1. Aims and state-of-the-art
In this article, we describe and analyze recent change in the frequency of use of hedged performatives (HPs) (Fraser Reference Fraser, Cole and Morgan1975) in spoken American English. Hedged performatives are combinations of a (semi-)modal verb and a performative verb, and they were first analyzed in detail by Fraser. The following are among the examples that he gives:
Fraser’s category of ‘performative’ verbs is broad: he classifies them in terms of eight types of ‘illocutionary acts’ (acts of asserting, evaluating, stipulating, requesting, suggesting, exercising authority, committing, acts reflecting speaker attitude), together resulting in a total of 275 ‘performative verbs’.Footnote 1 Fraser principally studies the impact of the modal verb on the illocutionary force of the performative verb that features in the HP: does it retain its illocutionary force, or does the addition of the modal mean that the illocutionary force of the performative gets hedged (hence the label used by Fraser to refer to the various combinations corresponding to I + (semi-)modal verb + performative verb)? Put differently, is there a difference in force between, for example, I must request and I request, and if so, how can this difference be accounted for?
HPs often appear in studies examining hedging within specific genres, including academic writing and political speeches (e.g. Hinkel Reference Hinkel1997; Meyer Reference Meyer, Markkaned and Schröder1997; Thue Vold Reference Thue Vold2006; Ponteretto Reference Ponteretto2018). Here they are mentioned in the list of hedging devices, with no detailed analysis being offered. A similar observation applies to Brown & Levinson’s book (Reference Brown and Levinson1987) on politeness, in which HPs serve as examples to illustrate the negative politeness strategy of hedging. In research on indirectness as well (e.g. Blum-Kulka & Olshtain Reference Blum-Kulka and Olshtain1984: 201; Blum-Kulka Reference Blum-Kulka1987: 134, 137–9; Leech Reference Leech2014: 165), HPs are included in the category of markers of indirectness. Finally, the concept of HPs is also mentioned and illustrated in empirical research on modal verbs, as in Palmer (Reference Palmer1990), Collins (Reference Collins2009), Coates (Reference Coates2014 [1983]), Johansson (Reference Johansson, Aarts, Close, Leech and Wallis2013), Cappelle, Depraetere & Lesuisse (Reference Cappelle, Depraetere and Lesuisse2019).
There are few detailed studies on HPs. Two papers offer a critical analysis of Fraser’s explanation of the ways in which the illocutionary force of the performative verb can be affected by the modal: Schneider (Reference Schneider, Kaltenböck, Mihatsch and Schneider2010) analyzes parenthetical HPs across several languages, including spoken English, French, Italian and Spanish. His discussion also moves beyond that of the impact of the modal on illocutionary force. Notably, he observes that HPs may sometimes apply to only a segment of the utterance, suggesting a phrasal rather than clausal scope. A more recent investigation by Panther & Thornburg (Reference Panther, Thornburg, Bolognesi, Brdar and Despot2019) makes use of conceptual framing and metonymic inferencing to explain the effect of the modal verb on the performative verb. Depraetere & Kaltenböck (Reference Depraetere and Kaltenböck2023) approach the topic of HPs from a different perspective. They are interested in the discursive functions of HPs in spoken English. First, they show that HPs are typically used in spoken English based on an analysis of I must HPs in the Corpus of Contemporary American English (COCA; Davies Reference Davies2008–), where it has a frequency of 21.19 pmw, against frequencies ranging from 2.83 pmw in the Academic subcorpus to 14.30 pmw in the Fiction subcorpus (Depraetere & Kaltenböck Reference Depraetere and Kaltenböck2023: 212).Footnote 2 Therefore, rather than looking at the HP-internal illocutionary force mechanism, they examine how HPs are exploited in spoken discourse. They argue that HPs can serve two main functions, that of downtoning and emphasizing the propositional content of the host clause (see examples (4) and (5) in section 2); they show how HPs can be used to preserve, give a boost to or challenge the speaker’s or hearer’s face. In other words, HPs do not only work as a hedging device in discourse.Footnote 3 Latouche (Reference Latouche2024) zooms in on the discourse marker uses of I must say and I have to say in spoken conversational British English (Spoken BNC2014). She shows that, through the process of grammaticalization, these HPs can serve functions such as floor-taking and introduction of a new discourse topic.
