LIGHT-Inferred: Yelp Reviews with Inferred Gender Labels for Bias Analysis

A text dataset from the Multi-Dimensional Gender Bias Classification project, containing Yelp reviews with automatically inferred gender labels. The dataset is part of a framework for decomposing gender bias along pragmatic and semantic dimensions, including bias from the speaker, the person spoken to, and the person spoken about. It was curated by authors including Alexander H. Miller and Jason Weston, with a model achieving 67.13% accuracy in binary gender prediction tasks.

Use Cases

• Training classifiers for gender bias detection based on inferred labels for the speaker, listener, and subject.
• Evaluating model performance on gender prediction tasks using the provided binary and ternary labels.
• Analyzing patterns of gender bias in conversational or review text as part of a multi-dimensional framework.

Strengths

• Provides labels for three distinct gender bias dimensions (ABOUT, TO, AS) based on a published framework.
• Includes confidence scores (binary_score, ternary_score) for the inferred gender labels.
• Model performance benchmark is provided, with a Transformer model achieving 67.13% accuracy.

Limitations

• Row count, file formats, and column details are unknown, limiting suitability assessment.
• Gender labels are inferred by a classifier rather than human-annotated, which may introduce errors.
• Last update date is unknown; freshness unverified.

Provenance

Source

Facebook AI Research (FAIR) and HuggingFace.

Collection Method

Text data from Yelp reviews with gender labels inferred by a classifier trained on other datasets.

Time Range

null

Freshness

Last updated date is unknown.

Geography

null