Aspect-Based Sentiment Analysis Using Adversarial BERT with Capsule Networks

Aspect-based sentiment analysis aims to predict the sentiment polarity of a given aspect in a sentence. The previous methods based on RNNs and attention mechanisms mainly have two problems: (1) The datasets for the aspect-based sentiment analysis task are small. Thus, the superiority of the neural network is not fully utilized. (2) The existing studies use brute force to accurately locate the one-to-one correspondence between the target words and sentiment words, and they also lack the ability to identify deep semantic relationships. To address these issues, we propose adversarial BERT with capsule networks. Specifically, the pre-trained BERT model with an adversarial training mechanism is introduced for semantic representation to solve the issue of insufficiently mining sentence semantic information due to an excessively small amount of training data. In addition, we apply the characteristics of the tensor neuron and dynamic routing mechanism of the capsule network to further explore the in-depth information of the sentence, which is conducive to accurately determining out the logical relationship between the target words and the sentiment words. Moreover, we deploy label smoothing regularization to reduce overfitting by preventing a network from assigning the full probability to each training example during training. As far as we know, this paper is the first innovative attempt to solve the challenge of capturing aspect words and their corresponding modified content bound to small sample data learning in aspect-based sentiment analysis using the adversarial training mechanism of the capsule network. Through experiments on three benchmark datasets, and compared with the state-of-the-art baselines, the extensive results show that our model achieves competitive improvements of up to 1.01%, 0.22%, and 1.68%, respectively.