing rank-one edges in the TN topology does not change the expression of the TN. This means that if any element in the FCTN rank is equal to one, the corresponding edge can be harmlessly eliminated from the \\\"fully-connected\\\" topology. For instance, a \\\"fully-connected\\\" topology with the rank\\n(R 1,2 , 1, • • • , 1, R 2,3 , 1, • • • , 1, R N -2,N -1 , R N -1,N ) can be converted into a \\\"chain\\\" topology with rank (R 1,2 , R 2,3 , • • • , R N -1,N\\n) in this manner. This fact can be formally stated as follows.\\nProperty 1 [16] There exists a one-to-one correspondence between the TN structure and FCTN rank.\\nThis work is inspired by Property 1, searching for a more compact TN structure by optimizing the FCTN rank.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"An Efficient Method for TN-SS\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We propose an efficient method to solve the TN-SS problem from a regularized modeling perspective. Unlike the existing \\\"sampling-evaluation\\\" framework, the main idea of the proposed method is to optimize the TN structure simultaneously during the computation of TN cores, thereby eliminating the need for repetitive structure evaluations and greatly decreasing the computational cost.\"\n },\n {\n \"figure_ref\": [\n \"fig_0\"\n ],\n \"heading\": \"SVDinsTN\",\n \"publication_ref\": [\n \"b36\"\n ],\n \"table_ref\": [],\n \"text\": \"We start with the definition of the following SVDinsTN.\\nDefinition 1 (SVDinsTN) Let X ∈ R I1×I2ו••×I N be an N th-order tensor such that X (i 1 , i 2 , • • • , i N ) = R1,2 r1,2=1 R1,3 r1,3=1 • • • R 1,N r 1,N =1 R2,3 r2,3=1 • • • R 2,N r 2,N =1 • • • R N -1,N r N -1,N =1 S 1,2 (r 1,2 , r 1,2 )S 1,3 (r 1,3 , r 1,3 ) • • • S 1,N (r 1,N , r 1,N ) S 2,3 (r 2,3 , r 2,3 ) • • • S 2,N (r 2,N , r 2,N ) • • • S N -1,N (r N -1,N , r N -1,N ) G 1 (i 1 , r 1,2 , r 1,3 , • • • , r 1,N ) G 2 (r 1,2 , i 2 , r 2,3 , • • • , r 2,N ) • • • G k (r 1,k , r 2,k , • • • , r k-1,k , i k , r k,k+1 , • • • , r k,N ) • • • G N (r 1,N , r 2,N , • • • , r N -1,N , i N ),(1)\\nwhere\\nG k ∈ R R 1,k ×R 2,k ו••×R k-1,k ×I k ×R k,k+1 ו••×R k,N\\nfor ∀k ∈ K N are N th-order tensors and called TN cores, and S t,l ∈ R R t,l ×R t,l for ∀(t, l) ∈ TL N are diagonal matrices. Then we call (1) an SVDinspired TN decomposition (SVDinsTN) of X , denoted by X = STN(G, S), where G denotes {G k |k ∈ K N } and S denotes {S t,l |(t, l) ∈ TL N }. Remark 1 (SVDinsTN & SVD) SVDinsTN extends upon the \\\"core&diagonal factor&core\\\" form of SVD to higher-order cases (see Figure 1(a) & (b)), incorporating the idea of determining rank through non-zero elements in the diagonal factor. In particular, SVDinsTN can degrade into SVD in second-order cases (matrices) when TN cores satisfy orthogonality.\\nRemark 2 (SVDinsTN & FCTN) SVDinsTN builds upon FCTN decomposition [37] but can reveal the FCTN rank (TN structure). It achieves this by inserting diagonal factors between any two TN cores in FCTN decomposition and leveraging the number of non-zero elements in the diagonal factors to determine the FCTN rank. In particular, SVDinsTN can transform into a TN (includes but is not limited to FCTN) decomposition by merging the diagonal factors into TN cores through the tensor contraction operation.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Regularized Method for TN-SS\",\n \"publication_ref\": [\n \"b9\",\n \"b1\",\n \"b5\",\n \"b5\"\n ],\n \"table_ref\": [],\n \"text\": \"We present an SVDinsTN-based regularized method, which updates TN cores and diagonal factors simultaneously, and imposes a sparse operator to induce the sparsity of diagonal factors to reveal a compact TN structure.\\nWe consider an ℓ 1 -norm-based operator for diagonal factors S and Tikhonov regularization [10] for TN cores G. The ℓ 1 -norm-based operator is used to promote the sparsity of S, and the Tikhonov regularization is used to constrict the feasible range of G. Mathematically, the proposed model can be formulated as\\nmin G,S 1 2 ∥X -STN(G, S)∥ 2 F + µ 2 k∈K N ∥G k ∥ 2 F + (t,l)∈TL N λ t,l ∥S t,l ∥ 1 ,(2)\\nwhere λ t,l > 0 for ∀(t, l) ∈ TL N and µ > 0 are regularization parameters, G denotes {G k |k ∈ K N }, and\\nS denotes {S t,l |(t, l) ∈ TL N }.\\nWe use the PAM-based algorithm [2] to solve (2), whose solution is obtained by alternately updating\\n G k = argmin G k 1 2 ∥X -STN(G, S)∥ 2 F + µ 2 ∥G k ∥ 2 F + ρ 2 ∥G k -Ĝk ∥ 2 F , ∀k ∈ K N , S t,l = argmin S t,l 1 2 ∥X -STN(G, S)∥ 2 F + λ t,l ∥S t,l ∥ 1 + ρ 2 ∥S t,l -Ŝt,l ∥ 2 F , ∀(t, l) ∈ TL N ,(3)\\nwhere ρ > 0 is a proximal parameter (we fix ρ = 0.001), and Ĝk and Ŝt,l are the solutions of the G k -subproblem and S t,l -subproblem at the previous iteration, respectively.\\n1) Update G k for ∀k ∈ K N : Solving the G k -subproblem requires fixing the other TN cores and diagonal factors. To address this, we use M k to denote the matrix obtained by performing tensor contraction and unfolding operations on all diagonal factors and TN cores except G k . Then, the G k -subproblem can be rewritten as follows:\\nmin G k (k) 1 2 ∥X (k) -G k(k) M k ∥ 2 F + µ 2 ∥G k(k) ∥ 2 F + ρ 2 ∥G k(k) -Ĝk(k) ∥ 2 F .(4)\\nThe objective function of ( 4) is differentiable, and thus its solution can be obtained by\\nG k(k) = X (k) M T k + ρ Ĝk(k) M k M T k + (µ + ρ)I -1 .(5)\\n2) Update S t,l for ∀(t, l) ∈ TL N : Solving the S t,l -subproblem requires fixing the other diagonal factors and TN cores. In a similar fashion, we use H t,l to denote the matrix obtained by performing tensor contraction and unfolding operations on all TN cores and diagonal factors except S t,l . Then, the S t,lsubproblem can be rewritten as follows:\\nmin s t,l 1 2 ∥x -H t,l s t,l ∥ 2 F + λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F ,(6)\\nwhere x = vec(X ) and s t,l = diag(S t,l ). We use an alternating direction method of multipliers (ADMM) [6] to solve (6) and rewrite it as follows:\\nmin s t,l 1 2 ∥x -H t,l q t,l ∥ 2 F + λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F s.t. s t,l -q t,l = 0,(7)\\nAlgorithm 1 PAM-based algorithm to optimize model (2).\\nInput: A tensor X ∈ R I 1 ×I 2 ו••×I N and a parameter γ. Initialization: Initialize S t,l and R t,l by the initialization scheme in Section 3.3 and let\\nβ t,l = 1 for ∀(t, l) ∈ TLN ; let G k = 1/ √ I k ones(R 1,k , R 2,k , • • • , R k-1,k , I k , R k,k+1 , • • • , R k,N\\n) for ∀k ∈ KN and µ = 1. 1: while not converged do 2:\\nLet X = X and λ t,l = γ max(S t,l )(ρ + β t,l ).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"3:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Update G k by ( 5) and let G k = Fold(G k (k) , k).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"4:\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"for i = 1 to 5 do 5:\\nUpdate q t,l , s t,l , and p t,l by (10). 6: end for 7:\\nDelete the zero elements in s t,l , let S t,l = diag(s t,l ), and define the size of s t,l as R t,l . 8:\\nDelete the corresponding dimensions of G k and let X = STN(G, S). 9:\\nCheck the convergence condition: ∥X -X ∥ F ∥ X ∥ F < 10 -5 . 10: end while Output: G k for ∀k ∈ KN , and S t,l and R t,l for ∀(t, l) ∈ TLN .\\nwhere q t,l is a auxiliary variable. The augmented Lagrangian function of ( 7) can be expressed as the following concise form:\\nL β t,l (s t,l , q t,l , p t,l ) = 1 2 ∥x -H t,l q t,l ∥ 2 F + λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F + β t,l 2 s t,l -q t,l + p t,l β t,l 2 F ,(8)\\nwhere p t,l is the Lagrangian multiplier and β t,l is the penalty parameter. Within the ADMM framework, q t,l , s t,l , and p t,l can be solved by alternately updating\\n q t,l = argmin q t,l 1 2 ∥x -H t,l q t,l ∥ 2 F + β t,l 2 s t,l -q t,l + p t,l β t,l2\\nF , s t,l = argmin s t,l λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F + β t,l 2 s t,l -q t,l + p t,l β t,l2\\nF\\n, p t,l = p t,l + β t,l (s t,l -q t,l ).\\nThat is,\\n q t,l = H T t,l H t,l + β t,l I -1 H T t,l x + β t,l s t,l + p t,l , s t,l = shrink ρŝ t,l + β t,l q t,l -p t,l ρ + β t,l , λ t,l ρ + β t,l , p t,l = p t,l + β t,l (s t,l -q t,l ),(10)\\nwhere shrink(a, b) = max(a -b, 0) + min(a + b, 0).\\nWe describe the pseudocode to optimize model (2) in Algorithm 1. Below, we provide the computational complexity and prove a convergence guarantee for the developed algorithm.\\nComputational complexity. For simplicity, we let the size of the N th-order tensor X be\\nI × I × • • • × I and the initial rank be (R, R, • • • , R) satisfied R ≤ I. The computational cost involves updat- ing G and S, resulting in costs of O N N k=2 I k R k(N -k)+k-1 + N I N -1 R 2(N -1) + N 3 IR N and O N 2 N k=2 I k R k(N -k)+k +N 4 IR N +N 2 RI 2N , respectively. In summary, the computational cost at each iteration is O N 2 N k=2 I k R k(N -k)+k +N 4 IR N +N 2 RI 2N .\\nTheorem 1 (Convergence) The sequence generated by Algorithm 1, denoted by {G (s) , S (s) } s∈N , converges to a critical point of the optimization problem in (2).\\nThe numerical convergence is shown in Section 4.3.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\"\n ],\n \"heading\": \"Initialization Scheme\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"SVDinsTN would encounter high computational cost in the first several iterations if the TN rank R t,l for ∀(t, l) ∈ TL N are initialized with large values. This is due to the utilization of a \\\"fully-connected\\\" topology as a starting point. To solve this issue, we design a novel initialization scheme, which can effectively reduce the initial values of the TN rank.\\nWe first give an upper bound for R t,l in Theorem 2, by which we then design an initialization scheme for both R t,l and diagonal factors S t,l ∈ R R t,l ×R t,l for ∀(t, l) ∈ TL N .\\nTheorem 2 Let X ∈ R I1×I2ו••×I N be an N th-order tensor, then there exists an SVDinsTN (1) with the TN rank R t,l ≤ min(rank(X (t) ), rank(X (l) )) for ∀(t, l) ∈ TL N .\\nTheorem 2 implies that min(rank(X (t) ),rank(X (l) )) can be the initial value of R t,l . For real-world data, this value is usually embodied by the rank of mode-(t, l) slices2 of X . Hence, we initialize R t,l and S t,l by virtue of truncated matrix SVD of mode-(t, l) slices of X , which consists of the following two steps.\\nStep 1: We first calculate the mean of all mode-(t, l) slices of X and denote it by X t,l . Then we perform matrix SVD on X t,l to obtain s t,l ∈ R min(It,I l ) , whose elements are singular values of X t,l .\\nStep 2: We first let s t,l = shrink s t,l , γ max(s t,l ) abs(s t,l )+ε and delete zero elements in s t,l . Then we let S t,l = diag(s t,l ) and define the size of s t,l as R t,l .\\nIn practical applications, the shrink operation in Step 2 effectively reduces the initial value of R t,l by projecting very small singular values in s t,l to zero. As a result, the challenge of high computational costs in the first several iterations of SVDinsTN can be effectively addressed (see Figure 2 for a numerical illustration).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Numerical Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this section, we present numerical experiments on both synthetic and real-world data to evaluate the performance of the proposed SVDinsTN. The primary objective is to validate the following three Claims: A: SVDinsTN can reveal a customized TN structure that aligns with the unique structure of a given tensor. B: SVDinsTN can greatly reduce time costs while achieving a comparable representation ability to state-of-the-art TN-SS methods. Moreover, SVDinsTN can also surpass existing tensor decomposition methods with pre-defined topologies regarding representation ability. C: SVDinsTN can outperform existing tensor decomposition methods in the tensor completion task, highlighting its effectiveness as a valuable tool in applications.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments for validating Claim A\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_3\"\n ],\n \"text\": \"First of all, we conduct experiments to validate Claim A. Real-world data lacks a true TN structure, and thus we consider synthetic data in this experiment.\\nData generation. We first randomly generate\\nG k ∈ R R 1,k ×R 2,k ו••×R k-1,k ×I k ×R k,k+1 ו••×R k,N for ∀k ∈ K N and S t,l ∈ R R t,l ×R t,l\\nfor ∀(t, l) ∈ TL N , whose elements are taken from a uniform distribution between 0 and 1. Then we obtain the synthetic tensor by X = STN(G, S).\\nExperiment setting. We test both fourth-order tensors of size 16 × 18 × 20 × 22 and fifth-order tensors of size 14 × 16 × 18 × 20 × 22, and consider five different kinds of TN structures. For each structure, we conduct 100 independent tests and regenerate the synthetic data to ensure reliable and unbiased results. The ability of the proposed SVDinsTN to reveal TN structure is measured by the success rate of the output structures, defined as S T /T × 100%, where T = 100 is the total number of tests and S T is the number of tests that accurately output the true TN structure. In all tests, the parameter γ is set to 0.0015.\\nResult analysis. Table 2 presents the success rate of the output TN structures obtained by the proposed SVDinsTN in 100 independent tests on fourth-order and fifth-order tensors. It can be observed that the proposed SVDinsTN consistently yields high success rates of over 95% in all test cases. Notably, in approximately half of the test cases, the success rates reach a perfect score of 100%. Moreover, it is also worth mentioning that in the test on fifth-order tensors, we consider two isomorphic topologies: the \\\"ring\\\" topology and the \\\"five-star\\\" topology. These two topologies are both the \\\"ring\\\" topology (TR Success rate 100% 98% 96% 97% 100% \\nG 1 → G 2 → G 3 → G 4 → G 5 → G 1 and G 1 → G 3 → G 5 → G 2 → G 4 → G 1 , respectively.\\nIt can be seen that despite the isomorphism, the proposed SVDinsTN can accurately identify the correct permutation for each topology.\"\n },\n {\n \"figure_ref\": [\n \"fig_3\",\n \"fig_3\"\n ],\n \"heading\": \"Experiments for Validating Claim B\",\n \"publication_ref\": [\n \"b34\",\n \"b36\",\n \"b10\",\n \"b12\",\n \"b14\",\n \"b13\"\n ],\n \"table_ref\": [\n \"tab_4\",\n \"tab_6\"\n ],\n \"text\": \"We conduct experiments to validate Claim B. We consider both real-world data and synthetic data, and use different methods to represent it in this experiment.\\nExperiment setting. We test three light field data3 , named Bunny, Knights, and Truck, which are fifth-order tensors of size 40 × 60 × 3 × 9 × 9 (spatial height × spatial width × color channel × vertical grid × horizontal grid). We employ six representative methods as the compared baselines, including two methods with pre-defined topology: TRALS [35] and FCTNALS [37], and four TN-SS methods: TNGreedy [11], TNGA [13], TNLS [15], and TNALE [14]. We represent the test light field data by different methods and calculate the corresponding compression ratio (CR) to achieve a certain reconstruction error (RE) bound. The CR is defined as F G /F X × 100%, where F G is the number of elements of TN cores used to represent a tensor and F X is the number of total elements of the original tensor. The RE is defined as ∥X -X ∥ F /∥X ∥ F , where X is the original data and X is the reconstructed data. In all tests, we select the parameter γ from the interval [10 -7 , 10 -3 ].\\nResult analysis. Table 3 reports the CR and run time of different methods on fifth-order light field data. The results show that the proposed SVDinsTN achieves significantly lower CR than TRALS and FCTNALS, which are methods with pre-defined topology. This indicates that SVDinsTN can obtain a more compact structure than the one pre-defined. Besides, while SVDinsTN requires the determination of diagonal matrices alongside TN cores, its iterative process generates progressively simpler structures, enhancing the computational efficiency. Consequently, SVDinsTN demonstrates faster performance compared to TRALS and FCTNALS. Furthermore, compared to the TN-SS methods, the proposed SVDinsTN achieves a substantial speed improvement while maintaining a comparable level of CR. Remarkably, SVDinsTN achieves an acceleration of approximately 100 ∼ 1000 times over TNGA, TNLS, and TNALE. This is because TNGreedy, TNGA, TNLS, and TNALE adopt the \\\"sampling-evaluation\\\" framework, necessitating a significant number of repeated structure evaluations. However, SVDinsTN introduces a regularized modeling framework, requiring only a single evaluation.\\nImpact of the initialization scheme. We analyze the impact of the shrink operation in the initialization scheme under this experimental setting. In Figure 2, we present the run time comparison of the initial five iterations between SVDinsTN with the shrink operation and SVDinsTN without the shrink operation in the initialization scheme. As observed, the designed initialization scheme greatly reduces the computational costs in the first several iterations. These findings validate the analysis presented in Section 3.3.\\nHigher-order cases. We analyze whether the proposed SVDinsTN still performs well on higher-order tensors. We randomly generate 6th-, 8th-, and 10th-order tensors by using the same procedure in Section ??.\\nThe size of each tensor mode is randomly selected from {5, 6, 7, 8}, the edge number of each TN is randomly selected from {6, 8, 10}, and the rank of each edge is randomly selected from {2, 3}. For each tensor order, we randomly generate 5 tensors. We compare SVDinsTN and baseline methods in terms of CR and run time when reaching the RE bound of 0.01, and show the results in Table 4. As observed, SVDinsTN is applicable to higher orders beyond 5, and even up to 10. The behind rational is the truncated SVD used in initialization restricts the initial values of the rank for each edge to a relatively small range, thus improving computational and storage efficiency (see Figure 2). As the iterations progress, the sparsity regularization in the model leads to progressively simpler learned structures, further boosting efficiency. \"\n },\n {\n \"figure_ref\": [\n \"fig_4\",\n \"fig_5\"\n ],\n \"heading\": \"Experiments for Validating Claim C\",\n \"publication_ref\": [\n \"b33\",\n \"b25\",\n \"b2\",\n \"b29\",\n \"b24\"\n ],\n \"table_ref\": [\n \"tab_7\"\n ],\n \"text\": \"We conduct experiments to validate Claim C. We employ the proposed SVDinsTN to a fundamental application, i.e., tensor completion (TC), and compare it with the state-of-the-art tensor decompositionbased TC methods. Given an incomplete observation tensor\\nF ∈ R I1×I2ו••×I N of X ∈ R I1×I2ו••×I N ,\\nthe proposed TC method first updates G and S by Algorithm 1, and then updates the target tensor X as follows: X = P Ω c ((STN(G, S) + ρ X )/(1 + ρ)) + P Ω (F), where Ω is the index set of the known elements, P Ω (X ) is a projection operator which projects the elements in Ω to themselves and all others to zeros, X\\nis the result at the previous iteration, and the initial X is F.\\nExperiment setting. We test four color videos 4 , named Bunny, News, Salesman, and Silent, which are fourth-order tensors of size 144 × 176 × 3 × 50 (spatial height×spatial width×color channel×frame). We employ five baseline methods for comparison, named FBCP [34], TMac [26], TMacTT [3], TRLRF [30], and TW [25], respectively. All of them are based on different tensor decompositions. We set the missing ratio (MR) to 90%, which is defined as the ratio of the number of missing elements to the total number of elements. We evaluate the reconstructed quality by the mean peak signal-to-noise ratio (MPSNR) computed across all frames. In all tests, the parameter γ is set to 0.0003.\\nResult analysis. Table 5 reports MPSNR and run time obtained by different TC methods. As observed, the proposed SVDinsTN consistently achieves the highest MPSNR values among all utilized TC methods across all test color videos. Besides, SVDinsTN exhibits comparable run time to the baseline methods.\\nFigure 3 displays the reconstructed images and their corresponding residual images at the 25th frame of News. We observe that the proposed SVDinsTN outperforms the baseline methods in terms of visual quality, particularly with respect to background cleanliness and local details (e.g., \\\"dancer\\\") recovery.\\nNumerical convergence. In Theorem 1, we provide a theoretical convergence guarantee for the proposed method. Here, we numerically verify the convergence. Figure 4 presents the relative change in the reconstructed color videos at each iteration compared to their respective previous iterations. We observe that the values of the relative change achieved by the proposed method decrease and gradually tend to zero as the number of iterations increases. This justifies the numerical convergence of the proposed method. Here the relative change is defined as ∥X c -X p ∥ F /∥X p ∥ F , and X c and X p are the results of the current iteration and its previous iteration.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Summary. We propose a novel TN paradigm, called SVDinsTN, which enables us to solve the challenging TN-SS problem from a regularized modeling perspective. This perspective renders our model highly amenable to easy solutions, allowing us to leverage well-established optimization algorithms for tackling the regularized model. As a result, the proposed method achieves approximately 100 ∼ 1000 times acceleration compared to the state-of-the-art TN-SS methods with a comparable representation ability. Additionally, SVDinsTN demonstrates its effectiveness as a valuable tool in practical applications.\\nLimitations. In existing research on the TN-SS problem, two challenging issues remain open. One is the computationally consuming issue, and the other is the theoretical guarantee of the optimal TN structure. SVDinsTN provides a new regularized modeling perspective to fundamentally address the computationally consuming issue and proves theorems regarding the upper bound of the TN rank and the convergence. However, the theoretical guarantee of the optimal TN structure is still an open problem. Solving this issue will be the direction of our future work.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"We would like to express our gratitude to Guillaume Rabusseau for his valuable assistance in correcting the experimental results of \\\"TN-Greedy\\\".\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":""},"doi":{"kind":"string","value":""},"references":{"kind":"list like","value":[{"authors":"A Anandkumar; R Ge; D Hsu; S M Kakade; M Telgarsky","journal":"The Journal of Machine Learning Research","ref_id":"b0","title":"Tensor decompositions for learning latent variable models","year":"2014"},{"authors":"H Attouch; J Bolte; P Redont; A Soubeyran","journal":"Mathematics of Operations Research","ref_id":"b1","title":"Proximal alternating minimization and projection methods for nonconvex problems: An approach based on the Kurdyka-Łojasiewicz inequality","year":"2010"},{"authors":"J A Bengua; H N Phien; H D Tuan; M N Do","journal":"IEEE Transactions on Image Processing","ref_id":"b2","title":"Efficient tensor completion for color image and video recovery: Low-rank tensor train","year":"2017"},{"authors":"A Cichocki; N Lee; I Oseledets; A.-H Phan; Q Zhao; D P Mandic","journal":"Foundations and Trends® in Machine Learning","ref_id":"b3","title":"Tensor networks for dimensionality reduction and large-scale optimization: Part 1 low-rank tensor decompositions","year":"2016"},{"authors":"A Cichocki; D Mandic; L De Lathauwer; G Zhou; Q Zhao; C Caiafa; H A Phan","journal":"IEEE Signal Processing Magazine","ref_id":"b4","title":"Tensor decompositions for signal processing applications: From two-way to multiway component analysis","year":"2015"},{"authors":"D Gabay; B Mercier","journal":"Computers and Mathematics with Applications","ref_id":"b5","title":"A dual algorithm for the solution of nonlinear variational problems via finite element approximation","year":"1976"},{"authors":"S Garnerone; T R De Oliveira; P Zanardi","journal":"Physical Review A","ref_id":"b6","title":"Typicality in random matrix product states","year":"2010"},{"authors":"M Ghadiri; M Fahrbach; G Fu; V Mirrokni","journal":"","ref_id":"b7","title":"Approximately optimal core shapes for tensor decompositions","year":"2023"},{"authors":"I Glasser; R Sweke; N Pancotti; J Eisert; J I Cirac","journal":"","ref_id":"b8","title":"Expressive power of tensor-network factorizations for probabilistic modeling","year":"2019"},{"authors":"G H Golub; P C Hansen; D P O'leary","journal":"SIAM journal on matrix analysis and applications","ref_id":"b9","title":"Tikhonov regularization and total least squares","year":"1999"},{"authors":"M Hashemizadeh; M Liu; J Miller; G Rabusseau","journal":"","ref_id":"b10","title":"Adaptive learning of tensor network structures","year":"2020"},{"authors":"C J Hillar; L.-H Lim","journal":"Journal of the ACM","ref_id":"b11","title":"Most tensor problems are NP-hard","year":"2013"},{"authors":"C Li; Z Sun","journal":"","ref_id":"b12","title":"Evolutionary topology search for tensor network decomposition","year":"2020"},{"authors":"C Li; J Zeng; C Li; C Caiafa; Q Zhao","journal":"","ref_id":"b13","title":"Alternating local enumeration (tnale): Solving tensor network structure search with fewer evaluations","year":"2023"},{"authors":"C Li; J Zeng; Z Tao; Q Zhao","journal":"","ref_id":"b14","title":"Permutation search of tensor network structures via local sampling","year":"2022"},{"authors":"C Li; Q Zhao","journal":"","ref_id":"b15","title":"Is rank minimization of the essence to learn tensor network structure?","year":"2021"},{"authors":"N Li; Y Pan; Y Chen; Z Ding; D Zhao; Z Xu","journal":"Complex & Intelligent Systems","ref_id":"b16","title":"Heuristic rank selection with progressively searching tensor ring network","year":"2022"},{"authors":"Y Liu; Y Lu; W Ou; Z Long; C Zhu","journal":"","ref_id":"b17","title":"Adaptively topological tensor network for multi-view subspace clustering","year":"2023"},{"authors":"Y Luo; X.-L Zhao; D Meng; T.-X Jiang","journal":"","ref_id":"b18","title":"Hlrtf: Hierarchical low-rank tensor factorization for inverse problems in multi-dimensional imaging","year":"2022"},{"authors":"C Nie; H Wang; L Tian","journal":"","ref_id":"b19","title":"Adaptive tensor networks decomposition","year":"2021"},{"authors":"R Orús","journal":"Annals of Physics","ref_id":"b20","title":"A practical introduction to tensor networks: Matrix product states and projected entangled pair states","year":"2014"},{"authors":"I V Oseledets","journal":"SIAM Journal on Scientific Computing","ref_id":"b21","title":"Tensor-train decomposition","year":"2011"},{"authors":"P Rai; Y Wang; S Guo; G Chen; D Dunson; L Carin","journal":"","ref_id":"b22","title":"Scalable bayesian low-rank decomposition of incomplete multiway tensors","year":"2014"},{"authors":"F Sedighin; A Cichocki; A.-H Phan","journal":"IEEE Journal of Selected Topics in Signal Processing","ref_id":"b23","title":"Adaptive rank selection for tensor ring decomposition","year":"2021"},{"authors":"Z.-C Wu; T.-Z Huang; L.-J Deng; H.-X Dou; D Meng","journal":"","ref_id":"b24","title":"Tensor wheel decomposition and its tensor completion application","year":"2022"},{"authors":"Y Xu; R Hao; W Yin; Z Su","journal":"Inverse Problems and Imaging","ref_id":"b25","title":"Parallel matrix factorization for low-rank tensor completion","year":"2015"},{"authors":"R Yamamoto; H Hontani; A Imakura; T Yokota","journal":"","ref_id":"b26","title":"Fast algorithm for low-rank tensor completion in delay-embedded space","year":"2022"},{"authors":"K Ye; L.-H Lim","journal":"","ref_id":"b27","title":"Tensor network ranks","year":"2018"},{"authors":"L Yuan; J Cao; X Zhao; Q Wu; Q Zhao","journal":"","ref_id":"b28","title":"Higher-dimension tensor completion via low-rank tensor ring decomposition","year":"2018"},{"authors":"L Yuan; C Li; D Mandic; J Cao; Q Zhao","journal":"","ref_id":"b29","title":"Tensor ring decomposition with rank minimization on latent space: An efficient approach for tensor completion","year":"2019"},{"authors":"L Yuan; Q Zhao; L Gui; J Cao","journal":"Signal Processing: Image Communication","ref_id":"b30","title":"High-order tensor completion via gradient-based optimization under tensor train format","year":"2019"},{"authors":"S Zhang; L Wang; L Zhang; H Huang","journal":"","ref_id":"b31","title":"Learning tensor low-rank prior for hyperspectral image reconstruction","year":"2021"},{"authors":"X Zhang; X Yuan; L Carin","journal":"","ref_id":"b32","title":"Nonlocal low-rank tensor factor analysis for image restoration","year":"2018"},{"authors":"Q Zhao; L Zhang; A Cichocki","journal":"IEEE Transactions on Pattern Analysis and Machine Intelligence","ref_id":"b33","title":"Bayesian CP factorization of incomplete tensors with automatic rank determination","year":"2015"},{"authors":"Q Zhao; G Zhou; S Xie; L Zhang; A Cichocki","journal":"","ref_id":"b34","title":"Tensor ring decomposition","year":"2016"},{"authors":"Y.-B Zheng; T.-Z Huang; X.-L Zhao; T.-X Jiang; T.-Y Ji; T.-H Ma","journal":"Information Sciences","ref_id":"b35","title":"Tensor N-tubal rank and its convex relaxation for low-rank tensor recovery","year":"2020"},{"authors":"Y.-B Zheng; T.-Z Huang; X.-L Zhao; Q Zhao; T.-X Jiang","journal":"","ref_id":"b36","title":"Fully-connected tensor network decomposition and its application to higher-order tensor completion","year":"2021"}],"string":"[\n {\n \"authors\": \"A Anandkumar; R Ge; D Hsu; S M Kakade; M Telgarsky\",\n \"journal\": \"The Journal of Machine Learning Research\",\n \"ref_id\": \"b0\",\n \"title\": \"Tensor decompositions for learning latent variable models\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"H Attouch; J Bolte; P Redont; A Soubeyran\",\n \"journal\": \"Mathematics of Operations Research\",\n \"ref_id\": \"b1\",\n \"title\": \"Proximal alternating minimization and projection methods for nonconvex problems: An approach based on the Kurdyka-Łojasiewicz inequality\",\n \"year\": \"2010\"\n },\n {\n \"authors\": \"J A Bengua; H N Phien; H D Tuan; M N Do\",\n \"journal\": \"IEEE Transactions on Image Processing\",\n \"ref_id\": \"b2\",\n \"title\": \"Efficient tensor completion for color image and video recovery: Low-rank tensor train\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"A Cichocki; N Lee; I Oseledets; A.-H Phan; Q Zhao; D P Mandic\",\n \"journal\": \"Foundations and Trends® in Machine Learning\",\n \"ref_id\": \"b3\",\n \"title\": \"Tensor networks for dimensionality reduction and large-scale optimization: Part 1 low-rank tensor decompositions\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"A Cichocki; D Mandic; L De Lathauwer; G Zhou; Q Zhao; C Caiafa; H A Phan\",\n \"journal\": \"IEEE Signal Processing Magazine\",\n \"ref_id\": \"b4\",\n \"title\": \"Tensor decompositions for signal processing applications: From two-way to multiway component analysis\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"D Gabay; B Mercier\",\n \"journal\": \"Computers and Mathematics with Applications\",\n \"ref_id\": \"b5\",\n \"title\": \"A dual algorithm for the solution of nonlinear variational problems via finite element approximation\",\n \"year\": \"1976\"\n },\n {\n \"authors\": \"S Garnerone; T R De Oliveira; P Zanardi\",\n \"journal\": \"Physical Review A\",\n \"ref_id\": \"b6\",\n \"title\": \"Typicality in random matrix product states\",\n \"year\": \"2010\"\n },\n {\n \"authors\": \"M Ghadiri; M Fahrbach; G Fu; V Mirrokni\",\n \"journal\": \"\",\n \"ref_id\": \"b7\",\n \"title\": \"Approximately optimal core shapes for tensor decompositions\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"I Glasser; R Sweke; N Pancotti; J Eisert; J I Cirac\",\n \"journal\": \"\",\n \"ref_id\": \"b8\",\n \"title\": \"Expressive power of tensor-network factorizations for probabilistic modeling\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"G H Golub; P C Hansen; D P O'leary\",\n \"journal\": \"SIAM journal on matrix analysis and applications\",\n \"ref_id\": \"b9\",\n \"title\": \"Tikhonov regularization and total least squares\",\n \"year\": \"1999\"\n },\n {\n \"authors\": \"M Hashemizadeh; M Liu; J Miller; G Rabusseau\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Adaptive learning of tensor network structures\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"C J Hillar; L.-H Lim\",\n \"journal\": \"Journal of the ACM\",\n \"ref_id\": \"b11\",\n \"title\": \"Most tensor problems are NP-hard\",\n \"year\": \"2013\"\n },\n {\n \"authors\": \"C Li; Z Sun\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"Evolutionary topology search for tensor network decomposition\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"C Li; J Zeng; C Li; C Caiafa; Q Zhao\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"Alternating local enumeration (tnale): Solving tensor network structure search with fewer evaluations\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"C Li; J Zeng; Z Tao; Q Zhao\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Permutation search of tensor network structures via local sampling\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"C Li; Q Zhao\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Is rank minimization of the essence to learn tensor network structure?\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"N Li; Y Pan; Y Chen; Z Ding; D Zhao; Z Xu\",\n \"journal\": \"Complex & Intelligent Systems\",\n \"ref_id\": \"b16\",\n \"title\": \"Heuristic rank selection with progressively searching tensor ring network\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Y Liu; Y Lu; W Ou; Z Long; C Zhu\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Adaptively topological tensor network for multi-view subspace clustering\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Y Luo; X.-L Zhao; D Meng; T.-X Jiang\",\n \"journal\": \"\",\n \"ref_id\": \"b18\",\n \"title\": \"Hlrtf: Hierarchical low-rank tensor factorization for inverse problems in multi-dimensional imaging\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"C Nie; H Wang; L Tian\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Adaptive tensor networks decomposition\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"R Orús\",\n \"journal\": \"Annals of Physics\",\n \"ref_id\": \"b20\",\n \"title\": \"A practical introduction to tensor networks: Matrix product states and projected entangled pair states\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"I V Oseledets\",\n \"journal\": \"SIAM Journal on Scientific Computing\",\n \"ref_id\": \"b21\",\n \"title\": \"Tensor-train decomposition\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"P Rai; Y Wang; S Guo; G Chen; D Dunson; L Carin\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"Scalable bayesian low-rank decomposition of incomplete multiway tensors\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"F Sedighin; A Cichocki; A.-H Phan\",\n \"journal\": \"IEEE Journal of Selected Topics in Signal Processing\",\n \"ref_id\": \"b23\",\n \"title\": \"Adaptive rank selection for tensor ring decomposition\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Z.-C Wu; T.-Z Huang; L.-J Deng; H.-X Dou; D Meng\",\n \"journal\": \"\",\n \"ref_id\": \"b24\",\n \"title\": \"Tensor wheel decomposition and its tensor completion application\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Y Xu; R Hao; W Yin; Z Su\",\n \"journal\": \"Inverse Problems and Imaging\",\n \"ref_id\": \"b25\",\n \"title\": \"Parallel matrix factorization for low-rank tensor completion\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"R Yamamoto; H Hontani; A Imakura; T Yokota\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Fast algorithm for low-rank tensor completion in delay-embedded space\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"K Ye; L.-H Lim\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Tensor network ranks\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"L Yuan; J Cao; X Zhao; Q Wu; Q Zhao\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Higher-dimension tensor completion via low-rank tensor ring decomposition\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"L Yuan; C Li; D Mandic; J Cao; Q Zhao\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Tensor ring decomposition with rank minimization on latent space: An efficient approach for tensor completion\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"L Yuan; Q Zhao; L Gui; J Cao\",\n \"journal\": \"Signal Processing: Image Communication\",\n \"ref_id\": \"b30\",\n \"title\": \"High-order tensor completion via gradient-based optimization under tensor train format\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"S Zhang; L Wang; L Zhang; H Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b31\",\n \"title\": \"Learning tensor low-rank prior for hyperspectral image reconstruction\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"X Zhang; X Yuan; L Carin\",\n \"journal\": \"\",\n \"ref_id\": \"b32\",\n \"title\": \"Nonlocal low-rank tensor factor analysis for image restoration\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Q Zhao; L Zhang; A Cichocki\",\n \"journal\": \"IEEE Transactions on Pattern Analysis and Machine Intelligence\",\n \"ref_id\": \"b33\",\n \"title\": \"Bayesian CP factorization of incomplete tensors with automatic rank determination\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"Q Zhao; G Zhou; S Xie; L Zhang; A Cichocki\",\n \"journal\": \"\",\n \"ref_id\": \"b34\",\n \"title\": \"Tensor ring decomposition\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Y.-B Zheng; T.-Z Huang; X.-L Zhao; T.-X Jiang; T.-Y Ji; T.-H Ma\",\n \"journal\": \"Information Sciences\",\n \"ref_id\": \"b35\",\n \"title\": \"Tensor N-tubal rank and its convex relaxation for low-rank tensor recovery\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Y.-B Zheng; T.-Z Huang; X.-L Zhao; Q Zhao; T.-X Jiang\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Fully-connected tensor network decomposition and its application to higher-order tensor completion\",\n \"year\": \"2021\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,187.8,680.53,236.41,20.09],"formula_id":"formula_0","formula_text":"Z(i 1 , i 3 , j 3 ) = I2 i2=1 I4 i4=1 X (i 1 , i 2 , i 3 , i 4 )Y(i 2 , i 4 , j 3 )."},{"formula_coordinates":[4,97.2,219.04,417.6,33.04],"formula_id":"formula_1","formula_text":"X ∈ R I1×I2ו••×I N by N small-sized N th-order cores denoted by G k ∈ R R 1,k ×R 2,k ו••×R k-1,k ×I k ×R k,k+1 ו••×R k,N for k ∈ K N"},{"formula_coordinates":[4,97.2,253.34,419.34,21.98],"formula_id":"formula_2","formula_text":"X = FCTN(G 1 , G 2 , • • • , G N ) and define the FCTN rank as the vector (R 1,2 , R 1,3 , • • • , R 1,N , R 2,3 , • • • , R 2,N , • • • , R N -1,N ) ∈ R N (N -1)/2 ."},{"formula_coordinates":[4,97.2,309.31,417.6,20.56],"formula_id":"formula_3","formula_text":"(R 1,2 , 1, • • • , 1, R 2,3 , 1, • • • , 1, R N -2,N -1 , R N -1,N ) can be converted into a \"chain\" topology with rank (R 1,2 , R 2,3 , • • • , R N -1,N"},{"formula_coordinates":[4,97.2,526.89,418.27,146.73],"formula_id":"formula_4","formula_text":"Definition 1 (SVDinsTN) Let X ∈ R I1×I2ו••×I N be an N th-order tensor such that X (i 1 , i 2 , • • • , i N ) = R1,2 r1,2=1 R1,3 r1,3=1 • • • R 1,N r 1,N =1 R2,3 r2,3=1 • • • R 2,N r 2,N =1 • • • R N -1,N r N -1,N =1 S 1,2 (r 1,2 , r 1,2 )S 1,3 (r 1,3 , r 1,3 ) • • • S 1,N (r 1,N , r 1,N ) S 2,3 (r 2,3 , r 2,3 ) • • • S 2,N (r 2,N , r 2,N ) • • • S N -1,N (r N -1,N , r N -1,N ) G 1 (i 1 , r 1,2 , r 1,3 , • • • , r 1,N ) G 2 (r 1,2 , i 2 , r 2,3 , • • • , r 2,N ) • • • G k (r 1,k , r 2,k , • • • , r k-1,k , i k , r k,k+1 , • • • , r k,N ) • • • G N (r 1,N , r 2,N , • • • , r N -1,N , i N ),(1)"},{"formula_coordinates":[4,124.19,678.89,195.83,11.23],"formula_id":"formula_5","formula_text":"G k ∈ R R 1,k ×R 2,k ו••×R k-1,k ×I k ×R k,k+1 ו••×R k,N"},{"formula_coordinates":[5,148.74,328.75,366.73,22.31],"formula_id":"formula_6","formula_text":"min G,S 1 2 ∥X -STN(G, S)∥ 2 F + µ 2 k∈K N ∥G k ∥ 2 F + (t,l)∈TL N λ t,l ∥S t,l ∥ 1 ,(2)"},{"formula_coordinates":[5,97.2,366.79,123.76,9.68],"formula_id":"formula_7","formula_text":"S denotes {S t,l |(t, l) ∈ TL N }."},{"formula_coordinates":[5,124.88,395.93,390.59,51.51],"formula_id":"formula_8","formula_text":" G k = argmin G k 1 2 ∥X -STN(G, S)∥ 2 F + µ 2 ∥G k ∥ 2 F + ρ 2 ∥G k -Ĝk ∥ 2 F , ∀k ∈ K N , S t,l = argmin S t,l 1 2 ∥X -STN(G, S)∥ 2 F + λ t,l ∥S t,l ∥ 1 + ρ 2 ∥S t,l -Ŝt,l ∥ 2 F , ∀(t, l) ∈ TL N ,(3)"},{"formula_coordinates":[5,163.21,525.95,352.25,22.51],"formula_id":"formula_9","formula_text":"min G k (k) 1 2 ∥X (k) -G k(k) M k ∥ 2 F + µ 2 ∥G k(k) ∥ 2 F + ρ 2 ∥G k(k) -Ĝk(k) ∥ 2 F .(4)"},{"formula_coordinates":[5,192.43,566.67,323.04,14.58],"formula_id":"formula_10","formula_text":"G k(k) = X (k) M T k + ρ Ĝk(k) M k M T k + (µ + ρ)I -1 .(5)"},{"formula_coordinates":[5,189.72,636.92,325.74,22.31],"formula_id":"formula_11","formula_text":"min s t,l 1 2 ∥x -H t,l s t,l ∥ 2 F + λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F ,(6)"},{"formula_coordinates":[5,195.33,687.54,320.14,37.01],"formula_id":"formula_12","formula_text":"min s t,l 1 2 ∥x -H t,l q t,l ∥ 2 F + λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F s.t. s t,l -q t,l = 0,(7)"},{"formula_coordinates":[6,112.44,101.34,403.1,19.4],"formula_id":"formula_13","formula_text":"β t,l = 1 for ∀(t, l) ∈ TLN ; let G k = 1/ √ I k ones(R 1,k , R 2,k , • • • , R k-1,k , I k , R k,k+1 , • • • , R k,N"},{"formula_coordinates":[6,165.21,288.15,350.26,47.7],"formula_id":"formula_14","formula_text":"L β t,l (s t,l , q t,l , p t,l ) = 1 2 ∥x -H t,l q t,l ∥ 2 F + λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F + β t,l 2 s t,l -q t,l + p t,l β t,l 2 F ,(8)"},{"formula_coordinates":[6,153.5,365.99,257.95,64.02],"formula_id":"formula_15","formula_text":" q t,l = argmin q t,l 1 2 ∥x -H t,l q t,l ∥ 2 F + β t,l 2 s t,l -q t,l + p t,l β t,l2"},{"formula_coordinates":[6,162.36,372.94,289.53,45.32],"formula_id":"formula_16","formula_text":"F , s t,l = argmin s t,l λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F + β t,l 2 s t,l -q t,l + p t,l β t,l2"},{"formula_coordinates":[6,447.91,410.53,5.05,6.12],"formula_id":"formula_17","formula_text":"F"},{"formula_coordinates":[6,194.55,447.85,320.91,57.17],"formula_id":"formula_19","formula_text":" q t,l = H T t,l H t,l + β t,l I -1 H T t,l x + β t,l s t,l + p t,l , s t,l = shrink ρŝ t,l + β t,l q t,l -p t,l ρ + β t,l , λ t,l ρ + β t,l , p t,l = p t,l + β t,l (s t,l -q t,l ),(10)"},{"formula_coordinates":[6,97.2,552.87,419.26,63.83],"formula_id":"formula_20","formula_text":"I × I × • • • × I and the initial rank be (R, R, • • • , R) satisfied R ≤ I. The computational cost involves updat- ing G and S, resulting in costs of O N N k=2 I k R k(N -k)+k-1 + N I N -1 R 2(N -1) + N 3 IR N and O N 2 N k=2 I k R k(N -k)+k +N 4 IR N +N 2 RI 2N , respectively. In summary, the computational cost at each iteration is O N 2 N k=2 I k R k(N -k)+k +N 4 IR N +N 2 RI 2N ."},{"formula_coordinates":[7,97.2,529.37,417.77,22.13],"formula_id":"formula_21","formula_text":"G k ∈ R R 1,k ×R 2,k ו••×R k-1,k ×I k ×R k,k+1 ו••×R k,N for ∀k ∈ K N and S t,l ∈ R R t,l ×R t,l"},{"formula_coordinates":[8,97.2,513.29,417.6,20.56],"formula_id":"formula_22","formula_text":"G 1 → G 2 → G 3 → G 4 → G 5 → G 1 and G 1 → G 3 → G 5 → G 2 → G 4 → G 1 , respectively."},{"formula_coordinates":[10,342.27,416.07,173.77,10.87],"formula_id":"formula_23","formula_text":"F ∈ R I1×I2ו••×I N of X ∈ R I1×I2ו••×I N ,"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 187.8,\n 680.53,\n 236.41,\n 20.09\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"Z(i 1 , i 3 , j 3 ) = I2 i2=1 I4 i4=1 X (i 1 , i 2 , i 3 , i 4 )Y(i 2 , i 4 , j 3 ).\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 97.2,\n 219.04,\n 417.6,\n 33.04\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"X ∈ R I1×I2ו••×I N by N small-sized N th-order cores denoted by G k ∈ R R 1,k ×R 2,k ו••×R k-1,k ×I k ×R k,k+1 ו••×R k,N for k ∈ K N\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 97.2,\n 253.34,\n 419.34,\n 21.98\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"X = FCTN(G 1 , G 2 , • • • , G N ) and define the FCTN rank as the vector (R 1,2 , R 1,3 , • • • , R 1,N , R 2,3 , • • • , R 2,N , • • • , R N -1,N ) ∈ R N (N -1)/2 .\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 97.2,\n 309.31,\n 417.6,\n 20.56\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"(R 1,2 , 1, • • • , 1, R 2,3 , 1, • • • , 1, R N -2,N -1 , R N -1,N ) can be converted into a \\\"chain\\\" topology with rank (R 1,2 , R 2,3 , • • • , R N -1,N\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 97.2,\n 526.89,\n 418.27,\n 146.73\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"Definition 1 (SVDinsTN) Let X ∈ R I1×I2ו••×I N be an N th-order tensor such that X (i 1 , i 2 , • • • , i N ) = R1,2 r1,2=1 R1,3 r1,3=1 • • • R 1,N r 1,N =1 R2,3 r2,3=1 • • • R 2,N r 2,N =1 • • • R N -1,N r N -1,N =1 S 1,2 (r 1,2 , r 1,2 )S 1,3 (r 1,3 , r 1,3 ) • • • S 1,N (r 1,N , r 1,N ) S 2,3 (r 2,3 , r 2,3 ) • • • S 2,N (r 2,N , r 2,N ) • • • S N -1,N (r N -1,N , r N -1,N ) G 1 (i 1 , r 1,2 , r 1,3 , • • • , r 1,N ) G 2 (r 1,2 , i 2 , r 2,3 , • • • , r 2,N ) • • • G k (r 1,k , r 2,k , • • • , r k-1,k , i k , r k,k+1 , • • • , r k,N ) • • • G N (r 1,N , r 2,N , • • • , r N -1,N , i N ),(1)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 124.19,\n 678.89,\n 195.83,\n 11.23\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"G k ∈ R R 1,k ×R 2,k ו••×R k-1,k ×I k ×R k,k+1 ו••×R k,N\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 148.74,\n 328.75,\n 366.73,\n 22.31\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"min G,S 1 2 ∥X -STN(G, S)∥ 2 F + µ 2 k∈K N ∥G k ∥ 2 F + (t,l)∈TL N λ t,l ∥S t,l ∥ 1 ,(2)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 97.2,\n 366.79,\n 123.76,\n 9.68\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"S denotes {S t,l |(t, l) ∈ TL N }.\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 124.88,\n 395.93,\n 390.59,\n 51.51\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \" G k = argmin G k 1 2 ∥X -STN(G, S)∥ 2 F + µ 2 ∥G k ∥ 2 F + ρ 2 ∥G k -Ĝk ∥ 2 F , ∀k ∈ K N , S t,l = argmin S t,l 1 2 ∥X -STN(G, S)∥ 2 F + λ t,l ∥S t,l ∥ 1 + ρ 2 ∥S t,l -Ŝt,l ∥ 2 F , ∀(t, l) ∈ TL N ,(3)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 163.21,\n 525.95,\n 352.25,\n 22.51\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"min G k (k) 1 2 ∥X (k) -G k(k) M k ∥ 2 F + µ 2 ∥G k(k) ∥ 2 F + ρ 2 ∥G k(k) -Ĝk(k) ∥ 2 F .(4)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 192.43,\n 566.67,\n 323.04,\n 14.58\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"G k(k) = X (k) M T k + ρ Ĝk(k) M k M T k + (µ + ρ)I -1 .(5)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 189.72,\n 636.92,\n 325.74,\n 22.31\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"min s t,l 1 2 ∥x -H t,l s t,l ∥ 2 F + λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F ,(6)\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 195.33,\n 687.54,\n 320.14,\n 37.01\n ],\n \"formula_id\": \"formula_12\",\n \"formula_text\": \"min s t,l 1 2 ∥x -H t,l q t,l ∥ 2 F + λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F s.t. s t,l -q t,l = 0,(7)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 112.44,\n 101.34,\n 403.1,\n 19.4\n ],\n \"formula_id\": \"formula_13\",\n \"formula_text\": \"β t,l = 1 for ∀(t, l) ∈ TLN ; let G k = 1/ √ I k ones(R 1,k , R 2,k , • • • , R k-1,k , I k , R k,k+1 , • • • , R k,N\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 165.21,\n 288.15,\n 350.26,\n 47.7\n ],\n \"formula_id\": \"formula_14\",\n \"formula_text\": \"L β t,l (s t,l , q t,l , p t,l ) = 1 2 ∥x -H t,l q t,l ∥ 2 F + λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F + β t,l 2 s t,l -q t,l + p t,l β t,l 2 F ,(8)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 153.5,\n 365.99,\n 257.95,\n 64.02\n ],\n \"formula_id\": \"formula_15\",\n \"formula_text\": \" q t,l = argmin q t,l 1 2 ∥x -H t,l q t,l ∥ 2 F + β t,l 2 s t,l -q t,l + p t,l β t,l2\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 162.36,\n 372.94,\n 289.53,\n 45.32\n ],\n \"formula_id\": \"formula_16\",\n \"formula_text\": \"F , s t,l = argmin s t,l λ t,l ∥s t,l ∥ 1 + ρ 2 ∥s t,l -ŝt,l ∥ 2 F + β t,l 2 s t,l -q t,l + p t,l β t,l2\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 447.91,\n 410.53,\n 5.05,\n 6.12\n ],\n \"formula_id\": \"formula_17\",\n \"formula_text\": \"F\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 194.55,\n 447.85,\n 320.91,\n 57.17\n ],\n \"formula_id\": \"formula_19\",\n \"formula_text\": \" q t,l = H T t,l H t,l + β t,l I -1 H T t,l x + β t,l s t,l + p t,l , s t,l = shrink ρŝ t,l + β t,l q t,l -p t,l ρ + β t,l , λ t,l ρ + β t,l , p t,l = p t,l + β t,l (s t,l -q t,l ),(10)\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 97.2,\n 552.87,\n 419.26,\n 63.83\n ],\n \"formula_id\": \"formula_20\",\n \"formula_text\": \"I × I × • • • × I and the initial rank be (R, R, • • • , R) satisfied R ≤ I. The computational cost involves updat- ing G and S, resulting in costs of O N N k=2 I k R k(N -k)+k-1 + N I N -1 R 2(N -1) + N 3 IR N and O N 2 N k=2 I k R k(N -k)+k +N 4 IR N +N 2 RI 2N , respectively. In summary, the computational cost at each iteration is O N 2 N k=2 I k R k(N -k)+k +N 4 IR N +N 2 RI 2N .\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 97.2,\n 529.37,\n 417.77,\n 22.13\n ],\n \"formula_id\": \"formula_21\",\n \"formula_text\": \"G k ∈ R R 1,k ×R 2,k ו••×R k-1,k ×I k ×R k,k+1 ו••×R k,N for ∀k ∈ K N and S t,l ∈ R R t,l ×R t,l\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 97.2,\n 513.29,\n 417.6,\n 20.56\n ],\n \"formula_id\": \"formula_22\",\n \"formula_text\": \"G 1 → G 2 → G 3 → G 4 → G 5 → G 1 and G 1 → G 3 → G 5 → G 2 → G 4 → G 1 , respectively.\"\n },\n {\n \"formula_coordinates\": [\n 10,\n 342.27,\n 416.07,\n 173.77,\n 10.87\n ],\n \"formula_id\": \"formula_23\",\n \"formula_text\": \"F ∈ R I1×I2ו••×I N of X ∈ R I1×I2ו••×I N ,\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"SVDinsTN: A Tensor Network Paradigm for Efficient Structure Search from Regularized Modeling Perspective"},"abstract":{"kind":"string","value":"Tensor network (TN) representation is a powerful technique for computer vision and machine learning. TN structure search (TN-SS) aims to search for a customized structure to achieve a compact representation, which is a challenging NP-hard problem. Recent \"sampling-evaluation\"-based methods require sampling an extensive collection of structures and evaluating them one by one, resulting in prohibitively high computational costs. To address this issue, we propose a novel TN paradigm, named SVD-inspired TN decomposition (SVDinsTN), which allows us to efficiently solve the TN-SS problem from a regularized modeling perspective, eliminating the repeated structure evaluations. To be specific, by inserting a diagonal factor for each edge of the fully-connected TN, SVDinsTN allows us to calculate TN cores and diagonal factors simultaneously, with the factor sparsity revealing a compact TN structure. In theory, we prove a convergence guarantee for the proposed method. Experimental results demonstrate that the proposed method achieves approximately 100 ∼ 1000 times acceleration compared to the state-of-the-art TN-SS methods while maintaining a comparable representation ability."},"authors":{"kind":"string","value":"Yu-Bang Zheng; Xi-Le Zhao; Junhua Zeng; Chao Li; Qibin Zhao; Heng-Chao Li; Ting-Zhu Huang"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: (a) A graphical illustration of SVD. (b) A graphical illustration of SVD-inspired TN decomposition on a fifth-order tensor. (c) Comparison of the compression ratio (↓) and run time (↓) of different methods on a fifth-order light field image Knights, where the reconstruction error bound is set to 0.05, TRALS[35] and FCTNALS[37] are methods with pre-defined topologies, and TNGreedy[11], TNGA[13], TNLS[15], and TNALE[14] are TN-SS methods (please see more results in Table3).","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 1 (1Figure 1(b) shows a graphical illustration of SVDinsTN on a fifth-order tensor. SVDinsTN includes TN cores and diagonal factors, enables it to leverage the sparsity of diagonal factors to reveal a compact TN structure and utilize TN cores (merged with diagonal factors) to represent a tensor.","figure_data":"","figure_id":"fig_1","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Run time comparison of the first five iterations between SVDinsTN and SVDinsTN without the shrink operation in the initialization scheme.","figure_data":"","figure_id":"fig_3","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 3 :3Figure 3: Reconstructed images and residual images obtained by different methods on the 25th frame of News. Here the residual image is the average absolute difference between the reconstructed image and the ground truth over R, G, and B channels.","figure_data":"","figure_id":"fig_4","figure_label":"3","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: Relative change curves with respect to the iteration number on four test color videos.Here the relative change is defined as ∥X c -X p ∥ F /∥X p ∥ F , and X c and X p are the results of the current iteration and its previous iteration.","figure_data":"","figure_id":"fig_5","figure_label":"4","figure_type":"figure"},{"figure_caption":"","figure_data":"","figure_id":"tab_0","figure_label":"","figure_type":"table"},{"figure_caption":": Several operations and their interpretations returns a column vector formed from the elements on the main diagonal of X when the input variable is a diagonal matrix, and diag(x) returns a diagonal matrix whose main diagonal is formed from the elements of x when the input variable is a column vector. ones ones(I1, I2, • • • , IN ) returns an I1 × I2 × • • • × IN tensor, whose elements are all equal to 1.","figure_data":"Operation Interpretationdiag diag(X) zeros zeros(I1, I2, • • • , IN ) returns an I1 × I2 × • • • × IN tensor, whose elements are all equal to 0.vecvec(X ) returns a column vector by lexicographical reordering of the elements of X .","figure_id":"tab_2","figure_label":"","figure_type":"table"},{"figure_caption":"Performance of SVDinsTN on TN structure revealing under 100 independent tests.","figure_data":"","figure_id":"tab_3","figure_label":"2","figure_type":"table"},{"figure_caption":"Comparison of the CR (↓) and run time (×1000s, ↓) of different methods on light field data.","figure_data":"DataMethodRE bound: 0.01RE bound: 0.05RE bound: 0.1CRTimeCRTimeCRTimeTRALS [35]60.5%13.5417.4%0.4715.31%0.118BunnyFCTNALS [37] TNGreedy [11]65.1% 26.1%13.08 11.0220.9% 6.32%0.473 1.0213.93% 2.34%0.041 0.362TNGA [13]27.9%10145.01%180.32.25%12.52TNLS [15]24.3%14024.26%63.702.16%24.53TNALE [14]26.3%144.54.52%18.362.26%3.064SVDinsTN22.4%0.7456.92%0.0292.66%0.005TRALS [35]74.7%10.3126.9%3.8359.15%0.423KnightsFCTNALS [37] TNGreedy [11]73.5% 32.1%12.35 12.5320.9% 7.55%0.619 1.3663.93% 3.50%0.014 0.481TNGA [13]38.7%912.95.01%140.22.44%12.52TNLS [15]27.3%12864.73%75.512.15%5.320TNALE [14]27.6%266.44.52%25.052.10%3.386SVDinsTN32.0%1.5485.64%0.1042.76%0.019TRALS [35]62.8%17.6222.6%1.7386.00%0.090TruckFCTNALS [37] TNGreedy [11]69.3% 26.9%7.735 6.67620.9% 7.26%2.953 1.2593.93% 3.35%0.159 0.488TNGA [13]27.9%10295.01%170.32.85%14.83TNLS [15]26.4%992.64.99%119.82.57%19.35TNALE [14]24.7%239.35.77%19.542.90%5.160SVDinsTN23.5%1.0516.42%0.1522.83%0.023","figure_id":"tab_4","figure_label":"3","figure_type":"table"},{"figure_caption":"Comparison of the CR (↓) and run time (×1000s, ↓) of different methods when RE bound is 0.01. The result is the average value of 5 independent experiments and \"-\" indicates \"out of memory\".","figure_data":"Method6th-order8th-order10th-orderCRTimeCRTimeCRTimeTRALS [35]1.35%0.0060.064%0.034--FCTNALS [37]2.13%0.002----TNGreedy [11]0.88%0.1670.016%2.6250.0008%45.39TNGA [13]0.94%3.8250.024%51.40--TNLS [15]1.11%0.6730.038%59.83--TNALE [14]1.65%0.2010.047%19.96--SVDinsTN1.13%0.0020.016%0.0170.0007%0.608","figure_id":"tab_6","figure_label":"4","figure_type":"table"},{"figure_caption":"Comparison of MPSNR (↑) and run time (second, ↓) of different TC methods on color videos. MPSNR Time MPSNR Time MPSNR Time MPSNR Time MPSNR Time MPSNR Time Bunny 28.402 1731 28.211 1203 29.523 453.7 29.163 486.7 30.729 1497 32.401 691.3 News 28.234 1720 27.882 340.4 28.714 535.9 28.857 978.1 30.027 1426 31.643 932.4 Salesman 29.077 1783 28.469 353.6 29.534 656.4 28.288 689.3 30.621 1148 31.684 769.5 Silent 30.126 1453 30.599 316.2 30.647 1305 31.081 453.2 31.731 1232 32.706 532.3","figure_data":"VideoFBCP [34]TMac [26]TMacTT [3]TRLRF [30]TW [25]SVDinsTNObservedFBCP [34] TMac [26] TMacTT [3] TRLRF [30]TW [25]SVDinsTN Ground truth0 00. 1 0.10. 2 0.20. 3 0.30. 4 0.40. 5 0.50. 6 0.60. 7 0.70. 8 0.80. 9 0.91 1","figure_id":"tab_7","figure_label":"5","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: (a) A graphical illustration of SVD. (b) A graphical illustration of SVD-inspired TN decomposition on a fifth-order tensor. (c) Comparison of the compression ratio (↓) and run time (↓) of different methods on a fifth-order light field image Knights, where the reconstruction error bound is set to 0.05, TRALS[35] and FCTNALS[37] are methods with pre-defined topologies, and TNGreedy[11], TNGA[13], TNLS[15], and TNALE[14] are TN-SS methods (please see more results in Table3).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 1 (1Figure 1(b) shows a graphical illustration of SVDinsTN on a fifth-order tensor. SVDinsTN includes TN cores and diagonal factors, enables it to leverage the sparsity of diagonal factors to reveal a compact TN structure and utilize TN cores (merged with diagonal factors) to represent a tensor.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Run time comparison of the first five iterations between SVDinsTN and SVDinsTN without the shrink operation in the initialization scheme.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 3 :3Figure 3: Reconstructed images and residual images obtained by different methods on the 25th frame of News. Here the residual image is the average absolute difference between the reconstructed image and the ground truth over R, G, and B channels.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: Relative change curves with respect to the iteration number on four test color videos.Here the relative change is defined as ∥X c -X p ∥ F /∥X p ∥ F , and X c and X p are the results of the current iteration and its previous iteration.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \": Several operations and their interpretations returns a column vector formed from the elements on the main diagonal of X when the input variable is a diagonal matrix, and diag(x) returns a diagonal matrix whose main diagonal is formed from the elements of x when the input variable is a column vector. ones ones(I1, I2, • • • , IN ) returns an I1 × I2 × • • • × IN tensor, whose elements are all equal to 1.\",\n \"figure_data\": \"Operation Interpretationdiag diag(X) zeros zeros(I1, I2, • • • , IN ) returns an I1 × I2 × • • • × IN tensor, whose elements are all equal to 0.vecvec(X ) returns a column vector by lexicographical reordering of the elements of X .\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Performance of SVDinsTN on TN structure revealing under 100 independent tests.\",\n \"figure_data\": \"\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison of the CR (↓) and run time (×1000s, ↓) of different methods on light field data.\",\n \"figure_data\": \"DataMethodRE bound: 0.01RE bound: 0.05RE bound: 0.1CRTimeCRTimeCRTimeTRALS [35]60.5%13.5417.4%0.4715.31%0.118BunnyFCTNALS [37] TNGreedy [11]65.1% 26.1%13.08 11.0220.9% 6.32%0.473 1.0213.93% 2.34%0.041 0.362TNGA [13]27.9%10145.01%180.32.25%12.52TNLS [15]24.3%14024.26%63.702.16%24.53TNALE [14]26.3%144.54.52%18.362.26%3.064SVDinsTN22.4%0.7456.92%0.0292.66%0.005TRALS [35]74.7%10.3126.9%3.8359.15%0.423KnightsFCTNALS [37] TNGreedy [11]73.5% 32.1%12.35 12.5320.9% 7.55%0.619 1.3663.93% 3.50%0.014 0.481TNGA [13]38.7%912.95.01%140.22.44%12.52TNLS [15]27.3%12864.73%75.512.15%5.320TNALE [14]27.6%266.44.52%25.052.10%3.386SVDinsTN32.0%1.5485.64%0.1042.76%0.019TRALS [35]62.8%17.6222.6%1.7386.00%0.090TruckFCTNALS [37] TNGreedy [11]69.3% 26.9%7.735 6.67620.9% 7.26%2.953 1.2593.93% 3.35%0.159 0.488TNGA [13]27.9%10295.01%170.32.85%14.83TNLS [15]26.4%992.64.99%119.82.57%19.35TNALE [14]24.7%239.35.77%19.542.90%5.160SVDinsTN23.5%1.0516.42%0.1522.83%0.023\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison of the CR (↓) and run time (×1000s, ↓) of different methods when RE bound is 0.01. The result is the average value of 5 independent experiments and \\\"-\\\" indicates \\\"out of memory\\\".\",\n \"figure_data\": \"Method6th-order8th-order10th-orderCRTimeCRTimeCRTimeTRALS [35]1.35%0.0060.064%0.034--FCTNALS [37]2.13%0.002----TNGreedy [11]0.88%0.1670.016%2.6250.0008%45.39TNGA [13]0.94%3.8250.024%51.40--TNLS [15]1.11%0.6730.038%59.83--TNALE [14]1.65%0.2010.047%19.96--SVDinsTN1.13%0.0020.016%0.0170.0007%0.608\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison of MPSNR (↑) and run time (second, ↓) of different TC methods on color videos. MPSNR Time MPSNR Time MPSNR Time MPSNR Time MPSNR Time MPSNR Time Bunny 28.402 1731 28.211 1203 29.523 453.7 29.163 486.7 30.729 1497 32.401 691.3 News 28.234 1720 27.882 340.4 28.714 535.9 28.857 978.1 30.027 1426 31.643 932.4 Salesman 29.077 1783 28.469 353.6 29.534 656.4 28.288 689.3 30.621 1148 31.684 769.5 Silent 30.126 1453 30.599 316.2 30.647 1305 31.081 453.2 31.731 1232 32.706 532.3\",\n \"figure_data\": \"VideoFBCP [34]TMac [26]TMacTT [3]TRLRF [30]TW [25]SVDinsTNObservedFBCP [34] TMac [26] TMacTT [3] TRLRF [30]TW [25]SVDinsTN Ground truth0 00. 1 0.10. 2 0.20. 3 0.30. 4 0.40. 5 0.50. 6 0.60. 7 0.70. 8 0.80. 9 0.91 1\",\n \"figure_id\": \"tab_7\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"[12,14]\", \"Explanation\": \"The cited works provide a theoretical foundation for the challenging and NP-hard problem of TN structure search, which the citing paper aims to address in its research on the same topic.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work introduces the TNGA method, which the citing paper adopts in their research to solve the TN-SS problem and compute TN cores.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work presents the TNALE method, which the citing paper uses in their study to solve the TN-SS problem and compute TN cores.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[11]\", \"Explanation\": \"The cited work introduces the TNGreedy method, which the citing paper references in their research to provide a method for solving the TN-SS problem.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work presents the TNGA method, which the citing paper uses in their study to provide a method for solving the TN-SS problem.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[15]\", \"Explanation\": \"The cited work introduces the TNLS method, which the citing paper references in their research to provide a method for solving the TN-SS problem.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work presents the TNALE method, which the citing paper uses in their study to provide a method for solving the TN-SS problem.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[11]\", \"Explanation\": \"The cited work introduces the greedy algorithm as a heuristic search method for sampling candidate structures in TN-SS methods, which the citing paper adopts in its research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work presents the genetic algorithm as a heuristic search method for sampling candidate structures in TN-SS methods, which the citing paper adopts in its research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work introduces the alternating local enumeration algorithm as a heuristic search method for sampling candidate structures in TN-SS methods, which the citing paper adopts in its research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[37]\", \"Explanation\": \"The cited work provides a review of operations on tensors, which the citing paper uses to define and perform various operations on tensors in the context of the research conducted.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[4]\", \"Explanation\": \"The cited work provides the definition of tensor networks and the operations used in contracting modes, which the citing paper adopts in representing higher-order data using TN cores.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[16]\", \"Explanation\": \"The cited work provides the foundational concept of a one-to-one correspondence between the TN structure and FCTN rank, which the citing paper uses to inspire their research on optimizing the TN structure by optimizing the FCTN rank.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[37]\", \"Explanation\": \"The cited work on FCTN decomposition provides the basis for the insertion of diagonal factors between TN cores in the citing paper to determine the FCTN rank and reveal the TN structure.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[10]\", \"Explanation\": \"The cited work introduces the concept of Tikhonov regularization, which the citing paper adopts in the proposed model to constrain the feasible range of TN cores G.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[2]\", \"Explanation\": \"The cited work provides the algorithm for solving the optimization problem in the citing paper, which the citing paper adopts in their research to update the parameters in the model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"[6]\", \"Explanation\": \"The cited work introduces the ADMM algorithm, which the citing paper adopts to solve the S t,l subproblem in the optimization process of the model.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[35]\", \"Explanation\": \"The cited work provides the real-world data used in the experiment to test the light field data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[37]\", \"Explanation\": \"The cited work provides the real-world data used in the experiment to test the light field data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[11]\", \"Explanation\": \"The cited work provides the method used to represent the test light field data in the experiment.