The paper introduces the GerFuN dataset, consisting of five German-language novels fully annotated for character coreference, comprising a total of 450,000 tokens. Using a semi-manual pipeline, we first pre-annotated the novels using a LLM, and then manually corrected the annotations in the INCEpTION tool. The annotation guidelines, which build on existing approaches but are made more explicit and refined, are presented in the paper and released alongside the dataset. Finally, we evaluate LLMs on GerFuN, which surpass previous pipelines and exhibit near-human accuracy on prototypical cases of particular interest to literary studies.
Hilger et al. (Tue,) studied this question.