Publications

2026

1. MLingualFC: Evaluating Jailbreak Vulnerabilities in Multilingual Vision-Language Models

   Mamta . , and Oana Cocarascu

   In Proceedings of the MeLLM Workshop at ACL 2026 , 2026

2. 

   TinyAttack: Exploring Stylistic Vulnerabilities in Large Language Models

   Mamta . , Bogdan Grecu , and Oana Cocarascu

   In Findings of the Association for Computational Linguistics: ACL 2026 , 2026

   Abs PDF

   Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of NLP tasks. However, their robustness to stylistic variations in text remains underexplored. In this paper, we introduce TinyAttack, a framework that systematically exploits stylistic vulnerabilities in LLMs through controlled style transfer. Our approach reveals that even subtle stylistic changes can significantly alter model predictions, raising important concerns about the robustness of LLMs in real-world deployment.

3. 

   Improving Dialect Robustness in Large Language Models via LoRA and Mixture-of-Experts

   Sanjh Maheshwari , Aniket Singh Rajpoot , Oana Cocarascu , and 1 more author

   In Proceedings of the 13th Workshop on NLP for Similar Languages, Varieties and Dialects , Mar 2026

   Abs PDF

   Despite the success of large language models (LLMs) in a wide range of applications, it has been shown that their performance varies across English dialects. Differences among English dialects are reflected in vocabulary, syntax, and writing style, and can adversely affect model performance. Several studies evaluate the dialect robustness of LLMs, yet research on enhancing their robustness to dialectal variation remains limited. In this paper, we propose two parameter-efficient frameworks for improving dialectal robustness in LLMs: DialectFusion where we train separate LoRA layers for each dialect and apply different LoRA merging methods, and DialectMoE which is built on top of Mixture of Experts LoRA and introduces multiple LoRA-based experts to the feed-forward layer to internally model the dialectal dependencies. Our comprehensive analysis on five open-source LLMs for sentiment and sarcasm tasks in zero- and few-shot settings shows that our proposed approaches enhance the dialect robustness of LLMs and outperforms instruct and LoRA fine-tuning based approaches.

4. 

   Indic-TunedLens: Interpreting Multilingual Models in Indian Languages

   Mihir Panchal , Deeksha Varshney , Mamta . , and 1 more author

   In Proceedings of the 13th Workshop on NLP for Similar Languages, Varieties and Dialects , Mar 2026

   Abs PDF

   Multilingual large language models (LLMs) are increasingly deployed in linguistically diverse regions like India, yet most interpretability tools remain tailored to English. Prior work reveals that LLMs often operate in English centric representation spaces, making cross lingual interpretability a pressing concern. We introduce Indic-TunedLens, a novel interpretability framework specifically for Indian languages that learns shared affine transformations. Unlike the standard Logit Lens, which directly decodes intermediate activations, Indic-TunedLens adjusts hidden states for each target language, aligning them with the target output distributions to enable more faithful decoding of model representations. We evaluate our framework on 10 Indian languages using the MMLU benchmark and find that it significantly improves over SOTA interpretability methods, especially for morphologically rich, low resource languages.

5. 

   Federated Model Synchronization for Diagnostic Redefinition through a Novel Selective Parameter Unlearning

   Mayank Kumar Kundalwal , Mamta Mamta, Deepak Mishra , and 1 more author

   In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , Mar 2026

   Abs PDF

   We propose a novel federated learning framework for medical image diagnosis that incorporates selective parameter unlearning. Our method enables federated model synchronization while allowing diagnostic models to selectively forget outdated or sensitive information, improving both privacy and model adaptability in clinical settings.

2025

1. 

