Previous studies on question generation from videos have mostly focused on generating questions about common objects and attributes and hence are not entity-centric. In this work, we focus on the generation of entity-centric information-seeking questions from videos. Such a system could be useful for video-based learning, recommending “People Also Ask” questions, video-based chatbots, and fact-checking. Our work addresses three key challenges: identifying question-worthy information, linking it to entities, and effectively utilizing multimodal signals. Further, to the best of our knowledge, there does not exist a large-scale dataset for this task. Most video question generation datasets are on TV shows, movies, or human activities or lack entity-centric information-seeking questions. Hence, we contribute a diverse dataset of YouTube videos, VideoQuestions, consisting of 411 videos with 2265 manually annotated questions. We further propose a model architecture combining Transformers, rich context signals (titles, transcripts, captions, embeddings), and a combination of cross-entropy and contrastive loss function to encourage entity-centric question generation. Our best method yields BLEU, ROUGE, CIDEr, and METEOR scores of 71.3, 78.6, 7.31, and 81.9, respectively, demonstrating practical usability. We make the code and dataset publicly available.

Lack of diverse perspectives causes neutrality bias in Wikipedia content leading to millions of worldwide readers getting exposed by potentially inaccurate information. Hence, neutrality bias detection and mitigation is a critical problem. Although previous studies have proposed effective solutions for English, no work exists for Indian languages. First, we contribute two large datasets, mWIKIBIAS and mWNC, covering 8 languages, for the bias detection and mitigation tasks respectively. Next, we investigate the effectiveness of popular multilingual Transformer-based models for the two tasks by modeling detection as a binary classification problem and mitigation as a style transfer problem. We make the code and data publicly available.

We focus on the novel problem of persona based dialogue generation for comic strips. Dialogs in comic strips is a unique and unexplored area where every strip contains utterances from various characters with each one building upon the previous utterances and the associated visual scene. Previous works like DialoGPT, PersonaGPT and other dialog generation models encode two-party dialogues and do not account for the visual information. To the best of our knowledge we are the first to propose the paradigm of multimodal persona based dialogue generation. We contribute a novel dataset, ComSet, consisting of 54K strips, harvested from 13 popular comics available online. Further, we propose a multimodal persona-based architecture, MPDialog, to generate dialogues for the next panel in the strip which decreases the perplexity score by ~10 points over strong dialogue generation baseline models. We demonstrate that there is still ample opportunity for improvement, highlighting the importance of building stronger dialogue systems that are able to generate persona-consistent dialogues and understand the context through various modalities.

Syntactic parsing is the task of assigning a syntactic structure to a sentence. There are two popular syntactic parsing methods: constituency and dependency parsing. Recent works have used syntactic embeddings based on constituency trees, incremental top-down parsing, and other word syntactic features for brain activity prediction given the text stimuli to study how the syntax structure is represented in the brain’s language network. However, the effectiveness of dependency parse trees or the relative predictive power of the various syntax parsers across brain areas, especially for the listening task, is yet unexplored. In this study, we investigate the predictive power of the brain encoding models in three settings: (i) individual performance of the constituency and dependency syntactic parsing based embedding methods, (ii) efficacy of these syntactic parsing based embedding methods when controlling for basic syntactic signals, (iii) relative effectiveness of each of the syntactic embedding methods when controlling for the other. Further, we explore the relative importance of syntactic information (from these syntactic embedding methods) versus semantic information using BERT embeddings. We find that constituency parsers help explain activations in the temporal lobe and middle-frontal gyrus, while dependency parsers better encode syntactic structure in the angular gyrus and posterior cingulate cortex. Although semantic signals from BERT are more effective compared to any of the syntactic features or embedding methods, syntactic embedding methods explain additional variance for a few brain regions.

The use of large language models (LLMs) in natural language processing (NLP) tasks is rapidly increasing, leading to changes in how researchers approach problems in the field. To fully utilize these models’ abilities, a better understanding of their behavior for different input protocols is required. With LLMs, users can directly interact with the models through a text-based interface to define and solve various tasks. Hence, understanding the conversational abilities of these LLMs, which may not have been specifically trained for dialog modeling, is also important. This study examines different approaches for building dialog systems using LLMs by considering various aspects of the prompt. As part of prompt tuning, we experiment with various ways of providing instructions, exemplars, current query and additional context. The research also analyzes the representations of dialog history that have the optimal usable-information density. Based on the findings, the paper suggests more compact ways of providing dialog history information while ensuring good performance and reducing model’s inference-API costs. The research contributes to a better understanding of how LLMs can be effectively used for building interactive systems.

