Aija Veldre Beldavs argued about the daina world the way a folklorist in 2001 could: she quotes a song, and reads outward from it. What was not available to her — and is now — is asking whether the song she quoted was representative.
The Institute of Mathematics and Computer Science at the University of Latvia now runs a searchable, lemmatised corpus of Latvju dainas at korpuss.lv: 46,587 song texts, 921,508 words, recorded between 1770 and 1903, each one still carrying Krišjānis Barons’ own thematic classification. This page takes her claims one at a time, restates each as something that can be counted, and reports what the count says.
Confirming, qualifying and occasionally correcting a scholar using evidence she did not have is how a field advances. It is not an audit and there is no scoreboard here. The most interesting results are the ones where a claim comes back in a different shape than she stated it, and the single most interesting result is a claim the corpus turns out to have no standing to judge at all.
What this corpus is, and is not. It is Volume 2 of Latvju dainas — one complete book of the six Barons published, dated 1903 — and not a sample drawn across the edition. Volume 2 is the youth songs: adornment, the wreath, the ring, learning to work, the dowry, reputation, girls’ and boys’ songs, night visits, old maids and bachelors. Every classified song in it carries a section code beginning 2.4, and exactly one strays into 2.5.
So childhood, the wedding itself, married life, death and the mythological songs are in the other five volumes and are simply not here. That matters for every number below, and it is the reason one of the ten tests returns out of reach rather than an answer.
Two further limits, both worth stating because they are invisible in the data. The printed volume holds 47,781 texts; this corpus has 46,576, because the Latgalian-language songs were excluded when it was built — so it is not quite complete even for its own volume, and it is missing a dialect region. And 46,587 songs out of Barons’ ~217,996 is roughly a fifth of the whole.
Large enough for distributional claims, and far too small to prove an absence. “Not attested here” never means “does not exist.”
The ten claims
The quatrain
ConfirmedThe basic text unit of the daina is a semantically self-sufficient quatrain, the predominant form in Baron’s classic publication. Chapter V, The Daina-Song
Measured two ways that share no method. Counting total syllables per song needs no sense of where the lines break; counting lines does. They agree.
| Form | By syllables | By line count |
|---|---|---|
| Quatrain | 89.2% | 87.5% |
| Six-line song | 4.8% | 5.5% |
| Free-standing distich | 0.1% | 0.2% |
Nearly nine songs in ten. The distich, which she treats as the quotable sub-unit and which is often quoted as a proverb, almost never stands alone as a whole song.
One detail the count makes visible that she does not mention: the mode is 30–31 syllables, not the 32 a quatrain of eight-syllable lines would give. That discrepancy is the subject of the next claim.
The eight-syllable line
RefinedStructurally each eight‑syllable daina is most often organized into two parallel equivalences of nature and culture, of the universal and the particular, of an address and a response. Chapter V, The Daina-Song
Trochaic tetrameter — four beats, eight syllables — is assumed throughout the chapter. It holds, but not alone.
| Syllables in a line | Share | |
|---|---|---|
| 8 | 56.2% | |
| 7 | 33.3% | |
| 6 | 8.6% | |
| everything else | 1.9% |
The seven-syllable line is not noise. It is three times every other non-standard length put together, and eight and seven between them are 89.5% of all lines. The daina line is a four-beat trochaic line whose last beat is very often a single syllable — and that variant is exactly what produces the 30–31 syllable mode in the previous claim. Her assumption holds; the picture is one degree less regular than “eight-syllable daina” implies.
Where the formulas live
RefinedThe first half of a daina distich is more likely to be the formulaic source, while the second is the improvisatory lyric target. Chapter VI, Singing, Buzzing, and Weaving
“Formulaic” here means shared with genuinely different songs: a run of four words counts as a formula only if it turns up under two different Barons themes. That control matters enormously — see the method note at the end.
As she states it, it is not supported. At the level of the verse line, distich completers are slightly more shared than openers — 68.7% against 65.8%, on 76,504 and 91,428 lines. The clean opener-formulaic / completer-improvised split is not there.
As a front-loading effect, it is strongly supported. The first tenth of a song is 2.3× more formulaic than everything after it: 14.2% of four-word windows against 6.2%.
