abstracts
listlengths
2
2
id_1
stringlengths
9
14
id_2
stringlengths
9
14
pair_id
stringlengths
20
25
[ " The ever-growing diversity of pre-training text corpora has equipped language\nmodels with generalization capabilities across various downstream tasks.\nHowever, such diverse datasets are often too large for academic budgets; hence,\nmost research on Transformer architectures, training procedures, optimizers,\ne...
2304.08442
2205.12491
2304.08442_2205.12491
[ " Large language models (LLMs) have demonstrated remarkable prowess in language\nunderstanding and generation. Advancing from foundation LLMs to\ninstructionfollowing LLMs, instruction tuning plays a vital role in aligning\nLLMs to human preferences. However, the existing LLMs are usually focused on\nEnglish, lead...
2306.10968
2206.11349
2306.10968_2206.11349
[ " Several pre-training objectives, such as masked language modeling (MLM), have\nbeen proposed to pre-train language models (e.g. BERT) with the aim of learning\nbetter language representations. However, to the best of our knowledge, no\nprevious work so far has investigated how different pre-training objectives\n...
2203.10415
2210.14389
2203.10415_2210.14389
[ " Given a natural language statement, how to verify its veracity against a\nlarge-scale textual knowledge source like Wikipedia? Most existing neural\nmodels make predictions without giving clues about which part of a false claim\ngoes wrong. In this paper, we propose LOREN, an approach for interpretable fact\nver...
2012.13577
2302.06426
2012.13577_2302.06426
[ " We propose DiffCSE, an unsupervised contrastive learning framework for\nlearning sentence embeddings. DiffCSE learns sentence embeddings that are\nsensitive to the difference between the original sentence and an edited\nsentence, where the edited sentence is obtained by stochastically masking out\nthe original s...
2204.10298
2211.05172
2204.10298_2211.05172
[ " When humans design cost or goal specifications for robots, they often produce\nspecifications that are ambiguous, underspecified, or beyond planners' ability\nto solve. In these cases, corrections provide a valuable tool for\nhuman-in-the-loop robot control. Corrections might take the form of new goal\nspecifica...
2204.05186
2303.10583
2204.05186_2303.10583
[ " Pre-training methods with contrastive learning objectives have shown\nremarkable success in dialog understanding tasks. However, current contrastive\nlearning solely considers the self-augmented dialog samples as positive samples\nand treats all other dialog samples as negative ones, which enforces dissimilar\nr...
2209.06638
2205.15868
2209.06638_2205.15868
[ " Recent progress in language model pre-training has achieved a great success\nvia leveraging large-scale unstructured textual data. However, it is still a\nchallenge to apply pre-training on structured tabular data due to the absence\nof large-scale high-quality tabular data. In this paper, we propose TAPEX to\ns...
2107.07653
2204.12679
2107.07653_2204.12679
[ " Recent advances in NLP are brought by a range of large-scale pretrained\nlanguage models (PLMs). These PLMs have brought significant performance gains\nfor a range of NLP tasks, circumventing the need to customize complex designs\nfor specific tasks. However, most current work focus on finetuning PLMs on a\ndoma...
2211.03154
2204.06518
2211.03154_2204.06518
[ " Adapter-tuning is a paradigm that transfers a pretrained language model to\ndownstream tasks by adding and tuning a small number of new parameters.\nPreviously proposed adapter architectures are all feed-forward neural networks.\nIn this paper, we investigate the effectiveness of using tiny-attention --\ni.e., a...
2211.01979
2306.04050
2211.01979_2306.04050
[ " The paper describes the open Russian medical language understanding benchmark\ncovering several task types (classification, question answering, natural\nlanguage inference, named entity recognition) on a number of novel text sets.\nGiven the sensitive nature of the data in healthcare, such a benchmark\npartially...
2201.06499
2201.01209
2201.06499_2201.01209
[ " Multimodal knowledge graph completion (MKGC) aims to predict missing entities\nin MKGs. Previous works usually share relation representation across\nmodalities. This results in mutual interference between modalities during\ntraining, since for a pair of entities, the relation from one modality probably\ncontradi...
