2. Historical overview of English verbal negation

Jespersen (Reference Jespersen1917: 4) observes that negation is often subject to a ‘curious fluctuation’:

This has become known as Jespersen’s Cycle. English negation has already undergone the cycle once. Old English (OE) started out with the negator ne preceding the finite verb, as in (1) (cf. e.g. Jespersen Reference Jespersen1917: 9; Denison Reference Denison1993: 449; Laing Reference Laing2002: 298; Fischer et al. Reference Fischer, De Smet and van der Wurff2017: 157).

When combining with forms of the auxiliaries beon/wesan (‘be’) and habban (‘have’) as well as with pre-modals like wile (‘will’) and some frequent verbs like witan (‘know’), ne cliticised onto the verb, ne + wile > nele, see (2) (cf. e.g. Denison Reference Denison1993: 449; Fischer et al. Reference Fischer, De Smet and van der Wurff2017: 157), which Jespersen (Reference Jespersen1917: 9) classifies as weakening.

OE allowed negative concord, i.e. multiple negators in the same clause, which did not cancel each other out. Thus, ne often appeared together with other negative elements which strengthened the negation. One of these was nawith (‘no wight’), originally a negative pronoun, as in (3), which later reduced to nouht and eventually to not and became a negative adverb, as in (4) (cf. e.g. Denison Reference Denison1993: 449; Laing Reference Laing2002: 299–300; Fischer et al. Reference Fischer, De Smet and van der Wurff2017: 157).

In Middle English (ME), ne was increasingly dropped and not remained as the only negator, such as in (5) and (6), which completed the cycle (cf. e.g. Jespersen Reference Jespersen1917: 9; Denison Reference Denison1993: 450; Laing Reference Laing2002: 299; Fischer et al. Reference Fischer, De Smet and van der Wurff2017: 157–8).

While not could initially follow any finite verb,Footnote ³ be it an auxiliary, as in (5), or a lexical verb, as in (6), between c. 1500 and 1700, the pattern finite lexical verb + not gradually fell out of use, except with some highly frequent verbs with which it survived longer (cf. e.g. Ellegård Reference Ellegård1953: 200). The standard pattern in negative declarative sentences became finite operator + not. Where no other auxiliary was present, do became the obligatory operator (cf. e.g. Jespersen Reference Jespersen1917: 10–11; Ellegård Reference Ellegård1953: 162; Visser Reference Visser1969: 1534–6; Strang Reference Strang1970: 151; Denison Reference Denison1993: 451; Fischer et al. Reference Fischer, De Smet and van der Wurff2017: 130–1, 158). In the late sixteenth or early seventeenth century, not began to contract and to cliticise onto the operator, as in (7) (cf. e.g. Jespersen Reference Jespersen1917: 117; Denison Reference Denison1993: 309). Contraction was rare at first but caught on in the nineteenth century (cf. Daugs Reference Daugs, Hilpert, Cappelle and Depraetere2021: 26; Hejná & Walkden Reference Hejná and Walkden2022: 79; Nakamura Reference Nakamura2023; Schneider Reference Schneider2023: 25).

While we may simply note that this marks the beginning of a second spin around Jespersen’s Cycle, there are actually good reasons to look more closely at the alternating variants, i.e. full-form not and contracted n’t, as there are indications that (while variation persists) speakers are associating not and n’t with different cotextual factors.

Using data from the Corpus of Contemporary American English (COCA; Davies Reference Davies2008–) and collostructional methods (see e.g. Stefanowitsch & Flach Reference Stefanowitsch, Flach, Pastor and Colson2020 for an overview), Daugs & Lorenz (Reference Daugs and Lorenz2024) show that won’t is distributionally skewed towards inanimate subjects (e.g. it, there, that) and stative verbs (e.g. be, happen, matter), thereby predominantly conveying ‘prediction’. By contrast, will not typically collocates with animate first-person subjects (I and we) and verbs conjointly indicating ‘unwillingness’ (e.g. accept, tolerate, permit). Uses of n’t have furthermore been shown to be associated with informal spoken language and fiction as well as with syntactically independent contexts, with monosyllabic lexical verbs and verbs encoding mental processes as well as with different kinds of modality than their full-form counterparts (cf. Bergs Reference Bergs, Trousdale and Gisborne2008: 122; Biber et al. Reference Biber, Johansson, Leech, Conrad and Finegan1999: 1129–32; Szmrecsanyi Reference Szmrecsanyi2003: 302–9; Daugs Reference Daugs, Hilpert, Cappelle and Depraetere2021: 32, 40–1).

