Multi-Perspective Evidence Synthesis and Reasoning for Unsupervised Multimodal Entity Linking
MSR-MEL is a framework for unsupervised Multimodal Entity Linking that constructs comprehensive evidence offline and reasons over it online using Large Language Models (LLMs). It synthesizes instance, group, lexical, and statistical evidence. The online stage uses an LLM to re-rank candidates based on multi-perspective signals to achieve superior entity linking performance.