Current data augmentation methods for sentiment analysis predominantly employ random text transformations, which fail to preserve the semantic importance of emotion-bearing words and often generate low-quality training samples. To address these limitations, we present SentiAug, a unified framework that combines adaptive keyword replacement with confidence-guided self-training for robust sentiment classification. Our approach introduces two key innovations: (1) Keyword Replacement Data Augmentation (KRDA), which employs a self-attention mechanism to identify emotionally salient keywords and performs semantically-aware replacements that preserve sentiment polarity; and (2) Confidence-guided Self-Training Selection (CSTS), a threshold-based framework that dynamically filters and labels augmented samples based on model confidence scores. We further propose the Augmentation Rate (AR), a novel metric to quantify the semantic diversity of augmented datasets, providing theoretical insights into model robustness. Extensive experiments across multiple benchmark datasets demonstrate that SentiAug achieves state-of-the-art performance, improving sentiment classification accuracy.