Coarse-to-Fine Retriever for Better Open-Domain Question Answering

Abstract

The retriever-reader framework has been widely used in open-domain question answering with great success. Current studies show that better retrieval can greatly improve the performance of final answer extraction and may replace the reader stage. Considering the limited computing resources and the great progress that has been made in reading comprehension, we continue to use the retriever-reader framework and focus on efficient retrieval. In this paper, we propose a new coarse-to-fine retrieval method to take away the semantic noise left by coarse-grained filtering. In particular, we join a fine-grained retriever after the passages generated by the coarse-grained retriever, making all sentences in the passage match more closely with the question. Meanwhile, we use contrastive learning to construct dense vector representation for fine-grained retriever. Experiments on the QA dataset show that our model outperforms the most mainstream model greatly by about 11.7% which importantly gets more out of the operation of retriever.