Entity Linking datasets¶

These loaders all follow the same shape as algorithms.el's convention: dataset1 (mentions) carries context=(text, start, end) and no description, dataset2 (the KB) carries a label and, where the source provides one, a description. A mention with no resolvable gold entity (NIL) is skipped entirely rather than kept with an empty target — the same convention used by every loader on this page.

AIDA-CoNLL¶

AidaConllDataset links named-entity mentions in Reuters newswire text (the CoNLL 2003 NER corpus) to Wikidata QIDs, via the Hugging Face Hub republish cyanic-selkie/aida-conll-yago-wikidata (the original release's licensing is murky; this republish is cc-by-sa-3.0). KB entities' descriptions are fetched live from Wikipedia's MediaWiki API and disk-cached. Has a native train/validation/test split:

from linkingtk.datasets import AidaConllDataset

dataset = AidaConllDataset()
mentions, kb, ground_truth = dataset.load()
train_pairs, test_pairs, val_pairs = dataset.load_splits()

Zeshel¶

ZeshelDataset is BLINK's own zero-shot EL benchmark (Logeswaran et al. 2019): mentions inside a Wikia (fandom) page's text, linked to other Wikia pages, with train/validation/test domains kept disjoint so the model must resolve entities from description text alone. Loaded from the Hugging Face Hub republish naist-nlp/zeshel. See BlinkLinker. Has a native train/validation/test split; load_splits() skips the ~1.2GB entity dictionary download entirely:

from linkingtk.datasets import ZeshelDataset

dataset = ZeshelDataset()
mentions, kb, ground_truth = dataset.load()
train_pairs, test_pairs, val_pairs = dataset.load_splits()

DaMuEL¶

DamuelDataset links Wikipedia mentions across 53 languages to Wikidata QIDs (Kubeša & Straka, 2023), published on LINDAT/CLARIAH-CZ as one large tar per language (278MB-26.3GB). This loader streams the tar and stops after max_parts of its 500 shuffled .xz shards, so it never materializes a full language in memory:

from linkingtk.datasets import DamuelDataset

dataset = DamuelDataset(language="en", max_parts=2)
mentions, kb, ground_truth = dataset.load()

Raise max_parts for denser ground truth, especially on lower-resource languages. DaMuEL publishes no native train/dev/test split, so there's no load_splits() here.

LCQuAD 2.0¶

Lcquad2Dataset derives entity mentions from a large-scale KGQA benchmark (~30,000 crowd-sourced questions, each paired with a SPARQL query over Wikidata) that itself carries no mention-span annotations: every wd:Qxxx referenced in a row's sparql_wikidata is a gold entity, and its text span is recovered by matching the entity's live Wikidata label against the question text (~81% resolve this way; the rest are skipped, same NIL-skip convention as the other loaders on this page). KB labels/descriptions are fetched live via Wikidata's wbgetentities API. Has a native train/test split (no validation split):

from linkingtk.datasets import Lcquad2Dataset

dataset = Lcquad2Dataset()
mentions, kb, ground_truth = dataset.load()
train_pairs, test_pairs, val_pairs = dataset.load_splits()  # val_pairs is always []

Unlike the other loaders here, load_splits() can't skip the network label lookup — resolving a mention's span requires each candidate entity's Wikidata label, not just its id — so it only skips building the full KB Entity list.