Research On Key Techniques Of Knowledge Acquisition For Knowledge Graph Population

Knowledge graph is a kind of knowledge base that uses a graph-structured data model to integrate data.It converts unstructured data in the real world into triple structured knowledge.The interpretable structured semantic knowledge is an important factor in the construction of intelligent machines,and it is also a basic data resource for the development of artificial intelligence applications.Although many existing knowledge graphs have tens of millions or even billions of triples,they still have a large degree of incompleteness,which seriously affects the performance of related applications.As an important means to improve the completeness of knowledge graph,knowledge acquisition has received extensive attention from researchers and has developed into one of the key research directions in the field of knowledge graph.Knowledge acquisition aims to obtain triples from data sources.There are two main types of current data sources:one type is natural language text containing a large amount of information;the other type is the knowledge graph containing a large amount of hidden information.According to these two types of data sources,knowledge acquisition usually includes the following three key technologies:entity link with knowledge graph,relation extraction for knowledge graph and knowledge reasoning on knowledge graph.Among them,the first technique aims to find entities from the text.The second technique aims to identify the relations between entities in the text and form triples.The third technique aims to dig out hidden triples based on the known triples in the knowledge graph.Although researchers have achieved certain results in knowledge acquisition research,the existing research results still have many shortcomings in the use of the characteristics of the knowledge graph and the design of the model.Therefore,in view of the shortcomings of existing research results,this thesis has deeply studied the three key technologies of knowledge acquisition and achieved the following research results:(1)To entity linking with knowledge graph,an entity linking method that combines co-attention mechanism with graph convolutional network is proposed.The existing neural network-based entity linking methods ignore the semantic gap between the sequential entity mention's context,which affects the accuracy of entity linking.It is observed that the importance of each word in the entity mention's context is different,and the content in the entity's context is also different.In order to solve the above problems,this thesis proposes an entity linking method that combines co-attention mechanism with graph convolutional network.This method uses the co-attention mechanism to establish the correlation between the entity mention's context and the entity's context,in order to narrow the semantic gap between the two,and model the important content of the two simultaneously.On this basis,a context-aware graph convolutional network is proposed to learn the significant graph topological feature of the entity.The experimental results on five public datasets show that the precision,recall and F1 value of our method are higher than the comparison method.(2)To relation extraction for knowledge graph,a distant supervised relation extraction method based on encoder-decoder framework is proposed.Existing distant supervised relation extraction methods often ignore the fact that one entity pair may have multiple relations,which affects the accuracy of relation extraction.It is observed that multiple relations between one entity pair often have dependencies.In order to solve the above problems,this thesis proposes a distant supervised relation extraction method based on encoder-decoder framework.This method uses convolutional neural network to extract the sentence bag feature corresponding to the given entity pair at the encoder side.At the decoder side,the long short-term memory network is used to predict relations of the given entity pair and model the dependencies between them in the form of conditional probability.In addition,the attention mechanism is introduced into the framework to highlight sentence features corresponding to unpredicted relations.For the problem that the encoder-decoder framework cannot be trained when it is directly used for relation extraction,a measurement called Information Quantity is proposed.This measurement quantifies the information contained in each training sentence bag for each of its relations,and determine the order of relations in descending order according to the amounts of information.The experimental results on the widely used public real dataset show that our method is significantly better than comparison method in terms of precision,recall and precision@N.(3)To knowledge reasoning on knowledge graph,a method of path-augmented knowledge reasoning based on convolutional neural network is proposed.Existing knowledge reasoning methods based on convolutional neural network ignore the path information between entities,which seriously affects the accuracy of knowledge reasoning.It is observed that from a local and global perspective respectively,the importance of each path to the relation is different.In order to solve the above problems,this thesis proposes a path-augmented knowledge reasoning method based on convolutional neural network.This method introduces path information in the process of knowledge reasoning,and considers the local and global importance of the path.In particular,in order to calculate the local importance of the path,a method based on attention mechanism is proposed.In order to calculate the global importance of the path,a degree-guided inverse path frequency index is designed.The experimental results on four public datasets show that our method performs better than comparison method on two common evaluation tasks,namely link prediction and triple classification.