This research introduces a few-shot learning approach for detecting rare pandemic-related emotions in social media, suggesting an effective method for overcoming class imbalance.
Key Points
IAPN achieved a macro-F1 improvement of 4.8 points, enhancing emotion detection for rare classes.
The model combines idiom-aware semantic enrichment with contextual XLM-RoBERTa embeddings to classify tweets.
A total of 480,000 COVID-19 tweets were organized into episodic tasks reflecting real-world emotion distribution.
Integration of metric learning with idiomatic knowledge significantly boosts capture of subtle emotional signals.