DESIGN AND ANALYSIS OF FAKE REVIEW DETECTION IN URDU LANGUAGE USING TRANSFORMER-AUGMENTED RoBERTa + LSTM HYBRID MACHINE LEARNING MODEL
DOI:
https://doi.org/10.70917/ijcisim-2026-5310Keywords:
LSTM, RoBERTa, Fake Reviews, Bi-directional, Transformer BasedAbstract
The development of internet-based platforms has had a major effect on how consumers make choices when shopping and how they view businesses. The increased availability of information about products via the same digital channels has resulted in an increase in fake user reviews, deceptive content intended to sway popular opinion regarding products. Though advancements have been made in identifying these fake reviews for popular languages, such as English, advances in lower-resourced languages, such as Urdu, have a long way to go. To fill this gap, we proposed a new deep learning method based on the linguistic strengths of RoBERTa, trained specifically on Urdu, combined with the temporal modeling capability of a Bi-LSTM model. This method exploits the high contextual value of the RoBERTa (Urdu) text model and the Bi-LSTM model's unique ability to learn sequences to improve accuracy in detecting fraudulent text patterns. The hybrid model outperformed baseline and single model techniques considerably (93.2% accuracy, 93.3% F1 score). This research demonstrates the potential of transformer-based models in detecting fraudulent user reviews in low-resource language contexts, and provides a starting point for future study on how to use this type of model to combat digital misinformation that is prevalent in South Asia