Improving Hate Speech Classification Through Ensemble Learning and Explainable AI Techniques

Abstract

Identifying offensive and discriminatory content, commonly referred to as hate speech, within textual data is a critical task. This study addresses the task of identifying hate speech in textual data, focusing on the challenge of selecting optimal word embedding methods and classifiers. Leveraging the Google Jigsaw dataset, the research employs explainable artificial intelligence (XAI) for hate speech detection. Following preprocessing, which includes converting text to lowercase, removing punctuation, extra whitespace, numbers, and non-ASCII characters, a thorough analysis reveals high-frequency words. The research extensively compares three-word embedding techniques—CountVectorizer, GloVe, and bidirectional encoder representations from transformers (BERT)—in combination with two machine learning models (support vector classifier and logistic regression) and four deep learning models [artificial neural network (ANN), recurrent neural network (RNN), bidirectional gated recurrent unit (Bi-GRU), bidirectional long-short term memory (Bi-LSTM)] for hate speech detection. The fusion of BERT with a bidirectional gated recurrent unit (Bi-GRU) achieved an impressive accuracy of 92%, and an ensemble of the top-performing models further improves accuracy by nearly 2%. To enhance result interpretability, the study employs XAI techniques such as local interpretable model agnostic explanations (LIME) and Shapley additive explanations (SHAP) on the top-performing ensembled model to provide insights into its predictions. The paper concludes by suggesting potential future research directions, including exploring additional embedding techniques and models, addressing dataset generalizability, improving interpretability methods, and considering computational resource constraints.