BanglaNirTox: A Large-scale Parallel Corpus for Explainable AI in Bengali Text Detoxification

Bengali toxic language online is a significant but understudied problem: The paper identifies toxic language as "prevalent" in Bengali online environments with "few effective precautions," highlighting a gap in responsible AI efforts for non-English languages.

A new dataset of 68,041 toxic Bengali sentences has been created to address the resource gap: BanglaNirTox provides paired toxic and detoxified sentences with explanations, offering researchers concrete tools to develop better content moderation systems for Bengali-speaking communities.

Combining AI reasoning with optimization techniques improves detoxification quality: The study found that using "Pareto-optimized LLMs with Chain-of-Thought prompting significantly enhance the quality and consistency" of detoxification, suggesting that thoughtful AI design choices matter for responsible content moderation.

Explainability is built into the solution from the start: The dataset includes "reasonings" alongside detoxified text, meaning the AI doesn't just remove toxic content—it explains why changes were made, supporting transparency and trust in automated moderation systems.

This work addresses a language equity issue in AI safety: By focusing on Bengali rather than only high-resource languages like English, the research helps ensure that responsible AI protections extend to underserved linguistic communities.