2210.08821
2305.19709
2210.08821_2305.19709
[ " We study learning from user feedback for extractive question answering by\nsimulating feedback using supervised data. We cast the problem as contextual\nbandit learning, and analyze the characteristics of several learning scenarios\nwith focus on reducing data annotation. We show that systems initially trained\n...
2203.10079
2006.01245
2203.10079_2006.01245
[ " Pre-trained neural Language Models (PTLM), such as CodeBERT, are recently\nused in software engineering as models pre-trained on large source code\ncorpora. Their knowledge is transferred to downstream tasks (e.g. code clone\ndetection) via fine-tuning. In natural language processing (NLP), other\nalternatives f...
2204.08653
2305.02156
2204.08653_2305.02156
[ " Retrieval-augmented Neural Machine Translation models have been successful in\nmany translation scenarios. Different from previous works that make use of\nmutually similar but redundant translation memories~(TMs), we propose a new\nretrieval-augmented NMT to model contrastively retrieved translation memories\nth...
2212.03140
2208.00748
2212.03140_2208.00748
[ " Tables are often created with hierarchies, but existing works on table\nreasoning mainly focus on flat tables and neglect hierarchical tables.\nHierarchical tables challenge existing methods by hierarchical indexing, as\nwell as implicit relationships of calculation and semantics. This work presents\nHiTab, a fr...
2108.06712
2204.06518
2108.06712_2204.06518
[ " Instruction tuning is an emergent paradigm in NLP wherein natural language\ninstructions are leveraged with language models to induce zero-shot performance\non unseen tasks. Instructions have been shown to enable good performance on\nunseen tasks and datasets in both large and small language models. Dialogue is\...
2205.12673
2010.12676
2205.12673_2010.12676
[ " Contrastive learning models have achieved great success in unsupervised\nvisual representation learning, which maximize the similarities between feature\nrepresentations of different views of the same image, while minimize the\nsimilarities between feature representations of views of different images. In\ntext s...
2109.03481
2106.00903
2109.03481_2106.00903
[ " While neural text-to-speech systems perform remarkably well in high-resource\nscenarios, they cannot be applied to the majority of the over 6,000 spoken\nlanguages in the world due to a lack of appropriate training data. In this\nwork, we use embeddings derived from articulatory vectors rather than\nembeddings d...
2203.03191
2303.10368
2203.03191_2303.10368
[ " The widespread dissemination of toxic online posts is increasingly damaging\nto society. However, research on detecting toxic language in Chinese has lagged\nsignificantly. Existing datasets lack fine-grained annotation of toxic types\nand expressions, and ignore the samples with indirect toxicity. In addition, ...
2305.04446
2304.11164
2305.04446_2304.11164
[ " Reliable methods for automatic readability assessment have the potential to\nimpact a variety of fields, ranging from machine translation to self-informed\nlearning. Recently, large language models for the German language (such as\nGBERT and GPT-2-Wechsel) have become available, allowing to develop Deep\nLearnin...
2209.04299
2205.02014
2209.04299_2205.02014
[ " Self-attention is a key enabler of state-of-art accuracy for various\ntransformer-based Natural Language Processing models. This attention mechanism\ncalculates a correlation score for each word with respect to the other words in\na sentence. Commonly, only a small subset of words highly correlates with the\nwor...
2204.03227
2203.10545
2204.03227_2203.10545
[ " Recently, Language Models (LMs) instruction-tuned on multiple tasks, also\nknown as multitask-prompted fine-tuning (MT), have shown the capability to\ngeneralize to unseen tasks. Previous work has shown that scaling the number of\ntraining tasks is the key component in making stronger MT LMs. In this work, we\nr...
2302.03202
2205.02035
2302.03202_2205.02035
[ " For most natural language processing tasks, the dominant practice is to\nfinetune large pretrained transformer models (e.g., BERT) using smaller\ndownstream datasets. Despite the success of this approach, it remains unclear\nto what extent these gains are attributable to the massive background corpora\nemployed ...
