Natural language understanding (NLU) tasks are typically defined by creating an annotated dataset in which each utterance is encountered once. Such data does not resemble real-world natural language interactions in which certain utterances are encountered frequently, others rarely. For deployed NLU systems this is a vital problem, since the underlying machine learning (ML) models are often fine-tuned on typical NLU data, and then applied to real-world data with a very different distribution. Such systems need to maintain interpretation consistency for both high-frequency utterances and low-frequency utterances. We propose an alternative strategy that explicitly uses utterance frequency in training data to learn models that are more robust to unknown distributions. We present a methodology to simulate utterance usage in two public NLU corpora and create new corpora with head, body and tail segments. We evaluate several methods for joint intent classification and named entity recognition (IC-NER), and use two domain generalization approaches that we adapt to NER. The proposed approaches demonstrate upto 7.02% relative improvement in semantic accuracy over baselines on the tail data. We provide insights as to why the proposed approaches work and show that the reasons for observed improvements do not align with those reported in previous work.

A principle behind dozens of attribution methods is to take the prediction difference between before-and-after an input feature (here, a token) is removed as its attribution. A popular Input Marginalization (IM) method (Kim et al., 2020) uses BERT to replace a token, yielding more plausible counterfactuals. While Kim et al., 2020 reported that IM is effective, we find this conclusion not convincing as the Deletion-BERT metric used in their paper is biased towards IM. Importantly, this bias exists in Deletion-based metrics, including Insertion, Sufficiency, and Comprehensiveness. Furthermore, our rigorous evaluation using 6 metrics and 3 datasets finds no evidence that IM is better than a Leave-One-Out (LOO) baseline. We find two reasons why IM is not better than LOO: (1) deleting a single word from the input only marginally reduces a classifier’s accuracy; and (2) a highly predictable word is always given near-zero attribution, regardless of its true importance to the classifier. In contrast, making LIME samples more natural via BERT consistently improves LIME accuracy under several ROAR metrics.

Cross-lingual transfer learning has proven useful in a variety of Natural Language (NLP) tasks, but it is understudied in the context of legal NLP, and not at all in Legal Judgment Prediction (LJP). We explore transfer learning techniques on LJP using the trilingual Swiss-Judgment-Prediction (SJP) dataset, including cases written in three languages. We find that Cross-Lingual Transfer (CLT) improves the overall results across languages, especially when we use adapter-based fine-tuning. Finally, we further improve the model’s performance by augmenting the training dataset with machine-translated versions of the original documents, using a 3× larger training corpus. Further on, we perform an analysis exploring the effect of cross-domain and cross-regional transfer, i.e., train a model across domains (legal areas), or regions. We find that in both settings (legal areas, origin regions), models trained across all groups perform overall better, while they also have improved results in the worst-case scenarios. Finally, we report improved results when we ambitiously apply cross-jurisdiction transfer, where we further augment our dataset with Indian legal cases.

Though recent works have focused on modeling high resource languages, the area is still unexplored for low resource languages like Bengali and Hindi. We propose an end to end trainable memory efficient CNN architecture named CoCNN to handle specific characteristics such as high inflection, morphological richness, flexible word order and phonetical spelling errors of Bengali and Hindi. In particular, we introduce two learnable convolutional sub-models at word and at sentence level that are end to end trainable. We show that state-of-the-art (SOTA) Transformer models including pretrained BERT do not necessarily yield the best performance for Bengali and Hindi. CoCNN outperforms pretrained BERT with 16X less parameters and achieves much better performance than SOTA LSTMs on multiple real-world datasets. This is the first study on the effectiveness of different architectures from Convolution, Recurrent, and Transformer neural net paradigm for modeling Bengali and Hindi.

Recent language generative models are mostly trained on large-scale datasets, while in some real scenarios, the training datasets are often expensive to obtain and would be small-scale. In this paper we investigate the challenging task of less-data constrained generation, especially when the generated news headlines are short yet expected by readers to keep readable and informative simultaneously. We highlight the key information modeling task and propose a novel duality fine-tuning method by formally defining the probabilistic duality constraints between key information prediction and headline generation tasks. The proposed method can capture more information from limited data, build connections between separate tasks, and is suitable for less-data constrained generation tasks. Furthermore, the method can leverage various pre-trained generative regimes, e.g., autoregressive and encoder-decoder models. We conduct extensive experiments to demonstrate that our method is effective and efficient to achieve improved performance in terms of language modeling metric and informativeness correctness metric on two public datasets.

Mitigating bias in training on biased datasets is an important open problem. Several techniques have been proposed, however the typical evaluation regime is very limited, considering very narrow data conditions. For instance, the effect of target class imbalance and stereotyping is under-studied. To address this gap, we examine the performance of various debiasing methods across multiple tasks, spanning binary classification (Twitter sentiment), multi-class classification (profession prediction), and regression (valence prediction). Through extensive experimentation, we find that data conditions have a strong influence on relative model performance, and that general conclusions cannot be drawn about method efficacy when evaluating only on standard datasets, as is current practice in fairness research.

Knowledge graph question answering (KGQA) based on information retrieval aims to answer a question by retrieving answer from a large-scale knowledge graph. Most existing methods first roughly retrieve the knowledge subgraphs (KSG) that may contain candidate answer, and then search for the exact answer in the KSG. However, the KSG may contain thousands of candidate nodes since the knowledge graph involved in querying is often of large scale, thus decreasing the performance of answer selection. To tackle this problem, we first propose to partition the retrieved KSG to several smaller sub-KSGs via a new subgraph partition algorithm and then present a graph-augmented learning to rank model to select the top-ranked sub-KSGs from them. Our proposed model combines a novel subgraph matching networks to capture global interactions in both question and subgraphs and an Enhanced Bilateral Multi-Perspective Matching model to capture local interactions. Finally, we apply an answer selection model on the full KSG and the top-ranked sub-KSGs respectively to validate the effectiveness of our proposed graph-augmented learning to rank method. The experimental results on multiple benchmark datasets have demonstrated the effectiveness of our approach.

Capitalise on deep learning models, offering Natural Language Processing (NLP) solutions as a part of the Machine Learning as a Service (MLaaS) has generated handsome revenues. At the same time, it is known that the creation of these lucrative deep models is non-trivial. Therefore, protecting these inventions’ intellectual property rights (IPR) from being abused, stolen and plagiarized is vital. This paper proposes a practical approach for the IPR protection on recurrent neural networks (RNN) without all the bells and whistles of existing IPR solutions. Particularly, we introduce the Gatekeeper concept that resembles the recurrent nature in RNN architecture to embed keys. Also, we design the model training scheme in a way such that the protected RNN model will retain its original performance iff a genuine key is presented. Extensive experiments showed that our protection scheme is robust and effective against ambiguity and removal attacks in both white-box and black-box protection schemes on different RNN variants. Code is available at https://github.com/zhiqin1998/RecurrentIPR.

Word associations are among the most common paradigms to study the human mental lexicon. While their structure and types of associations have been well studied, surprisingly little attention has been given to the question of why participants produce the observed associations. Answering this question would not only advance understanding of human cognition, but could also aid machines in learning and representing basic commonsense knowledge. This paper introduces a large, crowd-sourced data set of English word associations with explanations, labeled with high-level relation types. We present an analysis of the provided explanations, and design several tasks to probe to what extent current pre-trained language models capture the underlying relations. Our experiments show that models struggle to capture the diversity of human associations, suggesting WAX is a rich benchmark for commonsense modeling and generation.

Multimodal sentiment analysis (MSA) is an important way of observing mental activities with the help of data captured from multiple modalities. However, due to the recording or transmission error, some modalities may include incomplete data. Most existing works that address missing modalities usually assume a particular modality is completely missing and seldom consider a mixture of missing across multiple modalities. In this paper, we propose a simple yet effective meta-sampling approach for multimodal sentiment analysis with missing modalities, namely Missing Modality-based Meta Sampling (M3S). To be specific, M3S formulates a missing modality sampling strategy into the modal agnostic meta-learning (MAML) framework. M3S can be treated as an efficient add-on training component on existing models and significantly improve their performances on multimodal data with a mixture of missing modalities. We conduct experiments on IEMOCAP, SIMS and CMU-MOSI datasets, and superior performance is achieved compared with recent state-of-the-art methods.

This paper focuses on the generation of natural language questions based on SPARQL queries, with an emphasis on conversational use cases (follow-up question-answering). It studies what can be achieved so far based on current deep learning models (namely pretrained T5 and BART models). To do so, 4 knowledge-based QA corpora have been homogenized for the task and a new challenge set is introduced. A first series of experiments analyzes the impact of different training setups, while a second series seeks to understand what is still difficult for these models. The results from automatic metrics and human evaluation show that simple questions and frequent templates of SPARQL queries are usually well processed whereas complex questions and conversational dimensions (coreferences and ellipses) are still difficult to handle. The experimental material is publicly available on https://github.com/Orange-OpenSource/sparql-to-text .

