Research On Entity Relation Extraction By Combining Knowledge Bases And Context Information

With the development of Internet and the arrival of Big Data,the amount of biomedical literature has grown rapidly.How to mine and organize entity relation from unstructured information becomes imperious demands.In the field of biomedical,the task of protein-protein interaction(PPI)relation extraction requires extracting interacting protein entity pair from unstructured text,which is of great significance in precision medicine,the development of diseases,controlling cellular homeostasis and so on.In addition,biomedical knowledge bases contain large amount of structural information about entity-relation triples,which can help recognize PPI relation in complicated semantic environment.This paper explores the method of combining knowledge base and context information and mainly studies the PPI relation extraction task.Main research contents are list as follows:(1)Research on PPI extraction based on entity representationsThis paper applies knowledge representation learning model to learn large amount of entity-relation triples in knowledge bases,achieving entity representations and relation representations.Then deep learning models are applied to combine the entity representations and context information,building PPI extraction system which combines entity representations and context information.Experimental results demonstrate that entity representations can effectively increase the capacity of capturing entity related information for a model and improve the precision of PPI extraction.(2)Research on PPI extraction based on relation representationsBased on the relation presentation about an entity pair in KB,this paper uses attention mechanism to extract context information about entity relation,building PPI extraction model which combines relation representations and context features.Experimental results demonstrate that relation embeddings contain directive information and provide important features about PPI.Attention mechanism can further capture the context features about the entity relation and improve the performance of PPI extraction.(3)Research on PPI extraction based on memory networkMemory network consists of hierarchical attention mechanisms,which is more powerful to extract the global information and captures more important features from sequences.This paper uses memory networks to hierarchically extract global information about entities from context and to further integrate with relation representations,building deep PPI extraction model combining knowledge representations and context information.Experimental results demonstrate that memory network could take advantage of the prior knowledge to extract global information in text and multi-level structure helps memory network improve the performance of PPI extraction.Researches of our paper can not only significantly improve the system performance for PPI extraction,but also provide a novel method of using prior knowledge,which is universal and popularized to other tasks.