Low-resource historical named entity recognition in the era of large language modelsŠmajdek, Uroš (Avtor) Munda, Tina (Avtor) Marolt, Matija (Avtor) digital humanitieshistorical documentsnamed entity recognitionlarge language modelsWe present a comparative evaluation of encoder-based and generative approaches for named entity recognition (NER) on historical Slovene parliamentary records. Using a manually annotated subset of the Carniolan Provincial Assembly Corpus Kranjska 1.0, we evaluate a range of locally deployable open-weight models within a unified multishot prompting framework. Our experiments compare traditional supervised encoderbased systems with modern instruction-tuned generative models across both strict and fuzzy evaluation settings. The results show that larger generative models achieve the best overall performance and demonstrate strong multilingual transfer capabilities, while smaller encoder-based systems remain highly competitive in terms of efficiency, inference speed, and person recognition accuracy. However, all models struggle with the extraction of more ambiguous entity types in historical texts, such as locations and organizations.20262026-10-07 11:48:28Drugo189467UDK: 004.8:81'33:930.2COBISS_ID: 293950979OceCobissID: 292667395sl