Emotion recognition in conversation (ERC) has attracted much attention in recent years for its necessity in widespread applications. With the development of graph neural network (GNN), recent state-of-the-art ERC models mostly use GNN to embed the intrinsic structure information of a conversation into the utterance features. In this paper, we propose a novel GNN-based model for ERC, namely S+PAGE, to better capture the speaker and position-aware conversation structure information. Specifically, we add the relative positional encoding and speaker dependency encoding in the representations of edge weights and edge types respectively to acquire a more reasonable aggregation algorithm for ERC. Besides, a two-stream conversational Transformer is presented to extract both the self and inter-speaker contextual features for each utterance. Extensive experiments are conducted on four ERC benchmarks with state-of-the-art models employed as baselines for comparison, whose results demonstrate the superiority of our model.

Grammatical error correction (GEC) systems are a useful tool for assessing a learner’s writing ability. These systems allow the grammatical proficiency of a candidate’s text to be assessed without requiring an examiner or teacher to read the text. A simple summary of a candidate’s ability can be measured by the total number of edits between the input text and the GEC system output: the fewer the edits the better the candidate. With advances in deep learning, GEC systems have become increasingly powerful and accurate. However, deep learning systems are susceptible to adversarial attacks, in which a small change at the input can cause large, undesired changes at the output. In the context of GEC for automated assessment, the aim of an attack can be to deceive the system into not correcting (concealing) grammatical errors to create the perception of higher language ability. An interesting aspect of adversarial attacks in this scenario is that the attack needs to be simple as it must be applied by, for example, a learner of English. The form of realistic attack examined in this work is appending the same phrase to each input sentence: a concatenative universal attack. The candidate only needs to learn a single attack phrase. State-of-the-art GEC systems are found to be susceptible to this form of simple attack, which transfers to different test sets as well as system architectures,

The use of deep neural models for diagnosis prediction from clinical text has shown promising results. However, in clinical practice such models must not only be accurate, but provide doctors with interpretable and helpful results. We introduce ProtoPatient, a novel method based on prototypical networks and label-wise attention with both of these abilities. ProtoPatient makes predictions based on parts of the text that are similar to prototypical patients—providing justifications that doctors understand. We evaluate the model on two publicly available clinical datasets and show that it outperforms existing baselines. Quantitative and qualitative evaluations with medical doctors further demonstrate that the model provides valuable explanations for clinical decision support.

Related works used indexes like CKA and variants of CCA to measure the similarity of cross-lingual representations in multilingual language models. In this paper, we argue that assumptions of CKA/CCA align poorly with one of the motivating goals of cross-lingual learning analysis, i.e., explaining zero-shot cross-lingual transfer. We highlight what valuable aspects of cross-lingual similarity these indexes fail to capture and provide a motivating case study demonstrating the problem empirically. Then, we introduce Average Neuron-Wise Correlation (ANC) as a straightforward alternative that is exempt from the difficulties of CKA/CCA and is good specifically in a cross-lingual context. Finally, we use ANC to construct evidence that the previously introduced “first align, then predict” pattern takes place not only in masked language models (MLMs) but also in multilingual models with causal language modeling objectives (CLMs). Moreover, we show that the pattern extends to the scaled versions of the MLMs and CLMs (up to 85x original mBERT). Our code is publicly available at https://github.com/TartuNLP/xsim

Given the challenges and complexities introduced while dealing with Dialect Arabic (DA) variations, Transformer based models, e.g., BERT, outperformed other models in dealing with the DA identification task. However, to fine-tune these models, a large corpus is required. Getting a large number high quality labeled examples for some Dialect Arabic classes is challenging and time-consuming. In this paper, we address the Dialect Arabic Identification task. We extend the transformer-based models, ARBERT and MARBERT, with unlabeled data in a generative adversarial setting using Semi-Supervised Generative Adversarial Networks (SS-GAN). Our model enabled producing high-quality embeddings for the Dialect Arabic examples and aided the model to better generalize for the downstream classification task given few labeled examples. Experimental results showed that our model reached better performance and faster convergence when only a few labeled examples are available.

Word embeddings and pre-trained language models have become essential technical elements in natural language processing. While the general practice is to use or fine-tune publicly available models, there are significant advantages in creating or pre-training unique models that match the domain. The performance of the models degrades as language changes or evolves continuously, but the high cost of model building inhibits regular re-training, especially for the language models. This study proposes an efficient way to detect time-series performance degradation of word embeddings and pre-trained language models by calculating the degree of semantic shift. Monitoring performance through the proposed method supports decision-making as to whether a model should be re-trained. The experiments demonstrated that the proposed method can identify time-series performance degradation in two datasets, Japanese and English. The source code is available at https://github.com/Nikkei/semantic-shift-stability.

We present a novel approach for text sanitization, which is the task of editing a document to mask all (direct and indirect) personal identifiers and thereby conceal the identity of the individuals(s) mentioned in the text. In contrast to previous work, the approach relies on explicit measures of privacy risk, making it possible to explicitly control the trade-off between privacy protection and data utility. The approach proceeds in three steps. A neural, privacy-enhanced entity recognizer is first employed to detect and classify potential personal identifiers. We then determine which entities, or combination of entities, are likely to pose a re-identification risk through a range of privacy risk assessment measures. We present three such measures of privacy risk, respectively based on (1) span probabilities derived from a BERT language model, (2) web search queries and (3) a classifier trained on labelled data. Finally, a linear optimization solver decides which entities to mask to minimize the semantic loss while simultaneously ensuring that the estimated privacy risk remains under a given threshold. We evaluate the approach both in the absence and presence of manually annotated data. Our results highlight the potential of the approach, as well as issues specific types of personal data can introduce to the process.

Keywords or keyphrases are often used to highlight a document’s domains or main topics. Unsupervised keyphrase extraction (UKE) has always been highly anticipated because no labeled data is needed to train a model. This paper proposes an augmented graph-based unsupervised model to identify keyphrases from a document by integrating graph and deep learning methods. The proposed model utilizes mutual attention extracted from the pre-trained BERT model to build the candidate graph and augments the graph with global and local context nodes to improve the performance. The proposed model is evaluated on four publicly available datasets against thirteen UKE baselines. The results show that the proposed model is an effective and robust UKE model for long and short documents. Our source code is available on GitHub.

With a knowledge graph and a set of if-then rules, can we reason about the conclusions given a set of observations? In this work, we formalize this question as the cognitive inference problem, and introduce the Cognitive Knowledge Graph (CogKG) that unifies two representations of heterogeneous symbolic knowledge: expert rules and relational facts. We propose a general framework in which the unified knowledge representations can perform both learning and reasoning. Specifically, we implement the above framework in two settings, depending on the availability of labeled data. When no labeled data are available for training, the framework can directly utilize symbolic knowledge as the decision basis and perform reasoning. When labeled data become available, the framework casts symbolic knowledge as a trainable neural architecture and optimizes the connection weights among neurons through gradient descent. Empirical study on two clinical diagnosis benchmarks demonstrates the superiority of the proposed method over time-tested knowledge-driven and data-driven methods, showing the great potential of the proposed method in unifying heterogeneous symbolic knowledge, i.e., expert rules and relational facts, as the substrate of machine learning and reasoning models.

Pre-trained transformer-based language models have achieved state-of-the-art performance in many areas of NLP. It is still an open question whether the models are capable of integrating syntax and semantics in language processing like humans. This paper investigates if models and humans construct argument hierarchy similarly with the effects from telicity, agency, and individuation, using the Chinese structure “NP1+BA/BEI+NP2+VP”. We present both humans and six transformer-based models with prepared sentences and analyze their preference between BA (view NP1 as an agent) and BEI (NP2 as an agent). It is found that the models and humans respond to (non-)agentive features in telic context and atelic feature very similarly. However, the models show insufficient sensitivity to both pragmatic function in expressing undesirable events and different individuation degrees represented by human common nouns vs. proper names. By contrast, humans rely heavily on these cues to establish the thematic relation between two arguments NP1 and NP2. Furthermore, the models tend to interpret the subject as an agent, which is not the case for humans who align agents independently of subject position in Mandarin Chinese.

