GAP is a gender-balanced dataset containing 8,908 coreference-labeled pairs of (ambiguous pronoun, antecedent name), sampled from Wikipedia and released by Google AI Language for the evaluation of coreference resolution in practical applications.

http://goo.gl/language/gap-coreference

Coreference resolution is an important task for natural language understanding and the resolution of ambiguous pronouns a longstanding challenge. Nonetheless, existing corpora do not capture ambiguous pronouns in sufficient volume or diversity to accurately indicate the practical utility of models.

Google AI Language's GAP dataset is an evaluation benchmark comprising 8,908 coreference-labeled pairs of (ambiguous pronoun, antecedent name), sampled from Wikipedia to provide diverse coverage of challenges posed by real-world text. Importantly, GAP is gender-balanced to address the gender bias in coreference systems noted in our and other's analysis.

More details are available in our paper (which should be cited if you use or discuss GAP in your work):

@inproceedings{webster2018gap,
  title =     {Mind the GAP: A Balanced Corpus of Gendered Ambiguou},
  author =    {Webster, Kellie and Recasens, Marta and Axelrod, Vera and Baldridge, Jason},
  booktitle = {Transactions of the ACL},
  year =      {2018},
  pages =     {to appear},
}

The GAP dataset release comprises three .tsv files, each with eleven columns.

The files are:

• test 4,000 pairs, to be used for official evaluation
• development 4,000 pairs, may be used for model development
• validation 908 pairs, may be used for parameter tuning

The columns contain:

^ Please note that systems should detect mentions for inference automatically, and access labeled spans only to output predictions.

^^ Please also note that there are two task settings, snippet-context in which the URL column may not be used, and page-context where the URL, and the denoted Wikipedia page, may be used.

Performance on GAP may be benchmarked against the syntactic parallelism baseline from our above paper on the test set:

where the metrics are F1 score on Masculine and Feminine examples, Overall, and a Bias factor calculated as F / M.

To contact us, please use [email protected]