5,505 languages · 371 families · 199,652 words

The Word That Gives Itself Away

Here are forty words from a language you have almost certainly never heard of, and the forty meanings they carry, in a separate pile. Nothing connects the two columns except sound. Match them.

The board

language

Loading the board.

the words

the meanings

you 0 of 0 the model — guessing —

Words are written in ASJPcode, the 41-letter alphabet the source database uses, with an IPA reading underneath. The model saw no language of this one's family. Forty words to forty meanings, one each: a random pairing gets exactly one right, on average.

Companion film, 2:38. Every pitch in it is one of the numbers on this page. The law is a single multiplication: hertz is 4400 times accuracy, which puts chance, one meaning in forty, at exactly 110 hertz. That is the drone, and it is under the whole film from the first second, because chance never moves. The ladder descends through it as each leak is closed. The free null is then held for twelve seconds against it, and because 2.4933 per cent is not quite 2.5000 the two of them beat, once every three and a half seconds, which is the sound of a null landing on chance and not quite on it. Then the forty enter loudest first, each at its own accuracy, and the twelve that give nothing away sit at or below the drone. The rhythm at the start is a real wordlist: one pluck per word, spaced by that word’s own length.
research/the-word-that-gives-itself-away/film/check-encoded.mjs cuts those spans out of the delivered MP4, decodes them through AAC and measures them, so the caption is checked against what your speaker actually does. 9 of 9.

That is the whole question, and it is older than linguistics as a discipline: is the link between a word and its meaning arbitrary? The orthodox answer, since Saussure, is yes, and the orthodox answer is nearly right. This page is about the part where it is not, and about how easy it is to overstate that part.

What the model is not allowed to see

The trap in a question like this is relatedness. English nose, German Nase and Sanskrit nāsā all carry an n, and that is not sound symbolism, it is one word inherited three times. A model trained on Indo-European and tested on Indo-European would score handsomely and would have measured nothing but descent. Every number here is produced under one rule: the model never sees the test language, and never sees a single language of the test language's family. Each of the 371 families is withheld in turn, and every family is given equal weight in the training set, so that Austronesian and Atlantic-Congo, which between them supply 37% of the data, cannot carry a result on their own.

The model is deliberately plain: a naive Bayes classifier over the character one-, two- and three-grams of the word, with word edges marked. It knows no phonology. It has never been told that n is a nasal. The only thing it can learn is which letter sequences go with which of the forty meanings.

The ladder, and where it lands

Each rung below removes one more way for the answer to leak in. The last two rungs are nulls: the meanings are shuffled so that no sound-meaning link exists at all. If the pipeline scored above chance there, everything above it would be an artefact.

How much of it is just the length of the word

A word's length is not sound symbolism, and it carries meaning anyway. Words for I and you are short nearly everywhere; words for breasts and mountain are long. A model given nothing but the number of letters in the word, and no letters, already beats chance.

Which of the forty talk

An average across forty meanings is a poor description of what is happening, because the meanings are not alike. A handful give themselves away repeatedly. Most give away nothing, and several are worse than a coin.

Amber: how often the model names this meaning correctly, out of forty, with the family withheld. Blue tick: the same model on the same words with the meanings shuffled among words of equal length, which is what length alone is worth for this meaning. The grey line is a one-in-forty guess.

Which sounds, and in whose mouths

The classifier is a blunt instrument; it says a word leaks without saying what leaks out of it. The table below asks the narrower question directly, one sound at a time, with two guards. Each of the 371 families contributes exactly one number, so no family can vote twice. And the comparison is not against how common the sound is in general but against how often it would turn up by chance in a word of that length, drawn from that same language's other words, which is the only way to keep the test from quietly measuring length again.

meaning	sound	in	expected	ratio	families	areas	Blasi 2016

Pick a row above.

Three keys other people cut

A result like this is exactly the kind that a pipeline can manufacture, so the useful question is not whether the numbers are internally consistent but whether they agree with work done by other people, on other data, with other statistics. Three chances to be caught:

1. A list published in 2016, and how much of it comes back

2. A ranking published in 2010, on a different version of the database

3. A ratio published in 2021, which predicts how wrong this page is

What was thrown away, and why

Where this disagrees with the literature

In 2026 Frederic Blum re-ran the largest of these surveys with an explicit phylogenetic covariance and an explicit spatial covariance in the model, and most of the published associations did not survive. What did survive, in his words, were the pronominal nasals and the lateral in the word for tongue. The controls on this page are cruder than his: withholding a whole family is a hard cut, but it treats every family as equally distant from every other, and it treats a language's neighbours as independent of it unless they are explicitly removed. That is why the neighbour rungs and the macroarea rung are on the ladder, and why the honest reading of the headline is the bottom of its interval rather than the top.

The check

Every number above is recomputed from the committed artefacts, and the list below is generated from them rather than typed.

The offline verifier is research/the-word-that-gives-itself-away/verify.mjs. Run it with --mutate and it corrupts its own inputs on purpose to prove the checks can go red. The browser check is research/the-word-that-gives-itself-away/verify-page.mjs.

What this does not say

It does not say the model is the best available estimator of the leak. It is a deliberately plain one, chosen before the sweep and kept after it even though the sweep found better. Every variant tried is in the table below, and the whole reasonable space lands within about one point of the headline.
It does not say the sound-meaning link is causal, or that it comes from iconicity rather than from shared human physiology, from nursery vocabulary, or from ancient contact too deep for any family classification to see. It measures that the link is there after relatedness inside a family is removed. Deeper relatedness than that is precisely what nobody can rule out.
It does not say anything about the 60 items of the hundred-item list that ASJP fills in for only a minority of its languages, nor about vocabulary outside basic vocabulary, where the answer may be quite different and is not measured here.
ASJPcode throws away vowel length, nasalisation, tone, retroflexion and the voicing distinction in several places. Anything carried by those is invisible to this page, and its absence pushes the estimate down, not up.
The playable board is one language at a time and your own score on it is worth nothing statistically. It is there so the task is a thing you have done rather than a number you were handed.

Every variant tried

arm	free choice	forced matching	top 5