Research On Relation Extraction Method Based On Semantic Dependency Graph

In the era of artificial intelligence,many aspects of daily life are moving towards intelligence,including web search,smart recommendation,and intelligent question and answer.In the process of advancing intelligence,machines need to have a similar knowledge reserve with humans.Therefore,how to build a perfect and reliable knowledge graph becomes a very valuable work.Extracting the relationship between entities from the text is an important part in the process of constructing the knowledge graph,and it is also one of the difficulties.The relationship extraction method of an entity pair can be divided into two types according to whether there is context information.One is relationship extraction.In a sentence text containing a pair of relational entities to be extracted and the relationship to be expressed does not change,the relationship of the entity pairs is extracted according to the information of the sentence.Currently,most methods use distant supervisory learning,and multi-instance learning methods are used to reduce the impact of noise data.The other is relational reasoning.In a text composed of a plurality of declarative sentences and the relationship between the entities changes as the facts are described,the questions are answered based on all factual information,the questions relating to the entities and relationships in the facts.At present,most methods use the feature information of the sentence level,then fuse the sentence information,and finally give the answer together with the problem information.After the basis of the basic technology learning,this paper focuses on the research and proposes the semantic feature extraction method between the entities in the sentence and the relationship reasoning method in the word layer.The main research contents are as follows:(1)For the current method of extracting sentence features,only the structural features of sentences can be extracted,and the semantic features of sentences are neglected,which affects the accuracy of the model.By introducing the shortest path of semantic dependent graphs and extracting semantic features between entities,Semantic dependency graph relation extraction method.In this paper,the shortest path between entities extracted from the semantic dependence graph is used as the input of the neural network model,and a bidirectional circular convolutional attention neural network model(BLCANN)is constructed.The minimum dependent unit feature information of the form<word,dependency,word>is extracted in the model.Thus,the same or similar feature information is extracted in sentences having the same semantics and different expressions.Finally,experiments show that the accuracy of this method is significantly improved compared with the baseline method.(2)In recent years,graph network models have been proposed and widely used.Aiming at the problem of graph network construction oriented to reasoning,this paper proposes a word-level relation graph network model by introducing semantic dependency graphs to represent the association information between words.Compared with the sentence-level relational reasoning model,the model can reflect the relationship between entities in relational reasoning in a fine-grained manner.The relational graph network model based on semantic dependency graph takes words as nodes,and the inter-word dependent information of semantic dependent graphs is used as the side information of nodes and nodes,and the whole fact information is incorporated into the graph network step by sentence.At the same time,in order to reduce the update calculation amount of the basic complete graph under the condition of retaining the key information between nodes,this paper proposes a node update operation based on information aggregation by introducing the Master node into the graph.The method aggregates all the word node information and updates the Master node,and uses the updated Master node information as the basis for updating each word node,thereby reducing the calculation amount when updating the node.Finally,experiments show that the relation graph network based on semantic dependency graph can improve the accuracy compared with the basic method.(3)Chinesize and improved Pydial system,the system gathers a large number of attractions,restaurants and hotels information around the Qiandao Lake scenic area,and can provide users with some simple restaurants and hotels information consultation.In the system,the BLCANN model is used to extract the relationships between entities for the catering and hotel information captured from the network and to store the information in the database.Secondly,in the question-and-answer action speculation stage,the relational graph network model based on the semantic dependency graph is combined with the overall dialogue process to speculate the next answer action.