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

Abstract

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