The Impact of Incorporating Back Translation in Data Augmentation for Subjectivity Classification

In text classification tasks, limited training data often leads to poor model performance, making data augmentation essential for improving robustness and generalization. This study explores the effectiveness of combining Easy Data Augmentation (EDA) with back-translation and examines the impact of different intermediary languages on classification outcomes. Experiments are conducted on the SST-2 (sentiment classification) and SUBJ (subjective/objective classification) datasets, using GPT-3.5-turbo and Google for back-translation. Intermediary languages include French, Malay, Chinese, and Japanese. Results show that back-translation significantly enhances performance, especially with training sizes of 500 and 2000. Deep learning models like BERT, LSTM, and CNN benefit consistently across languages, while traditional models such as Random Forest are more sensitive to language choice and may experience performance drops. Notably, Chinese yields the best results for SST-2, while Japanese performs best for SUBJ. BERT achieves the highest overall accuracy, but models like LSTM and CNN show greater relative improvement with augmented data.