Nested Named Entity Recognition for Historical Swedish Texts: Evaluating Task-Specific Models and Large Language Models on the SWENER-1800 Dataset

This thesis investigates the performance of task-specific and general purpose models on the Nested Named Entity Recognition (NNER) task in historical Swedish texts. Using the SWENER-1800 dataset, an annotated corpus that displays the linguistic complexity of 18th- and 19th-century Swedish, the study compares a State-of-the-Art task-specific model (BINDER) against a selection of Large Language Models (GPT4-o, DeepSeek R1, and LLama 3), under zero-shot, few-shot, and fine-tuned conditions where applicable. The analysis evaluates each model's ability to detect nested (and flat) entities in a low-resource, linguistically specific and diverse setting. Results indicate that that while LLMs offer flexibility and reasonable performance in certain configurations, task-specific architectures like BINDER still outperform them, especially in capturing nested structures and, sometimes, rare entity types. Confusion matrix analysis reveals consistent semantic confusions among similar entity classes and highlights the structural limitations of LLMs in handling (deeply) nested entities. This work reinforces the value of traditional NLP methodologies for domain-specific, structured tasks.