Similar observations have been made for the enclitics ’d and ’ll. Nesselhauf (Reference Nesselhauf and Hundt2014), for example, claims that ’ll developed the meaning ‘spontaneous decision’ in Late Modern British English, which is hardly ever expressed with the full form. Daugs (Reference Daugs2022) supports this finding with data from American English, but proposes that the meaning is not directly expressed by the enclitic but by specific subject and verb collocations that are more likely to co-occur with ’ll than with will.

The key conclusions drawn from these results are that (i) the contractions’ and full forms’ symbolic associations (i.e. the form–meaning correspondences) are not inherent properties of their nodes in the network, but they emerge dynamically from and are continuously updated through the interaction of all associations that each pattern evokes; and (ii) the distinction between contractions and their full forms is not categorical but gradient, reflecting the strength of their respective associations (cf. Daugs Reference Daugs2022: 242–5; Daugs & Lorenz Reference Daugs and Lorenz2024: 18–23).

3. British–American differences and commonalities

In how far various modal constructions have evolved and changed along different paths in British and American English and whether speakers’ cognitive associations with these constructions differ between varieties is currently difficult to determine due to the dearth of studies on the topic. What we do know is that will and would are the most frequent modals in both British and American English, together constituting over 40 per cent of all uses of core modals in written English in the twentieth century (based on data from the Brown corpora provided by Leech Reference Leech, Facchinetti, Palmer and Krug2003: 228, 2013: 101; Mair Reference Mair2006: 101 as well as Leech et al. Reference Leech, Hundt, Mair and Smith2009: 74). Their share rises to over 70 per cent once we look at spoken language (based on British data provided by Leech et al. Reference Leech, Hundt, Mair and Smith2009: 78). Over the course of the twentieth century, usage of the modals seems to have been on the decline in both varieties – yet not at equal rates.

Table 1 shows that in the Brown corpora (written English), will declined at a faster rate in American English between 1961 and 1991 while it was would which declined faster in British English. This may be at least partially explained by the faster spread of going to in American English (cf. Biber et al. Reference Biber, Johansson, Leech, Conrad and Finegan1999: 488; Tottie Reference Tottie2002: 158; Leech Reference Leech, Facchinetti, Palmer and Krug2003: 229). As a consequence, will may have become associated with more formal language in American English as evidenced by going to and will being equally frequent in informal American English, but will being more frequent in formal American English (cf. Mair Reference Mair2006: 99; see also Leech Reference Leech, Marín-Arrese, Carretero, Hita and van der Auwera2013: 112) – an effect not found in British English.

Table 1. Changes in usage frequency of will and would in written British and American English (based on data from the Brown corpora provided by Leech Reference Leech, Facchinetti, Palmer and Krug2003: 228, 2013: 101; Mair Reference Mair2006: 101 as well as Leech et al. Reference Leech, Hundt, Mair and Smith2009: 74)

Besides medium/register, the most influential determinant of diverging usage rates of will and would in British and American English may be competition with shall/should. While today any remaining uses of shall are strongly associated with British English (cf. Bergs Reference Bergs, Trousdale and Gisborne2008: 118), Kytö’s (Reference Kytö1991) analysis of will/shall and would/should variation in Early Modern English indicates that this has not always been the case. Between 1570 and 1640, American writers were far more likely to opt for shall or should instead of will or would with first-person subjects than British writers (cf. Kytö Reference Kytö1991: 334). By this time, shall was already strongly associated with first-person subjects in American English, an association which subsequently also developed in British English (cf. Kytö Reference Kytö1991: 334–6). The revival of the subjunctive in American English does not seem to have had an effect on the usage frequency of will, though (despite it being interchangeable with the subjunctive in some contexts). In the subjunctive-triggering context analysed by Schlüter (Reference Schlüter2009), no such effect is evident.

