Answers are grounded in this paper and cite their sources.
Fine-tuning models identifies depression emotions in Indonesian tweets, suggesting improved mental health insights.
Emotion detection in social media, particularly in the context of mental health, is a growing area of interest in natural language processing (NLP). This study aims to fine-tune Transformer-based models—IndoBERT, RoBERTa, and DistilBERT—to detect emotions associated with depression in Indonesian-language tweets. Unlike previous studies that focus solely on general sentiment analysis or use limited feature extraction methods, this research integrates domain-specific keywords curated from mental health literature and expert references to collect 5,271 Indonesian tweets. The preprocessing steps are systematically designed and supported by existing NLP practices to ensure consistency and noise reduction. A comparative evaluation using various feature extraction techniques—such as TF-IDF, Bag of Words, and Word2Vec—is conducted with traditional machine learning classifiers under a 10-fold cross-validation scheme. Results show that DistilBERT paired with TF-IDF and Support Vector Machine achieves the highest F1-score of 0.699 and accuracy of 0.705, outperforming IndoBERT and RoBERTa in most configurations. While these results are modest compared to benchmarks in high-resource language settings, they represent an incremental yet meaningful step forward in the underexplored context of Indonesian mental health discourse on social media. Limitations related to dataset scope and platform dependency (Twitter/X) are acknowledged, with future work directed at expanding to multi-platform and multilingual settings. This study contributes to a deeper understanding of emotional expressions tied to depression in low-resource language environments and highlights the need for more rigorous, domain-aware approaches in mental health NLP research.
No takes yet. Share an insight, caveat, or question.
Martono et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: