Improving Human Happiness Analysis Based on Transfer Learning: Algorithm Development and Validation.

Area of Science:

• Natural Language Processing
• Sentiment Analysis
• Affective Computing

Background:

• Happiness, a subjective emotional state, significantly impacts human life quality.
• Understanding happiness is crucial for advancing sentiment analysis and emotional intelligence.
• This research focuses on two key facets of happiness: Agency and Sociality.

Purpose of the Study:

• To develop and evaluate a novel model for detecting and analyzing the Agency and Sociality aspects of happiness.
• To expand conceptual definitions of happiness and enhance the understanding of human emotions.
• To improve the accuracy and effectiveness of sentiment analysis in the domain of positive emotions.

Main Methods:

• Proposed a short happiness detection model utilizing transfer learning with Bidirectional Encoder Representations from Transformers (BERT).
• Developed a semantically enhanced language model, happyBERT, by retraining BERT on unlabeled data.
• Employed fine-tuning of BERT and happyBERT for text classification, integrating models with a voting strategy and pseudo-data.

Main Results:

• The proposed approach demonstrated superior performance on the happyDB dataset compared to baseline methods.
• Achieved high accuracy and F1 scores in predicting the Agency aspect (0.8653 accuracy, 0.9126 F1).
• Achieved high accuracy and F1 scores in predicting the Sociality aspect (0.9367 accuracy, 0.9491 F1).

Conclusions:

• The developed approach is highly effective for analyzing happiness, achieving state-of-the-art performance.
• Transfer learning significantly enhances the capabilities of happiness analysis models.
• The findings contribute to a deeper understanding of emotional nuances in sentiment analysis.