abstracts list | id_1 string | id_2 string | pair_id string |
|---|---|---|---|
[
" Large Language Models (LLMs, e.g., ChatGPT) have shown impressive zero- and\nfew-shot capabilities in Named Entity Recognition (NER). However, these models\ncan only be accessed via online APIs, which may cause data leak and\nnon-reproducible problems. In this paper, we propose VicunaNER, a zero/few-shot\nNER fr... | 2305.03253 | 2211.10511 | 2305.03253_2211.10511 |
[
" As one of the challenging NLP tasks, designing math word problem (MWP)\nsolvers has attracted increasing research attention for the past few years. In\nprevious work, models designed by taking into account the properties of the\nbinary tree structure of mathematical expressions at the output side have\nachieved ... | 2210.12432 | 2112.15471 | 2210.12432_2112.15471 |
[
" How to learn a universal facial representation that boosts all face analysis\ntasks? This paper takes one step toward this goal. In this paper, we study the\ntransfer performance of pre-trained models on face analysis tasks and introduce\na framework, called FaRL, for general Facial Representation Learning in a\... | 2112.03109 | 2105.03949 | 2112.03109_2105.03949 |
[
" Predicting stock prices from textual information is a challenging task due to\nthe uncertainty of the market and the difficulty understanding the natural\nlanguage from a machine's perspective. Previous researches focus mostly on\nsentiment extraction based on single news. However, the stocks on the financial\nm... | 2107.10941 | 2305.17678 | 2107.10941_2305.17678 |
[
" In this work, we pursue a unified paradigm for multimodal pretraining to\nbreak the scaffolds of complex task/modality-specific customization. We propose\nOFA, a Task-Agnostic and Modality-Agnostic framework that supports Task\nComprehensiveness. OFA unifies a diverse set of cross-modal and unimodal tasks,\nincl... | 2202.03052 | 2304.04616 | 2202.03052_2304.04616 |
[
" Transformers are the most eminent architectures used for a vast range of\nNatural Language Processing tasks. These models are pre-trained over a large\ntext corpus and are meant to serve state-of-the-art results over tasks like\ntext classification. In this work, we conduct a comparative study between\nmonolingu... | 2204.08669 | 2204.09595 | 2204.08669_2204.09595 |
[
" As generic machine translation (MT) quality has improved, the need for\ntargeted benchmarks that explore fine-grained aspects of quality has increased.\nIn particular, gender accuracy in translation can have implications in terms of\noutput fluency, translation accuracy, and ethics. In this paper, we introduce\n... | 2211.01355 | 2110.07524 | 2211.01355_2110.07524 |
[
" Writing is, by nature, a strategic, adaptive, and more importantly, an\niterative process. A crucial part of writing is editing and revising the text.\nPrevious works on text revision have focused on defining edit intention\ntaxonomies within a single domain or developing computational models with a\nsingle leve... | 2203.03802 | 2203.07523 | 2203.03802_2203.07523 |
[
" In recent years, hate speech has gained great relevance in social networks\nand other virtual media because of its intensity and its relationship with\nviolent acts against members of protected groups. Due to the great amount of\ncontent generated by users, great effort has been made in the research and\ndevelop... | 2210.00465 | 2210.11359 | 2210.00465_2210.11359 |
[
" In this paper, we develop a new multi-singer Chinese neural singing voice\nsynthesis (SVS) system named WeSinger. To improve the accuracy and naturalness\nof synthesized singing voice, we design several specifical modules and\ntechniques: 1) A deep bi-directional LSTM-based duration model with multi-scale\nrhyth... | 2203.10750 | 2208.00635 | 2203.10750_2208.00635 |
[
" In recent years, monitoring hate speech and offensive language on social\nmedia platforms has become paramount due to its widespread usage among all age\ngroups, races, and ethnicities. Consequently, there have been substantial\nresearch efforts towards automated detection of such content using Natural\nLanguage... | 2202.06218 | 2203.10254 | 2202.06218_2203.10254 |
[
" In recent years, multilingual machine translation models have achieved\npromising performance on low-resource language pairs by sharing information\nbetween similar languages, thus enabling zero-shot translation. To overcome the\n\"curse of multilinguality\", these models often opt for scaling up the number of\n... | 2210.11621 | 1905.09153 | 2210.11621_1905.09153 |
[
