A Research Of Multi-source Character Attributes Data Fusion

In recent years,Knowledge Graph has been widely used in intelligent questionanswering,character relationship analysis,search engine and other scenarios,covering the fields of finance,Internet,medical treatment,government affairs,entertainment and so on.Structured data such as Wikipedia,Baidu Baike,and other industry databases are the preferred data sources for constructing character knowledge graph.However,compared with structured data,the Internet stores more unstructured data sources,which contains a huge amount of character attribute information.How to accurately extract character information from unstructured data sources for constructing character knowledge graph is a hot research issue in the field of character knowledge graph construction.Therefore,based on an in-depth analysis of existing related research,this thesis proposes a character attributes extraction method of graph convolutional networks based on dependency tree pruning,entity alignment method of Knowledge Graph based on embedded expression,and attribute alignment method based on attribute type,which implements the process of accurately extracting character attributes from unstructured data sources and constructing a knowledge graph.Specific research is as follows:1)Because existing dependency tree-based relationship extraction models have the problem that the critical information is trimmed due to excessive trimming of the dependency tree,this thesis proposes a graph convolution network relationship extraction model based on dependency tree pruning.First,the model prunes the dependency tree with the entity and the shortest dependency path as the center to obtain the adjacency matrix of the pruned dependency tree.Subsequently,the word vectors and adjacency matrices of the sentences are used as inputs to the graph convolutional network.The goal is to get the implicit expression of each word.Then use the attention mechanism based on entity type embedding and position to get the sentence expression,and finally classify the relationship.Experiments show that the model has achieved good results on the SemEval dataset and the person attribute dataset.From the perspective of pruning subtree extension and negation,this thesis shows that the dependency tree pruning method can effectively improve the relationship extraction effect.At the same time,from the perspective of statistical results of entity types,this thesis explains that the attention mechanism based on entity type embedding can effectively distinguish entity types and replace named entity recognition.2)Attribute triples are an important part of the knowledge graph,but most of the current entity alignment methods don't use this information.In response to this issue,this thesis proposes a knowledge graph entity alignment model based on embedded expression.This model uses the combination of character embedding and attribute type embedding to introduce attribute value information.Simultaneously,in the structure embedding,the graph attention network introduces the structural information of the knowledge graph to obtain enhanced entity embedding.Experiments show that the model works well on the cross-language entity alignment dataset and the same language entity alignment dataset.3)For attribute alignment,this thesis proposes an attribute alignment method based on the attribute type.This method obtains attribute similarity from two aspects: attribute name and attribute value.For the attribute name similarity calculation,semantic information is introduced by adding the calculated word vector similarity.The attribute value similarity is calculated by corresponding methods based on three different attribute categories.Experiments show that the model works well in attribute alignment,accuracy score reached 98%.It is found that the semantic and attribute value information can reduce the number of misaligned attribute pairs and improve the accuracy of attribute alignment.