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

TL;DR

This paper introduces BanglaNirTox, a large-scale parallel corpus for Bengali text detoxification, and demonstrates how Pareto-optimized LLMs with Chain-of-Thought prompting improve detoxification quality.

Contribution

It presents a novel pipeline combining Pareto-optimized LLMs and CoT prompting, along with a new dataset for Bengali text detoxification, addressing resource scarcity in low-resource language detoxification.

Findings

Pareto-optimized LLMs with CoT improve detoxification quality

The BanglaNirTox dataset contains 68,041 toxic sentences with labels and detoxified paraphrases

Enhanced consistency in Bengali text detoxification results

Abstract

Toxic language in Bengali remains prevalent, especially in online environments, with few effective precautions against it. Although text detoxification has seen progress in high-resource languages, Bengali remains underexplored due to limited resources. In this paper, we propose a novel pipeline for Bengali text detoxification that combines Pareto class-optimized large language models (LLMs) and Chain-of-Thought (CoT) prompting to generate detoxified sentences. To support this effort, we construct BanglaNirTox, an artificially generated parallel corpus of 68,041 toxic Bengali sentences with class-wise toxicity labels, reasonings, and detoxified paraphrases, using Pareto-optimized LLMs evaluated on random samples. The resulting BanglaNirTox dataset is used to fine-tune language models to produce better detoxified versions of Bengali sentences. Our findings show that Pareto-optimized LLMs with CoT prompting significantly enhance the quality and consistency of Bengali text detoxification.