1. Introduction
Specialisation is a specific outcome of competition between two or more functionally equivalent expressions, where one expression is clearly preferred over the other in a specific linguistic context. The phenomenon has been investigated in variationist studies (e.g. Tagliamonte Reference Tagliamonte, Lindquist and Mair2004; Torres Cacoullos & Walker Reference Torres Cacoullos and Walker2009; Denis & Tagliamonte Reference Denis and Tagliamonte2017), in grammaticalisation studies (e.g. Hopper Reference Hopper, Heine and Traugott1991) and in Diachronic Construction Grammar (e.g. Hilpert Reference Hilpert2012; De Smet et al. Reference De Smet, D’hoedt, Fonteyn and van Goethem2018). The competition between will and be going to is especially often investigated: Denis & Tagliamonte (Reference Denis and Tagliamonte2017) look at the competition between will and be going to for the expression of future reference in North American English and find that be going to specialises for interrogatives and protasis clauses, while will specialises for apodosis clauses. Hilpert (Reference Hilpert2012), using distinctive collexeme analysis, further finds that in British English be going to specialises for collocation with say.
One problem with these studies is that they underestimate more fine-grained functional differences between will and be going to. While both expressions have future reference, they differ in the degree to which they express different futural (intention, prediction) as well as non-futural meanings. Will is much more frequent with ‘prediction’ than with intention (Nesselhauf Reference Nesselhauf2012) and additionally occurs with meanings that be going to is unavailable with, e.g. willingness, volition and dynamic modality. The above-mentioned studies do not differentiate between these different meanings when attesting specialisation. This is problematic because what might seem to be a specialisation for a specific context might actually be due to functional differences.
Recent studies are more sensitive to such fine-grained differences in meaning. Leclercq & Depraetere (Reference Leclercq and Depraetere2022), for instance, compare be able to, can and could in terms of different meanings they express (epistemic, ability, opportunity etc.). They find that be able to specialises for the expression of what they call ‘actualised opportunity’ (e.g. We were able to see the Eiffel Tower). Hilpert & Flach (Reference Hilpert and Flach2021) find that collocational preferences relatively robustly predict the difference between deontic and epistemic uses of must and may.
There have also been attempts to offer motivations for specialisation. A motivation often invoked in grammaticalisation studies is retention of distributional preferences. Torres & Cacoullos (Reference Torres Cacoullos and Walker2009), for instance, argue that the fact that in present-day English be going to occurs more often than will with epistemic stance (I think) is because such contexts were needed initially to jump-start its grammaticalisation and are retained still. Another motivation is differences in connotation or construals: Palmer (Reference Palmer1990) argues that will always implies contingency, which explains its higher frequency in the apodosis of if-clauses. And lastly, statistical pre-emption has been proposed as a motivation, which means that speakers choose one expression over another in a specific linguistic context because it is more idiomatic (e.g. #Explain me this vs Explain this to me) (Goldberg Reference Goldberg2019).
In this article, I test the value and feasibility of the different motivations for specialisation: construal differences (henceforth construal pre-emption), statistical pre-emption and also a new motivation I propose, which I call polysemous pre-emption. I define polysemous pre-emption as the semasiological counterpart to statistical pre-emption that considers whether better entrenched polysemes of an expression may pre-empt the use of that expression with less entrenched polysemes in particular contexts of use. For a case study, I look at the competition between will and must with two specific meanings: the deontic ‘command’ meaning (e.g. You must listen to me!) and the epistemic ‘inference’ meaning (e.g. You must be tired). For each meaning, I identify collexemes that will or must specialise for (i.e. where one is much more frequent than the other) and test in how far each of the three types of pre-emption may account for specialisation.
The article is structured as followed: section 2 gives an overview of the three motivations for specialisation I want to explore (statistical, construal and polysemous pre-emption). Section 3 gives an overview of the semantics of will and must and specifically attests the extent to which they compete for expression of the meanings ‘command’ and ‘inference’. Section 4.1 introduces the methodology. Section 4.2 presents the results of the investigation. Section 5 discusses the implications of the results.
2. Motivations for specialisation
In this section, I give more in-depth characterisations of the three types of pre-emption I investigate in this article and their shortcomings: construal, statistical and polysemous pre- emption. I do not look at pre-emption that follows from retention of earlier distributional preferences, which was mentioned in the introduction, because this is mainly helpful when looking at syntactic contexts and in this article I mainly focus on specialisation for specific meanings and collexemes.
The first type of pre-emption is what I call construal pre-emption (or differences in construal and connotation). This motivation can be linked to the isomorphism principle which holds that differences in form should also signal differences in meaning (Goldberg Reference Goldberg1995: 67; Croft Reference Croft2001: 108ff.). In cognitive linguistics, especially, this is taken to mean that two forms, even when they seem to have the same communicative function, differ in how they construe this function. You’re welcome and No worries can both serve the communicative function of responding to thanks but imply different construals: You’re welcome foregrounds the appreciation of the addressee, while No worries foregrounds the minimisation of the favour (Brinton Reference Brinton2021). Thus, if either of the expressions specialises and starts to occur only in specific contexts or with specific speakers, this specialisation would most likely be motivated by the differences in construal.
One problem with construal pre-emption is that it cannot explain specialisation with competing expressions that genuinely seem to have the same coded meaning. These are sometimes called ‘alternations’ or ‘near-synonyms’. ‘Near synonymy’ often occurs with more schematic patterns and more abstract meanings such as the dative alternation (e.g. Perek Reference Perek2012), particle placement alternation (Cappelle Reference Cappelle2006) and some modal expressions (Hilpert & Flach Reference Hilpert and Flach2021). Since alternations are assumed to be true functional equivalents, construal differences cannot account for specialisation in these cases, because the competing expressions do not imply a different construal.
The second type of pre-emption, statistical pre-emption, might offer a solution here. Goldberg (Reference Goldberg2019) sums up the idea nicely in the following quote:
Native speakers have learned a more conventional way to express the intended messages in context. We favor formulations that have become entrenched through previous exposure for expressing our intended messages, because we implicitly understand those formulations to represent the ‘right’ way to use our language. (Goldberg Reference Goldberg2019: 75)
Statistical pre-emption is an onomasiological notion (see also ‘onomasiological salience’, Geeraerts Reference Geeraerts and Schmid2017). It means that an expression is better entrenched with a specific meaning and a specific context than competing expressions with the same meaning and the same context. Specialisation in the dative-alternation is good example of statistical pre-emption: the ditransitive and to-dative can both be used to refer to some sort of transfer. While many verbs (give, tell) occur with both patterns, some verbs (e.g. explain) almost exclusively occur with the to-dative. Goldberg argues that the use of the ditransitive with explain (#Explain me something) is pre-empted by the better entrenchment of explain with the functionally equivalent to-dative expression (Explain something to me).
Statistical pre-emption is a useful concept for explaining specialisation when construal cannot (or cannot alone) account for it, but it is also circular to a certain extent, as it takes itself as evidence. It holds that an expression is selected over a competing expression, because this has always been the case in the speaker’s experience. While this might be true in some cases, there may be cases of specialisation that can also be explained by other entrenchment phenomena.
The third type of pre-emption I want to discuss, polysemous pre-emption, might, I suggest, be such an entrenchment phenomenon and could be seen as both an additional and alternative motivation for specialisation. Polysemous pre-emption can be seen as the semasiological equivalent to statistical pre-emption (see also Geeraerts Reference Geeraerts and Schmid2017 on ‘semasiological salience’). It means that an expression might be pre-empted from being used with a meaning (or polyseme) in a specific context because that expression is already very well entrenched in that context with another meaning (or polyseme). The speaker might thus opt to use a different expression for a given meaning.
Polysemous pre-emption has not been investigated as a motivation for specialisation, but it is related to other phenomena, specifically ambiguity resolution and ambiguity avoidance, which have been subject to some investigation.
Ambiguity resolution refers to how listeners process lexically ambiguous expressions. There are many models for ambiguity resolution, some arguing that different meanings are processed in parallel (e.g. Rodd, Johnsrude & Davis Reference Rodd, Johnsrude and Davis2010) and others arguing for the possibility of sequential processing (e.g. Giora Reference Giora2003). What both types of approaches agree on is that in the absence of contextual cues, listeners will always resolve lexical ambiguities by selecting the more frequent of the two meanings (Rodd Reference Rodd, Rueschemeyer and Gaskell2018). For instance, if one hears the sentence John broke my record without any context, it is predicted that listeners will select the more frequent meaning (probably record ‘high score’), rather than the less frequent meaning (record ‘CD’).
Ambiguity avoidance is a controversial phenomenon; its proponents hold that speakers avoid using expressions that are ambiguous if they can and instead opt for non-ambiguous expressions. While evidence for syntactic ambiguity avoidance is mixed (Rohdenburg Reference Rohdenburg2021; Zehentner Reference Zehentner2022), there is some clearer evidence for lexical ambiguity avoidance (Ferreira & Dell Reference Ferreira and Dell2000). An ambiguity avoidance account would predict that speakers would avoid saying Mary broke my record and instead say Mary broke my CD to avoid potential ambiguities.
Polysemous pre-emption is similar to ambiguity avoidance in its effect, but conceptually they are distinct: ambiguity avoidance implies that speakers are aware of potential ambiguities and are actively trying to avoid them, regardless of how well entrenched the given meanings are. This may well be the case. Polysemous pre-emption, on the other hand, holds that speakers are prevented from accessing an ambiguous expression to express a specific meaning in a specific context when it is more strongly associated in that context with a different meaning. It would predict that in a situation where John has broken Mary’s CD, Mary would say John broke my CD rather than John broke my record, not merely because record is ambiguous in general, but because record, specifically in the context of breaking, is better entrenched with the meaning ‘high score’ than with the meaning ‘CD’.
Polysemous pre-emption is further motivated by assumptions about the organisation of linguistic knowledge in the brain. Cognitivist approaches argue that linguistic knowledge is stored in a network; forms and meanings are connected by associative links if they frequently co-occur, to the extent that forms and meanings might be stored in a single node in the network as constructions (e.g. Goldberg Reference Goldberg2006; Bybee Reference Bybee2010; Diessel Reference Diessel2019; Schmid Reference Schmid2020). Links and nodes that are frequently activated have a certain degree of entrenchment (Langacker Reference Langacker2013) or resting-level activation (Hudson Reference Hudson2010: 73), which means that they can be more easily retrieved. Online language processing is guided both by resting-level activation and also by activation patterns that emerge from the context of use. It is assumed that different meanings of an expression are stored together because of their association with the expression. If one of the meanings is especially frequent, it will be much better entrenched with the expression, more easily activated and thus better associated with it. The idea is that this very well-entrenched meaning could pre-empt the use of the expression with less entrenched meanings to avoid ambiguity.
3. Competition for the expression of ‘command’ and ‘inference’: will and must
To test the extent to which polysemous pre-emption would be an equally good explanation for specialisation as construal and statistical pre-emption, I will look at instances of specialisation in the case of will and must, with the meanings ‘command’ and ‘inference’ respectively.
Will and must, as most modal verbs, express a variety of different modal meanings (e.g. Visser Reference Visser1969; Palmer Reference Palmer1990):
-
• deontic meanings: ‘command’ (You must/will leave now), ‘obligation’ (Kids must respect their parents), ‘plan’ (The train will arrive soon),Footnote 1 ‘intention’ (We will go to the cinema later) and ‘willingness’ (Will you help me with this?);
-
• epistemic meanings: ‘inference’ (He must/will know about it) and ‘prediction’ (It will rain);
-
• dynamic meanings: ‘genericity’ (Oil will float on water), ‘circumstantial possibility’ (You will be seen).
In the following two sections, I will explore the degree of conceptual overlap between will and must with the meanings ‘command’ and ‘inference’ to determine whether they are similar enough in meaning to qualify for statistical pre-emption.
3.1. ‘Command’ will and must
‘Command’ is a deontic meaning that takes the speaker as the director or the ‘deontic source’ and demands an agent to do their bidding (Huddleston & Pullum et al. Reference Huddleston and Pullum2002: 183; Leech Reference Leech2004: 78; Collins Reference Collins2009: 134). Both must and will have been associated with this use.

