ylmmcl at Multilingual Text Detoxification 2025: Lexicon-Guided Detoxification and Classifier-Gated Rewriting

2025-07-24

AI Summary

This work addresses multilingual text detoxification—removing toxic content while preserving semantic fidelity and linguistic fluency. We propose a dictionary-guided, classifier-gated iterative sequence-to-sequence framework. Our approach innovatively integrates a multilingual toxic lexicon for fine-grained, explicit toxicity annotation and introduces a learnable classifier gating mechanism to dynamically control the iterative rewriting process, thereby enhancing cross-lingual toxicity detection accuracy and detoxification robustness. The model is built upon the s-nlp/mt0-xl-detox-orpo architecture and fine-tuned using the ORPO optimization strategy. Experiments show that our method achieves 0.922 STA and 0.612 average Jaccard score on the development and test sets, respectively, with an xCOMET score exceeding 0.787. It ranks ninth overall, significantly outperforming existing baselines.

Technology Category

• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

In this work, we introduce our solution for the Multilingual Text Detoxification Task in the PAN-2025 competition for the ylmmcl team: a robust multilingual text detoxification pipeline that integrates lexicon-guided tagging, a fine-tuned sequence-to-sequence model (s-nlp/mt0-xl-detox-orpo) and an iterative classifier-based gatekeeping mechanism. Our approach departs from prior unsupervised or monolingual pipelines by leveraging explicit toxic word annotation via the multilingual_toxic_lexicon to guide detoxification with greater precision and cross-lingual generalization. Our final model achieves the highest STA (0.922) from our previous attempts, and an average official J score of 0.612 for toxic inputs in both the development and test sets. It also achieved xCOMET scores of 0.793 (dev) and 0.787 (test). This performance outperforms baseline and backtranslation methods across multiple languages, and shows strong generalization in high-resource settings (English, Russian, French). Despite some trade-offs in SIM, the model demonstrates consistent improvements in detoxification strength. In the competition, our team achieved ninth place with a score of 0.612.

Problem

• Multilingual text detoxification using lexicon-guided tagging
• Improving detoxification precision with cross-lingual generalization
• Combining sequence-to-sequence models and classifier-gated mechanisms

Innovation

• Lexicon-guided toxic word annotation
• Fine-tuned sequence-to-sequence model
• Classifier-gated iterative rewriting