Semantic learning and attention dynamics for behavioral classification in police narratives

Abstract

The proactive identification of behavioral health incidents concerns from police reports is a critical yet underexplored area. The law enforcement officers provide follow-up services to improve community life by manually analyzing and identifying generated public narrative reports after the 911 incident calls. Therefore, automatically identifying these behavioral health calls from public narrative reports helps reduce the current manual labor-intensive identification process for law enforcement officers. In this work, we introduce a novel, multi-faceted approach that combines manual expert annotations, natural language processing (NLP), and cutting-edge machine learning strategies to classify and understand these incidents within police narratives efficiently. Our proposed method retrieves relevant narrative reports utilizing domain knowledge as behavioral health cues as terms/keywords. Our approach automatically extracts different categories of behavioral health cues by utilizing the limited domain knowledge enabled behavioral health terms using a cosine similarity-based thresholding approach. The behavioral health classification model employs an automatic attention-aware feature representation of terms/keywords, categories, and narrative reports to identify behavioral health cases with an accuracy of 85%. Extensive evaluation shows that our proposed model outperforms all the state-of-the-art models by approximately 4%.