SemEval-2025 Task 10
Multilingual Narratives in Online News
SemEval-2025 Task 10: Multilingual Characterization and Extraction of Narratives from Online News
Purdue University Fort Wayne, 2025
Built multilingual transformer-based pipelines for analyzing narrative framing in online news across five languages (English, Georgian, German, Greek, and Turkish). This work demonstrated how representation choices and evaluation design significantly impact model behavior in politically sensitive text.
Task Components
- Entity-Role Framing: Classify how entities are framed (hero, victim, villain) in news narratives
- Narrative Classification: Identify overarching narrative themes in news articles
- Narrative Extraction: Extract specific narrative elements and their relationships
Technical Approach
- XLM-RoBERTa and multilingual BERT architectures
- Cross-lingual transfer learning strategies
- Language-specific fine-tuning with shared representations
- Systematic evaluation across linguistic and cultural contexts
Results
- Achieved approximately 8× improvement over baseline performance
- Demonstrated effective cross-lingual transfer for narrative understanding
- Identified challenges in handling cultural and linguistic nuances in framing
Insights
- Representation choices critically affect model interpretations of politically sensitive content
- Evaluation design must account for cultural context and linguistic variation
- Cross-lingual models can transfer narrative understanding across languages with proper tuning
Skills & Tools
Python, Transformers, XLM-RoBERTa, BERT, Multilingual NLP, Cross-lingual Transfer, Narrative Analysis