2209.14389
2305.20010
2209.14389_2305.20010
[ " Given the ubiquitous nature of numbers in text, reasoning with numbers to\nperform simple calculations is an important skill of AI systems. While many\ndatasets and models have been developed to this end, state-of-the-art AI\nsystems are brittle; failing to perform the underlying mathematical reasoning\nwhen the...
2204.05660
2205.04605
2204.05660_2205.04605
[ " Semantic communication in the 6G era has been deemed a promising\ncommunication paradigm to break through the bottleneck of traditional\ncommunications. However, its applications for the multi-user scenario,\nespecially the broadcasting case, remain under-explored. To effectively exploit\nthe benefits enabled by...
2209.09425
2204.04748
2209.09425_2204.04748
[ " Although large conversational AI models such as OpenAI's ChatGPT have\ndemonstrated great potential, we question whether such models can guarantee\nfactual accuracy. Recently, technology companies such as Microsoft and Google\nhave announced new services which aim to combine search engines with\nconversational A...
2304.11076
2306.08401
2304.11076_2306.08401
[ " Despite the effectiveness of utilizing the BERT model for document ranking,\nthe high computational cost of such approaches limits their uses. To this end,\nthis paper first empirically investigates the effectiveness of two knowledge\ndistillation models on the document ranking task. In addition, on top of the\n...
2009.07531
2202.03829
2009.07531_2202.03829
[ " The growing size of neural language models has led to increased attention in\nmodel compression. The two predominant approaches are pruning, which gradually\nremoves weights from a pre-trained model, and distillation, which trains a\nsmaller compact model to match a larger one. Pruning methods can significantly\...
2204.00408
2203.10012
2204.00408_2203.10012
[ " With the increasing number of clinical trial reports generated every day, it\nis becoming hard to keep up with novel discoveries that inform evidence-based\nhealthcare recommendations. To help automate this process and assist medical\nexperts, NLP solutions are being developed. This motivated the SemEval-2023\nT...
2304.13180
2209.10918
2304.13180_2209.10918
[ " Automated reasoning with unstructured natural text is a key requirement for\nmany potential applications of NLP and for developing robust AI systems.\nRecently, Language Models (LMs) have demonstrated complex reasoning capacities\neven without any finetuning. However, existing evaluation for automated\nreasoning...
2306.07934
2112.08804
2306.07934_2112.08804
[ " Linking computational natural language processing (NLP) models and neural\nresponses to language in the human brain on the one hand facilitates the effort\ntowards disentangling the neural representations underpinning language\nperception, on the other hand provides neurolinguistics evidence to evaluate\nand imp...
2303.14871
2205.06439
2303.14871_2205.06439
[ " Summarization systems make numerous \"decisions\" about summary properties\nduring inference, e.g. degree of copying, specificity and length of outputs,\netc. However, these are implicitly encoded within model parameters and specific\nstyles cannot be enforced. To address this, we introduce HydraSum, a new\nsumm...
2110.04400
2305.13281
2110.04400_2305.13281
[ " Background: Electronic Health Records hold detailed longitudinal information\nabout each patient's health status and general clinical history, a large\nportion of which is stored within the unstructured text. Existing approaches\nfocus mostly on structured data and a subset of single-domain outcomes. We\nexplore...
2212.08072
2205.01703
2212.08072_2205.01703
[ " Parameter-efficient fine-tuning (PEFT) of pre-trained language models has\nrecently demonstrated remarkable achievements, effectively matching the\nperformance of full fine-tuning while utilizing significantly fewer trainable\nparameters, and consequently addressing the storage and communication\nconstraints. No...
2305.16742
2203.05948
2305.16742_2203.05948
[ " Text summarization aims to condense long documents and retain key\ninformation. Critical to the success of a summarization model is the faithful\ninference of latent representations of words or tokens in the source documents.\nMost recent models infer the latent representations with a transformer encoder,\nwhich...
2203.07586
2107.12708
2203.07586_2107.12708
[ " Keeping the performance of language technologies optimal as time passes is of\ngreat practical interest. We study temporal effects on model performance on\ndownstream language tasks, establishing a nuanced terminology for such\ndiscussion and identifying factors essential to conduct a robust study. We\npresent e...
