KESHEM: Knowledge Enabled Short Health Misinformation Detection Framework

Fei Liu, Yibo Li, Meiyun Zuo

Abstract: Health misinformation detection is a challenging but urgent problem in the field of information governance. In recent years, some studies have utilized long-form text detection models for this task, producing some promising early results. However, we found that most health information online is a short text, especially knowledge-based information. Meanwhile, the explainability of detection results is as important as the detection accuracy. There is no appropriate explainable short health misinformation detection model currently. To address these issues, we propose a novel Knowledge Enabled Short HEalth Misinformation detection framework, called KESHEM. This method extracts abundant knowledge from multiple, multi-form, and dynamically updated knowledge graphs (KGs) as supplementary material and effectively represents semantic features of the information contents and the external knowledge by powerful language models. KG-attention is then applied to distinguish the effects of each external knowledge for the information credibility reasoning and enhance the model’s explainability. We build a credible Chinese short text dataset for better evaluation and future research. Extensive experiments demonstrate that KESHEM significantly outperforms competing methods and accurately identifies important knowledge that explains the veracity of short health information.