Rule-Enhanced Pharmaceutical Instructions Information Extraction Based on Deep Learning

Pharmaceutical instructions information extraction is to transform the core content of drug manuals into formatted data, including entity recognition and relationship extraction, which is an essential value for rational drug use and drug management. The task is challenging due to the various formats of drug manuals and the noise of watermarks and stamps on the original images. For this reason, this paper proposes the Rule-enhanced Drug Description Information Extraction based on deep learning method, which is a pipeline-based method that includes three steps of data preprocessing, text recognition and sorting, and information extraction. The results of the CHIP2023-Recognition and Entity Relationship Extraction of Pharmaceutical Instruction task show that the method proposed in this paper is able to extract the information embedded in pharmaceutical instructions more effectively, and achieves first place in this evaluation. Significantly, it achieves the first place in the evaluation and is significantly better than the second place. However, there are obvious things that could be improved in this method, as the pipeline approach contains many different modules and processing steps, which makes optimization very difficult. In the future, we try end-to-end extraction of pharmaceutical instruction information on the basis of multimodal large models.