foudil/HU-Nat

Naturalness Spectrum

An 8-level ordinal dataset for English-text naturalness: the degree to which a text conforms to the distributional and stylistic regularities of fluent, idiomatic native English, independent of its semantic content. Spans the spectrum from severely disfluent statistical-MT output, through learner English at three CEFR proficiency bands, to professionally edited native prose at three reading levels.

To our knowledge, this is the first dataset explicitly labelled for naturalness as a continuous ordinal property. It is constructed by composing three existing corpora whose texts can be ordered by provenance, rather than by collecting fresh human ratings — no new human annotation was solicited.

Dataset summary

Size	65,868 sentences (52,694 train / 13,174 test, stratified 80/20)
Label space	Integer 0–7, ordinal (higher = more natural)
Sentence length	5–60 words (filtered)
Language	English
Label provenance	By source corpus (no per-sentence annotation)
Constituent corpora	WMT06, W&I+LOCNESS, OneStopEnglish

The 8-level spectrum

Label	Source	Description	Train
0	WMT06 (human fluency scores 1–2)	Very disfluent MT output	1,947
1	WMT06 (human fluency score 3)	Mediocre MT output	2,560
2	W&I CEFR A	Beginner learner English	6,990
3	W&I CEFR B	Intermediate learner English	9,099
4	W&I CEFR C	Advanced learner English	7,786
5	OneStopEnglish Elementary	Simplified native English	8,102
6	OneStopEnglish Intermediate	Standard native English	8,195
7	OneStopEnglish Advanced	Sophisticated native English	8,015

The spectrum groups naturally into three macro-bands — MT (0–1), Learner (2–4), Native (5–7) — and the ordering within each band is established by construction from the source labels.

Usage

from datasets import load_dataset

ds = load_dataset("foudil/HU-Nat")
print(ds)
# DatasetDict({
#     train: Dataset({features: ['text', 'label', 'source'], num_rows: 52694}),
#     test:  Dataset({features: ['text', 'label', 'source'], num_rows: 13174}),
# })

print(ds["train"][0])
# {'text': '...', 'label': 6, 'source': 'OSE_Intermediate'}

Data fields

Field	Type	Description
text	string	English sentences
label	int8	Ordinal naturalness label, 0–7
source	string	Source corpus tag (e.g. WMT06_1-2, WI_CEFR_B, OSE_Advanced)

Splits

Split	Sentences
train	52,694
test	13,174

Dataset creation

Curation rationale

Existing approaches to text naturalness fall into two camps, each with limitations:

• MT-evaluation metrics (COMET, BLEURT) conflate naturalness with adequacy and require a reference translation, so they cannot score isolated text.
• Language-model perplexity is a one-dimensional proxy that rewards low-entropy text independent of distributional match, and inherits the bias of its training corpus.

No corpus provides continuous naturalness ratings, and no off-the-shelf encoder produces naturalness-aware representations. The key construction insight is that while no single corpus provides graded naturalness ratings, several corpora exist whose texts can be ordered by provenance: machine-translated output is, by definition, less natural than text written by native speakers; CEFR-graded learner essays sit between these extremes; reading-level rewrites of the same native articles preserve content while varying surface form.

Source data

WMT06 (labels 0–1). Human fluency judgements from the WMT 2006 shared task on machine-translation evaluation. We restrict to English-target language pairs (DE→EN, ES→EN, FR→EN). MT output was rated on a 1–5 scale; we retain only outputs with scores ≤ 3 (the disfluent end). Sentences with score ≤ 2 are assigned label 0; sentences with score 3 are assigned label 1.

W&I+LOCNESS (labels 2–4). The Writing & Improvements corpus from the BEA-2019 shared task: learner essays written by non-native English speakers at CEFR proficiency levels A (beginner), B (intermediate), and C (advanced). Original (pre-correction) text is used, since the learner errors — inconsistent tense, missing articles, non-idiomatic phrasing — constitute the naturalness signal of interest.

OneStopEnglish (labels 5–7). The OneStopEnglish corpus: parallel rewrites of the same news articles at three reading levels (Elementary, Intermediate, Advanced). The Advanced level approximates the original adult-directed native article; Elementary and Intermediate are simplifications produced by skilled editors for language learners. All three are produced by native speakers and remain grammatical; the ordering reflects their derivational relationship from the Advanced source, not a quality judgement.

Annotations

No per-sentence annotation was solicited. Labels are assigned by source provenance, following the rationale above. Within-corpus labels (WMT06 fluency scores, CEFR proficiency tags, OSE reading levels) are inherited from the source datasets.

WMT06 fluency scores are inherently noisy: on the 999 multiply-annotated sentences (5.6 % of WMT06 data; avg. 2.07 annotations/sentence), exact inter-annotator agreement ranges from 62.1 % (score 1) to 66.9 % (score 3). The ordinal framework partially absorbs this noise via distance-weighted training objectives.

Preprocessing

• Sentence tokenisation (NLTK) applied uniformly to all sources, so that OneStopEnglish (whose raw entries are paragraphs or articles) is brought to the same granularity as WMT06 and W&I (already sentence-level).
• Length filter: sentences of 5–60 words retained. This removes 1,781 sentences (2.6 %) and substantially reduces length confounds across sources.
• Stratified 80/20 train/test split on the 8-level label.