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[13]\", \"Explanation\": \"The cited work provides the method used to represent the test light field data in the experiment.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[15]\", \"Explanation\": \"The cited work provides the method used to represent the test light field data in the experiment.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[14]\", \"Explanation\": \"The cited work provides the method used to represent the test light field data in the experiment.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[34]\", \"Explanation\": \"The cited work, FBCP, provides a baseline method for comparison in the study of tensor decomposition in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[26]\", \"Explanation\": \"The cited work, TMac, is a baseline method for comparison in the study of tensor decomposition in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[3]\", \"Explanation\": \"The cited work, TMacTT, is a baseline method for comparison in the study of tensor decomposition in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[30]\", \"Explanation\": \"The cited work, TRLRF, is a baseline method for comparison in the study of tensor decomposition in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"[25]\", \"Explanation\": \"The cited work, TW, is a baseline method for comparison in the study of tensor decomposition in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[34]\", \"Explanation\": \"The data source for the study in the citing paper is the FBCP method.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[26]\", \"Explanation\": \"The data source for the study in the citing paper is the TMac method.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[3]\", \"Explanation\": \"The data source for the study in the citing paper is the TMacTT method.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[30]\", \"Explanation\": \"The data source for the study in the citing paper is the TRLRF method.\"}, {\"Category\": \"Data Source\", \"Citation\": \"[25]\", \"Explanation\": \"The data source for the study in the citing paper is the TW method.\"}]"}}},{"rowIdx":98,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":["fig_1","fig_2"],"heading":"Introduction","publication_ref":["b29","b32","b26","b23","b37","b51","b34","b20","b39","b39","b34","b20","b28","b54","b52"],"table_ref":["tab_5"],"text":"The named entity recognition (NER) task aims to locate and classify entity spans in a given text into predefined entity types. It is widely used for many downstream applications, such as relation extraction and question answering. Deep neural networks have made significant progress on this task leveraging large-scale human-annotated data for training. However, fine-grained token-level annotation makes it costly to collect enough high-quality labeled data, especially for low-resource languages. Such scenarios motivate the research on zero-shot cross-lingual NER, which attempts to leverage labeled data in a rich-resource source language to solve the NER task in a target language without annotated data.\nRecent attempts at cross-lingual NER can be roughly categorized from two aspects: learning language-independent features via feature alignment (Huang et al., 2019;Keung et al., 2019) and learning language-specific features from automatically labeled target-language data (Wu et al., 2020b,a). Despite bringing great success to crosslingual NER, the former line of research misses exploiting language-specific features and thus shows substandard performance, especially when transferring to distant languages, e.g., from English to Arabic (Fu et al., 2023). Hence, a series of studies focuses on the latter category, which typically creates pseudo-labeled target-language data and uses it to perform conventional supervised learning or teacher-student learning. For example, as shown in Fig 1 (a), earlier studies (Ehrmann et al., 2011;Mayhew et al., 2017;Xie et al., 2018;Jain et al., 2019), such as TMP (Jain et al., 2019), first translate labeled data in the source language and then perform label projection. Recently, several approaches have utilized a weak model, which could be an NER model either trained on the source language's labeled data as in TSL (Wu et al., 2020b), or further finetuned on the generated translation data as in UniTrans (Wu et al., 2020a), to annotate the unlabeled target-language data for improvement, as shown in Fig 1(b) andFig 1(c). Unfortunately, these methods inevitably suffer from the label noise induced by inaccurate translation and label projection, or the weak model's limited capability. Although some methods are proposed to mitigate the label noise problem by additionally training an instance selector (Liang et al., 2021;Chen et al., 2021) or designing heuristic rules for data selection (Ni et al., 2017), they independently manipulate either the translation data (D trans ) (Ni et al., 2017) or the target-language data (D tgt ) pseudo-labeled by NER models trained in the source language (Liang et al., 2021;Chen et al., 2021). Hence, all these methods ignore the complementary characteristics between both for denoising. Particularly, from the text view, D tgt is collected from a natural text distribution of the target-language data, while D trans can be regarded as a way of data augmentation to provide more lexicon variants. From the labeling function view, labels of D trans are obtained via the label projection algorithm, which have little association with those of D tgt generated by NER models.\nWith such consideration, we propose a modelcollaboration-based denoising scheme, which incorporates models trained on both data sources to mutually denoise the pseudo-labels of both data sources in an iterative way. As shown in Fig 1(d), we first leverage M tgt trained on the pseudolabeled target-language data D tgt to denoise the translation data annotated by label projection. In this way, the learned model M trans will be less affected by noise in the translation data. We then employ the improved M trans to re-label the targetlanguage unlabeled data D tgt . It is expected that there is less noise in the relabeled data, and thus we can produce a more powerful M tgt . We perform this procedure for several iterations, so that all the involved data sources and models can be improved in an upward spiral.\nMoreover, borrowing the idea from anomaly detection (Gu et al., 2019) that a given data point's neighborhood information can be used to measure its anomalism, here we find that the similar tokens in the feature space can also collaborate for denois- ing. Previous studies (Zhai and Wu, 2019;Xu et al., 2020) have shown that instances with the same label are more likely to locate close to each other in the representation space. Our intuition is that, if a token's label conflicts a lot with labels of other tokens in its neighborhood, then this label is probably noisy. Therefore, we further propose an instancecollaboration-based denoising strategy to explore the neighborhood structure of each token for denoising, as shown in Figure 2. Specifically, we utilize the label consistency of each token's neighborhood in the representation space to re-weight the soft-labeled examples in knowledge distillation.\nWe integrate the instance-collaboration-based denoising strategy into the model-collaboration-based denoising scheme and propose a Collaborative Label Denoising framework, i.e., CoLaDa, for cross-lingual NER. We conduct extensive experiments on two popular benchmarks covering six languages for evaluation. Experimental results show that our method outperforms existing state-of-theart methods. Qualitative and quantitative analyses further demonstrate the effectiveness of our framework in reducing the data noise."},{"figure_ref":[],"heading":"Problem Formulation","publication_ref":[],"table_ref":[],"text":"Here we take the typical sequence labeling formulation for the named entity recognition task. Given a sequence with L tokens x = (x 1 , . . . , x L ) as the input text, an NER system is excepted to assign each token x i with a label y i .\nIn this paper, we assume to have the labeled training data D src = {(x s , y s )} in the source language, the unlabeled data D tgt = {x u } from the target language, and translation data D trans = {(x t , y t )} obtained by data projection from D src . Our goal is to train an NER model M that can generalize well to the target language utilizing these resources. "},{"figure_ref":[],"heading":"CoLaDa Framework","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Instance Collaboration for Denoising","publication_ref":["b54","b52"],"table_ref":[],"text":"Previous work (Zhai and Wu, 2019;Xu et al., 2020) indicates that tokens with the same labels are more likely to locate close to each other in the representation space of a deep neural network. If the label of a given token is inconsistent with lots of its neighbors, this token would be isolated from other tokens with the same label in the feature space, and hence its label is more likely to be noisy. Therefore, we propose instance-collaboration-based denoising, which evaluates the reliability of a given token's label by measuring the label consistency of its neighborhood, and then uses the reliability score to weight the noisy labels from a teacher model M for knowledge distillation on data D = {x}. Noisy labels are expected to have lower weights than clean ones.\nCreate a memory bank. We leverage the feature extractor F of the NER model M to obtain the hid-\nden representations h = {h i } L i=1 of each sentence x = {x i } L i=1 ∈ D: h = F(x).(1)\nWe then construct a memory bank B D = {h} to store the hidden representations of all tokens in D.\nCompute label consistency. Given a token x i , we retrieve its K-nearest neighbors N K (x i ) in B D using cosine similarity. Let p i denote the soft label (i.e., the probability distribution over the entity label set) assigned by the teacher model M for x i . We measure the label consistency of x i , i.e., λ(p i ; x i ), by calculating the fraction of x i 's neighbors that are assigned with the same labels as\nx i in N K (x i ): λ(p i ; x i ) = 1 K x j ∈N k (x i ) I(ŷ j = ŷi ),(2)\nwhere ŷi = arg max (p i ) is the pseudo entity label corresponding to the maximum probability in p i . Similarly, ŷj is the pseudo entity label corresponding to x j . I is the indicator function.\nProduce a reliability score. We use the label consistency λ(p i ; x i ) to compute the reliability score of the soft label p i , which is further used as the weight of p i during model learning (see 3.2).\nConsidering that different entity types may contain different levels of label noise and show different statistics on label consistency, here we present a class-adaptive reliability score for weighting:\nw(p i ; x i ) = Sigmoid (α (λ(p i ; x i ) -µ(ŷ i ))) ,(3)\nwhere µ(ŷ i ) denote the mean of all λ(p j ; x j ) where arg max (p j ) = ŷi and x j ∈ D. α > 0 is a hyper-parameter that controls the sharpness of the weighting strategy. If α → 0, all tokens have equal weights. If α → ∞, tokens whose label consistency is larger than the average label consistency w.r.t. its pseudo label will be weighted with 1 and those with smaller consistency will be dropped."},{"figure_ref":[],"heading":"Model Collaboration for Denoising","publication_ref":[],"table_ref":[],"text":"Here we elaborate on the details of the two noiserobust training processes. Algorithm 1 depicts the overall training procedure of CoLaDa.\nNoise-robust training on translation data. Assuming the availability of a collaborator M tgt 2 trained on pseudo-labeled target-language data D tgt , here we focus on leveraging M tgt to reduce the label noise in the translation data D trans = {(x t , y t )}, with which we further deliver a more powerful model M trans . Specifically, given a sentence (x t , y t ) ∈ D trans , we first obtain the soft label p u i of each x t i ∈ x t from the collaborator M tgt . Then, we take both the one hot label y t i and the soft label p u i as the supervision to train the model M trans .3 Denote the output probability distribution of M trans for x t i as pt i . The loss function w.r.t. x t is defined as:\nL x t = 1 L L i=1 β 1 CE(p t i , p u i ) + β 2 CE(p t i , y t i ) ,(4)\nwhere CE(•, •) denotes the cross-entropy loss, L is the sentence length, β 1 and β 2 are weighting scalars. Here we further incorporate the instancecollaboration-based denoising strategy (3.1) to provide a token-level reliability evaluation to the supervision from the collaborator M tgt via:\nβ 1 (x t i ) ← β 1 * w(p u i , x t i ),(5)\nwhere w(p u i , x t i ) is calculated by Eq. ( 3). Noise-robust training on target-language unlabeled data. Here we leverage M trans obtained via the above noise-robust training on translation data to provide high-quality supervision for Algorithm 1 Pseudo code of CoLaDa. on Dtrans = {(x t , y t )} to get the predictions Dtrans = {(x t , p u )} 5:\nGet w for (x t , p u ) ∈ Dtrans with\nM (i-1)\ntgt , Eq.(3) 6:\nTrain\nM (i)\ntrans with loss on (x t , y t , p u , w), Eq.( 4) 7:\n# Step 2: Noise-robust training on Dtgt 8:\nInference\nM (i) trans on Dtgt = {x u } to get the predic- tions Dtgt = {(x u , p t )} 9: Get w ′ for (x u , p t ) ∈ Dtgt with M (i)\ntrans , Eq.(3) 10:\nTrain\nM (i)\ntgt with loss on (x u , p t , w ′ ), Eq.( 6) 11: end for Output: an NER model M (T ) tgt .\nD tgt = {x u }. By performing knowledge distillation on D tgt , the student model M tgt is supposed to benefit from the unlabeled data drawn from the real text distribution in the target language with the knowledge from the teacher model M trans .\nSpecifically, given a sentence x u ∈ D tgt , we first utilize M trans to predict soft label p t i for each token x u i ∈ x u . Then, we integrate the instancecollaboration-based denoising technique into the learning process. The loss function w.r.t. x u to train the student model M tgt can be formulated as:\nL x u = 1 L L i=1 w(p t i , x u i ) • CE(p u i , p t i ),(6)\nwhere pu i denotes the output probability distribution of M tgt for the i-th token x u i and w(p t i , x u i ) is calculated by Eq. (3)."},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Experiment Settings","publication_ref":["b42","b43","b40","b43","b42","b49","b24","b53"],"table_ref":[],"text":"Datasets We conduct experiments on two standard cross-lingual NER benchmarks: CoNLL (Tjong Kim Sang, 2002;Tjong Kim Sang and De Meulder, 2003) and WikiAnn (Pan et al., 2017). CoNLL contains four languages: English (en) and German (de) from the CoNLL-2003 4 NER shared task (Tjong Kim Sang and De Meulder, 2003), and Spanish (es) and Dutch (nl) from the CoNLL-20025 NER shared task (Tjong Kim Sang, 2002). This dataset is annotated with four entity types: PER, LOC, ORG, and MISC. WikiAnn contains an English dataset and datasets in three non-western languages: Arabic (ar), Hindi (hi), and Chinese (zh). Each dataset is annotated with 3 entity types: PER, LOC, and ORG. All datasets are annotated with the BIO tagging scheme. We use the train, development, and test splits as previous work (Wu and Dredze, 2019;Wu et al., 2020a).\nWe take English as the source language and other languages as the target language, respectively. We remove the labels of the training data for the target language and take it as the unlabeled target language data. For the CoNLL benchmark, we use the word-to-word translation data provided in UniTrans (Wu et al., 2020a) for a fair comparison. For the WikiAnn benchmark, we translate the source data to the target language with the public M2M100 (Fan et al., 2020) translation system and conduct label projection with the marker-based alignment algorithm as Yang et al. (2022).\nEvaluation The entity-level micro-F1 on test set of the target language is used as the evaluation metric. We report the mean value of 5 runs with different seeds for all the experiments."},{"figure_ref":[],"heading":"Implementation Details","publication_ref":["b45","b31","b49","b57","b22","b36","b49","b32"],"table_ref":[],"text":"For the base NER model, we stack a linear classifier with softmax over a base encoder such as mBERT. We implement our framework with Pytorch 1.7.16 , the Hugging-Face transformer library (Wolf et al., 2020), and use FAISS (Johnson et al., 2019) for embedding retrieval. Following Wu and Dredze (2019) and Zhou et al. (2022), we use the multilingual BERT base model (Devlin et al., 2019) and XLM-R (Conneau et al., 2020) large model as our base encoders. Most of our hyper-parameters are set following Wu et al. (2020a). We use AdamW (Loshchilov and Hutter, 2019) as optimizer and train the model on source NER data with the learning rate of 5e-5 for 3 epochs. The dropout rate is 0.1. For teacher-student learning, we train the model with a learning rate of 2e-5 for 10 epochs. We freeze the bottom three layers as Wu and Dredze (2019). Following Keung et al. (2019), we choose other hyper-parameters according to the target language dev set. We set K in Eq. ( 2) to 500 and α in Eq. (3) to 6. For the first iteration, we start with an NER model trained on the source-language data to denoise the translation data with β 1 and β 2 in Eq. ( 5) setting to 0.5. For the following iterations, β 1 is set to 0.9 and β 2 is set to 0.1. The maximum number of iterations is 8. CoLaDa outperforms prior methods with both encoders, achieving a significant improvement of 2.70 F1 scores on average for CoNLL and 10.14 F1 scores on average for WikiAnn with XLM-R as the encoder. This well demonstrates the effectiveness of our approach. Interestingly, CoLaDa shows more significant superiority when transferring to distant target languages in WikiAnn. The knowledge distillation based baselines (i.e., TSL, RIKD, MTMT) struggle on distant languages such as Chinese (zh) due to the noisy predictions from the weak teacher model M src trained in the source language. UniTrans, which is developed with the same data sources as ours, shows poor performance, especially in distant languages such as Arabic (ar). We conjecture that the problem of label noise is even more critical in these distant languages. Our CoLaDa can better handle noise in both translation data and unlabeled target-language data, thus leading to significant performance gains."},{"figure_ref":[],"heading":"Analysis","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Ablation Study","publication_ref":[],"table_ref":["tab_3"],"text":"To further validate the effectiveness of each mechanism in the proposed framework, we introduce the following variants of CoLaDa in an ablation study: 1) CoLaDa w/o instance collaboration, where we directly set the reliability score in Eq. ( 3) to 1 for all tokens. 2) CoLaDa w/o translation data denoise, where we set β 1 in Eq. ( 4) to 0. 3) CoLaDa w/o iteratively denoise, where we remove the iterative enhancement and only conduct the denoising process for one iteration. 4) CoLaDa w/o model collaboration, where we set β 1 in Eq. ( 4) to 0, remove the iteration mechanism, and directly take the model finetuned on D trans as the teacher model to train a student model with instance-collaborationbased denoising on D tgt . 5) CoLaDa w/o instance & model collaboration, which further drops the instance-collaboration-based denoising from 4). Table 3 shows the ablation results. We can draw some in-depth conclusions as follows. 1) CoLaDa outperforms CoLaDa w/o instance collaboration, which highlights the effectiveness of leveraging neighborhood information to reduce label noise in knowledge distillation.\n2) CoLaDa outperforms CoLaDa w/o translation data denoise, which emphasizes the importance of using the collaborator M tgt to refine labels of translation data, especially in distant languages where the translation data is noisier (e.g., 12.8 F1 drop on Arabic and 5.7 F1 drop on Chinese).\n3) CoLaDa outperforms CoLaDa w/o iteratively denoise, which indicates the necessity of iterative learning: models obtained from the previous iteration should be re-used as the collaborator to further improve label quality in the next iteration.\n4) CoLaDa w/o instance & model collaboration, which eliminates all denoising strategies from CoLaDa, leads to a significant performance drop, demonstrating the essentiality of label denoising for cross-lingual NER."},{"figure_ref":["fig_5","fig_5"],"heading":"Analysis of Model Collaboration","publication_ref":[],"table_ref":[],"text":"Here we attempt to understand how the two models, i.e., M trans and M tgt , collaboratively improve each other. As shown in Figure 4, F1 scores of M trans and M tgt consistently improve as iterations go on, and finally converge at the last iteration. This indicates that both models benefit from the proposed model collaboration scheme. Two reasons are speculated: i) An improved M tgt can provide more accurate labels on the translation data, which further help to improve M trans via noise-robust learning on such translation data. For example, at the initial step (i = 0), the F1 score of the model M 0 trans trained on the original translation labels is 50.0. With the additional supervision from the collaborator M 0 tgt , M 1 trans achieves a performance gain of 5.7 F1. ii) An improved M trans predicts pseudo labels with higher quality on the target-language unlabeled data, which further benefits the learning of M tgt . As in Figure 4, the quality of pseudo-labeled D tgt (the green line) grows as M trans improves. In this way, both M trans and M tgt are providing more and more reliable labels for each other to learn as the iterations progress."},{"figure_ref":["fig_6"],"heading":"Analysis of Instance Collaboration","publication_ref":["b38"],"table_ref":[],"text":"This subsection dives into the working mechanism of the instance-collaboration-based denoising.\nReliability scores v.s. label quality. To study the relationship between reliability score and label quality, we partition tokens in the target-language unlabeled data, x i ∈ D tgt into several bins according to their reliability scores w(p t i , x i ) calculated via M\n(1) trans . Then, we compute the tokenlevel F1 over each bin by comparing pseudo labels ŷi = arg max(p t i ) to the ground-truth ones. As shown in Figure 5, the label quality is proportional to the reliability score, which well demonstrates the effectiveness of our instance-collaboration-based denoising strategy. Analysis of Label Consistency. We also study the characteristics of label consistency w.r.t. different entity types and representation spaces of the memory bank. Figure 6 shows the results. We can draw some in-depth observations as follows. i) Clean tokens show a larger average consistency than noisy tokens w.r.t. all entity types, demonstrating the effectiveness of our label consistency based denoising strategy again.\nii) Different entity types lead to different distributions of label consistency, which validates the necessity of our design for class-adaptive reliability score for weighting as Eq.(3).\niii) Label consistencies calculated with token representations from the upper layers are generally larger than those corresponding to the bottom layers. Also, the label consistency gap between clean tokens and noisy tokens gets larger from the bottom to the top (e.g., the gap between two orange lines). This may be attributed to the fact that representations from upper layers are more taskspecific (Muller et al., 2021), hence they can better discriminate between noisy and clean tokens. Original English:\nThere have been many tenants, including The Sunday Times[ORG] '' and The Daily News[ORG] ''."},{"figure_ref":[],"heading":"Inaccurate alignment boundary","publication_ref":[],"table_ref":[],"text":"The symbols《 and 》are corner brackets used to enclose the name of a newspaper, and other works.\nOriginal Translation: 有许多租户，包括 《 星期日时报[ORG] 》 和 《 每日新闻[ORG] 》。 Translate-train: 有许多租户，包括 《 星期日时报[ORG] 》 和 《 每日新闻[ORG] 》。 Ours: 有许多租户，包括 《 星期日时报[ORG] 》 和 《 每日新闻[ORG] 》。 Original English: It is found in Democratic Republic of Congo[ORG] , Kenya[LOC] , Tanzania[LOC] …."},{"figure_ref":["fig_7"],"heading":"Label noise in English data","publication_ref":[],"table_ref":[],"text":"Wrong entity type in original English data for \"Democratic Republic of China\".\nOriginal Translation: 它位于刚果民主共和国[ORG] ，肯尼亚[LOC] ，坦桑尼亚[LOC] … Translate-train: 它位于刚果民主共和国[ORG] ，肯尼亚[LOC] ，坦桑尼亚[LOC] … Ours: 它位于刚果民主共和国[LOC] ， 肯尼亚[LOC] ，坦桑尼亚[LOC] …\nFigure 8: Case study on translation data in Chinese. The blue (red) texts denote the correct (incorrect) entity labels.\nThe original translation lines display the translation texts and labels obtained by data projection. Translate-train and Ours illustrate the predictions from the translate-train method (M 0 trans ) and our CoLaDa, respectively.\nChoice of K for neighborhood information.\nFigure 7 shows the performance of CoLaDa using different K in Eq. ( 2). Generally speaking, CoLaDa is robust to the choice of K. Any value for K > 0 leads to a better performance compared with removing the instance collaboration, i.e., K = 0. A smaller K may lead to a slight performance drop due to limited neighborhood information."},{"figure_ref":[],"heading":"Case Study","publication_ref":[],"table_ref":[],"text":"To better illustrate the kinds of label noise presented in the data and the capability of CoLaDa to address such noise, we conduct a case study on the Chinese translation data from the WikiAnn English data. As shown in Figure 8, there are three typical cases of noisy labels in the translation data: noisy labels induced by inaccurate translations, alignment errors, and annotation errors in the original source-language data. 7 Figure 8 shows that the translate-train model, finetuned on the original translation data, overfits the noisy labels. However, CoLaDa is less affected by such noise and makes correct predictions."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b41","b44","b39","b22","b21","b49","b50","b32","b20","b25","b26","b51","b37","b35","b53","b39","b19","b27","b33","b20","b34","b56","b57"],"table_ref":[],"text":"Prior work on cross-lingual NER mainly falls into two major categories: feature-based and data-based transfer.\nFeature-based These methods learn languageindependent features so that the model trained on the source language can directly adapt to the target language. Earlier work exploits word clusters (Täckström et al., 2012), gazetteers (Zirikly 7 Due to the short entity context information in many sentences in WikiAnn, the translation quality of entity mentions with M2M100 is less than satisfactory on the dataset. and Hagiwara, 2015), Wikifier features (Tsai et al., 2016), and cross-lingual word embedding (Ni et al., 2017), etc. More recently, with the fast growth of multilingual pre-trained language models (Devlin et al., 2019;Conneau et al., 2020) and their promising results on cross-lingual transfer (Wu and Dredze, 2019), lots of studies build upon such pre-trained models and further promote the learning of language-independent features via meta-learning (Wu et al., 2020c), contrastive alignment (Wu and Dredze, 2020), adversarial learning (Keung et al., 2019;Chen et al., 2021), and by integrating other resources (Fetahu et al., 2022). Despite the great success, they mostly ignore language-specific features, which are especially important when transferring to distant languages (Fu et al., 2023).\nData-based These approaches learn languagespecific features via automatically labeled targetlanguage data and can be further divided into translation-based and knowledge distillation-based methods.\nTranslation-based methods first translate the source-language data to the target language, then perform label projection from the source side to the target side. Some prior studies have proposed to use cheap translation such as word-to-word (Xie et al., 2018) or phrase-to-phrase (Mayhew et al., 2017) translation. Jain et al. (2019) propose an entity projection algorithm to utilize the Google translation system. Recently, Liu et al. (2021) and Yang et al. (2022) propose to translate sentences with pre-defined markers for label projection. And Ni et al. (2017) design heuristic rules to select highquality translation data. However, both data noise and artifacts (Artetxe et al., 2020) in the transla-tion data still limit the performance of such methods (García-Ferrero et al., 2022).\nKnowledge distillation-based methods train a student model on unlabeled target-language data with the soft labels from a teacher model (Wu et al., 2020b). Li et al. (2022) improve the single task based teacher-student learning with entity similarity as an auxiliary task. To mitigate the label noise from the teacher model, Chen et al. (2021) propose AdvPicker, which trains a language discriminator to select the less language-dependent unlabeled data for knowledge distillation; Liang et al. (2021) design a reinforcement learning algorithm to train an instance selector according to features such as model confidence to select reliable pseudo labels iteratively.\nWhile most previous work leverage either translation data or unlabeled data, UniTrans (Wu et al., 2020a) utilizes the model trained on translation data to perform teacher-student learning on unlabeled data. But it still suffers from the data noise problem. More recently, consistency training (Zheng et al., 2021;Zhou et al., 2022) has also been explored to leverage both unlabeled data and translation data without explicit label annotation.\nTo the best of our knowledge, we are the first to propose a unified denoising framework to handle data noise in both translation and unlabeled data collaboratively from the model and instance levels for cross-lingual NER."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"To address the problem of label noise in crosslingual NER, this paper presents CoLaDa, a collaborative label denoising framework. We propose a model-collaboration-based denoising scheme to make two models trained on different data sources to denoise the labels of each other and hence promote each other's learning. We further propose an instance-collaboration-based strategy that collaboratively considers the label consistency among similar tokens in the feature space to re-weight the noisy labels assigned by a teacher model in knowledge distillation. By integrating the instancecollaboration strategy into the model-collaboration denoising scheme, our final framework CoLada achieves superior performance over prior start-ofthe-art methods by benefiting from better handling the data noise."},{"figure_ref":[],"heading":"Limitations","publication_ref":["b18"],"table_ref":[],"text":"Our framework relies on the availability of translation system and unlabeled data in the target language, which can not be applied to languages without any unlabeled text or translation text. The knowledge distillation step requires a certain amount of unlabeled text, while it may struggle in cases where only few hundreds of unlabeled sentences are available. It would be interesting to combine our label denoising framework with data augmentation techniques in such scenarios. Besides, the boarder application to other low-resource languages, such as MasakhaNER 2.0 (Adelani et al., 2022), and other cross-lingual sequence labeling tasks are left for exploration in future work. "},{"figure_ref":[],"heading":"A Appendix","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"A.1 Dataset Statistics","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"A.2 Other Implementation Details","publication_ref":[],"table_ref":[],"text":"All experiments are conducted on a Tesla V100 (32GB). The total of trainable parameters (M trans and M tgt ) for our model with mBERT-base-cased as the encoder is 172M and the training time is about 35 mins for one iteration. With XLM-Rlarge as our base encoder, the total of trainable parameters are 822M and the training takes about 90 mins for one iteration."},{"figure_ref":[],"heading":"A.3 Baselines","publication_ref":["b49","b21","b50","b32","b20","b35","b55","b34","b33","b56","b57"],"table_ref":[],"text":"We consider the following start-of-the-art baselines: mBERT (Wu and Dredze, 2019) and XLM-R (Conneau et al., 2020) directly train an NER model on the labeled data in the source language, with mBERT and XLM-R as the basic encoder, respectively.\nBERT-align (Wu and Dredze, 2020) tries to explicitly add word-level contrastive alignment loss to enhance the mBERT representation.\nAdvCE (Keung et al., 2019) exploits adversarial learning on source-and target-language text to avoid learning language-specific information.\nAdvPicker (Chen et al., 2021) leverages adversarial learning to learn language-shared features and then selects the less language-specific sentences in target-language unlabeled text for knowledge distillation.\nMulDA (Liu et al., 2021) proposes the labeled sequence translation method for data projection from source-language NER data, a generative model is further applied to augment more diverse examples in the target language.\nUniTrans (Wu et al., 2020a) unifies model-and translation-data-based-transfer via knowledge distillation.\nTOF (Zhang et al., 2021) leverages the labeled data for machine reading comprehension task on target language to help the NER task in crosslingual transfer.\nTSL (Wu et al., 2020b) proposes knowledge distillation to use unlabeled target-language data for cross-lingual NER.\nRIKD (Liang et al., 2021) proposes a reinforcement learning algorithm to iteratively select reliable pseudo-labels for knowledge distillation.\nMTMT (Li et al., 2022) proposes multi-task multi-teacher knowledge distillation, which further leverages the entity similarity task.\nxTune (Zheng et al., 2021) leverages unlabeled translation text and other word-level data augmentation techniques for consistency training.\nConNER (Zhou et al., 2022) conducts spanlevel consistency training on unlabeled targetlanguage data using translation and further applies dropout-based consistency training on the sourcelanguage data."}],"string":"[\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_2\"\n ],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b29\",\n \"b32\",\n \"b26\",\n \"b23\",\n \"b37\",\n \"b51\",\n \"b34\",\n \"b20\",\n \"b39\",\n \"b39\",\n \"b34\",\n \"b20\",\n \"b28\",\n \"b54\",\n \"b52\"\n ],\n \"table_ref\": [\n \"tab_5\"\n ],\n \"text\": \"The named entity recognition (NER) task aims to locate and classify entity spans in a given text into predefined entity types. It is widely used for many downstream applications, such as relation extraction and question answering. Deep neural networks have made significant progress on this task leveraging large-scale human-annotated data for training. However, fine-grained token-level annotation makes it costly to collect enough high-quality labeled data, especially for low-resource languages. Such scenarios motivate the research on zero-shot cross-lingual NER, which attempts to leverage labeled data in a rich-resource source language to solve the NER task in a target language without annotated data.\\nRecent attempts at cross-lingual NER can be roughly categorized from two aspects: learning language-independent features via feature alignment (Huang et al., 2019;Keung et al., 2019) and learning language-specific features from automatically labeled target-language data (Wu et al., 2020b,a). Despite bringing great success to crosslingual NER, the former line of research misses exploiting language-specific features and thus shows substandard performance, especially when transferring to distant languages, e.g., from English to Arabic (Fu et al., 2023). Hence, a series of studies focuses on the latter category, which typically creates pseudo-labeled target-language data and uses it to perform conventional supervised learning or teacher-student learning. For example, as shown in Fig 1 (a), earlier studies (Ehrmann et al., 2011;Mayhew et al., 2017;Xie et al., 2018;Jain et al., 2019), such as TMP (Jain et al., 2019), first translate labeled data in the source language and then perform label projection. Recently, several approaches have utilized a weak model, which could be an NER model either trained on the source language's labeled data as in TSL (Wu et al., 2020b), or further finetuned on the generated translation data as in UniTrans (Wu et al., 2020a), to annotate the unlabeled target-language data for improvement, as shown in Fig 1(b) andFig 1(c). Unfortunately, these methods inevitably suffer from the label noise induced by inaccurate translation and label projection, or the weak model's limited capability. Although some methods are proposed to mitigate the label noise problem by additionally training an instance selector (Liang et al., 2021;Chen et al., 2021) or designing heuristic rules for data selection (Ni et al., 2017), they independently manipulate either the translation data (D trans ) (Ni et al., 2017) or the target-language data (D tgt ) pseudo-labeled by NER models trained in the source language (Liang et al., 2021;Chen et al., 2021). Hence, all these methods ignore the complementary characteristics between both for denoising. Particularly, from the text view, D tgt is collected from a natural text distribution of the target-language data, while D trans can be regarded as a way of data augmentation to provide more lexicon variants. From the labeling function view, labels of D trans are obtained via the label projection algorithm, which have little association with those of D tgt generated by NER models.\\nWith such consideration, we propose a modelcollaboration-based denoising scheme, which incorporates models trained on both data sources to mutually denoise the pseudo-labels of both data sources in an iterative way. As shown in Fig 1(d), we first leverage M tgt trained on the pseudolabeled target-language data D tgt to denoise the translation data annotated by label projection. In this way, the learned model M trans will be less affected by noise in the translation data. We then employ the improved M trans to re-label the targetlanguage unlabeled data D tgt . It is expected that there is less noise in the relabeled data, and thus we can produce a more powerful M tgt . We perform this procedure for several iterations, so that all the involved data sources and models can be improved in an upward spiral.\\nMoreover, borrowing the idea from anomaly detection (Gu et al., 2019) that a given data point's neighborhood information can be used to measure its anomalism, here we find that the similar tokens in the feature space can also collaborate for denois- ing. Previous studies (Zhai and Wu, 2019;Xu et al., 2020) have shown that instances with the same label are more likely to locate close to each other in the representation space. Our intuition is that, if a token's label conflicts a lot with labels of other tokens in its neighborhood, then this label is probably noisy. Therefore, we further propose an instancecollaboration-based denoising strategy to explore the neighborhood structure of each token for denoising, as shown in Figure 2. Specifically, we utilize the label consistency of each token's neighborhood in the representation space to re-weight the soft-labeled examples in knowledge distillation.\\nWe integrate the instance-collaboration-based denoising strategy into the model-collaboration-based denoising scheme and propose a Collaborative Label Denoising framework, i.e., CoLaDa, for cross-lingual NER. We conduct extensive experiments on two popular benchmarks covering six languages for evaluation. Experimental results show that our method outperforms existing state-of-theart methods. Qualitative and quantitative analyses further demonstrate the effectiveness of our framework in reducing the data noise.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Problem Formulation\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Here we take the typical sequence labeling formulation for the named entity recognition task. Given a sequence with L tokens x = (x 1 , . . . , x L ) as the input text, an NER system is excepted to assign each token x i with a label y i .\\nIn this paper, we assume to have the labeled training data D src = {(x s , y s )} in the source language, the unlabeled data D tgt = {x u } from the target language, and translation data D trans = {(x t , y t )} obtained by data projection from D src . Our goal is to train an NER model M that can generalize well to the target language utilizing these resources. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"CoLaDa Framework\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Instance Collaboration for Denoising\",\n \"publication_ref\": [\n \"b54\",\n \"b52\"\n ],\n \"table_ref\": [],\n \"text\": \"Previous work (Zhai and Wu, 2019;Xu et al., 2020) indicates that tokens with the same labels are more likely to locate close to each other in the representation space of a deep neural network. If the label of a given token is inconsistent with lots of its neighbors, this token would be isolated from other tokens with the same label in the feature space, and hence its label is more likely to be noisy. Therefore, we propose instance-collaboration-based denoising, which evaluates the reliability of a given token's label by measuring the label consistency of its neighborhood, and then uses the reliability score to weight the noisy labels from a teacher model M for knowledge distillation on data D = {x}. Noisy labels are expected to have lower weights than clean ones.\\nCreate a memory bank. We leverage the feature extractor F of the NER model M to obtain the hid-\\nden representations h = {h i } L i=1 of each sentence x = {x i } L i=1 ∈ D: h = F(x).(1)\\nWe then construct a memory bank B D = {h} to store the hidden representations of all tokens in D.\\nCompute label consistency. Given a token x i , we retrieve its K-nearest neighbors N K (x i ) in B D using cosine similarity. Let p i denote the soft label (i.e., the probability distribution over the entity label set) assigned by the teacher model M for x i . We measure the label consistency of x i , i.e., λ(p i ; x i ), by calculating the fraction of x i 's neighbors that are assigned with the same labels as\\nx i in N K (x i ): λ(p i ; x i ) = 1 K x j ∈N k (x i ) I(ŷ j = ŷi ),(2)\\nwhere ŷi = arg max (p i ) is the pseudo entity label corresponding to the maximum probability in p i . Similarly, ŷj is the pseudo entity label corresponding to x j . I is the indicator function.\\nProduce a reliability score. We use the label consistency λ(p i ; x i ) to compute the reliability score of the soft label p i , which is further used as the weight of p i during model learning (see 3.2).\\nConsidering that different entity types may contain different levels of label noise and show different statistics on label consistency, here we present a class-adaptive reliability score for weighting:\\nw(p i ; x i ) = Sigmoid (α (λ(p i ; x i ) -µ(ŷ i ))) ,(3)\\nwhere µ(ŷ i ) denote the mean of all λ(p j ; x j ) where arg max (p j ) = ŷi and x j ∈ D. α > 0 is a hyper-parameter that controls the sharpness of the weighting strategy. If α → 0, all tokens have equal weights. If α → ∞, tokens whose label consistency is larger than the average label consistency w.r.t. its pseudo label will be weighted with 1 and those with smaller consistency will be dropped.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Model Collaboration for Denoising\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Here we elaborate on the details of the two noiserobust training processes. Algorithm 1 depicts the overall training procedure of CoLaDa.\\nNoise-robust training on translation data. Assuming the availability of a collaborator M tgt 2 trained on pseudo-labeled target-language data D tgt , here we focus on leveraging M tgt to reduce the label noise in the translation data D trans = {(x t , y t )}, with which we further deliver a more powerful model M trans . Specifically, given a sentence (x t , y t ) ∈ D trans , we first obtain the soft label p u i of each x t i ∈ x t from the collaborator M tgt . Then, we take both the one hot label y t i and the soft label p u i as the supervision to train the model M trans .3 Denote the output probability distribution of M trans for x t i as pt i . The loss function w.r.t. x t is defined as:\\nL x t = 1 L L i=1 β 1 CE(p t i , p u i ) + β 2 CE(p t i , y t i ) ,(4)\\nwhere CE(•, •) denotes the cross-entropy loss, L is the sentence length, β 1 and β 2 are weighting scalars. Here we further incorporate the instancecollaboration-based denoising strategy (3.1) to provide a token-level reliability evaluation to the supervision from the collaborator M tgt via:\\nβ 1 (x t i ) ← β 1 * w(p u i , x t i ),(5)\\nwhere w(p u i , x t i ) is calculated by Eq. ( 3). Noise-robust training on target-language unlabeled data. Here we leverage M trans obtained via the above noise-robust training on translation data to provide high-quality supervision for Algorithm 1 Pseudo code of CoLaDa. on Dtrans = {(x t , y t )} to get the predictions Dtrans = {(x t , p u )} 5:\\nGet w for (x t , p u ) ∈ Dtrans with\\nM (i-1)\\ntgt , Eq.(3) 6:\\nTrain\\nM (i)\\ntrans with loss on (x t , y t , p u , w), Eq.( 4) 7:\\n# Step 2: Noise-robust training on Dtgt 8:\\nInference\\nM (i) trans on Dtgt = {x u } to get the predic- tions Dtgt = {(x u , p t )} 9: Get w ′ for (x u , p t ) ∈ Dtgt with M (i)\\ntrans , Eq.(3) 10:\\nTrain\\nM (i)\\ntgt with loss on (x u , p t , w ′ ), Eq.( 6) 11: end for Output: an NER model M (T ) tgt .\\nD tgt = {x u }. By performing knowledge distillation on D tgt , the student model M tgt is supposed to benefit from the unlabeled data drawn from the real text distribution in the target language with the knowledge from the teacher model M trans .\\nSpecifically, given a sentence x u ∈ D tgt , we first utilize M trans to predict soft label p t i for each token x u i ∈ x u . Then, we integrate the instancecollaboration-based denoising technique into the learning process. The loss function w.r.t. x u to train the student model M tgt can be formulated as:\\nL x u = 1 L L i=1 w(p t i , x u i ) • CE(p u i , p t i ),(6)\\nwhere pu i denotes the output probability distribution of M tgt for the i-th token x u i and w(p t i , x u i ) is calculated by Eq. (3).\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiment Settings\",\n \"publication_ref\": [\n \"b42\",\n \"b43\",\n \"b40\",\n \"b43\",\n \"b42\",\n \"b49\",\n \"b24\",\n \"b53\"\n ],\n \"table_ref\": [],\n \"text\": \"Datasets We conduct experiments on two standard cross-lingual NER benchmarks: CoNLL (Tjong Kim Sang, 2002;Tjong Kim Sang and De Meulder, 2003) and WikiAnn (Pan et al., 2017). CoNLL contains four languages: English (en) and German (de) from the CoNLL-2003 4 NER shared task (Tjong Kim Sang and De Meulder, 2003), and Spanish (es) and Dutch (nl) from the CoNLL-20025 NER shared task (Tjong Kim Sang, 2002). This dataset is annotated with four entity types: PER, LOC, ORG, and MISC. WikiAnn contains an English dataset and datasets in three non-western languages: Arabic (ar), Hindi (hi), and Chinese (zh). Each dataset is annotated with 3 entity types: PER, LOC, and ORG. All datasets are annotated with the BIO tagging scheme. We use the train, development, and test splits as previous work (Wu and Dredze, 2019;Wu et al., 2020a).\\nWe take English as the source language and other languages as the target language, respectively. We remove the labels of the training data for the target language and take it as the unlabeled target language data. For the CoNLL benchmark, we use the word-to-word translation data provided in UniTrans (Wu et al., 2020a) for a fair comparison. For the WikiAnn benchmark, we translate the source data to the target language with the public M2M100 (Fan et al., 2020) translation system and conduct label projection with the marker-based alignment algorithm as Yang et al. (2022).\\nEvaluation The entity-level micro-F1 on test set of the target language is used as the evaluation metric. We report the mean value of 5 runs with different seeds for all the experiments.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Implementation Details\",\n \"publication_ref\": [\n \"b45\",\n \"b31\",\n \"b49\",\n \"b57\",\n \"b22\",\n \"b36\",\n \"b49\",\n \"b32\"\n ],\n \"table_ref\": [],\n \"text\": \"For the base NER model, we stack a linear classifier with softmax over a base encoder such as mBERT. We implement our framework with Pytorch 1.7.16 , the Hugging-Face transformer library (Wolf et al., 2020), and use FAISS (Johnson et al., 2019) for embedding retrieval. Following Wu and Dredze (2019) and Zhou et al. (2022), we use the multilingual BERT base model (Devlin et al., 2019) and XLM-R (Conneau et al., 2020) large model as our base encoders. Most of our hyper-parameters are set following Wu et al. (2020a). We use AdamW (Loshchilov and Hutter, 2019) as optimizer and train the model on source NER data with the learning rate of 5e-5 for 3 epochs. The dropout rate is 0.1. For teacher-student learning, we train the model with a learning rate of 2e-5 for 10 epochs. We freeze the bottom three layers as Wu and Dredze (2019). Following Keung et al. (2019), we choose other hyper-parameters according to the target language dev set. We set K in Eq. ( 2) to 500 and α in Eq. (3) to 6. For the first iteration, we start with an NER model trained on the source-language data to denoise the translation data with β 1 and β 2 in Eq. ( 5) setting to 0.5. For the following iterations, β 1 is set to 0.9 and β 2 is set to 0.1. The maximum number of iterations is 8. CoLaDa outperforms prior methods with both encoders, achieving a significant improvement of 2.70 F1 scores on average for CoNLL and 10.14 F1 scores on average for WikiAnn with XLM-R as the encoder. This well demonstrates the effectiveness of our approach. Interestingly, CoLaDa shows more significant superiority when transferring to distant target languages in WikiAnn. The knowledge distillation based baselines (i.e., TSL, RIKD, MTMT) struggle on distant languages such as Chinese (zh) due to the noisy predictions from the weak teacher model M src trained in the source language. UniTrans, which is developed with the same data sources as ours, shows poor performance, especially in distant languages such as Arabic (ar). We conjecture that the problem of label noise is even more critical in these distant languages. Our CoLaDa can better handle noise in both translation data and unlabeled target-language data, thus leading to significant performance gains.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Ablation Study\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_3\"\n ],\n \"text\": \"To further validate the effectiveness of each mechanism in the proposed framework, we introduce the following variants of CoLaDa in an ablation study: 1) CoLaDa w/o instance collaboration, where we directly set the reliability score in Eq. ( 3) to 1 for all tokens. 2) CoLaDa w/o translation data denoise, where we set β 1 in Eq. ( 4) to 0. 3) CoLaDa w/o iteratively denoise, where we remove the iterative enhancement and only conduct the denoising process for one iteration. 4) CoLaDa w/o model collaboration, where we set β 1 in Eq. ( 4) to 0, remove the iteration mechanism, and directly take the model finetuned on D trans as the teacher model to train a student model with instance-collaborationbased denoising on D tgt . 5) CoLaDa w/o instance & model collaboration, which further drops the instance-collaboration-based denoising from 4). Table 3 shows the ablation results. We can draw some in-depth conclusions as follows. 1) CoLaDa outperforms CoLaDa w/o instance collaboration, which highlights the effectiveness of leveraging neighborhood information to reduce label noise in knowledge distillation.\\n2) CoLaDa outperforms CoLaDa w/o translation data denoise, which emphasizes the importance of using the collaborator M tgt to refine labels of translation data, especially in distant languages where the translation data is noisier (e.g., 12.8 F1 drop on Arabic and 5.7 F1 drop on Chinese).\\n3) CoLaDa outperforms CoLaDa w/o iteratively denoise, which indicates the necessity of iterative learning: models obtained from the previous iteration should be re-used as the collaborator to further improve label quality in the next iteration.\\n4) CoLaDa w/o instance & model collaboration, which eliminates all denoising strategies from CoLaDa, leads to a significant performance drop, demonstrating the essentiality of label denoising for cross-lingual NER.\"\n },\n {\n \"figure_ref\": [\n \"fig_5\",\n \"fig_5\"\n ],\n \"heading\": \"Analysis of Model Collaboration\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Here we attempt to understand how the two models, i.e., M trans and M tgt , collaboratively improve each other. As shown in Figure 4, F1 scores of M trans and M tgt consistently improve as iterations go on, and finally converge at the last iteration. This indicates that both models benefit from the proposed model collaboration scheme. Two reasons are speculated: i) An improved M tgt can provide more accurate labels on the translation data, which further help to improve M trans via noise-robust learning on such translation data. For example, at the initial step (i = 0), the F1 score of the model M 0 trans trained on the original translation labels is 50.0. With the additional supervision from the collaborator M 0 tgt , M 1 trans achieves a performance gain of 5.7 F1. ii) An improved M trans predicts pseudo labels with higher quality on the target-language unlabeled data, which further benefits the learning of M tgt . As in Figure 4, the quality of pseudo-labeled D tgt (the green line) grows as M trans improves. In this way, both M trans and M tgt are providing more and more reliable labels for each other to learn as the iterations progress.\"\n },\n {\n \"figure_ref\": [\n \"fig_6\"\n ],\n \"heading\": \"Analysis of Instance Collaboration\",\n \"publication_ref\": [\n \"b38\"\n ],\n \"table_ref\": [],\n \"text\": \"This subsection dives into the working mechanism of the instance-collaboration-based denoising.\\nReliability scores v.s. label quality. To study the relationship between reliability score and label quality, we partition tokens in the target-language unlabeled data, x i ∈ D tgt into several bins according to their reliability scores w(p t i , x i ) calculated via M\\n(1) trans . Then, we compute the tokenlevel F1 over each bin by comparing pseudo labels ŷi = arg max(p t i ) to the ground-truth ones. As shown in Figure 5, the label quality is proportional to the reliability score, which well demonstrates the effectiveness of our instance-collaboration-based denoising strategy. Analysis of Label Consistency. We also study the characteristics of label consistency w.r.t. different entity types and representation spaces of the memory bank. Figure 6 shows the results. We can draw some in-depth observations as follows. i) Clean tokens show a larger average consistency than noisy tokens w.r.t. all entity types, demonstrating the effectiveness of our label consistency based denoising strategy again.\\nii) Different entity types lead to different distributions of label consistency, which validates the necessity of our design for class-adaptive reliability score for weighting as Eq.(3).\\niii) Label consistencies calculated with token representations from the upper layers are generally larger than those corresponding to the bottom layers. Also, the label consistency gap between clean tokens and noisy tokens gets larger from the bottom to the top (e.g., the gap between two orange lines). This may be attributed to the fact that representations from upper layers are more taskspecific (Muller et al., 2021), hence they can better discriminate between noisy and clean tokens. Original English:\\nThere have been many tenants, including The Sunday Times[ORG] '' and The Daily News[ORG] ''.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Inaccurate alignment boundary\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The symbols《 and 》are corner brackets used to enclose the name of a newspaper, and other works.\\nOriginal Translation: 有许多租户，包括 《 星期日时报[ORG] 》 和 《 每日新闻[ORG] 》。 Translate-train: 有许多租户，包括 《 星期日时报[ORG] 》 和 《 每日新闻[ORG] 》。 Ours: 有许多租户，包括 《 星期日时报[ORG] 》 和 《 每日新闻[ORG] 》。 Original English: It is found in Democratic Republic of Congo[ORG] , Kenya[LOC] , Tanzania[LOC] ….\"\n },\n {\n \"figure_ref\": [\n \"fig_7\"\n ],\n \"heading\": \"Label noise in English data\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Wrong entity type in original English data for \\\"Democratic Republic of China\\\".\\nOriginal Translation: 它位于刚果民主共和国[ORG] ，肯尼亚[LOC] ，坦桑尼亚[LOC] … Translate-train: 它位于刚果民主共和国[ORG] ，肯尼亚[LOC] ，坦桑尼亚[LOC] … Ours: 它位于刚果民主共和国[LOC] ， 肯尼亚[LOC] ，坦桑尼亚[LOC] …\\nFigure 8: Case study on translation data in Chinese. The blue (red) texts denote the correct (incorrect) entity labels.\\nThe original translation lines display the translation texts and labels obtained by data projection. Translate-train and Ours illustrate the predictions from the translate-train method (M 0 trans ) and our CoLaDa, respectively.\\nChoice of K for neighborhood information.\\nFigure 7 shows the performance of CoLaDa using different K in Eq. ( 2). Generally speaking, CoLaDa is robust to the choice of K. Any value for K > 0 leads to a better performance compared with removing the instance collaboration, i.e., K = 0. A smaller K may lead to a slight performance drop due to limited neighborhood information.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Case Study\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To better illustrate the kinds of label noise presented in the data and the capability of CoLaDa to address such noise, we conduct a case study on the Chinese translation data from the WikiAnn English data. As shown in Figure 8, there are three typical cases of noisy labels in the translation data: noisy labels induced by inaccurate translations, alignment errors, and annotation errors in the original source-language data. 7 Figure 8 shows that the translate-train model, finetuned on the original translation data, overfits the noisy labels. However, CoLaDa is less affected by such noise and makes correct predictions.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b41\",\n \"b44\",\n \"b39\",\n \"b22\",\n \"b21\",\n \"b49\",\n \"b50\",\n \"b32\",\n \"b20\",\n \"b25\",\n \"b26\",\n \"b51\",\n \"b37\",\n \"b35\",\n \"b53\",\n \"b39\",\n \"b19\",\n \"b27\",\n \"b33\",\n \"b20\",\n \"b34\",\n \"b56\",\n \"b57\"\n ],\n \"table_ref\": [],\n \"text\": \"Prior work on cross-lingual NER mainly falls into two major categories: feature-based and data-based transfer.\\nFeature-based These methods learn languageindependent features so that the model trained on the source language can directly adapt to the target language. Earlier work exploits word clusters (Täckström et al., 2012), gazetteers (Zirikly 7 Due to the short entity context information in many sentences in WikiAnn, the translation quality of entity mentions with M2M100 is less than satisfactory on the dataset. and Hagiwara, 2015), Wikifier features (Tsai et al., 2016), and cross-lingual word embedding (Ni et al., 2017), etc. More recently, with the fast growth of multilingual pre-trained language models (Devlin et al., 2019;Conneau et al., 2020) and their promising results on cross-lingual transfer (Wu and Dredze, 2019), lots of studies build upon such pre-trained models and further promote the learning of language-independent features via meta-learning (Wu et al., 2020c), contrastive alignment (Wu and Dredze, 2020), adversarial learning (Keung et al., 2019;Chen et al., 2021), and by integrating other resources (Fetahu et al., 2022). Despite the great success, they mostly ignore language-specific features, which are especially important when transferring to distant languages (Fu et al., 2023).\\nData-based These approaches learn languagespecific features via automatically labeled targetlanguage data and can be further divided into translation-based and knowledge distillation-based methods.\\nTranslation-based methods first translate the source-language data to the target language, then perform label projection from the source side to the target side. Some prior studies have proposed to use cheap translation such as word-to-word (Xie et al., 2018) or phrase-to-phrase (Mayhew et al., 2017) translation. Jain et al. (2019) propose an entity projection algorithm to utilize the Google translation system. Recently, Liu et al. (2021) and Yang et al. (2022) propose to translate sentences with pre-defined markers for label projection. And Ni et al. (2017) design heuristic rules to select highquality translation data. However, both data noise and artifacts (Artetxe et al., 2020) in the transla-tion data still limit the performance of such methods (García-Ferrero et al., 2022).\\nKnowledge distillation-based methods train a student model on unlabeled target-language data with the soft labels from a teacher model (Wu et al., 2020b). Li et al. (2022) improve the single task based teacher-student learning with entity similarity as an auxiliary task. To mitigate the label noise from the teacher model, Chen et al. (2021) propose AdvPicker, which trains a language discriminator to select the less language-dependent unlabeled data for knowledge distillation; Liang et al. (2021) design a reinforcement learning algorithm to train an instance selector according to features such as model confidence to select reliable pseudo labels iteratively.\\nWhile most previous work leverage either translation data or unlabeled data, UniTrans (Wu et al., 2020a) utilizes the model trained on translation data to perform teacher-student learning on unlabeled data. But it still suffers from the data noise problem. More recently, consistency training (Zheng et al., 2021;Zhou et al., 2022) has also been explored to leverage both unlabeled data and translation data without explicit label annotation.\\nTo the best of our knowledge, we are the first to propose a unified denoising framework to handle data noise in both translation and unlabeled data collaboratively from the model and instance levels for cross-lingual NER.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"To address the problem of label noise in crosslingual NER, this paper presents CoLaDa, a collaborative label denoising framework. We propose a model-collaboration-based denoising scheme to make two models trained on different data sources to denoise the labels of each other and hence promote each other's learning. We further propose an instance-collaboration-based strategy that collaboratively considers the label consistency among similar tokens in the feature space to re-weight the noisy labels assigned by a teacher model in knowledge distillation. By integrating the instancecollaboration strategy into the model-collaboration denoising scheme, our final framework CoLada achieves superior performance over prior start-ofthe-art methods by benefiting from better handling the data noise.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Limitations\",\n \"publication_ref\": [\n \"b18\"\n ],\n \"table_ref\": [],\n \"text\": \"Our framework relies on the availability of translation system and unlabeled data in the target language, which can not be applied to languages without any unlabeled text or translation text. The knowledge distillation step requires a certain amount of unlabeled text, while it may struggle in cases where only few hundreds of unlabeled sentences are available. It would be interesting to combine our label denoising framework with data augmentation techniques in such scenarios. Besides, the boarder application to other low-resource languages, such as MasakhaNER 2.0 (Adelani et al., 2022), and other cross-lingual sequence labeling tasks are left for exploration in future work. \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A Appendix\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.1 Dataset Statistics\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.2 Other Implementation Details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"All experiments are conducted on a Tesla V100 (32GB). The total of trainable parameters (M trans and M tgt ) for our model with mBERT-base-cased as the encoder is 172M and the training time is about 35 mins for one iteration. With XLM-Rlarge as our base encoder, the total of trainable parameters are 822M and the training takes about 90 mins for one iteration.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"A.3 Baselines\",\n \"publication_ref\": [\n \"b49\",\n \"b21\",\n \"b50\",\n \"b32\",\n \"b20\",\n \"b35\",\n \"b55\",\n \"b34\",\n \"b33\",\n \"b56\",\n \"b57\"\n ],\n \"table_ref\": [],\n \"text\": \"We consider the following start-of-the-art baselines: mBERT (Wu and Dredze, 2019) and XLM-R (Conneau et al., 2020) directly train an NER model on the labeled data in the source language, with mBERT and XLM-R as the basic encoder, respectively.\\nBERT-align (Wu and Dredze, 2020) tries to explicitly add word-level contrastive alignment loss to enhance the mBERT representation.\\nAdvCE (Keung et al., 2019) exploits adversarial learning on source-and target-language text to avoid learning language-specific information.\\nAdvPicker (Chen et al., 2021) leverages adversarial learning to learn language-shared features and then selects the less language-specific sentences in target-language unlabeled text for knowledge distillation.\\nMulDA (Liu et al., 2021) proposes the labeled sequence translation method for data projection from source-language NER data, a generative model is further applied to augment more diverse examples in the target language.\\nUniTrans (Wu et al., 2020a) unifies model-and translation-data-based-transfer via knowledge distillation.\\nTOF (Zhang et al., 2021) leverages the labeled data for machine reading comprehension task on target language to help the NER task in crosslingual transfer.\\nTSL (Wu et al., 2020b) proposes knowledge distillation to use unlabeled target-language data for cross-lingual NER.\\nRIKD (Liang et al., 2021) proposes a reinforcement learning algorithm to iteratively select reliable pseudo-labels for knowledge distillation.\\nMTMT (Li et al., 2022) proposes multi-task multi-teacher knowledge distillation, which further leverages the entity similarity task.