While there has been substantial progress in developing systems to automate fact-checking, they still lack credibility in the eyes of the users. Thus, an interesting approach has emerged: to perform automatic fact-checking by verifying whether an input claim has been previously fact-checked by professional fact-checkers and to return back an article that explains their decision. This is a sensible approach as people trust manual fact-checking, and as many claims are repeated multiple times. Yet, a major issue when building such systems is the small number of known tweet–verifying article pairs available for training. Here, we aim to bridge this gap by making use of crowd fact-checking, i.e., mining claims in social media for which users have responded with a link to a fact-checking article. In particular, we mine a large-scale collection of 330,000 tweets paired with a corresponding fact-checking article. We further propose an end-to-end framework to learn from this noisy data based on modified self-adaptive training, in a distant supervision scenario. Our experiments on the CLEF’21 CheckThat! test set show improvements over the state of the art by two points absolute. Our code and datasets are available at https://github.com/mhardalov/crowdchecked-claims

Social media often serves as a breeding ground for various hateful and offensive content. Identifying such content on social media is crucial due to its impact on the race, gender, or religion in an unprejudiced society. However, while there is extensive research in hate speech detection in English, there is a gap in hateful content detection in low-resource languages like Bengali. Besides, a current trend on social media is the use of Romanized Bengali for regular interactions. To overcome the existing research’s limitations, in this study, we develop an annotated dataset of 10K Bengali posts consisting of 5K actual and 5K Romanized Bengali tweets. We implement several baseline models for the classification of such hateful posts. We further explore the interlingual transfer mechanism to boost classification performance. Finally, we perform an in-depth error analysis by looking into the misclassified posts by the models. While training actual and Romanized datasets separately, we observe that XLM-Roberta performs the best. Further, we witness that on joint training and few-shot training, MuRIL outperforms other models by interpreting the semantic expressions better. We make our code and dataset public for others.

We present a novel architecture for explainable modeling of task-oriented dialogues with discrete latent variables to represent dialogue actions. Our model is based on variational recurrent neural networks (VRNN) and requires no explicit annotation of semantic information. Unlike previous works, our approach models the system and user turns separately and performs database query modeling, which makes the model applicable to task-oriented dialogues while producing easily interpretable action latent variables. We show that our model outperforms previous approaches with less supervision in terms of perplexity and BLEU on three datasets, and we propose a way to measure dialogue success without the need for expert annotation. Finally, we propose a novel way to explain semantics of the latent variables with respect to system actions.

Recent progress in language model pre-training has led to important improvements in Named Entity Recognition (NER). Nonetheless, this progress has been mainly tested in well-formatted documents such as news, Wikipedia, or scientific articles. In social media the landscape is different, in which it adds another layer of complexity due to its noisy and dynamic nature. In this paper, we focus on NER in Twitter, one of the largest social media platforms, and construct a new NER dataset, TweetNER7, which contains seven entity types annotated over 11,382 tweets from September 2019 to August 2021. The dataset was constructed by carefully distributing the tweets over time and taking representative trends as a basis. Along with the dataset, we provide a set of language model baselines and perform an analysis on the language model performance on the task, especially analyzing the impact of different time periods. In particular, we focus on three important temporal aspects in our analysis: short-term degradation of NER models over time, strategies to fine-tune a language model over different periods, and self-labeling as an alternative to lack of recently-labeled data. TweetNER7 is released publicly (https://huggingface.co/datasets/tner/tweetner7) along with the models fine-tuned on it (NER models have been integrated into TweetNLP and can be found at https://github.com/asahi417/tner/tree/master/examples/tweetner7_paper).

Reasoning with preconditions such as “glass can be used for drinking water unless the glass is shattered” remains an open problem for language models. The main challenge lies in the scarcity of preconditions data and the model’s lack of support for such reasoning. We present PInKS , Preconditioned Commonsense Inference with WeaK Supervision, an improved model for reasoning with preconditions through minimum supervision. We show, empirically and theoretically, that PInKS improves the results on benchmarks focused on reasoning with the preconditions of commonsense knowledge (up to 40% Macro-F1 scores). We further investigate PInKS through PAC-Bayesian informativeness analysis, precision measures, and ablation study.

Open-Domain Generative Question Answering has achieved impressive performance in English by combining document-level retrieval with answer generation. These approaches, which we refer to as GenQA, can generate complete sentences, effectively answering both factoid and non-factoid questions. In this paper, we extend to the multilingual and cross-lingual settings. For this purpose, we first introduce GenTyDiQA, an extension of the TyDiQA dataset with well-formed and complete answers for Arabic, Bengali, English, Japanese, and Russian. Based on GenTyDiQA, we design a cross-lingual generative model that produces full-sentence answers by exploiting passages written in multiple languages, including languages different from the question. Our cross-lingual generative system outperforms answer sentence selection baselines for all 5 languages and monolingual generative pipelines for three out of five languages studied.

In recent years, top-down neural models have achieved significant success in text-level discourse parsing. Nevertheless, they still suffer from the top-down error propagation issue, especially when the performance on the upper-level tree nodes is terrible. In this research, we aim to learn from the correlations in between EDUs directly to shorten the hierarchical distance of the RST structure to alleviate the above problem. Specifically, we contribute a joint top-down framework that learns from both discourse dependency and constituency parsing through one shared encoder and two independent decoders. Moreover, we also explore a constituency-to-dependency conversion scheme tailored for the Chinese discourse corpus to ensure the high quality of the joint learning process. Our experimental results on CDTB show that the dependency information we use well heightens the understanding of the rhetorical structure, especially for the upper-level tree layers.

We aim to develop methods for understanding how multimedia news exposure can affect people’s emotional responses, and we especially focus on news content related to gun violence, a very important yet polarizing issue in the U.S. We created the dataset NEmo+ by significantly extending the U.S. gun violence news-to-emotions dataset, BU-NEmo, from 320 to 1,297 news headline and lead image pairings and collecting 38,910 annotations in a large crowdsourcing experiment. In curating the NEmo+ dataset, we developed methods to identify news items that will trigger similar versus divergent emotional responses. For news items that trigger similar emotional responses, we compiled them into the NEmo+-Consensus dataset. We benchmark models on this dataset that predict a person’s dominant emotional response toward the target news item (single-label prediction). On the full NEmo+ dataset, containing news items that would lead to both differing and similar emotional responses, we also benchmark models for the novel task of predicting the distribution of evoked emotional responses in humans when presented with multi-modal news content. Our single-label and multi-label prediction models outperform baselines by large margins across several metrics.

Data augmentation techniques have been proven useful in many applications in NLP fields. Most augmentations are task-specific, and cannot be used as a general-purpose tool. In our work, we present AugCSE, a unified framework to utilize diverse sets of data augmentations to achieve a better, general-purpose, sentence embedding model. Building upon the latest sentence embedding models, our approach uses a simple antagonistic discriminator that differentiates the augmentation types. With the finetuning objective borrowed from domain adaptation, we show that diverse augmentations, which often lead to conflicting contrastive signals, can be tamed to produce a better and more robust sentence representation. Our methods achieve state-of-the-art results on downstream transfer tasks and perform competitively on semantic textual similarity tasks, using only unsupervised data.

Complex conversation settings such as persuasion involve communicating changes in attitude or behavior, so users’ perspectives need to be addressed, even when not directly related to the topic. In this work, we contribute a novel modular dialogue system framework that seamlessly integrates factual information and social content into persuasive dialogue. Our framework is generalizable to any dialogue tasks that have mixed social and task contents. We conducted a study that compared user evaluations of our framework versus a baseline end-to-end generation model. We found our model was evaluated to be more favorable in all dimensions including competence and friendliness compared to the baseline model which does not explicitly handle social content or factual questions.

Multimodal aspect-based sentiment analysis (MABSA) aims to extract the aspect terms from text and image pairs, and then analyze their corresponding sentiment. Recent studies typically use either a pipeline method or a unified transformer based on a cross-attention mechanism. However, these methods fail to explicitly and effectively incorporate the alignment between text and image. Supervised finetuning of the universal transformers for MABSA still requires a certain number of aligned image-text pairs. This study proposes a dual-encoder transformer with cross-modal alignment (DTCA). Two auxiliary tasks, including text-only extraction and text-patch alignment are introduced to enhance cross-attention performance. To align text and image, we propose an unsupervised approach which minimizes the Wasserstein distance between both modalities, forcing both encoders to produce more appropriate representations for the final extraction. Experimental results on two benchmarks demonstrate that DTCA consistently outperforms existing methods.

Recently, emerging approaches have been proposed to deal with robotic navigation problems, especially vision-and-language navigation task which is one of the most realistic indoor navigation challenge tasks. This task can be modelled as a sequential decision-making problem, which is suitable to be solved by deep reinforcement learning. Unfortunately, the observations provided from the simulator in this task are not fully observable states, which exacerbate the difficulty of implementing reinforcement learning. To deal with this challenge, this paper presents a novel method, called as attentive variational state tracker (AVAST), a variational approach to approximate belief state distribution for the construction of a reinforced navigator. The variational approach is introduced to improve generalization to the unseen environment which barely achieved by traditional deterministic state tracker. In order to stabilize the learning procedure, a fine-tuning process using policy optimization is proposed. From the experimental results, the proposed AVAST does improve the generalization relative to previous works in vision-and-language navigation task. A significant performance is achieved without requiring any additional exploration in the unseen environment.

