Sentiment analysis has become an indispensable tool for understanding consumer perceptions, yet its reliability is undermined by implicit negation, which triggers polarity shifts that reverse the intended sentiment. Despite progress in transformer-based and deep learning models, existing approaches remain limited in their ability to capture such subtle semantic inversions. We propose a hybrid deep learning architecture that combines an LSTM-enhanced BERT encoder with contextual back-translation and sequence-to-sequence perturbations for robust data augmentation. FastText and BERT embeddings are jointly leveraged with attention mechanisms to capture both long-range dependencies and nuanced contextual cues. The model was trained and validated on multiple mobile review datasets, and its performance was benchmarked against strong state-of-the-art baselines. Evaluation employed accuracy, Cohen’s kappa, and Matthews’s correlation coefficient, with statistical significance assessed through ANOVA and Kruskal–Wallis H statistic tests.The proposed framework consistently outperformed baseline models across all datasets, achieving superior scores on accuracy, kappa, and MCC. Statistical analyses confirmed that these improvements were significant. Visual inspection via attention heatmaps revealed enhanced sensitivity to negation cues, while box–whisker plots demonstrated greater robustness, with higher medians and reduced variance.These findings establish that integrating hybrid architectures with targeted augmentation strategies markedly improves the detection of polarity shifts induced by implicit negation. The approach enhances both accuracy and stability, offering a pathway toward more reliable sentiment analysis in linguistically complex contexts. Future research may extend this work by combining multiple augmentation strategies to further improve generalization across domains and languages.