DPR@DravidianLangTech 2026: Transformer-Based Abusive Content Detection for Tamil Text Targeting Women on Social Media

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Advanced, quick

Summary

A study investigated transformer-based language models for detecting abusive and gender-directed hate speech within Tamil text on social media platforms. This task is particularly challenging due to the Tamil language's complex morphological structure, diverse dialects, transliteration variations, and contextual usage. Researchers fine-tuned and tested five prominent transformer models: mBERT, MuRIL, XLM-RoBERTa, IndicBERT, and Tamil-BERT, utilizing the DravidianLangTech 2026 shared task dataset. Experimental results demonstrated that Tamil-BERT achieved the highest accuracy, reaching 80.72%. Its superior performance is attributed to its Tamil-specific pretraining and advanced morphological analysis capabilities. The developed system secured the 5th rank on the DravidianLangTech 2026 shared task challenge leaderboard, with its source code and fine-tuned models made open-source and publicly accessible.

Key takeaway

For NLP engineers developing hate speech detection systems for low-resource or morphologically complex languages like Tamil, you should prioritize language-specific pretrained models. Tamil-BERT's 80.72% accuracy demonstrates that specialized pretraining and morphological analysis significantly outperform general multilingual models. Consider integrating the open-source Tamil-BERT model and its fine-tuned weights to enhance your system's effectiveness and accelerate deployment.

Key insights

Tamil-specific transformer models excel at detecting abusive content in morphologically complex languages.

Principles

• Language-specific pretraining improves performance.
• Morphological analysis is key for complex languages.
• Shared task datasets drive model evaluation.

Method

Fine-tuning mBERT, MuRIL, XLM-RoBERTa, IndicBERT, and Tamil-BERT on the DravidianLangTech 2026 dataset for abusive content detection.

In practice

• Use Tamil-BERT for Tamil hate speech detection.
• Access open-source models for implementation.
• Evaluate models on DravidianLangTech 2026 data.

Topics

• Abusive Content Detection
• Tamil Language Processing
• Transformer Models
• Tamil-BERT
• DravidianLangTech 2026
• Gender-Directed Hate Speech

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Multi-Stage Training for Abusive Comment Detection in Indic Languages — A multi-stage pipeline combining language preprocessing and model ensembles effectively detects abusive Indic language comments while minimizing false positives.
• A Comparative Study of Pretrained Transformer Models for Quranic ASR: Speech Representations, Label Formats, and Dataset Composition — Fine-tuning pretrained Transformers with specific speech representations and label formats significantly improves Quranic ASR accuracy and efficiency.
• Sentiment Analysis with BERT: The Heart of the Transformer, The Soul of the Text — BERT's bidirectional context understanding via Transformers significantly advanced sentiment analysis and NLP tasks.
• Comparison of Modern Multilingual Text Embedding Techniques for Hate Speech Detection Task — Supervised models with multilingual embeddings and gradient boosted trees excel at hate speech detection.
• Can Small GenAI Language Models Rival Large Language Models in Understanding Application Behavior? — Small GenAI models can rival larger LLMs in malware detection, offering efficiency for resource-constrained environments.
• Anatomy of Data Repositories for the Analysis and Detection of Toxicity in Portuguese — Lexical diversity and annotation practices significantly impact cross-domain generalization in hate speech detection datasets.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.