1. Introduction
Verbal formatives such as have and get have been described as ‘semi-lexical’, having a status intermediate between purely functional and purely lexical categories (e.g. van Riemsdijk Reference van Riemsdijk1998; Emonds Reference Emonds, Corver and van Riemsdijk2001; Alexiadou Reference Alexiadou2012). On the one hand, they behave like lexical verbs in most varieties of English in that they do not pass tests for auxiliary status (e.g. those of Huddleston & Pullum Reference Huddleston and Pullum2002: 92–102). On the other, they lack the semantic specificity of lexical verbs, being compatible with a diverse range of interpretations, as illustrated for have below:
There are limits to this freedom, however, as shown by examples such as the following, where the oddness indicated by ‘#’ seems to be due to a requirement to construe the subject of have as an affected experiencer:
An adequate analysis of have must therefore be able not only to rule in examples such as (1) but also to rule out those such as (2). The existing literature contains a variety of approaches to the syntax and semantics of have, but a common theme in many of them is ‘underspecification’; that is, they claim that have differs from lexical verbs in lacking lexical specifications of meaning and/or syntactic properties, allowing it to be compatible with a wide range of uses. On the other hand, have cannot be too underspecified, on pain of permitting examples such as (2). The primary aim of this article is to propose a system of argument linking – and an analysis of have within it – that licenses examples such as (1) in the same way that it licenses argument structures for lexical verbs (e.g. possess, break, contain, and gain, respectively for (1a–d)) and correspondingly rules out examples such as (2) on the same basis that lexical transitives do not allow ‘pure’ patient, goal, or theme subjects. A particular contribution of the paper is a description and analysis of ‘affectee’ examples such as (1d), which differ in important ways from ‘affected experiencer’ examples such as (1b).
Section 2 presents an analysis of verbal argument linking, inspired by various previous works (e.g. Jackendoff Reference Jackendoff1990; Neeleman & Van de Koot Reference Neeleman and Van de Koot2002a,b; Reinhart Reference Reinhart2002, Reference Reinhart2016), which captures a number of generalisations about the repertoire of verbs in English. Section 3 applies this analysis to have, proposing that have is underspecified both syntactically and semantically, and shows that this proposal can account for the major interpretations of have in (1). Section 4 compares the proposal with two recent underspecification proposals by Kim (Reference Kim2012) and Myler (Reference Myler2016), showing that they face particular problems with ruling in ‘affectee’ examples such as (1d) while ruling out examples such as (2). Section 5 is the conclusion.
2. Argument structure in syntax and semantics
2.1. Syntax: D-selectors and s-sets
In this paper, I adopt the basic approach of Neeleman & Van de Koot (Reference Neeleman and Van de Koot2002a), which advances a system of syntactic representation that accounts for five shared properties of grammatical dependencies: c-command by the antecedent, obligatoriness of the antecedent, uniqueness of the antecedent, non-uniqueness of dependents, and locality (cf. Koster Reference Koster1987). In particular, I adopt their view that verb-argument dependencies are established by a ‘selectional function’ (a lexical property of the verb) that is copied upwards node by node until it comes to immediately dominate an appropriate argument. Accordingly, a sentence with a transitive verb (i.e. one taking two determiner phrase [DP] arguments) can be analyzed as in (3b), where the verbal node contains an ordered set (henceforth ‘s-set’) of selectional functions (henceforth ‘D-selectors’):
The analysis in (3b) departs from that of Neeleman & Van de Koot in two main ways. First, D-selectors combine the functions of thematic selection and c-selection (for which they posit separate, but linked, functions).Footnote 1 Second, the ordering of D-selectors within a single node (for which they posit a separate ‘ordering tier’) is determined on the basis of a binary-valued ‘theta-feature’ (for related ideas and discussion, see Reinhart Reference Reinhart2002, Reference Reinhart2016; Marelj Reference Marelj, Malicka-Kleparska and Bloch-Trojnar2019; and references cited in Levin & Rappaport Hovav Reference Levin and Hovav2005: 44–47). The binary nature of this feature can be motivated by a number of observations. First, there is a distinction between transitive verbs that allow passives and those that do not, which is usually assumed to correspond to a distinction between verbs that take an external argument (unergatives) and those that do not (unaccusatives) (e.g. Perlmutter & Postal Reference Perlmutter, Postal, Perlmutter and Rosen1984; Pesetsky Reference Pesetsky1995). Second, within the unergative group, there is a distinction between verbs that allow by-phrases in passives and nominalisations and those that do not. This is illustrated for passives of transitive verbs in (4), and for nominalisations of intransitive verbs in (5):Footnote 2
If we take the type of auxiliary used in perfect/past constructions in languages such as Dutch as indirect evidence for unergative vs. unaccusative status in English, the transitive verbs that allow passive in (4a–h) can be identified as unergative, as their Dutch equivalents use hebben ‘to have’ in perfects, while those in (4i, j) can be identified as unaccusative because some Dutch verbs with similar meanings use zijn ‘to be’ (see, e.g. Zaenen Reference Zaenen and Pustejovsky1993; Lieber & Baayen Reference Lieber and Baayen1997; Hoekstra Reference Hoekstra1999).Footnote 3 The class of intransitives also shows a three-way distinction. Alongside clear unergatives such as (5a), which both permit by-phrases and use hebben in Dutch, and clear unaccusatives such as (5c), which disallow by-phrases and use zijn in Dutch, there is a third category, exemplified by (5b), that disallows by-phrases but uses hebben in Dutch (Zaenen Reference Zaenen and Pustejovsky1993: 131). This category corresponds to what Levin & Rappaport Hovav (Reference Levin and Hovav1995) call ‘internally caused verbs’ and Reinhart (Reference Reinhart2002, Reference Reinhart2016) calls ‘theme unergatives’.
Thus, there is grammatical justification for a three-way distinction, which can naturally be characterized in terms of a binary feature that may be present or absent on a given D-selector, as expressed in Table 1. Thus, an ‘external argument’ under the present proposal is simply a ±θ-marked D-selector that is initial in its s-set.
