Quality Estimation for Post-Editing (QE4PE)

For more details on QE4PE, see our paper and our Github repository

Gabriele Sarti • Vilém Zouhar • Malvina Nissim • Grzegorz Chrupała • Ana Guerberof Arenas • Arianna Bisazza

Abstract TBD

Dataset Summary

This dataset provides a convenient access to the processed main and pretask splits of the QE4PE dataset. A sample of challenging documents extracted from WMT23 evaluation data were machine translated from English to Italian and Dutch using NLLB 3.3B, and post-edited by 12 translators per direction across 4 highlighting modalities employing various word-level quality estimation (QE) strategies to present translators with potential errors during the editing. Additional details are provided in the main task readme and in our paper. During the post-editing, behavioral data (keystrokes, pauses and editing times) were collected using the GroTE online platform.

We publicly release the granular editing logs alongside the processed dataset to foster new research on the usability of word-level QE strategies in modern post-editing workflows.

News 📢

October 2024: The QE4PE dataset is released on the HuggingFace Hub! 🎉

Repository Structure

The repository is organized as follows:

qe4pe/
├── questionnaires/ # Configs and results for pre- and post-task questionnaires for translators
│   ├── pretask_results.csv # Results of the pretask questionnaire, corresponding to the `pretask_questionnaire` configuration
│   ├── posttask_highlight_results.csv # Results of the posttask questionnaire for highlighted modalities, corresponding to the `posttask_highlight_questionnaire` configuration
│   ├── posttask_no_highlight_results.csv # Results of the posttask questionnaire for the `no_highlight` modality, corresponding to the `posttask_no_highlight_questionnaire` configuration
│   └── ... # Configurations reporting the exact questionnaires questions and options.
├── setup/
│   ├── highlights/ # Outputs of word-level QE strategies used to setup highlighted spans in the tasks
│   ├── processed/ # Intermediate outputs of the selection process for the main task
│   └── wmt23/ # Original collection of WMT23 sources and machine-translated outputs
└── task/
    ├── example/ # Example folder with task structure 
    ├── main/ # Main task data, logs, outputs and guidelines
    │   ├── ...
    │   ├── processed_main.csv # Processed main task data, corresponds to the `main` configuration
    │   └── README.md # Details about the main task
    └── pretask/ # Pretask data, logs, outputs and guidelines
        ├── ...
        ├── processed_pretask.csv # Processed pretask data, corresponds to the `pretask` configuration
        └── README.md # Details about the pretask

Languages

The language data of QE4PE is in English (BCP-47 en), Italian (BCP-47 it) and Dutch (BCP-47 nl).

Dataset Structure

Data Instances

The dataset contains two configurations, corresponding to the two tasks: main and pretask. main contains the full data collected during the main task and analyzed during our experiments. pretask contains the data collected in the initial verification phase, before the main task begins.

Data Fields

A single entry in the dataframe represents a segment (~sentence) in the dataset, that was machine-translated and post-edited by a professional translator. The following fields are contained in the training set:

Field	Description
Identification
unit_id	The full entry identifier. Format: qe4pe-{task_id}-{src_lang}-{tgt_lang}-{doc_id}-{segment_in_doc_id}-{translator_main_task_id}.
wmt_id	Identifier of the sentence in the original WMT23 dataset.
wmt_category	Category of the document: biomedical or social
doc_id	The index of the document in the current configuration of the QE4PE dataset containing the current segment.
segment_in_doc_id	The index of the segment inside the current document.
segment_id	The index of the segment in the current configurations (i.e. concatenating all segments from all documents in order)
translator_pretask_id	The identifier for the translator according to the pretask format before modality assignments: tXX.
translator_main_id	The identifier for the translator according to the main task format after modality assignments: {highlight_modality}_tXX.
src_lang	The source language of the segment. For QE4PE, this is always English (eng)
tgt_lang	The target language of the segment: either Italian (ita) or Dutch (nld).
highlight_modality	The highlighting modality used for the segment. Values: no_highlight, oracle, supervised, unsupervised.
Text statistics
src_num_chars	Length of the source segment in number of characters.
mt_num_chars	Length of the machine-translated segment in number of characters.
pe_num_chars	Length of the post-edited segment in number of characters.
src_num_words	Length of the source segment in number of words.
mt_num_words	Length of the machine-translated segment in number of words.
pe_num_words	Length of the post-edited segment in number of words.
Edits statistics
num_words_insert	Number of post-editing insertions computed using jiwer.
num_words_delete	Number of post-editing deletions computed using jiwer.
num_words_substitute	Number of post-editing substitutions computed using jiwer.
num_words_unchanged	Number of post-editing hits computed using jiwer.
tot_words_edits	Total of all edit types for the sentence.
wer	Word Error Rate score computed between mt_text and pe_text using jiwer.
num_chars_insert	Number of post-editing insertions computed using jiwer.
num_chars_delete	Number of post-editing deletions computed using jiwer.
num_chars_substitute	Number of post-editing substitutions computed using jiwer.
num_chars_unchanged	Number of post-editing hits computed using jiwer.
tot_chars_edits	Total of all edit types for the sentence.
cer	Character Error Rate score computed between mt_text and pe_text using jiwer.
Translation quality
mt_bleu_max	Max BLEU score between mt_text and all pe_text for the corresponding segment using SacreBLEU with default parameters.
mt_bleu_min	Min BLEU score between mt_text and all pe_text for the corresponding segment using SacreBLEU with default parameters.
mt_bleu_mean	Mean BLEU score between mt_text and all pe_text for the corresponding segment using SacreBLEU with default parameters.
mt_bleu_std	Standard deviation of BLEU scores between mt_text and all pe_text for the corresponding segment using SacreBLEU with default parameters.
mt_chrf_max	Max chrF score between mt_text and all pe_text for the corresponding segment using SacreBLEU with default parameters.
mt_chrf_min	Min chrF score between mt_text and all pe_text for the corresponding segment using SacreBLEU with default parameters.
mt_chrf_mean	Mean chrF score between mt_text and all pe_text for the corresponding segment using SacreBLEU with default parameters.
mt_chrf_std	Standard deviation of chrF scores between mt_text and all pe_text for the corresponding segment using SacreBLEU with default parameters.
mt_ter_max	Max TER score between mt_text and all pe_text for the corresponding segment using SacreBLEU with default parameters.
mt_ter_min	Min TER score between mt_text and all pe_text for the corresponding segment using SacreBLEU with default parameters.
mt_ter_mean	Mean TER score between mt_text and all pe_text for the corresponding segment using SacreBLEU with default parameters.
mt_ter_std	Standard deviation of TER scores between mt_text and all pe_text for the corresponding segment using SacreBLEU with default parameters.
mt_comet_max	Max COMET sentence-level score for the mt_text and all pe_text for the corresponding segment using Unbabel/wmt22-comet-da with default parameters.
mt_comet_min	Min COMET sentence-level score for the mt_text and all pe_text for the corresponding segment using Unbabel/wmt22-comet-da with default parameters.
mt_comet_mean	Mean COMET sentence-level score for the mt_text and all pe_text for the corresponding segment using Unbabel/wmt22-comet-da with default parameters.
mt_comet_std	Standard deviation of COMET sentence-level scores for the mt_text and all pe_text for the corresponding segment using Unbabel/wmt22-comet-da with default parameters.
mt_xcomet_qe	Unbabel/XCOMET-XXL sentence-level quality estimation score for the mt_text.
mt_xcomet_errors	List of error spans detected by Unbabel/XCOMET-XXL for the mt_text.
pe_bleu_max	Max BLEU score between pe_text and all other pe_text for the corresponding segment using SacreBLEU with default parameters.
pe_bleu_min	Min BLEU score between pe_text and all other pe_text for the corresponding segment using SacreBLEU with default parameters.
pe_bleu_mean	Mean BLEU score between pe_text and all other pe_text for the corresponding segment using SacreBLEU with default parameters.
