Deepthreatexplainer: a united explainable predictor for threat comments identification on Twitter

Deepthreatexplainer: a united explainable predictor for threat comments identification on Twitter

연구 분야: Safety

학회: Social Network Analysis and Mining

Identification of threatening comments on social media platforms has recently gained attention. Prior approaches have addressed this task in some low-resource languages but the interpretability of results was not studied. In addition, approaches in the English language are minimal. To support explainable predictive inference, this research proposes an inherently explainable model for threat comment identification on Twitter. The proposed system incorporates the strengths of Bayesian logistic regression with optimal variational capacity and facilitates the estimation of salient features. Furthermore, the Optimal Variational-Bayesian Logistic Regression (OVB-LR) model can handle the limited labeled dataset issue, achieving the highest performance in classification. The proposed framework automatically mines the threat-related context in language and provides intrinsic explainability for its prediction. This is achieved by posterior-probability approximation, and feature weight calculation to select salient features. For evaluation, a new dataset containing English tweets is designed for threat comment identification. The performance of the proposed framework is evaluated on the threat dataset, and compared with four classical Machine Learning (ML) models (logistic regression, random forest, support vector machine, and k-nearest neighbors) using two feature extraction methods: ELMo embeddings and word uni-gram. The results exhibit that the proposed framework achieves benchmark performance and outperforms four ML models, achieving 81.25% accuracy, 80.85% f1-score for threat class, and 81.24% macro f1-score stably on the newly designed dataset. Furthermore, the OVB-LR model demonstrates comparable interpretations and selects important features that align with features inferred by two post-hoc: Shapley Additive Explanations (SHAP) and Accelerated Model-agnostic Explanations (AcME) Explainable Artificial Intelligence methods. The findings have practical implications for commercial applications and future research.