Reasoning stability improvement for knowledge graph question answering based on gated attention and contrastive graph neural networks

Graph Neural Networks (GNNs) have recently shown strong potential in modeling structured data and performing multi-hop reasoning, making them effective for knowledge graph question answering (KGQA) tasks. By recursively aggregating semantic information over graph structures, GNNs capture higher-order dependencies among entities and support complex reasoning processes. However, existing GNN-based approaches still face challenges when handling complex queries, particularly due to unstable attention distributions and noisy relational propagation, which can negatively affect reasoning reliability.To address these challenges, we propose an enhanced GNN-based reasoning framework. The framework integrates a gating mechanism and exponential moving average (EMA)–based attention smoothing into a multi-head attention architecture. The gating mechanism is designed to adaptively balance query semantics and structural signals from neighboring nodes, thereby mitigating the influence of redundant relations. During multi-hop message passing, EMA smoothing is applied to stabilize attention dynamics.In addition, a contrastive learning objective is introduced to improve representation discriminability between correct answer entities and irrelevant candidates.Experiments on two public benchmark datasets, WebQSP and ComplexWebQuestions (CWQ), show that the proposed method outperforms representative GNN-based baselines, achieving notable gains in both F1 and Hits@1 for multi-hop reasoning tasks. Ablation studies further indicate that each component contributes to the overall effectiveness of the proposed framework.