2111.12790
2205.12771
2111.12790_2205.12771
[ " Unsupervised sentence representation learning is one of the fundamental\nproblems in natural language processing with various downstream applications.\nRecently, contrastive learning has been widely adopted which derives\nhigh-quality sentence representations by pulling similar semantics closer and\npushing diss...
2305.16726
2204.12811
2305.16726_2204.12811
[ " Importance: Social determinants of health (SDOH) are known to be associated\nwith increased risk of suicidal behaviors, but few studies utilized SDOH from\nunstructured electronic health record (EHR) notes.\n Objective: To investigate associations between suicide and recent SDOH,\nidentified using structured an...
2212.05546
2210.17027
2212.05546_2210.17027
[ " Neuron analysis provides insights into how knowledge is structured in\nrepresentations and discovers the role of neurons in the network. In addition\nto developing an understanding of our models, neuron analysis enables various\napplications such as debiasing, domain adaptation and architectural search. We\npres...
2305.17073
2209.07084
2305.17073_2209.07084
[ " We present a new fact-checking benchmark, Check-COVID, that requires systems\nto verify claims about COVID-19 from news using evidence from scientific\narticles. This approach to fact-checking is particularly challenging as it\nrequires checking internet text written in everyday language against evidence\nfrom j...
2305.18265
2303.10311
2305.18265_2303.10311
[ " Language models have become very popular recently and many claims have been\nmade about their abilities, including for commonsense reasoning. Given the\nincreasingly better results of current language models on previous static\nbenchmarks for commonsense reasoning, we explore an alternative dialectical\nevaluati...
2304.11164
2210.08817
2304.11164_2210.08817
[ " Few-shot named entity recognition (NER) targets generalizing to unseen labels\nand/or domains with few labeled examples. Existing metric learning methods\ncompute token-level similarities between query and support sets, but are not\nable to fully incorporate label semantics into modeling. To address this issue,\...
2211.04337
2204.03035
2211.04337_2204.03035
[ " The Natural Language for Optimization (NL4Opt) Competition was created to\ninvestigate methods of extracting the meaning and formulation of an\noptimization problem based on its text description. Specifically, the goal of\nthe competition is to increase the accessibility and usability of optimization\nsolvers by...
2303.08233
2204.05610
2303.08233_2204.05610
[ " Recent approaches of computer vision utilize deep learning methods as they\nperform quite well if training and testing domains follow the same underlying\ndata distribution. However, it has been shown that minor variations in the\nimages that occur when using these methods in the real world can lead to\nunpredic...
2201.11794
2202.12205
2201.11794_2202.12205
[ " Large pretrained Transformer language models have been shown to exhibit\nzero-shot generalization, i.e. they can perform a wide variety of tasks that\nthey were not explicitly trained on. However, the architectures and pretraining\nobjectives used across state-of-the-art models differ significantly, and there\nh...
2204.05832
2105.08481
2204.05832_2105.08481
[ " Large language models generate fluent texts and can follow natural language\ninstructions to solve a wide range of tasks without task-specific training.\nNevertheless, it is notoriously difficult to control their generation to\nsatisfy the various constraints required by different applications. In this\nwork, we...
2304.14293
2209.09480
2304.14293_2209.09480
[ " A significant share of political discourse occurs online on social media\nplatforms. Policymakers and researchers try to understand the role of social\nmedia design in shaping the quality of political discourse around the globe. In\nthe past decades, scholarship on political discourse theory has produced\ndistin...
2302.09540
2110.08345
2302.09540_2110.08345
[ " Finding word boundaries in continuous speech is challenging as there is\nlittle or no equivalent of a 'space' delimiter between words. Popular Bayesian\nnon-parametric models for text segmentation use a Dirichlet process to jointly\nsegment sentences and build a lexicon of word types. We introduce DP-Parse,\nwhi...
2206.11332
2203.09148
2206.11332_2203.09148
[ " Chain-of-Thought (CoT) prompting can dramatically improve the multi-step\nreasoning abilities of large language models (LLMs). CoT explicitly encourages\nthe LLM to generate intermediate rationales for solving a problem, by providing\na series of reasoning steps in the demonstrations. Despite its success, there\...
