Leveraging Decoupled Composition in Large Language Models for Knowledge Base Question Answering

Traditional knowledge base question answering (KBQA) models rely on large amounts of annotated data and complex training architectures, making it difficult for them to be adapted to new knowledge bases. Recently, large language models (LLMs) have demonstrated remarkable capabilities in code generation and generalization. However, they face challenges in directly generating knowledge base query statements due to the complexity of inference structures and the diversity of knowledge semantics. To address these issues, this paper proposes a knowledge base question answering framework, DC-KBQA, that integrates large language models. The framework employs a decoupling and combination strategy, breaking down the task of generating knowledge base query statements into the retrieval and combination of reasoning structures and semantic components. Extensive experiments on public datasets show that this framework improves the performance of knowledge base query statements and enhances the model's generalization ability.