In general, both imply a high degree of force on the part of the speaker, so that the addressee is not merely invited or requested to do something, but instead it is presumed that the speaker has the authority to force compliance (Palmer Reference Palmer1990: 73).
Some researchers argue that will and must contrast in terms of politeness, although there does not seem to be a consensus on which of the two is the politer one. Some argue that must is more polite than will (Coates Reference Coates1983: 183; Perkins Reference Perkins1983: 45). Others hold commands with will are quite mild, especially in the context of superiors giving directions to their servants; this is however more seen in contrast to the imperative, rather than in contrast to must, as is evident in (2a) (Bain Reference Bain1904: 169 in Visser Reference Visser1969: 1695ff.). Given instances such as (1b), Visser instead argues that will can express various degrees of ‘mildness’ or ‘sternness’ depending on the contexts (2b–c):

Although will can apparently express different degrees of force, there are some contexts in which will seems to imply too much force to be felicitous. These are contexts, in which must can occur to express a ‘passionate’ appeal (3) or else advice (4) (Visser Reference Visser1969: 1806f.)


Thus, although will might be infelicitous in some contexts because it can imply a higher degree of force, will and must are largely interchangeable when it comes to the expression of commands.
3.2. ‘Inference’ will and must
‘Inference’ is an epistemic meaning, where the speaker signals that a proposition is predictable from something else, either past experience or present evidence. Both will and must are associated with this meaning and express a high degree of certainty and are thus thought to be relatively interchangeable (Coates Reference Coates1983: 177; Huddleston & Pullum et al. Reference Huddleston and Pullum2002: 189; Leech Reference Leech2004: 86; Collins Reference Collins2009: 127).

Despite their similarities, they are supposed to differ in terms of the kind of evidence and the type of conclusion implied in the inference. It is argued that will often expresses more of an assumption or expectation that something is the case, usually on the grounds of previous knowledge. Must in contrast rather expresses ‘the only possible conclusion’ (Palmer Reference Palmer1990: 57) based on present evidence (Huddleston & Pullum et al. Reference Huddleston and Pullum2002: 189, Perkins Reference Perkins1983: 45). As such, will can be semantically weaker compared to must.

Palmer argues that only (6a) would be felicitous in the context of present evidence, while (6b) is infelicitous. This is because the lights being on seems to be such compelling evidence that it is unclear why the speaker would not express a higher commitment to their inference by using must.
While must cannot always be replaced by will, will usually can be replaced by must without much change in meaning, as can be seen in (7) (Leech Reference Leech2004: 86).