Large pretrained multilingual models, trained on dozens of languages, have delivered promising results due to cross-lingual learning capabilities on a variety of language tasks. Further adapting these models to specific languages, especially ones unseen during pre-training, is an important goal toward expanding the coverage of language technologies. In this study, we show how we can use language phylogenetic information to improve cross-lingual transfer leveraging closely related languages in a structured, linguistically-informed manner. We perform adapter-based training on languages from diverse language families (Germanic, Uralic, Tupian, Uto-Aztecan) and evaluate on both syntactic and semantic tasks, obtaining more than 20% relative performance improvements over strong commonly used baselines, especially on languages unseen during pre-training.

The concerning rise of hateful content on online platforms has increased the attention towards automatic hate speech detection, commonly formulated as a supervised classification task. State-of-the-art deep learning-based approaches usually require a substantial amount of labeled resources for training. However, annotating hate speech resources is expensive, time-consuming, and often harmful to the annotators. This creates a pressing need to transfer knowledge from the existing labeled resources to low-resource hate speech corpora with the goal of improving system performance. For this, neighborhood-based frameworks have been shown to be effective. However, they have limited flexibility. In our paper, we propose a novel training strategy that allows flexible modeling of the relative proximity of neighbors retrieved from a resource-rich corpus to learn the amount of transfer. In particular, we incorporate neighborhood information with Optimal Transport, which permits exploiting the geometry of the data embedding space. By aligning the joint embedding and label distributions of neighbors, we demonstrate substantial improvements over strong baselines, in low-resource scenarios, on different publicly available hate speech corpora.

Text autoencoders are often used for unsupervised conditional text generation by applying mappings in the latent space to change attributes to the desired values. Recently, Mai et al. (2020) proposed Emb2Emb, a method to learn these mappings in the embedding space of an autoencoder. However, their method is restricted to autoencoders with a single-vector embedding, which limits how much information can be retained. We address this issue by extending their method to Bag-of-Vectors Autoencoders (BoV-AEs), which encode the text into a variable-size bag of vectors that grows with the size of the text, as in attention-based models. This allows to encode and reconstruct much longer texts than standard autoencoders. Analogous to conventional autoencoders, we propose regularization techniques that facilitate learning meaningful operations in the latent space. Finally, we adapt Emb2Emb for a training scheme that learns to map an input bag to an output bag, including a novel loss function and neural architecture. Our empirical evaluations on unsupervised sentiment transfer show that our method performs substantially better than a standard autoencoder.

Conversational Recommender System (CRS), which aims to recommend high-quality items to users through interactive conversations, has gained great research interest recently. A CRS is usually composed of a recommendation module and a generation module. In the previous work, these two modules are loosely connected in the model training and are shallowly integrated during inference, where a simple switching or copy mechanism is adopted to incorporate recommended items into generated responses. Moreover, the current end-to-end neural models trained on small crowd-sourcing datasets (e.g., 10K dialogs in the ReDial dataset) tend to overfit and have poor chit-chat ability. In this work, we propose a novel unified framework that integrates recommendation into the dialog (RecInDial) generation by introducing a vocabulary pointer. To tackle the low-resource issue in CRS, we finetune the large-scale pretrained language models to generate fluent and diverse responses, and introduce a knowledge-aware bias learned from an entity-oriented knowledge graph to enhance the recommendation performance. Furthermore, we propose to evaluate the CRS models in an end-to-end manner, which can reflect the overall performance of the entire system rather than the performance of individual modules, compared to the separate evaluations of the two modules used in previous work. Experiments on the benchmark dataset ReDial show our RecInDial model significantly surpasses the state-of-the-art methods. More extensive analyses show the effectiveness of our model.

This paper introduces a new method, SummVD, for automatic unsupervised extractive summarization. This method is based on singular value decomposition, a linear method in the number of words, in order to reduce the dimensionality of word embeddings and propose a representation of words on a small number of dimensions, each representing a hidden topic. It also uses word clustering to reduce the vocabulary size. This representation, specific to one document, reduces the noise brought by several dimensions of the embeddings that are useless in a restricted context. It is followed by a linear sentence extraction heuristic. This makes SummVD an efficient method for text summarization. We evaluate SummVD using several corpora of different nature (news, scientific articles, social network). Our method outperforms in effectiveness recent extractive approaches. Moreover, SummVD requires low resources, in terms of data and computing power. So it can be run on long single documents such as scientific papers as much as large multi-document corpora and is fast enough to be used in live summarization systems.

Current language models achieve low perplexity but their resulting generations still suffer from toxic responses, repetitiveness, and contradictions. The standard language modeling setup fails to address these issues. In this paper, we introduce a new architecture, Director, that consists of a unified generator-classifier with both a language modeling and a classification head for each output token. Training is conducted jointly using both standard language modeling data, and data labeled with desirable and undesirable sequences. Experiments in several settings show that the model has competitive training and decoding speed compared to standard language models while yielding superior results, avoiding undesirable behaviors while maintaining generation quality. It also outperforms existing model guiding approaches in terms of both accuracy and efficiency. Our code is made publicly available.

In this paper we study how to measure stereotypical bias in pre-trained vision-language models. We leverage a recently released text-only dataset, StereoSet, which covers a wide range of stereotypical bias, and extend it into a vision-language probing dataset called VLStereoSet to measure stereotypical bias in vision-language models. We analyze the differences between text and image and propose a probing task that detects bias by evaluating a model’s tendency to pick stereotypical statements as captions for anti-stereotypical images. We further define several metrics to measure both a vision-language model’s overall stereotypical bias and its intra-modal and inter-modal bias. Experiments on six representative pre-trained vision-language models demonstrate that stereotypical biases clearly exist in most of these models and across all four bias categories, with gender bias slightly more evident. Further analysis using gender bias data and two vision-language models also suggest that both intra-modal and inter-modal bias exist.

We investigate the effect of varying citation context window sizes on model performance in citation intent classification. Prior studies have been limited to the application of fixed-size contiguous citation contexts or the use of manually curated citation contexts. We introduce a new automated unsupervised approach for the selection of a dynamic-size and potentially non-contiguous citation context, which utilises the transformer-based document representations and embedding similarities. Our experiments show that the addition of non-contiguous citing sentences improves performance beyond previous results. Evalu- ating on the (1) domain-specific (ACL-ARC) and (2) the multi-disciplinary (SDP-ACT) dataset demonstrates that the inclusion of additional context beyond the citing sentence significantly improves the citation classifi- cation model’s performance, irrespective of the dataset’s domain. We release the datasets and the source code used for the experiments at: https://github.com/oacore/dynamic_citation_context

Word embeddings learned using the distributional hypothesis (e.g., GloVe, Word2vec) do not capture the affective dimensions of valence, arousal, and dominance, which are present inherently in words. We present a novel retrofitting method for updating embeddings of words for their affective meaning. It learns a non-linear transformation function that maps pre-trained embeddings to an affective vector space, in a representation learning setting. We investigate word embeddings for their capacity to cluster emotion-bearing words. The affective embeddings learned by our method achieve better inter-cluster and intra-cluster distance for words having the same emotions, as evaluated through different cluster quality metrics. For the downstream tasks on sentiment analysis and sarcasm detection, simple classification models, viz. SVM and Attention Net, learned using our affective embeddings perform better than their pre-trained counterparts (more than 1.5% improvement in F1-score) and other benchmarks. Furthermore, the difference in performance is more pronounced in limited data setting.

Encoder-decoder architecture is widely adopted for sequence-to-sequence modeling tasks. For machine translation, despite the evolution from long short-term memory networks to Transformer networks, plus the introduction and development of attention mechanism, encoder-decoder is still the de facto neural network architecture for state-of-the-art models. While the motivation for decoding information from some hidden space is straightforward, the strict separation of the encoding and decoding steps into an encoder and a decoder in the model architecture is not necessarily a must. Compared to the task of autoregressive language modeling in the target language, machine translation simply has an additional source sentence as context. Given the fact that neural language models nowadays can already handle rather long contexts in the target language, it is natural to ask whether simply concatenating the source and target sentences and training a language model to do translation would work. In this work, we investigate the aforementioned concept for machine translation. Specifically, we experiment with bilingual translation, translation with additional target monolingual data, and multilingual translation. In all cases, this alternative approach performs on par with the baseline encoder-decoder Transformer, suggesting that an encoder-decoder architecture might be redundant for neural machine translation.

