GAND encompasses gender ambiguous (w.r.t. a specific referent) natural data and is a benchmarking resource for evaluating gender in machine translation.
Dataset split:
train: 4037 examples
validation: 505 examples
test: 505 examples
To build GAND, data from two different sources were used: C4 (Common Crawl) and Open Subtitles.
- C4: We used the en subset from the allenai/c4 dataset, which is available on Hugging Face. C4 contains 364,868,892 texts (with 456 tokens per text on average in the first 1,000 texts of the randomly shuffled dataset).
- Open Subtitles: We used the monolingual English data from the Open Subtitles corpus, compiled within the OPUS project. The Open Subtitles corpus contains 2,739,528 texts (in this case, a text corresponds to a single subtitle; 8 tokens per subtitle on average in the first 1,000 subtitles of the randomly shuffled dataset).
More information on the construction of GAND is available on this GitHub repository: https://github.com/jhacken/GAND/
GAND encompasses purely gender-ambiguous (w.r.t. to a specific referent) natural data and is a benchmarking resource for evaluating gender (bias) in machine translation.
To build GAND, data from two different sources were used: C4 (Common Crawl) and Open Subtitles.
C4: We used the en subset from the allenai/c4 dataset, which is available on Hugging Face. C4 contains 364,868,892 texts (with 456 tokens per text on average in the first 1,000 texts of the randomly shuffled dataset).
Open Subtitles: We used the monolingual English data from the Open Subtitles corpus, compiled within the OPUS project. The Open Subtitles corpus contains 2,739,528 texts (in this case, a text corresponds to a single subtitle; 8 tokens per subtitle on average in the first 1,000 subtitles of the randomly shuffled dataset).
More information on the specific construction of GAND can be found in the GitHub repository: https://github.com/jhacken/GAND/
