Source-linked AI summary
RocketQAv2: A Joint Training Method for Dense Passage Retrieval and Passage Re-ranking
Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, Qiaoqiao She, Hua Wu, Haifeng Wang, Ji-Rong Wen
TL;DR
Passage retrieval and re-ranking both affect final performance, yet their usual training procedures are mismatched. RocketQAv2 jointly trains them with dynamic listwise distillation and hybrid data augmentation, and reports effectiveness on MSMARCO and Natural Questions. The approach adaptively transfers relevance information between the two modules during training.
Problem
Retriever and re-ranker training methods differ, limiting straightforward joint optimization despite both modules contributing to final performance.
Method
RocketQAv2 jointly trains both modules with unified listwise learning, dynamic soft-label distillation, and hybrid data augmentation.
Results
Experiments show the approach is effective on both MSMARCO and Natural Questions.
Takeaways & Limitations
The unified architecture supports mutual improvement between the retriever and re-ranker and provides a possibility for end-to-end training of the retrieval architecture.
Takeaways & Limitations
The training procedure can benefit from initializing the retriever and re-ranker with trained models, which slightly improves results.
Abstract
from arXiv · showhide
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.
1 Introduction
Passage retrieval and re-ranking jointly determine retrieval quality, but they are typically trained with different objectives. RocketQAv2 addresses this mismatch through joint listwise training, dynamic distillation, and hybrid data augmentation.
- Dense retrieval represents queries and passages as low-dimensional vectors, while re-ranking subsequently improves retrieved results with a re-ranker.
- The retriever and re-ranker contribute jointly to final performance, motivating correlated training for mutual improvement.
- The retriever is usually trained listwise with in-batch negatives, whereas the re-ranker is commonly trained pointwise or pairwise.
- RocketQAv2 unifies training by applying a listwise approach to both modules and transferring relevance information between them dynamically.
- Dynamic listwise distillation and hybrid data augmentation are the paper's two major technical contributions.
- Extensive experiments evaluate the proposed approach on both MSMARCO and Natural Questions.
2 Related Work
Related work establishes dense retrieval as a strong alternative to sparse retrieval and emphasizes the dependence between retrieval and re-ranking. RocketQAv2 differs by jointly learning the dense retriever and re-ranker in one architecture.
- Dense passage retrieval has outperformed traditional sparse methods such as TF-IDF and BM25 on passage retrieval.
- Dense retrievers are commonly trained either through self-supervised pre-training or by fine-tuning pre-trained language models on labeled data.
- Passage retrieval and re-ranking are highly related and dependent, with each module benefiting from information associated with the other.
- Unlike prior efforts, RocketQAv2 uses a joint learning architecture for the dense passage retriever and the re-ranker.
3 Methodology
RocketQAv2 jointly trains dense retrieval and passage re-ranking through unified listwise learning, dynamic mutual distillation, and hybrid data augmentation. The procedure transfers relevance information between retriever and re-ranker while constructing diverse candidate lists for training.
- 3 Methodology: RocketQAv2 unifies listwise training for the dual-encoder retriever and cross-encoder re-ranker to support joint optimization.Both modules compute relevance scores over candidate passage lists rather than using separate training paradigms.
- 3.2 Dynamic Listwise Distillation: Dynamic listwise distillation adaptively transfers relevance information between the retriever and re-ranker by reducing differences between their relevance distributions.Both modules are updated during distillation instead of freezing the re-ranker, enabling mutual improvement.
- 3.2 Dynamic Listwise Distillation: The re-ranker receives supervised cross-entropy guidance while KL-divergence aligns its listwise relevance distribution with the retriever’s distribution.The final loss combines the KL-divergence and supervised cross-entropy objectives.
- 3.3 Hybrid Data Augmentation: Hybrid data augmentation creates candidate lists from undenoised and denoised instances, including ground-truth positives, randomly sampled hard negatives, and high-confidence denoised positives.The RocketQA retriever retrieves top-n passages, while the RocketQA re-ranker removes low-confidence predicted negatives and identifies high-confidence positives.
- 3.4 Training Procedure: The training procedure initializes the retriever and re-ranker, generates hybrid-augmented training data, performs dynamic listwise distillation, and then applies both models as an inference pipeline.The models can be initialized from trained RocketQA components, which the paper reports can slightly improve the initial optimization stage.
- 3.5 Discussion: Compared with RocketQA, RocketQAv2 adds joint retriever–re-ranker training with soft relevance labels, simplifying distillation and enabling potential end-to-end training of the dense retrieval architecture.The approach reuses RocketQA’s network architecture and important training tricks while adding dynamic listwise distillation.
4 Experiments
This section covers the experimental settings, main results, ablation study, and detailed analysis.
- The experiments begin by describing the experimental settings.
- The paper then reports its main experimental results.
- The analysis includes ablation studies and detailed analysis.
4.1 Experimental Setup
The experiments use MSMARCO and Natural Questions, evaluate retrieval with MRR and Recall@k, and specify ERNIE-based model and training settings.
