abstracts list | id_1 string | id_2 string | pair_id string |
|---|---|---|---|
[
" Text-to-image generative models have achieved unprecedented success in\ngenerating high-quality images based on natural language descriptions. However,\nit is shown that these models tend to favor specific social groups when\nprompted with neutral text descriptions (e.g., 'a photo of a lawyer').\nFollowing Zhao ... | 2210.15230 | 2302.05110 | 2210.15230_2302.05110 |
[
" In scientific research, the method is an indispensable means to solve\nscientific problems and a critical research object. With the advancement of\nsciences, many scientific methods are being proposed, modified, and used in\nacademic literature. The authors describe details of the method in the abstract\nand bod... | 2209.03687 | 2210.03329 | 2209.03687_2210.03329 |
[
" Multimodal pre-training with text, layout, and image has made significant\nprogress for Visually Rich Document Understanding (VRDU), especially the\nfixed-layout documents such as scanned document images. While, there are still\na large number of digital documents where the layout information is not fixed\nand n... | 2110.08518 | 2306.00622 | 2110.08518_2306.00622 |
[
" Technology-assisted review (TAR) refers to iterative active learning\nworkflows for document review in high recall retrieval (HRR) tasks. TAR\nresearch and most commercial TAR software have applied linear models such as\nlogistic regression to lexical features. Transformer-based models with\nsupervised tuning ar... | 2105.01044 | 2203.03312 | 2105.01044_2203.03312 |
[
" The therapeutic working alliance is an important predictor of the outcome of\nthe psychotherapy treatment. In practice, the working alliance is estimated\nfrom a set of scoring questionnaires in an inventory that both the patient and\nthe therapists fill out. In this work, we propose an analytical framework of\n... | 2204.05522 | 2212.14882 | 2204.05522_2212.14882 |
[
" Responsing with image has been recognized as an important capability for an\nintelligent conversational agent. Yet existing works only focus on exploring\nthe multimodal dialogue models which depend on retrieval-based methods, but\nneglecting generation methods. To fill in the gaps, we first present a\nmultimoda... | 2110.08515 | 2204.07980 | 2110.08515_2204.07980 |
[
" Visual information extraction (VIE) plays an important role in Document\nIntelligence. Generally, it is divided into two tasks: semantic entity\nrecognition (SER) and relation extraction (RE). Recently, pre-trained models\nfor documents have achieved substantial progress in VIE, particularly in SER.\nHowever, mo... | 2304.10759 | 2004.10037 | 2304.10759_2004.10037 |
[
" Scaling up language models has been shown to predictably improve performance\nand sample efficiency on a wide range of downstream tasks. This paper instead\ndiscusses an unpredictable phenomenon that we refer to as emergent abilities of\nlarge language models. We consider an ability to be emergent if it is not\n... | 2206.07682 | 2204.04629 | 2206.07682_2204.04629 |
[
" Recently, context-dependent text-to-SQL semantic parsing which translates\nnatural language into SQL in an interaction process has attracted a lot of\nattention. Previous works leverage context-dependence information either from\ninteraction history utterances or the previous predicted SQL queries but fail\nin t... | 2203.07376 | 2204.10185 | 2203.07376_2204.10185 |
[
" In-context learning (ICL), teaching a large language model (LLM) to perform a\ntask with few-shot demonstrations rather than adjusting the model parameters,\nhas emerged as a strong paradigm for using LLMs. While early studies primarily\nused a fixed or random set of demonstrations for all test queries, recent\n... | 2305.14128 | 2103.06370 | 2305.14128_2103.06370 |
[
" Tuning pre-trained language models (PLMs) with task-specific prompts has been\na promising approach for text classification. Particularly, previous studies\nsuggest that prompt-tuning has remarkable superiority in the low-data scenario\nover the generic fine-tuning methods with extra classifiers. The core idea o... | 2108.02035 | 2210.10305 | 2108.02035_2210.10305 |
[
" Hate speech detection models are typically evaluated on held-out test sets.\nHowever, this risks painting an incomplete and potentially misleading picture\nof model performance because of increasingly well-documented systematic gaps\nand biases in hate speech datasets. To enable more targeted diagnostic\ninsight... | 2206.09917 | 2210.01613 | 2206.09917_2210.01613 |
[
