Offensive Text Detection using Temporal Convolutional Networks

Abstract

Offensive text refers to the use of language in a rude, insulting or hateful way that may upset people. More specifically, hate speech refers to the numerous means of expression that is intended to spread, provoke, promote or justify hatred, violence or discrimination against a person or a group of persons for various reasons. There is an increase in the amount of offensive and hate speech with increasing access to internet. Automated systems of offensive text detection have largely been developed on traditional machine learning techniques such as SVM, Logistic Regression, and the use of deep neural networks such a convolution neural networks (CNN), Long Short Term Memory (LSTM), Gated Recurrent Units (GRU). However, LSTM/GRU models have a large memory footprint and suffer from vanishing memory problem. A more recent development in neural networks is the Temporal Convolutional Network (TCN). In this paper, we investigate TCN in the domain of text classification. We also make use of FastText, GloVe and Paragram SL999 as feature space and compare results with RNN based models such as LSTM and GRU. Our experiments on a dataset of 24k annotated tweets indicate that TCN marginally outperforms RNN based models.