Fusion-Enhanced Bidirectional Multimodal Entity Linking Model

Multimodal knowledge graphs rely on multimodal entity linking (MEL) to align entity mentions across diverse modalities with structured knowledge bases. However, existing MEL methods face significant challenges in complex real-world scenarios, primarily due to semantic gaps between heterogeneous modalities, severe textual noise, and the presence of multiple non-salient visual entities. Conventional approaches, relying on indiscriminate feature fusion or end-to-end generative large language models (LLMs), lack explicit mechanisms to prevent cross-modal noise propagation, rendering them vulnerable to hallucinations and representational degradation. To address these critical limitations, we propose a Fusion-Enhanced Bidirectional Multimodal Entity Linking Model (FILLED) designed to systematically integrate feature enhancement, dynamic evaluation, and noise distillation. Specifically, the framework introduces: (i) a feature pyramid bidirectional fusion network to enrich visual representations; (ii) a phrase-aware text encoder to capture multi-granularity semantics; (iii) an LLM-guided dynamic gating mechanism to explicitly evaluate the modality-specific signal-to-noise ratio; and (iv) a bottleneck fusion module optimized with contrastive learning to execute physical noise filtration and enforce cross-modal semantic consistency. Extensive experiments on three benchmark datasets (Wiki-MEL, WikiDiverse, and Noisy-MEL) demonstrate that FILLED achieves competitive performance on standard corpora while establishing new state-of-the-art results on the extreme-interference Noisy-MEL dataset. Notably, on the highly perturbed noisy subset, FILLED achieves a Top-1 accuracy of 57.8% and a Mean Reciprocal Rank (MRR) of 83.6%, outperforming the strongest baseline by 3.2% and 1.1%, respectively. These metrics substantiate the proposed model’s exceptional robustness and long-tail disambiguation capability in highly noisy environments.