Multi-Label Text Categorization Model Based on Semanteme and Label Knowledge Guidance
Chunhui Zhao, Jinxin Si, Fuyong Sun, Shijie Gao, Xin Yi He, Zehao Yu, Jiangbo Yin, Yueyuan Wang · 2024
Since electric power text has strong professional knowledge and industry characteristics, traditional text classification model has the problems of extracting features not accurate enough and neglecting domain knowledge guidance, this paper proposes a multi-label text classification model for electric power based on Semanteme and knowledge guidance. The text features are extracted using Bert pre-trained language model, the public semantics are extracted based on the text set corresponding to each label as the guiding information of the labeled text, and the label feature vectors are trained using graph-attention neural network to get the deep semantics of the labels. The label feature vectors are used as knowledge guides in classification to get the text features based on the label semantic attention, and finally they are fed into the multilayer perceptual machine to realize classification. The experiment proves that the method proposed in this paper obtains improvement in classification accuracy, F1 value and other selected indexes, which can effectively improve the performance of automatic power text classification.