Model Card for old-english-lemmatizer
A character-level Transformer encoder–decoder that lemmatises Old English
(c. 450–1150 CE) and jointly predicts a Universal Dependencies (UPOS) tag. It
maps the character sequence of an inflected form to the character sequence of its
lemma, processing words in isolation (context-free). This is the best (improved)
model behind the two companion publications, reaching 67.58 % lemma accuracy
and 82.75 % POS accuracy on the held-out corpus-only test set.
Model Details
Model Description
- Developed by: Javier Martín Arista (Nerthus Project, Universidad de La Rioja)
- Funded by: Grant PID2023-149762NB-100 (Agencia Estatal de Investigación); grant PRXX24/00108 (Ministerio de Ciencia, Innovación y Universidades)
- Model type: Character-level Transformer encoder–decoder with a joint UPOS classification head (sequence-to-sequence character transduction)
- Language(s): Old English (
ang)
- License: MIT (source code and released weights only; not the training corpora)
- Reference version: Where the two companion papers differ, the conference paper (LaTeLL 2026) is authoritative.
Model Sources
- Repository: https://huggingface.co/Nerthus-Project/old-english-lemmatizer
- Paper (conference, reference): Martín Arista, J. (2026). Data Curation, Annotation Quality, and Error Patterns in Old English Automatic Lemmatisation. LaTeLL 2026.
- Paper (journal): Martín Arista, J. (2026). Neural Architectures for Old English Lemmatisation: A Comparative Study of Corpus-Based and Dictionary-Augmented Approaches. Procesamiento del Lenguaje Natural (SEPLN).
Uses
Direct Use
Lemmatisation and UPOS tagging of individual Old English word forms for corpus
linguistics and digital-humanities workflows: concordance generation, frequency
analysis, and dictionary linking. Input is an inflected form; output is
(lemma, UPOS).
Downstream Use
As a preprocessing component feeding search/concordancing tools, editorial
workflows, or lexical-resource linking over digitised Old English texts.
Out-of-Scope Use
- Sentence-level or context-dependent disambiguation of homographs (the model is
context-free; e.g. it cannot resolve þā demonstrative vs. adverb from context).
- Languages or historical stages other than Old English.
- Named-entity-aware lemmatisation (proper nouns are the weakest category).
Bias, Risks, and Limitations
- Processes words in isolation; no syntactic context.
- Errors concentrate in strong-verb ablaut (~25 %), spelling variation (~20 %),
and prefix (
ge-) handling (~15 %); proper nouns and hapax legomena are hardest.
- The gold data contain ~11.3 % inconsistently lemmatised word types, which caps
achievable accuracy (human agreement ≈ 90–95 %).
- Evaluated only on a held-out split of the YCOE; behaviour on independently
collected texts or other editions may differ.
Recommendations
Treat outputs as high-quality suggestions to be validated for ambiguous forms,
proper nouns, and rare words. For homographs requiring context, add
post-processing or human review.
How to Get Started with the Model
The weights (model.pt) are a raw PyTorch state_dict; load them with the
bundled code (this is not a transformers AutoModel).
1# pip install -r requirements.txt (torch>=2.0)
2from lemmatizer import OldEnglishLemmatizer
3
4lem = OldEnglishLemmatizer.from_pretrained(".") # folder with config.json, char_vocab.json, model.pt
5
6lem.lemmatize("cyninges")
7# ('cyning', 'NOUN')
8
9lem.lemmatize_batch(["cwæð", "folce", "cyning"])
10# [('cweþan', 'VERB'), ('folc', 'NOUN'), ('cyning', 'NOUN')]
11
12lem.lemmatize_text("se cyning cwæð to his folce")
13# [{'word': 'se', 'lemma': ..., 'pos': ...}, ...]
Training Details
Training Data
Trained on the corpus-only condition: word–lemma–POS triples extracted from
the York–Toronto–Helsinki Parsed Corpus of Old English (YCOE, prose + poetry).
The dictionaries and corpora cannot be redistributed and are not included
here; obtain them from their official providers:
Dataset statistics (corpus-only, after harmonisation): 105,063 annotated
instances; 84,050 / 10,506 / 10,507 train/dev/test (80/10/10, seed 42); 90,752
unique word types; 37,627 unique lemmas; 28.3 % identity transformations; 4.4 %
unknown POS; 11.7 % ge- prefix; 11.3 % inconsistent types.
