Combining Biaffine Model and Constraints Inference for Chinese Clinical Temporal Relation Extraction

The extraction of clinical events and their temporal relation from electronic medical records (EMRs) is crucial and plays a significant role in the development of various intelligent clinical applications. Nevertheless, achieving precise extraction of such information from Chinese electronic medical records (CEMRs) presents a formidable challenge due to the limited availability of Chinese language resources in this field. To address this challenge, we create a dataset comprising clinical events and their temporal relations extracted from CEMRs. Previous methods for extracting clinical events and temporal relations typically relied on sequential pipeline models, which involve initially identifying events and then training classifiers to recognize temporal relations between them. However, this step-by-step approach can result in the accumulation of errors at each stage. Therefore, we propose a joint extraction model utilizing a biaffine architecture to simultaneously extract clinical events and temporal relations. To enhance the model’s performance, we incorporate constraints related to relatedness and irreversibility, resulting in an efficient approach for extracting temporal relations from CEMRs. Our joint extraction model performs admirably on the Chinese dataset we constructed.