Paper presented at KONVENS 2025

BamNLP at the German conference on computational linguistics in Hildesheim

Recently, we have attended KONVENS, the annual conference organized by German computational linguistics research societies. Johannes Schäfer presented our work on German Affective Narrative Generation where we investigate the adaptation of an English large language model prompting approach to the German language. Affective narrative generation aims to solve the task explaining different emotions being evoked in ambiguous event descriptions by generating contexts that make a particular interpretation more clear. We test different possibilities of localizing the English approach to German in order to create a dataset for contextualized emotion analysis. The results show that a substantial degree of localization is required, however, using localized German prompts does not neccessarily lead to the most natural output. While the multilingual large language model performs similar in German than in English at easier components of the task, it struggles with more difficult components, such as shaping emotions precisely in the added German contexts.

Johannes Schäfer is part of our research staff working in the project CEAT. The goal of this project is to understand how emotions are expressed in text and how we can computationally analyze them in modeling event description. With the abovementioned work we address the essential aspect of contexts and how they can influence the emotion interpretation.