Research On Biomedical Entity Relation Extraction

Biomedical domain knowledge mainly comes from the biomedical literatures and biomedical databases.Biomedical entity relations in biomedical literatures and databases represent domain knowledge.Entity relations transform abstract linguistic information in literatures into structured entity information.It is easy for biomedical researchers to acquire domain knowledge and to implement automatic processing of biomedical information.It helps the development of research tools and information revolution in clinics.Therefore,entity relation extraction is an important research area to struct domain knowledge and to discover new biomedical knowledge.Current research focuses on entity relation extraction from huge amount of biomedical literatures.However,entity relation extraction from structured biomedical knowledge is still under-estimated.This dissertation follows the research of biomedical entity relation extraction.Biomedical topic relation model is developed to extract entity relations related to literature topics.Build biomedical knowledge network with topic relations and other biomedical databases and implement biomedical entity discovery method based on link prediction and graph neural network to find entity implicit relations from structured database.This dissertation achieves these contributions to this topic:First,to solve the lack of relevance between extracted relations and literature topics,we develop a topic relation extraction method.Build topic distribution model and entity relation extraction model and extract biomedical entity topic relations from certain literature with strong topic relations.Topic relations are biomedical relations extracted from one literature and are relevant to the topic of this literature.Biomedical relations extracted from huge literature sets cannot build relevance to literatures.This dissertation builds a topic entity distribution model within literature space and turns topic into entity attributes of distribution with a value.Combined with patternbased relation extraction model,it becomes topic relation extraction model.Relations extracted with topic relation extraction model is close to its source literature.The literature is the proof for the extracted relation.The close connection between literature and topic relations can be used to build literature retrieval model with relation query.The experiment proves that the topic relation extraction model can achieve similar accuracy and provide the topic of the literature.Second,to solve the lack of information in multi-type entity knowledge bases and the lack of research on relation extraction from biomedical knowledge bases,we develop a method to build multi-type entity knowledge network and implicit relation extraction method upon it.Build biomedical knowledge network,implement link prediction based entity relation extraction method and extract entity implicit relations from biomedical knowledge network.Take disease entities and gene entities for example.We develop a method to combine databases with same type of entities and build a one-type entity knowledge network.Introducing entity relation databases and topic relations and build connection among entities from databases with different entity types.Combine all databases and it forms cross-type entity knowledge network.biomedical knowledge network includes many kinds of entities and their attributes and relations.Entities with pathways in biomedical knowledge network may have relations,and this kind of relation is defined as implicit relations.Implement retrieval method and implicit relation extraction method in biomedical knowledge network.The experiment shows that implicit relations do exist in biomedical knowledge network and can be proved by recent papers.Third,to solve the insufficiency of research on relation extraction in biomedical knowledge network and domain knowledge rarely used in relation extraction in literatures,we implement biomedical entity relation discovery model based on graph neural network.Build siamese graph neural network model,discover implicit biomedical entity relations from biomedical knowledge network implemented in this dissertation,and assist entity relation extraction from texts.Biomedical knowledge network has a non-Euclid space with multiple types of entities.There are few researches on relation extraction method for knowledge base.The biomedical knowledge network potentially contains lots of unknown relations.Meanwhile,relation extraction models with external domain knowledge is rare.And the external knowledge can be helpful.This dissertation uses Siamese network and graph neural network to learn entity information from biomedical knowledge network and uses entity features to determine the relations.With position embedding and word embedding,extract relations from literatures.The experiment shows that biomedical knowledge network entity relation discovery model has better results than link prediction based method and other methods.