\\nxTune (Zheng et al., 2021) leverages unlabeled translation text and other word-level data augmentation techniques for consistency training.\\nConNER (Zhou et al., 2022) conducts spanlevel consistency training on unlabeled targetlanguage data using translation and further applies dropout-based consistency training on the sourcelanguage data.\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-05-24"},"doi":{"kind":"string","value":"10.18653/v1/2020.emnlp-main.618"},"references":{"kind":"list like","value":[{"authors":"","journal":"Wu and Dredze","ref_id":"b0","title":"mBERT based methods","year":"2019"},{"authors":"( Advce; Keung","journal":"","ref_id":"b1","title":"","year":"2019"},{"authors":" Tsl (wu","journal":"","ref_id":"b2","title":"","year":"2020"},{"authors":"( Unitrans; Wu","journal":"","ref_id":"b3","title":"","year":"2020"},{"authors":" Zhang","journal":"","ref_id":"b4","title":"","year":"2021"},{"authors":" Advpicker (chen","journal":"","ref_id":"b5","title":"","year":"2021"},{"authors":" Rikd (liang","journal":"","ref_id":"b6","title":"","year":"2021"},{"authors":" Mtmt (li","journal":"CoLaDa","ref_id":"b7","title":"","year":"2022"},{"authors":" Liu","journal":"XLM-R based methods","ref_id":"b8","title":"","year":"2021"},{"authors":"Zheng ","journal":"","ref_id":"b9","title":"","year":"2021"},{"authors":"( Conner; Zhou","journal":"","ref_id":"b10","title":"CoNLL. Method ar hi zh avg mBERT based methods: BERT-align (Wu and Dredze","year":"2020"},{"authors":" Tsl (wu","journal":"","ref_id":"b11","title":"","year":"2020"},{"authors":" Rikd (liang","journal":"","ref_id":"b12","title":"","year":"2021"},{"authors":" Mtmt (li","journal":"","ref_id":"b13","title":"","year":"2022"},{"authors":"† ( Unitrans; Wu","journal":"CoLaDa","ref_id":"b14","title":"","year":"2020"},{"authors":"","journal":"XLM-R based methods","ref_id":"b15","title":"","year":"2020"},{"authors":"( Conner; Zhou","journal":"CoLaDa","ref_id":"b16","title":"","year":"2022"},{"authors":"","journal":"","ref_id":"b17","title":"start-of-the-art baselines as follows: i) feature alignment based methods: mBERT","year":"2019"},{"authors":"David Adelani; Graham Neubig; Sebastian Ruder; Shruti Rijhwani; Michael Beukman; Chester Palen-Michel; Constantine Lignos; Jesujoba Alabi; Shamsuddeen Muhammad; Peter Nabende; M Cheikh; Andiswa Bamba Dione; Rooweither Bukula; Mabuya; F P Bonaventure; Blessing Dossou; Happy Sibanda; Jonathan Buzaaba; Godson Mukiibi; Derguene Kalipe; Amelia Mbaye; Fatoumata Taylor; Chris Kabore; Anuoluwapo Chinenye Emezue; Perez Aremu; Catherine Ogayo; Edwin Gitau; Victoire Munkoh-Buabeng; Memdjokam Koagne; Auguste Allahsera; Tebogo Tapo; Vukosi Macucwa; Mboning Marivate; Tajuddeen Tchiaze Elvis; Tosin Gwadabe; Orevaoghene Adewumi; Joyce Ahia; Neo Nakatumba-Nabende; Ignatius Lerato Mokono; Chiamaka Ezeani; Chukwuneke; Oluwaseun Mofetoluwa; Gilles Adeyemi; Idris Quentin Hacheme; Odunayo Abdulmumin; Oreen Ogundepo; Tatiana Yousuf; Dietrich Moteu; Klakow","journal":"Association for Computational Linguistics","ref_id":"b18","title":"MasakhaNER 2.0: Africa-centric transfer learning for named entity recognition","year":"2022"},{"authors":"Mikel Artetxe; Gorka Labaka; Eneko Agirre","journal":"Association for Computational Linguistics","ref_id":"b19","title":"Translation artifacts in cross-lingual transfer learning","year":"2020"},{"authors":"Weile Chen; Huiqiang Jiang; Qianhui Wu; Börje F Karlsson; Yi Guan","journal":"Association for Computational Linguistics","ref_id":"b20","title":"AdvPicker: Effectively Leveraging Unlabeled Data via Adversarial Discriminator for Cross-Lingual NER","year":"2021"},{"authors":"Alexis Conneau; Kartikay Khandelwal; Naman Goyal; Vishrav Chaudhary; Guillaume Wenzek; Francisco Guzmán; Edouard Grave; Myle Ott; Luke Zettlemoyer; Veselin Stoyanov","journal":"Association for Computational Linguistics","ref_id":"b21","title":"Unsupervised cross-lingual representation learning at scale","year":"2020"},{"authors":"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova","journal":"Association for Computational Linguistics","ref_id":"b22","title":"BERT: Pre-training of deep bidirectional transformers for language understanding","year":"2019"},{"authors":"Maud Ehrmann; Marco Turchi; Ralf Steinberger","journal":"Association for Computational Linguistics","ref_id":"b23","title":"Building a multilingual named entityannotated corpus using annotation projection","year":"2011"},{"authors":"Angela Fan; Shruti Bhosale; Holger Schwenk; Zhiyi Ma; Ahmed El-Kishky; Siddharth Goyal; Mandeep Baines; Onur Çelebi; Guillaume Wenzek; Vishrav Chaudhary; Naman Goyal; Tom Birch; Vitaliy Liptchinsky; Sergey Edunov; Edouard Grave; Michael Auli; Armand Joulin","journal":"J. Mach. Learn. Res","ref_id":"b24","title":"Beyond english-centric multilingual machine translation","year":"2020"},{"authors":"Besnik Fetahu; Anjie Fang; Oleg Rokhlenko; Shervin Malmasi","journal":"Association for Computational Linguistics","ref_id":"b25","title":"Dynamic gazetteer integration in multilingual models for cross-lingual and cross-domain named entity recognition","year":"2022"},{"authors":"Yingwen Fu; Nankai Lin; Boyu Chen; Ziyu Yang; Shengyi Jiang","journal":"IEEE/ACM Transactions on Audio, Speech, and Language Processing","ref_id":"b26","title":"Cross-lingual named entity recognition for heterogenous languages","year":"2023"},{"authors":"Iker García-Ferrero; Rodrigo Agerri; German Rigau","journal":"","ref_id":"b27","title":"Model and data transfer for cross-lingual sequence labelling in zero-resource settings","year":"2022"},{"authors":"Xiaoyi Gu; Leman Akoglu; Alessandro Rinaldo","journal":"","ref_id":"b28","title":"Statistical analysis of nearest neighbor methods for anomaly detection","year":"2019-12-08"},{"authors":"Lifu Huang; Ji Heng; Jonathan May","journal":"Association for Computational Linguistics","ref_id":"b29","title":"Crosslingual multi-level adversarial transfer to enhance low-resource name tagging","year":"2019"},{"authors":"Alankar Jain; Bhargavi Paranjape; Zachary C ","journal":"Association for Computational Linguistics","ref_id":"b30","title":"Lipton. 2019. Entity projection via machine translation for cross-lingual NER","year":""},{"authors":"Jeff Johnson; Matthijs Douze; Hervé Jégou","journal":"IEEE Transactions on Big Data","ref_id":"b31","title":"Billion-scale similarity search with GPUs","year":"2019"},{"authors":"Phillip Keung; Yichao Lu; Vikas Bhardwaj","journal":"Association for Computational Linguistics","ref_id":"b32","title":"Adversarial learning with contextual embeddings for zero-resource cross-lingual classification and NER","year":"2019"},{"authors":"Zhuoran Li; Chunming Hu; Xiaohui Guo; Junfan Chen; Wenyi Qin; Richong Zhang","journal":"Association for Computational Linguistics","ref_id":"b33","title":"An unsupervised multiple-task and multiple-teacher model for cross-lingual named entity recognition","year":"2022"},{"authors":"Shining Liang; Ming Gong; Jian Pei; Linjun Shou; Wanli Zuo; Xianglin Zuo; Daxin Jiang","journal":"Association for Computing Machinery","ref_id":"b34","title":"Reinforced iterative knowledge distillation for crosslingual named entity recognition","year":"2021"},{"authors":"Linlin Liu; Bosheng Ding; Lidong Bing; Shafiq Joty; Luo Si; Chunyan Miao","journal":"","ref_id":"b35","title":"MulDA: A multilingual data augmentation framework for lowresource cross-lingual NER","year":"2021"},{"authors":"Ilya Loshchilov; Frank Hutter","journal":"","ref_id":"b36","title":"Decoupled weight decay regularization","year":"2019-05-06"},{"authors":"Stephen Mayhew; Chen-Tse Tsai; Dan Roth","journal":"Association for Computational Linguistics","ref_id":"b37","title":"Cheap translation for cross-lingual named entity recognition","year":"2017"},{"authors":"Benjamin Muller; Yanai Elazar; Benoît Sagot; Djamé Seddah","journal":"Association for Computational Linguistics","ref_id":"b38","title":"First align, then predict: Understanding the cross-lingual ability of multilingual BERT","year":"2021"},{"authors":"Jian Ni; Georgiana Dinu; Radu Florian","journal":"Association for Computational Linguistics","ref_id":"b39","title":"Weakly supervised cross-lingual named entity recognition via effective annotation and representation projection","year":"2017"},{"authors":"Xiaoman Pan; Boliang Zhang; Jonathan May; Joel Nothman; Kevin Knight; Heng Ji","journal":"","ref_id":"b40","title":"Cross-lingual name tagging and linking for 282 languages","year":"2017"},{"authors":"Oscar Täckström; Ryan Mcdonald; Jakob Uszkoreit","journal":"Association for Computational Linguistics","ref_id":"b41","title":"Cross-lingual word clusters for direct transfer of linguistic structure","year":"2012"},{"authors":"Erik F Tjong; Kim Sang","journal":"","ref_id":"b42","title":"Introduction to the CoNLL-2002 shared task: Language-independent named entity recognition","year":"2002"},{"authors":"Erik F Tjong; Kim Sang; Fien De; Meulder ","journal":"","ref_id":"b43","title":"Introduction to the conll-2003 shared task: Languageindependent named entity recognition","year":"2003"},{"authors":"Chen-Tse Tsai; Stephen Mayhew; Dan Roth","journal":"Association for Computational Linguistics","ref_id":"b44","title":"Cross-lingual named entity recognition via wikification","year":"2016"},{"authors":"Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Remi Louf; Morgan Funtowicz; Joe Davison; Sam Shleifer; Clara Patrick Von Platen; Yacine Ma; Julien Jernite; Canwen Plu; Teven Xu; Sylvain Le Scao; Mariama Gugger; Quentin Drame; Alexander Lhoest; Rush","journal":"Association for Computational Linguistics","ref_id":"b45","title":"Transformers: State-of-the-art natural language processing","year":"2020"},{"authors":"Qianhui Wu; Zijia Lin; Börje F Karlsson; Biqing Huang; Jianguang Lou; ; ","journal":"","ref_id":"b46","title":"Unitrans : Unifying model transfer and data transfer for cross-lingual named entity recognition with unlabeled data","year":"2020"},{"authors":"Qianhui Wu; Zijia Lin; Börje F Karlsson; Jian-Guang Lou; Biqing Huang","journal":"Association for Computational Linguistics","ref_id":"b47","title":"Single-/multi-source cross-lingual NER via teacher-student learning on unlabeled data in target language","year":"2020"},{"authors":"Qianhui Wu; Zijia Lin; Guoxin Wang; Hui Chen; Börje F Karlsson; Biqing Huang; Chin-Yew Lin","journal":"AAAI Press","ref_id":"b48","title":"Enhanced meta-learning for cross-lingual named entity recognition with minimal resources","year":"2020-02-07"},{"authors":"Shijie Wu; Mark Dredze","journal":"Association for Computational Linguistics","ref_id":"b49","title":"Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT","year":"2019"},{"authors":"Shijie Wu; Mark Dredze","journal":"Association for Computational Linguistics","ref_id":"b50","title":"Do explicit alignments robustly improve multilingual encoders","year":"2020"},{"authors":"Jiateng Xie; Zhilin Yang; Graham Neubig; Noah A Smith; Jaime Carbonell","journal":"Association for Computational Linguistics","ref_id":"b51","title":"Neural crosslingual named entity recognition with minimal resources","year":"2018"},{"authors":"Hong Xu; Keqing He; Yuanmeng Yan; Sihong Liu; Zijun Liu; Weiran Xu","journal":"International Committee on Computational Linguistics","ref_id":"b52","title":"A deep generative distance-based classifier for out-of-domain detection with mahalanobis space","year":"2020"},{"authors":"Jian Yang; Shaohan Huang; Shuming Ma; Yuwei Yin; Li Dong; Dongdong Zhang; Hongcheng Guo; Zhoujun Li; Furu Wei","journal":"","ref_id":"b53","title":"Crop: Zero-shot crosslingual named entity recognition with multilingual labeled sequence translation","year":"2022"},{"authors":"Andrew Zhai; Hao-Yu Wu","journal":"BMVA Press","ref_id":"b54","title":"Classification is a strong baseline for deep metric learning","year":"2019-09-09"},{"authors":"Ying Zhang; Fandong Meng; Yufeng Chen; Jinan Xu; Jie Zhou","journal":"Association for Computational Linguistics","ref_id":"b55","title":"Target-oriented fine-tuning for zero-resource named entity recognition","year":"2021"},{"authors":"Bo Zheng; Li Dong; Shaohan Huang; Wenhui Wang; Zewen Chi; Saksham Singhal; Wanxiang Che; Ting Liu; Xia Song; Furu Wei","journal":"Association for Computational Linguistics","ref_id":"b56","title":"Consistency regularization for cross-lingual fine-tuning","year":"2021"},{"authors":"Ran Zhou; Xin Li; Lidong Bing; Erik Cambria; Luo Si; Chunyan Miao","journal":"","ref_id":"b57","title":"ConNER: Consistency training for cross-lingual named entity recognition","year":"2022"},{"authors":"Ayah Zirikly; Masato Hagiwara","journal":"Association for Computational Linguistics","ref_id":"b58","title":"Cross-lingual transfer of named entity recognizers without parallel corpora","year":"2015"}],"string":"[\n {\n \"authors\": \"\",\n \"journal\": \"Wu and Dredze\",\n \"ref_id\": \"b0\",\n \"title\": \"mBERT based methods\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"( Advce; Keung\",\n \"journal\": \"\",\n \"ref_id\": \"b1\",\n \"title\": \"\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \" Tsl (wu\",\n \"journal\": \"\",\n \"ref_id\": \"b2\",\n \"title\": \"\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"( Unitrans; Wu\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \" Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b4\",\n \"title\": \"\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \" Advpicker (chen\",\n \"journal\": \"\",\n \"ref_id\": \"b5\",\n \"title\": \"\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \" Rikd (liang\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \" Mtmt (li\",\n \"journal\": \"CoLaDa\",\n \"ref_id\": \"b7\",\n \"title\": \"\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \" Liu\",\n \"journal\": \"XLM-R based methods\",\n \"ref_id\": \"b8\",\n \"title\": \"\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Zheng \",\n \"journal\": \"\",\n \"ref_id\": \"b9\",\n \"title\": \"\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"( Conner; Zhou\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"CoNLL. Method ar hi zh avg mBERT based methods: BERT-align (Wu and Dredze\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \" Tsl (wu\",\n \"journal\": \"\",\n \"ref_id\": \"b11\",\n \"title\": \"\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \" Rikd (liang\",\n \"journal\": \"\",\n \"ref_id\": \"b12\",\n \"title\": \"\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \" Mtmt (li\",\n \"journal\": \"\",\n \"ref_id\": \"b13\",\n \"title\": \"\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"† ( Unitrans; Wu\",\n \"journal\": \"CoLaDa\",\n \"ref_id\": \"b14\",\n \"title\": \"\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"XLM-R based methods\",\n \"ref_id\": \"b15\",\n \"title\": \"\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"( Conner; Zhou\",\n \"journal\": \"CoLaDa\",\n \"ref_id\": \"b16\",\n \"title\": \"\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"start-of-the-art baselines as follows: i) feature alignment based methods: mBERT\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"David Adelani; Graham Neubig; Sebastian Ruder; Shruti Rijhwani; Michael Beukman; Chester Palen-Michel; Constantine Lignos; Jesujoba Alabi; Shamsuddeen Muhammad; Peter Nabende; M Cheikh; Andiswa Bamba Dione; Rooweither Bukula; Mabuya; F P Bonaventure; Blessing Dossou; Happy Sibanda; Jonathan Buzaaba; Godson Mukiibi; Derguene Kalipe; Amelia Mbaye; Fatoumata Taylor; Chris Kabore; Anuoluwapo Chinenye Emezue; Perez Aremu; Catherine Ogayo; Edwin Gitau; Victoire Munkoh-Buabeng; Memdjokam Koagne; Auguste Allahsera; Tebogo Tapo; Vukosi Macucwa; Mboning Marivate; Tajuddeen Tchiaze Elvis; Tosin Gwadabe; Orevaoghene Adewumi; Joyce Ahia; Neo Nakatumba-Nabende; Ignatius Lerato Mokono; Chiamaka Ezeani; Chukwuneke; Oluwaseun Mofetoluwa; Gilles Adeyemi; Idris Quentin Hacheme; Odunayo Abdulmumin; Oreen Ogundepo; Tatiana Yousuf; Dietrich Moteu; Klakow\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b18\",\n \"title\": \"MasakhaNER 2.0: Africa-centric transfer learning for named entity recognition\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Mikel Artetxe; Gorka Labaka; Eneko Agirre\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b19\",\n \"title\": \"Translation artifacts in cross-lingual transfer learning\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Weile Chen; Huiqiang Jiang; Qianhui Wu; Börje F Karlsson; Yi Guan\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b20\",\n \"title\": \"AdvPicker: Effectively Leveraging Unlabeled Data via Adversarial Discriminator for Cross-Lingual NER\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Alexis Conneau; Kartikay Khandelwal; Naman Goyal; Vishrav Chaudhary; Guillaume Wenzek; Francisco Guzmán; Edouard Grave; Myle Ott; Luke Zettlemoyer; Veselin Stoyanov\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b21\",\n \"title\": \"Unsupervised cross-lingual representation learning at scale\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jacob Devlin; Ming-Wei Chang; Kenton Lee; Kristina Toutanova\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b22\",\n \"title\": \"BERT: Pre-training of deep bidirectional transformers for language understanding\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Maud Ehrmann; Marco Turchi; Ralf Steinberger\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b23\",\n \"title\": \"Building a multilingual named entityannotated corpus using annotation projection\",\n \"year\": \"2011\"\n },\n {\n \"authors\": \"Angela Fan; Shruti Bhosale; Holger Schwenk; Zhiyi Ma; Ahmed El-Kishky; Siddharth Goyal; Mandeep Baines; Onur Çelebi; Guillaume Wenzek; Vishrav Chaudhary; Naman Goyal; Tom Birch; Vitaliy Liptchinsky; Sergey Edunov; Edouard Grave; Michael Auli; Armand Joulin\",\n \"journal\": \"J. Mach. Learn. Res\",\n \"ref_id\": \"b24\",\n \"title\": \"Beyond english-centric multilingual machine translation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Besnik Fetahu; Anjie Fang; Oleg Rokhlenko; Shervin Malmasi\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b25\",\n \"title\": \"Dynamic gazetteer integration in multilingual models for cross-lingual and cross-domain named entity recognition\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Yingwen Fu; Nankai Lin; Boyu Chen; Ziyu Yang; Shengyi Jiang\",\n \"journal\": \"IEEE/ACM Transactions on Audio, Speech, and Language Processing\",\n \"ref_id\": \"b26\",\n \"title\": \"Cross-lingual named entity recognition for heterogenous languages\",\n \"year\": \"2023\"\n },\n {\n \"authors\": \"Iker García-Ferrero; Rodrigo Agerri; German Rigau\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Model and data transfer for cross-lingual sequence labelling in zero-resource settings\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Xiaoyi Gu; Leman Akoglu; Alessandro Rinaldo\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Statistical analysis of nearest neighbor methods for anomaly detection\",\n \"year\": \"2019-12-08\"\n },\n {\n \"authors\": \"Lifu Huang; Ji Heng; Jonathan May\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b29\",\n \"title\": \"Crosslingual multi-level adversarial transfer to enhance low-resource name tagging\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Alankar Jain; Bhargavi Paranjape; Zachary C \",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b30\",\n \"title\": \"Lipton. 2019. Entity projection via machine translation for cross-lingual NER\",\n \"year\": \"\"\n },\n {\n \"authors\": \"Jeff Johnson; Matthijs Douze; Hervé Jégou\",\n \"journal\": \"IEEE Transactions on Big Data\",\n \"ref_id\": \"b31\",\n \"title\": \"Billion-scale similarity search with GPUs\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Phillip Keung; Yichao Lu; Vikas Bhardwaj\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b32\",\n \"title\": \"Adversarial learning with contextual embeddings for zero-resource cross-lingual classification and NER\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Zhuoran Li; Chunming Hu; Xiaohui Guo; Junfan Chen; Wenyi Qin; Richong Zhang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b33\",\n \"title\": \"An unsupervised multiple-task and multiple-teacher model for cross-lingual named entity recognition\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Shining Liang; Ming Gong; Jian Pei; Linjun Shou; Wanli Zuo; Xianglin Zuo; Daxin Jiang\",\n \"journal\": \"Association for Computing Machinery\",\n \"ref_id\": \"b34\",\n \"title\": \"Reinforced iterative knowledge distillation for crosslingual named entity recognition\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Linlin Liu; Bosheng Ding; Lidong Bing; Shafiq Joty; Luo Si; Chunyan Miao\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"MulDA: A multilingual data augmentation framework for lowresource cross-lingual NER\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ilya Loshchilov; Frank Hutter\",\n \"journal\": \"\",\n \"ref_id\": \"b36\",\n \"title\": \"Decoupled weight decay regularization\",\n \"year\": \"2019-05-06\"\n },\n {\n \"authors\": \"Stephen Mayhew; Chen-Tse Tsai; Dan Roth\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b37\",\n \"title\": \"Cheap translation for cross-lingual named entity recognition\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Benjamin Muller; Yanai Elazar; Benoît Sagot; Djamé Seddah\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b38\",\n \"title\": \"First align, then predict: Understanding the cross-lingual ability of multilingual BERT\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Jian Ni; Georgiana Dinu; Radu Florian\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b39\",\n \"title\": \"Weakly supervised cross-lingual named entity recognition via effective annotation and representation projection\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Xiaoman Pan; Boliang Zhang; Jonathan May; Joel Nothman; Kevin Knight; Heng Ji\",\n \"journal\": \"\",\n \"ref_id\": \"b40\",\n \"title\": \"Cross-lingual name tagging and linking for 282 languages\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Oscar Täckström; Ryan Mcdonald; Jakob Uszkoreit\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b41\",\n \"title\": \"Cross-lingual word clusters for direct transfer of linguistic structure\",\n \"year\": \"2012\"\n },\n {\n \"authors\": \"Erik F Tjong; Kim Sang\",\n \"journal\": \"\",\n \"ref_id\": \"b42\",\n \"title\": \"Introduction to the CoNLL-2002 shared task: Language-independent named entity recognition\",\n \"year\": \"2002\"\n },\n {\n \"authors\": \"Erik F Tjong; Kim Sang; Fien De; Meulder \",\n \"journal\": \"\",\n \"ref_id\": \"b43\",\n \"title\": \"Introduction to the conll-2003 shared task: Languageindependent named entity recognition\",\n \"year\": \"2003\"\n },\n {\n \"authors\": \"Chen-Tse Tsai; Stephen Mayhew; Dan Roth\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b44\",\n \"title\": \"Cross-lingual named entity recognition via wikification\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"Thomas Wolf; Lysandre Debut; Victor Sanh; Julien Chaumond; Clement Delangue; Anthony Moi; Pierric Cistac; Tim Rault; Remi Louf; Morgan Funtowicz; Joe Davison; Sam Shleifer; Clara Patrick Von Platen; Yacine Ma; Julien Jernite; Canwen Plu; Teven Xu; Sylvain Le Scao; Mariama Gugger; Quentin Drame; Alexander Lhoest; Rush\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b45\",\n \"title\": \"Transformers: State-of-the-art natural language processing\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Qianhui Wu; Zijia Lin; Börje F Karlsson; Biqing Huang; Jianguang Lou; ; \",\n \"journal\": \"\",\n \"ref_id\": \"b46\",\n \"title\": \"Unitrans : Unifying model transfer and data transfer for cross-lingual named entity recognition with unlabeled data\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Qianhui Wu; Zijia Lin; Börje F Karlsson; Jian-Guang Lou; Biqing Huang\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b47\",\n \"title\": \"Single-/multi-source cross-lingual NER via teacher-student learning on unlabeled data in target language\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Qianhui Wu; Zijia Lin; Guoxin Wang; Hui Chen; Börje F Karlsson; Biqing Huang; Chin-Yew Lin\",\n \"journal\": \"AAAI Press\",\n \"ref_id\": \"b48\",\n \"title\": \"Enhanced meta-learning for cross-lingual named entity recognition with minimal resources\",\n \"year\": \"2020-02-07\"\n },\n {\n \"authors\": \"Shijie Wu; Mark Dredze\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b49\",\n \"title\": \"Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Shijie Wu; Mark Dredze\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b50\",\n \"title\": \"Do explicit alignments robustly improve multilingual encoders\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jiateng Xie; Zhilin Yang; Graham Neubig; Noah A Smith; Jaime Carbonell\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b51\",\n \"title\": \"Neural crosslingual named entity recognition with minimal resources\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Hong Xu; Keqing He; Yuanmeng Yan; Sihong Liu; Zijun Liu; Weiran Xu\",\n \"journal\": \"International Committee on Computational Linguistics\",\n \"ref_id\": \"b52\",\n \"title\": \"A deep generative distance-based classifier for out-of-domain detection with mahalanobis space\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Jian Yang; Shaohan Huang; Shuming Ma; Yuwei Yin; Li Dong; Dongdong Zhang; Hongcheng Guo; Zhoujun Li; Furu Wei\",\n \"journal\": \"\",\n \"ref_id\": \"b53\",\n \"title\": \"Crop: Zero-shot crosslingual named entity recognition with multilingual labeled sequence translation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Andrew Zhai; Hao-Yu Wu\",\n \"journal\": \"BMVA Press\",\n \"ref_id\": \"b54\",\n \"title\": \"Classification is a strong baseline for deep metric learning\",\n \"year\": \"2019-09-09\"\n },\n {\n \"authors\": \"Ying Zhang; Fandong Meng; Yufeng Chen; Jinan Xu; Jie Zhou\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b55\",\n \"title\": \"Target-oriented fine-tuning for zero-resource named entity recognition\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Bo Zheng; Li Dong; Shaohan Huang; Wenhui Wang; Zewen Chi; Saksham Singhal; Wanxiang Che; Ting Liu; Xia Song; Furu Wei\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b56\",\n \"title\": \"Consistency regularization for cross-lingual fine-tuning\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Ran Zhou; Xin Li; Lidong Bing; Erik Cambria; Luo Si; Chunyan Miao\",\n \"journal\": \"\",\n \"ref_id\": \"b57\",\n \"title\": \"ConNER: Consistency training for cross-lingual named entity recognition\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Ayah Zirikly; Masato Hagiwara\",\n \"journal\": \"Association for Computational Linguistics\",\n \"ref_id\": \"b58\",\n \"title\": \"Cross-lingual transfer of named entity recognizers without parallel corpora\",\n \"year\": \"2015\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[3,70.87,358.11,454.27,417.51],"formula_id":"formula_0","formula_text":"den representations h = {h i } L i=1 of each sentence x = {x i } L i=1 ∈ D: h = F(x).(1)"},{"formula_coordinates":[3,306.14,503.58,219,60.58],"formula_id":"formula_1","formula_text":"x i in N K (x i ): λ(p i ; x i ) = 1 K x j ∈N k (x i ) I(ŷ j = ŷi ),(2)"},{"formula_coordinates":[3,315.66,745.96,209.48,23.36],"formula_id":"formula_2","formula_text":"w(p i ; x i ) = Sigmoid (α (λ(p i ; x i ) -µ(ŷ i ))) ,(3)"},{"formula_coordinates":[4,78.25,457.21,211.62,44.01],"formula_id":"formula_3","formula_text":"L x t = 1 L L i=1 β 1 CE(p t i , p u i ) + β 2 CE(p t i , y t i ) ,(4)"},{"formula_coordinates":[4,123.75,594.81,166.12,14.19],"formula_id":"formula_4","formula_text":"β 1 (x t i ) ← β 1 * w(p u i , x t i ),(5)"},{"formula_coordinates":[4,461.99,166.39,28.85,10.4],"formula_id":"formula_5","formula_text":"M (i-1)"},{"formula_coordinates":[4,356.18,179.71,19.44,10.4],"formula_id":"formula_6","formula_text":"M (i)"},{"formula_coordinates":[4,309.93,201.9,215.97,34.89],"formula_id":"formula_7","formula_text":"M (i) trans on Dtgt = {x u } to get the predic- tions Dtgt = {(x u , p t )} 9: Get w ′ for (x u , p t ) ∈ Dtgt with M (i)"},{"formula_coordinates":[4,356.88,238.49,19.44,10.4],"formula_id":"formula_8","formula_text":"M (i)"},{"formula_coordinates":[4,334.03,458.95,191.11,33.71],"formula_id":"formula_9","formula_text":"L x u = 1 L L i=1 w(p t i , x u i ) • CE(p u i , p t i ),(6)"},{"formula_coordinates":[8,82.53,131.39,287.08,41.64],"formula_id":"formula_10","formula_text":"Original Translation: 有许多租户，包括 《 星期日时报[ORG] 》 和 《 每日新闻[ORG] 》。 Translate-train: 有许多租户，包括 《 星期日时报[ORG] 》 和 《 每日新闻[ORG] 》。 Ours: 有许多租户，包括 《 星期日时报[ORG] 》 和 《 每日新闻[ORG] 》。 Original English: It is found in Democratic Republic of Congo[ORG] , Kenya[LOC] , Tanzania[LOC] …."},{"formula_coordinates":[8,82.53,176.86,247.05,30.26],"formula_id":"formula_11","formula_text":"Original Translation: 它位于刚果民主共和国[ORG] ，肯尼亚[LOC] ，坦桑尼亚[LOC] … Translate-train: 它位于刚果民主共和国[ORG] ，肯尼亚[LOC] ，坦桑尼亚[LOC] … Ours: 它位于刚果民主共和国[LOC] ， 肯尼亚[LOC] ，坦桑尼亚[LOC] …"}],"string":"[\n {\n \"formula_coordinates\": [\n 3,\n 70.87,\n 358.11,\n 454.27,\n 417.51\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \"den representations h = {h i } L i=1 of each sentence x = {x i } L i=1 ∈ D: h = F(x).(1)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 306.14,\n 503.58,\n 219,\n 60.58\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"x i in N K (x i ): λ(p i ; x i ) = 1 K x j ∈N k (x i ) I(ŷ j = ŷi ),(2)\"\n },\n {\n \"formula_coordinates\": [\n 3,\n 315.66,\n 745.96,\n 209.48,\n 23.36\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"w(p i ; x i ) = Sigmoid (α (λ(p i ; x i ) -µ(ŷ i ))) ,(3)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 78.25,\n 457.21,\n 211.62,\n 44.01\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"L x t = 1 L L i=1 β 1 CE(p t i , p u i ) + β 2 CE(p t i , y t i ) ,(4)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 123.75,\n 594.81,\n 166.12,\n 14.19\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"β 1 (x t i ) ← β 1 * w(p u i , x t i ),(5)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 461.99,\n 166.39,\n 28.85,\n 10.4\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"M (i-1)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 356.18,\n 179.71,\n 19.44,\n 10.4\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"M (i)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 309.93,\n 201.9,\n 215.97,\n 34.89\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"M (i) trans on Dtgt = {x u } to get the predic- tions Dtgt = {(x u , p t )} 9: Get w ′ for (x u , p t ) ∈ Dtgt with M (i)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 356.88,\n 238.49,\n 19.44,\n 10.4\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"M (i)\"\n },\n {\n \"formula_coordinates\": [\n 4,\n 334.03,\n 458.95,\n 191.11,\n 33.71\n ],\n \"formula_id\": \"formula_9\",\n \"formula_text\": \"L x u = 1 L L i=1 w(p t i , x u i ) • CE(p u i , p t i ),(6)\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 82.53,\n 131.39,\n 287.08,\n 41.64\n ],\n \"formula_id\": \"formula_10\",\n \"formula_text\": \"Original Translation: 有许多租户，包括 《 星期日时报[ORG] 》 和 《 每日新闻[ORG] 》。 Translate-train: 有许多租户，包括 《 星期日时报[ORG] 》 和 《 每日新闻[ORG] 》。 Ours: 有许多租户，包括 《 星期日时报[ORG] 》 和 《 每日新闻[ORG] 》。 Original English: It is found in Democratic Republic of Congo[ORG] , Kenya[LOC] , Tanzania[LOC] ….\"\n },\n {\n \"formula_coordinates\": [\n 8,\n 82.53,\n 176.86,\n 247.05,\n 30.26\n ],\n \"formula_id\": \"formula_11\",\n \"formula_text\": \"Original Translation: 它位于刚果民主共和国[ORG] ，肯尼亚[LOC] ，坦桑尼亚[LOC] … Translate-train: 它位于刚果民主共和国[ORG] ，肯尼亚[LOC] ，坦桑尼亚[LOC] … Ours: 它位于刚果民主共和国[LOC] ， 肯尼亚[LOC] ，坦桑尼亚[LOC] …\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"CoLaDa: A Collaborative Label Denoising Framework for Cross-lingual Named Entity Recognition"},"abstract":{"kind":"string","value":"Cross-lingual named entity recognition (NER) aims to train an NER system that generalizes well to a target language by leveraging labeled data in a given source language. Previous work alleviates the data scarcity problem by translating source-language labeled data or performing knowledge distillation on target-language unlabeled data. However, these methods may suffer from label noise due to the automatic labeling process. In this paper, we propose CoLaDa, a Collaborative Label Denoising Framework, to address this problem. Specifically, we first explore a model-collaboration-based denoising scheme that enables models trained on different data sources to collaboratively denoise pseudo labels used by each other. We then present an instance-collaboration-based strategy that considers the label consistency of each token's neighborhood in the representation space for denoising. Experiments on different benchmark datasets show that the proposed CoLaDa achieves superior results compared to previous methods, especially when generalizing to distant languages."},"authors":{"kind":"string","value":"Tingting Ma; Qianhui Wu; Huiqiang Jiang; Börje F Karlsson; Tiejun Zhao; Chin-Yew Lin"},"figures":{"kind":"list like","value":[{"figure_caption":"Figure 1 :1Figure 1: Comparison between previous methods (a/b/c) and our CoLaDa at the i-th iteration (d) CoLaDa starts at M 0 tgt and performs denoising iteratively. D src : Sourcelanguage labeled data. D trans : Translation data. D tgt : Target-language unlabeled data with pseudo-labels generated by NER models. M src/trans/tgt : NER model learned on D src/trans/tgt .","figure_data":"","figure_id":"fig_1","figure_label":"1","figure_type":"figure"},{"figure_caption":"Figure 2 :2Figure 2: Illustration of the instance collaboration for denoising. Different colors depict different entity types.","figure_data":"","figure_id":"fig_2","figure_label":"2","figure_type":"figure"},{"figure_caption":"Figure 33Figure 3 depicts an overview of the CoLaDa framework. It is an iterative model-collaboration-based denoising framework which consists of two steps: noise-robust learning on translation data and noiserobust learning on unlabeled target-language data. An instance-collaboration-based denoising strategy (Sec 3.1) is then integrated into the modelcollaboration-based denoising procedure (Sec 3.2).","figure_data":"","figure_id":"fig_3","figure_label":"3","figure_type":"figure"},{"figure_caption":"Input: an NER model Msrc trained on Dsrc, translation data Dtrans, the unlabeled data Dtgt, the maximum iteration T.1: M (0)tgt ← Msrc ▷ Initialization 2: for i = 1, 2, . . ., T do 3:# Step 1: Noise-robust training on Dtrans 4","figure_data":"","figure_id":"fig_4","figure_label":"","figure_type":"figure"},{"figure_caption":"Figure 4 :4Figure 4: F1 scores of M (i) trans and M (i) tgt on the test data and the target-language unlabeled data. M 0 tgt : model trained on source-language data. M 0 trans : model trained on original translation data.","figure_data":"","figure_id":"fig_5","figure_label":"4","figure_type":"figure"},{"figure_caption":"Figure 5 :5Figure 5: Illustration of the relationship between reliability score and label quality.","figure_data":"","figure_id":"fig_6","figure_label":"5","figure_type":"figure"},{"figure_caption":"F1Figure 7 :7Figure 7: F1 scores of CoLaDa with different K for neighborhood information on German dataset.","figure_data":"","figure_id":"fig_7","figure_label":"7","figure_type":"figure"},{"figure_caption":"' Duchy of Aquitaine[LOC] '' ' -William V[PER] ( 995-1030 ) Inaccurate translation Translate \"Duchy of Aquitaine\" into \"爱因公爵\" (Duke of Ain) incorrectly. Original Translation: ** \" 爱因公爵[LOC] \" -威廉五世[PER] (995) Translate-train: ** \" 爱因公爵[LOC] \" -威廉五世[PER] (995) Ours: ** \" 爱因公爵[PER] \" -威廉五世[PER] (995)","figure_data":"","figure_id":"fig_8","figure_label":"","figure_type":"figure"},{"figure_caption":"","figure_data":"𝒊-th iterationStep 1𝒉ො 𝑦 𝑖𝑤 𝑝 𝑖 𝑢 , 𝑥 𝑖 𝑡𝛽 1{𝒙 𝑡 , 𝒚 𝑡 }𝑥 𝑖 𝑡ℳ 𝑡𝑔𝑡 (𝑖-1)𝑝 𝑖 𝑢{𝒙 𝑡 , 𝒑 𝑢 }ℳ 𝑡𝑟𝑎𝑛𝑠 (𝑖)to Step 2Translation data 𝓓 𝒕𝒓𝒂𝒏𝒔Instance collaboration𝒉ො 𝑦 𝑖𝑤 𝑝 𝑖 𝑡 , 𝑥 𝑖 𝑢{𝒙 𝑢 }𝑥 𝑖 𝑢ℳ 𝑡𝑟𝑎𝑛𝑠 (𝑖)𝑝 𝑖 𝑡{𝒙 𝑢 , 𝒑 𝑡 }ℳ 𝑡𝑔𝑡 (𝑖)to 𝒊+1 iterationPseudo-labeled target-language data","figure_id":"tab_1","figure_label":"","figure_type":"table"},{"figure_caption":": model trained on D trans /D tgt at i-th iteration.","figure_data":"Figure 3: Framework of CoLaDa, which is an iterative model-collaboration process with two steps: 1) Step 1:noise-robust training on translation data with the collaborator M (i-1) tgt , 2) Step 2: noise-robust training on unlabeledtarget-language data with the collaborator M (i) trans . The instance-collaboration is used to re-weight the noisylabels from a teacher model in both steps. M (i) trans /M (i)","figure_id":"tab_2","figure_label":"","figure_type":"table"},{"figure_caption":"Ablation study on CoNLL and WikiAnn.","figure_data":"MethoddeesnlarhizhCoLaDa77.30 80.43 85.09 54.26 72.42 60.771) CoLaDa w/o instance collaboration76.08 79.94 83.86 50.98 71.31 59.642) CoLaDa w/o translation data denoise76.17 79.22 83.10 41.41 71.10 55.043) CoLaDa w/o iteratively denoise75.77 79.64 83.50 47.82 71.31 57.644) CoLaDa w/o model collaboration75.64 78.99 82.98 46.51 71.09 55.255) CoLaDa w/o instance & model collaboration 74.54 79.94 82.97 42.33 70.39 55.55","figure_id":"tab_3","figure_label":"3","figure_type":"table"},{"figure_caption":"1 reports the dataset statistics for CoNLL and WikiAnn. Table A.1: Dataset statistics. N S : the number of sentences, N E : the number of entities.","figure_data":"LanguageStatisticTrainDevTestEnglish (en)NS14,0423,2523,454(CoNLL-2003)NE23,4995,9425,648German (de)NS12,1672,8753,009(CoNLL-2003)NE11,8514,8333,673Spanish (es)NS8,4051,9261,524(CoNLL-2002)NE18,7984,3513,558Dutch (nl)NS15,8362,8955,202(CoNLL-2002)NE13,3442,6163,941English (en)NS20,000 10,000 10,000(WikiAnn)NE27,931 14,146 13,958Arabic (ar)NS20,000 10,000 10,000(WikiAnn)NE22,501 11,267 11,259Hindi (hi)NS5,0001,0001,000(WikiAnn)NE6,1241,2261,228Chinese (zh)NS20,000 10,000 10,000(WikiAnn)NE24,135 12,017 12,049","figure_id":"tab_5","figure_label":"A","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Figure 1 :1Figure 1: Comparison between previous methods (a/b/c) and our CoLaDa at the i-th iteration (d) CoLaDa starts at M 0 tgt and performs denoising iteratively. D src : Sourcelanguage labeled data. D trans : Translation data. D tgt : Target-language unlabeled data with pseudo-labels generated by NER models. M src/trans/tgt : NER model learned on D src/trans/tgt .\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 2 :2Figure 2: Illustration of the instance collaboration for denoising. Different colors depict different entity types.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 33Figure 3 depicts an overview of the CoLaDa framework. It is an iterative model-collaboration-based denoising framework which consists of two steps: noise-robust learning on translation data and noiserobust learning on unlabeled target-language data. An instance-collaboration-based denoising strategy (Sec 3.1) is then integrated into the modelcollaboration-based denoising procedure (Sec 3.2).\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Input: an NER model Msrc trained on Dsrc, translation data Dtrans, the unlabeled data Dtgt, the maximum iteration T.1: M (0)tgt ← Msrc ▷ Initialization 2: for i = 1, 2, . . ., T do 3:# Step 1: Noise-robust training on Dtrans 4\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 4 :4Figure 4: F1 scores of M (i) trans and M (i) tgt on the test data and the target-language unlabeled data. M 0 tgt : model trained on source-language data. M 0 trans : model trained on original translation data.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_5\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Figure 5 :5Figure 5: Illustration of the relationship between reliability score and label quality.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"F1Figure 7 :7Figure 7: F1 scores of CoLaDa with different K for neighborhood information on German dataset.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"7\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"' Duchy of Aquitaine[LOC] '' ' -William V[PER] ( 995-1030 ) Inaccurate translation Translate \\\"Duchy of Aquitaine\\\" into \\\"爱因公爵\\\" (Duke of Ain) incorrectly. Original Translation: ** \\\" 爱因公爵[LOC] \\\" -威廉五世[PER] (995) Translate-train: ** \\\" 爱因公爵[LOC] \\\" -威廉五世[PER] (995) Ours: ** \\\" 爱因公爵[PER] \\\" -威廉五世[PER] (995)\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_8\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"\",\n \"figure_data\": \"𝒊-th iterationStep 1𝒉ො 𝑦 𝑖𝑤 𝑝 𝑖 𝑢 , 𝑥 𝑖 𝑡𝛽 1{𝒙 𝑡 , 𝒚 𝑡 }𝑥 𝑖 𝑡ℳ 𝑡𝑔𝑡 (𝑖-1)𝑝 𝑖 𝑢{𝒙 𝑡 , 𝒑 𝑢 }ℳ 𝑡𝑟𝑎𝑛𝑠 (𝑖)to Step 2Translation data 𝓓 𝒕𝒓𝒂𝒏𝒔Instance collaboration𝒉ො 𝑦 𝑖𝑤 𝑝 𝑖 𝑡 , 𝑥 𝑖 𝑢{𝒙 𝑢 }𝑥 𝑖 𝑢ℳ 𝑡𝑟𝑎𝑛𝑠 (𝑖)𝑝 𝑖 𝑡{𝒙 𝑢 , 𝒑 𝑡 }ℳ 𝑡𝑔𝑡 (𝑖)to 𝒊+1 iterationPseudo-labeled target-language data\",\n \"figure_id\": \"tab_1\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \": model trained on D trans /D tgt at i-th iteration.\",\n \"figure_data\": \"Figure 3: Framework of CoLaDa, which is an iterative model-collaboration process with two steps: 1) Step 1:noise-robust training on translation data with the collaborator M (i-1) tgt , 2) Step 2: noise-robust training on unlabeledtarget-language data with the collaborator M (i) trans . The instance-collaboration is used to re-weight the noisylabels from a teacher model in both steps. M (i) trans /M (i)\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Ablation study on CoNLL and WikiAnn.\",\n \"figure_data\": \"MethoddeesnlarhizhCoLaDa77.30 80.43 85.09 54.26 72.42 60.771) CoLaDa w/o instance collaboration76.08 79.94 83.86 50.98 71.31 59.642) CoLaDa w/o translation data denoise76.17 79.22 83.10 41.41 71.10 55.043) CoLaDa w/o iteratively denoise75.77 79.64 83.50 47.82 71.31 57.644) CoLaDa w/o model collaboration75.64 78.99 82.98 46.51 71.09 55.255) CoLaDa w/o instance & model collaboration 74.54 79.94 82.97 42.33 70.39 55.55\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"1 reports the dataset statistics for CoNLL and WikiAnn. Table A.1: Dataset statistics. N S : the number of sentences, N E : the number of entities.\",\n \"figure_data\": \"LanguageStatisticTrainDevTestEnglish (en)NS14,0423,2523,454(CoNLL-2003)NE23,4995,9425,648German (de)NS12,1672,8753,009(CoNLL-2003)NE11,8514,8333,673Spanish (es)NS8,4051,9261,524(CoNLL-2002)NE18,7984,3513,558Dutch (nl)NS15,8362,8955,202(CoNLL-2002)NE13,3442,6163,941English (en)NS20,000 10,000 10,000(WikiAnn)NE27,931 14,146 13,958Arabic (ar)NS20,000 10,000 10,000(WikiAnn)NE22,501 11,267 11,259Hindi (hi)NS5,0001,0001,000(WikiAnn)NE6,1241,2261,228Chinese (zh)NS20,000 10,000 10,000(WikiAnn)NE24,135 12,017 12,049\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"A\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu et al., 2020a)\", \"Explanation\": \"The cited work, UniTrans, is utilized in the citing paper to finetune a weak model for annotation of unlabeled target-language data, providing a methodological basis for the research conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liang et al., 2021)\", \"Explanation\": \"The cited work by Liang et al. provides a method for training an instance selector to mitigate label noise in the data, which the citing paper adopts in its research to address the issue of label noise in the data.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chen et al., 2021)\", \"Explanation\": \"The cited work by Chen et al. presents a method for data selection to mitigate label noise in the data, which the citing paper uses to design heuristic rules for data selection in their research to address the issue of label noise in the data.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ni et al., 2017)\", \"Explanation\": \"The cited work by Ni et al. proposes a method for manipulating the translation data to mitigate label noise in the data, which the citing paper extends by further exploring the complementary characteristics between the translation data and the target-language data for denoising.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Liang et al., 2021)\", \"Explanation\": \"The cited work by Liang et al. provides a method for generating pseudo-labels in the source language to mitigate label noise in the data, which the citing paper extends by further exploring the complementary characteristics between the translation data and the target-language data for denoising.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Chen et al., 2021)\", \"Explanation\": \"The cited work by Chen et al. presents a method for data selection to mitigate label noise in the data, which the citing paper extends by further exploring the complementary characteristics between the translation data and the target-language data for denoising.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Gu et al., 2019)\", \"Explanation\": \"The cited work introduces the concept of anomaly detection in the feature space, which the citing paper adopts to measure the anomalism of data points in the context of denoising translation data.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Zhai and Wu, 2019)\", \"Explanation\": \"The cited work by Zhai and Wu (2019) has shown that instances with the same label are more likely to locate close to each other in the representation space, which the citing paper extends to support the use of similar tokens in the feature space for denoising translation data.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Xu et al., 2020)\", \"Explanation\": \"The cited work by Xu et al. (2020) has also shown the use of similar tokens in the feature space for denoising translation data, which the citing paper further extends to support the use of this method in the context of re-labeling target-language unlabeled data.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhai and Wu, 2019)\", \"Explanation\": \"The cited work by Zhai and Wu (2019) provides a method for evaluating the reliability of token labels based on label consistency in the feature space of a deep neural network, which the citing paper adopts in their research on instance-collaboration-based denoising.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xu et al., 2020)\", \"Explanation\": \"The cited work by Xu et al. (2020) also contributes to the method of evaluating label reliability in the feature space of a deep neural network, which the citing paper builds upon in their research on instance-collaboration-based denoising.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Tjong Kim Sang, 2002;Tjong Kim Sang and De Meulder, 2003)\", \"Explanation\": \"The cited works are the source of the CoNLL dataset, which the citing paper uses in their research on cross-lingual NER.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Pan et al., 2017)\", \"Explanation\": \"The cited work is the source of the WikiAnn dataset, which the citing paper uses in their research on cross-lingual NER.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wu et al., 2020a)\", \"Explanation\": \"The cited work provides the word-to-word translation data used in the CoNLL benchmark for a fair comparison in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Fan et al., 2020)\", \"Explanation\": \"The cited work is used to translate the source data to the target language in the WikiAnn benchmark, providing the data used for label projection in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Yang et al., 2022)\", \"Explanation\": \"The cited work is referenced for the marker-based alignment algorithm used in the label projection for the WikiAnn benchmark in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu and Dredze, 2019)\", \"Explanation\": \"The cited work by Wu and Dredze (2019) provides the base encoders used in the citing paper for the implementation of the NER model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Conneau et al., 2020)\", \"Explanation\": \"The cited work by Conneau et al. (2020) contributes the XLM-R large model as a base encoder in the implementation of the NER model in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Devlin et al., 2019)\", \"Explanation\": \"The cited work by Devlin et al. (2019) is the source of the multilingual BERT base model used in the base NER model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Loshchilov and Hutter, 2019)\", \"Explanation\": \"The cited work by Loschilov and Hutter (2019) provides the AdamW optimizer used in the training of the base NER model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu et al., 2020a)\", \"Explanation\": \"The cited work by Wu et al. (2020a) provides the majority of the hyper-parameters used in the training of the base NER model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Johnson et al., 2019)\", \"Explanation\": \"The cited work by Johnson et al. (2019) contributes the FAISS library for embedding retrieval in the implementation of the NER model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wolf et al., 2020)\", \"Explanation\": \"The cited work by Wolf et al. (2020) provides the Hugging-Face transformer library used in the implementation of the NER model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu et al., 2020a)\", \"Explanation\": \"The cited work by Wu et al. (2020a) provides the base encoders and most of the hyper-parameters used in the training of the base NER model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu and Dredze, 2019)\", \"Explanation\": \"The cited work by Wu and Dredze (2019) provides the base encoders and the training hyper-parameters used in the base NER model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Conneau et al., 2020)\", \"Explanation\": \"The cited work by Conneau et al. (2020) contributes the XLM-R large model as a base encoder in the implementation of the NER model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Devlin et al., 2019)\", \"Explanation\": \"The cited work by Devlin et al. (2019) is the source of the multilingual BERT base model used in the base NER model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Loshchilov and Hutter, 2019)\", \"Explanation\": \"The cited work by Loschilov and Hutter (2019) provides the AdamW optimizer used in the training of the base NER model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu et al., 2020a)\", \"Explanation\": \"The cited work by Wu et al. (2020a) provides the majority of the hyper-parameters used in the training of the base NER model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Johnson et al., 2019)\", \"Explanation\": \"The cited work by Johnson et al. (2019) contributes the FAISS library for embedding retrieval in the implementation of the NER model in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wolf et al., 2020)\", \"Explanation\": \"The cited work by Wolf et al. (2020) provides the Hugging-Face transformer library used in the implementation of the NER model in the citing paper.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Keung et al., 2019)\", \"Explanation\": \"The cited work by Keung et al. provides a set of hyper-parameters that the citing paper uses in their research, which helps to support the findings and conclusions of the study.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(CoNLL and WikiAnn)\", \"Explanation\": \"The cited datasets, CoNLL and WikiAnn, are the source of the data used in the research conducted in the citing paper. The reliance on these datasets is crucial for the study and its outcomes.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(XLM-R as the encoder)\", \"Explanation\": \"The citing paper extends the research by using the XLM-R encoder in their model, exploring a new dimension and potentially leading to further improvements in the study.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Muller et al., 2021)\", \"Explanation\": \"The cited work by Muller et al. provides evidence that representations from upper layers are more task-specific, which supports the claim made in the citing paper about the better discrimination between noisy and clean tokens in these layers.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(T\\u00e4ckstr\\u00f6m et al., 2012)\", \"Explanation\": \"The cited work provides the idea of using word clusters as a language-independent feature for cross-lingual NER.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Zirikly and Hagiwara, 2015)\", \"Explanation\": \"The cited work introduces the use of gazetteers as a feature for cross-lingual NER.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Tsai et al., 2016)\", \"Explanation\": \"The cited work suggests the use of Wikifier features as a method for cross-lingual NER.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Ni et al., 2017)\", \"Explanation\": \"The cited work presents the idea of using cross-lingual word embedding as a feature for cross-lingual NER.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Devlin et al., 2019)\", \"Explanation\": \"The cited work by Devlin et al. (2019) provides the foundation for the development of multilingual pre-trained language models, which are then utilized in the citing paper to promote the learning of language-independent features through meta-learning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Conneau et al., 2020)\", \"Explanation\": \"The cited work by Conneau et al. (2020) contributes to the development of multilingual pre-trained language models, which are then used in the citing paper to further promote the learning of language-independent features via meta-learning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu and Dredze, 2019)\", \"Explanation\": \"The cited work by Wu and Dredze (2019) highlights the promising results of cross-lingual transfer with multilingual pre-trained language models, which the citing paper builds upon to promote the learning of language-independent features through meta-learning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu et al., 2020c)\", \"Explanation\": \"The cited work by Wu et al. (2020c) provides a method for learning language-independent features through meta-learning, which the citing paper adopts to further promote the learning of such features.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu and Dredze, 2020)\", \"Explanation\": \"The cited work by Wu and Dredze (2020) presents a method for learning language-independent features through contrastive alignment, which the citing paper builds upon to further promote the learning of such features.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Keung et al., 2019)\", \"Explanation\": \"The cited work by Keung et al. (2019) introduces a method for learning language-independent features through adversarial learning, which the citing paper adopts to further promote the learning of such features.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chen et al., 2021)\", \"Explanation\": \"The cited work by Chen et al. (2021) presents a method for learning language-independent features through adversarial learning, which the citing paper builds upon to further promote the learning of such features.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Fetahu et al., 2022)\", \"Explanation\": \"The cited work by Fetahu et al. (2022) discusses the integration of other resources to promote the learning of language-independent features, which the citing paper adopts to further advance the learning of such features.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Fu et al., 2023)\", \"Explanation\": \"The cited work by Fu et al. (2023) highlights the importance of language-specific features in cross-lingual transfer, which the citing paper acknowledges in the context of learning language-independent features.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xie et al., 2018)\", \"Explanation\": \"The cited work proposes a word-to-word translation method that the citing paper adopts to translate text in a specific way.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Mayhew et al., 2017)\", \"Explanation\": \"The cited work proposes a phrase-to-phrase translation method that the citing paper uses to translate text in a specific way.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Jain et al., 2019)\", \"Explanation\": \"The cited work proposes an entity projection algorithm that the citing paper utilizes to translate text using the Google translation system.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2021)\", \"Explanation\": \"The cited work proposes a method to translate sentences with pre-defined markers for label projection, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Yang et al., 2022)\", \"Explanation\": \"The cited work proposes a method to translate sentences with pre-defined markers for label projection, which the citing paper uses in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ni et al., 2017)\", \"Explanation\": \"The cited work designs heuristic rules to select high-quality translation data, which the citing paper utilizes in their research to improve the quality of translation data.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Garc\\u00eda-Ferrero et al., 2022)\", \"Explanation\": \"The cited work highlights the limitations of data noise and artifacts in translation data, which the citing paper acknowledges in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu et al., 2020b)\", \"Explanation\": \"The cited work proposes a knowledge distillation-based method to train a student model on unlabeled target-language data with soft labels from a teacher model, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Li et al., 2022)\", \"Explanation\": \"The cited work improves single task-based teacher-student learning with entity similarity as an auxiliary task, which the citing paper builds upon in their research to improve the method.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chen et al., 2022)\", \"Explanation\": \"The cited work addresses the issue of label noise from the teacher model in knowledge distillation-based methods, which the citing paper builds upon in their research to mitigate the problem.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(2021)\", \"Explanation\": \"The cited work by Liang et al. (2021) provides a reinforcement learning algorithm for training an instance selector to select reliable pseudo labels iteratively, which the citing paper adopts as a methodological basis for their own research on cross-lingual NER.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Wu et al., 2020a)\", \"Explanation\": \"The cited work by UniTrans (Wu et al., 2020a) utilizes a model trained on translation data to perform teacher-student learning on unlabeled data, which the citing paper extends to a unified denoising framework for cross-lingual NER.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Zheng et al., 2021;Zhou et al., 2022)\", \"Explanation\": \"The cited works by Zheng et al. (2021) and Zhou et al. (2022) have explored the use of consistency training in leveraging unlabeled data and translation data for cross-lingual NER, providing supporting evidence for the citing paper to build upon in their own research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Adelani et al., 2022)\", \"Explanation\": \"The cited work on MasakhaNER 2.0 provides a potential application of the label denoising framework to other low-resource languages, which the citing paper plans to explore in future work.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Wu and Dredze, 2019)\", \"Explanation\": \"The cited work, mBERT, serves as a base model for the citing paper in the development of a NER model in the source language.\"}, {\"Category\": \"Supporting Evidence\", \"Citation\": \"(Conneau et al., 2020)\", \"Explanation\": \"The cited work, XLM-R, is another base model that the citing paper uses in the development of a NER model in the source language.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu and Dredze, 2020)\", \"Explanation\": \"The cited work, BERT-align, introduces the word-level contrastive alignment loss to enhance the mBERT representation, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Keung et al., 2019)\", \"Explanation\": \"The cited work, AdvCE, exploits adversarial learning on source-and target-language text to avoid language-specific information, which the citing paper leverages in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chen et al., 2021)\", \"Explanation\": \"The cited work, AdvPicker, leverages adversarial learning to learn language-shared features and select less language-specific sentences in target-language unlabeled text for knowledge distillation, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2021)\", \"Explanation\": \"The cited work, MulDA, proposes the labeled sequence translation method for data projection from source-language NER data, which the citing paper builds upon in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Wu et al., 2020a)\", \"Explanation\": \"The cited work, UniTrans, unifies model-and translation-data-based transfer via knowledge distillation, which the citing paper builds upon in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2021)\", \"Explanation\": \"TOF leverages the labeled data for machine reading comprehension task on target language to help the NER task in crosslingual transfer, providing a method for cross-lingual transfer in NER tasks.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wu et al., 2020b)\", \"Explanation\": \"TSL proposes knowledge distillation to use unlabeled target-language data for cross-lingual NER, indicating the use of unlabeled data in cross-lingual NER.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liang et al., 2021)\", \"Explanation\": \"RIKD proposes a reinforcement learning algorithm to iteratively select reliable pseudo-labels for knowledge distillation, providing a method for knowledge distillation in cross-lingual NER.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Li et al., 2022)\", \"Explanation\": \"MTMT proposes multi-task multi-teacher knowledge distillation, which further leverages the entity similarity task, extending the research on cross-lingual NER by exploring new tasks in the field.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zheng et al., 2021)\", \"Explanation\": \"xTune leverages unlabeled translation text and other word-level data augmentation techniques for consistency training, providing a method for consistency training in cross-lingual NER.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhou et al., 2022)\", \"Explanation\": \"ConNER conducts span-level consistency training on unlabeled target-language data using translation and further applies dropout-based consistency training on the source-language data, providing a method for consistency training in cross-lingual NER.\"}]"}}},{"rowIdx":99,"cells":{"sections":{"kind":"list like","value":[{"figure_ref":[],"heading":"Introduction","publication_ref":["b44","b19","b28","b38","b26","b34","b25","b37","b21","b33","b6","b16","b18","b27","b12"],"table_ref":["tab_0"],"text":"Semantic segmentation, aiming at assigning semantic labels for each image pixel, is a fundamental and long-standing goal of both the computer vision (CV) and remote sensing (RS) fields (Zhu et al., 2017;Kuznietsov et al., 2017;Mou and Zhu, 2018;Xiong et al., 2022). With the advent of deep learning techniques, RGB image-based semantic segmentation has attracted great research attention, and significant progress has been made on performance (Long et al., 2015;Xie et al., 2021;Liu et al., 2021). Despite the richness of texture information in RGB images, semantic segmentation models often face challenges in extracting discriminative features from them, owing to the inherent limitations of 2D RGB representations. Considering this problem, multimodal image segmentation is becoming more and more popular in both the CV and RS communities. Multi-modal data such as RGB-Depth (RGB-D) (Xiong et al., 2021) and RGB-Thermal (RGB-T) (Li et al., 2019), contain richer information compared with traditional RGB images, and are widely used for many tasks (Wang and Neumann, 2018;Cheng et al., 2017;Jiang et al., 2018b;Xiong et al., 2020b). Multi-modal data, which incorporates additional modalities beyond RGB, has shown promising results in achieving significantly improved performance compared to using RGB data alone. Remote sensing (RS) data presents several rich data modalities beyond RGB images, such as hyperspectral (HS), multi-spectral (MS), light detection and ranging (LiDAR), normalized digital surface model (nDSM), and synthetic aperture radar (SAR). As a result, multi-modal or multi-sensor data fusion has become a vital area of research in the remote sensing community.\nThere are various modalities widely used in RS for different Earth observation applications. The development of multi-modal benchmark datasets is critical for advancing research in multi-modal learning methods. However, for semantic segmentation on remote sensing (RS) data, existing multi-modal benchmark datasets have several limitations, given the application areas and data acquisition cost. The foremost limitation is the limited spatial resolutions of the available multi-modal datasets. While Sentinel-1 and Sentinel-2 images are widely-used satellite data with multiple modalities, their low spatial resolution fails to provide detailed land-use and land-cover information. The second limitation pertains to the limited geospatial coverage of existing multi-modal datasets. Although the ISPRS Potsdam and Vaihingen datasets have a higher spatial resolution of 5 cm (0.05m), obtaining such high-resolution images for large-scale real-world applications could be cost-prohibitive. The third limitation of the current multi-modal benchmark datasets is the lack of a unified benchmark platform to enable comprehensive and fair comparison of different multi-modal learning methods. have proven that using the geometric information in nDSM leads to a higher segmentation performance on remote sensing data. However, as presented in Table 1, existing multi-modal datasets that contain RGB and nDSM pairs are relatively small. This makes it difficult to faithfully compare the effectiveness of different types of multi-modal representation learning methods. These limitations motivate us to build a new multi-modal semantic segmentation dataset to enable a fair and unified evaluation of different multimodal segmentation methods. The proposed Geometry-Aware Multi-modal Segmentation (GAMUS) dataset contains images with a resolution of 0.33m, which is higher enough to be used in many real-world applications. The GAMUS dataset contains two data modalities: RGB images and normalized digital surface models (nDSM). Since nDSM indicates the height of ground objects, we use the height modality to represent the nDSM data. nDSM has been broadly provided by many cities owing to its importance in 3D city modeling. Unlike depth images in CV datasets, different types of land covers usually have unique height attributes. Thus, the geometric information contained in height maps is highly correlated to the semantic information (Kunwar, 2019;Mahmud et al., 2020). In other words, compared with other modalities, height information from nDSM has great potential in improving the segmentation performance of high-resolution remote sensing images. Although extensive works have been proposed to make better use of the diverse information contained in different modalities (Hong et al., 2021), there is still a lack of comprehensive benchmarking of existing multi-modal learning methods on RS data. There are many widely used strategies for multi-modal data fusion, including early fusion, featurelevel fusion, late-fusion, and Transformer-based token fusion. However, it is still not clear which one is more suitable for the pixel-level semantic segmentation task on RGB-H (Height) data. Thus, the great potential of the nDSM modality is heavily overlooked by existing works.