融合对抗训练与ERNIE的短文本情感分析模型

使用深度学习技术进行文本情感分类是近年来自然语言处理领域的研究热点,好的文本表示是提升深度学习模型分类性能的关键因素。由于短文本蕴含情感信息较少、训练时易受噪声干扰,因此提出一种融合对抗训练的文本情感分析模型PERNIE RCNN。该模型使用ERNIE预训练模型对输入文本进行向量化,初步提取文本的情感特征。随后在ERNIE预训练模型的输出向量上添加噪声扰动,对原始样本进行对抗攻击生成对抗样本,并将生成的对抗样本送入分类模型进行对抗训练,提高模型面临噪声攻击时的鲁棒性。实验结果表明, PERNIE RCNN模型的文本分类性能更好,泛化能力更优。

A Short Text Affective Analysis Model Combining Adversary Training and ERNIE

Text sentiment classification using deep learning techniques is a hot research topic in the field of natural language processing in recent years, and good text representation is a key factor in improving the classification performance of deep learning models. A text sentiment analysis model PERNIE RCNN that includes adversarial training is proposed, as short texts contain little sentiment information and are susceptible to noise interference during training. The model uses the ERNIE pre-trained model to vectorize the input text and initially extract the sentiment features of the text. The model then adds noise perturbations to the output vector of the ERNIE pre-training model to generate adversarial samples against the original samples, and feeds the generated adversarial samples into the classification model for adversarial training to improve the robustness of the model against noise attacks. The experimental results show that the PERNIE RCNN model has better text classification performance and better generalisation ability.