Abstract
Online abusive language detection (ALD) has become a societal issue ofincreasing importance in recent years. Several previous works in online ALDfocused on solving a single abusive language problem in a single domain, likeTwitter, and have not been successfully transferable to the general ALD task ordomain. In this paper, we introduce a new generic ALD framework, MACAS, whichis capable of addressing several types of ALD tasks across different domains.Our generic framework covers multi-aspect abusive language embeddings thatrepresent the target and content aspects of abusive language and applies atextual graph embedding that analyses the user's linguistic behaviour. Then, wepropose and use the cross-attention gate flow mechanism to embrace multipleaspects of abusive language. Quantitative and qualitative evaluation resultsshow that our ALD algorithm rivals or exceeds the six state-of-the-art ALDalgorithms across seven ALD datasets covering multiple aspects of abusivelanguage and different online community domains.