Dense Passage Retrieval for Open-Domain Question Answering

Traditional open-domain question answering (QA) systems often rely on sparse retrieval methods like BM25 to find relevant documents. While effective, these methods struggle with semantic matching—such as identifying that 'bad guy' and 'villain' are related. This paper investigates whether a dense retrieval system, which uses learned vector representations, can outperform sparse methods without requiring complex pretraining objectives.

The authors propose the Dense Passage Retriever (DPR), which uses a dual-encoder architecture based on BERT. The system maps questions and passages into a shared, continuous vector space. During training, the model is optimized to maximize the inner product between the vectors of a question and its corresponding relevant passage. To make training efficient and effective, the authors utilize 'in-batch negatives,' where positive passages from other questions in the same training batch serve as negative examples for the current question.

DPR significantly outperforms BM25 across a range of open-domain QA benchmarks. For instance, on the Natural Questions dataset, DPR achieves a top-20 retrieval accuracy of 78.4%, compared to 59.1% for BM25. The authors also show that the system is sample-efficient; even with only 1,000 training examples, the dense retriever surpasses the performance of traditional keyword-based systems. Furthermore, they demonstrate that higher retrieval precision directly translates to improved end-to-end QA accuracy, establishing new state-of-the-art results on several benchmarks.

This work simplifies the pipeline for open-domain QA by showing that high-quality retrieval can be achieved through straightforward fine-tuning of pretrained language models. By proving that dense representations can effectively capture semantic relationships that sparse models miss, the authors provide a robust, scalable framework for information retrieval that does not require the computationally expensive pretraining previously thought necessary for dense retrieval.