RocketQAv2: A Joint Training Method for Dense Passage Retrieval and Passage Re-ranking

TL;DR

This work tackles the challenge of jointly training dense passage retrievers and passage re-rankers. It introduces RocketQAv2, which uses dynamic listwise distillation to align the relevance distributions of a dual-encoder retriever and a cross-encoder re-ranker, optimized with a loss $\ ext{L}_{\text{final}} = \\text{L}_{\text{KL}} + \\text{L}_{\text{sup}}$ and guided by soft labels. A hybrid data augmentation strategy generates diverse, hard-negative-rich training lists to support robust listwise learning. Empirical results on MSMARCO and Natural Questions show substantial gains over strong baselines, demonstrating effective end-to-end training and mutual improvement between components.

Abstract

In various natural language processing tasks, passage retrieval and passage re-ranking are two key procedures in finding and ranking relevant information. Since both the two procedures contribute to the final performance, it is important to jointly optimize them in order to achieve mutual improvement. In this paper, we propose a novel joint training approach for dense passage retrieval and passage re-ranking. A major contribution is that we introduce the dynamic listwise distillation, where we design a unified listwise training approach for both the retriever and the re-ranker. During the dynamic distillation, the retriever and the re-ranker can be adaptively improved according to each other's relevance information. We also propose a hybrid data augmentation strategy to construct diverse training instances for listwise training approach. Extensive experiments show the effectiveness of our approach on both MSMARCO and Natural Questions datasets. Our code is available at https://github.com/PaddlePaddle/RocketQA.

Figures (4)

• Figure 1: The illustration of dynamic listwise distillation in our approach.
• Figure 2: The illustration of hybrid data augmentation.
• Figure 3: The overall joint training architecture of RocketQAv2.
• Figure 4: MRR@10 results of passage retrieval and passage re-ranking with different numbers of instances per query on MSMARCO. Note that instances per query contain one positive instance, and the rest are hard negatives.