Unified Framework for NER and RE Tasks with Strong Interpretability on Chinese Medicine Instruction Parsing

Abstract

Medicine instructions and related medical guide-lines could serve as valuable knowledge source to serious medicine Knowledge graph (KG) construction. To formulate these nonstructural texts into a well-defined structure, a pre-cise and efficient way to extract information from medicine instructions is needed. Two basic tasks of the extraction, i.e. the Naming Entity Recognition (NER) task and the Relation Extraction (RE) task, have been well studied in traditional way (such as templates) as well as deep neural networks (such as BERT-based models). However, sound interpretability is key to serious medical scenario, which almost rules out the direct application of deep neural networks. In this work, we propose a unified framework for NER task and RE task on Chinese medicine instructions, in which hierarchical dynamic templates serve as basic tools for both tasks. These templates are precise, flexible and strong-interpretable, and they are initially manually constructed and easily expanded in a bootstrap session with the help of distant supervision and BERT-based models. In another point of view, BERT-based models are effectively distilled by rules and manual check, resulting into a cluster of strong-interpretable dynamic templates. At the end of the day, a Chinese medical KG based on the exacted parse result of more than 30 thousand medical instructions was established, and then partly applied to a prescription pre-audit system, which led to more than 10% improvement to the prescription audit accuracy in eight primary hospitals.