Skip to navigation – Site map

HomeIssues16PapersDiachronic Emergence of Zipf-like...


Diachronic Emergence of Zipf-like Patterns in Construction-Specific Frequency Distributions: A Quantitative Study of the Way Too Construction

Quentin Feltgen


The linguistic constituents that come to express path are often co-opted into more abstract schematic constructions through a diachronic process known as constructionalization. This applies to the way too intensifier construction, whose origin I retrace to a former away too construction. Moreover, the emergence of a schematic construction is associated with an evolution of its type frequency (the total number of different linguistic units that fill the free slot of the construction). Although the evolution of the token frequency throughout a constructionalization process has already been studied from a quantitative perspective, the type frequency has not received such an empirical characterization. I show that the type and token frequencies are related with one another through a Zipf’s law, whose coefficient varies with time, and which sorts the different types according to their collocate frequency in a way that is robust both across time and across varieties of English.

Top of page

Full text


1The lexical expression of path in languages has often led to further semantic developments that have resulted in complex, schematic constructions, through a process known as constructionalization (Traugott & Trousdale [2013], Hilpert [2013a]; see also Rostila [2004] for an earlier discussion), that is, the process through which a new cognitively entrenched form-meaning pair emerges in the language of a community of speakers. For instance, the motion-manner way construction in English (Israel [1996], Cotte [2012], Perek [2018]) or the preposition en voie de construction in French (De Mulder [2019]) build upon the lexical element way and voie respectively, which are close translations of each other (and derived from the same Indo-European root ultimately). The resulting constructions, [{V} {one’s} way] for the motion-manner construction and [en voie de {N}] for the complex proposition, are said schematic, because they define a schema including a free slot to be filled (respectively {V} and {N}) that offers paradigmatic variation.

2Constructionalizations leading to such partially abstract schemas offer a special challenge to the study of language change, which has already been partly addressed within the grammaticalization framework (Hopper & Traugott [1991]), in which the mechanisms at work in these changes, such as reanalysis, pragmatic inference, and phonetic reduction have already been identified and discussed. From a quantitative point of view, the rise of token frequency associated with a grammaticalization process is known to follow a well-identified pattern, the S-curve (Kroch [1989], Croft [2000], Nevalainen [2000], Fagard [2009], Mair [2013], Rissanen [2013], Fagard & Combettes [2013], Feltgen et al. [2017], Danino et al. [2020]). The S-curve is nevertheless not specific to grammaticalization or even constructionalization, and is a more general feature of language change (Weinreich et al. [1968], Blythe & Croft [2012], Ghanbarnejad et al. [2014], Nevalainen [2015]). According to this pattern, the frequency increases slowly at first, then accelerates, then slows down again and reaches a plateau, with the further characteristic that this successive increase and decrease in momentum should be roughly symmetrical, so that the maximum increase of frequency per unit of time should be reached about halfway in the process.

3However, if the S-curve describes frequency rise over time, interpreting it is not straightforward. Indeed, the frequency rise can be due to a number of factors: Ogura & Wang [1996] have distinguished the S-diffusion (over speakers) and the W-diffusion (over words or, more generally, types) both assumed to obey an S-curve. There also remains the likely possibility that the frequency encompasses more than speaker diffusion and type diffusion: as they get entrenched, some linguistic forms can just be produced more often. Anyway, the S-curves extracted from corpus data typically conflate these sources of increase, and reflect a rise in token frequency (the raw number of occurrences of the new form in the corpus), even though in some instances, different contexts can be considered separately (Kroch [1989]).

4The type frequency (the number of different types this form is found to be associated with in the corpus), even though it is known to play a crucial role in grammaticalization processes (Bybee & Thompson [1997]), is usually left aside from these quantitative corpus-based studies. It is nevertheless all the more important since the semantics of a schematic construction is deeply related to the types it can host (Barðdal [2008], Perek [2016]) so that an increase in type frequency is associated with a semantic widening of the construction, a feature which is again characteristic (but not unique) of grammaticalization. It seems therefore critical to tackle this dual increase of frequency of a constructionalizing construction from a quantitative perspective. In this paper, I shall offer new ways to characterize these two measures of frequencies and how they relate to each other over time. The sociolinguistic aspect of the propagation of a change over the community of speakers will be, however, left unaddressed.

5From now on, I will concentrate on the way too construction, which is used as an intensifier in Present-Day English: e.g., from the COCA corpus (Davies [2008-]): “That’s how I learned that the pain biologique, a loaf that sounded way too crunchy-granola for my tastes, was actually exotic and elegant”. This construction is schematic in that it admits a free slot (the {ADJ/AV} in [way too {ADJ/ADV}]), which can either be fit by an adjective (e.g. big), an adverb (much, many), or as attested by the example above, a noun or any part of speech which gets coerced into an adjectival reading (Lauwers & Willems [2011]). This possibility for coercion is a clear sign that the construction is nowadays cognitively entrenched in the constructicon of the speakers, that is, in their mental repertory of constructions.

6The first occurrences of the construction are consistently found in the 1930s-1940s. This recent development allows us to accurately retrace the whole constructionalization process from its very beginning using the corpus data. In the first section of the paper, I will investigate the possible source of this construction: a phonetic reduction of an away too construction will be my main hypothesis, and I will discuss the possible role played by other constructions (the every way intensifier construction and the give way formulaic verb phrase) on the emergence of way too. In the second section, I will quantitatively characterize the rise of the token frequency as well as the rise of the type frequency, and I will show the relationship between them through the Heaps-Herdan law. Finally, in the third section, I will investigate the organization of the different types of the construction in terms of their own individual token frequencies, showing that they obey a Zipf-like distribution. This suggests that the construction is not only characterized by its type and token frequencies, but also by the characteristic and robust organization of its types into a cohesive and constrained relationship. To support this claim, I will more specifically consider the robustness of the Zipf ranking of the types over different English varieties.

7Throughout this study, I rely on data from four different corpora: the COCA (Davies [2008-]), the COHA (Davies [2010-]), the Google Books Ngram corpus as accessed through the interface designed by Davies [2011-]) and the GloWbE corpus (Davies [2013]). All these corpora have different uses and limitations. The COCA corpus is large, and covers a lot of colloquial sources, which makes it excellent to obtain an accurate picture of the Present-Day use of the way too construction. I used it chiefly in Section 1.1. to sketch a broad portrait of the construction’s usage, as well as in Section 3.1. to derive the most complete possible Zipf’s law associated with the construction. Given its limited temporal horizon (1990-2019), however, it is unsuitable for a diachronic investigation. For this purpose, the COHA corpus is one of the most reliable databases available. It has been cleverly designed, and all ‘hits’ of a query can be checked manually. It covers all decades from 1810-1819 to 2000-2009, even though not all decades are equally represented. The main limitation of this corpus is its size, which is relatively small, and often insufficient to conduct quantitative analyses in a satisfyingly robust way: 400 million words spread over 20 decades, to be compared with the COCA corpus and its 1 billion words for the last three decades.

8The Google Books Ngram corpus seems to offer the advantages of both corpora: it encompasses a very large number of sources (155 billion words) and covers the English language from the sixteenth century. I mostly avoided using it however, except to occasionally pick up an attested example to illustrate some point under discussion. The reason for this is that Google Books Ngram is extremely unreliable: the data is made of n-gram counts (for each n-gram, and each year, the data provides the number of hits), so that the context cannot be checked. It also encompasses a large number of OCR errors. On top of that, the date associated with the texts is too often incorrect: a novel published within the last ten years can be dated from the 1950s for no reason, while recent academic reprints of past century works will be listed under the date of the reprint, often in multiple editions. Some texts are repeated over a dozen times. Therefore, I deemed it better to rely on a much smaller, but also much more reliable and controlled corpus. Finally, I only used the GloWbE corpus for the last Section 3.3., to check the robustness of the Zipf ranking over the different varieties of English that it covers.

1. Origin and diachronic development of the way too construction

1.1. Broad characteristics of the way too construction

9The way too construction is an intensifier construction which can take two kinds of arguments, adverbs (1) or adjectives (2), although the different types of adverbs that can be used with the construction form a very restricted set. There is no reason to consider these two uses of the construction as two separate constructions: they can, for instance, work together in conjunction (3), and as we shall see, they collectively obey cohesive frequency patterns over time. The construction semantically carries an overt judgment of excess over the argument it is applied to, and can most often be glossed by ‘excessively’. Although the subjective valence of the construction is usually negative, the construction can also be found in an emphatic context (4).

