Operation aware learning improves knowledge graph question answering
OpCredit: Operation-Aware Policy Learning for Agentic Knowledge Graph Question Answering
Computation and Language
Summary
Answering questions by navigating knowledge graphs involves choosing steps or operations that help find the right answer. The problem is that successful answers can include useless steps, and failed attempts can have useful ones, making it hard to judge each step's value. The authors created a method called OpCredit that looks at both the final outcome and the immediate effect of each step to better learn which steps help. They tested this approach on multiple datasets and found it consistently improved answering questions using knowledge graphs.
What this means in practice
- •For data engineers: Improve the accuracy of systems that answer complex questions by better learning which graph operations lead to correct answers.
- •For ai software developers: Build smarter question answering agents that more effectively evaluate individual steps in a sequence to navigate knowledge graphs.
Authors
Yucheng Li, Mingyuan Li, Xiaxia Wang, Gong Cheng
Abstract
Agentic knowledge graph question answering(KGQA) relies on sequences of executable operations, but trajectory-level rewards obscure their individual contributions. Successful trajectories may contain ineffective operations, while failed trajectories may contain useful ones. To assess individual operations, we propose OpCredit, an operationaware policy-learning framework that combines downstream outcome credit with immediate transition credit: Branch Outcome Credit (BOC) reuses ordinary rollouts to compare terminal task rewards observed after different operations under matched navigation contexts; Frontier Transition Credit (FTC) measures operation-induced changes in answer-conditioned frontier potential. The two complementary signals jointly supervise the corresponding decision spans during training, while inference uses only the learned policy and standard KG interface. Experiments on ComplexWebQuestions, WebQSP, and GrailQA across three backbones demonstrate consistent improvements in KGQA performance.