Mental health supportive response matching via multi-view graph attention on concept interaction graphs

Abstract Selecting an appropriate supportive response for a mental health question post is a key step toward scalable mental health question answering, yet it remains difficult due to semantic mismatch, noisy informal language, and the need to align supportive intent beyond surface lexical overlap. We propose MF-GAT, a novel multi view graph attention matching framework that constructs a Concept Interaction Graph (CIG) to explicitly encode post response concept alignments and their interaction structure. MF-GAT learns three complementary evidence streams, including local interaction features, multi view fusion features, and global context features, and applies view specific graph attention to propagate and reweight informative relational signals over the CIG. A gated fusion module then adaptively integrates the view representations into a unified matching vector for prediction. We evaluate MF-GAT both as a pair classification model and as a retrieval ranking model for selecting the best support from a candidate pool, reporting Accuracy and F1 together with standard ranking metrics including MRR and NDCG. On the MHQA benchmark, MF-GAT achieves 0.95 Accuracy and 0.85 F1, outperforming BERT (0.89, 0.67), CIG-GCN (0.92, 0.79), ARC-II (0.85, 0.60), and MatchPyramid (0.82, 0.59). These results show that novel multi view interaction graph modeling with attention based propagation improves both supportive response classification and practical retrieval quality for mental health support selection.