(1) I’ve spent way too many nights in a bar where men think there’s glamour in a marginalized status.

(2) I think some of the comments here are way too harsh.

(3) I can agreed (sic) to some point that Yes – our government is way too big and some are earning way too much.

(4) That’s amazing! That is sexy and way too cute. So pretty!

10There also exists a sister construction with a similar use, the [way {ADJ/ADV}] construction. This sister construction also acts as an intensifier, but is semantically different in that the absence of too does not entail a negative quality of the intensification, which can therefore be emphatic (5) or neutral (6). Furthermore, this sister construction is often found with adjectives in comparative forms (7), and can associate with more (8), while the way too construction cannot. The form way much is occasionally found, but mostly in comparative contexts.

(5) Ty was way hot, totally gorgeous.

(6) That’s the fun thing; I’m not that fighter anymore and Amanda is not that fighter anymore too. I believe this is going to be a way different fight.

(7) Frankly, I think this could have been said in a way funnier and less emotionally damaging way.

(8) Having good Wisdom and Charisma will give you way more EXP if you work your way through sometimes-obscure dialogue choices.

11This sister construction, although most likely related to the way too construction, is semantically different, and does not show the same attachment preferences. It also follows its own diachronic pattern of development, arising in the 1990s only. For this reason, I will not devote more attention to the intensifier lone way construction in this study.

1.2. Earliest attestations of the way too construction

12As is often the case with emerging constructions, the first attested occurrences of the way too construction are debatable. The earliest one that I found in the COHA corpus dates from the very late nineteenth century (9):

(9) Pass over them documents for Cherokee Hall, an’ don’t hold out nothin’ onto us. We-alls is ’way too peevish to stand any offishul gaieties to-day. (1897)

13The quote is from a novel, Wolfville: Episodes of Cowboy Life, written by Alfred Henry Lewis in Kansas and published in 1893 (the 1897 date of the occurrence is the one provided by the COHA). The peculiarity of this novel is that it is narrated by a fictional character, the Old Cattleman, whose language is characterized by marked Southern specificities (he is supposed to come from Tennessee). Furthermore, the quote featuring the way too occurrence is a piece of dialogue, highlighting the highly colloquial nature of the construction. However, the way element in the occurrence is preceded by an apostrophe, to indicate that this is, instead, a phonetic reduction of away. Therefore, this first occurrence is not strictly speaking an occurrence of the way too construction, but rather an occurrence of another, possibly related, away too construction, as I shall discuss later on.

14The second occurrence (10) that I found in the COHA is from the March 1939 issue of the Harper’s magazine, within a three-page short story by Preston Quadland titled ‘Setter and Terrier’. Here again, the construction is to be found in a dialogue (between the two dogs that give this short story its title), but it is not especifically marked. It is shortly followed by a third occurrence (11) from the story “Vortex Blaster” published in the five (and last) issue of the Comet pulp magazine, in July 1941. An apostrophe indicates here as well that the way element of the way too construction is rather a phonetic contraction of away. Finally, way too much is recorded for the first time in the COHA in 1951 (12), in a dialogue of the novel The Cruel Sea by the British writer Nicholas J.T. Monsarrat, according to the COHA listing. It is rather surprising to find a British novel in the COHA corpus, devoted to American English. This is, however, an error from the corpus metadata. The excerpt is nowhere to be found in that novel; instead, it belongs to The Origin of Evil, published in 1951 under the pseudonym Ellery Queen by two American writers.

(10) “My dear fellow, your ears are way too keen for a gentleman’s gentleman,” remarked Terrier. (1939)

(11) The activity of the vortex stayed high, ’way too high. The tiny control room of the flitter grew hotter and hotter. (1941)

(12) “A mistake,” said the detective. “It happens all the time. Either they use way too much or way too little (1951)

15An earlier instance of way too much can be found in the Google Books corpus in the editorial of the Farm Life magazine (Vol. 46, Issue 4), dated from 1927, which also features an explicit attrition of away (13). This particular occurrence is surprising, because the context is clearly not colloquial, and yet it features a marked phonetic reduction.

(13) So drastic a reduction is questionable, but at the least, the Department expects to see a considerable area shifted from cotton to corn and like crops – which, it is believed, will give us ’way too much of these. (1927)

16From the 1950s, the construction becomes more and more widespread. The quantitative study of this frequency rise shall be the focus of Section 2. I shall only mention two major milestones in that development: the appearance of the sister way construction (without too, and discussed above), attested in the COHA from 1994 (see (14), with numerous other instances with other types in the late 1990s and the 2000s) and the appearance of occurrences that elaborate on the nominal nature of the way too construction by turning it into a full Noun Phrase (see (15) and (16) from the COCA corpus). This latter use of the way too construction seems fairly recent and is not to be found in the COHA corpus.

(14) These cars weren’t show pieces. Just by looking at them, I could see they had character. “Way cool,” I said to Ray. (1994)

(15) And there Mother and I sat through most of my ten nights at home. We were a good way too far gone in life before TV to become the instant zombies that most later Americans are, at a flicker of the tube. (2011)

(16) Okay, am I being punked?’ Cause you’re likin’ me a little way too much. (2018)

1.3. Possible sources of the development of the way too construction

17As we have seen, there is evidence that the construction way too could be the result of the phonetic reduction of an earlier construction away too, similar in meaning and use. As such, this hypothesis deserves further consideration. Nonetheless, concurrent hypotheses might shed light on the constructionalization of the way too construction and shall be investigated as well.

1.3.1. The away too construction

18Two among the first three attested examples of the way too constructions are phonetic attritions of a rather close construction, away too. That, and the fact that the examples resulting from this attrition are indistinguishable from the occurrences of the way too construction in their use and meaning, concur with the idea that way too is but the phonetic reduction of the away too construction.

19The away too construction in itself is only marginally attested – which could be due to the scarcity of the coverage of the relevant periods. There are, however, clear examples of that construction (17-20) in the first part of the twentieth century. Therefore, this would be consistent with the rise of the way too construction happening by the end of that period, as seen with examples (10) to (12). Interestingly, all these examples are taken from dialogues, highlighting the colloquial and oral nature of the construction, a trait crucially shared by the way too construction.

(17) He said, ‘Well, your line looks pretty good; but, heavens alive! your prices are away too high.’ (1905)

(18) She examined the shoes critically. “They seem to fit,” she said, “but they are away too fine to walk country roads.” (1909)

(19) “You just stop, Missy!” she cried. “You’re away too smart, trying to get folks in here, and ruin my George’s chances […]” (1918)

(20) “We know we’re on Andover Road but it’s too dark to travel much further.” “I do agree wi’ ye! Away and away too dark and dreary, to be out on the long road. …” (1945)

20Beside these four attestations, there seems to be no further occurrences of this construction in the COHA. Chronologically, it fits with the earliest attestations of way too: away too could have been used in the very late nineteenth century (I found one occurrence dated from 1905), remaining in the American English language until the mid-twentieth century, in colloquial contexts only throughout this whole period. Past the 1940s, the construction is dropped in favor of its phonetic reduction, way too.

21The construction away too is part of a higher-level node of away + ADV constructions, and seems closely derived from the neighboring far + ADV intensifier construction. The far too construction in particular is much more frequent than the way too one (a fortiori more than the away too one) and is attested over the whole period covered by the COHA, from 1811 (21) to 2009 (22):

(21) Vivid lightning may strike and melt me; a false friend’s compunction is far too weak a fire to make me bend. (1811)

(22) She’d simply been relieved Ada finally decided to get help running the house. The work was far too much for the older woman, suffering from arthritis. (2009)

22The away + ADV construction might have been influenced from this construction, as suggested by the numerous instances of the variant far and away + ADV (far and away superior, from 1891, far and away ahead, from 1896, far and away beyond, from 1899, far and away more from 1905), including far and away too from 1939:

(23) He gestured toward those which had been marked for lambing, and said brusquely: “Yes, but the lambs so far have been poorly, some misformed, many dead. […] And the ewes are laboring too hard, look at my kalat – there’s far and away too much blood –” (1939)

23However, away + ADV is well attested before the first occurrences of the far and away construction: 1819 for away behind, 1822 for away below and away beyond, 1823 for away off, 1833 for away back, 1847 for away above, 1858 for away ahead, 1898 for away deep and away high, etc. These different members of the away + ADV paradigm of constructions are mirrored by a closely equivalent way + ADV paradigm:

(24) “And while he’s about it,” said Alice, “you might ask him to make a little list of some of the new music. I’ve got way behind the times, being without a piano so long. Tell him not any VERY difficult pieces, you know.” (1896)

(25) Then we stand watches, now in the cavernous fire rooms or alongside the glittering, slithering engines’ way below the water line, now on the high and lofty bridge under the contemplative stars. (1916)

(26) The archaeologist to-day summons to his aid the science of language, studies into the origin of civilization and the comparison of the different races of men, and derives from each and all of these concurrent testimony as to a vast, shadowy, and profound antiquity for man, one stretching way beyond the dawn of history, far into the very night of time. (1885)

(27) “Well,” ses Mary, “thar’s a eend to my jurney to the north. I couldn’t think of gwine a step without Prissy to take care of the child; and spose I was to git sick, too, way off’ mong strangers – what would I do without Prissy?” (1848)

(28) -- But, O dear, I can’t find out half so much about it now, here in this great city of Portland, where all the Governors live, as I could six months ago among the bear traps and log houses in our town, way back in the woods. (1833)

(29) And Cousin Kate, excuse the expression, but your reception to them is way above par. (1879)

(30) “She can’t come. I know who the woman is. They tried to get her when Squire Payne’s sister died last week. Aunt Sylvy told me about it. She was engaged ’way ahead.” (1894)

(31) Last night I woke up and I heard the children crying. They were all asleep but I heard’ em crying, way deep in my head – Jimmy and Walter and Mary, all of ’em crying inside my head. (1928)

(32) “Bes’ one ev’ I hear, he clim up an’ down same as a circus man. […] He clim way high up, an’ holler:’ Goin’ to heavum, goin’ to heavum, goin’ to heavum NOW. Hallelujah, praise my Lawd!’ […]” (1914)

24Crucially, all these first instances of the way + ADV paradigm members appear later than their counterparts from the away + ADV paradigm, so that they remain consistent with an away > way reduction. They are, furthermore, quite often found in a colloquial context ((24), (27), (30), (32) are dialogue excerpts, (29) and (31) are from theater plays), one being explicitly indicated to be a phonetic attrition of away (30). Last, the earliest instance of this way + ADV paradigm (27) shows marked Southern US English features (e.g. the use of ‘gwine’), just as the way too peevish occurrence was also from a markedly Southern dialect. These two occurrences therefore point to a Southern American origin of the phonetic reduction.

25This comforts away too as a major source of the way too construction. This conclusion is strengthened by the fact that no away too instances are found past 1945 and the emergence of the way too construction.

26A few caveats must however be highlighted. First, away too much and away too many are not attested in the COHA, while way too much and way too many are two of the most frequent types of the way too construction. Although way too many does not appear in the COHA corpus before 1980, possibly indicative of a later development of the construction, way too much is attested as early as 1951 (albeit rather infrequent until the 1980s). Moreover, the far and away + more/much occurrences are indicative that the away + ADV was in principle compatible with more and much, although this might have been problematic in most prosodic contexts. Indeed, much is especially found after verb phrases, and away productively combines with an open-ended set of verbs to alter their meaning. Therefore, the potential ‘intensifier’ reading, in these contexts, is always superseded by a more standard ‘verb modifier’ reading, as in the following example, where the [give away] + [too much information] parsing is far more likely than a [give] + [away too much information] one:

(33) But slowly he felt he had given away too much information about certain things, and not enough about others, and this bothered him. (1928)

27The phonetic reduction away > way in the away + ADV construction might therefore have favored the use of this new way + ADV construction in contexts in which the intensifier reading was bound to be suppressed before. This might for instance explain why this phonetical attrition is forced even in registers that are not colloquial, as featured in (13).

28Second, away too is different from the other members of the away + ADV construction: it appears later on, and too, unlike all the other members of the construction, is not a spatial adverbial. Furthermore, it introduces a further level of schematicity (away too + ADJ/ADV) that is absent from all the other members of the paradigm. This late appearance, however, might well be coincidental, given the scarcity of the corpus data and the marked colloquial nature of the construction. Besides, way too, as a contraction of away too, does not appear later than the others way + ADV types.

29Third, away too remained rather infrequent and marked throughout its period of use, to the point that one can wonder whether it has ever been entrenched at all. As such, it does not seem to offer a pool of occurrences reliable enough to give birth to another construction. This reserve can be nonetheless nuanced: again, it is hard to assess the extent of the construction’s usage, given its mostly colloquial nature. Moreover, the way too construction followed a frequency rise and entrenchment process of its own, as I will show in Section 2, so that there is no need for its source to have been itself as much entrenched and rooted in the English language use.

1.3.2. Other possible sources of the way too construction

30I shall now briefly discuss two additional sources which might have favored the emergence of the way too construction as such. Indeed, even though the main hypothesis of an away origin seems well supported, the processes which give rise to a constructionalization are often highly complex, and multiple sources might play a role in this emergence (Van der Welde et al. [2015]).

31The first of these other two sources dates from as early as the seventeenth century, and is still in use nowadays: the every way + ADJ/ADV construction. The earliest occurrence I have found of it is from Google Books, in the translation of a French proverb in the French-English dictionary of Randle Cotgrave:

(34) Il luy fit tenir la mule. He overruled, overcrowed, over mastered him; he was every way too good for him; also, he made him dance attendance, or stay long, for him. (1611)

32This construction is also found with much, for instance in The Countess of Pembroke’s Arcadia by Sir Philip Sidney (35; I can’t explain the duplication of too which is present in several different printings of the book), or with well (36) in Of the Thundering Legion by William Whiston:

(35) But too far I find my passion, yet honest passion hath guided mee; the caus is every way too too much unanswearable. (1655)

(36) This Objection, therefore, is not so much level’d at this particular History of Josephus, as at the intire History of the Jewish Nation : Which yet is every way too well supported to be at all shaken by such a negative Argument, as the bare Omission of one History, by a few later Heathen Historians, can amount to. (1726)

33This every way too construct was still found in the nineteenth century (3 occurrences in the COHA up to 1866) but it seems to have disappeared before the rise of the way too construction. Therefore, the every way construction is unlikely to have exerted any influence on the constructionalization of way too. What is more, the every way construction is actually more versatile than the later way too one and can accommodate not only too, but a wide variety of adjectives, mostly meliorative (every way desirable, suitable, favorable, admirable, competent, satisfactory, similar, worthy, probable, respectable, appropriate, advisable, agreeable, preferable, unexceptionable...), so that every way too is only a specific and not exceptionally frequent type of this intensifier construction. It could be used in comparative contexts as well (every way more, every was as good as you are, every way better, etc.). The last occurrence I could find of this construction in the COHA is dated from 1957 (37), but most occurrences are dated from the nineteenth century and the construction had already mostly disappeared by then. The construction only survives in a variant form, the more explicit in every way construction.

(37) his results would be more uniform and every way more satisfactory, were he to make the king of his haras a direct descendant of a high-type Arabian stallion (1957)

34Besides, the array of adjectives that could easily combine with this construction is indicative of mostly emphatic, subjective judgements, which is at odds with the semantic orientation of the later way too construction. Therefore, there are three main reasons that rule out an every way too > way too derivation: i) there is a chronological gap between the disappearance of the former and the emergence of the latter; ii) the way too construction shows an applicability restriction with too, while the every way construction does not; iii) the two constructions are semantically dissimilar.

35Nonetheless, the every way construction might have played a more indirect role in the emergence of way too. Indeed, it might have contributed to infuse way with a valence towards an intensifier use. The survival of this every way construction in its in every way form might also have contributed towards the development of a way (without too) variant of the way too construction.