" Understanding human tasks through video observations is an essential\ncapability of intelligent agents. The challenges of such capability lie in the\ndifficulty of generating a detailed understanding of situated actions, their\neffects on object states (i.e., state changes), and their causal dependencies.\nThese... | 2210.03929 | 2207.00691 | 2210.03929_2207.00691 |
[
" Recent advancements in large language models (LLMs) have led to the\ndevelopment of highly potent models like OpenAI's ChatGPT. These models have\nexhibited exceptional performance in a variety of tasks, such as question\nanswering, essay composition, and code generation. However, their effectiveness\nin the hea... | 2303.04360 | 1908.06629 | 2303.04360_1908.06629 |
[
" A core process in human cognition is analogical mapping: the ability to\nidentify a similar relational structure between different situations. We\nintroduce a novel task, Visual Analogies of Situation Recognition, adapting the\nclassical word-analogy task into the visual domain. Given a triplet of images,\nthe t... | 2212.04542 | 2302.10198 | 2212.04542_2302.10198 |
[
" In the past few years, it has become increasingly evident that deep neural\nnetworks are not resilient enough to withstand adversarial perturbations in\ninput data, leaving them vulnerable to attack. Various authors have proposed\nstrong adversarial attacks for computer vision and Natural Language Processing\n(N... | 2203.06414 | 2302.11466 | 2203.06414_2302.11466 |
[
" Knowledge graphs (KG) are essential background knowledge providers in many\ntasks. When designing models for KG-related tasks, one of the key tasks is to\ndevise the Knowledge Representation and Fusion (KRF) module that learns the\nrepresentation of elements from KGs and fuses them with task representations.\nWh... | 2303.03922 | 2204.07693 | 2303.03922_2204.07693 |
[
" In this paper, we explore the usability of different natural language\nprocessing models for the sentiment analysis of social media applied to\nfinancial market prediction, using the cryptocurrency domain as a reference. We\nstudy how the different sentiment metrics are correlated with the price\nmovements of Bi... | 2204.10185 | 2305.14483 | 2204.10185_2305.14483 |
[
" While counterfactual data augmentation offers a promising step towards robust\ngeneralization in natural language processing, producing a set of\ncounterfactuals that offer valuable inductive bias for models remains a\nchallenge. Most existing approaches for producing counterfactuals, manual or\nautomated, rely ... | 2210.12365 | 2211.10017 | 2210.12365_2211.10017 |
[
" Analogies play a central role in human commonsense reasoning. The ability to\nrecognize analogies such as \"eye is to seeing what ear is to hearing\",\nsometimes referred to as analogical proportions, shape how we structure\nknowledge and understand language. Surprisingly, however, the task of\nidentifying such ... | 2105.04949 | 2302.04460 | 2105.04949_2302.04460 |
[
" We introduce LAVIS, an open-source deep learning library for LAnguage-VISion\nresearch and applications. LAVIS aims to serve as a one-stop comprehensive\nlibrary that brings recent advancements in the language-vision field accessible\nfor researchers and practitioners, as well as fertilizing future research and\... | 2209.09019 | 2304.00958 | 2209.09019_2304.00958 |
[
" Recently, Neural Topic Models (NTMs) inspired by variational autoencoders\nhave obtained increasingly research interest due to their promising results on\ntext analysis. However, it is usually hard for existing NTMs to achieve good\ndocument representation and coherent/diverse topics at the same time. Moreover,\... | 2008.13537 | 2212.09535 | 2008.13537_2212.09535 |
[
" We study the problem of building text classifiers with little or no training\ndata, commonly known as zero and few-shot text classification. In recent years,\nan approach based on neural textual entailment models has been found to give\nstrong results on a diverse range of tasks. In this work, we show that with\... | 2203.14655 | 2109.01156 | 2203.14655_2109.01156 |
[
" Recent advances in vision-and-language modeling have seen the development of\nTransformer architectures that achieve remarkable performance on multimodal\nreasoning tasks. Yet, the exact capabilities of these black-box models are\nstill poorly understood. While much of previous work has focused on studying\nthei... | 2210.12079 | 2203.04212 | 2210.12079_2203.04212 |
[
" This paper explores new frontiers in agricultural natural language processing\nby investigating the effectiveness of using food-related text corpora for\npretraining transformer-based language models. In particular, we focus on the\ntask of semantic matching, which involves establishing mappings between food\nde... | 2306.11892 | 2204.12191 | 2306.11892_2204.12191 |
