Research On Text Classification Tasks Integrating Label Informatio

With the rapid development of the Internet,a large amount of non-standard text data is flooding all kinds of platforms,and text classification can help people quickly review and efficiently categorize text.Text classification has single-label text classification and multilabel text classification,and how to fuse label information for better extraction of text features has been a hot topic of research,so this paper addresses this issue as follows:(1)Existing single-label text classification tasks fail to focus better on words with distinct category features in sentences,as well as fail to highlight text topic features with the semantic information of labels.To address these problems,this paper proposes a hierarchical graph attention network text classification model(HGAT-Label)that fuses semantic information of labels.The model fuses random initialized target vectors and maximum pooled target vectors to find out words of obvious categories in sentences by the graph attention mechanism;in addition,the label information of all text annotations is vectorized and interacted and fused with feature information of the text to highlight text topic features.The experimental results show that learning the important word information representation and fusing the semantic information of tags is effective.(2)Although the HGAT-Label model extracts text classification features more accurately with the help of label semantic information,it is only applicable to single-label text classification and cannot handle finer-grained multi-label text classification.The existing multi-label text classification also has the following problems: firstly,it fails to extract comprehensive text features;secondly,it does not consider the connection between global labels.To address the above problems,this paper proposes a multi-label text classification model(CS-GAT)based on convolutional neural network-self-attention mechanism and graph attention network.The model fuses local and global semantic features of text on the one hand;on the other hand,it exploits the association between global tags using graph attention network.The experimental results show the effectiveness of enriching semantic features of text and mining the correlation between global labels.(3)Although the CS-GAT model considers the text and label feature information,it ignores the interaction effect of both in the feature extraction process;in addition,the longtail distribution of labels and the setting of adjacency matrix can lead to a lot of noise.To address the above problems,this paper proposes a label attention network model based on local and global semantic feature guidance on the basis of CS-GAT model.The model establishes the connection between text and labels through the label graph attention mechanism;on the other hand,it sets weights and thresholds to reduce the noise impact caused by the label co-occurrence pattern.The experimental results show that establishing the connection between text and labels and reducing the noise effect can improve the classification effect of the model.