Text Classification Model Based on Multi-head self-attention mechanism and BiGRU
Fang Fang, Xuegang Hu, Jianhua Shu, Pin Wang, Tongping Shen, Fangfang Li
Abstract
Deep learning promotes the development of natural language processing quickly. However, there are still many problems in natural language processing, such as semantic diversity and so on. To solve these problems, this paper proposes a text classification model based on multi-head self-attention mechanism and BiGRU. Firstly, the BERT model is used to generate input sequence of tokens for word embedding layer. Secondly, in order to capture the relation of word dependence in sentences, multi-head self-attention mechanism is used to calculate feature vector weights. And bidirectional GRU is used to extract feature and optimize classification model in order to improve calculate speed. The experiment shows that using this proposed model for text classification can improve the accuracy, recall and F1 score also were higher than traditional Benchmark algorithm, such as CNN, time consuming is lower than BERT model. And word embedding on BERT is better than word2vec.