Telling classical from reformed Armenian by counting one letter

Armenian is printed in two orthographies with the same alphabet, and a book's title is no guide to which one its pages use. One letter is: the reform left ւ only one place to stand, so counting it anywhere else tells the two apart, on a page or in a single name

11 October 2026 · 11 min read · Stephen Masters

For most of its history Armenian had one spelling. The alphabet Mesrop Mashtots devised at the start of the fifth century was made for the classical language, Grabar, and its spelling stayed put while the spoken language moved on. By the nineteenth century there were two modern literary standards, both a long way from Grabar: Eastern Armenian, centred on Tiflis in the Russian Empire and spoken across Persia, and Western Armenian, centred on Constantinople in the Ottoman Empire. They differed in grammar and pronunciation, but both were written in the same classical spelling, including letters that had stopped being pronounced.

Soviet Armenia changed that. A reform drafted largely by the linguist Manuk Abeghyan and adopted in 1922, then partly revised in 1940, respelled the language closer to how Eastern Armenian was spoken. It was part of the literacy campaign: fewer silent letters, fewer spellings to memorise. It applied wherever Soviet power did, and independent Armenia kept it after 1991.

Most of the Armenian world was outside Soviet power, and most of it never adopted the reform. The Western Armenian communities scattered by the genocide of 1915 rebuilt their schools and presses in Beirut, Aleppo, Paris and the Americas, in classical spelling. Iran’s Armenians speak Eastern Armenian, and they kept the classical spelling too. So did the Eastern Armenian politicians and writers of the short-lived First Republic who went into exile after 1920. For many of these communities, keeping the old spelling was also a way of not accepting a Soviet decision about their language. The newer diaspora, people who left Soviet and post-Soviet Armenia, mostly brought the reformed spelling with them.

The result is that dialect and spelling don’t line up. Eastern Armenian is printed in both orthographies, depending on where and when. Western Armenian is almost always classical, except where Soviet publishers reset Western authors in reformed spelling. A collection like the Armenian Institute’s, gathered from across the diaspora over a century, holds all of these on the same shelves.

A goal of the next version of our Armenian model is to read the books the current one has never seen: everything printed in classical spelling, before the reform and outside Soviet Armenia since, in either dialect. To train and test on that material we first had to find it, among several hundred Armenian works on Wikisource. The obvious approach was to read the index titles. The titles turned out to be no help.

They mislead in both directions. Kajaznuni’s Ազգ և հայրենիք is in classical spelling throughout, but its title carries the reformed ligature և, because whoever typed it in typed it that way. The Soviet Yerevan editions of Western Armenian authors (Zohrab, Varoujan, Terian, Metsarents) were reset in reformed spelling, whatever the authors wrote. A title is a piece of metadata somebody typed. The page is what the printer set. So we stopped reading titles and read the pages instead, looking for a test that anyone could run on a page of text.

That test turned out to be one letter.

What the reform changed, and what it did not

The reform added no letters and removed none. Classical and reformed Armenian share an alphabet and, in Unicode, a code block. What changed is where the letters go:

A classifier therefore can’t look for a letter that only one orthography has. It has to look at distribution, and the question is which distribution separates the two cleanly enough to trust.

Four candidates, one winner

We measured four markers on pages we already knew the answer for. That meant every reformed harvest we had, plus a handful of sampled pages from works we knew were classical. Each marker was counted per thousand Armenian letters, so that long and short pages compare.

The two-letter եւ against the ligature և. This was the marker we expected to work, and the one that failed outright. It is a habit of the press, not of the orthography. Հայկական տպագրութիւն, a history of Armenian printing set in impeccable classical spelling, uses the ligature almost everywhere. Across the classical works the census eventually found, some presses prefer the two letters and some the ligature, and the split is close to even. We still record it, because it matters for scoring (we now treat the two forms as one when measuring accuracy), but it can’t decide anything.

է inside a word, where reformed spelling writes ե. A silent final յ after a vowel. Both move in the right direction, and both overlap at the edges. A word-final յ is also perfectly reformed in հայ and թեյ. Once the fourth marker has been counted, neither adds anything.

ւ outside the ու digraph. This one separates the two with room to spare, and the reason is structural rather than statistical. In classical spelling ւ (U+0582, yiwn) does several jobs: half of the digraph ու for u, the v in աւ and եւ, and the glide in իւ. The reform gave each of the other jobs to another letter, and left ւ exactly one place to stand: the second half of ու. The ligature և is a separate code point (U+0587) and doesn’t contain a U+0582 at all. So in properly reformed text, a ւ that doesn’t follow an ո isn’t rare. It shouldn’t exist.

A rule rather than a tendency is what you want to build a classifier on. On the reformed harvests the free ւ barely registers, and the little there is comes from quotation and apparatus. On classical pages it is everywhere, because the commonest abstract-noun suffix, -ութիւն, carries one, and so does “and” wherever the press writes it as two letters. Between the two groups there was a wide empty band with no work in it.

The classifier

The test fits in a regular expression with a lookbehind:

import re

ARMENIAN_LETTER = re.compile(r"[Ա-Ֆա-և]")
FREE_YIWN = re.compile(r"(?<![ոՈ])ւ")   # ւ that is not the second half of ու

def free_yiwn_per_thousand(text: str) -> float:
    letters = len(ARMENIAN_LETTER.findall(text))
    return 1000 * len(FREE_YIWN.findall(text)) / letters if letters else 0.0

The trainer’s orthography module adds three verdicts and a floor. At 8 or more per thousand, the text is classical. At 2 or fewer, it’s reformed. Anything between is mixed. Below 500 Armenian letters it declines to answer, because a short sample says nothing about a rate. The thresholds sit well inside the empty band, so a page has to be genuinely unusual to land near one.

