Transfer Learning for Historical Texts : Exploring Gender and Work in Old Swedish Manuscripts

Sammanfattning

Historical document digitization for 16th–17th century Swedish manuscripts presents significant challenges due to archaic letterforms, document degradation, and limited labeled training data. This thesis investigates the effectiveness of transfer learning approaches for adapting modern Handwritten Text Recognition (HTR) systems to historical Swedish manuscripts, focusing on tran- scription accuracy for archival datasets such as the “Gender and Work” collection. We compare two distinct approaches: (1) fine-tuning a flexible open-source pipeline (Loghi HTR) using progressive data scaling strategies (5k–200k lines) and selective layer-freezing techniques, and (2) evaluating a domain-pretrained system (Swedish Lion) for zero-shot performance on historical documents. Our experiments show that transfer learning substantially improves recognition accuracy, reduc- ing Character Error Rate (CER) from over 90% in zero-shot scenarios to approximately 58% for Loghi HTR through systematic fine-tuning. The domain-pretrained Swedish Lion system achieves superior performance with 29.08% CER and 62.68% Word Error Rate (WER) without additional training, highlighting the advantages of domain-specific pretraining. Through controlled experiments on data scaling, freezing strategies, and decoding approaches (CTC vs. Word Beam Search), we identify diminishing returns beyond 100k samples, optimal con- figurations for computational efficiency, and trade-offs between accuracy and inference speed. We also report error characteristics and compute requirements relevant for deployment. The findings contribute insights into transfer learning dynamics for historical HTR and provide practical guidance for cultural-heritage digitization projects. By enabling more accurate tran- scription of Swedish manuscripts, this work supports future digital humanities research on early modern society.