Categorisation of D-selectors in terms of ±θ
The postulation of a single binary-valued theta-feature allows for a natural characterisation of the range of permissible verbal argument structures in English. A first relevant observation is that the maximum number of DP arguments per verb in English (and many other languages) is three. This fact can be captured straightforwardly under the assumption that the members of an s-set must be distinct (e.g. Neeleman & Van de Koot Reference Neeleman and Van de Koot2002a: 542; Reinhart Reference Reinhart2002: 264). Thus, alongside the proposed s-sets for intransitive and monotransitive verbs in (6) and (7), we have s-sets for ditransitives as in (8), which represent the maximum number of D-selectors for a single s-set:Footnote 4
Table 1 only determines the θ-specification of the initial D-selector in an s-set, which is the most relevant for our purposes. As for the feature values associated with non-initial D-selectors, I assume that these can be determined in part by the distribution of of-prepositional phrases (PPs) in nominalisations. While of is thematically underspecified in that it can introduce agents, causers, experiencers, patients and inclusors, there are some cases in which of-PPs are unable to represent a particular argument:
I suggest that of-PPs target the most underspecified D-selector in an s-set. The effects of this are clearest in (9a, b), assuming that the s-set of the underlying verb is as shown: the of-PP is only compatible with the final, θ-unspecified D-selector (under an event reading of the nominal).Footnote 5 Examples such as (9c, d) – and (5a) above – can be handled if English has a rule of ‘+θ-impoverishment’, which deletes the +θ specification of a D-selector (as indicated by parentheses in (9)), provided that this does not create a conflict with the distinctness requirement on D-selectors. Thus, by process of elimination, the non-initial D-selectors that cannot be targeted by of must be specified as –θ.Footnote 6
2.2. Semantics: Conceptual structures and linking rules
2.2.1. The interpretation of ±θ
I propose that ±θ restricts the kind of argument positions to which a DP may be linked in a conceptual structure (CS) in the sense of Jackendoff (Reference Jackendoff1990), as determined by the linking rules below:Footnote 7
Thus, the semantic content associated with ±θ is relational, in that this feature is always associated with a two-place function in CS. The terms ‘exclusive’ and ‘inclusive’ are intended to capture the intuition that certain relations conceive of two arguments as separate or independent of one another, while others conceive of two arguments as in a kind of ‘inclusion’ relation, either literally or metaphorically. The contrast is perhaps clearest between agentive verbs such as such as kill, which involve one individual (the agent) acting on another individual (the patient) external to it, and verbs such as include, which involve two individuals considered to form a ‘whole’ of which one is a part. Arguably, however, inclusion should be understood in a broader sense. For example, following Belvin (Reference Belvin1996), I assume that the notion of ‘inclusion’ should be extended to take in ‘proximal’ relations, as in the case of verbs such as contain. Similarly, given the bifurcation between ‘external’ and ‘internal’ causation referred to above, it seems reasonable to think of internal causation as involving a kind of inclusion relation; hence the assignment of –θ to subjects of verbs such as bleed, where we might think of the relevant relation as that between the causal mechanism inside a person and the person themselves. Furthermore, the fact that certain uses of object experiencer verbs such as concern are incompatible with passives (as in (4k) above) suggests that, in such cases, the subject and object are in a kind of mental inclusion relation (cf. Manzini & Franco Reference Manzini and Franco2016). For example, in (4k), it is not the cleaning rota itself but a mental representation internal to Corin that is the causer of the emotion.
The conception of unergative subjects expressed in (10i, ii) can be contrasted with that of proposals making reference to notions such as ‘cause’ and ‘mental state’ (e.g. Reinhart Reference Reinhart2002, Reference Reinhart2016). The main problem for such proposals involves the linking properties of verbs such as precede, follow, border and accompany, which behave like unergatives (i.e. permit by-phrases and take hebben in Dutch) but whose initial argument cannot be characterised as a cause or a mental state holder (see e.g. Davis & Koenig Reference Davis and Koenig2000: 59). Under the present analysis, such verbs can be distinguished from contain and include in terms of the conditions in (10i, ii): while contain and include crucially involve the internal structure of the subject argument, precede and follow rather refer to the external relation that the subject entertains with the object.
2.2.2. Force transmission and mental state relevance
Let us now consider the way in which a transitive s-set can be associated with a CS. Note that the rules in (10i, ii) may apply either to a single ±θ-marked D-selector or to such a selector immediately followed by another D-selector. In principle, then, an s-set <D+θ, D> may be linked to a CS either by the application of the rules in (10i, iii), or by the application of rule (10i) alone. Some evidence that the second, more ‘economical’ linking is correct can be provided by considering a number of restrictions that hold of transitive verbs. For example, consider the sentences in (11a, b) and one potential CS that could represent them in (11c), based on that proposed by Jackendoff (Reference Jackendoff1990: 91–95):Footnote 8 , Footnote 9
While (11c) seems roughly adequate as a representation of the meaning of (11b), it falls short as a representation of (11a) because transitive change-of-state verbs such as break generally require something like ‘direct causation’ (e.g. Fodor Reference Fodor1970 and much subsequent work). That is, in (11a) it is generally understood that the ball must have come into forceful contact with the bookcase, while (11b) could describe a situation in which someone puts a ball into a complex machine, starting a chain of events that culminates in a baseball bat swinging against the bookcase (see e.g. Rappaport Hovav & Levin Reference Rappaport Hovav and Levin2001; Neeleman & Van de Koot Reference Neeleman, Van de Koot and Everaert2012 and references cited).
