NANA · MAMA · PAPA · TATA across 5,501 languages

That the word for mother resembles itself across unrelated languages is one of the oldest observations in comparative linguistics. It is also one of the best explained: Jakobson settled it in 1960 — the labial nasal is the first thing an infant produces with the mouth busy nursing, and the adult picks up that babble and hands it back as a name.

None of that is disputed here. What was missing was not the explanation: it was the measurement.

The procedure

Vowels are stripped from 100,507 lexical forms in Lexibank — 5,501 languages, 298 families — and each consonant is replaced by its sound class. mama → MM. ama → M. madre → MTR.

Over that skeleton, the question is whether the nurture-core concepts — mother, father, baby, grandmother, grandfather — draw on a repertoire of four codes: N (nana, ana), M (mama, ama), P (papa, apa, baba) and T (tata, ata).

The answer, and its margin

Affirmative, with a wide margin.

Among words formed from a single consonant class, the quartet takes 80.9 % in nurture against 48.0 % in ordinary vocabulary. Chance over ten classes would give 40 %.

The difference is 32.9 points, with a Cohen’s h of 0.71, and it repeats within 40 of 45 families compared against themselves (paired Wilcoxon p = 8.4·10⁻¹⁰, Cliff’s delta 0.78).

And the quartet does not merely rise: it expels. The velar class K, the most frequent single-class code in ordinary vocabulary at 18.1 %, falls to 6.7 %.

The correction to the headline, placed here and not buried below

The quartet does not rise as a block.

M, N and P go from 30.9 % to 63.0 % and carry almost the whole effect. T does not move: 17.2 % → 17.9 %, h = 0.02.

T is not a class nurture prefers: it is a class nurture does not expel, while it does expel K, S and R. What needs explaining is an avoidance, not a preference.

It does not depend on the class table

A result that existed only under one way of grouping consonants would not be a result. It was repeated under seven readings: Dolgopolsky as Lexibank ships it, Dolgopolsky recomputed with LingPy, SCA in two readings, ASJP, PanPhon articulatory features, and raw IPA with nothing merged — 201 distinct consonants, where chance gives the quartet 8 %.

Across all seven, the nurture − control difference holds between +31 and +34 points.

Nor does it depend on the large families. Removing Indo-European and Austronesian entirely leaves +30.7 points. And drawing one language per family, so that all weigh equally, leaves +31.8 — positive in 100 % of 500 draws.

The negative result: there is no infant cluster

Against expectation, and this is the part that interests me most.

With eleven concepts, 4,563 languages, the language as counting unit and attempts equalised, the infant falls ten points below its own ordinary vocabulary — 28.6 % against 38.7 %, h = −0.21 — in 24 of 27 families, and with no more reduplication than the control.

The reason is structural and can be read off the code length:

Languages do not name the infant with a primitive word. They derive it.

One effect does survive: transfer. When the infant does use the repertoire, it tends to take the code of the parents in its own language — 18.7 % against 14.5 % expected, p = 0.0005, h = 0.11: real and small. The code does not belong to the infant; it arrives on loan.

And a corpus gap that conditions all of the above

Lexibank records the term of reference, not the term of address. Of its 1,740,092 records, nine carry a vocative or address-term marking. And those nine indicate that the vocative is a different word, and closer to the repertoire.

That is: the affective register — what a mother says to the child in her arms — is not in the comparative databases. The figures in this work are a floor and not a ceiling, and no comparative corpus of that register exists for the vast majority of the languages treated.

A note on method

A phenomenon as visible as mama / papa invites apophenia: there will always be languages that fit.

So every claim is tested against a comparison block fixed before computing, and those that depend on family size are redone with the family — not the language — as unit. A cluster is not a historical family, nor a protoform, nor a rule: the languages inside may be unrelated to one another, and those outside are exceptions to nothing.

Every cluster is reported with two figures: how many languages and how many independent families. The second is what separates a convergence cluster (50–60 families) from a branch-inheritance cluster (4–7).

This work was done, among other reasons, as proof that the instrument recovers what is already known before being asked for anything new.

The paper: The Repertoire of Nurture (PDF) · Versión en español