Aspect-Pair Supervised Contrastive Learning for aspect-based sentiment analysis

Aspect-based sentiment analysis (ABSA) is a fine-grained sentiment classification task, which identifies the sentiment polarity of a specific aspect in a sentence. In general, the syntactic dependency and semantic information between aspects and their contexts are modeled using deep neural networks. However, most of the existing methods treat specific aspects in a sentence independently, while ignoring the sentiment relationships between multiple aspects. In this work, we propose an Aspect-Pair Supervised Contrastive Learning (APSCL) model to capture the latent relationships between multiple aspects in the sentiment subspace. Through experiments, we approve that in the embedding space, the representation discrepancy of aspect-pairs in the same relation category is narrowed while the embedding representation of aspect-pairs in different relation categories is pushed away. Then the aspect feature representation is enhanced through the relationship optimization between aspects. Furthermore, the relation categories between aspects are established in terms of the existing label attributes of aspects, and no additional corpus is needed. The extensive experiments on public datasets SemEval 2014 and MAMs show that the proposed framework APSCL is able to improve up to 2.29% on accuracy and 4.18% on F1 score over top-1 baseline models. Moreover, our framework can also be adapted to other benchmark models.