ando55/WikiSQE_experiment

Dataset Card for WikiSQE\_experiment

Dataset Description

• —Repository: https://github.com/ken-ando/WikiSQE
• —Paper: https://arxiv.org/abs/2305.05928 (AAAI 2024)

Dataset Summary

WikiSQE_experiment is the official evaluation split for WikiSQE: A Large‑Scale Dataset for Sentence Quality Estimation in Wikipedia.

While the parent dataset (ando55/WikiSQE) contains every sentence flagged with a quality problem in the full edit history of English Wikipedia, this repo provides the exact train/validation/test partitions used in the AAAI 2024 paper. It offers ≈ 8.3 million sentences organised as:

• —*27 dataset groups*** (20 frequent quality labels + 5 Quality type categories + 2 Coarse groups)
• —3 standard splits per group (train, val, test) – for example citation/train, citation/val, …

Each split blends labeled and unlabeled sentences at a 1 : 1 ratio to support semi-supervised and positive/negative training paradigms.

Need the full dump? Head to https://huggingface.co/datasets/ando55/WikiSQE.

Dataset Structure

Groups (27)

Notes

• —`all` contains a random subset uniformly sampled from the entire WikiSQE corpus. Use it when you want a representative slice without downloading the full 3.4 M‑sentence dump.
• —`sac` contains a composite set randomly drawn from the three fine‑grained categories `disputed claim`, `information addition`, and `syntactic or semantic revision`. It was introduced in the paper to study sentence‑level action classification.

Split sizes

Data Fields

Download & Usage

1 — Download the Parquet snapshot

# Install (if you haven't already)
pip install --upgrade datasets huggingface_hub

from huggingface_hub import snapshot_download

repo_dir = snapshot_download(
    repo_id="ando55/WikiSQE_experiment",  # this repo
    repo_type="dataset",
    local_dir="WikiSQE_experiment_parquet",
    local_dir_use_symlinks=False,
)
print("Saved at:", repo_dir)

This grabs all 27 configs (each providing train, val, test) in their native Parquet format.

2 — Load a split on‑the‑fly

Streaming access without a full download:

from datasets import load_dataset

ds = load_dataset(
    "ando55/WikiSQE_experiment",
    name="citation",   # choose any config
    split="train",
    streaming=True
)

3 — (Optionally) Convert Parquet → CSV

The downloaded files are in Parquet format. By converting them to CSV, they can be used for various purposes.

import pyarrow.dataset as ds, pyarrow.csv as pv, pyarrow as pa, pathlib

src = pathlib.Path("WikiSQE_experiment_parquet")
dst = pathlib.Path("WikiSQE_experiment_csv"); dst.mkdir(exist_ok=True)

for pq in src.rglob("*.parquet"):
    cfg   = pq.parent.name  # config name
    split = pq.stem         # train/val/test
    print(cfg, split)
    out   = dst / f"{cfg}_{split}.csv"
    first = not out.exists()
    dset  = ds.dataset(str(pq))
    with out.open("ab") as f, pv.CSVWriter(
            f, dset.schema,
            write_options=pv.WriteOptions(include_header=first)) as w:
        for batch in dset.to_batches():
            w.write_table(pa.Table.from_batches([batch]))

Citation

@inproceedings{ando-etal-2024-wikisqe,
  title     = {{WikiSQE}: A Large-Scale Dataset for Sentence Quality Estimation in Wikipedia},
  author    = {Ando, Kenichiro and Sekine, Satoshi and Komachi, Mamoru},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  year      = {2024},
  volume    = {38},
  number    = {16},
  pages     = {17656--17663},
  address   = {Vancouver, Canada},
  publisher = {Association for the Advancement of Artificial Intelligence}
}

Happy experimenting! 🚀