Research And Application Of Entity Alignment Technology Based On Association Information Modeling

With the development of artificial intelligence technology,numerous socially valuable applications such as human-machine dialogue,hobby recommendation,and character portrait are becoming truly viable.Knowledge graph,as the foundation of these applications,plays an important role in highly intelligent social services.Entity alignment,as a crucial component of knowledge graph construction,greatly affects the performance of the final application.However,few works have paid attention to the problem of coarse perception granularity,which can lead to the omission of important neighborhood information.In addition,most works that extract structural and attribute information from the neighborhood for entity semantic representation and ignores the shared neighbor information(aligned entity pairs)in the neighborhood.To address these issues,this thesis conducts research on entity alignment technology based on modeling of association information.The main contributions of this paper are as follows:(1)To address the problem of coarse perception granularity,this thesis proposes a fine-grained perception module,which utilizes the attention mechanism and the gating mechanism to control the flow of information to differentiate the importance of each neighbor entity in the neighborhood at a fine granularity.The module captures the similarity to achieve attribute reflection and determine the entity’s own attribute features.Experimental results show that the improved model’s the direct hit rate Hits@1 is improved by 0.38% to 0.64% compared to Neighborhood Matching Network(NMN)in three datasets of DBP15 K.(2)To address the problem of inaccurate expression of entity pair distances in the semantic space due to the shared neighbor information being ignored,this thesis proposes a shared neighbor information interaction module,which models the shared neighbor information in the neighborhood and interacts with the distance between entity pairs in low-dimensional vector space to correct the entity pair’s distance.Experimental results show that on the ZH-EN(ZH→EN,using entities in Chinese data as source entities and entities in English data as candidate entities)dataset,the improved model’s the direct hit rate Hits@1 and the top ten hit rate Hits@10 are improved by 2.06% and2.82%,respectively,compared to NMN.On the JA-EN(JA→EN)dataset,Hits@1 and Hits@10 are improved by 2.19% and 2.94%,respectively,compared to NMN.On the FR-EN(FR → EN)dataset,the two indicators also improved by 1.12% and 1.45%,respectively.(3)The question-answering system based on an extensible knowledge graph is realized.The system includes functions such as graph fusion,intelligent question-answering,information retrieval,and data statistics.To improve system usability,the system is optimized using thread pooling technology,which greatly increases system performance.