Reproduce “Multilingual and Explainable Text Detoxification with Parallel Corpora” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Daryna Dementieva, Nikolay Babakov, Amit Ronen, Abinew Ali Ayele, Naquee Rizwan, Florian Schneider, Xintong Wang, Seid Muhie Yimam and 6 more

This research introduces new parallel detoxification corpora for five languages: German, Chinese, Arabic, Hindi, and Amharic, providing resources for culturally diverse toxicity mitigation. The study performs a detailed explainability analysis of how toxicity manifests across different linguistic and cultural contexts and proposes a Chain-of-Thoughts prompting method, which improves text detoxification performance by leveraging these explainability insights.

Key insights

  • Toxicity manifests differently across languages and cultures, with variations in tone, language type, implied sentiment, and required transformation (e.g., simple word removal in Arabic vs. substantial paraphrasing in Chinese).
  • Explainability analysis revealed that toxic content is often characterized by high toxicity levels, aggressive tone, and vulgar or insulting language, while non-toxic paraphrases tend to be informative or colloquial.
  • Cultural context significantly influences what constitutes toxicity; for example, religious references are common toxic elements in Amharic, while animal metaphors carry derogatory connotations in Chinese.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.