\nConsidering this problem, in this work, we propose a simple yet effective Transformer-based Intermediary Multimodal Fusion (TIMF) Module. TIMF utilizes an intermediary learnable token to fuse the features of RGB and height modalities via the self-attention mechanism. To provide comprehensive benchmarking results, we conduct extensive experiments to evaluate Convolutional Neural Networks (CNN) and Transformer-based methods and different variants of fusion strategies. We believe that the benchmark dataset and released codebase can foster future research on evaluating and developing new multi-modal learning methods for Earth observation applications. We summarize the main contributions of this paper below. The remainder of this paper is structured as follows. Section 2 provides an overview of the related methods of multi-modal learning and existing multi-modal datasets for remote sensing (RS) data. Section 3 presents detailed information about the proposed GAMUS dataset. In Section 4, we describe the CNN and Transformer-based multi-modal learning methods with different fusion strategies, along with the designed Transformer-based Intermediary Multi-modal Fusion (TIMF) module. Finally, in Section 5, we present and analyze the benchmarking results obtained through comprehensive and fair comparisons of the proposed models on the GAMUS dataset."},{"figure_ref":[],"heading":"Related Work","publication_ref":["b41","b3","b9","b5","b4","b8","b7","b42","b10","b29","b22","b14","b6","b31","b32"],"table_ref":[],"text":"Multi-modal Representation Learning for Computer Vision Multi-modal feature learning is commonly studied in RGB-D image-based scene recognition (Yuan et al., 2019), semantic segmentation (Cao et al., 2021;Ha et al., 2017;Chen et al., 2020Chen et al., , 2021)), object detection (Gupta et al., 2014;Fan et al., 2020), and action recognition (Zhang et al., 2016). Multi-modal fusion is the key to designing multi-branch deep networks. Existing methods usually conduct fusion at the image level (early fusion) or the feature level (intermediate fusion) (Jiang et al., 2018a). FuseNet (Hazirbas et al., 2016) and RedNet (Jiang et al., 2018a) summed RGB and depth features to obtain multi-modal representations for RGB-D semantic segmentation. Multilevel feature fusion (Park et al., 2017) was designed to extend the residual learning idea of RefineNet (Lin et al., 2017) for RGB-D image segmentation. Similarly, ACNet (Hu et al., 2019) and Gated Fusion Net (Cheng et al., 2017) were proposed to adaptively fuse features of different modalities for image segmentation. PSTNet (Shivakumar et al., 2020) and RTFNet (Sun et al., 2019) were proposed to utilize long wave infrared (LWIR) imagery as a viable supporting modality for semantic segmentation using multi-modal learning networks."},{"figure_ref":[],"heading":"Multi-modal Representation Learning for Remote Sensing","publication_ref":["b17","b0","b1","b24","b43"],"table_ref":[],"text":"For remote sensing data, (Kampffmeyer et al., 2016) designed a multi-modal deep network with an early deep fusion architecture by stacking all modalities as the input. A promising performance has been achieved on the semantic segmentation of urban images. Audebert et al. (Audebert et al., 2017) proposed to combine the optical and OpenStreetMap data using a two-stream multi-modal learning network to improve the segmentation performance. Exploring the combination of Multi-spectral images (MSI) and Lidar is closer to our work. Audebert et al. (Audebert et al., 2018) introduced a SegNet-based multi-modal fusion architecture for the segmentation of urban scenes. Crossmodality learning (CML) was investigated by (Hong et al., 2020b,a). In (Hong et al., 2020a), a cross-fusion strategy was proposed for learning multi-modal features with more balanced weight contributions from different modalities. Liu et al. (Liu et al., 2019) designed a high-order conditional random field (CRF) based method for the fusion of optical and Lidar predictions in a late-fusion manner. G2GNet (Zheng et al., 2021) was proposed to combine the complementary merits of RGB and auxiliary modality data using the introduced gather-to-guide module."},{"figure_ref":[],"heading":"Remote Sensing Datasets for Multi-modal Semantic Segmentation","publication_ref":["b30","b2","b20"],"table_ref":["tab_0"],"text":"The two widely used high-resolution semantic segmentation benchmark datasets in remote sensing are the ISPRS Potsdam and Vaihingen datasets (Rottensteiner et al., 2014). As compared in Table 1, The US3D (Bosch et al., 2019) dataset is collected from public data. Although it is large in volume (more than 700 GB) and contains data for many different tasks, it is not designed dedicated to multi-modal segmentation tasks. In contrast, the GAMUS dataset is smaller in volume, easier to use, and covers more cities. We provide a standard dataloader, detailed instructions, strong baselines, and extensive benchmark results for multi-modal learning. The DFC 19 (Le Saux et al., 2019) dataset is derived from US3D, which is smaller in scale. \n.111止--------一------ 1111111■■ 一------一-一-------1111 •• _____-一----- 600 400 200 0 .11111111■■---一一一一一一一-一- 1111111■■■-一一---- 111111■■-----------一-"},{"figure_ref":[],"heading":"The GAMUS Dataset","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":["fig_1"],"heading":"Data collection","publication_ref":[],"table_ref":[],"text":"High-resolution orthophotos, semantic maps, and nDSM (height) are derived and processed from open Data DC catalog (https://opendata.dc.gov) and open data in the Philadelphia region (https://www.opendataphilly.org). As illustrated in Fig. 2, nDSMs are derived from Lidar point clouds. Firstly, noises in the point clouds are removed. Then the height values are rasterized into DSMs (digital surface models) with all the points, and DTMs (digital terrain models) with only the ground points. Finally, subtracting DTM from DSM gives nDSM. The classified point clouds are also ingredients for semantic maps when land cover maps are not available from open data sources. This is simply done by rasterizing the class labels. All the processed data are aligned and cropped into patches to make the final dataset."},{"figure_ref":["fig_1","fig_3","fig_6","fig_7"],"heading":"Statistics of the GAMUS dataset","publication_ref":[],"table_ref":[],"text":"The introduced GAMUS dataset contains 11,507 tiles collected from five different cities: Oklahoma, Washington, D.C., Philadelphia, Jacksonville, and New York City. These image tiles are collected using the aforementioned data collection process, as shown in Fig. 2. Each RGB image tile has a corresponding nDSM map with a spatial size of 1024×1024. We split all the image tiles into three subsets: the training set with 6,304 tiles, the validation set with 1,059 tiles, and the test set with 4,144 tiles. All the image pixels are annotated with six different land cover types, including 1. ground; 2. low-vegetation; 3. building; 4. water; 5. road; 6. tree. The height statistics provided in the nDSM are displayed in Fig. 3. It can be seen from this figure, there is an obvious long-tailed distribution of the heights, i.e., the number of pixels with lower height values is significantly more than those with higher height values. We also show the statistics of the spatial distribution by averaging the height values at each image pixel across the whole dataset. We can see that the spatial distributions are different for the train, validation, and test subsets.\nFurthermore, we also display the height distributions of different cities in Fig. 4. For all the cities, the height values obey clear long-tailed distributions. We also visualize the spatial patterns of the height maps by averaging all the height values for each position. We can see that the spatial patterns of different cities are quite different. In Fig. 5, we present the distributions of semantic labels for different cities. This analysis reveals significant differences in the distribution of semantic objects across different cities."},{"figure_ref":[],"heading":"Methods","publication_ref":["b39"],"table_ref":[],"text":"Existing deep learning models for multi-modal semantic segmentation can be divided into CNN-based and Transformerbased networks. For CNN-based methods, the fusion in different layers has a great influence on the performance. For Transformer-based methods, the self-attention mechanism can be used to fuse multi-modal features at a token level (Xu et al., 2022), which can be more effective than CNN-based methods."},{"figure_ref":["fig_8"],"heading":"CNN-based Fusion Methods","publication_ref":["b3","b10","b9","b32"],"table_ref":[],"text":"For CNN-based fusion methods, we evaluate the performance using five different training paradigms as shown in Fig. 6. To formally define these training paradigms, we first introduce the notations of input modalities and networks. We use 𝑿 𝑟𝑔𝑏 ∈ ℝ 3×𝐻×𝑊 and 𝑿 ℎ ∈ ℝ 𝐻×𝑊 to denote the RGB input and height from nDSM, respectively. For the sake of \n(𝑋 𝑟𝑔𝑏 )) + 𝑓 𝑑 𝐖 ℎ (𝑓 𝑒 𝐖 ℎ (𝑋 ℎ )).\nSince the height information in the nDSM data is highly similar to the geometric information in the depth images, we select three RGB-D multi-modal learning methods for benchmarking. ShapeConv (Cao et al., 2021), VCD (Xiong et al., 2020a) aims at designing a cross-modality guided encoder to better fuse RGB features and depth information. FuseNet (Hazirbas et al., 2016) focuses on designing better multi-scale feature fusion architectures. Considering the diversity of different multi-modal architecture designs, we also choose two RGB-T segmentation methods for performance evaluation on our GAMUS dataset. Multi-spectral fusion networks (MFNet) (Ha et al., 2017) takes RGB images and IR images as input and fuse them in the feature level for multi-modal learning. RTFNet (Sun et al., 2019) uses two separate encoders for RGB and thermal modalities and fuses their intermediate features progressively to learn multimodal representations.\nFor the performance evaluation of multiple modalities, we adapt three methods (MFNet, RTFNet, and FuseNet) to five types of multi-modal fusion paradigms to compare the performance. Specifically, for single modality settings, we modify them by removing one input modality and only remain a single RGB or height modality. For the earlyfusion method, we change the model by simply stacking the RGB and height modality together as the multi-modal input. As for the feature-level fusion, we use the same network architectures as described in the original papers of the eight compared methods. For the late fusion setting, we only fuse the predicted segmentation results of each modality at the last layer of the segmentation head. Toward fair comparisons, we exploit the exact same network architectures and training hyper-parameters for methods trained in different experimental settings."},{"figure_ref":["fig_9","fig_9"],"heading":"Transformer-based Fusion Methods","publication_ref":["b23","b34","b23"],"table_ref":[],"text":"CMX (Liu et al., 2022) introduces a cross-modal feature rectification module to calibrate the feature of one modality from another in both spatial-and channel-wise dimensions. SegFormer (Xie et al., 2021) is used as its baseline segmentation architecture. Considering its effectiveness, we take CMX as our baseline method and adapt it to five different fusion strategies for performance comparison. Fig. 7 presents six types of fusion methods. The first five fusion methods including single modality, early fusion, late fusion, and cross-modal feature fusion (Liu et al., 2022) are similar to the five CNN-based fusion strategies. Different from these five fusion methods, in this work, we propose the intermediary feature fusion module (illustrated in Fig. 7 (f)), which is described in the following section.\nTo formally describe these Transformer-based fusion models, we first give a brief introduction to the widelyused vision Transformer (ViT). Given the input data 𝒙, a Transformer encoder first projects the image patches into patch embeddings (denoted as 𝑝(⋅)) and add position embeddings 𝒙 𝑝𝑜𝑠 to enhance the position information. Then the input will further go through alternating layers of multihead self-attention (MSA) and multi-layer perception (MLP) blocks. Layer Normalization (LN) and residual connection are applied before and after every block. The computation process of a Transformer block TL(𝒙) can be expressed as\n𝒙 0 = 𝑝(𝒙) + 𝒙 𝑝𝑜𝑠 , 𝒛 ′ 𝑘 = 𝒛 𝑘-1 + MSA(LN(𝒙 𝑘-1 )), 𝒛 𝑘 = 𝒛 ′ 𝑘 + MLP(LN(𝒛 ′ 𝑘 )),(1)\nwhere 𝒛 𝑘 is the output features, and 𝑘 is the index of blocks.\nLet 𝐹 𝑒 𝑚 (⋅) and 𝐹 𝑑 𝑚 (⋅) be the encoder and decoder of the Segformer () networks, where 𝑚 could be either the RGB or the height modality. The single modality model with RGB input can be defined as 𝑦 * 𝑠 = 𝐹 𝑑 (𝐹 𝑒 (𝑋 𝑟𝑔𝑏 )). The computation for the height modality can be defined by simply replacing 𝑋 𝑟𝑔𝑏 with 𝑋 𝑑 . The early fusion strategy can be expressed as 𝑦 * 𝑠 = 𝐹 𝑑 (𝐹 𝑒 (𝑋 𝑟𝑔𝑏 ∥ 𝑋 𝑑 )). Similarly, the late fusion model is formulated as\nTransformer Q K V Transformer Q K V V K Q Transformer Transformer V K Q Q K V Transformer Q K V Transformer Transformer Q K V(\n𝑦 * 𝑠 = 𝐹 𝑑 𝑟𝑔𝑏 (𝐹 𝑒 𝑟𝑔𝑏 (𝑋 𝑟𝑔𝑏 )) + 𝐹 𝑑 ℎ (𝐹 𝑒 𝑑 (𝑋 ℎ )).\n(2)\nThe Transformer-based cross feature fusion is conducted by exchanging queries of different modalities in the MSA blocks. The other computation steps are the same as a Transformer block. For the sake of simplicity, we omit them and only present the computation of MSA, which can be expressed as:\n𝒛 𝑟𝑔𝑏 ⟵ MSA(LN(𝑿 ℎ , 𝑿 𝑟𝑔𝑏 , 𝑿 𝑟𝑔𝑏 ⏟⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏟ 𝑄, 𝐾, 𝑉\n)),\n𝒛 ℎ ⟵ MSA(LN(𝑿 𝑟𝑔𝑏 , 𝑿 ℎ , 𝑿 ℎ ⏟⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏟ 𝑄, 𝐾, 𝑉\n)).\n(3)\nBy this means, the multi-modal features are fused via cross-modal self-attention modules, which is more flexible than layer-wise feature fusion."},{"figure_ref":["fig_9"],"heading":"Intermediary Multi-modal Fusion","publication_ref":["b23"],"table_ref":[],"text":"Different from existing fusion strategies, in this work, we propose a novel Transformer-based intermediary multimodal fusion method, i.e., the TIMF module. TIMF exploits an additional intermediary token 𝑴 to adaptively combine RGB and height features at a token level via the self-attention mechanism. To be more specific, TIMF is designed in a hierarchical manner. As illustrated in Fig. 7 (f), the intermediary token first extracts global multi-modal features by feeding the intermediary token into an MSA module, which is similar to the CLS token in ViT. Next, the output intermediary token 𝑴 ′ is concatenated to the tokens of each modality respectively. Then, single-block (𝑘 = 1) Transformer layers are used to further fuse the intermediary token 𝑴 ′ with modality-specific tokens via self-attention modules. Formally, the first stage of TIMF takes 𝑴 and tokens of both modalities as input to a singleblock Transformer layer, which can be formulated as\n𝒛 𝑎 = TL(𝑝(𝑿 𝑟𝑔𝑏 ) ∥ 𝑝(𝑿 𝑑 ) ∥ 𝑀), 𝒛 𝑎 = (𝒛 𝑟𝑔𝑏 ) ∥ (𝒛 ℎ ) ∥ 𝑀 ′ ),(4)\nwhere 𝒛 𝑎 is the output feature. ∥ is the concatenation operation. By slicing 𝒛 𝑎 according to the number of tokens, we can obtain the output features of RGB modality 𝒛 𝑟𝑔𝑏 ), height modality 𝒛 ℎ ), and the output intermediary token 𝑴 ′ . Then, the final output features of the TIMF module are obtained by fusing 𝑴 ′ with these two modalities individually using two Transformer layers. This can be formulated as\n𝒛 ′ 𝑟𝑔𝑏 = TL 𝑟𝑔𝑏 (𝑝(𝒛 𝑟𝑔𝑏 ) ∥ 𝑴 ′ ), 𝒛 ′ 𝑟𝑔𝑏 = (𝒛 𝑟𝑔𝑏 ) ∥ 𝑴 ′ 𝑟𝑔𝑏 ), 𝒛 ′ ℎ = TL ℎ (𝑝(𝒛 ℎ ) ∥ 𝑴 ′ ), 𝒛 ′ ℎ = ((𝒛 ℎ ) ∥ 𝑴 ′ ℎ ).\n(5)\nIn practice, TIMF module can be used in a plug-andplay manner. We simply replace the TIMF module with the feature fusion module in CMX (Liu et al., 2022). "},{"figure_ref":[],"heading":"Experiments","publication_ref":[],"table_ref":[],"text":"The constructed dataset GAMUS in this paper can be publicly accessed at https://github.com/EarthNets/Dataset4EO. In this dataset, we have explicitly provided the official splits for training, validation, and test subsets. Providing a well-defined data set and the official split is crucial for reproducibility."},{"figure_ref":[],"heading":"Implementation details","publication_ref":[],"table_ref":[],"text":"All the evaluated models are implemented using Pytorch and run on eight GeForce RTX 3090 GPUs. We use the public codebase provided by the original paper. For all the supervised multi-modal segmentation methods, 100 epochs are used for training. The input images are resized to 512×512. The batch sizes are set to 16, and we use the same learning rate and optimizer as described in the original paper. For data augmentation, random flip and random crop are used for all the methods towards a fair comparison. On the GAMUS dataset, 50 epochs are used for the model training. The implemented dataset loading code can be found in https: //github.com/EarthNets/Dataset4EO. For all the experiments, the mean Intersection over Union (mIoU) and IoU of each class are used as the evaluation metrics.\nTo ensure reproducibility, we choose to make all the used source codes publicly available at https://github.com/ EarthNets/RSI-MMSegmentation."},{"figure_ref":[],"heading":"Multi-modal Learning Analysis","publication_ref":[],"table_ref":["tab_4","tab_5","tab_3","tab_4","tab_5"],"text":"To analyze the benefits of using multiple modalities for semantic segmentation, we compare the experimental results of using different data modalities as input in Table 2, Table 3, andTable 4. Specifically, for CNN-based fusion methods, we present the accuracy and mIoU results in Table 2 andTable 3. As for the Transformer-based methods, we provide the mIoU results of six different types of modality inputs in Table 4. In the following, we analyze and discuss these extensive results from five different aspects."},{"figure_ref":[],"heading":"Benefits of Multi-modal Learning","publication_ref":[],"table_ref":["tab_4","tab_5"],"text":"Despite the fact that early-fusion is the simplest multi-modal learning method, stacking RGB and height map as a four-channel input, can still obtain clearly better results than using a single RGB modality. As presented in Table 3, for CNN-based methods, using early-fusion, MFNet can improve the mIoU of RGB modality from 46.8% to 49.6%. FuseNet can obtain a 9% improvement than only using the RGB data. These results reveal the effectiveness of multi-modal learning for the segmentation of RS images. Furthermore, by exploiting more sophisticated feature-level multi-modal fusion methods, the performance can be further improved. For example, MFNet with a feature-fusion strategy can obtain a 6% improvement to using a single RGB modality. However, surprisingly, for FuseNet, the performances of the simple early-fusion methods may outperform feature-level fusion methods. In general, from the results, we can clearly see that using both the RGB and height modalities can obtain much better results compared to using single modalities.\nComparing the results of using RGB and RGBH data, tree and building are the semantic classes with the greatest performance boost when the extra height modality is used. This makes sense because trees and buildings are objects with clearly higher height values. Comparing the results of using height and RGBH data, there is significant performance improvement. Specifically, ground and vegetation are the semantic classes with the greatest performance boost when the extra RGB modality is used. As ground and vegetation always have the lowest height data, they are not distinguishable if RGB texture is not used. The results also support that fusing multiple modalities is the key to improving the segmentation performance of remote sensing data.\nFor Transformer-based multi-modal fusion methods, similar conclusions can be derived. As presented in Table 4, a simple early-fusion method can boost the mIoU of using the single RGB data from 69% to 75%. Using the proposed TIMF module can achieve a 10% performance improvement, which is significant for segmentation tasks. When looking "},{"figure_ref":[],"heading":"Height Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Early Fusion Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Feature Fusion Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Late Fusion Results","publication_ref":[],"table_ref":[],"text":"Water Tree Building Ground Road Vegetation into the results of different semantic classes, it can be found that the performance of five of the classes can be improved obviously."},{"figure_ref":["fig_10","fig_11"],"heading":"Comparison of Different Fusion Strategies","publication_ref":[],"table_ref":["tab_4","tab_5"],"text":"When comparing the performance of using the RGB data and height data, we can find that in most cases, using RGB performs better than using the height modality for the pixelwise segmentation task. The reason is that RGB can provide rich textures of different semantic objects, which are important for learning discriminative representations. Nevertheless, using the height modality can clearly obtain better performance on the building and tree classes, as they are more sensitive to the geometry information. Thus, by designing effective multi-modal learning methods, complementary features can be learned and the performance can be largely improved.\nFor CNN-based methods, among different multi-modal fusion strategies, we find that feature-fusion and late-fusion are generally more effective than early-fusion. This makes sense because more diverse representations can be learned from complementary modalities with more sophisticated fusion methods. By comparing the results in Table 3, we can see that, the performance of late-fusion is relatively higher than early-fusion and feature-fusion methods in general. Some qualitative results are visualized in Fig. 8. The segmentation results of using different multi-modal fusion strategies are visualized for a clear comparison. It can be seen that feature-fusion and late-fusion models can obtain greatly better segmentation results.\nAs for Transformer-based methods, it can be seen from the results in Table 4 that late-fusion can even outperform the cross-attention-based fusion method. This shares a similar conclusion with CNN-based methods. Late-fusion works surprisingly well for the RGBH segmentation task. These insights can be helpful for future research to design more effective multi-modal learning models. For Transformerbased fusion methods, we visualize the segmentation maps of different multi-modal fusion strategies in Fig. 9."},{"figure_ref":[],"heading":"Comparison between CNN and Transformer-based","publication_ref":["b23","b34"],"table_ref":[],"text":"Methods. From the experimental results, it can be clearly seen that Transformer-based segmentation models can achieve much better performance than CNN-based methods. Even using a single RGB modality, the performance is much higher than CNN-based methods. We attribute the performance improvement to two reasons. One is that CMX (Liu et al., 2022) using SegFormer (Xie et al., 2021) as a baseline can achieve better performance owing to its global spatial context and self-attention mechanism for learning better representations. The second reason is that Transformerbased methods can fuse multi-modal features at a more flexible token level."},{"figure_ref":[],"heading":"Comparisons of State-of-the-art Multi-modal Learning Models.","publication_ref":["b3"],"table_ref":["tab_6"],"text":"We compare seven existing methods that are designed for the multi-modal segmentation task. The results in Table 5 indicate that designing better feature-fusion methods is useful to improve the segmentation performance. The results from ShapeConv (Cao et al., 2021) and VCD (Xiong et al., 2020a) reveal that making better use of the geometry information in the height modality can help improve the segmentation performance. Compared with existing methods, TIMF module achieves the best mIoU performance, which demonstrates the effectiveness of the proposed method.\nThe multi-modal semantic segmentation task is still under-explored, and the value of the extra height modality is still not fully utilized. More research efforts are required to further improve the performance."},{"figure_ref":[],"heading":"Conclusion","publication_ref":[],"table_ref":[],"text":"In this work, we focus on the multi-modal segmentation task, where two modalities (RGB and nDSM (height)) are jointly used to improve the segmentation performance. It is still an under-explored field in remote sensing due to the lack of large-scale datasets and unified benchmarks. This leads to difficulties in comparing the effectiveness of different algorithms. Thus, it is still not clear which type of fusion method is suitable for remote sensing data. To cope with these problems, in this work, we introduce a new remote-sensing "},{"figure_ref":[],"heading":"Height Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Early Fusion Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Feature Fusion Results","publication_ref":[],"table_ref":[],"text":""},{"figure_ref":[],"heading":"Late Fusion Results","publication_ref":[],"table_ref":[],"text":"Water Tree Building Ground Road Vegetation Intermediary Fusion Results is an important step towards a fair and comprehensive benchmarking on multi-modal learning with RGB and geometric modality in remote sensing and earth observation."},{"figure_ref":[],"heading":"Acknowledgement","publication_ref":[],"table_ref":[],"text":"The work is jointly supported by German Federal Ministry for Economic Affairs and Climate Action in the framework of the \"national center of excellence ML4Earth\" (grant number: 50EE2201C), by the German Federal Ministry of Education and Research (BMBF) in the framework of the international future AI lab \"AI4EO -Artificial Intelligence for Earth Observation: Reasoning, Uncertainties, Ethics and Beyond\" (grant number: 01DD20001) and by the Helmholtz Association through the Framework of the Helmholtz Excellent Professorship \"Data Science in Earth Observation -Big Data Fusion for Urban Research\"(grant number: W2-W3-100)."}],"string":"[\n {\n \"figure_ref\": [],\n \"heading\": \"Introduction\",\n \"publication_ref\": [\n \"b44\",\n \"b19\",\n \"b28\",\n \"b38\",\n \"b26\",\n \"b34\",\n \"b25\",\n \"b37\",\n \"b21\",\n \"b33\",\n \"b6\",\n \"b16\",\n \"b18\",\n \"b27\",\n \"b12\"\n ],\n \"table_ref\": [\n \"tab_0\"\n ],\n \"text\": \"Semantic segmentation, aiming at assigning semantic labels for each image pixel, is a fundamental and long-standing goal of both the computer vision (CV) and remote sensing (RS) fields (Zhu et al., 2017;Kuznietsov et al., 2017;Mou and Zhu, 2018;Xiong et al., 2022). With the advent of deep learning techniques, RGB image-based semantic segmentation has attracted great research attention, and significant progress has been made on performance (Long et al., 2015;Xie et al., 2021;Liu et al., 2021). Despite the richness of texture information in RGB images, semantic segmentation models often face challenges in extracting discriminative features from them, owing to the inherent limitations of 2D RGB representations. Considering this problem, multimodal image segmentation is becoming more and more popular in both the CV and RS communities. Multi-modal data such as RGB-Depth (RGB-D) (Xiong et al., 2021) and RGB-Thermal (RGB-T) (Li et al., 2019), contain richer information compared with traditional RGB images, and are widely used for many tasks (Wang and Neumann, 2018;Cheng et al., 2017;Jiang et al., 2018b;Xiong et al., 2020b). Multi-modal data, which incorporates additional modalities beyond RGB, has shown promising results in achieving significantly improved performance compared to using RGB data alone. Remote sensing (RS) data presents several rich data modalities beyond RGB images, such as hyperspectral (HS), multi-spectral (MS), light detection and ranging (LiDAR), normalized digital surface model (nDSM), and synthetic aperture radar (SAR). As a result, multi-modal or multi-sensor data fusion has become a vital area of research in the remote sensing community.\\nThere are various modalities widely used in RS for different Earth observation applications. The development of multi-modal benchmark datasets is critical for advancing research in multi-modal learning methods. However, for semantic segmentation on remote sensing (RS) data, existing multi-modal benchmark datasets have several limitations, given the application areas and data acquisition cost. The foremost limitation is the limited spatial resolutions of the available multi-modal datasets. While Sentinel-1 and Sentinel-2 images are widely-used satellite data with multiple modalities, their low spatial resolution fails to provide detailed land-use and land-cover information. The second limitation pertains to the limited geospatial coverage of existing multi-modal datasets. Although the ISPRS Potsdam and Vaihingen datasets have a higher spatial resolution of 5 cm (0.05m), obtaining such high-resolution images for large-scale real-world applications could be cost-prohibitive. The third limitation of the current multi-modal benchmark datasets is the lack of a unified benchmark platform to enable comprehensive and fair comparison of different multi-modal learning methods. have proven that using the geometric information in nDSM leads to a higher segmentation performance on remote sensing data. However, as presented in Table 1, existing multi-modal datasets that contain RGB and nDSM pairs are relatively small. This makes it difficult to faithfully compare the effectiveness of different types of multi-modal representation learning methods. These limitations motivate us to build a new multi-modal semantic segmentation dataset to enable a fair and unified evaluation of different multimodal segmentation methods. The proposed Geometry-Aware Multi-modal Segmentation (GAMUS) dataset contains images with a resolution of 0.33m, which is higher enough to be used in many real-world applications. The GAMUS dataset contains two data modalities: RGB images and normalized digital surface models (nDSM). Since nDSM indicates the height of ground objects, we use the height modality to represent the nDSM data. nDSM has been broadly provided by many cities owing to its importance in 3D city modeling. Unlike depth images in CV datasets, different types of land covers usually have unique height attributes. Thus, the geometric information contained in height maps is highly correlated to the semantic information (Kunwar, 2019;Mahmud et al., 2020). In other words, compared with other modalities, height information from nDSM has great potential in improving the segmentation performance of high-resolution remote sensing images. Although extensive works have been proposed to make better use of the diverse information contained in different modalities (Hong et al., 2021), there is still a lack of comprehensive benchmarking of existing multi-modal learning methods on RS data. There are many widely used strategies for multi-modal data fusion, including early fusion, featurelevel fusion, late-fusion, and Transformer-based token fusion. However, it is still not clear which one is more suitable for the pixel-level semantic segmentation task on RGB-H (Height) data. Thus, the great potential of the nDSM modality is heavily overlooked by existing works.\\nConsidering this problem, in this work, we propose a simple yet effective Transformer-based Intermediary Multimodal Fusion (TIMF) Module. TIMF utilizes an intermediary learnable token to fuse the features of RGB and height modalities via the self-attention mechanism. To provide comprehensive benchmarking results, we conduct extensive experiments to evaluate Convolutional Neural Networks (CNN) and Transformer-based methods and different variants of fusion strategies. We believe that the benchmark dataset and released codebase can foster future research on evaluating and developing new multi-modal learning methods for Earth observation applications. We summarize the main contributions of this paper below. The remainder of this paper is structured as follows. Section 2 provides an overview of the related methods of multi-modal learning and existing multi-modal datasets for remote sensing (RS) data. Section 3 presents detailed information about the proposed GAMUS dataset. In Section 4, we describe the CNN and Transformer-based multi-modal learning methods with different fusion strategies, along with the designed Transformer-based Intermediary Multi-modal Fusion (TIMF) module. Finally, in Section 5, we present and analyze the benchmarking results obtained through comprehensive and fair comparisons of the proposed models on the GAMUS dataset.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Related Work\",\n \"publication_ref\": [\n \"b41\",\n \"b3\",\n \"b9\",\n \"b5\",\n \"b4\",\n \"b8\",\n \"b7\",\n \"b42\",\n \"b10\",\n \"b29\",\n \"b22\",\n \"b14\",\n \"b6\",\n \"b31\",\n \"b32\"\n ],\n \"table_ref\": [],\n \"text\": \"Multi-modal Representation Learning for Computer Vision Multi-modal feature learning is commonly studied in RGB-D image-based scene recognition (Yuan et al., 2019), semantic segmentation (Cao et al., 2021;Ha et al., 2017;Chen et al., 2020Chen et al., , 2021)), object detection (Gupta et al., 2014;Fan et al., 2020), and action recognition (Zhang et al., 2016). Multi-modal fusion is the key to designing multi-branch deep networks. Existing methods usually conduct fusion at the image level (early fusion) or the feature level (intermediate fusion) (Jiang et al., 2018a). FuseNet (Hazirbas et al., 2016) and RedNet (Jiang et al., 2018a) summed RGB and depth features to obtain multi-modal representations for RGB-D semantic segmentation. Multilevel feature fusion (Park et al., 2017) was designed to extend the residual learning idea of RefineNet (Lin et al., 2017) for RGB-D image segmentation. Similarly, ACNet (Hu et al., 2019) and Gated Fusion Net (Cheng et al., 2017) were proposed to adaptively fuse features of different modalities for image segmentation. PSTNet (Shivakumar et al., 2020) and RTFNet (Sun et al., 2019) were proposed to utilize long wave infrared (LWIR) imagery as a viable supporting modality for semantic segmentation using multi-modal learning networks.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Multi-modal Representation Learning for Remote Sensing\",\n \"publication_ref\": [\n \"b17\",\n \"b0\",\n \"b1\",\n \"b24\",\n \"b43\"\n ],\n \"table_ref\": [],\n \"text\": \"For remote sensing data, (Kampffmeyer et al., 2016) designed a multi-modal deep network with an early deep fusion architecture by stacking all modalities as the input. A promising performance has been achieved on the semantic segmentation of urban images. Audebert et al. (Audebert et al., 2017) proposed to combine the optical and OpenStreetMap data using a two-stream multi-modal learning network to improve the segmentation performance. Exploring the combination of Multi-spectral images (MSI) and Lidar is closer to our work. Audebert et al. (Audebert et al., 2018) introduced a SegNet-based multi-modal fusion architecture for the segmentation of urban scenes. Crossmodality learning (CML) was investigated by (Hong et al., 2020b,a). In (Hong et al., 2020a), a cross-fusion strategy was proposed for learning multi-modal features with more balanced weight contributions from different modalities. Liu et al. (Liu et al., 2019) designed a high-order conditional random field (CRF) based method for the fusion of optical and Lidar predictions in a late-fusion manner. G2GNet (Zheng et al., 2021) was proposed to combine the complementary merits of RGB and auxiliary modality data using the introduced gather-to-guide module.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Remote Sensing Datasets for Multi-modal Semantic Segmentation\",\n \"publication_ref\": [\n \"b30\",\n \"b2\",\n \"b20\"\n ],\n \"table_ref\": [\n \"tab_0\"\n ],\n \"text\": \"The two widely used high-resolution semantic segmentation benchmark datasets in remote sensing are the ISPRS Potsdam and Vaihingen datasets (Rottensteiner et al., 2014). As compared in Table 1, The US3D (Bosch et al., 2019) dataset is collected from public data. Although it is large in volume (more than 700 GB) and contains data for many different tasks, it is not designed dedicated to multi-modal segmentation tasks. In contrast, the GAMUS dataset is smaller in volume, easier to use, and covers more cities. We provide a standard dataloader, detailed instructions, strong baselines, and extensive benchmark results for multi-modal learning. The DFC 19 (Le Saux et al., 2019) dataset is derived from US3D, which is smaller in scale. \\n.111止--------一------ 1111111■■ 一------一-一-------1111 •• _____-一----- 600 400 200 0 .11111111■■---一一一一一一一-一- 1111111■■■-一一---- 111111■■-----------一-\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"The GAMUS Dataset\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [\n \"fig_1\"\n ],\n \"heading\": \"Data collection\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"High-resolution orthophotos, semantic maps, and nDSM (height) are derived and processed from open Data DC catalog (https://opendata.dc.gov) and open data in the Philadelphia region (https://www.opendataphilly.org). As illustrated in Fig. 2, nDSMs are derived from Lidar point clouds. Firstly, noises in the point clouds are removed. Then the height values are rasterized into DSMs (digital surface models) with all the points, and DTMs (digital terrain models) with only the ground points. Finally, subtracting DTM from DSM gives nDSM. The classified point clouds are also ingredients for semantic maps when land cover maps are not available from open data sources. This is simply done by rasterizing the class labels. All the processed data are aligned and cropped into patches to make the final dataset.\"\n },\n {\n \"figure_ref\": [\n \"fig_1\",\n \"fig_3\",\n \"fig_6\",\n \"fig_7\"\n ],\n \"heading\": \"Statistics of the GAMUS dataset\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The introduced GAMUS dataset contains 11,507 tiles collected from five different cities: Oklahoma, Washington, D.C., Philadelphia, Jacksonville, and New York City. These image tiles are collected using the aforementioned data collection process, as shown in Fig. 2. Each RGB image tile has a corresponding nDSM map with a spatial size of 1024×1024. We split all the image tiles into three subsets: the training set with 6,304 tiles, the validation set with 1,059 tiles, and the test set with 4,144 tiles. All the image pixels are annotated with six different land cover types, including 1. ground; 2. low-vegetation; 3. building; 4. water; 5. road; 6. tree. The height statistics provided in the nDSM are displayed in Fig. 3. It can be seen from this figure, there is an obvious long-tailed distribution of the heights, i.e., the number of pixels with lower height values is significantly more than those with higher height values. We also show the statistics of the spatial distribution by averaging the height values at each image pixel across the whole dataset. We can see that the spatial distributions are different for the train, validation, and test subsets.\\nFurthermore, we also display the height distributions of different cities in Fig. 4. For all the cities, the height values obey clear long-tailed distributions. We also visualize the spatial patterns of the height maps by averaging all the height values for each position. We can see that the spatial patterns of different cities are quite different. In Fig. 5, we present the distributions of semantic labels for different cities. This analysis reveals significant differences in the distribution of semantic objects across different cities.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Methods\",\n \"publication_ref\": [\n \"b39\"\n ],\n \"table_ref\": [],\n \"text\": \"Existing deep learning models for multi-modal semantic segmentation can be divided into CNN-based and Transformerbased networks. For CNN-based methods, the fusion in different layers has a great influence on the performance. For Transformer-based methods, the self-attention mechanism can be used to fuse multi-modal features at a token level (Xu et al., 2022), which can be more effective than CNN-based methods.\"\n },\n {\n \"figure_ref\": [\n \"fig_8\"\n ],\n \"heading\": \"CNN-based Fusion Methods\",\n \"publication_ref\": [\n \"b3\",\n \"b10\",\n \"b9\",\n \"b32\"\n ],\n \"table_ref\": [],\n \"text\": \"For CNN-based fusion methods, we evaluate the performance using five different training paradigms as shown in Fig. 6. To formally define these training paradigms, we first introduce the notations of input modalities and networks. We use 𝑿 𝑟𝑔𝑏 ∈ ℝ 3×𝐻×𝑊 and 𝑿 ℎ ∈ ℝ 𝐻×𝑊 to denote the RGB input and height from nDSM, respectively. For the sake of \\n(𝑋 𝑟𝑔𝑏 )) + 𝑓 𝑑 𝐖 ℎ (𝑓 𝑒 𝐖 ℎ (𝑋 ℎ )).\\nSince the height information in the nDSM data is highly similar to the geometric information in the depth images, we select three RGB-D multi-modal learning methods for benchmarking. ShapeConv (Cao et al., 2021), VCD (Xiong et al., 2020a) aims at designing a cross-modality guided encoder to better fuse RGB features and depth information. FuseNet (Hazirbas et al., 2016) focuses on designing better multi-scale feature fusion architectures. Considering the diversity of different multi-modal architecture designs, we also choose two RGB-T segmentation methods for performance evaluation on our GAMUS dataset. Multi-spectral fusion networks (MFNet) (Ha et al., 2017) takes RGB images and IR images as input and fuse them in the feature level for multi-modal learning. RTFNet (Sun et al., 2019) uses two separate encoders for RGB and thermal modalities and fuses their intermediate features progressively to learn multimodal representations.\\nFor the performance evaluation of multiple modalities, we adapt three methods (MFNet, RTFNet, and FuseNet) to five types of multi-modal fusion paradigms to compare the performance. Specifically, for single modality settings, we modify them by removing one input modality and only remain a single RGB or height modality. For the earlyfusion method, we change the model by simply stacking the RGB and height modality together as the multi-modal input. As for the feature-level fusion, we use the same network architectures as described in the original papers of the eight compared methods. For the late fusion setting, we only fuse the predicted segmentation results of each modality at the last layer of the segmentation head. Toward fair comparisons, we exploit the exact same network architectures and training hyper-parameters for methods trained in different experimental settings.\"\n },\n {\n \"figure_ref\": [\n \"fig_9\",\n \"fig_9\"\n ],\n \"heading\": \"Transformer-based Fusion Methods\",\n \"publication_ref\": [\n \"b23\",\n \"b34\",\n \"b23\"\n ],\n \"table_ref\": [],\n \"text\": \"CMX (Liu et al., 2022) introduces a cross-modal feature rectification module to calibrate the feature of one modality from another in both spatial-and channel-wise dimensions. SegFormer (Xie et al., 2021) is used as its baseline segmentation architecture. Considering its effectiveness, we take CMX as our baseline method and adapt it to five different fusion strategies for performance comparison. Fig. 7 presents six types of fusion methods. The first five fusion methods including single modality, early fusion, late fusion, and cross-modal feature fusion (Liu et al., 2022) are similar to the five CNN-based fusion strategies. Different from these five fusion methods, in this work, we propose the intermediary feature fusion module (illustrated in Fig. 7 (f)), which is described in the following section.\\nTo formally describe these Transformer-based fusion models, we first give a brief introduction to the widelyused vision Transformer (ViT). Given the input data 𝒙, a Transformer encoder first projects the image patches into patch embeddings (denoted as 𝑝(⋅)) and add position embeddings 𝒙 𝑝𝑜𝑠 to enhance the position information. Then the input will further go through alternating layers of multihead self-attention (MSA) and multi-layer perception (MLP) blocks. Layer Normalization (LN) and residual connection are applied before and after every block. The computation process of a Transformer block TL(𝒙) can be expressed as\\n𝒙 0 = 𝑝(𝒙) + 𝒙 𝑝𝑜𝑠 , 𝒛 ′ 𝑘 = 𝒛 𝑘-1 + MSA(LN(𝒙 𝑘-1 )), 𝒛 𝑘 = 𝒛 ′ 𝑘 + MLP(LN(𝒛 ′ 𝑘 )),(1)\\nwhere 𝒛 𝑘 is the output features, and 𝑘 is the index of blocks.\\nLet 𝐹 𝑒 𝑚 (⋅) and 𝐹 𝑑 𝑚 (⋅) be the encoder and decoder of the Segformer () networks, where 𝑚 could be either the RGB or the height modality. The single modality model with RGB input can be defined as 𝑦 * 𝑠 = 𝐹 𝑑 (𝐹 𝑒 (𝑋 𝑟𝑔𝑏 )). The computation for the height modality can be defined by simply replacing 𝑋 𝑟𝑔𝑏 with 𝑋 𝑑 . The early fusion strategy can be expressed as 𝑦 * 𝑠 = 𝐹 𝑑 (𝐹 𝑒 (𝑋 𝑟𝑔𝑏 ∥ 𝑋 𝑑 )). Similarly, the late fusion model is formulated as\\nTransformer Q K V Transformer Q K V V K Q Transformer Transformer V K Q Q K V Transformer Q K V Transformer Transformer Q K V(\\n𝑦 * 𝑠 = 𝐹 𝑑 𝑟𝑔𝑏 (𝐹 𝑒 𝑟𝑔𝑏 (𝑋 𝑟𝑔𝑏 )) + 𝐹 𝑑 ℎ (𝐹 𝑒 𝑑 (𝑋 ℎ )).\\n(2)\\nThe Transformer-based cross feature fusion is conducted by exchanging queries of different modalities in the MSA blocks. The other computation steps are the same as a Transformer block. For the sake of simplicity, we omit them and only present the computation of MSA, which can be expressed as:\\n𝒛 𝑟𝑔𝑏 ⟵ MSA(LN(𝑿 ℎ , 𝑿 𝑟𝑔𝑏 , 𝑿 𝑟𝑔𝑏 ⏟⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏟ 𝑄, 𝐾, 𝑉\\n)),\\n𝒛 ℎ ⟵ MSA(LN(𝑿 𝑟𝑔𝑏 , 𝑿 ℎ , 𝑿 ℎ ⏟⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏟ 𝑄, 𝐾, 𝑉\\n)).\\n(3)\\nBy this means, the multi-modal features are fused via cross-modal self-attention modules, which is more flexible than layer-wise feature fusion.\"\n },\n {\n \"figure_ref\": [\n \"fig_9\"\n ],\n \"heading\": \"Intermediary Multi-modal Fusion\",\n \"publication_ref\": [\n \"b23\"\n ],\n \"table_ref\": [],\n \"text\": \"Different from existing fusion strategies, in this work, we propose a novel Transformer-based intermediary multimodal fusion method, i.e., the TIMF module. TIMF exploits an additional intermediary token 𝑴 to adaptively combine RGB and height features at a token level via the self-attention mechanism. To be more specific, TIMF is designed in a hierarchical manner. As illustrated in Fig. 7 (f), the intermediary token first extracts global multi-modal features by feeding the intermediary token into an MSA module, which is similar to the CLS token in ViT. Next, the output intermediary token 𝑴 ′ is concatenated to the tokens of each modality respectively. Then, single-block (𝑘 = 1) Transformer layers are used to further fuse the intermediary token 𝑴 ′ with modality-specific tokens via self-attention modules. Formally, the first stage of TIMF takes 𝑴 and tokens of both modalities as input to a singleblock Transformer layer, which can be formulated as\\n𝒛 𝑎 = TL(𝑝(𝑿 𝑟𝑔𝑏 ) ∥ 𝑝(𝑿 𝑑 ) ∥ 𝑀), 𝒛 𝑎 = (𝒛 𝑟𝑔𝑏 ) ∥ (𝒛 ℎ ) ∥ 𝑀 ′ ),(4)\\nwhere 𝒛 𝑎 is the output feature. ∥ is the concatenation operation. By slicing 𝒛 𝑎 according to the number of tokens, we can obtain the output features of RGB modality 𝒛 𝑟𝑔𝑏 ), height modality 𝒛 ℎ ), and the output intermediary token 𝑴 ′ . Then, the final output features of the TIMF module are obtained by fusing 𝑴 ′ with these two modalities individually using two Transformer layers. This can be formulated as\\n𝒛 ′ 𝑟𝑔𝑏 = TL 𝑟𝑔𝑏 (𝑝(𝒛 𝑟𝑔𝑏 ) ∥ 𝑴 ′ ), 𝒛 ′ 𝑟𝑔𝑏 = (𝒛 𝑟𝑔𝑏 ) ∥ 𝑴 ′ 𝑟𝑔𝑏 ), 𝒛 ′ ℎ = TL ℎ (𝑝(𝒛 ℎ ) ∥ 𝑴 ′ ), 𝒛 ′ ℎ = ((𝒛 ℎ ) ∥ 𝑴 ′ ℎ ).\\n(5)\\nIn practice, TIMF module can be used in a plug-andplay manner. We simply replace the TIMF module with the feature fusion module in CMX (Liu et al., 2022). \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Experiments\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The constructed dataset GAMUS in this paper can be publicly accessed at https://github.com/EarthNets/Dataset4EO. In this dataset, we have explicitly provided the official splits for training, validation, and test subsets. Providing a well-defined data set and the official split is crucial for reproducibility.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Implementation details\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"All the evaluated models are implemented using Pytorch and run on eight GeForce RTX 3090 GPUs. We use the public codebase provided by the original paper. For all the supervised multi-modal segmentation methods, 100 epochs are used for training. The input images are resized to 512×512. The batch sizes are set to 16, and we use the same learning rate and optimizer as described in the original paper. For data augmentation, random flip and random crop are used for all the methods towards a fair comparison. On the GAMUS dataset, 50 epochs are used for the model training. The implemented dataset loading code can be found in https: //github.com/EarthNets/Dataset4EO. For all the experiments, the mean Intersection over Union (mIoU) and IoU of each class are used as the evaluation metrics.\\nTo ensure reproducibility, we choose to make all the used source codes publicly available at https://github.com/ EarthNets/RSI-MMSegmentation.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Multi-modal Learning Analysis\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\",\n \"tab_5\",\n \"tab_3\",\n \"tab_4\",\n \"tab_5\"\n ],\n \"text\": \"To analyze the benefits of using multiple modalities for semantic segmentation, we compare the experimental results of using different data modalities as input in Table 2, Table 3, andTable 4. Specifically, for CNN-based fusion methods, we present the accuracy and mIoU results in Table 2 andTable 3. As for the Transformer-based methods, we provide the mIoU results of six different types of modality inputs in Table 4. In the following, we analyze and discuss these extensive results from five different aspects.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Benefits of Multi-modal Learning\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\",\n \"tab_5\"\n ],\n \"text\": \"Despite the fact that early-fusion is the simplest multi-modal learning method, stacking RGB and height map as a four-channel input, can still obtain clearly better results than using a single RGB modality. As presented in Table 3, for CNN-based methods, using early-fusion, MFNet can improve the mIoU of RGB modality from 46.8% to 49.6%. FuseNet can obtain a 9% improvement than only using the RGB data. These results reveal the effectiveness of multi-modal learning for the segmentation of RS images. Furthermore, by exploiting more sophisticated feature-level multi-modal fusion methods, the performance can be further improved. For example, MFNet with a feature-fusion strategy can obtain a 6% improvement to using a single RGB modality. However, surprisingly, for FuseNet, the performances of the simple early-fusion methods may outperform feature-level fusion methods. In general, from the results, we can clearly see that using both the RGB and height modalities can obtain much better results compared to using single modalities.\\nComparing the results of using RGB and RGBH data, tree and building are the semantic classes with the greatest performance boost when the extra height modality is used. This makes sense because trees and buildings are objects with clearly higher height values. Comparing the results of using height and RGBH data, there is significant performance improvement. Specifically, ground and vegetation are the semantic classes with the greatest performance boost when the extra RGB modality is used. As ground and vegetation always have the lowest height data, they are not distinguishable if RGB texture is not used. The results also support that fusing multiple modalities is the key to improving the segmentation performance of remote sensing data.\\nFor Transformer-based multi-modal fusion methods, similar conclusions can be derived. As presented in Table 4, a simple early-fusion method can boost the mIoU of using the single RGB data from 69% to 75%. Using the proposed TIMF module can achieve a 10% performance improvement, which is significant for segmentation tasks. When looking \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Height Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Early Fusion Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Feature Fusion Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Late Fusion Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Water Tree Building Ground Road Vegetation into the results of different semantic classes, it can be found that the performance of five of the classes can be improved obviously.\"\n },\n {\n \"figure_ref\": [\n \"fig_10\",\n \"fig_11\"\n ],\n \"heading\": \"Comparison of Different Fusion Strategies\",\n \"publication_ref\": [],\n \"table_ref\": [\n \"tab_4\",\n \"tab_5\"\n ],\n \"text\": \"When comparing the performance of using the RGB data and height data, we can find that in most cases, using RGB performs better than using the height modality for the pixelwise segmentation task. The reason is that RGB can provide rich textures of different semantic objects, which are important for learning discriminative representations. Nevertheless, using the height modality can clearly obtain better performance on the building and tree classes, as they are more sensitive to the geometry information. Thus, by designing effective multi-modal learning methods, complementary features can be learned and the performance can be largely improved.\\nFor CNN-based methods, among different multi-modal fusion strategies, we find that feature-fusion and late-fusion are generally more effective than early-fusion. This makes sense because more diverse representations can be learned from complementary modalities with more sophisticated fusion methods. By comparing the results in Table 3, we can see that, the performance of late-fusion is relatively higher than early-fusion and feature-fusion methods in general. Some qualitative results are visualized in Fig. 8. The segmentation results of using different multi-modal fusion strategies are visualized for a clear comparison. It can be seen that feature-fusion and late-fusion models can obtain greatly better segmentation results.\\nAs for Transformer-based methods, it can be seen from the results in Table 4 that late-fusion can even outperform the cross-attention-based fusion method. This shares a similar conclusion with CNN-based methods. Late-fusion works surprisingly well for the RGBH segmentation task. These insights can be helpful for future research to design more effective multi-modal learning models. For Transformerbased fusion methods, we visualize the segmentation maps of different multi-modal fusion strategies in Fig. 9.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Comparison between CNN and Transformer-based\",\n \"publication_ref\": [\n \"b23\",\n \"b34\"\n ],\n \"table_ref\": [],\n \"text\": \"Methods. From the experimental results, it can be clearly seen that Transformer-based segmentation models can achieve much better performance than CNN-based methods. Even using a single RGB modality, the performance is much higher than CNN-based methods. We attribute the performance improvement to two reasons. One is that CMX (Liu et al., 2022) using SegFormer (Xie et al., 2021) as a baseline can achieve better performance owing to its global spatial context and self-attention mechanism for learning better representations. The second reason is that Transformerbased methods can fuse multi-modal features at a more flexible token level.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Comparisons of State-of-the-art Multi-modal Learning Models.\",\n \"publication_ref\": [\n \"b3\"\n ],\n \"table_ref\": [\n \"tab_6\"\n ],\n \"text\": \"We compare seven existing methods that are designed for the multi-modal segmentation task. The results in Table 5 indicate that designing better feature-fusion methods is useful to improve the segmentation performance. The results from ShapeConv (Cao et al., 2021) and VCD (Xiong et al., 2020a) reveal that making better use of the geometry information in the height modality can help improve the segmentation performance. Compared with existing methods, TIMF module achieves the best mIoU performance, which demonstrates the effectiveness of the proposed method.\\nThe multi-modal semantic segmentation task is still under-explored, and the value of the extra height modality is still not fully utilized. More research efforts are required to further improve the performance.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Conclusion\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"In this work, we focus on the multi-modal segmentation task, where two modalities (RGB and nDSM (height)) are jointly used to improve the segmentation performance. It is still an under-explored field in remote sensing due to the lack of large-scale datasets and unified benchmarks. This leads to difficulties in comparing the effectiveness of different algorithms. Thus, it is still not clear which type of fusion method is suitable for remote sensing data. To cope with these problems, in this work, we introduce a new remote-sensing \"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Height Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Early Fusion Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Feature Fusion Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Late Fusion Results\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"Water Tree Building Ground Road Vegetation Intermediary Fusion Results is an important step towards a fair and comprehensive benchmarking on multi-modal learning with RGB and geometric modality in remote sensing and earth observation.\"\n },\n {\n \"figure_ref\": [],\n \"heading\": \"Acknowledgement\",\n \"publication_ref\": [],\n \"table_ref\": [],\n \"text\": \"The work is jointly supported by German Federal Ministry for Economic Affairs and Climate Action in the framework of the \\\"national center of excellence ML4Earth\\\" (grant number: 50EE2201C), by the German Federal Ministry of Education and Research (BMBF) in the framework of the international future AI lab \\\"AI4EO -Artificial Intelligence for Earth Observation: Reasoning, Uncertainties, Ethics and Beyond\\\" (grant number: 01DD20001) and by the Helmholtz Association through the Framework of the Helmholtz Excellent Professorship \\\"Data Science in Earth Observation -Big Data Fusion for Urban Research\\\"(grant number: W2-W3-100).