Syntactic Enhanced Euphemisms Identification Based on Graph Convolution Networks and Dependency Parsing
Xinru Huang, Jiapeng Zhao, Jinqiao Shi, Yanwei Sun, Xuebin Wang, Liyan Shen · 2023
To evade content moderation policies, social media users adopt euphemisms (also called algospeak, which refers to a benign-looking alternative to sensitive content) to obfuscate their true intentions. Euphemisms identification task aims at understanding the confusing expressions and interpreting the true meanings of euphemisms, which plays a key role in recent euphemisms analysis. The leading euphemisms identification method defines this problem as a text classification task, which establishes a mapping between euphemisms and corresponding real-meaning words in a specific area through fine-grained classification. However, the traditional classification methods leverage explicit textual semantic features, where specific terms and oral swearing in user speech from forums may introduce bias to the classification results on specific areas and limit the performance of euphemisms identification. In this work, we propose SyGCN for euphemisms identification task. Specifically, SyGCN captures the explicit syntactic features by using dependency parsing tree as extended features and integrates syntactic features and semantic features into the Graph Convolution Network to enhance the identification performance. We conduct experiments on three sensitive categories and demonstrate the effectiveness of our method. The experimental results show that our model can effectively utilize the syntactic structure information of dependency parsing to improve the performance of euphemisms identification.