36The second possible source is based on the idea of reanalysis in specific ‘switching’ contexts (Heine [2002], Diewald [2002], [2006]). Indeed, the verb give way (to NP), attested since the seventeenth century at least, with a meaning close to ‘yield’, displays a frequency peak roughly over the whole nineteenth century, that is, by the time when the away > way attrition was taking place. This verb could be intensified with too much (38), inducing a processing gap between the give way verb and the to-complement, therefore favoring, to some extent, an interference with an alternative reading, in which way too much would be the direct object of give. Furthermore, the addition of too much could make the indirect object facultative. Although this is not attested in the COHA, there are some rare attestations of it in Google Books (39, Harper’s Bazaar magazine, Volume 100). In this context, the give way meaning is obscured due to the absence of a definite complement, making a transitive reading more accessible. Furthermore, pronouns used in place of the indirect complement could split the give way phrase, allowing for interferences with a ditransitive reading (40):

(38) I see plainly, Harry, that you have some scruples, and I caution you against giving way too much to them. (1846)

(39) We can apply the brakes a bit, try not to give way too much, hide a few things maybe, but it doesn’t make any difference either way to what really matters. (1967)

(40) Reason should confirm Our hearts’ emotions, ere we give them way. (1883)

37These interferences could have combined in a hypothetical give + Pronoun + way too much context, which is not attested, except in a clear way too construction reading (41, from Google Books, Automatic Indexing: A State-of-the-Art Report, National Bureau of Standards Monograph 91):

(41) Carlson raises the further critical questions of overassignment and false drops and suggests that: “a simple machine processing of titles would give us way too much or practically nothing.” (1965)

38There are several problems with this scenario, besides its lack of empirical support. First, give way is usually found in unambiguous semantic contexts. Even though the syntactic context might induce a reanalysis, that is, interferences with alternative readings which would parse way too much as a unit, separated from give, the semantic context is most often more than enough to re-establish the usual give way reading. Moreover, if the way too construction is not yet entrenched, it seems difficult to assign a specific meaning to way too much as a unit.

39It is remarkable, nonetheless, that in the Hansard corpus, there is a frequency peak of give way too much starting from the 1940s to the 1970s, which coincides chronologically with the time during which the way too construction became more frequent. Furthermore, some of them, unlike the earlier occurrences of give way too much featured in the COHA, do show some level of semantic ambiguity (42):

(42) Gentleman suggested, namely that if you provisionally give way too much, you may find it extremely difficult to get it back again when it is adjusted (1939)

40However, give way too much remains extremely formulaic in these parliamentary debates, so the probability that they could have led to a reanalysis is weak. Also, and much more critically, we have seen that the way too construction had most likely an American English origin, while the frequency peak of give way too much appears specific to British English. Therefore, although it cannot entirely be ruled out that this source may have played any influence in the development of the way too construction, there is no conclusive empirical evidence in favor of this hypothesis.

2. Frequency rise of the construction

41We now turn to a detailed, quantitative investigation of the frequency rise of the construction. The request used in the COHA interface was way too *, and gave 601 hits. I cleaned up the occurrences manually, keeping those in the ’way form (2 of them in total). In the end, there are 393 occurrences of the way too construction, spread over 112 types, from 1897 to 2009, the last year covered by the COHA corpus.

2.1. Token frequency rise

42Language changes, and especially constructionalizations, are most often associated with an S-curve pattern of frequency rise (Kroch [1989], Ghanbarnejad et al. [2014], Feltgen et al. [2017]). More precisely, this pattern is associated with the token frequency of the construction (the raw count of occurrences per decade, usually scaled by the number of words in the corpus for the corresponding decade), even though historically, it has first been propounded to account for the number of adopters of the change (Osgood & Sebeok [1954], Weinreich et al. [1968]). Although the S-curve pattern can be associated with a variety of mathematical forms (mostly, the Gaussian cumulative distribution function and the sigmoid), the specific choice of form matters little (they are empirically mostly indistinguishable), and the sigmoid is often favored for its convenience. The token frequency f is therefore modelled according to the following function:

where fmin is the starting value of the sigmoid (usually 0), fmax the frequency reached at the end of the rise, and α and t0 two free parameters, respectively accounting for the steepness of the frequency rise and the time at which the curve reaches its mid-point.

43Traditionally, the logit transform of the frequency data has been used to extract the S-curve parameters (Kroch [1989], Feltgen et al. [2017]). This approach is straightforward in that the logit transform of a sigmoid is a linear function, therefore easy to fit to recover the α and t0 parameters. However, this approach supposes to choose the two endpoints of the S-curve (corresponding to fmin and fmax), which is problematic in several ways: they are often chosen arbitrarily (although Feltgen et al. [2017] offered a way to automatize this selection), the parameter fit is over-sensitive to this choice, and finally, if the process is still ongoing by the end of the time period covered by the corpus, there might be no proper end point to rely on. This is typically the case with the way too construction.

44Therefore, I chose to keep the two end points as free parameters and to rely on a least squares fit that minimizes the sum of squared errors (SSE), using the lsqcurvefit function in Matlab. However, the range of the data that should be fit still needs to be determined: taking into account all data points is potentially problematic, especially because the S-curve pattern can be preceded by a long latency period during which the frequency is non-zero but mostly stagnant (Feltgen et al. [2017]). To select the most relevant range, I tried all possible starting points and performed the fit for each resulting data set, keeping in the end the fit with the smallest SSE, scaled by the number of data points. This procedure selected the 1950-1959 decade as the best starting point, which seems appropriate. Figure 1 shows the resulting sigmoid fit of the data on the 1950-2009 period:

Figure 1. Token frequency evolution of the way too construction over the 1950-2009 period, from the COHA corpus

Figure 1. Token frequency evolution of the way too construction over the 1950-2009 period, from the COHA corpus

45The corresponding slope α is then equal to 2.01 and t0 is equal to 5.1, which means that the mid-point of the process is reached at the beginning of the 1990-1999 decade. Furthermore, the fit predicts that the S-curve will plateau at 9.0 tokens per million words, so that the process has nearly reached its peak by the end of the time period covered by the COHA (7.7 hits per million words). This is consistent with the picture that emerges out of a simple Google Books Ngram Viewer query for the time period 1950-2019 (Figure 2), in which it is apparent that the process has reached its end by the very last decade. Unfortunately, a quantitative comparison of the raw frequency is compromised by the uncontrolled nature of the Google Books data. Finally, these predictions are robust with respect to a different choice of data range (i.e. encompassing earlier data points while fitting the sigmoid).

Figure 2. Token frequency evolution of the way too construction over the 1950-2019 period from a Google Nooks Ngram Viewer query (with a smoothing of 3)

Figure 2. Token frequency evolution of the way too construction over the 1950-2019 period from a Google Nooks Ngram Viewer query (with a smoothing of 3)

46To furthermore assess the reliability of the S-curve pattern, I performed the same analysis on the data from the COCA corpus, which covers three decades (1990-2019). Since three data points is insufficient to fit the sigmoid function, I partitioned the data into intervals of 5 years (from 1990-1994 to 2015-2019). This partition is warranted by the relatively higher token frequency of way too in this period. As evidenced on Figure 3, the sigmoid fit captures well the overall behavior of the data and the plateau is indeed reached by the end of the sampled period. The associated slope α and inflexion time t0 are respectively equal to 1.79 and 0.6, the latter corresponding to the middle of the 1980-1989 decade.

Figure 3. Token frequency evolution of the way too construction over the 1990-2019 period, from the COCA corpus

Figure 3. Token frequency evolution of the way too construction over the 1990-2019 period, from the COCA corpus

47Since the COHA and COCA corpora differ in their design (the former being chiefly composed of literary texts, the latter also encompassing oral productions), a straightforward comparison of the token frequencies is not meaningful; therefore, the α, fmin and fmax parameters cannot be readily compared. However, the period on which the process occurs is expected to be the same for both corpora. To see whether the two fits match in this regard provides thus a good test of the robustness of the pattern: it would indicate that the S-curve has a reliable predictive power, both backward or forward in time. To perform such a match, I rescaled on Figure 4 the two curves so that they both go from 0 to 1, setting aside the frequency differences that arise from the specific nature of the two corpora, and extrapolated on the whole 1940-2029 range, beyond the range over which they have been fitted. Even though the agreement is not perfect, the COCA-based fit remarkably predicts that the frequency rise must have taken off in the 1960s-1970s, while the COHA-based fit successfully predicts that the frequency rise will come at a term by the 2010-2019 decade, which is empirically confirmed in the data from the COCA.

Figure 4. Rescaled S-curve patterns of the process, as extrapolated from a sigmoidal fit of both the COHA (straight line with diamonds) and COCA data (dashed line). The vertical lines show the data range over which each curve has been fitted

Figure 4. Rescaled S-curve patterns of the process, as extrapolated from a sigmoidal fit of both the COHA (straight line with diamonds) and COCA data (dashed line). The vertical lines show the data range over which each curve has been fitted

48The S-curve pattern suffers nonetheless from a limitation: while it predicts a plateau at the end of a frequency rise, a small decline can be observed on both the COCA data and the Google Ngram query. Such a non-monotonic behavior of frequency is beyond the scope of the S-curve and would require a different pattern to be captured.

