ABSA: Aspect-based sentiment Analysis aims to extract emotions from text concerning attributes of products or services. However, the training with labeled data is scarce, and difficult to preserve tiny contextual information, which introduces complications for ABSA. A novel solution for the existing problems of ABSA is proposed in this paper. To combat data scarcity, we utilize large-scale language models such as GPT for contextual, synthetic text generation (data augmentation techniques), which are evaluated on semantic similarity using BERTScore. In conclusion, to improve our model’s generalization ability to different forms of sentiment expressions, we propose a method incorporated into global learning in both the forward and backward phases. Our key contributions are: developing a synthetic text generation framework to mitigate data scarcity, validating the quality of generated texts using BERTScore, and improving ABSA performance, especially in multilingual and low-resource settings. Experimental results demonstrate that our approach significantly advances sentiment classification accuracy and offers a robust strategy for data-centric sentiment analysis.