Thus, although must and will are not perfect equivalents in the expression of inference, they are sufficiently similar to compete in at least some contexts, especially given that will can usually be replaced by must.
4. Case study: ‘Command’ and ‘inference’ will and must
4.1. Data collection and analysis
To investigate cases of specialisation in ‘command’ and ‘inference’ will and must and possible forms of pre-emption I have carried out a quantitative corpus analysis using the final period of the Corpus of Late Modern English (CLMET 3.1; De Smet et al. Reference De Smet, Flach, Tyrkkö and Diller2015) (1850–1920) and only included the speech-based genres, novels and drama. The size of this subset is around 12 million words. I have opted for a historical rather than present-day corpus mainly because must is in decline in present-day spoken English in all of its uses because of competition with have to and got to (Close & Aarts Reference Close, Aarts, Lenker, Huber and Mailhammer2010). Further, directive speech acts and strong expressions of ‘command’ specifically seem to be in decline in present-day English, probably because present-day British and American societies are less hierarchically organised than those of the eighteenth to early twentieth century (Kohnen Reference Kohnen and Jucker2008: 303; Nesselhauf Reference Nesselhauf2012: 122).
The corpus was searched for instances of will and must in affirmative contexts with second-person subjects, using the following query: You {must, will} Verb. Looking at such a specific context means that contextual differences between will and must will be smaller and this increases the likelihood of finding instances where will and must are in genuine competition. I have particularly opted for second-person subjects because these are likely to occur both with the ‘command’ and ‘inference’ meaning. This search has yielded 2,031 instances. Instances that co-occurred with have were excluded as they were often used for modal perfects, leaving 1,935 instances.
In a first step, I carried out a first annotation of the semantics of the expressions, differentiating between ‘inference’, ‘command’ and ‘other’.
An instance was annotated as denoting ‘inference’ if it implied
-
(a) an evaluation by the speaker;
-
(b) the proposition making reference to present time (not the past or the future);
-
(c) and/or the proposition being predictable to some degree from the speaker’s knowledge.
An instance was annotated as denoting ‘command’ if it implied
-
(a) the speaker being the authority that demands that a future event will happen;
-
(b) the addressee being able/willing? to carry out the future event; and/or
-
(c) the addressee not being thought to have a say in the matter.
‘Other’ was assigned if neither ‘inference’ nor ‘command’ could be diagnosed. Table 1 gives an overview of the number of instances per meaning.
Table 1. Instances of will and must per meaning

In a second step, I carried out a meaning-specific distinctive collexeme analysis to determine instances of specialisation, i.e. where one expression occurs more often with a given verb than the other. Distinctive collexeme analysis identifies collocational differences between two near-synonymous expressions. By looking at the frequency with which each of the alternating expressions occurs with a collocate compared to their occurrence with other collocates, the expected frequency for each expression with this collocate can be calculated. If one expression occurs more often than expected with a collocate, that collocate can be said to be attracted by this expression. To see whether this attraction is statistically significant, an association measure can be applied, which can be used to measure collostructional strength, i.e. the strength of association between the expression and the collocate (Stefanowitsch Reference Stefanowitsch, Hoffmann and Trousdale2013). Table 2 gives an overview of the frequencies calculated in distinctive collexeme analysis.
Table 2. Relevant frequency types in distinctive collexeme analysis

Because I am interested in specialisation with particular meanings (command, inference) in a very constrained context (second-person subject, positive), I have adjusted the distinctive collexeme analysis to reflect these constraints, which can be seen in table 3. I used z-score as an association measure to calculate collostructional strength. Specialisation was attested when the significance of attraction was p ≥ 0.05.
Table 3. Adjusted frequency types for distinctive collexeme analysis

To identify cases of specialisation, I extracted the verbs that will and must occur with in their ‘command’ and ‘inference’ meanings, added up number of instances per verb and calculated the collostructional strength for each of the verbs that will and must both occur with. A case of specialisation was identified, when a verb was significantly attracted to one of the modals according to the z-score.
With the meaning ‘command’, 146 different verbs occurred with will and must, 52 uniquely with will and 82 uniquely with must. With the meaning ‘inference’, 23 different verbs occurred with will and must, 3 uniquely with will and 9 uniquely with must. Footnote 2 Table 4 shows the cases of specialisation, i.e. the verbs that are significantly attracted to both of the modals, with the meaning ‘command’; table 5 does the same for ‘inference’.
Table 4. Specialisation with ‘command’

*p<0.05.
Table 5. Specialisation with ‘inference’

*p<0.05.
To investigate the possibility of polysemous pre-emption, a second round of semantic annotation was carried out. Here, I looked at instances of the verbs involved in specialisation that had been coded as ‘other’ in the first round and further specified their meaning according to the present-day semantics of will and must (Nuyts Reference Nuyts and Frawley2006; Palmer Reference Palmer1990; Collins Reference Collins2009; Huddleston & Pullum et al. Reference Huddleston and Pullum2002). Will was annotated for prediction, dynamicity, willingness, plan and intention (including volition). Must was annotated for intention, dynamicity and obligation (including circumstantiality).
Table 6 gives an overview of all the meanings that were annotated for, the annotation principles and some examples.
Table 6. Annotation principles

