I couldn’t help but wonder: do modals and negation attract?

2.1 Scope and suppletion

In the simplest terms, what is described in a negated or modal clause is marked as neither having happened nor happening at the time of speaking, i.e. both independently shift a proposition to irrealis or false (cf. e.g. Givón 1984: 321). This is why Hopper and Thompson (1980) classify both as Transitivity lowering—the action has not affected the patient. There may be a possibility for it to do so in the future—even an obligation—but the actual event happening is not part of the proposition of the clause.

Semantics get more complex where the two are combined. When a modal and a negator co-occur, the negator can either have narrow scope, which only extends over the proposition (in which case the modal scopes over the negator), as illustrated by (5) or it can have wide scope extending over the modal, as in (6) (cf. Klima 1964; De Haan 1997: 12; Palmer 1997: 137; Quirk et al. 1985: 794–796; Radden 2009: 170).

While many other languages use word-order to distinguish between these (cf. e.g. De Haan 1997: 86–109), English does not.^[1] Instead, which reading(s) is/are permitted, depends on the modal and to some degree on modal flavour (i.e. deontic vs. epistemic modality). Thus negation is most likely to have narrow scope when combined with should, might, must and epistemic may, while it tends to have wide scope with can, could and deontic may (cf. Coates 1983; De Haan 1997: 60–67; Francis and Iatridou 2020: 285; Palmer 1997: 136; Quirk et al. 1985: 794–797; Radden 2009: 174). Quirk et al. (1985: 795–797) argue that the distinction between narrow and wide scope “is neutralised” in the case of will and would as both interpretations are semantically equivalent. Coates (1983: 64) assigns should to this ‘scope neutralised’ category, too. De Haan (1997: 60–67) lists alternative—and presumably much rarer—scope interpretations which are possible with some modals. He claims that deontic can not actually permits both narrow and wide scope interpretations of negation while cannot as well as contracted can’t only permit a wide-scope reading of negation (De Haan 1997: 62).

This may prompt the question whether negation rate and scope are linked; i.e. whether modal–negation pairs with more flexible interpretations occur more frequently than ones with restricted scope interpretations. However, scope is unlikely to be a good predictor of negation rate, not least because there are additional complicating factors; namely that not all semantic contexts are equally frequent and that speakers often adopt suppletion strategies; i.e. “the difference between wide scope interpretation of the negation and narrow scope interpretation is expressed by means of different modal elements” (De Haan 1997: 58). If speakers, for instance, wish to express ‘not necessary’ they cannot use a necessity modal, such as must, due to scope restrictions. Instead, they opt for a possibility modal like may (cf. Palmer 1997: 135; see also Radden 2009: 170). Suppletion may explain some of the differences in use of between negated and non-negated modals (cf. e.g. Bybee 2010: 151–164; Leclercq 2022).

Negation rates determined by Mindt (1995; grey bars in Figure 1; based on a corpus consisting primarily of British fiction) suggest, on the one hand, that would and should with their potentially neutralised scope interpretations have higher negation rates than scope-restricted might and must. On the other hand, negation rates of will and shall are not elevated despite neutralised scope and could, with which negation is restricted to narrow scope, has some of the highest negation rates. Moreover, more nuanced data provided by Römer (2004; yellow bars in Figure 1; based on the spoken section of the British National Corpus) shows that the frequency of can’t far exceeds that of cannot, despite the fact that cannot is possibly less restricted in scope than the contracted form.

Figure 1:

Negation rates of different modals as determined by Mindt (1995: 176) and Römer (2004: 189).

Finally, language change also affects scope—some older sources still list combinations of scope and modal flavour which are missing from later lists and, of course, there are differences between British and American English (cf. Tottie 1985). Finally, core modals are not only replaced by other core modals, but also by semi-modals as well as lexical expressions of modality etc. (cf. also Leech 2013). All of these factors make it unlikely that flexibility of scope has an influence on negation rates of modals.

