基于ERNIE-BiGRU模型的中文文本分类方法

针对新闻文本分类方法中词向量的表示无法很好地保留字在句子中的信息及其多义性,利用知识增强的语义表示(ERNIE)预训练模型,根据上下文计算出字的向量表示,在保留该字上下文信息的同时也能根据字的多义性进行调整,增强了字的语义表示。在ERNIE模型后增加了双向门限循环单元(Bi GRU),将训练后的词向量作为Bi GRU的输入进行训练,得到文本分类结果。实验表明,该模型在新浪新闻的公开数据集THUCNews上的精确率为94. 32%,召回率为94. 12%,F1值为0. 942 2,在中文文本分类任务中具有良好的性能。

Chinese-text Classification Method Based on ERNIEBiGRU

In the news text classification method,the representation of word vectors cannot well preserve the information of the words in the sentence and its ambiguity.Using ERNIE pre-trained model,the vector of words is calculated according to the context.While retaining the context information of the word,it can also be adjusted according to the ambiguity of the word,which enhances the semantic representation of the word.A BiGRU layer is innovatively added after the ERNIE model,and the trained word vector is used as the input of the BiGRU for training to obtain the text classification result.The experiments show that the accuracy of the model on the public data set THUCNews of Sina News is 94.32%,the loss rate is 94.12%,and the F1 value is 0.9422,which has good performance in Chinese text classification tasks.