HCapsNet: A Text Classification Model Based on Hierarchical Capsule Network

In text classification tasks, RNNs are usually used to establish global relationships. However, RNNs have the problems that the semantic information coding of key words is not prominent and cannot be calculated in parallel. In addition, hierarchical information of text is usually ignored during feature extraction. Aiming at the above problems, a text classification model based on hierarchical capsule network (HCapsNet) is proposed. In order to capture the hierarchical features, text is divided into granularities and constantly aggregate according to the characteristics of the data. A parallel LSTM network fused with self-attention is utilized to complete the encoding of multiple natural sentences. Then, we construct sentence features into sentence capsules to extract richer semantic information. The spatial relationship between sentence capsule as part and chapter capsule as whole is established by dynamic routing algorithm. Our experiments show that HCapsNet gives better results compared with the state-of-the-art methods on six public data sets.