Detecting Fake Reviews in Roman Urdu Using Transformer Based Language Models

Reviews on online platforms face growing threats from deceptive content published by malicious users, which affects marketplace integrity. While widely used in South Asian e-commerce, Roman Urdu remains less explored due to non-standard conventions of spellings and frequent code mixing with English which weaken standard NLP pipelines. This paper introduces a Roman Urdu fake review detection corpus, RU-FRDC, which contains 5,026 samples annotated into fake and real classes. The dataset shows a realistic 2.53:1 imbalance ratio, containing 3,602 real and 1,424 fake instances. To counter evaluation biases, we propose a leakage-safe protocol which removes duplicates and enforces disjoint train (2,947), validation (328), and test (1,751) splits. Using this protocol, we evaluate lexical baselines against multiple fine-tuned transformers. Our best model, TF-IDF with Logistic Regression, achieved the highest overall efficacy with 0.9175 accuracy, weighted F1 score 0.9136, and a macro F1 score of 0.8878. Importantly, it balances precision and recall by maintaining a remarkably low false positive count1 (FP=7) at the expense of 143 false negatives, which demonstrates a conservative minority class flagging behavior. This close outcome is plausible for Roman Urdu review text because reviews are often short and sentiment heavy, and fake reviews commonly reuse a limited set of promotional templates. Under such conditions, TF IDF models can capture repeated phrases and common deception patterns effectively, especially when evaluation is leakage-safe and duplicates are controlled. Among the fine-tuned transformer networks, the multilingual encoder XLM-RoBERTa (XLM-RoBERTa-base) achieves the highest performance with a classification accuracy of 0.9143 and a weighted F1 score of 0.9096, which is followed closely by bert base multilingual cased at 0.9109 accuracy and a weighted F1 score of 0.9061.