Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mitigating the Impact of False Negatives in Dense Retrieval with Contrastive Confidence Regularization

Shiqi Wang, Yeqin Zhang, Cam-Tu Nguyen

TL;DR

This work tackles false negatives in dense retrieval for open-domain QA by extending the Noise Contrastive Estimation framework with a Contrastive Confidence Regularizer (CCR) inspired by CORES$^2$. The authors prove theoretical guarantees (Theorems 1–2) and develop a model-agnostic Passage Sieve to filter noisy negatives using a CCR-trained model, improving downstream retrieval performance. The approach is demonstrated on three datasets (NQ, TQ, MS-pas) and shown to outperform state-of-the-art dense retrievers, with robustness to limited batch size and fewer negatives. The methods are applicable to any NCE-based dense retriever, offering practical improvements in retrieval accuracy and training efficiency for QA systems.

Abstract

In open-domain Question Answering (QA), dense retrieval is crucial for finding relevant passages for answer generation. Typically, contrastive learning is used to train a retrieval model that maps passages and queries to the same semantic space. The objective is to make similar ones closer and dissimilar ones further apart. However, training such a system is challenging due to the false negative issue, where relevant passages may be missed during data annotation. Hard negative sampling, which is commonly used to improve contrastive learning, can introduce more noise in training. This is because hard negatives are those closer to a given query, and thus more likely to be false negatives. To address this issue, we propose a novel contrastive confidence regularizer for Noise Contrastive Estimation (NCE) loss, a commonly used loss for dense retrieval. Our analysis shows that the regularizer helps dense retrieval models be more robust against false negatives with a theoretical guarantee. Additionally, we propose a model-agnostic method to filter out noisy negative passages in the dataset, improving any downstream dense retrieval models. Through experiments on three datasets, we demonstrate that our method achieves better retrieval performance in comparison to existing state-of-the-art dense retrieval systems.

Mitigating the Impact of False Negatives in Dense Retrieval with Contrastive Confidence Regularization

TL;DR

This work tackles false negatives in dense retrieval for open-domain QA by extending the Noise Contrastive Estimation framework with a Contrastive Confidence Regularizer (CCR) inspired by CORES. The authors prove theoretical guarantees (Theorems 1–2) and develop a model-agnostic Passage Sieve to filter noisy negatives using a CCR-trained model, improving downstream retrieval performance. The approach is demonstrated on three datasets (NQ, TQ, MS-pas) and shown to outperform state-of-the-art dense retrievers, with robustness to limited batch size and fewer negatives. The methods are applicable to any NCE-based dense retriever, offering practical improvements in retrieval accuracy and training efficiency for QA systems.

Abstract

In open-domain Question Answering (QA), dense retrieval is crucial for finding relevant passages for answer generation. Typically, contrastive learning is used to train a retrieval model that maps passages and queries to the same semantic space. The objective is to make similar ones closer and dissimilar ones further apart. However, training such a system is challenging due to the false negative issue, where relevant passages may be missed during data annotation. Hard negative sampling, which is commonly used to improve contrastive learning, can introduce more noise in training. This is because hard negatives are those closer to a given query, and thus more likely to be false negatives. To address this issue, we propose a novel contrastive confidence regularizer for Noise Contrastive Estimation (NCE) loss, a commonly used loss for dense retrieval. Our analysis shows that the regularizer helps dense retrieval models be more robust against false negatives with a theoretical guarantee. Additionally, we propose a model-agnostic method to filter out noisy negative passages in the dataset, improving any downstream dense retrieval models. Through experiments on three datasets, we demonstrate that our method achieves better retrieval performance in comparison to existing state-of-the-art dense retrieval systems.
Paper Structure (39 sections, 1 theorem, 28 equations, 4 figures, 6 tables, 1 algorithm)

This paper contains 39 sections, 1 theorem, 28 equations, 4 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Dense Retrieval
  4. Label-Noise Robust Machine Learning
  5. Preliminaries
  6. Problem Formalization
  7. Dual-Encoders for Dense Retrieval
  8. Peer Loss and Confidence Regularization
  9. Contrastive Confidence Regularizer
  10. Analysis on Simulated Data
  11. Theoretical Analysis
  12. Passage Sieve Method
  13. Experiment
  14. Experimental Setup
  15. Datasets
  16. ...and 24 more sections

Key Result

Lemma 1

Algorithm algo:passage_sieve ensures that a negative sample $p^-$ in hard negatives will NOT be selected into the sieved dataset $\mathcal{D}^*$ if its score $f(p^-,q)$ given by the model f (eq. eq:nce) is more than a random guess, i.e. its similarity score after softmax is bigger than the average v

Figures (4)

  • Figure 1: The effect of the contrastive confidence regularizer. Four figures show analysis from four experiments with all the same settings but with different values of beta. All of them are continuously trained based on the same pre-trained vanilla DPR. The y-axis value represents the NCE loss value between queries and different types of passages. It shows that normal DPR training will cause different distributions to be squeezed together while $\ell_{CCR}$ has the potential to help distinguish false negatives from real negatives.
  • Figure 2: Performance of AR2+passage sieve with different hyper-parameter $\beta$ on NQ test sets. Overall, the passage sieve algorithm is not sensitive to $\beta$.
  • Figure 3: Sieve out rate during AR2+passage sieve training on NQ dataset
  • Figure 4: Learning curve of MS-pas development set and TQ test set.

Theorems & Definitions (3)

  • Lemma 1
  • proof
  • proof