Hate Speech Detection in Audio Using SHAP - An Explainable AI

Abstract

Hate speech detection is a process of recognition of communication media such as text, audio, and/or video, if it contains hatred and/or encourages violence towards a person or a community of people. This is usually based on prejudice against ’protected characteristics’ such as their ethnicity, gender, sexual orientation, religion, age and so on. Complex and sophisticated classifiers based hate speech detection systems are available in the literature. However, the characteristics exhibited by explainable artificial intelligence techniques demonstrated versatile capabilities. This potential is due to the complex classifiers presenting themselves as black-box in nature hence limiting the social acceptability and usability of the developed systems. In this study, video datasets for English and Kiswahili languages were manually collected from YouTube, converted to audio, and used to detect hate speech. Ensemble based classification algorithms have been used for implementation of hate speech detection system. Random Forest classifier recorded an accuracy of 95.8% for English language while for Kiswahili language, Extreme Gradient Boosting classifier achieved an accuracy of 91.8%. To explain the results achieved by these classifiers, in terms of how specific audio-based features contributed to the overall detection of hate speech, SHapley Additive exPlanations technique (SHAP) is used. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024.