Ontology-Guided Inference Boosts Knowledge Graph QA Accuracy

Answering complex natural language questions by reasoning over knowledge graphs (KGQA) is challenging, especially for multi-hop questions that require traversing several facts. Existing methods often struggle with an exploding search space filled with irrelevant paths and fail to satisfy the semantic constraints embedded in complex queries. This new research proposes OPI, an Ontology-Guided Evidence Path Inference framework, to overcome these limitations. OPI introduces a relation-centric ontology graph that explicitly captures the type constraints between entities linked by relations. This ontology provides a structured way to understand answer-side constraints. The framework employs a bidirectional retrieval mechanism, combining topic-side expansion with answer-side matching, which effectively prunes noisy paths. Furthermore, OPI incorporates an iterative refinement strategy. This step re-evaluates retrieved paths and candidate answers within the context of the original question, filtering out paths that are type-compatible but semantically irrelevant. Empirical results on standard benchmarks like WebQSP and CWQ show OPI substantially reduces the search space and significantly improves accuracy (Hit@1/F1 scores), demonstrating its effectiveness in building more reliable KGQA systems.