New Metric Measures Directional Bias Amplification in Image Captioning
Researchers propose Directional Bias Amplification in Captioning (DBAC), a metric that identifies how image captioning models amplify biases from training data. Unlike prior metrics like Leakage in Captioning (LIC), DBAC pinpoints the source of bias amplification and is less sensitive to sentence encoder choice. The work addresses a gap in bias measurement for captioning datasets, where classification-focused metrics fail to capture language semantics.
Key facts
- DBAC is a language-aware and directional metric for bias amplification in image captioning.
- It improves upon LIC by identifying the source of bias amplification.
- DBAC is less sensitive to sentence encoders than LIC.
- Existing metrics like BA (MALS) and DPA are ineffective for captioning datasets.
- Bias amplification occurs when models worsen biases present in training data.
- The research is published on arXiv with identifier 2503.07878.
- The paper is a replacement cross, indicating a revised version.
Entities
Institutions
- arXiv