pe_bleu_std	Standard deviation of BLEU scores between pe_text and all other pe_text for the corresponding segment using SacreBLEU with default parameters.
pe_chrf_max	Max chrF score between pe_text and all other pe_text for the corresponding segment using SacreBLEU with default parameters.
pe_chrf_min	Min chrF score between pe_text and all other pe_text for the corresponding segment using SacreBLEU with default parameters.
pe_chrf_mean	Mean chrF score between pe_text and all other pe_text for the corresponding segment using SacreBLEU with default parameters.
pe_chrf_std	Standard deviation of chrF scores between pe_text and all other pe_text for the corresponding segment using SacreBLEU with default parameters.
pe_ter_max	Max TER score between pe_text and all other pe_text for the corresponding segment using SacreBLEU with default parameters.
pe_ter_min	Min TER score between pe_text and all other pe_text for the corresponding segment using SacreBLEU with default parameters.
pe_ter_mean	Mean TER score between pe_text and all other pe_text for the corresponding segment using SacreBLEU with default parameters.
pe_ter_std	Standard deviation of TER scores between pe_text and all other pe_text for the corresponding segment using SacreBLEU with default parameters.
pe_comet_max	Max COMET sentence-level score for the pe_text and all other pe_text for the corresponding segment using Unbabel/wmt22-comet-da with default parameters.
pe_comet_min	Min COMET sentence-level score for the pe_text and all other pe_text for the corresponding segment using Unbabel/wmt22-comet-da with default parameters.
pe_comet_mean	Mean COMET sentence-level score for the pe_text and all other pe_text for the corresponding segment using Unbabel/wmt22-comet-da with default parameters.
pe_comet_std	Standard deviation of COMET sentence-level scores for the pe_text and all other pe_text for the corresponding segment using Unbabel/wmt22-comet-da with default parameters.
pe_xcomet_qe	Unbabel/XCOMET-XXL sentence-level quality estimation score for the pe_text.
pe_xcomet_errors	List of error spans detected by Unbabel/XCOMET-XXL for the pe_text.
mt_pe_bleu	BLEU score computed between mt_text and pe_text using SacreBLEU with default parameters.
mt_pe_chrf	chrF score computed between mt_text and pe_text using SacreBLEU with default parameters.
mt_pe_comet	COMET sentence-level score for the mt_text and pe_text using COMET
mt_xcomet_xxl_qe	XCOMET-XXL sentence-level quality estimation score for the mt_text.
pe_xcomet_xxl_qe	XCOMET-XXL sentence-level quality estimation score for the pe_text.
Behavioral data
doc_num_edits	Total number of edits performed by the translator on the current document. Only the last edit outputs are considered valid.
doc_edit_order	Index corresponding to the current document edit order. If equal to doc_id, the document was edited in the given order.
doc_edit_time	Total editing time for the current document in seconds (from start to end, no times ignored)
doc_edit_time_filtered	Total editing time for the current document in seconds (from start to end, >5m pauses between logged actions ignored)
doc_keys_per_min	Keystrokes per minute computed for the current document using doc_edit_time_filtered.
doc_chars_per_min	Characters per minute computed for the current document using doc_edit_time_filtered.
doc_words_per_min	Words per minute computed for the current document using doc_edit_time_filtered.
segment_num_edits	Total number of edits performed by the translator on the current segment. Only edits for the last edit of the doc are considered valid.
segment_edit_order	Index corresponding to the current segment edit order (only first enter action counts). If equal to segment_in_doc_id, the segment was edited in the given order.
segment_edit_time	Total editing time for the current segment in seconds (summed time between enter-exit blocks)
segment_edit_time_filtered	Total editing time for the current segment in seconds (>5m pauses between logged actions ignored).
segment_keys_per_min	Keystrokes per minute computed for the current segment using segment_edit_time_filtered.
segment_chars_per_min	Characters per minute computed for the current segment using segment_edit_time_filtered.
segment_words_per_min	Words per minute computed for the current segment using segment_edit_time_filtered.
num_enter_actions	Number of enter actions (focus on textbox) performed by the translator on the current segment during post-editing.
remove_highlights	If True, the Clear Highlights button was pressed for this segment (always false for no_highlight modality).
Texts and annotations
src_text	The original source segment from WMT23 requiring translation.
mt_text	Output of the NLLB-3.3B model when translating src_text into tgt_lang (default config, 5 beams)
mt_text_highlighted	Highlighted version of mt_text with potential errors according to the highlight_modality.
pe_text	Post-edited version of mt_text produced by a professional translator with highlight_modality.
mt_pe_word_aligned	Aligned visual representation of word-level edit operations (I = Insertion, D = Deletion, S = Substitution) (replace \\n with \n to show the three aligned rows).
mt_pe_char_aligned	Aligned visual representation of character-level edit operations (I = Insertion, D = Deletion, S = Substitution) (replace \\n with \n to show the three aligned rows).
highlights	List of dictionaries for highlighted spans with error severity and position, matching XCOMET format for word-level error annotations.

