Deep Neural Network Based Entity Relation Extraction

With the advent of the era of big data,the scale of information on the Internet expends sharply,how to effectively use these information involves the important information extraction technology in the field of natural language processing.Relation extraction is a very important task in information extraction,which aims at finding the semantic relationships of entity pairs in sentences.This task has great significance for automatic question and answering,medical informatics,ontology learning and so on.This paper focuses on chemical disease relation(CDR)extraction by making full use of semantic information,dependency information and prior knowledge.The main contents are as follows:(1)Research of CDR extraction based on semantic informationThis thesis uses relation instance construction method and pool the instances into two groups at intra-and inter-sentence level,respectively.Relation instance construction method could effective make a balance between positive and negative instances and train high performance classifiers.This paper proposes an extended context semantic representation based on weight and gate mechanism.Two neural network models are used to obtain the semantic representation of the two entities contexts,then the semantic representations are combined by weight and gate mechanism for classification.Experiments show that this method could capture better semantic representation than traditional methods which based on word sequence and position information,and improve the performance of CDR extraction.(2)Research of CDR extraction based on semantic and dependency informationThis thesis further explores the influence of dependency information on CDR extraction.A dependency-based neural network model is constructed by extracting the dependency information along the shortest dependency path(SDP).Then we use model fusion and weight adjustment methods to effectively integrate semantic information and dependency information.Experiments show that the performance of CDR extraction model based on integrating semantic and dependency information is better than the performance of system based on semantic and dependency information alone.(3)Research of CDR extraction based on integrating prior knowledgeThere is a large number of knowledge bases(KBs)in the biomedical field,and these KBs could help us extraction the relations between biomedical entities.This paper uses knowledge representation method to learn the prior knowledge in KBs and acquire the knowledge representation.Then,the attention mechanism is employed to incorporate knowledge representation,semantic and dependency information,and constructs high performance CDR extraction system.Experiments show that prior knowledge could help us identify CDR effectively,and the fusion of prior knowledge,semantic information and dependency information could further improve the performance of CDR extraction.These researches can significantly improve the system performance for CDR extraction,and it also can be more universal and popularized to other relation extraction tasks in other fields.