Anes Abdelouahab Mahammedi, Abdelghani Bouziane, Yasser Yahiaoui, Djelloul Bouchiha · Research Square 2026 · 2026
DOI: 10.21203/rs.3.rs-11323391/v1
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Arabic entity linking loses a large share of mentions before disambiguation begins. We measure that loss with a layered audit of the public Arabic surface-form resources (Wikidata labels and aliases, ParaNames, Wikipedia titles, redirects and a held-out anchor dictionary). The instrument is exact-match lookup over 200,000 Arabic Wikipedia link mentions, which attributes every retrieved entity to a resource. Only 17.7% of the entities with an English Wikipedia article carry an Arabic label, and almost all of them have an Arabic article: Wikidata’s Arabic naming stops where Arabic Wikipedia stops. For entities that have one, all resources combined retrieve the correct entity for 74.4% of the mentions that differ from their target’s title, on a benchmark that draws each mention–entity pair once. Weighted by frequency in text, the figure is 86.8%. At rank 10 the shortfall is coverage, not ranking. Anchor text supplies almost all of it, and an identity predicts the ceiling from dictionary size exactly: the share of pairs the dictionary has recorded, plus a small constant on the rest. Stripping the definite article is the most valuable normalisation step. A hand analysis of the failures separates unrecorded name forms, which dominate for persons, from benchmark artefacts. The cheap remedies it suggests recover 0.1 to 0.4 points; character n-gram and dense string matchers recover 66% and 72% of the shortfall at 100 candidates per mention. On Mewsli-9 Arabic the same resources match the published alias-table baseline. Resources and code are released.
No comments yet — start the discussion below.