Scholarlink

<p>To address the issues of insufficient understanding and reasoning bias in existing methods for Temporal Knowledge Graph Question Answering (TKGQA) tasks when handling complex temporal logic and implicit temporal constraints, we propose Temporal Relation-Aware Alignment (TempRAA), a novel TKGQA approach, designed to enhance Large Language Models (LLMs) reasoning in time-sensitive question answering tasks. The framework integrates a Self-Feedback Question Rewriting (SFQR) module, a Temporal Relation-Aware Retriever (TRAR), and QA Reasoning, aiming to effectively incorporate temporal knowledge from temporal knowledge graphs (TKGs) into LLMs. Firstly, leveraging the self-evaluation and correction capabilities of LLMs, the SFQR rewrites implicit temporal information in questions into explicit timestamps, thereby enhancing the precision of temporal semantic alignment between questions and the TKGs and effectively mitigating the temporal hallucination issues that LLMs may generate when processing ambiguous temporal expressions. Secondly, the TRAR module introduces a Temporal Relation Attention (TRA) mechanism and an adaptive time filtering function to jointly model attention weights among questions, facts, and temporal relations, thereby achieving dynamic weighted ranking of candidate facts. Finally, during the QA Reasoning phase, the model utilizes the rewritten questions and retrieved temporal facts as contextual inputs to generate final answers through fine-tuned LLMs. Experimental results on two benchmark datasets, MultiTQ and TimeQuestions, demonstrate that the proposed method outperforms baseline methods on different types of temporal questions, highlighting its effectiveness in enhancing the temporal reasoning ability of LLMs. Ablation studies further verify the contribution of each core module to the overall performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

TempRAA: temporal relation-aware alignment for enhancing LLMs reasoning in time-sensitive knowledge graph question answering

  • 期刊论文

  • 发表日期：

    2026年1月19日

查看全文
  • Lianhong Ding,
  • Na Ding,
  • Peng Shi

摘要

To address the issues of insufficient understanding and reasoning bias in existing methods for Temporal Knowledge Graph Question Answering (TKGQA) tasks when handling complex temporal logic and implicit temporal constraints, we propose Temporal Relation-Aware Alignment (TempRAA), a novel TKGQA approach, designed to enhance Large Language Models (LLMs) reasoning in time-sensitive question answering tasks. The framework integrates a Self-Feedback Question Rewriting (SFQR) module, a Temporal Relation-Aware Retriever (TRAR), and QA Reasoning, aiming to effectively incorporate temporal knowledge from temporal knowledge graphs (TKGs) into LLMs. Firstly, leveraging the self-evaluation and correction capabilities of LLMs, the SFQR rewrites implicit temporal information in questions into explicit timestamps, thereby enhancing the precision of temporal semantic alignment between questions and the TKGs and effectively mitigating the temporal hallucination issues that LLMs may generate when processing ambiguous temporal expressions. Secondly, the TRAR module introduces a Temporal Relation Attention (TRA) mechanism and an adaptive time filtering function to jointly model attention weights among questions, facts, and temporal relations, thereby achieving dynamic weighted ranking of candidate facts. Finally, during the QA Reasoning phase, the model utilizes the rewritten questions and retrieved temporal facts as contextual inputs to generate final answers through fine-tuned LLMs. Experimental results on two benchmark datasets, MultiTQ and TimeQuestions, demonstrate that the proposed method outperforms baseline methods on different types of temporal questions, highlighting its effectiveness in enhancing the temporal reasoning ability of LLMs. Ablation studies further verify the contribution of each core module to the overall performance.