The words that pay off first

This is the headline of thread two: the one corpus-based, expert-validated CEFR profile for Ukrainian (Synchak et al. 2025, the PULS platform), joined onto this frequency list per lemma. A co-author granted permission to redistribute the per-lemma A1 to C1 levels; the join is POS-aware, so a homograph is dropped rather than mislabeled. Each lemma's level is its onset level, the earliest at which the profile introduces it.

It is a dated snapshot, not the final map. The profile is still being extended (the live list already carries more B2 and C1 lemmas than the published paper), so treat these levels as validated but provisional.

6,403 of the top 12,000 lemmas now carry a validated CEFR level, A1 to C1

The validated map above needs the authors' permission to redistribute. This overlay is the fully open alternative: it labels our lemmas from the learn-ukrainian open course (Krisztian Koos, CC-BY-SA), an expert selection based on Ukraine's 2024 State Standard, no permission needed. A lemma's level here is the first course lesson that introduces it, so treat it as a rough teaching-order signal, not a proficiency label.

Its role now is a cross-check, not a fallback. The validated overlay above measures the two against each other on the full overlap; this card keeps the open overlay's own calibration against the published gold points. Neither the open map alone nor the paper's frozen tables is the finished thing a learner deserves, and closing that gap is why this series exists.

2,325 of the top 12,000 lemmas carry an open CEFR label, A1 to B2

The field census below used to list the "1000 i 1 slovo" A1 lexical minimum (Kseniia Borodin and Oksana Turkevych, School of Ukrainian Language and Culture, Ukrainian Catholic University) as walled: licensed CC-BY-4.0 but published only as a prose PDF an extractor could not reach, so the one independent, empirically grounded A1 guideline for Ukrainian could not be joined and the overlay's rigor model called that missing consensus a seam. This page asked the authors for a machine-readable table, and co-author Kseniia Borodin sent the word list herself. This card is that list, joined onto the frequency list and republished as attributed open data under the guide's own license.

The join is honest about its grain. The authors' unit is the published entry (a single word, an aspect pair, a masculine and feminine pair, sometimes a whole phrase), the list carries no part of speech, and the join is by lemma string at the best rank, coarser than the POS-aware overlays above. Phrase entries are counted, never force-joined, and a beginner word the corpus does not rank stays honestly off-list.

838 of the 929 unique A1 entries join the top 12,000, reaching 833 distinct lemmas

Of the 806 joined lemmas the validated PULS map also covers, 75.7 percent are validated A1 and 96.8 percent sit at A1 or A2. Two independent efforts, a corpus-validated profile and an expert lexical minimum, converge on what beginner Ukrainian is. The handful above A2 is mostly the price of a coarser join: автомат the A1 ticket machine shares a spelling with the C1 rifle, and this list cannot tell them apart without a part of speech.

The A1 card above closed a seam. Its provenance note also recorded an open one, in plain words: the A2 volume of the same series existed, but no machine-readable A2 list had been shared. That volume has since been published open access under CC-BY-4.0 by Kseniia Borodin and Olesia Lazarenko at the Europa-Universitat Viadrina, so the seam closes by publication rather than by favor. This card is its 1,000-entry register, extracted from the publication's own PDF, joined onto the frequency list the same coarse way, and republished as attributed open data.

Two things become answerable that were not before. Because this is the same authors' next volume, the two lists together measure what one CEFR step actually costs, in frequency ranks rather than in word counts. And because the A2 book also prints its whole vocabulary regrouped under its own 51 topic headings, this is the first source on this page that can say WHICH semantic fields a news-heavy frequency list misses, not merely how many words it misses.

710 of the 1,000 published A2 entries join the top 12,000, reaching 709 distinct lemmas

Of the 673 joined lemmas the validated PULS map also covers, 70.1 percent sit at A1 or A2 and 29.9 percent sit above A2, nearly all at B1. That is a looser agreement than the A1 card found, and it should be: A2 is where two expert efforts start to disagree about the boundary, and the coarse lemma-string join adds noise of its own. The page reports the disagreement rather than tuning it away.

The three cards above each read one source against the frequency list. This one reads two sources against each other. Ukrainian now has both a corpus-informed teaching syllabus, the A1 and A2 volumes of "1000 i 1 slovo", and a corpus-validated descriptive profile, the Ukrainian Vocabulary Profile on PULS. They were built by different methods to answer different questions, and until now nobody had joined them lemma by lemma. This card does, on the frequency spine, and classifies every cell.

The point is not to score either one. A syllabus decides what to teach on Tuesday and a profile records where a word first turns up in graded use; they are allowed to differ, and the interesting reading is where and how. The grid below is the whole join. Choose a cell and the words in it are listed underneath, most frequent first.

10,366 lemmas in the union of the two sources, 1,200 of them placed at the same level by both

The card above says 229 word entries of the A2 minimum do not rank in this list's top 12,000, and reads that as the ceiling on learning by frequency. That reading rests on something untested. This list is built on UberText 2.0's COMBINED corpus, and the largest thing in that corpus by a wide margin is news. So an off-list word might be rare in Ukrainian, or it might be ordinary and simply absent from journalism, which does not often discuss forks, pumpkins or neckties. Those two readings give opposite advice to anyone building a word list, and nothing above can tell them apart.