These semantic meanings are broadly conceptualised as belonging to two groups of modal meaning:
-
(I) ‘epistemic’: involves the evaluation by a speaker [inference, prediction, dynamic], or
-
(II) ‘directive’: implies an authority committed to bringing the future event about [willingness, plan, intention, obligation].
4.2. Results
In this section, I investigate which of the three types of pre-emption – construal, statistical and polysemous pre-emption – can best explain the identified cases of specialisation. I do this in two steps: in a first step, I investigate whether must and will can be used interchangeably with a given verb without a difference in meaning to test for construal pre-emption. If they cannot, this is taken as an indication of a case of construal pre-emption and I offer arguments for why must and will offer different construal and how this explains the infelicity. In a second step, I investigate if, in addition to or instead of construal pre-emption, there might be a case for polysemous pre-emption. For this, I look at other meanings that will and must occur with in conjunction with a given verb, to see if there is strong association between will/must and this verb with another meaning which might have a pre-emptive effect. I argue that only if neither construal nor polysemous pre-emption can be attested, statistical pre-emption should be appealed to.
4.2.1. Specialisation of ‘inference’ must be over will be
The investigation shows that for the meaning ‘inference’ there is a clear preference for be to occur with must (54 instances) rather than with will (5 instances).
Potential for construal pre-emption: in some instances, must and will are interchangeable, as can be seen in (8) (modal outside the curly brackets is the actually attested modal). These instances usually involve the inference of an emotional state such as tiredness, hunger or stress (must 19 instances, will 1):

The inference of an emotional state is usually made on the basis of present evidence (she had been to see Phoebe, such a night as this), which naturally favours must. Will is also possible in (8a–c) and is also attested on its own once with an emotional state, although this instance may be an instance of ‘prediction’ as well (see 8d). While will can occur in these instances, it seems to offer a different construal, as it seems to imply that the speaker is not only drawing on the present evidence but also on their own experience.
Outside emotional state predicates, replacing must with will is questionable to impossible, as can be seen in (9)

In (9a–b) will seems odd because the given evidence is very much in focus and clearly related to the inference, which might conflict with the association of will with prior knowledge. In (9c–d), however, will seems to be infelicitous. This might be because the evidence is so pressing that it only allows one possible conclusion, rendering the more tentative will, which rather expresses a confidence based on past experience, infelicitous.
Will be also occurs outside emotional states and can be replaced by must as well with relatively little change in meaning:

Examples (10a–b) seem overall more assumptive than conclusive, which might explain why will is used instead of must. In (10a), it is unlikely that there are very clear indications that the addressees want to have a quiet talk with each other. In (10b), similarly, there are no strong indications in the context that the addressee can ride the horses; the speaker is merely being polite.
Overall, construal pre-emption seems to motivate the specialisation of must for be, because in many instances must cannot be replaced by will and even in instances where it can, will seems to offer a different construal.
Potential for polysemous pre-emption: I now test whether in addition to construal pre-emption, there might be a case for polysemous pre-emption as well.
Figures 1a and b show the proportion of must be and will be instances, respectively, across the different polysemes associated with them. Must be most commonly occurs with ‘inference’, but it also occurs frequently with the meanings ‘command’ (23%) and ‘obligation’ (25%). Thus, inferential must does not seem to be polysemously pre-empted from occurring with be. Will be, on the hand, predominantly occurs with ‘prediction’ (68%); most of the other meanings are very low in frequency. ‘Inference’ and ‘dynamic’ are especially low at just about 2 per cent each. Predictive will be is not only relatively better entrenched than the other meanings of will, it is also very well entrenched: predictive will be occurs 139 times, which is much more frequent than inferential must be at 54 instances. This distribution suggests that speakers have a very strong association between the use of will be and ‘prediction’ and that the strength of their association might have a pre-emptive effect on the use of will with be with other polysemes.

Figure 1a. Distribution of meanings of all instances of must be

Figure 1b. Distribution of meanings of all instances of will be
This shows that while construal pre-emption can explain the specialisation of inferential must be, polysemous pre-emption can be an additional cognitive factor contributing to this specialisation.
4.2.2. Specialisation of ‘inference’ must know over will know
The investigation shows that must know with ‘inference’ (34 instances) is a lot more common than will know (5 instances) and thus suggests specialisation for inferential must know.
Potential for construal pre-emption: many inferential must know instances (60%) express an inference based on assumption and can thus largely be replaced by will without much change in meaning. Here, the two would be in genuine competition:

In (11a–d), must know seems to express an assumption on the part of the speaker about the addressee’s knowledge rather than a conclusion based on evidence. None of the examples mentions any present indication; instead, vague epistemic stances (surely, I think, I submit, best) make the examples seem more assumptive than conclusive. Thus, must can be easily replaced by will in all the above instances. There might still be slight construal differences between must and will, though. With will being purely assumptive and based on past knowledge it seems to express a higher degree of confidence and speaker involvement than must: in (11a), for instance, must invokes some connotation of conclusion and thus creates some distance between speaker and addressee, whereas will does not.
Some instances of inferential must know cannot easily be replaced by will.

In (12a) will might replace must as the evidence for the inference is rather more like past knowledge than a clear present indication. In (12b–d), will seems impossible, because the examples either imply that the speaker really does see their inference as the only possible conclusion (12b–c) or else points to some emotional aggravation on the part of the speaker (12d).
Will know can in general be replaced by must; instances usually express an assumption on the part of the speaker, which can be expressed by both must and will.

