Improving Text Embedding Fairness: Training-Free Bias Correction

This research explores a novel method for mitigating bias in text embeddings, a critical area for fair AI development. The training-free approach offers a potential advantage in terms of efficiency and ease of implementation.

• •Focuses on correcting mean bias, a common issue in text embeddings.
• •Employs a training-free renormalization technique, offering practical advantages.
• •Evaluates the approach using the MMTEB benchmark.