Somnath Banerjee

Mithun Das, Somnath Banerjee, Animesh Mukherjee

Abusive language is a growing concern in many social media platforms. Repeated exposure to abusive speech has created physiological effects on the target users. Thus, the problem of abusive language should be addressed in all forms for online peace and safety. While extensive research exists in abusive speech detection, most studies focus on English. Recently, many smearing incidents have occurred in India, which provoked diverse forms of abusive speech in online space in various languages based on the geographic location. Therefore it is essential to deal with such malicious content. In this paper, to bridge the gap, we demonstrate a large-scale analysis of multilingual abusive speech in Indic languages. We examine different interlingual transfer mechanisms and observe the performance of various multilingual models for abusive speech detection for eight different Indic languages. We also experiment to show how robust these models are on adversarial attacks. Finally, we conduct an in-depth error analysis by looking into the models' misclassified posts across various settings. We have made our code and models public for other researchers.

Mithun Das, Somnath Banerjee, Punyajoy Saha et al.

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.

Rima Hazra, Agnik Saha, Somnath Banerjee et al.

Community Question Answering (CQA) platforms steadily gain popularity as they provide users with fast responses to their queries. The swiftness of these responses is contingent on a mixture of query-specific and user-related elements. This paper scrutinizes these contributing factors within the context of six highly popular CQA platforms, identified through their standout answering speed. Our investigation reveals a correlation between the time taken to yield the first response to a question and several variables: the metadata, the formulation of the questions, and the level of interaction among users. Additionally, by employing conventional machine learning models to analyze these metadata and patterns of user interaction, we endeavor to predict which queries will receive their initial responses promptly.

Sagnik Basu, Subhrajit Mitra, Aman Juneja et al.

Recent research points toward LLMs being manipulated through adversarial and seemingly benign inputs, resulting in harmful, biased, or policy-violating outputs. In this paper, we study an underexplored issue concerning harmful and toxic mathematical word problems. We show that math questions, particularly those framed as natural language narratives, can serve as a subtle medium for propagating biased, unethical, or psychologically harmful content, with heightened risks in educational settings involving children. To support a systematic study of this phenomenon, we introduce ToxicGSM, a dataset of 1.9k arithmetic problems in which harmful or sensitive context is embedded while preserving mathematically well-defined reasoning tasks. Using this dataset, we audit the behaviour of existing LLMs and analyse the trade-offs between safety enforcement and mathematical correctness. We further propose SafeMath -- a safety alignment technique that reduces harmful outputs while maintaining, and in some cases improving, mathematical reasoning performance. Our results highlight the importance of disentangling linguistic harm from math reasoning and demonstrate that effective safety alignment need not come at the cost of accuracy. We release the source code and dataset at https://github.com/Swagnick99/SafeMath/tree/main.

Mithun Das, Somnath Banerjee, Animesh Mukherjee

Social media platforms often act as breeding grounds for various forms of trolling or malicious content targeting users or communities. One way of trolling users is by creating memes, which in most cases unites an image with a short piece of text embedded on top of it. The situation is more complex for multilingual(e.g., Tamil) memes due to the lack of benchmark datasets and models. We explore several models to detect Troll memes in Tamil based on the shared task, "Troll Meme Classification in DravidianLangTech2022" at ACL-2022. We observe while the text-based model MURIL performs better for Non-troll meme classification, the image-based model VGG16 performs better for Troll-meme classification. Further fusing these two modalities help us achieve stable outcomes in both classes. Our fusion model achieved a 0.561 weighted average F1 score and ranked second in this task.

Somnath Banerjee, Pranav Jha, Rima Hazra et al.

LLMs deployed multilingually are often audited via English explanations for non-English inputs. We evaluate extractive explanations ''where the model identifies input token spans as evidence alongside a generated rationale'' and uncover a systematic trade-off: English-pivot explanations can achieve higher span agreement with human rationales while their evidence becomes less causally grounded in the model's prediction, as measured by both comprehensiveness and sufficiency. Across 3 tasks, 5~languages, and 2~multilingual LLM families, we find that English explanations frequently produce fluent but loosely anchored rationales, with comprehensiveness degrading by up to 5.7x relative to native-language conditions - even as task accuracy remains stable across settings. For socially nuanced classification, English pivots also fail to preserve pragmatic cues, reducing both faithfulness and span agreement. We recommend auditing explanations in the input language, reporting multi-faceted faithfulness metrics beyond lexical overlap, and treating English rationales as communication summaries rather than faithful decision traces.

Rima Hazra, Bikram Ghuku, Ilona Marchenko et al.

Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation, failing to capture whether a model is simultaneously pedagogically effective and safe across student-tutor interaction. We argue that tutoring safety is fundamentally different from conventional LLM safety: the primary risk is not toxic content but the quiet erosion of learning through answer over-disclosure, misconception reinforcement, and the abdication of scaffolding. To systematically study this failure mode, we introduce SafeTutors, a benchmark that jointly evaluates safety and pedagogy across mathematics, physics, and chemistry. SafeTutors is organized around a theoretically grounded risk taxonomy comprising 11 harm dimensions and 48 sub-risks drawn from learning-science literature. We uncover that all models show broad harm; scale doesn't reliably help; and multi-turn dialogue worsens behavior, with pedagogical failures rising from 17.7% to 77.8%. Harms also vary by subject, so mitigations must be discipline-aware, and single-turn "safe/helpful" results can mask systematic tutor failure over extended interaction.

Somnath Banerjee, Sayan Layek, Pratyush Chatterjee et al.

Ensuring consistent safety across multiple languages remains a significant challenge for large language models (LLMs). We introduce Soteria, a lightweight yet powerful strategy that locates and minimally adjusts the "functional heads" most responsible for harmful content generation in each language. By altering only a fraction of parameters, Soteria drastically reduces policy violations without sacrificing overall model performance, even in low-resource settings. To rigorously evaluate our approach, we also present XThreatBench, a specialized multilingual dataset capturing fine-grained harmful behaviors drawn from real policy guidelines. Experiments with leading open-source LLMs (e.g., Llama, Qwen, Mistral) show that Soteria consistently improves safety metrics across high-, mid-, and low-resource languages. These findings highlight a promising path toward scalable, linguistically attuned, and ethically aligned LLMs worldwide.

Somnath Banerjee, Avik Dutta, Aaditya Agrawal et al.

With the AI revolution in place, the trend for building automated systems to support professionals in different domains such as the open source software systems, healthcare systems, banking systems, transportation systems and many others have become increasingly prominent. A crucial requirement in the automation of support tools for such systems is the early identification of named entities, which serves as a foundation for developing specialized functionalities. However, due to the specific nature of each domain, different technical terminologies and specialized languages, expert annotation of available data becomes expensive and challenging. In light of these challenges, this paper proposes a novel named entity recognition (NER) technique specifically tailored for the open-source software systems. Our approach aims to address the scarcity of annotated software data by employing a comprehensive two-step distantly supervised annotation process. This process strategically leverages language heuristics, unique lookup tables, external knowledge sources, and an active learning approach. By harnessing these powerful techniques, we not only enhance model performance but also effectively mitigate the limitations associated with cost and the scarcity of expert annotators. It is noteworthy that our model significantly outperforms the state-of-the-art LLMs by a substantial margin. We also show the effectiveness of NER in the downstream task of relation extraction.

Sayantan Adak, Somnath Banerjee, Rajarshi Mandal et al.

Online memes are a powerful yet challenging medium for content moderation, often masking harmful intent behind humor, irony, or cultural symbolism. Conventional moderation systems "especially those relying on explicit text" frequently fail to recognize such subtle or implicit harm. We introduce MemeSense, an adaptive framework designed to generate socially grounded interventions for harmful memes by combining visual and textual understanding with curated, semantically aligned examples enriched with commonsense cues. This enables the model to detect nuanced complexed threats like misogyny, stereotyping, or vulgarity "even in memes lacking overt language". Across multiple benchmark datasets, MemeSense outperforms state-of-the-art methods, achieving up to 35% higher semantic similarity and 9% improvement in BERTScore for non-textual memes, and notable gains for text-rich memes as well. These results highlight MemeSense as a promising step toward safer, more context-aware AI systems for real-world content moderation. Code and data available at: https://github.com/sayantan11995/MemeSense

Somnath Banerjee, Sayan Layek, Rima Hazra et al.

In this study, we tackle a growing concern around the safety and ethical use of large language models (LLMs). Despite their potential, these models can be tricked into producing harmful or unethical content through various sophisticated methods, including 'jailbreaking' techniques and targeted manipulation. Our work zeroes in on a specific issue: to what extent LLMs can be led astray by asking them to generate responses that are instruction-centric such as a pseudocode, a program or a software snippet as opposed to vanilla text. To investigate this question, we introduce TechHazardQA, a dataset containing complex queries which should be answered in both text and instruction-centric formats (e.g., pseudocodes), aimed at identifying triggers for unethical responses. We query a series of LLMs -- Llama-2-13b, Llama-2-7b, Mistral-V2 and Mistral 8X7B -- and ask them to generate both text and instruction-centric responses. For evaluation we report the harmfulness score metric as well as judgements from GPT-4 and humans. Overall, we observe that asking LLMs to produce instruction-centric responses enhances the unethical response generation by ~2-38% across the models. As an additional objective, we investigate the impact of model editing using the ROME technique, which further increases the propensity for generating undesirable content. In particular, asking edited LLMs to generate instruction-centric responses further increases the unethical response generation by ~3-16% across the different models.