Training Procedure
Preprocessing
Extract triples from parsed trees (prose columns LEMMA–WORD–POS, poetry
WORD–LEMMA–POS); map all POS labels to UPOS (unresolved → X); Unicode-normalise
and case-fold while preserving macrons; normalise 221 parenthetical (ge)
lemmas; remove exact-duplicate triples; 80/10/10 random split with seed 42.
Training Hyperparameters
- Training regime: fp32
- Epochs: 80 (best at epoch 74)
- Batch size: 24 (effective 72, gradient accumulation 3)
- Optimizer: AdamW, weight decay 0.01
- Learning rate: 1.5e-4, cosine annealing with warm restarts (range 1e-5–1.5e-4, 4 cycles, peaks at epochs 22/41/58/74), 5 % warmup
- Label smoothing: 0.15
- POS loss weight: 0.2
- Data augmentation: character-level (substitution
þ↔ð, æ↔a/e, i↔y; deletion; adjacent swap; duplication), p=0.2 per word, re-applied each epoch
- Seed: 42
Evaluation
Testing Data, Factors & Metrics
- Testing data: YCOE corpus-only held-out test split (10,507 instances).
- Metrics: exact-match lemma accuracy (after normalisation) and exact-match UPOS accuracy.
Results
| Model | Dev lemma | Test lemma | Test POS |
|---|
| This model (Transformer + augmentation + cyclical LR) | 68.68 % | 67.58 % | 82.75 % |
| Transformer baseline | 66.65 % | — | — |
| Seq2Seq (BiLSTM + attention) baseline | 53.15 % | — | — |
Ablations (test accuracy): Word2Vec context fusion 63.53 % (−4.05); Word2Vec +
character concat 49.37 % (−12.67); BPE 500-merge 54.66 % (−12.92).
Approximate accuracy by POS: ADP ~85 %, CCONJ ~83 %, DET ~80 %, ADV ~75 %,
NOUN ~68 %, ADJ ~65 %, VERB ~55 %, PROPN ~45 %.
Summary
A corpus-only, character-level Transformer with morphology-aware augmentation and
cyclical learning rates gives the best reported accuracy for Old English
lemmatisation on the YCOE; word embeddings and subword tokenisation both hurt.
Technical Specifications
Model Architecture and Objective
Character-level Transformer encoder–decoder (d_model 512, 8 heads, 4 encoder +
4 decoder layers, feed-forward 1024, dropout 0.15, sinusoidal positional
encoding) with an auxiliary UPOS classification head. Character vocabulary of 140
symbols (incl. þ ð æ ƿ, macron vowels, Latin script, digits, punctuation, and 4
special tokens); 18 UPOS tags; ≈ 26.5 M parameters. Objective: character-level
cross-entropy for the lemma plus cross-entropy for the UPOS tag. Released weights
are used in single-word (context-free) mode.
Reconciliation note. The papers describe the model with a "98-symbol"
vocabulary and ≈ 21.5 M parameters (the plain Transformer baseline). The
released best checkpoint ships a 140-entry operational vocabulary and has
≈ 26.5 M parameters. Evaluation figures are unaffected. See
NOTES_ON_PUBLICATIONS.md.
Compute Infrastructure
- Software: Python ≥ 3.9, PyTorch ≥ 2.0.
Citation
Please cite both works.
BibTeX:
1@inproceedings{MartinArista2026LaTeLL,
2 title = {Data Curation, Annotation Quality, and Error Patterns in Old English Automatic Lemmatisation},
3 author = {Mart{\'i}n Arista, Javier},
4 booktitle = {Proceedings of LaTeLL 2026 (Workshop on Language Technologies for Historical and Ancient Languages)},
5 year = {2026}
6}
7
8@article{MartinArista2026PLN,
9 title = {Neural Architectures for Old English Lemmatisation: A Comparative Study of Corpus-Based and Dictionary-Augmented Approaches},
10 author = {Mart{\'i}n Arista, Javier},
11 journal = {Procesamiento del Lenguaje Natural},
12 year = {2026}
13}
Model Card Authors
Javier Martín Arista (Nerthus Project, Universidad de La Rioja).
Model Card Contact
