The Learnability Hierarchy of News Values: What Makes Some Journalistic Concepts Harder to Classify?

This study analyzes the performance patterns of BERT-based classifiers trained to identify news values in Danish journalism, revealing a systematic learnability hierarchy among journalistic concepts. We trained multilabel classifiers on 59,108 LLM-annotated sentences across 10 news values and 62 subcategories, using perturbation-based analysis to examine linguistic decision-making patterns. Results demonstrate three distinct performance tiers. High-performing classifiers like ’unexpectedness’ rely on consistent surface markers ("omvendt," "anderledes"), achieving reliable automated detection. Mid-tier classifiers such as ’personalization’ and ’timeliness’ show context-dependent but learnable patterns. Low-performing classifiers like ’eliteness’ struggle due to complex pragmatic reasoning requirements and severe class imbalance. Word importance analysis reveals that successful classifiers attribute significance to single words, while struggling classifiers distribute importance across many words per sentence. These findings suggest that computational journalism requires methodological pluralism""combining transformer models with sentiment analysis and named entity recognition based on concept complexity rather than applying uniform approaches to all news values.