<![CDATA[Burnout has become an important psychological concern that is increasingly discussed through social media, providing valuable textual data for understanding public experiences of emotional exhaustion, workplace pressure, and coping. This study analyzes sentiment in burnout related YouTube comments using DistilIndoBERT and examines the contribution of back translation to classification performance. The initial dataset consisted of 2,931 comments collected from 5 YouTube videos published between 2021 and 2025 was subjected to a data quality audit that removed exact duplicates, promotional content, spam, and nonmeaningful comments, resulting in 2,829 relevant records. Sentiment labels were assigned through a semi automated process and reviewed by the researchers into positive, neutral, and negative categories. The dataset was divided using stratified sampling into 70% training data, 15% validation data, and 15% test data. Back translation was applied exclusively to the positive and neutral classes in the training set to prevent data leakage, expanding the training data from 1,980 to 3,003 records. Negative sentiment was dominant, accounting for 1,430 comments or 50.55%, followed by neutral sentiment with 846 comments or 29.90% and positive sentiment with 553 comments or 19.55%. DistilIndoBERT achieved 82.4% accuracy, 81.9% macro precision, 81.5% macro recall, and 81.6% macro F1 score on the original dataset. After augmentation, the respective scores increased to 87.1%, 86.8%, 86.2%, and 86.4%. These observed improvements demonstrate the potential of training focused back translation to strengthen DistilIndoBERT classification of burnout discourse.]]>
Copyrights © 2026