Transformer-based pretrained models like BERT, GPT-2 and T5 have been finetuned for a large number of natural language processing (NLP) tasks, and have been shown to be very effective. However, while finetuning, what changes across layers in these models with respect to pretrained checkpoints is under-studied. Further, how robust are these models to perturbations in input text? Does the robustness vary depending on the NLP task for which the models have been finetuned? While there exists some work on studying robustness of BERT finetuned for a few NLP tasks, there is no rigorous study which compares this robustness across encoder only, decoder only and encoder-decoder models. In this paper, we characterize changes between pretrained and finetuned language model representations across layers using two metrics: CKA and STIR. Further, we study the robustness of three language models (BERT, GPT-2 and T5) with eight different text perturbations on classification tasks from General Language Understanding Evaluation (GLUE) benchmark, and generation tasks like summarization, free-form generation and question generation. GPT-2 representations are more robust than BERT and T5 across multiple types of input perturbation. Although models exhibit good robustness broadly, dropping nouns, verbs or changing characters are the most impactful. Overall, this study provides valuable insights into perturbation-specific weaknesses of popular Transformer-based models which should be kept in mind when passing inputs.

Multiple business scenarios require an automated generation of descriptive human-readable text from structured input data. This has resulted into substantial work on fact-to-text generation systems recently. Unfortunately, previous work on fact-to-text (F2T) generation has focused primarily on English mainly due to the high availability of relevant datasets. Only recently, the problem of cross-lingual fact-to-text (XF2T) was proposed for generation across multiple languages alongwith a dataset, XAlign for eight languages. However, there has been no rigorous work on the actual XF2T generation problem. We extend XAlign dataset with annotated data for four more languages: Punjabi, Malayalam, Assamese and Oriya. We conduct an extensive study using popular Transformer-based text generation models on our extended multi-lingual dataset, which we call XAlignV2. Further, we investigate the performance of different text generation strategies: multiple variations of pretraining, fact-aware embeddings and structure-aware input encoding. Our extensive experiments show that a multi-lingual mT5 model which uses fact-aware embeddings with structure-aware input encoding leads to best results (30.90 BLEU, 55.12 METEOR and 59.17 chrF++) across the twelve languages. We make our code, dataset and model publicly available, and hope that this will help advance further research in this critical area.

Although many pretrained models exist for text or images, there have been relatively fewer attempts to train representations specifically for dialog understanding. Prior works usually relied on finetuned representations based on generic text representation models like BERT or GPT-2. But such language modeling pretraining objectives do not take the structural information of conversational text into consideration. Although generative dialog models can learn structural features too, we argue that the structure-unaware word-by-word generation is not suitable for effective conversation modeling. We empirically demonstrate that such representations do not perform consistently across various dialog understanding tasks. Hence, we propose a structure-aware Mutual Information based loss-function DMI (Discourse Mutual Information) for training dialog-representation models, that additionally captures the inherent uncertainty in response prediction. Extensive evaluation on nine diverse dialog modeling tasks shows that our proposed DMI-based models outperform strong baselines by significant margins.

Several popular Transformer based language models have been found to be successful for text-driven brain encoding. However, existing literature leverages only pretrained text Transformer models and has not explored the efficacy of task-specific learned Transformer representations. In this work, we explore transfer learning from representations learned for ten popular natural language processing tasks (two syntactic and eight semantic) for predicting brain responses from two diverse datasets: Pereira (subjects reading sentences from paragraphs) and Narratives (subjects listening to the spoken stories). Encoding models based on task features are used to predict activity in different regions across the whole brain. Features from coreference resolution, NER, and shallow syntax parsing explain greater variance for the reading activity. On the other hand, for the listening activity, tasks such as paraphrase generation, summarization, and natural language inference show better encoding performance. Experiments across all 10 task representations provide the following cognitive insights: (i) language left hemisphere has higher predictive brain activity versus language right hemisphere, (ii) posterior medial cortex, temporo-parieto-occipital junction, dorsal frontal lobe have higher correlation versus early auditory and auditory association cortex, (iii) syntactic and semantic tasks display a good predictive performance across brain regions for reading and listening stimuli resp.

