wikidata-rdf2vec-evaluation

Evaluation data of the RDF2Vec embedding release of the Wikidata truthy dump of 2026-04-24: the temporal long-tail holdout, KGrEaT-ready vectors of every family we trained or embedded, and the raw results behind the paper's tables. The embedding tables are the wikidata-rdf2vec-* repositories of this organisation; the code is at https://anonymous.4open.science/r/wikidata-rdf2vec-pipeline-7410/. Anonymized for double-blind review. Licence: CC0 1.0, as Wikidata.

Temporal long-tail holdout (holdout/)

holdout.parquet: one row per item and property: entity (item IRI), prop (P31, P106, P17 or P136), label (class value), split (train: value already in the 2026-04-24 snapshot; test: value only in the 2026-10-07 dump), occ (triples with the item in the 2026-04-24 snapshot), bin (1-4, 5-19, >=20), has_label, has_desc.

facts_20260424.parquet, facts_20261007.parquet: all item-to-item facts of the four properties plus owl:sameAs (redirects of merged items) in the two truthy dumps (subject, prop, object), from which the holdout is built. The 2026-10-07 dump will not stay on the Wikidata mirrors, so the facts are kept here.

KGrEaT-ready vectors (kgreat_exports/)

<family>.tsv in KGrEaT's embedding format (entity IRI, then the vector, tab-separated) for the KGrEaT target entities; <family>.missing.txt lists target entities without a vector. Families: RDF2Vec GPU (torchrec_*) and CPU (gensim_*), TransE retrained on the snapshot with and without frequency filter (transe_l1_2026*, _l2: L2-normalized), text encoders over labels, descriptions or verbalized neighbourhoods (jina_*, qwen3_*, tarte_*) and graph-text concatenations (concat_*). Copy them to kg/wikidata/embedding/ of a KGrEaT checkout with our patch to rerun the evaluation. Third-party embeddings (PyTorch-BigGraph TransE 2019, Wikidata Vectors, embeddings.cc) are not redistributed; the code repository contains the scripts that build their exports from the public sources (tools/export_kgreat_embeddings.py, tools/export_embeddingscc.py).

Results (results/)

kgreat/: one TSV per KGrEaT task (24) with every family, evaluation mode (KNOWN, COMMON, ALL), estimator and metric of the final run, as KGrEaT scores it. kgreat_fallback/: the same, with ALL for classification and regression recomputed with the fallbacks for uncovered entities described in the paper (majority class, mean label; the paper's tables use these), and the task populations in fallback_populations.csv. holdout_cells.csv, holdout_pooled.csv: holdout accuracy per family, property and frequency bin, and pooled over properties. kgreat_task_coverage.csv: share of each task's entities with a vector, per family (the Cov columns); kgreat_target_coverage.csv: KGrEaT target items with a vector, per family (the coverage row).

To regenerate Tables 5 and 9 of the paper and its rankings and tests with the code repository:

python tools/make_kgreat_tables.py results/kgreat_fallback out results/kgreat_task_coverage.csv \
    results/kgreat_target_coverage.csv
python tools/kgreat_friedman.py results/kgreat_fallback --mode ALL --out ranking.md \
    --families torchrec_d200 transe_l1_l2 embeddingscc_keci jina_512

SHA-256 checksums of every file: SHA256SUMS.