Somnath Banerjee, Sayan Layek, Hari Shrawgi et al.

As LLMs are increasingly deployed in global applications, the importance of cultural sensitivity becomes paramount, ensuring that users from diverse backgrounds feel respected and understood. Cultural harm can arise when these models fail to align with specific cultural norms, resulting in misrepresentations or violations of cultural values. This work addresses the challenges of ensuring cultural sensitivity in LLMs, especially in small-parameter models that often lack the extensive training data needed to capture global cultural nuances. We present two key contributions: (1) A cultural harm test dataset, created to assess model outputs across different cultural contexts through scenarios that expose potential cultural insensitivities, and (2) A culturally aligned preference dataset, aimed at restoring cultural sensitivity through fine-tuning based on feedback from diverse annotators. These datasets facilitate the evaluation and enhancement of LLMs, ensuring their ethical and safe deployment across different cultural landscapes. Our results show that integrating culturally aligned feedback leads to a marked improvement in model behavior, significantly reducing the likelihood of generating culturally insensitive or harmful content. Ultimately, this work paves the way for more inclusive and respectful AI systems, fostering a future where LLMs can safely and ethically navigate the complexities of diverse cultural landscapes.

Somnath Banerjee, Avik Dutta, Sayan Layek et al.

In this paper, we delve into the advancement of domain-specific Large Language Models (LLMs) with a focus on their application in software development. We introduce DevAssistLlama, a model developed through instruction tuning, to assist developers in processing software-related natural language queries. This model, a variant of instruction tuned LLM, is particularly adept at handling intricate technical documentation, enhancing developer capability in software specific tasks. The creation of DevAssistLlama involved constructing an extensive instruction dataset from various software systems, enabling effective handling of Named Entity Recognition (NER), Relation Extraction (RE), and Link Prediction (LP). Our results demonstrate DevAssistLlama's superior capabilities in these tasks, in comparison with other models including ChatGPT. This research not only highlights the potential of specialized LLMs in software development also the pioneer LLM for this domain.

Sayantan Adak, Pratyush Chatterjee, Somnath Banerjee et al.

Present day LLMs face the challenge of managing affordance-based safety risks-situations where outputs inadvertently facilitate harmful actions due to overlooked logical implications. Traditional safety solutions, such as scalar outcome-based reward models, parameter tuning, or heuristic decoding strategies, lack the granularity and proactive nature needed to reliably detect and intervene during subtle yet crucial reasoning steps. Addressing this fundamental gap, we introduce AURA, an innovative, multi-layered framework centered around Process Reward Models (PRMs), providing comprehensive, step level evaluations across logical coherence and safety-awareness. Our framework seamlessly combines introspective self-critique, fine-grained PRM assessments, and adaptive safety-aware decoding to dynamically and proactively guide models toward safer reasoning trajectories. Empirical evidence clearly demonstrates that this approach significantly surpasses existing methods, significantly improving the logical integrity and affordance-sensitive safety of model outputs. This research represents a pivotal step toward safer, more responsible, and contextually aware AI, setting a new benchmark for alignment-sensitive applications.

Somnath Banerjee, Pratyush Chatterjee, Shanu Kumar et al.

Recent advancements in LLMs have raised significant safety concerns, particularly when dealing with code-mixed inputs and outputs. Our study systematically investigates the increased susceptibility of LLMs to produce unsafe outputs from code-mixed prompts compared to monolingual English prompts. Utilizing explainability methods, we dissect the internal attribution shifts causing model's harmful behaviors. In addition, we explore cultural dimensions by distinguishing between universally unsafe and culturally-specific unsafe queries. This paper presents novel experimental insights, clarifying the mechanisms driving this phenomenon.

Somnath Banerjee, Sayan Layek, Soham Tripathy et al.

