基于多段落排序的机器阅读理解研究

针对多段落的机器阅读理解问题，在双向注意力流（BiDAF）模型的基础上，结合双向长短期记忆网络（BiLSTM）和self-attention机制构建了多段落排序BiDAF（PR-BiDAF）模型，利用该模型定位答案所在的段落，然后在预测段落中寻找最终答案的始末位置。实验结果表明，相较于BiDAF模型，本文提出的PR-BiDAF模型的段落选择正确率、BLEU4指标及ROUGE-L指标分别提高了约13%、6%和4%。

Machine reading comprehension based on multi-passage ranking

To solve the problem of machine reading comprehension of multi-paragraphs, we propose a model named PR-BiDAF which uses BiLSTM and self-attention based on the bi-directional attention flow (BiDAF) model. The model is used to locate the paragraph in which the answer is located, and then to find the beginning and end of the final answer in the prediction paragraph. Experiments show that, compared with the BiDAF model, the paragraph selection accuracy, BLEU4 index and ROUGE-L index of the PR-BiDAF model proposed in this paper are increased by about 13%, 6% and 4% respectively.