ThoughtForest-KGQA: A Multi-Chain Tree Search for Knowledge Graph Reasoning

  • Xingrun Quan
  • , Yongkang Zhou
  • , Junjie Yao*
  • *Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Most multi-hop Knowledge Graph Question Answering (KGQA) methods utilize fixed pruning strategies that, while efficient, critically impair the diversity of answer paths and fail to discover complex or less common correct answers. To address these limitations, this paper introduces ThoughtForest-KGQA, a novel multi-chain tree search algorithm. The method employs a dual-level reinforcement learning framework where a local-level agent optimizes individual reasoning chains by capturing fine-grained semantic details in the knowledge graph. Concurrently, a global-level agent strategically coordinates the simultaneous exploration of multiple chains. Comprehensive evaluations conducted across two distinct KGQA benchmarks reveal that this approach identifies a broader spectrum of correct answers, setting a new state-of-the-art in the field.

Original languageEnglish
Title of host publicationCIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management
PublisherAssociation for Computing Machinery, Inc
Pages5156-5160
Number of pages5
ISBN (Electronic)9798400720406
DOIs
StatePublished - 10 Nov 2025
Event34th ACM International Conference on Information and Knowledge Management, CIKM 2025 - Seoul, Korea, Republic of
Duration: 10 Nov 202514 Nov 2025

Publication series

NameCIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management

Conference

Conference34th ACM International Conference on Information and Knowledge Management, CIKM 2025
Country/TerritoryKorea, Republic of
CitySeoul
Period10/11/2514/11/25

Keywords

  • knowledge graph
  • multi-hop reasoning
  • reinforcement learning.

Fingerprint

Dive into the research topics of 'ThoughtForest-KGQA: A Multi-Chain Tree Search for Knowledge Graph Reasoning'. Together they form a unique fingerprint.

Cite this