Deteksi Stres Berbasis Teks pada Dreaddit Menggunakan Fine Tuning DeBERTa-v3

Abstract

Mental health has become an important issue in the digital era, as psychological expressions are increasingly reflected through social media posts such as Reddit. This study uses the publicly available Dreaddit dataset containing Reddit user texts labeled with stress categories. The objective is to compare two text-based stress detection approaches: fine-tuning the transformer model DeBERTa-v3 and the classical TF-IDF LinearSVC method. Both approaches are implemented as binary classification systems to automatically distinguish stress and non-stress texts. The research workflow includes data preprocessing, tokenization, model training, validation, and evaluation using Accuracy, Precision, Recall, F1-score, and AUROC metrics. DeBERTa-v3 is fine-tuned using contextual representations with a self-attention mechanism, while TF-IDF LinearSVC relies on statistical n-gram weighting. Experimental results show that DeBERTa-v3 achieves superior performance with an Accuracy of 0.830, Precision of 0.802, Recall of 0.889, F1-score of 0.843, and AUROC of 0.918. Meanwhile, TF-IDF LinearSVC obtains an Accuracy of 0.732, Precision of 0.722, Recall of 0.783, F1-score of 0.751, and AUROC of 0.817. The experiments were conducted with consistent training configurations, data splits, and evaluation procedures to ensure a fair comparison. The confusion matrix analysis indicates that DeBERTa-v3 produces fewer false positives and false negatives, demonstrating stronger capability in recognizing implicit stress expressions. These findings highlight the advantages of transformer-based models in capturing emotional and semantic context and indicate the potential for real-time deployment in social-media-based mental health monitoring systems.