Safety-aligned language models often exhibit fragile and imbalanced safety mechanisms, increasing the likelihood of generating unsafe content. In addition, incorporating new knowledge through editing techniques to language models can further compromise safety. To address these issues, we propose SafeInfer, a context-adaptive, decoding-time safety alignment strategy for generating safe responses to user queries. SafeInfer comprises two phases: the safety amplification phase, which employs safe demonstration examples to adjust the model's hidden states and increase the likelihood of safer outputs, and the safety-guided decoding phase, which influences token selection based on safety-optimized distributions, ensuring the generated content complies with ethical guidelines. Further, we present HarmEval, a novel benchmark for extensive safety evaluations, designed to address potential misuse scenarios in accordance with the policies of leading AI tech giants.

Rima Hazra, Sayan Layek, Somnath Banerjee et al.

Ensuring the safe alignment of large language models (LLMs) with human values is critical as they become integral to applications like translation and question answering. Current alignment methods struggle with dynamic user intentions and complex objectives, making models vulnerable to generating harmful content. We propose Safety Arithmetic, a training-free framework enhancing LLM safety across different scenarios: Base models, Supervised fine-tuned models (SFT), and Edited models. Safety Arithmetic involves Harm Direction Removal to avoid harmful content and Safety Alignment to promote safe responses. Additionally, we present NoIntentEdit, a dataset highlighting edit instances that could compromise model safety if used unintentionally. Our experiments show that Safety Arithmetic significantly improves safety measures, reduces over-safety, and maintains model utility, outperforming existing methods in ensuring safe content generation.

Somnath Banerjee, Avik Halder, Rajarshi Mandal et al.

The integration of pretrained language models (PLMs) like BERT and GPT has revolutionized NLP, particularly for English, but it has also created linguistic imbalances. This paper strategically identifies the need for linguistic equity by examining several knowledge editing techniques in multilingual contexts. We evaluate the performance of models such as Mistral, TowerInstruct, OpenHathi, Tamil-Llama, and Kan-Llama across languages including English, German, French, Italian, Spanish, Hindi, Tamil, and Kannada. Our research identifies significant discrepancies in normal and merged models concerning cross-lingual consistency. We employ strategies like 'each language for itself' (ELFI) and 'each language for others' (ELFO) to stress-test these models. Our findings demonstrate the potential for LLMs to overcome linguistic barriers, laying the groundwork for future research in achieving linguistic inclusivity in AI technologies.

Somnath Banerjee, Maulindu Sarkar, Punyajoy Saha et al.

Recently, influence functions present an apparatus for achieving explainability for deep neural models by quantifying the perturbation of individual train instances that might impact a test prediction. Our objectives in this paper are twofold. First we incorporate influence functions as a feedback into the model to improve its performance. Second, in a dataset extension exercise, using influence functions to automatically identify data points that have been initially `silver' annotated by some existing method and need to be cross-checked (and corrected) by annotators to improve the model performance. To meet these objectives, in this paper, we introduce InfFeed, which uses influence functions to compute the influential instances for a target instance. Toward the first objective, we adjust the label of the target instance based on its influencer(s) label. In doing this, InfFeed outperforms the state-of-the-art baselines (including LLMs) by a maximum macro F1-score margin of almost 4% for hate speech classification, 3.5% for stance classification, and 3% for irony and 2% for sarcasm detection. Toward the second objective we show that manually re-annotating only those silver annotated data points in the extension set that have a negative influence can immensely improve the model performance bringing it very close to the scenario where all the data points in the extension set have gold labels. This allows for huge reduction of the number of data points that need to be manually annotated since out of the silver annotated extension dataset, the influence function scheme picks up ~1/1000 points that need manual correction.

Somnath Banerjee, Amruit Sahoo, Sayan Layek et al.

In the continuously advancing AI landscape, crafting context-rich and meaningful responses via Large Language Models (LLMs) is essential. Researchers are becoming more aware of the challenges that LLMs with fewer parameters encounter when trying to provide suitable answers to open-ended questions. To address these hurdles, the integration of cutting-edge strategies, augmentation of rich external domain knowledge to LLMs, offers significant improvements. This paper introduces a novel framework that combines graph-driven context retrieval in conjunction to knowledge graphs based enhancement, honing the proficiency of LLMs, especially in domain specific community question answering platforms like AskUbuntu, Unix, and ServerFault. We conduct experiments on various LLMs with different parameter sizes to evaluate their ability to ground knowledge and determine factual accuracy in answers to open-ended questions. Our methodology GraphContextGen consistently outperforms dominant text-based retrieval systems, demonstrating its robustness and adaptability to a larger number of use cases. This advancement highlights the importance of pairing context rich data retrieval with LLMs, offering a renewed approach to knowledge sourcing and generation in AI systems. We also show that, due to rich contextual data retrieval, the crucial entities, along with the generated answer, remain factually coherent with the gold answer.

