Explainable threatening tweets identification using abstract meaning representation graph and pattern structure approaches

Abstract

Social media platforms have become an integral part of people's lives, generating a huge amount of data. Threatening speech detection has gained serious attention in recent years. To control the dissemination of such content on social media, the design of an artificial intelligence-based pipeline is an active area of research. The majority of prior studies handled under-resourced languages for detecting threatening speech and very limited research is observed on the English language. Furthermore, there is no work on the explainability of prediction inference for the said task. This study proposes an inherently interpretable threatening speech detection framework for the Twitter platform. The strengths of Pattern Structures and Abstract Meaning Representations graphs are explored first time to support classification and explainability tasks for threatening speech detection. The proposed system automatically mines the context of threatening speech and provides intermediate and final explanations for the prediction inference. A new English corpus is built for the experimental evaluation of the proposed framework for the Twitter platform. The experimental evaluation demonstrates that the proposed framework outperformed the standard baselines and obtained benchmark performance with an accuracy of 73.15% stably on the newly built corpus. In addition, the intermediate and final interpretations offered by the inherently explainable proposed framework provide meaningful and trustworthy prediction inference. The findings of this study have several implications for regulating social media and future research.