" We propose a simple yet effective embedding model to learn quaternion\nembeddings for entities and relations in knowledge graphs. Our model aims to\nenhance correlations between head and tail entities given a relation within the\nQuaternion space with Hamilton product. The model achieves this goal by further\nas... | 2009.12517 | 2205.09501 | 2009.12517_2205.09501 |
[
" The task of multi-document summarization (MDS) aims at models that, given\nmultiple documents as input, are able to generate a summary that combines\ndisperse information, originally spread across these documents. Accordingly, it\nis expected that both reference summaries in MDS datasets, as well as system\nsumm... | 2210.12688 | 2306.05406 | 2210.12688_2306.05406 |
[
" The emotion recognition in conversation (ERC) task aims to predict the\nemotion label of an utterance in a conversation. Since the dependencies between\nspeakers are complex and dynamic, which consist of intra- and inter-speaker\ndependencies, the modeling of speaker-specific information is a vital role in\nERC.... | 2206.03173 | 2204.06889 | 2206.03173_2204.06889 |
[
" Gaze behaviour has been used as a way to gather cognitive information for a\nnumber of years. In this paper, we discuss the use of gaze behaviour in solving\ndifferent tasks in natural language processing (NLP) without having to record\nit at test time. This is because the collection of gaze behaviour is a costl... | 2112.15471 | 2202.07959 | 2112.15471_2202.07959 |
[
" Speaker diarization (SD) is typically used with an automatic speech\nrecognition (ASR) system to ascribe speaker labels to recognized words. The\nconventional approach reconciles outputs from independently optimized ASR and\nSD systems, where the SD system typically uses only acoustic information to\nidentify th... | 2306.09313 | 2210.06432 | 2306.09313_2210.06432 |
[
" Being able to rank the similarity of short text segments is an interesting\nbonus feature of neural machine translation. Translation-based similarity\nmeasures include direct and pivot translation probability, as well as\ntranslation cross-likelihood, which has not been studied so far. We analyze\nthese measures... | 2204.13692 | 2304.02015 | 2204.13692_2304.02015 |
[
" Automatically generating compilable programs with (or without) natural\nlanguage descriptions has always been a touchstone problem for computational\nlinguistics and automated software engineering. Existing deep-learning\napproaches model code generation as text generation, either constrained by\ngrammar structu... | 2203.05132 | 2201.09745 | 2203.05132_2201.09745 |
[
" Variation in speech is often quantified by comparing phonetic transcriptions\nof the same utterance. However, manually transcribing speech is time-consuming\nand error prone. As an alternative, therefore, we investigate the extraction of\nacoustic embeddings from several self-supervised neural models. We use the... | 2011.12649 | 2203.10430 | 2011.12649_2203.10430 |
[
" Contextualized word embeddings have demonstrated state-of-the-art performance\nin various natural language processing tasks including those that concern\nhistorical semantic change. However, language models such as BERT was trained\nprimarily on contemporary corpus data. To investigate whether training on\nhisto... | 2202.03612 | 2302.14035 | 2202.03612_2302.14035 |
[
" Multi-encoder models are a broad family of context-aware neural machine\ntranslation systems that aim to improve translation quality by encoding\ndocument-level contextual information alongside the current sentence. The\ncontext encoding is undertaken by contextual parameters, trained on\ndocument-level data. In... | 2103.17151 | 2304.02468 | 2103.17151_2304.02468 |
[
" In this paper, we formulate a novel task to synthesize speech in sync with a\nsilent pre-recorded video, denoted as automatic voice over (AVO). Unlike\ntraditional speech synthesis, AVO seeks to generate not only human-sounding\nspeech, but also perfect lip-speech synchronization. A natural solution to AVO\nis t... | 2110.03342 | 2109.11087 | 2110.03342_2109.11087 |
[
" Scientific document classification is a critical task for a wide range of\napplications, but the cost of obtaining massive amounts of human-labeled data\ncan be prohibitive. To address this challenge, we propose a weakly-supervised\napproach for scientific document classification using label names only. In\nscie... | 2306.07193 | 2202.07962 | 2306.07193_2202.07962 |
[
" A Large Language Model (LLM) is an artificial intelligence system that has\nbeen trained on vast amounts of natural language data, enabling it to generate\nhuman-like responses to written or spoken language input. GPT-3.5 is an example\nof an LLM that supports a conversational agent called ChatGPT. In this work,... | 2305.04400 | 2202.03543 | 2305.04400_2202.03543 |