So she was right about where formula lives and wrong about the unit. It is the song’s opening, not the distich’s first line.
And that result independently confirms a second claim of hers, made from a completely different kind of evidence. She quotes the singer Ansis Bērziņš: that singers used known verses until they warmed up, and improvised after. The corpus puts the warm-up in the opening lines. One measurement, two claims — one refined, one confirmed.
Oak and linden
ConfirmedAmongst other trees, these folk songs most often mention the oak and linden tree. Chapter IV, Myth, Belief Systems, and Cosmology
| Tree | Songs | |
|---|---|---|
| ozols oak | 1,795 | |
| liepa linden | 1,036 | |
| bērzs birch | 989 | |
| ābele apple tree | 768 | |
| priede pine | 457 | |
| ieva bird cherry | 436 | |
| egle spruce | 352 | |
| then kārkls 269, vītols 244, apse 209, kļava 153, lazda 144, osis 121, alksnis 15 | ||
Oak and linden lead every other species, exactly as she says. Worth being precise about what this does not show: she reads the pair as the gendered landscape of the songs, oak masculine and linden feminine, and frequency cannot confirm a reading. It can only confirm that the pair she built the reading on is the pair the songs actually reach for most.
Gossip as social control
ConfirmedBehavior is usually controlled by the nonresponsive, limited aggression of gossip (aprunāšana) or at the extreme by all-out aggression of curse, witchcraft, and ostracism. Chapters III and VIII
This is the one claim where her citations and the corpus provably join. Her own Latvian gloss of Barons LD 8282–9080 — “ļaužu valodām, slavu un neslavu, paļām, nievāšanu, mēlnesību” — is corpus section 2.4.6.1 word for word.
That section holds 5,011 songs: one in nine of the whole corpus, and the third-largest section in Barons’ entire youth branch. There are more songs about what people say than about the wreath and the ring together. And the vocabulary is concentrated there rather than diffused across the collection:
| Word | In 2.4.6.1 | Elsewhere | Lift |
|---|---|---|---|
| mēle tongue, tale-bearer | 5.1% | 0.2% | 26× |
| valoda talk | 42.9% | 1.9% | 23× |
| nievāt, nicināt scorn | 4.5% | 0.2% | 21× |
| paļas blame | 1.1% | 0.1% | 19× |
| ļaudis people, “what people say” | 49.4% | 2.8% | 18× |
Confirmed, and stronger than she states it. She argued from close reading that gossip is a principal mechanism of social control in this world. Barons had agreed structurally, by giving reputation its own section. The corpus shows the section is one of the largest in youth, with the sharpest lexical concentration of any test on this page.
The wreath and honour
RefinedThe young woman’s symbol of honor is her wreath or crown, primarily a sign of vital young womanhood and marriageable status and secondarily, most likely with Christian influence, of her purity and virginity. Chapter IX, Research Concepts and Tools
This one came back differently than expected, and the difference is worth more than a confirmation would have been.
The wreath is everywhere. Vaiņags appears in 4,598 songs, one in ten — more than any tree, any deity, or any other single object in the corpus. The word gods, honour, is not: 963 songs, 2.1%. And the two barely co-occur. Honour appears in 2.8% of wreath songs against a 2.1% base rate: a lift of 1.3×, which is nothing beside the 18–26× lifts of the previous claim.
So the equation is real, but it is not lexical. It lives where Barons put it — in a section title that spells the whole moral syllogism out:
2.4.4.2.1. Meita tura godu, sargā vaiņagu, netura godu, zaudē vaiņagu. “A girl keeps her honour, guards her wreath; fails to keep her honour, loses her wreath.”
— and it lives in what the songs narrate: a wreath given, guarded, wilted, floated down a river, lost. The songs enact the symbol instead of glossing it. Reading that equation out of them is exactly the work close reading does and counting cannot. On this claim her method sees something the corpus method would have walked straight past.
Whose songs survived
ConfirmedHistorical sources going back several hundred years generally attribute daina-singing to women and instrumental playing to men, a common enough Eurasian division. Chapter VI; and Chapter I on women as song leaders
Barons filed girls’ songs and boys’ songs as separate sections, which makes the asymmetry countable.
| Section | Songs |
|---|---|
| 2.4.7.1 Meitu dziesmas girls’ songs | 9,142 |
| 2.4.7.2 Puišu dziesmas boys’ songs | 3,691 |
Two and a half to one, and the girls’ songs are the single largest section in the corpus. Consistent with her claim and with the historical sources she cites.
