Synergizing Global Structure and Local Evidence: A Discriminative Question Answering Framework Based on Knowledge Graphs

Abstract

With the recent progress of Large Language Models (LLMs), there has been paradigm-shifting improvement in natural language understanding, significantly outperforming traditional Knowledge Graph Question Answering (KGQA) methods. However, when deployed in specialized vertical domains, domain knowledge deficits of general-purpose LLMs lead to factual hallucinations. Moreover, their black-box generative nature lacks explicit logic chains with poor reasoning interpretability. To address these serious issues, we propose a new Retrieve-and-Discriminate framework that combines the global structure injection technique with local evidence reasoning for better factual accuracy as well as explainability. In particular, to address the lack of domain knowledge, we employ ComplEx embeddings for precise candidate selection and propose a Knowledge Embedding Injection method by soft prompts, fusing continuous graph structure directly into the LLM. We explicitly build multi-hop reasoning path as local evidence to clarify the reasoning logic. Lastly, we adopt a discriminative instruction tuning approach rather than openended generation and essentially restrict hallucination. Our extensive experiments show that our approach reaches 82.2% Hits@1 on the general WebQSP benchmark, and a dominant 90.8% on the industrial Mecha-QA dataset.