Dense Passage Retriever

DPR is a dual-encoder dense retrieval model — separate BERT encoders for questions and passages, trained with a contrastive objective so that a question’s vector lands near the passage that answers it. Published by Karpukhin et al. (EMNLP 2020) at Facebook AI Research, it was one of the results that established dense retrieval as a serious alternative to lexical search for open-domain question answering.

In this vault it matters less as an architecture — it is a Bi-Encoder, structurally — than as the canonical cautionary example of in-domain success that does not transfer.


The Cautionary Result

DPR is trained on Natural Questions. As reported in Improving Zero-Shot Ranking with Vespa Hybrid Search:

  • In-domain on NQ: strong performance
  • Zero-shot on MS MARCO: underperforms BM25

And the pattern generalizes — the BEIR leaderboard shows dense embedding models trained on NQ substantially underperforming BM25 across nearly all BEIR datasets.

The two corpora are not exotic relative to one another. NQ is Wikipedia passages with 9.2-word queries; MS MARCO is web-search passages with 5.9-word queries. Similar enough that transfer ought to work, different enough that it doesn’t. That narrowness of the gap is the point: domain shift does not have to be dramatic to break a single-vector model.

Why a Dual Encoder Is Exposed

DPR pools each passage into one fixed vector, offline, with no knowledge of the query it will eventually be matched against. The compression forces a training-time decision about which distinctions to keep — made against the training query distribution. Move the distribution and the discarded distinctions may be the ones now needed.

This is the general argument developed in Zero-Shot Retrieval and Bi-Encoder, with architectures that defer matching to query time (ColBERT, Late Interaction) or stay in term space (Learned Sparse Retrieval) proposed as the more transferable alternatives.

Datasets

Articles

Source

  • Dense Passage Retrieval for Open-Domain Question Answering — Karpukhin, Oğuz, Min, Lewis, Wu, Edunov, Chen, Yih (EMNLP 2020), arXiv:2004.04906