The census that uses it samples four body pages from every work with enough proofread pages. It spreads them through the book and skips the first tenth, where title pages and prefaces live. It cleans each page with the same code the harvester uses, so the classifier sees exactly the text the trainer would. It caches the verdicts, so that adding a work doesn’t mean re-reading the rest.

What the census found

It found far more classical material than the titles had suggested, including two full-resolution sources we hadn’t known were classical. Before it could, it exposed a bug: every poetry index on Wikisource had been harvesting as empty. hy.wikisource wraps a verse page in a template, our cleaner strips templates whole, and the poem went with the wrapper. The census couldn’t classify a page with no letters on it. That made the gap visible for the first time, and the fix brought several thousand proofread pages of verse back into every harvest.

The verdicts also explained themselves in the places they were awkward.

The mixed band is reformed print with something old inside it. Soviet editions that quote Grabar at length. Collected works whose apparatus is in classical spelling. A few pages of facsimile. Nothing in the mixed band turned out to be a classical source, which is the outcome you want from a band labelled “don’t use this for either”.

Grabar trips the marker, correctly. The seventeenth- and eighteenth-century song collections in modern critical editions came back classical, and they are classical orthography. They just aren’t the modern Armenian in classical spelling that we’re trying to teach the model, so they’re left out of the harvest. The classifier answers the question it was asked, about spelling, not the one we sometimes meant, about the language.

The Western authors’ Soviet editions are reformed, as feared. Classical spelling on Wikisource means diaspora and émigré printing, in both dialects. On the Western side there are Toranian’s books from Aleppo, Yesayan, Shushanian, and Baronian’s Ազգային Ջոջեր in its original spelling. On the Eastern side there are Khatisian, Kajaznuni and Aghbalyan, First Republic figures who went on writing in classical spelling in exile.

A handful of works sit just over the classical line. Each is probably a reformed edition quoting classical text at length, or a collection containing both. We treat them as mixed: text supply for the language model, not image crops for the recogniser, until a page-level pass says otherwise.

The same question about a single name

A rate needs a page. A library catalogue often has only a title or a name, and the question still matters there. A reader searching for Boghosian should find both Պօղոսեան and Պողոսյան, and knowing which orthography a record is in tells a search system how to fold it. So the same idea went into tetrak-translit, our transliteration library, rebuilt for strings too short to have a rate.

Its detector works on presence rather than rate, and adds tells that only make sense for a short string:

• Classical if there is any free ւ, any եա or եօ digraph (which reformed spelling writes յա, յո), or any է inside a word. On a page, inner է adds nothing once ւ has been counted. On a single name like Գէորգ it may be the only tell there is.
• Reformed if there is none of those, but there is a յ between a consonant and a vowel (the reformed -յան of Պողոսյան) or the ligature և.
• Indeterminate if there is nothing at all. Արմեն is spelled the same way under both, and the honest answer is that the string doesn’t say.

from tetrak_translit import detect

detect("Պօղոսեան").orthography   # 'classical'
detect("Պողոսյան").orthography   # 'reformed'
detect("Գէորգ").orthography      # 'classical'
detect("Արմեն").orthography      # 'indeterminate'

Three answers rather than two is the important part. A classifier that must always pick a side will pick one for Արմեն, and whatever relies on it downstream will treat a guess as a fact.

Where it goes wrong

The test reads how a text was spelled. That isn’t always how it was printed, and three cases fall in the gap.

Reformed text typed with two letters. The reformed “and” is the ligature և. A lot of reformed Armenian typed on keyboards that make the ligature awkward to enter is written եւ instead. Every one of those contains a free ւ. On a page, the trainer’s rate absorbs a few of them, but text that uses եւ every time can drift into the mixed band. On a single string, tetrak-translit calls it classical. Wikisource’s proofreaders set what the page prints, so the census never met the problem. A catalogue typed by hand will. For now we treat that as a known gap rather than guess.

Quotation. A reformed book that quotes Grabar contains classical spelling, and the classifier reports it. That’s correct about the text and unhelpful about the book. The mixed verdict is how the page-level version copes. The string-level version has no rate to fall back on, and a quoted classical title in a reformed record is called classical.

The ligature in a classical title. The case we started with: a title-only string like Ազգ և հայրենիք has no classical tell and one reformed one, so the detector calls it reformed. The book is classical. A detector working on a title can only be as right as the title, which is where this began.

If you are doing the same for another script

Armenian isn’t the only script printed in two spellings. Russian before and after 1918 sets the same problem with different details, and so does any language whose spelling was reformed while its old books stayed on the shelves. The lessons carry over:

• Read the text, not the metadata. Titles and catalogue records are typed by people at a later date. The page is the evidence.
• Look for a rule, not a frequency. The marker worth having is a character the reform made impossible in one context, not one it made less common. Impossible gives you an empty band to put thresholds in.
• Have a third answer. Mixed on a page and indeterminate on a string. They aren’t failures of the classifier. They’re the cases where the honest answer is that the text doesn’t say.
• Measure the markers you expected to work, and write down why they didn’t. The եւ/և habit looked like the obvious test. Both of our modules keep a note in their docstrings saying why it isn’t, so nobody spends an afternoon rediscovering that.

• armenian
• orthography
• unicode
• training
• project-notes