A way of expressing this difference is by adopting a version of Jackendoff’s (Reference Jackendoff1990) proposal that CSs consist of two linked ‘tiers’ expressing distinct aspects of a sentence’s meaning. For example, the CS in (11c) can be modified as in (12) to incorporate the requirement of direct causation: the outermost cause function on the uppermost (henceforth ‘matrix’) tier is linked to a function ftr (for ‘force transmission’) on the lower (henceforth ‘dependent’) tier, of which the ball and the bookcase are both arguments:Footnote 10 , Footnote 11
A plausible way of understanding the contribution of ftr in (12) is as meaning ‘The ball transmits force starting from the causing event and continuing to the “being-broken” state predicated of the bookcase’.Footnote 12
Now, recall that the verb break was assigned the s-set <D+θ, D> on the basis of its permitting a by-phrase in passives. Given the linking rule in (10i), the two D-selectors in this s-set can now be related to the two argument positions of ftr. If, instead, rules (10i+iii) could apply separately to the D-selectors in the s-set, we would predict that direct causation (qua force transmission) could be circumvented. Accordingly, I propose that the general economy condition in (13) holds of s-sets:
To obtain the desired result, it is also necessary to assume that all D-selectors in a given s-set must be targeted by at least one linking rule (a kind of Full Interpretation requirement; Chomsky Reference Chomsky1986).
Agentive verbs require a different treatment, as they do not strictly require direct causation. For example, Neeleman & Van de Koot (Reference Neeleman, Van de Koot and Everaert2012) observe that in (14) (adapted from Katz Reference Katz1970), the gunsmith does not have to have directly attacked the sheriff for the sentence to be true or acceptable; he could have deliberately tampered with a gun that was supposed to fire blanks, and in that case would be considered to be accountable for the sheriff’s death:
Neeleman & Van de Koot argue that the requirement holding of a causative verb with a sentient subject (or agent) is not direct causation, but ‘accountability’. A way of encoding this requirement under the present analysis is to posit a distinct dependent function msr encoding ‘mental state relevance’ (cf. the [+m] feature of Reinhart Reference Reinhart2002, Reference Reinhart2016):
In accordance with the interpretation of dependent functions proposed above, [msr (x, y)] can be read as follows: ‘the gunsmith’s mental state is relevant both as part of the causing event and as part of the “being-dead” state that is the minimal eventuality containing the sheriff’.
Mental state relevance is also presumably the crucial factor in the linking of other types of transitive verbs that have a sentient subject, such as those in (16):
As it is not crucial here, I leave open whether these verb types are to be licensed by a single msr function or by distinct (sub)functions referring to different types of mental state relation. (I return to examples such as (16c) below.)
2.2.3. Verbs of motion and transfer: msr, aff, and anc
Transitive verbs of motion, as in (17), illustrate the need for dependent functions beyond ftr and msr:
As discussed by various authors (e.g. Carter Reference Carter, Levin, Tenny and linking1988; Jackendoff Reference Jackendoff1990: 261; Gruber Reference Gruber, Baltin and Collins2001: 262–264), verbs expressing pure motion toward a goal require theme>goal order, as in (17a); thus, there are no verbs such as *bereach with the opposite linking, as in (17a’). The example (17b, b’) illustrates that the opposite ordering holds for verbs expressing change of possession; that is, they require recipient(goal)>theme order. I would like to argue that these generalisations can be understood in terms of the Economy Condition on S-Sets (ECOS) in (13), together with the following additional condition on linking:
Recall that in Section 2.1 the verb reach was assigned the s-set <D, D–θ>, while receive was assigned <D+θ, D>. An important difference between these two s-sets is that <D+θ, D> can be licensed by a single linking rule, (10i) (which links it to msr), while <D, D–θ> requires two linking rules, (10i+iii), to apply. Now, suppose that both reach and receive have a matrix CS along the lines of (19):
The Initial Linking Condition (ILC) in (18) requires one of x and y in (19) to be linked to an outermost function. Since this cannot be satisfied by the matrix tier alone, it forces the presence of a dependent tier function linked to become. Plausibly, in the case of receive the relevant function is msr, as suggested above. Thus, (20a) would have the CS in (20b):
I take msr here to mean ‘Rhea’s mental state is relevant both as part of the becoming event and as part of the possession state that is the minimal eventuality containing Rhea’. The independent justification for linking msr to become here is that, for some person x to receive some item y, it is normally the case that x’s mental state changes in relation to y.Footnote 13 Of course, we can imagine an event in which someone puts the ravioli in Rhea’s handbag without her knowledge, but calling this a receiving event would be somewhat misleading.
In the case of reach, however, neither msr nor ftr seems appropriate: an event of reaching neither requires force transmission between its two arguments nor sentience of its initial argument. What, then, is the relevant interpretative property of the subject of reach? Important here, I think, is the fact that the subject, but not the object, of reach passes the classic ‘What happened to X’ test for ‘affectee’ status (e.g. Halliday Reference Halliday1968: 196; Jackendoff Reference Jackendoff1990: 111):
Accordingly, I posit a dependent function aff(ectee) whose single argument is interpreted as affected by the eventuality to which it is linked, where, roughly following Beavers (Reference Beavers2011: 339), I take ‘affectedness’ to refer to a change (either actual or likely) of state or location. Thus, I propose that (22a) has the CS in (22b):
The aff function in (22b) has the effect of allowing an argument of the embedded state introduced by be to also be understood as a participant in the matrix become event. In terms of linking, aff further ensures that the initial D-selector of <D, D–θ> satisfies the ILC: this D-selector is linked via rule (10iii) to the single argument of aff in (22b), which is an ‘outermost’ function.
Let us now consider how the unattested ‘inverse’ linkings in (17a’, b’) can be ruled out. Suppose, for example, that *acceive has the s-set <D, D+θ>. This could in principle be linked to the CS in (23b):
However, this would require the application of two linking rules, (10i+iii), to the s-set. By contrast, the s-set <D+θ, D> requires only a single application of (10i) to be linked to the CS in (20b) above. Hence, by the ECOS, <D+θ, D> blocks <D, D+θ>.
