DIPE: a diagnosis-assisted inquiry point extractor towards medical dialogues

Automatic knowledge extraction from medical dialogues has emerged as an increasingly significant problem in modern medical care. However, diagnosis characteristics of medical texts and imbalanced distribution of item categories within inquiry points are ignored in traditional methods used for medical information extraction, resulting in unsatisfactory performance. In this paper, we propose a Diagnosis-assisted Inquiry Point Extractor (DIPE), where a novel hierarchical attention mechanism, named WSWC (Word-Sentence-Window-Context), is devised to simulate diagnosis-oriented inference and further effectively captures semantic correlation in utterances. Additionally, we construct an imbalance-aware loss function to mitigate the imbalanced distribution of entity categories within inquiry points by assigning weights based on the disparity in sample counts for each category. Experimental results on two public datasets demonstrate that DIPE is an effective solution for inquiry point extraction in medical dialogues.