ImageChat: Conversations About Image Contents with Gender Bias Annotations

ImageChat contains conversations discussing the contents of an image, annotated for gender bias. The dataset is part of the Multi-Dimensional Gender Bias Classification framework, which decomposes bias along pragmatic and semantic dimensions. It was created by Xiang Zhang, Junbo Zhao, and Yann LeCun, with data sourced from HuggingFace and the Courant Institute of Mathematical Sciences.

Use Cases

• Train a model for gender bias classification based on the annotated pragmatic dimensions described.
• Analyze conversational patterns about image contents based on the described caption and gender fields.
• Benchmark model performance in binary gender prediction tasks, as referenced by the 67.13% accuracy baseline.
• Study the relationship between image descriptions and gendered language in dialogue.

Strengths

• Includes a gender column derived from explicit male and female boolean annotations.
• Part of a larger framework with seven automatically annotated datasets for gender bias analysis.
• Based on a published methodology from a 2020 Transformer model benchmark.

Limitations

• Row count, file formats, and column details beyond the description are unknown.
• Last update date is unknown; freshness unverified.
• The description suggests the dataset may be focused on gendered examples, which could limit generalizability.

Provenance

Source

HuggingFace and the Courant Institute of Mathematical Sciences

Collection Method

Conversations were generated using an image captioning system to identify image contents and select gendered examples.

Time Range

null

Freshness

null

Geography

null