Chinese Toxic Language Mitigation via Sentiment Polarity Consistent Rewrites
This work addresses the challenge of improving online interaction quality in Chinese by mitigating toxic language without distorting sentiment, though it is incremental as it builds on existing detoxification methods with a new dataset.
The paper tackles the problem of detoxifying offensive language in Chinese while preserving the speaker's original intent, especially in subtle cases like emojis and homophones, by introducing ToxiRewriteCN, a dataset of 1,556 annotated triplets, and evaluating 17 LLMs, finding that all models struggle with balancing safety and emotional fidelity in complex scenarios.
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.