Multi-task CNN for Abusive Language Detection

Abstract

Abusive language detection serves to ensure a compelling user experience via high-quality content. Different sub-categories of abusive language are closely related, with most aggressive comments containing personal attacks and toxic content and vice versa. We set a multi-task learning framework to detect different types of abusive content in a mental health forum to address this feature. Each classification task is treated as a subclass in a multi-class classification problem, with shared knowledge used for three related tasks: attack, aggression, and toxicity. Experimental results on three sub-types of Wikipedia abusive language datasets show that our framework can improve the net F1-score by 7.1%, 5.6%, and 2.7% in the attack, aggressive, and toxicity detection. Our experiments identified multi tasking framework act as an effective method in abusive language detection.