2.2. Type frequency rise

49As way too is a schematic construction, it can host different arguments (much, long, out of control, prehistoric, etc.). An important aspect of the development of the construction is therefore to consider over time the number of such arguments (Perek [2016], Perek [2018]), a quantity which is known as the ‘type frequency’ of the construction (Bybee & Thompson [1997]). While the rise in token frequency of a constructionalizing construction is well documented, little is known regarding the pattern followed by the type frequency. I therefore plot on Figure 5 the evolution of type frequency over time:

Figure 5. Type frequency evolution of the way too construction over the 1920-2009 period

Figure 5. Type frequency evolution of the way too construction over the 1920-2009 period

50Interestingly, the pattern of type frequency rise and that of the token frequency rise look closely alike. Computing the Pearson correlation coefficient between type frequency and the scaled token frequency displayed on Figure 1, a coefficient of = 0.996 (p = 2e-20) is found, which is strikingly high.

51There is, however, a fundamental issue with type frequency: type frequency is expected to scale with the number of occurrences, in a way which is known not to be of the proportional kind. There already are methods to account for this scaling: to overcome the varying size of the corpus across decades, Perek [2018] creates random samples of the sub-corpora of each decade, of a size equal to the smallest of such sub-corpora, computes the type frequency for each of these random samples, and averages across them to obtain a scaled type frequency for each decade. However, the issue goes beyond that. Indeed, this makes the implicit assumption that, for equal sub-corpus size, the type frequency of a construction is readily comparable from one decade to another. Nonetheless, the token frequency itself of the construction can change, and the type frequency will scale with this token frequency. So not only does one have to account for the varying coverage of each decade, but also to account for this type-token frequency relationship.

52To do so, I rely on the Heaps-Herdan law (Herdan [1960]), which states that the vocabulary size V of a given text varies with the size of that text n according to the following relationship:

where K and β are two free parameters that need to be determined. This law is related to the Zipf’s law (Gerlach & Altmann [2013], Font-Clos et al. [2013]). Anticipating the next section, the frequency distribution over the different types of the way too construction is Zipfian; therefore, one could in principle apply the Heaps-Herdan law to it. To calibrate the two parameters, I focused on the data of the last decade, primarily because it is the decade associated with the highest number of hits (228 occurrences). Then, for each value of n between 1 and 228, I drew n occurrences randomly out of the 228, and computed the corresponding type frequency V over this random sample. This V value should therefore be comprised between 1 and 84, the total number of types found within these 228 occurrences. 30 of these random drawings were performed for each n value, and the type frequency was averaged over these 30 drawings. That way, I was able to compute the Heaps-Herdan profile of the construction for this particular decade, and extract the two parameters with a simple linear fit of log V vs. log n. The empirically constructed profile and the fit are both displayed in Figure 6:

Figure 6. Empirically retrieved Heaps-Herdan law for the way too construction over the 2000-2009 decade

Figure 6. Empirically retrieved Heaps-Herdan law for the way too construction over the 2000-2009 decade

53Then, I compared the type frequency of each decade to the type frequency which would be predicted with our newly calibrated Heaps-Herdan law, based on the number n of occurrences of this decade (Figure 5). The agreement between the two is striking –the Pearson correlation coefficient r is equal to 0.9998 (p = 7e-33). This is impressive since the number of occurrences does not contain any information in principle on the number of types, and yet the latter can be predicted from the former with a high degree of precision (ironically, the biggest deviation comes from the datapoint with which the law has been calibrated; this actually comes from the Heaps-Herdan fit itself, which does not capture accurately the high n part of the curve). Therefore, we can consider that the type frequency pattern is derived from the token frequency pattern of the constructionalizing construction.

3. Emergence of a Zipf-like pattern in the frequency distribution

54Schematic constructions like way too, which can host a variety of arguments, usually show some preference for some specific arguments, a preference that can evolve with time as the semantic scope of the construction expands (Perek [2016]). This preference is reflected through the wide differences in token frequencies of the different types: way too long is for instance eight times more frequent than way too good in the COCA corpus, despite good having a twice higher frequency overall.

55This distribution of frequencies over the different types is known to follow a precise pattern, that of the Zipf’s law (Ellis et al. [2014]), which also holds at the scale of the whole language. According to the Zipf’s law, if all items of a language are ranked according to their frequency, then their frequency f should be a function of that rank r:

with α a free parameter. Equivalently, the Zipf’s law amounts to a linear relationship between the logarithm of these two quantities.

56The same law (with a probably different value for the α parameter) can be retrieved within the free slot of a schematic construction. However, little is known regarding the diachronic emergence and stability of this law. Although a complete study of these questions is beyond the scope of this paper, the specific case of way too will shed precious light on them.

3.1. Zipf-like distribution of the frequencies of the types

57The way too construction is associated with a total of 10,074 hits in the COCA spread among 941 types (for the way too * query). Although I did not clean up manually each occurrence due to their large number, I removed the spurious types (993 initially). The ten most frequent types (accounting for 56% of the total token frequency) are much, many, long, early, far, high, late, fast, big, hard. Some of these arguments are reminiscent of the spatial origin of the construction (long, far, high), and the presence of time-related terms (long as well, early, late) is unsurprising given the common space > time metaphoric extension (Heine et al. [1991]).

58To fit the law (Figure 7), I removed all the hapax legomena; the obtained for the linear fit is equal to 0.989. The α coefficient is equal to 1.24, which is significantly higher than the usual coefficient found in English texts (Gelbukh & Sidorov [2001]); however, any comment on this observation should be avoided for reasons detailed below. In conclusion, the token frequency distribution over the different types clearly follows a Zipf’s law.

59This latter analysis is reminiscent, but not equivalent, to the analysis of a construction in terms of collostructions (Stefanowitsch & Gries [2003], Hilpert [2013b]). According to collostructional analyses, a schematic construction shows distinct preferences for some types, which collocate more frequently than others with the construction, so that the collocation is significantly greater than what one could have expected from the overall frequency of the type in the corpus alone. For instance, way too is frequently associated with expensive, even though expensive is not particularly frequent; on the contrary, full, despite being approximately four times more frequent in the COCA corpus than expensive, is also about seven times less frequently associated with the way too construction than expensive. The Zipf characterization of the construction is different in that it does not tell which individual associations are specifically significant and characteristic of the construction; it provides, however, a global view of the profile of preferences that fundamentally specifies the construction’s use.

Figure 7. Zipf’s law associated with the way too construction, based on data retrieved from the COCA corpus

Figure 7. Zipf’s law associated with the way too construction, based on data retrieved from the COCA corpus

3.2. Diachronic development of the Zipf-like distribution

60Here I shall consider three questions: is the Zipf’s law followed for each decade? Does the α parameter evolve with time? Finally, are the ranks conserved with time?

3.2.1. Does the Zipf’s law hold for each decade of the constructionalization process?

61Unfortunately, an accurate answer to these questions is made difficult due to the scarcity of the available data: in the COHA corpus, the type frequency for each decade from 1930-1939 to 2000-2009 is respectively equal to 1, 2, 4, 4, 9, 17, 51, 84. If furthermore we remove the hapax legomena, then there is only data for the four last decades, with type frequencies (hapax legomena remove) respectively equal to 1, 6, 19 and 32. Therefore, the Zipf’s law can only be meaningfully fitted for the last three decades only. The results of these fits are summarized on Table 1. For all these three decades, the token frequency distribution over types behaves in a Zipf-like way, as attested by the high values of the linear fit.

Table 1. Zipf’s law associated with the way too construction for the three decades 1980-1989, 1990-1999 and 2000-2009, based on data retrieved from the COHA corpus


Type frequency

Type frequency (no hapax)

Zipf’s α coefficient
















62Since the way too construction starts to be consistently attested from the 1950s, it is impossible to know on the sole basis of the COHA data whether the emergence of the Zipf’s law is concomitant to the frequency take-off of the construction. However, as soon as the number of attested types that are not hapax legomena is high enough to allow for a rank analysis, here the 1980-1989 decade, the frequency distribution over types is consistent with a Zipf’s law. Given that the frequency is only slowly rising by this time, this is clear empirical evidence that the Zipfian organization emerges very early during the constructionalization process.

63Note that the Zipf’s α coefficient found for the last decade here is widely different from the Zipf’s coefficient previously computed over the COCA data (0.84 vs. 1.24), and is significantly below the Zipf’s α coefficient found for entire texts. There are two possible explanations here. The first one is that the COCA data relies on considerably more colloquial sources, while the COHA data chiefly come from literary sources. Therefore, different registers might be associated with different α coefficients. The second explanation is that, as we shall see below, the empirical determination of the α coefficient is widely dependent on the sample size. Therefore, any direct comparison between these coefficients for different sample sizes is compromised.