Data Splits

config	split
main	train	8100 (51 docs i.e. 324 sents x 25 translators)
pretask	train	950 (6 docs i.e. 38 sents x 25 translators)
pretask_questionnaire	train	26 (all translators, including replaced/replacements)
posttask_highlight_questionnaire	train	19 (all translators for highlight modalities + 1 replacement)
posttask_no_highlight_questionnaire	train	6 (all translators for no_highlight modality)

Train Split

The train split contains the totality of triplets (or pairs, when translation from scratch is performed) annotated with behavioral data produced during the translation.

The following is an example of the subject oracle_t1 post-editing for segment 3 of doc20 in the eng-nld direction of the main task. The fields mt_pe_word_aligned and mt_pe_char_aligned are shown over three lines to provide a visual understanding of their contents.

{
    # Identification
    "unit_id": "qe4pe-main-eng-nld-20-3-oracle_t1",
    "wmt_id": "doc5",
    "wmt_category": "biomedical",
    "doc_id": 20,
    "segment_in_doc_id": 3,
    "segment_id": 129,
    "translator_pretask_id": "t4",
    "translator_main_id": "oracle_t1",
    "src_lang": "eng",
    "tgt_lang": "nld",
    "highlight_modality": "oracle",
    # Text statistics
    "src_num_chars": 104,
    "mt_num_chars": 136,
    "pe_num_chars": 106,
    "src_num_words": 15,
    "mt_num_words": 16,
    "pe_num_words": 16,
    # Edits statistics
    "num_words_insert": 0,
    "num_words_delete": 0,
    "num_words_substitute": 1,
    "num_words_unchanged": 15,
    "tot_words_edits": 1,
    "wer": 0.0625,
    "num_chars_insert": 0,
    "num_chars_delete": 0,
    "num_chars_substitute": 6,
    "num_chars_unchanged": 100,
    "tot_chars_edits": 6,
    "cer": 0.0566,
    # Translation quality
    "mt_bleu_max": 100.0,
    "mt_bleu_min": 7.159,
    "mt_bleu_mean": 68.687,
    "mt_bleu_std": 31.287,
    "mt_chrf_max": 100.0,
    "mt_chrf_min": 45.374,
    "mt_chrf_mean": 83.683,
    "mt_chrf_std": 16.754,
    "mt_ter_max": 100.0,
    "mt_ter_min": 0.0,
    "mt_ter_mean": 23.912,
    "mt_ter_std": 29.274,
    "mt_comet_max": 0.977,
    "mt_comet_min": 0.837,
    "mt_comet_mean": 0.94,
    "mt_comet_std": 0.042,
    "mt_xcomet_qe": 0.985,
    "mt_xcomet_errors": "[]",
    "pe_bleu_max": 100.0,
    "pe_bleu_min": 11.644,
    "pe_bleu_mean": 61.335,
    "pe_bleu_std": 28.617,
    "pe_chrf_max": 100.0,
    "pe_chrf_min": 53.0,
    "pe_chrf_mean": 79.173,
    "pe_chrf_std": 13.679,
    "pe_ter_max": 100.0,
    "pe_ter_min": 0.0,
    "pe_ter_mean": 28.814,
    "pe_ter_std": 28.827,
    "pe_comet_max": 0.977,
    "pe_comet_min": 0.851,
    "pe_comet_mean": 0.937,
    "pe_comet_std": 0.035,
    "pe_xcomet_qe": 0.984,
    "pe_xcomet_errors": "[]",
    # Behavioral data
    "doc_num_edits": 103,
    "doc_edit_order": 20,
    "doc_edit_time": 118,
    "doc_edit_time_filtered": 118,
    "doc_keys_per_min": 52.37,
    "doc_chars_per_min": 584.24,
    "doc_words_per_min": 79.83,
    "segment_num_edits": 9,
