Audio-based Toxic Language Classification using Self-attentive Convolutional Neural Network
Midia Yousefi, Dimitra Emmanouilidou
Abstract
The monumental increase in online social interaction activities such as social networking or online gaming is often riddled by hostile or aggressive behavior that can lead to unsolicited manifestations of cyberbullying or harassment. In this work, we develop an audio-based toxic language classifier using self-attentive Convolutional Neural Networks (CNNs). As definitions of hostility or toxicity can vary depending on the platform or application, in this work we take a more general approach for identifying toxic utterances, one that does not depend on individual lexicon terms, but rather considers the entire acoustical context of the short verse or utterance. In the proposed architecture, the self-attention mechanism captures the temporal dependency of the verbal content by summarizing all the relevant information from different regions of the utterance. The proposed audio-based self-attentive CNN model is evaluated on a public and an internal dataset and achieves 75% accuracy, 79% precision, and 80 % recall in identifying toxic speech recordings.