Considerations for using the data

Confound audit

Because the spectrum is assembled from heterogeneous corpora, a model could in principle achieve high ordinal correlation by classifying source rather than naturalness. Three of the strongest candidate shortcuts have been audited (see article for details):

Confound	Probe	Finding
Sentence length	Distribution overlap after filtering	Substantial overlap across all 8 levels
Semantic content	Pairwise cosine in all-mpnet-base-v2 space, within- vs. cross-source	Within−cross gap ≤ +0.06, distributions overlap heavily
Affective content	Pairwise cosine in a GoEmotions-trained SupCon encoder (foudil/lens-emotion-encoder)	Within−cross gap ≤ +0.06, distributions overlap heavily

This bounds — but does not eliminate — the space of plausible shortcuts. Surface features not measured here (punctuation, capitalisation, MT-pipeline orthographic artefacts) could carry residual signal. Some of these arguably are part of the naturalness signal under the working definition; others may be confounds.

Known limitations and biases

• English only. Cross-lingual transfer is not supported by this dataset.

• Register coverage. Spans MT output, classroom learner writing, and graded news prose. Literary, poetic, and rhetorically ornate registers are not represented.

• Native-band side-signal. The OneStopEnglish ordering co-varies with lexical and syntactic sophistication. Models trained on this dataset may learn sophistication as a side-signal aligned with high-naturalness scores. This conflates simple-but-natural prose with simplification artefacts.

• MT label noise. WMT06 fluency judgements have moderate inter-annotator agreement (~62–67 %). The MT/Mediocre-MT boundary is noisier than the Learner and Native boundaries.

• Non-uniform spacing. The integer label scale assigns unit distance between all adjacent steps, but actual naturalness gaps are much larger between macro-bands (MT/Learner/Native) than within. Models recover this empirically; raw label arithmetic on the integer scale should be avoided.

• Demographic representation. Inherits the demographic skew of the source corpora (W&I learners are predominantly European; OneStopEnglish is UK-focused news; WMT06 is European-language news translation).

• Reddit, social-media, and dialectal English are absent. The spectrum does not cover informal native registers.

Personal and sensitive information

The constituent corpora are publicly released for research. No new personal information is introduced. W&I+LOCNESS learner essays were collected under the original BEA-2019 shared-task consent; no demographic identifiers beyond CEFR proficiency are retained here.

Additional information

Source data licences and attribution

This dataset re-uses material derived from three publicly available corpora. Each source remains governed by its original licence or terms of use. Users must comply with the source terms in addition to any terms attached to this aggregated release.

• WMT06: Koehn & Monz, “Manual and Automatic Evaluation of Machine Translation between European Languages,” WMT 2006. The WMT06 shared-task page describes the data as publicly available for machine translation and MT evaluation research.
• W&I+LOCNESS: Bryant, Felice, Andersen & Briscoe, “The BEA-2019 Shared Task on Grammatical Error Correction,” BEA 2019. The BEA-2019 release states that W&I+LOCNESS and related corpora are subject to non-commercial-use licence terms.
• OneStopEnglish: Vajjala & Lučić, “OneStopEnglish corpus: A new corpus for automatic readability assessment and text simplification,” BEA 2018. The corpus is released under CC BY-SA 4.0.

Because the constituent corpora have different terms, this aggregated dataset is not intended to relicense the source corpora. Downstream users must preserve attribution to all three sources and comply with each source’s licence or terms of use.

Citation information

If you use this dataset, please cite both the aggregating article and the three constituent corpora:

@inproceedings{koehn-monz-2006-manual,
  title     = {Manual and Automatic Evaluation of Machine Translation between {E}uropean Languages},
  author    = {Koehn, Philipp and Monz, Christof},
  booktitle = {Proceedings on the Workshop on Statistical Machine Translation},
  year      = {2006},
  month     = jun,
  address   = {New York City},
  publisher = {Association for Computational Linguistics},
  pages     = {102--121},
  url       = {https://aclanthology.org/W06-3114/}
}

@inproceedings{bryant-etal-2019-bea,
  title     = {The {BEA}-2019 Shared Task on Grammatical Error Correction},
  author    = {Bryant, Christopher and Felice, Mariano and Andersen, {\O}istein E. and Briscoe, Ted},
  booktitle = {Proceedings of the Fourteenth Workshop on Innovative Use of NLP for Building Educational Applications},
  year      = {2019},
  month     = aug,
  address   = {Florence, Italy},
  publisher = {Association for Computational Linguistics},
  pages     = {52--75},
  doi       = {10.18653/v1/W19-4406},
  url       = {https://aclanthology.org/W19-4406/}
}

@inproceedings{vajjala-lucic-2018-onestopenglish,
  title     = {{OneStopEnglish} corpus: A new corpus for automatic readability assessment and text simplification},
  author    = {Vajjala, Sowmya and Lu{\v{c}}i{\'c}, Ivana},
  booktitle = {Proceedings of the Thirteenth Workshop on Innovative Use of NLP for Building Educational Applications},
  year      = {2018},
  month     = jun,
  address   = {New Orleans, Louisiana},
  publisher = {Association for Computational Linguistics},
  pages     = {297--304},
  doi       = {10.18653/v1/W18-0535},
  url       = {https://aclanthology.org/W18-0535/}
}

Models trained on this dataset

• foudil/lens-naturalness-encoder — DeBERTa-v3-base + LoRA, ordinal SupCon fine-tuned, 128-d embedding space where cosine similarity = naturalness proximity.