\"\n }\n]","isConversational":false},"pub_date":{"kind":"string","value":"2023-05-24"},"doi":{"kind":"string","value":"10.21227/c6tm-vw12"},"references":{"kind":"list like","value":[{"authors":"N Audebert; B Le Saux; S Lefèvre","journal":"","ref_id":"b0","title":"Joint learning from earth observation and openstreetmap data to get faster better semantic maps","year":"2017"},{"authors":"N Audebert; B Le Saux; S Lefèvre","journal":"ISPRS journal of photogrammetry and remote sensing","ref_id":"b1","title":"Beyond rgb: Very high resolution urban remote sensing with multimodal deep networks","year":"2018"},{"authors":"M Bosch; K Foster; G Christie; S Wang; G D Hager; M Brown","journal":"IEEE","ref_id":"b2","title":"Semantic stereo for incidental satellite images","year":"2019"},{"authors":"J Cao; H Leng; D Lischinski; D Cohen-Or; C Tu; Y Li","journal":"","ref_id":"b3","title":"Shapeconv: Shape-aware convolutional layer for indoor rgb-d semantic segmentation","year":"2021"},{"authors":"L Z Chen; Z Lin; Z Wang; Y L Yang; M M Cheng","journal":"IEEE Transactions on Image Processing","ref_id":"b4","title":"Spatial information guided convolution for real-time rgbd semantic segmentation","year":"2021"},{"authors":"X Chen; K Y Lin; J Wang; W Wu; C Qian; H Li; G Zeng","journal":"Springer","ref_id":"b5","title":"Bi-directional cross-modality feature propagation with separation-andaggregation gate for rgb-d semantic segmentation","year":"2020"},{"authors":"Y Cheng; R Cai; Z Li; X Zhao; K Huang","journal":"","ref_id":"b6","title":"Locality-sensitive deconvolution networks with gated fusion for RGB-D indoor semantic segmentation","year":"2017"},{"authors":"D P Fan; Z Lin; Z Zhang; M Zhu; M M Cheng","journal":"IEEE Transactions on Neural Networks and Learning Systems","ref_id":"b7","title":"Rethinking rgbd salient object detection: Models, data sets, and large-scale benchmarks","year":"2020"},{"authors":"S Gupta; R Girshick; P Arbeláez; J Malik","journal":"Springer","ref_id":"b8","title":"Learning rich features from rgb-d images for object detection and segmentation","year":"2014"},{"authors":"Q Ha; K Watanabe; T Karasawa; Y Ushiku; T Harada","journal":"IEEE","ref_id":"b9","title":"Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes","year":"2017"},{"authors":"C Hazirbas; L Ma; C Domokos; D Cremers","journal":"","ref_id":"b10","title":"Fusenet: Incorporating depth into semantic segmentation via fusion-based cnn architecture","year":"2016"},{"authors":"D Hong; J Chanussot; N Yokoya; J Kang; X X Zhu","journal":"IEEE Geoscience and Remote Sensing Letters","ref_id":"b11","title":"Learningshared cross-modality representation using multispectral-lidar and hyperspectral data","year":"2020"},{"authors":"D Hong; J Hu; J Yao; J Chanussot; X X Zhu","journal":"ISPRS Journal of Photogrammetry and Remote Sensing","ref_id":"b12","title":"Multimodal remote sensing benchmark datasets for land cover classification with a shared and specific feature learning model","year":"2021"},{"authors":"D Hong; N Yokoya; G S Xia; J Chanussot; X X Zhu","journal":"ISPRS Journal of Photogrammetry and Remote Sensing","ref_id":"b13","title":"Xmodalnet: A semi-supervised deep cross-modal network for classification of remote sensing data","year":"2020"},{"authors":"X Hu; K Yang; L Fei; K Wang","journal":"","ref_id":"b14","title":"Acnet: Attention based network to exploit complementary features for RGBD semantic segmentation","year":"2019"},{"authors":"J Jiang; L Zheng; F Luo; Z Zhang","journal":"","ref_id":"b15","title":"Rednet: Residual encoderdecoder network for indoor RGB-D semantic segmentation","year":"2018"},{"authors":"M X Jiang; C Deng; M M Zhang; J S Shan; H Zhang","journal":"Complexity","ref_id":"b16","title":"Multimodal deep feature fusion (mmdff) for rgb-d tracking","year":"2018"},{"authors":"M Kampffmeyer; A B Salberg; R Jenssen","journal":"","ref_id":"b17","title":"Semantic segmentation of small objects and modeling of uncertainty in urban remote sensing images using deep convolutional neural networks","year":"2016"},{"authors":"S Kunwar","journal":"IEEE","ref_id":"b18","title":"U-Net ensemble for semantic and height estimation using coarse-map initialization","year":"2019"},{"authors":"Y Kuznietsov; J Stuckler; B Leibe","journal":"","ref_id":"b19","title":"Semi-supervised deep learning for monocular depth map prediction","year":"2017"},{"authors":"Le Saux; B Yokoya; N Hänsch; R Brown; M ","journal":"","ref_id":"b20","title":"Data fusion contest 2019 (dfc","year":"2019"},{"authors":"C Li; X Liang; Y Lu; N Zhao; J Tang","journal":"Pattern Recognition","ref_id":"b21","title":"RGB-T object tracking: benchmark and baseline","year":"2019"},{"authors":"G Lin; A Milan; C Shen; I Reid","journal":"","ref_id":"b22","title":"RefineNet: Multi-path refinement networks for high-resolution semantic segmentation","year":"2017"},{"authors":"H Liu; J Zhang; K Yang; X Hu; R Stiefelhagen","journal":"","ref_id":"b23","title":"Cmx: Crossmodal fusion for RGB-X semantic segmentation with transformers","year":"2022"},{"authors":"Y Liu; S Piramanayagam; S T Monteiro; E Saber","journal":"Journal of Applied Remote Sensing","ref_id":"b24","title":"Semantic segmentation of multisensor remote sensing imagery with deep convnets and higher-order conditional random fields","year":"2019"},{"authors":"Z Liu; Y Lin; Y Cao; H Hu; Y Wei; Z Zhang; S Lin; B Guo","journal":"","ref_id":"b25","title":"Swin Transformer: Hierarchical vision transformer using shifted windows","year":"2021"},{"authors":"J Long; E Shelhamer; T Darrell","journal":"","ref_id":"b26","title":"Fully convolutional networks for semantic segmentation","year":"2015"},{"authors":"J Mahmud; T Price; A Bapat; J M Frahm","journal":"","ref_id":"b27","title":"Boundary-aware 3d building reconstruction from a single overhead image","year":"2020"},{"authors":"L Mou; X X Zhu","journal":"","ref_id":"b28","title":"Im2height: Height estimation from single monocular imagery via fully residual convolutional-deconvolutional network","year":"2018"},{"authors":"S J Park; K S Hong; S Lee","journal":"","ref_id":"b29","title":"Rdfnet: Rgb-d multi-level residual feature fusion for indoor semantic segmentation","year":"2017"},{"authors":"F Rottensteiner; G Sohn; M Gerke; J D Wegner","journal":"ISPRS: Leopoldshöhe","ref_id":"b30","title":"Isprs semantic labeling contest","year":"2014"},{"authors":"S S Shivakumar; N Rodrigues; A Zhou; I D Miller; V Kumar; C J Taylor","journal":"IEEE","ref_id":"b31","title":"Pst900: Rgb-thermal calibration, dataset and segmentation network","year":"2020"},{"authors":"Y Sun; W Zuo; M Liu","journal":"IEEE Robotics and Automation Letters","ref_id":"b32","title":"Rtfnet: Rgb-thermal fusion network for semantic segmentation of urban scenes","year":"2019"},{"authors":"W Wang; U Neumann","journal":"","ref_id":"b33","title":"Depth-aware CNN for RGB-D segmentation","year":"2018-09-08"},{"authors":"E Xie; W Wang; Z Yu; A Anandkumar; J M Alvarez; P Luo","journal":"Advances in Neural Information Processing Systems","ref_id":"b34","title":"Segformer: Simple and efficient design for semantic segmentation with transformers","year":"2021"},{"authors":"Z Xiong; Y Yuan; N Guo; Q Wang","journal":"","ref_id":"b35","title":"Variational contextdeformable convnets for indoor scene parsing","year":"2020"},{"authors":"Z Xiong; Y Yuan; Q Wang","journal":"Neurocomputing","ref_id":"b36","title":"Msn: Modality separation networks for rgb-d scene recognition","year":"2020"},{"authors":"Z Xiong; Y Yuan; Q Wang","journal":"IEEE Transactions on Image Processing","ref_id":"b37","title":"Ask: Adaptively selecting key local features for rgb-d scene recognition","year":"2021"},{"authors":"Z Xiong; F Zhang; Y Wang; Y Shi; X X Zhu","journal":"","ref_id":"b38","title":"EarthNets: Empowering AI in earth observation","year":"2022"},{"authors":"P Xu; X Zhu; D A Clifton","journal":"","ref_id":"b39","title":"Multimodal learning with transformers: A survey","year":"2022"},{"authors":"N Yokoya; P Ghamisi; J Xia; S Sukhanov; R Heremans; I Tankoyeu; B Bechtel; B Le Saux; G Moser; D Tuia","journal":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","ref_id":"b40","title":"Open data for global multimodal land use classification: Outcome of the 2017 ieee grss data fusion contest","year":"2018"},{"authors":"Y Yuan; Z Xiong; Q Wang","journal":"","ref_id":"b41","title":"Acm: Adaptive cross-modal graph convolutional neural networks for rgb-d scene recognition","year":"2019"},{"authors":"J Zhang; W Li; P O Ogunbona; P Wang; C Tang","journal":"Pattern Recognition","ref_id":"b42","title":"Rgb-d-based action recognition datasets: A survey","year":"2016"},{"authors":"X Zheng; X Wu; L Huan; W He; H Zhang","journal":"IEEE Transactions on Geoscience and Remote Sensing","ref_id":"b43","title":"A gather-to-guide network for remote sensing semantic segmentation of rgb and auxiliary image","year":"2021"},{"authors":"X X Zhu; D Tuia; L Mou; G S Xia; L Zhang; F Xu; F Fraundorfer","journal":"IEEE Geoscience and Remote Sensing Magazine","ref_id":"b44","title":"Deep learning in remote sensing: A comprehensive review and list of resources","year":"2017"}],"string":"[\n {\n \"authors\": \"N Audebert; B Le Saux; S Lefèvre\",\n \"journal\": \"\",\n \"ref_id\": \"b0\",\n \"title\": \"Joint learning from earth observation and openstreetmap data to get faster better semantic maps\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"N Audebert; B Le Saux; S Lefèvre\",\n \"journal\": \"ISPRS journal of photogrammetry and remote sensing\",\n \"ref_id\": \"b1\",\n \"title\": \"Beyond rgb: Very high resolution urban remote sensing with multimodal deep networks\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"M Bosch; K Foster; G Christie; S Wang; G D Hager; M Brown\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b2\",\n \"title\": \"Semantic stereo for incidental satellite images\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"J Cao; H Leng; D Lischinski; D Cohen-Or; C Tu; Y Li\",\n \"journal\": \"\",\n \"ref_id\": \"b3\",\n \"title\": \"Shapeconv: Shape-aware convolutional layer for indoor rgb-d semantic segmentation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"L Z Chen; Z Lin; Z Wang; Y L Yang; M M Cheng\",\n \"journal\": \"IEEE Transactions on Image Processing\",\n \"ref_id\": \"b4\",\n \"title\": \"Spatial information guided convolution for real-time rgbd semantic segmentation\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"X Chen; K Y Lin; J Wang; W Wu; C Qian; H Li; G Zeng\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b5\",\n \"title\": \"Bi-directional cross-modality feature propagation with separation-andaggregation gate for rgb-d semantic segmentation\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Y Cheng; R Cai; Z Li; X Zhao; K Huang\",\n \"journal\": \"\",\n \"ref_id\": \"b6\",\n \"title\": \"Locality-sensitive deconvolution networks with gated fusion for RGB-D indoor semantic segmentation\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"D P Fan; Z Lin; Z Zhang; M Zhu; M M Cheng\",\n \"journal\": \"IEEE Transactions on Neural Networks and Learning Systems\",\n \"ref_id\": \"b7\",\n \"title\": \"Rethinking rgbd salient object detection: Models, data sets, and large-scale benchmarks\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"S Gupta; R Girshick; P Arbeláez; J Malik\",\n \"journal\": \"Springer\",\n \"ref_id\": \"b8\",\n \"title\": \"Learning rich features from rgb-d images for object detection and segmentation\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"Q Ha; K Watanabe; T Karasawa; Y Ushiku; T Harada\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b9\",\n \"title\": \"Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"C Hazirbas; L Ma; C Domokos; D Cremers\",\n \"journal\": \"\",\n \"ref_id\": \"b10\",\n \"title\": \"Fusenet: Incorporating depth into semantic segmentation via fusion-based cnn architecture\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"D Hong; J Chanussot; N Yokoya; J Kang; X X Zhu\",\n \"journal\": \"IEEE Geoscience and Remote Sensing Letters\",\n \"ref_id\": \"b11\",\n \"title\": \"Learningshared cross-modality representation using multispectral-lidar and hyperspectral data\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"D Hong; J Hu; J Yao; J Chanussot; X X Zhu\",\n \"journal\": \"ISPRS Journal of Photogrammetry and Remote Sensing\",\n \"ref_id\": \"b12\",\n \"title\": \"Multimodal remote sensing benchmark datasets for land cover classification with a shared and specific feature learning model\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"D Hong; N Yokoya; G S Xia; J Chanussot; X X Zhu\",\n \"journal\": \"ISPRS Journal of Photogrammetry and Remote Sensing\",\n \"ref_id\": \"b13\",\n \"title\": \"Xmodalnet: A semi-supervised deep cross-modal network for classification of remote sensing data\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"X Hu; K Yang; L Fei; K Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b14\",\n \"title\": \"Acnet: Attention based network to exploit complementary features for RGBD semantic segmentation\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"J Jiang; L Zheng; F Luo; Z Zhang\",\n \"journal\": \"\",\n \"ref_id\": \"b15\",\n \"title\": \"Rednet: Residual encoderdecoder network for indoor RGB-D semantic segmentation\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"M X Jiang; C Deng; M M Zhang; J S Shan; H Zhang\",\n \"journal\": \"Complexity\",\n \"ref_id\": \"b16\",\n \"title\": \"Multimodal deep feature fusion (mmdff) for rgb-d tracking\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"M Kampffmeyer; A B Salberg; R Jenssen\",\n \"journal\": \"\",\n \"ref_id\": \"b17\",\n \"title\": \"Semantic segmentation of small objects and modeling of uncertainty in urban remote sensing images using deep convolutional neural networks\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"S Kunwar\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b18\",\n \"title\": \"U-Net ensemble for semantic and height estimation using coarse-map initialization\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Y Kuznietsov; J Stuckler; B Leibe\",\n \"journal\": \"\",\n \"ref_id\": \"b19\",\n \"title\": \"Semi-supervised deep learning for monocular depth map prediction\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"Le Saux; B Yokoya; N Hänsch; R Brown; M \",\n \"journal\": \"\",\n \"ref_id\": \"b20\",\n \"title\": \"Data fusion contest 2019 (dfc\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"C Li; X Liang; Y Lu; N Zhao; J Tang\",\n \"journal\": \"Pattern Recognition\",\n \"ref_id\": \"b21\",\n \"title\": \"RGB-T object tracking: benchmark and baseline\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"G Lin; A Milan; C Shen; I Reid\",\n \"journal\": \"\",\n \"ref_id\": \"b22\",\n \"title\": \"RefineNet: Multi-path refinement networks for high-resolution semantic segmentation\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"H Liu; J Zhang; K Yang; X Hu; R Stiefelhagen\",\n \"journal\": \"\",\n \"ref_id\": \"b23\",\n \"title\": \"Cmx: Crossmodal fusion for RGB-X semantic segmentation with transformers\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"Y Liu; S Piramanayagam; S T Monteiro; E Saber\",\n \"journal\": \"Journal of Applied Remote Sensing\",\n \"ref_id\": \"b24\",\n \"title\": \"Semantic segmentation of multisensor remote sensing imagery with deep convnets and higher-order conditional random fields\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"Z Liu; Y Lin; Y Cao; H Hu; Y Wei; Z Zhang; S Lin; B Guo\",\n \"journal\": \"\",\n \"ref_id\": \"b25\",\n \"title\": \"Swin Transformer: Hierarchical vision transformer using shifted windows\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"J Long; E Shelhamer; T Darrell\",\n \"journal\": \"\",\n \"ref_id\": \"b26\",\n \"title\": \"Fully convolutional networks for semantic segmentation\",\n \"year\": \"2015\"\n },\n {\n \"authors\": \"J Mahmud; T Price; A Bapat; J M Frahm\",\n \"journal\": \"\",\n \"ref_id\": \"b27\",\n \"title\": \"Boundary-aware 3d building reconstruction from a single overhead image\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"L Mou; X X Zhu\",\n \"journal\": \"\",\n \"ref_id\": \"b28\",\n \"title\": \"Im2height: Height estimation from single monocular imagery via fully residual convolutional-deconvolutional network\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"S J Park; K S Hong; S Lee\",\n \"journal\": \"\",\n \"ref_id\": \"b29\",\n \"title\": \"Rdfnet: Rgb-d multi-level residual feature fusion for indoor semantic segmentation\",\n \"year\": \"2017\"\n },\n {\n \"authors\": \"F Rottensteiner; G Sohn; M Gerke; J D Wegner\",\n \"journal\": \"ISPRS: Leopoldshöhe\",\n \"ref_id\": \"b30\",\n \"title\": \"Isprs semantic labeling contest\",\n \"year\": \"2014\"\n },\n {\n \"authors\": \"S S Shivakumar; N Rodrigues; A Zhou; I D Miller; V Kumar; C J Taylor\",\n \"journal\": \"IEEE\",\n \"ref_id\": \"b31\",\n \"title\": \"Pst900: Rgb-thermal calibration, dataset and segmentation network\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Y Sun; W Zuo; M Liu\",\n \"journal\": \"IEEE Robotics and Automation Letters\",\n \"ref_id\": \"b32\",\n \"title\": \"Rtfnet: Rgb-thermal fusion network for semantic segmentation of urban scenes\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"W Wang; U Neumann\",\n \"journal\": \"\",\n \"ref_id\": \"b33\",\n \"title\": \"Depth-aware CNN for RGB-D segmentation\",\n \"year\": \"2018-09-08\"\n },\n {\n \"authors\": \"E Xie; W Wang; Z Yu; A Anandkumar; J M Alvarez; P Luo\",\n \"journal\": \"Advances in Neural Information Processing Systems\",\n \"ref_id\": \"b34\",\n \"title\": \"Segformer: Simple and efficient design for semantic segmentation with transformers\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Z Xiong; Y Yuan; N Guo; Q Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b35\",\n \"title\": \"Variational contextdeformable convnets for indoor scene parsing\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Z Xiong; Y Yuan; Q Wang\",\n \"journal\": \"Neurocomputing\",\n \"ref_id\": \"b36\",\n \"title\": \"Msn: Modality separation networks for rgb-d scene recognition\",\n \"year\": \"2020\"\n },\n {\n \"authors\": \"Z Xiong; Y Yuan; Q Wang\",\n \"journal\": \"IEEE Transactions on Image Processing\",\n \"ref_id\": \"b37\",\n \"title\": \"Ask: Adaptively selecting key local features for rgb-d scene recognition\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"Z Xiong; F Zhang; Y Wang; Y Shi; X X Zhu\",\n \"journal\": \"\",\n \"ref_id\": \"b38\",\n \"title\": \"EarthNets: Empowering AI in earth observation\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"P Xu; X Zhu; D A Clifton\",\n \"journal\": \"\",\n \"ref_id\": \"b39\",\n \"title\": \"Multimodal learning with transformers: A survey\",\n \"year\": \"2022\"\n },\n {\n \"authors\": \"N Yokoya; P Ghamisi; J Xia; S Sukhanov; R Heremans; I Tankoyeu; B Bechtel; B Le Saux; G Moser; D Tuia\",\n \"journal\": \"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing\",\n \"ref_id\": \"b40\",\n \"title\": \"Open data for global multimodal land use classification: Outcome of the 2017 ieee grss data fusion contest\",\n \"year\": \"2018\"\n },\n {\n \"authors\": \"Y Yuan; Z Xiong; Q Wang\",\n \"journal\": \"\",\n \"ref_id\": \"b41\",\n \"title\": \"Acm: Adaptive cross-modal graph convolutional neural networks for rgb-d scene recognition\",\n \"year\": \"2019\"\n },\n {\n \"authors\": \"J Zhang; W Li; P O Ogunbona; P Wang; C Tang\",\n \"journal\": \"Pattern Recognition\",\n \"ref_id\": \"b42\",\n \"title\": \"Rgb-d-based action recognition datasets: A survey\",\n \"year\": \"2016\"\n },\n {\n \"authors\": \"X Zheng; X Wu; L Huan; W He; H Zhang\",\n \"journal\": \"IEEE Transactions on Geoscience and Remote Sensing\",\n \"ref_id\": \"b43\",\n \"title\": \"A gather-to-guide network for remote sensing semantic segmentation of rgb and auxiliary image\",\n \"year\": \"2021\"\n },\n {\n \"authors\": \"X X Zhu; D Tuia; L Mou; G S Xia; L Zhang; F Xu; F Fraundorfer\",\n \"journal\": \"IEEE Geoscience and Remote Sensing Magazine\",\n \"ref_id\": \"b44\",\n \"title\": \"Deep learning in remote sensing: A comprehensive review and list of resources\",\n \"year\": \"2017\"\n }\n]","isConversational":false},"formulas":{"kind":"list like","value":[{"formula_coordinates":[4,314.61,281.48,203.37,64.99],"formula_id":"formula_0","formula_text":".111止--------一------ 1111111■■ 一------一-一-------1111 •• _____-一----- 600 400 200 0 .11111111■■---一一一一一一一-一- 1111111■■■-一一---- 111111■■-----------一-"},{"formula_coordinates":[5,401.39,241.79,105.1,15.83],"formula_id":"formula_1","formula_text":"(𝑋 𝑟𝑔𝑏 )) + 𝑓 𝑑 𝐖 ℎ (𝑓 𝑒 𝐖 ℎ (𝑋 ℎ ))."},{"formula_coordinates":[6,331.51,635.41,212.46,44.07],"formula_id":"formula_2","formula_text":"𝒙 0 = 𝑝(𝒙) + 𝒙 𝑝𝑜𝑠 , 𝒛 ′ 𝑘 = 𝒛 𝑘-1 + MSA(LN(𝒙 𝑘-1 )), 𝒛 𝑘 = 𝒛 ′ 𝑘 + MLP(LN(𝒛 ′ 𝑘 )),(1)"},{"formula_coordinates":[7,99.11,74.9,429.74,184.08],"formula_id":"formula_3","formula_text":"Transformer Q K V Transformer Q K V V K Q Transformer Transformer V K Q Q K V Transformer Q K V Transformer Transformer Q K V("},{"formula_coordinates":[7,76.21,428.61,151.9,14.07],"formula_id":"formula_4","formula_text":"𝑦 * 𝑠 = 𝐹 𝑑 𝑟𝑔𝑏 (𝐹 𝑒 𝑟𝑔𝑏 (𝑋 𝑟𝑔𝑏 )) + 𝐹 𝑑 ℎ (𝐹 𝑒 𝑑 (𝑋 ℎ ))."},{"formula_coordinates":[7,76.21,530.17,138.07,29.75],"formula_id":"formula_5","formula_text":"𝒛 𝑟𝑔𝑏 ⟵ MSA(LN(𝑿 ℎ , 𝑿 𝑟𝑔𝑏 , 𝑿 𝑟𝑔𝑏 ⏟⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏟ 𝑄, 𝐾, 𝑉"},{"formula_coordinates":[7,76.21,565.53,125.62,29.75],"formula_id":"formula_6","formula_text":"𝒛 ℎ ⟵ MSA(LN(𝑿 𝑟𝑔𝑏 , 𝑿 ℎ , 𝑿 ℎ ⏟⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏟ 𝑄, 𝐾, 𝑉"},{"formula_coordinates":[7,331.51,501.59,212.46,27.56],"formula_id":"formula_7","formula_text":"𝒛 𝑎 = TL(𝑝(𝑿 𝑟𝑔𝑏 ) ∥ 𝑝(𝑿 𝑑 ) ∥ 𝑀), 𝒛 𝑎 = (𝒛 𝑟𝑔𝑏 ) ∥ (𝒛 ℎ ) ∥ 𝑀 ′ ),(4)"},{"formula_coordinates":[7,331.51,631.35,113.7,64.7],"formula_id":"formula_8","formula_text":"𝒛 ′ 𝑟𝑔𝑏 = TL 𝑟𝑔𝑏 (𝑝(𝒛 𝑟𝑔𝑏 ) ∥ 𝑴 ′ ), 𝒛 ′ 𝑟𝑔𝑏 = (𝒛 𝑟𝑔𝑏 ) ∥ 𝑴 ′ 𝑟𝑔𝑏 ), 𝒛 ′ ℎ = TL ℎ (𝑝(𝒛 ℎ ) ∥ 𝑴 ′ ), 𝒛 ′ ℎ = ((𝒛 ℎ ) ∥ 𝑴 ′ ℎ )."}],"string":"[\n {\n \"formula_coordinates\": [\n 4,\n 314.61,\n 281.48,\n 203.37,\n 64.99\n ],\n \"formula_id\": \"formula_0\",\n \"formula_text\": \".111止--------一------ 1111111■■ 一------一-一-------1111 •• _____-一----- 600 400 200 0 .11111111■■---一一一一一一一-一- 1111111■■■-一一---- 111111■■-----------一-\"\n },\n {\n \"formula_coordinates\": [\n 5,\n 401.39,\n 241.79,\n 105.1,\n 15.83\n ],\n \"formula_id\": \"formula_1\",\n \"formula_text\": \"(𝑋 𝑟𝑔𝑏 )) + 𝑓 𝑑 𝐖 ℎ (𝑓 𝑒 𝐖 ℎ (𝑋 ℎ )).\"\n },\n {\n \"formula_coordinates\": [\n 6,\n 331.51,\n 635.41,\n 212.46,\n 44.07\n ],\n \"formula_id\": \"formula_2\",\n \"formula_text\": \"𝒙 0 = 𝑝(𝒙) + 𝒙 𝑝𝑜𝑠 , 𝒛 ′ 𝑘 = 𝒛 𝑘-1 + MSA(LN(𝒙 𝑘-1 )), 𝒛 𝑘 = 𝒛 ′ 𝑘 + MLP(LN(𝒛 ′ 𝑘 )),(1)\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 99.11,\n 74.9,\n 429.74,\n 184.08\n ],\n \"formula_id\": \"formula_3\",\n \"formula_text\": \"Transformer Q K V Transformer Q K V V K Q Transformer Transformer V K Q Q K V Transformer Q K V Transformer Transformer Q K V(\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 76.21,\n 428.61,\n 151.9,\n 14.07\n ],\n \"formula_id\": \"formula_4\",\n \"formula_text\": \"𝑦 * 𝑠 = 𝐹 𝑑 𝑟𝑔𝑏 (𝐹 𝑒 𝑟𝑔𝑏 (𝑋 𝑟𝑔𝑏 )) + 𝐹 𝑑 ℎ (𝐹 𝑒 𝑑 (𝑋 ℎ )).\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 76.21,\n 530.17,\n 138.07,\n 29.75\n ],\n \"formula_id\": \"formula_5\",\n \"formula_text\": \"𝒛 𝑟𝑔𝑏 ⟵ MSA(LN(𝑿 ℎ , 𝑿 𝑟𝑔𝑏 , 𝑿 𝑟𝑔𝑏 ⏟⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏟ 𝑄, 𝐾, 𝑉\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 76.21,\n 565.53,\n 125.62,\n 29.75\n ],\n \"formula_id\": \"formula_6\",\n \"formula_text\": \"𝒛 ℎ ⟵ MSA(LN(𝑿 𝑟𝑔𝑏 , 𝑿 ℎ , 𝑿 ℎ ⏟⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏟ 𝑄, 𝐾, 𝑉\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 331.51,\n 501.59,\n 212.46,\n 27.56\n ],\n \"formula_id\": \"formula_7\",\n \"formula_text\": \"𝒛 𝑎 = TL(𝑝(𝑿 𝑟𝑔𝑏 ) ∥ 𝑝(𝑿 𝑑 ) ∥ 𝑀), 𝒛 𝑎 = (𝒛 𝑟𝑔𝑏 ) ∥ (𝒛 ℎ ) ∥ 𝑀 ′ ),(4)\"\n },\n {\n \"formula_coordinates\": [\n 7,\n 331.51,\n 631.35,\n 113.7,\n 64.7\n ],\n \"formula_id\": \"formula_8\",\n \"formula_text\": \"𝒛 ′ 𝑟𝑔𝑏 = TL 𝑟𝑔𝑏 (𝑝(𝒛 𝑟𝑔𝑏 ) ∥ 𝑴 ′ ), 𝒛 ′ 𝑟𝑔𝑏 = (𝒛 𝑟𝑔𝑏 ) ∥ 𝑴 ′ 𝑟𝑔𝑏 ), 𝒛 ′ ℎ = TL ℎ (𝑝(𝒛 ℎ ) ∥ 𝑴 ′ ), 𝒛 ′ ℎ = ((𝒛 ℎ ) ∥ 𝑴 ′ ℎ ).\"\n }\n]","isConversational":false},"title":{"kind":"string","value":"GAMUS: A Geometry-aware Multi-modal Semantic Segmentation Benchmark for Remote Sensing Data"},"abstract":{"kind":"string","value":"Geometric information in the normalized digital surface models (nDSM) is highly correlated with the semantic class of the land cover. Exploiting two modalities (RGB and nDSM (height)) jointly has great potential to improve the segmentation performance. However, it is still an under-explored field in remote sensing due to the following challenges. First, the scales of existing datasets are relatively small and the diversity of existing datasets is limited, which restricts the ability of validation. Second, there is a lack of unified benchmarks for performance assessment, which leads to difficulties in comparing the effectiveness of different models. Last, sophisticated multi-modal semantic segmentation methods have not been deeply explored for remote sensing data. To cope with these challenges, in this paper, we introduce a new remote-sensing benchmark dataset for multi-modal semantic segmentation based on RGB-Height (RGB-H) data. Towards a fair and comprehensive analysis of existing methods, the proposed benchmark consists of 1) a large-scale dataset including co-registered RGB and nDSM pairs and pixel-wise semantic labels; 2) a comprehensive evaluation and analysis of existing multi-modal fusion strategies for both convolutional and Transformer-based networks on remote sensing data. Furthermore, we propose a novel and effective Transformer-based intermediary multi-modal fusion (TIMF) module to improve the semantic segmentation performance through adaptive token-level multi-modal fusion. The designed benchmark can foster future research on developing new methods for multi-modal learning on remote sensing data. Extensive analyses of those methods are conducted and valuable insights are provided through the experimental results. Code for the benchmark and baselines can be accessed at https://github.com/EarthNets/RSI-MMSegmentation."},"authors":{"kind":"string","value":"Zhitong Xiong; Sining Chen; Yi Wang; Lichao Mou; Xiao Xiang Zhu"},"figures":{"kind":"list like","value":[{"figure_caption":"Fig. 1 :1Fig. 1: Example images of the GAMUS dataset. Images from left to right are the RGB modality, the nDSM modality, the blending visualization image, and the segmentation label.","figure_data":"","figure_id":"fig_0","figure_label":"1","figure_type":"figure"},{"figure_caption":"Fig. 2 :2Fig. 2: Data collection and processing of the GAMUS dataset.","figure_data":"","figure_id":"fig_1","figure_label":"2","figure_type":"figure"},{"figure_caption":"compared and analyzed on remote sensing data. which enables a fair and comprehensive performance comparison and evaluation; 3. A novel Transformer-based Intermediary Multi-modal Fusion (TIMF) module is proposed for the adaptive fusion of RGB and Height data, which achieves stateof-the-art segmentation performance. 4. The proposed datasets and benchmarking results can provide useful insights and spark novel ideas for developing new multi-modal segmentation methods for RS data.","figure_data":"","figure_id":"fig_2","figure_label":"","figure_type":"figure"},{"figure_caption":"Fig. 3 :3Fig. 3: Statistics of height values of the GAMUS dataset. Both the histograms and spatial distributions are presented. A long-tailed distribution can be clearly observed.","figure_data":"","figure_id":"fig_3","figure_label":"3","figure_type":"figure"},{"figure_caption":"the Potsdam dataset contains 38 tiles with about 6,000×6,000 image resolution. Usually, 24 tiles are used for training and the rest 14 tiles for testing. The Vaihingen dataset contains 33 tiles with an image resolution of about 2,500×2,000. These tiles are officially split into two subsets, of which 16 tiles are used for training and 17 tiles for testing. The DFC19 dataset is a remote sensing dataset that contains 2783 images of 1024X1024 resolution. It is designed for the Deep Learning for Semantic Segmentation of Urban Scenes challenge, which aims to advance the state-of-the-art in semantic segmentation of urban scenes using multi-spectral and LiDAR data. The GeoNRW dataset is a remote sensing dataset that contains 7783 images with pixel-level annotation. It includes 10 different land cover classes. The dataset includes RGB and nDSM data. The Zeebruges dataset(Yokoya et al., 2018) was acquired using an airborne platform flying over the urban and harbor areas of Zeebruges, Belgium. The dataset contains seven separate tiles and each tile is with a 10,000 × 10,000 image resolution. Five tiles are used for training, and the remaining two tiles are for testing. Hong et al.(Hong et al., 2021) introduced a multi-modal dataset with Hyperspectral (HS), SAR, and nDSMs for the classification of remote sensing data.","figure_data":"","figure_id":"fig_4","figure_label":"","figure_type":"figure"},{"figure_caption":"Fig. 4 :4Fig. 4: Statistics of height values of the GAMUS dataset. Both the histograms and spatial distributions are presented. A longtailed distribution can be clearly observed.","figure_data":"","figure_id":"fig_6","figure_label":"4","figure_type":"figure"},{"figure_caption":"Fig. 5 :5Fig. 5: Statistics of the semantic class of the GAMUS dataset.","figure_data":"","figure_id":"fig_7","figure_label":"5","figure_type":"figure"},{"figure_caption":"Fig. 6 :6Fig.6: Illustration of five different multi-modal fusion paradigms, including 1) single modality, where only the RGB modality is used; 2) 1) single modality, where only the height modality is used; 3) multi-modal early-fusion, where image-level fusion is conducted; 4) multi-modal feature fusion, where features of different modalities are fused; 5) multi-modal late-fusion, where segmentation results from different modalities are combined.","figure_data":"","figure_id":"fig_8","figure_label":"6","figure_type":"figure"},{"figure_caption":"Fig. 7 :7Fig.7: Illustration of six different multi-modal fusion paradigms, including 1) single modality, where only the RGB modality is used; 2) single modality, where only the height modality is used; 3) multi-modal early-fusion, where image-level fusion is conducted; 4) multi-modal cross feature fusion, where features of different modalities are fused via cross-attention mechanism; 5) multi-modal late-fusion, where segmentation results from different modalities are combined; 6) Intermediary fusion, where the proposed TIMF module is illustrated.","figure_data":"","figure_id":"fig_9","figure_label":"7","figure_type":"figure"},{"figure_caption":"Fig. 8 :8Fig. 8: some qualitative visualization examples of segmentation results on the GAMUS dataset. The segmentation results of using different multi-modal fusion strategies are visualized for a clear comparison.","figure_data":"","figure_id":"fig_10","figure_label":"8","figure_type":"figure"},{"figure_caption":"Fig. 9 :9Fig. 9: some qualitative visualization examples of segmentation results on the GAMUS dataset. The segmentation results of using different multi-modal fusion strategies are visualized for a clear comparison.","figure_data":"","figure_id":"fig_11","figure_label":"9","figure_type":"figure"},{"figure_caption":"Statistics comparison with existing RGB-H datasets. Note that the test set of DFC 19 is not publicly accessible.","figure_data":"Datasets#Total TilesTile Size#Cities#Training #Validation#Test#Class LabelsPotsdam (Rottensteiner et al., 2014)386000 × 6000Single24 tiles014 tiles6Vahingen (Rottensteiner et al., 2014)332500 × 2000Single16 tiles017 tiles5DFC 19 (Le Saux et al., 2019)27831024 × 1024Multiple2783 tiles100 tiles-6GeoNRW (Bosch et al., 2019)332500 × 2000Multiple16 tiles017 tiles5Zeebruges (Yokoya et al., 2018)910,000 × 10,000Single5 tiles02 tiles8Augsburg (Hong et al., 2021)1332 × 485Single761 pixels077,533 pixels7GAMUS (Ours)11,5071024 × 1024Multiple 6304 tiles1059 tiles4144 tiles6","figure_id":"tab_0","figure_label":"1","figure_type":"table"},{"figure_caption":"simplicity, the encoder and decoder sub-networks for RGB and height modality are denoted by 𝑓 𝑒","figure_data":"𝐖 𝑟𝑔𝑏, 𝑓 𝑑 𝐖 𝑟𝑔𝑏and 𝑓 𝑒 𝐖 ℎ,𝑓 𝑑 𝐖 ℎ be formally expressed as follows. . With these notations, the five training paradigms can1. Single Modality (RGB).𝑦 * 𝑠 = 𝑓 𝑑 𝐖 𝑟𝑔𝑏(𝑓 𝑒 𝐖 𝑟𝑔𝑏(𝑋 𝑟𝑔𝑏 )).2. Single Modality (Height).𝑦 * 𝑠 = 𝑓 𝑑 𝐖 ℎ(𝑓 𝑒 𝐖 ℎ(𝑋 ℎ )).3. Early Multi-Modal Fusion.𝑦 * 𝑠 = 𝑓 𝑑 𝐖 (𝑓 𝑒 𝐖 (𝑋 𝑟𝑔𝑏 ∥ 𝑋 ℎ )).4. Feature-level Multi-Modal Fusion.𝑦 * 𝑠 = 𝑓 𝑑 𝐖 (𝑓 𝑒 𝐖 𝑟𝑔𝑏(𝑋 𝑟𝑔𝑏 ) ∥ 𝑓 𝑒 𝐖 ℎ(𝑋 ℎ )).5. Late Multi-Modal Fusion.𝑦 * 𝑠 = 𝑓 𝑑 𝐖 𝑟𝑔𝑏(𝑓 𝑒 𝐖 𝑟𝑔𝑏","figure_id":"tab_2","figure_label":"","figure_type":"table"},{"figure_caption":"Comparison results (Acc) of different multi-modal fusion methods on the GAMUS dataset for supervised semantic segmentation.","figure_data":"ParadigmMethodsModality Ground Vegetation Building WaterRoadTreemAccSingle ModalityMFNet (Ha et al., 2017) RTFNet (Sun et al., 2019) FuseNet (Hazirbas et al., 2016)RGB RGB RGB0.7030 0.7370 0.37530.4251 0.5980 0.51040.7696 0.8873 0.87240.3510 0.4242 0.7947 0.5779 0.2144 0.6236 0.8666 0.6545 0.1045 0.6375 0.7887 0.5481Single ModalityMFNet (Ha et al., 2017) RTFNet (Sun et al., 2019) FuseNet (Hazirbas et al., 2016)Height Height Height0.6881 0.8223 0.78210.5304 0.5537 0.42080.7601 0.8708 0.81730.4078 0.4911 0.7502 0.6046 0.2513 0.6764 0.8186 0.6655 0.6912 0.7455 0.4715 0.6547MFNet (Ha et al., 2017)RGBH0.69680.57490.78750.4673 0.4917 0.8005 0.6365EarlyRTFNet (Sun et al., 2019)RGBH0.79550.65210.87060.4695 0.6516 0.8135 0.7088FusionFuseNet (Hazirbas et al., 2016)RGBH0.80840.41780.90490.6758 0.8179 0.6807 0.7176MFNet (Ha et al., 2017)RGBH0.72110.59450.80200.5041 0.5298 0.8530 0.6674FeatureRTFNet (Sun et al., 2019)RGBH0.71900.67320.87740.5975 0.7057 0.8533 0.7377FusionFuseNet (Hazirbas et al., 2016)RGBH0.47160.75580.91090.3184 0.6221 0.9444 0.6705Late FusionMFNet (Ha et al., 2017) RTFNet (Sun et al., 2019) FuseNet (Hazirbas et al., 2016)RGBH RGBH RGBH0.7491 0.7798 0.72390.6037 0.6195 0.31850.8517 0.8941 0.95260.5530 0.5774 0.7831 0.6863 0.5141 0.6953 0.8087 0.7186 0.0758 0.8571 0.8194 0.6245","figure_id":"tab_3","figure_label":"2","figure_type":"table"},{"figure_caption":"Comparison results (IoU) of different multi-modal fusion methods on the GAMUS dataset for supervised semantic segmentation.","figure_data":"ParadigmMethodsModality Ground Vegetation Building WaterRoadTreemIoUSingle ModalityMFNet (Ha et al., 2017) RTFNet (Sun et al., 2019) FuseNet (Hazirbas et al., 2016)RGB RGB RGB0.5560 0.6337 0.56670.3776 0.4326 0.34390.6776 0.7474 0.56520.2713 0.3036 0.6231 0.4682 0.2197 0.5286 0.7026 0.5441 0.3958 0.4298 0.4326 0.4557Single ModalityMFNet (Ha et al., 2017) RTFNet (Sun et al., 2019) FuseNet (Hazirbas et al., 2016)Height Height Height0.5199 0.5887 0.37530.3393 0.3912 0.51040.7382 0.7522 0.87240.1918 0.2479 0.7064 0.4572 0.1807 0.4752 0.7206 0.5181 0.1045 0.6375 0.7887 0.5481MFNet (Ha et al., 2017)RGBH0.57730.39900.73450.2511 0.3073 0.7056 0.4958EarlyRTFNet (Sun et al., 2019)RGBH0.62180.45020.74570.3771 0.5240 0.7083 0.5712FusionFuseNet (Hazirbas et al., 2016)RGBH0.63590.37020.69140.5321 0.4336 0.6289 0.5487MFNet (Ha et al., 2017)RGBH0.60340.44800.76970.2563 0.3347 0.7517 0.5273FeatureRTFNet (Sun et al., 2019)RGBH0.60100.44310.75920.4507 0.5190 0.7226 0.5826FusionFuseNet (Hazirbas et al., 2016)RGBH0.43180.41860.75140.2979 0.3937 0.6773 0.4951Late FusionMFNet (Ha et al., 2017) RTFNet (Sun et al., 2019) FuseNet (Hazirbas et al., 2016)RGBH RGBH RGBH0.6278 0.6209 0.56240.4606 0.4384 0.30420.8031 0.7576 0.76990.2997 0.4013 0.7831 0.5626 0.4537 0.5494 0.7162 0.5894 0.0748 0.3975 0.7396 0.4747","figure_id":"tab_4","figure_label":"3","figure_type":"table"},{"figure_caption":"Comparison results (IoU) of different fusion strategies on the GAMUS dataset for Transformer-based semantic segmentation methods.","figure_data":"ParadigmMethodsModality Ground Vegetation BuildingWaterRoadTreemIoUSingle ModalityCMX (Liu et al., 2022)RGB0.72650.60470.76530.6988 0.6859 0.6790 0.6934Single ModalityCMX (Liu et al., 2022)Height0.66720.48160.82710.4203 0.5893 0.8024 0.6313Early FusionCMX (Liu et al., 2022)RGBH0.79280.66540.82360.6876 0.7273 0.8001 0.7495Late FusionCMX (Liu et al., 2022)RGBH0.79760.67630.84380.6827 0.7281 0.8249 0.7589Cross Feature FusionCMX (Liu et al., 2022)RGBH0.78270.67100.84530.6770 0.7127 0.8250 0.7523Intermediary FusionTIMF (Ours)RGBH0.80230.67970.84520.6955 0.7368 0.8232 0.7638","figure_id":"tab_5","figure_label":"4","figure_type":"table"},{"figure_caption":"Comparison results (IoU) of different fusion strategies on the GAMUS dataset for supervised semantic segmentation.","figure_data":"MethodsModality Ground Vegetation BuildingWaterRoadTreemIoUMFNet (Ha et al., 2017)RGBH0.60340.44800.76970.2563 0.3347 0.7517 0.5273RTFNet (Sun et al., 2019)RGBH0.60100.44310.75920.4507 0.5190 0.7226 0.5826FuseNet (Hazirbas et al., 2016)RGBH0.43180.41860.75140.2979 0.3937 0.6773 0.4951MFNet-ShapeConv (Cao et al., 2021)RGBH0.61720.46550.76930.2190 0.5485 0.7320 0.5586MFNet-VCD (Xiong et al., 2020a)RGBH0.64010.45600.78010.4595 0.5203 0.7261 0.5970CMX (Liu et al., 2022)RGBH0.78270.67100.84530.6770 0.7127 0.8250 0.7523TIMF (Ours)RGBH0.80230.67970.84520.6955 0.7368 0.8232 0.7638","figure_id":"tab_6","figure_label":"5","figure_type":"table"}],"string":"[\n {\n \"figure_caption\": \"Fig. 1 :1Fig. 1: Example images of the GAMUS dataset. Images from left to right are the RGB modality, the nDSM modality, the blending visualization image, and the segmentation label.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 2 :2Fig. 2: Data collection and processing of the GAMUS dataset.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_1\",\n \"figure_label\": \"2\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"compared and analyzed on remote sensing data. which enables a fair and comprehensive performance comparison and evaluation; 3. A novel Transformer-based Intermediary Multi-modal Fusion (TIMF) module is proposed for the adaptive fusion of RGB and Height data, which achieves stateof-the-art segmentation performance. 4. The proposed datasets and benchmarking results can provide useful insights and spark novel ideas for developing new multi-modal segmentation methods for RS data.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 3 :3Fig. 3: Statistics of height values of the GAMUS dataset. Both the histograms and spatial distributions are presented. A long-tailed distribution can be clearly observed.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_3\",\n \"figure_label\": \"3\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"the Potsdam dataset contains 38 tiles with about 6,000×6,000 image resolution. Usually, 24 tiles are used for training and the rest 14 tiles for testing. The Vaihingen dataset contains 33 tiles with an image resolution of about 2,500×2,000. These tiles are officially split into two subsets, of which 16 tiles are used for training and 17 tiles for testing. The DFC19 dataset is a remote sensing dataset that contains 2783 images of 1024X1024 resolution. It is designed for the Deep Learning for Semantic Segmentation of Urban Scenes challenge, which aims to advance the state-of-the-art in semantic segmentation of urban scenes using multi-spectral and LiDAR data. The GeoNRW dataset is a remote sensing dataset that contains 7783 images with pixel-level annotation. It includes 10 different land cover classes. The dataset includes RGB and nDSM data. The Zeebruges dataset(Yokoya et al., 2018) was acquired using an airborne platform flying over the urban and harbor areas of Zeebruges, Belgium. The dataset contains seven separate tiles and each tile is with a 10,000 × 10,000 image resolution. Five tiles are used for training, and the remaining two tiles are for testing. Hong et al.(Hong et al., 2021) introduced a multi-modal dataset with Hyperspectral (HS), SAR, and nDSMs for the classification of remote sensing data.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_4\",\n \"figure_label\": \"\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 4 :4Fig. 4: Statistics of height values of the GAMUS dataset. Both the histograms and spatial distributions are presented. A longtailed distribution can be clearly observed.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_6\",\n \"figure_label\": \"4\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 5 :5Fig. 5: Statistics of the semantic class of the GAMUS dataset.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_7\",\n \"figure_label\": \"5\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 6 :6Fig.6: Illustration of five different multi-modal fusion paradigms, including 1) single modality, where only the RGB modality is used; 2) 1) single modality, where only the height modality is used; 3) multi-modal early-fusion, where image-level fusion is conducted; 4) multi-modal feature fusion, where features of different modalities are fused; 5) multi-modal late-fusion, where segmentation results from different modalities are combined.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_8\",\n \"figure_label\": \"6\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 7 :7Fig.7: Illustration of six different multi-modal fusion paradigms, including 1) single modality, where only the RGB modality is used; 2) single modality, where only the height modality is used; 3) multi-modal early-fusion, where image-level fusion is conducted; 4) multi-modal cross feature fusion, where features of different modalities are fused via cross-attention mechanism; 5) multi-modal late-fusion, where segmentation results from different modalities are combined; 6) Intermediary fusion, where the proposed TIMF module is illustrated.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_9\",\n \"figure_label\": \"7\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 8 :8Fig. 8: some qualitative visualization examples of segmentation results on the GAMUS dataset. The segmentation results of using different multi-modal fusion strategies are visualized for a clear comparison.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_10\",\n \"figure_label\": \"8\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Fig. 9 :9Fig. 9: some qualitative visualization examples of segmentation results on the GAMUS dataset. The segmentation results of using different multi-modal fusion strategies are visualized for a clear comparison.\",\n \"figure_data\": \"\",\n \"figure_id\": \"fig_11\",\n \"figure_label\": \"9\",\n \"figure_type\": \"figure\"\n },\n {\n \"figure_caption\": \"Statistics comparison with existing RGB-H datasets. Note that the test set of DFC 19 is not publicly accessible.\",\n \"figure_data\": \"Datasets#Total TilesTile Size#Cities#Training #Validation#Test#Class LabelsPotsdam (Rottensteiner et al., 2014)386000 × 6000Single24 tiles014 tiles6Vahingen (Rottensteiner et al., 2014)332500 × 2000Single16 tiles017 tiles5DFC 19 (Le Saux et al., 2019)27831024 × 1024Multiple2783 tiles100 tiles-6GeoNRW (Bosch et al., 2019)332500 × 2000Multiple16 tiles017 tiles5Zeebruges (Yokoya et al., 2018)910,000 × 10,000Single5 tiles02 tiles8Augsburg (Hong et al., 2021)1332 × 485Single761 pixels077,533 pixels7GAMUS (Ours)11,5071024 × 1024Multiple 6304 tiles1059 tiles4144 tiles6\",\n \"figure_id\": \"tab_0\",\n \"figure_label\": \"1\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"simplicity, the encoder and decoder sub-networks for RGB and height modality are denoted by 𝑓 𝑒\",\n \"figure_data\": \"𝐖 𝑟𝑔𝑏, 𝑓 𝑑 𝐖 𝑟𝑔𝑏and 𝑓 𝑒 𝐖 ℎ,𝑓 𝑑 𝐖 ℎ be formally expressed as follows. . With these notations, the five training paradigms can1. Single Modality (RGB).𝑦 * 𝑠 = 𝑓 𝑑 𝐖 𝑟𝑔𝑏(𝑓 𝑒 𝐖 𝑟𝑔𝑏(𝑋 𝑟𝑔𝑏 )).2. Single Modality (Height).𝑦 * 𝑠 = 𝑓 𝑑 𝐖 ℎ(𝑓 𝑒 𝐖 ℎ(𝑋 ℎ )).3. Early Multi-Modal Fusion.𝑦 * 𝑠 = 𝑓 𝑑 𝐖 (𝑓 𝑒 𝐖 (𝑋 𝑟𝑔𝑏 ∥ 𝑋 ℎ )).4. Feature-level Multi-Modal Fusion.𝑦 * 𝑠 = 𝑓 𝑑 𝐖 (𝑓 𝑒 𝐖 𝑟𝑔𝑏(𝑋 𝑟𝑔𝑏 ) ∥ 𝑓 𝑒 𝐖 ℎ(𝑋 ℎ )).5. Late Multi-Modal Fusion.𝑦 * 𝑠 = 𝑓 𝑑 𝐖 𝑟𝑔𝑏(𝑓 𝑒 𝐖 𝑟𝑔𝑏\",\n \"figure_id\": \"tab_2\",\n \"figure_label\": \"\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison results (Acc) of different multi-modal fusion methods on the GAMUS dataset for supervised semantic segmentation.\",\n \"figure_data\": \"ParadigmMethodsModality Ground Vegetation Building WaterRoadTreemAccSingle ModalityMFNet (Ha et al., 2017) RTFNet (Sun et al., 2019) FuseNet (Hazirbas et al., 2016)RGB RGB RGB0.7030 0.7370 0.37530.4251 0.5980 0.51040.7696 0.8873 0.87240.3510 0.4242 0.7947 0.5779 0.2144 0.6236 0.8666 0.6545 0.1045 0.6375 0.7887 0.5481Single ModalityMFNet (Ha et al., 2017) RTFNet (Sun et al., 2019) FuseNet (Hazirbas et al., 2016)Height Height Height0.6881 0.8223 0.78210.5304 0.5537 0.42080.7601 0.8708 0.81730.4078 0.4911 0.7502 0.6046 0.2513 0.6764 0.8186 0.6655 0.6912 0.7455 0.4715 0.6547MFNet (Ha et al., 2017)RGBH0.69680.57490.78750.4673 0.4917 0.8005 0.6365EarlyRTFNet (Sun et al., 2019)RGBH0.79550.65210.87060.4695 0.6516 0.8135 0.7088FusionFuseNet (Hazirbas et al., 2016)RGBH0.80840.41780.90490.6758 0.8179 0.6807 0.7176MFNet (Ha et al., 2017)RGBH0.72110.59450.80200.5041 0.5298 0.8530 0.6674FeatureRTFNet (Sun et al., 2019)RGBH0.71900.67320.87740.5975 0.7057 0.8533 0.7377FusionFuseNet (Hazirbas et al., 2016)RGBH0.47160.75580.91090.3184 0.6221 0.9444 0.6705Late FusionMFNet (Ha et al., 2017) RTFNet (Sun et al., 2019) FuseNet (Hazirbas et al., 2016)RGBH RGBH RGBH0.7491 0.7798 0.72390.6037 0.6195 0.31850.8517 0.8941 0.95260.5530 0.5774 0.7831 0.6863 0.5141 0.6953 0.8087 0.7186 0.0758 0.8571 0.8194 0.6245\",\n \"figure_id\": \"tab_3\",\n \"figure_label\": \"2\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison results (IoU) of different multi-modal fusion methods on the GAMUS dataset for supervised semantic segmentation.\",\n \"figure_data\": \"ParadigmMethodsModality Ground Vegetation Building WaterRoadTreemIoUSingle ModalityMFNet (Ha et al., 2017) RTFNet (Sun et al., 2019) FuseNet (Hazirbas et al., 2016)RGB RGB RGB0.5560 0.6337 0.56670.3776 0.4326 0.34390.6776 0.7474 0.56520.2713 0.3036 0.6231 0.4682 0.2197 0.5286 0.7026 0.5441 0.3958 0.4298 0.4326 0.4557Single ModalityMFNet (Ha et al., 2017) RTFNet (Sun et al., 2019) FuseNet (Hazirbas et al., 2016)Height Height Height0.5199 0.5887 0.37530.3393 0.3912 0.51040.7382 0.7522 0.87240.1918 0.2479 0.7064 0.4572 0.1807 0.4752 0.7206 0.5181 0.1045 0.6375 0.7887 0.5481MFNet (Ha et al., 2017)RGBH0.57730.39900.73450.2511 0.3073 0.7056 0.4958EarlyRTFNet (Sun et al., 2019)RGBH0.62180.45020.74570.3771 0.5240 0.7083 0.5712FusionFuseNet (Hazirbas et al., 2016)RGBH0.63590.37020.69140.5321 0.4336 0.6289 0.5487MFNet (Ha et al., 2017)RGBH0.60340.44800.76970.2563 0.3347 0.7517 0.5273FeatureRTFNet (Sun et al., 2019)RGBH0.60100.44310.75920.4507 0.5190 0.7226 0.5826FusionFuseNet (Hazirbas et al., 2016)RGBH0.43180.41860.75140.2979 0.3937 0.6773 0.4951Late FusionMFNet (Ha et al., 2017) RTFNet (Sun et al., 2019) FuseNet (Hazirbas et al., 2016)RGBH RGBH RGBH0.6278 0.6209 0.56240.4606 0.4384 0.30420.8031 0.7576 0.76990.2997 0.4013 0.7831 0.5626 0.4537 0.5494 0.7162 0.5894 0.0748 0.3975 0.7396 0.4747\",\n \"figure_id\": \"tab_4\",\n \"figure_label\": \"3\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison results (IoU) of different fusion strategies on the GAMUS dataset for Transformer-based semantic segmentation methods.\",\n \"figure_data\": \"ParadigmMethodsModality Ground Vegetation BuildingWaterRoadTreemIoUSingle ModalityCMX (Liu et al., 2022)RGB0.72650.60470.76530.6988 0.6859 0.6790 0.6934Single ModalityCMX (Liu et al., 2022)Height0.66720.48160.82710.4203 0.5893 0.8024 0.6313Early FusionCMX (Liu et al., 2022)RGBH0.79280.66540.82360.6876 0.7273 0.8001 0.7495Late FusionCMX (Liu et al., 2022)RGBH0.79760.67630.84380.6827 0.7281 0.8249 0.7589Cross Feature FusionCMX (Liu et al., 2022)RGBH0.78270.67100.84530.6770 0.7127 0.8250 0.7523Intermediary FusionTIMF (Ours)RGBH0.80230.67970.84520.6955 0.7368 0.8232 0.7638\",\n \"figure_id\": \"tab_5\",\n \"figure_label\": \"4\",\n \"figure_type\": \"table\"\n },\n {\n \"figure_caption\": \"Comparison results (IoU) of different fusion strategies on the GAMUS dataset for supervised semantic segmentation.\",\n \"figure_data\": \"MethodsModality Ground Vegetation BuildingWaterRoadTreemIoUMFNet (Ha et al., 2017)RGBH0.60340.44800.76970.2563 0.3347 0.7517 0.5273RTFNet (Sun et al., 2019)RGBH0.60100.44310.75920.4507 0.5190 0.7226 0.5826FuseNet (Hazirbas et al., 2016)RGBH0.43180.41860.75140.2979 0.3937 0.6773 0.4951MFNet-ShapeConv (Cao et al., 2021)RGBH0.61720.46550.76930.2190 0.5485 0.7320 0.5586MFNet-VCD (Xiong et al., 2020a)RGBH0.64010.45600.78010.4595 0.5203 0.7261 0.5970CMX (Liu et al., 2022)RGBH0.78270.67100.84530.6770 0.7127 0.8250 0.7523TIMF (Ours)RGBH0.80230.67970.84520.6955 0.7368 0.8232 0.7638\",\n \"figure_id\": \"tab_6\",\n \"figure_label\": \"5\",\n \"figure_type\": \"table\"\n }\n]","isConversational":false},"citation_data":{"kind":"string","value":"[{\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhu et al., 2017)\", \"Explanation\": \"The cited work by Zhu et al. (2017) is referenced to highlight the importance of semantic segmentation in the fields of computer vision and remote sensing, providing a foundational basis for the citing paper to build upon.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kuznietsov et al., 2017)\", \"Explanation\": \"The cited work by Kuznietsov et al. (2017) is referenced to further emphasize the significance of semantic segmentation in the fields of computer vision and remote sensing, providing a methodological basis for the citing paper to follow.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Mou and Zhu, 2018)\", \"Explanation\": \"The cited work by Mou and Zhu (2018) is referenced to highlight the advancements in RGB image-based semantic segmentation research, providing a methodological basis for the citing paper to build upon.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xiong et al., 2022)\", \"Explanation\": \"The cited work by Xiong et al. (2022) is referenced to showcase the recent progress in RGB image-based semantic segmentation research, providing a methodological basis for the citing paper to reference.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Long et al., 2015)\", \"Explanation\": \"The cited work by Long et al. (2015) is referenced to highlight the use of deep learning techniques in RGB image-based semantic segmentation research, providing a methodological basis for the citing paper to follow.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xie et al., 2021)\", \"Explanation\": \"The cited work by Xie et al. (2021) is referenced to showcase the advancements in RGB image-based semantic segmentation research using deep learning techniques, providing a methodological basis for the citing paper to reference.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2021)\", \"Explanation\": \"The cited work by Liu et al. (2021) is referenced to highlight the recent progress in RGB image-based semantic segmentation research using deep learning techniques, providing a methodological basis for the citing paper to follow.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Xiong et al., 2021)\", \"Explanation\": \"The cited work by Xiong et al. (2021) provides the RGB-Depth (RGB-D) data modality that the citing paper utilizes in their research on multi-modal data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Li et al., 2019)\", \"Explanation\": \"The cited work by Li et al. (2019) contributes the RGB-Thermal (RGB-T) data modality to the citing paper for their study on multi-modal data.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Wang and Neumann, 2018)\", \"Explanation\": \"The cited work by Wang and Neumann (2018) extends the research on multi-modal data by exploring the use of RGB images in various tasks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Cheng et al., 2017)\", \"Explanation\": \"The cited work by Cheng et al. (2017) continues the research on multi-modal data by studying the use of RGB images in different applications.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Jiang et al., 2018b)\", \"Explanation\": \"The cited work by Jiang et al. (2018b) expands the research on multi-modal data by exploring the use of RGB images in various tasks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Xiong et al., 2020b)\", \"Explanation\": \"The cited work by Xiong et al. (2020b) extends the research on multi-modal data by studying the use of RGB images in different applications.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Wang and Neumann, 2018)\", \"Explanation\": \"The cited work by Wang and Neumann (2018) provides the RGB images that the citing paper utilizes in their research on multi-modal data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Cheng et al., 2017)\", \"Explanation\": \"The cited work by Cheng et al. (2017) contributes the RGB images that the citing paper uses in their study on multi-modal data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Jiang et al., 2018b)\", \"Explanation\": \"The cited work by Jiang et al. (2018b) provides the RGB images that the citing paper utilizes in their research on multi-modal data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Xiong et al., 2020b)\", \"Explanation\": \"The cited work by Xiong et al. (2020b) contributes the RGB images that the citing paper uses in their study on multi-modal data.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Kunwar, 2019)\", \"Explanation\": \"The cited work is the source of the height information used in the nDSM data, which is a crucial element in the proposed GAMUS dataset for multi-modal semantic segmentation.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Mahmud et al., 2020)\", \"Explanation\": \"The cited work is the source of the nDSM data, which is a key data modality in the GAMUS dataset for multi-modal semantic segmentation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Hong et al., 2021)\", \"Explanation\": \"The cited work by Hong et al. provides a foundation for the development of multi-modal learning methods in the field of Earth observation applications, which the citing paper builds upon to address the problem of data fusion in the context of pixel-level semantic segmentation.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Yuan et al., 2019)\", \"Explanation\": \"The cited work on scene recognition provides a methodological basis for the study of multi-modal feature learning in RGB-D images in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Cao et al., 2021)\", \"Explanation\": \"The cited work on semantic segmentation contributes to the design of multi-branch deep networks for multi-modal feature learning in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Ha et al., 2017)\", \"Explanation\": \"The cited work on semantic segmentation provides methodological insights for the design of multi-branch deep networks in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chen et al., 2020)\", \"Explanation\": \"The cited work on semantic segmentation offers methodological guidance for the design of multi-branch deep networks in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Chen et al., 2021)\", \"Explanation\": \"The cited work on semantic segmentation contributes to the design of multi-branch deep networks for multi-modal feature learning in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Gupta et al., 2014)\", \"Explanation\": \"The cited work on object detection provides methodological insights for the design of multi-branch deep networks in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Fan et al., 2020)\", \"Explanation\": \"The cited work on object detection contributes to the design of multi-branch deep networks for multi-modal feature learning in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zhang et al., 2016)\", \"Explanation\": \"The cited work on action recognition provides methodological guidance for the design of multi-branch deep networks in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Hazirbas et al., 2016)\", \"Explanation\": \"The cited work on RGB-D semantic segmentation is a data source for the design of multi-branch deep networks in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Jiang et al., 2018a)\", \"Explanation\": \"The cited work on RGB-D semantic segmentation is a data source for the design of multi-branch deep networks in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Park et al., 2017)\", \"Explanation\": \"The cited work on RGB-D image segmentation is a data source for the design of multi-branch deep networks in the citing paper.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Lin et al., 2017)\", \"Explanation\": \"The cited work on RGB-D image segmentation is a data source for the design of multi-branch deep networks in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Hu et al., 2019)\", \"Explanation\": \"The cited work ACNet provides a method for adaptively fusing features of different modalities for image segmentation, which the citing paper adopts in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Cheng et al., 2017)\", \"Explanation\": \"The cited work Gated Fusion Net also contributes a method for adaptively fusing features of different modalities for image segmentation, which the citing paper may have considered in their research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Shivakumar et al., 2020)\", \"Explanation\": \"The cited work PSTNet extends the research on utilizing LWIR imagery as a supporting modality for semantic segmentation by proposing a new method for multi-modal learning networks.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Sun et al., 2019)\", \"Explanation\": \"The cited work RTFNet also extends the research on utilizing LWIR imagery as a supporting modality for semantic segmentation by proposing a new method for multi-modal learning networks.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Kampffmeyer et al., 2016)\", \"Explanation\": \"The cited work by Kampffmeyer et al. (2016) serves as a methodological basis for the citing paper, as it designed a multi-modal deep network with an early deep fusion architecture that the citing paper can build upon for semantic segmentation of urban images.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Audebert et al., 2017)\", \"Explanation\": \"The cited work by Audebert et al. (2017) extends the research on combining optical and OpenStreetMap data by proposing a two-stream multi-modal learning network to improve segmentation performance, which the citing paper can further explore in their research.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Audebert et al., 2018)\", \"Explanation\": \"The cited work by Audebert et al. (2018) introduces a SegNet-based multi-modal fusion architecture for urban scene segmentation, which the citing paper can build upon to further investigate the combination of Multi-spectral images (MSI) and Lidar in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Hong et al., 2020a)\", \"Explanation\": \"The cited work by Hong et al. (2020a) proposes a cross-fusion strategy for learning multi-modal features with more balanced weight contributions from different modalities, which the citing paper can adopt in their research to improve the fusion of optical and Lidar predictions in a late-fusion manner.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Liu et al., 2019)\", \"Explanation\": \"The cited work by Liu et al. (2019) provides a high-order conditional random field (CRF) based method for the fusion of optical and Lidar predictions in a late-fusion manner, which the citing paper can use as a data source to further explore the fusion of optical and Lidar data in their research.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Zheng et al., 2021)\", \"Explanation\": \"The cited work introduces the gather-to-guide module, which the citing paper adopts to combine the complementary merits of RGB and auxiliary modality data in their research.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Rottensteiner et al., 2014)\", \"Explanation\": \"The cited work provides the benchmark datasets of ISPRS Potsdam and Vaihingen for high-resolution semantic segmentation in remote sensing.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Bosch et al., 2019)\", \"Explanation\": \"The cited work introduces the US3D dataset, which is larger in volume and covers more tasks but is not designed for multi-modal segmentation tasks. The citing paper extends the research by providing a standard dataloader and detailed instructions for multi-modal learning on the GAMUS dataset.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Le Saux et al., 2019)\", \"Explanation\": \"The cited work introduces the DFC 19 dataset, which is derived from the US3D dataset and is smaller in scale. The citing paper uses this dataset to provide benchmark results for multi-modal learning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xu et al., 2022)\", \"Explanation\": \"The cited work introduces the use of the self-attention mechanism for multi-modal feature fusion at the token level, which the citing paper adopts in their research to improve the performance of their model.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Cao et al., 2021)\", \"Explanation\": \"The cited work by Cao et al. serves as a methodological basis for the design of a cross-modality guided encoder in the RGB-D multi-modal learning methods used for benchmarking in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xiong et al., 2020a)\", \"Explanation\": \"The cited work by Xiong et al. is also used as a methodological basis for the design of a cross-modality guided encoder in the RGB-D multi-modal learning methods for benchmarking in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Hazirbas et al., 2016)\", \"Explanation\": \"The cited work by Hazirbas et al. provides a design for better multi-scale feature fusion architectures that the citing paper adopts in their research to improve the performance of the model.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Ha et al., 2017)\", \"Explanation\": \"The cited work by Ha et al. introduces the multi-spectral fusion networks (MFNet) method, which the citing paper extends by using it to evaluate the performance of the GAMUS dataset in multi-modal learning.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Sun et al., 2019)\", \"Explanation\": \"The cited work by Sun et al. presents the RTFNet method, which the citing paper builds upon to use two separate encoders for RGB and thermal modalities and progressively fuse their intermediate features to learn multimodal representations.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The cited work introduces a cross-modal feature rectification module that the citing paper adopts to calibrate features in both spatial and channel dimensions.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Xie et al., 2021)\", \"Explanation\": \"The cited work is used as the baseline segmentation architecture in the citing paper, indicating an extension or continuation of the research in the field of segmentation.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The cited work is further adapted in the citing paper to five different fusion strategies for performance comparison, showing an extension or continuation of the research in the field of fusion methods.\"}, {\"Category\": \"Data Source\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The cited work is the source of the cross-modal feature fusion method used in the citing paper, providing a foundational element for the study conducted in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The cited work introduces a new fusion method that the citing paper proposes, specifically the intermediary feature fusion module, which is described in the citing paper.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The cited work provides the feature fusion module that the citing paper uses in the TIMF module to perform feature extraction and fusion in the plug-and-play manner.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Liu et al., 2022)\", \"Explanation\": \"The cited work, CMX, uses SegFormer as a baseline for its segmentation model, which contributes to the performance improvement by leveraging the global spatial context and self-attention mechanism for better feature learning.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xie et al., 2021)\", \"Explanation\": \"The cited work, SegFormer, serves as a baseline for the CMX model, providing the global spatial context and self-attention mechanism for better feature learning and performance improvement.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Cao et al., 2021)\", \"Explanation\": \"The ShapeConv method is cited as a way to demonstrate the importance of using geometry information in the height modality to improve segmentation performance.\"}, {\"Category\": \"Methodological Basis\", \"Citation\": \"(Xiong et al., 2020a)\", \"Explanation\": \"The VCD method is cited to highlight the effectiveness of using height modality to improve segmentation performance.\"}, {\"Category\": \"Extension or Continuation\", \"Citation\": \"(Cao et al., 2021)\", \"Explanation\": \"The TIMF module is presented as a new method that builds upon existing feature-fusion methods to achieve the best mIoU performance in the multi-modal segmentation task.\"}]"}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":0,"numItemsPerPage":100,"numTotalItems":19454,"offset":0,"length":100}},"jwt":"eyJhbGciOiJFZERTQSJ9.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTc3NjUzMTIwMSwic3ViIjoiL2RhdGFzZXRzL0JhY2hOZ29IL1BhcnNlZEFyeGl2UGFwZXJzIiwiZXhwIjoxNzc2NTM0ODAxLCJpc3MiOiJodHRwczovL2h1Z2dpbmdmYWNlLmNvIn0.NPlvXMXW9v9lt6v3nniOJ7kL8NkTrBNYz_5IZVvYTTV0HDpYRD1GkgNI1iFlGchVPboKyk_jtUCzxpay86mcDA","displayUrls":true,"splitSizeSummaries":[{"config":"default","split":"train","numRows":19454,"numBytesParquet":536920578}]},"discussionsStats":{"closed":0,"open":2,"total":2},"fullWidth":true,"hasGatedAccess":true,"hasFullAccess":true,"isEmbedded":false,"savedQueries":{"community":[],"user":[]}}">
sections listlengths 0 910 | pub_date stringclasses 722
values | doi stringlengths 0 570 | references listlengths 0 835 | formulas listlengths 0 679 | title stringlengths 0 235 | abstract stringlengths 0 7.77k | authors stringlengths 0 11.9k | figures listlengths 0 270 | citation_data stringlengths 2 160k ⌀ |
|---|---|---|---|---|---|---|---|---|---|
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b3",
"b37",
"b4",
"b11",
"b14",
"b34",
"b13",
"b25",
"b19",
"b42"
],
"table_ref": [],
"text": "Pre-trained language models (PLMs) have quickly become a staple in t... | 2023-10-23 | 10.18653/v1/2021.findings-emnlp.410 | [
{
"authors": "Alan Ansell; Maria Edoardo; Jonas Ponti; Sebastian Pfeiffer; Goran Ruder; Ivan Glavaš; Anna Vulić; Korhonen",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "MAD-G: Multilingual adapter generation for efficient cross-lingual transfer",
"year": "20... | [
{
"formula_coordinates": [
4,
78.41,
688.88,
201.98,
25.55
],
"formula_id": "formula_0",
"formula_text": "RIPL(S AL , S PL ) = AUC(S AL ) -AUC(S PL ) 1 -AUC(S PL )"
}
] | Parameter-Efficient Language Model Tuning with Active Learning in Low-Resource Settings | Pre-trained language models (PLMs) have ignited a surge in demand for effective finetuning techniques, particularly in low-resource domains and languages. Active learning (AL), a set of algorithms designed to decrease labeling costs by minimizing label complexity, has shown promise in confronting the labeling bottlenec... | Josip Jukić; Jan Šnajder Takelab | [
{
"figure_caption": "Figure 3 :3Figure3: AL learning curves compared with random sampling on the SUBJ dataset. The first and the second rows show learning curves for adapters without and with TAPT, respectively. The third row shows learning curves for FFT, without and with TAPT. The results are averaged over fi... | [{"Category": "Methodological Basis", "Citation": "(Cohn et al., 1996)", "Explanation": "The cited work introduces the concept of active learning as a potential solution to the challenge of data labeling in low-resource settings, which the citing paper builds upon in its research on efficient finetuning methods for PLM... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b12",
"b36",
"b15",
"b34",
"b2",
"b3",
"b8"
],
"table_ref": [],
"text": "Document comprehension involves interpreting words that can alter the meaning of the text based on their pla... | 2024-01-22 | 10.18653/v1/D19-1345 | [
{
"authors": "Shilpa Arora; Mahesh Joshi; Carolyn Rosé",
"journal": "",
"ref_id": "b0",
"title": "Identifying types of claims in online customer reviews",
"year": "2009"
},
{
"authors": "Iz Beltagy; Matthew E Peters; Arman Cohan",
"journal": "",
"ref_id": "b1",
"title": "Long... | [] | Connecting the Dots: What Graph-Based Text Representations Work Best for Text Classification using Graph Neural Networks? | Given the success of Graph Neural Networks (GNNs) for structure-aware machine learning, many studies have explored their use for text classification, but mostly in specific domains with limited data characteristics. Moreover, some strategies prior to GNNs relied on graph mining and classical machine learning, making it... | Margarita Bugueño; Gerard De Melo | [
{
"figure_caption": "1http://derekgreene.com/bbc/ 2 https://zenodo.org/HNDrecord",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 1 :1Figure 1: Graph Construction Methods. Given the input text \"Start working! The soone... | [{"Category": "Methodological Basis", "Citation": "(Castillo et al., 2015)", "Explanation": "The cited work provides a foundation for the use of graphs in text classification tasks, as it discusses the applicability and effectiveness of graphs in broader settings."}, {"Category": "Extension or Continuation", "Citation"... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b10",
"b5",
"b11",
"b12",
"b13",
"b14",
"b15",
"b16",
"b17",
... | 2023-07 | 10.1177/0361198105191800112 | [
{
"authors": "",
"journal": "OECD",
"ref_id": "b0",
"title": "The Economic Consequences of Outdoor Air Pollution",
"year": "2016"
},
{
"authors": "R N Annavarapu; S Kathi",
"journal": "Environmental pollution",
"ref_id": "b1",
"title": "Cognitive disorders in children associa... | [
{
"formula_coordinates": [
4,
147.2,
542.39,
325.78,
11.3
],
"formula_id": "formula_0",
"formula_text": "M = S + N(M, S, N ∈ R N c ×SR )(1)"
},
{
"formula_coordinates": [
4,
147.2,
605.5,
325.78,
14.13
],
"formula_id": "form... | Real-Time Idling Vehicles Detection Using Combined Audio-Visual Deep Learning | Combustion vehicle emissions contribute to poor air quality and release greenhouse gases into the atmosphere, and vehicle pollution has been associated with numerous adverse health effects. Roadways with extensive waiting and/or passenger drop-off, such as schools and hospital drop-off zones, can result in a high incid... | Xiwen Li; Tristalee Mangin; Surojit Saha; Rehman Mohammed; Evan Blanchard; Dillon Tang; Henry Poppe; Nathan Searle; Ouk Choi; Kerry Kelly; Ross Whitaker | [
{
"figure_caption": "Figure 1 .1Figure 1. Proposed System Design. The yellow arrow collects vehicle motion, engine sound, and pollutant concentrations. The red arrow represents data transmission to the computer. The green arrow denotes sending the predicted idling status to the displays. The blue arrow represen... | [{"Category": "Methodological Basis", "Citation": "[1]", "Explanation": "The cited work provides the global data on the impact of poor air quality on human health, which serves as a methodological basis for the citing paper to analyze the effects of vehicle pollution on health and healthcare costs."}, {"Category": "Sup... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b1",
"b30",
"b53",
"b29",
"b33"
],
"table_ref": [],
"text": "The meaning and import of an utterance are often underdetermined by the utterance itself. Human interpretation involves making infer... | 2023-10-25 | 10.18653/v1/D15-1008 | [
{
"authors": "Elliott Ash; Germain Gauthier; Philine Widmer",
"journal": "",
"ref_id": "b0",
"title": "Relatio: Text semantics capture political and economic narratives",
"year": "2022"
},
{
"authors": "Kent Bach",
"journal": "",
"ref_id": "b1",
"title": "Conversational impli... | [
{
"formula_coordinates": [
2,
306.43,
160.53,
194.83,
27.4
],
"formula_id": "formula_0",
"formula_text": "### INPUT: { utterance } OUTPUT: { inference 1 } | { inference 2 } | . . ."
