Research On Interpretable Knowledge Graph Reasoning Technology

In order to effectively organize and express the knowledge in massive Internet data,knowledge graphs emerge as the times require.However,most knowledge graphs are still very sparse and incomplete,which affect their practical utility in downstream tasks.In fact,a lot of missing knowledge can be mined and reasoned out based on the existing knowledge in the knowledge graphs.As an important method to complete knowledge graphs,knowledge graph reasoning technology has received extensive attention from academia and industry.Knowledge graph reasoning technology aims at mining the potential hidden knowledge from the existing knowledge in knowledge graphs by computer reasoning.Generally,the reasoning process which only starts from the existing knowledge in the knowledge graphs is called closed-world knowledge graph reasoning;correspondingly,the reasoning process which needs further external information is called open-world knowledge graph reasoning.In addition,in the scenarios of disease inference,case reasoning,and assistant decision-making,the interpretability of knowledge graph reasoning results has also attracted more attention.Therefore,this thesis explores the interpretable knowledge graph reasoning technology in both closed and open world.For interpretable closed-world knowledge graph reasoning,the existing reasoning method based on reinforcement learning relies heavily on artificial reward engineering,which has high labor cost and limits the applicable scenarios and generalization performance.In addition,these existing methods deal with different reasoning tasks independently and do not make full use of the related information between tasks.In order to solve the above problems,this thesis proposes a collaborative reasoning framework based on generative adversarial imitation learning to enhance the existing reasoning methods based on reinforcement learning.On the one hand,the proposed framework can automatically sample and imitate reasoning demonstrations,and then adaptively adjust its reward function and reasoning policy;on the other hand,the proposed framework can further improve its reasoning performance with the help of the transfer between multiple related reasoning tasks.Experimental results on public benchmarks show that the proposed framework effectively improves the reasoning performance of existing methods based on reinforcement learning,and relieves the dependence on artificial reward engineering.For interpretable open-world knowledge graph reasoning,the existing reasoning methods based on embeddings are still limited in interpretability and multi-step reasoning ability,and the multi-step reasoning mode chain contained in the knowledge graph is not considered.To solve the above problems,this thesis proposes an open-world knowledge graph reasoning method based on reinforcement learning.This method can effectively fuse multi-source open-world knowledge and explore multi-step reasoning patterns containing uncertain open-world knowledge.Experiments on benchmarks show that compared with the existing methods,the proposed method can achieve comparable reasoning performance and make the reasoning results interpretable.