Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Guiding Graph: Explicit Query-Structure Guidance for Knowledge Graph Question Answering

AbstractLarge Language Models (LLMs) can match local semantic clues in knowledge graphs (KGs), but complex KGQA often requires satisfying the full dependency structure of a question. Existing retrieval-based methods therefore risk producing reasoning paths that look plausible at each step but violate the query’s global intent. We propose Guiding Graph, a schema-light query-structure interface that turns KGQA from open-ended semantic traversal into structure-constrained query completion. A Guiding Graph exposes the entities, placeholders, and global dependencies that KG exploration must preserve, serving as an operational interface between natural language and graph traversal. The interface is training-free by default through prompt-based construction, and can optionally be strengthened with a cold-start fine-tuned constructor trained from limited bootstrapped supervision. Given the constructed Guiding Graph, exploration is governed by target selection from the query structure, structure-guided pruning before semantic matching, and target-scoped semantic alignment under the current Guiding Graph triple. Experiments on WebQSP, CWQ, and a domain-specific agricultural KG show that explicit query-structure guidance improves accuracy, robustness, and efficiency, with especially clear benefits for multi-hop reasoning and smaller LLMs.

More from EMNLP 2026

All EMNLP 2026 papers →