Rima Hazra, Sayan Layek, Somnath Banerjee et al.

In the rapidly advancing field of artificial intelligence, the concept of Red-Teaming or Jailbreaking large language models (LLMs) has emerged as a crucial area of study. This approach is especially significant in terms of assessing and enhancing the safety and robustness of these models. This paper investigates the intricate consequences of such modifications through model editing, uncovering a complex relationship between enhancing model accuracy and preserving its ethical integrity. Our in-depth analysis reveals a striking paradox: while injecting accurate information is crucial for model reliability, it can paradoxically destabilize the model's foundational framework, resulting in unpredictable and potentially unsafe behaviors. Additionally, we propose a benchmark dataset NicheHazardQA to investigate this unsafe behavior both within the same and cross topical domain. This aspect of our research sheds light on how the edits, impact the model's safety metrics and guardrails. Our findings show that model editing serves as a cost-effective tool for topical red-teaming by methodically applying targeted edits and evaluating the resultant model behavior.

Mithun Das, Somnath Banerjee, Punyajoy Saha

Online hatred is a growing concern on many social media platforms. To address this issue, different social media platforms have introduced moderation policies for such content. They also employ moderators who can check the posts violating moderation policies and take appropriate action. Academicians in the abusive language research domain also perform various studies to detect such content better. Although there is extensive research in abusive language detection in English, there is a lacuna in abusive language detection in low resource languages like Hindi, Urdu etc. In this FIRE 2021 shared task - "HASOC- Abusive and Threatening language detection in Urdu" the organizers propose an abusive language detection dataset in Urdu along with threatening language detection. In this paper, we explored several machine learning models such as XGboost, LGBM, m-BERT based models for abusive and threatening content detection in Urdu based on the shared task. We observed the Transformer model specifically trained on abusive language dataset in Arabic helps in getting the best performance. Our model came First for both abusive and threatening content detection with an F1scoreof 0.88 and 0.54, respectively.

Somnath Banerjee, Maulindu Sarkar, Nancy Agrawal et al.

Hate speech is considered to be one of the major issues currently plaguing online social media. Repeated and repetitive exposure to hate speech has been shown to create physiological effects on the target users. Thus, hate speech, in all its forms, should be addressed on these platforms in order to maintain good health. In this paper, we explored several Transformer based machine learning models for the detection of hate speech and offensive content in English and Indo-Aryan languages at FIRE 2021. We explore several models such as mBERT, XLMR-large, XLMR-base by team name "Super Mario". Our models came 2nd position in Code-Mixed Data set (Macro F1: 0.7107), 2nd position in Hindi two-class classification(Macro F1: 0.7797), 4th in English four-class category (Macro F1: 0.8006) and 12th in English two-class category (Macro F1: 0.6447).

Somnath Banerjee, Sudip Kumar Naskar, Paolo Rosso et al.

Question classification (QC) is a prime constituent of automated question answering system. The work presented here demonstrates that the combination of multiple models achieve better classification performance than those obtained with existing individual models for the question classification task in Bengali. We have exploited state-of-the-art multiple model combination techniques, i.e., ensemble, stacking and voting, to increase QC accuracy. Lexical, syntactic and semantic features of Bengali questions are used for four well-known classifiers, namely Naïve Bayes, kernel Naïve Bayes, Rule Induction, and Decision Tree, which serve as our base learners. Single-layer question-class taxonomy with 8 coarse-grained classes is extended to two-layer taxonomy by adding 69 fine-grained classes. We carried out the experiments both on single-layer and two-layer taxonomies. Experimental results confirmed that classifier combination approaches outperform single classifier classification approaches by 4.02% for coarse-grained question classes. Overall, the stacking approach produces the best results for fine-grained classification and achieves 87.79% of accuracy. The approach presented here could be used in other Indo-Aryan or Indic languages to develop a question answering system.

Somnath Banerjee, Sahar Ghannay, Sophie Rosset et al.

This paper describes the participation of LIMSI UPV team in SemEval-2020 Task 9: Sentiment Analysis for Code-Mixed Social Media Text. The proposed approach competed in SentiMix Hindi-English subtask, that addresses the problem of predicting the sentiment of a given Hindi-English code-mixed tweet. We propose Recurrent Convolutional Neural Network that combines both the recurrent neural network and the convolutional network to better capture the semantics of the text, for code-mixed sentiment analysis. The proposed system obtained 0.69 (best run) in terms of F1 score on the given test data and achieved the 9th place (Codalab username: somban) in the SentiMix Hindi-English subtask.