[
" Here we consider some well-known facts in syntax from a physics perspective,\nallowing us to establish equivalences between both fields with many\nconsequences. Mainly, we observe that the operation MERGE, put forward by N.\nChomsky in 1995, can be interpreted as a physical information coarse-graining.\nThus, ME... | 1708.01525 | 2206.14774 | 1708.01525_2206.14774 |
[
" Today's VQA models still tend to capture superficial linguistic correlations\nin the training set and fail to generalize to the test set with different QA\ndistributions. To reduce these language biases, recent VQA works introduce an\nauxiliary question-only model to regularize the training of targeted VQA model... | 2110.01013 | 2203.05325 | 2110.01013_2203.05325 |
[
" Transformer models can use two fundamentally different kinds of information:\ninformation stored in weights during training, and information provided\n``in-context'' at inference time. In this work, we show that transformers\nexhibit different inductive biases in how they represent and generalize from\nthe infor... | 2210.05675 | 2305.04676 | 2210.05675_2305.04676 |
[
" Semi-parametric models, which augment generation with retrieval, have led to\nimpressive results in language modeling and machine translation, due to their\nability to retrieve fine-grained information from a datastore of examples. One\nof the most prominent approaches, $k$NN-MT, exhibits strong domain adaptatio... | 2205.12230 | 2302.07727 | 2205.12230_2302.07727 |
[
" Extractive Question Answering (EQA) is one of the most important tasks in\nMachine Reading Comprehension (MRC), which can be solved by fine-tuning the\nspan selecting heads of Pre-trained Language Models (PLMs). However, most\nexisting approaches for MRC may perform poorly in the few-shot learning\nscenario. To ... | 2205.03071 | 2212.11311 | 2205.03071_2212.11311 |
[
" In this paper, we present an end-to-end joint entity and relation extraction\napproach based on transformer-based language models. We apply the model to the\ntask of linking mathematical symbols to their descriptions in LaTeX documents.\nIn contrast to existing approaches, which perform entity and relation\nextr... | 2203.05325 | 2102.01373 | 2203.05325_2102.01373 |
[
" Building on current work on multilingual hate speech (e.g., Ousidhoum et al.\n(2019)) and hate speech reduction (e.g., Sap et al. (2020)), we present\nXTREMESPEECH, a new hate speech dataset containing 20,297 social media passages\nfrom Brazil, Germany, India and Kenya. The key novelty is that we directly\ninvol... | 2203.11764 | 2110.14566 | 2203.11764_2110.14566 |
[
" We launch EVA, a vision-centric foundation model to explore the limits of\nvisual representation at scale using only publicly accessible data. EVA is a\nvanilla ViT pre-trained to reconstruct the masked out image-text aligned vision\nfeatures conditioned on visible image patches. Via this pretext task, we can\ne... | 2211.07636 | 2204.00885 | 2211.07636_2204.00885 |
[
" Large repositories of products, patents and scientific papers offer an\nopportunity for building systems that scour millions of ideas and help users\ndiscover inspirations. However, idea descriptions are typically in the form of\nunstructured text, lacking key structure that is required for supporting\ncreative ... | 2102.09761 | 2306.05642 | 2102.09761_2306.05642 |
[
" A common approach for testing fairness issues in text-based classifiers is\nthrough the use of counterfactuals: does the classifier output change if a\nsensitive attribute in the input is changed? Existing counterfactual generation\nmethods typically rely on wordlists or templates, producing simple\ncounterfactu... | 2206.13757 | 2205.03966 | 2206.13757_2205.03966 |
[
" Dialogue contradiction is a critical issue in open-domain dialogue systems.\nThe contextualization nature of conversations makes dialogue contradiction\ndetection rather challenging. In this work, we propose a benchmark for\nContradiction Detection in Chinese Conversations, namely CDConv. It contains\n12K multi-... | 2210.08511 | 2204.09817 | 2210.08511_2204.09817 |
[
" We present a new framework AMOS that pretrains text encoders with an\nAdversarial learning curriculum via a Mixture Of Signals from multiple\nauxiliary generators. Following ELECTRA-style pretraining, the main encoder is\ntrained as a discriminator to detect replaced tokens generated by auxiliary\nmasked languag... | 2204.03243 | 2211.00792 | 2204.03243_2211.00792 |
[
