Research On Entity And Relation Extraction Technology Based On Deep Learning

With the increase in the number of Internet users and the popularity of mobile devices,the amount of network activity has led to the exponential growth of data on the Internet.A large proportion of the massive data is unstructured text data,which contains rich valuable information.Entity and relation extraction aims to extract entities and their relations from unstructured texts and express them in the form of triples,so as to present the key information in structured forms,which can meet people's fine-grained information needs.It is the basis of downstream knowledge graph,user portrait,question answering system and other applications.Entity and relation extraction can be divided into pipeline method and joint method.Relation extraction on entity annotations is also called relation classification.The existing relation classification model still has deficiencies in the expression and utilization of features.Improving the effect of relation classification is of great significance in both pipeline method and joint method.The joint method can extract all the triples in the text at one time.However,due to the variable number of entities and relations in the sentence,and the existence of entity overlap,entity redundancy and other problems,the extraction effect of entity and relation is seriously affected.This thesis conducts in-depth analysis and research on the problems existing in relation classification and joint entity and relation extraction,and the main work is as follows:(1)Aiming at the problem that the expression of features is not accurate and the utilization is not comprehensive in the relation classification,a relation classification model integrating multi-entity information is proposed.In the aspect of feature expression,the pre-trained BERT model is used as the feature extractor.The pre-trained BERT model contains rich semantic information and has stronger feature expression ability.In terms of feature utilization,the result of relation classification depends not only on sentence feature,but also on the information of two entities,especially the type feature and dependency relation feature of entities.The type information of entities can reduce the scope of the judgment of the relation and help predict the semantic relation between entities.The dependency paths of entities often contain important information that reflects the relations between entities.The sentence vector,entity vector and entity dependency vector encoded by BERT are merged into the final relation feature,so as to achieve the purpose of accurate expression and fusion of multiple features.Experimental results showed that the model incorporating multi-entity information was 1-17 percentage points higher than the baseline models.(2)In the joint entity and relation extraction,existing models cannot solve the problem of entity overlap and entity redundancy simultaneously,which leads to poor extraction effect.Therefore,a relation-oriented joint extraction model is proposed.The model firstly extracts all the relation types implied in the sentence,and then integrates the extracted relation types into the entity recognition module to identify the entity pairs corresponding to the relation types.In order to avoid the problem of entity redundancy,the priori knowledge of relation type is used to reduce the attention to irrelevant entities.In order to solve the problem of entity overlap,the binary pointer network was used to mark the corresponding entity pairs of the extracted relation types,and finally all the triples in the sentence were extracted.Experiments verify that the joint extraction model based on relation-oriented can effectively solve the problems of entity overlap and entity redundancy simultaneously.An improvement of 1-28 percentage points than the baseline models on F1 score.