Research And Application Of Joint Extraction Model Of Entity And Relation Based On Deep Learning

With the prosperity of mobile Internet,Internet information has grown exponentially,with a significant proportion of text messages.Named entity recognition and relation extraction can help people to extract entities and their relations from massive texts.It is the basic of domain knowledge graph construction,human-machine conversation and other tasks in the field of natural language processing,and is significantly valuable.Entity recognition and relation extraction are regarded as two independent sub-tasks in the traditional pipeline model,leading to error propagation,information redundancy,and the failure to establish two sub-tasks dependencies.These problems will affect the final extraction performance.In order to solve these defects in the pipeline model,the entity and relation joint extraction model is studied based on deep learning in this thesis.The disadvantages in related works about entity recognition,relation extraction and their joint extraction are investigated.Two joint extraction models in view of parameter sharing and tagging strategy innovation are proposed.The main works are listed as follows:1.To solve the ignored inherent dependence of two subtasks in the pipeline model,a joint extraction model with clause information fusion is proposed.The model is based on parameter sharing,using the pre-trained language model BERT as a shared encoding layer for entity recognition and relation extraction to establish their connection.Entity pair encoding information and clause information are combined to improve the performance of relationship classification.Finally,the performance of entity recognition and relationship extraction on the Co NLL04 dataset achieves 89.2% and 71.5% respectively in term of F1,which proves the effectiveness of the proposed model.2.Aiming at information redundancy,a joint extraction model with novel decomposition strategy is proposed.Joint extraction task is transformed into two subtasks,which are head entity recognition,tail entity and relation extraction.In the training stage,category imbalance problem caused by tagging strategy is alleviated by means of bias weights.In the prediction stage,the performance is improved greatly based on sentence semantic relationship.In comparison to other joint extraction models,the best performance is achieved,which is 88.6% in term of F1 on the NYT dataset.3.Two joint extraction models are developed for music domain-oriented entity and relation extraction,by which music text is transformed into structured triplets and knowledge graph in music domain can be constructed.