Hybrid and Supervised NER for Old English: Methods, Results, and Replication

Author : Javier Martín Arista, Darío Metola Rodríguez, Daniel B. Morris

Abstract :We study NER for Old English, a setting where spelling drifts, inflection and compounding are pervasive, and annotated material is limited, thus disrupting tokenisation, lemmatisation and span decisions (Pettersson, 2016; Piotrowski, 2012). Using the UD conformant OEDT resource (≈50k tokens) with BIO splits, we keep the token segmentation and evaluate a unified tagset for five types: PERSON, PLACE, ORGANISATION, RELIGIOUS, and OBJECT. Scoring uses exact-span precision, recall, and F1; token-level F1 is reported. We compare two systems. The supervised baseline is a Stanza sequence tagger that couples word representations with character level context and trains directly on the BIO files with fixed UD segmentation (Qi et al., 2020; Lample et al., 2016; Ma & Hovy, 2016). The hybrid method fuses three sources: morphology- and syntax-sensitive priors from UD features and dependencies; a zero-shot pass from a contemporary English transformer NER model with subword-to-token aggregation (Devlin et al., 2019); and a cautious ecclesiastical lexicon targeting religious names and offices. Spans are merged by confidence with a mild preference for longer, coherent mentions and mapped to the unified label set. The Stanza model achieves 89.69 entity-level F1 on the test set (precision 91.53, recall 87.91) and 89.29 token-level F1. By class, RELIGIOUS is near saturation, PERSON is robust, PLACE lags due to historical toponyms and compounding, and ORGANISATION/ OBJECT are constrained by sparsity. The hybrid covers 85% of gold entities with the correct type; among covered items, 78% receive exact boundaries, yielding 66% exact-span recall overall. The morphology-aware cues promote tighter spans, the transformer adds recall but favours short segments, and the ecclesiastical component contributes high-precision matches within scope. We provide a UD-aligned, reproducible workflow, a strong baseline for Old English NER, and a compact hybrid that improves coverage and boundary fidelity. Remaining gaps cluster in toponyms and institutional names. We outline extensions to byte- or character-level transformers, multi-task learning with morphology, document-level coherence via linking, and broader ParCorOEv3 coverage. We conclude that lightweight morphological priors and conservative domain knowledge complement neural taggers in early English

Keywords :Old English NER, named entity recognition, Universal Dependencies, morphology, hybrid models

Conference Name :International Conference on General Artificial Intelligence (ICGAI - 25)

Conference Place Shanghai, China

Conference Date 20th Dec 2025