This study investigates whether a high-quality, 19-label named entity recogniser for medieval Latin charters can be constructed using only a few hundred annotated sentences. The authors introduce "semantic scaffolding," an innovation that utilizes richly descriptive English label phrases as prompts to activate latent multilingual knowledge within the model. This is paired with a custom span-based architecture utilizing XLM-ROBERTa-large, 4-head attention pooling to handle long property descriptions, and a hybrid loss system including Asymmetric Focal-Dice and InfoNCE contrastive terms. Results demonstrate that semantic scaffolding enables fine-tuned GLiNER to reach 80.8% overlap F1, while the custom architecture achieves 83.4% overlap F1 using only 298 training sentences. Significantly, the paper provides an empirical demonstration that domain-specific pre-training on medieval Latin offers no performance advantage once task-specific fine-tuning is applied. While the model excels at frequent categories like PER (95.7% F1) and LOC (93.5% F1), challenges persist for rare, position-dependent legal categories such as LEG (53.1% F1) and TRANS (52.6% F1).

This paper explores the automated extraction of job titles from unstructured historical job advertisements, using a corpus of digitized German-language newspapers from 1850-1950. The study addresses the challenges of working with unstructured, OCR-processed historical data, contrasting with contemporary approaches that often use structured, digitally-born datasets when dealing with this text type. We compare four extraction methods: a dictionary-based approach, a rule-based approach, a named entity recognition (NER) mode, and a text-generation method. The NER approach, trained on manually annotated data, achieved the highest F1 score (0.944 using transformers model trained on GPU, 0.884 model trained on CPU), demonstrating its flexibility and ability to correctly identify job titles. The text-generation approach performs similarly (0.920). However, the rule-based (0.69) and dictionary-based (0.632) methods reach relatively high F1 Scores as well, while offering the advantage of not requiring extensive labeling of training data. The results highlight the complexities of extracting meaningful job titles from historical texts, with implications for further research into labor market trends and occupational history.