Publications

2025

1. ACL 25

   DPO Kernels: A Semantically-Aware, Kernel-Enhanced, and Divergence-Rich Paradigm for Direct Preference Optimization

   Amitava Das, Sameer Trivedy, Deepansh Khanna, and 8 more authors

   In Proceedings of ACL 2025, 2025

   A*

   Abs PDF

   We present DPO Kernels, combining kernel density estimation with divergence-based objectives for direct preference optimization in large language models. By incorporating semantics via custom kernel functions, DPO Kernels significantly improve the alignment between model outputs and human preferences. Experiments on public preference datasets show gains of 3-5% over existing DPO approaches, demonstrating improved generation quality and user satisfaction.

2. ACL 25

   YINYANG-ALIGN: Benchmarking Contradictory Objectives and Proposing Multi-Objective Optimization based DPO for Text-to-Image Alignment

   Amitava Das, Yogesh Narsupalli, Gautam Singh, and 5 more authors

   In Proceedings of ACL 2025, 2025

   A*

   Abs PDF

   YINYANG-ALIGN is a novel benchmark and training framework for text-to-image alignment that concurrently optimizes for coherence and diversity, conceptually contradictory objectives. We propose multi-objective DPO training that balances these via Pareto optimality. Evaluated on standard T2I datasets, YINYANG-ALIGN achieves higher user preference and diversity metrics without sacrificing fidelity.

3. COLING 25

   Exploring the Abilities of Large Language Models to Solve Proportional Analogies via Knowledge-Enhanced Prompting

   T. Wijesiriwardene, R. Wickramarachchi, S. R. Vennam, and 5 more authors

   In Proceedings of COLING 2025, 2025

   B

   Abs PDF

   This work investigates LLMs’ capability on analogical reasoning (e.g., A:B :: C:? queries) with knowledge-enhanced prompts. We design knowledge retrieval modules that inject relevant facts into prompts, enhancing model performance by  12% on proportional analogy datasets. Analyses reveal improvements when symbolic knowledge complements contextual embeddings.

4. ICCV 25

   PECCAVI - Visual Paraphrase Attack Safe and Distortion-Free Image Watermarking Technique for AI-Generated Images

   Siddharth Dixit, Adeel Aziz, Vivek Sharma, and 4 more authors

   In ICCV 2025 (submitted), 2025

   A*

   Abs

   PECCAVI introduces a novel distortion-free watermarking scheme resilient to visual paraphrasing attacks, where an image’s semantics are preserved but styling is changed. Through dual-space embedding and signal-preserving transformations, PECCAVI achieves over 90% watermark robustness under various paraphrasing attack scenarios, while maintaining visual fidelity.

2024

1. AAAI 2024

   Overview of Factify5WQA: Fact Verification through 5W Question-Answering

   Suryavardan Suresh, Anku Rani, Parth Patwa, and 6 more authors

   2024

   Abs

   Researchers have found that fake news spreads much faster than real news, especially on platforms used widely by younger demographics. Manual fact-checking is labor-intensive and can’t match the pace of misinformation. The Factify5WQA shared task addresses this by promoting an automated, explainable fact-verification pipeline where claims are paired with supporting documents and evaluated through 5W-based (Who, What, When, Where, Why) question answering. Answers are evaluated using BLEU followed by accuracy metrics. With a  35% improvement over random baseline, the best-performing system achieved 69.56% accuracy.

2. TrustNLP 25

   “Sorry, Come Again?” Prompting - Enhancing Comprehension and Diminishing Hallucination with [PAUSE]-Injected Optimal Paraphrasing

   Vipula Rawte, S. M. T. Tonmoy, S. M. Zaman, and 4 more authors

   2024

   Abs

   Hallucination remains a critical weakness of large language models (LLMs). We propose the “Sorry, Come Again” (SCA) prompting framework, which combines optimal paraphrasing and strategic insertion of [PAUSE] tokens to slow down generation and promote deeper processing. We analyze linguistic traits such as readability, formality, and concreteness across prompts on 21 LLMs, and show that low clarity leads to more hallucinations. We then identify optimal paraphrases using integrated gradients, and fine-tune models with [PAUSE]-injected inputs. This dual approach notably improves coherence and factuality in generated outputs.