[
" Being able to efficiently retrieve the required building information is\ncritical for construction project stakeholders to carry out their engineering\nand management activities. Natural language interface (NLI) systems are\nemerging as a time and cost-effective way to query Building Information Models\n(BIMs). ... | 2303.15116 | 2204.10994 | 2303.15116_2204.10994 |
[
" Deep learning models have achieved great success in many fields, yet they are\nvulnerable to adversarial examples. This paper follows a causal perspective to\nlook into the adversarial vulnerability and proposes Causal Intervention by\nSemantic Smoothing (CISS), a novel framework towards robustness against natur... | 2205.12331 | 2201.08239 | 2205.12331_2201.08239 |
[
" We present a method to formulate algorithm discovery as program search, and\napply it to discover optimization algorithms for deep neural network training.\nWe leverage efficient search techniques to explore an infinite and sparse\nprogram space. To bridge the large generalization gap between proxy and target\nt... | 2302.06675 | 2209.12604 | 2302.06675_2209.12604 |
[
" Large amounts of training data are one of the major reasons for the high\nperformance of state-of-the-art NLP models. But what exactly in the training\ndata causes a model to make a certain prediction? We seek to answer this\nquestion by providing a language for describing how training data influences\npredictio... | 2207.14251 | 2204.08198 | 2207.14251_2204.08198 |
[
" Change captioning is to describe the semantic change between a pair of\nsimilar images in natural language. It is more challenging than general image\ncaptioning, because it requires capturing fine-grained change information while\nbeing immune to irrelevant viewpoint changes, and solving syntax ambiguity in\nch... | 2303.03171 | 2204.09145 | 2303.03171_2204.09145 |
[
" We propose a general and efficient framework to control auto-regressive\ngeneration models with NeurAlly-Decomposed Oracle (NADO). Given a pre-trained\nbase language model and a sequence-level boolean oracle function, we propose to\ndecompose the oracle function into token-level guidance to steer the base model\... | 2205.14219 | 2208.11857 | 2205.14219_2208.11857 |
[
" The notion of equality (identity) is simple and ubiquitous, making it a key\ncase study for broader questions about the representations supporting abstract\nrelational reasoning. Previous work suggested that neural networks were not\nsuitable models of human relational reasoning because they could not represent\... | 2006.07968 | 2208.04347 | 2006.07968_2208.04347 |
[
" Natural language understanding (NLU) models often rely on dataset biases\nrather than intended task-relevant features to achieve high performance on\nspecific datasets. As a result, these models perform poorly on datasets outside\nthe training distribution. Some recent studies address this issue by reducing\nthe... | 2212.05421 | 2104.07091 | 2212.05421_2104.07091 |
[
" While large pretrained Transformer models have proven highly capable at\ntackling natural language tasks, handling long sequence inputs continues to be\na significant challenge. One such task is long input summarization, where\ninputs are longer than the maximum input context of most pretrained models.\nThrough ... | 2208.04347 | 2208.04415 | 2208.04347_2208.04415 |
[
" Transformer-based sequence-to-sequence architectures, while achieving\nstate-of-the-art results on a large number of NLP tasks, can still suffer from\noverfitting during training. In practice, this is usually countered either by\napplying regularization methods (e.g. dropout, L2-regularization) or by\nproviding ... | 2109.07276 | 2210.07783 | 2109.07276_2210.07783 |
[
" While counterfactual data augmentation offers a promising step towards robust\ngeneralization in natural language processing, producing a set of\ncounterfactuals that offer valuable inductive bias for models remains a\nchallenge. Most existing approaches for producing counterfactuals, manual or\nautomated, rely ... | 2210.12365 | 2302.05110 | 2210.12365_2302.05110 |
[
" We propose LLMA, an LLM accelerator to losslessly speed up Large Language\nModel (LLM) inference with references. LLMA is motivated by the observation\nthat there are abundant identical text spans between the decoding result by an\nLLM and the reference that is available in many real world scenarios (e.g.,\nretr... | 2304.04487 | 2205.03092 | 2304.04487_2205.03092 |
[
