When Correct Edges Cannot Be Verified: A Provenance Gap in Incomplete KGQA and a Provenance-Favoring Completion Policy
A new research paper published on arXiv explores the challenges in Knowledge Graph Question Answering (KGQA), specifically focusing on incomplete knowledge graphs where missing information needs to be inferred. The study reveals a significant gap between textual verifiability and actual correctness of inferred edges, with a large percentage of correct edges lacking direct textual support. This suggests that current methods often measure provenance rather than accuracy. To address this, the paper introduces TGComplete, a policy that prioritizes provenance by abstaining from admitting edges without textual evidence, leading to higher edge precision and stricter faithfulness at the cost of some recall. AI
IMPACT This research reframes KGQA by prioritizing verifiable provenance over inferred correctness, potentially impacting how AI systems handle uncertain information.