3. arxiv 24

   The What, Why, and How of Context Length Extension Techniques in Large Language Models – A Detailed Survey

   Saurav Pawar, S. M. T. Islam Tonmoy, S. M. Zaman, and 3 more authors

   2024

   Abs

   Large language models (LLMs) excel in text comprehension and generation, but struggle with limited context windows. This survey dissects the motivations behind context-length extension, evaluates existing techniques—including memory augmentation, sliding-window attention, recurrence, positional interpolation, and retrieval-based methods—examines current evaluation standards, and highlights unresolved challenges. It serves as a comprehensive resource for researchers navigating the evolving long-context landscape.

4. arxiv 24

   A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models

   S. M. T. Islam Tonmoy, S. M. Mehedi Zaman, Vinija Jain, and 4 more authors

   2024

   Abs

   As LLMs advance, hallucination—the generation of content that seems factual but is ungrounded—remains a critical obstacle to safe deployment. This survey reviews over 32 mitigation methods, including retrieval-augmented generation (RAG), knowledge retrieval, CoNLI, and CoVe. We propose a taxonomy structured around dataset utilization, task type, feedback mechanisms, and retriever architecture, and discuss strengths, weaknesses, and open challenges.

5. arxiv 24

   Beyond the Limits: A Survey of Techniques to Extend the Context Length in Large Language Models

   Xindi Wang, Mahsa Salmani, Parsa Omidi, and 3 more authors

   2024

   Abs

   Recently, large language models (LLMs) have shown remarkable capabilities in understanding, reasoning, and generating language, but their ability to process long input sequences is constrained by computational and memory costs. This survey comprehensively reviews techniques for extending context length—such as architectural modifications, altered positional encodings, and novel attention mechanisms—across training, fine-tuning, and inference stages. We classify methods, discuss their trade-offs, and highlight limitations and future research directions.

6. arxiv 24

   A Survey on Hallucination in Large Vision-Language Models

   Hanchao Liu, Wenyuan Xue, Yifei Chen, and 6 more authors

   2024

   Abs

   Large vision-language models (LVLMs) have advanced multi-modal AI applications, but often suffer from “hallucination”—a misalignment between visual input and textual output. In this survey, we clarify hallucination in LVLMs, categorize different symptoms, review existing benchmarks and evaluation methods, analyze root causes (e.g., training data, model components), and summarize mitigation techniques. We conclude with open challenges and future directions for reducing multimodal hallucinations.

7. EMNLP 24

   Counter Turing Test (CT2): Investigating AI-Generated Text Detection for Hindi

   Isha Kavathekar, Anku Rani, Ananya Chamoli, and 3 more authors

   In Findings of EMNLP 2024, 2024

   Abs PDF

   We introduce the Hindi AI Detectability Index (ADIhi), a benchmark suite and evaluation metric for detecting AI-generated Hindi text. We compile a corpus of 20K human-written and LLM-generated texts and evaluate detectors including token-level, syntactic, and stylometric features. Results indicate existing detectors struggle on code-mixed content and slang; using ADIhi, we rank LLMs and propose improvements tailored for Hindi.

8. EACL 24

   On the Relationship between Sentence Analogy Identification and Sentence Structure Encoding in Large Language Models

   T. Wijesiriwardene, R. Wickramarachchi, Aishwarya Reganti, and 4 more authors

   In EACL 2024 Findings, 2024

   A

   Abs

   This study examines whether pre-trained LLMs capture sentence structural features necessary for solving analogies (e.g., “The cat sat on the mat” : “The dog lay in the yard”). We design probing tasks and evaluate models like BERT and LLaMA. Results suggest strong correlation between structural encoding and analogy accuracy; fine-tuning on structural signals improves performance by  8%.

9. Neurocomputing 24

   KI-MAG: A knowledge-infused abstractive question answering system in medical domain

   A. Zafar, S.K. Sahoo, H. Bhardawaj, and 2 more authors

   Neurocomputing, 2024

   Abs DOI

   The KI-MAG system integrates BART-based abstractive QA with medical knowledge graphs, improving answer precision by  15%. Evaluation on synthetic and real medical QA shows increased semantic correctness and reduced hallucinations by embedding entities from UMLS and PharmKG into its generation pipeline.

10. Applied Intelligence 24

    Are my answers medically accurate? Exploiting medical knowledge graphs for medical question answering

    A. Zafar, D. Varshney, S.K. Sahoo, and 2 more authors

    Applied Intelligence, 2024

    Abs DOI

    This work introduces a knowledge-enabled QA model leveraging Medical Entity Scoring (MES) and Context Relevance Scoring (CRS) to enrich answers with medical knowledge from UMLS and PharmKG. Evaluations on MASH-QA & COVID-QA show state-of-the-art results in both automated and human assessment, achieving improved factuality and relevance.

