DepressionEmo: A novel dataset for multilabel classification of depression emotions

Area of Science:

• Computational linguistics
• Mental health informatics
• Affective computing

Background:

• Negative emotional states are linked to adverse mental health outcomes.
• Analyzing emotions in text is crucial for understanding mental health.
• Existing datasets may not fully capture the nuances of depression-related emotions.

Purpose of the Study:

• Introduce the novel DepressionEmo dataset for depression-related emotion detection.
• Facilitate research on the correlation between emotions and linguistic patterns in depression.
• Evaluate various machine learning and deep learning models for emotion classification.

Main Methods:

• Developed the DepressionEmo dataset with 6037 Reddit posts, using zero-shot classification and human validation.
• Performed correlation analysis between emotions and linguistic features.
• Implemented and compared machine learning (SVM, XGBoost, LightGBM) and deep learning (BERT, BART, GAN-BERT, T5) models.

Main Results:

• The DepressionEmo dataset demonstrated acceptable inter-rater reliability.
• BERT (bert-base-uncased) achieved an F1 Macro score of 0.76, matching BART but with higher efficiency.
• Suicide intent showed the highest F1 Macro score, highlighting the dataset's utility in identifying critical emotional states.

Conclusions:

• The DepressionEmo dataset provides a valuable resource for studying emotions in individuals with depression.
• Text analysis using models like BERT can effectively identify depression-related emotions.
• Further research can leverage this dataset to improve mental health monitoring and support.