pywsd-datasets

Unified Word Sense Disambiguation benchmark datasets, normalized to modern wn lexicon sense IDs (oewn:2024 for English, OMW for other languages).

Companion to pywsd ≥ 1.3.0.

What's shipped (v0.2)

English, test-only Raganato all-words benchmark:

Config	Instances	OEWN 2024 coverage
en-senseval2-aw	2,282	99.43 %
en-senseval3-aw	1,850	99.51 %
en-semeval2007-aw	455	99.78 %
en-semeval2013-aw	1,644	100.00 %
en-semeval2015-aw	1,022	99.32 %

English, training corpora (via UFSAC v2.1):

Config	Split	OEWN 2024 coverage
en-semcor	train	see coverage_report
en-wngt	train	see coverage_report
en-masc	train	see coverage_report
en-senseval2_ls	train + test	lexical-sample
en-senseval3_ls	train + test	lexical-sample
en-semeval2007_t17_ls	test	lexical-sample

Run python -m pywsd_datasets.scripts.coverage_report locally to get up-to-date OEWN resolution rates after rebuilding.

Install

pip install pywsd-datasets

Use via HuggingFace datasets

from datasets import load_dataset

# Raganato all-words evaluation set
ds = load_dataset("alvations/pywsd-datasets", "en-senseval2-aw")

# SemCor training data
ds = load_dataset("alvations/pywsd-datasets", "en-semcor")

ds["test"][0] if "test" in ds else ds["train"][0]
# {'instance_id': 'd000.s000.t000', 'dataset': 'senseval2_aw',
#  'split': 'test', 'lang': 'en',
#  'tokens': ['The', 'art', 'of', 'change-ringing', ...],
#  'target_idx': 1, 'target_lemma': 'art', 'target_pos': 'n',
#  'source_sense_id': 'art%1:09:00::',
#  'source_sense_system': 'pwn_sensekey_3.0',
#  'sense_ids_wordnet': ['oewn-05646832-n'],
#  'wordnet_lexicon': 'oewn:2024', ...}

Use via the loader package

from pywsd_datasets.loaders.raganato import iter_instances as iter_raganato
from pywsd_datasets.loaders.ufsac import iter_instances as iter_ufsac

for inst in iter_raganato("senseval2"):
    print(inst.target_lemma, inst.sense_ids_wordnet)

for inst in iter_ufsac("semcor", "/path/to/ufsac-public-2.1"):
    print(inst.target_lemma, inst.sense_ids_wordnet)

Rebuild locally

pip install pywsd-datasets[dev]

# Raganato only (always works, ~1 MB fetch from our GH release mirror)
python -m pywsd_datasets.scripts.build_all

# With UFSAC corpora — download ufsac-public-2.1 separately (see below)
python -m pywsd_datasets.scripts.build_all \
    --ufsac-root ~/.cache/pywsd-datasets/ufsac/ufsac-public-2.1

# Coverage report across every built parquet:
python -m pywsd_datasets.scripts.coverage_report

UFSAC download

UFSAC v2.1 is distributed as a single Google Drive tarball (ufsac-public-2.1.tar.xz, ~196 MB). Fetch with gdown:

pip install gdown
mkdir -p ~/.cache/pywsd-datasets/ufsac
gdown 'https://drive.google.com/uc?id=1kwBMIDBTf6heRno9bdLvF-DahSLHIZyV' \
    -O ~/.cache/pywsd-datasets/ufsac/ufsac-public-2.1.tar.xz
cd ~/.cache/pywsd-datasets/ufsac && tar -xf ufsac-public-2.1.tar.xz

Schema

Every row follows WSDInstance:

instance_id, dataset, split, task, lang,
tokens[], pos_tags[], lemmas[],
target_idx, target_lemma, target_pos,
source_sense_id, source_sense_system,
sense_ids_wordnet[], wordnet_lexicon,
doc_id, sent_id

sense_ids_wordnet is list-valued to handle multi-gold instances and any PWN-3.0 key that splits into multiple OEWN 2024 synsets.

Multilingual / XL-WSD / BabelNet — deferred

loaders/xl_wsd.py exists as a stub and raises NotImplementedError. mappers/babelnet_to_wn.py is similarly unused. Why:

• XL-WSD uses BabelNet synset IDs as gold labels; resolving them to modern wn lexicon IDs requires the BabelNet → PWN 3.0 bridge file, which is distributed only with a BabelNet academic license.
• XL-WSD itself is CC-BY-NC 4.0 — we don't redistribute the data.

Reviving this path requires (a) a BabelNet license, (b) loading bn_to_wn.txt via babelnet_to_wn.load_bn_to_pwn3_map(), (c) selecting per-language OMW lexicons via mappers.omw_lookup.lexicon_for(lang), then (d) chaining through pwn3_to_oewn.pwn3_sensekey_to_wn(key, lexicon=...). All four pieces are in place — wiring them is blocked on the BabelNet mapping file. See the module docstrings for details.

Roadmap

• v0.2 (this release): Raganato all-words evaluation + UFSAC training corpora (SemCor, WNGT, MASC, Senseval lexical-sample).
• v0.3 (planned): WiC (CC-BY-NC — loader-only), CoarseWSD-20.
• Deferred: XL-WSD multilingual (needs BabelNet academic license).

Citation

If you use these datasets please cite the original sources:

• Raganato, Camacho-Collados, Navigli (2017). Word Sense Disambiguation: A Unified Evaluation Framework and Empirical Comparison. EACL.
• Vial, Lecouteux, Schwab (2018). UFSAC: Unification of Sense Annotated Corpora and Tools. LREC.
• Plus the specific evaluation or training set paper (Senseval-2 / 3, SemEval-2007 T17, SemEval-2013 T12, SemEval-2015 T13, SemCor, WNGT/Princeton Gloss Corpus, MASC).

License

MIT for the code. Each dataset keeps its original license — see the source papers. Raganato bundle and SemEval shared-task data are research-unrestricted; UFSAC is MIT.

Sense-ID mapping details

PWN 3.0 sense keys are resolved against OEWN 2024 via wn.compat.sensekey. The few percent of keys that fail to resolve are typically WN 3.0 synsets that OEWN split, merged, or removed — those rows ship with an empty sense_ids_wordnet list so the coverage report can flag them. Background:

• Kaf (2023). Mapping Wordnets on the Fly with Permanent Sense Keys. arXiv:2303.01847.

Known issues

• The upstream Raganato zip at http://lcl.uniroma1.it/wsdeval/ serves a mismatched TLS cert; our loader prefers the mirror on this repo's GitHub release assets and falls back to the original URL over HTTP.
• UFSAC v2.1 is distributed as a Google Drive tarball; the loader assumes you have it unpacked locally. A future release may mirror it.