Semantic-preserved Augmentation with Reliability-aware Fine-tuning for Aspect Category Sentiment Analysis

Data scarcity is one of the challenges faced by aspect category sentiment analysis (ACSA) due to limited labeled data. While recent studies leverage large language models (LLMs) with handcrafted prompts for data augmentation, these approaches often fail to preserve the semantics of the original text. We introduce a semantics-preserved, linguistically diverse data augmentation approach for ACSA that employs structured prompt templates to guide LLMs in generating predefined content. To further enhance semantic consistency, a cosine-similarity-based filtering mechanism ensures that augmented sentences remain faithful to their original meanings. Beyond data augmentation, we propose a reliability-aware fine-tuning strategy that reweights the training objective using a reliability score that combines token-level correctness and sequence-level confidence. Experimental results demonstrate that our method improves performance across benchmark datasets compared with strong baselines.