Scientific paper summarization is always challenging in Natural Language Processing (NLP) since it is hard to collect summaries from such long and complicated text. We observe that previous works tend to extract summaries from the head of the paper, resulting in information incompleteness. In this work, we present SAPGraph to utilize paper structure for solving this problem. SAPGraph is a scientific paper extractive summarization framework based on a structure-aware heterogeneous graph, which models the document into a graph with three kinds of nodes and edges based on structure information of facets and knowledge. Additionally, we provide a large-scale dataset of COVID-19-related papers, CORD-SUM. Experiments on CORD-SUM and ArXiv datasets show that SAPGraph generates more comprehensive and valuable summaries compared to previous works.

Communicating efficiently in natural language requires that we often leave information implicit, especially in spontaneous speech. This frequently results in phenomena of incompleteness, such as omitted references, that pose challenges for language processing. In this survey paper, we review the state of the art in research regarding the automatic processing of such implicit references in dialog scenarios, discuss weaknesses with respect to inconsistencies in task definitions and terminologies, and outline directions for future work. Among others, these include a unification of existing tasks, addressing data scarcity, and taking into account model and annotator uncertainties.

In pace with developments in the research field of artificial intelligence, knowledge graphs (KGs) have attracted a surge of interest from both academia and industry. As a representation of semantic relations between entities, KGs have proven to be particularly relevant for natural language processing (NLP), experiencing a rapid spread and wide adoption within recent years. Given the increasing amount of research work in this area, several KG-related approaches have been surveyed in the NLP research community. However, a comprehensive study that categorizes established topics and reviews the maturity of individual research streams remains absent to this day. Contributing to closing this gap, we systematically analyzed 507 papers from the literature on KGs in NLP. Our survey encompasses a multifaceted review of tasks, research types, and contributions. As a result, we present a structured overview of the research landscape, provide a taxonomy of tasks, summarize our findings, and highlight directions for future work.

Automated reporting has the potential to assist radiologists with the time-consuming procedure of generating text radiology reports. Most existing approaches generate the report directly from the radiology image, however we observe that the resulting reports exhibit realistic style but lack clinical accuracy. Therefore, we propose a two-step pipeline that subdivides the problem into factual triple extraction followed by free-text report generation. The first step comprises supervised extraction of clinically relevant structured information from the image, expressed as triples of the form (entity1, relation, entity2). In the second step, these triples are input to condition the generation of the radiology report. In particular, we focus our work on Chest X-Ray (CXR) radiology report generation. The proposed framework shows state-of-the-art results on the MIMIC-CXR dataset according to most of the standard text generation metrics that we employ (BLEU, METEOR, ROUGE) and to clinical accuracy metrics (recall, precision and F1 assessed using the CheXpert labeler), also giving a 23% reduction in the total number of errors and a 29% reduction in critical clinical errors as assessed by expert human evaluation. In future, this solution can easily integrate more advanced model architectures - to both improve the triple extraction and the report generation - and can be applied to other complex image captioning tasks, such as those found in the medical domain.

Models based on large-pretrained language models, such as S(entence)BERT, provide effective and efficient sentence embeddings that show high correlation to human similarity ratings, but lack interpretability. On the other hand, graph metrics for graph-based meaning representations (e.g., Abstract Meaning Representation, AMR) can make explicit the semantic aspects in which two sentences are similar. However, such metrics tend to be slow, rely on parsers, and do not reach state-of-the-art performance when rating sentence similarity. In this work, we aim at the best of both worlds, by learning to induce Semantically Structured Sentence BERT embeddings (S³BERT). Our S³BERT embeddings are composed of explainable sub-embeddings that emphasize various sentence meaning features (e.g., semantic roles, negation, or quantification). We show how to i) learn a decomposition of the sentence embeddings into meaning features, through approximation of a suite of interpretable semantic AMR graph metrics, and how to ii) preserve the overall power of the neural embeddings by controlling the decomposition learning process with a second objective that enforces consistency with the similarity ratings of an SBERT teacher model. In our experimental studies, we show that our approach offers interpretability – while preserving the effectiveness and efficiency of the neural sentence embeddings.

Knowledge graphs are increasingly used in a plethora of downstream tasks or in the augmentation of statistical models to improve factuality. However, social biases are engraved in these representations and propagate downstream. We conducted a critical analysis of literature concerning biases at different steps of a knowledge graph lifecycle. We investigated factors introducing bias, as well as the biases that are rendered by knowledge graphs and their embedded versions afterward. Limitations of existing measurement and mitigation strategies are discussed and paths forward are proposed.

Knowledge graph embedding (KGE) has been well-studied in general domains, but has not been examined for food computing. To fill this gap, we perform knowledge representation learning over a food knowledge graph (KG). We employ a pre-trained language model to encode entities and relations, thus emphasizing contextual information in food KGs. The model is trained on two tasks – predicting a masked entity from a given triple from the KG and predicting the plausibility of a triple. Analysis of food substitutions helps in dietary choices for enabling healthier eating behaviors. Previous work in food substitutions mainly focuses on semantic similarity while ignoring the context. It is also hard to evaluate the substitutions due to the lack of an adequate validation set, and further, the evaluation is subjective based on perceived purpose. To tackle this problem, we propose a collection of adversarial sample generation strategies for different food substitutions over our learnt KGE. We propose multiple strategies to generate high quality context-aware recipe and ingredient substitutions and also provide generalized ingredient substitutions to meet different user needs. The effectiveness and efficiency of the proposed knowledge graph learning method and the following attack strategies are verified by extensive evaluations on a large-scale food KG.

Speakers repeat constructions frequently in dialogue. Due to their peculiar information-theoretic properties, repetitions can be thought of as a strategy for cost-effective communication. In this study, we focus on the repetition of lexicalised constructions—i.e., recurring multi-word units—in English open-domain spoken dialogues. We hypothesise that speakers use construction repetition to mitigate information rate, leading to an overall decrease in utterance information content over the course of a dialogue. We conduct a quantitative analysis, measuring the information content of constructions and that of their containing utterances, estimating information content with an adaptive neural language model. We observe that construction usage lowers the information content of utterances. This facilitating effect (i) increases throughout dialogues, (ii) is boosted by repetition, (iii) grows as a function of repetition frequency and density, and (iv) is stronger for repetitions of referential constructions.

As we begin to see low-powered computing paradigms (Neuromorphic Computing, Spiking Neural Networks, etc.) becoming more popular, learning binary word embeddings has become increasingly important for supporting NLP applications at the edge. Existing binary word embeddings are mostly derived from pretrained real-valued embeddings through different simple transformations, which often break the semantic consistency and the so-called “arithmetic” properties learned by the original, real-valued embeddings. This paper aims to address this limitation by introducing a new approach to learn binary embeddings from scratch, preserving the semantic relationships between words as well as the arithmetic properties of the embeddings themselves. To achieve this, we propose a novel genetic algorithm to learn the relationships between words from existing word analogy data-sets, carefully making sure that the arithmetic properties of the relationships are preserved. Evaluating our generated 16, 32, and 64-bit binary word embeddings on Mikolov’s word analogy task shows that more than 95% of the time, the best fit for the analogy is ranked in the top 5 most similar words in terms of cosine similarity.

Despite recent progress in video and language representation learning, the weak or sparse correspondence between the two modalities remains a bottleneck in the area. Most video-language models are trained via pair-level loss to predict whether a pair of video and text is aligned. However, even in paired video-text segments, only a subset of the frames are semantically relevant to the corresponding text, with the remainder representing noise; where the ratio of noisy frames is higher for longer videos. We propose FineCo (Fine-grained Contrastive Loss for Frame Sampling), an approach to better learn video and language representations with a fine-grained contrastive objective operating on video frames. It helps distil a video by selecting the frames that are semantically equivalent to the text, improving cross-modal correspondence. Building on the well established VideoCLIP model as a starting point, FineCo achieves state-of-the-art performance on YouCookII, a text-video retrieval benchmark with long videos. FineCo also achieves competitive results on text-video retrieval (MSR-VTT), and video question answering datasets (MSR-VTT QA and MSR-VTT MC) with shorter videos.

Recent methods based on pre-trained language models have exhibited superior performance over tabular tasks (e.g., tabular NLI), despite showing inherent problems such as not using the right evidence and inconsistent predictions across inputs while reasoning over the tabular data (Gupta et al., 2021). In this work, we utilize Pattern-Exploiting Training (PET) (i.e., strategic MLM) on pre-trained language models to strengthen these tabular reasoning models’ pre-existing knowledge and reasoning abilities. Our upgraded model exhibits a superior understanding of knowledge facts and tabular reasoning compared to current baselines. Additionally, we demonstrate that such models are more effective for underlying downstream tasks of tabular inference on INFOTABS. Furthermore, we show our model’s robustness against adversarial sets generated through various character and word level perturbations.