" Large pre-trained language models (PLMs) have demonstrated strong performance\non natural language understanding (NLU) tasks through fine-tuning. However,\nfine-tuned models still suffer from overconfident predictions, especially in\nout-of-domain settings. In this paper, we tackle the problem of calibrating\nfi... | 2305.19249 | 2104.08661 | 2305.19249_2104.08661 |
[
" Named Entity Recognition (NER) is a foundational NLP task that aims to\nprovide class labels like Person, Location, Organisation, Time, and Number to\nwords in free text. Named Entities can also be multi-word expressions where the\nadditional I-O-B annotation information helps label them during the NER\nannotati... | 2204.13743 | 2112.02512 | 2204.13743_2112.02512 |
[
" The open-ended Visual Question Answering (VQA) task requires AI models to\njointly reason over visual and natural language inputs using world knowledge.\nRecently, pre-trained Language Models (PLM) such as GPT-3 have been applied to\nthe task and shown to be powerful world knowledge sources. However, these\nmeth... | 2305.18842 | 2206.00826 | 2305.18842_2206.00826 |
[
" OpenAI has recently released GPT-4 (a.k.a. ChatGPT plus), which is\ndemonstrated to be one small step for generative AI (GAI), but one giant leap\nfor artificial general intelligence (AGI). Since its official release in\nNovember 2022, ChatGPT has quickly attracted numerous users with extensive\nmedia coverage. ... | 2304.06488 | 2109.04500 | 2304.06488_2109.04500 |
[
" Active learning (AL) is a prominent technique for reducing the annotation\neffort required for training machine learning models. Deep learning offers a\nsolution for several essential obstacles to deploying AL in practice but\nintroduces many others. One of such problems is the excessive computational\nresources... | 2205.03598 | 2112.03572 | 2205.03598_2112.03572 |
[
" Since the first end-to-end neural coreference resolution model was\nintroduced, many extensions to the model have been proposed, ranging from using\nhigher-order inference to directly optimizing evaluation metrics using\nreinforcement learning. Despite improving the coreference resolution\nperformance by a large... | 2107.01700 | 2209.14958 | 2107.01700_2209.14958 |
[
" Large-scale pre-trained language models have been shown to be helpful in\nimproving the naturalness of text-to-speech (TTS) models by enabling them to\nproduce more naturalistic prosodic patterns. However, these models are usually\nword-level or sup-phoneme-level and jointly trained with phonemes, making them\ni... | 2301.08810 | 2206.05511 | 2301.08810_2206.05511 |
[
" This paper presents Z-Code++, a new pre-trained language model optimized for\nabstractive text summarization. The model extends the state of the art\nencoder-decoder model using three techniques. First, we use a two-phase\npre-training process to improve model's performance on low-resource\nsummarization tasks. ... | 2208.09770 | 2206.05975 | 2208.09770_2206.05975 |
[
" Disentangling content and speaking style information is essential for\nzero-shot non-parallel voice conversion (VC). Our previous study investigated a\nnovel framework with disentangled sequential variational autoencoder (DSVAE) as\nthe backbone for information decomposition. We have demonstrated that\nsimultane... | 2205.05227 | 2209.10887 | 2205.05227_2209.10887 |
[
" Recent parameter-efficient language model tuning (PELT) methods manage to\nmatch the performance of fine-tuning with much fewer trainable parameters and\nperform especially well when training data is limited. However, different PELT\nmethods may perform rather differently on the same task, making it nontrivial\n... | 2110.07577 | 2305.10250 | 2110.07577_2305.10250 |
[
" This article presents morphologically-annotated Yemeni, Sudanese, Iraqi, and\nLibyan Arabic dialects Lisan corpora. Lisan features around 1.2 million tokens.\nWe collected the content of the corpora from several social media platforms.\nThe Yemeni corpus (~ 1.05M tokens) was collected automatically from Twitter.... | 2212.06468 | 2205.15219 | 2212.06468_2205.15219 |
[
" Large amounts of training data are one of the major reasons for the high\nperformance of state-of-the-art NLP models. But what exactly in the training\ndata causes a model to make a certain prediction? We seek to answer this\nquestion by providing a language for describing how training data influences\npredictio... | 2207.14251 | 2210.04191 | 2207.14251_2210.04191 |
[
" Summarization datasets are often assembled either by scraping naturally\noccurring public-domain summaries -- which are nearly always in\ndifficult-to-work-with technical domains -- or by using approximate heuristics\nto extract them from everyday text -- which frequently yields unfaithful\nsummaries. In this wo... | 2205.11465 | 2004.05964 | 2205.11465_2004.05964 |