Concerning contracted forms of the modals, we see ’ll appearing as a contracted form of will in writing from the sixteenth century onwards (cf. Schneider Reference Schneider2023: 16). The contraction ’d as a short form of would, however, doesn’t appear until much later. Before the nineteenth century ’d almost exclusively occurs as a representation of the past tense and past participle -ed suffixes (cf. Schneider Reference Schneider2023: 10).Footnote ⁴ Both ’ll and ’d are initially rare in writing. By the twentieth century, ’ll has caught up with will in British novels, yet the frequency of would is still more than four times higher than that of ’d (cf. Schneider Reference Schneider2023: 16). Results for American English vary. On the one hand, ’ll has been shown to be less frequent in twentieth-century American English than in British English (cf. Algeo Reference Algeo2006: 23; Szmrecsanyi Reference Szmrecsanyi2003: 302), on the other hand, data with pronominal subjects shows high contraction rates in twentieth-century American English (cf. Daugs Reference Daugs2022: 234).Footnote ⁵ Interestingly, Nesselhauf (Reference Nesselhauf, Hundt, Nesselhauf and Biewer2007: 291) indicates that a rise in the use of shall in nineteenth-century British English does not seem to have come at the expense of will but at that of reduced ’ll (this is, however, only the case in one of two datasets representing British fiction; cf. Nesselhauf Reference Nesselhauf, Hundt, Nesselhauf and Biewer2007: 295).

After contractions of the negator not (e.g. can’t) emerged in the late sixteenth or early seventeenth century (cf. Jespersen Reference Jespersen1917: 117; Denison Reference Denison1993: 309), they were much less frequent than their full-form counterparts for about 200 years. Not before the twentieth century do they surpass the full forms in frequency (cf. Millar Reference Millar2009: 211; Daugs Reference Daugs, Hilpert, Cappelle and Depraetere2021: 26; Hejná & Walkden Reference Hejná and Walkden2022: 79; Nakamura Reference Nakamura2023; Schneider Reference Schneider2023: 25).

Crucially, modal contraction and negative contraction are incompatible (e.g. *he’lln’t go; though some combinations of modal + contracted negator trigger a different kind of contraction of the modal, as in won’t). While this theoretically still leaves speakers with three options to choose from, e.g. he will not go, he won’t go and he’ll not go, ’ll and particularly ’d are hardly used in negative contexts – if so, it is mostly in British English (cf. Yaeger-Dror et al. Reference Yaeger-Dror2002: 99; Szmrecsanyi Reference Szmrecsanyi2003: 304; Nesselhauf Reference Nesselhauf, Hundt, Nesselhauf and Biewer2007: 292; Varela Pérez Reference Varela Pérez, Aarts, Close, Leech and Wallis2013: 267; Schneider Reference Schneider2023: 17). Algeo (Reference Algeo2006: 23), for instance, notes that won’t is 36 times as frequent as ’ll not in British English, while it is 346 times (!) more frequent than ’ll not in American English. He further finds that ‘’d not (representing both would not and had not) occurs 4 times as often in British texts as in American’ (Algeo Reference Algeo2006: 24). It seems that ’ll not is a regional feature of certain British dialects (cf. Tagliamonte & Smith Reference Tagliamonte and Smith2002: 268). Interestingly, nineteenth-century rates of ’ll not seem to have been higher than those in the twentieth century (cf. Nesselhauf Reference Nesselhauf, Hundt, Nesselhauf and Biewer2007: 296; Schneider Reference Schneider2023: 17). Concerning the use of full forms, Algeo (Reference Algeo2006: 22–3) suggests that they are more typical of American English, while in López-Couso & Pérez-Guerra’s (Reference López-Couso, Pérez-Guerra, De Smet, Petré and Szmrecsanyi2023: 12) multifactorial analysis, variety (British vs American) was not found to be a significant predictor of negative contraction versus use of the full form. Unfortunately, most analyses are binary, contrasting only two of the options, i.e. either negative contraction versus auxiliary contraction or full form versus negative contraction. Moreover, there are actually further options worth exploring, one of them being negation by means of never: while not-negation rates of ’ll and ’d are low, these contracted modals actually attract never more strongly than the full forms of the modals do (cf. Tagliamonte & Smith Reference Tagliamonte and Smith2002: 268; Flach Reference Flach2020a: 752; Schneider Reference Schneider2023: 17).

