From Decomposition to Validation: A Multi-stage Framework for Temporal Knowledge Graph Question Answering with LLMs

ACL ARR 2026 January Submission3812 Authors

Keywords: Temporal Knowledge Graph Question Ansewring

Abstract: Temporal Knowledge Graph Question Answering (TKGQA) requires reasoning over time-dependent facts. However, existing approaches suffer from two fundamental limitations: (1) general-purpose text retrievers often fail to align the entity, relation, and temporal mentions in the question with TKG facts and (2) large language models (LLMs) remain unreliable in temporal reasoning, frequently exhibiting the hallucination problem. To address these challenges, we propose TempReasoner, a multi-stage framework for TKGQA. TempReasoner decomposes complex temporal questions into structured sub-questions, retrieves temporally relevant facts using a trained TKG-Retriever, filters noisy evidence through LLM-generated alignment signals, and generates answers along with supporting proofs. A dedicated validation module further refines uncertain predictions by issuing verification queries, improving both answer accuracy and reliability. Experiments on two widely used benchmarks, MultiTQ and CronQuestions, show that TempReasoner consistently delivers strong performance. In particular, TempReasoner achieves substantial improvements on multi-hop temporal questions, demonstrating its effectiveness in temporal reasoning and evidence grounding. Additional analysis further confirms the efficiency of its retrieval component and consistent model-agnostic generalization.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: Temporal knowledge Graph Question Ansewring

Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: English

Submission Number: 3812