The massive increase in online social interaction activities such as social networking and online gaming is frequently marred by hostile or aggressive behavior, which can result in uninvited manifestations of cyberbullying or harassment. In this paper, we use self-attentive Convolutional Neural Networks to build an audio-based toxic language classifier (CNNs). Because definitions of hostility or toxicity differ depending on the platform or application, we take a more general approach to identifying toxic utterances in this work, one that does not rely on individual lexicon terms, but rather takes into account the entire acoustical context of the short verse or utterance. The self-attention mechanism in the proposed architecture captures the temporal dependency of verbal content by summarizing all relevant information from different regions of the utterance. On a public and an internal dataset, the proposed audio-based self-attentive CNN model achieves 75% accuracy, 79% precision, and 80% recall in identifying toxicspeech recordings.

Keywords : Toxic Language Detection, Self-Attention, Hate Speech, Sentiment Detection, Cyberbullying.