What it cannot show is who sang. A song in a girl’s voice can be sung by anyone, and the collectors’ own preferences are inside this number too. It is evidence about the archive — which is the only thing a corpus is ever evidence about.
The missing mother insult
ConfirmedBurgen notes that absence of mother insults is characteristic of northern Europe. … This is in contrast to the Russian where most of the swear words are oriented toward the recipient’s family. Chapter VI, on nerātnās dainas
An insult at your mother has to be addressed at your mother, and Latvian marks that with a possessive. So: second-person possessive plus “mother”, against first-person.
| Phrase | Occurrences |
|---|---|
| mana / mūsu māte my / our mother | 871 |
| tava / jūsu māte your mother | 23 |
Thirty-eight to one. In 46,587 songs, the phrase “your mother” occurs twenty-three times in all. The daina mother is almost always the singer’s own, invoked as authority, teacher and protector rather than aimed at an opponent. Latvian insult in these songs goes at the opponent directly: at their looks, their work, their horse.
Laima’s verbs
RefinedLaima is the most powerful goddess, associated with verbs of creation, destiny, and fortune: laist, likt, lemt, vēlēt, rakstīt. Chapter IV, quoting Elza Kokare (1992)
A collocation claim, and collocation is the one thing a corpus is unambiguously good at. Do those five verbs occur more often in the 1,247 songs naming Laima than in the corpus at large?
| Verb | In Laima songs | Elsewhere | Lift |
|---|---|---|---|
| lemt to decree a fate | 2.7% | 0.1% | 46× |
| vēlēt to wish, to grant | 2.2% | 0.2% | 9.5× |
| likt to lay, to ordain | 10.3% | 1.3% | 7.7× |
| laist to let, to release | 0.8% | 1.3% | 0.6× |
| rakstīt to write, to pattern | 0.3% | 1.2% | 0.3× |
| controls — dot, iet, ņemt: 0.9×, 1.1×, 1.2× | |||
Three of the five, strongly. Two, not at all. The controls sit at 1.1×, which is what “no association” looks like, so the three-verb result is not an artefact of Laima songs simply being wordier.
Lemt, to decree a fate, is the signature: at 46× it hardly occurs in this corpus except beside her name. But rakstīt is the finding. Kokare read it as “to write”, a verb of divine inscription. In Barons’ youth branch it is a weaving word — he has a whole section called Adīšana, rakstīšana, šūšana — so its high frequency here is girls patterning mittens, not Laima writing fates. The claim is not wrong so much as drawn from a part of the collection this corpus does not hold.
One caveat, and it is not a small one: Laima the goddess and laime, “fortune, luck”, are the same word. Nothing here separates them. Arguably the songs do not either.
The female orientation
Out of reachMost of the deities remain female, the male gods representing different aspects of sky and atmospheric phenomenon (dark sky — Jods, thunder — Pērkons, wind — Vējš). Chapter IV, Myth, Belief Systems, and Cosmology
This is the test that comes back least comfortably, and the reason it does is worth more than the number.
| Figure | Songs | |
|---|---|---|
| Dievs | 2,703 | |
| saule the sun | 1,541 | |
| Laima | 1,247 | |
| mēness the moon | 268 | |
| Velns | 207 | |
| Māra | 93 | |
| Pērkons | 26 |
Taken at face value that inverts her claim. It should not be taken at face value, for three reasons that all point the same way.
Barons filed the mythological songs elsewhere, so the songs that would carry her argument are precisely the ones this corpus does not contain. Dievs in nineteenth-century recordings is both the old sky god and the Christian God, and no count separates them. And Dievs had travelled into everyday speech — paldies Dievam, Dievs dod — in a way Māra never did, so a frequency list here partly measures how far a name reached into idiom rather than how central it was to belief.
This one is on the page because a careless run of exactly this test would produce a confidently wrong answer, in a form that looks like data. Recording that is more useful than the number would have been.