2212.10001
2205.06072
2212.10001_2205.06072
[ " This paper discusses OpenAIs ChatGPT, a generative pre-trained transformer,\nwhich uses natural language processing to fulfill text-based user requests\n(i.e., a chatbot). The history and principles behind ChatGPT and similar models\nare discussed. This technology is then discussed in relation to its potential\n...
2303.13367
2109.02707
2303.13367_2109.02707
[ " Text-to-image models offer unprecedented freedom to guide creation through\nnatural language. Yet, it is unclear how such freedom can be exercised to\ngenerate images of specific unique concepts, modify their appearance, or\ncompose them in new roles and novel scenes. In other words, we ask: how can we\nuse lang...
2208.01618
2205.00241
2208.01618_2205.00241
[ " Synthesizing QA pairs with a question generator (QG) on the target domain has\nbecome a popular approach for domain adaptation of question answering (QA)\nmodels. Since synthetic questions are often noisy in practice, existing work\nadapts scores from a pretrained QA (or QG) model as criteria to select\nhigh-qua...
2203.08926
2305.11411
2203.08926_2305.11411
[ " Social media platforms have transformed traditional communication methods by\nallowing users worldwide to communicate instantly, openly, and frequently.\nPeople use social media to express their opinion and share their personal\nstories and struggles. Negative feelings that express hardship, thoughts of\ndeath, ...
2201.10515
2203.02966
2201.10515_2203.02966
[ " Fine-tuning large language models for different tasks can be costly and\ninefficient, and even methods that reduce the number of tuned parameters still\nrequire full gradient-based optimization. We propose HyperTuning, a novel\napproach to model adaptation that uses a hypermodel to generate task-specific\nparame...
2211.12485
2212.14518
2211.12485_2212.14518
[ " This paper introduces Doc2Bot, a novel dataset for building machines that\nhelp users seek information via conversations. This is of particular interest\nfor companies and organizations that own a large number of manuals or\ninstruction books. Despite its potential, the nature of our task poses several\nchalleng...
2210.11060
1708.08615
2210.11060_1708.08615
[ " Entity alignment (EA) is a fundamental data integration task that identifies\nequivalent entities between different knowledge graphs (KGs). Temporal\nKnowledge graphs (TKGs) extend traditional knowledge graphs by introducing\ntimestamps, which have received increasing attention. State-of-the-art\ntime-aware EA s...
2302.00796
2305.11442
2302.00796_2305.11442
[ " In this paper, we propose a novel method based on character\nsequence-to-sequence models to correct documents already processed with Optical\nCharacter Recognition (OCR) systems. The main contribution of this paper is a\nset of strategies to accurately process strings much longer than the ones used\nto train the...
2109.06264
2305.16504
2109.06264_2305.16504
[ " Multi-lingual speech recognition aims to distinguish linguistic expressions\nin different languages and integrate acoustic processing simultaneously. In\ncontrast, current multi-lingual speech recognition research follows a\nlanguage-aware paradigm, mainly targeted to improve recognition performance\nrather than...
2302.13750
2110.08352
2302.13750_2110.08352
[ " The massive amounts of digitized historical documents acquired over the last\ndecades naturally lend themselves to automatic processing and exploration.\nResearch work seeking to automatically process facsimiles and extract\ninformation thereby are multiplying with, as a first essential step, document\nlayout an...
2002.06144
2305.01795
2002.06144_2305.01795
[ " A method to perform offline and online speaker diarization for an unlimited\nnumber of speakers is described in this paper. End-to-end neural diarization\n(EEND) has achieved overlap-aware speaker diarization by formulating it as a\nmulti-label classification problem. It has also been extended for a flexible\nnu...
2206.02432
2302.11989
2206.02432_2302.11989
[ " Generative modeling has been the dominant approach for large-scale\npretraining and zero-shot generalization. In this work, we challenge this\nconvention by showing that discriminative approaches perform substantially\nbetter than generative ones on a large number of NLP tasks. Technically, we\ntrain a single di...
