Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·24d ago·5 min read
YNU-HPCC at SemEval-2025 Task 11: Bridging the Gap in Text-Based Emotion Using Multiple Prediction Headers
TL;DR
YNU-HPCC's team achieved a ranking score of 0.44 in SemEval-2025 Task 11 by utilizing a RoBERTa model with a single prediction head for text-based emotion recognition. Their findings indicate that a single head outperforms multiple heads for simultaneous emotion predictions.
✦ Why It Matters
Consider using a single prediction head in your emotion recognition models for improved accuracy.
Key Takeaways
- Utilize the RoBERTa model for emotion recognition tasks.
- Implement a single prediction head for better performance.
- Translate datasets uniformly to enhance model training.
- Analyze attention mechanisms to improve prediction accuracy.
Full Summary
YNU-HPCC participated in SemEval-2025 Task 11, focusing on text-based emotion recognition. They employed the RoBERTa model, an advanced version of BERT that uses a Transformer architecture, to enhance emotion prediction.
The team modified the model to utilize a single output head, allowing it to predict one emotion at a time. Their approach yielded a competitive ranking score of 0.44 across all languages.
Notably, they translated the entire dataset into English using Google Translate, which improved processing efficiency. Analysis revealed that using a single prediction head was more effective than employing six heads for predicting multiple emotions simultaneously.
Additionally, training on a uniformly translated dataset led to better performance compared to the original dataset.