Research On Named Entity Recognition And Entity Relationship Extraction For Document Corpus

In recent years,the technologies of Knowledge Graph have attracted wide attention from industry and academia.Relevant researches have been applied in some downstream tasks like personal recommendation system and intelligent question answering system.Among these technologies,named entity recognition and entity relationship extraction are the core technologies for Knowledge Graph construction.In general,the original corpus of knowledge graph is documents.However,most current researches on entity recognition and relation extraction only focus on sentence-level models.Documents contain more complete contents of entity attributes and relationships between entities,whereas models that are only sentence-oriented ignore the global information of document.Therefore,this study mainly focuses on document-level named entity recognition and document-level relationship extraction model.Explore the method of extracting global information of document feature representation other than word embedding representation.The main work of this article includes:(1)Two attention mechanisms for capturing global entity knowledge of a document are proposed,and a document-level NER model is constructed based on them.The goal of the two attention mechanisms is to effectively capture the remote entity knowledge in the document,filter irrelevant global features,and optimize the effect of NER model.In addition,the advantages of the document-level NER model and the advantages of BILSTM encoder over Transformer encoder are discussed through empirical research.(2)A document level entity relationship extraction model based on entity mention pooling method is proposed.Four entity mention pooling methods are proposed to extract the entity representation of specific entity pairs in the document.In addition,external features such as relative distance representation and entity category are introduced.Through empirical research,the pooling methods,external features and are verified and analyzed.By comparing with other models,the effectiveness of this model is proved.(3)Based on parameter sharing strategy,a joint model framework for document-level entity and relation extraction is proposed.Through setting control experiments,it is proved that the joint model outperforms the single task model in both NER task and relation extraction task.