Interestingly, the diachrony of modals has been extensively addressed (to quote just a few publications, Myhill (Reference Myhill1996), Leech (Reference Leech, Facchinetti, Krug and Palmer2003), Leech et al. (Reference Leech, Hundt, Mair and Smith2009), Leech & Mair (Reference Mair, Leech, Aarts, McMahon and Hinrichs2021), Mair (Reference Mair and Collins2015), Daugs (Reference Daugs, Hiltunen, McVeigh and Säily2017), Kranich (Reference Kranich, Kranich and Breban2021)). While it is important to bear in mind that genre and/or the corpus used impacts on the trends observed (see e.g. Leech’s (Reference Leech2011) response to Millar’s (Reference Millar2009) critique of Leech (Reference Leech, Facchinetti, Krug and Palmer2003), in which he defends the view that the modals are overall declining in use, and Daugs (Reference Daugs, Hiltunen, McVeigh and Säily2017)), the relative change (i.e. percentage change from the 1900s to the 2000s) in the Corpus of Historical American English (COHA; Davies Reference Davies2010–) reported by Daugs (Reference Daugs, Hiltunen, McVeigh and Säily2017) is as follows: shall declines by 94.87 percent, may by 68.36 percent, must by 64.96 percent, should by 55.39 percent, might by 27.09 percent, will by 15.13 percent. The frequency of can increases by 23.15 percent, would by 19.6 percent, could by 49.37 percent. Interestingly, lexical verbs show a far more drastic change, with have to and need to increasing by 709.49 percent and 2,728.18 percent respectively. While Leech’s and Mair’s studies concern two closer data points (1961 vs. 1991–2), they report roughly similar trends for written American English data based on the Brown versus Frown corpora, but with a few differences. Regarding the three verbs that are of interest to us, they note a decrease by 34.4 percent for must, but more stable trends for have to, which increases by a mere 1.1 percent, and for can, which decreases by a small 1.5 percent.
The above discussions about the diachrony of modals have been enriched by research within Construction Grammar, such as Hilpert (Reference Hilpert2013, Reference Hilpert2016), Cappelle & Depraetere (Reference Cappelle and Depraetere2016), Cappelle, Depraetere & Lesuisse (Reference Cappelle, Depraetere and Lesuisse2019). Hilpert has argued that there are modal auxiliary constructions and that, accordingly, diachronic change may manifest itself as well in ‘changes that affect only the relative strength of connections in the constructional network’ (2016: 17). Hilpert (Reference Hilpert2013) analyzes the shifting collocates of can, could, may, might, must, shall, should, will and would, with a more in-depth analysis of may being offered in Hilpert (Reference Hilpert2016). Cappelle & Depraetere (Reference Cappelle and Depraetere2016) and Cappelle, Depraetere & Lesuisse (Reference Cappelle, Depraetere and Lesuisse2019) look for highly collocating and high-frequency n-grams up to five words containing have to, must, need to and should in the British National Corpus (BNC) to identify constructions, and hedged performatives feature in the list of modal constructions that they identify (Cappelle & Depraetere Reference Cappelle and Depraetere2016: 93; Cappelle, Depraetere & Lesuisse Reference Cappelle, Depraetere and Lesuisse2019: 232–3). Daugs (Reference Daugs, Hilpert, Cappelle and Depraetere2021) argues that modal contractions are constructions, and not simply variants of their uncontracted forms. He describes the trends in the development of ‘can’t V’, ‘won’t V’ and the enclitic (would) ‘SUBJ’d V’ and analyzes changes in the collostructional behavior of the contracted versus non-contracted forms. HPs have not been subjected to similar, detailed diachronic or constructionist analyses. This observation notwithstanding, a few corpus studies on modality have made observations that pertain to the ways in which HPs have developed, without necessarily using the label ‘HP’ to refer to the expression examined. Kranich (Reference Kranich, Kranich and Breban2021) is one of them. Her paper is more generally concerned with diachronic change in the use of must and may between 1960 and 2000–9. One of the constructions she zooms in on is ‘we + may + verb of speaking/thinking/reasoning’ (2021: 277–82), which she argues is on its way to be lost. More specifically, she investigates changes in the frequency of use of we may combined with one of 47 selected verbs of speaking/thinking/reasoning, retrieved from COHA. Close & Aarts (Reference Close, Aarts, Lenker, Huber and Mailhamme2010) found that in the Diachronic Corpus of Present-Day Spoken English of British English (DCPSE, late 1960s until early 1990s), the use of the modal must has declined by 55 percent, and the use of I must x (which they categorize as ‘performative modality’) by a very similar 56 percent. Tagliamonte & D’Arcy (Reference Tagliamonte and D’Arcy2007) offer a sociolinguistic analysis of the use of necessity modals in Canadian English (Toronto) in age groups ranging from 7 to over 60. They show that while have to is mainly used to express deontic modality, must is ‘nearly categorically’ used in ‘formulaic expressions’ (2007: 73) like I must say, admit, confess, again, without actually mentioning the concept of HP. Smith & Busse (Reference Smith and Busse2019), a conference presentation, is the only study that has addressed the diachrony of HPs under that heading; they report on changes in the use of ‘obligational’ (I must/have to/should/(’ve) got to/ta/ought to/need to) HPs in spoken COCA. They show that there is a shift away from I must + speech act verb towards I have to, and that there is only a minor increase in I need to + speech act verb compared to the overall expansion of need to in COCA. They also observe that admit, ask, say and tell are the most common speech-act verbs, and that there is considerable frequency variation among these verbs.
