🤖 AI Summary
Current dense passage retrieval (DPR) training paradigms typically rely on a single positive passage per query, overlooking the realistic many-to-one semantic relationship between a query and multiple relevant passages. This work presents the first systematic investigation into multi-positive training for DPR and proposes a multi-positive contrastive learning framework built upon the dual-encoder architecture: for each query, it jointly optimizes representations against all annotated positive passages, thereby strengthening negative contrast and enhancing semantic discriminability. Experiments demonstrate that our method significantly improves retrieval accuracy—achieving MRR@10 gains of 1.2–2.8 points—on standard benchmarks including MSMARCO and Natural Questions, without increasing GPU memory consumption. It maintains training stability even under small batch sizes, enabling efficient single-GPU training. The core contribution lies in empirically validating and realizing the effectiveness and practicality of multi-positive supervision for dense retrieval modeling.
📝 Abstract
By leveraging a dual encoder architecture, Dense Passage Retrieval (DPR) has outperformed traditional sparse retrieval algorithms such as BM25 in terms of passage retrieval accuracy. Recently proposed methods have further enhanced DPR's performance. However, these models typically pair each question with only one positive passage during training, and the effect of associating multiple positive passages has not been examined. In this paper, we explore the performance of DPR when additional positive passages are incorporated during training. Experimental results show that equipping each question with multiple positive passages consistently improves retrieval accuracy, even when using a significantly smaller batch size, which enables training on a single GPU.