Recent research has revealed undesirable biases in NLP data and models. However, these efforts focus of social disparities in West, and are not directly portable to other geo-cultural contexts. In this paper, we focus on NLP fairness in the context of India. We start with a brief account of the prominent axes of social disparities in India. We build resources for fairness evaluation in the Indian context and use them to demonstrate prediction biases along some of the axes. We then delve deeper into social stereotypes for Region and Religion, demonstrating its prevalence in corpora and models. Finally, we outline a holistic research agenda to re-contextualize NLP fairness research for the Indian context, accounting for Indian societal context, bridging technological gaps in NLP capabilities and resources, and adapting to Indian cultural values. While we focus on India, this framework can be generalized to other geo-cultural contexts.

While neural methods for text-to-speech (TTS) have shown great advances in modeling multiple speakers, even in zero-shot settings, the amount of data needed for those approaches is generally not feasible for the vast majority of the world’s over 6,000 spoken languages. In this work, we bring together the tasks of zero-shot voice cloning and multilingual low-resource TTS. Using the language agnostic meta learning (LAML) procedure and modifications to a TTS encoder, we show that it is possible for a system to learn speaking a new language using just 5 minutes of training data while retaining the ability to infer the voice of even unseen speakers in the newly learned language. We show the success of our proposed approach in terms of intelligibility, naturalness and similarity to target speaker using objective metrics as well as human studies and provide our code and trained models open source.

IR models using a pretrained language model significantly outperform lexical approaches like BM25. In particular, SPLADE, which encodes texts to sparse vectors, is an effective model for practical use because it shows robustness to out-of-domain datasets. However, SPLADE still struggles with exact matching of low-frequency words in training data. In addition, domain shifts in vocabulary and word frequencies deteriorate the IR performance of SPLADE. Because supervision data are scarce in the target domain, addressing the domain shifts without supervision data is necessary. This paper proposes an unsupervised domain adaptation method by filling vocabulary and word-frequency gaps. First, we expand a vocabulary and execute continual pretraining with a masked language model on a corpus of the target domain. Then, we multiply SPLADE-encoded sparse vectors by inverse document frequency weights to consider the importance of documents with low-frequency words. We conducted experiments using our method on datasets with a large vocabulary gap from a source domain. We show that our method outperforms the present state-of-the-art domain adaptation method. In addition, our method achieves state-of-the-art results, combined with BM25.

Though some recent works focus on injecting sentiment knowledge into pre-trained language models, they usually design mask and reconstruction tasks in the post-training phase. This paper aims to integrate sentiment knowledge in the fine-tuning stage. To achieve this goal, we propose two sentiment-aware auxiliary tasks named sentiment word selection and conditional sentiment prediction and, correspondingly, integrate them into the objective of the downstream task. The first task learns to select the correct sentiment words from the given options. The second task predicts the overall sentiment polarity, with the sentiment polarity of the word given as prior knowledge. In addition, two label combination methods are investigated to unify multiple types of labels in each auxiliary task. Experimental results demonstrate that our approach consistently outperforms baselines (achieving a new state-of-the-art) and is complementary to existing sentiment-enhanced post-trained models.

Large pre-trained language models (LMs) have demonstrated the ability to obtain good performance on downstream tasks with limited examples in cross-lingual settings. However, this was mostly studied for relatively resource-rich languages, where at least enough unlabeled data is available to be included in pre-training a multilingual language model. In this paper, we explore the problem of cross-lingual transfer in unseen languages, where no unlabeled data is available for pre-training a model. We use a downstream sentiment analysis task across 12 languages, including 8 unseen languages, to analyze the effectiveness of several few-shot learning strategies across the three major types of model architectures and their learning dynamics. We also compare strategies for selecting languages for transfer and contrast findings across languages seen in pre-training compared to those that are not. Our findings contribute to the body of knowledge on cross-lingual models for low-resource settings that is paramount to increasing coverage, diversity, and equity in access to NLP technology. We show that, in few-shot learning, linguistically similar and geographically similar languages are useful for cross-lingual adaptation, but taking the context from a mixture of random source languages is surprisingly more effective. We also compare different model architectures and show that the encoder-only model, XLM-R, gives the best downstream task performance.

Existing self-supervised learning strategies are constrained to either a limited set of objectives or generic downstream tasks that predominantly target uni-modal applications. This has isolated progress for imperative multi-modal applications that are diverse in terms of complexity and domain-affinity, such as meme analysis. Here, we introduce two self-supervised pre-training methods, namely Ext-PIE-Net and MM-SimCLR that (i) employ off-the-shelf multi-modal hate-speech data during pre-training and (ii) perform self-supervised learning by incorporating multiple specialized pretext tasks, effectively catering to the required complex multi-modal representation learning for meme analysis. We experiment with different self-supervision strategies, including potential variants that could help learn rich cross-modality representations and evaluate using popular linear probing on the Hateful Memes task. The proposed solutions strongly compete with the fully supervised baseline via label-efficient training while distinctly outperforming them on all three tasks of the Memotion challenge with 0.18%, 23.64%, and 0.93% performance gain, respectively. Further, we demonstrate the generalizability of the proposed solutions by reporting competitive performance on the HarMeme task. Finally, we empirically establish the quality of the learned representations by analyzing task-specific learning, using fewer labeled training samples, and arguing that the complexity of the self-supervision strategy and downstream task at hand are correlated. Our efforts highlight the requirement of better multi-modal self-supervision methods involving specialized pretext tasks for efficient fine-tuning and generalizable performance.

Vision-language models can encode societal biases and stereotypes, but there are challenges to measuring and mitigating these multimodal harms due to lacking measurement robustness and feature degradation. To address these challenges, we investigate bias measures and apply ranking metrics for image-text representations. We then investigate debiasing methods and show that prepending learned embeddings to text queries that are jointly trained with adversarial debiasing and a contrastive loss, reduces various bias measures with minimal degradation to the image-text representation.

Linguistic disparity in the NLP world is a problem that has been widely acknowledged recently. However, different facets of this problem, or the reasons behind this disparity are seldom discussed within the NLP community. This paper provides a comprehensive analysis of the disparity that exists within the languages of the world. We show that simply categorising languages considering data availability may not be always correct. Using an existing language categorisation based on speaker population and vitality, we analyse the distribution of language data resources, amount of NLP/CL research, inclusion in multilingual web-based platforms and the inclusion in pre-trained multilingual models. We show that many languages do not get covered in these resources or platforms, and even within the languages belonging to the same language group, there is wide disparity. We analyse the impact of family, geographical location, GDP and the speaker population of languages and provide possible reasons for this disparity, along with some suggestions to overcome the same.

Automatic Readability Assessment aims at assigning a complexity level to a given text, which could help improve the accessibility to information in specific domains, such as the administrative one. In this paper, we investigate the behavior of a Neural Pairwise Ranking Model (NPRM) for sentence-level readability assessment of Italian administrative texts. To deal with data scarcity, we experiment with cross-lingual, cross- and in-domain approaches, and test our models on Admin-It, a new parallel corpus in the Italian administrative language, containing sentences simplified using three different rewriting strategies. We show that NPRMs are effective in zero-shot scenarios (~0.78 ranking accuracy), especially with ranking pairs containing simplifications produced by overall rewriting at the sentence-level, and that the best results are obtained by adding in-domain data (achieving perfect performance for such sentence pairs). Finally, we investigate where NPRMs failed, showing that the characteristics of the training data, rather than its size, have a bigger effect on a model’s performance.

Automatic language processing is used frequently in the Human Resources (HR) sector for automated candidate sourcing and evaluation of resumes. These models often use pre-trained language models where it is difficult to know if possible biases exist. Recently, Mutual Information (MI) methods have demonstrated notable performance in obtaining representations agnostic to sensitive variables such as gender or ethnicity. However, accessing these variables can sometimes be challenging, and their use is prohibited in some jurisdictions. These factors can make detecting and mitigating biases challenging. In this context, we propose to minimize the MI between a candidate’s name and a latent representation of their CV or short biography. This method may mitigate bias from sensitive variables without requiring the collection of these variables. We evaluate this methodology by first projecting the name representation into a smaller space to prevent potential MI minimization problems in high dimensions.

Negation is poorly captured by current language models, although the extent of this problem is not widely understood. We introduce a natural language inference (NLI) test suite to enable probing the capabilities of NLP methods, with the aim of understanding sub-clausal negation. The test suite contains premise–hypothesis pairs where the premise contains sub-clausal negation and the hypothesis is constructed by making minimal modifications to the premise in order to reflect different possible interpretations. Aside from adopting standard NLI labels, our test suite is systematically constructed under a rigorous linguistic framework. It includes annotation of negation types and constructions grounded in linguistic theory, as well as the operations used to construct hypotheses. This facilitates fine-grained analysis of model performance. We conduct experiments using pre-trained language models to demonstrate that our test suite is more challenging than existing benchmarks focused on negation, and show how our annotation supports a deeper understanding of the current NLI capabilities in terms of negation and quantification.

