Enhancing knowledge base question answering with AMR-driven subgraph retrieval

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Knowledge Representation

Background:

• Knowledge Base Question Answering (KBQA) uses knowledge graphs (KGs) for factual retrieval.
• Graph Neural Network (GNN) methods often rely on subgraph retrieval, but struggle with semantic gaps between natural language and structured data.
• Current retrieval methods lack precision, leading to inconsistent subgraph quality and poor interpretability, hindering QA performance.

Purpose of the Study:

• To propose a novel framework for KBQA that addresses the limitations of existing subgraph retrieval techniques.
• To improve the quality and interpretability of subgraph retrieval for enhanced QA performance.
• To leverage Abstract Meaning Representation (AMR) for more accurate semantic understanding in KBQA.

Main Methods:

• Parsing natural language questions into Abstract Meaning Representation (AMR) graphs to capture core semantics.
• Extracting canonical predicate-argument structures from AMR graphs to filter and refine noisy subgraphs.
• Utilizing AMR structure to inform GNN adjacency matrices and generate interpretable reasoning chains.

Main Results:

• Significant improvement in subgraph quality and retrieval performance.
• Demonstrated competitive results on the WebQSP and CWQ KBQA benchmarks.
• Enhanced interpretability of the reasoning process within the GNN-based QA system.

Conclusions:

• The proposed AMR-driven framework effectively bridges the semantic gap in KBQA.
• This approach leads to superior subgraph quality, better retrieval, and competitive QA performance.
• AMR provides a valuable mechanism for improving interpretability and precision in GNN-based KBQA.