Noise-Robust Dense Retrieval via Contrastive Alignment Post Training
This addresses robustness issues in dense retrieval systems for noisy query scenarios, though it is an incremental improvement over existing methods.
The paper tackles the problem of dense retrieval models being brittle to noisy queries by proposing CAPOT, a post-training fine-tuning method that improves robustness without requiring index regeneration or training set optimization. The method achieved similar impact to data augmentation on noisy variants of MSMARCO, Natural Questions, and Trivia QA passage retrieval.
The success of contextual word representations and advances in neural information retrieval have made dense vector-based retrieval a standard approach for passage and document ranking. While effective and efficient, dual-encoders are brittle to variations in query distributions and noisy queries. Data augmentation can make models more robust but introduces overhead to training set generation and requires retraining and index regeneration. We present Contrastive Alignment POst Training (CAPOT), a highly efficient finetuning method that improves model robustness without requiring index regeneration, the training set optimization, or alteration. CAPOT enables robust retrieval by freezing the document encoder while the query encoder learns to align noisy queries with their unaltered root. We evaluate CAPOT noisy variants of MSMARCO, Natural Questions, and Trivia QA passage retrieval, finding CAPOT has a similar impact as data augmentation with none of its overhead.