2.1 Origins and development of the SDH

‘Semantic determinism’ is defined here as any hypothesis concerning verbal argument structure whereby there is a mapping from (some level of) a verb’s semantic representation to the syntactic form of its arguments, such that syntactic representation of those arguments does not feature in the verb’s lexical entry.Footnote ^[8]

Semantic constraints on argument realisation were present in early transformational generative grammar (TGG) (Chomsky Reference Chomsky1965), but they were co-constraints on ‘lexical insertion’. A lexical predicate of category V (say, sense) both subcategorised for the number and syntactic category of its complements (7a), and also specified ‘selectional features’, which placed constraints on the inherent semantic properties of the head of all its arguments (7b):

Note that, in isolation, a subcategorisation frame such as (7a) is sufficient to license an unbounded set of complements, as long as their categorial features match those stipulated in the frame. However, the addition of the features (7b) constrains the set of argument tokens licensed for a particular position to a certain subset, in order to putatively block the generation of a sentence like danger senses Walt.Footnote ^[9]

The development of the Government and Binding (GB) theory (Chomsky Reference Chomsky1981) crucially changed the relationship between selection and subcategorisation, by establishing the principle that the former is a necessary condition on the latter, following Stowell (Reference Stowell1981), whose ‘$\unicode[STIX]{x1D703}$-grid’ listed verbal arguments in the form of semantic (‘theta’) roles. Most subsequent GB work made the de facto assumption that theta roles are sufficient determinants of the number and category of syntactic arguments; finally, the UTAH (Uniformity of Theta Assignment Hypothesis) of Baker (Reference Baker1988) put in place a semantic constraint on their mapping to syntactic positions.Footnote ^[10]

Given well-documented problems with theta roles, which in GB were essentially conceived of as a universal list of semantic categories, some researchers (e.g. Jackendoff Reference Jackendoff1987, Rappaport & Levin Reference Rappaport and Levin1988) have concluded that theta roles are not primitives of lexical entries, but are better understood as derived notions, being defined over lexical structure itself. In the TGG framework, ‘lexical decompositions’ have been formalised syntactically (following Hale & Keyser Reference Hale, Keyser, Hale and Keyser1993), with primitive predicates such as cause and have analysed as heads and projecting into structures defining argument positions, while in lexicalist alternatives (e.g. Levin & Rappaport Hovav Reference Levin and Rappaport Hovav1995), such predicates combine into ‘event schemata’, which define semantic argument positions, which in turn putatively map to syntactic positions.Footnote ^[11] But the criterial underlying claim of both approaches is equivalent: abstract semantic elements define ‘configurations’ from which the syntactic expression of surface arguments can be derived.

2.2 Implications of the SDH for argument realisation

Jackendoff (Reference Jackendoff1987) argues that if theta roles are relational notions defined over a verb’s conceptual structure (an autonomous level of representation), then they must be determined by a verb’s individual meaning. He points out that there are many verbs whose theta roles cannot be captured by conventional labels (e.g. the direct objects of pass, jump, climb), and there is the related fact that ‘selectional restrictions’ can often only be stated in terms of highly specific semantic types (a liquid in the case of drink, a sum of money with pay).Footnote ^[12]

Such a conclusion sits uncomfortably with the SDH, even a version based on lexical decomposition. Since the semantic representation is supposed to derive the properties of syntactic arguments, it must embody semantic constraints in such a way that syntactic properties ‘fall out’ from them.Footnote ^[13] Unlike for Jackendoff (or Chomsky Reference Chomsky1965 for that matter), the constraints must generalise: they should define sets of syntactic argument structures (and the lexical membership of those sets) independently of the individuating semantic properties of their members (verbs). This necessity arises from the obvious fact that, although there are thousands of lexically distinct verbs, there are relatively very few syntactically distinct argument structures. Much of the literature on semantic roles has therefore been devoted to tackling the problem of defining the nature of roles in such a way as to adequately predict the mapping.

The price that is paid for this gain in generality, however, is the loss of (potentially criterial) semantic distinctions between verbs which may result in an ‘event schema’ that fares no better than a list of theta roles. To see why, compare the paired configurations of theta roles in (8) with the corresponding event schemata in (9), which on ‘strong’ versions of the SDH map directly to the ditransitive double object and oblique structures respectively:

While the roles in (8) are simply stipulated, the roles of the variables in (9) are defined in relation to the semantic ‘predicates’ of which they are ‘arguments’. But the roles themselves are equivalent: for example, y is the entity that is ‘caused to have z’ in (9a), but the entity to which z is ‘caused to go’ in (9b); in other words, y is a Recipient in (9a), but a Goal in (9b). Hence the empirical adequacy of an event schema in predicting the mapping from semantics to syntax is still valid only to the extent that the roles it defines accurately capture the salient semantic properties of all the verbs in a syntactically-defined set.

Moreover, it is not enough to say that roles are just relations defined by a predicate: they also invoke certain semantic argument types (‘conceptual categories’ in Jackendoff’s terms). For example, Recipients must be conceptually distinct from Goals, so there must be inherent semantic properties that distinguish them, namely that Recipients must be animate (capable of ‘receiving’), whereas Goals (understood as the ‘endpoint of a path’) need not be, and the Theme of (8a) (z in (9a)) must be an entity with properties enabling it to be ‘received’. In this way, both configurations in (8) and (9) must determine ‘selection’ of candidate argument tokens according to what semantic properties they bear. Thus the SDH entails ‘selectional restrictions’ which constrain licit argument tokens to a subset of the syntactic types licensed by subcategorisation alone, just as in Chomsky (Reference Chomsky1965).