The above overview shows that even though the ‘hedged performative’ has gained solid ground as a concept, detailed empirical analyses remain limited. Their internal speech-act mechanism and their discourse functions have been addressed. Construction Grammar has demonstrated that modal constructions exist, and while isolated cases of HPs are noted (though not always explicitly labeled as ‘HPs’ or as ‘constructions’) in both CxG and broader corpus research on the development of modal verbs, a comprehensive diachronic CxG analysis is still absent. The current article will address this gap by investigating the distribution of HPs in unscripted and scripted spoken American English and examining recent change in their frequency of use. Our analysis will be informed by data from COCA, the TV Corpus (Davies Reference Davies2019a) and the Movie Corpus (Davies Reference Davies2019b) and will zoom in on three modal verbs (must, have to and can) and a subset of performative verbs that are strong collocates in these HP contexts (I must x, I have to x and I can x).
2. HPs as a constructional network
In this study, we adopt a Construction Grammar (CxG) approach (Goldberg Reference Goldberg2006; Hilpert Reference Hilpert2019) to investigate HPs. According to CxG, language is made up of stored form–meaning pairings called ‘constructions’, which vary from fully lexical micro-constructions to fully schematic macro-constructions, the latter being generalizations over lower-level semi-lexical constructions or fully lexical instantiations. Construction status can be claimed on two grounds: (i) idiosyncratic form or function and/or (ii) sufficient frequency. We argue that HPs qualify as constructions on both grounds, and further make a case for treating HPs as a constructional network at different levels of abstraction.
Firstly, as pointed out above, Depraetere & Kaltenböck (Reference Depraetere and Kaltenböck2023) show that HPs serve specific functions in discourse, working either as downtoners or as emphasizers of the propositional content of the host clause. This is illustrated by examples (4) and (5):
In (4), the HP functions as a downtoner to hedge the disagreement expressed by that’s debatable, while I must say in (5) is an emphasizer that underscores the (positive) state of affairs in the host clause (I was intellectually ready for it). In that way, they are a means for speakers to preserve, boost or potentially also challenge their own or the addressee’s face. In other words, HPs demonstrate specific discursive functions that go beyond the compositional meaning of the individual lexical items.
Secondly, collocational patterns further substantiate their constructional status on frequency grounds. Modals, in particular in the string ‘I + MODAL’, collocate strongly with performative verbs. Table 1 shows the ten most frequent collocates for I must in COCA, and the strongest ones are all performative verbs (in bold) that consistently yield Mutual Information scores above 3 (i.e. the threshold value for strong collocates; see Hunston Reference Hunston2002). Note that similar results hold for can and have to.
Table 1. List of the ten most frequent verb collocates of I must in COCA, search string [I must _v], listed in descending MI score order, performative verbs in bold
The collocational behavior of HPs also extends to the subject pronoun. As can be seen in table 2 for must and say, the first-person singular pronoun I is not only the most frequent subject collocate, it is also the only one that has collocational status. For all HPs investigated in this article, I is consistently a strong subject collocate, and the strongest one for 11 out of 15 HPs. This is in line with the standard definition of a performative: a performative verb is typically used in the present tense and with a first-person singular subject (Austin Reference Austin1962: 56). As a performative verb is a verb that makes explicit an illocutionary force, and as a performative realizes the illocutionary act that it refers to, it is indeed only in a context with a first-person singular that the utterance of a performative verb itself has the capacity to make the corresponding action happen.
Table 2. List of the ten most frequent pronoun collocates of must say in COCA, search string [PRON must say]
The consistent collocational behavior of the fully lexical instances of HPs indicates their entrenchment in speakers’ mental lexicons and reinforces the view of HPs as constructions, at least at the lexical level.
From a usage-based perspective, the evidence presented above suggests that HPs form a constructional network at multiple levels of abstraction. Given (i) the broadly unified face-strategy discourse function(s) served by HPs, and (ii) the assumption that each modal contributes distinct, idiosyncratic meaning within HP contexts, we propose that speakers abstract away from these entrenched micro-constructions at two additional, more general levels. As a result, we posit a constructional network at three levels of abstraction:
-
(i) At the macro-level, we find the overarching abstract schema [I + MODAL + Vperf].Footnote 4
-
(ii) At the meso-level, modal-specific constructions that instantiate the general schema by specifying the modal verb, such as [I must Vperf] or [I can Vperf].
-
(iii) At the micro-level, we find fixed lexicalized strings such as [I must say], [I have to admit], or [I can assure].
Figure 1 represents part of this constructional network, with three modal-specific meso-constructions, each instantiated by five micro-constructions. This subset of the larger HP constructional network also forms the focus of our diachronic investigation in this article (see section 3 for the rationale behind the choice of specific modals and performative verbs). As will be made clear in the following section, this hierarchical network of constructions also provides an analytical framework to assess diachronic development at different levels of abstraction.
