Multilingual Scam Detection for English, Tamil, and Tanglish Using Traditional Machine Learning and XLM-R

Scam-oriented spam messages increasingly appear in more than one linguistic form, including native-script Tamil, Tamil written with the Latin alphabet (Tanglish), and English. This variation creates a practical text-classification problem because messages with similar intent can differ substantially at the character and token level. This study presents a multilingual scam-message detection framework that compares three conventional TF-IDF-based classifiers—Logistic Regression, Linear Support Vector Machine (SVM), and Random Forest—with the multilingual XLMR transformer. The evaluation uses a dataset of 2,120 messages divided into 1,484 training, 318 validation, and 318 test messages. The test partition contains 150 Tamil, 150 Tanglish, and 18 English messages. Preprocessing preserves the distinction between Tamil script and Tanglish rather than translating both forms into English. Unigram and bigram TF-IDF features are used for the conventional models, while XLM-R is fine-tuned with a maximum sequence length of 128 tokens. On the current test split, all four evaluated models obtain 1.0000 accuracy, precision, recall, and F1-score. However, a near-duplicate analysis identifies substantial overlap between messages across partitions: 5,211 pairs have a similarity of at least 0.90, and 1,084 pairs reach at least 0.95. In an examination of 300 test messages, 265 have a training message with similarity of at least 0.90. These findings indicate that the measured perfect scores reflect the constructed benchmark and should not be interpreted as proof of equivalent real-world generalization. The study therefore emphasizes both the feasibility of multilingual scam detection and the need for template-disjoint evaluation with larger, naturally occurring Tamil, Tanglish, and English data.