Table-to-text generation has been widely studied in the Natural Language Processing community in the recent years. We give a new perspective to this problem by incorporating signals from both tables as well as associated images to generate relevant text. While tables contain a structured list of facts, images are a rich source of unstructured visual information. For example, in the tourism domain, images can be used to infer knowledge such as the type of landmark (e.g., church), its architecture (e.g., Ancient Roman), and composition (e.g., white marble). Therefore, in this paper, we introduce the novel task of Vision-augmented Table-To-Text Generation (VisToT, defined as follows: given a table and an associated image, produce a descriptive sentence conditioned on the multimodal input. For the task, we present a novel multimodal table-to-text dataset, WikiLandmarks, covering 73,084 unique world landmarks. Further, we also present a competitive architecture, namely, VT3 that generates accurate sentences conditioned on the image and table pairs. Through extensive analyses and experiments, we show that visual cues from images are helpful in (i) inferring missing information from incomplete or sparse tables, and (ii) strengthening the importance of useful information from noisy tables for natural language generation. We make the code and data publicly available.

Can we build multi-view decoders that can decode concepts from brain recordings corresponding to any view (picture, sentence, word cloud) of stimuli? Can we build a system that can use brain recordings to automatically describe what a subject is watching using keywords or sentences? How about a system that can automatically extract important keywords from sentences that a subject is reading? Previous brain decoding efforts have focused only on single view analysis and hence cannot help us build such systems. As a first step toward building such systems, inspired by Natural Language Processing literature on multi-lingual and cross-lingual modeling, we propose two novel brain decoding setups: (1) multi-view decoding (MVD) and (2) cross-view decoding (CVD). In MVD, the goal is to build an MV decoder that can take brain recordings for any view as input and predict the concept. In CVD, the goal is to train a model which takes brain recordings for one view as input and decodes a semantic vector representation of another view. Specifically, we study practically useful CVD tasks like image captioning, image tagging, keyword extraction, and sentence formation. Our extensive experiments lead to MVD models with ~0.68 average pairwise accuracy across view pairs, and also CVD models with ~0.8 average pairwise accuracy across tasks. Analysis of the contribution of different brain networks reveals exciting cognitive insights: (1) Models trained on picture or sentence view of stimuli are better MV decoders than a model trained on word cloud view. (2) Our extensive analysis across 9 broad regions, 11 language sub-regions and 16 visual sub-regions of the brain help us localize, for the first time, the parts of the brain involved in cross-view tasks like image captioning, image tagging, sentence formation and keyword extraction. We make the code publicly available.

Brain encoding aims at reconstructing fMRI brain activity given a stimulus. There exists a plethora of neural encoding models which study brain encoding for single mode stimuli: visual (pretrained CNNs) or text (pretrained language models). Few recent papers have also obtained separate visual and text representation models and performed late-fusion using simple heuristics. However, previous work has failed to explore the co-attentive multi-modal modeling for visual and text reasoning. In this paper, we systematically explore the efficacy of image and multi-modal Transformers for brain encoding. Extensive experiments on two popular datasets, BOLD5000 and Pereira, provide the following insights. (1) We find that VisualBERT, a multi-modal Transformer, significantly outperforms previously proposed single-mode CNNs, image Transformers as well as other previously proposed multi-modal models, thereby establishing new state-of-the-art. (2) The regions such as LPTG, LMTG, LIFG, and STS which have dual functionalities for language and vision, have higher correlation with multi-modal models which reinforces the fact that these models are good at mimicing the human brain behavior. (3) The supremacy of visio-linguistic models raises the question of whether the responses elicited in the visual regions are affected implicitly by linguistic processing even when passively viewing images. Future fMRI tasks can verify this computational insight in an appropriate experimental setting. We make our code publicly available.