One of the challenges of developing a summarization model arises from the difficulty in measuring the factual inconsistency of the generated text. In this study, we reinterpret the decoder overconfidence-regularizing objective suggested in (Miao et al., 2021) as a hallucination risk measurement to better estimate the quality of generated summaries. We propose a reference-free metric, HaRiM+, which only requires an off-the-shelf summarization model to compute the hallucination risk based on token likelihoods. Deploying it requires no additional training of models or ad-hoc modules, which usually need alignment to human judgments. For summary-quality estimation, HaRiM+ records state-of-the-art correlation to human judgment on three summary-quality annotation sets: FRANK, QAGS, and SummEval. We hope that our work, which merits the use of summarization models, facilitates the progress of both automated evaluation and generation of summary.

The peer-review system has primarily remained the central process of all science communications. However, research has shown that the process manifests a power-imbalance scenario where the reviewer enjoys a position where their comments can be overly critical and wilfully obtuse without being held accountable. This brings into question the sanctity of the peer-review process, turning it into a fraught and traumatic experience for authors. A little more effort to still remain critical but be constructive in the feedback would help foster a progressive outcome from the peer-review process. In this paper, we argue to intervene at the step where this power imbalance actually begins in the system. To this end, we develop the first dataset of peer-review comments with their real-valued harshness scores. We build our dataset by using the popular Best-Worst-Scaling mechanism. We show the utility of our dataset for text moderation in peer reviews to make review reports less hurtful and more welcoming. We release our dataset and associated codes in https://github.com/Tirthankar-Ghosal/moderating-peer-review-harshness. Our research is one step towards helping create constructive peer-review reports.

Word order choices during sentence production can be primed by preceding sentences. In this work, we test the DUAL MECHANISM hypothesis that priming is driven by multiple different sources. Using a Hindi corpus of text productions, we model lexical priming with an n-gram cache model, and we capture more abstract syntactic priming with an adaptive neural language model. We permute the preverbal constituents of corpus sentences and then use a logistic regression model to predict which sentences actually occurred in the corpus against artificially generated meaning-equivalent variants. Our results indicate that lexical priming and lexically-independent syntactic priming affect complementary sets of verb classes. By showing that different priming influences are separable from one another, our results support the hypothesis that multiple different cognitive mechanisms underlie priming.

In this work, we study the importance of content frequency on abstractive summarization, where we define the content as “semantic units.” We propose a two-stage training framework to let the model automatically learn the frequency of each semantic unit in the source text. Our model is trained in an unsupervised manner since the frequency information can be inferred from source text only. During inference, our model identifies sentences with high-frequency semantic units and utilizes frequency information to generate summaries from the filtered sentences. Our model performance on the CNN/Daily Mail summarization task outperforms the other unsupervised methods under the same settings. Furthermore, we achieve competitive ROUGE scores with far fewer model parameters compared to several large-scale pre-trained models. Our model can be trained under low-resource language settings and thus can serve as a potential solution for real-world applications where pre-trained models are not applicable.

With the increasing use of influencing incongruent news headlines for spreading fake news, detecting incongruent news articles has become an important research challenge. Most of the earlier studies on incongruity detection focus on estimating the similarity between the headline and the encoding of the body or its summary. However, most of these methods fail to handle incongruent news articles created with embedded noise. Motivated by the above issue, this paper proposes a Multi-head Attention Dual Summary (MADS) based method which generates two types of summaries that capture the congruent and incongruent parts in the body separately. From various experimental setups over three publicly available datasets, it is evident that the proposed model outperforms the state-of-the-art baseline counterparts.

Dialogue Act Classification (DAC) that determines the communicative intention of an utterance has been investigated widely over the years as a standalone task. But the emotional state of the speaker has a considerable effect on its pragmatic content. Sentiment as a human behavior is also closely related to emotion and one aids in the better understanding of the other. Thus, their role in identification of DAs needs to be explored. As a first step, we extend the newly released multi-modal EMOTyDA dataset to enclose sentiment tags for each utterance. In order to incorporate these multiple aspects, we propose a Dual Attention Mechanism (DAM) based multi-modal, multi-tasking conversational framework. The DAM module encompasses intra-modal and interactive inter-modal attentions with multiple loss optimization at various hierarchies to fuse multiple modalities efficiently and learn generalized features across all the tasks. Additionally, to counter the class-imbalance issue in dialogues, we introduce a 2-step Deferred Optimisation Schedule (DOS) that involves Meta-Net (MN) learning and deferred re-weighting where the former helps to learn an explicit weighting function from data automatically and the latter deploys a re-weighted multi-task loss with a smaller learning rate. Empirically, we establish that the joint optimisation of multi-modal DAC, SA and ER tasks along with the incorporation of 2-step DOS and MN learning produces better results compared to its different counterparts and outperforms state-of-the-art model.

Typical financial documents consist of tables, texts, and numbers. Given sufficient training data, large language models (LM) can learn the tabular structures and perform numerical reasoning well in question answering (QA). However, their performances fall significantly when data and computational resources are limited. This study improves this performance drop by infusing explicit tabular structures through a graph neural network (GNN). We proposed a model developed from the baseline of a financial QA dataset named TAT-QA. The baseline model, TagOp, consists of answer span (evidence) extraction and numerical reasoning modules. As our main contributions, we introduced two components to the model: a GNN-based evidence extraction module for tables and an improved numerical reasoning module. The latter provides a solution to TagOp’s arithmetic calculation problem specific to operations requiring number ordering, such as subtraction and division, which account for a large portion of numerical reasoning. Our evaluation shows that the graph module has the advantage in low-resource settings, while the improved numerical reasoning significantly outperforms the baseline model.

Recently, pre-trained transformer-based models have achieved great success in the task of definition generation (DG). However, previous encoder-decoder models lack effective representation learning to contain full semantic components of the given word, which leads to generating under-specific definitions. To address this problem, we propose a novel contrastive learning method, encouraging the model to capture more detailed semantic representations from the definition sequence encoding. According to both automatic and manual evaluation, the experimental results on three mainstream benchmarks demonstrate that the proposed method could generate more specific and high-quality definitions compared with several state-of-the-art models.

Many NLP main tasks benefit from an accurate understanding of temporal expressions, e.g., text summarization, question answering, and information retrieval. This paper introduces Hengam, an adversarially trained transformer for Persian temporal tagging outperforming state-of-the-art approaches on a diverse and manually created dataset. We create Hengam in the following concrete steps: (1) we develop HengamTagger, an extensible rule-based tool that can extract temporal expressions from a set of diverse language-specific patterns for any language of interest. (2) We apply HengamTagger to annotate temporal tags in a large and diverse Persian text collection (covering both formal and informal contexts) to be used as weakly labeled data. (3) We introduce an adversarially trained transformer model on HengamCorpus that can generalize over the HengamTagger’s rules. We create HengamGold, the first high-quality gold standard for Persian temporal tagging. Our trained adversarial HengamTransformer not only achieves the best performance in terms of the F1-score (a type F1-Score of 95.42 and a partial F1-Score of 91.60) but also successfully deals with language ambiguities and incorrect spellings. Our code, data, and models are publicly available at https://github.com/kargaranamir/Hengam.

In visual question answering (VQA), a machine must answer a question given an associated image. Recently, accessibility researchers have explored whether VQA can be deployed in a real-world setting where users with visual impairments learn about their environment by capturing their visual surroundings and asking questions. However, most of the existing benchmarking datasets for VQA focus on machine “understanding” and it remains unclear how progress on those datasets corresponds to improvements in this real-world use case. We aim to answer this question by evaluating discrepancies between machine “understanding” datasets (VQA-v2) and accessibility datasets (VizWiz) by evaluating a variety of VQA models. Based on our findings, we discuss opportunities and challenges in VQA for accessibility and suggest directions for future work.

Task-oriented conversational agents are gaining immense popularity and success in a wide range of tasks, from flight ticket booking to online shopping. However, the existing systems presume that end-users will always have a pre-determined and servable task goal, which results in dialogue failure in hostile scenarios, such as goal unavailability. On the other hand, human agents accomplish users’ tasks even in a large number of goal unavailability scenarios by persuading them towards a very similar and servable goal. Motivated by the limitation, we propose and build a novel end-to-end multi-modal persuasive dialogue system incorporated with a personalized persuasive module aided goal controller and goal persuader. The goal controller recognizes goal conflicting/unavailability scenarios and formulates a new goal, while the goal persuader persuades users using a personalized persuasive strategy identified through dialogue context. We also present a novel automatic evaluation metric called Persuasiveness Measurement Rate (PMeR) for quantifying the persuasive capability of a conversational agent. The obtained improvements (both quantitative and qualitative) firmly establish the superiority and need of the proposed context-guided, personalized persuasive virtual agent over existing traditional task-oriented virtual agents. Furthermore, we also curated a multi-modal persuasive conversational dialogue corpus annotated with intent, slot, sentiment, and dialogue act for e-commerce domain.

