Research on causality extraction algorithm for medical text based on BERT and graph attention network

This paper proposes the BERT-CGAT algorithm, designed for causality extraction from medical text by integrating BERT with a graph attention network. This approach enhances the accuracy of causality extraction by leveraging BERT’s semantic feature extraction capabilities alongside the relationship modeling advantages of graph neural networks, addressing the complexities inherent in medical texts. The process begins with the construction of a causal graph, followed by fine-tuning the BERT model on medical text to obtain more accurate entity embeddings. Next, a knowledge fusion channel is employed to integrate text encoding information with the causal structure, which is then fed into the graph attention network. Utilizing a multi-head attention mechanism, the BERT-CGAT model processes information from various subspaces in parallel, thereby improving its ability to capture complex semantic relationships within medical texts. A dual-channel decoding layer is designed to synchronize the extraction of entities and inter-entity causal relationships. On a self-constructed diabetes causal entity dataset, the method demonstrates high accuracy and recall rates, validating the effectiveness of the BERT-CGAT model.