[
" Inferring meta information about tables, such as column headers or\nrelationships between columns, is an active research topic in data management\nas we find many tables are missing some of this information. In this paper, we\nstudy the problem of annotating table columns (i.e., predicting column types\nand the ... | 2104.01785 | 2211.09527 | 2104.01785_2211.09527 |
[
" Despite recent concerns about undesirable behaviors generated by large\nlanguage models (LLMs), including non-factual, biased, and hateful language, we\nfind LLMs are inherent multi-task language checkers based on their latent\nrepresentations of natural and social knowledge. We present an interpretable,\nunifie... | 2304.03728 | 2108.10015 | 2304.03728_2108.10015 |
[
" YourTTS brings the power of a multilingual approach to the task of zero-shot\nmulti-speaker TTS. Our method builds upon the VITS model and adds several novel\nmodifications for zero-shot multi-speaker and multilingual training. We\nachieved state-of-the-art (SOTA) results in zero-shot multi-speaker TTS and\nresu... | 2112.02418 | 2305.14483 | 2112.02418_2305.14483 |
[
" Recent advances in deep learning have relied heavily on the use of large\nTransformers due to their ability to learn at scale. However, the core building\nblock of Transformers, the attention operator, exhibits quadratic cost in\nsequence length, limiting the amount of context accessible. Existing\nsubquadratic ... | 2302.10866 | 2105.14762 | 2302.10866_2105.14762 |
[
" With the advance of language models, privacy protection is receiving more\nattention. Training data extraction is therefore of great importance, as it can\nserve as a potential tool to assess privacy leakage. However, due to the\ndifficulty of this task, most of the existing methods are proof-of-concept and\nsti... | 2302.04460 | 2211.04054 | 2302.04460_2211.04054 |
[
" Structure information extraction refers to the task of extracting structured\ntext fields from web pages, such as extracting a product offer from a shopping\npage including product title, description, brand and price. It is an important\nresearch topic which has been widely studied in document understanding and ... | 2202.00217 | 2302.13136 | 2202.00217_2302.13136 |
[
" In this paper we present VDTTS, a Visually-Driven Text-to-Speech model.\nMotivated by dubbing, VDTTS takes advantage of video frames as an additional\ninput alongside text, and generates speech that matches the video signal. We\ndemonstrate how this allows VDTTS to, unlike plain TTS models, generate speech\nthat... | 2111.10139 | 2303.02399 | 2111.10139_2303.02399 |
[
" Previous literature has proved that Pretrained Language Models (PLMs) can\nstore factual knowledge. However, we find that facts stored in the PLMs are not\nalways correct. It motivates us to explore a fundamental question: How do we\ncalibrate factual knowledge in PLMs without re-training from scratch? In this\n... | 2210.03329 | 2112.08688 | 2210.03329_2112.08688 |
[
" Dialogue systems are usually categorized into two types, open-domain and\ntask-oriented. The first one focuses on chatting with users and making them\nengage in the conversations, where selecting a proper topic to fit the dialogue\ncontext is essential for a successful dialogue. The other one focuses on a\nspeci... | 2204.10591 | 2205.00355 | 2204.10591_2205.00355 |
[
" Contextually aware intelligent agents are often required to understand the\nusers and their surroundings in real-time. Our goal is to build Artificial\nIntelligence (AI) systems that can assist children in their learning process.\nWithin such complex frameworks, Spoken Dialogue Systems (SDS) are crucial\nbuildin... | 2205.04006 | 2112.03572 | 2205.04006_2112.03572 |
[
" We introduce SummScreen, a summarization dataset comprised of pairs of TV\nseries transcripts and human written recaps. The dataset provides a challenging\ntestbed for abstractive summarization for several reasons. Plot details are\noften expressed indirectly in character dialogues and may be scattered across\nt... | 2104.07091 | 2201.08542 | 2104.07091_2201.08542 |
[
" When people answer questions about a specific situation, e.g., \"I cheated on\nmy mid-term exam last week. Was that wrong?\", cognitive science suggests that\nthey form a mental picture of that situation before answering. While we do not\nknow how language models (LMs) answer such questions, we conjecture that t... | 2112.08656 | 2203.09161 | 2112.08656_2203.09161 |
[
" Sentiment analysis is one of the most widely studied applications in NLP, but\nmost work focuses on languages with large amounts of data. We introduce the\nfirst large-scale human-annotated Twitter sentiment dataset for the four most\nwidely spoken languages in Nigeria (Hausa, Igbo, Nigerian-Pidgin, and\nYor\\`u... | 2201.08277 | 2204.06252 | 2201.08277_2204.06252 |
