MTDOT: A Multilingual Translation-Based Data Augmentation Technique for Offensive Content Identification in Tamil Text Data

The posting of offensive content in regional languages has increased as a result the accessibility low-cost internet and widespread use online social media. Despite large number comments available online, only small percentage them are offensive, resulting an unequal distribution non-offensive comments. Due to this class imbalance, classifiers may be biased toward with most samples, i.e., class. To address Multilingual Translation-based Data augmentation technique for Offensive identification Tamil text data (MTDOT) is proposed work. MTDOT method applied HASOC’21, which dataset. obtain balanced dataset, each comment augmented using multi-level back translation English Malayalam intermediate languages. Another dataset generated by employing single-level Malayalam, Kannada, Telugu While both approaches equally effective, back-translation approach produces more diverse data, evident from BLEU score. work achieved promising improvement F1-score over widely used SMOTE balancing 65%.