Dataset Card for gender-bias-PE data

Dataset Description

The gender-bias-PE dataset contains the post-edits and associated behavioural data of the human-centered experiments presented in the paper: What the Harm? Quantifying the Tangible Impact of Gender Bias in Machine Translation with a Human-centered Study accepted at EMNLP 2024.

The dataset allows to study the impact of gender bias in Machine Translation (MT) via human-centered measures like post-editing effort (i.e. temporal and technical). The English source sentences contained in the dataset have been automatically translated with Google Translate into Italian, Spanish, and German. 88 human participants were tasked with the light post-editing of the MT output, primarly ensure correct gender translation in the target language.

The data represent multilanguage, multiuser, and multidataset conditions. See the associated paper for full details.

Dataset Structure

Data Config

• all: load all the 3 language pairs in one single dataset instance
• en-it: load the en-it portion of the dataset
• en-es: load the en-es portion of the dataset
• en-de: load the en-de portion of the dataset

Source Data

The dataset is created based on a subset of parallel en-* sentences extracted from the existing MT-GenEval corpus. We adapted several target references translation from the original corpus (see Appendix B.1.2 in Savoldi et al., (2024)).

⚠️ The release of the MuST-SHE sentences and post-edits is temporarily suspended is temporarily suspended pending clarification of the new policy adopted by TED for the use of its proprietary data.

Data Fields

lang: target language

• it
• es
• de

dataset: original dataset

• mtgen_un: subset of MTGenEval with unambiguous gender in the source
• mtgen_a: subset of MTGenEval with ambiguous gender in the source

user_type: type of user that carried out the post-edit

• professional: experienced translator
• student: unexperienced user

original_id: unique sentence identifier from the original dataset

gender: gender expressed in the original reference translation and in the post-edited sentence

• F: feminine
• M: masculine

segment: source English sentence

tgt: target reference translation

raw_word_count: number of source words

time_to_edit: time to edit in milliseconds

suggestion: Google Translate output

secs_per_word: time to edit per source word

parsed_time_to_edit: time to edit as duration

last_translation: post-edited output

HTER: sentence-level HTER score

Annotation

The post-edits were carried out following dedicated guidelines, which are available at https://github.com/bsavoldi/post-edit_guidelines

License

The MTGenvEval-based portion of the dataset is released with the same CC-BY-SA-3.0 license as the original corpus.

Citation

The dataset is associated with a paper accepted at EMNLP 2024. Please cite the paper when referencing the gender-bias-pe corpus as:

@misc{savoldi2024whattheharm,
      title={{What the Harm? Quantifying the Tangible Impact of Gender Bias in Machine Translation with a Human-centered Study}}, 
      author={Savoldi, Beatrice  and
      Papi, Sara  and
      Negri, Matteo and
      Guerberof Ana  and
      Bentivogli, Luisa},
      year={2024},
      booktitle = {Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing},
      month = {nov},
      year = {2023},
      address = {Miami},
      publisher = {Association for Computational Linguistics}
}

Contribution

Thanks to @Bsavoldi for adding this dataset.