Dense Passage Retrieval for Open-Domain Question Answering

Open-domain QA retrieval used TF-IDF/BM25, which fails on semantic paraphrase and requires hand-tuned sparse statistics.

A dual encoder (two independent BERT towers, no token-level interaction) maps a question and a passage to a single dense vector; retrieval is nearest-neighbour search over the passage index with a contrastive in-batch-negatives objective. Document embeddings are computed OFFLINE once per corpus; only the query needs an encoder pass at query time.

Outperforms a strong Lucene-BM25 system by 9%-19% absolute top-20 passage retrieval accuracy on a range of open-domain QA datasets, and lifts end-to-end QA to new state of the art on multiple benchmarks. Baseline (Lucene-BM25) and metric (top-20 accuracy) are both stated; no hardware or latency disclosed.

Venue EMNLP 2020 (arXiv comment). DPR is the template for essentially every dense retriever since: separate encoders, in-batch negatives, offline document embedding. THE ARCHITECTURAL SERVING ARGUMENT: because documents are embedded OFFLINE, the serving cost of a dense index is (a) a one-time corpus embedding pass and (b) a query encoder forward plus an ANN scan - the expensive transformer never runs per document at query time. That asymmetry versus cross-encoders, which re-encode every candidate, is why dense retrieval won. The memory cost is the index, which is why dimensionality-reduction work (pap-mm-matryoshka) matters so much here. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass): venue_track set to 'main' from the ACL Anthology bib index (anthology.bib, 131650 entries) by normalized title match. Evidence: anthology id 2020.emnlp-main.550, booktitle 'Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)'. The Anthology id is the authoritative signal here and it is unambiguous, so this is an additive clarification of a track, not a change of venue - `venue` is untouched and no status change is implied. The reason this matters for inference readers: a Findings paper and a main-track paper at the same venue are not the same kind of evidence, and until this field existed the repo could not say which was which.