Unveiling ideological extremes in parliamentary debates using transformer-based language models

This study explores how natural language processing can uncover ideological orientations in European parliamentary debates. Using speeches from 28 countries, transformer-based models (BERT, RoBERTa, and DistilBERT) were fine-tuned for binary classification and evaluated by accuracy and F1-score. RoBERTa and DistilBERT slightly outperformed BERT, and DistilBERT offered a strong balance between efficiency and accuracy. Principal Component Analysis and attention score analysis revealed cross-national linguistic and ideological patterns. The findings advance understanding of political discourse through computational linguistics, highlighting the potential of transformer models for analyzing ideological diversity in large-scale parliamentary data.