Think Before You Link: Rarity, Reasoning, and Retrieval in Multilingual Entity Linking
Rare entities expose failures that popularity-based tests do not catch.
The paper adds knowledge-graph structure measures to identify poorly documented or weakly connected entities, finding rare cases missed by pageview-style metrics. Across those slices, state-of-the-art multilingual multimodal entity linking accuracy drops 15.4% to 39.9%. Its training-free system has a reasoning-capable vision-language model search Wikipedia iteratively and use gathered evidence. Retrieval and reasoning work best together, improving MERLIN by 6.9% overall and up to 23.3% on rare-entity slices. HF Daily Papers' note
The paper adds knowledge-graph structure measures to identify poorly documented or weakly connected entities, finding rare cases missed by pageview-style metrics. Across those slices, state-of-the-art multilingual multimodal entity linking accuracy drops 15.4% to 39.9%. Its training-free system has a reasoning-capable vision-language model search Wikipedia iteratively and use gathered evidence. Retrieval and reasoning work best together, improving MERLIN by 6.9% overall and up to 23.3% on rare-entity slices. HF Daily Papers' note
score 5