Improving Multilingual Entity Linking via Knowledge-Graph Rarity Metrics
September 8, 2026
New structural metrics for entity rarity reveal accuracy drops of 15.4-39.9% in multimodal entity linking compared to popularity-based metrics. A training-free framework allows vision-language models to iteratively search and reason over Wikipedia to gather dynamic evidence for these rare entities.
HOW THIS AFFECTS YOU
●
builderYou can implement a reasoning-based retrieval framework to improve performance on long-tail entities without retraining.
●
researcherYou can use knowledge-graph structural metrics to identify more nuanced failure modes in entity linking.