The exclusion of *bereach requires a little more discussion, given the linking properties of locative verbs such as contain. If the CS of contain-type verbs corresponds roughly to the embedded be state in (22b), then the ILC would predict that they should have theme>location order, contrary to fact. Suppose, in part following a proposal of Neeleman & Van de Koot (Reference Neeleman and Van de Koot2002b), that contain-type verbs involve a dependent function anc (for ‘anchor’) that takes the location as its first argument (cf. the ‘proximal zone’ of Belvin Reference Belvin1996: 83):
Why, then, could there not be a verb such as *bereach, with the s-set <D–θ, D> and the CS in (25b)? Indeed, this possibility would be expected to block verbs with the linking of reach under the ECOS, because linking <D–θ, D> to (25b) would require only the application of rule (10ii):
Intuitively, the difference has to do with the fact that receiving primarily involves a (likely) change in the mental state of the recipient, whereas reaching involves a change in the location of the theme but not in the internal properties or location of the goal. For example, Rhea could receive the rambutans without any literal change in the internal properties or location of the rambutans having taken place (for example, if the previous owner simply signs them over to Rhea). By contrast, in order for the rambutans to reach the restaurant, they literally have to undergo a change of location. Thus, the become component of receive is fundamentally an event of (expected) mental state change (hence linked to msr), while the become component of reach is fundamentally an event of change of location (hence linked to aff). I will therefore assume that anc is restricted to occurring with eventualities in which the relevant locative relation holds throughout the eventuality, and hence can at best be linked to be in (25b).Footnote 14 , Footnote 15
At this point, it is worth noting that the postulation of aff alongside ftr does not pose a risk of ‘overriding’ the linking possibilities provided by ftr. First, the well-known fact that we do not find patient>agent verbs alongside agent>patient verbs can be accounted for in terms of the ILC. Consider the potential CS for the unattested (26a) in (26b):
In (26b), aff is linked to the embedded become event, as it makes sense to describe this event as a change of state but not the outermost cause event; thus, the linking in (26a) would violate the ILC (the bookcase would not be an argument of an outermost function).
A second relevant case involves non-causative verbs that entail force transmission, such as hit. Notable here is the fact that both the subject and object pass the test for affectee status:
Suppose that hit involves a theme-goal relation, as in (28ci) (e.g. Fillmore Reference Fillmore, Jacobs and Rosenbaum1970; Jackendoff Reference Jackendoff1990). The observations in (27) can then be taken to indicate that either of the dependent tiers in (28cii, iii) is possible in principle, as either would satisfy the ILC:
However, a hitting event fundamentally involves force transfer between its two participants. Thus, the dependent function ftr, as in (28civ), is another plausible option. Given that options (28cii, iii) would require two applications of linking rules, and (28civ) only one application of (10i), (28civ) blocks the other options under the ECOS. As well as accounting for the fact that hit, unlike reach, is compatible with passive by-phrases, this captures the absence of verbs with the meaning of hit but the opposite linking, such as *behit in (28b).
Finally in this subsection, consider examples such as (29):
Verbs such as gain and lose can denote alienable possession, if they have a sentient subject, or inalienable possession otherwise – here, a kind of part-whole relation – as in (29). In this case, the subject, but not the object, passes the affectee test:
This suggests the following analysis:
The analysis in (31b) combines aspects of those of contain and reach. As with contain, the two D-selectors of gain are linked by rule (10ii) to the two arguments of anc in the corresponding CS. As with reach, the initial D–θ-selector is linked by rule (10iii) to the argument position of aff, which is itself linked to the outermost matrix function become, satisfying the ILC.Footnote 16 Furthermore, the proposal captures the fact that there are probably no verbs with the meanings of gain and lose but with the order theme>inclusor.Footnote 17
3. The interpretations of have
3.1. The lexical specification of have
Section 2 outlined a system of argument linking with the following five main features:
In this subsection, I will argue that the major meanings of the semi-lexical verb have can be accounted for within this system if have is lexically specified as follows:Footnote 19
That is, have is underspecified in two senses. First, it has a transitive s-set whose first D-selector is compatible with either value of ±θ.Footnote 20 Given the association of initial ±θ with unergativity, this captures Myler’s (Reference Myler2016: 336–343) observation that have is transitive (or unergative), not unaccusative (pace Freeze Reference Freeze1992, among other works). Second, have lacks a lexically specified CS altogether. Thus, when have occurs in a syntactic structure, it must somehow be associated with a CS in order to obtain an interpretation. The system as proposed thus far predicts that this may be achieved by applying one or more of the linking rules in (10), repeated below:
In principle, the function-argument structure introduced by one of the rules could be a matrix or a dependent CS. I will argue that both of these possibilities are instantiated by have. Before showing how this works for each of the major interpretations of have, it will be necessary to go into more detail on how the linking rules in (34i–iii) establish semantic interpretations of syntactic constituents. First, it is necessary to provide principles determining the way in which the CS associated with a syntactic node is derived from the CSs of that node’s daughters. Consider, for example, CS-inheritance by a transitive verb phrase (VP) from its daughters:
The relationship between the CS of the higher V node and its daughters can be captured by the following principles, a default inheritance rule in (36) and an inheritance rule referring to a satisfied D-selector in (37):
Let us now consider how these principles can be applied to the major interpretations of have.