[
" Vision-and-Language Navigation (VLN) is the task that requires an agent to\nnavigate through the environment based on natural language instructions. At\neach step, the agent takes the next action by selecting from a set of navigable\nlocations. In this paper, we aim to take one step further and explore whether\n... | 2304.04907 | 2303.14956 | 2304.04907_2303.14956 |
[
" When humans cooperate, they frequently coordinate their activity through both\nverbal communication and non-verbal actions, using this information to infer a\nshared goal and plan. How can we model this inferential ability? In this paper,\nwe introduce a model of a cooperative team where one agent, the principal... | 2306.16207 | 2305.14635 | 2306.16207_2305.14635 |
[
" Out-of-Domain (OOD) intent detection is important for practical dialog\nsystems. To alleviate the issue of lacking OOD training samples, some works\npropose synthesizing pseudo OOD samples and directly assigning one-hot OOD\nlabels to these pseudo samples. However, these one-hot labels introduce noises\nto the t... | 2211.05561 | 2212.08120 | 2211.05561_2212.08120 |
[
" Coreference resolution -- which is a crucial task for understanding discourse\nand language at large -- has yet to witness widespread benefits from large\nlanguage models (LLMs). Moreover, coreference resolution systems largely rely\non supervised labels, which are highly expensive and difficult to annotate,\nth... | 2205.07407 | 2210.03588 | 2205.07407_2210.03588 |
[
" The success of ChatGPT has recently attracted numerous efforts to replicate\nit, with instruction-tuning strategies being a key factor in achieving\nremarkable results. Instruction-tuning not only significantly enhances the\nmodel's performance and generalization but also makes the model's generated\nresults mor... | 2303.14742 | 2205.10479 | 2303.14742_2205.10479 |
[
" Large language models (LLMs) learn not only natural text generation abilities\nbut also social biases against different demographic groups from real-world\ndata. This poses a critical risk when deploying LLM-based applications.\nExisting research and resources are not readily applicable in South Korea due\nto th... | 2305.17701 | 2111.13854 | 2305.17701_2111.13854 |
[
" Finetuning large pre-trained language models with a task-specific head has\nadvanced the state-of-the-art on many natural language understanding\nbenchmarks. However, models with a task-specific head require a lot of training\ndata, making them susceptible to learning and exploiting dataset-specific\nsuperficial... | 2205.09295 | 2205.11308 | 2205.09295_2205.11308 |
[
" The remarkable success of transformers in the field of natural language\nprocessing has sparked the interest of the speech-processing community, leading\nto an exploration of their potential for modeling long-range dependencies\nwithin speech sequences. Recently, transformers have gained prominence across\nvario... | 2303.11607 | 2305.13198 | 2303.11607_2305.13198 |
[
" Despite recent concerns about undesirable behaviors generated by large\nlanguage models (LLMs), including non-factual, biased, and hateful language, we\nfind LLMs are inherent multi-task language checkers based on their latent\nrepresentations of natural and social knowledge. We present an interpretable,\nunifie... | 2304.03728 | 2205.12331 | 2304.03728_2205.12331 |
[
" A big convergence of model architectures across language, vision, speech, and\nmultimodal is emerging. However, under the same name \"Transformers\", the above\nareas use different implementations for better performance, e.g.,\nPost-LayerNorm for BERT, and Pre-LayerNorm for GPT and vision Transformers. We\ncall ... | 2210.06423 | 2006.08328 | 2210.06423_2006.08328 |
[
" The attention mechanism is considered the backbone of the widely-used\nTransformer architecture. It contextualizes the input by computing\ninput-specific attention matrices. We find that this mechanism, while powerful\nand elegant, is not as important as typically thought for pretrained language\nmodels. We intr... | 2211.03495 | 2305.15334 | 2211.03495_2305.15334 |
[
" Moral norms vary across cultures. A recent line of work suggests that English\nlarge language models contain human-like moral biases, but these studies\ntypically do not examine moral variation in a diverse cultural setting. We\ninvestigate the extent to which monolingual English language models contain\nknowled... | 2306.01857 | 2110.00976 | 2306.01857_2110.00976 |
[
