Enhancing Knowledge Base Question Answering with AMR-Driven Subgraph Retrieval

ACL ARR 2025 February Submission8332 Authors

Abstract: Knowledge Base Question Answering (KBQA) seeks to provide answers to natural language questions by utilizing pertinent triples from knowledge graphs (KGs). The mainstream methods of KBQA involve the use of graph neural networks for the reasoning and rely on subgraph retrieval to reduce the complexity. However, current retrieval methods predominantly align question text with graph relations, leading to inconsistent subgraph quality and limited interpretability, thereby impeding QA performance. Here, we proposed a subgraph retrieval method based on Abstract Meaning Representation (AMR) to captures deep semantic structures, enhance retrieval precision and optimize the reasoning by leveraging the structural similarity of AMR to KGs. Additionally, we construct reasoning chains in AMR form to enhance interpretability. Experiments on the WebQSP and CWQ datasets demonstrated that the integrating of AMR enhances retrieval performance, improves the subgraph quality, and achieves competitive KBQA performance and interpretable reasoning.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: AMR (Abstract Meaning Representation),Subgraph Retrieval,Knowledge Base Question Answering (KBQA),Reasoning Interpretability

Contribution Types: Model analysis & interpretability

Languages Studied: Python

Submission Number: 8332