23 October 2026 in SSRN Electronic Journal
Grounded Multimodal Named Entity Recognition (GMNER) aims to recognize textual entities and ground them to corresponding visual regions, requiring effective integration of heterogeneous multimodal information. However, existing approaches often rely on independent entity--region alignment and overlook structured inter-entity dependencies, leading to ambiguous grounding and redundant region assignments, particularly in multi-entity scenarios. To address these challenges, we propose ANCOR, a structured framework for GMNER based on anchor-guided grounding. ANCOR follows a three-stage design that progressively exploits complementary textual and visual information. First, a boundary-guided multimodal entity recognition stage produces structured entity predictions. Second, a type-aware comparative visual prior generation stage leverages fine-grained entity types to construct discriminative visual priors for grounding. Third, an anchor-guided grounding mechanism models structured inter-entity dependencies through anchor-guided exclusion during training, where a confidently grounded entity is selected as an anchor to guide the disambiguation of others. Extensive experiments on benchmark GMNER datasets demonstrate that ANCOR consistently outperforms existing methods, particularly in challenging multi-entity scenarios. These results demonstrate the effectiveness of explicitly modeling cross-entity dependencies for accurate and robust multimodal entity grounding.