2.2 do-support

Besides semantics, there are a range of syntactic changes which could potentially have influenced modal-negation rates. The first of these is the emergence of do-support, or operator do, between 1500 and 1700. Within this timeframe, English shifted from mostly ‘finite verb + not’, as in (7), to predominantly ‘(finite) operator + not (+ lexical verb)’ in negative declarative sentences, as in (8) (cf. Denison 1993: 451; Ellegård 1953; Jespersen 1917: 10–11; Strang 1970: 151; Visser 1969: §1440–1441).

Ellegård (1953) shows that the rate of do-support in negative declarative sentences (in which no other operator is present) rises from 5 % in 1500 to 75 % in 1700. In negative questions, do spreads even faster. When no other auxiliary is present, these are almost invariably formed with do by 1700 (cf. Ellegård 1953).^[2] Warner (1993: 221–222) speculates that the parallel development of the modals and operator do “strongly suggests that [the two] are interconnected”.

In fact, we have evidence that modals functioning as operators paved the way for operator do. Denison (1993: 467–468) analyses six Paston Letters, written in the late fifteenth and early sixteenth centuries, “when periphrastic do was in infrequent use and had by no means reached its peak, let alone been regulated” (Denison 1993: 467). He finds that the chance of any operator being present is significantly higher in negated than in affirmed sentences and concludes that this “suggests that the essence of the process now labelled as the ‘regulation of do’—namely the tendency for negatives […] to contain an operator—had already begun long before do itself was a statistically important element in the language, possibly even before periphrastic do first arose” (Denison 1993: 467–468). Budts and Petré (2020) corroborate this in an analysis of Early English Books Online (EEBO), covering the timespan 1580–1700. They confirm that “periphrastic do became distributionally more and more similar to the modal auxiliaries (and, possibly, also the reverse)” (Budts and Petré 2020: 346).

Warner (1993: 209) suggests that this reverse influence consisted in an early effect on the emergence of the modals as a class of auxiliaries. Budts (2022) provides tentative suggestions as to what a later influence of do on the modals could have looked like. Her neural network analysis of late sixteenth-century uses of do as well as modals uncovers functional overlap between them. Particularly will shows “extensive overlap with do” (Budts 2022: 357). Her results further suggest that negated do and can in combination with perception and cognition verbs are functionally interchangeable, as “the inability (can not) to perceive something inevitably leads to the perception not taking place (do not)” (Budts 2022: 355). She concludes that, under the influence of other modals, do briefly expressed epistemic modality in sixteenth century affirmative declaratives and further that

the loss of affirmative do seems to suggest that the story of periphrastic do and the modals in the 17th century is one of divergence rather than attraction. Perhaps affirmative do had grown so similar to the modals, especially to will, that the two forms entered in a competition that pushed do to the fringes of the paradigm, allowing it to preserve its auxiliary syntax but forcing it to let go of its modal semantics. (Budts 2022: 359)

The present study allows us to assess whether there are changes in the rate of modal-negation after 1700 which point to a divergence from or further convergence with do.

2.3 Decline in core modals

The second syntactic reorganisation process which may have triggered changes in modal-negation rates is the decline in core modals and the concomitant rise of semi-modals such as need to and want to over the course of the twentieth century (cf. Leech 2003, 2013; Leech et al. 2009). As already pointed out in Section 2.1 above, speakers today have the option to use semi-modals in contexts where core modals would have dominated earlier. Leech (2003) shows that this leads to a shift in LModE. He analyses the frequency of modals in the Brown family of corpora and finds significant decreases in the frequency of core modals. Rates of decline range from 2.7 % for will to 43.7 % for shall. Yet, in British English can and could do not seem to be affected by this shift; they are actually on the rise.

It is however possible that this process has not affected negated and affirmative contexts in the same way. If affirmative uses of modals are lost at a different rate than negated uses, this would have an effect on negation rates. The present analysis investigates whether we see any evidence of this happening.

