---
language:
- en
license: cc-by-sa-4.0
pretty_name: "Coarse-Grained and Fine-Grained WiC Datasets"
task_categories:
- text-classification
task_ids:
- semantic-similarity-classification
tags:
- arxiv:2609.20593
- word-in-context
- wic
- word-sense-disambiguation
- lexical-semantics
- lexical-ambiguity
- wordnet
- coarsewsd-20
size_categories:
- 1K<n<10K
configs:
- config_name: coarse_grained
  data_files:
  - split: test
    path: data/coarse_grained/test.jsonl
- config_name: fine_grained
  data_files:
  - split: test
    path: data/fine_grained/test.jsonl
---

# Coarse-Grained and Fine-Grained WiC Datasets

This is the official dataset repository of [WiC is Not WSD: A Study on LLMs and Lexical Ambiguity Resolution](https://arxiv.org/abs/2609.20593), accepted to the AACL 2026 main conference.

## About
![Example of a WiC instance where sense granularity changes the answer](main_figure.png)

This repository contains two Word-in-Context (WiC) evaluation datasets. Given two sentences that contain the same
target word, the task is to decide whether the word has the **same meaning** in both (`True`) or **different
meanings** (`False`). The two datasets differ in the granularity of the sense distinctions they encode, as
illustrated above for the word *club*:

- **Coarse-grained WiC**: 2,798 sentence pairs over 20 ambiguous nouns, built from CoarseWSD-20 by pairing
  sentences with the same target word and comparing their coarse-grained sense labels.
- **Fine-grained WiC**: 1,185 sentence pairs over 924 target words, newly constructed from WordNet 3.1 noun and verb
  entries. Each instance also includes the WordNet definition of the target word in each sentence.

Both datasets are evaluation sets and are released as a single `test` split.

## Dataset

The two configs differ in the granularity of the sense distinctions they encode:

| Config | Instances | Target words | Distinct senses | Unique sentences | True | False |
|---|---|---|---|---|---|---|
| `coarse_grained` | 2,798 | 20 | 53 | 5,594 | 1,399 | 1,399 |
| `fine_grained` | 1,185 | 924 | 1,655 | 2,266 | 594 | 591 |

Both datasets are intended for **evaluation only** and are provided as a single `test` split.

## Loading

```python
from datasets import load_dataset

coarse = load_dataset("cardiffnlp/GrainWiC", "coarse_grained", split="test")
fine   = load_dataset("cardiffnlp/GrainWiC", "fine_grained", split="test")
```


### `coarse_grained`

Constructed automatically from **CoarseWSD-20** (Loureiro et al., 2021), a coarse-grained WSD dataset
covering 20 ambiguous nouns. Sentences containing the same target word are paired and the WiC label
is derived by comparing their coarse-grained sense labels: same sense label gives `True`, otherwise `False`.
The label distribution is exactly balanced.

| Field | Type | Description |
|---|---|---|
| `id` | int | Row index (0-based). |
| `word` | string | Target word (one of 20 nouns). |
| `target_index` | string | Position of the target word in `sentence1` and `sentence2`, as `index1-index2`. Indices are 0-based and count whitespace-separated tokens, following the original WiC data format. |
| `sentence1` | string | First context: a sentence from CoarseWSD-20. Text is lower-cased and pre-tokenised, with tokens separated by single spaces. |
| `sentence2` | string | Second context. |
| `label` | string | `"True"` if the target word has the same coarse sense in both sentences, else `"False"`. |

Per-word statistics (from the paper, Table 8):

| Word | Senses | Pairs | True | False |
|---|---|---|---|---|
| apple | 2 | 424 | 212 | 212 |
| arm | 2 | 70 | 35 | 35 |
| bank | 2 | 46 | 23 | 23 |
| bass | 3 | 508 | 254 | 254 |
| bow | 3 | 106 | 53 | 53 |
| chair | 2 | 40 | 20 | 20 |
| club | 3 | 96 | 48 | 48 |
| crane | 2 | 78 | 39 | 39 |
| deck | 2 | 14 | 7 | 7 |
| digit | 2 | 10 | 5 | 5 |
| hood | 3 | 38 | 19 | 19 |
| java | 2 | 824 | 412 | 412 |
| mole | 5 | 70 | 35 | 35 |
| pitcher | 2 | 26 | 13 | 13 |
| pound | 2 | 20 | 10 | 10 |
| seal | 4 | 108 | 54 | 54 |
| spring | 3 | 182 | 91 | 91 |
| square | 4 | 82 | 41 | 41 |
| trunk | 3 | 38 | 19 | 19 |
| yard | 2 | 18 | 9 | 9 |
| **Total** | **53** | **2,798** | **1,399** | **1,399** |

