Dense Passage Retrieval for Open-Domain Question Answering

Inside the paper

The problem

Open-domain question answering usually first retrieves a small set of passages and then asks a reader to extract an answer. Keyword retrieval can miss relevant passages when the question and evidence use different words.

The idea

The authors train separate neural encoders for questions and passages so that useful pairs have similar dense representations. Precomputed passage vectors make it possible to search the collection before running a more expensive reader.

How it works

• Encode each question and passage into a vector using separate BERT-based encoders, and score a pair with their inner product.
• Train on questions paired with relevant passages, contrasting each positive passage with negative examples, including passages associated with other questions in the batch.
• Encode the passage collection ahead of time and use maximum inner product search to retrieve candidates for a new question.
• A reader examines the retrieved passages and predicts an answer span, allowing the authors to test the retriever within a complete question-answering system.

What it showed

The paper reports higher passage retrieval accuracy than its BM25 baseline across the evaluated question-answering datasets. It also reports end-to-end answer accuracy and investigates training choices, including the contribution of negative passages.

Read it for

Read Section 3 for the encoders and training objective, then Tables 2 and 3 in Section 5 for retrieval results and training comparisons. Section 6 covers the reader; the dataset inventory in Section 4 can be skimmed initially.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.