Linguistic annotation of high-stakes narrative data is often constrained by data confidentiality, domain expertise, and the lack of large-scale multi-annotator pipelines. We present a human-in-the-loop framework for auditing annotation discrepancies in crash narratives, combining structured labels, narrative-based annotation, and expert adjudication. Using 9,387 crash reports, we conduct a multi-layer analysis of disagreement across annotation sources. Nearly half of the records (49.4%) exhibit discrepancies between structured and narrative labels, driven mainly by unsupported structured assignments. In contrast, narrative-based annotation achieves near-perfect agreement with adjudication (𝜅 = 0.990), indicating strong consistency when grounded in textual evidence. We introduce a taxonomy of discrepancies, showing refinement opportunities and missing details are the most common, while linguistic factors such as hedging and underspecification contribute to ambiguity. We further show that annotator-reported uncertainty strongly predicts annotation difficulty, with uncertain records nearly nine times more likely to disagree with structured labels. These findings highlight limitations of administrative coding and support a scalable, uncertainty-guided annotation paradigm for restricted-access domains.

Text classification in low-resource languages has become increasingly important due to the rapid growth of user-generated digital content. While multitask learning has long been studied in NLP, the use of LLMs for multitask text classification in low-resource languages such as Bengali remains underexplored. Although LLMs are inherently multilingual and multitasking, their effectiveness in structured multitask classification settings for Bengali has not been systematically evaluated. In this work, we investigate how LLMs can be leveraged for multitask Bengali text classification across five domains: sentiment analysis, aggressive text detection, fake news detection, news categorization, and emotion analysis. We compare in-context learning strategies—including zero-shot, one-shot, and chain-of-thought prompting—with parameter-efficient fine-tuning approaches. Our findings show that CoT prompting does not consistently improve performance and often degrades performance, highlighting the instability of prompt-based adaptation in low-resource settings with limited pretraining exposure. Moreover, reasoning-optimized models such as DeepSeek-R1 exhibit substantial performance drops, indicating that enhanced reasoning capabilities alone cannot overcome the challenges posed by low-resource settings. Among the evaluated mLLMs, Gemma-3-4B demonstrates the most stable and balanced cross-task performance under both in-context learning and parameter-efficient fine-tuning, making it a strong backbone candidate for multitask Bengali text classification. These results provide empirical evidence on the limitations of prompting and the advantages of lightweight fine-tuning for low-resource multilingual NLP.

The rise of social media has significantly facilitated the rapid spread of hate speech. Detecting hate speech for content moderation is challenging, especially in low-resource languages (LRLs) like Telugu. Although some progress has been noticed in hate speech detection in Telegu concerning unimodal (text or image) in recent years, there is a lack of research on hate speech detection based on multimodal content detection (specifically using audio and text). In this regard, DravidianLangTech has arranged a shared task to address this challenge. This work explored three machine learning (ML), three deep learning (DL), and seven transformer-based models that integrate text and audio modalities using cross-modal attention for hate speech detection. The evaluation results demonstrate that mBERT achieved the highest F-1 score of 49.68% using text. However, the proposed multimodal attention-based approach with Whisper-small+TeluguBERT-3 achieved an F-1 score of 43 68%, which helped us achieve a rank of 3rd in the shared task competition.

Semantic textual relatedness is crucial to Natural Language Processing (NLP). Methodologies often exhibit superior performance in high-resource languages such as English compared to low-resource ones like Marathi, Telugu, and Spanish. This study leverages various machine learning (ML) approaches, including Support Vector Regression (SVR) and Random Forest, deep learning (DL) techniques such as Siamese Neural Networks, and transformer-based models such as MiniLM-L6-v2, Marathi-sbert, Telugu-sentence-bert-nli, and Roberta-bne-sentiment-analysis-es, to assess semantic relatedness across English, Marathi, Telugu, and Spanish. The developed transformer-based methods notably outperformed other models in determining semantic textual relatedness across these languages, achieving a Spearman correlation coefficient of 0.822 (for English), 0.870 (for Marathi), 0.820 (for Telugu), and 0.677 (for Spanish). These results led to our work attaining rankings of 22th (for English), 11th (for Marathi), 11th (for Telegu) and 14th (for Spanish), respectively.