" State-of-the-art keyphrase generation methods generally depend on large\nannotated datasets, limiting their performance in domains with limited\nannotated data. To overcome this challenge, we design a data-oriented approach\nthat first identifies salient information using retrieval-based corpus-level\nstatistics... | 2203.08118 | 2302.04459 | 2203.08118_2302.04459 |
[
" In order to interpret the communicative intents of an utterance, it needs to\nbe grounded in something that is outside of language; that is, grounded in\nworld modalities. In this paper, we argue that dialogue clarification\nmechanisms make explicit the process of interpreting the communicative intents\nof the s... | 2104.08964 | 2208.10806 | 2104.08964_2208.10806 |
[
" Recent research has made impressive progress in large-scale multimodal\npre-training. In the context of the rapid growth of model size, it is necessary\nto seek efficient and flexible methods other than finetuning. In this paper, we\npropose to use prompt vectors to align the modalities. Our method achieves\ncom... | 2203.08055 | 2304.10145 | 2203.08055_2304.10145 |
[
" Although existing neural network approaches have achieved great success on\nChinese spelling correction, there is still room to improve. The model is\nrequired to avoid over-correction and to distinguish a correct token from its\nphonological and visually similar ones. In this paper, we propose an\nerror-guided ... | 2301.06323 | 2209.02970 | 2301.06323_2209.02970 |
[
" Sequential sentence classification deals with the categorisation of sentences\nbased on their content and context. Applied to scientific texts, it enables the\nautomatic structuring of research papers and the improvement of academic search\nengines. However, previous work has not investigated the potential of tr... | 2102.06008 | 2211.12707 | 2102.06008_2211.12707 |
[
" As language models grow in popularity, it becomes increasingly important to\nclearly measure all possible markers of demographic identity in order to avoid\nperpetuating existing societal harms. Many datasets for measuring bias\ncurrently exist, but they are restricted in their coverage of demographic axes\nand ... | 2205.09209 | 2208.10244 | 2205.09209_2208.10244 |
[
" Sentiment analysis (SA) is the automated process of detecting and\nunderstanding the emotions conveyed through written text. Over the past decade,\nSA has gained significant popularity in the field of Natural Language\nProcessing (NLP). With the widespread use of social media and online platforms,\nSA has become... | 2305.14842 | 2211.01267 | 2305.14842_2211.01267 |
[
" Incorporating factual knowledge into pre-trained language models (PLM) such\nas BERT is an emerging trend in recent NLP studies. However, most of the\nexisting methods combine the external knowledge integration module with a\nmodified pre-training loss and re-implement the pre-training process on the\nlarge-scal... | 2109.04223 | 2204.03930 | 2109.04223_2204.03930 |
[
" Automatic Speech Recognition (ASR) systems are known to exhibit difficulties\nwhen transcribing children's speech. This can mainly be attributed to the\nabsence of large children's speech corpora to train robust ASR models and the\nresulting domain mismatch when decoding children's speech with systems trained\no... | 2206.09396 | 2110.08443 | 2206.09396_2110.08443 |
[
" The punctuation restoration task aims to correctly punctuate the output\ntranscriptions of automatic speech recognition systems. Previous punctuation\nmodels, either using text only or demanding the corresponding audio, tend to be\nconstrained by real scenes, where unpunctuated sentences are a mixture of those\n... | 2202.00468 | 2305.10703 | 2202.00468_2305.10703 |
[
" On June 24, 2018, Turkey conducted a highly consequential election in which\nthe Turkish people elected their president and parliament in the first election\nunder a new presidential system. During the election period, the Turkish people\nextensively shared their political opinions on Twitter. One aspect of\npol... | 2005.09649 | 2212.03657 | 2005.09649_2212.03657 |
[
" In this paper we present statistical analysis of English texts from\nWikipedia. We try to address the issue of language complexity empirically by\ncomparing the simple English Wikipedia (Simple) to comparable samples of the\nmain English Wikipedia (Main). Simple is supposed to use a more simplified\nlanguage wit... | 1204.2765 | 2109.03564 | 1204.2765_2109.03564 |
[
" Sign language recognition and translation first uses a recognition module to\ngenerate glosses from sign language videos and then employs a translation\nmodule to translate glosses into spoken sentences. Most existing works focus on\nthe recognition step, while paying less attention to sign language translation.... | 2204.05953 | 2202.01709 | 2204.05953_2202.01709 |
