Emotion Analysis Base on Capsule Network

Fine-grained sentiment analysis is a research hotspot in the field of natural language processing. Among them, short text sentiment analysis because of the less content. It is difficult to obtain better results using only a single feature extraction network. The article proposes a new hybrid model that uses BERT as the word vector layer for word vector mapping and uses bidirectional gated recurrent network and self-attention capsule network in turn for feature extraction, which solves the problem of the single capsule network's weak ability to extract time series information. At the same time, we propose a self-attention routing algorithm, which can be calculated in parallel to improve model speed. The experimental results on the Emobank sentiment regression dataset show that the mixed structure achieves the best results in the three dimensions of emotion valence, arousal and dominance, and the model is faster than the EM dynamic routing algorithm training.