Aspect category sentiment classification via document-level GAN and POS information

The purpose of aspect-category sentiment classification (ACSC) is to determine the sentiment polarity of the predefined aspect category from the texts. Current methods for ACSC have two main limitations. Since the aspect categories are not presented in the given texts, the establishment of relation between the aspect-category and its sentiment opinion expression is challenging using the widely-applied aspect-term sentiment classification approaches. Besides, the aspect-category-related information on document level are ignored during processing. In this work, we focus on dealing with the part-of-speech information based on gated-activation functions. Furthermore, two graph attention networks (GANs) are employed to exploit the document-level sentiment of both the entity and the attribute (intra-entity sentiment tendency and intra-attribute sentiment tendency). The aspect-category detection (ACD) is taken as a auxiliary task to capture the relevant semantic information. Besides, contrastive learning is receiving an increasing amount of interest due to its success in self-supervised representation learning in the field of NLP. By performing contrastive learning, representations of positive examples are drawn closer while those of negative samples are distanced. Comparing with the baseline methods, experimental results reveal that our model achieves the state-of-the-art performance in ACSC tasks.