2.4 Negative contraction

The final diachronic change to consider is the emergence of negative contraction. In the sixteenth century, some modal-negation combinations take on a different form: they are contracted and the vowel in not is elided (cf. e.g. Jespersen 1917: 116; Denison 1993: 309).^[3] For about 300 years, the contractions are much less frequent than the full forms.

Daugs (2021: 26) searches the fiction subcorpus of the Corpus of Historical American English (COHA) for can/will + negation and finds that both undergo the same shift: Early in the nineteenth century, contracted forms make up around 20 % of the data; this rate climbs steadily until it reaches around 80 % by the beginning of the twentieth century and then plateaus there. In more formal genres, the shift occurs later. In TIME magazine, for instance, contractions of can, could, will, would or should + n’t only surpass the full forms by the 1990s (cf. Millar 2009: 211; see also Daugs 2021: 44).

Nakamura (2023) provides evidence which suggests that the spread of the contraction may have been phonologically governed, with the contraction occurring first in forms where the resulting final consponant cluster consists exclusively of /nt/ (i.e. don’t, can’t, won’t and shan’t) and spreading later to forms where it results in longer consonant clusters (i.e. doesn’t, didn’t, couldn’t, wouldn’t, shouldn’t, needn’t and mustn’t; see also data by Hejná and Walkden 2022: 77–79 which supports this claim). Mayn’t seems to constitute an exception as with may the contraction never really caught on (cf. Zwicky and Pullum 1983: 507; Nakamura 2023).

There is further evidence that variation between contracted n’t and full form not is systematic: Firstly, contracted n’t is associated with informal spoken language and fiction (cf. e.g. Bergs 2008: 122; Biber et al. 1999: 1129–1132; Szmrecsanyi 2003: 302, 304). Secondly, speakers are much less likely to use won’t in syntactically dependent contexts than in independent ones—an indication that speakers may opt for the full form in syntactically more complex environments, in line with Rohdenburg’s (1996) complexity principle (cf. also Szmrecsanyi 2003: 309). Thirdly, the contracted forms can’t and won’t attract monosyllabic lexical verbs as well as verbs encoding mental processes (cf. Daugs 2021: 32, 40). Finally, contractions are associated with different kinds of modality than the full forms (cf. Daugs 2021: 40–41).

Due to its idiosyncrasies, Zwicky and Pullum (1983) classify n’t as an inflectional affix. Daugs (2021: 45, 47) interprets his findings as evidence that can’t and won’t may be separate constructions. Bergs (2008: 134) makes a similar argument for shall not versus shan’t—he cautiously talks about new constructions developing “in some sense”. At the very least, this means that empirical analyses need to take into consideration that for several modals, there may be two kinds of not-negation, which are preferred in different contexts. Therefore, the last leg of the analysis investigates whether we see differences in the attraction between modals and contracted n’t versus full-form not.

4.1 Baseline frequencies

In order to determine whether modals and not-negation attract in English, we first need to know their frequency in each of the four periods under investigation. This first leg of the analysis provides baseline rates against which we can later compare the combined use of modals and not.

Figure 2 shows the relative frequency of not per million words, which seems to be on the rise. However, normalisation per million words may not be the most adequate solution as this does not account for changes in writing styles. Eighteenth and nineteenth century novels are known to be rich in nominal structures and poor in verb phrases, as exemplified by (11).

Figure 2:

Relative frequency of not in British fiction.

If we assume that most not-negation is clausal negation (i.e. scopes over the entire clause) and occurs in the verb phrase and, furthermore, that even constituent negation, which only scopes over a single constituent, such as in (12) and (13), rarely occurs more than once per clause, normalising the data per clause or verb phrase (VP) appears to be a more suitable option for diachronic comparisons.

As the corpora are not parsed, the number of VPs had to be estimated. To do so, 200 tokens of lexical verbs (including copula) were randomly retrieved from each corpus. The context to the left of these sample verbs was cut off and the context to their right was cut off before the next lexical verb. Thus only the lexical verb and all words up to the next lexical verb in the text remained.^[9] The result were four samples representing distances between lexical verbs.

