RRRA: Resampling and Reranking through a Retriever Adapter

BongsuKim, Seungho Choi, Sihyun Park, Minsang Kim, Chansol Park

Keywords: Dense Retrieval, Hard Negative Sampling, False Negative Modeling, Resampling and Reranking

TL;DR: We propose RRRA, a dense retrieval framework with a learnable adapter to detect false negatives. By reweighting top-k hard negatives in training and reranking at inference, RRRA improves robustness and precision over Bi-Encoder baselines.

Abstract: In dense retrieval, effective training hinges on selecting high-quality hard negatives while avoiding false negatives. Recent methods apply heuristics based on positive document scores to identify hard negatives, improving both performance and interpretability. However, these global, example-agnostic strategies often miss instance-specific false negatives. To address this, we propose a learnable adapter module that monitors Bi-Encoder representations to estimate whether a hard negative is likely to be a false negative. This probability is modeled in a dynamic and context-aware manner, enabling fine-grained, query-specific judgments. The predicted scores are used in two downstream components: (1) resampling, by reweighting the top-k hard negatives used in training, and (2) reranking, where top-k retrieved documents are reordered at inference. Empirical results on standard benchmarks show that our adapter-enhanced framework consistently outperforms existing Bi-Encoder–based negative sampling baselines, underscoring the benefit of explicit false negative modeling in dense retrieval.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 23643