Transformer-Based Model for Identifying Ethnic Languages in Bangladesh

Bangladesh alone is home to more than 50 ethnic groups who speak at least 35 different languages, demonstrating the country's cultural and ethnic diversity. There are many more ethnic groups with diverse languages in the world. These languages continue to be underrepresented in the digital sphere, lacking developments like translation, digital preservation, and systematic identification. This study presents a novel empirical evaluation of machine learning and transformer-based models for ethnic language transliteration identification, specifically focusing on three leading ethnic languages: Chakma, Garo, and Marma. The study explores the performance of traditional machine learning classifiers with two feature extraction techniques (TF-IDF and CountVectorizer) and compares them against BERT-based models on an imbalanced dataset. The evaluation metrics, such as f1-score, g-mean, precision, recall, accuracy, and AUC scores, are utilized to assess the performance of the classifiers. The results indicate that Multinomial Naive Bayes (MNB), Support Vector Machines (SVM), and Logistic Regression (LoR) work best with CountVectorizer, whereas Random Forest (RF), XGBoost (XGB), and SVM perform best with TF-IDF on the ethnic dataset. Notably, the empirical evaluation reveals that transformer-based classifiers significantly outperformed traditional machine-learning classifiers with an impressive f1-score of. 99 in the case of ethnic language identification.