},
{
"formula_coordinates": [
7,
331.04,
371.46,
193.97,... | Natural Language Decompositions of Implicit Content Enable Better Text Representations | When people interpret text, they rely on inferences that go beyond the observed language itself. Inspired by this observation, we introduce a method for the analysis of text that takes implicitly communicated content explicitly into account. We use a large language model to produce sets of propositions that are inferen... | Alexander Hoyle; Rupak Sarkar; Pranav Goel; Philip Resnik | [
{
"figure_caption": "Federallands and waters must not be opened up to fossil fuel extraction. Public lands are national treasures that should be protected for future generations, not auctioned off to the fossil fuel industry's highest bidders.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label... | [{"Category": "Methodological Basis", "Citation": "(Bach, 1994)", "Explanation": "The cited work by Bach (1994) provides foundational theories and methods for understanding the meaning and import of utterances, which the citing paper builds upon in its research on human interpretation of text data."}, {"Category": "Met... |
[
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [
"b29",
"b28",
"b71",
"b8",
"b42",
"b65",
"b13",
"b73",
"b59",
"b2",
"b73",
"b22",
"b16",
"b24",
"b56",
"b32",
"b12",
"b6",
... | 2024-03-11 | 10.48550/arXiv.1106.6251 | [
{
"authors": "Ekin Akyürek; Tolga Bolukbasi; Frederick Liu; Binbin Xiong; Ian Tenney; Jacob Andreas; Kelvin Guu",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Towards tracing knowledge in language models back to the training data",
"year": "2022-12"
},
{
... | [
{
"formula_coordinates": [
3,
242.49,
536.67,
262.18,
8.99
],
"formula_id": "formula_0",
"formula_text": "kGLM(x) := W κ(x, X) + b,(1)"
},
{
"formula_coordinates": [
3,
244.64,
617.53,
260.03,
8.99
],
"formula_id": "formula_... | FAITHFUL AND EFFICIENT EXPLANATIONS FOR NEU-RAL NETWORKS VIA NEURAL TANGENT KERNEL SUR-ROGATE MODELS | A recent trend in explainable AI research has focused on surrogate modeling, where neural networks are approximated as simpler ML algorithms such as kernel machines. A second trend has been to utilize kernel functions in various explainby-example or data attribution tasks. In this work, we combine these two trends to a... | Andrew Engel; Zhichao Wang; Natalie S Frank; Ioana Dumitriu; Sutanay Choudhury; Anand Sarwate; Tony Chiang | [
{
"figure_caption": "Figure 1 :1Figure 1: Linear Realization of Bert-base Model. Each panel shows a linearization of a Bert-base transfer model, initialized from a different seed. An invertible mapping is fit between the kGLM and NN to transform the kGLM's final activations to the NN's, described in Appendix L.... | [{"Category": "Methodological Basis", "Citation": "(Leavitt & Morcos, 2020)", "Explanation": "The cited work by Leavitt and Morcos (2020) highlights the importance of explainability in the field of deep neural networks, which serves as a foundational basis for the citing paper to address the same issue."}, {"Category":... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b2"
],
"table_ref": [],
"text": "Generative Flow Networks (GFlowNets, GFNs; Bengio et al., 2021a,b) are probabilistic models over discrete sample spaces with a compositional structure. They are also stochastic sequent... | 2023-08-29 | 10.48550/arXiv.2302.05446 | [
{
"authors": "E Bengio; M Jain; M Korablyov; D Precup; Y Bengio",
"journal": "",
"ref_id": "b0",
"title": "Flow network based generative models for non-iterative diverse candidate generation",
"year": "2021"
},
{
"authors": "Y Bengio; T Deleu; E J Hu; S Lahlou; M Tiwari; E Bengio; T Dele... | [] | torchgfn: A PyTorch GFlowNet library | The growing popularity of generative flow networks (GFlowNets or GFNs) from a range of researchers with diverse backgrounds and areas of expertise necessitates a library which facilitates the testing of new features such as training losses that can be easily compared to standard benchmark implementations, or on a set o... | Salem Lahlou; Joseph D Viviano; Mila Victor Schmidt; Yoshua Bengio | [
{
"figure_caption": "Figure 1 :1Figure 1: Hierarchy of the codebase for the v1 release. States and Actions are top-level abstractions used to interface between the stateless Environments and Containers, which are generic objects used by the remainder of the codebase. Containers are utilized by both Samplers and... | [{"Category": "Supporting Evidence", "Citation": "(Lahlou et al., 2023)", "Explanation": "The cited work introduces the concept of GFlowNet and its continuous variant, which the citing paper builds upon in the development of the torchgfn library for fast prototyping of GFlowNet related algorithms in PyTorch."}, {"Categ... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b15",
"b17",
"b33",
"b16",
"b45",
"b13",
"b7",
"b9",
"b14",
"b3",
"b12",
"b40",
"b22",
"b1",
"b37",
"b4"
],
"table_ref": [],
"t... | 2023-11-05 | 10.18653/v1/2022.naacl-main.135 | [
{
"authors": "Eneko Agirre; Carmen Banea; Daniel Cer; Mona Diab; Aitor González-Agirre; Rada Mihalcea; German Rigau; Janyce Wiebe",
"journal": "",
"ref_id": "b0",
"title": "Semeval-2016 task 1: Semantic textual similarity, monolingual and cross-lingual evaluation",
"year": "2016"
},
{
"a... | [
{
"formula_coordinates": [
3,
70.87,
581.19,
189.71,
10.82
],
"formula_id": "formula_0",
"formula_text": "• Enc(s 1 ⊖ s 2 ) ≈ f diff (Enc(s 1 ), Enc(s 2 ))"
},
{
"formula_coordinates": [
4,
69.59,
125.99,
216.04,
56.63
],
"f... | Bridging Continuous and Discrete Spaces: Interpretable Sentence Representation Learning via Compositional Operations | Traditional sentence embedding models encode sentences into vector representations to capture useful properties such as the semantic similarity between sentences. However, in addition to similarity, sentence semantics can also be interpreted via compositional operations such as sentence fusion or difference. It is uncl... | James Y Huang; Wenlin Yao; Kaiqiang Song; Hongming Zhang; Muhao Chen; Dong Yu | [
{
"figure_caption": " 34.1 51.0 28.1 45.0 Model performance on four textual generation tasks for interpretability evaluation. Unsup. Contr. and Sup. Contr. represents Unsupervised and Supervised Contrastive baselines respectively. We report ROUGE-1/2/L scores.",
"figure_data": "SRoBERTa49.0 21.9 39.0 39.1 1... | [{"Category": "Methodological Basis", "Citation": "(Conneau et al., 2018)", "Explanation": "The cited work by Conneau et al. provides a method for probing individual linguistic properties in sentence embeddings, which the citing paper adopts to better understand the information encoded in the sentence representation sp... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3"
],
"table_ref": [],
"text": "The priority focus of modern face recognition studies has been in-line with that of representation learning studies: amplifying the representative p... | 2023-05-24 | [
{
"authors": "Yaobin Zhang; Weihong Deng; Yaoyao Zhong; Jiani Hu; Xian Li; Dongyue Zhao; Dongchao Wen",
"journal": "",
"ref_id": "b0",
"title": "Adaptive label noise cleaning with meta-supervision for deep face recognition",
"year": "2021"
},
{
"authors": "Boxiao Liu; Guanglu Song; Manyu... | [
{
"formula_coordinates": [
2,
335.22,
550.21,
209.89,
30.32
],
"formula_id": "formula_0",
"formula_text": "L cls = 1 N N i=1 -log e cos θy i e cos θy i + C j̸ =yi e cos θj ,(1)"
},
{
"formula_coordinates": [
3,
112.11,
689.93,
174.25,
... | FaceFusion: Exploiting Full Spectrum of Multiple Datasets | The size of training dataset is known to be among the most dominating aspects of training high-performance face recognition embedding model. Building a large dataset from scratch could be cumbersome and time-intensive, while combining multiple already-built datasets poses the risk of introducing large amount of label n... | Chiyoung Song; Dongjae Lee; Naver Cloud | [
{
"figure_caption": "Figure 1 :1Figure 1: Overview of FaceFusion. L cls,k and L cls shares the same class proxies. While L cls,k limits the softmax calculations to each dataset, L cls merges the class proxies of same identity, and removes the barriers between the datasets. GRL reverses the direction of gradient... | [{"Category": "Methodological Basis", "Citation": "[1]", "Explanation": "The cited work on rectifier models provides a method for adjusting identities in a way that is relevant to the citing paper."}, {"Category": "Methodological Basis", "Citation": "[2]", "Explanation": "The cited work on posterior data cleaning provi... | |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b39",
"b1",
"b8",
"b33",
"b31",
"b38",
"b2",
"b17",
"b34",
"b34",
"b30"
],
"table_ref": [],
"text": "Tracking entity states in procedural ... | 10.18653/v1/2020.findings-emnlp.91 | [
{
"authors": "Antoine Bosselut; Omer Levy; Ari Holtzman; Corin Ennis; Dieter Fox; Yejin Choi",
"journal": "",
"ref_id": "b0",
"title": "Simulating action dynamics with neural process networks",
"year": "2017"
},
{
"authors": "Antoine Bosselut; Omer Levy; Ari Holtzman; Corin Ennis; Dieter... | [] | OPENPI2.0: An Improved Dataset for Entity Tracking in Texts | Much text describes a changing world (e.g., procedures, stories, newswires), and understanding them requires tracking how entities change. An earlier dataset, OPENPI, provided crowdsourced annotations of entity state changes in text. However, a major limitation was that those annotations were free-form and did not iden... | Li Zhang; Hainiu Xu; Abhinav Kommula; Chris Callison-Burch; Niket Tandon | [
{
"figure_caption": "Figure 1 :1Figure 1: For each step in a procedure, OPENPI annotates the state change of attributes of entities. Our OPENPI2.0 additionally (shown in red boxes and texts) canonicalizes the entities and attributes and includes their salience scores.",
"figure_data": "",
"figure_id": "... | [{"Category": "Methodological Basis", "Citation": "(Weston et al., 2015)", "Explanation": "The cited work by Weston et al. provides a foundational method for tracking entity states in procedural texts, which the citing paper builds upon in its own research."}, {"Category": "Methodological Basis", "Citation": "(Bosselut... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b16",
"b10",
"b3"
],
"table_ref": [],
"text": "Pretrained large language models (LLMs) have recently seen widespread adoption by users worldwide due to their capabilities with generative tasks ranging from... | 2024-02-13 | 10.18653/v1/2020.acl-main.485 | [
{
"authors": "Abubakar Abid; Maheen Farooqi; James Zou",
"journal": "Nature Machine Intelligence",
"ref_id": "b0",
"title": "Large language models associate muslims with violence",
"year": "2021"
},
{
"authors": "Kabir Ahuja; Harshita Diddee; Rishav Hada; Millicent Ochieng; Krithika Rame... | [
{
"formula_coordinates": [
5,
330.4,
78.67,
169.74,
113.25
],
"formula_id": "formula_0",
"formula_text": "CS(ci, cj) = 1 if ci = cj, 0 otherwise Con CS(t) = CS(cKB, ci) Non CS(t) = 1 n c∈C non CS(cKB, c) ∆ CS(t) = Con CS -Non CS Non CS Cst CS(t) = 1 n(n -1) n i=1 n j=1,... | This Land is {Your, My} Land: Evaluating Geopolitical Bias in Language Models through Territorial Disputes | Do the Spratly Islands belong to China, the Philippines, or Vietnam? A pretrained large language model (LLM) may answer differently if asked in the languages of each claimant country: Chinese, Tagalog, or Vietnamese. This contrasts with a multilingual human, who would likely answer consistently. In this paper, we show ... | Bryan Li; Samar Haider; Chris Callison-Burch | [
{
"figure_caption": "Figure 2 :2Figure 2: Illustration of comparisons made for the CS metrics. KB CS, Control CS, and Non-control CS all compare between the KB country and a response, while Consistency CS compares between responses.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "2",
... | [{"Category": "Supporting Evidence", "Citation": "(Petroni et al., 2019)", "Explanation": "The cited work by Petroni et al. (2019) provides evidence that LLMs do internalize some relational knowledge, which is a foundational element for the discussion in the citing paper about the limitations of LLMs in terms of their ... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b2",
"b4",
"b20",
"b0"
],
"table_ref": [],
"text": "It is well known that any domain gap between training and test data hurts the performance of machine learning models in general, and object detecto... | [
{
"authors": "",
"journal": "",
"ref_id": "b0",
"title": "MMDetection3D: Open-MMLab next-generation platform for general 3D object detection",
"year": "2020"
},
{
"authors": "Jeevan Devaranjan; Amlan Kar; Sanja Fidler",
"journal": "",
"ref_id": "b1",
"title": "Meta-Sim2: Unsu... | [] | Realistically distributing object placements in synthetic training data improves the performance of vision-based object detection models | When training object detection models on synthetic data, it is important to make the distribution of synthetic data as close as possible to the distribution of real data. We investigate specifically the impact of object placement distribution, keeping all other aspects of synthetic data fixed. Our experiment, training ... | Setareh Dabiri; Vasileios Lioutas; Berend Zwartsenberg; Yunpeng Liu; Matthew Niedoba; Xiaoxuan Liang; Dylan Green; Jonathan Wilder Lavington; Frank Wood; Adam Ścibior; Inverted Ai | [
{
"figure_caption": "Figure 1 .1Figure 1. Sample training set images generated using CARLA. We compare the baseline placement (left) with a realistic one (right).",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 .2Fi... | [{"Category": "Methodological Basis", "Citation": "[3]", "Explanation": "The cited work, CARLA driving simulator, is used as a source of training data in the citing paper to generate synthetic data for the study of the impact of object placement distribution on the performance of vision models in driving contexts."}, {... | ||
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b4",
"b12",
"b26",
"b33",
"b30",
"b28",
"b19"
],
"table_ref": [],
"text": "Fact-checking is an essential task in natural language processing, focusing on evaluating the accuracy of ... | 2024-04-01 | 10.18653/v1/2023.emnlp-demo.10 | [
{
"authors": "Mubashara Akhtar; Oana Cocarascu; Elena Simperl",
"journal": "",
"ref_id": "b0",
"title": "Pubhealthtab: A public health table-based dataset for evidence-based fact checking",
"year": "2022"
},
{
"authors": "Firoj Alam; Stefano Cresci; Tanmoy Chakraborty; Fabrizio Silvestri... | [] | Self-Checker: Plug-and-Play Modules for Fact-Checking with Large Language Models | Fact-checking is an essential task in NLP that is commonly utilized to validate the factual accuracy of a piece of text. Previous approaches mainly involve the resource-intensive process of fine-tuning pre-trained language models on specific datasets. In addition, there is a notable gap in datasets that focus on factch... | Miaoran Li; Baolin Peng; Michel Galley; Jianfeng Gao; Zhu Zhang | [
{
"figure_caption": "Figure 2 :2Figure 2: Overview of SELF-CHECKER. The framework consists of four plug-and-play modules: (1) claim processor, (2) query generator, (3) evidence seeker, and (4) verdict counselor.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figu... | [{"Category": "Methodological Basis", "Citation": "(Brown et al., 2020)", "Explanation": "The cited work, GPT-3, is a large language model that the citing paper uses as a basis for their research on fact-checking and the generation of false information in LLMs."}, {"Category": "Extension or Continuation", "Citation": "... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b30",
"b2",
"b31",
"b19",
"b48",
"b34",
"b18"
],
"table_ref": [],
"text": "Retrieval-augmented language models, which integrate non-parametric dense retrieval with autoregressive ne... | 2023-05-24 | 10.18653/v1/P17-1171 | [
{
"authors": "Uri Alon; Frank Xu; Junxian He; Sudipta Sengupta; Dan Roth; Graham Neubig",
"journal": "",
"ref_id": "b0",
"title": "Neuro-symbolic language modeling with automaton-augmented retrieval",
"year": "2022"
},
{
"authors": " Pmlr",
"journal": "",
"ref_id": "b1",
"tit... | [
{
"formula_coordinates": [
3,
77.93,
96.49,
204.14,
26.03
],
"formula_id": "formula_0",
"formula_text": "P KN N (w t |q t ) ∝ (k i ,v i ) 1 wt=v i exp(-d(k i , q t ))"
},
{
"formula_coordinates": [
3,
70.87,
198.97,
218.27,
14.63
... | kNN-LM Does Not Improve Open-ended Text Generation | In this paper, we study the generation quality of interpolation-based retrieval-augmented language models (LMs). These methods, best exemplified by the kNN-LM (Khandelwal et al., 2020), interpolate the LM's predicted distribution of the next word with a distribution formed from the most relevant retrievals for a given ... | Shufan Wang; Yixiao Song; Andrew Drozdov; Aparna Garimella; Varun Manjunatha; Mohit Iyyer | [
{
"figure_caption": "Figure 1 :1Figure1: The plot presents how many times each type of generations (kNN-LM or GPT-2) is chosen by the evaluators. The dark area in each bar shows that the choices were made confidently. The light area represents the choices between kNN-LM and GPT-2 that were hard but the evaluato... | [{"Category": "Methodological Basis", "Citation": "(Metzler et al., 2022)", "Explanation": "The cited work by Metzler et al. provides a strong empirical validation of retrieval-augmented language models, which the citing paper builds upon to study interpolation-based LMs."}, {"Category": "Methodological Basis", "Citati... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b7",
"b5",
"b5",
"b6",
"b6",
"b3",
"b4",
"b2",
"b9"
],
"table_ref": [],
"text": "The success of large language models (LLMs) at generating human-like text has spurred a ... | 10.1609/aaai.v34i05.6239 | [
{
"authors": "",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Proceedings of the Fifth Black-boxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP",
"year": "2022"
},
{
"authors": "Yonatan Bisk; Rowan Zellers; Ronan Le Bras; Jianfeng G... | [] | Testing Causal Models of Word Meaning in GPT-3 and -4 | Large Language Models (LLMs) have driven extraordinary improvements in NLP. However, it is unclear how such models represent lexical concepts-i.e., the meanings of the words they use. This paper evaluates the lexical representations of GPT-3 and GPT-4 through the lens of HIPE theory, a theory of concept representations... | Sam Musker; Ellie Pavlick | [
{
"figure_caption": "One day Jane wanted to wipe up a water spill on the kitchen floor[...]. The object consisted of a bundle of plastic bags attached to a 4-foot long stick. [...] pressed it against the water spill.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "... | [{"Category": "Methodological Basis", "Citation": "(Forbes et al., 2019)", "Explanation": "The cited work by Forbes et al. provides a method for measuring the extent to which LLMs have good representations of word meanings by focusing on physical objects and their properties and affordances."}, {"Category": "Methodolog... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b10",
"b11",
"b12",
"b13",
"b14",
"b15",
"b16",
"b17",
"b18",... | 2023-05-28 | 10.1007/978-3-319-21903-5_8 | [
{
"authors": "S E Schaeffer",
"journal": "Computer science review",
"ref_id": "b0",
"title": "Graph clustering",
"year": "2007"
},
{
"authors": "Y Zhou; H Cheng; J X Yu",
"journal": "",
"ref_id": "b1",
"title": "Graph clustering based on structural/attribute similarities",
... | [
{
"formula_coordinates": [
4,
123,
65.75,
265.34,
24.74
],
"formula_id": "formula_0",
"formula_text": "Hψ (x) = - 2 2m ∇ 2 + v(x) ψ (x) = Eψ (x)(1)"
},
{
"formula_coordinates": [
4,
165.72,
186.62,
222.63,
24.44
],
"formula_... | Graph Analysis Using a GPU-based Parallel Algorithm: Quantum Clustering | The article introduces a new method for applying Quantum Clustering to graph structures. Quantum Clustering (QC) is a novel densitybased unsupervised learning method that determines cluster centers by constructing a potential function. In this method, we use the Graph Gradient Descent algorithm to find the centers of c... | Zhe Wang; Zhijie He; Ding Liu | [
{
"figure_caption": "Fig. 1 A1Fig. 1 A classification of Graph Clustering;",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Algorithm 11Calculating the Potential function for each data point Require: graph : graph structure d... | [{"Category": "Methodological Basis", "Citation": "[1]", "Explanation": "The cited work introduces the concept of graph clustering and its application in various fields, providing a methodological basis for the citing paper to discuss the techniques and applications of graph clustering."}, {"Category": "Supporting Evid... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [],
"table_ref": [],
"text": "As the volume of biomedical literature continues to grow, biomedical entity linking and event extraction tasks have received increasingly more at-1 Equal contribution. Binding to laminin-1"
},
{
... | 2023-05-24 | 10.18653/v1/2021.naacl-main.205 | [
{
"authors": "Rico Angell; Sunil Monath; Nishant Mohan; Andrew Yadav; Mccallum",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Clusteringbased inference for biomedical entity linking",
"year": "2021"
},
{
"authors": "Amos Bairoch; Rolf Apweiler; Cathy... | [
{
"formula_coordinates": [
3,
305.82,
436.44,
98.02,
13.87
],
"formula_id": "formula_0",
"formula_text": "x E = [x E 1 , x E 2 , ..., x E n ]"
},
{
"formula_coordinates": [
4,
426.06,
117.54,
98.35,
13.87
],
"formula_id": "f... | Iteratively Improving Biomedical Entity Linking and Event Extraction via Hard Expectation-Maximization | Biomedical entity linking and event extraction are two crucial tasks to support text understanding and retrieval in the biomedical domain. These two tasks intrinsically benefit each other: entity linking disambiguates the biomedical concepts by referring to external knowledge bases and the domain knowledge further prov... | Xiaochu Li; Minqian Liu; Zhiyang Xu; Lifu Huang | [
{
"figure_caption": "TheCSC-1, which associates with ICP-1, directly binds to the BIR-1.GO:0001817Process that modulates the frequencey, rate, or extent of production of",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "induces... | [{"Category": "Methodological Basis", "Citation": "(Kim et al., 2009)", "Explanation": "The cited work by Kim et al. provides a method for retrieving and organizing information related to gene functions, bio-molecule relations, and bio-molecule behaviors from unstructured texts, which the citing paper adopts in its res... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b28",
"b6",
"b40"
],
"table_ref": [],
"text": "Scientific Opinion Summarization provides a succinct synopsis for scientific documents and helps readers recap salient information and understand the professi... | 2023-11-13 | 10.18653/v1/2021.emnlp-main.528 | [
{
"authors": "Reinald Kim Amplayo; Stefanos Angelidis; Mirella Lapata",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Aspect-controllable opinion summarization",
"year": "2021"
},
{
"authors": "Reinald Kim Amplayo; Stefanos Angelidis; Mirella Lapata",
... | [] | Scientific Opinion Summarization: Meta-review Generation with Checklist-guided Iterative Introspection | Opinions in the scientific domain can be divergent, leading to controversy or consensus among reviewers. However, current opinion summarization datasets mostly focus on product review domains, which do not account for this variability under the assumption that the input opinions are non-controversial. To address this g... | Qi Zeng; Mankeerat Sidhu; Hou Pong Chan; Lu Wang; Heng Ji | [
{
"figure_caption": "Figure 5 :5Figure 5: We show the meta-reviews from human, vanilla, CGI 2 , and CGI 2 without iterative runs for the same paper. The yellow background indicates hallucinated content. The green background indicates redundant content.",
"figure_data": "",
"figure_id": "fig_0",
"fig... | [{"Category": "Methodological Basis", "Citation": "(Hu and Liu, 2006)", "Explanation": "The cited work by Hu and Liu (2006) provides a methodological approach for identifying representative and consensus opinions in product reviews, which the citing paper may adapt or build upon in the context of scientific opinion sum... |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b21",
"b22",
"b12",
"b16",
"b4",
"b35",
"b22"
],
"table_ref": [],
"text": "With the rapid development of social platforms and digital devices, more and more video... | [
{
"authors": "",
"journal": "Palaskar et al",
"ref_id": "b0",
"title": "How2 R-1 R-2 R-L B-1 B-2 B-3 B-4 METEOR CIDEr HA",
"year": "0128"
},
{
"authors": "Gilles Degottex; John Kane; Thomas Drugman; Tuomo Raitio; Stefan Scherer",
"journal": "IEEE",
"ref_id": "b1",
"title": "C... | [
{
"formula_coordinates": [
3,
351.12,
643.75,
174.02,
30.99
],
"formula_id": "formula_0",
"formula_text": "C ŷ = argmax y j ∈C P Θ (y j | Z),(1)"
},
{
"formula_coordinates": [
3,
388.76,
699.38,
136.39,
10.69
],
"formula_id"... | Denoising Bottleneck with Mutual Information Maximization for Video Multimodal Fusion | Video multimodal fusion aims to integrate multimodal signals in videos, such as visual, audio and text, to make a complementary prediction with multiple modalities contents. However, unlike other image-text multimodal tasks, video has longer multimodal sequences with more redundancy and noise in both visual and audio m... | Shaoxiang Wu; Damai Dai; Ziwei Qin; Tianyu Liu; Binghuai Lin; Yunbo Cao; Zhifang Sui | [
{
"figure_caption": "Figure 1 :1Figure 1: An example of redundancy and noise in a video. As illustrated, consecutive frames have high cosine similarity, which results in a problem of redundancy. In addition, useless information like distracting background and weak alignment between frames and transcripts compos... | [{"Category": "Methodological Basis", "Citation": "(Liu et al., 2020)", "Explanation": "The cited work by Liu et al. introduces a fusion forget gate to control the flow of information between multimodal sequences, which the citing paper adopts as a method to address the problem of redundancy and noise in video multimod... | ||
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b4",
"b49",
"b6",
"b27",
"b19",
"b51",
"b19",
"b51"
],
"table_ref": [],
"text": "Survival analysis is a typical statistical task for tracking occurrence of the event of intere... | 2023-05-24 | [
{
"authors": " Antolini",
"journal": "",
"ref_id": "b0",
"title": "",
"year": "2005"
},
{
"authors": "Laura Antolini; Patrizia Boracchi; Elia Biganzoli",
"journal": "Stats in Medicine",
"ref_id": "b1",
"title": "A time-dependent discrimination index for survival data",
"y... | [
{
"formula_coordinates": [
2,
390.32,
101.84,
167.68,
9.65
],
"formula_id": "formula_0",
"formula_text": "p(t|x) = P r(t x = t|x) (1)"
},
{
"formula_coordinates": [
2,
386.17,
160.46,
171.83,
38.61
],
"formula_id": "formula_... | Learning Survival Distribution with Implicit Survival Function | Survival analysis aims at modeling the relationship between covariates and event occurrence with some untracked (censored) samples. In implementation, existing methods model the survival distribution with strong assumptions or in a discrete time space for likelihood estimation with censorship, which leads to weak gener... | Yu Ling; Weimin Tan; Bo Yan | [
{
"figure_caption": "Figure 1 :1Figure 1: Brief framework of ISF. (a) ISF takes sample x and time t as input, and predicts conditional hazard rate ĥ(t|x). (b) Based on estimated conditional hazard rates, we can derive survival distribution p(t|x) through numerical integration.",
"figure_data": "",
"figu... | [{"Category": "Supporting Evidence", "Citation": "[Courtiol et al., 2019]", "Explanation": "The cited work provides a medical example of using survival analysis to model the death probability of diseases, which supports the claim in the citing paper about the use of survival analysis in medical situations."}, {"Categor... | |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b16",
"b26",
"b15",
"b1",
"b16",
"b30",
"b17",
"b14",
"b8",
"b12",
"b28",
"b29",
"b24",
"b31",
"b9",
"b17",
"b9",
... | 2023-11-14 | [
{
"authors": "Huda Alamri; Vincent Cartillier; Abhishek Das; Jue Wang; Anoop Cherian; Irfan Essa; Dhruv Batra; Tim K Marks; Chiori Hori; Peter Anderson",
"journal": "",
"ref_id": "b0",
"title": "Audio visual scene-aware dialog",
"year": "2019"
},
{
"authors": "Satanjeev Banerjee; Alon La... | [
{
"formula_coordinates": [
5,
104.63,
71.86,
390.54,
129.94
],
"formula_id": "formula_0",
"formula_text": "B L E U -1 B L E U -2 B L E U -3 B L E U -4 R O U G E -1 R O U G E -2 R O U G E -L M E T E O R C H R F + + B E R T S C O R E 0.0 0.2 0.4 0.6 0.8 1.0 score Metrics ... | Evaluate What You Can't Evaluate: Unassessable Quality for Generated Response | LLMs (large language models) like ChatGPT have demonstrated exceptional language comprehension and generation abilities. While reference-free evaluators grounded in LLMs exhibit superior human alignment compared to traditional reference-based evaluators, the utilization of such evaluators poses several challenges. Refe... | Yongkang Liu; Shi Feng; Daling Wang; Yifei Zhang; Hinrich Schütze | [
{
"figure_caption": "Figure 1 :1Figure 1: Examples of ChatGPT assigning high probabilities for unreasonable responses. The correct response semantic for this example is unique. The reference response is I checked and her birthday should be June 6, 1975.",
"figure_data": "",
"figure_id": "fig_0",
"fi... | [{"Category": "Methodological Basis", "Citation": "(Papineni et al., 2002)", "Explanation": "The cited work by Papineni et al. (2002) introduces the BLEU metric, which is a reference-based evaluation metric for assessing the quality of response generation in the context of the citing paper."}, {"Category": "Methodologi... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b10",
"b14",
"b2",
"b17",
"b1",
"b6",
"b18",
"b8"
],
"table_ref": [],
"text": "The goal of information extraction (IE) is to learn structure from unstructured documents. Exist... | 2023-11-17 | 10.3115/v1/P15-1034 | [
{
"authors": "Gabor Angeli; Melvin Jose ; Johnson Premkumar; Christopher D Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Leveraging linguistic structure for open domain information extraction",
"year": "2015"
},
{
"authors": "David Bamman; No... | [] | InteractiveIE: Towards Assessing the Strength of Human-AI Collaboration in Improving the Performance of Information Extraction | Learning template-based information extraction (IE) from documents is a crucial yet difficult task. Prior template-based IE approaches assume foreknowledge of the domain's templates. However, many real-world IE scenarios do not have pre-defined schemas. To "figureout-as you go" requires a solution with zero or minimal ... | Ishani Mondal; Michelle Yuan; Aparna Garimella; Francis Ferraro; Andrew Blair-Stanek; Benjamin Van Durme; Jordan Boyd-Graber | [
{
"figure_caption": "Figure 1 :1Figure 1: shows Human-AI interactions in InteractiveIE which consists of three main components: Preprocessing View, Explorer View and Document-Level Cluster view. Through this interface, the humans would be able to modify the bird's eye view of a corpus by altering the questions ... | [{"Category": "Methodological Basis", "Citation": "(Li et al., 2022)", "Explanation": "The cited work by Li et al. provides a method for understanding patterns and behaviors in the world, which the citing paper adopts to help analysts in their work."}, {"Category": "Methodological Basis", "Citation": "(M\u00f3ra et al.... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b12",
"b10",
"b7",
"b22",
"b15",
"b28",
"b3"
],
"table_ref": [],
"text": "As the volume of scientific publishing increases, it is becoming crucial to develop more sophisticated anal... | 2023-05-24 | 10.18653/v1/D19-1371 | [
{
"authors": "",
"journal": "Iz",
"ref_id": "b0",
"title": "",
"year": ""
},
{
"authors": "Kyle Beltagy; Arman Lo; Cohan",
"journal": "Association for Computational Linguistics",
"ref_id": "b1",
"title": "SciB-ERT: A pretrained language model for scientific text",
"year"... | [
{
"formula_coordinates": [
12,
104.07,
330.16,
118.83,
25.58
],
"formula_id": "formula_0",
"formula_text": "-(x 1 , x 2 , . . . x n ); -q ϕ (z|x) = N (z; µ, σ 2 I)."
}
] | Complex Mathematical Symbol Definition Structures: A Dataset and Model for Coordination Resolution in Definition Extraction | Mathematical symbol definition extraction is important for improving scholarly reading interfaces and scholarly information extraction (IE). However, the task poses several challenges: math symbols are difficult to process as they are not composed of natural language morphemes; and scholarly papers often contain senten... | Anna Martin-Boyle; Andrew Head; Kyle Lo; Risham Sidhu; Marti A Hearst; Dongyeop Kang | [
{
"figure_caption": "Figure 5 :5Figure 5: (a) The macro F1 score based on the number of symbols in the sample, and (b) the difference in scores calculated by subtracting baseline F1 scores from TaDDEx.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "5",
"figure_type": "figure"
},
... | [{"Category": "Methodological Basis", "Citation": "(Head et al., 2021)", "Explanation": "The cited work provides a reading interface (ScholarPhi) that could be used for mathematical symbol definition extraction, which the citing paper adopts as a potential use case for the task of symbol definition extraction."}, {"Cat... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b28",
"b14",
"b37",
"b40",
"b24",
"b13",
"b39",
"b49"
],
"table_ref": [],
"text": "Annotator disagreement is a common challenge in NLP (Leonardelli et al., 2021;Fornaciari et ... | 10.18653/v1/2020.alw-1.21 | [
{
"authors": "Sohail Akhtar; Valerio Basile; Viviana Patti",
"journal": "",
"ref_id": "b0",
"title": "Whose opinions matter? perspective-aware models to identify opinions of hate speech victims in abusive language detection",
"year": "2021"
},
{
"authors": "Al Hala; Maximilian Kuwatly; G... | [
{
"formula_coordinates": [
3,
99.95,
743.25,
160.11,
33.71
],
"formula_id": "formula_0",
"formula_text": "θ * = arg max θ E i=1 log P (y i |x i , a i ; θ)"
},
{
"formula_coordinates": [
3,
350.38,
591.52,
129.3,
31
],
"formu... | You Are What You Annotate: Towards Better Models through Annotator Representations | Annotator disagreement is ubiquitous in natural language processing (NLP) tasks. There are multiple reasons for such disagreements, including the subjectivity of the task, difficult cases, unclear guidelines, and so on. Rather than simply aggregating labels to obtain data annotations, we instead try to directly model t... | Naihao Deng; Xinliang Frederick Zhang; Siyang Liu; Winston Wu; Lu Wang; Rada Mihalcea | [
{
"figure_caption": "Ea : Text embedding and weighted annotation embedding.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": ", etc. For instance, during the collection process of the CommitmentBank dataset, De Marneffe et al. ... | [{"Category": "Supporting Evidence", "Citation": "(Leonardelli et al., 2021)", "Explanation": "The cited work by Leonardelli et al. provides a discussion on the common challenge of annotator disagreement in NLP, which serves as a foundational point for the citing paper to build upon in its own research."}, {"Category":... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b6",
"b20",
"b31",
"b29"
],
"table_ref": [],
"text": "3D pose estimation and object classification are challenging but important tasks in computer vision. In realworld applications such as autonomous... | 2023-06-05 | [
{
"authors": "Yutong Bai; Angtian Wang; Adam Kortylewski; Alan Yuille",
"journal": "",
"ref_id": "b0",
"title": "Coke: Localized contrastive learning for robust keypoint detection",
"year": "2020"
},
{
"authors": "Eran Borenstein; Shimon Ullman",
"journal": "IEEE Trans. Pattern Anal.... | [
{
"formula_coordinates": [
3,
117.11,
455.59,
72.16,
11.23
],
"formula_id": "formula_0",
"formula_text": "C = {C r ∈ R c } R"
},
{
"formula_coordinates": [
3,
57.56,
600.98,
228.8,
20.14
],
"formula_id": "formula_1",
"fo... | Robust 3D-aware Object Classification via Discriminative Render-and-Compare | In real-world applications, it is essential to jointly estimate the 3D object pose and class label of objects, i.e., to perform 3D-aware classification. While current approaches for either image classification or pose estimation can be extended to 3D-aware classification, we observe that they are inherently limited: 1)... | Artur Jesslen; Guofeng Zhang; Angtian Wang; Alan Yuille; Adam Kortylewski | [
{
"figure_caption": "Figure 1 .1Figure 1. 3D-aware classification of a partially occluded car (a). Neither a feed-forward neural network (b) nor a render-andcompare approach (c) produce satisfying results as they only predict one task correctly but fail at the other. Our RCNet model (d)handles both tasks robust... | [{"Category": "Methodological Basis", "Citation": "[25]", "Explanation": "The cited work introduces a render-and-compare approach for 3D pose estimation, which the citing paper adopts in their study of 3D-aware object classification."}, {"Category": "Methodological Basis", "Citation": "[25]", "Explanation": "The cited ... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b11"
],
"table_ref": [],
"text": "Video super-resolution (VSR) is the process of changing from low-resolution (LR) video to high-resolution (HR) video. Currently, VSR is divided into traditional VSR and real-world VSR... | 2024-01-01 | [
{
"authors": "Y Blau; R Mechrez; R Timofte; T Michaeli; L Zelnik-Manor",
"journal": "",
"ref_id": "b0",
"title": "The 2018 PIRM challenge on perceptual image super-resolution",
"year": "2018"
},
{
"authors": "K C Chan; X Wang; K Yu; C Dong; C C Loy",
"journal": "",
"ref_id": "b1"... | [
{
"formula_coordinates": [
3,
381.94,
432.46,
176.06,
9.65
],
"formula_id": "formula_0",
"formula_text": "x = M • x i + (1 -M ) • x j ,(1)"
},
{
"formula_coordinates": [
3,
382.97,
449.86,
175.03,
9.65
],
"formula_id": "form... | NegVSR: Augmenting Negatives for Generalized Noise Modeling in Real-world Video Super-Resolution | The capability of video super-resolution (VSR) to synthesize high-resolution (HR) video from ideal datasets has been demonstrated in many works. However, applying the VSR model to real-world video with unknown and complex degradation remains a challenging task. First, existing degradation metrics in most VSR methods ar... | Yexing Song; Meilin Wang; Zhijing Yang; Xiaoyu Xian; Yukai Shi | [
{
"figure_caption": "Figure 1 :1Figure 1: The overview of the proposed NegVSR. (a) Our approach initially extracts noise sequence N sq through window sequence C in an unsupervised manner. The motion of C occurs within the OOD video noise dataset V od . Subsequently, it mixes N sq and LR video V lr to create nov... | [{"Category": "Methodological Basis", "Citation": "(Chan et al. 2022b)", "Explanation": "The cited work by Chan et al. provides a classification of VSR into traditional and real-world VSR, which the citing paper adopts in its research to better understand the VSR process and the challenges associated with it."}, {"Cate... | |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b13",
"b35"
],
"table_ref": [],
"text": "Diffusion models (Sohl-Dickstein et al., 2015b;Ho et al., 2020;Song et al., 2020) have shown remarkable performance in image generation and attracted h... | 2023-06-14 | [
{
"authors": "Jacob Austin; Daniel D Johnson; Jonathan Ho; Daniel Tarlow; Rianne Van Den; Jacob Berg; Daniel D Austin; Jonathan Johnson; Daniel Ho; Rianne Tarlow; Van Den; Berg",
"journal": "",
"ref_id": "b0",
"title": "Structured denoising diffusion models in discrete state-spaces",
"year": "... | [
{
"formula_coordinates": [
2,
325.57,
166.65,
194.46,
19.34
],
"formula_id": "formula_0",
"formula_text": "q(x t |x t-1 ) = N (x t ; 1 -β t • x t-1 ; β t I) (1)"
},
{
"formula_coordinates": [
2,
351.69,
230.82,
168.34,
26.03
],
... | A Survey of Diffusion Models in Natural Language Processing | This survey paper provides a comprehensive review of the use of diffusion models in natural language processing (NLP). Diffusion models are a class of mathematical models that aim to capture the diffusion of information or signals across a network or manifold. In NLP, diffusion models have been used in a variety of app... | Hao Zou; Zae Myung Kim; Dongyeop Kang | [
{
"figure_caption": "Figure 1 :1Figure 1: The yearly number of both published and preprinted papers on diffusion models for NLP. For year 2023, the blue bar shows the number collected until the end of April 2023, and the dashed gray bar shows the estimated number for the whole year.",
"figure_data": "",
... | [{"Category": "Methodological Basis", "Citation": "(Sohl-Dickstein et al., 2015b)", "Explanation": "The cited work by Sohl-Dickstein et al. provides a foundational methodology for diffusion models in the field of image generation, which the citing paper builds upon in the context of natural language processing."}, {"Ca... | |
[
{
"figure_ref": [
"fig_5",
"fig_5",
"fig_5"
],
"heading": "Introduction",
"publication_ref": [
"b44",
"b43",
"b26",
"b28",
"b27",
"b49",
"b52",
"b15",
"b62",
"b62",
"b11",
"b28",
"b23",
"b0",
... | 2023-05-24 | [
{
"authors": "Max Bain; Arsha Nagrani; Gül Varol; Andrew Zisserman",
"journal": "",
"ref_id": "b0",
"title": "Frozen in time: A joint video and image encoder for end-to-end retrieval",
"year": "2021"
},
{
"authors": "Yogesh Balaji; Martin Renqiang Min; Bing Bai; Rama Chellappa; Hans Pete... | [
{
"formula_coordinates": [
5,
114.4,
384.42,
390.26,
45.04
],
"formula_id": "formula_0",
"formula_text": "Q = {(m i , y (i,t) )|i = 1, 2, . . . , H•W } pairs from the n-th view n = 1, 2, . . . , N = {(m (i,n) , y (i,n,t) = √ ᾱy (i,n,0) + √ 1 -ᾱt ϵ i )|i=1, 2, . . . , H•... | T1: Scaling Diffusion Probabilistic Fields to High-Resolution on Unified Visual Modalities | Diffusion Probabilistic Field (DPF) [63] models the distribution of continuous functions defined over metric spaces. While DPF shows great potential for unifying data generation of various modalities including images, videos, and 3D geometry, it does not scale to a higher data resolution. This can be attributed to the ... | Kangfu Mei; Mo Zhou; Vishal M Patel | [
{
"figure_caption": "Figure 1 :1Figure 1: Illustration of the field models' capability of modeling visual content distributions. The underlying data distribution is simplified into the 1-D space for demonstration. The score network learns the distribution through the attention among coordinate-signal pairs, whi... | [{"Category": "Methodological Basis", "Citation": "[27]", "Explanation": "The cited work on diffusion models provides the forward and backward processes that the citing paper adopts in their image generation model."}, {"Category": "Extension or Continuation", "Citation": "[29,28]", "Explanation": "The cited work on mul... | |
[
{
"figure_ref": [
"fig_0"
],
"heading": "I. INTRODUCTION",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b7",
"b9",
"b10",
"b11",
"b12",
"b13",
"b1",
"b3",
"b7",
"b8",
"b10",
... | 2023-07-25 | 10.1145/3511808.3557289 | [
{
"authors": "H Zhang; E Yuan; W Guo; Z He; J Qin; H Guo; B Chen; X Li; R Tang",
"journal": "ACM",
"ref_id": "b0",
"title": "Disentangling past-future modeling in sequential recommendation via dual networks",
"year": "2022"
},
{
"authors": "B Hidasi; A Karatzoglou; L Baltrunas; D Tikk",
... | [
{
"formula_coordinates": [
3,
311.98,
173.06,
251.06,
22.49
],
"formula_id": "formula_0",
"formula_text": "X u = x 1 → x 2 → • • • → x |X u | , chronologically records |X u |"
},
{
"formula_coordinates": [
3,
396.76,
309.12,
166.28,
... | TriMLP: Revenge of a MLP-like Architecture in Sequential Recommendation | Sequential recommenders concentrate on modeling the transmission patterns shrouded in sequences of historical user-item interactive behaviors (or referred as token) and inferring dynamic preferences over candidate items. Fueled by diverse advanced neural network architectures like RNN, CNN and Transformer, existing met... | Yiheng Jiang; Yuanbo Xu; Yongjian Yang; Funing Yang; Pengyang Wang; Hui Xiong | [
{
"figure_caption": "Fig. 1 .1Fig.1. Accuracy/Efficiency traded-off on QB-Video. Along the vertical axis, the higher, the better recommendation performance; along the horizontal axis, the more left, the less inference cost.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_... | [{"Category": "Methodological Basis", "Citation": "[1]", "Explanation": "The cited work introduces the concept of sequential recommendation and provides a framework for understanding the process of mining dependencies among tokens and inferring preferences over time, which the citing paper builds upon in its research o... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b29",
"b2",
"b32",
"b24",
"b10",
"b0"
],
"table_ref": [],
"text": "Generalization to unseen tasks has been explored and investigated on zero-/few-shot NLP tasks by performing multi-task l... | 2023-05-24 | 10.1109/CVPR.2017.670 | [
{
"authors": "Jean-Baptiste Alayrac; Jeff Donahue; Pauline Luc; Antoine Miech; Iain Barr; Yana Hasson; Karel Lenc; Arthur Mensch; Katie Millican; Malcolm Reynolds",
"journal": "",
"ref_id": "b0",
"title": "Flamingo: a visual language model for few-shot learning",
"year": "2022"
},
{
"aut... | [
{
"formula_coordinates": [
5,
70.87,
240.81,
101.04,
10.63
],
"formula_id": "formula_0",
"formula_text": "{(V 1 , l 1 ), ..., (V m , l m )}"
},
{
"formula_coordinates": [
6,
77.48,
74.77,
440.32,
18.97
],
"formula_id": "form... | GRILL: Grounded Vision-language Pre-training via Aligning Text and Image Regions | Generalization to unseen tasks is an important ability for few-shot learners to achieve better zero-/few-shot performance on diverse tasks. However, such generalization to visionlanguage tasks including grounding and generation tasks has been under-explored; existing few-shot VL models struggle to handle tasks that inv... | Woojeong Jin; Subhabrata Mukherjee; Yu Cheng; Yelong Shen; Weizhu Chen; Ahmed Hassan Awadallah; Damien Jose; Xiang Ren | [
{
"figure_caption": "Figure 2 :2Figure2: Illustration of GRILL. Our model is a sequence-to-sequence transformer that uses a vision transformer (ViT)(Dosovitskiy et al., 2021;Liu et al., 2021) to process images with patch embeddings, where each patch represents a fixed-size region of the image. We replace the re... | [{"Category": "Methodological Basis", "Citation": "(Sanh et al., 2021)", "Explanation": "The cited work by Sanh et al. provides a method of performing multi-task learning with task-specific prompts, which the citing paper adopts to explore generalization to unseen tasks in zero-/few-shot NLP tasks."}, {"Category": "Ext... |
[
{
"figure_ref": [
"fig_1"
],
"heading": "INTRODUCTION",
"publication_ref": [
"b8",
"b19",
"b34",
"b4",
"b22",
"b27",
"b5",
"b28",
"b15",
"b3",
"b39",
"b30",
"b38",
"b33",
"b16",
"b41",
"b12",
... | 2023-08-11 | 10.1145/3583780.3614999 | [
{
"authors": "Åke Björck",
"journal": "Handbook of numerical analysis",
"ref_id": "b0",
"title": "Least squares methods",
"year": "1990"
},
{
"authors": "C John; Butcher",
"journal": "J. Comput. Appl. Math",
"ref_id": "b1",
"title": "Numerical methods for ordinary differenti... | [
{
"formula_coordinates": [
2,
96.2,
99.07,
176.81,
44.93
],
"formula_id": "formula_0",
"formula_text": "𝑥 ! 𝑥 !\"# 𝑥 !\"$ 𝑥 !\"% 𝑥 !\"& 𝑒 ! 𝑒 !\"# 𝑒 !\"$ 𝑒 !\"% 𝑒 !\"&"
},
{
"formula_coordinates": [
3,
96.17,
212.59,
198.41,
... | Optimal Linear Subspace Search: Learning to Construct Fast and High-Quality Schedulers for Diffusion Models | In recent years, diffusion models have become the most popular and powerful methods in the field of image synthesis, even rivaling human artists in artistic creativity. However, the key issue currently limiting the application of diffusion models is its extremely slow generation process. Although several methods were p... | Zhongjie Duan; Chengyu Wang; Cen Chen; Jun Huang; Weining Qian | [
{
"figure_caption": "Figure 1 :1Figure1: The complete generation process of diffusion models consists of hundreds of steps for gradual denoising. Diffusion schedulers speed up this process by skipping some steps but may make destructive changes to images.",
"figure_data": "",
"figure_id": "fig_1",
"... | [{"Category": "Methodological Basis", "Citation": "[40]", "Explanation": "The cited work highlights the slow sampling procedure in diffusion models, which the citing paper uses as a basis for discussing the limitations of the framework in terms of practicability."}, {"Category": "Methodological Basis", "Citation": "[31... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b4",
"b18",
"b6",
"b7",
"b9",
"b26",
"b15",
"b30",
"b33",
"b29",
"b33",
"b19",
"b19",
"b16",
"b24",
"b25",
"b14",
"b3",
"... | 2023-11-15 | 10.1145/3442188.3445922 | [
{
"authors": "Emily M Bender; Timnit Gebru; Angelina Mcmillan-Major; Shmargaret Shmitchell",
"journal": "Association for Computing Machinery",
"ref_id": "b0",
"title": "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?",
"year": "2021"
},
{
"authors": "Yoav Benjamini;... | [] | Emergent Inabilities? Inverse Scaling Over the Course of Pretraining | Does inverse scaling only occur as a function of model size, or can it also occur over the course of training? We carry out an exploratory study investigating whether the performance of language models on specific tasks can decrease (while general performance remains high) during training on the language modeling task.... | James A Michaelov; Benjamin K Bergen | [
{
"figure_caption": "Figure 1 :1Figure1: Performance of the 8 Pythia(Biderman et al., 2023) models at 8 stages over the course of training at the two multiple-choice variants of TRUTHFULQA(Lin et al., 2022) and the 10 multiple-choice winners of the Inverse Scaling Prize(McKenzie et al., 2023b).",
"figure_da... | [{"Category": "Supporting Evidence", "Citation": "(Brown et al., 2020)", "Explanation": "The cited work by Brown et al. provides evidence that increased number of model parameters and training dataset size positively impact model performance, which supports the claim made in the citing paper that bigger is usually bett... |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b24",
"b1",
"b13",
"b34",
"b38",
"b3",
"b40",
"b14",
"b6",
"b31"
],
"table_ref": [],
"text": "Text-based question-answering datasets derive answ... | 2023-05-24 | 10.18653/v1/2020.emnlp-main.19 | [
{
"authors": "Joshua Ainslie; Santiago Ontanon; Chris Alberti; Vaclav Cvicek; Zachary Fisher; Philip Pham; Anirudh Ravula; Sumit Sanghai; Qifan Wang; Li Yang",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "ETC: Encoding long and structured inputs in transformers",... | [
{
"formula_coordinates": [
5,
355.06,
286.04,
66.61,
10.63
],
"formula_id": "formula_0",
"formula_text": "h q = BERT(Q),"
},
{
"formula_coordinates": [
5,
318.85,
302.58,
206.29,
27.17
],
"formula_id": "formula_1",
"form... | TACR: A Table-alignment-based Cell-selection and Reasoning Model for Hybrid Question-Answering | Hybrid Question-Answering (HQA), which targets reasoning over tables and passages linked from table cells, has witnessed significant research in recent years. A common challenge in HQA and other passage-table QA datasets is that it is generally unrealistic to iterate over all table rows, columns, and linked passages to... | Jian Wu; Yicheng Xu; Yan Gao; Jian-Guang Lou; Börje F Karlsson; Manabu Okumura | [
{
"figure_caption": "Figure 1 :1Figure 1: Example from the HybridQA dataset. The top sentence is the original question, and words in different colors show different parts of questions required for reasoning in different modalities. the two headers in blue-dashed boxes are column names aligned with the given que... | [{"Category": "Data Source", "Citation": "(Rajpurkar et al., 2016)", "Explanation": "The cited work by Rajpurkar et al. (2016) provides a text-based question-answering dataset that is used as a foundational element in the citing paper for deriving answers based on reasoning over given passages."}, {"Category": "Data So... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b3",
"b21",
"b25",
"b4",
"b21",
"b17",
"b23",
"b19",
"b25",
"b24",
"b4",
"b24",
"b27",
"b22",
"b9",
"b18",
"b13",
"b2",
"... | 2023-05-24 | [
{
"authors": "Chris Burges; Tal Shaked; Erin Renshaw; Ari Lazier; Matt Deeds; Nicole Hamilton; Greg Hullender",
"journal": "",
"ref_id": "b0",
"title": "Learning to rank using gradient descent",
"year": "2005"
},
{
"authors": "N Craswell; B Mitra; E Yilmaz; D Campos",
"journal": "",
... | [
{
"formula_coordinates": [
2,
399.45,
353.73,
87.27,
10.63
],
"formula_id": "formula_0",
"formula_text": "s i = f (q, d i , r i , D)."
