【摘要】 目前机器阅读理解的注意力网络主要基于LSTM或GRU，由于RNNs的本质，训练和推理非常耗时。此外，这些模型使用粗粒度的注意机制来定位答案的边界。为了解决上述问题，论文提出了一个双重的简单递归单元（DSRU）和一个细粒度的指针网络，并且提出了一种新的机器阅读注意网络（FGAN），旨在回答给定的叙事段落的问题，该网络性能良好，且耗时较短。在Stanford Question Dataset(1.1)上展示了模型的有效性，该单一模型在开发集上获得了85.1的F1分。此外，在2.0上进行的辅助实验表明，该单一模型在开发集上取得了65.95的F1分，这个分数高于两个提出的基线，但略低于最佳基线。更多还原

【Abstract】 The attention network of machine reading comprehension is mainly based on LSTM or GRU. Due to the nature of RNNs,training and reasoning are time-consuming. In addition,these models use coarse-grained attention mechanisms to locate the boundaries of the answers. To solve the above problems,a dual simple recursive unit(DSRU)and a fine-grained pointer network are proposed. A new machine reading attention network(FGAN)is proposed to answer the questions of a given narrative passage. The effectiveness of the model is demonstrated on the Stanford Question Dataset(1.1),and the single model gained 85.1 F1 points in the development set. In addition,auxiliary experiments on 2.0 shows that the single model achieved a F1 score of 65.95 on the development set,which is higher than the two proposed baselines but slightly below the optimal baseline.更多还原