4. Data and method

In order to explore the cognitive associations speakers have with different – particularly negated – modal constructions with will and would, how these have changed over the course of the nineteenth and twentieth centuries as well as potential differences between British and American speakers’ behaviour, we need large amounts of comparable diachronic data, preferably from a consistent genre. We therefore restricted our analysis to prose fiction. For American English, we opted for the Corpus of Historical American English (Davies Reference Davies2010–), which provides nineteenth-century and twentieth-century fiction. For British English, complete coverage of both centuries is harder to achieve and we had to combine several corpora. Two larger corpora, i.e. Nineteenth-Century Fiction (Karlin & Keymer Reference Karlin and Keymer1999–2000) and the imaginative prose subdomain of the British National Corpus (1995), leave a gap in the first half of the twentieth century, which we had to fill. To do so, we used the imaginative prose section of BLOB-1931, a member of the Brown family of corpora (Leech & Smith Reference Leech and Smith2005), as well as a collection of British novels freely available through Project Gutenberg (gutenberg.org). Table 2 lists the time-periods covered by the individual corpora as well as their sizes.

Table 2. American and British corpora used in the present study

From these corpora, we extracted all tokens of the following pattern:

This means we extracted all negated and non-negated occurrences of will and would with pronominal subjects from the corpora. Crucially, contraction of the modal and contraction of the negator are for the most part incompatible, as illustrated in table 3 – except in won’t, where a phonetically reduced form of will combines with contracted n’t (cf. Schneider Reference Schneider2023: 25). Thus, the predominant use of ’d or ’ll in a specific environment precludes the use of n’t in that environment.

Table 3. Modal-negation contractions

In total, we obtained 967,149 datapoints, c. 743,500 from American English and 223,500 from British English. Each token was then coded for the following factors which transpired to be significant predictors in our previous work on modal negation and contraction (cf. Daugs Reference Daugs2022; Schneider Reference Schneider2023; Daugs & Lorenz Reference Daugs and Lorenz2024):

• variety – British English or American English

• year of publication

• modal – will or would

• negation – affirmation or negation

• negator – not, n’t or never

• contraction of the modal – full form or contracted; for statistical purposes wo in won’t was treated as a full form of will

• subject – I,you/thou, s/he, it, we or they

• semantics of the lexical verb – for this purpose, a tagger based on the WordNet lexical database (Princeton University 2010) was used. It assigns every verbal lemma the most frequent sense of that lemma, i.e. body, change, cognition, communication competition, consumption, contact, creation, emotion, motion, perception, possession, social, stative or weather (Schneider Reference Schneider2022).

Graphs were generated in R (2023); trend curves are based on generalised additive models (GAMs).

5. Analysis and results

5.1. General trends

A first look at the frequency distribution of the data reveals that previous assumptions about developments in the use of core modals are not born out. Figure 1 shows the relative frequency of all tokens of will and would combined (whether full or contracted, negated or affirmed). Instead of the predicted loss of these two core modals, we see that usage increases over the course of the nineteenth century and plateaus for most of the twentieth century – with the exception of will in American English, which actually declines again.Footnote ⁹ The factors most likely responsible for this discrepancy between our results and those of Leech (Reference Leech, Marín-Arrese, Carretero, Hita and van der Auwera2013) are genre and pronominal subjects, though similar results obtained by Schneider (Reference Schneider2023) based on a wider range of subjects suggest that genre is the stronger contender: core modals may have been retained at higher rates in prose fiction than in other written registers.

Figure 1. Relative frequencies of will and would in British and American English

Figure 2 shows changes in the negation rate over time – or rather lack thereof, as negation rates are remarkably consistent. This indicates that the changes in usage frequency evident in figure 1 are distributed equally across affirmed and negated tokens. In both varieties, negation rates are around 15 per cent for will and just over 17 per cent for would. Only in nineteenth-century British English is the negation rate for will slightly higher (18.4 per cent). Compared to the average negation rate of English verb phrases, which lies at around 7 per cent (cf. Schneider Reference Schneider2023: 14), negation rates of will and would are highly elevated.

