Coded Hate Speech Detection via Contextual Information

Depeng Xu, Shuhan Yuan, Yueyang Wang, Angela Uchechukwu Nwude, Lu Zhang, Anna Zajicek, Xintao Wu

Abstract: Hate speech on online social media seriously affects the experience of common users. Many online social media platforms deploy automatic hate speech detection programs to filter out hateful content. To evade detection, coded words have been used to represent the targeted groups in hate speech. For example, on Twitter, “Google” is used to indicate African-Americans, and “Skittles” is used to indicate Muslim. As a result, it would be difficult to determine whether a hateful text including “Google” targets African-Americans or the search engine. In this paper, we develop a coded hate speech detection framework, called CODE, to detect hate speech by judging whether coded words like Google or Skittles are used in the coded meaning or not. Based on a proposed two-layer structure, CODE is able to detect the hateful texts with observed coded words as well as newly emerged coded words. Experimental results on a Twitter dataset show the effectiveness of our approach.