Research And Application Of Knowledge Fusion In The Construction Of Knowledge Graph

Knowledge graph is a semantic network that describes the relationship between entities in a formal way.During the construction of the knowledge graph,there are some problems such as information islands and insufficient knowledge coverage in the knowledge graph formed by a single data source.Knowledge fusion technology can be used to integrate multi-source data to improve the richness and accuracy of the knowledge graph,therefore,it is significantly valuable.There are two sub-tasks in the knowledge fusion: entity alignment and entity linking,both of which are used to solve the disambiguation of entities,but the ambiguous entities are different.Entity alignment aims at structured entities in heterogeneous knowledge bases;Entity Linking aims at the mentioned entities in the text.At present,entity alignment task faces the problems like information asymmetry between entities and the undifferentiated importance of attributes to entities,etc.Entity linking task faces the problems like error propagation between modules and separate encoding of mentions and candidates,and so on.In order to solve the above problems,some works are carried out in this thesis as follows:1.Aiming at the problems of entity alignment,a model with attribute attention based on Siamese network is proposed.The entity pairs in the task are all from the knowledge base,and their structure is relatively similar.Therefore,the Siamese network is developed for sharing the weighted parameters during the encoding of entity pairs;In the Siamese sub-network,two encodings are respectively used to learn the holistic of entity and the importance of attribute to entity.Experiments show that the performance on the own Chinese dataset achieves 96.52% in term of F1,which verifies the effectiveness of the proposed method.2.Aiming at the problems of entity linking,an end-to-end model based on joint encoding is proposed.In most entity liking tasks,entity detection and entity disambiguation are divided into modules,leading to error propagation.Meanwhile,mentions and candidates are usually encoded separately,and then calculating the similarity of the two vectors,resulting in the inability to learn the association between entities during the encoding process.Therefore,an end-to-end method is adopted to reduce error propagation.In addition,a joint encoding approach is developed for learning the close semantic correlation between mentions and candidates from the bottom encoding layer.Experiments show that the performance on the public Chinese dataset achieves 75.36% in term of F1,which is higher than other comparative models.3.The proposed model in this thesis is developed for knowledge fusion module in the construction of knowledge graph in music domain,enriching the knowledge base,and verifying the feasibility of the model by displaying the knowledge graph in music domain.