OpenMEL: Unsupervised Multimodal Entity Linking Using Noise-Free Expanded Queries and Global Coherence

Multimodal Entity Linking (MEL), which involves disambiguating a mention composed of multimodal inputs to a multimodal knowledge base (KB), has gained increasing attention. Although existing MEL approaches using supervised learning show promising performance, they depend heavily on large-scale labeled training data, which is expensive to obtain for each new scenario. Unsupervised learning MEL methods, on the other hand, typically consist of two main steps. In the first multimodal data encoding step, these methods either assume that the multimodal data inputs are of high quality or attempt to filter out the noisy modality. In the second entity ranking step, they employ a bipartite graph to model the relationships only between mentions and entities. However, unsupervised methods face challenges in both steps. In the first step, data quality issues arise, including limited context in textual inputs and noise in the corresponding images. Moreover, in the second step, the bipartite graph fails to capture coherence between highly correlated entities within the KB, which offers clues on shared domains among entities. This limitation hinders effective retrieval of the target entity. To address these issues, we propose a novel unsupervised learning framework, OpenMEL, for solving the MEL task. We enhance the textual modality contextual information by incorporating full context comprehension and general knowledge, and generates three levels of visual inputs for further adaptive selection to handle noise. To capture global entity coherence, we construct a tree cover structure, defining it as a maximum spanning tree with bounded nodes to meet the MEL objective. We then introduce a greedy algorithm with theoretical guarantees to solve this problem. Experimental results on three public benchmark datasets show that OpenMEL outperforms various state-of-the-art baselines.