文本阅读理解的快速多粒度推断深度神经网络

Fast multi-granularity inference deep neural networks for text reading comprehension

机器阅读理解任务(MRC)是自然语言处理领域的重要研究方向,通过深度学习网络来进行机器阅读理解课题研究已成为目前的主流方法。考虑到深度网络中的计算冗余与同质性现象,本文提出了一个快速多粒度推断深度神经网络(FMG)。FMG模型在纵向上以卷积神经网络和注意力机制为基本底层架构,横向上以多粒度的文章文本表征与问题表征分层交互融合,共同实现答案的推断。实验结果表明,多粒度推断机制在提高模型表现上具有一定的有效性,且相比于经典循环神经网络,模型实现了训练速度上的进一步提升。

Machine reading comprehension task(MRC)is an important research direction in the field of natural language processing.Applying deep learning network to research machine reading comprehension has become the mainstream methods at present.Considering the computational redundancy and homogeneity in depth network,a fast multi-granularity inference deep neural network(FMG)is proposed in this paper.The FMG model takes convolutional neural network and attention mechanism as the basic underlying architecture vertically,and multi-granularity passage representation and question representation are interacted in a hierarchical interactive way horizontally,so as to jointly realize the inference of answers.The experimental results show that the multi-granularity inference mechanism is effective in improving the performance of the model,and the model improves the training speed compared with the classical recurrent neural networks model.