grc-macronizer-char: a small char-level model for Ancient Greek vowel-length annotation
This model predicts the phonemic length (long/short/undetermined) of the three Ancient Greek "dichrona" -- alpha (α), iota (ι), and upsilon (υ) -- whose length is not disambiguated by the standard Greek script. It was trained on Ancient Greek text automatically macronized by the rule-based grc-macronizer, applied to the Opera Graeca Adnotata corpus (Celano 2024), and is intended as a complementary, corpus-general alternative to that rule-based system.
Architecture
A small transformer encoder (~0.9M parameters) operating over three character-level "planes" per input position:
- plane 1 (letter): which of the 24 Greek letters (final/medial sigma folded together), or space, or "other" (punctuation/digits/foreign chars)
- plane 2 (diacritic): the combination of accent/breathing/diaeresis marks on that letter, if any (a small vocabulary fit from the training data)
- target (macron): none / short / long -- predicted only at positions that are genuine ambiguous dichrona (diphthong members and circumflexed vowels are excluded, matching the rule-based system's own definition)
During training, the diacritic plane is randomly masked (accents stripped) for a fraction of characters, so the model also learns to macronize unaccented input.
Training data
Training labels come directly from the rule-based macronizer's own output: wherever it marks a dichronon (it does not guess -- see the macronizer paper), that becomes a training label; dichrona it left unmarked are excluded from the loss entirely (never treated as a negative "short" or "long" example), since the rule-based system is known to be incomplete rather than wrong where it commits.
Evaluation
Evaluated on the full Norma Syllabarum Graecarum
benchmark, spanning all 17 manually-annotated works (prose and verse; 2,957
gold-annotated positions). Scored with the identical harness
(macron_model/eval_norma.py) against the bundled rule-based macronizer,
counting a left-unmarked dichronon as an implicit "short" guess for both
systems (default_short_accuracy below) so abstention isn't unfairly
penalized either way:
| system | accuracy (unmarked defaults to short) | raw accuracy | unmarked / total |
|---|---|---|---|
| rule-based grc-macronizer | 86.84% | 53.77% | 1,300 / 2,957 |
| this model | 89.89% | 89.79% | 9 / 2,957 |
The model outperforms the rule-based system on the fair metric while resolving essentially every position (99.7% coverage vs. 56%), since it learned to generalize past the cases the rule-based system couldn't commit to. The margin is statistically significant (McNemar's test, continuity-corrected, χ²=10.04, p=0.0015).
Two additional training runs with different random seeds confirm this is not a lucky draw: defaults-to-short accuracy across three seeds averages 89.36±0.78% (88.47% / 89.72% / 89.89%, this checkpoint), with every seed beating the rule-based teacher.
Usage
import sys
sys.path.insert(0, "path/to/grc-macronizer/macron_model") # for predict.py
from predict import MacronPredictor
predictor = MacronPredictor("path/to/downloaded/checkpoint")
macronized = predictor.macronize("ανθρωπος ανηρ")
# -> "α^νθρωπος α_νηρ"
MacronPredictor expects the checkpoint directory (containing
config.json/model.safetensors) plus a sibling or parent diacritic_vocab.json
(see predict.py for the exact lookup logic). The predict.py/tokenizer.py
source lives in grc-macronizer/macron_model.
Citation
If you use this model, please cite:
Thörn Cleland, Albin and Eric Cullhed (forthcoming). Automatic Annotation of Ancient Greek Vowel Length.
License
GNU GPL v3, matching grc-macronizer (the training-data generator this model's code ships alongside).