关键词: 分层图注意力网络, 单标签文本分类, 邻接矩阵, 标签信息

Abstract: Currently, there are two main limitations in single-label text classification tasks based on hierarchical graph attention networks. First, it cannot effectively extract text features. Second, there are few studies that highlight text features through the connection between text and labels. To address these two issues, a hierarchical graph attention network text classification model that integrates label information is proposed. The model constructs an adjacency matrix based on the relevance between sentence keywords and topics, and then uses word-level graph attention network to obtain vector representations of sentences. This method is based on randomly initialized target vectors and utilizes maximum pooling to extract specific target vectors for sentences, making the obtained sentence vectors have more obvious category features. After the word-level graph attention layer, a sentence-level graph attention network is used to obtain new text representations with word weight information, and pooling layers are used to obtain feature information for the text. On the other hand, GloVe pre-trained word vectors are used to initialize vector representations for all text labels, which are then interacted and fused with the feature information of the text to reduce the loss of original features, obtaining feature representations that are distinct from different texts. Experimental results on five public datasets (R52, R8, 20NG, Ohsumed, and MR) show that the classification accuracy of the model significantly exceeds other mainstream baseline models.

Key words: hierarchical graph attention network, single label text classification, adjacency matrix, label information ,