[
" Unsupervised domain adaptation (UDA) is the task of modifying a statistical\nmodel trained on labeled data from a source domain to achieve better\nperformance on data from a target domain, with access to only unlabeled data in\nthe target domain. Existing state-of-the-art UDA approaches use neural networks\nto l... | 1905.09153 | 2110.03215 | 1905.09153_2110.03215 |
[
" Despite exciting progress in causal language models, the expressiveness of\nthe representations is largely limited due to poor discrimination ability. To\nremedy this issue, we present ContraCLM, a novel contrastive learning framework\nat both token-level and sequence-level. We assess ContraCLM on a variety of\n... | 2210.01185 | 2304.10946 | 2210.01185_2304.10946 |
[
" We present PanGu-Coder, a pretrained decoder-only language model adopting the\nPanGu-Alpha architecture for text-to-code generation, i.e. the synthesis of\nprogramming language solutions given a natural language problem description. We\ntrain PanGu-Coder using a two-stage strategy: the first stage employs Causal... | 2207.11280 | 2305.14233 | 2207.11280_2305.14233 |
[
" Text classification aims to assign labels to textual units by making use of\nglobal information. Recent studies have applied graph neural network (GNN) to\ncapture the global word co-occurrence in a corpus. Existing approaches require\nthat all the nodes (training and test) in a graph are present during training... | 2206.00265 | 2201.08808 | 2206.00265_2201.08808 |
[
" The detection of hate speech in political discourse is a critical issue, and\nthis becomes even more challenging in low-resource languages. To address this\nissue, we introduce a new dataset named IEHate, which contains 11,457 manually\nannotated Hindi tweets related to the Indian Assembly Election Campaign from... | 2306.14764 | 2110.06263 | 2306.14764_2110.06263 |
[
" Embodied control requires agents to leverage multi-modal pre-training to\nquickly learn how to act in new environments, where video demonstrations\ncontain visual and motion details needed for low-level perception and control,\nand language instructions support generalization with abstract, symbolic\nstructures.... | 2304.09448 | 2203.17152 | 2304.09448_2203.17152 |
[
" We propose PolyVoice, a language model-based framework for speech-to-speech\ntranslation (S2ST) system. Our framework consists of two language models: a\ntranslation language model and a speech synthesis language model. We use\ndiscretized speech units, which are generated in a fully unsupervised way, and\nthus ... | 2306.02982 | 1204.2765 | 2306.02982_1204.2765 |
[
" Large transformer-based models are able to perform in-context few-shot\nlearning, without being explicitly trained for it. This observation raises the\nquestion: what aspects of the training regime lead to this emergent behavior?\nHere, we show that this behavior is driven by the distributions of the training\nd... | 2205.05055 | 2209.15323 | 2205.05055_2209.15323 |
[
" In this paper, we develop a new multi-singer Chinese neural singing voice\nsynthesis (SVS) system named WeSinger. To improve the accuracy and naturalness\nof synthesized singing voice, we design several specifical modules and\ntechniques: 1) A deep bi-directional LSTM-based duration model with multi-scale\nrhyth... | 2203.10750 | 2304.09333 | 2203.10750_2304.09333 |
[
" Variation in speech is often quantified by comparing phonetic transcriptions\nof the same utterance. However, manually transcribing speech is time-consuming\nand error prone. As an alternative, therefore, we investigate the extraction of\nacoustic embeddings from several self-supervised neural models. We use the... | 2011.12649 | 2112.07210 | 2011.12649_2112.07210 |
[
" Training data attribution (TDA) methods offer to trace a model's prediction\non any given example back to specific influential training examples. Existing\napproaches do so by assigning a scalar influence score to each training\nexample, under a simplifying assumption that influence is additive. But in\nreality,... | 2303.08114 | 2204.02261 | 2303.08114_2204.02261 |
[
" This paper shows how to use large-scale pre-trained language models to\nextract character roles from narrative texts without training data. Queried\nwith a zero-shot question-answering prompt, GPT-3 can identify the hero,\nvillain, and victim in diverse domains: newspaper articles, movie plot\nsummaries, and pol... | 2205.07557 | 2301.01181 | 2205.07557_2301.01181 |
[
" With the ever-growing size of pretrained models (PMs), fine-tuning them has\nbecome more expensive and resource-hungry. As a remedy, low-rank adapters\n(LoRA) keep the main pretrained weights of the model frozen and just introduce\nsome learnable truncated SVD modules (so-called LoRA blocks) to the model.\nWhile... | 2210.07558 | 2305.09137 | 2210.07558_2305.09137 |