What the corpus cannot answer
Roughly a third of the generalisations in the dissertation cannot be tested against any amount of text, and publishing that list is part of the result rather than an omission. Testing only the tractable claims and staying quiet about the rest would misrepresent both the method and her book.
Ethnomusicological — the corpus holds no melody
| “Most recitative songs have a range of five or six tones corresponding to the number of strings on the kokle.” |
| “The Latvians usually sing in two voices and drag out the last syllable, so that some do a bass-like drone.” |
| “Kurzeme has the most widespread drone singing, while Zemgale has the fewest surviving examples.” |
| Kursīte’s “polymetric, mixed trochaic-dactyllic meter particularly characteristic of apdziedāšanās” — syllable counts cannot see feet. What can be said: the taunt songs are 56.0% eight-syllable against 56.2% elsewhere, which is no difference at all. |
Regional — the corpus carries no region and no date per song
Every Kurzeme-versus-Latgale claim in the dissertation is out of reach for this reason alone, whatever else is true of it.
Outside this branch of Barons
| “Contest songs … are scattered throughout major collections, but most are concentrated in wedding, Midsummer, and work party groupings.” Directly testable against the full taxonomy; those three groupings are not in this corpus. The strongest candidate for a future test. |
| “Most often witches are mentioned in the dainas in connection with Midsummer.” |
| “Barons, for the most part, chose not to separate out the mythological dainas from the others.” A claim about the whole taxonomy, which needs the whole taxonomy. |
About performance, not about texts
“The singing contests were generally women against women, mixed, or men against women.” The texts survive; the events do not.
How this was done
Everything above runs on a single local snapshot of the corpus, harvested once and analysed offline — korpuss.lv is a public academic service run by a research institute, not a commercial API, and firing a query per question at it would be poor manners. The code is a few hundred lines of Python with no dependencies.
Three decisions that changed answers. Each of these produced a wrong result first, which is the only reason they are worth writing down.
Count windows, not tokens. Measuring formulaic density per word carries an edge artefact: a word at either end of a song sits inside one four-word window while a word in the middle sits inside four, so both ends are mechanically depressed. The first run of claim 3 showed a dramatic collapse at song endings that was almost entirely measurement.
Control for variants. This corpus is variant-rich — the same song is often recorded several times with a word changed — so “shared with another song” can just mean “written down twice.” Requiring a four-word sequence to cross two of Barons’ themes cuts “formulaic” from 27.2% of sequences to 1.7%, and makes the opening effect sharper, not weaker. The loose measure was understating the real finding.
Recover the unpunctuated line breaks. The printed edition does not always set a comma at a line end, so splitting on punctuation silently turns quatrains into three- and two-line songs. Uncorrected, claim 1 read 50% rather than 87%. The tell was a large population of fifteen- and sixteen-syllable “lines”, which is twice a daina line and therefore impossible.
And one that did not. Editorial variant readings are printed inline — Kam, meitiņa ( māsiņa), skaista augi. Those parentheses are the editor’s record of variation, not words anyone sang, and they are stripped before every count on this page.
What does not join
She cites Barons numbers — LD 23679 — ninety-three times. The corpus carries archive identifiers instead. Her exact cited song usually cannot be retrieved. Tested directly on her Apgāzās vācelīte quote: the corpus holds the same opening formula completed differently — hers ends with quarrelling sisters-in-law, the corpus variant with a girl losing her honour. So every test here matches on text and theme, never on citation. Her themes join perfectly; her references do not.
A note on the point of this
Her text is hers, and it is reproduced unaltered on this site. Everything on this page is apparatus, added twenty-five years later by someone who is not a folklorist, and marked as such throughout.
The obvious question about a page like this is whether the counting was ever going to tell us anything the reading had not. On the evidence above, sometimes: the front-loading of formula in claim 3, the scale of the gossip section in claim 5, and the split in Kokare’s verb list in claim 9 are all things nobody could have seen from close reading, because they are facts about tens of thousands of songs at once. And sometimes not: claim 6 is a case where her reading is plainly right and the counting is plainly blind, and claim 10 is a case where the counting would have been confidently wrong if anyone had trusted it.
That seems like the honest summary. The corpus is good at scale and bad at meaning. She was working the other way round, and she got most of it right.