The idea in plain language.
Train question and passage encoders to place matching text close together.
How it works
One encoder represents a question and another represents candidate passages. Training brings a question closer to passages that contain its answer and separates it from negatives. At retrieval time, a vector index returns passages with similar embeddings. A downstream answerer can then use that retrieved context. The paper compares this approach with a strong sparse retrieval baseline.
What to keep in mind
Retrieval can miss relevant passages or return misleading context. The quality and coverage of the indexed corpus remain central to answer quality.
Source: Dense Passage Retrieval for Open-Domain Question Answering. The original manuscript contains the methods, experiments, figures, and references. An arXiv posting date may follow an earlier conference publication. Read the linked record for version history.