Improving Contrastive Learning of Sentence Embeddings with Case-Augmented Positives and Retrieved Negatives

Summary: Ranking has always been a top concern in information retrieval research, with the lexical matching signal dominating the ad-hoc retrieval process. However, relying solely on this signal can lead to vocabulary mismatch. In recent years, Dense Retrieval (DR) models have gained attention for better ranking performance, although their improvement heavily depends on sampling methods. Researchers have proposed two new training strategies, STAR and ADORE, which have shown significant improvements over existing methods in optimizing ranking performance.