一种基于图注意力网络的短文本分类方法

A short text classification method based on graph attention network

针对现有文本分类方法存在的不足,提出了使用图注意力网络进行短文本的分类。首先根据短文本所包含的字和词汇创建一个图结构的数据,然后通过图注意网络赋予同一邻域内各节点不同的权重来减弱噪音数据的影响并且捕获节点间的依赖信息,最后经过最大池化层以生成图级别的语义表示用于类别预测。实验结果表明,该方法相比于CNN、BiLSTM、BiLSTM-MP、FastText、RCNN准确率更高。

To solve the problems with the existing text classification methoods,the method of graph attention networks is proposed for short text classification.First,a graph structure data is created based on the words and vocabulary contained in the short text.Then the graph attention network is used to assign different weights to each node in the same neighborhood to reduce the influence of noisy data and capture the dependency information between nodes.Finally,graph-level semantic representations are generated through the maximum pooling layers for category prediction.Experimental results show that this method achieves better accuracy than CNN,BiLSTM,BiLSTM-MP,FastText,and RCNN.