From Fiction to Fact: Fine-Grained Emotion Classification in COVID-19 Newspaper Discourse

ACL ARR 2025 February Submission4655 Authors

Abstract: This study examines how a computation-al literary studies (CLS) emotion classifi-cation framework can be adapted to ana-lyze newspaper discourse on COVID-19. We developed and tested single-layer and dual-layer BERT models to classify emo-tions at two levels: 9 primary emotion families (Level 1) and 87 subcategories (Level 2). Using 7,498 sentences from German newspapers, data sparsity di-rected our focus to the 10 most common Level-2 emotions. Our results revealed varied model performances across emo-tion categories. The single-layer model exhibited more consistent performance and a stronger correlation with emotion frequency. In contrast, the dual-layer model excelled at distinguishing specific emotions like interest, curiosity, and hope, although with greater variability. Both models struggled to recognize more complex emotions such as LOVE, DIS-GUST, and AMBIVALENCE. Our results un-derscore the complexities and potential of automated emotion detection in media discourse, highlighting the need for do-main-specific classification methods.

Paper Type: Long

Research Area: Computational Social Science and Cultural Analytics

Research Area Keywords: emotion detection and analysis, quantitative analyses of news and/or social media

Contribution Types: Model analysis & interpretability, Data analysis

Languages Studied: German

Submission Number: 4655