HunEmBERT (Political Emotion & Sentiment)

Emotions and sentiment play a crucial role in communication, and particularly in political communication, but manually annotating large-scale corpora is prohibitively slow and costly. Traditional lexicon-based sentiment approaches fail to capture the nuanced, context-dependent expression of political emotions and value-laden arguments. A scalable, domain-adapted solution was needed to support both research and applied analysis in this area.

We developed a fine-tuned variant of the Hungarian BERT model adapted specifically for Hungarian political texts, which also served as a preliminary study for our later multilingual solutions. The model was trained for both sentiment detection and emotion classification, using techniques such as learning-rate scheduling, early stopping, and systematic error analysis, with particular attention to non-literal meanings, idioms, irony, and sarcasm. By combining domain adaptation with transformer fine-tuning, HunEmBERT achieves robust performance even in noisy, real-world data.

The resulting model enables stable automatic recognition of emotions and value-laden arguments in Hungarian political discourse. It provides a reliable foundation for analyzing affective dimensions of political texts at scale, with applications ranging from academic research to media and policy analysis.
Results have been validated in peer-reviewed publications, and the model has since been applied in various practical settings, highlighting its relevance beyond academic evaluation.