<p>Sentiment analysis is a computational method for detecting and evaluating opinions or attitudes within text. As digital data continues to grow, businesses are increasingly utilizing it to assess customer feedback. To address the limitations of traditional machine learning methods, particularly for code-mixed languages, we propose an unsupervised approach. This paper introduces an unsupervised sentiment classification framework, Hawk–Dove Dynamics with GAN Integration (HD–GAN), which combines game theory and generative modeling to enhance sentiment understanding. The proposed method models sentiment interaction as a Hawk–Dove game, where contextual and emotional agents compete and adapt until reaching a stable equilibrium that reflects the dominant sentiment polarity. To strengthen representation learning, a Generative Adversarial Network (GAN) is employed to generate synthetic sentiment features that improve data balance and reduce overfitting. This dual mechanism enables HD–GAN to effectively capture nuanced sentiment variations in code-mixed, multilingual, and imbalanced datasets without relying on labeled data. Experimental results demonstrate that HD–GAN outperforms recent transformer-based baselines, achieving 92.1% accuracy, 91.6% F1-score, and 91.8% recall, thereby validating the efficacy of integrating game-theoretic dynamics with adversarial learning for robust and interpretable sentiment classification</p>