# Dense Passage Retrieval for Open-Domain Question Answering

**Authors:** Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Ledell Wu, Sergey Edunov, Danqi Chen, Wen-tau Yih  
**arXiv:** [2004.04906](https://arxiv.org/abs/2004.04906)  
**Date:** 2020-04-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2004.04906v1/

## Key points

- Paper introduces a novel approach to open-domain question answering
- Dense retriever outperforms Lucene-BM25 system by 9%-19% in top-20 passage retrieval accuracy
- Effectiveness of approach evaluated across various open-domain QA datasets
- End-to-end QA system incorporating dense retriever achieves state-of-the-art results on multiple benchmarks
- Innovative method improves passage retrieval efficiency and highlights potential of dense representations in enhancing overall QA system performance

## Summary

The paper "Dense Passage Retrieval for Open-Domain Question Answering" by Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Ledell Wu, Sergey Edunov, Danqi Chen and Wen-tau Yih introduces a novel approach to open-domain question answering. Traditional methods rely on sparse vector space models like TF-IDF or BM25 for passage retrieval. However, the authors show that efficient retrieval can be achieved using dense representations alone. By utilizing embeddings learned from a small set of questions and passages through a dual-encoder framework, they develop a dense retriever that outperforms a strong Lucene-BM25 system by 9%-19% in top-20 passage retrieval accuracy. The study evaluates the effectiveness of their approach across various open-domain QA datasets and demonstrates its superiority over existing systems. Notably, the end-to-end QA system incorporating the dense retriever achieves state-of-the-art results on multiple benchmarks. This innovative method not only improves passage retrieval efficiency in open-domain question answering but also highlights the potential of dense representations in enhancing overall QA system performance. The findings presented in this work contribute significantly to advancing research in natural language processing and information retrieval.