UncCy: neural parsing of New Testament Greek manuscripts via collation-based annotation transfer

Abstract Ancient Greek manuscripts present major challenges for computational analysis due to the lack of diacritics, orthographic variation, and scribal conventions. Consequently, current state-of-the-art parsers perform poorly at lemmatization and morphological analysis. In this paper, we address this challenge by developing a new method that automatically collates annotated printed editions and manuscript traditions to generate silver-annotated datasets, using New Testament uncial manuscripts written in Koine Greek as our use case. After training a neural parser on this dataset, called UncCy (Uncial-Spacy), we demonstrate on a manually annotated manuscript that our parser reaches 93.6% performance on lemmatization, 96.9% on morphology, and 98.8% on POS, which corresponds to a gain of +40 pp in lemmatization, +34 pp in morphology, and +8 pp in Part-of-Speech (POS) compared to the current state-of-the-art parsers. This study thus provides the first production-ready parser for the study of New Testament uncials, opening the perspective of studying hundreds of currently unexploited transcriptions. All the resources generated during this study are fully available as Open-Source software and datasets.