Chance word resemblances

Why do linguists distrust a striking word match between two unrelated languages?

Persian bad means bad. Mbabaram, an Australian language with no plausible historical contact with English, has dog meaning dog. These are not approximate matches; they are direct hits in both sound and meaning, and the second one is famous precisely because it looks impossible.

The natural reaction is that such a match must mean something — shared ancestry, ancient contact, a lost migration. Historical linguists react the opposite way: a single striking match is treated as almost worthless evidence, and a proposal built on a list of them is dismissed. That looks like stubbornness. Understanding why it is not requires asking what a match is actually competing against.

Ask first what the alternative hypothesis is. Not "no explanation" — the alternative is chance, and chance is not a shrug. It makes a quantitative prediction, and that prediction is the thing a proposed match has to beat.

So estimate it. A language has a modest inventory of consonants and vowels. Consider a simple consonant-vowel-consonant word: with a couple of dozen consonants and a handful of vowels, and allowing the loose matching that comparisons in practice use — b counted as close enough to p, one vowel to a neighbouring vowel — the chance that two arbitrary words sound similar is not tiny. It might be one in a few hundred rather than one in a million.

Now ask how many chances there are. This is the step that does most of the work, and it is the same arithmetic that makes shared birthdays in a room surprising. You are not testing one pair. Compare a 200-item basic vocabulary list against another language's 200 items and you have tens of thousands of pairings, and in practice comparisons are looser still: a word may be allowed to match a semantically related item rather than an identical one, or any member of a small family of related meanings. Multiply a modest per-pair probability by that many trials and the expected number of accidental matches is not zero — it is several. Don Ringe's work in the early 1990s put this on a formal footing, showing that the yield of a comparison has to be measured against a calculated chance baseline, not judged by eye.

Which reframes the whole problem. Finding a handful of good matches between two languages is not evidence of relationship, because that is exactly what unrelated languages produce. And the number of language pairs on Earth is enormous, so somewhere out there a spectacular coincidence like dog is not merely possible but expected. Our attention is drawn to it precisely because it is extreme, which is the sampling bias that makes coincidences feel meaningful.

Then what does count? Ask what chance cannot produce. A random match is a one-off: b happened to line up with b in this word and nothing follows. Genuine common ancestry leaves a different signature — the same sound correspondence turning up again and again, across many words, in a regular pattern. Latin p answers to English f in pes/foot, pater/father, piscis/fish, pellis/fell. That recurrence is what the comparative method actually looks for, and its probability under chance falls off steeply with each additional instance, in a way a single dramatic match never does.

This has a striking consequence in the other direction: real cognates often do not look alike at all. Armenian erku and Latin duo both mean two, and their relationship is secure — not because they resemble each other, which they barely do, but because the sound changes linking them recur across the rest of the Armenian vocabulary. Similarity is neither necessary nor sufficient. Recurrence is the evidence.

Two further filters follow from the same logic. Compare basic vocabulary — body parts, low numerals, kinship, common verbs — because those resist borrowing, so a match there is less likely to be contact rather than descent. And weight shared irregular morphology very heavily: an idiosyncratic paradigm matching across two languages, such as the suppletion in be/was or an irregular plural pattern, is astronomically unlikely to be reinvented independently.

The distrust is not distrust of similarity as such; it is a refusal to evaluate evidence without its baseline. A comparison yielding thirty matches between two languages sounds impressive until you compute that unrelated languages of that structure would yield about thirty by chance — at which point the finding is exactly the null result.

This is the standing objection to mass comparison, the method of assembling large tables of look-alike words across many languages and reading relatedness off their volume. Its critics' central complaint is not that the data are fabricated but that no chance baseline is ever computed, so the method cannot distinguish a real family from the noise floor. Defenders reply that the aggregate pattern carries signal a pairwise calculation misses, and the dispute is genuinely unresolved at the level of statistical method, though the mainstream position is firmly that recurrent correspondences are what establish a family.

Worth adding: this is why time depth is a hard limit. Sound change keeps accumulating, so after roughly eight to ten thousand years the number of surviving regular correspondences sinks toward the chance level, and the method stops being able to tell relationship from coincidence — not because deeper relationships do not exist, but because the evidence for them has been erased.

• How a chance baseline is actually computed for a given pair of languages, and why the estimates are contested.
• Why irregular morphology is treated as the strongest single class of evidence for relatedness.
• The time-depth ceiling: what, if anything, can be said about relationships older than the comparative method can reach.

Every term the collection defines is gathered in the glossary.