Enhancing relation classification based on GCN Multi-hop Knowledge Graph Question Answering model of knowledge reasoning

Existing multi-hop knowledge graph question answering approaches primarily infer answers by predicting sequential relation paths or aggregating latent graph features, achieving some success, but the former is difficult to optimize and the latter lacks interpretability. Therefore, this paper proposes a multi-hop question answering model named ReGCN, which stands for GCN-augmented relation classification for knowledge reasoning. ReGCN selects appropriate relations for entity hopping in multiple steps. The ReGCN model consists of three parts: an encoding module that uses BERT to process questions and core terms, along with a dictionary-based method to encode the knowledge graph; a GCN node update module that refines the representations of candidate answer subgraphs through graph convolutions; and a joint scoring module that combines attention mechanisms and graph structural information to compute tail entity scores, selecting the highest-scoring one as the answer. Additionally, we have constructed a cardiovascular disease knowledge graph (CvdKG) dataset to verify the feasibility of ReGCN for multi-hop reasoning on medical knowledge graphs. The ReGCN model achieved a 2-hop hits@1 accuracy of 72.2% and 100% on the public datasets WebQSP and MetaQA, respectively, showing a 0.8% improvement over the baseline model on WebQSP. On our dataset CvdKGQA, the 2-hop hits@1 accuracy rate was 99.6%, which is 0.7% higher than the baseline model. Notably, for 3-hop questions in our dataset, there was a 1.9% improvement. Analyzing the experimental results, ReGCN is shown to be effective in intelligent question answering over knowledge graphs for complex multi-hop problems with densely related candidate answer subgraphs.