Figure 3 visualises the distances between lexical verbs in the samples. Both the grey beans and the white boxes at their core provide information about the distribution of the data. The horizontal black line at the cinched waistline of the boxes indicates the median: 50 % of verbs were further apart and 50 % were closer together. Between the median and each end of the box lay 25 % of the data. Thus the box shows the inner 50 % of the data. In the case of the beans, girth indicates quantity: the wider the bean is in a particular area, the more data falls within that area. Overall, we see that the distance between lexical verbs is quite variable and that mean distances (indicated in red) have changed over time with lexical verbs in the eighteenth century on average being further apart than in other centuries.

Figure 3:

Bean plots indicating variation in the distance between lexical verbs in each period.

Based on these samples, the number of VPs in each corpus could be extrapolated. This was done by doing 500 random draws of 100 words from each sample. The mean number of lexical verbs in these draws was used as the base for estimating the number of VPs. With the help of this estimate, the percentage of VPs which is negated was determined for each corpus; i.e. the frequency of not was normalised per 100 VPs. This way of normalising the data rules out that increases in not are only due to clauses becoming shorter.

Figure 4 shows the result. Now, the slight dip in the negation rate in the eighteenth century, that was visible in Figure 2, has disappeared, indicating that it was due to eighteenth-century clauses being longer than those in the centuries before and after and authors therefore having fewer opportunities to use not. Figure 4 still suggests a slight upward trend in the negation rate until the nineteenth century, though.

Figure 4:

Frequency of not normalised per 100 verb phrases.

We can now turn to the modals. To render the data comparable to twentieth-century frequencies determined by Leech (2003, 2013 it is normalised per million words besides normalisation per 100 VPs. Figure 5 confirms the well-known decline in core modals (cf. Leech 2003, 2013) and provides evidence that it already began before the twentieth century. When normalised per million words (left panel), the data suggests a decline since EModE. When normalised per 100 VPs, however, the decline turns out to be a LModE phenomenon.

Figure 5:

Frequency of modal verbs normalised per million words (left) as well as per 100 verb phrases (right).

Yet the modals are unlikely to all exhibit a common behaviour and so need to be investigated separately. Figure 6 shows the development in the frequency of each modal. In the case of shall, should and may, usage frequency has been on the decline since EModE. Might seems to have begun its decline a century or two later. Not all core modals are on the decline, though. Frequencies of can are relatively stable and use of could even increased in the twentieth century. The most pronounced changes are those undergone by will and would. We witness their decline as well as the accompanying rise of the contracted forms ’ll and ’d. In the case of ’ll, its ascent comes at the expense of the full form of the modal (even with the two forms combined, by the twentieth century, the frequency of will/’ll is declining). By contrast, usage of full-form would remained unaffected by the rise of contracted ’d until the twentieth century and the combined frequency of ’d and would is stable (if normalised per 100 VPs) or even on the rise (if normalised per million words).

Figure 6:

Frequency of core modals normalised per million words (left) as well as per 100 verb phrases (right).

There are also some significant deviations from the frequencies reported by Leech (2003: 228). Could, might and would (if ’d is treated as a variant of would and not as a modal verb in its own right, which is how it was treated by Leech 2003: 226) are much more frequent here than in Leech’s data^[10] while may occurs much less frequently in the novels than in the corpora used by Leech (2003) which comprise several genres (LOB/FLOB; see Leech 2003: 224).

4.2 Modal negation: frequencies

Let us now turn to the central question of whether modals and negation attract, which is addressed by Figure 7. At the bottom of the graph, just above the 5 % mark, is a white line that indicates the average negation rate for VPs which do not contain a core modal. This rate was estimated with the help of the two main datasets—the modal dataset and the not dataset—as well as the random draws providing information about the number of VPs in the corpora, described in Section 4.1. The shaded area around the mean is the standard deviation, determined with the help of Chang’s (2013) Summary SE function in R. The data indicates that mean not-negation rates have risen from 5.6 % in EModE to 6.3 % in the twentieth century.

Figure 7:

not-negation rates of core modals.

