Research On Entity Relation Extraction Technology Based On Graph Neural Network

With the vigorous development of online content such as web pages and social media,how to automatically acquire knowledge from massive unstructured text sources is an urgent problem to be solved.The task of named entity recognition(NER)and relation extraction(RE)came into being in this context,aiming at identifying named entities in unstructured text sources and mining the semantic relationship between entity pairs,and realizing the mapping from unstructured text language materials to structured knowledge.Named entity recognition and relation extraction are the core tasks in the automatic construction of large-scale knowledge graph,and also provide knowledge support for applications such as question answering system and search engine.At present,entity relation extraction methods are generally divided into intra-sentence relation extraction and inter-sentence relation extraction.The traditional entity relation extraction in sentences often adopts pipeline method,which has the problems of cumulative error propagation and lack of interactive information of entity relation.Therefore,the current entity relation extraction task mainly adopts joint extraction method.In practical applications,the source data generally appear in the form of long text,and it is often necessary to reason multiple sentences to speculate the fact of entity relationship.Therefore,in view of the above situation,this thesis proposes two models based on intra-sentence entity relation extraction and dialogue cross-sentence relation extraction.The research contents and innovations are summarized as follows :1.Aiming at the problem of triple overlapping relation detection in entity relation joint extraction task,this thesis proposes an entity relation joint extraction model based on relation self-adaptive(which is called MA-DCGCN).The model uses the multi-head attention mechanism to assign non-exclusive probability space for each relationship type between entities,and calculate independent attention weight.And the interaction between entities and relations can be captured by dense connection graph convolutional network(DCGCN)to obtain deeper graph structure information.In order to prove the superior performance of the model in this thesis,sufficient experiments are carried out on two widely used public datasets NYT and Web NLG.The results show that compared with other mainstream methods,the F1 score of the model in this thesis increases by 1.8 %on both data sets,which effectively improves the detection performance of overlapping triples.2.Aiming at the problems of high frequency and low information density of personal pronouns in dialogue cross-sentence relation extraction,this thesis proposes a dialogue level cross-sentence relation extraction model based on position-aware refinement mechanism(which is called PAR-DRE).Firstly,a heterogeneous mention dialogue graph is constructed by introducing pronoun reference nodes,and the multi-granularity nodes enrich the graphical reasoning information.Then,the proposed position-aware refinement mechanism is used to process each heterogeneous mention dialog graph to obtain the mention representation containing location-aware information,which makes it more discriminative.Finally,on the entity dialogue graph after merging mention nodes,the path information between entity pairs is fused,and the above entity representation is aggregated into the final entity pair representation,which is input into the standard multi-label classifier for relation label prediction.Experimental results on Dialog RE show that our method improves the performance of F1 by 1.25% compared with the current mainstream methods.