Dataset Card for DivEMT

For more details on DivEMT, see our EMNLP 2022 Paper and our Github repository

Gabriele Sarti • Arianna Bisazza • Ana Guerberof Arenas • Antonio Toral

We introduce DivEMT, the first publicly available post-editing study of Neural Machine Translation (NMT) over a typologically diverse set of target languages. Using a strictly controlled setup, 18 professional translators were instructed to translate or post-edit the same set of English documents into Arabic, Dutch, Italian, Turkish, Ukrainian, and Vietnamese. During the process, their edits, keystrokes, editing times and pauses were recorded, enabling an in-depth, cross-lingual evaluation of NMT quality and post-editing effectiveness. Using this new dataset, we assess the impact of two state-of-the-art NMT systems, Google Translate and the multilingual mBART-50 model, on translation productivity. We find that post-editing is consistently faster than translation from scratch. However, the magnitude of productivity gains varies widely across systems and languages, highlighting major disparities in post-editing effectiveness for languages at different degrees of typological relatedness to English, even when controlling for system architecture and training data size. We publicly release the complete dataset including all collected behavioral data, to foster new research on the translation capabilities of NMT systems for typologically diverse languages.

Dataset Summary

This dataset contains the processed warmup and main splits of the DivEMT dataset. A sample of documents extracted from the Flores-101 corpus were either translated from scratch or post-edited from an existing automatic translation by a total of 18 professional translators across six typologically diverse languages (Arabic, Dutch, Italian, Turkish, Ukrainian, Vietnamese). During the translation, behavioral data (keystrokes, pauses, editing times) were collected using the PET platform.

We publicly release the processed dataset including all collected behavioural data, to foster new research on the ability of state-of-the-art NMT systems to generate text in typologically diverse languages.

News 🎉

February, 2023: The DivEMT dataset now contains linguistic annotations (*_annotations fields) computed with Stanza and word-level quality estimation tags (src_wmt22_qe, mt_wmt22_qe) obtained using the same scripts adopted for the WMT22 QE Task 2.

Languages

The language data of DivEMT is in English (BCP-47 en), Italian (BCP-47 it), Dutch (BCP-47 nl), Arabic (BCP-47 ar), Turkish (BCP-47 tr), Ukrainian (BCP-47 uk) and Vietnamese (BCP-47 vi)

Dataset Structure

Data Instances

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

Data Fields

The following fields are contained in the training set:

Field	Description
unit_id	The full entry identifier. Format: flores101-{config}-{lang}-{doc_id}-{modality}-{sent_in_doc_num}
flores_id	Index of the sentence in the original Flores-101 dataset
item_id	The sentence identifier. The first digits of the number represent the document containing the sentence, while the last digit of the number represents the sentence position inside the document. Documents can contain from 3 to 5 contiguous sentences each.
subject_id	The identifier for the translator performing the translation from scratch or post-editing task. Values: t1, t2 or t3.
lang_id	Language identifier for the sentence, using Flores-101 three-letter format (e.g. ara, nld)
doc_id	Document identifier for the sentence
task_type	The modality of the translation task. Values: ht (translation from scratch), pe1 (post-editing Google Translate translations), pe2 (post-editing mBART 1-to-50 translations).
translation_type	Either ht for from scratch or pe for post-editing
src_len_chr	Length of the English source text in number of characters
mt_len_chr	Length of the machine translation in number of characters (NaN for ht)
tgt_len_chr	Length of the target text in number of characters
src_len_wrd	Length of the English source text in number of words
mt_len_wrd	Length of the machine translation in number of words (NaN for ht)
tgt_len_wrd	Length of the target text in number of words
edit_time	Total editing time for the translation in seconds.
k_total	Total number of keystrokes for the translation.
k_letter	Total number of letter keystrokes for the translation.
k_digit	Total number of digit keystrokes for the translation.
k_white	Total number of whitespace keystrokes for the translation.
k_symbol	Total number of symbol (punctuation, etc.) keystrokes for the translation.
k_nav	Total number of navigation keystrokes (left-right arrows, mouse clicks) for the translation.
k_erase	Total number of erase keystrokes (backspace, cancel) for the translation.
k_copy	Total number of copy (Ctrl + C) actions during the translation.
k_cut	Total number of cut (Ctrl + X) actions during the translation.
k_paste	Total number of paste (Ctrl + V) actions during the translation.
k_do	Total number of Enter actions during the translation.
n_pause_geq_300	Number of pauses of 300ms or more during the translation.
len_pause_geq_300	Total duration of pauses of 300ms or more, in milliseconds.
n_pause_geq_1000	Number of pauses of 1s or more during the translation.
len_pause_geq_1000	Total duration of pauses of 1000ms or more, in milliseconds.
event_time	Total time summed across all translation events, should be comparable to edit_time in most cases.
num_annotations	Number of times the translator focused the textbox for performing the translation of the sentence during the translation session. E.g. 1 means the translation was performed once and never revised.
n_insert	Number of post-editing insertions (empty for modality ht) computed using the tercom library.
n_delete	Number of post-editing deletions (empty for modality ht) computed using the tercom library.
n_substitute	Number of post-editing substitutions (empty for modality ht) computed using the tercom library.
n_shift	Number of post-editing shifts (empty for modality ht) computed using the tercom library.
tot_shifted_words	Total amount of shifted words from all shifts present in the sentence.
tot_edits	Total of all edit types for the sentence.
hter	Human-mediated Translation Edit Rate score computed between MT and post-edited TGT (empty for modality ht) using the tercom library.
cer	Character-level HTER score computed between MT and post-edited TGT (empty for modality ht) using CharacTER.
bleu	Sentence-level BLEU score between MT and post-edited TGT (empty for modality ht) computed using the SacreBLEU library with default parameters.
chrf	Sentence-level chrF score between MT and post-edited TGT (empty for modality ht) computed using the SacreBLEU library with default parameters.
time_s	Edit time expressed in seconds.
time_m	Edit time expressed in minutes.
time_h	Edit time expressed in hours.
time_per_char	Edit time per source character, expressed in seconds.
time_per_word	Edit time per source word, expressed in seconds.
key_per_char	Proportion of keys per character needed to perform the translation.
words_per_hour	Amount of source words translated or post-edited per hour.
words_per_minute	Amount of source words translated or post-edited per minute.
per_subject_visit_order	Id denoting the order in which the translator accessed documents. 1 correspond to the first accessed document.
src_text	The original source sentence extracted from Wikinews, wikibooks or wikivoyage.
mt_text	Missing if tasktype is ht. Otherwise, contains the automatically-translated sentence before post-editing.
tgt_text	Final sentence produced by the translator (either via translation from scratch of sl_text or post-editing mt_text)
aligned_edit	Aligned visual representation of REF (mt_text), HYP (tl_text) and edit operations (I = Insertion, D = Deletion, S = Substitution) performed on the field. Replace \\n with \n to show the three aligned rows.
src_tokens	List of tokens obtained tokenizing src_text with Stanza using default params.
src_annotations	List of lists (one per src_tokens token) containing dictionaries (one per word, >1 for mwt) with pos, ner and other info parsed by Stanza
mt_tokens	List of tokens obtained tokenizing mt_text with Stanza using default params.
mt_annotations	List of lists (one per mt_tokens token) containing dictionaries (one per word, >1 for mwt) with pos, ner and other info parsed by Stanza
tgt_tokens	List of tokens obtained tokenizing tgt_text with Stanza using default params.
tgt_annotations	List of lists (one per tgt_tokens token) containing dictionaries (one per word, >1 for mwt) with pos, ner and other info parsed by Stanza