11. Intelligent Information Systems 24

    KIMedQA: towards building knowledge-enhanced medical QA models

    A. Zafar, S.K. Sahoo, D. Varshney, and 2 more authors

    Journal of Intelligent Information Systems, 2024

    Abs DOI

    KIMedQA integrates pruned medical knowledge graphs with context and query representations using pre-trained LMs to generate informed answers. Evaluated on MASH-QA and COVID-QA, the model establishes new state-of-the-art performance by selectively filtering relevant KG subgraphs.

2023

1. ACL 25

   SEPSIS: I Can Catch Your Lies – A New Paradigm for Deception Detection

   Anku Rani, Dwip Dalal, Shreya Gautam, and 5 more authors

   2023

   Abs

   Deception is the intentional practice of twisting information. This research explores three forms: lies of omission, lies of commission, and lies of influence, focusing primarily on lies of omission. We introduce a novel NLP framework supported by an annotated dataset of 876,784 samples from merged fake-news sources and Times of India headlines. Each sample is labeled for omission type (e.g., speculation, bias), “color” of lie (black/white/gray/red), intent (influence, prestige), and topic (political, educational, religious, etc.). Our multi-task learning pipeline, which merges fine-tuned language models, attains an F1 score of 0.87 across all annotation layers. We also examine correlations between omission-based lies and propaganda techniques—e.g., loaded language often signals opinion—and will release both model and dataset under an open license.

2. arxiv 24

   A Survey of Hallucination in Large Foundation Models

   Vipula Rawte, Amit Sheth, and Amitava Das

   2023

   Abs

   Hallucination in foundation models (text, vision, audio, video) refers to output that is plausible but factually mistaken. This paper provides a comprehensive overview: defining hallucination types (intrinsic/extrinsic), categorizing causes, evaluating mitigation strategies, and offering a unified framework for assessing and addressing hallucination across modalities.

3. arxiv 23

   Exploring the Relationship between LLM Hallucinations and Prompt Linguistic Nuances: Readability, Formality, and Concreteness

   Vipula Rawte, Prachi Priya, S. M. T. Islam Tonmoy, and 3 more authors

   2023

   Abs

   As large language models (LLMs) have advanced, they display a form of “hallucination”—producing plausible yet factually incorrect information. In this exploratory study, we examine how prompt characteristics—readability, formality, and concreteness—influence hallucination rates. We compile an annotated dataset derived from New York Times event-related tweets and evaluate hallucination occurrences across multiple LLMs. Our findings suggest that prompts with higher formality and concreteness significantly reduce hallucination, while the impact of readability is mixed. Results highlight the importance of prompt design in mitigating factual errors in LLM outputs.

4. arxiv 23

   RESTORE: Graph Embedding Assessment Through Reconstruction

   Hiu Ying Yip, Chandra Ravuru, Neeloy Banerjee, and 4 more authors

   2023

   Abs

   Following the success of word embeddings, graph embeddings (GEs) have grown in popularity—but intrinsic evaluation methods are lacking. We propose RESTORE, an intrinsic evaluation framework that assesses graph embeddings via their ability to reconstruct the original graph’s topology and semantics. We evaluate on multiple embedding families (factorization-, random-walk-, and deep learning-based) using the CommonSense Knowledge Graph. Results show (1) deep-learning embeddings (SDNE) perform better on multi-hop topological reconstruction (mAP of 0.54 for 2 hops, 0.35 for 3 hops), and (2) factorization-based methods (HOPE) better capture semantic similarity. Our framework offers insights into embedding fidelity beyond downstream task performance.

5. arxiv 23

   A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions

   Lei Huang, Weijiang Yu, Weitao Ma, and 8 more authors

   2023

   Abs

   Given the generic and open-ended nature of LLMs, hallucinations present unique challenges compared to task-specific models. In this survey, we propose an innovative taxonomy of hallucinations, analyze contributing factors, and provide a thorough overview of detection methods and benchmarks. We further assess mitigation strategies, identify limitations in retrieval-augmented LLMs, and discuss future directions to build more robust models.

6. EMNLP 23

   FACTIFY3M: A benchmark for multimodal fact verification with explainability through 5W Question-Answering

   Vipula Rawte and Amitava Das

   In Proceedings of EMNLP 2023, 2023

   Oral, A*

   Abs

   We present FACTIFY3M, a multimodal factual verification benchmark that combines images, text, and explanations grounded in 5W QA. The dataset includes 50K multimodal claims with annotated 5WQA pairs and supporting sources. Baseline models using joint vision-language transformers achieve 74.5% accuracy, but leave room for improvement; the 5W explanations improve interpretability and diagnostic evaluation.

