Research On Relation Extraction Algorithm For Knowledge Graph

Knowledge graph,as a current efficient information management technology,can improve the performance of search engines and question and answer systems,and has a wide range of research prospects.The construction of knowledge graphs is the basis of applications,and relationship extraction,as one of the main tasks of information extraction,can provide entityrelationship triples for knowledge graph construction and is a key technology for knowledge graph construction.The remotely supervised relationship extraction method solves the problem of needing large amount of annotated data with supervision by automatically aligning data to a remote knowledge base,and is suitable for application to the field of knowledge graph construction,but there is a large amount of noisy data in the annotated corpus acquired by remote supervision,which will affect the performance of the model.The construction of knowledge graphs is a process of continuous improvement.Using the semantic information of knowledge graphs themselves can improve the performance of relationship extraction and provide higher quality entity-relationship triples.In this dissertation,we study the relationship extraction for knowledge graphs based on the knowledge of relationship extraction and knowledge graphs,combined with the requirements of knowledge graphs and related technologies.The main research work in this dissertation is described as follows.1)A relationship extraction model(EA+PCNN+ATT)using Entity Attention(EA)mechanism is proposed,based on the Entity Attention mechanism to capture the influence of different words on the relational expression of an utterance,in conjunction with PCNN(segmental convolutional neural network)to extract the textual features of an utterance,enabling the model to extract comprehensive semantic features,in addition to adding a sentence-level In addition,an attention mechanism is added at the sentence level to reduce the influence of mislabelling on the relational representation.In addition,a sentence-level attention mechanism is added to reduce the impact of mislabelling on relational expressions.Compared with existing methods,the impact of different words on relational expressions can be noticed and the performance of the relational extraction model is improved.2)Proposed Reinforcement learning based relationship extraction model(PCNN+RL).Since the remotely supervised data alignment process generates noisy data,remotely supervised relationship extraction models generally use training data at the bag(packet)level and use coarse-grained supervised signals,which will reduce the performance of the model.The reinforcement learning-based instance selector selects the statements that best express the bag relationship labels as the training data for the relationship classifier.Compared with existing methods,the noise and coarse-grained problem of remotely supervised relation extraction is solved,providing high-quality entity-relationship triples for knowledge graph construction.3)The proposed relationship extraction model combined with knowledge graph embedding(KG+PCNN),the construction of knowledge graph is a process of continuous improvement and update,the knowledge graph itself contains rich semantic information,using this information can extract the implicit relationships in the sentences.KG+PCNN first uses PCNN to extract semantic features,while using the knowledge graph embedding model to obtain knowledge representation,the two models The two models are jointly learned to effectively exploit the prior information of the knowledge graph and improve the performance of the relation extraction model.Compared with existing methods,the joint training of the two models using global loss functions enables the two models to contribute to each other and improve the performance of the models.Experiments on a remotely supervised benchmark dataset demonstrate the effectiveness of this approach.