[
" Existing large language models (LLMs) can only afford fix-sized inputs due to\nthe input length limit, preventing them from utilizing rich long-context\ninformation from past inputs. To address this, we propose a framework, Language\nModels Augmented with Long-Term Memory (LongMem), which enables LLMs to\nmemori... | 2306.07174 | 2205.15485 | 2306.07174_2205.15485 |
[
" With a growing number of BERTology work analyzing different components of\npre-trained language models, we extend this line of research through an\nin-depth analysis of discourse information in pre-trained and fine-tuned\nlanguage models. We move beyond prior work along three dimensions: First, we\ndescribe a no... | 2204.04289 | 2304.07849 | 2204.04289_2304.07849 |
[
" Pre-trained Transformers currently dominate most NLP tasks. They impose,\nhowever, limits on the maximum input length (512 sub-words in BERT), which are\ntoo restrictive in the legal domain. Even sparse-attention models, such as\nLongformer and BigBird, which increase the maximum input length to 4,096\nsub-words... | 2211.00974 | 2206.05802 | 2211.00974_2206.05802 |
[
" Advances in Large Language Models (LLMs) have led to remarkable capabilities,\nyet their inner mechanisms remain largely unknown. To understand these models,\nwe need to unravel the functions of individual neurons and their contribution\nto the network. This paper introduces a novel automated approach designed t... | 2305.19911 | 2110.14532 | 2305.19911_2110.14532 |
[
" Radiology report generation (RRG) aims to describe automatically a radiology\nimage with human-like language and could potentially support the work of\nradiologists, reducing the burden of manual reporting. Previous approaches\noften adopt an encoder-decoder architecture and focus on single-modal feature\nlearni... | 2207.04818 | 2301.02427 | 2207.04818_2301.02427 |
[
" Sequence modeling has important applications in natural language processing\nand computer vision. Recently, the transformer-based models have shown strong\nperformance on various sequence modeling tasks, which rely on attention to\ncapture pairwise token relations, and position embedding to inject positional\nin... | 2305.04749 | 2305.16157 | 2305.04749_2305.16157 |
[
" Training dense passage representations via contrastive learning has been\nshown effective for Open-Domain Passage Retrieval (ODPR). Existing studies\nfocus on further optimizing by improving negative sampling strategy or extra\npretraining. However, these studies keep unknown in capturing passage with\ninternal ... | 2110.07524 | 2305.16885 | 2110.07524_2305.16885 |
[
" Although the problem of hallucinations in neural machine translation (NMT)\nhas received some attention, research on this highly pathological phenomenon\nlacks solid ground. Previous work has been limited in several ways: it often\nresorts to artificial settings where the problem is amplified, it disregards\nsom... | 2208.05309 | 2209.07430 | 2208.05309_2209.07430 |
[
" Dual encoders and cross encoders have been widely used for image-text\nretrieval. Between the two, the dual encoder encodes the image and text\nindependently followed by a dot product, while the cross encoder jointly feeds\nimage and text as the input and performs dense multi-modal fusion. These two\narchitectur... | 2203.05465 | 2108.13556 | 2203.05465_2108.13556 |
[
" Today's VQA models still tend to capture superficial linguistic correlations\nin the training set and fail to generalize to the test set with different QA\ndistributions. To reduce these language biases, recent VQA works introduce an\nauxiliary question-only model to regularize the training of targeted VQA model... | 2110.01013 | 2202.08005 | 2110.01013_2202.08005 |
[
" Coordinated disinformation campaigns are used to influence social media\nusers, potentially leading to offline violence. In this study, we introduce a\ngeneral methodology to uncover coordinated messaging through analysis of user\nparleys on Parler. The proposed method constructs a user-to-user coordination\nnet... | 2109.00945 | 2305.19383 | 2109.00945_2305.19383 |
[
" Gender-neutral pronouns have recently been introduced in many languages to a)\ninclude non-binary people and b) as a generic singular. Recent results from\npsycholinguistics suggest that gender-neutral pronouns (in Swedish) are not\nassociated with human processing difficulties. This, we show, is in sharp\ncontr... | 2204.10281 | 2211.11446 | 2204.10281_2211.11446 |
[
