Publication

Abstract : Concomitant with the ubiquity of digitization, the wide-open gate, for unfettered use of offensive content in digital platforms, is quite apparent. The veil of anonymity, offered by the digital platforms, has been misused by miscreants who engage in inappropriate, discourteous and rude conversations. The platform, which was meant for people to share their views, ideas constructively and collaborate for the collective betterment of society, is bombarded with offensive tweets with ulterior motives. There have been multiple efforts by various stakeholders to identify and classify such posts automatically employing different algorithms but the changing texting styles by different users have made this a challenging task. The Government of India (GoI) on August 2018 goaded the social media barons to voluntarily enforce sufficient guidelines for the content on their platforms, as the extant measures were rudimentary and inept. While it has been established that the LSTM (Long Short-Term Memory) and the GRU (Gated Recurrent Unit) have been the state-of-the-art sequence modeling neural networks, the efficacy of Temporal Convolutional Network (TCN), which has been proposed as a viable alternate to LSTM/GRU for sequence modeling has not yet been explored for text (offensive) classification. In our work, we have evaluated the performance of TCN to identify and classify offensive language based on the intensity of its offensive content along with the conventional Convolutional Neural Network (CNN) and the state-of-the-art sequence modeling neural networks LSTM and GRU. Unlike LSTM and GRU, TCN exploits parallelism and is able to retain long range history with dilated convolutions and residual blocks. In addition, the TCN classifier was assessed for hate speech, aggression and harassment datasets. In all three datasets, the TCN set new benchmark scores (weighted F1). © BEIESP.