Optimizing multi-hop question answering reasoning through time-aware path construction

Abstract Multi-hop question answering (MHQA) requires models to perform reasoning across multiple text segments to answer complex questions. However, existing methods often overlook temporal dependencies between facts during reasoning path construction, resulting in illogical inference chains and reduced reasoning accuracy. This study enhances the modeling of time-sensitive information in MHQA to optimize the construction of reasoning paths, thereby improving the inference accuracy and robustness of QA systems in real-world scientific scenarios. The objective is to achieve a dynamic balance between reasoning precision and computational efficiency. A Time-Aware Graph Attention Mechanism is first introduced to dynamically construct entity-relation graphs within documents, incorporating timestamp features to capture event sequences and temporal dependencies. Then, a Dual-Level Temporal Path Selector is proposed to identify optimal reasoning paths at both the sentence and entity levels, ensuring temporal consistency and semantic completeness throughout the reasoning process. Experiments are conducted on two benchmark MHQA datasets, HotpotQA and MuSiQue-full, evaluating the model’s accuracy, path coherence, and temporal consistency. Results show that the proposed model achieves a Joint Exact Match score of 74.3% on HotpotQA, outperforming the current best-performing model Pathformer (71.5%) by 2.8%. On MuSiQue-full, the model attains an Answer F1 score of 87.1%, exceeding DecompT5’s 85.2%, and reaches a temporal consistency score of 91.3. These findings confirm the effectiveness of the proposed time-aware path construction strategy in enhancing MHQA performance. The proposed time-enhanced reasoning approach significantly improves logical coherence, temporal reasoning, and adaptive efficiency across varying reasoning depths, demonstrating an effective balance between performance gains and manageable computational costs. This study provides valuable insights for developing temporally-aware QA systems applicable to real-world scenarios such as historical analysis and legal reasoning.