3.2. Causative have
Causative uses of have involve one of three complement types, which I will refer to as ‘bare VP’, ‘passive VP’, and ‘gerund-participial VP’ (the latter term coming from Huddleston & Pullum Reference Huddleston and Pullum2002):
Here I will focus on the bare VP type in (38a), returning to aspects of (38c) in particular below. A major question here concerns the status of the post-have DP: is it a raised object of have, a base-generated object of have or a surface subject of hose? Here I take the position that (38a) involves a raising-to-object structure, assuming a ternary-branching structure as in (39):
This enables us to maintain the idea that have is consistently a monotransitive verb – that is, it has two D-selectors in its s-set. Now consider the proposed analysis in (40):
Neeleman & Van de Koot (Reference Neeleman and Van de Koot2002a) propose that A-trace (indicated by t <D> in (40)) is a lexical item that both satisfies a D-selector in its mother and introduces its own selector that percolates to its mother. In (40b), the percolated (unsatisfied) D-selector of the lower V undergoes identification with the second D-selector of have, with the identified D-selector being satisfied by Helga. Now, crucially, the inheritance condition in (37) allows for a mismatch between D-selection and linking to a CS position. That is, the diacritic ‘#’ indicates that the relevant selector immediately dominates a DP, but this does not force this DP to be linked to the same CS position that the D-selector is linked to.Footnote 21 This gives the effect of raising to object: all that the DP does in its ‘raised’ position is satisfy the D-selector of its mother, while the satisfied D-selector links another daughter – the lower VP in (39) and (40b) – to the relevant CS position. This accounts for the fact that the post-have DP may be an idiom chunk or (slightly marginally) an expletive:Footnote 22
3.3. Affected experiencer have
Affected experiencer uses of have are structurally similar to causative uses but differ in that the subject is understood as being (at least potentially) affected by the event denoted by the embedded VP (e.g. Bosse, Bruening & Yamada Reference Bosse, Bruening and Yamada2012). A further difference is that affected experiencer uses normally require a pronoun coreferential with the subject of have to occur within the embedded VP (e.g. Brunson & Cowper Reference Brunson, Cowper, Dick, Gomesh and Wilson1992); thus, Harry can more naturally be understood as an affected experiencer in (42) than in (38):
I propose that affected experiencer have also involves the raising-to-object structure in (39). The major difference is in the way that the CS of have is licensed. Consider the proposed analysis in (43):
As in (40), have in the syntax has the s-set <D+θ, D>. Linking rule (10i) /(34i) applies, linking this s-set to the CS-function msr. As this is a dependent function, there must be a matrix function to which it can be linked. I assume the following principle applies:
The requirement in (44) can now be satisfied if the default inheritance rule in (36) applies, in effect licensing the entire CS of the lower VP as part of the CS of the higher VP. Notice that, given that the lower CS already implies a mental state relation between the subject of have and the embedded VP, the mutual entailment requirement imposed by (36) is satisfied. A further desirable consequence of the licensing principle in (44) is that it accounts for the pronoun requirement. This is because (44ii) requires the subject of have to be contained in the matrix CS.
3.4. Possessive have
In the case of possessive have, there is no obvious source for the possessive CS component in the complement of have, which is simply the DP interpreted as the possessum. For example, in the alienable possession example in (45), we need a way to get from the CS content of horse, which is a Thing in Jackendoff’s (Reference Jackendoff1990) terminology, to the matrix CS introduced by be, which is a State:
There are various imaginable ways to do this, such as Myler’s (Reference Myler2016) proposal that the possessor interpretation comes from a Poss category inside the possessum DP. Given that Poss would always be silent in English, however, I would like to suggest an alternative based on the default inheritance rule in (36). Notice that, since the required relation between α and β is not identity, but mutual entailment (modulo type-shifting), this leaves open the possibility that the CS of the higher node could include that of the lower node as a proper part, provided that mutual entailment holds. Indeed, mutual entailment arguably does hold in the revised structure (46), in which the matrix CS of have contains the general loc function rather than loc poss :
Assuming that we can think of horse as existentially quantified, then the requirement for licensing the be-structure in the upper node is that ‘there is a horse’ and ‘there is a horse that is somewhere’ are in a mutual entailment relation, which is arguably the case.
3.5. Locative have
Given that (46) involves the general loc function, we can extend this analysis to locative have-sentences as follows:
The main difference between possessive and locative have-sentences is then that in the latter, the loc function of have is specified for a particular locative interpretation, coincident with that introduced by the preposition. I assume that the PP is syntactically a modifier that selects for its host via a Mod function (Neeleman, Van de Koot & Doetjes Reference Neeleman, Van de Koot and Doetjes2004). Its contribution to the CS of its mother is made via the default inheritance rule in (36).
3.6. Affectee have
The final major interpretation to be considered in this subsection is what I call the affectee interpretation of have, illustrated in (48):
In the previous literature, this usage has been relatively neglected, though examples are mentioned by Brugman (Reference Brugman1988: 146) and Belvin (Reference Belvin1996: 56–59). Despite this neglect, I believe such examples are particularly revealing about the correct analysis of have. Note, first, that in the examples in (48), the subject has two types of interpretative property: it is a kind of goal or source, and it is understood to be affected by the event. Examples in which the subject is a pure goal or source are ill-formed (parentheses around ‘#’ indicate that the sentence may be good on other interpretations; I return to this matter below):
Likewise, examples in which the subject is a pure affectee (patient) or theme, as in (50a, b) are also ill-formed. Furthermore, examples such as (50c) show that a locative/inclusion relation with some other argument in the clause is not enough; the corresponding theme argument must be the DP ‘complement’ of have. Compare (50d), where a locative relation is not required because the sentient subject can be licensed as an affected experiencer by the msr function:Footnote 23 , Footnote 24 , Footnote 25
That is, the interpretation of the subject of have in (48) has both a locative component and an affectee component. In this sense, it is parallel to the interpretation of verbs such as gain and lose when they take an inanimate subject.Footnote 26 Accordingly, I propose the analysis in (51):
The way in which have obtains its CS interpretation is the most complex of all the usages examined so far. As in the affected experiencer usage, it involves default inheritance of the matrix CS of the complement of have. As in the possessive and locative usages, it involves the application of rule (10ii)/(34ii) to the s-set of have (<D–θ, D>). Finally, in order to satisfy the ILC, it involves the application of rule (10iii)/(34iii) to the initial role of have (D–θ), licensing the dependent function aff.
Notice now that the lexical specification of have rules out the ill-formed usages of have in (49) and (50). First, a pure goal/source interpretation of the subject of have, as in (49), is ruled out because this would violate the ILC. This is because the anc function requires that the ‘anchoring’ relation it denotes must hold throughout the eventuality to which it is linked; thus, it must be linked to the be state and not to the matrix become event. Second, a pure affectee interpretation of the subject of have, as in (50a–c), is ruled out because the anc relation is necessary to license the –θ specification of the initial D-selector (Full Interpretation) and to ensure that the two D-selectors represent co-arguments of a single function (Co-Argumenthood, fn. 18).