Summarization of legal case judgement documents is a challenging problem in Legal NLP. However, not much analyses exist on how different families of summarization models (e.g., extractive vs. abstractive) perform when applied to legal case documents. This question is particularly important since many recent transformer-based abstractive summarization models have restrictions on the number of input tokens, and legal documents are known to be very long. Also, it is an open question on how best to evaluate legal case document summarization systems. In this paper, we carry out extensive experiments with several extractive and abstractive summarization methods (both supervised and unsupervised) over three legal summarization datasets that we have developed. Our analyses, that includes evaluation by law practitioners, lead to several interesting insights on legal summarization in specific and long document summarization in general.

The current classification methods for relation extraction (RE) generally utilize pre-trained language models (PLMs) and have achieved superior results. However, such methods directly treat relation labels as class numbers, therefore they ignore the semantics of relation labels. Recently, prompt-based fine-tuning has been proposed and attracted much attention. This kind of methods insert templates into the input and convert the classification task to a (masked) language modeling problem. With this inspiration, we propose a novel method Fine-tuning with Prompt Curriculum (FPC) for RE, with two distinctive characteristics: the relation prompt learning, introducing an auxiliary prompt-based fine-tuning task to make the model capture the semantics of relation labels; the prompt learning curriculum, a fine-tuning procedure including an increasingly difficult task to adapt the model to the difficult multi-task setting. We have conducted extensive experiments on four widely used RE benchmarks under fully supervised and low-resource settings. The experimental results show that FPC can significantly outperform the existing methods and obtain the new state-of-the-art results.

Different linguistic expressions can conceptualize the same event from different viewpoints by emphasizing certain participants over others. Here, we investigate a case where this has social consequences: how do linguistic expressions of gender-based violence (GBV) influence who we perceive as responsible? We build on previous psycholinguistic research in this area and conduct a large-scale perception survey of GBV descriptions automatically extracted from a corpus of Italian newspapers. We then train regression models that predict the salience of GBV participants with respect to different dimensions of perceived responsibility. Our best model (fine-tuned BERT) shows solid overall performance, with large differences between dimensions and participants: salient _focus_ is more predictable than salient _blame_, and perpetrators’ salience is more predictable than victims’ salience. Experiments with ridge regression models using different representations show that features based on linguistic theory similarly to word-based features. Overall, we show that different linguistic choices do trigger different perceptions of responsibility, and that such perceptions can be modelled automatically. This work can be a core instrument to raise awareness of the consequences of different perspectivizations in the general public and in news producers alike.

The task of event extraction (EE) aims to find the events and event-related argument information from the text and represent them in a structured format. Most previous works try to solve the problem by separately identifying multiple substructures and aggregating them to get the complete event structure. The problem with the methods is that it fails to identify all the interdependencies among the event participants (event-triggers, arguments, and roles). In this paper, we represent each event record in a unique tuple format that contains trigger phrase, trigger type, argument phrase, and corresponding role information. Our proposed pointer network-based encoder-decoder model generates an event tuple in each time step by exploiting the interactions among event participants and presenting a truly end-to-end solution to the EE task. We evaluate our model on the ACE2005 dataset, and experimental results demonstrate the effectiveness of our model by achieving competitive performance compared to the state-of-the-art methods.

There is an ongoing debate of whether neural network can grasp the quasi-regularities in languages like humans. In a typical quasi-regularity task, English past tense inflections, the neural network model has long been criticized that it learns only to generalize the most frequent pattern, but not the regular pattern, thus can not learn the abstract categories of regular and irregular and is dissimilar to human performance. In this work, we train a set of transformer models with different settings to examine their behavior on this task. The models achieved high accuracy on unseen regular verbs and some accuracy on unseen irregular verbs. The models’ performance on the regulars are heavily affected by type frequency and ratio but not token frequency and ratio, and vice versa for the irregulars. The different behaviors on the regulars and irregulars suggest that the models have some degree of symbolic learning on the regularity of the verbs. In addition, the models are weakly correlated with human behavior on nonce verbs. Although the transformer model exhibits some level of learning on the abstract category of verb regularity, its performance does not fit human data well suggesting that it might not be a good cognitive model.

Neural autoregressive sequence models smear the probability among many possible sequences including degenerate ones, such as empty or repetitive sequences. In this work, we tackle one specific case where the model assigns a high probability to unreasonably short sequences. We define the oversmoothing rate to quantify this issue. After confirming the high degree of oversmoothing in neural machine translation, we propose to explicitly minimize the oversmoothing rate during training. We conduct a set of experiments to study the effect of the proposed regularization on both model distribution and decoding performance. We use a neural machine translation task as the testbed and consider three different datasets of varying size. Our experiments reveal three major findings. First, we can control the oversmoothing rate of the model by tuning the strength of the regularization. Second, by enhancing the oversmoothing loss contribution, the probability and the rank of eos token decrease heavily at positions where it is not supposed to be. Third, the proposed regularization impacts the outcome of beam search especially when a large beam is used. The degradation of translation quality (measured in BLEU) with a large beam significantly lessens with lower oversmoothing rate, but the degradation compared to smaller beam sizes remains to exist. From these observations, we conclude that the high degree of oversmoothing is the main reason behind the degenerate case of overly probable short sequences in a neural autoregressive model.

Neural metrics have achieved impressive correlation with human judgements in the evaluation of machine translation systems, but before we can safely optimise towards such metrics, we should be aware of (and ideally eliminate) biases toward bad translations that receive high scores. Our experiments show that sample-based Minimum Bayes Risk decoding can be used to explore and quantify such weaknesses. When applying this strategy to COMET for en-de and de-en, we find that COMET models are not sensitive enough to discrepancies in numbers and named entities. We further show that these biases are hard to fully remove by simply training on additional synthetic data and release our code and data for facilitating further experiments.

Authorship attribution is the task of identifying the author of a given text. The key is finding representations that can differentiate between authors. Existing approaches typically use manually designed features that capture a dataset’s content and style, but these approaches are dataset-dependent and yield inconsistent performance across corpora. In this work, we propose to learn author-specific representations by fine-tuning pre-trained generic language representations with a contrastive objective (Contra-X). We show that Contra-X learns representations that form highly separable clusters for different authors. It advances the state-of-the-art on multiple human and machine authorship attribution benchmarks, enabling improvements of up to 6.8% over cross-entropy fine-tuning. However, we find that Contra-X improves overall accuracy at the cost of sacrificing performance for some authors. Resolving this tension will be an important direction for future work. To the best of our knowledge, we are the first to integrate contrastive learning with pre-trained language model fine-tuning for authorship attribution.

We present a novel method for higher-order dependency parsing which takes advantage of the general form of score functions written as arc-polynomials, a general framework which encompasses common higher-order score functions, and includes new ones. This method is based on non-linear optimization techniques, namely coordinate ascent and genetic search where we iteratively update a candidate parse. Updates are formulated as gradient-based operations, and are efficiently computed by auto-differentiation libraries. Experiments show that this method obtains results matching the recent state-of-the-art second order parsers on three standard datasets.

Questions regarding implicitness, ambiguity and underspecification are crucial for understanding the task validity and ethical concerns of multimodal image+text systems, yet have received little attention to date. This position paper maps out a conceptual framework to address this gap, focusing on systems which generate images depicting scenes from scene descriptions. In doing so, we account for how texts and images convey meaning differently. We outline a set of core challenges concerning textual and visual ambiguity, as well as risks that may be amplified by ambiguous and underspecified elements. We propose and discuss strategies for addressing these challenges, including generating visually ambiguous images, and generating a set of diverse images.

One characteristic that makes humans superior to modern artificially intelligent models is the ability to interpret images beyond what is visually apparent. Consider the following two natural language search queries – (i) “a queue of customers patiently waiting to buy ice cream” and (ii) “a queue of tourists going to see a famous Mughal architecture in India”. Interpreting these queries requires one to reason with (i) Commonsense such as interpreting people as customers or tourists, actions as waiting to buy or going to see; and (ii) Fact or world knowledge associated with named visual entities, for example, whether the store in the image sells ice cream or whether the landmark in the image is a Mughal architecture located in India. Such reasoning goes beyond just visual recognition. To enable both commonsense and factual reasoning in the image search, we present a unified framework namely Knowledge Retrieval-Augmented Multimodal Transformer (KRAMT) that treats the named visual entities in an image as a gateway to encyclopedic knowledge and leverages them along with natural language query to ground relevant knowledge. Further, KRAMT seamlessly integrates visual content and grounded knowledge to learn alignment between images and search queries. This unified framework is then used to perform image search requiring commonsense and factual reasoning. The retrieval performance of KRAMT is evaluated and compared with related approaches on a new dataset we introduce – namely COFAR.