Figure 2. Negation rates of will and would in British and American English. Point size is proportional to the sum total frequency per million words for each pair per year

5.2. Modal contractions

We will now narrow in on uses of will and would in which either the modal or the negator – if present – is contracted. The top panels in figure 3 show overall contraction rates. As is typical for robust language change in progress, the observable trend has an S-curve shape, indicating how a variant diffuses and gains ground on its competing alternative (cf. e.g. Blythe & Croft Reference Blythe and Croft2012; Nevalainen Reference Nevalainen and Sanchez-Stockhammer2015). In the case of will, the beginning of the twentieth century marks the point when contracted ’ll becomes more frequent in American fiction than full will. Contraction rates of would are lower, reaching only around 20 per cent by the beginning of the twentieth century. Developments in British English are similar, though in the case of would British English seems to reach the steep phase in the S-curve slightly later.

Figure 3. Contractions of will and would in British and American English

First idiosyncrasies of the contracted forms ’ll and ’d become evident when we turn to contraction rates in negated contexts (second row in figure 3). These are far below the overall contraction rates. This indicates that in both varieties a split has emerged in that will/wo and would are associated with negation while ’ll and ’d are used in affirmative contexts. This is particularly strongly evident in the case of twentieth-century will; in both varieties, reduced ’ll only has a 4 per cent chance of being negated while negation rates of will/wo are nine times higher (34 per cent in BrE and 37 per cent in AmE).

Interestingly, though, this split is not found with the adverbial negator never. In this context, usage rates of ’ll and ’d are even higher than they are overall (compare the first and third row of panels in figure 3), which indicates that ’ll and ’d indeed attract never.

Finally, the last panel in figure 3 contrasts ’ll not with won’t in order to assess whether the former is a predominantly British variant. The graph suggests that this had not been the case in the nineteenth century but that an American dispreference for the construction may have developed over the course of the twentieth century.

5.3. Configural Frequency Analysis

In order to probe for further relevant usage patterns, we employ Configural Frequency Analysis (CFA). CFA is a multivariate extension of the more widely known χ²-test for larger than 2x2 contingency tables (cf. Krauth & Lienert Reference Krauth and Lienert1973; von Eye & Wiedermann Reference Eye and Wiedermann2021). As an exploratory method, it aims at detecting patterns in categorical data by assessing what configurations of variables constitute types (where observed frequencies are significantly higher than expected frequencies) or antitypes (where observed frequencies are significantly lower than expected frequencies). Importantly, CFA is geared towards identifying local contributions to describing the structure in the data. In other words, it detects and ranks the configurations (or cells in a large contingency table) that add most to the variability.

Due to it being a relative of the χ²-test, CFA suffers from some of the same flaws and is subject to similar limitations. Firstly, p-values are dependent on sample size. Therefore, when running a CFA on close to a million datapoints, we are bound to a find a large number of types and antitypes. We will counter this to some degree by ranking types based on their contribution to overall χ², which is less susceptible to token frequencies.Footnote ¹⁰

Secondly, empty cells in the configuration table, i.e. non-observed configurations, can be problematic. These fall into two categories: structural zeros (more problematic) and empirical zeros (less problematic).

Structural zeros represent configurations which for logical, grammatical or other reasons cannot occur. Thus, statistically expected values need to be adjusted accordingly. In our case, this applies to the combined occurrence of contracted modals with a contracted negator, as in *he’lln’t, which we know cannot (yet) occur. This configuration has been ‘blanked out’ (cf. von Eye & Mair Reference Eye and Mair2007), so that the CFA does not generate expected values for configurations with *’lln’t or *’dn’t.

Empirical zeros, on the other hand, are configurations which could have been observed, had more data been available (cf. von Eye et al. Reference Eye, Mair and Mun2010: 42). These are less problematic for exploratory analyses (like the present one) than they are for significance testing (cf. Winter Reference Winter2019: 277–9). Nevertheless, we made efforts to reduce these, by (a) excluding tokens with weather verbs from the analysis, as they were extremely rare and (b) by merging the annual datapoints into larger periods using Variability-based Neighbour-Clustering (VNC; cf. Hilpert & Gries Reference Hilpert and Gries2009). Based on the annual rate of modal contraction across both varieties, this algorithm identifies the similarity between adjacent years by means of the standard deviation. After outlier-detection, the algorithm segmented the S-curve into three periods (P1: 1810–31, P2: 1832–1904; 1905–93); see figure 4. The three horizontal bars represent the respective mean rate of modal contraction for each VNC-determined period (P1: 9.5%, P2: 25.1%, P3: 50.1%). Consequently, we use binned years of publication in the CFA instead of exact years of publication.

