| Knowledge graph(KG)is the most advanced and popular structured knowledge storage method nowadays.Knowledge graph question answering(KGQA)is the task that gives the answer entities in the KG according to the natural language question,thereby greatly expanding the scope of human information acquisition and improving the efficiency of information obtainment.After years of academic research and practice,the KGQA task has achieved ideal results on simple single-hop problems,but there are still many problems unsolved for complex multi-hop reasoning problems.Inference path labeling is often lacking in KGQA task data sets,and the existing weakly supervised learning strategies cannot fully eliminate the misdirection of wrong reference inference paths to model inference learning.The key of the mainstream information retrieval modeling methods is to control the scale of the candidate inference path set through effective relation filtering.However,iterative relation reasoning is usually carried out according to the overall semantic information of the question,ignoring the locality of the relation in the question and the value of the fine-grained information used for relation filtering.Focusing on the two goals of "reducing the set size of candidate paths" and "eliminating wrong reference inference paths",this paper proposes fine-grained relation inference KGQA model and weakly supervised learning strategy for fine-grained scoring models.Firstly,it improves the ability of relation filtering and path filtering from the perspective of model structure.On this basis,the ability of the model to distinguish wrong reference inference paths is improved from the perspective of supervision strategy.In this paper,for the first time,the fine-grained information of relation and question is used in the process of relational inference to fully model human reasoning process.Three fine-grained scores are proposed for relational filtering,effectively reducing the scale of candidate reasoning paths.Using the relation filtering and path filtering capabilities of the fine-grained scoring model,the sampling and training strategy of weakly supervised learning is refined from three different search spaces to eliminate the wrong reference inference paths and further reduce the size of the candidate paths.In this paper,CWQ 1.1 public data set in KGQA task is selected for full experimental verification.Compared with existing KGQA models,the effect of the fine-grained relation inference KGQA model proposed in this paper is significantly improved,and the model’s effect is further significantly improved after supervised training with weakly supervised learning strategies we proposed.Moreover,detailed ablation experiments are designed and implemented,which fully verify the effectiveness of the model structure and the supervised strategies proposed in this paper. |