Figure 1. The HP constructional network under investigation
3. Aims and research questions
The aim of this article is to provide a diachronic analysis of the HP constructional network sketched out above in spoken American English. As pointed out in section 1, extensive studies exist on diachronic changes in modal verbs, but the corresponding diachronic behavior of HPs remains underexplored, while we have reason to believe that HPs may behave differently from the modal verbs on which they are formed. Building on the constructionist perspective outlined in section 2, we seek to systematically investigate how the different levels within the HP constructional network have evolved over time through four research questions:
-
• RQ1. Do we observe a uniform trend at the meso-level, which would be indicative of a general macro-trend, or do the different meso-constructions show divergent trends?
-
• RQ2. At the meso-level, do the modal-specific constructions mirror the frequency changes of their respective modal verbs, or do they exhibit distinct trajectories over time?
-
• RQ3. At the micro-level, do the individual micro-constructions follow the trends of their respective meso-level constructions, or is there evidence of idiosyncratic variation related to individual performative verbs or their illocutionary categories?
-
• RQ4. Overall, pertaining to register, do we find the same diachronic trends in scripted and unscripted speech?
4. Data and methods
4.1 The corpora: scripted and unscripted spoken data
To investigate the diachronics of HPs in spoken American English, we rely on data from three corpora: the spoken subpart of COCA, the TV Corpus, and the Movie Corpus. The spoken section of COCA is made up of 127 million words of largely unscripted conversations from TV and radio programmes, spanning from 1990 to 2019. The American subsection of the TV Corpus contains 265 million words of scripts from TV series, with data ranging from 1950 to 2019. Similarly, the American subsection of the Movie Corpus consists of 153 million words of scripted movie dialogues, ranging from 1930 to 2019. To ensure a consistent timespan across the TV and Movie datasets, we restricted the Movie Corpus data to the period from 1950 onwards.
While one may object to the use of scripted data to represent spoken English, the TV and Movie corpora have in fact been shown to be good proxies for spoken English. Davies (Reference Davies2019a, Reference Davies2019b), based on studies such as Brysbaert & New (Reference Brysbaert and New2009), Van Heuven et al. (Reference Van Heuven, Mandera, Keuleers and Brysbaert2014), and Brysbaert et al. (Reference Brysbaert, Keuleers and New2011), notes that, despite being scripted, the language in TV shows and movies aligns more closely with native speaker intuitions than transcripts of everyday conversation. Forchini (Reference Forchini, Friginal and Hardy2020) further substantiates this through a multidimensional analysis that compares movie dialogue to spoken English, concluding that movie dialogue is highly representative of face-to-face conversation. Jucker (Reference Jucker2021) adds nuance to these findings by comparing two types of markers of orality, namely contractions and inserts (i.e. response forms, discourse markers, hesitators and interjections), between the Santa Barbara Corpus (SBC), which represents spontaneous speech, and the TV and Movie corpora. He finds that while contractions are overrepresented in the TV and Movie corpora, inserts are more frequent in spontaneous speech.
All in all, while the TV and Movie corpora may not be ‘a perfect substitute’ (Davies Reference Davies2021: 34) for actual spoken language, they do appear to be reliable approximations that can provide valuable insights into spoken language. That said, as our analysis also includes data from COCA, we will also consider the potential register difference between scripted and unscripted data.
4.2. The choice of HPs
The focus of our study lies on a selection of HPs, which are defined by Fraser as performative sentences that contain a modal or a semi-modal, that is, ‘I + (semi-)modal + performative verb’. A large set of modal expressions can be used in HPs, including core modals such as can, could, might, shall, should, will, would and must, ‘semi-modals’ such as have to, as well as periphrastic forms like be able to and verbs such as wish, want to or intend to (Fraser Reference Fraser, Cole and Morgan1975). In this study, we focus on must, have to, and can. The primary motivation for this selection is their distinct diachronic trajectories: taking Daugs’ (Reference Daugs, Hiltunen, McVeigh and Säily2017) study on the 1900–2000 time span, and Leech & Mair’s (Reference Mair, Leech, Aarts, McMahon and Hinrichs2021) study on the 1960s–90s time span as a reference point (see section 1), must has been declining in use while have to shows a dramatic increase in Daugs (Reference Daugs, Hiltunen, McVeigh and Säily2017) but barely any change in Leech & Mair (Reference Mair, Leech, Aarts, McMahon and Hinrichs2021), while can shows a moderate increase over a century (Daugs Reference Daugs, Hiltunen, McVeigh and Säily2017) and a rather stable trend over the 1960s–90s time period (Leech & Mair Reference Mair, Leech, Aarts, McMahon and Hinrichs2021) (see section 1). This is of particular interest to us as we seek to verify whether the diachronic trend of the modal is reflected in the diachronic trend of the HPs in which they feature. Additionally, these three modals also represent modals of possibility (can) and necessity (must, have to), as well as core modals (must, can) and a semi-modal (have to), reflecting both semantic and syntactic diversity found among modal expressions (see, e.g., Depraetere & Langford Reference Depraetere and Langford2019).
