Knowledge-injected Stepwise Reasoning on Complex KBQA

While open-domain question answering provides extensive coverage, knowledge base question answering (KBQA) is distinguished by its reliability and accuracy, leveraging structured and verified knowledge sources. However, complex KBQA grapples with the intricacies of question structures and semantics, necessitating advanced natural language understanding and generalization capabilities. The vast search space, coupled with limited supervisory signals during the reasoning process, further intensifies these challenges, making consistent reasoning across diverse question answering scenarios a formidable task. To address these issues, we introduce the Knowledge-injected Stepwise Reasoning framework. Utilizing the KoPL introduced by the KQA Pro dataset, KiSR decomposes reasoning into sequential steps, each focusing on specific parts of the KoPL program. This approach not only reduces the demands on the model’s reasoning capabilities but also addresses the issue of sparse intermediate signals. Crucially, the reasoning process dynamically integrates relevant external knowledge at every step, refining the model’s understanding and predictions. Empirical evaluations underscore KiSR’s competitive performance across a range of complex KBQA tasks, outperforming state-of-theart baselines and demonstrating its effectiveness in harnessing structured knowledge for precise and reliable question answering.