OCR for Pre-Revolutionary Russian: Reading the 1917 Press with AI

Take a page of a Petrograd daily from 1917 and feed it to a modern Russian OCR engine. The scan may be decent, the type may be legible to any reader of Russian — and the output will still be quietly wrong, line after line. The reason is not image quality. It is that the engine has never really seen the language on the page.

Russian spelling was reformed in 1918. Everything printed before that — the entire imperial period, and the whole revolutionary year of 1917 — uses an orthography that modern models were barely trained on. For historians of the revolution, for genealogists, for local and family researchers, this one detail decides whether a digitization project produces a searchable corpus or an expensive pile of image files.

Why 1917 breaks modern models

Four letters no longer exist. Pre-reform Russian uses ѣ (yat), і (decimal i), ѳ (fita) and ѵ (izhitsa). A model trained on modern text has three bad options when it meets them: replace them with the modern look-alike (silently modernizing the text), confuse them with a visually similar letter (ѣ often becomes ъ or б), or fail the whole word. All three happen — and the worst part is that the output still looks like plausible Russian, so the errors hide.

Every word ends differently. Words ending in a consonant carried a final hard sign: миръ, объ, впередъ. That single character inflates the character set's frequency profile and trips language models that "correct" toward modern spelling.

The grammar differs in small, systematic ways. Adjective endings (-аго, -яго, -ыя), pronoun forms (ея, онѣ) — a modern language prior treats them as typos and "fixes" them. An engine with a strong built-in language model can be worse here than a dumber one: it confidently rewrites what it does not recognize.

And all of it is printed on 1917 paper. Wartime and revolutionary newsprint was cheap and acidic; type was worn; columns were narrow and dense; many surviving copies exist only as microfilm made decades ago. Each of these alone degrades OCR. Together they define the task.

What actually works

The honest tool landscape for pre-reform Russian print in 2026:

Tesseract ships a modern-Russian model. It has no official pre-reform model; community-trained data for old orthography exists and is worth testing on clean book print, but on newsprint it was not competitive in our experience.

Transkribus and eScriptorium/Kraken — the trainable platforms — are the serious route. Public Cyrillic models are oriented mostly toward handwriting; for pre-reform print you should expect to train or fine-tune on your own material. Several dozen corrected pages of ground truth is a realistic starting point, and the handwriting-recognition lessons about ground-truth discipline apply unchanged.

Commercial cloud OCR (Google, Microsoft) reads the shapes of pre-reform text surprisingly well on good scans, but normalizes and stumbles unpredictably on the abolished letters — and you cannot fine-tune it. It can, however, serve a smarter role: as a teacher.

Fine-tuned transformer OCR is where we landed. For our own corpus of a 1917 socialist daily, we distilled a commercial engine's output on high-quality pages into training data, corrected it, and fine-tuned a TrOCR-family model for pre-reform newsprint. The result reads at roughly 90% character accuracy on held-out pages — not publication-grade transcription, but reliably good enough for full-text search, frequency analysis, and locating any quote for human verification. We described the verification side of such pipelines in the confidence-calibration case study: the model must tell you where it is unsure, so a human looks exactly there.

The orthography decision every project must make

Digitizing pre-reform text forces a choice most projects discover too late: what spelling does your searchable text use?

Keep two layers. The diplomatic layer preserves the page exactly — ѣ, і, ѳ, final ъ — because that is the historical source, and scholars quote sources, not modernizations. The normalized layer converts to modern spelling by rule (the reform was mercifully systematic), because every real user — a historian, a genealogist searching a family name, a student — types modern Russian into the search box. The conversion is one-directional and cheap; storing both costs almost nothing. Projects that saved only the modernized text have destroyed evidence; projects that saved only the diplomatic text have built archives nobody can search.

This dual-layer corpus is exactly what enabled our historical newspapers research project and the database built around a single transformative year — and once the text layer exists, AI research agents can work across it at scale, reading every issue so the researcher reads only what matters.

A practical recipe

1. Sample first. OCR ten representative pages with whatever engine is at hand. Count real character errors on 500 characters. Below ~90%, plan for fine-tuning.
2. Fix the scans before the model. 300-400 DPI from originals where possible; deskew; separate columns explicitly — column bleed is a top error source in dense newsprint.
3. Build ground truth in the old orthography. Never let annotators modernize. 30-60 corrected pages is a workable first training set for print.
4. Fine-tune, measure on held-out pages, iterate. Character error rate on your material is the only number that matters — engine leaderboards do not transfer to 1917 newsprint.
5. Generate the normalized layer by rule, and keep both.
6. Route confidence to humans. Low-confidence lines get eyes; everything cited gets eyes. The model accelerates reading; it does not replace the reader.

Who needs this

Historians of the revolution working through the 1917 press. Genealogists — metrical books, civil records and obituaries are all pre-reform. Local historians with provincial gazettes. Families holding letters and postcards in the old spelling. Publishers preparing memoirs written before the reform.

Have pre-1918 Russian material — newspapers, records, letters — that resists every tool you have tried? This is precisely the kind of corpus we build pipelines for. Get in touch and we will run a feasibility sample on your own pages.