Figure 4. The overall development of contractions relative to their full forms in British and American English with VNC-determined stages

Several studies (presumably) assume a nested structure in the data and therefore employ Hierarchical Configural Frequency Analysis (HCFA), an alternative to CFA (see e.g. Hoffmann Reference Hoffmann2019; Hilpert et al. Reference Hilpert2021), where variables are recursively eliminated by essentially running multiple CFAs to help determine the minimal adequate configuration (cf. Gries Reference Gries2009: 248). In the present context, we opted for a more theory-driven approach and ran a CFA with exactly the configurations that cater to the variables we predicted to have an impact rather than finding the ‘best’ configuration via automated selection (cf. Winter Reference Winter2019: 277–9).

Accordingly, the following variables entered the CFA: period, variety, modal, contraction, negation (affirmation, not, n’t, never), semantics, subject. The analysis was performed with the help of the cfa package in R (Mair et al. Reference Mair, Funke, Harloff and von Eye2024). Out of the possible 7,056 configurations (after blank-out), 5,258 occurred. Of these, 150 were classified as types, 34 as antitypes. In order to provide a more structured overview of the results, the discussion will proceed period by period.

Table 4 shows the ten strongest types for the first period. All of them stem from British English. The CFA, in fact, only declares a single configuration in American English a type for this period. This skew partially owes to the fact that before 1832 (i.e. in the timespan encompassed by the first period) will and would were significantly more frequent in British English than in American English – a contrast that reversed in the twentieth century (third period).Footnote ¹¹

Table 4. Strongest types in the first period (<1832)

The fact that only the full forms will and would form types that stand out for the period before 1831 is not surprising. It confirms that the contractions ’ll and ’d were comparatively rare in the early nineteenth century. In this respect, will and would show similar usage patterns. What differentiates them is that would is associated with uses in the third-person singular (i.e. it forms almost exclusively types with third-person-singular subjects) while all types with will (bar one) have I or you as the subject. The single exception is it will V_stative (e.g. be). Of the few negated types of this period, only I will not + V_{communication} makes it into the top ten.

The period 1832–1904 is a transitional one. While British authors still use significantly more will and would than their American counterparts (see fn. 11), we also see American types characterising this period (in fact, almost half of the second-period types are from American English, i.e. 28 out of 60). What we also find now are types with contracted ’ll – though not yet with ’d. Types with contracted ’ll occur in both varieties and almost exclusively have first-person-singular subjects. The strongest such type is I’ll + V_{communication} (BrE); see table 5. Interestingly, while none of the configurations from the first period were antitypes, there are seven antitypes from the second period. Five of these are uses of contracted ’d, all in American English, which puts into doubt that American speakers were the innovators when it comes to contracting would to ’d.

Table 5. Strongest types in the second period (1832–1903)

By the twentieth century, usage frequencies have reversed; see table 6: American English has higher rates of will and would than British EnglishFootnote ¹² and the majority of twentieth-century types are American (78 out of 86). Some of the types are familiar, but there are also many unfamiliar ones among the top ten. Types with contracted ’ll predominate. ’ll exclusively forms types with first- and second-person subjects and mostly with affirmation. There are, however, three negated types with ’ll, none of which have made it into the top ten. They are:

This indicates that ’ll never is associated with cognitive verbs such as think, know, forget and believe. Moreover, that the configurations in (9) are the only negated types with ’ll in any of the three periods and that they are types in AmE leads us to conclude that ’ll not does not stand out in the CFA as a predominantly British phenomenon.

The third period is also the first one in which we find types with ’d. The strongest of these is I’d + V_emotion (e.g. like, worry, wish), as seen in table 6, but it also forms types with cognition verbs (not among the top ten). While full-form would is still associated with third-person subjects, ’d is so strongly associated with emotion and cognition that it is more likely to have I or you as its subject.

Table 6. Strongest types in the third period (>1903)

Finally, all types with n’t come from the third period. Table 7 shows that wouldn’t contrasts with other uses of would in terms of the subjects it is associated with. Like ’d, it forms types with verbs of cognition and emotion and consequently also with first and second-person subjects instead of would’s usual third-person subjects.

Table 7. Types with n’t