Dataset Card for WordNet-WSD

Dataset Summary

WordNet-WSD is a dataset for word sense disambiguation (WSD).

Each example contains:

• a sentence or short text (text),
• a target word to disambiguate (word),
• a list of candidate WordNet senses (definitions), and
• the index of the correct sense in that list (sense_index).

The dataset is distributed in Parquet format as dataset.parquet.

Supported Tasks and Leaderboards

• Task: word sense disambiguation framed as per-example multi-class classification.
• Input: (text, word, definitions)
• Label: sense_index (index into the row-specific definitions list)

Because the candidate sense set changes per row, this is not a single global-classification problem over a fixed label vocabulary.

Languages

• English (en)

Dataset Structure

Data Instances

An example instance:

{
    "text": "A dog is opening a drawer and climbing inside",
    "word": "dog",
    "definitions": [
        {
            "definition": "a member of the genus Canis ... domesticated by man",
            "lemma": "dog",
            "pos": "n",
            "synset_key": "dog%1:05:00::"
        },
        {
            "definition": "a dull unattractive unpleasant girl or woman",
            "lemma": "dog",
            "pos": "n",
            "synset_key": "dog%1:18:01::"
        }
    ],
    "sense_index": 0
}

Data Fields

• text (string): The input sentence or short text containing the target word.
• word (string): The target lexical item to disambiguate.
• definitions (list[object]): Candidate senses for the target word.
• definitions[].definition (string): Gloss for the candidate sense.
• definitions[].lemma (string): Lemma associated with the candidate sense.
• definitions[].pos (string): Part of speech (for example n, v, a, r).
• definitions[].synset_key (string): WordNet sense key identifier.
• sense_index (int64): Index of the correct sense in definitions.

Data Splits

The dataset is split into three partitions (see notebooks/stratify.ipynb):

Split	File	Size (approx.)	Share
train	train.parquet	~30,201	80%
validation	validation.parquet	~2,863	10%
test	test.parquet	~2,864	10%

Splits are stratified by word so that each target word's frequency distribution is preserved across all three partitions. Words with fewer than 10 total occurrences in the dataset cannot be safely stratified (a 20% sample would yield fewer than 2 representatives, which is required for the subsequent 50/50 val/test split); those rows are placed entirely in the training set.

Split Creation Procedure

1. Compute per-word occurrence counts.
2. Separate rare words (count < 10) from the stratifiable subset.
3. Apply train_test_split with test_size=0.2 and stratify=word on the stratifiable subset → train + temp.
4. Apply a second train_test_split with test_size=0.5 and stratify=word on temp → validation + test.
5. Append all rare-word rows to train and shuffle.

Both splits use random_state=42 for reproducibility.

Dataset Creation

Curation Rationale

The dataset supports lexical semantic modeling and evaluation of WSD systems by pairing context (text) with target terms and WordNet 2.1 candidate senses.

Source Data

Context sentences are drawn from two established semantic similarity benchmarks, redistributed under their respective license terms (see Licensing Information):

• SICK-R — Marelli et al. (2014), distributed under CC BY-NC-SA 3.0;
• STS-B — Cer et al. (2017), as distributed via the GLUE benchmark (Wang et al., 2019).

The preparation pipeline (see notebook notebooks/prepare-dataset.ipynb) reads JSON files, flattens nested request/response structures, and writes a normalized Parquet table.

Annotations

• Annotation type: machine-generated labels.
• Label format: row-local index into candidate senses.

Personal and Sensitive Information

No personal data collection process is described for this dataset. Content appears to be short lexical/sentence examples, but downstream users should still run their own privacy review before production use.

Dataset Statistics

• Total examples: 35,928
• Unique target words: 3,300
• Candidate senses per example: min 2, mean 7.85, max 75
• Consistency check: all sense_index values are valid indexes within the corresponding definitions list.

Considerations for Using the Data

Social Impact and Intended Use

Intended for:

• WSD model training and evaluation,
• lexical semantic research,
• benchmarking disambiguation strategies with dynamic candidate inventories.

Not intended as a standalone factual knowledge source.

