Event Extraction of Chinese Electronic Medical Records Based on BiGRU-CRF
Siyuan Ma, Longlong Cheng, Shuo Huang, Cui BingJian · 2021 4th International Conference on Artificial Intelligence and Pattern Recognition · 2021
Electronic medical record (EMR) is the product of medical information system, which contains rich medical knowledge, and plays an important role in medical question answering and assistant decision-making [1]. However, as an unstructured text, EMRs cannot be directly utilized by machine. Therefore, how to use information extraction technology to obtain a large amount of accurate knowledge, which is closely related to patients, from EMRs has become a research hotspot. In order to solve the problem of free-text extraction in EMRs, this paper proposes an event extraction model, which finds the corresponding short sentences from the long texts by trigger words, and then recognizes the entities of the short sentences. The method is as follows: Firstly, the text is preprocessed by word segmentation and vectorization, and then the word frequency co-occurrence method is used to identify the trigger words related to events. Secondly, using ID3 algorithm to evaluate the validity of the trigger words. Finally, locate short sentences by triggering words, and using the BiGRU (Bidirectional Gate Recurrent Unit)-CRF (Conditional Random Field) algorithm for sequence labeling to complete the task of event extraction. The result shows that the single-task independent extraction model based on BiGRU-CRF is superior to other deep learning models in the recognition effect of the three tasks: the F1 value of "primary site of tumor" is 96.43%, the F1 value of "primary lesion size" is 98.04%, and the F1 value of "metastatic site" is 88%. Moreover, through experimental comparison, it is found that the single-task independent extraction effect proposed in this paper is better than the multi-task joint extraction effect.