Entity Pair Recognition using Semantic Enrichment and Adversarial Training for Chinese Drug Knowledge Extraction
Feng Gao, Lunsheng Zhou, Jinguang Gu · 2021
Existing knowledge extraction methods in pharmacy often use natural language processing tools and deep learning model to identify drug entities and extract their relationships from drug instructions, thus obtaining drug-drug or drug-disease knowledge. However, sentences in drug instructions may contain multiple drug-related entities, and existing methods lack the capability of identifying valid the "drug-drug" or "drug-disease" entity pairs. This will introduce significant noise data in the subsequent tasks such as entity relationship extraction and knowledge graph construction. Meanwhile, some mentions in the sentence can have hierarchical relations even if they do not form valid entity pairs, such information is also crucial to knowledge extraction. To solve these two problems, this paper proposes an entity pair verification model based on entity semantic enhancement and adversarial training. Through the experiment on more than 2000 kinds of drug instructions data, the experimental results show that the F1 value of the model for entity pair verification is up to 98.65%, which is up to 9.37% compared with the existing methods.