7. EMNLP 23

   The Troubling Emergence of Hallucination in Large Language Models - An Extensive Definition, Quantification, and Prescriptive Remediations

   Vipula Rawte, S. Chakraborty, A. Pathak, and 5 more authors

   In Proceedings of EMNLP 2023, 2023

   Oral, A*

   Abs DOI PDF

   This paper introduces a formal definition and taxonomy for LLM hallucinations, proposes automatic metrics for hallucination quantification (e.g., hallucination density, factual error scores), and presents remediations such as knowledge grounding, self-critique modules, and multi-step prompting. In experiments across GPT-3, BERT, and open-source LLMs, these countermeasures reduce hallucination by up to 35%.

8. EMNLP 23

   Counter Turing Test (CT2): AI-Generated Text Detection is Not as Easy as You May Think

   Shreya Chakraborty, Aman Chadha, Aishwarya Sheth, and 1 more author

   In EMNLP 2023, 2023

   A*

   Abs PDF

   This paper demonstrates that standard token- and syntax-based detectors poorly generalize to AI-generated text in Hindi and code-mixed domains. We propose CT2 detection models that leverage stylometric and structural features, improving cross-domain robustness by 12%. We release CT2 dataset and models for future research.

9. ACL 23

   ANALOGICAL - A New Benchmark for Analogy of Long Text for Large Language Models

   T. Wijesiriwardene, R. Wickramarachchi, B. Gajera, and 6 more authors

   In Proceedings of ACL 2023, 2023

   Findings, A*

   Abs

   We introduce ANALOGICAL, the first benchmark targeting analogy reasoning in long-form text. It includes 5K human-crafted long-text analogies across narrative domains. Evaluation of LLMs (GPT-3, LLaMA) shows <40% accuracy. We analyze failure modes and propose structure-aware embeddings that improve performance by  15%.

10. ACL 23

    FACTIFY-5WQA: 5W Aspect-based Fact Verification through Question Answering

    A. Rani, S. M. Tonmoy, D. Dalal, and 5 more authors

    In Proceedings of ACL 2023, 2023

    Oral, A*

    Abs

    We propose FACTIFY-5WQA, a fact verification model that formulates evidence retrieval and claim assessment as an answer generation to “who/what/when/where/why” questions. Evaluated on FEVER and MultiFC, our model shows 5-8% gains over retrieval-based baselines, and provides interpretable outputs via generated 5W answers.

11. ACL 23

    CONFLATOR: Incorporating Switching Point-based Rotatory Positional Encodings for Code-Mixed Language Modeling

    M. Mohammed, S. Kandukuri, N. Gupta, and 5 more authors

    In 6th Workshop on Computational Approaches to Linguistic Code-Switching (AACL), ACL 2023, 2023

    A

    Abs

    CONFLATOR introduces a rotary positional encoding sensitive to code-switching points, enhancing modeling for bilingual text. Evaluated on English-Hindi-Bengali data, it reduces perplexity by 12% compared to standard RoPE and improves downstream tasks like language identification and translation.

12. Intelligent Systems 23

    Why Do We Need Neurosymbolic AI to Model Pragmatic Analogies?

    T. Wijesiriwardene, A. Sheth, V.L. Shalin, and 1 more author

    IEEE Intelligent Systems, 2023

    Abs DOI

    This article argues that pragmatic analogy understanding requires both symbolic logical reasoning and neural modeling. Neurosymbolic AI, combining deep learning and symbolic methods, is essential to model nuanced analogies in human communication.

2022

1. AAAI 22

   PESTO: Switching Point-Based Dynamic and Relative Positional Encoding for Code-Mixed Languages

   M. Ali, S. T. Kandukuri, S. Manduru, and 2 more authors

   In AAAI 2022, 2022

   Student Abstract, A*

   Abs

   PESTO designs dynamic positional encodings guided by code-switching points and integrates relative positions within mixed segments. This enhances modeling of bilingual utterances in English-Hindi and English-Bengali corpora, improving language modeling perplexity by 8% over baselines.

2. AAAI 22

   Memotion Analysis through the Lens of Joint Embedding

   N. Gunti, S. Ramamoorthy, P. Patwa, and 1 more author

   In AAAI 2022, 2022

   Student Abstract, A*

   Abs

   We present a joint embedding framework combining textual and visual features to analyze memes—detecting sentiment, humor, sarcasm, and offensiveness. On a curated English-Hindi meme dataset, our approach achieves over 70% accuracy across multiple tasks, outperforming text-only baselines by 12%.

