Συγκριτική μελέτη αλγορίθμων βαθιάς μάθησης
Comparative study of deep learning algorithms
View/
Keywords
Ανάλυση συναισθημάτων ; Bert ; Tokenization ; Padding ; Auroc ; F1-Score ; Μηχανική μάθηση ; Ταξινόμηση κειμένων ; Προεπεξεργασία δεδομένων ; Ensemble modelsAbstract
This study examines sentiment classification in text using advanced machine learning models such as BERT, Ensemble of BERT Models, and a Linear Layer. Three CSV datasets containing multi-label sentiment annotations were utilized. The data were processed through cleaning, tokenization, padding, and splitting into training and evaluation sets.
The methodology included:
• Training multiple BERT models to improve accuracy.
• Computing metrics such as AUROC and F1-score for model evaluation.
• Applying an Ensemble technique to boost model performance collectively.
The results demonstrate an accuracy reaching 0.93 AUROC for emotions like "admiration,"
"gratitude," and "love," significantly outperforming traditional methods. While limitations were noted for emotions such as "pride" and "relief," the Ensemble approach mitigated the lower performance in these categories.