ECI: Effective Contrastive Information to Evaluate Hard-Negatives
Researchers have developed a new training-free method called Effective Contrastive Information (ECI) to evaluate hard-negative sources for dense retrieval systems. This diagnostic tool ranks candidate negative sources using frozen target-encoder embeddings, bypassing the need for fine-tuning and downstream evaluation. ECI's effectiveness was demonstrated on MS MARCO negative sources, where it successfully identified the highest-ranking negatives and showed stability across various perturbations. AI
IMPACT Introduces a novel diagnostic for improving dense retrieval models by optimizing negative sampling.