3. Exper Systems 22

   Fake spreader is narcissist; Real spreader is Machiavellian prediction of fake news diffusion using psycho-sociological facets

   S.P.Y.K.L., A. Das, and V. Pulabaigari

   Expert Systems with Applications, 2022

   Abs DOI

   We explore psycho-sociological characteristics of fake vs real news spreaders via personality traits (Narcissism, Machiavellianism). Leveraging Twitter diffusion data, a predictive model identifies fake spreaders exhibiting high narcissism and low Machiavellianism, while real spreaders show opposite patterns. Results demonstrate the model’s effectiveness in predicting spreader type.

2021

1. IEEE ICDE 21

   Hate is the New Infodemic: A Topic-aware Modeling of Hate Speech Diffusion on Twitter

   S. Dutta, S. Masud, Sa. Makkar, and 4 more authors

   In IEEE ICDE 2021, 2021

   A1

   Abs

   We introduce a topic-aware diffusion model to study the spread of hate speech on Twitter, treating it as an infodemic. By integrating topic modeling with user engagement features (retweets, followers), we identify topics that significantly amplify hate. Evaluation on large-scale event-based data shows our model predicts hateful tweet virality with 85% accuracy and enables targeted intervention strategies.

2. CICLING 21

   Ranking of Event-focused English Tweets Based on Relevance and Informativeness

   D. Rudrapal, A. Das, and B. Bhattacharya

   Computing and Systems, 2021

   Abs DOI

   A Twitter event may generate a huge, noisy stream of tweets containing both relevant updates and irrelevant chatter. This paper presents a learning-to-rank framework that exploits enriched features—including content-popularity, keyword distribution, sentiment, hashtag usage, and tweet length—to rank tweets for relevance and informativeness. Experiments on real-world datasets show that our method effectively filters out noise and surfaces informative tweets with robust performance.

2020

1. LREC 20

   Minority Positive Sampling for Switching Points - an Anecdote for the Code-Mixing Language Modeling

   V. Guptha, A. Chatterjee, P. Chopra, and 1 more author

   In LREC 2020, 2020

   Poster

   Abs

   This paper presents a sampling strategy to over-sample minority code-switching points in mixed-language corpora. Using English-Hindi data, we show that balanced sampling around switch points reduces perplexity by 7.4% and improves language modeling performance near bilingual boundaries.

2. Artificial Intelligence Tools 20

   Deep Learning Based Sentiment Analysis in a Code-Mixed English-Hindi and English-Bengali Social Media Corpus

   A. Jamatia, S. D. Swamy, B. Gambäck, and 1 more author

   International Journal on Artificial Intelligence Tools, 2020

   Abs DOI

   This work addresses sentiment analysis on code-mixed English-Hindi and English-Bengali content from social media, a challenging task due to informal style and mixed language use. Three deep neural architectures (BiLSTM-CNN, double BiLSTM, attention-based models) are compared to traditional classifiers. Results show deep models, especially BiLSTM-CNN and attention-based ones, outperform conventional methods by 20-60% in accuracy, demonstrating their effectiveness for code-mixed sentiment detection.

2019

1. CICLING 19

   Recognition of Partial Textual Entailment for Indian Social Media Text

   D. Rudrapal, A. Das, and B. Bhattacharya

   Computing and Systems, 2019

   Abs DOI

   This paper presents a system for detecting partial textual entailment in noisy code-mixed Indian social media texts (English-Hindi, English-Bengali). By adapting features like overlap, negation, and semantic roles, the model shows effectiveness in identifying entailment despite informal structure and colloquialism. It demonstrates the potential for improving downstream tasks like summarization and contradiction detection in Indian social media.

2. IEEE CIM

   Understanding the Psycho-Sociological Facets of Homophily in Social Network Communities

   R. S. Solomon, S. P. Y. K. L., A. Das, and 2 more authors

   IEEE Computational Intelligence Magazine, 2019

   Abs DOI

   We analyze how psycho-sociological traits like personality and values drive homophily in online communities. Using large-scale Twitter data, we map linguistic markers to Myers-Briggs and Schwartz value traits. Findings show significant clustering—communities form based on shared traits—even when content differs, underscoring psychological attraction as a foundational principle in social media ties.