2211.08099
2211.02519
2211.08099_2211.02519
[ " Using prompts to explore the knowledge contained within pre-trained language\nmodels for downstream tasks has now become an active topic. Current prompt\ntuning methods mostly convert the downstream tasks to masked language modeling\nproblems by adding cloze-style phrases and mapping all labels to verbalizations...
2210.12435
2201.05363
2210.12435_2201.05363
[ " Text embeddings are useful features in many applications such as semantic\nsearch and computing text similarity. Previous work typically trains models\ncustomized for different use cases, varying in dataset choice, training\nobjective and model architecture. In this work, we show that contrastive\npre-training o...
2201.10005
2207.14087
2201.10005_2207.14087
[ " Recent works that revealed the vulnerability of dialogue state tracking (DST)\nmodels to distributional shifts have made holistic comparisons on robustness\nand qualitative analyses increasingly important for understanding their\nrelative performance. We present our findings from standardized and\ncomprehensive ...
2112.08321
2306.03078
2112.08321_2306.03078
[ " Construction of human-curated annotated datasets for abstractive text\nsummarization (ATS) is very time-consuming and expensive because creating each\ninstance requires a human annotator to read a long document and compose a\nshorter summary that would preserve the key information relayed by the original\ndocume...
2301.03252
2210.04963
2301.03252_2210.04963
[ " In this paper, we introduce the Tree-of-Thought (ToT) framework, a novel\napproach aimed at improving the problem-solving capabilities of auto-regressive\nlarge language models (LLMs). The ToT technique is inspired by the human mind's\napproach for solving complex reasoning tasks through trial and error. In this...
2305.08291
2305.19280
2305.08291_2305.19280
[ " Factorized layers--operations parameterized by products of two or more\nmatrices--occur in a variety of deep learning contexts, including compressed\nmodel training, certain types of knowledge distillation, and multi-head\nself-attention architectures. We study how to initialize and regularize deep\nnets contain...
2105.01029
2206.13947
2105.01029_2206.13947
[ " The relevance of the Key Information Extraction (KIE) task is increasingly\nimportant in natural language processing problems. But there are still only a\nfew well-defined problems that serve as benchmarks for solutions in this area.\nTo bridge this gap, we introduce two new datasets (Kleister NDA and Kleister\n...
2105.05796
1906.01926
2105.05796_1906.01926
[ " Recent work has shown that language models (LMs) trained with multi-task\n\\textit{instructional learning} (MTIL) can solve diverse NLP tasks in zero- and\nfew-shot settings with improved performance compared to prompt tuning. MTIL\nillustrates that LMs can extract and use information about the task from\ninstru...
2210.11617
2306.00024
2210.11617_2306.00024
[ " Deep neural networks (DNNs) have achieved unprecedented success in the field\nof artificial intelligence (AI), including computer vision, natural language\nprocessing and speech recognition. However, their superior performance comes at\nthe considerable cost of computational complexity, which greatly hinders the...
2204.11786
2212.13196
2204.11786_2212.13196
[ " The Generative Pre-trained Transformer (GPT) represents a notable\nbreakthrough in the domain of natural language processing, which is propelling\nus toward the development of machines that can understand and communicate using\nlanguage in a manner that closely resembles that of humans. GPT is based on the\ntran...
2305.10435
2112.08321
2305.10435_2112.08321
[ " To explain NLP models a popular approach is to use importance measures, such\nas attention, which inform input tokens are important for making a prediction.\nHowever, an open question is how well these explanations accurately reflect a\nmodel's logic, a property called faithfulness.\n To answer this question, w...
2110.08412
2211.11152
2110.08412_2211.11152
[ " There is mounting evidence that existing neural network models, in particular\nthe very popular sequence-to-sequence architecture, struggle to systematically\ngeneralize to unseen compositions of seen components. We demonstrate that one\nof the reasons hindering compositional generalization relates to\nrepresent...
2110.04655
2201.03533
2110.04655_2201.03533
[ " Vision and language navigation (VLN) is a challenging visually-grounded\nlanguage understanding task. Given a natural language navigation instruction, a\nvisual agent interacts with a graph-based environment equipped with panorama\nimages and tries to follow the described route. Most prior work has been\nconduct...
