A Long-Text Classification Method of Chinese News Based on BERT and CNN

Text Classification is an important research area in natural language processing (NLP) that has received a considerable amount of scholarly attention recent years. However, real Chinese online news characterized by long text, large information and complex structure, which also reduces the accuracy text classification as result. To improve news, we propose BERT-based local feature convolutional network (LFCN) model including four novel modules. First, to address limitation Bidirectional Encoder Representations from Transformers (BERT) on length max input sequence, named Dynamic LEAD-n (DLn) method extract short texts within based traditional LEAD digest algorithm. In Text-Text (TTE) module, use BERT pretrained complete sentence-level vector representation capture global features using mechanism identify correlated words text. After that, CNN-based convolution (LFC) module such key phrases. Finally, vectors generated different operations over several periods are fused used predict category Experimental results show new further improves news.