Three of the curves fall on this band, indicating that the modals may, might and must have consistently had average negation rates and thus show no particular attraction towards negation. The rest of the modals can be split into three further groups: Group one—the contracted modals ’ll and ’d—repels not-negation.^[11] The second group is constituted by shall, should, will and would, which show elevated negation rates, often doubling or tripling the average rate. By the twentieth century, full-form will approximates the behaviour of the third and final group, where negation rates are more than four times that of the baseline. This group is made up of can and could.

Surprisingly, even the highest twentieth-century rates of not-negation lie consistently below those reported by Mindt (1995: 176; assuming that he treated will and ’ll as one type). Yet they are higher than those found by Römer (2004: 189) in the spoken section of the BNC. Mindt (1995) does not report which part of his database was used to determine rates of modal negation. Therefore, I can only surmise that the differences are due to genre effects.

Figure 8 provides a different view of the data. Here, numbers are shown as frequencies per 100 VPs. It shows that changes in the negation rate can be a symptom of an increase or decrease in the usage frequency of affirmatives. The strong increases in the negation rate of will, for instance, are due to stronger losses in the use of affirmative will than of not-negated will.

Figure 8:

Modal negation normalised per 100 VPs.

What we have seen so far is that not all core modals seem to attract not—though predominantly the more frequent ones do. What remains is to statistically assess the strength of the attraction as well as to take the opposite perspective: does not attract modals? Both of these will be answered in the following section.

4.3 Modal negation: association strengths

Analyses of collocations (i.e. associations between two lexemes or word-forms) and collostructions (i.e. associations between a construction and a lexeme) are generally based on a contingency table, such as Table 2, which cross-tabulates the frequencies of the two elements (cf. e.g. Stefanowitsch and Gries 2003). Cell a is the one that is of particular interest to the researcher as it represents the condition where both elements are present, in our case, both the modal and the negator. b, c and d are used to calculate association measures which assess the stregth of attraction or repulsion between two elements.

Usually, word₁ is given more prominence in the analysis in that all relationships are decribed from the perspective of word₁. In the present case, however, we are equally interested in the negator and the modal as both are Transitivity-lowering factors. This does neither mar nor complicate the analysis; it merely means that we will interpret some measures differently from the way they are commonly interpreted.^[12]

As a first step, we compute attraction (or forward transitional probability/cue availability, cf. Bates and MacWhinney 1987: 164). It is based on the formula (cf. e.g. Schmid and Küchenhoff 2013: 550; Levshina 2015: 228):

Once we fill the formula with values based on Table 3, we get a measure which indicates how strongly a core modal attracts not. Mathematically, attraction is the percentage of modal tokens which co-occur with not. It has already been discussed above; see Figure 7, which is repeated here as the left panel of Figure 9. Attraction is a directional measure of association as it quantifies how strongly the left element attracts the right one. The corresponding measure of attraction from right to left is often referred to as reliance (or backward transitional probability/cue reliability, cf. Bates and MacWhinney 1987: 164). It is calculated as follows (cf. e.g. Schmid and Küchenhoff 2013: 550; Levshina 2015: 228):

Figure 9:

Attraction and reliance.

Usually, reliance is interpreted as a measure of how strongly word₁, in our case the modal, relies on word₂. It does that here, too; yet for the present purposes it is more useful to interpret it as a measure of how strongly not attracts the modal. Reliance is the percentage of not which is accompanied by a specific modal. We see it in the right panel of Figure 9.

Reliance is much lower than attraction—it was bound to be, due to the fact that we are dealing with a one-to-many relationship. This means that we are looking at one negator versus eleven core modals and each instance of not can only co-occur with a single modal. Therefore all reliance scores together cannot exceed 100 %. Coincidentally, the curve showing the combined reliance of all core modals on not would be almost identical to the curve displaying the attraction scores of could in the left plot—it has declined from 31.3 % in EModE to 23.4 % in the twentieth century.

