Chinese Toxic Language Mitigation via Sentiment Polarity Consistent Rewrites

2025-05-21

AI Summary

This work addresses the challenge of preserving original intent and sentiment polarity while detoxifying implicit toxicity in Chinese online text—such as offensive implications conveyed via emojis, homophonic substitutions, or conversational context. We introduce ToxiRewriteCN, the first sentiment-polarity-aligned Chinese detoxification rewriting dataset. It comprises high-quality, real-world triplets (toxic input, toxicity span annotation, sentiment-aligned rewrite) across five scenario categories and incorporates an explicit sentiment consistency constraint—a novel contribution. Comprehensive evaluation across 17 state-of-the-art large language models—including Mixture-of-Experts architectures—reveals that existing models consistently sacrifice sentiment fidelity to achieve safety, exhibiting notably poor sentiment polarity alignment under implicit toxicity. ToxiRewriteCN is publicly released to advance controllable, sentiment-aware Chinese detoxification research.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Constraint Satisfaction and Optimization: Satisfiability Modulo Theories

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Large pretrained models with web data
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

Detoxifying offensive language while preserving the speaker's original intent is a challenging yet critical goal for improving the quality of online interactions. Although large language models (LLMs) show promise in rewriting toxic content, they often default to overly polite rewrites, distorting the emotional tone and communicative intent. This problem is especially acute in Chinese, where toxicity often arises implicitly through emojis, homophones, or discourse context. We present ToxiRewriteCN, the first Chinese detoxification dataset explicitly designed to preserve sentiment polarity. The dataset comprises 1,556 carefully annotated triplets, each containing a toxic sentence, a sentiment-aligned non-toxic rewrite, and labeled toxic spans. It covers five real-world scenarios: standard expressions, emoji-induced and homophonic toxicity, as well as single-turn and multi-turn dialogues. We evaluate 17 LLMs, including commercial and open-source models with variant architectures, across four dimensions: detoxification accuracy, fluency, content preservation, and sentiment polarity. Results show that while commercial and MoE models perform best overall, all models struggle to balance safety with emotional fidelity in more subtle or context-heavy settings such as emoji, homophone, and dialogue-based inputs. We release ToxiRewriteCN to support future research on controllable, sentiment-aware detoxification for Chinese.

Problem

• Mitigating toxic Chinese language while preserving sentiment polarity
• Addressing implicit toxicity in emojis, homophones, and discourse context
• Evaluating LLMs for balanced detoxification and emotional fidelity

Innovation

• First Chinese detoxification dataset preserving sentiment polarity
• Evaluates 17 LLMs across four key dimensions
• Covers five real-world toxic language scenarios