3.2.2. Does the way too construction obey the same laws throughout the whole process?

64Interestingly, the Zipf’s coefficient α resulting from the fit is increasing over time. Even though three data points is insufficient to empirically support this increase with a satisfying degree of certainty, this trend would indicate that the construction might become, at once, both applicable to a greater number of types, and tighter at the same time, as the most attested types become even more attested compared to the least attested types. Indeed, the higher the coefficient, the steeper the curve, and the more uneven the distribution of token frequencies among the different types.

65Now, recall that the data from all the decades affected by the constructionalization process seemed to conform to the Heaps-Herdan law derived from the data of the last decade. Since the Heaps-Herdan law and the Zipf’s law are closely related to each other, this would imply that the Zipf’s law is the same for all these decades. There is therefore a contradiction between these two results that needs to be resolved.

66First, we can consider whether a partial sampling of a data set changes the Zipf’s law. To do so, I randomly sampled 100 times the data associated with the last decade, and computed the average Zipf’s coefficient, averaged over the resulting data sets. This was done for each sample size between 30 and 228 (the size of the data sample associated to the last decade). The results, displayed on Figure 8, show the opposite of what was expected to explain the results of Table 1: the Zipf’s law coefficient increases with low sample sizes, so that, if the same Zipf’s law was identically followed in each decade, the α coefficient of the last decade would be the lowest. Furthermore, this wide dependence of the α coefficient on the sample size shows that, far from the large sample size asymptotical regime where the Zipf’s law and the Heaps-Herdan law are related, the relationship between these two laws becomes irrelevant, since the Heaps-Herdan is, by its very nature, independent of the sample size (although see below for a more nuanced stance on this question).

Figure 8. Zipf’s law coefficient computed from samples of the data from the 2000-2009 decade of the COHA

Figure 8. Zipf’s law coefficient computed from samples of the data from the 2000-2009 decade of the COHA

67The contradiction is thus somehow dissolved by the lack of equivalence between the Zipf’s law and the Heaps-Herdan law for small sample sizes. Yet, the observed variations of the α coefficient appear to be a true phenomenon, and not a side effect of smaller sizes of the data sets of the earlier decades: the Zipf’s law does change from decade to decade as the constructionalization process unfolds. The same sort of variations could be expected for the Heaps-Herdan law, so that predicting the type frequency of all decades based on the law calibrated for a single decade is far from granted.

68I therefore computed the Heap-Herdan law’s coefficient for each relevant decade (1980-1989, 1990-1999, 2000-2009, as the others do not show a sufficient number of types). Furthermore, to obtain a reference point so as to assess the relevant magnitude of the variation, I took two independent batches of 200 samples of the occurrences of the last decade, of a size respectively matching the token frequency associated with the other two earlier decades (26 and 116). I then computed the β coefficient over each sample the way I would have done for the actual data of these decades. This procedure allows to compute the average β coefficient for these sample sizes, based on the data of the last decade alone, and to extract 95% confidence intervals. The results are shown on Table 2. Note that the β coefficient extracted for the last decade and used for the fit of Figure 4 was equal to 0.76.

Table 2. Heap-Herdan law β coefficient associated with the way too construction for the two decades 1980-1989 and 1990-1999, and estimates of comparable sample size based on the data from 2000-2009, all retrieved from the COHA corpus


Token frequency

Type frequency


β (estimate based on the 2000-2009 data)







[0.77; 0.95]






[0.77; 0.84]

69Even though the β coefficient computed for the two previous decades deviates from the value found for the decade 2000-2009, it deviates in a way which is what would be expected if we were to compute this coefficient from a smaller subset of the same data, of a size that corresponds to the token frequencies of these two previous decades. Therefore, the empirical evidence is consistent with the assumption that the same Heaps-Herdan law holds for the entirety of the process, in contrast to the Zipf’s law which shows significant variation from one decade to another.

3.2.3. Is the rank of the different types conserved throughout the process?

70The 8 highest ranked items for each decade are displayed in Table 3. At first glance, it appears that some decisive semantic shifts are going on during the first decades of the first process, during which the frequency of the construction remains skulking. During the first three decades the construction is mostly associated with spatial terms (especially deep, present in the three first decades, and high, often used metaphorically), even though we know that the first two instances are not related to anything spatial (peevish and keen), so that the construction already applies to character traits (bad also falls in this category). Temporal terms appear in the decade 1970-1979: late, long (all occurrences are associated with a temporal meaning rather than a spatial one), early, but some spatial terms remain ranked highly (low, far). The two adverbs much and many become more prevalent from the 1980s (many is not attested before and much is only attested once earlier), and end up dominating the paradigm.

Table 3. Eight most frequent types of the way too construction, for each decade starting with 1940-1949. Stars indicate that the number of types for this decade is below eight









































































71For a more quantitative comparison, I compared the rankings of the types of each decade (from 1970-1979 to 1990-1999) to the rankings of these types in the later decades (up to 2000-2009). The Spearman correlation coefficients between these different rankings are displayed in Table 4. Here I chose the Spearman correlation instead of the Pearson correlation because the former is more suited to compare rankings. They show that the ranking of each decade correlates significantly with the ranking of the last decade. Correlations between the other decades is not significant (except for 1970-1979 and 1990-1999), which can be due to the scarcity of the data for these decades (comparing them amounts to compare two samples of rather small size). That they all correlate significantly to the last decade indicate that the ranking of the Zipf’s law has not been significantly altered.

Table 4. Spearman correlation coefficients between the rankings of different decades. (*): p < 0.05; (**): p < 0.01; (***): p < 0.001







0.85 (**)

0.88 (**)




0.85 (***)




0.81 (***)

3.3. Robustness of the Zipf-like distribution across varieties of English

72The last question is to know whether this Zipf’s law organization found in American English is also found in other varieties of English. Indeed, I have already argued that the way too construction most likely originates from the United States, so the construction has spread from there to other English-speaking countries. However, the way too construction could be adopted as a general schema, not as a complex frequency organization governed by a specific ranking between its different types. It is therefore worthwhile to investigate the robustness of this ranking over the different varieties of English.

73To do so, I rely on data from the GloWbE corpus, a corpus compiled over Internet text productions of various sorts (blogs mostly, but newspapers as well) from the years 2012-2013. The specificity of this corpus is that it is divided into 20 sub-corpora corresponding to as many varieties of English, and the interface performs all queries over the different sub-corpora separately to allow for direct comparisons. The following varieties of English are represented: United States, Canada, Great Britain, Ireland, Australia, New Zealand, India, Sri Lanka, Pakistan, Bangladesh, Singapore, Malaysia, Philippines, Hong Kong, South Africa, Nigeria, Ghana, Kenya, Tanzania, Jamaica.

74Variety-to-variety comparison has been performed in the same way as the decade-to-decade comparison described above: I took the United States variety of English as a reference point, identified the new ranks of all the types across the different varieties of English, and finally computed the Spearman correlation coefficient between the United States ranking and each of the other rankings. However, I removed all types of token frequency below 5 in the United States sub-corpus, retaining therefore a total of 199 types. Indeed, below 5, a lot of types would share the same token frequency, so that their relative ranking becomes irrelevant.

75The Spearman correlation coefficient is comprised between 0.76 (Great Britain) and 0.92 (Ghana). Each of them is highly significant (the highest p-value being equal to 6e-10). Therefore, all varieties share comparable frequency rankings over the types of the way too construction. This ranking is thus a definitory trait of the construction that is transmitted with the use of the construction.


76The way too intensifier construction starts to appear sporadically during the first half of the twentieth century and emerges fully in the 1970s, increasing in frequency until the Present Day, where it probably reached its culmination point. Its origins are rooted in an away too construction which never really took off, probably hindered by interfering parsing that would have applied the away part to the verb, since it is indeed a common verbal semantic modifier. Therefore, the phonetic reduction of away to way, of which there are several attestations, would have allowed the intensifier reading to emerge, and the construction use to become more widespread. It cannot be ruled out, however, that other sources might have played a role in this entrenchment, and the overall versatility of way in English makes this additional hypothesis rather likely. In any way, the way element of the way construction is perceived by the Present-Day speakers as identical to the lexical way, as attested by instances such as a long way too large, even though it might not have been felt as such by speakers in the late nineteenth century / early twentieth century which were innovating with this construction. For this reason, it seems improper to consider the entrenchment of way too as a phenomenon of grammaticalization; it is, however, by all accounts, a clear case of constructionalization.