In turn, our selection of performative verbs is based on frequency and collocational criteria. We first extracted all performative verbs occurring with I must, I have to and I can, with a minimum frequency of 30 and minimum MI score of 3 in the three corpora under investigation. These verbs are listed in table 3. Among those verbs, we then selected, for each modal, all verbs that occur in all three corpora. These are the verbs in bold, and this corresponds to 5 HPs per modal verb. Note also that the selection of verbs is exactly the same for must and have to, while only tell appears with all three modals. These data represent a total of 21,185 instances in the TV Corpus, 10,961 instances in the Movie Corpus, and 11,414 instances in COCA.Footnote 5
Table 3. Strongest performative verb collocates of I must, I have to and I can in the TV Corpus, Movie Corpus and COCA (minimal frequency 30, MI score higher than 3), listed in descending MI score order
4.3. Methods: regression modeling
Once our selection of HPs operated, we extracted all normalized frequencies (pmw) of each HP per decade and per corpus and set up two dataframes, serving to answer different research questions. The first dataframe, made up of 102 datapoints, serves RQ1 and RQ2 and seeks to verify whether HPs follow the same diachronic trend as the modals they are based on, records the normalized frequencies of each modal verb per decade, according to its overall use in the corpus, and its HP use. More specifically, it contains the following variables:
-
– LOGFQCY_CENTERED: the normalized frequency pmw, further (i) log-transformed to reduce skewness and normalize the distribution of frequency values, which are typically right-skewed in corpus data; and (ii) mean-centered to address multicollinearity between variables and interaction terms, which is common in models that include interactions with timeFootnote 6
-
– DECADE_CENTERED: represents the different decades as a numeric variable ranging from 1950 to 2010, which has also been mean-centered (i.e. centered around 1980, the mean of the dataset) to avoid issues of multicollinearity
-
– MODAL: represents the modal verb used in the HP (can, have to or must)
-
– USE: distinguishes between the modal used as HP or its overall use as a modal
-
– CORPUS: represents the corpus from which the data is drawn (COCA, TV, Movie)
The second dataframe, made up of 135 datapoints, focuses specifically on the HP uses to serve the subsequent RQs, and includes the following variables:
-
– LOGFQCY_CENTERED: the normalized frequency pmw logged and centered (see above)
-
– DECADE_CENTERED: numeric variable ranging from 1950 to 2010, centered (see above)
-
– MODAL: represents the modal verb used in the HP (can, have to or must)
-
– PV: represents the performative verb used in the HP (e.g. tell, assure, ask, promise)
-
– CORPUS: represents the corpus from which the data is drawn (COCA, TV, Movie)
-
– REGISTER: distinguishes between the scripted (TV and Movie) and unscripted data (COCA)
Several mixed-effects models were fitted in RFootnote 7 to answer the different research questions, each with the frequency variable as the dependent variable and the theoretically relevant independent variables, as will be detailed in the results section. One particularity of our data is that we have unbalanced datasets in terms of the decades represented, as COCA only spans data from the 1990s to 2010s, while the TV and Movie corpora span from the 1950s to 2010s. Since our primary focus is on the overall diachronic trends, the models for the first three research questions include CORPUS as a random effect in order to (i) account for idiosyncratic corpus-specific variation, and (ii) avoid skewing trends due to the imbalance in temporal representation.Footnote 8 Each model was assessed by checking the normal distribution of the residuals by means of a qqplot and a Shapiro-test for normality, as well as for multicollinearity by means of vif scores. The residuals of all models are normally distributed (p > 0.05 on the Shapiro tests), and all vif values fall below 2.5, which indicates that there is no collinearity between variables.
5. Results
5.1. Macro- and meso-level diachronic patterns: modals versus hedged performatives
The first step in our analysis is to sketch the diachronic trend of the meso- level HP constructions over time (RQ1), with a particular focus on whether HP meso-level constructions follow the same diachronic trend as their base modals (RQ2). Based on our first dataframe, we thus fitted the mixed-effects regression model in (6), which seeks to predict frequency as a response to the two- and three-way interactions between the decade, the use of the modal (modal use versus HP use), and the modal itself.
The output of the model is provided in Appendix 1 (supplementary material online), and reveals a statistically significant three-way interaction between DECADE, USE and MODAL, which already suggests that there is no unified trend at the macro-level HP construction, and that the diachronic trends of the meso-level HP constructions diverge from those of the modals they are formed with.Footnote 9 The effects plot in figure 2 visualizes this interaction in a way that makes the results straightforward to interpret.
Figure 2. Effects plot of frequency of modal versus HP use over time with 95 percent confidence bands
Backed up with the significance value of the regression model and a post-hoc pairwise comparison of the slopes between HP and modal uses for each modal, conducted with the emtrends() function from the emmeans package in R, we observe that:
-
(i) have to shows a stable frequency over time as a modal, but increases significantly over time as a HP (p = 0.0002).
-
(ii) must exhibits a significant decline over time, both as a modal and a HP, although the decline is significantly steeper as a HP than as a modal (p = 0.0027).
-
(iii) can remains stable in frequency over time, both as a modal and a HP, and there is no difference in diachronic trend between the two uses (p = 0.1436).
In other words, these results are in line with previously reported diachronic trends for modals in American English (see section 1), and especially with the trends reported by Mair & Leech (Reference Mair, Leech, Aarts, McMahon and Hinrichs2021) regarding the decline of must and the rather stable use of can and have to. More crucially for us, however, the HP uses of must and have to exhibit distinct trends from their base modal.Footnote 10 From a Construction Grammar (CxG) perspective, this idiosyncratic behavior further supports the view that HPs are not merely compositional expressions, but entrenched semi-lexicalized templates that behave independently of their base modal, most likely as a result of the specific pragmatic discourse functions that they serve. A second observation, perhaps not particularly surprising, is that we see diachronic differences at the meso-level between the different meso-constructions ([I must Vperf], [I have to Vperf], [I can Vperf]), which also shows that while the macro-construction [I + MODAL + Vperf] captures the general HP pattern, each modal-specific construction evolves idiosyncratically, reinforcing the existence of meso-level constructions.
While the idiosyncratic behavior of HP meso-constructions compared to their base modal is in line with some previously reported results, they do not entirely align with reported observations either. As noted in section 1, Tagliamonte & D’Arcy (Reference Tagliamonte and D’Arcy2007) report that must persists in Canadian English primarily in formulaic constructions such as I must say – that is, in HPs. By contrast, in our American data, HPs with must decline, and even more strongly than the modal use of must. We believe that this discrepancy can nonetheless be reconciled with our results when we consider the concurrent rise of the have to HP. Given that both modals express obligation, we can reasonably hypothesize a partial functional replacement of the I must HP by its I have to HP counterpart. This can in turn be motivated by the broader trend noted in modal use, namely the shift away from ‘authoritarian’ uses (must) towards less authoritarian forms (have to). Smith (Reference Smith, Facchinetti, Krug and Palmer2003: 259) puts it as follows:
must is a casualty of a changing society where increasing emphasis is being placed on equality of power, and the informality of discourse found in private conversation is becoming more acceptable, even usual, in official types of discourse. … Just as these conditions are likely to disfavour the use of must, they should correspondingly favour forms which express obligation less directly.
Thus, rather than entirely undermining the idea of conservation of HPs noted by Tagliamonte & D’Arcy (Reference Tagliamonte and D’Arcy2007), our data may suggest that such entrenched templates that serve specific discourse functions are shifting in degree of use, and thus degree of entrenchment, in response to changing discourse norms. In other words, what we observe is a restructuring of the constructional network: it seems that the meso-construction [I have to Vperf] becomes increasingly entrenched at the expense of the [I must Vperf] construction to respond to a more general discourse trend that shifts towards less authoritarian modality in spoken English.
Nonetheless, without considering the constructions at the micro-level, we do not know whether the trend at the meso-level captures a uniform trend at a lower level, or whether it is the result of idiosyncratic lower-level behavior. This is what the next section will seek to address.
5.2. Meso-level versus micro-level diachronic patterns
In order to assess whether the micro-level constructions, i.e. performative verb-specific constructions, follow the same diachronic trend as their corresponding meso-level construction (RQ3), we fitted three mixed-effects regression models, one for each modal, as represented in (7). Each model seeks to predict frequency as a response to the interaction between the decade and the performative verb (PV).
The output of all three models is provided in Appendix 2 (supplementary material online), and reveals for all three modals a statistically significant interaction between DECADE and PV, which already indicates that there is some degree of individual lexical behavior over time at the micro-level constructions. To explore these lexical idiosyncrasies, the results have been plotted by means of two complementary plots for each modal. The left-hand plot is the effects plot that shows the diachronic trend of each micro-construction, with each colored trend line representing a different micro-construction. The dashed black trendline indicates the overall trend of the meso-construction. The 95 percent confidence bands around the trendlines have been omitted for the legibility of the plot. Instead, the plot on the right represents, for each micro-construction, the slope of the corresponding trendline in the left plot with its 95 percent confidence interval. The slopes were extracted using the emtrends function from the emmeans package in R, which computes slopes for each PV based on the full model, including the interaction terms. That is, the plot represents how steeply each micro-construction changes in frequency over time. The 95 percent confidence intervals provide a visualization (i) of the potential differences between micro- constructions, and (ii) of how each compares to the overall meso-level trend (represented by the dashed line with its shaded confidence band). To properly test whether the differences in slopes are statistically significant, we conducted pairwise post-hoc comparisons using the emtrends function from the emmeans package in R. Figures 3, 4 and 5 respectively represent the plots for the [I must Vperf], [I have to Vperf] and [I can Vperf] micro-constructions.
