Abstract
Current subgraph retrieval methods generally fall into two categories: those that rely on semantic matching, which use only surface-level semantic information of relations and lack flexibility; and those based on personalized PageRank algorithms, which fail to leverage the semantic connection between the relation and the question, rendering them susceptible to noisy data. To address these issues, this paper introduces a novel retrieval model that employs hybrid semantics of relations and path representations. Specifically, hybrid semantics involves merging relational and entity information within a knowledge graph to extract the deep semantics of relations and enhance semantic representation by integrating it with the explicit descriptive text of the relations. Path representation merges the semantics of the current relation with those of preceding ones to form a complete path representation. This representation is then semantically matched with the question to compute a score, which determines whether the relation should form part of the subgraph. We integrated our subgraph retrieval model with the Neural Symbolic Machine (NSM) reasoning model and evaluated it on the publicly available CWQ and WebQSP datasets. The experimental results demonstrate that our method performs exceptionally well on these datasets, validating the efficacy of utilizing deep semantics and path representations for the retrieval of subgraphs in response to complex questions.