YauhenBichel/belarusian-verse

Belarusian stress, rhyme, frequency and agreement tables

Five lookup tables for working with Belarusian words as they sound, as they combine, and as people actually use them:

1. stress — 1,997,602 word forms with the position of the stressed vowel. Belarusian does not write stress, but it decides pronunciation (по́бач, not паба́ч), rhyme and how a line fits a melody.
2. forms — 2,006,179 word forms with part of speech, gender, case and number, which is enough to check whether an adjective agrees with its noun, or a subject with its verb.
3. rhymes — 74,578 rhyme endings grouping ~2 million forms by how they sound from the stressed vowel to the end. The first Belarusian rhyme dictionary in software.
4. frequency — 538,449 words with how often they appear in Belarusian Wikipedia (59.7M tokens). Frequency is what separates a word people sing from one only a dictionary knows.
5. data/hunspell/ — the official-orthography Belarusian spelling dictionary (Hunspell .dic/.aff), included so a checker can be built from this one download.

Each table is published twice: as Parquet with named, typed columns (data/<table>/, what the dataset viewer and load_dataset use) and as the original tab-separated .tsv.gz (data/be-*.tsv.gz, what the belarusian-verse library downloads). The two hold the same rows.

The tables are derived from the Belarusian Grammar Database (RELEASE-202601), which publishes stress as + after the stressed vowel and a three-part morphological tag per form. This dataset flattens that into files you can load with one line of Python, with no XML parsing and no Java.

Load

from datasets import load_dataset

stress = load_dataset("YauhenBichel/belarusian-verse", "stress", split="train")
stress[0]          # {'word': 'а', 'stress': [1]}

forms = load_dataset("YauhenBichel/belarusian-verse", "forms", split="train")
rhymes = load_dataset("YauhenBichel/belarusian-verse", "rhymes", split="train")
frequency = load_dataset("YauhenBichel/belarusian-verse", "frequency", split="train")

Or, without datasets:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download("YauhenBichel/belarusian-verse", "data/frequency/train-00000-of-00001.parquet",
                       repo_type="dataset")
freq = pd.read_parquet(path)       # columns: word, count, rank

Table	Columns
stress	word string · stress list of int8 — 1-based index of the stressed vowel; several values mean a homograph
forms	word string · readings list of string — POS:GENDER:CASE:NUMBER (see Format)
rhymes	ending string · words list of string · count int32
frequency	word string · count int64 · rank int32 (1 = most frequent)

The .tsv.gz files have no header row. A reader that guesses column names from the first line takes the first word for a header and drops that row — until September 2026 the Hub's own automatic conversion did exactly that (the frequency table lost «і», its most frequent word, and the rhyme table could not be converted). Use the Parquet files, or read the TSV with explicit column names.

Tools

If you want the checks rather than the raw tables, the companion library does the lookups, the stress rules and the rhyme grading for you:

github.com/YauhenBichel/belarusian-verse (Apache-2.0)

pip install belarusian-verse

from belarusian_verse import mark_stress, rhyme, check_agreement

mark_stress("Побач ты, і добра мне")      # 'По́бач ты, і до́бра мне'
rhyme("вадзе", "ідзе")                    # ('rich', 1.0)
check_agreement("тваіх вачам")["errors"]  # 1

It downloads these tables on first use and caches them.

Why this exists

While building a Belarusian song generator we found that, for Belarusian:

• there is no rhyme dictionary in software (only Minkin's printed one);
• LanguageTool's Belarusian module has ~66 rules and almost no morphology, so it cannot see that тваіх вачам is ungrammatical;
• spell checkers accept any real word, so they pass тваіх вачам and ідзе́ / зна́йдзе as a rhyme.

Everything needed to fix that was already inside GrammarDB. These tables are that content, in a form you can use directly.

