A Toxic Euphemism Detection framework for online social network based on Semantic Contrastive Learning and dual channel knowledge augmentation

Abstract

For real-time content moderation systems, detecting toxic euphemisms remains a significant challenge due to the lack of available annotated datasets and the ability to deeply identify euphemistic toxicity. In this paper, we proposed the TED-SCL framework (Toxic Euphemism Detection based on Semantic Contrastive Learning) to solve these problems. Firstly, we collected nearly 8 million comments and constructed a toxic euphemism dataset (TE-Dataset), which contains 18,971 comments, covering six topics and 424 PTETs (Potential Toxic Euphemism Terms). Next, we employed contrastive learning to separate toxic euphemism samples from harmless ones in semantic space and enhance the model’s ability to capture subtle differences. Lastly, we utilized a dual channel knowledge augmentation module to integrate background knowledge with toxic comments and improve the identification of toxic euphemisms. Experimental results demonstrate that TED-SCL outperforms existing SOTA in toxic euphemism detection tasks, achieving accuracy of 93.94%, recall of 93.36%, and F1 score of 93.23%. Furthermore, TED-SCL demonstrates better generalization, zero-shot capability, and greater robustness on different topics and datasets, which provides a new way for real-time content moderation systems to detect euphemistic and implicit toxicity effectively.