Attention-Based Bi-LSTM Network for Abusive Language Detection

Abstract

Prevention of abusive content in social media platforms has received a lot of attention in recent years. The problem is still a challenge due to the nuances present in the content posted and the tactics of users to pass through the abusive detection algorithms employed by the social media giants. This paper presents a robust deep learning model that takes the advantage of both character and word representations, to address the challenges associated with the text content of a social media post for detecting abusive content. The proposed model uses character CNN to capture morphological information and learns the important features to distinguish abusive content with the help of an Attention-based Bi-LSTM network. In addition, it uses the pooling layer to avoid spatial translations. We confined our model to only the text content of the social media posts due to the limitations of collecting user characteristics for existing datasets and to honour the privacy concerns of users. Our model is dataset agnostic, as revealed by the empirical results on the data sets of multiple social media platforms. It outperformed all previous methods for abusive language detection with text-only features.