近年来, 随着胶囊网络的广泛研究, 其在图像、语言等领域取得了重大进展.但胶囊网络存在参数多、训练时间长的缺点. 分组反馈路由机制是一种称为分组路由的监督路由策略，该策略将胶囊平均地分成若干组，胶囊局部共享转换权重，从而减少路由参数和计算复杂度，在图像分类领域取得较好效果.本文将胶囊分组方法运用于文本分类任务中，再引入胶囊压缩、静态路由机制，提出了一种新的文本分类模型CapsNet-GSR. 该模型通过胶囊分组在提取文本局部信息的同时减少参数，利用胶囊压缩和静态路由机制，进一步提高胶囊质量、降低参数数量. 在20 news文本分类数据集上的实验证明，其在参数数量和训练时间上有明显减少. 在AG’s news、TREC和20 news数据集上的实验表明，该模型在准确率上也有所提高.
The use of capsule network in the fields of image and language have progressed tremendously over the past few years, with the extensive research of the capsule network. However, the capsule network has the disadvantages of too many parameters and long training time. The group feedback routing mechanism is a supervised routing strategy called group-routing. The strategy divides the capsules into several groups evenly, and the capsules locally share the conversion weights, thereby reducing routing parameters and computational complexity, and achieving better results in image classification. In this paper, a new text classification model CapsNet-GSR(CapsNet-Grouped capsule based on Static Routing) is proposed based on the capsule grouping method, and the capsule compression and static routing mechanism are also introduced. The model uses capsule grouping to extract local information of the text while reducing parameters. In addition, it uses capsule compression and static routing mechanisms to further improve the quality of the capsule and reduce the number of parameters. The experiment on the 20 news text classification dataset proves that the number of parameters and training time of the proposed model decrease obviously. The experiment results on AG’s news, TREC, and 20 news datasets show that the accuracy of the proposed model is also improved.
引用本文格式： 朱海景,余谅,盛钟松,陈贵强,王争. 基于静态路由分组胶囊网络的文本分类模型[J]. 四川大学学报: 自然科学版, 2021, 58: 062001.复制