2203.13838
2203.14371
2203.13838_2203.14371
[ " A key trait of daily conversations between individuals is the ability to\nexpress empathy towards others, and exploring ways to implement empathy is a\ncrucial step towards human-like dialogue systems. Previous approaches on this\ntopic mainly focus on detecting and utilizing the user's emotion for generating\ne...
2109.05739
2206.07023
2109.05739_2206.07023
[ " Generative Pre-trained Transformer 4 (GPT-4) is the fourth-generation\nlanguage model in the GPT series, developed by OpenAI, which promises\nsignificant advancements in the field of natural language processing (NLP). In\nthis research article, we have discussed the features of GPT-4, its potential\napplications...
2305.03195
2110.06634
2305.03195_2110.06634
[ " This paper studies the multimedia problem of temporal sentence grounding\n(TSG), which aims to accurately determine the specific video segment in an\nuntrimmed video according to a given sentence query. Traditional TSG methods\nmainly follow the top-down or bottom-up framework and are not end-to-end. They\nsever...
2208.14882
2305.10010
2208.14882_2305.10010
[ " Although the vision-and-language pretraining (VLP) equipped cross-modal\nimage-text retrieval (ITR) has achieved remarkable progress in the past two\nyears, it suffers from a major drawback: the ever-increasing size of VLP models\nrestricts its deployment to real-world search scenarios (where the high latency\ni...
2207.01426
2209.00099
2207.01426_2209.00099
[ " Modern embedding-based metrics for evaluation of generated text generally\nfall into one of two paradigms: discriminative metrics that are trained to\ndirectly predict which outputs are of higher quality according to supervised\nhuman annotations, and generative metrics that are trained to evaluate text\nbased o...
2212.05726
2201.08904
2212.05726_2201.08904
[ " We propose a cross-modal attention distillation framework to train a\ndual-encoder model for vision-language understanding tasks, such as visual\nreasoning and visual question answering. Dual-encoder models have a faster\ninference speed than fusion-encoder models and enable the pre-computation of\nimages and te...
2112.08723
2205.13339
2112.08723_2205.13339
[ " Real-life applications, heavily relying on machine learning, such as dialog\nsystems, demand out-of-domain detection methods. Intent classification models\nshould be equipped with a mechanism to distinguish seen intents from unseen\nones so that the dialog agent is capable of rejecting the latter and avoiding\nu...
2101.03778
2203.12990
2101.03778_2203.12990
[ " Hedges play an important role in the management of conversational\ninteraction. In peer tutoring, they are notably used by tutors in dyads (pairs\nof interlocutors) experiencing low rapport to tone down the impact of\ninstructions and negative feedback. Pursuing the objective of building a\ntutoring agent that m...
2306.14911
2303.07624
2306.14911_2303.07624
[ " Meta-embedding (ME) learning is an emerging approach that attempts to learn\nmore accurate word embeddings given existing (source) word embeddings as the\nsole input.\n Due to their ability to incorporate semantics from multiple source embeddings\nin a compact manner with superior performance, ME learning has g...
2204.11660
2112.09174
2204.11660_2112.09174
[ " Recent advances in generative models for language have enabled the creation\nof convincing synthetic text or deepfake text. Prior work has demonstrated the\npotential for misuse of deepfake text to mislead content consumers. Therefore,\ndeepfake text detection, the task of discriminating between human and\nmachi...
2210.09421
2206.01134
2210.09421_2206.01134
[ " Neuron analysis provides insights into how knowledge is structured in\nrepresentations and discovers the role of neurons in the network. In addition\nto developing an understanding of our models, neuron analysis enables various\napplications such as debiasing, domain adaptation and architectural search. We\npres...
2305.17073
2109.00087
2305.17073_2109.00087
[ " Task-oriented dialogue is often decomposed into three tasks: understanding\nuser input, deciding actions, and generating a response. While such\ndecomposition might suggest a dedicated model for each sub-task, we find a\nsimple, unified approach leads to state-of-the-art performance on the MultiWOZ\ndataset. Sim...
