Leveraging Contrastive Learning Techniques for Enhanced Aspect-Based Sentiment Analysis

Abstract

Aspect-based sentiment analysis (ABSA) is crucial for evaluating opinions by offering detailed sentiment analysis of specific features in a text. This study compares two contrastive learning techniques to enhance ABSA: supervised contrastive learning based on sentiment analysis and unsupervised contrastive learning based on augmentation. Sentiment-based supervised contrastive learning differentiates between positive and negative samples using sentiment labels, while augmentation-based unsupervised contrastive learning employs data augmentation algorithms to generate positive examples. Experiments on three ABSA datasets show that both strategies significantly improve performance. Supervised contrastive learning based on sentiment labels, while augmentation-based unsupervised contrastive learning employs data augmentation algorithms to generate positive examples. Experiments on three ABSA datasets show that both strategies significantly improve performance. Supervised contrastive learning based on sentiment outperforms unsupervised contrastive learning based on augmentation in terms of performance gains. Transfer learning using pre-trained language models like BERT or GPT is investigated to enhance ABSA in a target domain with limited labelled data by leveraging knowledge from a source domain with ample labelled data. The study also explores the benefits of incorporating external knowledge sources such as sentiment lexicons or domain-specific resources to enhance ABSA performance. These findings contribute valuable insights to ABSA development by examining different learning methodologies and considering cutting-edge technologies. The results promise increased accuracy and granularity of sentiment analysis for specific features in a text, thereby improving opinion and review analysis in various fields.