Half the words appear exactly once
I set out to build a reader for Russian literature and needed to answer a practical question first: which words should somebody learn, and in what order? The answer I found was not the one I expected, and it makes most vocabulary advice for literature quietly useless.
The measurement
Fifteen public-domain texts, lemmatised by hand: 68,172 running words in 5,075 sentences, containing 7,746 distinct dictionary words.
Of those 7,746 lemmas, 3,824 — 49.4% — occur exactly once in the entire corpus. Learn one of them and you have bought yourself a single sentence, once, in fifteen books.
This is not a quirk of my selection. It is Zipf's law doing what Zipf's law does, and linguists have a word for these words — hapax legomena. What surprised me was the size of the effect at the scale a learner actually cares about. Nearly half of everything you meet is a word you will never meet again.
Why frequency lists fail here
The standard advice is to learn the most frequent words first. That is sound, and it is also where the advice usually stops — which leaves out the thing that decides whether you can read a page.
Reading research puts comfortable unassisted reading at roughly 95–98% of running words known (Hu & Nation). Coverage is a property of the text, not of the sentence in front of you, and sentences are what you actually read. A sentence with one unknown word is a sentence you can read. A sentence with two is usually a sentence you cannot.
So I measured both. Here is what the ladder buys, at each depth:
| Words known | Coverage of running text | Sentences fully readable |
|---|---|---|
| 500 | 67.3% | 23.5% |
| 1000 | 77.1% | 35.5% |
| 2000 | 82.8% | 54.7% |
| 3000 | 89.8% | 67.0% |
| 5000 | 95.9% | 83.2% |
| 7000 | 98.9% | 95.6% |
Look at the gap. At 500 words you know two-thirds of the running text and can read under a quarter of the sentences. At 2,000 words you hold 82.8% of the words and 54.7% of the sentences. Coverage flatters; sentences tell the truth. The two columns only converge near the top, which is why intermediate learners feel stuck long after the frequency lists say they should be fine.
A different ordering
If sentences are the unit that matters, then rank words by how many whole sentences each one makes readable, given everything already known. That is a greedy set-cover problem, and it produces a noticeably different list from raw frequency: words that are common but always sit beside other unknown words score badly, and words of middling frequency that finish otherwise-complete sentences score well.
Ordered that way, 5000 words reach 95.9% coverage and make 83.2% of the sentences fully readable. The same ordering, applied per author, is also more interesting than a global list: weighting each word by the share of its corpus-wide uses that fall inside one writer puts нос, шинель, чиновник, коллежский, асессор at the top of Gogol's list — which is, precisely, the bureaucratic world «Шинель» is written in.
The practical conclusion
Two things follow, and they point in opposite directions from the usual advice.
Do not try to learn your way in. The hapax half cannot be pre-learned, by anyone, ever. Waiting until you "know enough words" to start reading literature means waiting for a threshold that recedes as you approach it.
Read with the words attached instead. If every word carries its meaning where it stands, the 49.4% you will never meet again stop being a barrier and become what they always were — scenery. What you then learn from reading is the other half, in the order the books themselves impose, which is the only order that was ever going to matter.
Method, and where the numbers come from
The corpus is fifteen public-domain works. Lemmatisation and glossing were done by hand and checked against dictionaries — 15,261 inflected forms, each with its dictionary form, meaning and stress mark, since Russian does not write stress and printed books do not carry it. 273 proper names are excluded from the vocabulary counts. Every figure on this page is computed from the shipped data at build time rather than typed in, so it cannot drift away from the corpus it describes.
The texts themselves, with all of this applied, are at languagethroughliterature.com — seven Chekhov stories are free with no account. The word lists are published in the ordering described above, and there are free Anki decks per author built the same way. The complete ledger this essay draws on — every book counted, with a downloadable file — stands at Russian classics, counted.