Tackling Toxic Online Communication with Recurrent Capsule Networks

Abstract

Internet has provided everyone a platform to productively exchange ideas, learn new things and have meaningful conversation. To make online interactions fruitful it is necessary the user feels comfortable with sharing information without the menace of online hate which includes insults, personal attacks, identity hate, threats and so on. The first step to combating this problem would be the identification of such online behaviour. Framing the problem as text classification, we present a novel and versatile model in this paper which employs Recurrent Neural Network and Capsule network as its backbone and captures contextual information to a larger extent when learning word representations in the text. A series of experiments are conducted on Wikipedia’s talk page edits provided by Jigsaw in Kaggle’s toxic comment classification challenge. The experimental results show that the proposed model outperforms other traditional state-of-the-art models on the dataset, thereby proving the effectiveness of capsule networks for multi-label text classification. The superior performance of architecture is also confirmed by results obtained on traditional benchmark datasets such as AG News, IMDB Large Movie Review and Yelp Reviews data.