Multi task opinion enhanced hybrid BERT model for mental health analysis

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Mental Health Informatics

Background:

• Analyzing user-generated content for mental health insights is vital but challenging.
• Current models often overlook integrating diverse viewpoints, limiting comprehensive mental health assessment.
• Single-task learning approaches in text data analysis for mental health fail to capture the complexity of user emotions and perspectives.

Purpose of the Study:

• To develop a more thorough understanding of mental health by integrating user viewpoints and sentiments.
• To introduce the Opinion-Enhanced Hybrid BERT Model (Opinion-BERT) for simultaneous sentiment and status categorization.
• To enhance the accuracy of mental health assessments through multi-task learning.

Main Methods:

• Developed Opinion-BERT, a hybrid model utilizing multi-task learning for sentiment and status classification.
• Extracted and dynamically constructed opinion embeddings using TextBlob and SciPy to complement pre-trained BERT.
• Integrated opinion embeddings with BERT's contextual embeddings via CNN and BiGRU layers to capture local and sequential features.

Main Results:

• Opinion-BERT demonstrated superior performance compared to baseline models like BERT, RoBERTa, and DistilBERT.
• Achieved 96.77% accuracy in sentiment classification and 94.22% accuracy in status classification.
• Validated the crucial role of opinion-enhanced embeddings in improving multi-task learning performance for mental health analysis.

Conclusions:

• The Opinion-BERT model offers a more nuanced understanding of emotions and psychological states from user-generated text.
• Combining opinion and sentiment data within a multi-task learning framework significantly enhances mental health analysis.
• This approach highlights the potential for improved mental well-being assessment through advanced text data analysis techniques.