Coarse-to-Fine Document Ranking for Multi-Document Reading Comprehension with Answer-Completion

Abstract

Multi-document machine reading comprehension (MRC) has two characteristics compared with traditional MRC: 1) many documents are irrelevant to the question; 2) the length of the answer is relatively longer. However, in existing models, not only key ranking metrics at different granularity are ignored, but also few current methods can predict the complete answer as they mainly deal with the start and end token of each answer equally. To address these issues, we propose a model that can fuse coarse-to-fine ranking processes based on document chunks to distinguish various documents more effectively. Furthermore, we incorporate an answer-completion strategy to predict complete answers by modifying loss function. The experimental results show that our model for multi-document MRC makes a significant improvement with 7.4% and 13% respectively on Rouge-L and BLEU-4 score, in contrast with the current models on a public Chinese dataset, DuReader.