Data Splits

config	train
main	7740 (107 docs i.e. 430 sents x 18 translators)
warmup	360 (5 docs i.e. 20 sents x 18 translators)

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 t1 post-editing a machine translation produced by Google Translate (task_type pe1) taken from the train split for Turkish. The field aligned_edit is showed over three lines to provide a visual understanding of its contents.

{
    'unit_id': 'flores101-main-tur-46-pe1-3',
    'flores_id': 871,
    'item_id': 'flores101-main-463',
    'subject_id': 'tur_t1',
    'task_type': 'pe1',
    'translation_type': 'pe',
    'src_len_chr': 109,
    'mt_len_chr': 129.0,
    'tgt_len_chr': 120,
    'src_len_wrd': 17,
    'mt_len_wrd': 15.0,
    'tgt_len_wrd': 13,
    'edit_time': 11.762999534606934,
    'k_total': 31,
    'k_letter': 9,
    'k_digit': 0,
    'k_white': 0,
    'k_symbol': 0,
    'k_nav': 20,
    'k_erase': 2,
    'k_copy': 0,
    'k_cut': 0,
    'k_paste': 0,
    'k_do': 0,
    'n_pause_geq_300': 2,
    'len_pause_geq_300': 4986,
    'n_pause_geq_1000': 1,
    'len_pause_geq_1000': 4490,
    'event_time': 11763,
    'num_annotations': 2,
    'last_modification_time': 1643569484,
    'n_insert': 0.0,
    'n_delete': 2.0,
    'n_substitute': 1.0,
    'n_shift': 0.0,
    'tot_shifted_words': 0.0,
    'tot_edits': 3.0,
    'hter': 20.0,
    'cer': 0.10,
    'bleu': 0.0,
    'chrf': 2.569999933242798,
    'lang_id': 'tur',
    'doc_id': 46,
    'time_s': 11.762999534606934,
    'time_m': 0.1960500031709671,
    'time_h': 0.0032675000838935375,
    'time_per_char': 0.1079174280166626,
    'time_per_word': 0.6919412016868591,
    'key_per_char': 0.2844036817550659,
    'words_per_hour': 5202.75439453125,
    'words_per_minute': 86.71257019042969,
    'per_subject_visit_order': 201,
    'src_text': 'As one example, American citizens in the Middle East might face different situations from Europeans or Arabs.',
    'mt_text': "Bir örnek olarak, Orta Doğu'daki Amerikan vatandaşları, Avrupalılardan veya Araplardan farklı durumlarla karşı karşıya kalabilir.",
    'tgt_text': "Örneğin, Orta Doğu'daki Amerikan vatandaşları, Avrupalılardan veya Araplardan farklı durumlarla karşı karşıya kalabilir.",
    'aligned_edit': "REF:  bir örnek olarak,  orta doğu'daki amerikan vatandaşları, avrupalılardan veya araplardan farklı durumlarla karşı karşıya kalabilir.\\n
                     HYP:  *** ***** örneğin, orta doğu'daki amerikan vatandaşları, avrupalılardan veya araplardan farklı durumlarla karşı karşıya kalabilir.\\n
                     EVAL: D   D     S"
}

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

Dataset Creation

The dataset was parsed from PET XML files into CSV format using the scripts available in the DivEMT Github repository.

Those are adapted from the ones by Antonio Toral found at the following link: https://github.com/antot/postediting_novel_frontiers.

Additional Information

Dataset Curators

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

Citation Information

@inproceedings{sarti-etal-2022-divemt,
    title = "{D}iv{EMT}: Neural Machine Translation Post-Editing Effort Across Typologically Diverse Languages",
    author = "Sarti, Gabriele  and
      Bisazza, Arianna  and
      Guerberof-Arenas, Ana  and
      Toral, Antonio",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-main.532",
    pages = "7795--7816",
}