Attraction and reliance are sometimes criticised for being hard to interpret, as they quantify associations but do not provide any information about significance (cf. Schmid and Küchenhoff 2013: 550). For significance testing, we have to compare the associations measured by attraction and reliance to distributions elsewhere in the language. This requires that cell d in the tables above be filled, which is not trivial (cf. e.g. Schmid and Küchenhoff 2013: 551–544). As modals and negators do not float through texts at random but occur only in syntactically restricted positions, the VP is once again a better unit than the word to determine corpus size.

We can now employ collostructional analysis as developed by Stefanowitsch and Gries (2003) to determine the mutual attraction between each modal and not. The standard metric for this type of analysis is Fisher’s Exact Test. Yet, it may not be the most suitable measure in the present case. The test has been criticised for three things; firstly, for being difficult to compute—many applications return 0/infinity when the test is applied to large datasets;^[13] secondly, for only providing information about significance and not of effect size and, thirdly, for being sensitive to sample size. The latter means that the same degree of deviation from expected values can result in a non-significant scoring when numbers are small while receiving significant scores when numbers are large (see discussion between Schmid and Küchenhoff [2013] and Gries [2015]).

To remedy these issues, Gries (2023) has recently proposed both a simplification and an expansion of collostructional analysis. In terms of simplification, he suggests replacing Fisher’s Exact p with chi² residuals. The latter are obtained by performing a chi²-test on a complete table of all associations to be tested (Gries 2023: 358). R outputs the residuals with chisq.test()$residuals, but they can also easily be calculated: observed − expected / expected . Gries (2023: 364–365) argues that this switch from one metric to another makes collostructional analysis faster and more accessible, while at the same time solving the 0/infinity issue. Chi² residuals and Fisher’s Exact p are strongly correlated, which means that they provide users with the same information about associations (cf. Gries 2023: 359–360).

Gries’s (2023: 370–372) proposed expansion solves the second and third issues. He confirms that Fisher’s Exact p and chi² residuals “not only reflect frequency and association, but they also reflect frequency more than association” (Gries 2023: 372; empasis in the original). He advocates that for some applications of collostructional analysis, frequency and association be kept separate (and that dispersion be added to the mix of dimensions taken into consideration). Thus, he proposes using log Odds Ratio, which is bidirectional and not sensitive to sample size. As an added perk, Odds Ratio is simple to compute (cf. Schmid and Küchenhoff 2013: 554):

As its name implies, it compares two odds, which relate the probability of the observed to “the probability of what could also have happened given the full set of possibilities” (Schmid and Küchenhoff 2013: 553). Odds Ratio is a true effect size measure in the sense of Schmid and Küchenhoff (2013: 552–555). Chi² residuals will therefore be combined with Odds Ratios.^[14] I will provide more information about both measures as we discuss the results.

The left panel in Figure 10 shows the chi² (or Pearson) residuals. All curves above the x-axis belong to modals which co-occur more frequently with not than statistically expected, while all curves below the x-axis belong to modals which co-occur with not less frequently than statistically expected. Any residual higher than 3.25 or lower than −3.25, i.e. which falls outside of the narrow corridor around the x-axis which is indicated in the plot, corresponds to a significant result.^[15]

Figure 10:

Association between modals and not assessed with the help of chi² residuals (also known as Pearson residuals) and Odds Ratio.

The right panel in Figure 10 shows Odds Ratios. The following example illustrates how they can be interpreted: the Odds Ratio for will not/won’t in the twentieth-century data is four. This means that if we encounter a VP containing will, the odds that it also contains not are four times higher than the odds of a will-less VP to do so. And vice versa; if we encounter a not-negated VP, the odds that it also contains will are four times higher than the odds of a not-less VP to do so (cf. Schmid and Küchenhoff 2013: 554). An Odds Ratio of one indicates that there is no effect, while ratios below one indicate repulsion. The solid grey line added to the plot marks this dividing line between attraction and repulsion. The other lines indicate effect size. Anything below the dashed line is a small effect, points between the dashed and the dotted lines are medium effects and points above the dotted line represent large effects.^[16]

The bidirectional tests largely confirm the impression we gleaned from the unidirectional tests above:

1. Although declining, the attraction between can/could and not has been strong for the past 500 years (OddsRatio > 4.25, p < 0.001).

