Paper Summary: CIOL at SemEval-2025 Task 11

Multilingual Pre-trained Model Fusion for Text-based Emotion Recognition
Md. Iqramul Hoque, Mahfuz Ahmed Anik, Abdur Rahman, Azmine Toushik Wasi

Overview

This study addresses multilingual emotion detection challenges in SemEval-2025 Task 11, focusing on three tracks:

• Track A: Multi-label emotion classification.
• Track B: Emotion intensity prediction.
• Track C: Cross-lingual generalization.

The authors leverage language-specific transformer models to tackle overlapping emotions, intensity quantification, and low-resource language adaptation.

Key Contributions

• Models Used:
  • Track A: DistilRoBERTa (English), ruBERT (Russian), DehateBERT (Portuguese).
  • Track B: ruBERT (Russian), EmoBERT-CN (Chinese), XLM-Twitter-EmoEs (Spanish).
  • Track C: Multilingual EmotionBERT for cross-lingual transfer.
• Architecture: Transformer backbone + MLP classifier with language-specific tuning.

Results

• Track A:
  • Russian achieved the highest F1 (0.848), benefiting from emotion-rich pretraining.
  • Portuguese struggled (F1=0.2773) due to data scarcity.
• Track B:
  • Russian excelled (F1=0.8594), while Chinese (F1=0.483) faced intensity estimation challenges.
• Track C:
  • Cross-lingual adaptation underperformed (F1=0.26–0.31), highlighting syntactic divergence in low-resource languages.

Challenges & Insights

• Key Issues:
  • Overlapping emotions (e.g., anger vs. sadness).
  • Cultural/sarcastic nuances in emotion expression.
  • High computational demands and annotation inconsistencies.
• Error Analysis:
  • Misclassifications in ambiguous pairs (e.g., Portuguese "anger" confused with "joy").
  • Intensity prediction errors in morphologically complex languages (e.g., Chinese).

Future Directions

• Improve cross-lingual representations.
• Address data scarcity via synthetic data or multimodal integration.
• Enhance cultural and contextual calibration.

Ethical Considerations

• Risks include bias propagation, privacy concerns, and misuse in surveillance.
• Emphasizes transparency and culturally aware AI governance.

GitHub Relevance

This work provides a blueprint for multilingual emotion recognition systems, emphasizing transformer-based architectures, language-specific tuning, and ethical AI practices. Code and model configurations (e.g., hidden dimensions, dropout rates) can guide implementations in similar NLP tasks.

View Full Paper | SemEval-2025 Task 11