Recognizing Emotions from Crime Video Comments Based on Pretrained Language Models and Explainable AI

Crime is a pervasive problem that affects the security and stability of society, especially in regions facing socio-economic challenges like Bangladesh. Reviewing and categorizing crime video comments provides a significant solution, enabling authorities to monitor criminal activity, identify emerging patterns, and take proactive measures. This study proposes a method for recognizing emotions from crime video comments by developing a novel dataset of 39,188 entries from multiple YouTube sources such as Taalash, Investigation 360 degree, Crime Patrol, CID etc. which is publicly available in the repository https://github.com/cseku170202/Bangla-Crime-Sentiment-Analysis. This empirical study discovers the necessity for a bidirectional model to successfully capture the context and meaning of crime-related remarks, whereas unidirectional models often fall short. To address the context sensitivity of crime-related texts, this study explores various pretrained language models including BERT, DistilBERT, IndicBERT, BanglaBERT, mBERT, and XLM-RoBERTa. Among the examined models, XLM-RoBERTa outperforms others with a remarkable accuracy of 86.10%. To assure the interpretability of the models, this study utilizes the Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) methods from explainable AI which ensure proper understanding of the model’s reasoning.