A Text Classification Model via Multi-Level Semantic Features

Text classification is a major task of NLP (Natural Language Processing) and has been the focus attention for years. News as branch text characterized by complex structure, large amounts information long length, which in turn leads to decrease accuracy classification. To improve Chinese news texts, we present model based on multi-level semantic features. First, add category correlation coefficient TF-IDF (Term Frequency-Inverse Document Frequency) frequency concentration CHI (Chi-Square), extract keyword features with improved algorithm. Then, local TextCNN symmetric-channel global from BiLSTM attention. Finally, fuse three prediction categories. The results experiments THUCNews, LTNews MCNews show that our presented method highly accurate, 98.01%, 90.95% 94.24% accuracy, respectively. With parameters two magnitudes smaller than Bert, improvements relative baseline Bert+FC are 1.27%, 1.2%, 2.81%,