Norma Syllabarum Graecarum - A Benchmark for grc Syllabification and Vowel Length Annotation

We introduce Norma as a common benchmark for the evaluation and comparison of NLP tools concerning markup of two tasks for Ancient Greek (grc): (1) vowel length of dichronic vowels (alpha, iota, ypsilon) in open syllables (where they impact syllable weight) and (2) syllabification, both boundaries and weight. This means that the benchmark also indirectly tests handling of sandhi phenomena and, for verse, metre.

The benchmark includes 1228 lines of text by each of the below 14 authors, spanning many periods and genres of both verse and prose:

• Alcman (Louvre Partheneion 36-49, mix of aeolic, trochaic and dactylic cola, NB: contains digamma, ϝ)
• Bacchylides (Epinikion 5, anapaesto-iambic and dactylo-troachaic)
• Aeschylus (Supplices 1-10, anapaests)
• Sophocles (Oedipus Tyrannus 1-20, iambic trimeter)
• Euripides (Bacchae 1-20, iambic trimeter)
• Thucydides (Thucydidis historiae 1.1.1.1-1.2.2.1, prose)
• Aristophanes (Acharnenses 1-16, iambic trimeter, and 299 lines from responding lyric songs)
• Plato (Cratylus 383-384a5, prose)
• Dioscorides (Anthologia Graeca 5.55, elegiac distichs)
• Plutarch (Stoicos absurdiora poetis dicere 1057.C1-F1, prose)
• Epictetus (Enchiridion 1.1.1-1.5.1, prose)
• Origenes (Contra Celsum 1.1-1.20, prose)
• Nonnus (Dionysiaca 1-20, dactylic hexameter)
• Quintus (Posthomerica 1-20, dactylic hexameter)

The texts have been chosen so that none (as of 2025) are included in the Hypotactic corpus, which means that a model trained on Hypotactic will not have any unfair advantage. Except for Bacchylides and Aristophanes, the texts are excerpts from works and not full works.

The 299 lines of Aristophanes' lyric poetry are borrowed from my project Aristophanis Cantica. For further details, see that repository.

The repository also provides a validation set, data/validation.jsonl, containing 764 lines from Aristophanes' lyric poetry for the macronization task.

Markup Format

For the syllabification task the markup is:

• square brackets [] enclose heavy syllables
• curly brackets {} enclose light syllables

and for the macronization task the markup is:

• carets ^ follow short ambiguous vowels (α, ι, υ)
• underscores _ follow long ambiguous vowels (α, ι, υ).

In prose, for ease of comparison we consider the sentence to be the equivalent of the verse line, with the consequence that the final syllable of a sentence will be considered heavy if it ends with a consonant, even though the following sentence begins with a vowel; i.e. a prose brevis in longo.

With regard to vowel length, the benchmark is designed to test performance of open (light) syllables only. This means that all open syllables with ambiguous vowels (α, ι, υ) have them marked with carets or underscores, while this is only sporadically true for closed syllables (and in such cases mostly before mute-liquid clusters).

We also have an example of a word that avoids the normal rules of syllabification:

• ...[ῇ]{σι^ }{Σκα^}[μάν][δρου,] (Posthomerica, 10)

How to cite

This benchmark is part of ongoing research projects by Albin Thörn Cleland at Lund university and Eric Cullhed at Uppsala university. Its new syllabified and macronized editions of ancient texts are made available under the copyleft GNU GPL licence. If you use this benchmark in your own research, please cite it in the following way:

Thörn Cleland, Albin and Eric Cullhed (forthcoming). Automatic Annotation of Ancient Greek Vowel Length.