    "segment_edit_order": 3,
    "segment_edit_time": 9,
    "segment_edit_time_filtered": 9,
    "segment_keys_per_min": 60.0,
    "segment_chars_per_min": 906.67,
    "segment_words_per_min": 106.67,
    "num_enter_actions": 2,
    "remove_highlights": False,
    # Texts and annotations
    "src_text": "The speed of its emerging growth frequently outpaces the development of quality assurance and education.",
    "mt_text": "De snelheid van de opkomende groei is vaak sneller dan de ontwikkeling van kwaliteitsborging en onderwijs.",
    "mt_text_highlighted": "De snelheid van de opkomende groei is vaak <minor>sneller</minor> dan de ontwikkeling van kwaliteitsborging en <major>onderwijs.</major>",
    "pe_text": "De snelheid van de opkomende groei is vaak sneller dan de ontwikkeling van kwaliteitsborging en opleiding.",
    "mt_pe_word_aligned": "MT: De snelheid van de opkomende groei is vaak sneller dan de ontwikkeling van kwaliteitsborging en onderwijs.\n" \
                          "PE: De snelheid van de opkomende groei is vaak sneller dan de ontwikkeling van kwaliteitsborging en opleiding.\n" \
                          "                                                                                                             S",
    "mt_pe_char_aligned": "MT: De snelheid van de opkomende groei is vaak sneller dan de ontwikkeling van kwaliteitsborging en onderwijs.\n" \
                          "PE: De snelheid van de opkomende groei is vaak sneller dan de ontwikkeling van kwaliteitsborging en opleiding.\n" \
                          "                                                                                                     SS SS SS ",
    "highlights": "[{'text': 'sneller', 'severity': 'minor', 'start': 43, 'end': 50}, {'text': 'onderwijs.', 'severity': 'major', 'start': 96, 'end': 106}]"
}

The text is provided as-is, without further preprocessing or tokenization.

Dataset Creation

The datasets were parsed from GroTE inputs, logs and outputs for the QE4PE study, available in this repository. Processed dataframes using the qe4pe process_task_data command. Refer to the QE4PE Github repository for additional details. The overall structure and processing of the dataset were inspired by the DivEMT dataset.

Additional Information

Metric signatures

The following signatures correspond to the metrics reported in the processed dataframes:

# Computed using SacreBLEU: https://github.com/mjpost/sacrebleu
BLEU:   case:mixed|eff:yes|tok:13a|smooth:exp|version:2.3.1
ChrF:   case:mixed|eff:yes|nc:6|nw:0|space:no|version:2.3.1
TER:    case:lc|tok:tercom|norm:no|punct:yes|asian:no|version:2.3.1

# Computed using Unbabel COMET: https://github.com/Unbabel/COMET
Comet:  Python3.11.9|Comet2.2.2|fp32|Unbabel/wmt22-comet-da
XComet: Python3.10.12|Comet2.2.1|fp32|Unbabel/XCOMET-XXL

Dataset Curators

For problems related to this 🤗 Datasets version, please contact me at [email protected].

Citation Information

TODO