Fatemeh Zarisfi Kermani

Sentiment analysis in Persian texts poses a persistent challenge in the field of natural language processing (NLP) due to the unique linguistic features and structural complexities of the language. Existing methods for sentiment analysis often demand substantial computational resources because of their reliance on complex models with numerous parameters. Despite this, these methods frequently fail to achieve satisfactory levels of accuracy and generalization. Additionally, their performance deteriorates when confronted with unconventional or noisy data, limiting their efficiency in real-world applications. This research proposes a novel transformerbased model tailored for sentiment analysis in Persian texts. By reducing the number of model layers and employing adversarial training techniques, the proposed approach significantly enhances performance. The reduction in model parameters not only improves computational efficiency but also achieves superior accuracy and F1-score compared to existing approaches. Experimental evaluations reveal that the proposed model attains an accuracy of 96.84 and an F1-score of 96.83 on the “Taghche” dataset, and an accuracy of 90.72 and an F1-score of 90.69 on the “Snappfood” dataset. These results highlight the model’s clear advantage in developing lightweight, faster, and more effective sentiment analysis systems for practical and wide-ranging applications.