Emotion-cognitive reasoning integrated BERT for sentiment analysis of online public opinions on emergencies

Abstract

Sentiment analysis of online public opinions on emergencies (OPOEs) requires accurate and explainable results to facilitate a better understanding of public sentiment and effective crisis management, but it is challenging due to the complexity and diversity of emotions contained in OPOEs. In this paper, we propose an Emotion-Cognitive Reasoning integrated BERT (ECR-BERT) for sentiment analysis of OPOEs. ECR-BERT combines an emotion model and deep learning to provide reliable auxiliary knowledge to improve BERT. Specifically, we use the emotion model proposed by Ortony, Clore, and Collins (OCC) to build emotion-cognitive rules and perform emotion-cognitive reasoning to discover emotion-cognitive knowledge. To mitigate the impact of knowledge noise, we propose a novel self-adaptive fusion algorithm that provides a selection mechanism for the incorporation of knowledge. In addition, we utilize knowledge-enabled feature representation to efficiently exploit inferred knowledge. Our evaluation on four real-world OPOE datasets shows that ECR-BERT significantly outperforms other BERT-based models, achieving state-of-the-art results with an absolute average accuracy improvement of 0.82%, 1.74%, 0.98%, and 1.37% over BERT, respectively. In addition, ECR-BERT provides a detailed explanation of how sentiment polarity is derived from fine-grained emotion categories. The ablation study demonstrates the effectiveness of each technique. In conclusion, ECR-BERT is an excellent choice for sentiment analysis of OPOEs, providing accurate and explainable results for crisis management.