Further support for the analysis in (51) comes from a consideration of the different possible VP complements of have. As has been observed many times in the literature, when have takes a VP complement, whether the sentence as a whole is eventive or stative seems to be determined by its complement (e.g. Belvin Reference Belvin1996: 29–33; Myler Reference Myler2016: Chapter 4). In particular, when have takes a bare or passive VP complement, it behaves as eventive, while when it takes a gerund-participial VP complement, it behaves as stative (e.g. Harley Reference Harley, Guéron and Zribi-Hertz1998).Footnote 27 This can be illustrated by the fact that present tense have is interpreted as repetitive (hence eventive) in (52a, b) but can be interpreted as a single eventuality (hence stative) in (52c):
The eventive/stative difference correlates with differences in the potential thematic interpretations available to the syntactic arguments of have. In particular, the requirement for an affectee interpretation disappears when have takes a gerund-participial complement:
Supposing that the gerund-participial VP is stative by default, while the bare VP is eventive by default (i.e. in this case, is introduced by a become function), we can understand the contrast in (53) as being due to the ILC. In (53a), the outermost matrix function is be, to which anc may be linked, hence satisfying the ILC. In (53b), by contrast, the outermost matrix function is become, and hence anc is too deeply embedded to satisfy the ILC, which means that an affectee interpretation of the tree is required:
3.7. Summary
The major interpretations surveyed in this subsection can be summarised as in Table 2.Footnote 28
Interpretations of have
4. Comparisons
4.1. Previous analyses of have
The existing literature on the syntax and semantics of have is vast, and I cannot do it justice in the space permitted here. Nevertheless, I think certain useful comparisons can be made between the present proposal and recent proposals by Kim (Reference Kim2012) and Myler (Reference Myler2016) that also analyse have in terms of underspecification.Footnote 29 Whereas many earlier works on have restricted themselves to analysing one or two different usages, the analyses proposed by Kim and Myler are more ambitious and detailed than most, and hence represent the main competitors of the present analysis in their aim to cover the major interpretations of have (causative, affected experiencer, locative, possessive).Footnote 30
4.2. Kim (Reference Kim2012)
Kim argues, working within the assumptions of Distributed Morphology, that have spells out a combination of two functional heads in a local relation, namely Appl(icative) and v (‘little V’). Her analyses of causative, affected experiencer and possessive have are illustrated in (55a–c):
As indicated above, the interpretation of the structure spelled out by have depends on the specific type of Appl and v selected: v cause (causative) vs. v be (affected experiencer, possessive), and ‘high’ Appl (causative) vs. ‘peripheral’ Appl (affected experiencer) vs. ‘low’ Appl (where ‘high’ and ‘low’ are used in the sense of Pylkkänen Reference Pylkkänen2008).
A first criticism that could be made concerns the analysis of causative have. Kim (Reference Kim2012: 77) describes Appl in (55a) as ‘denot[ing] a relation between a causee, Mary, and the event vP’. As noted in Section 3, however, the post-have DP may be an idiom chunk or an expletive, and hence does not need to bear any semantic relation to the event vP. The postulation of Appl in this case is therefore essentially vacuous, as the only function it serves is to enable spell-out of have together with v. By contrast, under the present proposal causative have instantiates a raising-to-object structure in which the second D-selector of have can be satisfied by the post-have DP without requiring this DP to be semantically linked to the relevant CS position.
As in the present proposal, Kim’s analysis of possessive have takes the possessive semantics to be derived from the interpretation of the clausal structure rather than from the possessed DP (as under Myler’s analysis; see below). She tentatively proposes (2012: 72–73 fn. 4) that locative have, not shown in (55), may have a similar structure, though does not provide details on the position and role of the locative PP. While this suggestion seems plausible within the set of assumptions and framework that Kim adopts, I believe that it leads to problems concerning the analysis of affectee have, which neither Kim nor Myler discusses.
Recall the main observations about affectee have made in Section 3: the subject of have is interpreted as (likely to be) affected but does not need to be sentient, and a locative or inclusion relation must hold between the subject and object of have. From the point of view of Kim’s analysis, these observations suggest that a kind of low Appl is involved.Footnote 31 For example, (56a) below might have a structure along the lines of (56b) (where the structural position of the controlled VoiceP is left open, as it is not crucial):
The question now arises of how to ensure that the subject is ‘affected’; for example, that sentences such as (49, 50) above are not generated. Presumably this requires a specific type of low Appl that imposes an affectee interpretation on its specifier, just as, for example, the low Appl in English ditransitives generally requires its specifier to be an (intended) possessor. This, however, raises a problem: just as the low Appl seen in ditransitives can be argued to occur in monotransitives such as receive, we would expect the putative ‘affectee’ low Appl to also occur in monotransitives. While we do indeed have verbs such as gain and lose (with an inanimate subject), discussed in Section 3, that fit this profile, we do not have ‘inverses’ of force-transfer verbs such as hit (i.e. Gruber’s *behit). The non-existence of verbs such as *behit was accounted for in Section 3 in terms of the ECOS in (13): theme>goal order results in fewer applications of the linking rules than goal>theme order; hence, the former blocks the latter. By contrast, within DM and related frameworks, the order of arguments is determined ‘once and for all’ by the order in which functional categories are merged; thus, there is no sense in which the absence of *behit could be explained in terms of theme>goal order ‘blocking’ goal>theme order. Furthermore, under Kim’s analysis, the only difference between have and lexical verbs is the part of the syntactic structure that they spell out. Therefore, it is difficult to see how it could capture the relative freedom of have in argument orderings as compared with lexical verbs.
