Verbs which obey standard rules of conjugation in their native language are called regular verbs¹³. In the modern English language, regular verbs are conjugated into the simple past and past participial forms by appending the dental suffix -ed to the root (for instance, talk/talked/talked). Irregular verbs obey antiquated rules (sing/sang/sung) or in some cases, no rule at all (go/went)^14,15.

New verbs entering English universally obey the regular conjugation (google/googled/googled), and many irregular verbs eventually regularize. Regular verbs become irregular much more rarely: for every sneak that snuck in¹⁶, there are many more flews that flied out.

Although less than 3% of modern verbs are irregular, the ten most common verbs are all irregular (be, have, do, go, say, can, will, see, take, get). The irregular verbs are heavily biased towards high frequencies of occurrence^17,18. Linguists have suggested an evolutionary hypothesis underlying the frequency distribution of irregular verbs: uncommon irregular verbs tend to disappear more rapidly because they are less readily learned, and more rapidly forgotten^19,20.

In order to study this phenomenon quantitatively, we studied verb inflection beginning with Old English (the language of Beowulf, spoken circa 800 CE), continuing through Middle English (the language of Chaucer’s Canterbury Tales, spoken circa 1200 CE), and ending with Modern English, the language as it is spoken today. The modern -ed rule descends from Old English ‘weak’ conjugation, which applied to 3/4 of all Old English verbs²¹. The exceptions - ancestors of the modern irregulars - were mostly members of the so-called ‘strong’ verbs. There are 7 different classes of strong verbs with exemplars among the modern English irregulars, each with distinguishing markers that often include characteristic vowel shifts. Though stable coexistence of multiple rules is one possible outcome of rule dynamics, this is not what occurred in English verb inflection²². We therefore define regularity with respect to the modern -ed rule, and call all these exceptional forms ‘irregular’.

We consulted a large collection of grammar textbooks describing verb inflection in these earlier epochs, and hand annotated every irregular verb they described. (See Supplementary Information.) This provided us with a list of irregular verbs from ancestral forms of English. Eliminating verbs which were no longer part of Modern English, we compiled a list of 177 Old English irregular verbs which remain part of the language to this day. Of these 177 Old English irregulars, 145 remained irregular in Middle English, and 98 are still irregular in Modern English. Verbs such as help, grip, and laugh, which were once irregular, have become regular with the passing of time.

Next we obtained frequency data for all verbs by using the CELEX corpus, which contains 17.9 million words from a wide variety of textual sources²³. For each of our 177 verbs we calculated the frequency of occurrence among all verbs. We subdivided the frequency spectrum into six logarithmically spaced bins from 10⁻⁶ to 1. Figure 1a shows the number of irregular verbs in each frequency bin. There are only two verbs, be and have, in the highest frequency bin, whose mean frequency is >0.1. Both remain irregular to the present day. There are eleven irregular verbs in the second bin, with mean frequency between 0.01 and 0.1. These eleven verbs have all remained irregular from Old English to Modern English. In the third frequency bin, 0.001 to 0.01, we find that 37 irregulars of Old English all remained irregular in Middle English, but only 33 of them are irregular in Modern English. Four verbs in this frequency range, help, reach, walk, and work, underwent regularization. In the fourth frequency bin, 10⁻⁴ to 10⁻³, 65 irregulars of Old English have left 57 in Middle and 37 in Modern English. In the fifth frequency bin, 10⁻⁵ to 10⁻⁴, 50 irregulars of Old English have left 29 in Middle and 14 in Modern English. In the sixth frequency bin, 10⁻⁶ to 10⁻⁵, 12 irregulars of Old English decline to 9 in Middle and only one in Modern English: slink, a verb which aptly describes this quiet process of disappearance.

Plotting the number of irregular verbs against their frequency generates a unimodal distribution with a peak between 10⁻⁴ and 10⁻³. This unimodal distribution again demonstrates that irregular verbs are not an arbitrary subset of all verbs, because a random subset of verbs (such as all verbs that contain the letter ‘m’) would follow a power law distribution with a slope of three-fourths^24,25.

