Homophone-aware offensive language detection via semantic-phonetic collaboration

Abstract

The increasing use of implicit and obfuscated expressions poses significant challenges to offensive language detection in Chinese online platforms. In particular, users often exploit homophone substitutions to bypass keyword-based moderation, making traditional detection systems inadequate. This study addresses the problem of detecting offensive content masked through homophonic substitutions, which retain aggressive intent while altering character representations. Existing methods fall into two main categories: (1) semantic-only models, which struggle with phonetic manipulations due to their reliance on text features alone, and (2) auxiliary-enhanced models, which incorporate phonetic or syntactic signals but lack deep integration between modalities. To overcome these limitations, we propose a lightweight dual-branch model that separately encodes textual semantics and pinyin phonetics under a multi-view learning framework. A Dual-Branch Interactive Training strategy is introduced to enable dynamic cross-modal alignment via contrastive objectives, allowing each modality to mutually refine the other and enhance robustness to adversarial inputs. We conduct experiments on two benchmark datasets, COLD and SWSR, both of which are augmented with varying levels of homophone noise to simulate real-world evasion strategies. The proposed model outperforms all baseline models, achieving an average F1-score improvement of 6.3 % under high-noise conditions, while reducing inference latency and memory usage by more than 60 %, demonstrating both effectiveness and efficiency for real-time deployment. We will release the source code for further use by the communityhttps://github.com/hjhhlc/DBIT.