Dense Passage Retrieval for Open-Domain Question Answering

This paper tackles passage retrieval for open-domain question answering, where the standard approach had been sparse vector space methods such as TF-IDF or BM25. The authors show that retrieval can instead be implemented purely with dense representations, learning question and passage embeddings from only a small number of question-passage pairs using a simple dual-encoder framework.

Evaluated across a wide range of open-domain QA datasets, the dense retriever outperforms a strong Lucene-BM25 baseline by 9% to 19% absolute in top-20 passage retrieval accuracy. This improved retrieval in turn helped the end-to-end QA system establish new state-of-the-art results on multiple open-domain QA benchmarks, showing that learned dense retrieval can practically replace long-dominant sparse methods.