2005.00796
2210.03797
2005.00796_2210.03797
[ " Automatic chart to text summarization is an effective tool for the visually\nimpaired people along with providing precise insights of tabular data in\nnatural language to the user. A large and well-structured dataset is always a\nkey part for data driven models. In this paper, we propose ChartSumm: a\nlarge-scal...
2304.13620
2203.01976
2304.13620_2203.01976
[ " Large Language Models (LLMs) are known to memorize significant portions of\ntheir training data. Parts of this memorized content have been shown to be\nextractable by simply querying the model, which poses a privacy risk. We\npresent a novel approach which uses prompt-tuning to control the extraction\nrates of m...
2305.11759
2205.12854
2305.11759_2205.12854
[ " State-sponsored trolls are the main actors of influence campaigns on social\nmedia and automatic troll detection is important to combat misinformation at\nscale. Existing troll detection models are developed based on training data for\nknown campaigns (e.g.\\ the influence campaign by Russia's Internet Research\...
2303.07354
2212.13036
2303.07354_2212.13036
[ " As text generated by large language models proliferates, it becomes vital to\nunderstand how humans engage with such text, and whether or not they are able\nto detect when the text they are reading did not originate with a human writer.\nPrior work on human detection of generated text focuses on the case where a...
2212.12672
2106.14361
2212.12672_2106.14361
[ " Despite tremendous progress in automatic summarization, state-of-the-art\nmethods are predominantly trained to excel in summarizing short newswire\narticles, or documents with strong layout biases such as scientific articles or\ngovernment reports. Efficient techniques to summarize financial documents,\nincludin...
2210.12467
2110.13900
2210.12467_2110.13900
[ " This paper focuses on automatically generating the text of an ad, and the\ngoal is that the generated text can capture user interest for achieving higher\nclick-through rate (CTR). We propose CREATER, a CTR-driven advertising text\ngeneration approach, to generate ad texts based on high-quality user reviews.\nTo...
2205.08943
2203.07627
2205.08943_2203.07627
[ " Hope is characterized as openness of spirit toward the future, a desire,\nexpectation, and wish for something to happen or to be true that remarkably\naffects human's state of mind, emotions, behaviors, and decisions. Hope is\nusually associated with concepts of desired expectations and\npossibility/probability ...
2210.14136
2205.00034
2210.14136_2205.00034
[ " Class-based language models (LMs) have been long devised to address context\nsparsity in $n$-gram LMs. In this study, we revisit this approach in the\ncontext of neural LMs. We hypothesize that class-based prediction leads to an\nimplicit context aggregation for similar words and thus can improve\ngeneralization...
2203.10692
2202.07543
2203.10692_2202.07543
[ " Automated event detection from news corpora is a crucial task towards mining\nfast-evolving structured knowledge. As real-world events have different\ngranularities, from the top-level themes to key events and then to event\nmentions corresponding to concrete actions, there are generally two lines of\nresearch: ...
2206.04153
2212.06121
2206.04153_2212.06121
[ " The transducer architecture is becoming increasingly popular in the field of\nspeech recognition, because it is naturally streaming as well as high in\naccuracy. One of the drawbacks of transducer is that it is difficult to decode\nin a fast and parallel way due to an unconstrained number of symbols that can\nbe...
2211.00484
2205.02564
2211.00484_2205.02564
[ " Robustness evaluation against adversarial examples has become increasingly\nimportant to unveil the trustworthiness of the prevailing deep models in\nnatural language processing (NLP). However, in contrast to the computer vision\ndomain where the first-order projected gradient descent (PGD) is used as the\nbench...
2212.09254
2205.14140
2212.09254_2205.14140
[ " Sustaining coherent and engaging narratives requires dialogue or storytelling\nagents to understand how the personas of speakers or listeners ground the\nnarrative. Specifically, these agents must infer personas of their listeners to\nproduce statements that cater to their interests. They must also learn to\nmai...
2305.02364
2208.12995
2305.02364_2208.12995
[ " We apply transfer learning to the task of phoneme segmentation and\ndemonstrate the utility of representations learned in self-supervised\npre-training for the task. Our model extends transformer-style encoders with\nstrategically placed convolutions that manipulate features learned in\npre-training. Using the T...
2211.01461
2202.05451
2211.01461_2202.05451