" We present a method to formulate algorithm discovery as program search, and\napply it to discover optimization algorithms for deep neural network training.\nWe leverage efficient search techniques to explore an infinite and sparse\nprogram space. To bridge the large generalization gap between proxy and target\nt... | 2302.06675 | 2206.00856 | 2302.06675_2206.00856 |
[
" We present L3Cube-MahaCorpus a Marathi monolingual data set scraped from\ndifferent internet sources. We expand the existing Marathi monolingual corpus\nwith 24.8M sentences and 289M tokens. We further present, MahaBERT, MahaAlBERT,\nand MahaRoBerta all BERT-based masked language models, and MahaFT, the fast\nte... | 2202.01159 | 2208.04415 | 2202.01159_2208.04415 |
[
" Responsive teaching is a highly effective strategy that promotes student\nlearning. In math classrooms, teachers might \"funnel\" students towards a\nnormative answer or \"focus\" students to reflect on their own thinking,\ndeepening their understanding of math concepts. When teachers focus, they treat\nstudents... | 2208.04715 | 2105.01044 | 2208.04715_2105.01044 |
[
" This paper introduces a new dysarthric speech command dataset in Italian,\ncalled EasyCall corpus. The dataset consists of 21386 audio recordings from 24\nhealthy and 31 dysarthric speakers, whose individual degree of speech\nimpairment was assessed by neurologists through the Therapy Outcome Measure.\nThe corpu... | 2104.02542 | 2306.10790 | 2104.02542_2306.10790 |
[
" Being able to efficiently retrieve the required building information is\ncritical for construction project stakeholders to carry out their engineering\nand management activities. Natural language interface (NLI) systems are\nemerging as a time and cost-effective way to query Building Information Models\n(BIMs). ... | 2303.15116 | 2212.10898 | 2303.15116_2212.10898 |
[
" People always desire an embodied agent that can perform a task by\nunderstanding language instruction. Moreover, they also want to monitor and\nexpect agents to understand commands the way they expected. But, how to build\nsuch an embodied agent is still unclear. Recently, people can explore this\nproblem with t... | 2203.04637 | 2302.07138 | 2203.04637_2302.07138 |
[
" We present KPI-BERT, a system which employs novel methods of named entity\nrecognition (NER) and relation extraction (RE) to extract and link key\nperformance indicators (KPIs), e.g. \"revenue\" or \"interest expenses\", of\ncompanies from real-world German financial documents. Specifically, we\nintroduce an end... | 2208.02140 | 2304.09607 | 2208.02140_2304.09607 |
[
" Cross-Lingual Word Embeddings (CLWEs) encode words from two or more languages\nin a shared high-dimensional space in which vectors representing words with\nsimilar meaning (regardless of language) are closely located. Existing methods\nfor building high-quality CLWEs learn mappings that minimise the $\\ell_{2}$\... | 2104.04916 | 2301.04449 | 2104.04916_2301.04449 |
[
" Large language models (LMs), while powerful, are not immune to mistakes, but\ncan be difficult to retrain. Our goal is for an LM to continue to improve after\ndeployment, without retraining, using feedback from the user. Our approach\npairs an LM with (i) a growing memory of cases where the user identified an\no... | 2112.09737 | 2306.16207 | 2112.09737_2306.16207 |
[
" Stance detection deals with identifying an author's stance towards a target.\nMost existing stance detection models are limited because they do not consider\nrelevant contextual information which allows for inferring the stance\ncorrectly. Complementary context can be found in knowledge bases but\nintegrating th... | 2211.01874 | 2301.10896 | 2211.01874_2301.10896 |
[
" With the rapid increase of multimedia data, a large body of literature has\nemerged to work on multimodal summarization, the majority of which target at\nrefining salient information from textual and visual modalities to output a\npictorial summary with the most relevant images. Existing methods mostly focus\non... | 2109.05812 | 2305.13282 | 2109.05812_2305.13282 |
[
" Distant supervision assumes that any sentence containing the same entity\npairs reflects identical relationships. Previous works of distantly supervised\nrelation extraction (DSRE) task generally focus on sentence-level or bag-level\nde-noising techniques independently, neglecting the explicit interaction with\n... | 2202.13352 | 2105.03824 | 2202.13352_2105.03824 |
[
" Prompt tuning is a new few-shot transfer learning technique that only tunes\nthe learnable prompt for pre-trained vision and language models such as CLIP.\nHowever, existing prompt tuning methods tend to learn spurious or entangled\nrepresentations, which leads to poor generalization to unseen concepts. Towards\... | 2210.10362 | 2110.00976 | 2210.10362_2110.00976 |