2. The attraction between will/would and not has been on the rise for the past 200 years. By the twentieth century, it reached medium strength (OddsRatio > 2.47, p < 0.001).

3. The attraction between shall/should and not has been consistently weak but still highly significant (OddsRatio < 2.47, p < 0.001).

4. Must and not have shown neither attraction nor repulsion until the twentieth century, by which time they began to develop a (still very) weak repulsion (OddsRatio between 0.4 and 1, p < 0.001).

5. The relationship between may/might and not is one of weak repulsion (OddsRatio between 0.4 and 1), while the effect is consistently significant in the case of might, with may, it only reaches significance in the eighteenth and nineteenth centuries.

6. The repulsion between the contracted modal ’ll and not has increased. By the twentieth century, the effect is large (OddsRatio < 0.24, p < 0.001). The contracted modal ’d and not have strongly repelled ever since the form emerged in the nineteenth century (OddsRatio < 0.24, p < 0.001).

4.4 Contraction rates of not

In this section, we take a closer look at the two variants of not, namely full-form not and contracted n’t (assuming that n’t and much rarer ’nt are mere spelling variants). Figure 11 shows a familiar picture (cf. e.g. Daugs 2021: 26; Hejna and Walkden 2022; López-Couso and Pérez-Guerra 2023; Nakamura 2023): contraction of not was already an option in EModE, but did not really take off until the nineteenth century. While the present data suggests continuing rapid replacement of full-form not by n’t, a decade-by-decade analysis of not contraction rates in the Corpus of Historical American English (COHA, Davies 2010–) by Daugs (2021: 26) suggest that—at least in American English—the development follows an S-curve, with the rate of replacement plateauing by the middle of the twentieth century. As the BNC only covers the second half of the twentieth century, it is not possible to tell from the present data, whether this is the case in British English, too.

Figure 11:

Contraction rates of not overall (left) and in combination with specific core modals (right).

Each curve in the left panel, however, represents a rather mixed bag of collocates. The grey curve mostly represents negation with forms of be, have and do, which show strongly divergent contraction rates (compare, e.g., *amn’t, aren’t, isn’t; Hejná and Walkden 2022: 79).^[17] The same is true for the modals, as becomes evident from the right panel.

A look at each modal separately (Figure 11, right) shows interesting similarities and differences. Can’t, won’t, shan’t and mayn’t were the pioneers in terms of contraction. Until the eighteenth century, they were the only modals with rates of not-contraction exceeding 2 %. As predicted by Nakamura (2023), this is exactly the group where the resulting final consonant cluster consists exclusively of /nt/. Their initial ranking (won’t and can’t above shan’t and mayn’t) suggests that within this group, frequency had an effect on early adoption rates of contraction—either the frequency of the modals by themselves or the combined frequency of modal + not (see Figures 6 and 7).

Except with may, might and the contracted modals ’ll and ’d, n’t goes on to surpass full form not. With will and can, the cross-over happens first: here, contraction rates already reach c. 50 % by the nineteenth century (cf. also Daugs 2021: 26), while with all other modals, n’t does not surpass not until the twentieth century. Mayn’t and mightn’t, on the other hand, are declining in popularity and, by the twentieth century, mayn’t has become all but impossible (cf. also Leech et al. 2009: 81). These patterns suggest that from the nineteenth century onwards, frequency (of modal + not) had a broader influence on negative contraction rates among modals.

Overall, the results confirm patterns described elsewhere, particularly those by Hejná and Walkden (2022: 79).^[18] Yet, the contraction rates themselves are higher than those based on other sources (see, for instance, contraction rates in Early English Books Online, EEBO, as reported by Nakamura 2023) presumably owing to the informality of the genre (cf. López-Couso and Pérez-Guerra 2023).