   I-GUARD: Interpretability-Guided Parameter Optimization for Adversarial Defense

   Mamta , and Oana Cocarascu

   In Findings of the Association for Computational Linguistics: EMNLP 2025 , Nov 2025

   Abs PDF

   We propose I-GUARD, an interpretability-guided framework for adversarial defense in NLP. By leveraging model interpretability signals to identify and optimize the most adversarially vulnerable parameters, I-GUARD achieves robust defense without sacrificing clean accuracy. Our method identifies critical parameters through gradient-based attribution and applies targeted regularization to enhance model robustness against a wide range of adversarial attacks.

2. 

   FactEval: Evaluating the Robustness of Fact Verification Systems in the Era of Large Language Models

   Mamta Mamta, and Oana Cocarascu

   In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , Apr 2025

   Abs PDF

   We introduce FactEval, a comprehensive benchmark for evaluating the robustness of fact verification systems against large language models. We systematically study how LLMs can be used to generate challenging misinformation that evades state-of-the-art fact verification systems and propose evaluation protocols that better reflect real-world deployment conditions.

2024

1. 

   BiasWipe: Mitigating Unintended Bias in Text Classifiers through Model Interpretability

   Mamta , Rishikant Chigrupaatii , and Asif Ekbal

   In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , Apr 2024

   Abs HTML PDF

   We present BiasWipe, a model interpretability-based framework for identifying and mitigating unintended bias in text classifiers. By analyzing attribution scores to locate biased neurons and features, BiasWipe selectively suppresses biased representations while preserving task performance. Our approach demonstrates significant bias reduction across multiple datasets and demographic categories without requiring retraining.

2. 

   Aspect-Based Multimodal Mining: Unveiling Sentiments, Complaints, and Beyond in User-Generated Content

   Mamta , Gopendra Vikram Singh , Deepak Raju Kori , and 1 more author

   In Proceedings of the 32nd ACM International Conference on Multimedia , Apr 2024

   Abs PDF

   We propose a multimodal framework for aspect-based opinion mining in user-generated content, jointly modeling text and visual signals to detect sentiments, complaints, and other opinion types at the aspect level. Our approach leverages cross-modal attention to align textual aspects with corresponding visual cues, enabling richer understanding of user opinions in multimedia content.

3. 

   MultiSEAO-Mix: A Multimodal Multitask Framework for Sentiment, Emotion, Support, and Offensive Analysis in Code-Mixed Setting

   Mamta , Gopendra Vikram Singh , Atul Verma , and 2 more authors

   IEEE Transactions on Computational Social Systems, Apr 2024

   Abs PDF

   We propose MultiSEAO-Mix, a unified multimodal and multitask learning framework for jointly analyzing sentiment, emotion, social support, and offensive language in code-mixed social media text. By sharing representations across tasks and modalities, our framework achieves state-of-the-art results on multiple code-mixed benchmarks.

4. 

   Quality achhi hai (is good), satisfied! Towards aspect based sentiment analysis in code-mixed language

   Mamta , and Asif Ekbal

   Computer Speech and Language, Apr 2024

   Abs PDF

   We address the challenge of aspect-based sentiment analysis (ABSA) in code-mixed languages, focusing on Hindi-English (Hinglish) text. We introduce a new benchmark dataset and propose a cross-lingual transfer learning approach that adapts monolingual models to the code-mixed setting, achieving strong performance on aspect term extraction and sentiment classification.

5. 

   Atmosphere kamaal ka tha (was wonderful): A Multilingual Joint Learning Framework for Aspect Category Detection and Sentiment Classification

   Mamta , and Asif Ekbal

   IEEE Transactions on Computational Social Systems, Apr 2024

   Abs PDF

   We propose a multilingual joint learning framework for simultaneously performing aspect category detection and sentiment classification across multiple languages. Our model leverages cross-lingual representations and shared task-specific components to effectively transfer knowledge across languages, enabling robust aspect-level sentiment analysis in low-resource settings.

2023

1. 