So we built two frequency lists instead of one, from UberText's own fiction and news subcorpora: same collection, same cleaning, same lemmatizer, same sampling shape, both capped to the same token count. Genre is the only thing that differs. The ranks below are internally comparable and deliberately NOT comparable to the main list on this page, which a different lemmatizer produced.

News reaches 658 of the entries (70.3 percent). Fiction reaches 648 (69.2 percent). Almost the same amount, and not the same words: together they reach 744 (79.5 percent). A second genre of the same size buys 9.2 points that more of the first genre would not have bought. For a pedagogical target, corpus diversity is worth more than corpus size.

The exchange is close to symmetric, which is what makes it a real result rather than a bigger hammer: fiction reaches 30.9 percent of what news misses, and news reaches 33.3 percent of what fiction misses. Neither genre is better. They are looking at different parts of the language.

Read the two rows together and the mechanism is plain. Journalism supplies the institutional and the geographic; fiction supplies the domestic, the bodily and the sensory. Neither is a description of the language on its own.

Aspect as a relation, not a column

Ukrainian aspect is lexical: a perfective and its imperfective are two lemmas, so the table relates them instead of tagging one. For the 443 B1 verbs, Wiktionary lists a partner for 378, a hand-curated table of suppletive pairs (брати and взяти among them) is asserted on top, 0 of its pairs touching a B1 verb, and the corpus supplies every attested prefixed derivative. Each related verb keeps its own validated level, which makes 102 split-level pairs visible: a B1 verb whose partner the profile places elsewhere.

Government, collocations, and gated examples

Aggregated over 2,880,835 parsed tokens of UberText's Wikipedia subcorpus, the one whose text is license-compatible with this release. Government is verb + preposition + case with the share of the verb's nominal dependents; a pattern is shown at ten percent or more and three observations, and the raw counts stay in the download. 365 of the 443 B1 verbs have five or more observations, 285 have twenty or more. Wikipedia has a register: дякувати does not occur.

The join keys are the PULS headwords, in VESUM's conventions. The parser's lemmas are pymorphy3's, whose Ukrainian dictionary is VESUM-derived, and the convention check holds: 99.9 percent of the B1 headwords normalize to themselves (2 do not, кошти among them), under the one-percent benchmark the plan sets. 73 percent of the parsed content tokens land on a lemma of this list; the rest is reported, never merged into the ranks.

The connectives a lemma list cannot hold

A frequency list ranks lemmas, so a phrase is invisible to it. "тому що" is counted as one very common conjunction plus one very common pronoun, and nowhere as the connective it is. This table gives 142 expressions their own IDs: 63 discourse connectives from UK-DiMLex and 79 multiword entries of the two published lexical minimums, each counted over the same 4,694,436 word tokens of Wikipedia the usage tables were cut from.

The parts carry the validated levels, the unit does not: the profile levels words, not expressions. Reading the parts gives a floor, and for 87 of the 98 expressions whose every part is validated (88.8 percent) that floor is B1 or easier. A coverage figure computed word by word therefore counts them as already known, which is the gap this table exists to measure.

A floor is not a level. Таким чином is a frozen instrumental of a C1 noun and reads C1 part by part, which is as wrong for the unit as A1 would be. The column says what the parts are worth, and nothing more.

Two readings of one expression: the levels its parts carry, and the expression itself.

Trained on the 6,403 validated labels and asked about the 5,597 lemmas that have none. Proper nouns, function words, and lemmas the lexicon does not know are ruled out of scope first, so the model predicts for 4,218 content words. Held-out folds are grouped: lemmas that share a stem, an aspect link, or a spelling (3,394 groups) never straddle train and test, so a family cannot be memorized.

Under plain random folds the shipped model scores 41.5 percent exact against 40 under grouped folds; the grouped figure is the one reported everywhere on this page, and the gap is the size of the leakage a naive split would have hidden.

Three caveats the design owes. The single strongest feature in the French template, per-level frequencies from a graded learner corpus, does not exist for Ukrainian, so this model works from native frequency, morphology and a few open resources (Wiktionary senses and loanword flags, subtitle frequency, UniMorph families, the Wiktionary family graph) and its ceiling is structurally lower; the ablation ladder shows what each block buys. Membership in the open A1 and A2 lexical minima is computed and reported but deliberately not in the shipped model: it was the strongest weight, worth about nine points of exact match, and that lift was a graded word list read back as a feature, firing on most validated A1 labels and on almost none of the lemmas being predicted. Dispersion here is documents per occurrence; Gries DP across the five subcorpora needs per-subcorpus counts the shipped dictionary does not carry. And an LLM judgment is reserved as at most one measured feature and is not in this version.

These are candidates for the profile's authors, not levels. The validated overlay is untouched by any of them.