Language-specific embeddings of Old English with character-level processing

This article seeks to contribute to narrowing the gap between philological research and computational linguistics by providing a neural network model for Old English analysis. The article makes three key innovations: (i) the development of language-specific word embeddings derived directly from The Dictionary of Old English Corpus ; (ii) a comparative analysis of character-level versus word-level models that demonstrates the superior performance of character-level processing for morphologically rich historical languages; and (iii) a comprehensive Stanza-based pipeline that outperforms previous approaches to Old English parsing. Our model achieves 88.92% Unlabeled Attachment Score and 79.65% Labeled Attachment Score on dependency parsing tasks, which represents approximately 20 percentage point improvement over previous state-of-the-art multilingual approaches. The main conclusion of this work is that language-specific resources and character-level modeling are more effective for Old English processing than cross-linguistic transfer learning. This opens new avenues for computational research in historical linguistics and digital humanities.