Must seems slightly odd even in (13a–b), because the speaker seems to express a very high degree of confidence (at all events, of course) in the addressee’s knowledge without indicating any source of evidence. In (13c), however, must is infelicitous, because the context makes a conclusion reading impossible.
Must know and will know overall seem more functionally similar, as they are mostly interchangeable. There are, however, examples, where they are not interchangeable, pointing to construal differences between the two.
Potential for polysemous pre-emption: I now look at whether polysemous pre-emption may explain the specialisation for must know.
Figures 2a and b show the distribution of different polysemes with must know (figure 2a) and will know (figure 2b). Must know is clearly most frequent with ‘inference’ (72%) and thus strongly associated with it; it also occurs frequently with ‘intention’ (usually as if you must know) (21%). Will know is more frequent with ‘prediction’ (62%) than ‘inference’ (38%). Although predictive will know is more frequent than ‘inference’ will know, the difference between the two is not very big and the number of will know instances is quite low in general (12 instances).

Figure 2a. Distribution of meanings of all instances of must know

Figure 2b. Distribution of meanings of all instances of will know
The specialisation of inferential must know over ‘inference’ will know is thus unlikely to be motivated by polysemous pre-emption. Construal pre-emption seems the best possible option here.
4.2.3. Specialisation of ‘inference’ will remember over must remember
The investigation shows that will remember with ‘inference’ (25 instances) is more common than must remember (3 instances), which suggests a specialisation of ‘inference’ will remember.
Potential for construal pre-emption: will remember can be replaced by must remember in some instances but there seems to be a clear difference in construal. Will remember usually expresses the speaker’s confidence in the addressee remembering a certain event. In many cases, a topic is being discussed and you will remember is inserted parenthetically as a polite way to signal that the speaker is discussing shared knowledge. The tone of all will remember instances is rather casual and does not imply agitation on the part of the speaker.

In (14a–b), must remember can replace will remember but the construal seems very different. In (14a) for instance, the speaker is likely to be impatient, because they need to remind the addressee. The use of will signals a confident assumption; must on the other hand would have made the utterance sound more like a conclusion, which could be construed as seeming more agitated or even passive aggressive, e.g. since I wrote to you, you must remember it, don’t you?.
In most cases, must remember cannot replace will remember:

In (15a–b) must could be used instead of will but seems strange because the instances are clearly assumptive not conclusive (implicit in the that clause and the parenthetical use). The must alternatives also seem to imply more emotional involvement, which clashes with the general casualness of tone of the utterances. In (15c) must is clearly infelicitous because of the I trust matrix clause, which makes it clear that the speaker expresses their confidence, based on their own assumptions rather than anything external.
Must remember, on the other hand, can be easily replaced by will without much change in meaning, as can be seen in (16):

The specialisation of ‘inference’ will remember seems to be again motivated by construal pre- emption.
Potential for polysemous pre-emption: I now test whether in addition there are grounds for polysemous pre-emption as well.
The figures show the relative frequency of must remember (figure 3a) and will remember (figure 3b) across different polysemes. Figure 3a shows that must remember is much more frequent with ‘command’ (90%) than ‘inference’ (10%); ‘command’ must remember might thus polysemously pre-empt the use of ‘inference’ must remember. Figure 3b shows that ‘inference’ will remember is especially well entrenched (86%) and its use thus does not seem to be pre- empted by other polysemes.

Figure 3a. Distribution of meanings with must remember

Figure 3b. Distribution of meanings with will remember
This distribution seems to indicate that in addition to construal pre-emption, polysemous pre-emption could be a contributing factor in the specialisation of ‘inference’ will remember. Speakers associate must remember so much with the meaning ‘command’ that they avoid using it with ‘inference’ if they can.
4.2.4. Specialisation of ‘command’ must remember over will remember
The investigation shows that ‘command’ must remember is a lot more frequent (27 instances) than ‘command’ will remember (1 instance) and thus seems to be a case of specialisation.
Potential for construal pre-emption: must remember cannot be replaced by will remember when it expresses a reminder and can be paraphrased with ‘don’t forget’, which is the case in all but one instance; it thus does not express an authoritative ‘command’ but rather a reminder. In almost all cases you must remember takes a clausal complement.

In (17a–d) will cannot replace must, because it would be pragmatically too strong: in (17c), for instance, the speaker is trying to be diplomatic (perhaps – perhaps) and uses must remember to gently disagree with the addressee. Will could not express this gentle reminder and instead would sound very forceful.
The one instance where will can replace must is (18):

Example (18) differs from the instances in (17), because here must remember is followed by a to-infinitive clause rather than a finite clause. Must remember (that) is used as a reminder of fact whereas must remember (to) gives instructions for future conduct. While will was pragmatically too strong for the former, it is perfectly fine for the latter.
It is thus unsurprising that the only ‘command’ will remember instance also occurs with a to-infinitive clause (see 19).

Will and must thus show clear construal differences when it comes to the reminding of facts, where must suggests a polite invitation to remember and will a sharp warning, which explains the specialisation for must remember here.
Potential for polysemous pre-emption: looking at figures 3a and b again, it seems that polysemous pre-emption might play a role here as well. Will remember is very strongly associated with ‘inference’ (86%), which might explain why speakers avoid using it in the ‘command’ sense.
While ‘command’ and ‘inference’ on the surface seem like very different meanings, it seems that ‘inference’ will remember and ‘command’ must remember can fulfil the same communicative function or speech act, i.e. reminding, as can be seen in (20) and (21).


