Transformer transfer learning emotion detection model: synchronizing socially agreed and self-reported emotions in...

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Traditional emotion detection relies on small, human-annotated datasets, prone to annotator bias.
• Large self-reported emotion datasets offer scale but may miss nuanced social emotions.
• Existing methods face challenges in achieving robust emotion detection across varied contexts.

Purpose of the Study:

• To develop a novel transformer transfer learning approach for enhanced emotion detection.
• To integrate self-reported emotions with human-annotated social emotions for model training.
• To improve the generalizability and performance of emotion detection models.

Main Methods:

• A novel transformer transfer learning model was proposed, mimicking human developmental stages.
• The model first detects author self-reported emotions from large datasets.
• It then synchronizes with social emotions identified in human-annotated datasets.

Main Results:

• The proposed model demonstrated strong performance on a large, novel self-reported emotion dataset (n=3,654,544).
• Performance was validated across 10 previously published emotion detection datasets.
• The transfer learning approach effectively improved emotion detection capabilities.

Conclusions:

• The novel transformer transfer learning method offers a promising approach for robust emotion detection.
• Integrating self-reported and socially annotated emotions enhances model performance and context adaptability.
• This method addresses limitations of both small-scale annotation and large-scale self-reported data.