77The way too rise in token frequency follows a pattern that is consistent with an S-curve. Moreover, the type frequency follows a closely related pattern. As it turns out, the evolution of type frequency can be accurately predicted from the token frequency through the Heaps-Herdan law, calibrated on the last decade covered by the corpus (chosen as a reference point because it is the decade with the largest number of hits of the way too construction). This shows a tight relationship over time between the two quantities which had not been quantitatively evidenced before.

78This raises a number of questions. For instance, is the token frequency rise but the result of an S-curve diffusion over the types, or does the token frequency rise favor the recruitment of new types, as the construction becomes more widespread? First, although the data was too scarce to conduct such an analysis, it is very likely that the individual types also follow an S-curve in their own token frequencies, ruling out the fact that the S-curve is but a diffusion over types or W-diffusion (Ogura & Wang [1996]). Second, I think that either reading is a misleading conception of the ongoing constructionalization process. What is really emerging is an organization that both gives shape and defines the arising construction. Without such an organization, there could be neither type nor token frequencies rise: the token frequency rises precisely because the construction starts to make sense as a schema, which encompasses a variety of types all with their own and gradient collostructional affinities to the construction, and yet in such a way that the whole remains cohesive. That the earliest decades of the rise already obey the Heaps-Herdan law seems indeed indicative that the organization of the construction is concomitant to the very taking off of the frequency itself. Therefore, it would make little sense to try to disentangle the token and the type frequencies.

79This organizational view of the construction is supported by the empirical findings regarding the Zipf-like organization of the paradigm. Indeed, the token frequencies of the types of the way too construction collectively obey a Zipf’s law. This is both true in synchrony, as attested from the COCA corpus, and in diachrony: the Zipf’s law is attested for each relevant decade, and the ranking between types that it implies is mostly conserved throughout the constructionalization process. This Zipf’s law, however, is not static, and becomes steeper as time goes by: that is, the most frequent items become even more frequent with respect to the least frequent ones as the process goes on. Even though the application scope of the construction is widening, the construction becomes paradoxically more specialized overall.

80This study is, of course, not without limitations. First, the scarcity of the data associated with the way construction in the COHA corpus prevents the conclusions that I drew to be firm and definitive. The same analyses should therefore be conducted on other constructions that show a wider attestation over the whole period covered by the COHA corpus. Also, these analyses are, for the most part, innovative, and for the time being, one cannot distinguish which conclusions are specific to the way too construction and which are broadly applicable to the constructionalization of schematic constructions in general. Third, I have left aside the study of far too, which is a neighboring construction with similar use and meaning, and that is five times more frequent than way too in the COHA (yet they have similar frequencies in the COCA). Moreover, the ranking of its types is very similar to that of the way too construction. These two constructions probably interact (e.g., they might tie to different registers, literary for far too and colloquial for way too, so as to co-exist without competing), and a complete account of the way too diachronic development should include such a discussion. Finally, I have shown that the Zipf’s law and the Heaps-Herdan law are sensitive to the sample size in ways which are difficult to predict. In particular, the Zipf’s law coefficient shows a non-monotonic relation of its coefficient with the sample size, such that it remains elusive to identify its ‘correct’ value. Comparisons between decades, and between corpora, are made all the more difficult to interpret.

81This study could also be completed by further analyses, in particular relating to collostructional analyses. Indeed, in a Zipf’s law spirit, I have arranged the different types according to the number of tokens of their collocation with the construction. However, such a ranking could also be performed on the collostructional score of the items. It is nonetheless unlikely that this alternative ranking will be less robust than the Zipf’s ranking – it would only be so if the overall frequencies of these types vary wildly at a global scale. The Zipf’s ranking, however, albeit robust, is not constant. It might well be that the items with a high collostructional score early in the process climb up the ranking over time; in other words, the residual rank variation might be partially explained by the collostructional score of the item. In the same spirit, it would be interesting to see whether an early collocation of a type is predictive of its ranking in the end of the process, that is, whether there is any correlation between the final rank and the time of first attestation within the construction.

82The array of empirical results that I presented in this paper should at least hint at a fact that has been neglected: a schematic construction is not only a form-meaning pair, it is also an ecological niche for its different types with a specific organization of this own. This organization, however abstract, is as much a definitory trait of the construction as its meaning or its specific formal schema. The last section of this paper shows indeed that the ranking of the types is universal across all varieties of English. Therefore, this organization has to be acquired by language learners and becomes cognitively entrenched as part of the construction’s usage, as is attested in psycholinguistic studies (Ellis & Ferreira-Junior [2009], Ellis [2012], Torres-Martínez [2018]). This suggests that a schematic construction is more than a sign-meaning pair: it is a robust system of its own, that might further evolve with time, and that organizes language at its own local and limited level.

Top of page


Barðdal Jóhanna, 2008, Productivity: Evidence from Case and Argument Structure in Icelandic, Amsterdam/Philadelphia: John Benjamins Publishing Company, 121-150.

Blythe Richard A. & Croft William, 2012, “S-curves and the mechanisms of propagation in language change”, Annual Meeting of the Berkeley Linguistics Society, 23(1), Washington D.C.: Linguistics Society of America, 378-388.

Bybee Joan & Thompson Sandra, 1997, “Three frequency effects in syntax”, Language, 88(2), Berkeley: Berkeley Linguistics Society, 269-304.

Cotte Pierre, 2012, « One’s way ou le résultat d’avant », in Le Querler Nicole, Neveu Franck & Roussel Emmanuelle (Eds.), Relations, connexions, dépendances. Hommage au Professeur Claude Guimier, Rennes: Presses Universitaires de Rennes, 25-30.

Croft William, 2000, Explaining Language Change: An Evolutionary Approach, Harlow: Pearson Education Limited.

Danino Charlotte, Wolfsgruber Anne C. & Joffre Marie-Dominique, 2020, « Voilà en diachronie : perception, énonciation et courbe en S. », in Col Gilles, Danino Charlotte & Bikialo Stéphane (Eds.), Polysémie, usages et fonctions de « voilà », Berlin/Boston: Walter de Gruyter, 33-80.

Davies Mark, 2008-, The Corpus of Contemporary American English (COCA): 600 million words, 1990-present, available at

Davies Mark, 2010-, The Corpus of Historical American English (COHA): 400 million words, 1810-2009, available at

Davies Mark, 2011-, Google Books Corpus. (Based on Google Books n-grams), available at

Davies Mark, 2013, Corpus of Global Web-Based English: 1.9 billion words from speakers in 20 countries (GloWbE), available at

De Mulder Walter, 2019, « En voie de : du trajet spatial à l’aspect », Revue Romane. Langue et littérature. International Journal of Romance Languages and Literatures, 54(1), Amsterdam/Philadelphia: John Benjamins Publishing Company, 39-61.

Diewald Gabriele, 2002, “A model for relevant types of contexts in grammaticalization”, in Wischer, Ilse & Diewald Gabriele (Eds.), New Reflections on Grammaticalization, Amsterdam/Philadelphia: John Benjamins Publishing Company, 103-120.

Diewald Gabriele, 2006, “Context types in grammaticalization as constructions”, Constructions 1(9), 1-29.

Ellis Nick & Ferreira-Junior Fernando, 2009, “Construction Learning as a Function of Frequency, Frequency Distribution and Function”, The Modern Language Journal, 93, iii, Wiley-Blackwell, 370-385.

Ellis Nick, 2012, “Formulaic language and second language acquisition: Zipf and the phrasal teddy bear”, Annual review of applied linguistic, 32, Cambridge: Cambridge University Press, 17-44.

Ellis Nick, O’Donnell Matthew & Römer Ute, 2014, “Does Language Zipf Right Along? Investigating Robustness in the Latent Structure of Usage and Acquisition”, in Connor-Linton J. & Amoroso L.W. (Eds.), Measured Language: Quantitative Studies of Acquisition, Assessment, and Variation, Washington D.C.: Georgetown University Press, 33-50.

Fagard Benjamin, 2009, « Vers du français classique au français moderne ou contemporain : l’aboutissement d’une grammaticalisation », Revue Québecoise de Linguistique, 33(1), Montréal: Université du Québec, 5-29.

Fagard Benjamin & Combettes Bernard, 2013, « De en à dans : un simple remplacement ? Une étude diachronique », Langue française, 178(2), Paris: Armand Colin, 93-115.

