qe-vs-human

Human quality judgments of English–Ukrainian sentence pairs, collected to validate automatic Quality Estimation (QE) metrics against expert assessment.

This is the human side of the evaluation. It is the reference data behind two papers: the ensemble QE study (UNLP 2025) and the professional-translator reliability study (UNLP 2026). The machine side — 55 million pairs scored by six QE models — is released separately as lang-uk/FiftyFiveShades.

What is in it

19,955 judgments in a single all-data.csv. Each row is one rating of one sentence pair by one evaluator.

Two evaluator populations and two rating protocols are combined in the same table, distinguished by evaluator_type:

• Holistic — a single overall quality score on a continuous 0–100 scale.
• Two-stage — fluency and adequacy rated separately, also 0–100.

The 2025 study drew a stratified sample of 9,755 pairs rated by linguistics students. The 2026 study added eight professional translators rating 1,000 pairs each, split across the two protocols, with a repeated block for test–retest reliability. Timing columns are present because response latency turned out to matter: experts got measurably faster and more lenient over a session.

Columns

Column	Description
id, hash	Identifiers for the sentence pair
source	Provenance of the pair within the parent parallel corpus
en, uk	The English source and the Ukrainian translation being judged
evaluator	Anonymised evaluator identifier
evaluator_type	Which population and protocol this row belongs to
answer	Holistic quality score, 0–100
fluency	Fluency score, 0–100, in the two-stage protocol
adequacy	Adequacy score, 0–100, in the two-stage protocol
z_score	answer standardised per evaluator
z_score_fluency, z_score_adequacy	The two-stage scores standardised per evaluator
start_time, end_time	Timestamps bounding the individual judgment
duration	Seconds spent on the judgment
Unnamed: 0	Leftover pandas index from export; safe to ignore

Rows carry answer or the fluency/adequacy pair depending on protocol, not both — filter on evaluator_type before aggregating, or the two scales will be averaged together.

Loading

from datasets import load_dataset

ds = load_dataset("lang-uk/qe-vs-human", split="train")
print(ds[0])

import pandas as pd

df = pd.read_csv(
    "hf://datasets/lang-uk/qe-vs-human/all-data.csv", index_col=0
)
print(df.groupby("evaluator_type")["answer"].describe())

What it is useful for

• Validating QE metrics against human judgment on a mid-resource language pair
• Measuring inter-rater reliability and test–retest stability in MT evaluation
• Studying rater drift, since latency is recorded per judgment
• Calibrating LLM-as-a-judge setups against expert scores
• Filtering large parallel corpora using a human-anchored quality threshold

Findings this data supports

Professionals using the single holistic scale agreed with each other more than professionals using separate fluency and adequacy scales — the opposite of the usual expectation that decomposing a judgment improves agreement. Adequacy correlates strongly with holistic quality; fluency behaves as a largely independent dimension. The best ensemble of automatic metrics explained roughly 60% of the variance in expert ratings.

Citation

If you use qe-vs-human in your work, please cite:

@inproceedings{chaplynskyi-zakharov-2025-framework,
    title     = "A Framework for Large-Scale Parallel Corpus Evaluation: Ensemble Quality Estimation Models Versus Human Assessment",
    author    = "Chaplynskyi, Dmytro and
      Zakharov, Kyrylo",
    editor    = "Romanyshyn, Mariana",
    booktitle = "Proceedings of the Fourth Ukrainian Natural Language Processing Workshop (UNLP 2025)",
    month     = jul,
    year      = "2025",
    address   = "Vienna, Austria (online)",
    publisher = "Association for Computational Linguistics",
    url       = "https://aclanthology.org/2025.unlp-1.9/",
    doi       = "10.18653/v1/2025.unlp-1.9",
    pages     = "73--85",
    ISBN      = "979-8-89176-269-5"
}

@inproceedings{chaplynskyi-etal-2026-professional,
    title     = "Professional Translators Versus Quality Estimation Models: Reliability and Agreement in {E}nglish-{U}krainian Translation Evaluation",
    author    = "Chaplynskyi, Dmytro and
      Zakharov, Kyrylo and
      Ivashkevych, Lesia",
    editor    = "Romanyshyn, Mariana",
    booktitle = "Proceedings of the Fifth {U}krainian Natural Language Processing Conference ({UNLP} 2026)",
    month     = may,
    year      = "2026",
    address   = "Lviv, Ukraine",
    publisher = "Association for Computational Linguistics",
    url       = "https://aclanthology.org/2026.unlp-1.10/",
    pages     = "97--107",
    ISBN      = "979-8-89176-359-3"
}

Paper: https://aclanthology.org/2025.unlp-1.9/ DOI: https://doi.org/10.18653/v1/2025.unlp-1.9

Maintained by lang-uk - Dmytro Chaplynskyi (ORCID)