Selected Conferences

2026

Martín Arista, J.

International Conference LAnguage TEchnologies for Low-resource Languages

Abstract: Old English (c. 450–1150 CE), preserved in approximately three million words, is a historical language that poses distinctive problems for natural language processing: pervasive spelling variation, rich inflectional morphology, and heterogeneous annotated sources. This paper provides an analysis of the data curation pipeline, annotation quality, and linguistically motivated error patterns for Old English lemmatisation. We describe the extraction, harmonisation, and validation of training data from three heterogeneous historical sources, including a parsed corpus, a scholarly dictionary, and a historical dictionary, each with different formats, POS conventions, and coverage. Quantitative analysis shows that 11.3% of word types receive inconsistent lemma assignments, identity transformations account for 28.3% of training data, and the ge- prefix affects 11.7% of word types. Error analysis organised by morphological category shows that strong-verb ablaut accounts for approximately 25% of lemmatisation errors, spelling variation for 20%, and prefix handling for 15%. We contextualise these findings within the broader landscape of low-resource NLP and formulate practical guidelines for building lemmatisation systems for historical low-resource languages, including recommendations on annotation consistency, POS completeness, and source harmonisation.

2025

Martín Arista, J., Metola Rodriguez, D., & Morris, D. B.

Proceedings of the IEEE International Conference on Big Data, BigData, 5390-5399.

Abstract: Named EntityRecognition (NER) for Old English presents distinctive problems arising from orthographic variability, rich inflectional morphology, sparse annotation, and genre-specific lexica that render modern transfer approaches unreliable. This study presents a reproducible integrated pipeline integrating morphology-aware grammatical analysis via Stanza, cross-Germanic transfer learning from contemporary English and German transformer models, and pattern matching adapted to ecclesiastical discourse. A span-level fusion algorithm reconciles overlapping predictions by confidence and span length and normalises heterogeneous label sets to a unified five-category schema comprising PERSON, PLACE, ORGANISATION, RELIGIOUS, and OBJECT. Evaluated on prose from TheAnglo-Saxon Chronicle, the integrated system achieves 85 per cent entitycoverage and 78 per cent boundary accuracy and outperforms individual components. In parallel supervised experiments, a character language model-augmented Stanza NER tagger trained on the Old English OEDT corpus (approximately 50,000 words) reaches entity-level F1 of 89.16 on development data and 89.69 on held-out test material. It shows near-ceiling performance for RELIGIOUS entities (98-99 per cent F1) and excellent results for PERSON (~85 per cent F1). PLACE, ORGANISATION, and OBJECT remain difficult due to diachronic variation, annotation sparsity, and tagset inconsistencies. The study provides quantitative results, per-category confusion matrices and checkpoint stability analyses. This work establishes the first systematically evaluated NER baseline for Old English and demonstrates practical pathways for pre-modern corpus annotation.

2025

Martín Arista, J., Metola Rodriguez, D., & & Ojangure López, A. E.

Proceedings of the IEEE International Conference on Big Data, BigData, 5482-5489.

Abstract: This study presents acomparative evaluation of spaCy and Stanza for Old English, a morphologically rich, orthographically variable, low-resource language. We benchmark five configurations spanning character-level and biaffine models in Stanza, and scratch, tok2vec-pretrained, and scratch-transformer pipelines in spaCy, under consistent Universal Dependencies (UD) evaluation. Beyond reporting scores, we analyse why each system gains or fails across tokenisation, sentence segmentation, part-of-speech and morphology, lemmatisation, and dependency parsing. The key findings are that Stanza character-level stack scales strongly in structural accuracy with more annotated data, raising labelled attachment score (LAS) by 8.82 points from roughly 25K to 50K tokens, while spaCy benefits substantially from unsupervised, in-domain tok2vec pretraining at 20K tokens, especially in POS, morphology, and parsing. In contrast, lemmatisation and fine-grained XPOS remain brittle and do not improve monotonically, reflecting inventory instability and orthographic variation. We include training schedules, checkpoint choice, and an error analysis disentangle data growth from pipeline-specific inductive biases, and position the results against multilingual transfer baselines for Old English. Our conclusions align with broader evidence that character-aware models and UD-optimised parsers remain strong options for low-resource, morphologically rich languages, and that light in-domain pretraining can reduce sample complexity when large-scale masked language model training is infeasible.