Integrating commonsense knowledge with GPT embeddings for emotion classification.

Area of Science:

• Natural Language Processing
• Affective Computing
• Artificial Intelligence

Background:

• Traditional emotion recognition models struggle with subtle and implicit emotional cues due to insufficient contextual understanding or external knowledge.
• Accurate emotion detection is crucial for various applications, including human-computer interaction and mental health monitoring.

Purpose of the Study:

• To develop a new fusion-based paradigm for enhanced emotion classification in text.
• To integrate contextual semantics with common sense knowledge for more robust emotion recognition.

Main Methods:

• Utilized GPT-based embeddings for capturing contextual meaning.
• Integrated external knowledge graphs (ConceptNet, COMET) for common sense reasoning.
• Developed a fusion model combining semantic and commonsense information for emotion classification.

Main Results:

• The proposed model significantly outperforms current baselines on the GoEmotions dataset for multi-label emotion classification.
• The fusion approach effectively connects surface-level language with deeper emotional states.
• Achieved improved accuracy in identifying emotions across seven categories: Joy, Sadness, Anger, Fear, Surprise, Disgust, and Neutral.

Conclusions:

• Combining commonsense reasoning with contextual semantics is vital for accurate and reliable emotion recognition, especially for ambiguous expressions.
• The fusion paradigm offers a more human-like approach to understanding emotions in text.
• This research advances affective computing by enabling more accurate and versatile emotion classification.