4.3. Myler (Reference Myler2016)
4.3.1. The proposal
Like Kim, Myler adopts a DM-style framework and proposes that have spells out a combination of two functional heads, in this case Voice and v. He argues that the various uses of have are distinguished in terms of (i) the particular semantics assigned to Voice and (ii) the semantics of other functional heads lower in the clause. Thus, consider the structures that Myler posits for one type of causative have in (57a), affected experiencer have in (57b), possessive have in (57c), and locative have in (57d):
In contrast to Kim’s analysis, which does not put any explicit constraints on the occurrence of different ‘flavours’ of Appl and v, Myler argues, following Wood (Reference Wood2015: 30), that the interpretations associated with Voice are determined by linking rules (‘rules of allosemy’) that refer to the semantics of Voice’s complement (Myler Reference Myler2016: 255, 282):
Thus, in the case of have, the rules in (58i–iii) derive different subtypes of causative have, while the rule for ‘expletive’ Voice in (58iv) derives the remaining uses of have. In the latter case, Voice does not assign the subject of have an argument interpretation; rather, an unsaturated argument position is ‘passed up’ from lower in the clause (the head Free for affected experiencer have, the head Poss inside the possessum DP for possessive have, and the pronoun inside the locative PP for locative have).
4.3.2. Causers and engineers
Myler’s main reasons for proposing rule (58iii) can be illustrated by the examples in (59) (adapted from Myler Reference Myler2016: 281):
The contrast between (59a, b) suggests that, when causative have takes a bare VP complement (and is hence interpreted as a dynamic event), the subject must be animate or sentient. By contrast, (59c, d) show that causative have with a gerund-participial complement (which is interpreted statively) permits either a sentient or a non-sentient subject (e.g. Belvin Reference Belvin1996; Ritter & Rosen Reference Ritter and Rosen1997; Harley Reference Harley, Guéron and Zribi-Hertz1998; see also Marelj Reference Marelj, Malicka-Kleparska and Bloch-Trojnar2019 on aspectual restrictions). A second relevant observation concerns the interpretation of the subject in (59a). While the subject could be referred to as an agent, Myler treats it as bearing a distinct role, engineer, because of its interpretation in examples such as (60c) (2016: 582):
As Myler (Reference Myler2016: 581) notes, in (60c) the subject argument is not a direct causer of the falling-over event but rather ‘a sort of backstairs orchestrator of it’. Thus, while (60a, b) can describe an event in which I put my leg out and cause John to fall over, (60c) cannot.
While I would not question Myler’s description of the interpretation of (60c), I believe that it is not necessary to postulate what is, in effect, a construction-specific interpretation for the subject of eventive have. Rather, I believe that the properties of causative have can plausibly be described in terms of ‘blocking’ (e.g. DiSciullo & Williams Reference DiSciullo and Williams1987). There are broadly two types of situations in which one form is preferred over another to express the same meaning. One could be called ‘Gricean blocking’; for example, monoclausal causatives such as (60b) are normally preferred over periphrastic causatives such as (60a) (or, more relevantly, I used my leg to make John fall over) because they are briefer (e.g. McCawley Reference McCawley and Cole1978). A second type could be called ‘structural blocking’: Where two forms could occur in a given syntactic position to express the same meaning, the more highly specified form wins; an example is the competition between *goed and went. As has been noted (e.g. Katzir Reference Katzir2008: 22–23), violations of structural blocking are typically much less acceptable than violations of, say, Grice’s Maxim of Manner. Applying this logic to eventive have would thus explain why, while I used my leg to make John fall over can mean what (60b) can (while conveying implicatures as a result of the Maxim of Manner), (60c) cannot mean what (60b) can. More specifically, if causative make is lexically specified for cause and ftr, it is more highly specified than have, which has no lexical CS, and hence make should structurally block have to express the same meaning.Footnote 32
The structural blocking approach predicts that have should be able to take an inanimate subject where make is not available. This is the case in stative gerund-participial cases such as (60d) above, where replacing have with make would lead to ungrammaticality because make does not select for a gerund-participle. It furthermore extends to the cases in (61a, b) (note that (61a) is relatively acceptable under a stative interpretation, but have here would not be expected to be blocked by eventive make):Footnote 33
This analysis further predicts that eventive have should be able to express non-engineer causation – and hence potentially take an inanimate subject – provided that the causing event does not involve force transmission. Indeed, although gerund-participial complements are much more common in Internet searches, there are many examples of have + bare VP with an inanimate subject, of which I give some below:Footnote 34
Notably, in some of these cases replacing have with make leads to a perceptibly different meaning. In (62a), using make would suggest that the men in question are under compulsion to rub shoulders with royalty, and similarly in (62b), it would suggest that Mickael had no choice but to reconsider his career. In (62c), the difference is less stark, perhaps, but the use of have here still seems to suggest that the author chose to have the relevant thought, rather than being compelled. These cases, then, can be accounted for if the eventuality they describe does not involve force transmission, and hence have is able to denote a cause function with no dependent ftr, allowing it to circumvent blocking by make.
4.3.3. Locative have and expletive Voice
Myler (Reference Myler2016: 263) argues that, under the locative interpretation, the locative argument slot is introduced by the locative preposition and passed up as a lambda-abstract to the subject of have via expletive Voice (in part following Kratzer Reference Kratzer2009). His structure for locative have is repeated below:
By contrast, under the present analysis, repeated below, the locative meaning arises from assigning have the anc function, which in turn depends on a matrix be function that is derived from the CS of the complement DP through mutual entailment. This proposal and Myler’s thus make different predictions with respect to the obligatoriness of the locative PP: while it must always be structurally present under Myler’s proposal, under the present proposal it should be optional in principle, assuming that the loc function can be interpreted freely as a locative or a possessive. That is, the following analysis should be possible with a locative interpretation of have:
Some evidence that this is correct comes from dialogues such as the following, in which the PP can be omitted if the context makes it clear that a locative interpretation is intended:
Under Myler’s proposal, it would be necessary to assume a silent PP in Brenda’s response. However, both of the logical options are implausible. Consider first the possibility that it contains a null PP pro-form (PP-pro). Although there does not appear to have been much discussion of the issue in the literature (though see Cinque Reference Cinque1990: Chapter 3; Grosu Reference Grosu1996: 289; Szabolcsi & Den Dikken Reference Szabolcsi, Den Dikken, Cheng and Sybesma2002), it is unlikely that English has PP-pro, any more than it has NP/DP-pro. For instance, this would lead us to expect that no English verb has an obligatory overt PP complement, but verbs such as put, place and locate are clear counterexamples to this. Furthermore, given that PP-pro, if it occurred in examples such as (65), would have no internal structure, and would always have an overt coreferential antecedent (hence could not be independently bound), it is difficult to see what kind of positive evidence could be given for its presence.