Automatic text summarization has been widely studied as an important task in natural language processing. Traditionally, various feature engineering and machine learning based systems have been proposed for extractive as well as abstractive text summarization. Recently, deep learning based, specifically Transformer-based systems have been immensely popular. Summarization is a cognitively challenging task – extracting summary worthy sentences is laborious, and expressing semantics in brief when doing abstractive summarization is complicated. In this paper, we specifically look at the problem of summarizing scientific research papers from multiple domains. We differentiate between two types of summaries, namely, (a) LaySumm: A very short summary that captures the essence of the research paper in layman terms restricting overtly specific technical jargon and (b) LongSumm: A much longer detailed summary aimed at providing specific insights into various ideas touched upon in the paper. While leveraging latest Transformer-based models, our systems are simple, intuitive and based on how specific paper sections contribute to human summaries of the two types described above. Evaluations against gold standard summaries using ROUGE metrics prove the effectiveness of our approach. On blind test corpora, our system ranks first and third for the LongSumm and LaySumm tasks respectively.

Hoping for a large number of clicks and potentially high social shares, journalists of various news media outlets publish sensationalist headlines on social media. These headlines lure the readers to click on them and satisfy the curiosity gap in their mind. Low quality material pointed to by clickbaits leads to time wastage and annoyance for users. Even for enterprises publishing clickbaits, it hurts more than it helps as it erodes user trust, attracts wrong visitors, and produces negative signals for ranking algorithms. Hence, identifying and flagging clickbait titles is very essential. Previous work on clickbaits has majorly focused on binary classification of clickbait titles. However not all clickbaits are equally clickbaity. It is not only essential to identify a click-bait, but also to identify the intensity of the clickbait based on the strength of the clickbait. In this work, we model clickbait strength prediction as a regression problem. While previous methods have relied on traditional machine learning or vanilla recurrent neural networks, we rigorously investigate the use of transformers for clickbait strength prediction. On a benchmark dataset with ∼39K posts, our methods outperform all the existing methods in the Clickbait Challenge.

While extensive popularity of online social media platforms has made information dissemination faster, it has also resulted in widespread online abuse of different types like hate speech, offensive language, sexist and racist opinions, etc. Detection and curtailment of such abusive content is critical for avoiding its psychological impact on victim communities, and thereby preventing hate crimes. Previous works have focused on classifying user posts into various forms of abusive behavior. But there has hardly been any focus on estimating the severity of abuse and the target. In this paper, we present a first of the kind dataset with 7,601 posts from Gab which looks at online abuse from the perspective of presence of abuse, severity and target of abusive behavior. We also propose a system to address these tasks, obtaining an accuracy of ∼80% for abuse presence, ∼82% for abuse target prediction, and ∼65% for abuse severity prediction.

This paper describes our system (Fermi) for Task 5 of SemEval-2019: HatEval: Multilingual Detection of Hate Speech Against Immigrants and Women on Twitter. We participated in the subtask A for English and ranked first in the evaluation on the test set. We evaluate the quality of multiple sentence embeddings and explore multiple training models to evaluate the performance of simple yet effective embedding-ML combination algorithms. Our team - Fermi’s model achieved an accuracy of 65.00% for English language in task A. Our models, which use pretrained Universal Encoder sentence embeddings for transforming the input and SVM (with RBF kernel) for classification, scored first position (among 68) in the leaderboard on the test set for Subtask A in English language. In this paper we provide a detailed description of the approach, as well as the results obtained in the task.

This paper describes our system (Fermi) for Task 6: OffensEval: Identifying and Categorizing Offensive Language in Social Media of SemEval-2019. We participated in all the three sub-tasks within Task 6. We evaluate multiple sentence embeddings in conjunction with various supervised machine learning algorithms and evaluate the performance of simple yet effective embedding-ML combination algorithms. Our team Fermi’s model achieved an F1-score of 64.40%, 62.00% and 62.60% for sub-task A, B and C respectively on the official leaderboard. Our model for sub-task C which uses pre-trained ELMo embeddings for transforming the input and uses SVM (RBF kernel) for training, scored third position on the official leaderboard. Through the paper we provide a detailed description of the approach, as well as the results obtained for the task.

