Constraint-Awareness and Graph Reasoning for Temporal Question Answering

Abstract

Knowledge Graphs(KGs) have achieved notable improvements in Question Answering(QA) tasks. However, many facts in the real world are subject to time constraints. Although previous methods have made significant progress, they are still limited by the following: Firstly, in neural reasoning, constraints in questions are often only softly matched, overlooking their potential biasing effects on the reasoning process. Secondly, complex problems often cannot be resolved in a single step and require multi-hop reasoning across multiple nodes in knowledge graphs. Thirdly, a gap exists between the vector spaces of KG embeddings and question representations, and a simplistic hard concatenation or fusion of these two can lead to suboptimal solutions. To address these shortcomings, we propose a temporally aware QA method named Constraint-Awareness and Graph Reasoning. This method unfolds in three main stages: (i) understanding the question and its implicit temporal constraints to ensure a comprehensive grasp of the user’s query intent; (ii) multi-hop information propagation modeling designed to capture the deep multi-hop reasoning relationships between entities; and (iii) cross-heterogeneous information fusion, which adaptively unifies the integration of KGs and multi-hop reasoning paths with the vector representation of the question. Experiments demonstrate that our method outperforms a range of baselines.