The (a) examples in (20) and (21) are the actual instances in which ‘inference’ will and ‘command’ must occur, while the (b) instances are respective paraphrases with ‘command’ must and ‘inference’ will. It seems that both can be used to express a polite reminder, but each offers a unique construal: ‘inference’ will by making assumptions about the addressee’s memory; ‘command’ must by inviting the addressing to remember a fact.
4.2.5. Specialisation of ‘command’ must forgive and excuse over ‘command’ will forgive and excuse
The investigation shows that ‘command’ must forgive (13 instances) and ‘command’ must excuse (12 instances) are more common than the equivalents with will, as will occurs with neither to express the meaning of ‘command’. This is a case of specialisation for must. Forgive and excuse will be jointly discussed, as they are semantically similar and exhibit similar distributions.
Potential for construal pre-emption: it seems that neither ‘command’ must forgive nor must excuse can be replaced by will (see (22) and (23)):


Must forgive and must excuse have very similar functions. The speaker makes a semi-passionate appeal to the addressee to forgive or excuse them, often accompanied by a justification for the action they are asking forgiveness for. ‘Command’ must expresses a lower degree of force that nonetheless seems to carry a degree of urgency, thus rendering the effect of a passionate appeal. Will forgive/excuse is infelicitous in these instances because it does not have this emotional quality and instead only foregrounds the directive force the speaker exerts on the addressee, which seems inappropriate in the context of forgiveness and excusals. In (23b) for instance, will would make it seem as if the speaker demands to be excused rather than passionately asking (or even begging) to be excused.
Thus, the specialisation of ‘command’ must for forgive and excuse seems to be strongly motivated by the construal differences between must and will.
Potential for polysemous pre-emption: I now test whether polysemous pre-emption might also apply.
Figures 4a and 5a show that must forgive and must excuse exclusively occur with the meaning ‘command’. This can be explained by the nature of the meanings involved and the second-person context restriction: forgive and excuse are accomplishment verbs and are thus less likely to occur with ‘inference’ than state verbs are. Obligation is strictly speaking not impossible since utterances like you must forgive those that have wronged you or as a host, you must excuse some of your guest’s peculiarities may imply an authority other than the speaker. These types of utterances, however, might not occur often in speech-based genres, which might explain why they are unattested in this particular investigation.

Figure 4a. Distribution of meanings with must forgive

Figure 4b. Distribution of meanings with will forgive

Figure 5a. Distribution of meanings with must excuse

Figure 5b. Distribution of meanings with will excuse
On the other hand, figures 4b and 5b show that will forgive/excuse does occur with other meanings, namely ‘prediction’ and ‘willingness’: will forgive is especially strongly associated with ‘willingness’, while will excuse is strongly associated with ‘prediction’. This might indicate polysemous pre-emption on top of construal pre-emption. We can also see this with the will paraphrases in (22) and (23): while ‘command’ will is infelicitous here as mentioned earlier, a ‘prediction’ will reading is perfectly fine – and possibly – the one we would default to if the meaning of will is not specified. Additionally, a ‘prediction’ will paraphrase of (22) and (23) does not seem to alter the communicative intent, i.e. a pragmatically implied speech act in which the speaker asks the addressee for forgiveness or excusal. The underlying semantic differences, however, result in different construals: while must expresses a passionate appeal via the ‘command’ sense, ‘prediction’ will rather expresses a sombre confidence. This is also evident in (24), which gives instances of ‘prediction’ will forgive/excuse.

The instances in (24) all express ‘prediction’, as the speaker is expressing their confidence in being excused or forgiven by the addressee in due course. In all instances I know or I trust can be added, which does not change the meaning and highlights the predictive nature of will. It might seem as if will could also express a ‘command’ here, but this does not seem to be the case: adding I ask to the sentences in (24) does change our interpretation of will, as can be seen in (25):

In (25a) adding I trust does not alter the construal of asking for excusal (confidence in future excusal) clearly implied in (24c). In (25b), however, adding I ask does change the construal (firm ‘command’).
Similarly, will excuse/forgive can also express the implicit speech act asking for excusal/forgiveness, however, with the construal of banking on the addressee’s willingness, as can be seen in (26).

The instances in (25) all express ‘willingness’. The speaker wants the forgiveness or excusal by the addressee, but instead of expressing that they are confident of being forgiven, they are banking on the addressee’s willingness, i.e. their acceptance of complying with the speaker’s desire for forgiveness or excusal.
This indicates that while there is specialisation of ‘command’ must for forgive/excuse over ‘command’ will, it cannot be said that ‘command’ must forgive/excuse (25 instances) have specialised for the expression of the speech act ‘asking for forgiveness/excusal’, since ‘willingness’ will forgive/excuse (19 instances) and ‘prediction’ will forgive/excuse (15 instances) express the same speech act and are similarly frequent. Instead, they are different construals of the same speech act (asking for forgiveness or excusal exist) but differ in terms of directness, and by extension in terms of politeness and emotionality (see figure 6): ‘command’ is the most direct construal, because it ‘stands for the request scenario as a whole’ (Ruytenbeek Reference Ruytenbeek2021: 83). As such, while it is still polite (probably because forgiveness is something that can never truly be commanded but needs to be freely given by the addressee), it is less polite than the other two construals, because the ‘command’ construal foregrounds the deontic force, which in this case is interpreted as higher emotional involvement by the speaker. ‘Willingness’ is less direct, because it rather expresses a condition for the request to be fulfilled. As such, it is the most polite construal, precisely because it is indirect and activates the pre-stage of the request. ‘Prediction’ is similarly indirect, as it expresses the result of the request, but it is less polite precisely because the request is made by presuming the request will be granted (cf. Panther & Thornburg Reference Panther and Thornburg1998 in Ruytenbeek Reference Ruytenbeek2021: 83). The different effects of the three construals can be compared in figure 6.

