Dense Passage Retrieval paper

Definition: Dense Passage Retrieval for Open-Domain Question Answering presents DPR, a retriever that encodes questions and text passages separately as numerical vectors. It evaluates whether retrieved passages contain answers and how retrieval affects a reader that extracts an answer from those passages.

Key facts

• The ACL Anthology record identifies Vladimir Karpukhin and colleagues as the authors and dates the EMNLP paper to November 2020.
• The paper's experimental setup uses an English Wikipedia snapshot from 20 December 2018, split into non-overlapping passages of 100 words.
• As of 17 October 2026, the accompanying DPR repository documentation provides separate workflows for encoder training, passage embedding generation and retrieval validation.

How does the dual encoder retrieve passages?

A dual encoder uses separate neural networks to transform questions and passages into vectors, ordered lists of numerical values. In the paper's architecture, both networks start from BERT, a pretrained language model. Each produces a vector with 768 dimensions using the representation of the special [CLS] token.

The scoring rule below expresses the paper's equation in plain text. E_Q encodes a question, E_P encodes a passage, and the dot product multiplies corresponding vector components and adds the products.

sim(q, p) = dot(E_Q(q), E_P(p))

Passage vectors are computed ahead of a question and indexed with FAISS, a vector similarity search library. At query time, DPR encodes the question and retrieves passages with high dot-product scores. A separate reader then examines candidate passages to extract an answer span.

How are the encoders trained?

A positive passage supplies relevant evidence for a training question. Negative passages provide competing examples. The training section optimizes the probability of the positive passage relative to the negatives through a negative log-likelihood loss.

In-batch negatives reuse passages paired with other questions in the same training batch. The paper's selected setup combines these with one additional negative per question retrieved by BM25, a term-based ranking method. That additional passage matches question terms while excluding the answer.

The repository's training data format makes the distinction concrete: positive_ctxs, negative_ctxs and hard_negative_ctxs hold passage pools alongside question and answers. Positive passages include title and text fields; the negative pools use the same passage structure.

What does the evaluation measure?

Top-k retrieval accuracy is the fraction of questions for which at least one of the highest-ranked k passages contains an answer. In the paper's test results, the model trained separately on Natural Questions achieves 78.4% top-20 accuracy, compared with 59.1% for BM25. On SQuAD, the corresponding scores are 63.2% for DPR and 68.8% for BM25. The comparison therefore depends on the dataset.

The accompanying validation code finds the earliest answer-containing passage for each question and counts a hit at that rank and every larger cutoff. Its string-matching option compares token sequences without case distinctions; another option accepts regular expressions.

Answer extraction receives a separate evaluation. The same code computes exact match by normalizing predicted and reference answers, including lowercasing and removing punctuation and English articles. Retrieval accuracy and answer exact match describe different stages of the system.

Related terms

• Dense retrieval
• Sparse retrieval
• Text embeddings
• Document chunking
• BEIR benchmark

FAQ

Can you reproduce the paper with any DPR checkpoint?

The repository README distinguishes the original model from a March 2021 release trained with additional negatives mined using DPR itself. Comparing results requires identifying the checkpoint, training data and evaluation split.

Does the paper measure citations of business websites?

The paper's experiments evaluate passage retrieval and extractive question answering using a fixed Wikipedia collection. Citation frequency for business websites falls outside that evaluation.

Sources

1. ACL Anthology, "Dense Passage Retrieval for Open-Domain Question Answering", November 2020
2. Association for Computational Linguistics, "Dense Passage Retrieval for Open-Domain Question Answering", November 2020
3. Facebook Research, "Dense Passage Retrieval: README", undated
4. Facebook Research, "DPR answer validation utilities", undated

---

Check how a page reads to answer engines. Open the GEO score analyzer