Exploring the Use of Named Entity Recognition on Swedish House Excavation Reports

Sammanfattning

The increasing digitization of archaeological archives has created a significant need for efficient methods to extract structured data from unstructured documentations. This thesis explores the feasibility of using Large Language Models (LLMs) to perform Named Entity Recognition (NER) on Swedish house excavation reports. Two Swedish transformer-based models, bert-base-swedish-cased (an Bidirectional Encoder Representation from Transformers (BERT) model ) and roberta-large-1160k ( an Optimized BERT Pretraining Approach (RoBERTa) model) , were fine-tuned on a dataset of annotated archaeological reports. Their performance was measured and compared against the state-of-the-art commercial LLM, Gemini 3 Pro, in a zero-shot setting. The results showed that the fine-tuned models are highly capable of performing this task, with the smaller BERT model outperforming the larger RoBERTa, achieving a weighted F1-score of 0.838. In contrast, the commercial model struggled at extracting exact entity spans, being significantly outperformed by the BERT model. The final result seems to indicate that the use of language models for information extraction in Swedish archaeology is highly viable.