Research map / Label-efficient and robust learning / Recognition and long-tail

TriSampler: A Better Negative Sampling Principle for Dense Retrieval

AAAI 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Recommendation representation learning · 8 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes the quasi-triangular principle for negative sampling in dense retrieval and TriSampler, which samples negatives constrained to a triangular-like region around the query and positive document, improving AR2, RocketQA and ANCE on passage and document retrieval benchmarks.

Problem

Negative sampling for dense retrieval lacks a guiding principle for constructing negative candidates and designing sampling distributions; hard negatives risk false negatives while easy negatives give little signal.

Approach

A quasi-triangular principle constrains negatives so that query-negative similarity approximates query-positive similarity and negatives are closer to the positive than to the query; TriSampler builds candidates under this constraint and samples them with two distributions (a Gaussian-like similarity-match weight and a ReLU-based positive-proximity weight).

Key results

• AR2 + TriSampler reaches 86.5 R@20 / 90.7 R@100 on NQ and 41.4 MRR@10 on MS MARCO passage, above AR2 (86.0 / 90.1 / 39.5).
• On MS MARCO document dev, AR2 + TriSampler achieves 43.8 MRR@100 and 93.1 R@100, above AR2+SimANS (43.1 / 92.3).
• ANCE + TriSampler improves MS MARCO passage MRR@10 from 33.0 to 35.8 and RocketQA + TriSampler from 37.0 to 38.3.

Measured effects reported in the paper

Surprising findings

• TriSampler's per-batch training cost is slightly higher than SimANS, but faster convergence reduces total training time to optimal performance.
• TriSampler does not surpass LexMAE, which uses a reranker-distilled third stage.

Limitations

• Negative candidate construction adds preprocessing cost (110s vs 85s for SimANS on NQ).
• The triangular angle threshold of 60 degrees is set by design without a reported sweep in the text.
• Does not surpass LexMAE, which uses an off-the-shelf reranker for distillation.

Experimental setup

Datasets: MS MARCO passage Natural Questions TriviaQA MS MARCO document

Benchmarks: Natural Questions TriviaQA MS MARCO passage MS MARCO document

Metrics: MRR@10 Recall@20 Recall@100 Recall@50 Recall@1k

Baselines: BM25 DPR ANCE RocketQA ADORE SimANS ANCE-Tele coCondenser SimLM LexMAE STAR

Builds on: AR2 SimANS ANCE RocketQA DPR

Closest papers by meaning

• QuRe: Query-Relevant Retrieval through Hard Negative Sampling in Composed Image Retrieval (ICML 2025 · similarity 0.87)
• Mitigating the Impact of False Negative in Dense Retrieval with Contrastive Confidence Regularization (AAAI 2024 · similarity 0.87)
• Worse than Zero-shot? A Fact-Checking Dataset for Evaluating the Robustness of RAG Against Misleading Retrievals (NeurIPS 2025 · similarity 0.85)
• POQD: Performance-Oriented Query Decomposer for Multi-vector retrieval (ICML 2025 · similarity 0.84)
• Distribution-Driven Dense Retrieval: Modeling Many-to-One Query-Document Relationship (AAAI 2025 · similarity 0.84)
• SePer: Measure Retrieval Utility Through The Lens Of Semantic Perplexity Reduction (ICLR 2025 · similarity 0.84)