[
" We study dangling-aware entity alignment in knowledge graphs (KGs), which is\nan underexplored but important problem. As different KGs are naturally\nconstructed by different sets of entities, a KG commonly contains some dangling\nentities that cannot find counterparts in other KGs. Therefore, dangling-aware\nen... | 2205.02406 | 2207.03477 | 2205.02406_2207.03477 |
[
" Change captioning is to describe the semantic change between a pair of\nsimilar images in natural language. It is more challenging than general image\ncaptioning, because it requires capturing fine-grained change information while\nbeing immune to irrelevant viewpoint changes, and solving syntax ambiguity in\nch... | 2303.03171 | 2109.07830 | 2303.03171_2109.07830 |
[
" Prevailing deep models are single-purpose and overspecialize at individual\ntasks. However, when being extended to new tasks, they typically forget\npreviously learned skills and learn from scratch. We address this issue by\nintroducing SkillNet-NLU, a general-purpose model that stitches together\nexisting skill... | 2203.03312 | 2203.15827 | 2203.03312_2203.15827 |
[
" We present a unified Vision-Language pretrained Model (VLMo) that jointly\nlearns a dual encoder and a fusion encoder with a modular Transformer network.\nSpecifically, we introduce Mixture-of-Modality-Experts (MoME) Transformer,\nwhere each block contains a pool of modality-specific experts and a shared\nself-a... | 2111.02358 | 2212.09621 | 2111.02358_2212.09621 |
[
" Various applications of voice synthesis have been developed independently\ndespite the fact that they generate \"voice\" as output in common. In addition,\nthe majority of voice synthesis models currently rely on annotated audio data,\nbut it is crucial to scale them to self-supervised datasets in order to\neffe... | 2305.19269 | 2306.06601 | 2305.19269_2306.06601 |
[
" Customer reviews are vital for making purchasing decisions in the Information\nAge. Such reviews can be automatically summarized to provide the user with an\noverview of opinions. In this tutorial, we present various aspects of opinion\nsummarization that are useful for researchers and practitioners. First, we w... | 2206.01543 | 2205.12206 | 2206.01543_2205.12206 |
[
" Personal assistants, automatic speech recognizers and dialogue understanding\nsystems are becoming more critical in our interconnected digital world. A clear\nexample is air traffic control (ATC) communications. ATC aims at guiding\naircraft and controlling the airspace in a safe and optimal manner. These\nvoice... | 2211.04054 | 2111.00161 | 2211.04054_2111.00161 |
[
" The Visual Question Answering (VQA) task aspires to provide a meaningful\ntestbed for the development of AI models that can jointly reason over visual\nand natural language inputs. Despite a proliferation of VQA datasets, this goal\nis hindered by a set of common limitations. These include a reliance on\nrelativ... | 2206.01718 | 2205.11308 | 2206.01718_2205.11308 |
[
" Event extraction (EE) is an essential task of information extraction, which\naims to extract structured event information from unstructured text. Most prior\nwork focuses on extracting flat events while neglecting overlapped or nested\nones. A few models for overlapped and nested EE includes several successive\n... | 2209.02693 | 2305.04320 | 2209.02693_2305.04320 |
[
" One of the emerging research trends in natural language understanding is\nmachine reading comprehension (MRC) which is the task to find answers to human\nquestions based on textual data. Existing Vietnamese datasets for MRC research\nconcentrate solely on answerable questions. However, in reality, questions can\... | 2203.11400 | 2210.06104 | 2203.11400_2210.06104 |
[
" In this work, we combine the two paradigms: Federated Learning (FL) and\nContinual Learning (CL) for text classification task in cloud-edge continuum.\nThe objective of Federated Continual Learning (FCL) is to improve deep learning\nmodels over life time at each client by (relevant and efficient) knowledge\ntran... | 2210.06101 | 2205.05738 | 2210.06101_2205.05738 |
[
" Despite attempts to increase gender parity in politics, global efforts have\nstruggled to ensure equal female representation. This is likely tied to\nimplicit gender biases against women in authority. In this work, we present a\ncomprehensive study of gender biases that appear in online political\ndiscussion. To... | 2112.12014 | 2112.02512 | 2112.12014_2112.02512 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.