Research On Entity Recognition And Relation Extraction Based On Joint Learning

In knowledge graph construction,task-based dialogue robots and other natural language comprehensive application tasks,entity recognition and relation extraction,which are sub-tasks of information extraction,play a very important role,and are also the basic links of many other downstream natural language processing tasks.Methods for entity recognition and relation extraction model performance have also greatly contributed to improving upper-level tasks in the field of natural language processing.In addition,since the joint learning mode can improve the defects of redundant information and error accumulation in the pipeline mode,this paper studies the entity recognition and relation extraction model based on the joint learning mode from the sentence level and the document level respectively.The purpose is to improve the performance of entity recognition and relation extraction tasks.Aiming at the defect that a single encoder is not enough to capture the information required by two associated tasks in the same space.This paper proposes a sentence-level entity-relation joint learning model that fuses a Two-Encoders model and a minimum-risk training algorithm,where the Two-Encoders model obtains entity labels and relation labels by using a sequence encoder and a table encoder,and the sequence encoding uses table-guided attention to learn the work of the table encoder,and the core of the table encoder is to use a multi-directional multi-dimensional recurrent neural network to obtain multi-dimensional and multi-directional information.In addition,this paper also introduces a minimum risk training algorithm based on the Two-Encoders model to construct a global loss function that allows the entity module and the relation module to perform back-propagation together,so as to strengthen the connection between the entity module and the relation module.Finally,the experiments on the four sentence-level datasets ACE04,ACE05,Co NLL04 and ADE show that the sentence-level model proposed in this paper improves the F1 of the baseline model Two-Encoders by 3.8%,3.3%,1.8% and 2.3%,respectively,confirming that the global loss function helps to optimize the update of the parameters of the joint learning model,and the advantages of the Two-Encoders model on the task of entity and relation extraction.For most of the current document-level entity-relationship joint learning models,there is a problem of redundant output results.This paper proposes a joint learning model NJEREx for extracting entities and entity-level relations from documents.The model uses a fully differentiable high-order reasoning structure to iteratively update the span representation.This high-order coreference resolution model can improve baseline coreference.The digestion model is easy to predict the defects of locally consistent but globally inconsistent entity clusters,and the introduction of coarse factors is used to alleviate the large increase in matrix computation caused by higher-order structures.In addition,the relation classification module of the NJEREx model collects relational features through multi-instance learning multi-level representations combining global entity information and local mention information,and outputs entity-level relations.Finally,experiments on the document-level dataset Doc RED show that the NJEREx model improves the F1 by 3% on the basis of the baseline model JEREX,which verifies the effectiveness of high-order coreference resolution in improving model performance,and combining global entities and local mentions,the advantages of multi-level representations in relational classification modules.