What the tables can and cannot do — measured

Agreement. On the 2,570 Belarusian subject-verb agreement minimal pairs of MultiBLiMP 1.0 (CC-BY-4.0):

Check	Wrong sentence caught	Correct sentence flagged
neighbouring words only, using forms	34.0 %	19.8 %
a dependency parser (Stanza, UD Belarusian-HSE) alone	60.3 %	7.0 %
the parser finds subject and verb, forms decides whether they agree	57.7 %	4.3 %

The parser figures are optimistic: MultiBLiMP's sentences come from the treebank the parser was trained on. The lesson is that forms is a reliable judge of whether two words can agree, and a poor finder of which words belong together.

Spelling. On a uniform random sample of 4,000 GrammarDB forms the Hunspell dictionary in data/hunspell/ rejects 34 %, but nearly all of those are rare derivations nobody writes (дзяньдзівірыш, міжзярнятная). Weighted by use the gap is small:

Forms seen in Belarusian Wikipedia	Rejected by Hunspell
any	33.7 %
at least 10 times	8.8 %
at least 100 times	3.5 %
at least 1,000 times	0.9 % (verbs 0 %)

The reason is by design: the dictionary is GrammarDB's own 2008-spelling export, which by its documented rules leaves out forms whose only source is Піскуноў's 2012 dictionary («слухаўка», «смаленскі», «кампус» — checked in the release XML). What is left is mostly adjectives from place names (жытомірскай) and newer words, plus a few ordinary ones (слухаўка, гучыш, гадовы) — so consult forms before calling a word misspelled. Names are capitalised in Hunspell and lowercase in these tables: look up both. Hunspell remains the stricter judge of the 2008 standard, since GrammarDB also lists some variant spellings.

Stress of homographs. 20,754 forms in stress have more than one stressed vowel, and the list order carries no information about which is commoner — taking the first is a coin toss. Scored on the homographs of the Belarusian Homographs Stress Benchmark (CC BY-SA 4.0, created for BelVoice):

Choice	Common Voice sentences (2,960)	«Засценак Малінаўка», literary (2,311)	10×10, every stress equally often (1,480)
first value in stress	48.9 %	24.9 %	50.0 %
BelVoice's most frequent stress for 679 common homographs (LGPL-3.0), else the first value	82.9 %	68.8 %	50.0 %

The 10×10 set is built so that no frequency list can win: those words need the sentence. BelVoice reports a Gemma 4-31B prompt method at 99.7 % on the homographs of an ordinary text.

Format

stress — word, tab, the 1-based index of the stressed vowel (not character). Several values, comma separated, mean the form is written the same with different stress (20,754 rows), in no particular order.

побач	1
удваіх	3
вадзе	1,2

forms — word, tab, comma-separated readings POS:GENDER:CASE:NUMBER. 0 means the form does not distinguish that category. X:0:0:0 marks a form that also reads as something that does not decline (adverb, verb, particle) — useful to avoid false alarms, e.g. тут is usually the adverb "here" even though a rare noun is spelled the same.

вачам	N:N:D:P
тваіх	S:0:A:P,S:0:G:P,S:0:L:P
тут	    N:F:G:P,N:M:A:S,N:M:N:S,X:0:0:0
іду	    V:0:1:S
спявала	V:F:P:S,V:N:P:S

For a verb the third field is the person (1/2/3) for the present, future and imperative, 0 for the infinitive, or P for the past tense — where the second field holds the gender instead, because Belarusian past tense agrees in gender rather than person: «яна спявала», «ён спяваў», «яны спявалі». Many verbs share a form between the infinitive and the third person (сніцца), so a checker should know from context which one it is looking at.

POS letters follow GrammarDB: N noun, A adjective, S adjective-like pronoun, V verb … Cases: N G D A I L V. Numbers: S P. Genders: M F N.