Online Community Question Answering Forums (cQA) have gained massive popularity within recent years. The rise in users for such forums have led to the increase in the need for automated evaluation for question comprehension and fact evaluation of the answers provided by various participants in the forum. Our team, Fermi, participated in sub-task A of Task 8 at SemEval 2019 - which tackles the first problem in the pipeline of factual evaluation in cQA forums, i.e., deciding whether a posed question asks for a factual information, an opinion/advice or is just socializing. This information is highly useful in segregating factual questions from non-factual ones which highly helps in organizing the questions into useful categories and trims down the problem space for the next task in the pipeline for fact evaluation among the available answers. Our system uses the embeddings obtained from Universal Sentence Encoder combined with XGBoost for the classification sub-task A. We also evaluate other combinations of embeddings and off-the-shelf machine learning algorithms to demonstrate the efficacy of the various representations and their combinations. Our results across the evaluation test set gave an accuracy of 84% and received the first position in the final standings judged by the organizers.

Sexism, an injustice that subjects women and girls to enormous suffering, manifests in blatant as well as subtle ways. In the wake of growing documentation of experiences of sexism on the web, the automatic categorization of accounts of sexism has the potential to assist social scientists and policy makers in utilizing such data to study and counter sexism better. The existing work on sexism classification, which is different from sexism detection, has certain limitations in terms of the categories of sexism used and/or whether they can co-occur. To the best of our knowledge, this is the first work on the multi-label classification of sexism of any kind(s), and we contribute the largest dataset for sexism categorization. We develop a neural solution for this multi-label classification that can combine sentence representations obtained using models such as BERT with distributional and linguistic word embeddings using a flexible, hierarchical architecture involving recurrent components and optional convolutional ones. Further, we leverage unlabeled accounts of sexism to infuse domain-specific elements into our framework. The best proposed method outperforms several deep learning as well as traditional machine learning baselines by an appreciable margin.

Relation Schema Induction (RSI) is the problem of identifying type signatures of arguments of relations from unlabeled text. Most of the previous work in this area have focused only on binary RSI, i.e., inducing only the subject and object type signatures per relation. However, in practice, many relations are high-order, i.e., they have more than two arguments and inducing type signatures of all arguments is necessary. For example, in the sports domain, inducing a schema win(WinningPlayer, OpponentPlayer, Tournament, Location) is more informative than inducing just win(WinningPlayer, OpponentPlayer). We refer to this problem as Higher-order Relation Schema Induction (HRSI). In this paper, we propose Tensor Factorization with Back-off and Aggregation (TFBA), a novel framework for the HRSI problem. To the best of our knowledge, this is the first attempt at inducing higher-order relation schemata from unlabeled text. Using the experimental analysis on three real world datasets we show how TFBA helps in dealing with sparsity and induce higher-order schemata.

Translation Quality Estimation (QE) aims to estimate the quality of an automated machine translation (MT) output without any human intervention or reference translation. With the increasing use of MT systems in various cross-lingual applications, the need and applicability of QE systems is increasing. We study existing approaches and propose multiple neural network approaches for sentence-level QE, with a focus on MT outputs in Indian languages. For this, we also introduce five new datasets for four language pairs: two for English–Gujarati, and one each for English–Hindi, English–Telugu and English–Bengali, which includes one manually post-edited dataset for English– Gujarati. These Indian languages are spoken by around 689M speakers world-wide. We compare results obtained using our proposed models with multiple state-of-the-art systems including the winning system in the WMT17 shared task on QE and show that our proposed neural model which combines the discriminative power of carefully chosen features with Siamese Convolutional Neural Networks (CNNs) works best for all Indian language datasets.

Ideally a metric evaluating an abstract system summary should represent the extent to which the system-generated summary approximates the semantic inference conceived by the reader using a human-written reference summary. Most of the previous approaches relied upon word or syntactic sub-sequence overlap to evaluate system-generated summaries. Such metrics cannot evaluate the summary at semantic inference level. Through this work we introduce the metric of Semantic Similarity for Abstractive Summarization (SSAS), which leverages natural language inference and paraphrasing techniques to frame a novel approach to evaluate system summaries at semantic inference level. SSAS is based upon a weighted composition of quantities representing the level of agreement, contradiction, independence, paraphrasing, and optionally ROUGE score between a system-generated and a human-written summary.