Classifying Offensive Speech of Bangla Text and Analysis Using Explainable AI

The rapid rise of social networking websites and blogging sites not only provides freedom of expression or speech, but also allows people to express society-prohibited behaviors such as online harassment and cyberbullying, which are known as offensive speech or hate speech. Despite the fact that various research work has been done on detecting hate or abusive speech on social networking websites in the English language, the opportunities for research for detecting offensive or abusive speech in the Bengali language remain open due to the computational resource constraints or the lack of standard-labeled datasets for accurate or effective Natural Language Processing (NLP) of Bangla language. In this paper, an Explainable AI approach is used for analysis as well as for detecting offensive comments or speech in the Bengali language is proposed. Moreover, Convolutional Neural Network (CNN) model is used to extract and classify features. Since the Neural Network is time-consuming for extracting features from the dataset, our proposed approach allows people to save time and effort. In the dataset, we classified all user's comments from social media comment sections into four categories: religious, personal, geopolitical, and political. Our proposed model successfully detects Bangla offensive speeches from the dataset (Bengali Hate Speech Dataset) by evaluating Machine Learning algorithms like linear and tree-based models and Neural Networks like CNN, Bi-LSTM, Conv-LSTM, and SVM models. Moreover, we calculate scores for completeness and sufficiency to assess the quality of explanations in terms of fidelity, achieving the results with the accuracy of 78% score, significantly outperforming ML and DNN baselines.