End-to-End Training of Multi-Document Reader and Retriever for Open-Domain Question Answering

We present an end-to-end differentiable training method for retrieval-augmented open-domain question answering systems that combine information from multiple retrieved documents when generating answers.