对抗网络和BERT结合的电商平台评论短文本情感分类

文本情感分类领域性强,传统情感分类方法在多领域混合数据上效果不佳.为了提升多领域混合情境下文本情感分类的准确率,使用双向编码器表征模型(bidirectional encoder representations from transformers,BERT)得到短文本的初级表征;利用对抗网络提取与领域无关的情感特征;利...

Sentiment classification of e-commerce platform reviews based on adversarial network and bert

Text sentiment classification is highly domain-specific, and traditional sentiment classification methods do not work well on multi-domain mixed data. To improve the accuracy of text sentiment classification in multi-domain mixed context, we use bidirectional encoder representations from transformers (BERT) to obtain the primary representations of short texts. Adversarial networks are used to extract domain-independent sentiment features. Bidirectional long short-term (BiLSTM) is used to extract contextual features. The two extracted features are fused to form an emotion classification model based on confrontation network and Bert to improve the accuracy of emotion classification. Comparative experiments on public datasets show that the sentiment classification models based on adversarial networks and BERT have higher accuracy than the baseline models, with the accuracy reaching 95.25% and 93.61% on the two datasets, respectively, and have better performance on the datasets with large differences in domains, which initially validates the effectiveness of the sentiment classification models in multi-domain mixed scenarios. The proposed model is suitable for scenarios involving multiple domains such as real-time monitoring of products in e-commerce platforms.