Lang, AaronAaronLangJegan, RobinRobinJegan0000-0002-0388-7220Henrich, AndreasAndreasHenrich0000-0002-5074-32542025-11-122025-11-122025https://fis.uni-bamberg.de/handle/uniba/111251In contrast to similar tasks such as keyphrase prediction or text summarization, the automatic creation of marginalia has not been explored yet. This paper takes the first steps to advance research in this area. For our experiments, we used marginalia data extracted from German computer science textbooks and analyzed the marginalia’s properties. We utilized methods from the similar task of keyphrase prediction to test how suitable they are for the creation of marginalia and adapted the methods accord-ing to the observed marginalia properties and for the use with the German language. The results are evaluated quantitatively and quali-tatively. We also highlight limitations of frequently used measures such as F1 and describe why it is desirable to advance research on suit-able evaluation measures. We find that GPT-4o and mT5 perform best while for the non-LLM-based methods, SIFRank+ and TF-IDF show promising results.eng-conferenceobjecthttps://aclanthology.org/2025.konvens-1.20/