基于可分离卷积神经网络的文本分类  

Decomposable convolutional neural network for text classification

作　　者：严佩敏[1] 唐婉琪 Yan Peimin;Tang Wanqi(College of Communication and Information Engineering,Shanghai University,Shanghai 200444,China)

机构地区：[1]上海大学通信与信息工程学院,上海200444

出　　处：《电子测量技术》2020年第13期7-12,共6页Electronic Measurement Technology

摘　　要：近年来,卷积神经网络模型在文本分类中显示出了良好的应用前景,但该模型优势在于可以应用更深更广的卷积层来提取更丰富的语义特征,带来了昂贵的计算成本,并且在量级差异较大的数据集中不具备普适性。为了解决这一问题,提出了一种新型卷积网络结构,即用可分解两层卷积网络代替传统的文本卷积网络。一层词嵌入卷积层用来提取单词的词嵌入特征,另一层区域卷积层用来提取单词的上下文特征。模型在CPU上对多个数据集进行了测试,结果表明,该模型不仅降低了训练复杂度,在MR数据集上实现4 min 40 s的最短训练时间,而且在大数据集AG上准确率达到92.6,小数据集MR上达到83.0,这证明了模型在精度和鲁棒性方面都有良好的效果。此外,还进一步讨论了卷积核尺寸和数量对模型性能的影响,并且对以往的诸如SVM,CNN等经典文本分类模型做出了回顾,比较和总结,显现出了可分离卷积模型在算法复杂度和准确率方面的优越性。In recent years, Convolutional Neural Network(CNN) model has shown a good prospect in text classification. However, the advantage of this model is that it can apply deeper and wider convolution layer to extract more abundant semantic features, which brings expensive computing costs, and does not have universality in datasets with large magnitude differences. In order to solve this problem, this paper presents a new convolutional network structure, that is, replaces the traditional text convolutional network with a newly designed decomposable two-layers convolutional network. One embedding convolutional layer is used to extract word embedding features, and the other regional convolutional layer is used to abstract the n-gram features of every word. In the experiment, Decomposable Convolutional Neural Network(DCNN) proposed in the paper is tested on multiple datasets on CPU. The results show that this model is not only reduces the training complexity, achieves the shortest training time of 4 min 40 s on the MR dataset, but also achieves the accuracy of 92.6 on the large AG dataset and 83.0 on the small MR dataset, which proves that the model has good results in accuracy and robustness. And, this paper further discusses the effect of the size and number of convolution kernels on the model performance. In addition, the previous classical text classification models are reviewed and summarized, such as SVM, CNN, which shows the superiority of DCNN model in algorithm complexity and accuracy.

关 键 词：神经网络 深度学习 文本分类 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]