3. Intelligent Systems 19

   Deep Learning-Based Language Identification in English-Hindi-Bengali Code-Mixed Social Media Corpora

   A. Jamatia, A. Das, and B. Gambäck

   Journal of Intelligent Systems, 2019

   Abs DOI

   Word-level language identification in code-mixed social media text—English-Hindi and English-Bengali—is tackled using deep learning. Bidirectional LSTM models are compared with CRF classifiers. Results show BiLSTM models outperform CRF, with best model achieving  87% accuracy and F-score. This demonstrates deep architectures’ capacity to handle noisy and blended language data.

4. CICLING 19

   Deep Semantic Role Labeling for Tweets using 5W1H: Who, What, When, Where, Why and How

   K. Chakma, A. Das, and S. Debbarma

   Computing and Systems, 2019

   Abs DOI

   We implement a deep neural network with attention mechanism for semantic role labeling in the 5W1H style on tweets. The model achieves an F₁ score of 88.21%, outperforming previous Twitter SRL systems by about 28.7%. By simplifying traditional PropBank roles into journalistic 5W1H categories, we improve readability and performance for noisy social media text.

5. CICLING 19

   A New Approach for Twitter Event Summarization Based on Sentence Identification and Partial Textual Entailment

   D. Rudrapal, A. Das, and B. Bhattacharya

   Computing and Systems, 2019

   Abs DOI

   We propose a two-phase summarization pipeline: first ranking event-related tweets and extracting key sentences, then using Partial Textual Entailment (PTE) to reduce redundancy before generating an abstractive summary. Evaluation shows superior performance versus state-of-the-art baselines on real-time Twitter event data.

2018

1. CICLING 18

   Semantic Role Labeling of English Tweets

   D. Rudrapal and A. Das

   In CICLING 2018, 2018

   Oral, B

   Abs

   We propose an SRL system adapted to Twitter’s noisy text, using a CRF classifier with features from syntax, POS tagging, and lexical patterns. On a manually annotated tweet dataset, the system achieves 68.9% F₁, outperforming off-the-shelf SRL tools by ∼15%.

2. AAAI 18

   Consonant-Vowel Sequences as Subword Units for Code-Mixed Languages

   U. Kumar, V. S. Rana, C. Andrew, and 2 more authors

   In AAAI 2018, 2018

   Poster, A*

   Abs

   We propose using consonant-vowel (CV) substrings as subword units to model code-mixed text, addressing inconsistent orthography in English-Hindi mixing. We show CV-based segmentation yields a 9% reduction in perplexity over character-level models and improves downstream performance on translation and named-entity recognition tasks.

3. IJCIE 18

   Understanding the Selectional Preferences of the Twitter Mentions Network

   R. Solomon, S. P. Y. K. L. Srinivas, A. Narayan, and 1 more author

   International Journal of Computer and Information Engineering, 2018

   Abs

   We investigate the semantic selectional preferences (e.g., verb-object choices) in Twitter mention networks. Using corpus-based frequency analyses, we link linguistic behavior to mention-based social connections, highlighting semantic patterns in user mention contexts on Twitter.

4. CICLING 2018

   A 5W1H Based Annotation Scheme for Semantic Role Labeling of English Tweets

   K. Chakma and A. Das

   Computing and Systems, 2018

   Oral, B

   Abs DOI

   We introduce a 5W1H annotation schema for semantic role labeling in tweets, adapting journalistic question categories (Who, What, When, Where, Why, How) to social media contexts. Our annotated corpus supports improved semantic parsing in noisy text.

5. Computing and Systems 18

   Semantic Role Labelling of English Tweets

   D. Rudrapal and A. Das

   Computing and Systems, 2018

   Abs DOI

   This paper adapts traditional semantic role labeling (SRL) techniques to handle informal and noisy Twitter text. We build a dataset of annotated tweets and train a CRF-based SRL model, incorporating features like POS tags, lexical clusters, and dependency parses. The model achieves 72.3% F₁, demonstrating robustness in identifying predicate-argument structures in tweets despite irregular text patterns.

6. Information Systems 18

   Inducing Personalities and Values from Language Use in Social Network Communities

   U. Kumar, A. N. Reganti, T. Maheshwari, and 3 more authors

   Information Systems Frontier, 2018

   Abs DOI

   We present a framework to estimate personality traits (Big Five) and value orientations (Schwartz) from language use in Twitter communities. Using psycholinguistic dictionaries, LIWC features, and topic modeling, we derive community-level trait distributions. Results reveal distinct personality profiles driving language patterns—e.g., openness correlates with creative topic usage—suggesting language-driven personality inference at group scale.