" Changing how pre-trained models behave -- e.g., improving their performance\non a downstream task or mitigating biases learned during pre-training -- is a\ncommon practice when developing machine learning systems. In this work, we\npropose a new paradigm for steering the behavior of neural networks, centered\nar... | 2212.04089 | 2207.01484 | 2212.04089_2207.01484 |
[
" The lack of data and the difficulty of multimodal fusion have always been\nchallenges for multimodal emotion recognition (MER). In this paper, we propose\nto use pretrained models as upstream network, wav2vec 2.0 for audio modality\nand BERT for text modality, and finetune them in downstream task of MER to cope\... | 2302.13661 | 2103.17151 | 2302.13661_2103.17151 |
[
" The amount of labeled data to train models for speech tasks is limited for\nmost languages, however, the data scarcity is exacerbated for speech\ntranslation which requires labeled data covering two different languages. To\naddress this issue, we study a simple and effective approach to build speech\ntranslation... | 2210.10191 | 2202.02113 | 2210.10191_2202.02113 |
[
" Word-level adversarial attacks have shown success in NLP models, drastically\ndecreasing the performance of transformer-based models in recent years. As a\ncountermeasure, adversarial defense has been explored, but relatively few\nefforts have been made to detect adversarial examples. However, detecting\nadversa... | 2203.01677 | 2302.03668 | 2203.01677_2302.03668 |
[
" Neural named entity recognition (NER) models may easily encounter the\nover-confidence issue, which degrades the performance and calibration. Inspired\nby label smoothing and driven by the ambiguity of boundary annotation in NER\nengineering, we propose boundary smoothing as a regularization technique for\nspan-... | 2204.12031 | 2212.10173 | 2204.12031_2212.10173 |
[
" Tweets are the most concise form of communication in online social media,\nwherein a single tweet has the potential to make or break the discourse of the\nconversation. Online hate speech is more accessible than ever, and stifling its\npropagation is of utmost importance for social media companies and users for\... | 2206.08406 | 2110.06306 | 2206.08406_2110.06306 |
[
" We launch EVA, a vision-centric foundation model to explore the limits of\nvisual representation at scale using only publicly accessible data. EVA is a\nvanilla ViT pre-trained to reconstruct the masked out image-text aligned vision\nfeatures conditioned on visible image patches. Via this pretext task, we can\ne... | 2211.07636 | 2205.12811 | 2211.07636_2205.12811 |
[
" Transcripts of teaching episodes can be effective tools to understand\ndiscourse patterns in classroom instruction. According to most educational\nexperts, sustained classroom discourse is a critical component of equitable,\nengaging, and rich learning environments for students. This paper describes the\nTalkMov... | 2204.09652 | 2201.06496 | 2204.09652_2201.06496 |
[
" The multilingual Sentence-BERT (SBERT) models map different languages to\ncommon representation space and are useful for cross-language similarity and\nmining tasks. We propose a simple yet effective approach to convert vanilla\nmultilingual BERT models into multilingual sentence BERT models using synthetic\ncor... | 2304.11434 | 2112.15471 | 2304.11434_2112.15471 |
[
" Large language models (LLMs) exhibit remarkable performance across various\nNLP tasks. However, they often generate incorrect or hallucinated information,\nwhich hinders their practical applicability in real-world scenarios. Human\nfeedback has been shown to effectively enhance the factuality and quality of\ngen... | 2305.14002 | 2208.01006 | 2305.14002_2208.01006 |
[
" Future work sentences (FWS) are the particular sentences in academic papers\nthat contain the author's description of their proposed follow-up research\ndirection. This paper presents methods to automatically extract FWS from\nacademic papers and classify them according to the different future directions\nembodi... | 2212.13860 | 2204.02363 | 2212.13860_2204.02363 |
[
" With careful manipulation, malicious agents can reverse engineer private\ninformation encoded in pre-trained language models. Security concerns motivate\nthe development of quantum pre-training. In this work, we propose a highly\nPortable Quantum Language Model (PQLM) that can easily transmit information to\ndow... | 2210.03221 | 2211.00792 | 2210.03221_2211.00792 |
[
" We present, Naamapadam, the largest publicly available Named Entity\nRecognition (NER) dataset for the 11 major Indian languages from two language\nfamilies. The dataset contains more than 400k sentences annotated with a total\nof at least 100k entities from three standard entity categories (Person,\nLocation, a... | 2212.10168 | 2211.15199 | 2212.10168_2211.15199 |