Figure 3. Diachronic trendlines and slopes of [I must Vperf] micro-constructions
Figure 4. Diachronic trendlines and slopes of [I have to Vperf] micro-constructions
Figure 5. Diachronic trendlines and slopes of [I can Vperf] micro-constructions
When considering all three figures, the first general observation at the micro-level is that the lexical micro-level idiosyncrasies are, at least to some extent, modal-specific: While for must and have to, all micro-constructions follow the same downwards and upwards trend as the meso-level construction, the can micro-constructions show non-null diachronic trends, contrary to the overall stable [I can Vperf] trend, with [I can guarantee] and [I can confirm] going upwards and [I can promise], [I can assure] and [I can tell] going downwards. Regarding [I must Vperf] and [I have to Vperf], their micro-constructions seem to exhibit broadly complementary patterns: while the slopes are not exact mirrors of one another, there does seem to be a general shift suggesting a process of replacement. This is in fact particularly clear for say, which shows the strongest decline with must and the strongest rise with have to. Only tell is an exception to this trend, as it decreases steeply with must but without a significant uptick with have to. Footnote 11 Tell aside, it seems that the hypothesis put forward above regarding the replacement of I must HPs by I have to HPs, motivated by a discursive shift towards less authoritarian modal expressions, is corroborated by the trends observed at the lexical micro-level. In other words, the diachronic trajectories of must and have to HPs seem to be primarily driven by discursive constraints introduced by the modal verb itself, thus operating mainly at the meso-level. By contrast, can micro-level constructions display diverging trends, which rather suggests that changes of can HPs are driven primarily by individual lexical factors rather than by a unified meso-level discursive or functional shift. From a Construction Grammar perspective, these findings imply a nuanced constructional network organization and re-organization, where diachronic trajectories may be dictated by constraints operating at different levels in the constructional network.
Secondly, while the micro-constructions with must and have to show broadly similar trends, it is worth examining the lexical patterns more closely. It was argued in the previous section that must conveys a more authoritarian modality than have to, which has been put forward as an explanation for their opposite meso-level diachronic trends. The degree of authoritarian modality can, however, also vary at the micro-constructional level, influenced by the semantics of the PV. Fraser (Reference Fraser, Cole and Morgan1975: 191) classifies the PVs admit and confess as ‘asserting II’ verbs, which ‘place certain restrictions on successful performance of the act’ while he classifies the PVs say and tell as ‘asserting I’ verbs, which require ‘few, if any, conditions on the successful performance of the associated illocutionary act’. Further ‘asserting II’ examples are verbs like concede, which ‘entails that the speaker previously refused to assert the proposition’, and accuse, which ‘entails that the act specified has some pejorative sense associated with it’ (Reference Fraser, Cole and Morgan1975: 191). There is also a negative shade of meaning associated with admit and confess, as admit suggests that there is pressure to say something that one would perhaps be reluctant to say (see, e.g., Brinton Reference Brinton2017: 168–71) and in a similar way, confess involves the disclosure of something that is self-damaging. The difference between the ‘asserting I’ nature of say and tell versus the ‘asserting II’ nature of admit and confess can also be mapped onto differences in degree of authoritarian modality: admit and confess suggest a more hierarchical, and thus more authoritarian, relationship between speakers than say and tell, which are more neutral in that regard. Thus, if there is a trend towards ‘democratization’, we would expect admit and confess to decline less steeply than say and tell with must precisely because their shared authoritarian nature would lead them to persist together over time. By contrast, we expect admit and confess to increase less steeply with have to than say and tell, given have to’s less authoritarian profile. These predictions are partially corroborated by our data. With must, admit and confess indeed decline significantly less than say (p < 0.05 for both slopes). Admit and confess also seem to decline less than tell, since the confidence intervals don’t overlap in figure 3, but a pairwise comparison post-hoc test reveals that the difference falls just above significance (p = 0.0686 for confess vs. tell, and p = 0.1457 for admit vs. tell), suggesting a possible trend that does not reach statistical significance but that does suggest a similar pattern as that with say. The picture is less clear with have to: in line with our expectation, confess increases less steeply than say (p = 0.0067), but admit follows a similar upward trend as say (p = 0.7203), and tell does not show the expected steeper rise. One may thus wonder why [I have to admit] does not show a weaker upward trend, and why [I have to tell] does not show a stronger upward trend. Here, further research relying on collostructional analyses over a wider range of PVs could help to better understand and capture the effect of such low-level lexico-semantic constraints on diachronic trends.
5.3. Hedged performatives in scripted and unscripted spoken American English
While our data has so far been pooled together both scripted and unscripted spoken data to consider overall diachronic trends based on the assumption that the TV and Movie corpora are good proxies for spoken interactions (see section 3), we may however wonder whether there are register-specific trends, i.e. whether scripted and unscripted American spoken English pattern in the same way (RQ4). As COCA, which represents unscripted speech, spans only three decades, this analysis is based on the 1990s–2010s data only for all three corpora. To check whether there is a difference between the two registers, we ran the simple linear model in (8), which seeks to predict frequency as an interaction between decade, register and modal.Footnote 12
The output of this model is provided in Appendix 3 (supplementary material online). It shows that the only significant factor is the main effect of REGISTER, revealing that there is a significant difference between the two registers, which is however constant across time and modals. As this is a diachronic study where clear diachronic differences have been found between modals, the decade and modal variables have been kept in the model so as to visualize the data over time and per modal. Figure 6 thus plots the predicted trend for each register over time and per decade, bearing in mind that the two- and three-way interactions between decade, register and modal are not significant.
