Proceedings of the Annual Meeting of the Cognitive Science Society

Knowledge Graph Reasoning (KGR) is an effective way to ameliorate sparsity and incompleteness problems by inferring new knowledge based on existing knowledge. The probabilistic case-based reasoning (CBR) model can gather reasoning paths from similar entities and relations in KG, thus outperforming rule-based and embedding-based KGR methods. However, it is still limited by some problems, such as insufficient learning of similarity features and sparse intermediate activations. This paper proposes a \textbf{S}imilarity-\textbf{A}ware graph transformer-enhanced probabilistic CBR model for \textbf{KGR}, namely SA-KGR. The proposed model regards the reasoning task as the KG query answering and is composed of two phases. The first phase is similarity-aware graph transformer-based graph feature encoding, which equips the similarity matrix and Mixture-of-Expert network to obtain fine-grained similarity features that are more helpful for reasoning path generation. The second phase is similarity-enhanced probabilistic case-based reasoning, which can retrieve and infer query answers from the generated candidate paths to complete brain-like cognitive reasoning. Extensive experimental results on various benchmarks unambiguously demonstrate that the proposed SA-KGR model can obtain the state-of-the-art results of current CBR-based methods.