   Elevating Code-mixed Text Handling through Auditory Information of Words

   Mamta , Zishan Ahmad , and Asif Ekbal

   Conference on Empirical Methods in Natural Language Processing, Apr 2023

   Abs HTML PDF

   We propose a novel approach for code-mixed text analysis that incorporates auditory (phonetic) information of words alongside textual features. By leveraging phonetic embeddings derived from the pronunciation of words, our method better captures the acoustic characteristics of code-mixed speech, leading to improved performance on sentiment analysis and other downstream tasks.

2. 

   Transformer based multilingual joint learning framework for code-mixed and english sentiment analysis

   Mamta , and Asif Ekbal

   Journal of Intelligent Information Systems, Apr 2023

   Abs PDF

   We present a transformer-based multilingual joint learning framework that simultaneously handles code-mixed and standard English sentiment analysis. By sharing transformer layers across language varieties and employing task-specific fine-tuning, our approach achieves competitive performance on both code-mixed and monolingual benchmarks.

3. 

   Service is good, very good or excellent? towards aspect based sentiment intensity analysis

   Mamta , and Asif Ekbal

   In European Conference on Information Retrieval , Apr 2023

   Abs PDF

   We tackle the problem of aspect-based sentiment intensity analysis, going beyond binary or ternary sentiment labels to predict fine-grained sentiment intensity scores at the aspect level. We introduce a new dataset annotated with sentiment intensity scores and propose a regression-based model that captures the degree of sentiment expressed towards specific aspects in reviews.

2022

1. 

   Adversarial sample generation for aspect based sentiment classification

   Mamta Mamta, and Asif Ekbal

   In Findings of the Association for Computational Linguistics: AACL-IJCNLP 2022 , Apr 2022

   Abs HTML

   We propose an adversarial data augmentation framework for aspect-based sentiment classification. By generating targeted adversarial examples that preserve aspect mentions while altering sentiment-bearing words, we create challenging training samples that improve model robustness and generalization across domains.

2. 

   Exploring multi-lingual, multi-task, and adversarial learning for low-resource sentiment analysis

   Mamta , Asif Ekbal , and Pushpak Bhattacharyya

   Transactions on Asian and Low-Resource Language Information Processing, Apr 2022

   Abs PDF

   We explore the synergies between multilingual learning, multitask learning, and adversarial training for improving sentiment analysis in low-resource languages. Through extensive experiments, we demonstrate that combining these three paradigms consistently outperforms individual approaches, enabling effective knowledge transfer from high-resource languages to low-resource ones.

3. 

   HindiMD: A multi-domain corpora for low-resource sentiment analysis

   In Proceedings of the Thirteenth Language Resources and Evaluation Conference , Apr 2022

   Abs PDF

   We introduce HindiMD, a multi-domain sentiment analysis corpus for Hindi, covering domains such as product reviews, movie reviews, and news articles. We provide detailed annotation guidelines and baseline experiments demonstrating the challenging nature of cross-domain sentiment analysis in Hindi.

2020

1. 

   Multi-domain tweet corpora for sentiment analysis: resource creation and evaluation

   In Proceedings of the Twelfth Language Resources and Evaluation Conference , Apr 2020

   Abs HTML

   We present a large-scale multi-domain tweet corpus for sentiment analysis covering diverse topics including politics, sports, entertainment, and technology. The corpus includes fine-grained sentiment annotations and is evaluated using multiple baseline models, providing a valuable resource for cross-domain sentiment analysis research.

2. 

   Only text? only image? or both? Predicting sentiment of internet memes

   Pranati Behera , Mamta Mamta, and Asif Ekbal

   In Proceedings of the 17th International Conference on Natural Language Processing (ICON) , Apr 2020

   Abs PDF

   We investigate the relative contribution of text and image modalities for predicting the sentiment of internet memes. Through systematic ablation studies comparing text-only, image-only, and multimodal fusion approaches, we demonstrate that both modalities are complementary and that multimodal models significantly outperform unimodal baselines for meme sentiment analysis.