},
{
"formula_coordinates": [
2,
306.14,
559.09,
219.12,
17.74
],
"formula_id": "formula_... | Fusion-in-T5: Unifying Document Ranking Signals for Improved Information Retrieval | Common IR pipelines are typically cascade systems that may involve multiple rankers and/or fusion models to integrate different information step-by-step. In this paper, we propose a novel re-ranker named Fusion-in-T5 (FiT5), which integrates document text information, retrieval features, and global document information... | Shi Yu; Chenghao Fan; Chenyan Xiong; David Jin; Zhiyuan Liu; Zhenghao Liu | [
{
"figure_caption": "Figure 1 :1Figure 1: Architecture of Fusion-in-T5. The query, document, and ranking feature are filled in the input template to form the input. We use retrieval score as the ranking feature.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figu... | [{"Category": "Methodological Basis", "Citation": "(Croft et al., 2010)", "Explanation": "The cited work by Croft et al. provides a foundational definition of the information retrieval task, which the citing paper uses to frame its research on building models for the task."}, {"Category": "Data Source", "Citation": "(Y... | |
[
{
"figure_ref": [
"fig_0",
"fig_2"
],
"heading": "Introduction",
"publication_ref": [
"b7",
"b21",
"b19",
"b26",
"b21",
"b31",
"b26",
"b11",
"b14",
"b3",
"b13",
"b22",
"b25",
"b24"
],
"table_ref":... | 2023-11-02 | [
{
"authors": "P L Bartlett; M I Jordan; J D Mcauliffe",
"journal": "Journal of the American Statistical Association",
"ref_id": "b0",
"title": "Convexity, classification, and risk bounds",
"year": "2006"
},
{
"authors": "S Ben-David; J Blitzer; K Crammer; F Pereira",
"journal": "",
... | [
{
"formula_coordinates": [
2,
245.05,
265.03,
259.61,
9.96
],
"formula_id": "formula_0",
"formula_text": "R(f ) = E pte(x,y) [ℓ(f (x), y)],(1)"
},
{
"formula_coordinates": [
2,
308.24,
295.31,
99.5,
14.32
],
"formula_id": "f... | Generalizing Importance Weighting to A Universal Solver for Distribution Shift Problems | Distribution shift (DS) may have two levels: the distribution itself changes, and the support (i.e., the set where the probability density is non-zero) also changes. When considering the support change between the training and test distributions, there can be four cases: (i) they exactly match; (ii) the training suppor... | Tongtong Fang; Nan Lu; Gang Niu; Masashi Sugiyama | [
{
"figure_caption": "Figure 1 :1Figure 1: An illustration of the relationship between the training support and the test support.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "This is binary classification to distinguish re... | [{"Category": "Methodological Basis", "Citation": "(Quionero-Candela et al., 2009)", "Explanation": "The cited work by Quionero-Candela et al. (2009) is used to highlight the issue of distribution shift in deep supervised classification, which can lead to poor generalization in practice."}, {"Category": "Methodological... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b241",
"b253",
"b182",
"b241",
"b253",
"b253",
"b110",
"b111",
"b82",
"b126",
"b260",
"b84",
"b276",
"b58",
"b59",
"b227",
"b72",
... | 2023-05-24 | 10.1201/b17320 | [
{
"authors": "E A Abioye; M S Z Abidin; M S A Mahmud; S Buyamin; M H I Ishak; M K I Abd Rahman; A O Otuoze; P Onotu; M S A Ramli",
"journal": "Computers and Electronics in Agriculture",
"ref_id": "b0",
"title": "A review on monitoring and advanced control strategies for precision irrigation",
"y... | [
{
"formula_coordinates": [
4,
59.36,
334.63,
229.31,
24.14
],
"formula_id": "formula_0",
"formula_text": "min 𝜃 𝜆 𝑙 ⋅ ∑ (𝐱,𝑦)∈ 𝐿 𝑠𝑢𝑝 (𝐱, 𝑦, 𝜃)+𝜆 𝑢 ⋅ ∑ 𝐱∈ 𝑈 𝑢𝑛𝑠𝑢𝑝 (𝐱, 𝜃), (1)"
},
{
"formula_coordinates": [
5,
331.51,
... | Label-Efficient Learning in Agriculture: A Comprehensive Review | The past decade has witnessed many great successes of machine learning (ML) and deep learning (DL) applications in agricultural systems, including weed control, plant disease diagnosis, agricultural robotics, and precision livestock management. Despite tremendous progresses, one downside of such ML/DL models is that th... | Jiajia Li; Dong Chen; Xinda Qi; Zhaojian Li; Yanbo Huang; Daniel Morris; Xiaobo Tan | [
{
"figure_caption": "Figure 1 :1Figure 1: The PRISMA guideline flowchart used in this review. The figure first row illustrates initially selected articles based on the keywords that enhanced the initial filtering before other exclusion criteria are applied.",
"figure_data": "",
"figure_id": "fig_0",
... | [{"Category": "Methodological Basis", "Citation": "(Walter et al., 2017)", "Explanation": "The cited work by Walter et al. (2017) provides a comprehensive overview of the use of information and communication technologies in smart farming, which serves as a methodological basis for the citing paper to discuss the integr... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b24",
"b2",
"b40",
"b0",
"b33",
"b28",
"b8",
"b25",
"b26",
"b13",
"b18",
"b16"
],
"table_ref": [],
"text": "Deep neural networks (DNN) are ubiquitously... | 2023-05-24 | 10.3115/v1/D14-1181 | [
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel Ziegler; Jeffrey Wu; Clemens Winter; Ch... | [
{
"formula_coordinates": [
2,
492.26,
95.85,
6.27,
47.15
],
"formula_id": "formula_0",
"formula_text": "… … … …"
},
{
"formula_coordinates": [
3,
70.87,
732.81,
220.08,
14.63
],
"formula_id": "formula_1",
"formula_text":... | SELFOOD: Self-Supervised Out-Of-Distribution Detection via Learning to Rank | Deep neural classifiers trained with crossentropy loss (CE loss) often suffer from poor calibration, necessitating the task of out-ofdistribution (OOD) detection. Traditional supervised OOD detection methods require expensive manual annotation of in-distribution and OOD samples. To address the annotation bottleneck, we... | Dheeraj Mekala; Adithya Samavedhi; Chengyu Dong; Jingbo Shang | [
{
"figure_caption": "Figure 1 :1Figure 1: CE Loss and SELFOOD optimization for two documents D 1 , D 2 belonging to Sports and Arts. CE loss increases the scores corresponding to the Sports class for D 1 and Arts class for D 2 , implying an intradocument comparison. Instead, SELFOOD compares the softmax scores ... | [{"Category": "Methodological Basis", "Citation": "(Liu et al., 2019)", "Explanation": "The cited work by Liu et al. provides a deep neural network (DNN) model for text classification, which the citing paper adopts as a method for their research on OOD detection."}, {"Category": "Data Source", "Citation": "(Devlin et a... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b40",
"b28",
"b57",
"b59",
"b49",
"b60",
"b2",
"b11",
"b7",
"b50",
"b38",
"b39",
"b37",
"b1",
"b46",
"b39",
"b17",
"b10",
... | 2023-10-28 | 10.18653/v1/N19-1264 | [
{
"authors": "Kristjan Arumae; Fei Liu",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Guiding extractive summarization with question-answering rewards",
"year": "2019"
},
{
"authors": "Yuntao Bai; Andy Jones; Kamal Ndousse; Amanda Askell; Anna Chen; N... | [
{
"formula_coordinates": [
2,
198.41,
672.36,
62.9,
15.86
],
"formula_id": "formula_0",
"formula_text": "(n) 1 , S (n) 2 )} N n=1"
},
{
"formula_coordinates": [
2,
348.71,
734.77,
176.43,
36.25
],
"formula_id": "formula_1",
... | DecipherPref: Analyzing Influential Factors in Human Preference Judgments via GPT-4 | Human preference judgments are pivotal in guiding large language models (LLMs) to produce outputs that align with human values. Human evaluations are also used in summarization tasks to compare outputs from various systems, complementing existing automatic metrics. Despite their significance, however, there has been li... | Yebowen Hu; Kaiqiang Song; Sangwoo Cho; Xiaoyang Wang; Hassan Foroosh; Fei Liu | [
{
"figure_caption": "Figure 3 :3Figure 3: Two prompts for GPT-4 to assess whether a given summary is fluent (TOP) or clear (BOTTOM).",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "3",
"figure_type": "figure"
},
{
"figure_caption": "Figure 4 :4Figure 4: Two prompts for GPT-4 ... | [{"Category": "Methodological Basis", "Citation": "(Papineni et al., 2002)", "Explanation": "The cited work by Papineni et al. (2002) introduces the BLEU and ROUGE metrics for comparing system summaries with reference texts based on word overlap, which serves as a methodological basis for the citing paper in evaluating... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b48",
"b43",
"b51",
"b3",
"b27",
"b33",
"b37",
"b27",
"b44",
"b9",
"b11",
"b22",
"b3"
],
"table_ref": [],
"text": "The recent years have witnesse... | 2023-07-05 | [
{
"authors": "Vamsi Aribandi; Yi Tay; Tal Schuster; Jinfeng Rao; Huaixiu Steven Zheng; Sanket Vaibhav Mehta; Honglei Zhuang; Dara Vinh Q Tran; Jianmo Bahri; Ni",
"journal": "",
"ref_id": "b0",
"title": "Ext5: Towards extreme multi-task scaling for transfer learning",
"year": "2021"
},
{
... | [
{
"formula_coordinates": [
5,
153.57,
210.28,
66.8,
13.18
],
"formula_id": "formula_0",
"formula_text": "10 ¤110"
}
] | Mixture-of-Experts Meets Instruction Tuning: A Winning Combination for Large Language Models | Sparse Mixture-of-Experts (MoE) is a neural architecture design that can be utilized to add learnable parameters to Large Language Models (LLMs) without increasing inference cost. Instruction tuning is a technique for training LLMs to follow instructions. We advocate combining these two approaches, as we find that MoE ... | Sheng Shen; Le Hou; Yanqi Zhou; Nan Du; Shayne Longpre; Jason Wei; Hyung Won; Barret Zoph; William Fedus; Xinyun Chen; Tu Vu; Yuexin Wu; Wuyang Chen; Albert Webson; Yunxuan Li; Vincent Zhao; Hongkun Yu; Kurt Keutzer; Trevor Darrell; Denny Zhou; † Google | [
{
"figure_caption": "Figure 2 :2Figure 2: Average zero performance of FLAN-MOE models versus FLAN-T5 dense models for similar effective FLOPs per token over the 57 MMLU tasks and 23 BBH tasks.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"f... | [{"Category": "Methodological Basis", "Citation": "[49]", "Explanation": "The cited work introduces the transformer-based language models that the citing paper builds upon to enhance their performance in NLP tasks."}, {"Category": "Extension or Continuation", "Citation": "[44,52,4,28,34,38]", "Explanation": "The cited ... | |
[
{
"figure_ref": [
"fig_1",
"fig_1"
],
"heading": "I. INTRODUCTION",
"publication_ref": [
"b1",
"b0",
"b2",
"b4",
"b2",
"b6",
"b0",
"b7",
"b7"
],
"table_ref": [],
"text": "Egocentric videos are captured from a first-person ... | 2023-07-26 | [
{
"authors": "K C Chan; S Zhou; X Xu; C C Loy",
"journal": "",
"ref_id": "b0",
"title": "Investigating tradeoffs in real-world video super-resolution",
"year": "2022"
},
{
"authors": "",
"journal": "",
"ref_id": "b1",
"title": "Tips for extended gopro battery life + accessori... | [
{
"formula_coordinates": [
3,
326.87,
623.39,
236.17,
31.8
],
"formula_id": "formula_0",
"formula_text": "I (t) motion = t+⌊N/2⌋ j=t-⌊N/2⌋ r • I (j) clear + (1 -N • r) • I (t) clear (1)"
},
{
"formula_coordinates": [
4,
90.05,
389.09,
209... | EgoVSR: Towards High-Quality Egocentric Video Super-Resolution | Due to the limitations of capture devices and scenarios, egocentric videos frequently have low visual quality, mainly caused by high compression and severe motion blur. With the increasing application of egocentric videos, there is an urgent need to enhance the quality of these videos through super-resolution. However,... | Yichen Chi; Junhao Gu; Jiamiao Zhang; Wenming Yang; Yapeng Tian | [
{
"figure_caption": "VSR results from our method (top) and Real-BasicVSR[1] (bottom).",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 1 :1Fig. 1: Egocentric videos generally contain temporally changing visual scenes with... | [{"Category": "Supporting Evidence", "Citation": "[2]", "Explanation": "The cited work provides information on the limitations of wearable cameras in terms of image quality and battery life, which supports the discussion on the challenges of egocentric VSR in the citing paper."}, {"Category": "Methodological Basis", "C... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b40",
"b13",
"b10",
"b23",
"b19",
"b34",
"b17",
"b31",
"b0",
"b22",
"b4",
"b7",
"b41",
"b36",
"b11",
"b25",
"b32",
"b43",
... | 2023-11-03 | 10.1145/3442188.3445922 | [
{
"authors": "Sandhini Agarwal; Gretchen Krueger; Jack Clark; Alec Radford; Jong Wook Kim; Miles Brundage",
"journal": "",
"ref_id": "b0",
"title": "Evaluating clip: towards characterization of broader capabilities and downstream implications",
"year": "2021"
},
{
"authors": "Peter Ander... | [
{
"formula_coordinates": [
3,
315.02,
737.28,
209.99,
35.68
],
"formula_id": "formula_0",
"formula_text": "AccG,C = 1 N N i=1 1[S(c good i , ri, Ii) > S(c bad i , ri, Ii)],(1)"
},
{
"formula_coordinates": [
4,
120.2,
569.29,
169.53,
... | Gender Biases in Automatic Evaluation Metrics for Image Captioning | Model-based evaluation metrics (e.g., CLIP-Score and GPTScore) have demonstrated decent correlations with human judgments in various language generation tasks. However, their impact on fairness remains largely unexplored. It is widely recognized that pretrained models can inadvertently encode societal biases, thus empl... | Haoyi Qiu; Zi-Yi Dou; Tianlu Wang; Asli Celikyilmaz; Nanyun Peng | [
{
"figure_caption": "Good caption: a woman who is reading Bad caption: a man who is reading Reference: a photo of a woman who is reading CLIPScore 0",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Example... | [{"Category": "Supporting Evidence", "Citation": "(Zhang et al., 2019)", "Explanation": "The cited work by Zhang et al. (2019) introduces the BERTScore evaluation metric, which has shown promising performance in terms of correlation with human judgments in generation tasks. This work provides foundational evidence for ... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b14",
"b10",
"b27",
"b13",
"b24",
"b6",
"b6",
"b10",
"b1",
"b13",
"b27",
"b14",
"b8",
"b27",
"b13",
"b25",
"b18",
"b0",
"... | [
{
"authors": "Z Allen-Zhu; Y Li; Z Song",
"journal": "",
"ref_id": "b0",
"title": "A convergence theory for deep learning via overparameterization",
"year": "2019"
},
{
"authors": "M Arjovsky; S Chintala; L Bottou",
"journal": "",
"ref_id": "b1",
"title": "Wasserstein generat... | [
{
"formula_coordinates": [
3,
169.9,
169.99,
334.71,
39.72
],
"formula_id": "formula_0",
"formula_text": "f θ S is dF (Q θ S , P0) = sup g∈F E S lim sup m→∞ 1 m m j=1 g(z j , S) -Ex∼P 0 [g(x, S)] ,(1)"
},
{
"formula_coordinates": [
3,
134.47,
2... | On the Generalization of Diffusion Model | The diffusion probabilistic generative models are widely used to generate highquality data. Though they can synthetic data that does not exist in the training set, the rationale behind such generalization is still unexplored. In this paper, we formally define the generalization of the generative model, which is measure... | Mingyang Yi; Jiacheng Sun; Zhenguo Li | [
{
"figure_caption": "Figure 1 :1Figure 1: The first figure is the averaged distance ∥x t -x * t ∥ per dimension (3×32×32) over 50k samples of generated CIFAR10. The second figure randomly samples a batch of x t and x * t with the same x T = x * T and T = 50.",
"figure_data": "",
"figure_id": "fig_0",
... | [{"Category": "Methodological Basis", "Citation": "(Somepalli et al., 2022)", "Explanation": "The cited work provides empirical evidence of the diffusion model generating data that is combined with parts of the training set, which the citing paper uses to highlight the need for extrapolating in the application of the d... | ||
[
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [
"b32",
"b1",
"b31",
"b47",
"b26",
"b7",
"b35",
"b35",
"b24",
"b6"
],
"table_ref": [],
"text": "For safe autonomous driving, predicting a vehicle's future trajectory... | 2023-05-24 | [
{
"authors": "Inhwan Bae; Jin-Hwi Park; Hae-Gon Jeon",
"journal": "",
"ref_id": "b0",
"title": "Non-probability sampling network for stochastic human trajectory prediction",
"year": "2022"
},
{
"authors": "Alexander Barth; Uwe Franke",
"journal": "IEEE",
"ref_id": "b1",
"titl... | [
{
"formula_coordinates": [
4,
133.12,
93.32,
354.01,
138.5
],
"formula_id": "formula_0",
"formula_text": "Future GT Trajectory (𝐱𝐱 + ) Lane Graph 𝓖𝓖 Interaction edge 𝐳𝐳 e -/+ Interaction Feature (𝐡𝐡 𝑅𝑅 ) F Trajectory Samples (𝐘𝐘) Intention Feature ( 𝐡𝐡 𝐼�... | LEVERAGING FUTURE RELATIONSHIP REASONING FOR VEHICLE TRAJECTORY PREDICTION | Understanding the interaction between multiple agents is crucial for realistic vehicle trajectory prediction. Existing methods have attempted to infer the interaction from the observed past trajectories of agents using pooling, attention, or graph-based methods, which rely on a deterministic approach. However, these me... | Daehee Park; Hobin Ryu; Yunseo Yang; Jegyeong Cho; Jiwon Kim; Kuk-Jin Yoon | [
{
"figure_caption": "FigureFigure 2: Lane segments represented in different colors.",
"figure_data": "",
"figure_id": "fig_1",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "al. (2014)) or VAE (Kingma & Welling (2013)) have been employed to address this issue. GAN-ba... | [{"Category": "Methodological Basis", "Citation": "(Lin et al., 2000)", "Explanation": "The cited work by Lin et al. (2000) provides a heuristic method for predicting a vehicle's future trajectory, which the citing paper adopts as a foundational approach for early trajectory prediction models."}, {"Category": "Methodol... | |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b11",
"b16",
"b20",
"b42",
"b4",
"b21",
"b18",
"b26"
],
"table_ref": [],
"text": "Advances in multilingual natural language processing (NLP) technologies (D... | 2023-05-24 | 10.1162/tacl_a_00416 | [
{
"authors": "David Ifeoluwa Adelani; Jade Abbott; Graham Neubig; D' Daniel; Julia Souza; Constantine Kreutzer; Chester Lignos; Happy Palen-Michel; Shruti Buzaaba; Sebastian Rijhwani; Stephen Ruder; Israel Mayhew; Shamsuddeen H Abebe Azime; Chris Chinenye Muhammad; Joyce Emezue; Perez Nakatumba-Nabende; Aremu O... | [
{
"formula_coordinates": [
3,
339.65,
120.03,
185.49,
24.32
],
"formula_id": "formula_0",
"formula_text": "u l = performance l theoretical max performance (1)"
},
{
"formula_coordinates": [
3,
376.94,
349.88,
148.2,
29.59
],
... | GlobalBench: A Benchmark for Global Progress in Natural Language Processing | Despite the major advances in NLP, significant disparities in NLP system performance across languages still exist. Arguably, these are due to uneven resource allocation and sub-optimal incentives to work on less resourced languages. To track and further incentivize the global development of equitable language technolog... | Yueqi Song; Catherine Cui; Simran Khanuja; Pengfei Liu; Fahim Faisal; Alissa Ostapenko; Genta Indra Winata; Alham Fikri; Samuel Cahyawijaya; Yulia Tsvetkov; Antonios Anastasopoulos; Graham Neubig | [
{
"figure_caption": "Figure 1 :1Figure 1: GlobalBench Design: A leaderboard for each task is separately maintained. Each leaderboard contains a multi-faceted evaluation of submitted systems, along with a ranking of the most under-served languages. More details can be found in Section 2.",
"figure_data": "",... | [{"Category": "Extension or Continuation", "Citation": "(Dabre et al., 2020)", "Explanation": "The cited work by Dabre et al. (2020) is an important contribution to the field of multilingual NLP, and the citing paper extends the research by exploring the possibilities of NLP systems that benefit all people around the w... |
[
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [
"b46",
"b14",
"b40",
"b52",
"b80",
"b10",
"b59",
"b55",
"b26",
"b16",
"b73",
"b58",
"b74",
"b18"
],
"table_ref": [],
"text": "Pretrained (Ra... | 2024-03-26 | 10.18653/v1/2021.emnlp-main.397 | [
{
"authors": "",
"journal": "",
"ref_id": "b0",
"title": "K training instances containing responses that abruptly end in a colon",
"year": ""
},
{
"authors": " Werra",
"journal": "",
"ref_id": "b1",
"title": "We use the huggingface TRL",
"year": "2020"
},
{
"autho... | [
{
"formula_coordinates": [
3,
212.52,
464.59,
292.15,
22.24
],
"formula_id": "formula_0",
"formula_text": "J(θ) = max θ x∈X d πref (x) y∈Y R(x, y, ⋆)π θ (y|x) (1)"
},
{
"formula_coordinates": [
3,
140.42,
515.67,
364.24,
20.24
]... | LEFTOVER-LUNCH: ADVANTAGE-BASED OFFLINE REINFORCEMENT LEARNING FOR LANGUAGE MODELS | Reinforcement Learning with Human Feedback (RLHF) is the most prominent method for Language Model (LM) alignment. However, RLHF is an unstable and data-hungry process that continually requires new high-quality LM-generated data for finetuning. We introduce Advantage-Leftover Lunch RL (A-LOL), a new class of offline pol... | Ashutosh Baheti; Ximing Lu; Ronan Le Bras; Mark Riedl | [
{
"figure_caption": "Figure 1 :1Figure 1: Illustration of Advantage-Leftover Lunch RL in practice. We first supervised finetune the reference policy (π ref ) on the training data as a precursor to A-LOL training. Then, an external reward model is employed to train the value estimate layer (V πref ) on frozen π ... | [{"Category": "Methodological Basis", "Citation": "(Touvron et al., 2023b)", "Explanation": "The cited work introduces the use of RLHF in finetuning LMs, which serves as a methodological basis for the citing paper to build upon in their research on improving the quality and safety of LMs."}, {"Category": "Extension or ... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b3",
"b32",
"b1",
"b6"
],
"table_ref": [],
"text": "U.S. Supreme Court justices hear some of the most important cases in the country, resolving disagreements among lower courts, adjudicating the cons... | 2023-05-24 | 10.1016/j.artint.2020.103387 | [
{
"authors": "Bryant Michael L Anderson; Shuda Lee; Jon Li; Ben Go; Luoyan Sutandio; Zhou",
"journal": "",
"ref_id": "b0",
"title": "On the Types, Frequency, Uses and Characteristics of Metalanguage in Conversation",
"year": "2006"
},
{
"authors": "Katie Atkinson; Trevor Bench-Capon; Da... | [] | CuRIAM: Corpus re Interpretation and Metalanguage in U.S. Supreme Court Opinions | Most judicial decisions involve the interpretation of legal texts; as such, judicial opinion requires the use of language as a medium to comment on or draw attention to other language. Language used this way is called metalanguage. We develop an annotation schema for categorizing types of legal metalanguage and apply o... | Michael Kranzlein; Nathan Schneider; Kevin Tobia | [
{
"figure_caption": "(4) ...the term \"violation\" referred to [ D the \"[a]ct or instance of violating, or state of being violated].\" Webster's New International Dictionary 2846 (2d ed. 1949) (Webster's Second).",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "fig... | [{"Category": "Data Source", "Citation": "(Tobia, 2021)", "Explanation": "The cited work is a source of information for furthering legal and linguistic scholarship on judicial interpretation, which the citing paper may use as a reference for its own research."}, {"Category": "Data Source", "Citation": "(Go\u017ad\u017a... |
[
{
"figure_ref": [
"fig_0"
],
"heading": "",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b10",
"b11",
"b12",
"b13",
"b14",
"b11",
"b12",
"b10",
... | 2023-10-21 | 10.1080/01431168908903939 | [
{
"authors": "A Singh",
"journal": "International Journal of Remote Sensing",
"ref_id": "b0",
"title": "Review article digital change detection techniques using remotely-sensed data",
"year": "1989"
},
{
"authors": "H Chen; Z Shi",
"journal": "Remote Sensing",
"ref_id": "b1",
... | [
{
"formula_coordinates": [
5,
93.78,
531.85,
206.24,
12.73
],
"formula_id": "formula_1",
"formula_text": "I hr s , I lr s = swap(I hr d , I lr u , u, v, crop size),(3)"
},
{
"formula_coordinates": [
5,
48.96,
552.69,
251.06,
20.91
... | Continuous Cross-resolution Remote Sensing Image Change Detection | Most contemporary supervised Remote Sensing (RS) image Change Detection (CD) approaches are customized for equal-resolution bitemporal images. Real-world applications raise the need for cross-resolution change detection, aka, CD based on bitemporal images with different spatial resolutions. Given training samples of a ... | Hao Chen; Haotian Zhang; Keyan Chen; Chenyao Zhou; Song Chen; Zhengxia Zou; Zhenwei Shi; = 𝜙; 𝛿𝑥 𝑞; {𝒁; 𝑥 𝑞; C S 𝑞; } {𝒁; 𝑥 𝑞 | [
{
"figure_caption": "Fig. 1 .1Fig. 1. Illustration of continuous cross-resolution change detection, i.e., CD towards varying resolution difference ratios between the HR image and the relatively LR image.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},... | [{"Category": "Methodological Basis", "Citation": "[6]", "Explanation": "The cited work on convolutional neural networks (CNNs) is used as a methodological basis for the change detection techniques in the citing paper."}, {"Category": "Methodological Basis", "Citation": "[7]", "Explanation": "The cited work on vision t... |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b29",
"b41",
"b42",
"b52",
"b34",
"b28",
"b59",
"b23",
"b43",
"b58",
"b10",
"b0"
],
"table_ref": [],
"text": "Entity linking aims to... | 2023-05-24 | 10.48550/arxiv.2207.04108 | [
{
"authors": "Omar Adjali; Romaric Besançon; Olivier Ferret; Hervé Le Borgne; Brigitte Grau",
"journal": "Springer",
"ref_id": "b0",
"title": "Multimodal entity linking for tweets",
"year": "2020-04-14"
},
{
"authors": "Tom Ayoola; Joseph Fisher; Andrea Pierleoni",
"journal": "",
... | [
{
"formula_coordinates": [
8,
184.79,
392.02,
82.08,
15.24
],
"formula_id": "formula_0",
"formula_text": "K-1 j=0 exp(s t (r, e j ))"
},
{
"formula_coordinates": [
8,
106.17,
597.65,
150.02,
33.3
],
"formula_id": "formula_1"... | AMELI: Enhancing Multimodal Entity Linking with Fine-Grained Attributes | We propose attribute-aware multimodal entity linking, where the input is a mention described with a text and image, and the goal is to predict the corresponding target entity from a multimodal knowledge base (KB) where each entity is also described with a text description, a visual image and a set of attributes and val... | Barry Menglong; Yu Chen; Qifan Wang; Sijia Wang; Minqian Liu; Zhiyang Xu; ♠ Licheng; Yu ♡ Lifu | [
{
"figure_caption": ":Figure 1 :1Figure 1: An example for our attribute-aware multimodal entity linking. Left: review text and image; Right: product title, image, description, and attributes. In order to link the mention ASUS laptop to the target entity, we need to be aware of the attributes, e.g., memory and S... | [{"Category": "Methodological Basis", "Citation": "(Onoe and Durrett, 2020)", "Explanation": "The cited work by Onoe and Durrett (2020) provides a methodological basis for the citing paper by introducing techniques for entity linking in text."}, {"Category": "Methodological Basis", "Citation": "(Zhang et al., 2021b)", ... |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b18",
"b49",
"b11",
"b10",
"b44",
"b13",
"b24",
"b33",
"b31"
],
"table_ref": [],
"text": "Deep language representations have become the dominant form ... | 2023-06-01 | 10.1016/j.inffus.2019.12.012 | [
{
"authors": "Stefano Baccianella; Andrea Esuli; Fabrizio Sebastiani",
"journal": "European Language Resources Association (ELRA",
"ref_id": "b0",
"title": "SentiWordNet 3.0: An enhanced lexical resource for sentiment analysis and opinion mining",
"year": "2010"
},
{
"authors": "Alejandr... | [
{
"formula_coordinates": [
4,
70.87,
586.86,
218.27,
41.1
],
"formula_id": "formula_0",
"formula_text": "S c i = {s j,c i } m j=1 associated with each category c i ∈ C, where C = {c i } d i=1 ."
},
{
"formula_coordinates": [
4,
70.87,
643.01,
... | SENTECON: Leveraging Lexicons to Learn Human-Interpretable Language Representations | Although deep language representations have become the dominant form of language featurization in recent years, in many settings it is important to understand a model's decisionmaking process. This necessitates not only an interpretable model but also interpretable features. In particular, language must be featurized i... | Victoria Lin; Louis-Philippe Morency | [
{
"figure_caption": "Figure 1 :1Figure 1: A comparison of lexicon-based language representations and SENTECON. While lexicons encode wordlevel category counts, SENTECON parses whole sentences and encodes sentence-level category intensities.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": ... | [{"Category": "Methodological Basis", "Citation": "(Lin et al., 2020)", "Explanation": "The cited work by Lin et al. provides a method for understanding the relationship between language patterns and specific outcomes, which the citing paper builds upon in their research on affective computing, computational social sci... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b63",
"b12",
"b58",
"b70",
"b29",
"b88",
"b85",
"b43",
"b83",
"b90",
"b96",
"b31",
"b46",
"b8",
"b55",
"b77",
"b54",
"b67",
... | 2023-11-09 | 10.18653/v1/2021.acl-long.551 | [
{
"authors": "Ahmed Abdelali; Sabit Hassan; Hamdy Mubarak; Kareem Darwish; Younes Samih",
"journal": "",
"ref_id": "b0",
"title": "Pre-training bert on arabic tweets: Practical considerations",
"year": "2021"
},
{
"authors": "Muhammad Abdul-Mageed; Abdelrahim Elmadany; El Moatez; Billah ... | [
{
"formula_coordinates": [
4,
94.45,
173.79,
403.59,
29.62
],
"formula_id": "formula_0",
"formula_text": "M² R K K R K S Lev. R M R M R R ARETA R K D R"
},
{
"formula_coordinates": [
4,
251.72,
359.79,
262.9,
413.59
],
"form... | Advancements in Arabic Grammatical Error Detection and Correction: An Empirical Investigation | Grammatical error correction (GEC) is a wellexplored problem in English with many existing models and datasets. However, research on GEC in morphologically rich languages has been limited due to challenges such as data scarcity and language complexity. In this paper, we present the first results on Arabic GEC using two... | Bashar Alhafni; Inoue † Go; Christian Khairallah; Nizar Habash | [
{
"figure_caption": "Figure 1 :1Figure1: An example showing the differences between the alignments of the M 2 scorer, a standard Levenshtein distance, ARETA, and our proposed algorithm. The edit operations are keep (K), replace (R), insert (I), delete (D), merge (M), and split (S). Dotted lines between the erro... | [{"Category": "Supporting Evidence", "Citation": "(Ng et al., , 2014;;Bryant et al., 2019)", "Explanation": "The cited works provide a foundation for the organization of shared tasks in English grammatical error correction, which is a key factor in the development of SOTA GEC systems."}, {"Category": "Extension or Cont... |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b3",
"b21",
"b25",
"b19",
"b32",
"b28",
"b30",
"b0",
"b28",
"b19"
],
"table_ref": [],
"text": "Language models (LMs) are remarkably effective in... | 10.18653/v1/2020.emnlp-main.749 | [
{
"authors": "Sid Black; Gao Leo; Phil Wang; Connor Leahy; Stella Biderman",
"journal": "",
"ref_id": "b0",
"title": "GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow",
"year": "2021"
},
{
"authors": "Shuyang Cao; Lu Wang",
"journal": "",
"ref_id": "b1",... | [
{
"formula_coordinates": [
2,
116.48,
437.17,
127.04,
28.95
],
"formula_id": "formula_0",
"formula_text": "y t ∼ p θ (y t | c, x, y <t ) ∝ exp logit θ (y t | c, x, y <t )"
},
{
"formula_coordinates": [
2,
319.87,
89.02,
190.32,
43.5... | Trusting Your Evidence: Hallucinate Less with Context-aware Decoding | Language models (LMs) often struggle to pay enough attention to the input context, and generate texts that are unfaithful or contain hallucinations. To mitigate this issue, we present context-aware decoding (CAD), which follows a contrastive output distribution that amplifies the difference between the output probabili... | Weijia Shi; Xiaochuang Han; Mike Lewis; Yulia Tsvetkov; Luke Zettlemoyer; Scott Yih | [
{
"figure_caption": "Figure 1 :1Figure 1: An illustration of context-aware decoding.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: OPT models of varying sizes consistently benefit from CAD. The x-axis ... | [{"Category": "Methodological Basis", "Citation": "(Chan et al., 2022)", "Explanation": "The cited work provides a framework for understanding the role of context knowledge in language model generation, which the citing paper builds upon in their study of the balance between prior and context knowledge."}, {"Category":... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b11",
"b29",
"b27",
"b19",
"b42",
"b28",
"b40"
],
"table_ref": [],
"text": "Using natural language for image generation and manipulation is a straightforward and intuitive approach ... | 2023-06-05 | [
{
"authors": "Rameen Abdal; Yipeng Qin; Peter Wonka",
"journal": "",
"ref_id": "b0",
"title": "Image2stylegan: How to embed images into the stylegan latent space",
"year": "2019"
},
{
"authors": "Rameen Abdal; Yipeng Qin; Peter Wonka",
"journal": "",
"ref_id": "b1",
"title": ... | [
{
"formula_coordinates": [
4,
312.88,
616.65,
154.96,
17.17
],
"formula_id": "formula_0",
"formula_text": "q (x t | x t-1 ) := N √ 1 -β t x t-1 , β t I"
},
{
"formula_coordinates": [
5,
211.94,
90.83,
292.73,
23.61
],
"formu... | ChatFace: Chat-Guided Real Face Editing via Diffusion Latent Space Manipulation | Can you add a little smile to the person? I'll make this one look a little happier. Please add more smiles. Okay, smile is set to medium. I want to make the girl | Dongxu Yue; Qin Guo; Munan Ning; Jiaxi Cui; Yuesheng Zhu; Li Yuan | [
{
"figure_caption": "Figure 2 :2Figure 2: Overview of ChatFace inference pipeline. Large language model parsing queries from user for solving facial image editing tasks, which then enable the activation of corresponding facial attributes and control over the editing strength in diffusion semantic latent space."... | [{"Category": "Supporting Evidence", "Citation": "[12]", "Explanation": "The cited work on Generative Adversarial Networks (GANs) provides the foundational approach for image generation and manipulation in the citing paper."}, {"Category": "Supporting Evidence", "Citation": "[30]", "Explanation": "The cited work on CLI... | |
[
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [
"b0",
"b1",
"b0",
"b2",
"b3",
"b4",
"b5"
],
"table_ref": [],
"text": "Currency is a universally accepted medium of exchange that enables the trade of goods and services. Typically, i... | [
{
"authors": "",
"journal": "Investopedia",
"ref_id": "b0",
"title": "Currency",
"year": "2022-07-22"
},
{
"authors": "",
"journal": "CFI team",
"ref_id": "b1",
"title": "Currency",
"year": "2019-11-14"
},
{
"authors": " Fxssi",
"journal": "",
"ref_id": "b... | [
{
"formula_coordinates": [
4,
366.65,
727.27,
176.16,
19.28
],
"formula_id": "formula_0",
"formula_text": "𝑋 𝑛𝑜𝑟𝑚 = 𝑋 𝑐𝑢𝑟𝑟𝑒𝑛𝑡 -𝑋 𝑚𝑖𝑛 𝑋 𝑚𝑎𝑥 -𝑋 𝑚𝑖𝑛 (1)"
}
] | Applications of Machine Learning in Detecting Afghan Fake Banknotes | Fake currency, unauthorized imitation money lacking government approval, constitutes a form of fraud. Particularly in Afghanistan, the prevalence of fake currency poses significant challenges and detrimentally impacts the economy. While banks and commercial establishments employ authentication machines, the public lack... | Hamida Ashna; Ziaullah Momand | [
{
"figure_caption": "Fig. 1 .1Fig. 1. Flow of proposed method",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Fig. 3 .3Fig. 3. AFN banknote cropped two main features",
"figure_data": "",
"figure_id": "fig_1",
"fi... | [{"Category": "Data Source", "Citation": "[1]", "Explanation": "The cited work provides a historical overview of the evolution of currency, including its origins in Ancient Egypt and the current form of currency in use today."}, {"Category": "Data Source", "Citation": "[2]", "Explanation": "The cited work highlights th... | ||
[
{
"figure_ref": [
"fig_0",
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b36",
"b12",
"b30",
"b15",
"b17"
],
"table_ref": [],
"text": "Large language models (LLMs) have been shown to hallucinate, meaning that... | 2023-05-24 | 10.18653/v1/N19-1300 | [
{
"authors": "Hussam Alkaissi; I Samy; Mcfarlane",
"journal": "Cureus",
"ref_id": "b0",
"title": "Artificial hallucinations in chatgpt: implications in scientific writing",
"year": "2023"
},
{
"authors": "Yejin Bang; Samuel Cahyawijaya; Nayeon Lee; Wenliang Dai; Dan Su; Bryan Wilie; Hol... | [
{
"formula_coordinates": [
3,
319.65,
600.42,
205.36,
26.84
],
"formula_id": "formula_0",
"formula_text": "scoreT (C, a) = 1 n -1 n i=2 1(C tf (i) = \"true\"), (1)"
}
] | Mastering the ABCDs of Complex Questions: Answer-Based Claim Decomposition for Fine-grained Self-Evaluation | When answering complex questions, large language models (LLMs) may produce answers that do not satisfy all criteria of the question. While existing self-evaluation techniques aim to detect if such answers are correct, these techniques are unable to determine which criteria of the question are satisfied by the generated... | Nishant Balepur; Jie Huang; Samraj Moorjani; Hari Sundaram; Kevin Chen; Chuan Chang; Philip K Dick; George Mccaffrey; John Robert Rozanov | [
{
"figure_caption": "Figure 1 :1Figure 1: Using answer-based claim decomposition to verify ChatGPT's answer to an OBSCUREQA question.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Claim 1 : 3 :Figure 3 :133Figure 3: Qualit... | [{"Category": "Supporting Evidence", "Citation": "(Alkaissi and McFarlane, 2023)", "Explanation": "The cited work provides evidence that LLMs are capable of generating untruthful statements, which is a key factor in the discussion of LLMs and their potential to cause harm in decision-making."}, {"Category": "Supporting... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b75",
"b41",
"b41",
"b17",
"b61",
"b31",
"b65",
"b24"
],
"table_ref": [],
"text": "Metrics that capture human-like attributes of dialog agents can help inform dialog agents th... | 2023-09-16 | 10.1073/pnas.2218523120 | [
{
"authors": "Muhammad Abdul-Mageed; Anneke Buffone; Hao Peng; Salvatore Giorgi; Johannes C Eichstaedt; Lyle H Ungar",
"journal": "",
"ref_id": "b0",
"title": "Recognizing pathogenic empathy in social media",
"year": "2017"
},
{
"authors": "Gavin Abercrombie; Amanda Cercas Curry; Tanvi D... | [] | Psychological Metrics for Dialog System Evaluation | We present metrics for evaluating dialog systems through a psychologically-grounded "human" lens in which conversational agents express a diversity of both states (e.g., emotion) and traits (e.g., personality), just as people do. We present five interpretable metrics from established psychology that are fundamental to ... | Salvatore Giorgi; Shreya Havaldar; Ahmed ‡ Farhan; Zuhaib Akhtar; Shalaka Vaidya; Gary Pan; Lyle H Ungar; H Andrew Schwartz; João Sedoc | [
{
"figure_caption": "Figure 1 :1Figure1: Mauve score (traditional metric) and emotion matching (psychological metric) to evaluate two conversation snippets (turns). Humans rated the top response as highly appropriate and the bottom response as inappropriate. The dialog agent's response in both conversations rec... | [{"Category": "Methodological Basis", "Citation": "(Chen et al., 2021)", "Explanation": "The cited work by Chen et al. provides a discussion on the limitations of traditional automatic metrics in evaluating open-domain dialog systems, which the citing paper builds upon to highlight the need for a more comprehensive app... |
[
{
"figure_ref": [
"fig_0",
"fig_1",
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b38",
"b39",
"b42",
"b37",
"b18",
"b52",
"b7",
"b28",
"b1",
"b23",
"b45",
"b50",
"b51",
"b27"
],
... | 2023-07-30 | [
{
"authors": "Rahaf Aljundi; Francesca Babiloni; Mohamed Elhoseiny; Marcus Rohrbach; Tinne Tuytelaars",
"journal": "",
"ref_id": "b0",
"title": "Memory aware synapses: Learning what (not) to forget",
"year": "2018"
},
{
"authors": "Rahaf Aljundi; Punarjay Chakravarty; Tinne Tuytelaars",
... | [
{
"formula_coordinates": [
3,
62.4,
97.44,
466.18,
195.72
],
"formula_id": "formula_0",
"formula_text": "ℒ clf Dataset D i Module i B-E-E-R-S F i Stage-II : Multi-lingual modeling Rehearsal Set M i ℋ 1 Frozen F 1 F 2 F i φ 1 φ 2 φ i TDR•東京 ℒ clf"
},
{
"formula_c... | MRN: Multiplexed Routing Network for Incremental Multilingual Text Recognition | Multilingual text recognition (MLTR) systems typically focus on a fixed set of languages, which makes it difficult to handle newly added languages or adapt to everchanging data distribution. In this paper, we propose the Incremental MLTR (IMLTR) task in the context of incremental learning (IL), where different language... | Tianlun Zheng; Zhineng Chen; Bingchen Huang; Wei Zhang; Yu-Gang Jiang | [
{
"figure_caption": "Figure 1 .1Figure 1. Incremental multilingual text recognition (IMLTR) focuses on the practical scenario where different languages are introduced sequentially. The goal is to accurately recognize the newly introduced language while maintaining high recognition accuracy for previously seen l... | [{"Category": "Methodological Basis", "Citation": "[39,40,43,38,19,53]", "Explanation": "The cited works provide deep learning methods that have improved the accuracy of scene text recognition, which the citing paper leverages in its research on the task of reading text in natural scenes."}, {"Category": "Extension or ... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b5",
"b11",
"b16",
"b4",
"b14",
"b5",
"b10",
"b3",
"b0",
"b23",
"b24",
"b26",
"b24",
"b27"
],
"table_ref": [],
"text": "In recent years, Na... | 10.18653/v1/d15-1075 | [
{
"authors": "Armen Aghajanyan; Akshat Shrivastava; Anchit Gupta; Naman Goyal; Luke Zettlemoyer; Sonal Gupta",
"journal": "",
"ref_id": "b0",
"title": "Better fine-tuning by reducing representational collapse",
"year": "2021"
},
{
"authors": "R Samuel; Gabor Bowman; Christopher Angeli; C... | [
{
"formula_coordinates": [
3,
136.94,
485,
152.79,
19.75
],
"formula_id": "formula_0",
"formula_text": "θt+1 = θt -η ∂L (θt) ∂θt (1)"
},
{
"formula_coordinates": [
3,
105.34,
628.32,
184.4,
47.56
],
"formula_id": "formula_1"... | Bi-Drop: Enhancing Fine-tuning Generalization via Synchronous sub-net Estimation and Optimization | Pretrained language models have achieved remarkable success in natural language understanding. However, fine-tuning pretrained models on limited training data tends to overfit and thus diminish performance. This paper presents Bi-Drop, a fine-tuning strategy that selectively updates model parameters using gradients fro... | Shoujie Tong; Heming Xia; Damai Dai; Runxin Xu; Tianyu Liu; Binghuai Lin; Yunbo Cao; Zhifang Sui | [
{
"figure_caption": "Figure 2 :2Figure 2: An overall illustration of Bi-Drop. Bi-Drop splits each training step into three sub-steps: (1) Multiple forward propagations: each mini-batch sample goes through the forward pass multiple times (denoted as k) with dropout;(2) sub-net selection: an advanced strategy is ... | [{"Category": "Methodological Basis", "Citation": "(Phang et al., 2018)", "Explanation": "The cited work by Phang et al. (2018) provides foundational data and insights on the challenges of maintaining generalization performance in fine-tuning methods, which the citing paper builds upon to address the issue of overfitti... | |
[
{
"figure_ref": [
"fig_2"
],
"heading": "Introduction",
"publication_ref": [
"b22",
"b55",
"b71",
"b36",
"b59",
"b30",
"b72",
"b77",
"b33",
"b7",
"b14",
"b58",
"b36",
"b32"
],
"table_ref": [],
"text... | 2023-10-10 | 10.1007/978-3-031-19815-1_11 | [
{
"authors": "Youngmin Baek; Bado Lee; Dongyoon Han; Sangdoo Yun; Hwalsuk Lee",
"journal": "",
"ref_id": "b0",
"title": "Character region awareness for text detection",
"year": "2019"
},
{
"authors": "Leilani Battle; Peitong Duan; Zachery Miranda; Dana Mukusheva; Remco Chang; Michael Sto... | [] | UniChart: A Universal Vision-language Pretrained Model for Chart Comprehension and Reasoning | Charts are widely used for data analysis, providing visual representations and insights into complex data. To facilitate chart-based data analysis using natural language, several downstream tasks have been introduced recently such as chart question answering and chart summarization. However, existing methods for these ... | Ahmed Masry; Parsa Kavehzadeh; Xuan Long; Enamul Hoque; Shafiq Joty | [
{
"figure_caption": "Sum two leftmost values of gray line </answer> <open-ended question answering> What is the birth rate in the U.S. from 2005 to 2019? </answer> 90.0",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Characte... | [{"Category": "Methodological Basis", "Citation": "(Hoque et al., 2022)", "Explanation": "The cited work provides a general introduction to the use of information visualizations in data analysis, which serves as a methodological basis for the citing paper in understanding the importance of charts in data analysis."}, {... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b61",
"b58",
"b27",
"b4",
"b3",
"b2",
"b62",
"b44",
"b27",
"b31",
"b55",
"b4",
"b27",
"b55",
"b29",
"b40"
],
"table_ref": [],
"... | 10.18653/v1/D19-6004 | [
{
"authors": "Simon Baron-Cohen; Alan M Leslie; Uta Frith",
"journal": "Cognition",
"ref_id": "b0",
"title": "Does the autistic child have a \"theory of mind",
"year": "1985"
},
{
"authors": "Simon Baron-Cohen; Michelle O 'riordan; Valerie Stone; Rosie Jones; Kate Plaisted",
"journal... | [] | Clever Hans or Neural Theory of Mind? Stress Testing Social Reasoning in Large Language Models | The escalating debate on AI's capabilities warrants developing reliable metrics to assess machine "intelligence." Recently, many anecdotal examples were used to suggest that newer large language models (LLMs) like ChatGPT and GPT-4 exhibit Neural Theory-of-Mind (N-ToM); however, prior work reached conflicting conclusio... | Natalie Shapira; Mosh Levy; Seyed Hossein Alavi; Xuhui Zhou; Yejin Choi; Yoav Goldberg; Maarten Sap; Vered Shwartz | [
{
"figure_caption": "FalseBelief. In the false-belief examples from Kosinski (2023), the protagonist's belief about the content of the container is different from its actual contents. The examples are variants of the corresponding original tests from psychology, e.g. the unexpected contents examples are variant... | [{"Category": "Supporting Evidence", "Citation": "(2022)", "Explanation": "The cited work by (2022) shows that LLMs lack the ability to demonstrate N-ToM, which is a key finding that supports the claims made in the citing paper about the limitations of LLMs in this area."}, {"Category": "Extension or Continuation", "Ci... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b11",
"b12",
"b13",
"b14",
"b15",
"b16",
"b17",
"b18"
],
"t... | 2023-10-19 | 10.1109/TBDATA.2019.2921572 | [
{
"authors": " Openai",
"journal": "OpenAI",
"ref_id": "b0",
"title": "",
"year": "2023"
},
{
"authors": "Le Teven; Angela Scao; Christopher Fan; Ellie Akiki; Suzana Pavlick; Daniel Ilic; Roman Hesslow; Alexandra Castagné; François Sasha Luccioni; Matthias Yvon; Jonathan Gallé; Alexander... | [
{
"formula_coordinates": [
2,
70.87,
155.9,
441.11,
146.95
],
"formula_id": "formula_0",
"formula_text": "< Threshold, Continue! Complement Complement … … … … … Beam 𝒊 + 𝟏 Figure 1:"
},
{
"formula_coordinates": [
4,
284.4,
287.25,
143.5... | ALLIES: Prompting Large Language Model with Beam Search | With the advance of large language models (LLMs), the research field of LLM applications becomes more and more popular and the idea of constructing pipelines to accomplish complex tasks by stacking LLM API calls come true. However, this kind of methods face two limitations: narrow information coverage and low fault tol... | Hao Sun; Xiao Liu; Yeyun Gong; Yan Zhang; Daxin Jiang; Linjun Yang; Nan Duan | [
{
"figure_caption": "Figure 2 :2Figure 2: The abstract process of ALLIES.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "2",
"figure_type": "figure"
},
{
"figure_caption": "Figure 3 :3Figure 3: Performance comparison w.r.t. hyperparameters on NQ dataset.",
"figure_data":... | [{"Category": "Methodological Basis", "Citation": "[Brown et al., 2020]", "Explanation": "The cited work by Brown et al. introduces the in-context learning method, which the citing paper adopts to generate responses and answer queries using large language models."}, {"Category": "Extension or Continuation", "Citation":... |
[
{
"figure_ref": [
"fig_1"
],
"heading": "Introduction",
"publication_ref": [
"b44",
"b25",
"b13",
"b30",
"b12",
"b26",
"b55",
"b25",
"b37",
"b7",
"b28",
"b23",
"b20",
"b22",
"b21",
"b32",
"b4"... | 2023-05-24 | [
{
"authors": "Alaaeldin Ali; Hugo Touvron; Mathilde Caron; Piotr Bojanowski; Matthijs Douze; Armand Joulin; Ivan Laptev; Natalia Neverova; Gabriel Synnaeve; Jakob Verbeek",
"journal": "Advances in Neural Information Processing Systems",
"ref_id": "b0",
"title": "Xcit: Cross-covariance image transfor... | [
{
"formula_coordinates": [
4,
88.61,
166.31,
406.27,
188.29
],
"formula_id": "formula_0",
"formula_text": "H × W × 3 ! \" × # \" × D ! ! $ × # $ × 𝐷 \" ! %& × # %& × 𝐷 # ! '( × # '( × 𝐷 $ Dual Attention Block Multi-Head Partition Attention 𝐻 % × 𝑊 % × 𝐷 % 𝑘×𝑘 DW... | Dual Path Transformer with Partition Attention | This paper introduces a novel attention mechanism, called dual attention, which is both efficient and effective. The dual attention mechanism consists of two parallel components: local attention generated by Convolutional Neural Networks (CNNs) and long-range attention generated by Vision Transformers (ViTs). To addres... | Zhengkai Jiang; Liang Liu; Jiangning Zhang; Yabiao Wang; Mingang Chen; Chengjie Wang | [
{
"figure_caption": "Figure 2 .2Figure 2. Parameters vs. ImageNet Accuracy. DualFormers outperform state-of-the-art Vision Transformers while having fewer parameters and FLOPs. The model names, T, XS, S, and B, denote tiny, extra-small, small, and base, respectively.",
"figure_data": "",
"figure_id": "f... | [{"Category": "Methodological Basis", "Citation": "[26]", "Explanation": "The cited work on Transformers provides the foundational basis for the use of Vision Transformers (ViTs) in architecture design for computer vision tasks in the citing paper."}, {"Category": "Methodological Basis", "Citation": "[13]", "Explanatio... | |
[
{
"figure_ref": [
"fig_1",
"fig_1"
],
"heading": "Introduction",
"publication_ref": [
"b3",
"b52",
"b54",
"b53",
"b40",
"b12",
"b6",
"b32",
"b11",
"b42",
"b23",
"b38",
"b44",
"b2",
"b35",
"b38... | 2023-11-14 | 10.18653/v1/N19-1300 | [
{
"authors": "Meysam Alizadeh; Fabrizio Gilardi; Emma Hoes; Jonathan Klüser; Mael Kubli; Nahema Marchal",
"journal": "Journal of Quantitative Description: Digital Media",
"ref_id": "b0",
"title": "Content moderation as a political issue: The twitter discourse around trump's ban",
"year": "2022"
... | [
{
"formula_coordinates": [
6,
119.43,
572.75,
76.24,
14.27
],
"formula_id": "formula_0",
"formula_text": "s ′ ki = f (r ki , e ki )."
},
{
"formula_coordinates": [
7,
104.71,
583.18,
151.77,
30.56
],
"formula_id": "formula_1... | Using Natural Language Explanations to Rescale Human Judgments | The rise of large language models (LLMs) has brought a critical need for high-quality humanlabeled data, particularly for processes like human feedback and evaluation. A common practice is to label data via consensus annotation over crowdworker judgments. However, annotators' judgments for subjective tasks can differ i... | Manya Wadhwa Jifan; Chen Junyi; Jessy Li; Greg Durrett | [
{
"figure_caption": "Figure 1 :1Figure1: Overview of our method. By feeding explanations that annotators write into an LLM, we can rescale their coarse-grained judgment to a 100-point scale.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"fig... | [{"Category": "Supporting Evidence", "Citation": "(Brown et al., 2020)", "Explanation": "The cited work by Brown et al. (2020) highlights the shift towards more subjective tasks in the field of language models, which is a key focus of the citing paper."}, {"Category": "Supporting Evidence", "Citation": "(Ouyang et al.,... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b17",
"b21",
"b19",
"b0",
"b15",
"b42",
"b15",
"b5",
"b45",
"b35",
"b46",
"b29",
"b44",
"b8",
"b47",
"b35",
"b15"
],
"table_r... | 10.5281/zenodo.5297715 | [
{
"authors": "Jacob Austin; Daniel D Johnson; Jonathan Ho; Daniel Tarlow; Rianne Van Den; Berg",
"journal": "",
"ref_id": "b0",
"title": "Structured denoising diffusion models in discrete state-spaces",
"year": "2021"
},
{
"authors": "Nikita Balagansky; Daniil Gavrilov",
"journal": ... | [
{
"formula_coordinates": [
2,
328.22,
166.06,
176.08,
32.7
],
"formula_id": "formula_0",
"formula_text": "0 = logits-initialization(w c:c+B ) wc:c+B t = √ ᾱt wc:c+B 0 + √ 1 -ᾱt ϵ"
},
{
"formula_coordinates": [
2,
317.57,
359.6,
195.41,
... | David helps Goliath: Inference-Time Collaboration Between Small Specialized and Large General Diffusion LMs | Diffusion-based language models are emerging as a promising alternative to autoregressive LMs: they approach the competence of autoregressive LMs while offering nuanced controllability at inference time. While autoregressive LMs have benefited immensely from scaling and instruction-based learning, existing studies of d... | Xiaochuang Han; Sachin Kumar; Yulia Tsvetkov; Marjan Ghazvininejad | [
{
"figure_caption": "Figure 2 :2Figure 2: Training and decoding algorithms for SSD-2. The training algorithm describes the training objective at an arbitrary context length c. The decoding algorithm can be applied multiple rounds by appending the generation from one round to the context for the next. The decodi... | [{"Category": "Methodological Basis", "Citation": "(Ho et al., 2020)", "Explanation": "The cited work by Ho et al. (2020) serves as a methodological basis for diffusion-based generative models in continuously valued data such as images, audio, and video, which the citing paper extends to discrete text data."}, {"Catego... | |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b0",
"b23",
"b36",
"b1",
"b28",
"b10",
"b9",
"b50",
"b6",
"b31",
"b6"
],
"table_ref": [],
"text": "Tools to support research activities of... | 2023-12-01 | 10.1145/3589955 | [
{
"authors": "Tal August; Lucy Lu Wang; Jonathan Bragg; Marti A Hearst; Andrew Head; Kyle Lo",
"journal": "ACM Trans. Comput.-Hum. Interact. Just Accepted",
"ref_id": "b0",
"title": "Paper plain: Making medical research papers approachable to healthcare consumers with natural language processing",
... | [] | A Question Answering Framework for Decontextualizing User-facing Snippets from Scientific Documents | Many real-world applications (e.g., note taking, search) require extracting a sentence or paragraph from a document and showing that snippet to a human outside of the source document. Yet, users may find snippets difficult to understand as they lack context from the original document. In this work, we use language mode... | Benjamin Newman; Luca Soldaini; Raymond Fok; Arman Cohan; Kyle Lo | [
{
"figure_caption": "Figure 1 :1Figure 1: Illustration of two user-facing scenarios requiring snippet decontextualization. (Top) A citation graph explorer surfacing citation context snippets to explain relationships between papers. (Bottom) An AI research assistant providing snippets as attributions. Highlighte... | [{"Category": "Supporting Evidence", "Citation": "(August et al., 2023)", "Explanation": "The cited work by August et al. provides a specific example of how snippets can be used to help readers understand technical documents more efficiently."}, {"Category": "Supporting Evidence", "Citation": "(Fok et al., 2023b)", "Ex... |
[
{
"figure_ref": [
"fig_0"
],
"heading": "INTRODUCTION",
"publication_ref": [
"b24",
"b56",
"b9",
"b45",
"b62",
"b35",
"b40",
"b11",
"b26",
"b35",
"b26",
"b46",
"b43",
"b37",
"b73",
"b31",
"b30... | 2023-10-09 | 10.1609/aaai.v35i3.16370 | [
{
"authors": "Davide Abati; Jakub Tomczak; Tijmen Blankevoort; Simone Calderara; Rita Cucchiara; Babak Ehteshami Bejnordi",
"journal": "",
"ref_id": "b0",
"title": "Conditional channel gated networks for task-aware continual learning",
"year": "2020"
},
{
"authors": "Akiko Aizawa",
"... | [
{
"formula_coordinates": [
4,
372.67,
154.97,
97.65,
9.65
],
"formula_id": "formula_0",
"formula_text": "Q [θ ∈ R α (Q)] ≥ 1 -α."
},
{
"formula_coordinates": [
4,
213.38,
572.21,
48.03,
14.11
],
"formula_id": "formula_1",
... | IBCL: ZERO-SHOT MODEL GENERATION FOR TASK TRADE-OFFS IN CONTINUAL LEARNING | Like generic multi-task learning, continual learning has the nature of multiobjective optimization, and therefore faces a trade-off between the performance of different tasks. That is, to optimize for the current task distribution, it may need to compromise performance on some previous tasks. This means that there exis... | Pengyuan Lu; Michele Caprio; Eric Eaton; Insup Lee | [
{
"figure_caption": "Figure 1 :1Figure 1: The IBCL workflow. The orange polytopes are the geometric representations of FGCSs.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Top to bottom: Results on Cel... | [{"Category": "Methodological Basis", "Citation": "(Kendall et al., 2018)", "Explanation": "The cited work by Kendall et al. provides the foundational theory of multi-task learning, which the citing paper builds upon to address the multi-objective optimization problem in the context of lifelong or continual learning."}... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b16",
"b3",
"b12",
"b23",
"b7",
"b28",
"b30",
"b14",
"b8",
"b28",
"b24",
"b17"
],
"table_ref": [],
"text": "Chinese Spelling Correction (CSC) is a task... | 2023-05-24 | 10.18653/v1/2020.acl-main.81 | [
{
"authors": "Xingyi Cheng; Weidi Xu; Kunlong Chen; Shaohua Jiang; Feng Wang; Taifeng Wang; Wei Chu; Yuan Qi",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "SpellGCN: Incorporating phonological and visual similarities into language models for Chinese spelling chec... | [
{
"formula_coordinates": [
4,
84,
539.22,
205.74,
21.01
],
"formula_id": "formula_0",
"formula_text": "si = xi, 1 ≤i ≤ n initi-n, finali-n, n + 1 ≤i ≤ n + n .(1)"
},
{
"formula_coordinates": [
4,
320.05,
467.18,
180.9,
51.73
],
... | Disentangled Phonetic Representation for Chinese Spelling Correction | Chinese Spelling Correction (CSC) aims to detect and correct erroneous characters in Chinese texts. Although efforts have been made to introduce phonetic information (Hanyu Pinyin) in this task, they typically merge phonetic representations with character representations, which tends to weaken the representation effect... | Zihong Liang; Xiaojun Quan; Qifan Wang | [
{
"figure_caption": "Figure 1 :1Figure 1: The architecture of the proposed DORM, which consists of a phonetics-aware input sequence S, an encoder with separation mask, a pinyin-to-character objective, and a self-distillation module. X is the original input sentence, R is the pinyin sequence of X, Y is the corre... | [{"Category": "Methodological Basis", "Citation": "(Martins and Silva, 2004)", "Explanation": "The cited work by Martins and Silva (2004) provides a foundational method for Chinese spelling correction (CSC) that is adopted in the citing paper to address the task of detecting and correcting erroneous characters in Chine... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b7",
"b11",
"b23",
"b14",
"b27",
"b26",
"b28",
"b29",
"b30",
"b33",
"b13",
"b16",
"b20",
"b21",
"b3",
"b5",
"b8",
"b15",
... | 2023-05-24 | [
{
"authors": "A Gary; Edwin R Atkinson; Hancock",
"journal": "IEEE Trans. Image Process",
"ref_id": "b0",
"title": "Recovery of surface orientation from diffuse polarization",
"year": "2006"
},
{
"authors": "Yunhao Ba; Alex Gilbert; Franklin Wang; Jinfa Yang; Rui Chen; Yiqin Wang; Lei Y... | [
{
"formula_coordinates": [
4,
140.39,
285.8,
242.56,
22.54
],
"formula_id": "formula_0",
"formula_text": "I = (P 0 + P 45 + P 90 + E 135 ) 2(1)"
},
{
"formula_coordinates": [
6,
84.43,
283.28,
298.52,
20.75
],
"formula_id": ... | Polarimetric Imaging for Perception | Autonomous driving and advanced driver-assistance systems rely on a set of sensors and algorithms to perform the appropriate actions and provide alerts as a function of the driving scene. Typically, the sensors include color cameras, radar, lidar and ultrasonic sensors. Strikingly however, although light polarization i... | Michael Baltaxe; Tomer Pe'er; Dan Levi | [
{
"figure_caption": "Figure 1 :1Figure 1: Examples of collected data. Each row shows a different sample with RGB (left), AoLP (middle left), DoLP (middle right) and lidar projected on RGB (right). The cyclic color map in the AoLP images goes from red for 0 • to magenta for 179 • . In the DoLP images black corre... | [{"Category": "Methodological Basis", "Citation": "[8,12,24]", "Explanation": "The cited works provide a basis for free space detection in ADAS and autonomous vehicles, which the citing paper builds upon to plan and act appropriately on the road."}, {"Category": "Methodological Basis", "Citation": "[15,28]", "Explanati... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b50",
"b6",
"b49",
"b27",
"b7",
"b55",
"b49",
"b52",
"b43",
"b42"
],
"table_ref": [],
"text": "Transformer-based (Vaswani et al., 2017) language models (LMs) have ... | 2023-11-04 | 10.18653/v1/2021.deelio-1.3 | [
{
"authors": "Joshua Ainslie; Tao Lei; Santiago Michiel De Jong; Siddhartha Ontañón; Yury Brahma; David Zemlyanskiy; Mandy Uthus; James Guo; Yi Lee-Thorp; Yun-Hsuan Tay; Sumit Sung; Sanghai",
"journal": "",
"ref_id": "b0",
"title": "CoLT5: Faster long-range transformers with conditional computation... | [
{
"formula_coordinates": [
3,
311.01,
542.83,
208.54,
33.96
],
"formula_id": "formula_0",
"formula_text": "L = - 1 N n i=1 m i t=1 log p(x i t | x i 1 , . . . , x i t-1 , σ <i )."
