The retrained Old English Lemmatiser released by the Nerthus Project is a character-level Transformer encoder–decoder that maps an inflected Old English word form onto its lemma and, at the same time, predicts its Universal Dependencies part-of-speech (UPOS) tag. Trained on the lemmatised prose and poetry of the Project's own corpora and augmented with morphology-aware data, the improved model reaches 67.58% lemma accuracy and 82.75% POS accuracy on a held-out corpus-only test set, a substantial result for a language in which strong-verb ablaut, orthographic variation and the prefix ge- account for most of the residual errors. Some pending questions for future research include: words are processed in isolation, without syntactic context, and proper nouns and hapax legomena remain the hardest categories, while around eleven per cent of inconsistently lemmatised types in the gold data effectively cap the attainable accuracy. The lemmatiser is distributed under an MIT licence (code and weights only), is fully documented with respect to architecture, preprocessing and training procedure, and is immediately useful for concordancing, frequency analysis and dictionary linking in Old English corpus linguistics and the digital humanities more broadly.

The Universal Dependencies parser of Old English is a complete spaCy pipeline that performs tokenisation, part-of-speech tagging, morphological analysis, lemmatisation and dependency parsing of Old English text. Its development follows the comparison of neural architectures in which a pipeline with a monolingually pretrained tok2vec component consistently outperformed both a baseline configuration and a MobileBERT transformer, achieving 83.24% unlabelled attachment score (UAS) and 74.23% labelled attachment score (LAS) when trained on 20,000 words. Tagging tasks reach accuracies in the region of 85–90%, and the published scaling analysis, which shows steep gains between 1,000 and 5,000 words and diminishes returns beyond 10,000, provides an evidence base for resource planning in historical-language NLP. The parser is trained on data drawn from ParCorOEv3 and the OEDT treebank. Because it is packaged for spaCy, it can be installed directly from Hugging Face and run on arbitrary Old English text, which opens the way to the automatic annotation of the roughly three million words of extant Old English prose and verse.

The Old English Dependency Treebank (OEDT) is a Universal Dependencies treebank of some 50,000 words of Old English (68,689 CoNLL-U lines) whose texts are retrieved from ParCorOEv3, the open-access annotated parallel corpus of Old English and English compiled by the Nerthus Project. The selection balances vernacular prose (the Anglo-Saxon Chronicle A, the Laws and Ælfric's Catholic Homilies I) with translated material, namely Orosius and the Gospel of St Mark, so that the main registers of Old English prose are represented. One half of the treebank is a thoroughly revised version of the manual annotation produced for Sara Domínguez Barragáns doctoral dissertation on the UD parsing of Old English (University of La Rioja, 2024); the other half was annotated automatically and then revised by hand by the ParCorOE team. Annotation decisions specific to Old English are documented explicitly: multi-word tokens are not used, contractions of the negative ne with pronouns, verbs and adverbs are captured by the feature Polarity=Neg, lemmas follow the headword list of the lexical database Nerthus, and every sentence carries an English translation in its comment lines. With a standard split into training, development and test sets, the OEDT is both a gold-standard benchmark for parser evaluation and a searchable syntactic resource for the study of Old English word order, case and clause structure.