7. IEEE CIM

   Friends are the family we choose - Understanding the Psycho-Sociological Homophilic Nature amongst Friends, Relatives, and Colleagues in Twitter

   S. Solomon, S. P. Y. K. L. Srinivas, A. Das, and 2 more authors

   IEEE Computational Intelligence Magazine, 2018

   Abs

   This study investigates homophily—tendency to connect with similar individuals—on Twitter by comparing interactions between friends, colleagues, and relatives. Using psycho-sociological language features, we find similarity-based homophily is stronger among friends than other relationship types, suggesting platform connections often reflect chosen social alignment over familial or professional ties.

8. IEEE IS

   Revealing Psycholinguistic Dimensions of Communities in Social Networks

   T. Maheshwari, A. Reganti, U. Kumar, and 2 more authors

   IEEE Intelligent Systems, 2018

   Abs DOI

   We analyze linguistic data from social network communities to uncover underlying psychological dimensions. Using LIWC and personality value metrics, we reveal significant clustering in psycholinguistic profiles, demonstrating that language use reflects community-level personality traits, cognition, and affect.

2017

1. ASONAM 17

   Understanding the Psycho-Sociological Vulnerability of ISIS Patronizers on Twitter

   T. Maheshwari, A. Reganti, A. Das, and 2 more authors

   In ASONAM 2017, 2017

   Poster, A

   Abs

   This study investigates psychological and social traits of Twitter users who support ISIS. Using psycholinguistic features from LIWC and network centrality measures, we identify markers like high anxiety, social isolation, and echo-chamber engagement. The analysis highlights how psycho-social vulnerability contributes to extremist content endorsement and helps inform early detection frameworks.

2. NLDB 17

   Quotology - Reading Between the Lines of Quotations

   D. Rudrapal, A. Das, and B. Bhattacharya

   In NLDB 2017, 2017

   Poster, B

   Abs

   We present ’Quotology’, a framework for analyzing implied meanings in quotations. By extracting clauses and applying entailment checks, we detect implied nuance such as sarcasm or irony. Evaluation on a dataset of political quotations shows 78% accuracy in identifying implied sentiment, outperforming baseline sentiment analysis systems.

3. CSCW 17

   Socio-Ethnic Ingredients of Social Network Communities

   T. Maheshwari, A. Reganti, T. Chakraborty, and 1 more author

   In CSCW 2017, 2017

   Poster, A

   Abs

   This paper analyzes how socio-ethnic attributes (e.g., caste, religion, region) shape online community formation. Using unsupervised clustering of user metadata and chat content, we uncover that ethnic identity significantly influences community boundaries, more so than shared interests. These findings highlight the social stratification embedded in digital interactions.

4. EACL 17

   A Societal Sentiment Analysis: Predicting the Values and Ethics of Individuals by Analyzing Social Media Content

   T. Maheshwari, A. N. Reganti, S. Gupta, and 4 more authors

   In EACL 2017, 2017

   Oral, A

   Abs

   We propose a methodology to infer individuals’ ethical and value orientations by analyzing their social media text. Using linguistic cues and value lexica, we train classifiers predicting Schwartz’s values (e.g., benevolence, conformity). Evaluation on survey-validated users achieves 68% accuracy, suggesting language reflects underlying moral dimensions.

5. AAAI 17

   Semantic Interpretation of Social Network Communities

   A. N. Reganti, T. Maheshwari, U. Kumar, and 3 more authors

   In AAAI 2017, 2017

   Poster, A*

   Abs

   We introduce a semantic interpretation framework for characterizing social network communities. By mapping topic models to ontology concepts and psycholinguistic traits, we generate semantic labels (e.g., ’eco-activists’, ’tech enthusiasts’) for communities. A user study shows 82% agreement with expert-generated labels, highlighting the model’s utility in summarizing community semantics.

6. Information Systems 17

   Mining Human Psycholinguistic Behavior from Social Media

   T. Maheshwari, A. Reganti, U. Kumar, and 3 more authors

   Information Systems Frontier, 2017

   Abs DOI

   This paper examines how psycholinguistic traits—personality, sentiment, social orientation—can be mined from social media text. We apply computational linguistics tools, including LIWC and topic models, on Twitter datasets and validate findings against self-reported surveys. Results indicate strong alignment between inferred and actual traits, demonstrating linguistic analysis as a proxy for psychological profiling.

2016

1. LREC 16

   Comparing the Level of Code-Switching in Corpora

   B. Gambäck and A. Das

   In LREC 2016, 2016

   Poster, A

   Abs

   We compare multiple multilingual corpora on their degree of code-switching, using metrics like language mix density and switch frequency. Our analysis across nine corpora shows wide variation, and we propose a standardized framework for code-switching quantification. This aids in the selection and generation of corpora for modeling code-mixed language phenomena.

