LLM-Based Challenging Data Augmentation for Aspect-Based Sentiment Analysis

Aspect-based Sentiment Analysis is a fine-grained sentiment analysis task that aims to evaluate sentiment polarity towards particular aspects in given texts. This task often faces data scarcity issues, where training data remains limited due to high manual annotation costs. Data augmentation is considered as an effective method to expand small datasets. However, traditional data augmentation typically targets all training data, ignoring the difficulty variations between different samples. Augmenting simple data does not yield better quality data, but might instead incur increased training overhead. To tackle this challenge, this paper proposes an approach using Large Language Models for Challenging Data Augmentation (CDA) to generate higher quality labeled data. In contrast to conventional data augmentation that expands all data, CDA first identifies challenging samples from the training data, then incorporates Chain of Thought to specifically augment these challenging instances during the data augmentation process. This augmented data significantly improves model performance under limited training data conditions. The experimental results establish the efficacy of the proposed method.