Four of our six frequency bins, those between 10⁻⁶ and 10⁻², allow us to estimate the relative regularization rates of irregular verbs. Calculating the relative regularization rates of verbs of different frequencies is independent of time, which makes the dating of Old and Middle English irrelevant for this calculation. We can draw regularization rate versus frequency and fit a straight line in a log-log plot (Figure 1b). Comparing Old and Modern English we obtain a slope of about −0:51. Therefore, an irregular verb which is 100 times less frequent is regularized 10 times as fast. In other words, the half-life of irregular verbs is proportional to the square root of their frequency. Comparing Middle and Modern English we find a slope of about −0:48, consistent with the previous result. Both comparisons show that low frequency irregulars are selectively forgotten.

Figure 2a shows the exponential decay of the irregular verbs in the four frequency bins between 10⁻⁶ and 10⁻² as a function of time. From these data, which depend on the dating of Old and Middle English, we can estimate actual half-lives of the irregular verbs in different frequency bins. Irregular verbs that occur with a frequency between 10⁻⁶ and 10⁻⁵ have a half-life of about 300 years, while those with a frequency between 10⁻⁴ and 10⁻³ have a half-life of 2000 years. If we fit half-life versus frequency with a straight line in a log-log plot, we obtain a slope of 0.50, which again suggests that the half-life of irregular verbs is proportional to approximately the square root of their frequency (Figure 2b). It is noteworthy that various methods of fitting the data give the same results.

We cannot directly determine the regularization rate for frequency bins above 10⁻², because regularization is so slow that no event was observed in the time span of our data. But we can extrapolate. For instance, the half-life of verbs with frequencies between 10⁻² and 10⁻¹ should be 14,400 years. For these bins, the population is so small and the half-life so long that we may not see a regularization event in the lifetime of the English language.

To test whether the dynamics within individual competing rules were captured by our global analysis, we studied the decay of individual classes of strong verbs (e.g., hit/hit/hit, hurt/hurt/hurt; draw/drew/drawn, grow/grew/grown)²⁶. Although our resolution is limited by the small sample size, exponential decay is once again observed, with similar exponents. (See Supplementary Figure S1.) Like a Cheshire cat, dying rules vanish one instance at a time, leaving behind a unimodal frown.

Because adequate corpora of Old and Middle English do not exist, we have estimated the frequency of an irregular verb of Old and Middle English by the frequency of the corresponding (regular or irregular) verb of Modern English.²⁷ A large fraction of verbs would have had to change frequency by several orders of magnitude in order to interfere with the effects observed. To verify that large changes in frequency are rare, we compared frequency data from CELEX with frequencies drawn from the largest available corpus of Middle English texts²⁸. Out of fifty verbs, only five had frequency changes greater than a factor of 10. (See Supplementary Figure S2.)

Our analysis covers a vast period, spanning the Norman invasion and the invention of the printing press, but these events did not upset the dynamics of English regularization.

Therefore, it is possible to retrospectively trace the evolution of the irregular verbs, moving backwards in time from the observed Modern distribution and up through Middle and Old English. Going still further back in time allows us to explore the effects of completely undoing the frequency-dependent selective process which the irregular verbs have undergone. Eventually, the shape of the curve changes from unimodal to a power law decline with slope nearly −3/4 (Figure 3). This finding is remarkably consistent with the fact that random subsets of verbs (and of all types of words) exhibit such a Zipfian distribution. The observed irregular verb distribution is the result of selective pressure on a random collection of ancestral verbs.

We can also make predictions about the future of the past tense. By the time one verb from the set {begin, break, bring, buy, choose, draw, drink, drive, eat, fall} will regularize, five verbs from the set {bid, dive, heave, shear, shed, slay, slit, sow, sting, stink} will be regularized. If the current trends continue, only 83 of the 177 verbs studied will be irregular in 2500.

What will be the next irregular verb to regularize? Most likely it will be wed/wed/wed. Wed’s frequency is only 4.2 uses per million verbs, ranking at the very bottom of the modern irregulars. Indeed, it is already being replaced in many contexts by wed/wedded/wedded. Now is your last chance to be a newly-wed. The married couples of the future can only hope for wedded bliss.

Supplementary Online Information

Supplementary English Verbs