Example:

```json
{"word": "apple", "target_index": "6-18",
 "sentence1": "the local economy includes forestry , apple orchard and sheep and dairy farming .",
 "sentence2": "symbol is one of the four standard fonts available on all postscript - based printers , starting with apple 's original laserwriter ( 1985 ) .",
 "label": "False"}
```

### `fine_grained`

A newly constructed WiC-style dataset that follows the construction of the original WiC benchmark
(Pilehvar and Camacho-Collados, 2019) but is built from **WordNet 3.1** noun and verb entries, with
modifications intended to limit the effect of memorisation by language models. Each target word is
paired with its WordNet gloss, and sentence pairs are built from the usage examples of the senses.
Unlike standard WiC, each instance also carries the **sense definition** of the target word in each
sentence, so the dataset can be used both with and without explicit sense information.

| Field | Type | Description |
|---|---|---|
| `id` | int | Row index (0-based). |
| `word` | string | Target word (noun or verb lemma). |
| `sentence1` | string | First context (WordNet usage example). |
| `sentence2` | string | Second context (WordNet usage example). |
| `definition1` | string | WordNet definition of the target word's sense in `sentence1`. |
| `definition2` | string | WordNet definition of the target word's sense in `sentence2`. |
| `label` | string | `"True"` if `definition1` and `definition2` are the same sense, else `"False"`. |

Example:

```json
{"word": "absence",
 "sentence1": "he was surprised by the absence of any explanation",
 "sentence2": "he visited during my absence",
 "definition1": "the state of being absent",
 "definition2": "the time interval during which something or somebody is away",
 "label": "False"}
```

## Files

```
data/coarse_grained/test.jsonl   # 2,798 rows
data/fine_grained/test.jsonl     # 1,185 rows
```

Each file is JSON Lines (one JSON object per line, UTF-8). Labels in both configs are the strings
`"True"` / `"False"`.

## Notes

- Sentences in `coarse_grained` are lower-cased and tokenised (punctuation separated by spaces), as in CoarseWSD-20.
- `coarse_grained` contains one exact duplicate pair (`bank`, `target_index` `12-5`). It is kept so that the
  counts match the paper.
- In `coarse_grained`, `target_index` points at the target word in 2,797 of 2,798 rows. In one `club` row
  (`target_index` `85-19`) the first index is off because of a tokenisation irregularity in the source sentence.
- These are evaluation sets. There is no training split.

## Source datasets

- **CoarseWSD-20**: Daniel Loureiro, Kiamehr Rezaee, Mohammad Taher Pilehvar, and Jose Camacho-Collados. 2021.
  *Analysis and Evaluation of Language Models for Word Sense Disambiguation.* Computational Linguistics, 47(2):387–443.
- **WordNet 3.1**: George A. Miller. 1995. *WordNet: A Lexical Database for English.* Communications of the ACM;
  Christiane Fellbaum (ed.). 1998. *WordNet: An Electronic Lexical Database.* MIT Press.
- **WiC** (construction procedure): Mohammad Taher Pilehvar and Jose Camacho-Collados. 2019.
  *WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representations.* NAACL-HLT 2019.

## License and attribution

This dataset is released under [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/).

- The `coarse_grained` sentences and sense labels are derived from
  [CoarseWSD-20](https://github.com/danlou/bert-disambiguation) (Loureiro et al., 2021).
- The `fine_grained` definitions and example sentences come from WordNet 3.1 and are used under the
  [WordNet License](https://wordnet.princeton.edu/license-and-commercial-use).
  WordNet 3.1 Copyright 2011 by Princeton University. All rights reserved.

## Citation

If you use these datasets, please cite the paper:

```bibtex
@inproceedings{zhou2026wicnotwsd,
  title     = {WiC is Not WSD: A Study on LLMs and Lexical Ambiguity Resolution},
  author    = {Zhou, Yi and Rezaee, Kiamehr and Bollegala, Danushka and Pilehvar, Mohammad Taher and Camacho-Collados, Jose},
  booktitle = {Proceedings of the 2026 Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (AACL 2026)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2609.20593}
}
```

Please also cite CoarseWSD-20 when using the `coarse_grained` config, and WordNet when using the `fine_grained` config.