Roman Urdu sentiment analysis faces significant challenges due to transliteration inconsistencies, informal language usage, and the lack of labeled datasets. This study proposes a novel framework that addresses these challenges by combining advanced data preprocessing techniques and data augmentation strategies such as synonym replacement, back-translation, and random word insertion. These methods enhance dataset diversity, improving the model’s generalization ability. A rich Roman Urdu dataset was collected from diverse sources, including social media platforms (Facebook, Twitter, YouTube), blogs, forums, and e-commerce sites, to capture a wide range of user opinions. Three deep learning models, Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM), were evaluated for sentiment classification. The results show that the LSTM model outperforms the others with an accuracy of 94%, compared to 90% for RNN and 92% for GRU. The LSTM’s ability to capture long-term dependencies and contextual nuances in Roman Urdu text makes it the most effective model for this task, demonstrating a significant improvement over the traditional method.