Hate speech classification on social media using a service framework

Abstract

It is indeed a challenge for the existing 
machine learning approaches to segregate the 
hateful content from the one that is merely 
offensive. One prevalent reason for low 
accuracy of hate detection with the current 
methodologies is that these techniques treat 
hate classification as a multiclass problem. In 
this article, we present the hate identification 
on the social media as a multilabel problem. 
To this end, we propose a CNN-based service 
framework called “HateClassify” for labeling 
the social media contents as the hate speech, 
offensive, or nonoffensive. Results 
demonstrate that the multiclass classification 
accuracy for the CNN-based approaches 
particularly sequential CNN (SCNN) is 
competitive and even higher than certain 
state-of-the-art classifiers. Moreover, in the 
multilabel classification problem,
sufficiently high performance is exhibited by 
the SCNN among other CNN-based 
techniques. The results have shown that using 
multilabel classification instead of multiclass 
classification, hate speech detection is 
increased up to 20%.