Mutually improved dense retriever and GNN-based reader for arbitrary-hop open-domain question answering

Open-domain question answering (OpenQA) requires not only a high-precision reader, but also high-quality retrieval of related passages. Particularly, real-world OpenQA usually involves multi-hop retrieval and reading to deal with complex questions that need bridging information. In this paper, we investigate the mutual promotion of dense retrievers and Graph Neural Network-based readers to improve OpenQA. Specifically, we introduce an alternate training strategy where the scores of the dense retriever and the GNN-based reader are used as correction weights to enhance the performance of each other. We leverage off-the-shelf strong dense retrievers such as Dense Passage Retriever (DPR) and Multi-hop Dense Retriever for retrieval. For the reader, we extend the Asynchronous Multi-grained Graph Network (AMGN) by defining passage nodes and passage-level relationships to cater to the retrieval. It is worth mentioning that through the Recurrent Neural Networks based question reformulation mechanism in AMGN and appropriate preprocessing, the proposed training strategy can be free from the constraints of fixed-hop question answering. We evaluate the proposed framework on several prevalent OpenQA datasets, Natural Questions, TriviaQA, and HotpotQA, achieving competitive results compared with other published models. Extensive experimental analyses illustrate the effectiveness of enhanced passage-aware AMGN and mutual promotion.