Parsing Old English with Universal Dependencies. The Impact of Model Architectures and Dataset Sizes

This study evaluates the performance of Universal Dependencies (UD) parsing for Old English using three neural architectures across various dataset sizes. We compare a baseline spaCy pipeline, a pipeline with pretrained tok2vec component, and a MobileBERT transformer-based model on datasets ranging from 1,000 to 20,000 words. Our results demonstrate that the pretrained model consistently outperforms the alternatives, achieving 83.24% UAS and 74.23% LAS with the largest dataset. Performance analysis shows that basic tagging tasks reach 85-90% accuracy, while dependency parsing achieves approximately 75% accuracy. We observe significant improvements with increasing dataset size, though with diminishing returns beyond 10,000 words. The transformer-based approach underperforms in spite of its higher computational cost. This highlights the difficulties of applying modern NLP techniques to historical languages with limited training data. Our findings suggest that medium-complexity architectures with pretraining on raw text offer the optimal balance between performance and computational efficiency for Old English dependency parsing.