Quantifying the Intensity of Toxicity for Discussions and Speakers

Abstract

In this work, from YouTube News-show multimodal dataset with dyadic speakers having heated discussions, we analyze the toxicity through audio-visual signals. Firstly, as different speakers may contribute differently towards the toxicity, we propose a speaker-wise toxicity score revealing individual proportionate contribution. As discussions with disagreements may reflect some signals of toxicity, in order to identify discussions needing more attention we categorize discussions into binary high-low toxicity levels. By analyzing visual features, we show that the levels correlate with facial expressions as Upper Lid Raiser (associated with ‘surprise’), Dimpler (associated with ‘contempť), and Lip Corner Depressor (associated with ‘disgust’) remain statistically significant in separating high-low intensities of disrespect. Secondly, we investigate the impact of audio-based features such as pitch and intensity that can significantly elicit disrespect, and utilize the signals in classifying disrespect and non-disrespect samples by applying logistic regression model achieving 79.86% accuracy. Our findings shed light on the potential of utilizing audio-visual signals in adding important context towards understanding toxic discussions.