GENIE: Benchmark Datasets for GENder Inclusive English-to-Hindi Translation

The datasets we used include GAP, KnowRef, GeBioCorpus, WinoMT, WinoGender, StereoSet, Crows-Pair, and EEC. They contribute to the different linguistic and contextual challenges which are essential for the assessment of translation models. 

1. 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 problem solving in practical applications.
2. The KnowRef dataset is a benchmark designed for evaluating coreference resolution and natural language inference (NLI) systems. It highlights the importance of common sense comprehension and world knowledge. It was first presented in a 2019 study and consists of more than 8,000 annotated passages of text with ambiguous pronominal anaphora that are both difficult and represent the language use in everyday situations.
3. The GeBioCorpus dataset is a high-quality, non-synthetic, gender-balanced test set designed for machine translation assessment. It was introduced through the GeBioToolkit, a tool developed to extract multilingual parallel corpora from Wikipedia biographies, which ensures a balanced representation of genders.
4. The WinoMT dataset is a benchmark for assessing gender bias in machine translation (MT) systems, especially when translating from English to grammatical gender languages. It consists of 3,888 English sentences with a gendered pronoun and an occupation. The dataset is carefully balanced between male and female gender, as well as traditional and anti-stereotypical gender roles (e.g., female doctor versus female nurse).
5. The WinoGender dataset is a diagnostic tool designed to assess gender bias in coreference resolution systems. It comprises minimal pairs of sentences that differ only by the gender of a pronoun, allowing the assessment of whether systems exhibit systematic gender biases.
6. StereoSet dataset is a comprehensive English-language dataset designed to measure stereotypical biases in pretrained language models in four domains: gender, profession, race, and religion. It contains around 17,000 sentences assessing model preferences in relation to these biases.
7. The CrowS-Pairs dataset is a benchmark for assessing social biases in masked language models (MLMs). This dataset consists of 1,508 sentence pairs that focus on stereotypes associated with historically disadvantaged communities in the United States. Each pair contains one sentence that reflects a stereotype and another that does not, allowing a comparative evaluation of model biases.
8. The Equity Evaluation Corpus (EEC) dataset identifies biases associated with gender and race in natural language processing (NLP) systems, namely sentiment analysis. The data consists of 8,640 carefully crafted English sentences intended to evaluate the fairness of NLP algorithms.
9. The WinoBias dataset measures gender bias in coreference resolution methods. This dataset consists of 3,160 sentences divided equally between development and testing sets. Each sentence includes two entities referred to by their jobs as well as a gendered pronoun, which allows us to analyze whether coreference systems are biased when associating pronouns to stereotypical or anti-stereotypical vocations.