[
" We review the scholarly contributions that utilise Natural Language\nProcessing (NLP) techniques to support the design process. Using a heuristic\napproach, we gathered 223 articles that are published in 32 journals within the\nperiod 1991-present. We present state-of-the-art NLP in-and-for design research\nby r... | 2111.13827 | 2212.07769 | 2111.13827_2212.07769 |
[
" We take the first step to address the task of automatically generating\nshellcodes, i.e., small pieces of code used as a payload in the exploitation of\na software vulnerability, starting from natural language comments. We assemble\nand release a novel dataset (Shellcode_IA32), consisting of challenging but\ncom... | 2104.13100 | 2102.01065 | 2104.13100_2102.01065 |
[
" Compared to standard retrieval tasks, passage retrieval for conversational\nquestion answering (CQA) poses new challenges in understanding the current user\nquestion, as each question needs to be interpreted within the dialogue context.\nMoreover, it can be expensive to re-train well-established retrievers such ... | 2112.08558 | 2205.09073 | 2112.08558_2205.09073 |
[
" We present JOIST, an algorithm to train a streaming, cascaded, encoder\nend-to-end (E2E) model with both speech-text paired inputs, and text-only\nunpaired inputs. Unlike previous works, we explore joint training with both\nmodalities, rather than pre-training and fine-tuning. In addition, we explore\nJOIST usin... | 2210.07353 | 2110.07477 | 2210.07353_2110.07477 |
[
" Since a vast number of tables can be easily collected from web pages,\nspreadsheets, PDFs, and various other document types, a flurry of table\npre-training frameworks have been proposed following the success of text and\nimages, and they have achieved new state-of-the-arts on various tasks such as\ntable questi... | 2201.09745 | 2202.13093 | 2201.09745_2202.13093 |
[
" Natural Language Processing algorithms have made incredible progress, but\nthey still struggle when applied to out-of-distribution examples. We address a\nchallenging and underexplored version of this domain adaptation problem, where\nan algorithm is trained on several source domains, and then applied to example... | 2102.12206 | 2206.11871 | 2102.12206_2206.11871 |
[
" The increase in people's use of mobile messaging services has led to the\nspread of social engineering attacks like phishing, considering that spam text\nis one of the main factors in the dissemination of phishing attacks to steal\nsensitive data such as credit cards and passwords. In addition, rumors and\nincor... | 2110.15718 | 2210.02592 | 2110.15718_2210.02592 |
[
" Contrastive Vision-Language Pre-training, known as CLIP, has provided a new\nparadigm for learning visual representations using large-scale image-text\npairs. It shows impressive performance on downstream tasks by zero-shot\nknowledge transfer. To further enhance CLIP's adaption capability, existing\nmethods pro... | 2207.09519 | 2210.07353 | 2207.09519_2210.07353 |
[
" Question answering models commonly have access to two sources of \"knowledge\"\nduring inference time: (1) parametric knowledge - the factual knowledge encoded\nin the model weights, and (2) contextual knowledge - external knowledge (e.g.,\na Wikipedia passage) given to the model to generate a grounded answer. H... | 2211.05655 | 2206.08835 | 2211.05655_2206.08835 |
[
" Few-shot named entity recognition (NER) enables us to build a NER system for\na new domain using very few labeled examples. However, existing prototypical\nnetworks for this task suffer from roughly estimated label dependency and\nclosely distributed prototypes, thus often causing misclassifications. To\naddress... | 2208.08023 | 2203.14465 | 2208.08023_2203.14465 |
[
" Continual learning is one of the key components of human learning and a\nnecessary requirement of artificial intelligence. As dialogue can potentially\nspan infinitely many topics and tasks, a task-oriented dialogue system must\nhave the capability to continually learn, dynamically adapting to new\nchallenges wh... | 2204.05928 | 2204.08806 | 2204.05928_2204.08806 |
[
" Modern studies in radiograph representation learning rely on either\nself-supervision to encode invariant semantics or associated radiology reports\nto incorporate medical expertise, while the complementarity between them is\nbarely noticed. To explore this, we formulate the self- and report-completion\nas two c... | 2301.13155 | 2204.04263 | 2301.13155_2204.04263 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.