Use

import gzip

stress = {}
with gzip.open("data/be-stress.tsv.gz", "rt", encoding="utf-8") as fh:
    for line in fh:
        word, _, marks = line.rstrip("\n").partition("\t")
        stress[word] = tuple(int(m) for m in marks.split(","))

VOWELS = "аеёіоуыэюя"

def mark(word):
    """по́бач — put a combining acute after the stressed vowel."""
    positions = [i for i, c in enumerate(word) if c in VOWELS]
    options = stress.get(word.lower())
    if not options or len(positions) < 2:
        return word
    at = positions[options[0] - 1]
    return word[: at + 1] + "́" + word[at + 1 :]

print(mark("побач"))     # по́бач

Agreement in three lines: read forms, take the readings of the two words, and keep the pair only if some reading shares gender, case and number.

def agrees(a, b):
    return any(c1 == c2 and n1 == n2 and (g1 == "0" or g2 == "0" or g1 == g2)
               for _, g1, c1, n1 in a for _, g2, c2, n2 in b)

agrees(forms["тваіх"], forms["вачам"])   # False
agrees(forms["тваім"], forms["вачам"])   # True

тваіх (accusative/genitive/locative plural) against вачам (dative plural) → no shared reading → the phrase is wrong. тваім вачам → dative plural on both → correct.

rhymes — a rhyme ending, tab, the words that share it, space separated. A Belarusian rhyme matches from the last stressed vowel onward, with iotated vowels folded (ёй = ой) and the final consonant devoiced, which is why the stress table has to come first.

э	вадзе дзе ідзе мне не яе ...
ой	спакой маёй настрой ...

frequency — word, tab, how many times it appears in Belarusian Wikipedia.

не	187983
ідзе	3350
вадзе	1965

Rare derivations (кувадзе, зеляноч) simply do not appear, which is the point: without frequency a rhyme list is full of words nobody has ever sung.

data/stress-overrides.tsv — hand corrections that win over the lexicon: word, tab, stressed vowel index. It includes folk and dialect forms the modern dictionary does not carry (поліць, коліць, дочка, садочку, рве, from «Купалінка»).

Two Belarusian rules worth knowing

• ё is always stressed. If a word contains ё, that is the stress, and no lookup is needed.
• A word starting with ў is the same word as the one starting with у (ўдваіх = удваіх), but ў is not a vowel, so every stress index shifts down by one.

Rebuilding

build_stress_lexicon.py and build_grammar_index.py read the GrammarDB release zip directly, so a new release becomes new tables in about a minute. build_rhyme_index.py derives the rhyme groups from the stress table, and build_frequency_list.py counts words in a Wikipedia dump (about 15 minutes for Belarusian). See BUILD.md. The Parquet files are the same rows with named, typed columns.

Licence and attribution

CC BY-SA 4.0, inherited from the sources. The stress and morphology tables are derived from the Belarusian Grammar Database by Aleś Bułojčyk and Uładzimir Koščanka (https://github.com/Belarus/GrammarDB, https://bnkorpus.info/grammar.html), used under CC BY-SA 4.0. data/hunspell/ is the Belarusian spelling dictionary for the 2008 orthography by the same authors, redistributed from the LibreOffice dictionaries under CC BY-SA 4.0 or LGPLv3 (see data/hunspell/README_be_BY.txt). data/be-freq.tsv.gz is counted from the Belarusian Wikipedia dump of 1 September 2026, whose text is CC BY-SA 4.0 by its contributors. Anything built on these tables must keep the same licence and the attribution.

What is not here

Stress for homographs is not disambiguated by context: 20,754 forms carry more than one stress, in no particular order, and you must choose by meaning (see the measured table above for a frequency list that gets most of them right). There is no lemmatiser, no syntax, and no phonetic transcription (espeak-ng has a Belarusian voice if you need IPA).

The frequency list is Wikipedia, so it leans encyclopaedic: place names and dates are commoner here than in speech or song, and spoken words are under-counted. It is good enough to tell a real word from a dictionary curiosity, which is what it is used for. No Belarusian frequency list from speech or subtitles exists — the usual multilingual source covers 60+ languages and not this one.