2. CICLING 16

   Collecting and Annotating Indian Social Media Code-Mixed Corpora

   A. Jamatia, B. Gambäck, and A. Das

   In CICLING 2016, 2016

   Oral, B

   Abs

   We present the methodology for harvesting, filtering, and annotating a large-scale code-mixed English-Hindi corpus from Twitter and Facebook. We describe annotation guidelines for language tags, part-of-speech, and transliteration, and analyze corpus properties. The resulting corpus enables improved downstream models for code-mixed NLP.

2015

1. RANLP 15

   Part-of-Speech Tagging for Code-Mixed English-Hindi Twitter and Facebook Chat Messages

   A. Jamatia, B. Gambäck, and A. Das

   In RANLP 2015, 2015

   Oral, B

   Abs

   This paper addresses POS tagging in code-mixed English-Hindi chat data. We train CRF models with novel features such as transliteration consistency and bilingual context windows. Experiments show 88.4% tagging accuracy, outperforming monolingual baselines by  10%.

2014

1. INLG 14

   A Framework for Health Behavior Change using Companionable Robots

   B. Sarma, A. Das, and R. Nielsen

   In INLG 2014, 2014

   Poster, B

   Abs

   We propose a robot-based framework to encourage health behavior change via personalized dialogue, emotional rapport, and goal tracking. Using consumer health data and motivational interviewing techniques, the system adapts recommendations. A pilot with 30 participants shows improved engagement and self-reported activity confidence.

2. TAL 14

   Code-Mixing in Social Media Text: The Last Language Identification Frontier?

   Amitava Das and Björn Gambäck

   Traitement Automatique des Langues (TAL), 2014

   Abs

   This paper investigates the challenges of automatic language identification in code-mixed social media text (e.g., English-Hindi, English-Bengali). It introduces a hierarchical model combining character and token-level features using CRFs, achieving 90% accuracy. This work lays the groundwork for more nuanced NLP pipelines that can accurately process multilingual social media data.

2013

1. HCI 13

   Sentimental Eyes!

   A. Das and B. Gambäck

   In Proceedings of HCI 2013, 2013

   Poster, A

   Abs

   We introduce “Sentimental Eyes!”, a system that visualizes real-time sentiment from text streams via emotion-aware emoticons and color mapping. Applied to chat logs, users reported improved mood awareness and communication moderation. An evaluation with 50 users shows increased sentiment awareness and higher conversation satisfaction.

2012

1. CICLING 12

   The 5W Structure for Sentiment Summarization-Visualization-Tracking

   Amitava Das, Sivaji Bandyopadhyay, and Björn Gambäck

   In CICLING 2012, 2012

   Oral, B

   Abs

   We propose a framework that structures sentiment summaries around the “Who, What, When, Where, Why” dimensions. Applied to Twitter, this enables flexible query-based summarization and interactive visualization. Our evaluations suggest improved user comprehension and efficient tracking of sentiment trends over time.

2011

1. ACL 11

   PsychoSentiWordNet

   Amitava Das

   In Proceedings of ACL 2011, 2011

   Poster, A*

   Abs

   We introduce PsychoSentiWordNet, a sentiment lexicon augmented with psycho-linguistic attributes (e.g., concreteness, valence, arousal). By combining lexical and psychological dimensions, it enhances sentiment analysis, particularly in emotional intensity estimation. Evaluations on benchmark datasets show improved performance over SentiWordNet-based systems.

2. ACL 11

   Dr Sentiment Knows Everything!

   Amitava Das and Sivaji Bandyopadhyay

   In ACL/HLT 2011, 2011

   Demo, A*

   Abs

   This demo presents “Dr Sentiment”, a system that explains sentiment analysis via concept graphs derived from WordNet and SentiWordNet. It visualizes sentiment flow and reasoning paths. A usability study with users (N=40) shows improved trust and transparency compared to standard sentiment scores.

2010

1. ICON 10

   Can We Mimic Human Pragmatics Knowledge in a Computational Lexicon?

   Amitava Das

   In ICON 2010, 2010

   Oral, Best Student Paper

   Abs

   We explore incorporating pragmatic knowledge—e.g., implicature, presupposition—into a computational lexicon. We extend an existing lexical resource with pragmatics annotations and evaluate on inference tasks requiring context understanding. Results show improved performance in counterfactual and implicature detection, highlighting lexicon-based pragmatics’ potential.