Figure 6. Different construals of asking for forgiveness
5. Discussion
In this article, I investigated the feasibility of different motivations for specialisation. Specialisation was defined as a specific outcome of competition where one expression that functionally competes with another expression occurs significantly more often. In addition to two motivations for specialisation proposed by previous studies, i.e. construal pre-emption and statistical pre-emption, I suggested polysemous pre-emption as a possible motivation for specialisation. Polysemous pre-emption means that one expression is semasiologically (i.e. relative to its other meanings) less entrenched with the target meaning than its competitor, which might pre-empt its use. The main purpose of the article was to see in how far polysemous pre-emption could motivate specialisation in the competition of will and must for specific verbs. I looked at the meanings ‘command’ and ‘inference’, which both will and must express, although they exhibit differences in construal in some contexts. I identified six instances of specialisation and investigated possible motivations for this specialisation.
The investigation found that all instances of specialisation could be explained by construal pre- emption. However, it also showed that in some instances, polysemous pre-emption could be a possible motivation. I showed that in these cases the rejected expression had highly frequent polysemes, which indicate that the speaker better associates that expression with other meanings instead of the target meaning. This was especially conspicuous in the case of rejection of ‘command’ will forgive. When substituting will in an instance of ‘command’ must forgive (You must forgive me) speakers might default to a ‘prediction’ interpretation, which is a better entrenched polyseme here.
I want to highlight, though, that polysemous pre-emption is only seen as a possible, not the sole, contributing motivation for specialisation. This is because, despite their conceptual similarities, the small meaning differences between ‘command’ and ‘inference’ will and must often meant that the two expressions could not always be interchanged, when it came to a specific communicative intent. ‘Inference’ must be for instance could be replaced by ‘inference’ will be with the communicative function of making assumptions about the addressee’s emotional state but less so in other contexts. The construal differences that drive specialisation usually boil down to politeness and degree of force. ‘Command’ must forgive, for instance, serves the communicative function or speech act of asking for forgiveness by construing it as a passionate appeal. ‘Command’ will forgive was, for instance, completely unattested in the corpus, because the degree of force that will assumes here would be incompatible with the communicative function. This shows that even if two expressions have similar meanings they are not always similarly appropriate for the expression of the same communicative function.
It might seem as if polysemous pre-emption is not really a motivation for specialisation in the cases investigated, because construal pre-emption on its own seems able to account for specialisation. However, activation and selection of different meanings for production and comprehension is complex and multifaceted and it is likely that different motivations factor into it. Construal and polysemous pre-emption are, for instance, both captured in Giora’s concept of salience (Giora Reference Giora2003). Salience is a property of meanings at a given point in time, which is a function of frequency, familiarity, prototypicality, contextual appropriateness and priming; more salient meanings are more easily activated than less salient meanings. In the case of the rejection of ‘command’ will forgive it could be argued that ‘prediction’ will forgive is more salient not only by virtue of its contextual appropriateness, because it offers a more polite construal than ‘command’ for asking for forgiveness (~construal pre-emption), but also by virtue of frequency, because will is more frequent with the meaning ‘prediction’ than with ‘command’ (~polysemous pre-emption).
I have also found that two expressions that have different coded meanings could be used to express the same communicative function or speech act. Asking for forgiveness, for instance, could be construed as a passionate appeal (‘command’ must forgive), an indirect request (‘willingness’ will forgive) or confidence in future forgiveness (‘prediction’ will forgive). None of the construals has specialised for expressing forgiveness; instead, speakers can freely choose between the construals, depending on the degree of politeness or emotionality they want to express.
Through the qualitative analysis of will and must with ‘command’ and ‘inference’, the article also contributes to a better understanding of the respective similarities and differences between the two expressions. Across different contexts and meanings, I found that must can often express a degree of agitation, emotion and urgency that is in general absent from will. These tendencies can be observed both with ‘command’ and ‘inference’ must. In you must remember and you must forgive/excuse, for instance, ‘command’ must is only possible because it signals a degree of emotional involvement (almost desperation) which softens the otherwise strong deontic force of the ‘command’. ‘Inference’ you must remember on the one hand is less common precisely because of the sense of urgency it expresses (i.e. you must remember, there is no other possibility). Will on the other hand expresses a confidence both with ‘command’ and ‘inference’ meaning.
The article has thus delivered on testing the potential for polysemous pre-emption but cannot reliably show the size of its effect on specialisation, given that construal pre-emption could also be attested.