Text classification is a challenging problem which aims to identify the category of texts. In the process of training, word embeddings occupy a large part of parameters. Under the constraint of limited computing resources, it indirectly restricts the ability of subsequent network design. In order to reduce the number of parameters for constructing word embeddings, this paper explores compositional coding mechanism and proposes a compositional weighted coding method to replace the conventional embedding layer. Furthermore, inspired by the excellent performance of capsule network in image classification, we design a capsule network combined with our compositional weighted coding method for text classification. We also offer a new routing algorithm based on k-means clustering theory to thoroughly mine the relationship between capsules. Experiments conducted on eight challenging text classification datasets show that the proposed method achieves competitive accuracy compared to the state-of-the-art approach with significantly fewer parameters.

文本分类是一个具有挑战性的问题，旨在识别文本的类别。在训练过程中，词嵌入占据了很大一部分参数。在计算资源有限的约束下，间接制约了后续网络设计的能力。为了减少构建词嵌入的参数数量，本文探索了组合编码机制，并提出了一种组合加权编码方法来替代传统的嵌入层。此外，受胶囊网络在图像分类中的优异性能的启发，我们设计了一个胶囊网络，并结合我们的组合加权编码方法进行文本分类。我们还提供了一种基于 k-means 聚类理论的新路由算法，以彻底挖掘胶囊之间的关系。