Irregular verb survival

Why do the verbs a language uses most often stay irregular while rarer ones quietly become regular?

Line up the most-used verbs of English — be, have, do, go, say, make, take, come, see — and almost none forms its past tense with -ed. Now think of a verb you use rarely: chide, cleave, slink, thrive. Several had irregular pasts within recorded history and most speakers now regularise them without noticing.

The pattern looks upside down. Irregular forms are the awkward ones, so the obvious prediction is that they should be eliminated first where they are met most often. Instead frequency appears to protect irregularity. Something in that reasoning is wrong, and locating it tells us how a grammatical pattern survives.

Start with what an irregular form is, from the point of view of someone who has to produce one. Walked can be built: take the stem, add the ending — and you can do it for a verb you have never met, which is why to google immediately had a past tense. Went cannot be built from anything. It has to be retrieved from memory as a whole, and that memory has to have been laid down by hearing it.

That asymmetry is the engine. The regular pattern is a rule, available by default and costing nothing to extend. An irregular form is a stored item, and stored items require exposure to be acquired and maintained.

Now run a child through it. She meets a verb. If a stored past tense comes to hand, she produces it. If nothing comes — because she has not heard it often enough — she is not stuck: the default rule is sitting there. Notice what has happened. A failure of retrieval does not produce silence or an error the community rejects; it produces a perfectly usable word other people understand. The failure is survivable, which means it can spread.

Iterate over a generation. Verbs heard constantly are stored reliably by every learner, so the irregular passes on intact. Verbs heard seldom are stored patchily; some learners never acquire the irregular, produce the regular form as adults, and thereby reduce the next cohort's exposure further. That is a ratchet turning in one direction, because the regular form is reachable from ignorance and the irregular one is not.

Does the account predict anything measurable? It predicts that the rate at which irregular verbs regularise is governed by frequency of use, quantitatively rather than vaguely. A study by Lieberman, Michel, Jackson, Tang and Nowak, published in Nature in 2007, tracked the Old English strong verbs across the history of English and reported that regularisation rate scales roughly with the inverse square root of frequency — I am recalling that finding rather than deriving it. The arithmetic it implies is honest enough to state: a verb used a hundred times less often regularises about the square root of a hundred — ten times — faster. So the protection is real but sub-linear, which means even common irregulars sit on the same slope, merely further along it.

A second prediction works at the scale of one speaker. If irregulars are stored and regulars computed, over-regularisation errors — goed, bringed, catched — should appear in children who plainly have heard the correct forms, and should fall disproportionately on the less frequent irregulars. Both are reported.

What would refute this? Cleanly stated: find that among the Old English strong verbs, the ones that regularised were no rarer than the ones that survived — that the sorting was by phonological shape, or by nothing at all, rather than by frequency. Or find a language in which the most-used verbs are the most regular ones and the rare verbs carry the suppletive forms. Either observation and the storage-and-retrieval account is finished.

Three refinements, and one belief worth naming as false.

First, "frequency protects" is a statement about transmission, not difficulty. Nothing here says irregular forms are hard and get eliminated by an efficiency-seeking language. The mechanism is about whether a learner encounters a form often enough to store it, and travel is asymmetric only because the regular rule is available to someone who has learned nothing.

Second, the ratchet is not perfectly one-way. Analogy occasionally runs the other direction — dive acquired dove on the model of drive and drove, sneak acquired snuck — so irregulars can be created as well as lost, though far more rarely, and by attraction to a strong existing pattern rather than by invention. This is where I have simplified hardest.

Third, the underlying cognitive story is contested even though the frequency effect is not. The account I have given is a dual-route one: a symbolic default rule plus a store of memorised exceptions, associated with Steven Pinker. Single-route connectionist models, following Rumelhart and McClelland, reproduce the same frequency dependence and the same childhood over-regularisation from one associative network with no separate rule in it. The observations do not discriminate between them as easily as either side has sometimes claimed, so treat the mechanism as agreed at the level of frequency and exposure, and the architecture as open.

Now the false belief. It is routinely said that irregular verbs are decaying remnants and that English is gradually becoming regular — often with the prescriptive rider that we should hurry it along, or resist it. The first half is a real historical trend for the tail of rare irregulars. The second does not follow: the highest-frequency verbs show no sign of going, and be has been irregular throughout the recorded history of the language and its ancestors. A process that removes rare exceptions and preserves common ones is not converging on regularity but on a stable core of very frequent irregulars, which is roughly what English now has.

• Why suppletive forms such as go / went, drawn from two different verbs entirely, are found among the most frequent verbs of many languages.