Capsule Network Based on Multi-granularity Attention Model for Text Classification
Hao Wang, Jing Zhao · 2022
Text classification is a challenging task aimed at identifying categories of text. In order to improve the feature extraction capability of existing shallow text classification models and extract text information hierarchically from bottom to top, a multi-granularity attention model-based capsule network (LMAC) is proposed. This paper adopts a stacked dilated convolutional structure as an encoding module from the perspective of local correlation and long-term dependency. Capturing multi-granular semantic features of text. Combined with capsule network to build a classification model. A multi-granularity attention model based on dilated convolution is proposed, which fully mines multi-granularity semantic features of text by weighting and summing semantic features of different granularities. The effectiveness of the method is validated on 7 datasets for text classification tasks. Compared with the existing research work, the performance is improved significantly. It shows that the LMAC model can mine text semantic features more comprehensively and deeply, and improve the performance of text classification.