基于ERNIE的新闻标题文本分类

News title classification based on ERNIE

针对新闻标题文本分类任务,文章提出的EWLTC(ERNIE with Whole Layers for Text.Classification)模型旨在获取文本的不同层级的特征信息,通过抽取预训练模型ERNIE的12层Eneoder层输出中的CLS向量,利用注意力机制进行加权求和作为全连接层的输入。通过实验证明,EWLTC模型在中文新闻标题文本数据集上能够获得更为优秀的表现。

For the task of news title text classification,the EWLTC(ERNIE with Whole Layers for Text Classification)model proposed in this paper aims to obtain different levels of feature information of the text by extracting the CLS vectors from twelve Encoder layers of the pretrained model ERNIE and through the attention mechanism assigning different weights to each vector and summing them as the input into the fully connected layer.Through experiments it can prove that the EWLTC model can obtain better performance on the Chinese news title text dataset.