Papers by Somnath Banerjee
InfFeed: Influence Functions as a Feedback to Improve the Performance of Subjective Tasks (2024.lrec-main)
Copied to clipboard
| Challenge: | InfFeed uses influence functions to compute the influential instances for a target instance. |
| Approach: | They propose an apparatus that uses influence functions to compute the influential instances for a target instance. |
| Outcome: | The proposed model outperforms the state-of-the-art baselines by 4% for hate speech classification, 3.5% for stance classification, and 3% for irony and 2% for sarcasm detection. |
IR like a SIR: Sense-enhanced Information Retrieval for Multiple Languages (2021.emnlp-main)
Copied to clipboard
Rexhina Blloshmi, Tommaso Pasini, Niccolò Campolungo, Somnath Banerjee, Roberto Navigli, Gabriella Pasi
| Challenge: | Recent advances in contextualized embeddings have made ranking on non-English documents cumbersome . a novel multilingual query expansion mechanism provides sense definitions as additional semantic information for the query. |
| Approach: | They propose a multilingual query expansion mechanism that leverages word sense information to enhance the model's performance. |
| Outcome: | The proposed model performs better than its supervised and unsupervised alternatives across languages while being trained on English Robust04 data. |
Breaking Boundaries: Investigating the Effects of Model Editing on Cross-linguistic Performance (2025.naacl-industry)
Copied to clipboard
Somnath Banerjee, Avik Halder, Rajarshi Mandal, Sayan Layek, Ian Soboroff, Rima Hazra, Animesh Mukherjee
| Challenge: | Pretrained language models (PLMs) have revolutionized NLP but amplify linguistic inequities in multilingual applications. |
| Approach: | They evaluate pretrained language models including Mistral, TowerInstruct, OpenHathi, Tamil-Llama, and Kan-Lama across eight languages spanning high-resource and low-resourced settings. |
| Outcome: | The proposed models fail to bridge linguistic divides and are inefficient when compared to other models. |
Context Matters: Pushing the Boundaries of Open-Ended Answer Generation with Graph-Structured Knowledge Context (2024.emnlp-industry)
Copied to clipboard
| Challenge: | GraphContextGen outperforms dominant text-based retrieval systems in domain specific community question answering platforms like AskUbuntu, Unix, and ServerFault. |
| Approach: | They propose a framework that combines graph-driven context retrieval with knowledge graphs based enhancement to improve the proficiency of LLMs. |
| Outcome: | The proposed framework outperforms dominant text-based retrieval systems in open-ended questions. |
Sowing the Wind, Reaping the Whirlwind: The Impact of Editing Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) face challenges in maintaining accuracy due to the dynamic nature of world knowledge. |
| Approach: | They propose to use a benchmark dataset to investigate the effects of model edits on model safety metrics and guardrails. |
| Outcome: | The proposed dataset sheds light on how the edits, impact the model’s safety metrics and guardrails. |
Navigating the Cultural Kaleidoscope: A Hitchhiker’s Guide to Sensitivity in Large Language Models (2025.naacl-long)
Copied to clipboard
Somnath Banerjee, Sayan Layek, Hari Shrawgi, Rajarshi Mandal, Avik Halder, Shanu Kumar, Sagnik Basu, Parag Agrawal, Rima Hazra, Animesh Mukherjee
| Challenge: | Cultural harm arises when LLMs misrepresent or normalize values, identities, and practices in ways that conflict with the norms of diverse cultural groups. |
| Approach: | They propose a cultural harm test dataset and a preference dataset to assess model outputs across different cultural contexts. |
| Outcome: | The proposed model improves model behavior significantly reducing the likelihood of generating culturally insensitive or harmful content. |
Safety Arithmetic: A Framework for Test-time Safety Alignment of Language Models by Steering Parameters and Activations (2024.emnlp-main)
Copied to clipboard
| Challenge: | Current alignment methods struggle with dynamic user intentions and complex objectives, making models vulnerable to harmful content. |
| Approach: | They propose a training-free framework that enhances LLM safety across different scenarios. |
| Outcome: | The proposed framework significantly improves safety measures, reduces over-safety, and maintains model utility, outperforming existing methods in ensuring safe content generation. |
Soteria: Language-Specific Functional Parameter Steering for Multilingual Safety Alignment (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Soteria locates and minimally adjusts the “functional heads” most responsible for harmful content generation in each language. |
| Approach: | Soteria locates and minimally adjusts the "functional heads" responsible for harmful content generation in each language. |
| Outcome: | The proposed approach reduces harmful content generation in languages while preserving model performance. |
Hate Speech and Offensive Language Detection in Bengali (2022.aacl-main)
Copied to clipboard
| Challenge: | Existing research on hate speech detection in English does not cover low-resource languages like Bengali. |
| Approach: | They develop an annotated dataset of 10K Bengali posts consisting of 5K actual and 5K Romanized Bengali tweets. |
| Outcome: | The proposed model outperforms other models on training actual and romanized datasets by interpreting the semantic expressions better. |