Research On Knowledge Graph Completion Based On Reinforcement Learning

The current commonly used knowledge graphs all suffer from incompleteness due to a large number of missing facts,which seriously affects their ability to be applied in downstream tasks.The research on knowledge graph complementation can help improve the data quality and completeness of knowledge graphs,among them,reinforcement learning-based complementation methods have been widely studied for their inference accuracy and interpretability.However,the current reinforcement learning-based knowledge graph complementation methods have the following problems when applied to sparse knowledge graphs: insufficient amount of data in the inference process and low reliability of the inference path.To address these problems,this paper improves the framework of reinforcement learning to apply to the task of complementation on sparse knowledge graphs,with the following main work:(1)Aiming at the problems of insufficient amount of inference data and low reliability of inference paths in sparse knowledge graphs,a knowledge graph complementation method incorporating semantic information is proposed.The method first embeds the relations and entities in the knowledge graph into the vector space as an external environment for reinforcement learning training.Then in the path search phase of the agent,based on the semantic information of query relations and inference paths,the(relation,entity)pairs with high semantic similarity to both are selected to extend the corresponding action space.Finally,for different question and answer queries,the reliability of inference is evaluated based on the semantic similarity between inference paths and query relations.Experimental results on four sparse datasets,NBLL23 K,WD-singer,FB15K-237-10%,and FB15K-237-20%,show that the method proposed in this paper effectively exploits the semantic information of query relations and inference paths and outperforms related work on the inference task of query Q&A.On the FB15K-237-10% dataset,the SIKGR model outperformed the Dac KGR model in terms of MRR,Hits@3 and Hits@10 metrics by 0.9%,0.8% and 1%,respectively,and the SBS model outperformed the MRR and Hits@3metrics by 1.2% and 0.3%,respectively.Using ablation tests,the two optimization techniques’ efficacy on action space and reward functions was also confirmed.(2)Aiming at the problem of low reliability of inference paths in sparse knowledge graphs,a knowledge graph complementation method incorporating entity description information is proposed.The method first learns an entity description information encoder to encode the entity description information in the knowledge graph.It then fuses the triple structure information to jointly construct a vector representation of the entities,which in turn shapes the external environment for reinforcement learning.Finally,inference reliability is assessed based on the semantic similarity of the optimized inference paths to the query relations.The experimental results on two datasets,FB15K-237-10% and FB15K-237-20%,show that the inference performance of the proposed method outperforms related work.In comparison to the Dac KGR model,the EDIKGR model performed 0.6% better on the hit rate metrics Hits@3 and Hits@10 and 0.7% better on the ranking metric MRR on the FB15K-237-10% dataset.This demonstrates the entity description information encoder’s efficacy.