Multiple Perspective Answer Reranking for Multi-passage Reading Comprehension

This study focuses on multi-passage Machine Reading Comprehension (MRC) task. Prior work has shown that retriever, reader pipeline model could improve overall performance. However, the pipeline model relies heavily on retriever component since inferior retrieved documents would significantly degrade the performance. In this study, we proposed a new multi-perspective answer reranking technique that considers all documents to verify the confidence of candidate answers; such nuanced technique can carefully distinguish candidate answers to improve performance. Specifically, we rearrange the order of traditional pipeline model and make a posterior answer reranking instead of prior passage reranking. In addition, new proposed pre-trained language model BERT is also introduced here. Experiments with Chinese multi-passage dataset DuReader show that our model achieves competitive performance.