Research On Knowledge-Guided Entity Linking Method In The Medical Field

With the deepening of social informatization and the continuous development of informatization technology in the medical field,a large amount of electronic medical corpora has been accumulated,such as electronic medical records,medical literature,online medical question and answer data,etc.However,these data are not always clean.If we want to use these data to mine relevant medical knowledge,such as disease-symptom mining,it is necessary to normalize relevant entities and then we could use the existing knowledge base to mine and infer relevant information.The task of mapping non-canonical entities to normaliezd ones in the standard entity library is called the entity linking task.Compared with open domain entity linking,medical entity linking faces various difficulties.On the one hand,due to the diversity of non-standard words and noise in data sources,the recall rate of candidate entity generation is not high.On the other hand,the current entity linking does not make full use of the relevant knowledge in the domain,and only uses the surface similarity or semantic similarity between standard words and original words to match words.In view of the above problems,this paper mainly improves the effectiveness of entity linking from the following three aspects:· Textual Knowledge Mining and Candidate Entity Generation Through exploration of the characteristics of the dataset,we propose a text preprocessing method for clinical diagnosis and treatment data,which reduces the difference between the original word and the standard word and improves the effect of the text matching.At the same time,only a single text surface similarity method is usually used during the candidate entity recall period.This paper proposes a method that combines multiple surface similarity features,which effectively improves the recall rate.· Incorporating Lexical Information Entity Linking Method In addition to character-based semantic similarity,lexical information also plays an important role in entity linking.This paper proposes a method to incorporate lexical information based on the BERT model,and uses a data augumentation method and self-supervised comparative learning method to imporve the representation of original words and standard entities,and the effect of entity linking is improved.· Knowledge-graph Oriented Entity Linking Method After generating candidate entities for the current original word,it is necessary to score the similarity between the current original word and all its candidate entities.For the traditional method that only relies on the semantic similarity or word surface similarity of the text itself to score,this paper proposes a method that integrates knowledge in the knowledge graph by mapping entities to the knowledge graph,effectively improving the accuracy of entity matching.