Research And Application Of Knowledge Graph Embedding Based On Affine Transformation

Knowledge graph embeddings have been extensively studied as an important bridge connecting knowledge graphs and other application fields such as question answering systems,recommender systems,and information retrieval systems.In the knowledge graph,the relationship connecting two entities has the characteristics of symmetry,composition,inversion,etc.These characteristics are called relation patterns.Knowledge graphs contain various relation patterns that reveal universal rules of knowledge graphs.Mining new relation patterns and designing models that can build more relation patterns is one of the main research directions of knowledge graph embedding.The success of many knowledge graph embedding models stems from the modeling of various relational patterns.However,some relational patterns are still difficult to model simultaneously,such as non-commutative composition and multiple relation patterns.This paper studies distancebased embedding models from the perspective of relation pattern,and the main contributions include:1.This paper analyzes the existing distance-based knowledge graph embedding models.A distance-based knowledge graph embedding model AffE constructed using two-dimensional affine transformation is proposed.It is proved by theory and experiment that AffE can model six relation patterns,including inversion pattern,symmetric pattern,asymmetric pattern,non-commutative composition pattern,commutative composition pattern,and multiple relation pattern;this paper uses AffE to perform link prediction tasks on the public datasets WN18RR,FB15k-237,and YAGO3-10 and compares it with recent research work.2.This paper proposes a path regularization method by using the path information in the knowledge graph.Path regularization uses the relationship path between head and tail entities to constrain the relationship between them,helping the model to learn better embedding vectors;Path regularization uses a data-driven global filtering mechanism and an attention-based local filtering mechanism to filter out invalid relationship paths to avoid noise from invalid paths;The effect of path regularization is verified on the link prediction task.3.In this paper,the AffE model is applied to the knowledge graph question answering field,and the extended model AffE-KGQA is obtained.AffE-KGQA utilizes AffE's predictive ability for missing relationships to obtain multi-hop inference results from knowledge graphs;AffE-KGQA is used in the dataset MetaQA to perform multi-hop question answering and analyze its performance.The experimental results show that compared with the models known in this paper,AffE achieves the best results on all indicators on WN18RR and achieves the best results on almost all indicators on FB15k-237;Path regularization can improve the performance of the knowledge embedding model through effective path constraints.AffE further improves the results after using path regularization;In the domain of knowledge graph question answering,the AffE-KGQA model using AffE achieves better results than the EmbedKGQA model using ComplEx in both settings on the MetaQA dataset;The knowledge graph question answering model AffE-KGQA achieves results close to state-of-the-art models on the full knowledge graph of the MetaQA dataset and state-of-the-art results on the incomplete knowledge graph of the MetaQA dataset.