Figure 6. Diachronic trendlines of HPs per modal and per register with 95 percent confidence bands
A first thing to note is the large 95 percent confidence bands, which is in all likelihood the result of having far fewer data points once we (i) work with only three decades and (ii) separate the two registers. While the results are no longer significant for the considered timespan, the trends for each modal are entirely in line with those reported in section 4.2, with a relatively stable trend for can, a rise for have to and a decline for must, across both registers. Turning to our question of interest, what this also shows (and which is significant as a main effect), is that HPs are consistently more frequent in the scripted than in the unscripted register. We also considered what happens at the micro-level across registers, even if only by means of descriptive trends, to probe whether individual micro-constructions exhibit divergent diachronic trends. This is represented by the descriptive plot in figure 7, where each color represents a different micro-level HP construction, and the dashed line represents their normalized frequency pmw in scripted speech, while the solid line represents their normalized frequency pmw in unscripted speech.
Figure 7. Diachronic trendlines of micro-level HP constructions per register
At this level too, the scripted micro-level HP constructions are systematically more frequent than their unscripted counterparts, with similar diachronic trajectories. Thus, the trend of scripted HPs being more frequent than unscripted HPs appears to hold across the board at all three levels of analysis. This suggests that script writers have a biased perception of the frequency of use of HPs and overestimate how frequently they are actually used in spontaneous discourse. This is partially in line with the findings of Jucker (Reference Jucker2021), who observed a higher frequency of use of a specific marker of orality, i.e. contractions, in the scripted corpora, even though another feature (the use of inserts) occurred more often in spontaneous speech. It does, however, seem that regarding diachronic trends, scripted data are a good proxy to study the evolution of spoken language over time.
6. Conclusion
In this article, we examined the diachronics of hedged performatives (HPs) in spoken American English over the period 1950–2019, framed in a Constructionist approach. We more specifically studied the HP constructional network based on a selection of three meso-level HP constructions, [I have to Vperf], [I must Vperf] and [I can Vperf], each subsuming five PV-specific micro-constructions. By adopting a Construction Grammar (CxG) framework, we were able to carry out a systematic analysis at each level of abstraction, which allowed for nuanced insights into the levels at which language change seems to occur.
Returning to the four research questions that guided this study, the first finding of our study is that each meso-level construction displays distinctive diachronic trends, indicating that there is no unified pattern at the macro-level HP construction (RQ1), but that the constructional network is reshaped at more lexical levels of abstraction. Secondly, when considering the meso-level constructions, the [I have to Vperf] and [I must Vperf] constructions exhibit frequency trajectories over time that differ significantly from the frequency trajectories of their base modals, have to and must (RQ2). This idiosyncratic behavior substantiates their status as constructions rather than merely compositional patterns.
When considering the micro-level, must and have to HPs present consistent trends across their micro-level instantiations, albeit with varying degrees of steepness, while the can micro-level HPs exhibit diverging tendencies, with some increasing and others declining over time (RQ3). Thus, our analysis reveals that the reorganization within the constructional network occurs at different levels depending on the modal involved: without precluding idiosyncratic change at the micro-level, some change appears to be driven at the meso-level for must and have to, while for can, the evidence points towards change occurring at the micro-level without a unified meso-level trend.
Regarding [I have to Vperf] and [I must Vperf], we interpret the opposite trends that they exhibit as [I have to Vperf] gradually replacing [I must Vperf] in response to broader changes in discourse norms, notably a shift toward speech acts characterized by weaker authoritarian tone (Myhill Reference Myhill1996; Smith Reference Smith, Facchinetti, Krug and Palmer2003; Close & Aarts Reference Close, Aarts, Lenker, Huber and Mailhamme2010). By contrast, at the micro-level, the more authoritarian HPs [I must admit] and [I must confess] decline less than [I must tell] and [I must say] given that both the modal and PV more explicitly convey a hierarchical relationship where one party has authority over the other, and thus persist together longer over time.
Finally, we also considered whether register was an impacting factor when considering the use of HPs by comparing scripted versus unscripted speech over the 1990s–2010s period. This analysis reveals that while diachronic trends are largely consistent across registers, HPs are notably overrepresented in scripted speech (RQ4). This suggests that, although scripted speech, i.e. television and movie dialogues, inflate the frequency of HPs, it still serves as a reliable proxy for to study for diachronic linguistic studies.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S136067432510035X
Acknowledgments
We are grateful to Gunther Kaltenböck for his valuable feedback on an earlier draft of this article, as well as to the reviewers for their insightful comments.
Open access
