Ensemble of transformers for depression emotion classification.

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Mental Health Research

Background:

• Emotions are powerful motivators influencing behavior and social interactions.
• Co-occurring emotions are common in mental health conditions like depression.
• Analyzing multiple emotions in depression offers deeper insights than single-emotion analysis.

Purpose of the Study:

• To analyze co-occurring emotions in depression using artificial intelligence.
• To propose and evaluate a transformer-based ensemble model for predicting multiple emotional tendencies in depression.

Main Methods:

• Utilized the public DepressionEmo dataset of user posts.
• Developed a heterogeneous ensemble learning architecture using stacking and majority voting.
• Employed pre-trained transformer models including DistilBERT, RoBERTa, Mental-BERT, Mental-RoBERTa, and DeBERTa.

Main Results:

• The transformer ensemble model demonstrated improved prediction performance over individual transformers.
• Ensemble approaches showed consistently better results, especially in macro-averaged F1 scores under class imbalance.
• Stacking-FFNN achieved the highest performance with a score of 0.8121, highlighting the effectiveness of ensemble learning.

Conclusions:

• Transformer ensemble models are effective for identifying multiple emotions in textual psychological data.
• Ensemble learning significantly enhances depression emotion classification accuracy, particularly in imbalanced datasets.
• This approach offers a promising avenue for more accurate mental health analysis and intervention.