Feltgen Quentin, Fagard Benjamin & Nadal Jean-Pierre, 2017, “Frequency patterns of semantic change: corpus-based evidence of a near-critical dynamics in language change”, Royal Society open science, 4(11), London: The Royal Society.

Font-Clos Francesc, Boleda Gemma & Corral Àlvaro, 2013, “A scaling law beyond Zipf’s law and its relation to Heaps’ law”, New Journal of Physics, 15, Bristol, Philadelphia: IOP Publishing.

Gelbuck Alexander & Sidorov Grigori, 2001, “Zipf and Heaps laws’ coefficients depend on language”, in Gelbuck Alexander (Ed.), Computational Linguistics and Intelligent Text Processing. CICLing 2001. Lecture Notes in Computer Science, vol 2004, Berlin, Heidelberg: Springer.

Gerlach Martin & Altmann Eduardo G., 2013, “Stochastic model for the vocabulary growth in natural languages”, Physical Review X, 3(2), Ridge: American Physical Society.

Ghanbarnejad Fakhteh, Gerlach Martin, Miotto José M. & Altmann Eduardo G., 2014, “Extracting information from S-curves of language change”, Journal of The Royal Society Interface, 11(101), London: The Royal Society.

Heine Bernd, Claudi Ulrike & Hünnemeyer Friederike, 1991, Grammaticalization: A Conceptual Framework, Chicago: The University of Chicago Press.

Heine Bernd, 2002, “On the role of context in grammaticalization”, in Wischer Ilse & Diewald Gabriele (Eds.), New Reflections on Grammaticalization, Amsterdam/Philadelphia: John Benjamins Publishing Company, 83-101.

Herdan Gustav, 1960, Type-Token Mathematics, The Hague: Mouton & Cie.

Hilpert Martin, 2013a, Constructional Change in English: Developments in allomorphy, word formation, and syntax, Cambridge: Cambridge University Press.

Hilpert Martin, 2013b, “Collostructional analysis: Measuring associations between constructions and lexical elements”, in Glynn Dylan & Robinson Justyna A. (Eds.), Corpus Methods for Semantics: Quantitative studies in polysemy and synonymy, Amsterdam/Philadelphia: John Benjamins Publishing Company, 391-404.

Hopper Paul & Traugott Elizabeth Closs, 1991, Grammaticalization, Cambridge: Cambridge University Press.

Israel Michael, 1996, “The Way Constructions Grow”, in Goldberg Adele (Ed.), Conceptual Structure, Discourse and Language, Stanford: CSLI, 217-230.

Kroch Anthony, 1989, “Reflexes of grammar in patterns of language change”, Language variation and change, 1(3), Cambridge: Cambridge University Press, 199-244.

Lauwers Peter & Willems Dominique, 2011, “Coercion: Definition and challenges, current approaches, and new trends”, Linguistics, 49(6), Berlin: De Gruyter Mouton, 1219-1235.

Mair Christian, 2013, “Corpus linguistics and grammaticalisation theory: Statistics, frequencies, and beyond”, in Lindquist Hans & Mair Christian (Eds.), Corpus Approaches to Grammaticalisation in English, Amsterdam/Philadelphia: John Benjamins Publishing Company, 121-150.

Nevalainen Terttu, 2000, “Processes of supralocalisation and the rise of Standard ENglish in the early Modern period”, in Bermùdez-Otero Ricardo, Denison David, Hogg Richard M. & McCully C.B. (Eds.), Generative Theory and Corpus Studies: A Dialogue from 10 ICEHL, Berlin: Mouton de Gruyter, 329-372.

Nevalainen Terttu, 2015, “Descriptive adequacy of the S-curve model in diachronic studies of language change”, in Sanchez-Stockhammer Christina (Ed.), Can We Predict Language Change?, Helsinki: Varieng.

Ogura Mieko & Wang William S-Y., 1996, “Snowball effect in lexical diffusion: The Development of –s in the Third Person Singular Present Indicative in English”, in Britton Derek (Ed.), English historical linguistics 1994: papers from the 8th Int. Conf. on English Historical Linguistics (8. ICEHL), Edinburgh, 19–23 September, Amsterdam/Philadelphia: John Benjamins Publishing Company, 119-141.

Osgood Charles E. & Sebeok Thomas A., 1954, Psycholinguistics: a survey of theory and research problems, Baltimore: Waverly Press.

Perek Florent, 2016, “Using distributional semantics to study syntactic productivity in diachrony: A case study”, Linguistics, 54(1), Berlin: De Gruyter Mouton, 149-188.

Perek Florent, 2018, “Recent change in the productivity and schematicity of the way-construction: A distributional semantic analysis”, Corpus Linguistics and Linguistic Theory, 14(1), Berlin: De Gruyter Mouton, 65-97.

Rissanen Matti, 2013, “Grammaticalisation from side to side: On the development of beside(s)”, in Lindquist Hans & Mair Christian (Eds.), Corpus Approaches to Grammaticalisation in English, Amsterdam/Philadelphia: John Benjamins Publishing Company, 151-170.

Rostila Jouni, 2004, “Lexicalization as a way to grammaticalization”, in Karlsson Fred (Ed.), Proceedings of the 20th Scandinavian Conference of Linguistics, Helsinki, January 7-9, Helsinki: University of Helsinki.

Stefanowitsch Anatol & Gries Stephan Th., 2003, “Collostructions: Investigating the interaction of words and constructions”, International journal of corpus linguistics, 8(2), Amsterdam/Philadelphia: John Benjamins Publishing Company, 209-243.

Torres-Martinez Walter, 2018, “Applied Cognitive Construction Grammar: A usage-based approach to the teaching of phrasal verbs (and other constructions)”, European Journal of Applied Linguistics, 6(2), Berlin: De Gruyter Mouton, 279-314.

Traugott Elizabeth Closs & Trousdale Graeme, 2013, Constructionalization and constructional changes, Oxford: Oxford University Press.

Van Der Welde Freek, De Smet Hendrik & Ghesquière Lobke, 2015, “On multiple source constructions in language change”, in Van Der Welde Freek, De Smet Hendrik & Ghesquière Lobke (Eds.), On Multiple Source Constructions in Language Change, Amsterdam/Philadelphia: John Benjamins Publishing Company, 1-18.

Weinreich Uriel, Labov William & Herzog Maivin I., 1968, “On multiple source constructions in language change.” Empirical Foundations for a Theory of Language Change, New York: Columbia University, 95-195.

Top of page

List of illustrations

File image/png, 1.7k
Title Figure 1. Token frequency evolution of the way too construction over the 1950-2009 period, from the COHA corpus
File image/png, 25k
Title Figure 2. Token frequency evolution of the way too construction over the 1950-2019 period from a Google Nooks Ngram Viewer query (with a smoothing of 3)
File image/png, 17k
Title Figure 3. Token frequency evolution of the way too construction over the 1990-2019 period, from the COCA corpus
File image/png, 29k
Title Figure 4. Rescaled S-curve patterns of the process, as extrapolated from a sigmoidal fit of both the COHA (straight line with diamonds) and COCA data (dashed line). The vertical lines show the data range over which each curve has been fitted
File image/png, 39k
Title Figure 5. Type frequency evolution of the way too construction over the 1920-2009 period
File image/png, 30k
File image/png, 839 octets
Title Figure 6. Empirically retrieved Heaps-Herdan law for the way too construction over the 2000-2009 decade
File image/png, 22k
File image/png, 759 octets
Title Figure 7. Zipf’s law associated with the way too construction, based on data retrieved from the COCA corpus
File image/png, 24k
Title Figure 8. Zipf’s law coefficient computed from samples of the data from the 2000-2009 decade of the COHA
File image/png, 21k
Top of page


Electronic reference

Quentin Feltgen, Diachronic Emergence of Zipf-like Patterns in Construction-Specific Frequency Distributions: A Quantitative Study of the Way Too ConstructionLexis [Online], 16 | 2020, Online since 17 December 2020, connection on 07 June 2023. URL:; DOI:

Top of page

About the author

Quentin Feltgen

Institut du Cerveau et de la Moelle épinière – ICM, Sorbonne Universités, UPMC Univ Paris 06, Inserm U1127, CNRS UMR 7225, Paris, France

Top of page



Creative Commons - Attribution-NonCommercial-NoDerivatives 4.0 International - CC BY-NC-ND 4.0

Top of page
Search OpenEdition Search

You will be redirected to OpenEdition Search