Research And Application Of Reasoning Technology Based On Knowledge Graph

With the advent of the era of big data,the application of knowledge graphs in various fields has become more and more critical.However,it has a high dependence on source data,and there is often a problem of lack of information.The incompleteness of knowledge graphs greatly limits their use in industrial applications.How to reason through knowledge graphs,complete missing entities or relationships,and explore available information has become the focus of recent research.The researchers found that integrating reinforcement learning into knowledge graph reasoning can achieve excellent results.However,the conventional knowledge graph reinforcement learning reasoning algorithm has problems such as insufficient representation ability,no memory components,and no well-defined path quality,which makes the algorithm less efficient.To solve the above problems,the thesis makes the following contributions:1.Based on the problems of insufficient representation ability,invalid redundant action selection,and no memory components in conventional knowledge graph reinforcement learning inference algorithms,Chapter 3 of the thesis selects adaptive representation learning methods based on the original fact prediction scores of representation learning methods on the dataset.A more robust representation learning method is used to represent the reinforcement learning environment to enhance the representation ability of the algorithm;an action sampler is designed to reduce the ineffective and redundant action selection of the agent in the process of walking;Long Short-Term Memory is used as a memory component to encode historical information to increase model accuracy.In the link prediction task,compared with Deep Path,the result indicators Hits@1,Hits@3,MRR and MAP of this algorithm(RLKGR-ASM)increase by 1.3%,2.5%,2.1% and 1.9% respectively on the NELL-995 dataset,and 4.6%,6.3%,4.9% and 4.1% respectively on the FB15K-237 dataset;in the fact prediction task,the result index MAP of this algorithm is compared with Deep Path,which is higher in NELL-995 dataset with a 3.4% increase and 0.8% on the FB15K-237 dataset.2.In view of the problem that the quality of the reasoning path is not clearly defined in the conventional algorithm and in Chapter 3 of the thesis,that is,after each walk,the agent only obtains rewards through the artificially set reward function,which cannot reflect the findings of the current round of walks how good or bad the path is.A reinforcement learning knowledge graph reasoning method based on bilateral path quality assessment(RLKGR-BPQA)is proposed in fourth chapter of thesis.Set up a bilateral path quality evaluation module,and by crawling Wikipedia as external auxiliary information,calculate the semantic similarity between the head and tail entity description information keyword sets and path entities respectively,and then replace the original reward module after fitting and feed back to the agent to motivate its choice high quality path.The algorithm relies too much on auxiliary information,has a strong specialization,and performs well on the FB15K-237 data set with many long paths.In the link prediction task,compared with Deep Path,the result indicators of this algorithm(RLKGR-BPQA),Hits@1,Hits@3,MRR,and MAP,increase by 0.4%,1.3%,0.5%,and 0.8% respectively on the NELL-995 dataset,increased by 7.4%,7.5%,5.7%,and6.5% on the FB15K-237 dataset respectively;in the fact prediction task,the result indicators MAP of this algorithm increased by 1.1% on the FB15K-237 dataset compared with Deep Path.3.A knowledge graph completion and reasoning system based on the algorithm in Chapter 3 of the thesis is implemented,providing functions such as user registration and login,knowledge graph completion and reasoning,and using the NELL-N95 dataset for testing.