Bias, Risks, and Limitations

• Word and sense distributions are imbalanced.
• Since labels are machine-generated, systematic errors may exist for specific words or contexts.
• Candidate set size varies per row, making performance comparisons sensitive to evaluation protocol.

How to Load

from datasets import load_dataset

# Full dataset (single file)
dataset = load_dataset("parquet", data_files="dataset.parquet")
print(dataset["train"][0])

# Pre-split files
dataset = load_dataset(
    "parquet",
    data_files={
        "train": "train.parquet",
        "validation": "validation.parquet",
        "test": "test.parquet",
    },
)
print(dataset["train"][0])

Citation

If you use this dataset, cite this repository, the source corpora (SICK-R and STS-B), and WordNet as the underlying lexical resource.

@inproceedings{marelli2014sick,
    title={A {SICK} cure for the evaluation of compositional distributional semantic models},
    author={Marelli, Marco and Menini, Stefano and Baroni, Marco and Bentivogli, Luisa and Bernardi, Raffaella and Zamparelli, Roberto},
    booktitle={Proceedings of the Ninth International Conference on Language Resources and Evaluation ({LREC}'14)},
    pages={216--223},
    year={2014},
    address={Reykjavik, Iceland},
    url={https://aclanthology.org/L14-1314/}
}

@inproceedings{cer2017semeval,
    title={{SemEval}-2017 Task 1: Semantic Textual Similarity Multilingual and Crosslingual Focused Evaluation},
    author={Cer, Daniel and Diab, Mona and Agirre, Eneko and Lopez-Gazpio, I{\~n}igo and Specia, Lucia},
    booktitle={Proceedings of the 11th International Workshop on Semantic Evaluation ({SemEval}-2017)},
    pages={1--14},
    year={2017},
    address={Vancouver, Canada},
    doi={10.18653/v1/S17-2001},
    url={https://aclanthology.org/S17-2001/}
}

@inproceedings{wang2019glue,
    title={{GLUE}: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding},
    author={Wang, Alex and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel R.},
    booktitle={Proceedings of the International Conference on Learning Representations (ICLR)},
    year={2019},
    url={https://arxiv.org/abs/1804.07461}
}

@inproceedings{miller1995wordnet,
    title={WordNet: A Lexical Database for English},
    author={Miller, George A.},
    booktitle={Communications of the ACM},
    year={1995}
}

Licensing Information

The dataset (context sentences, candidate sense inventories, and silver-standard annotations) is released under the [Creative Commons Attribution-NonCommercial-ShareAlike 3.0 (CC BY-NC-SA 3.0)] (https://creativecommons.org/licenses/by-nc-sa/3.0/) license.

The context sentences are redistributed from existing resources, and the license of a redistribution cannot be more permissive than the license of its sources:

Content	Source	Applicable license
Context sentences (subset)	SICK-R — Marelli et al. (2014)	CC BY-NC-SA 3.0
Context sentences (subset)	STS-B — Cer et al. (2017), as distributed via GLUE (Wang et al., 2019)	Source-dependent; GLUE refers users to the original dataset licenses
Glosses and synset keys	WordNet 2.1 — Miller (1992), Fellbaum (1998)	WordNet 2.1 License (attribution required)
Original WSD annotations and dataset files	this repository	CC BY-NC-SA 3.0
Preparation and split code (notebooks/)	this repository	MIT

Because SICK-R is distributed under CC BY-NC-SA 3.0, the dataset as a whole — including the newly generated WSD annotations — is distributed under the same license (version 3.0, as required by the ShareAlike clause). The code artifacts (preparation and stratification notebooks) are released under the MIT license; this applies to the code only and does not extend to the data.

When using or redistributing this dataset you must:

1. comply with CC BY-NC-SA 3.0 (attribution, non-commercial use, share-alike);
2. attribute the source corpora: SICK-R (Marelli et al., 2014), STS-B (Cer et al., 2017), and GLUE (Wang et al., 2019) for its distribution of STS-B;
3. comply with the WordNet 2.1 license for the glosses and synset keys (see https://wordnetcode.princeton.edu/2.1/LICENSE), including attribution of WordNet and preservation of its copyright notice;
4. verify the licensing of the STS-B source data with respect to your intended use, since GLUE defers to the original dataset licenses.