基于Bi-LSTM和图注意力网络的多标签文本分类算法

MULTI LABEL TEXT CLASSIFICATION ALGORITHM BASED ONBI-LSTM AND GRAPH ATTENTION NETWORK

针对当前大多数分类算法忽略标签之间相关性的问题,提出一种基于双向长短时记忆(Bi-LSTM)网络和图注意力网络(Graph Attention Network,GAT)的多标签文本分类算法。使用词嵌入工具对文本序列和标签中的词向量进行预处理后的文本序列和标签分别输入到Bi-LSTM网络和GAT网络中;提取文本序列的上下文信息和全局特征,以及GAT网络捕获标签之间的相关性;将特征向量和标签相关性进行组合对标签文本分类任务进行预测。实验结果表明,所提算法通过有效关注标签之间的相关性使得文本分类任务的精度得以明显提高,在多个评估指标的测试结果优于其他对比方法。

Aimed at the phenomenon that most classification algorithms ignore the relationship among labels,a multi label text classification algorithm based on bi-directional long short-term memory(Bi-LSTM)network and graph attention network(GAT)is proposed.The word embedding tool was used to preprocess the word vectors in the text sequence and labels.The preprocessed text sequence and labels were input into the Bi-LSTM network and the GAT respectively.The context information and global features of the text sequence were extracted,and the GAT network captured the relationship among labels.The feature vector and the label relationship were combined to predict the task of tagging text.Experimental results show that the proposed algorithm can improve the accuracy of text classification task by effectively focusing on the relationship among labels,and its test results under multiple evaluation indicators are better than the results of other comparison methods.