- Datasets: Experiments use the MSMARCO and Natural Questions datasets for dense passage retrieval and passage re-ranking.
- Evaluation Metrics: Retrieval performance is evaluated with MRR and Recall@k.MRR measures the reciprocal rank of the first positive passage, while Recall@k measures whether top-k passages contain positives.
- Model Specifications: The retriever and re-ranker largely follow ERNIE-2.0 base, using dual-encoder and cross-encoder initializations.
- Implementation Details: Training uses PaddlePaddle on up to 32 NVIDIA Tesla V100 GPUs, with Adam, a 1e-5 learning rate, three epochs, and batch size 96.
- Experimental Results: Table 2 reports passage retrieval results on MSMARCO and Natural Questions, marking the best and second-best results distinctly.
4.2 Results on Passage Retrieval
RocketQAv2 is compared with sparse and dense retrieval baselines on MSMARCO and Natural Questions, where its retriever performs strongly and exceeds PAIR on selected metrics.
- Baselines: The comparison includes traditional and neural-enhanced sparse retrievers alongside dense retrievers such as DPR, ANCE, ColBERT, RocketQA, and PAIR.
- Results: RocketQAv2 retriever and PAIR outperform other baselines by a large margin.
- Results: RocketQAv2 outperforms PAIR on MRR@10 and Recall@5.The paper attributes this to dynamic listwise distillation enabling the retriever to capture re-ranker ability at top ranks.
- Method: RocketQAv2 jointly trains the retriever and re-ranker with listwise training, dynamic listwise distillation, and hybrid data augmentation.
- Results: RocketQAv2 significantly outperforms DPR-E despite using the same ERNIE base backbone.This comparison indicates that the backbone pre-trained language model is not the factor responsible for the improvement.
- Results: Sparse retrievers overall perform worse than dense retrievers, while COIL is the strongest sparse baseline.
4.3 Results on Passage Re-ranking
RocketQAv2 re-ranking is evaluated against multiple baselines on MSMARCO, including different candidate sources and retrieval settings, and achieves strong comparative performance.
- Baselines: The evaluation compares BM25, ColBERT, BERTlarge, RepBERT, Multi-stage, CAKD, ME-BERT, ME-HYBIRD, TFR-BERT, and RocketQA.
- Experimental Setup: RocketQAv2 re-ranking is tested with BM25 top-1000 candidates, RocketQA top-50 candidates, and RocketQAv2 top-50 candidates.
- Results: RocketQAv2 re-ranker significantly outperforms all competitive methods.The paper links this result to joint learning optimized to fit the retriever's relevance distribution through dynamic listwise distillation.
- Results: Using RocketQAv2 re-ranker with RocketQA retrieval results improves performance by 0.9 percentage point over the RocketQA re-ranker.
- Results: Applied to BM25's top 1000 candidates, the RocketQAv2 re-ranker performs significantly better than other base models and comparably to other large models.
4.4 Detailed Analysis
The analysis shows that dynamic listwise distillation and hybrid data augmentation materially affect joint-training performance. Larger instance lists improve both retrieval and re-ranking, while in-batch negatives alone do not.
- Dynamic or Static?: Static distillation underperforms dynamic listwise distillation when the re-ranker is frozen.Dynamic optimization allows relevance information to be learned between the two modules rather than transferred from a fixed teacher.
- Listwise or Pointwise?: Pointwise re-ranker training causes a performance drop compared with listwise training in the joint architecture.Listwise training better simulates the relevance distribution used during dynamic listwise distillation.
- The Effect of Denoised Instances: Removing denoised instances lowers MRR@10, showing that hybrid training data better represents the collection distribution.The improvement is especially observed on metrics evaluating top-ranked passages.
- The Number of Hard Negatives: More instances per query, consisting of one positive and additional hard negatives, improve passage retrieval and passage re-ranking.The study uses instance-list size as a proxy for the number of hard negatives.
- Incorporation of In-batch Negatives: Adding in-batch negatives does not produce further performance improvements during joint training.Although in-batch sampling increases the number of negatives for each query, the experiments report no observed gain.
5 Conclusion
RocketQAv2 jointly trains the dense passage retriever and passage re-ranker through dynamic listwise distillation and hybrid data augmentation. The authors report enhanced mutual improvement, a simplified training process, and an end-to-end training possibility for the retrieval architecture.
- 5 Conclusion: RocketQAv2 jointly trains the retriever and re-ranker using dynamic listwise distillation and hybrid data augmentation.The conclusion identifies these as the approach’s two important technical contributions.
- 5 Conclusion: The approach enhances mutual improvement between the retriever and re-ranker while simplifying training.The conclusion attributes both effects to the joint-training approach.
- 5 Conclusion: The unified architecture provides the possibility of training the entire retrieval architecture end to end.The authors describe this as the first joint training of the retriever and re-ranker in a unified architecture, to their knowledge.