Another alternative that should be considered is that (65) contains a full PP that has undergone ellipsis (i.e. ‘PF-deletion’; Merchant Reference Merchant2001). Unlike PP-pro, an elliptical PP would have internal structure that could in theory be probed by syntactic tests. Consider, in this light, the following examples:
As shown in (66a, b), VP-ellipsis and sluicing allow scope relations between the stranded and elided quantifiers – in particular, (66a, b) allow a reading in which sets of professors covary with students – which can be taken to indicate that the elided quantifiers are still syntactically present (e.g. Merchant Reference Merchant2001). By contrast, in (66c), in which the PP argument of give in the second conjunct is ‘missing’, the covarying reading is not available. This suggests that whatever mechanism is responsible for the missing PP in (66c) is not ellipsis in the sense of PF-deletion.
Thus, I conclude that examples such as (65) do not plausibly contain a PP at all, as this would have to be a specific type of elliptical PP or PP-pro that is excluded in other cases in English. Under the present proposal, the PP is expected to be optional in principle, as borne out by cases such as (65). The fact that the PP is apparently obligatory if contextual support is absent must be accounted for by assuming that possessive interpretations of the loc function have a default or unmarked status (perhaps along the lines of the ‘feature tree’ proposal of Harley & Ritter Reference Harley and Ritter2002: 485–486); the presence of a PP thus enables this unmarked interpretation to be overridden through inheritance of a more specific loc function.
4.3.4. Affectee have and overgeneration
A related question concerns how Myler might account for ‘affectee’ uses of have, considered in Section 3. For example, one might combine aspects of his analyses of affected experiencer have and locative have in (57b, d) above to analyse example (56a) above as in (67):
That is, the pronoun it in (67) would pass up a lambda-abstract to the hotel via expletive Voice, as in (57d) and (63), and a silent functional head Aff(ect) would introduce the affected interpretation, just as Free in (57b) introduces the affected experiencer interpretation. The analysis in (67) would require an extension of the passing-up mechanism, such that the lambda-abstract could pass not just through PP and PredP as in (57d) and (63), but also through VP. Indeed, this is required anyway, given examples such as The stadium has two pubs flanking it (Myler Reference Myler2016: 263), which Myler analyses as involving locative have.
Such an analysis, however, would face a problem of potential overgeneration. For example, it would generate not just acceptable affectee have examples such as in (67), but also unacceptable examples such as (68), repeated from (49b) and (50b):
As discussed in Section 3, affectee have appears to require the subject of have to be both location-like and affected; (68a) is then problematic because the subject is not plausibly affected, and (68b) because the subject is not location-like. This dual requirement was handled in Section 3 via the (independently motivated) ILC and the requirement of Full Interpretation applying to s-sets. Under a Myler-style analysis such as (67), however, it is not clear how either requirement could be ensured: as a reviewer suggests, it would seem necessary to stipulate that the passing-up of the lambda-abstract is impossible if a bare, non-participial VoiceP is embedded, which would require the use of a head such as Free (or Aff).
5. Conclusion
This paper has proposed an analysis of the major uses of (semi-lexical) have in terms of syntactic and semantic underspecification. In contrast to previous underspecification analyses, it accounts for a class of interpretations of have that have been neglected in the previous literature, namely affectee interpretations. Furthermore, by positing an explicit parallel between the linking mechanisms for have and those for lexical transitives more generally, it is able to account for various properties of have-constructions that are problematic for previous analyses, such as Kim (Reference Kim2012) and Myler (Reference Myler2016).
An obvious extension of the present proposal that suggests itself concerns the properties of get, also often described as a semi-lexical verb. If have is lexically specified as <Dαθ, D> with no CS, we naturally expect that there might be other semi-lexical verbs making use of the possibility of one or three D-selectors per s-set. Get fulfils this expectation, as it can occur with one, two or three DP arguments, as shown in (69a–c) together with suggested s-sets:
The arguments of get can also apparently be interpreted in all the ways discussed for have: in addition to the causative/agentive subjects in (69), it permits an affected experiencer subject as in (70a), an eventive locative subject in (70b) or an affectee subject as in (70c) (note that the affectedness requirement in (70b, c) follows from the ILC):
A further possible extension of the present proposal concerns the broader cross-linguistic syntax of possessive sentences, as discussed in particular in Myler (Reference Myler2016). A major point of variation is whether languages use (the equivalent of) have to express clausal possession, as in English, or (the equivalent of) be. For example, consider the possession sentences in (71) (both meaning ‘I have a book’):
It is notable that in ‘be-languages’ either the possessor or the possessum typically bears an extra marker: for example, a preposition or case-marking on the possessor, as in (71a), or a preposition on the possessum, as in (71b). While space precludes a fuller analysis here, one possibility is that selectional relations in cases such as (71) are ‘distributed’ between the verb be, which is semantically underspecified and contributes the s-set <D>, and the preposition or case-marker, which contributes the possessive semantics and selects upwards for a verbal projection (cf. Osborne Reference Osborne2019 on this possibility). For example, (71a) could then be analysed as follows (where I use arrows to indicate the direction of selection):
More work is required to make this a viable analysis, given the wealth of cross-linguistic observations contained in Myler’s work and elsewhere, but I think it shows that the present proposal raises interesting possibilities.
Acknowledgements
I would like to thank the three Journal of Linguistics referees (one of whom was Neil Myler) for their thorough and thoughtful reviews, which I believe have led to a much-improved article (though I alone am responsible for the results, of course). I would also like to thank audiences at Fudan University and Zhejiang University, where I presented some of this work in 2021 and 2022 respectively.
Competing interests
The author declares none.
Open access
