Skip to content
Book Open access

GraphMatch: A Dual-Channel Framework for Compositional Cross-Modal Retrieval

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 42 references

Abstract

Cross-modal retrieval has made remarkable strides with embedding-based methods that map different modalities into a unified space. However, these approaches fail on compositional queries that require understanding multiple entities and their relationships. The root cause is the flat feature paradigm, which compresses structured semantics into single vectors, irrevocably losing relational information. We present GraphMatch, a neural subgraph alignment framework that explicitly preserves and matches compositional structure. Our key innovations are threefold: (i) a dual-channel architecture that dynamically routes simple queries to efficient embedding matching and complex queries to structured subgraph alignment, (ii) a query-to-graph parser that transforms natural language into executable query subgraphs with entities, attributes, and relations, and (iii) a differentiable subgraph matching algorithm that learns to align query graphs with subgraphs through joint optimization of node semantics and structural consistency. This enables retrieval to understand not just what entities are present, but how they relate. Extensive experiments show GraphMatch achieves state-of-the-art on standard tasks while delivering a 19.2% absolute improvement on complex compositional queries, closing the gap between simple embedding matching and true relational understanding. Our work establishes neural subgraph alignment as a new paradigm for fine-grained cross-modal retrieval where structure matters.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.