},
{
"formula_coordinates": [
5,
102.42,
73.15,
409.75,
... | Adapting Language Models to Compress Contexts | Transformer-based language models (LMs) are powerful and widely-applicable tools, but their usefulness is constrained by a finite context window and the expensive computational cost of processing long text documents. We propose to adapt pre-trained LMs into AutoCompressors. These language models are capable of compress... | Alexis Chevalier; Alexander Wettig; Anirudh Ajith; Danqi Chen | [
{
"figure_caption": "Figure 2 :2Figure2: Perplexity on 2048 held-out tokens given different numbers of compressed tokens. Compression is performed on up to 3 segments of 2048 tokens. Ablations show that the different components of our finetuning strategy help boost performance and that stopgradients do not impa... | [{"Category": "Methodological Basis", "Citation": "(Vaswani et al., 2017)", "Explanation": "The cited work introduces the concept of language models, which serves as the basis for the research conducted in the citing paper on teaching pre-trained LMs to compress text into summary vectors."}, {"Category": "Supporting Ev... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b29",
"b12",
"b24",
"b50",
"b16",
"b38"
],
"table_ref": [],
"text": "Counterfactual generation, designed to eliminate spurious correlations in data, is a crucial technique used in causal ... | 2024-02-23 | 10.18653/v1/D15-1075 | [
{
"authors": "",
"journal": "Bibliographical References",
"ref_id": "b0",
"title": "",
"year": ""
},
{
"authors": "R Samuel; Gabor Bowman; Christopher Angeli; Christopher D Potts; Manning",
"journal": "Association for Computational Linguistics",
"ref_id": "b1",
"title": "A l... | [
{
"formula_coordinates": [
3,
86.28,
380.15,
204.59,
25.14
],
"formula_id": "formula_0",
"formula_text": "P (Y |X) = 1 ||XC ∪ XT || x i ∈(X C ∪X T ) P (Y |xi),(1)"
},
{
"formula_coordinates": [
3,
102.37,
417.78,
52.24,
9.65
],
... | Prompting Large Language Models for Counterfactual Generation: An Empirical Study | Large language models (LLMs) have made remarkable progress in a wide range of natural language understanding and generation tasks. However, their ability to generate counterfactuals has not been examined systematically. To bridge this gap, we present a comprehensive evaluation framework on various types of NLU tasks, w... | Yongqi Li; Mayi Xu; Xin Miao; Shen Zhou; Tieyun Qian | [
{
"figure_caption": "Figure 1 :1Figure 1: (a) Structural causal model of the SA task, (b) Intervention operation.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Left: The proposed framework for evaluati... | [{"Category": "Methodological Basis", "Citation": "(Pearl, 1993)", "Explanation": "The cited work by Pearl (1993) introduces the concept of causal intervention, which is a crucial technique for counterfactual generation. The citing paper adopts this technique to enhance the robustness and performance of neural network ... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b5",
"b29",
"b37",
"b29",
"b11",
"b14",
"b14"
],
"table_ref": [],
"text": "A wealth of information exists in the form of structured knowledge, such as movie information databases or... | 2023-07-11 | 10.1162/coli.07-034-R2 | [
{
"authors": "Oshin Agarwal; Mihir Kale; Heming Ge; Siamak Shakeri; Rami Al-Rfou",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "Machine translation aided bilingual data-to-text generation and semantic parsing",
"year": "2020"
},
{
"authors": "Ron Arts... | [
{
"formula_coordinates": [
4,
91.77,
421.03,
177.74,
16.59
],
"formula_id": "formula_0",
"formula_text": "L d ′ = -1 |d| |d| i=0 log p(d i |d 0 , ..., d i-1 , t)"
},
{
"formula_coordinates": [
4,
94.81,
555.7,
171.66,
16.59
],
... | Faithful Low-Resource Data-to-Text Generation through Cycle Training | Methods to generate text from structured data have advanced significantly in recent years, primarily due to fine-tuning of pre-trained language models on large datasets. However, such models can fail to produce output faithful to the input data, particularly on out-of-domain data. Sufficient annotated data is often not... | Zhuoer Wang; Marcus Collins; Nikhita Vedula; Simone Filice; Shervin Malmasi; Oleg Rokhlenko | [
{
"figure_caption": "Figure 1 :1Figure 1: Cycle Training of the Data-to-Text model and Text-to-Data model. For each cycle, the upper-level models are frozen to generate the intermediate text for the training of the lower-level models, that attempt to reconstruct the initial inputs (d, t denote initial inputs of... | [{"Category": "Data Source", "Citation": "WebNLG (Castro Ferreira et al., 2020)", "Explanation": "The cited work is a public dataset that the citing paper uses to tackle the data-to-text generation task for a variety of purposes."}, {"Category": "Data Source", "Citation": "ToTTo (Parikh et al., 2020)", "Explanation": "... |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b25",
"b35",
"b4",
"b20"
],
"table_ref": [],
"text": "As large language models (LLMs) are deployed widely, the need to keep their knowledge correct and up-to-date without massive r... | [
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell",
"journal": "",
"ref_id": "b0",
"title": "Language models are few-shot learners",
"year": "2020"
},
{
"authors": "Wei-Li... | [
{
"formula_coordinates": [
3,
306.14,
567.16,
220.08,
40
],
"formula_id": "formula_0",
"formula_text": "C = ⟨(s 1 , r 1 , o 1 ), . . . , (s n , r n , o n )⟩, we randomly sam- ple t ∈ {1, . . . , N } counterfactual edits in C."
},
{
"formula_coordinates": [
... | MQUAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questions | The information stored in large language models (LLMs) falls out of date quickly, and retraining from scratch is often not an option. This has recently given rise to a range of techniques for injecting new facts through updating model weights. Current evaluation paradigms are extremely limited, mainly validating the re... | Zexuan Zhong; Zhengxuan Wu; Christopher D Manning; Christopher Potts; Danqi Chen; Boris Johnson; Rishi Sunak | [
{
"figure_caption": "Figure 1 :1Figure 1: An example of our benchmark MQUAKE. Existing knowledge-editing methods often perform well at answering paraphrased questions of the edited fact but fail on multi-hop questions that are entailed consequences of the edited fact.",
"figure_data": "",
"figure_id": "... | [{"Category": "Supporting Evidence", "Citation": "(Sinitsin et al., 2020)", "Explanation": "The cited work highlights the importance of keeping language models up-to-date and correct, which is a key focus of the citing paper."}, {"Category": "Extension or Continuation", "Citation": "(Zhu et al., 2021;De Cao et al., 202... | ||
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b4",
"b5",
"b6",
"b13",
"b14",
"b15"
],
"table_ref": [],
"text": "In contemporary times, the Language Model (LM) [1; 2] has emerged as a pivotal player in the field of Natural Language Pr... | 2024-01-23 | [
{
"authors": "Jacob Devlin; Ming-Wei Chang; Kenton ; Lee Kristina; Toutanova ",
"journal": "",
"ref_id": "b0",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"year": "2019"
},
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbia... | [
{
"formula_coordinates": [
4,
108,
713.17,
66.91,
9.68
],
"formula_id": "formula_1",
"formula_text": "C 1 , ..., C N ∈ S."
}
] | Exploring Diverse In-Context Configurations for Image Captioning | After discovering that Language Models (LMs) can be good in-context few-shot learners, numerous strategies have been proposed to optimize in-context sequence configurations. Recently, researchers in Vision-Language (VL) domains also develop their few-shot learners, while they only use the simplest way, i.e. , randomly ... | Xu Yang; Yongliang Wu; Mingzhuo Yang; Haokun Chen; Xin Geng | [
{
"figure_caption": "Figure 2 :2Figure 2: Image selection strategies: (a) SIIR-CLIP, (b) SIIR-TAG, (c) DIIR-TT, (d) SICR-CLIP. 3 Configuring In-Context Sequences The in-context captioning can be treated as a vision-language conditioned text generation task. Given the multi-modal in-context sequence S = {(I 1 , ... | [{"Category": "Methodological Basis", "Citation": "[1; 2]", "Explanation": "The cited works are the Language Model (LM) that have emerged as a pivotal player in the field of Natural Language Processing (NLP), providing a unified approach to a range of diverse tasks by using a shared prompt paradigm."}, {"Category": "Ex... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b0"
],
"table_ref": [],
"text": "In-context learning (ICL) with large language models (LLMs) has shown great potential in performing a wide range of language tasks (Brown et al., 2020). ICL has the unique advantages o... | 2023-10-26 | [
{
"authors": "Tom Brown; Benjamin Mann; Nick Ryder; Melanie Subbiah; Jared D Kaplan; Prafulla Dhariwal; Arvind Neelakantan; Pranav Shyam; Girish Sastry; Amanda Askell; Sandhini Agarwal; Ariel Herbert-Voss; Gretchen Krueger; Tom Henighan; Rewon Child; Aditya Ramesh; Daniel Ziegler; Jeffrey Wu; Clemens Winter; Ch... | [
{
"formula_coordinates": [
3,
86.8,
368.05,
32.73,
16
],
"formula_id": "formula_0",
"formula_text": "(1) i , y(1)"
},
{
"formula_coordinates": [
3,
158.99,
368.05,
40.86,
16
],
"formula_id": "formula_1",
"formula_text": ... | Estimating Large Language Model Capabilities without Labeled Test Data | Large Language Models (LLMs) have the impressive ability to perform in-context learning (ICL) from only a few examples, but the success of ICL varies widely from task to task. Thus, it is important to quickly determine whether ICL is applicable to a new task, but directly evaluating ICL accuracy can be expensive in sit... | Harvey Yiyun Fu; Qinyuan Ye; Albert Xu; Xiang Ren; Robin Jia | [
{
"figure_caption": "Figure 2 :2Figure2: Bar graph of evaluation results (MAE) for all meta-models, baseline methods, and Oracle baselines of all 3 dataset collections with all 4 LLMs. We use the confidence vector as the meta-feature. Red/blue bars represent the meta-model/baseline evaluation results and the ho... | [{"Category": "Supporting Evidence", "Citation": "(Brown et al., 2020)", "Explanation": "The cited work by Brown et al. (2020) provides foundational evidence of the potential of in-context learning (ICL) with large language models (LLMs) in performing a wide range of language tasks, which supports the claims and hypoth... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b16",
"b27",
"b5",
"b10",
"b7",
"b0",
"b25",
"b23",
"b36",
"b24",
"b10",
"b40",
"b3",
"b5",
"b13",
"b23",
"b18",
"b38",
"... | 10.18653/v1/2022.acl-long.589 | [
{
"authors": "Iz Beltagy; Matthew E Peters; Arman Cohan",
"journal": "",
"ref_id": "b0",
"title": "Longformer: The long-document transformer",
"year": "2020"
},
{
"authors": "Amanda Bertsch; Uri Alon; Graham Neubig; Matthew R Gormley",
"journal": "",
"ref_id": "b1",
"title": ... | [
{
"formula_coordinates": [
2,
315.64,
75.72,
199.27,
71.08
],
"formula_id": "formula_0",
"formula_text": "Approach In→Out Enc Enc←Dec Dec Efficient Attention x → y ■ ■ Extract-Abstract xe → y □ ⋆ Dynamic Weight x → y □ ■ + ⋆ Divide-Conquer xi → yi □ □"
},
{
"for... | AWESOME: GPU Memory-constrained Long Document Summarization using Memory Mechanism and Global Salient Content | Long document summarization systems are critical for domains with lengthy and jargonladen text, yet they present significant challenges to researchers and developers with limited computing resources. Existing solutions mainly focus on efficient attentions or divideand-conquer strategies. The former reduces theoretical ... | Shuyang Cao; Lu Wang | [
{
"figure_caption": "Model",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: Running time (batch per second) of each model. A higher number of batches processed per second indicates a faster running speed. ... | [{"Category": "Methodological Basis", "Citation": "(Lewis et al., 2020)", "Explanation": "The cited work by Lewis et al. provides the performance benchmarks for large pre-trained transformer models in abstractive summarization, which the citing paper uses to evaluate the performance of their own model."}, {"Category": ... | |
[
{
"figure_ref": [],
"heading": "INTRODUCTION",
"publication_ref": [
"b16",
"b17",
"b2",
"b10"
],
"table_ref": [],
"text": "In search engines, the queries issued by users are mostly broad and vague [15,16]. This problem is extremely crucial for product search scenarios... | 2023-05-24 | 10.1145/3539618.3591900 | [
{
"authors": "Qingyao Ai; Daniel N Hill; S V N Vishwanathan; W Bruce Croft",
"journal": "ACM",
"ref_id": "b0",
"title": "A Zero Attention Model for Personalized Product Search",
"year": "2019-11-03"
},
{
"authors": "Qingyao Ai; Lakshmi Narayanan; Ramasamy ",
"journal": "ACM",
"re... | [
{
"formula_coordinates": [
5,
371.15,
416,
187.59,
21.99
],
"formula_id": "formula_0",
"formula_text": "IE(𝑞) = ∑︁ 𝑝 ∈ I (𝑞) -𝑃 (𝑝 |𝑞) log 2 𝑃 (𝑝 |𝑞)(1)"
},
{
"formula_coordinates": [
5,
366.14,
443.22,
189.43,
21.78
],... | JDsearch: A Personalized Product Search Dataset with Real Queries and Full Interactions | Recently, personalized product search attracts great attention and many models have been proposed. To evaluate the effectiveness of these models, previous studies mainly utilize the simulated Amazon recommendation dataset, which contains automatically generated queries and excludes cold users and tail products. We argu... | Jiongnan Liu; Zhicheng Dou; Guoyu Tang; Sulong Xu | [
{
"figure_caption": "4 https://github.com/rucliujn/JDsearch",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
},
{
"figure_caption": "Figure 1 :1Figure 1: The log-log distribution of product's interaction frequency",
"figure_data": "",
"figure... | [{"Category": "Data Source", "Citation": "[1-6, 9, 11]", "Explanation": "The cited works are the most widely-used personalized product search datasets based on real user behaviors, which serve as the foundational data for the research conducted in the citing paper on personalized product search."}, {"Category": "Data S... |
[
{
"figure_ref": [
"fig_0",
"fig_0",
"fig_0",
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b40",
"b50",
"b38",
"b66",
"b5",
"b57",
"b42",
"b90",
"b62",
"b63",
"b0",
"b73",
"b81",
... | 2023-05-24 | [
{
"authors": "Eric Arazo",
"journal": "",
"ref_id": "b0",
"title": "Pseudo-labeling and confirmation bias in deep semi-supervised learning",
"year": "2020"
},
{
"authors": "Philip Bachman; Ouais Alsharif; Doina Precup",
"journal": "NeurIPS",
"ref_id": "b1",
"title": "Learning... | [
{
"formula_coordinates": [
5,
42.11,
408.78,
238.07,
31.61
],
"formula_id": "formula_0",
"formula_text": "ℓ labeled cls = (x,y * )∼D l K k=1 Cls(y * , p k (y|x, b k-1 )),(1)"
},
{
"formula_coordinates": [
5,
42.11,
446.09,
233.83,
3... | Semi-Supervised and Long-Tailed Object Detection with CascadeMatch | This paper focuses on long-tailed object detection in the semi-supervised learning setting, which poses realistic challenges, but has rarely been studied in the literature. We propose a novel pseudolabeling-based detector called CascadeMatch. Our detector features a cascade network architecture, which has multi-stage d... | Yuhang Zang; Kaiyang Zhou; Chen Huang; Chen Change Loy | [
{
"figure_caption": "Fig. 1 :1Fig. 1: Motivation of our research. (a) The Average Precision (AP) and Average Recall (AR) curves, obtained using different fixed confidence thresholds (denoted by τ ). Clearly, none of the chosen thresholds gives the best trade-off. (b) The distribution of prediction scores for a ... | [{"Category": "Supporting Evidence", "Citation": "[40]", "Explanation": "The cited work, the COCO dataset, is used as a benchmark for evaluating the performance of semi-supervised object detectors in the citing paper."}, {"Category": "Supporting Evidence", "Citation": "[16]", "Explanation": "The cited work, LVIS, is a ... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b26",
"b30",
"b11",
"b49",
"b3",
"b50",
"b8"
],
"table_ref": [],
"text": "A core challenge in deploying NLP systems lies in managing temporal misalignment, where a model that is tra... | 10.18653/v1/2020.emnlp-main.550 | [
{
"authors": "Stephen Anastasios N Angelopoulos; Adam Bates; Lihua Fisch; Tal Lei; Schuster",
"journal": "",
"ref_id": "b0",
"title": "Conformal risk control",
"year": "2022"
},
{
"authors": "Hessam Bagherinezhad; Hannaneh Hajishirzi; Yejin Choi; Ali Farhadi",
"journal": "",
"re... | [] | Mitigating Temporal Misalignment by Discarding Outdated Facts | While large language models are able to retain vast amounts of world knowledge seen during pretraining, such knowledge is prone to going out of date and is nontrivial to update. Furthermore, these models are often used under temporal misalignment, tasked with answering questions about the present, despite having only b... | Michael J Q Zhang; Eunsol Choi | [
{
"figure_caption": "FactFigure 1 :1Figure1: We depict the critical timestamps at play in open-retrieval QA systems. In the example on the left, the temporal misalignment between when the system was trained and evaluated has no affect on the answer. On the right, the answer has changed, causing the system to ou... | [{"Category": "Methodological Basis", "Citation": "(Lazaridou et al., 2021)", "Explanation": "The cited work highlights the issue of temporal misalignment in deploying NLP systems, which the citing paper addresses by proposing methods to manage the misalignment."}, {"Category": "Supporting Evidence", "Citation": "(Luu ... | |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b10",
"b1",
"b18",
"b26",
"b11",
"b14",
"b22",
"b6",
"b3",
"b17",
"b25",
"b6",
"b17",
"b22",
"b10",
"b3",
"b17",
... | 2023-11-14 | 10.18653/v1/2020.nlp4convai-1.5 | [
{
"authors": "",
"journal": "L-BERT TAPT",
"ref_id": "b0",
"title": "",
"year": ""
},
{
"authors": "References Pavel Burnyshev; A Bout; Anfrey Bout; Valentin Malykh; Irina Piontkovskaya",
"journal": "",
"ref_id": "b1",
"title": "Infobert: Zeroshot approach to natural language... | [
{
"formula_coordinates": [
2,
360.78,
345.38,
164.36,
30.47
],
"formula_id": "formula_0",
"formula_text": "c n = 1 K x n,i ∈Sn f ϕ (x n,i )(1)"
},
{
"formula_coordinates": [
2,
313.41,
625.12,
211.73,
30.47
],
"formula_id": ... | Pre-training Intent-Aware Encoders for Zero-and Few-Shot Intent Classification | Intent classification (IC) plays an important role in task-oriented dialogue systems. However, IC models often generalize poorly when training without sufficient annotated examples for each user intent. We propose a novel pretraining method for text encoders that uses contrastive learning with intent psuedo-labels to p... | Mujeen Sung; James Gung; Elman Mansimov; Nikolaos Pappas; Raphael Shu; Salvatore Romeo; Yi Zhang; Vittorio Castelli | [
{
"figure_caption": "Figure 1 :1Figure 1: Overview of pre-training the intent-aware encoder (PIE). Given an utterance x 1 from pre-training corpus, we generate a pseudo intent name y pseudo",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figu... | [{"Category": "Methodological Basis", "Citation": "(Liu et al., 2019a)", "Explanation": "The cited work by Liu et al. provides a method for few-shot text classification that the citing paper adopts to tackle the challenge of data collection and re-training models in task-oriented dialogue systems."}, {"Category": "Meth... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b10",
"b7",
"b15",
"b17",
"b14",
"b12",
"b0",
"b7",
"b15",
"b17",
"b14",
"b12",
"b0",
"b17",
"b5",
"b7"
],
"table_ref": [],
"te... | [
{
"authors": "Anpei Chen; Zexiang Xu; Andreas Geiger; Jingyi Yu; Hao Su",
"journal": "",
"ref_id": "b0",
"title": "Tensorf: Tensorial radiance fields",
"year": "2022"
},
{
"authors": "Anpei Chen; Zexiang Xu; Fuqiang Zhao; Xiaoshuai Zhang; Fanbo Xiang; Jingyi Yu; Hao Su",
"journal": "... | [
{
"formula_coordinates": [
3,
183.78,
556.48,
320.89,
30.32
],
"formula_id": "formula_0",
"formula_text": "Ĉ(r) = N -1 i=0 T i (1 -exp(-σ i δ i ))c i , T i = exp(- i-1 j=0 σ j δ j ).(1)"
},
{
"formula_coordinates": [
3,
108,
624.92,
96.94... | OD-NeRF: Efficient Training of On-the-Fly Dynamic Neural Radiance Fields | Dynamic neural radiance fields (dynamic NeRFs) have demonstrated impressive results in novel view synthesis on 3D dynamic scenes. However, they often require complete video sequences for training followed by novel view synthesis, which is similar to playing back the recording of a dynamic 3D scene. In contrast, we prop... | Zhiwen Yan; Chen Li; Gim Hee Lee | [
{
"figure_caption": "Figure 1 :1Figure 1: We introduce the on-the-fly training(left) of dynamic NeRFs and the OD-NeRF model(right). In on-the-fly training, the dynamic NeRF is trained based on the current and previous training frames to synthesize novel views for the current time step. Our OD-NeRF leverages the... | [{"Category": "Methodological Basis", "Citation": "[11]", "Explanation": "The cited work introduces the concept of neural radiance fields (NeRF), which serves as the basis for the research conducted in the citing paper to develop a new 3D volumetric representation capable of synthesizing photo-realistic novel views fro... | ||
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b4",
"b22",
"b18",
"b14",
"b36",
"b5",
"b13",
"b8",
"b17"
],
"table_ref": [],
"text": "Document summarization aims to compress text material while retaining its most sal... | 2023-10-09 | 10.18653/v1/D19-1435 | [
{
"authors": "Afra Feyza Akyürek; Ekin Akyürek; Aman Madaan; Ashwin Kalyan; Peter Clark; Derry Wijaya; Niket Tandon",
"journal": "",
"ref_id": "b0",
"title": "Rl4f: Generating natural language feedback with reinforcement learning for repairing model outputs",
"year": "2023"
},
{
"authors... | [
{
"formula_coordinates": [
4,
103.17,
243.19,
186.69,
33.58
],
"formula_id": "formula_0",
"formula_text": "p S (y 0 | x) = m t=1 p S y 0 t | y 0 <t , x ,(1)"
},
{
"formula_coordinates": [
4,
306.14,
244.68,
62.6,
10.72
],
"f... | SummIt: Iterative Text Summarization via ChatGPT | Text summarization systems have made significant progress in recent years, but typically generate summaries in one single step. However, the one-shot summarization setting is sometimes inadequate, as the generated summary may contain hallucinations or overlook essential details related to the reader's interests. This p... | Haopeng Zhang; Xiao Liu; Jiawei Zhang | [
{
"figure_caption": "Figure 1 :1Figure 1: An illustration of the iterative summarization process. The summarizer continuously refines the summary according to self-feedback from the evaluator at each iteration.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figur... | [{"Category": "Methodological Basis", "Citation": "(Cheng and Lapata, 2016)", "Explanation": "The cited work by Cheng and Lapata (2016) provides a methodological basis for the development of end-to-end summarization systems by introducing the use of neural networks and pre-trained language models in summarization syste... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b25",
"b4",
"b1",
"b10",
"b15",
"b2",
"b3",
"b20",
"b14",
"b36",
"b0",
"b37"
],
"table_ref": [],
"text": "Autonomous driving is a rapidly developing fi... | [
{
"authors": "P Anderson; X He; C Buehler; D Teney; M Johnson; S Gould; L Zhang",
"journal": "",
"ref_id": "b0",
"title": "Bottom-up and top-down attention for image captioning and visual question answering",
"year": "2018"
},
{
"authors": "S Antol; A Agrawal; J Lu; M Mitchell; D Batra; ... | [
{
"formula_coordinates": [
3,
352.97,
376.64,
205.03,
23.23
],
"formula_id": "formula_0",
"formula_text": "θ = cos -1 (B 1 [: 2] -B 2 [: 2]) • V ego [: 2] ∥B 1 [: 2] -B 2 [: 2]∥∥V ego [: 2]∥ ,(1)"
},
{
"formula_coordinates": [
3,
420.54,
418.37... | NuScenes-QA: A Multi-Modal Visual Question Answering Benchmark for Autonomous Driving Scenario | We introduce a novel visual question answering (VQA) task in the context of autonomous driving, aiming to answer natural language questions based on street-view clues. Compared to traditional VQA tasks, VQA in autonomous driving scenario presents more challenges. Firstly, the raw visual data are multi-modal, including ... | Tianwen Qian; Jingjing Chen; Linhai Zhuo; Yang Jiao; Yu-Gang Jiang | [
{
"figure_caption": "Figure 2 :2Figure 2: Data construction flow of NuScenes-QA. First, the scene graphs are generated using the annotated object labels and 3D bounding boxes. Then, we design question templates manually, and instantiate the question-answer pairs with them. Finally, the generated data are filter... | [{"Category": "Data Source", "Citation": "(Liu et al. 2023)", "Explanation": "The cited work by Liu et al. is a source of data for the autonomous driving system, providing information on 3D object detection that is used in the system."}, {"Category": "Data Source", "Citation": "(Jiao et al. 2023b)", "Explanation": "The... | ||
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b31",
"b8",
"b4",
"b18",
"b32",
"b34",
"b33",
"b30",
"b7",
"b3",
"b36",
"b5",
"b3",
"b15",
"b3",
"b26",
"b14",
"b37",
"b2... | 2023-05-24 | [
{
"authors": "",
"journal": "FLOPs ↓ Throughput ↑ ViT-B",
"ref_id": "b0",
"title": "Model Efficiency Top1 Acc.(%) ↑ Resolution #Params",
"year": ""
},
{
"authors": "",
"journal": "",
"ref_id": "b1",
"title": "Comparison with baselines. Except for E-ViT, which undergoes traini... | [
{
"formula_coordinates": [
4,
101.83,
425.68,
184.54,
11.72
],
"formula_id": "formula_0",
"formula_text": "X = x i ∈ R d |i = 1, 2. . . , N ,(1)"
},
{
"formula_coordinates": [
4,
90.03,
489.24,
196.34,
9.65
],
"formula_id": ... | Predicting Token Impact Towards Efficient Vision Transformer | Token filtering to reduce irrelevant tokens prior to selfattention is a straightforward way to enable efficient vision Transformer. This is the first work to view token filtering from a feature selection perspective, where we weigh the importance of a token according to how much it can change the loss once masked. If t... | Hong Wang; Su Yang; Xiaoke Huang; Weishan Zhang | [
{
"figure_caption": "Figure 1 .1Figure 1. The overall training process: The two branches of the vision Transformer are in fact the same one, whose parameters are fixed during training. The delta loss and ρ refer to Eq. (7) and Eq. (8).",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
... | [{"Category": "Methodological Basis", "Citation": "[23]", "Explanation": "The cited work, Swin Transformer, is a method that the citing paper adopts to reduce the computational load in vision Transformer models by confining self-attention to a local neighborhood."}, {"Category": "Methodological Basis", "Citation": "[33... | |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b32",
"b40",
"b28",
"b38",
"b1",
"b15",
"b29",
"b39",
"b16",
"b0",
"b13",
"b5",
"b30",
"b5",
"b17",
"b17"
],
"ta... | 2023-10-24 | 10.18653/v1/D17-1303 | [
{
"authors": "Emanuele Bugliarello; Fangyu Liu; Jonas Pfeiffer; Siva Reddy; Desmond Elliott; Edoardo ; Maria Ponti; Ivan Vulić",
"journal": "",
"ref_id": "b0",
"title": "Iglue: A benchmark for transfer learning across modalities, tasks, and languages",
"year": "2022"
},
{
"authors": "Yen... | [
{
"formula_coordinates": [
3,
306.14,
500.51,
83.33,
10.69
],
"formula_id": "formula_0",
"formula_text": "T = {T 1 , ..., T N },"
},
{
"formula_coordinates": [
3,
386.27,
606.41,
138.87,
33.18
],
"formula_id": "formula_1",
... | Meta-learning For Vision-and-language Cross-lingual Transfer | Current pre-trained vision-language models (PVLMs) achieve excellent performance on a range of multi-modal datasets. Recent work aims at building multilingual versions of such models, and a range of multilingual multimodal datasets have been introduced for this purpose. However, current PVLMs typically perform poorly o... | Hanxu Hu; Frank Keller | [
{
"figure_caption": "Figure 1 :1Figure 1: Examples in IGLUE (Bugliarello et al., 2022) benchmark. The left example comes from MaRVL (Liu et al., 2021) dataset, and the right example comes from XVNLI dataset proposed in IGLUE.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figur... | [{"Category": "Methodological Basis", "Citation": "(Chen et al., 2020)", "Explanation": "The cited work by Chen et al. provides a pre-trained vision-language model that the citing paper adopts in their research to perform multi-modal tasks."}, {"Category": "Data Source", "Citation": "(Lu et al., 2019)", "Explanation": ... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b4",
"b5",
"b33",
"b16",
"b28",
"b27",
"b16",
"b26"
],
"table_ref": [],
"text": "Predicting and modeling sequences of events has become more sophisticated over the past decade... | 2023-05-24 | 10.18653/v1/P19-1470 | [
{
"authors": "Antoine Bosselut; Hannah Rashkin; Maarten Sap; Chaitanya Malaviya; Asli Celikyilmaz; Yejin Choi",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "COMET: Commonsense transformers for automatic knowledge graph construction",
"year": "2019"
},
{
... | [
{
"formula_coordinates": [
3,
323.24,
620.91,
201.9,
21.12
],
"formula_id": "formula_0",
"formula_text": "r(s, Ŝ) = max ŝ∈ Ŝ (max(E(s, ŝ), E(ŝ, s)))(1)"
},
{
"formula_coordinates": [
5,
334.93,
699.92,
50.39,
9.57
],
"formul... | Drafting Event Schemas using Language Models | Past work has studied event prediction and event language modeling, sometimes mediated through structured representations of knowledge in the form of event schemas. Such schemas can lead to explainable predictions and forecasting of unseen events given incomplete information. In this work, we look at the process of cre... | Anisha Gunjal; Greg Durrett | [
{
"figure_caption": "\"List causes and events that can happen over the course of a [d]? Causes of a [d]: 1.\" Causes and temporally-aided prompt \"List causes and events that can happen before, during and after a [d]? Causes of a [d]: 1.\" One-Shot Prompts Figure 2 depicts a sample of the one-shot prompts we us... | [{"Category": "Methodological Basis", "Citation": "(Chambers and Jurafsky, 2008)", "Explanation": "The cited work by Chambers and Jurafsky in 2008 introduced the concept of mining narrative schemas using predicate-role pairs, which has been adopted in later works for event prediction and cloze tasks."}, {"Category": "M... |
[
{
"figure_ref": [],
"heading": "I. INTRODUCTION",
"publication_ref": [
"b0",
"b1",
"b2",
"b4",
"b5",
"b6",
"b7",
"b10",
"b11",
"b12",
"b13",
"b14",
"b15",
"b16",
"b6",
"b17",
"b18",
"b19",
"... | 2023-05-24 | [
{
"authors": "M Wang; W Deng",
"journal": "Neurocomputing",
"ref_id": "b0",
"title": "Deep Face Recognition: A Survey",
"year": "2021"
},
{
"authors": "P Li; L Prieto; D Mery; P J Flynn",
"journal": "IEEE Transactions on Information Forensics and Security",
"ref_id": "b1",
"t... | [
{
"formula_coordinates": [
2,
311.98,
617,
71.2,
10.31
],
"formula_id": "formula_0",
"formula_text": "Q * = H(M (I))"
},
{
"formula_coordinates": [
3,
169.29,
490.68,
119.19,
13.38
],
"formula_id": "formula_1",
"formula_... | Optimization-Based Improvement of Face Image Quality Assessment Techniques | Contemporary face recognition (FR) models achieve near-ideal recognition performance in constrained settings, yet do not fully translate the performance to unconstrained (realworld) scenarios. To help improve the performance and stability of FR systems in such unconstrained settings, face image quality assessment (FIQA... | Žiga Babnik; Naser Damer; Vitomir Štruc | [
{
"figure_caption": "Fig. 1 .1Fig.1. Overview of the proposed method that consists of: Label Optimization and Transfer Learning. The label-optimization step incorporates information extracted from mated image pairs into quality scores precomputed with an existing FIQA technique. The transfer-learning step is th... | [{"Category": "Supporting Evidence", "Citation": "[1], [2]", "Explanation": "The cited works highlight the challenges of out-of-distribution data in real-world scenarios, which motivates the need for face image quality assessment techniques in the citing paper."}, {"Category": "Methodological Basis", "Citation": "[3]- ... | |
[
{
"figure_ref": [
"fig_0",
"fig_0",
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b12",
"b34",
"b24",
"b11",
"b54",
"b42",
"b42",
"b43"
],
"table_ref": [],
"text": "Recent advances in NLP primarily focus on t... | 10.1162/tacl_a_00416 | [
{
"authors": "David Ifeoluwa Adelani; Jade Abbott; Graham Neubig; D' Daniel; Julia Souza; Constantine Kreutzer; Chester Lignos; Happy Palen-Michel; Shruti Buzaaba; Sebastian Rijhwani; Stephen Ruder; Israel Mayhew; Shamsuddeen H Abebe Azime; Chris Chinenye Muhammad; Joyce Emezue; Perez Nakatumba-Nabende; Aremu O... | [
{
"formula_coordinates": [
2,
315.93,
91.73,
189.85,
58.29
],
"formula_id": "formula_0",
"formula_text": "XTREME ✓ XTREME-R ✓ XGLUE ✓ ✓ CrossFit ✓ ✓ MEGA* ✓ ✓ BUFFET ✓ ✓ ✓ ✓"
},
{
"formula_coordinates": [
19,
214.94,
452.98,
226.77,
... | BUFFET: Benchmarking Large Language Models for Few-shot Cross-lingual Transfer | Despite remarkable advancements in few-shot generalization in natural language processing, most models are developed and evaluated primarily in English. To facilitate research on few-shot cross-lingual transfer, we introduce a new benchmark, called BUFFET, which unifies 15 diverse tasks across 54 languages in a sequenc... | Akari Asai; Sneha Kudugunta; Xinyan Velocity; Terra Blevins; Hila Gonen; Machel Reid; Yulia Tsvetkov; Sebastian Ruder; Hannaneh Hajishirzi; Victor Sanh; Albert Webson; Colin Raffel; Stephen Bach; Lintang Sutawika; Zaid Alyafeai; Antoine Chaffin; Arnaud Stiegler; Arun Raja; Manan Dey; Saiful Bari; Canwen Xu; Urmish Thak... | [
{
"figure_caption": "Figure 1 :1Figure 1: BUFFET includes unified diverse tasks in the same format, covering many typologically diverse languages. It enables a fair comparison across models, transfer methods, and languages and facilitates largescale analysis across different learning setups.",
"figure_data"... | [{"Category": "Supporting Evidence", "Citation": "(Blasi et al., 2022)", "Explanation": "The cited work by Blasi et al. (2022) highlights the focus of recent advances in NLP on the English language, which serves as a foundational point for the citing paper to build upon."}, {"Category": "Supporting Evidence", "Citation... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b37",
"b17",
"b32",
"b4",
"b61",
"b42",
"b54",
"b44",
"b57",
"b31"
],
"table_ref": [],
"text": "Large language models (LLMs) have evolved considerably in size, arc... | 2023-11-19 | 10.18653/v1/2021.emnlp-main.98 | [
{
"authors": "Jimmy Ba; Jamie Ryan Kiros; Geoffrey E Hinton",
"journal": "",
"ref_id": "b0",
"title": "Layer normalization",
"year": "2016"
},
{
"authors": "Yoshua Bengio; Réjean Ducharme; Pascal Vincent; Christian Janvin",
"journal": "J. Mach. Learn. Res",
"ref_id": "b1",
"t... | [
{
"formula_coordinates": [
3,
105.69,
611.93,
184.17,
33.71
],
"formula_id": "formula_0",
"formula_text": "p(X) = N i=1 p (x i |x 1 , x 2 , ..., x i-1 )(1)"
},
{
"formula_coordinates": [
3,
310.71,
97.68,
214.43,
44.01
],
"f... | How To Train Your (Compressed) Large Language Model | With the increase in the size of large language models (LLMs), we need compression methods that can reduce the model size while preserving the generality and zero-shot promptability of the model. This goal is more ambitious than the typical compression setup, which reduces the model's size at the expense of specializin... | Ananya Harsh Jha; Tom Sherborne; Evan Pete Walsh; Dirk Groeneveld; Emma Strubell; Iz Beltagy | [
{
"figure_caption": "Figure 1 :1Figure 1: Truncated initialization configurations for layer pruning in a decoder-only language model. Highlighted layers (green) are removed. Our method and distillation baselines remove half of the layers according to each configuration. We retain the first and last layer as the... | [{"Category": "Methodological Basis", "Citation": "(Brown et al., 2020)", "Explanation": "The cited work by Brown et al. (2020) introduces the GPT-3 model, which is a key methodological basis for the citing paper in terms of the development of large language models and the use of in-context learning for task-specific f... |
[
{
"figure_ref": [
"fig_0",
"fig_1"
],
"heading": "Introduction",
"publication_ref": [
"b15",
"b12",
"b42",
"b44",
"b41",
"b87",
"b75",
"b7",
"b65",
"b37",
"b27",
"b30",
"b42",
"b4",
"b34",
"b2... | 10.48550/arXiv.2305.19068 | [
{
"authors": "Jiaxin Bai; Xin Liu; Weiqi Wang; Chen Luo; Yangqiu Song",
"journal": "",
"ref_id": "b0",
"title": "Complex query answering on eventuality knowledge graph with implicit logical constraints",
"year": "2023"
},
{
"authors": "Pratyay Banerjee; Chitta Baral",
"journal": "Ass... | [
{
"formula_coordinates": [
3,
306.14,
387.35,
168.84,
9.57
],
"formula_id": "formula_0",
"formula_text": "D = {(h, r, t)|h ∈ H, r ∈ R, t ∈ T }"
},
{
"formula_coordinates": [
3,
306.14,
439.6,
218.27,
25.07
],
"formula_id": "... | CAR: Conceptualization-Augmented Reasoner for Zero-Shot Commonsense Question Answering | The task of zero-shot commonsense question answering evaluates models on their capacity to reason about general scenarios beyond those presented in specific datasets. Existing approaches for tackling this task leverage external knowledge from CommonSense Knowledge Bases (CSKBs) by pre-training the model on synthetic QA... | Weiqi Wang; Tianqing Fang; Wenxuan Ding; Baixuan Xu; Xin Liu; Yangqiu Song; Antoine Bosselut | [
{
"figure_caption": "Figure 1 :1Figure 1: An example of constructing synthetic QA pairs from CSKB(Ma et al., 2021). The simple heuristic used in this process can result in false negative options.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
... | [{"Category": "Supporting Evidence", "Citation": "(Ma et al., 2021)", "Explanation": "The cited work by Ma et al. provides evidence that pre-trained language models struggle to generalize to distributionally different examples from their training sets, which supports the claim made in the citing paper about the limitat... | |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b46",
"b66",
"b35",
"b33",
"b37",
"b61",
"b54",
"b4",
"b45",
"b44",
"b43",
"b17",
"b9"
],
"table_ref": [],
"text": "Text clust... | 2023-11-03 | 10.18653/v1/2020.acl-main.372 | [
{
"authors": "Wenbin An; Feng Tian; Qinghua Zheng; Wei Ding; Qianying Wang; Ping Chen",
"journal": "",
"ref_id": "b0",
"title": "Generalized category discovery with decoupled prototypical network",
"year": "2022"
},
{
"authors": "Yuntao Bai; Saurav Kadavath; Sandipan Kundu; Amanda Askell... | [
{
"formula_coordinates": [
3,
144.66,
439.08,
145.21,
10.69
],
"formula_id": "formula_0",
"formula_text": "c j = P T (I T , t),(1)"
},
{
"formula_coordinates": [
3,
70.87,
622.74,
109.26,
14
],
"formula_id": "formula_1",
... | CLUSTERLLM: Large Language Models as a Guide for Text Clustering | We introduce CLUSTERLLM, a novel text clustering framework that leverages feedback from an instruction-tuned large language model, such as ChatGPT. Compared with traditional unsupervised methods that builds upon "small" embedders, CLUSTERLLM exhibits two intriguing advantages: (1) it enjoys the emergent capability of L... | Yuwei Zhang; Zihan Wang; Jingbo Shang | [
{
"figure_caption": "Figure 1 :1Figure 1: LLMs like ChatGPT are not applicable for text clustering directly because of the inaccessible embeddings. CLUSTERLLM resolves the dilemma by leveraging LLM as a guide on text clustering.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"fi... | [{"Category": "Methodological Basis", "Citation": "(MacQueen, 1967)", "Explanation": "The cited work by MacQueen (1967) introduces a clustering algorithm that is adopted in the citing paper to perform text clustering on pre-trained embedders."}, {"Category": "Data Source", "Citation": "(Muennighoff et al., 2022)", "Exp... |
[
{
"figure_ref": [
"fig_1",
"fig_1",
"fig_1"
],
"heading": "Introduction",
"publication_ref": [
"b32",
"b43",
"b10"
],
"table_ref": [],
"text": "Large Language Models (LLMs) have demonstrated remarkable performance in solving various natural language pr... | [
{
"authors": "H Stephen; Victor Bach; Zheng-Xin Sanh; Albert Yong; Colin Webson; Raffel; V Nihal; Abheesht Nayak; Taewoon Sharma; M Kim; Thibault Saiful Bari; Zaid Fevry; Manan Alyafeai; Andrea Dey; Zhiqing Santilli; Srulik Sun; Canwen Ben-David; Gunjan Xu; Han Chhablani; Jason Wang; Alan Fries; Maged S Al-Sha... | [
{
"formula_coordinates": [
3,
96.96,
175.77,
390.88,
54.56
],
"formula_id": "formula_0",
"formula_text": "Log-probability Mean ZLP s(x, y) = 1 |T | t log p(y|x, t) Probability Mean ZPM s(x, y) = 1 |T | t p(y|x, t) Majority Vote ZMV s(x, y) = t 1{arg max v∈Y p(y|x, t) = ... | Improving Probability-based Prompt Selection Through Unified Evaluation and Analysis | Previous works in prompt engineering for large language models have introduced different gradient-free probability-based prompt selection methods that aim to choose the optimal prompt among the candidates for a given task but have failed to provide a comprehensive and fair comparison between each other. In this paper, ... | Sohee Yang; Jonghyeon Kim; Joel Jang; Seonghyeon Ye; Hyunji Lee; Minjoon Seo; Yizhong Wang; Yeganeh Kordi; Swaroop Mishra; Alisa Liu; Noah A Smith; Daniel Khashabi; Hannaneh 2023 Hajishirzi; Bigscience Workshop; Teven Le Scao; Angela Fan; Christopher Akiki; Ellie Pavlick; Suzana Ilić; Daniel Hesslow; Roman Castagné; Sa... | [
{
"figure_caption": "Figure 1 :1Figure 1: (a) F1 of the prompts selected by different probability-based prompt selection methods, averaged across 13 datasets. Per-dataset F1 and accuracy are shown in Figure 9. The methods without super/subscripts are the existing methods (Table1), while those with super/subscri... | [{"Category": "Methodological Basis", "Citation": "(Sorensen et al., 2022)", "Explanation": "The cited work provides the foundational concept of Mutual Information (MI) that the citing paper utilizes in the development of prompt selection methods for large language models (LLMs). The discovery in the cited work is used... | ||
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b7",
"b8",
"b9",
"b10",
"b11",
"b12",
"b8",
"b13",
"b14",
"b15",
"... | 2023-05-24 | [
{
"authors": "J Liu; F Guo; Y Zhang; B Hou; H Zhou",
"journal": "Applied Intelligence",
"ref_id": "b0",
"title": "Defect classification on limited labeled samples with multiscale feature fusion and semi-supervised learning",
"year": "2021"
},
{
"authors": "J Wu; J Le; Z Xiao; F Zhang; L ... | [
{
"formula_coordinates": [
8,
56.88,
173.63,
331.5,
25.84
],
"formula_id": "formula_0",
"formula_text": "F G (y k ) = F 1 G (y k ) , F 2 G (y k ) , ..., F L G (y k ) and F T (y k ) = F 1 T (y k ) , F 2 T (y k ) , ."
},
{
"formula_coordinates": [
8,
5... | A T E X template Multiresolution Feature Guidance Based Transformer for Anomaly Detection | Anomaly detection is represented as an unsupervised learning to identify deviated images from normal images. In general, there are two main challenges of anomaly detection tasks, i.e., the class imbalance and the unexpectedness of anomalies. In this paper, we propose a multiresolution feature guidance method based on T... | Shuting Yan; Pingping Chen; Honghui Chen; Huan Mao; Feng Chen; Zhijian Lin | [
{
"figure_caption": "Fig. 11Fig. 1 Visual results from the MVTec AD datasets. Superimposed on the images are the anomaly localization map from GTrans. Red areas correspond to the located anomalies, whereas the blue areas indicate the normality regions.",
"figure_data": "",
"figure_id": "fig_0",
"fig... | [{"Category": "Supporting Evidence", "Citation": "[1][2][3]", "Explanation": "The cited works provide evidence of the use of industrial inspection techniques in anomaly detection, which supports the claim that anomaly detection techniques have been extensively studied in various research and application domains."}, {"C... | |
[
{
"figure_ref": [
"fig_0"
],
"heading": "",
"publication_ref": [
"b0",
"b0",
"b1",
"b2",
"b3",
"b4",
"b5",
"b6",
"b5",
"b7",
"b8",
"b8",
"b9",
"b10",
"b0",
"b11",
"b8",
"b9",
"b12"... | 2023-05-24 | 10.1007/sxxxxx-yyy-zzzz-1 | [
{
"authors": "L Wu; X He; X Wang; K Zhang; M Wang",
"journal": "IEEE Transactions on Knowledge and Data Engineering",
"ref_id": "b0",
"title": "A survey on accuracy-oriented neural recommendation: From collaborative filtering to information-rich recommendation",
"year": "2022"
},
{
"auth... | [
{
"formula_coordinates": [
3,
310.7,
161.7,
13.05,
12.7
],
"formula_id": "formula_0",
"formula_text": "C (l) p"
},
{
"formula_coordinates": [
3,
310.48,
169.76,
202.57,
58.99
],
"formula_id": "formula_1",
"formula_text":... | How Graph Convolutions Amplify Popularity Bias for Recommendation? | Graph convolutional networks (GCNs) have become prevalent in recommender system (RS) due to their superiority in modeling collaborative patterns. Although improving the overall accuracy, GCNs unfortunately amplify popularity bias -tail items are less likely to be recommended. This effect prevents the GCN-based RS from ... | Jiajia Chen; Jiancan Wu; Jiawei Chen; Xin Xin; Yong Li; Xiangnan He | [
{
"figure_caption": "Fig. 11Fig. 1 Performance change of LightGCN with different graph convolution layers on Gowalla. Recall@20 and TR@20 stand for the overall recall score and the ratio of tail items in the top-20 recommendation list, respectively.",
"figure_data": "",
"figure_id": "fig_0",
"figure... | [{"Category": "Methodological Basis", "Citation": "[6]", "Explanation": "The cited work LightGCN is the backbone model used in the citing paper, and the BPR loss is employed as the loss function in the GCN-based recommender system."}, {"Category": "Methodological Basis", "Citation": "[9]", "Explanation": "The cited wor... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b26",
"b37",
"b61",
"b6",
"b26",
"b3",
"b43",
"b53",
"b35",
"b16",
"b54",
"b56",
"b8",
"b40",
"b55",
"b40",
"b5",
"b42",
... | 2023-05-25 | 10.1109/WACV.2019.00137 | [
{
"authors": "Abien Fred; Agarap ",
"journal": "",
"ref_id": "b0",
"title": "Deep learning using rectified linear units (relu)",
"year": "2018"
},
{
"authors": "Anthreas Antoniou; Amos Storkey; Harrison Edwards",
"journal": "",
"ref_id": "b1",
"title": "Data Augmentation Gene... | [
{
"formula_coordinates": [
4,
253.36,
262.9,
251.3,
10.7
],
"formula_id": "formula_0",
"formula_text": "L s↔ t = D [f s (x), f t (x)] .(1)"
},
{
"formula_coordinates": [
4,
253.41,
314.13,
251.26,
9.65
],
"formula_id": "form... | HARD: Hard Augmentations for Robust Distillation | Knowledge distillation (KD) is a simple and successful method to transfer knowledge from a teacher to a student model solely based on functional activity. However, current KD has a few shortcomings: it has recently been shown that this method is unsuitable to transfer simple inductive biases like shift equivariance, st... | Arne F Nix; Max F Burg; Fabian H Sinz | [
{
"figure_caption": "HFigure 2 :2Figure2: Fitting the student, a three-layer ReLU MLP, to the teacher function, cos(x), for 10, 000 iterations. We show results for 10 random seeds (A-D) and the distribution of (augmented) training inputs as a normalized histogram (E-H). We compare baseline (no augmentations) wi... | [{"Category": "Supporting Evidence", "Citation": "[27,37,60]", "Explanation": "The cited works provide a foundation for the use of knowledge distillation methods in transfer learning scenarios, including model compression, continual learning, and neuroscience research."}, {"Category": "Extension or Continuation", "Cita... |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b5",
"b11",
"b8",
"b19",
"b6",
"b22",
"b26",
"b3",
"b28",
"b33",
"b11",
"b8",
"b7",
"b20",
"b17",
"b32",
"b16",
... | 10.18653/v1/D19-1651 | [
{
"authors": "Sangnie Bhardwaj; Samarth Aggarwal; Mausam Mausam",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "CaRB: A crowdsourced benchmark for open IE",
"year": "2019"
},
{
"authors": "Sangnie Bhardwaj; Samarth Aggarwal; Mausam Mausam",
"journa... | [] | PIVOINE: Instruction Tuning for Open-world Information Extraction | We consider the problem of Open-world Information Extraction (Open-world IE), which extracts comprehensive entity profiles from unstructured texts. Different from the conventional closed-world setting of Information Extraction (IE), Open-world IE considers a more general situation where entities and relations could be ... | Keming Lu; Xiaoman Pan; Kaiqiang Song; Hongming Zhang; Dong Yu; Jianshu Chen | [
{
"figure_caption": "Figure 1 :1Figure 1: Illustration of Open-world IE and its two main challenges: generalization to unseen instructions and out-of-ontology entities.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "LLM\"Pa... | [{"Category": "Methodological Basis", "Citation": "(Grishman, 2015)", "Explanation": "The cited work by Grishman (2015) provides a foundational understanding of information extraction (IE) and its various tasks, which the citing paper uses to structure its own research on IE."}, {"Category": "Supporting Evidence", "Cit... | |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b31",
"b25",
"b30",
"b21",
"b7",
"b28",
"b26",
"b34",
"b19",
"b33"
],
"table_ref": [],
"text": "Large language models (LLMs) are becoming mainstream and easily acc... | 2024-03-10 | 10.18653/v1/2021.acl-long.565 | [
{
"authors": "Elizabeth Clark; Tal August; Sofia Serrano; Nikita Haduong; Suchin Gururangan; Noah A Smith",
"journal": "Association for Computational Linguistics",
"ref_id": "b0",
"title": "All that's 'human' is not gold: Evaluating human evaluation of generated text",
"year": "2021"
},
{
... | [] | M4: Multi-Generator, Multi-Domain, and Multi-Lingual Black-Box Machine-Generated Text Detection | Large language models (LLMs) have demonstrated remarkable capability to generate fluent responses to a wide variety of user queries. However, this has also raised concerns about the potential misuse of such texts in journalism, education, and academia. In this study, we strive to create automated systems that can detec... | Yuxia Wang; Jonibek Mansurov; Petar Ivanov; Jinyan Su; Artem Shelmanov; Akim Tsvigun; Chenxi Whitehouse; † Osama; Mohammed Afzal; Tarek Mahmoud; Toru Sasaki; Thomas Arnold; Alham Fikri; Nizar Habash; Iryna Gurevych; Preslav Nakov; Mohamed Bin | [
{
"figure_caption": "Figure 1 :1Figure 1: Accuracy of cross-domain experiments: given generations from ChatGPT (top) or davinci (bottom), train on a single domain and test across domains across five detectors. (see more detail in Tables12 and 13)",
"figure_data": "",
"figure_id": "fig_0",
"figure_la... | [{"Category": "Data Source", "Citation": "(Mitchell et al., 2023)", "Explanation": "The cited work provides a benchmark dataset for evaluating the performance of human and machine-generated text classification models."}, {"Category": "Data Source", "Citation": "(Wang et al., 2024)", "Explanation": "The cited work serve... |
[
{
"figure_ref": [
"fig_0"
],
"heading": "Introduction",
"publication_ref": [
"b1",
"b21",
"b3",
"b21",
"b31",
"b16",
"b25",
"b16",
"b29",
"b16",
"b25"
],
"table_ref": [],
"text": "Large language models (LLMs) (Devlin... | 2023-11-06 | 10.18653/v1/N19-1423 | [
{
"authors": "Deborah A Dahl; Madeleine Bates; Michael Brown; William Fisher; Kate Hunicke-Smith; David Pallett; Christine Pao; Alexander Rudnicky; Elizabeth Shriberg",
"journal": "",
"ref_id": "b0",
"title": "Expanding the scope of the ATIS task: The ATIS-3 corpus",
"year": "1994-03-08"
},
... | [
{
"formula_coordinates": [
2,
313.81,
763.54,
211.33,
10.84
],
"formula_id": "formula_0",
"formula_text": "y test ∼ P LM (• | x 1 , y 1 , . . . , x K , y K , x test ) (1)"
},
{
"formula_coordinates": [
3,
105.52,
175.83,
130.66,
15.... | Coverage-based Example Selection for In-Context Learning | In-context learning (ICL), the ability of large language models to perform novel tasks by conditioning on a prompt with a few task examples, requires these examples to be informative about the test instance. The standard approach of independently ranking and selecting the most similar examples selects redundant example... | Shivanshu Gupta; Matt Gardner; Sameer Singh; Sweta Agrawal; Chunting Zhou; Mike Lewis; Luke Zettlemoyer; Marjan 2022 Ghazvininejad; Jacob Andreas; John Bufe; David Burkett; Charles Chen; Josh Clausman; Jean Crawford; Kate Crim; Jordan Deloach; Leah Dorner; Jason Eisner; Hao Fang; Alan Guo; David Hall; Kristin Hayes; Ke... | [
{
"figure_caption": "Figure 1 :1Figure 1: (a) Test input with salient aspects highlighted. (a) Independently selecting similar examples leads to redundancy and failure to demonstrate all salient aspects, in this case, the need to identify the manager. (b) Coverage-based selection using SET-BSR mitigates this by... | [{"Category": "Methodological Basis", "Citation": "(Devlin et al., 2019)", "Explanation": "The cited work by Devlin et al. introduces the concept of large language models (LLMs), which serves as the basis for the research conducted in the citing paper on in-context learning (ICL). The study builds upon the capabilities... |
[
{
"figure_ref": [
"fig_1"
],
"heading": "Introduction",
"publication_ref": [
"b11",
"b40",
"b6",
"b27",
"b22",
"b26",
"b31",
"b30",
"b39",
"b14",
"b2",
"b14",
"b19",
"b34",
"b3",
"b34",
"b19",... | 2023-05-24 | 10.18653/v1/2021.naacl-main.44 | [
{
"authors": "Rami Aly; Christos Christodoulopoulos; Oana Cocarascu; Zhijiang ",
"journal": "",
"ref_id": "b0",
"title": "Proceedings of the Fourth Workshop on Fact Extraction and VERification (FEVER). Association for Computational Linguistics",
"year": "2021"
},
{
"authors": "Raviteja A... | [
{
"formula_coordinates": [
2,
354.02,
731.04,
171.12,
11.62
],
"formula_id": "formula_0",
"formula_text": "Attr (y,A) = avg s∈y Attr (s,A)(2)"
},
{
"formula_coordinates": [
3,
98.17,
149.07,
191.7,
24.67
],
"formula_id": "fo... | PURR: Efficiently Editing Language Model Hallucinations by Denoising Language Model Corruptions | The remarkable capabilities of large language models have been accompanied by a persistent drawback: the generation of false and unsubstantiated claims commonly known as "hallucinations". To combat this issue, recent research has introduced approaches that involve editing and attributing the outputs of language models,... | Anthony Chen; Panupong Pasupat; Sameer Singh; Hongrae Lee; Kelvin Guu; James W Marshall; John Sutter's Carpenter; John Sutter; James W Marshall' | [
{
"figure_caption": "Training PURR. Given a seed query, we search for relevant evidence and summarize them into a claim which we corrupt. PURR is trained to denoise the corruption conditioned on the evidence.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "",
"figure_type": "figure"
... | [{"Category": "Supporting Evidence", "Citation": "(Bai et al., 2022)", "Explanation": "The cited work by Bai et al. provides a method for few-shot prompting of language models to perform editing tasks, which the citing paper uses to highlight the advantages of posthoc methods in text editing."}, {"Category": "Supportin... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b18",
"b19",
"b59",
"b41",
"b55",
"b23",
"b54",
"b52",
"b53",
"b30",
"b32",
"b47",
"b54",
"b12",
"b17",
"b21",
"b46",
"b36",
... | 2023-11-02 | [
{
"authors": "Anthony Michael Ahn; Noah Brohan; Yevgen Brown; Omar Chebotar; Byron Cortes; Chelsea David; Keerthana Finn; Karol Gopalakrishnan; Alex Hausman; Herzog",
"journal": "",
"ref_id": "b0",
"title": "Do as i can, not as i say: Grounding language in robotic affordances",
"year": "2022"
... | [
{
"formula_coordinates": [
20,
123.27,
219.32,
7.31,
100.88
],
"formula_id": "formula_0",
"formula_text": "→ → → → → → → → → → → → →"
},
{
"formula_coordinates": [
36,
123.17,
386.54,
148.16,
22.44
],
"formula_id": "formula_... | Leveraging Pre-trained Large Language Models to Construct and Utilize World Models for Model-based Task Planning | There is a growing interest in applying pre-trained large language models (LLMs) to planning problems. However, methods that use LLMs directly as planners are currently impractical due to several factors, including limited correctness of plans, strong reliance on feedback from interactions with simulators or even the a... | Lin Guan; Karthik Valmeekam; Sarath Sreedharan; Subbarao Kambhampati | [
{
"figure_caption": "Figure 1 :1Figure 1: An overview of our framework and existing methods that use LLMs directly as planners.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},
{
"figure_caption": "Figure 2 :2Figure 2: The prompt template for PDD... | [{"Category": "Methodological Basis", "Citation": "[19]", "Explanation": "The cited work by [19] has showcased remarkable performance in natural language processing tasks, which the citing paper leverages to improve the planning capabilities of LLMs."}, {"Category": "Methodological Basis", "Citation": "[1]", "Explanati... | |
[
{
"figure_ref": [
"fig_0",
"fig_1"
],
"heading": "Introduction",
"publication_ref": [
"b0",
"b22",
"b33",
"b3",
"b8",
"b18",
"b26",
"b31",
"b32",
"b11",
"b13",
"b10",
"b12",
"b13",
"b10",
"b12... | [
{
"authors": "A Anandkumar; R Ge; D Hsu; S M Kakade; M Telgarsky",
"journal": "The Journal of Machine Learning Research",
"ref_id": "b0",
"title": "Tensor decompositions for learning latent variable models",
"year": "2014"
},
{
"authors": "H Attouch; J Bolte; P Redont; A Soubeyran",
... | [
{
"formula_coordinates": [
3,
187.8,
680.53,
236.41,
20.09
],
"formula_id": "formula_0",
"formula_text": "Z(i 1 , i 3 , j 3 ) = I2 i2=1 I4 i4=1 X (i 1 , i 2 , i 3 , i 4 )Y(i 2 , i 4 , j 3 )."
},
{
"formula_coordinates": [
4,
97.2,
219.04,
... | SVDinsTN: A Tensor Network Paradigm for Efficient Structure Search from Regularized Modeling Perspective | Tensor network (TN) representation is a powerful technique for computer vision and machine learning. TN structure search (TN-SS) aims to search for a customized structure to achieve a compact representation, which is a challenging NP-hard problem. Recent "sampling-evaluation"-based methods require sampling an extensive... | Yu-Bang Zheng; Xi-Le Zhao; Junhua Zeng; Chao Li; Qibin Zhao; Heng-Chao Li; Ting-Zhu Huang | [
{
"figure_caption": "Figure 1 :1Figure 1: (a) A graphical illustration of SVD. (b) A graphical illustration of SVD-inspired TN decomposition on a fifth-order tensor. (c) Comparison of the compression ratio (↓) and run time (↓) of different methods on a fifth-order light field image Knights, where the reconstruc... | [{"Category": "Methodological Basis", "Citation": "[12,14]", "Explanation": "The cited works provide a theoretical foundation for the challenging and NP-hard problem of TN structure search, which the citing paper aims to address in its research on the same topic."}, {"Category": "Methodological Basis", "Citation": "[13... | ||
[
{
"figure_ref": [
"fig_1",
"fig_2"
],
"heading": "Introduction",
"publication_ref": [
"b29",
"b32",
"b26",
"b23",
"b37",
"b51",
"b34",
"b20",
"b39",
"b39",
"b34",
"b20",
"b28",
"b54",
"b52"
],
... | 2023-05-24 | 10.18653/v1/2020.emnlp-main.618 | [
{
"authors": "",
"journal": "Wu and Dredze",
"ref_id": "b0",
"title": "mBERT based methods",
"year": "2019"
},
{
"authors": "( Advce; Keung",
"journal": "",
"ref_id": "b1",
"title": "",
"year": "2019"
},
{
"authors": " Tsl (wu",
"journal": "",
"ref_id": "... | [
{
"formula_coordinates": [
3,
70.87,
358.11,
454.27,
417.51
],
"formula_id": "formula_0",
"formula_text": "den representations h = {h i } L i=1 of each sentence x = {x i } L i=1 ∈ D: h = F(x).(1)"
},
{
"formula_coordinates": [
3,
306.14,
503.58... | CoLaDa: A Collaborative Label Denoising Framework for Cross-lingual Named Entity Recognition | Cross-lingual named entity recognition (NER) aims to train an NER system that generalizes well to a target language by leveraging labeled data in a given source language. Previous work alleviates the data scarcity problem by translating source-language labeled data or performing knowledge distillation on target-languag... | Tingting Ma; Qianhui Wu; Huiqiang Jiang; Börje F Karlsson; Tiejun Zhao; Chin-Yew Lin | [
{
"figure_caption": "Figure 1 :1Figure 1: Comparison between previous methods (a/b/c) and our CoLaDa at the i-th iteration (d) CoLaDa starts at M 0 tgt and performs denoising iteratively. D src : Sourcelanguage labeled data. D trans : Translation data. D tgt : Target-language unlabeled data with pseudo-labels g... | [{"Category": "Methodological Basis", "Citation": "(Wu et al., 2020a)", "Explanation": "The cited work, UniTrans, is utilized in the citing paper to finetune a weak model for annotation of unlabeled target-language data, providing a methodological basis for the research conducted in the citing paper."}, {"Category": "M... |
[
{
"figure_ref": [],
"heading": "Introduction",
"publication_ref": [
"b44",
"b19",
"b28",
"b38",
"b26",
"b34",
"b25",
"b37",
"b21",
"b33",
"b6",
"b16",
"b18",
"b27",
"b12"
],
"table_ref": [
"tab_0"
... | 2023-05-24 | 10.21227/c6tm-vw12 | [
{
"authors": "N Audebert; B Le Saux; S Lefèvre",
"journal": "",
"ref_id": "b0",
"title": "Joint learning from earth observation and openstreetmap data to get faster better semantic maps",
"year": "2017"
},
{
"authors": "N Audebert; B Le Saux; S Lefèvre",
"journal": "ISPRS journal of ... | [
{
"formula_coordinates": [
4,
314.61,
281.48,
203.37,
64.99
],
"formula_id": "formula_0",
"formula_text": ".111止--------一------ 1111111■■ 一------一-一-------1111 •• _____-一----- 600 400 200 0 .11111111■■---一一一一一一一-一- 1111111■■■-一一---- 111111■■-----------一-"
},
{
"... | GAMUS: A Geometry-aware Multi-modal Semantic Segmentation Benchmark for Remote Sensing Data | Geometric information in the normalized digital surface models (nDSM) is highly correlated with the semantic class of the land cover. Exploiting two modalities (RGB and nDSM (height)) jointly has great potential to improve the segmentation performance. However, it is still an under-explored field in remote sensing due ... | Zhitong Xiong; Sining Chen; Yi Wang; Lichao Mou; Xiao Xiang Zhu | [
{
"figure_caption": "Fig. 1 :1Fig. 1: Example images of the GAMUS dataset. Images from left to right are the RGB modality, the nDSM modality, the blending visualization image, and the segmentation label.",
"figure_data": "",
"figure_id": "fig_0",
"figure_label": "1",
"figure_type": "figure"
},... | [{"Category": "Methodological Basis", "Citation": "(Zhu et al., 2017)", "Explanation": "The cited work by Zhu et al. (2017) is referenced to highlight the importance of semantic segmentation in the fields of computer vision and remote sensing, providing a foundational basis for the citing paper to build upon."}, {"Cate... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.