基于域对抗网络和BERT的跨领域文本情感分析

跨领域文本情感分析时,为了使抽取的共享情感特征能够捕获更多的句子语义信息特征,提出域对抗和BERT(bidirectional encoder representations from transformers)的深度网络模型。利用BERT结构抽取句子语义表示向量,通过卷积神经网络抽取句子的局部特征。通过使用域对抗神经网络使得不同领域抽取的特征表示尽量不可判别,即源领域和目标领域抽取的特征具有更多的相似性;通过在有情感标签的源领域数据集上训练情感分类器,期望该分类器在源领域和目标领域均能达到较好的情感分类效果。在亚马逊产品评论数据集上的试验结果表明,该方法具有良好的性能,能够更好地实现跨领域文本情感分类。

Cross-domain text sentiment classification based on domain-adversarialnetwork and BERT

In order to capture more sentence semantic information from the extracted shared sentiment features for cross-domain sentiment analysis, a deep network model based on domain adversarial mechanism and BERT (bidirectional encoder representations from transformers) was proposed. The model firstly used BERT to obtain the semantic representation vectors of sentences, and then extracted the local features of sentences with a convolutional neural network. A domain adversarial neural network was designed to make the representations of features extracted from different domains to be as indistinguishable as possible, that was, the features extracted from source domain and target domain had much more similarities; and a sentiment classifier was trained on the source domain dataset with sentiment labels, and it was expected that the trained sentiment classifier would have good classification performance in the source domain, and in the target domain. The experimental results on Amazon product reviews dataset showed that the proposed method achieved the expectation and was competent for achieving cross-domain text sentiment classification.