基于BERT和深层等长卷积的新闻标签分类

针对THUCNews的中文新闻文本标签分类任务，在BERT预训练语言模型的基础上，提出一种融合多层等长卷积和残差连接的新闻标签分类模型（DPCNN-BERT）。首先，通过查询中文向量表将新闻文本中的每个字转换为向量输入到BERT模型中以获取文本的全文上下文关系。然后，通过初始语义提取层和深层等长卷积来获取文本中的局部上下文关系。最后，通过单层全连接神经网络获得整个新闻文本的预测标签。将本文模型与卷积神经网络分类模型(TextCNN)、循环神经网络分类模型(TextRNN)等模型进行对比实验。实验结果表明，本文模型的预测准确率达到94.68%，F1值达到94.67%，优于对比模型，验证了本文提出模型的性能。

News Label Classification Based on BERT and Deep Equal Length Convolution

For the THUCNews’ Chinese news text label classification task, a news label classification model (DPCNN-BERT) that combines multi-layer equal-length convolution and residual connection based on BERT pre-training language model is proposed. Firstly, by querying the Chinese vector table, each word in the news text is converted into a vector and input into BERT model to get the full-text context of the text. Then, the local context relationship in the text is obtained through the initial semantic extraction layer and deep equal-length convolution. Finally, the predicted label of the entire news text is obtained through a single-layer fully connected neural network. The model proposed in this paper is compared with the convolutional Neural Network Classification Model (TextCNN), Recurrent Neural Network Classification Model (TextRNN) and other models. The experimental results show that the prediction accuracy of the model reaches 94.68%, and the F1 value reaches 94.67%, which is better than the comparison models. The performance of the model proposed in this paper is verified.