ConvAI2 Inferred: Persona-Based Dialogue Data with Gender Bias Labels

ConvAI2 contains persona-based conversations used for gender bias analysis. The dataset is part of the Multi-Dimensional Gender Bias Classification framework and includes inferred labels for gender of the speaker, addressee, and subject. It was curated by authors including Alexander H. Miller and others from Facebook AI Research.

Use Cases

• Training classifiers to predict gender bias categories based on inferred binary and ternary labels.
• Analyzing bias from the gender of the speaker (AS) and addressee (TO) in dialogue data.
• Benchmarking model performance against the reported 67.13% accuracy baseline for binary gender prediction.
• Studying persona-based conversations for pragmatic and semantic dimensions of bias.

Strengths

• Dataset is part of a published framework for multi-dimensional gender bias classification.
• Labels include binary and ternary classifications with associated confidence scores.
• Based on the ConvAI2 dataset, which contains structured persona-based conversations.

Limitations

• Row count, file formats, and column-level documentation are unknown.
• Gender labels are imputed by a classifier, which may introduce inference errors.
• Last update date is unknown; freshness unverified.

Provenance

Source

Facebook AI Research, derived from the ConvAI2 dataset.

Collection Method

Persona-based conversations were annotated, and gender labels were imputed using a classifier trained on related datasets.

Time Range

null

Freshness

null

Geography

null