<p>Aspect Sentiment Triplet Extraction (ASTE), a core fine-grained sentiment analysis task, faces severe class imbalance, especially in neutral sentiment labels, which undermines model performance. Existing data augmentation methods for ASTE suffer from three key limitations: semantic inconsistencies, lack of complex entity structures, and inability to generate samples with specific sentiment labels. Moreover, current approaches fail to dynamically regulate the volume of augmented data, leading to suboptimal class distribution balancing. To address these challenges, we propose A Gini-Adaptive Diffusion-based Data Augmentation Approach for Aspect Sentiment Triplet Extraction (DiffAste). To our knowledge, this represents the first application of diffusion models to ASTE data augmentation, where the proposed method systematically combines masked matrices with conditional diffusion processes to generate linguistically diverse templates while rigorously preserving original entity semantics and sentiment orientations. We introduce a semantic filtering mix-up strategy to mitigate generative-model-induced noise and employ a KL divergence constraint to ensure the generation of neutral sentiment samples. We further design a Gini-Adaptive Sample Controller to utilize the Gini coefficient, dynamically balancing sample generation. Through extensive experiments, our approach achieves significant improvements, elevating F1 scores by 1% to 2% across all four sub-datasets while reducing the Gini coefficient by up to 0.14, demonstrating its effectiveness in class imbalance mitigation.</p>