Coco Racial Bias Eval

coco-racial-bias-eval

Understanding and Evaluating Racial Biases in Image Captioning — Zhao et al. (2021) (arXiv:2106.08503, 2021)

What this evaluates

This protocol evaluates racial bias in image captioning models by measuring performance disparities between images containing lighter-skinned versus darker-skinned individuals. It probes whether models systematically generate lower-quality captions or exhibit different linguistic patterns for darker-skinned subjects compared to lighter-skinned ones, even when visual content is controlled.

Datasets

• COCO 2014 validation — total 10969; splits: test (10969)

Metrics

• CIDEr (primary) — range: [0, 1]
  • Computes TF-IDF weighted n-gram similarity between generated and reference captions. Higher values indicate better alignment with the human consensus. The paper reports the difference in scores between lighter and darker image subsets.
• BLEU — range: [0, 1]
  • Measures precision of n-gram matches between generated and reference captions. The paper reports the difference in scores between lighter and darker image subsets.
• SPICE — range: [0, 1]
  • Focuses on semantic scene graph matching (objects, attributes) rather than n-grams. The paper reports the difference in scores between lighter and darker image subsets.

Input / output format

Input: RGB image (or extracted CNN features) paired with a ground-truth caption.

Output: A single natural language caption string describing the image.

Scoring recipe

def compute_bias_disparity(predictions, golds, skin_tone_labels):
    lighter_mask = skin_tone_labels == 'lighter'
    darker_mask = skin_tone_labels == 'darker'
    score_lighter = compute_metric(predictions[lighter_mask], golds[lighter_mask])
    score_darker = compute_metric(predictions[darker_mask], golds[darker_mask])
    disparity = score_lighter - score_darker
    return disparity

Common pitfalls

• Reporting only aggregate metric scores without stratifying by skin tone, which masks underlying disparities.
• Assuming semantic metrics like SPICE will capture all bias, as the paper shows SPICE differences are negligible while n-gram metrics reveal significant gaps.
• Failing to control for visual content differences when analyzing caption vocabulary, leading to confounding effects from scene composition rather than skin tone.

Evidence (verbatim from paper)

To do so, we first assess the differences in BLEU [55], CIDEr [68] and SPICE [2] scores between captions on lighter and darker images. Both BLEU and CIDEr rely on n-gram matching with BLEU measuring precision and CIDEr the similarity between the generated caption and the "consensus" of manual captions.

Citation

@misc{zhao2021racialbias,
  title={Understanding and Evaluating Racial Biases in Image Captioning},
  author={Zhao et al. (2021)},
  year={2021},
  note={arXiv:2106.08503}
}

• arXiv: 2106.08503