3.1 The ‘strong’ hypothesis

Pinker’s (Reference Pinker1989) hypothesis can be schematised as follows:

As can be seen, there are two kinds of mapping rule: one maps a particular event schema to a particular syntactic structure, and another converts one schema into the other. In order to explain why the second rule is not completely productive (why some verbs do not entail both structures), Pinker proposes various semantic subclasses of ditransitive verb, membership of a certain subclass being a condition for the triggering of the rule.

Pinker’s proposal amounts to a claim that all alternating ditransitives manifest a systematic difference of meaning between their use in the oblique and their use in the double object. However, Rappaport Hovav & Levin (Reference Rappaport Hovav and Levin2008) demonstrate that a ‘caused motion’ semantics entails a Path argument, which surfaces as a PP with a variable head, or as the locative pro-form where, but that this is manifested only with a subset of the alternating ditransitives (examples adapted from theirs):

By contrast, verbs in Pinker’s other subclasses (‘giving’, ‘future having’, and ‘communication’) do not have a Path argument: the alternating verbs allow only PPs headed by to, and these indirect objects cannot be questioned with where:

Rappaport Hovav & Levin conclude that these other verbs have only a ‘caused possession’ meaning even in their oblique form.

But even this conclusion is too strong, for three main reasons. First, it is predicted that the indirect objects of all verbs based on a ‘caused possession’ schema will bear the role of Recipient, yet unlike the ‘verbs of giving’ (14a), neither show (included in Pinker’s ‘communication’ subclass) nor any of the ‘verbs of future having’ (14b) have this entailment:

The by-phrase is anomalous in (14b) because, as sporadically observed in the literature, although giving something to somebody entails that they receive it, a person cannot receive something simply by having it shown, offered, promised, or owed to them. Therefore a ‘caused possession’ sense for these verbs is unsubstantiated.Footnote ^[15]

Second, Pinker’s ‘future having’ and ‘communication’ subclasses are not syntactically valid because they are not, in fact, semantically coherent. For example, although causation is a basic component of the event schema, owe is a stative, not causative verb (Walt tried to offer/promise/#owe Skyler her freedom). Moreover, the ‘communication’ verbs show a lack of homogeneity among the semantic argument types associated with the direct object. Show patterns with the ‘giving’ verbs in being felicitous with NPs denoting concrete entities (15a), but the other verbs included in the subclass are not (15b):

Although the indirect object referent of show is inferred to see something (not receive it), the verb itself implies a physical act involving physical entities, just like the ‘giving’ verbs, and not a communicative one. It is therefore unsurprising that it makes a poor fit with the verbs in (15b).

However, the argument types linked to the direct object of these verbs do not suffice to define a coherent semantic set either:

Even in (16c) where all three verbs are non-deviant with the same direct object token, the NP is interpreted differently: a story is a message in oral form with tell, but it must be in written form for read, and its role in relation to the indirect object referent differs for teach, since it is not merely perceived but mentally acquired by him. These facts substantiate Jackendoff’s (Reference Jackendoff1987) argument that there are many verbs whose semantic argument types cannot be generalised by appealing to uniform semantic roles.

Finally, the mapping hypothesis is undermined by cases where semantic coherence exists, but it fails to entail syntactic coherence. Although there are near-synonyms of alternating ditransitives, they follow distinct syntactic patterns:

As we see in (20), this phenomenon is also manifested in the case of ‘verbs of giving’. Therefore, only those verbs that are actually, not potentially, in the ‘giving’ subclass satisfy the input conditions for the rules in (10), since only they have a ‘caused possession’ entailment, belong to a semantically coherent set, and undergo the ‘right’ alternation. Hence the ‘rules’ are vacuous.

3.2 The ‘weak’ hypothesis

As already noted, Rappaport Hovav & Levin (Reference Rappaport Hovav and Levin2008) (henceforth RH&L) argue, contra Pinker, that give instantiates only ‘caused possession’ in both syntactic variants, i.e. they assume the mapping rule (21):

Note that, even if this rule were not vacuous (applying only to an already-enumerated subset of ditransitives), it would represent a significant weakening of the SDH, since there is no longer any isomorphic mapping to syntactic structure from event structure. This notwithstanding, RH&L’s paper is remarkable for its implicit claim that ‘all uses’ of give are captured by the schema, since the verb’s individual meaning (its ‘root’ in their terms) is said to be coextensive with it (RH&L: 134–135). This section critically examines the validity of this claim.

According to RH&L, all of the following therefore involve the same ‘possessional’ meaning (examples based on theirs):

Note that these examples include two (22h, i) they acknowledge as ‘idioms’, but none of them are identified as problematic for their argument that all uses of give ‘involve caused possession’. Presumably in an attempt to claim semantic uniformity for even the most metaphorical expressions, they deny that the sense of physical transfer in (22a) betrays the basic elements of the verb’s semantics (it is said to be ‘illusory’), because the ‘possessum’ may be an ‘abstract entity’, in which case the inference of physical transfer fails, as in (22b).

Given that RH&L are defending the SDH (‘a verb’s own meaning plays a key role in its argument realization options’), what are the constraints that they assume? Although their paper nowhere states these explicitly (they appear to assume no constraints at all on either subject or direct object position), their argument clearly hinges on the properties of the indirect object: if give is a ‘caused possession’ verb, the variable y in (21) must bear the role of Possessor (and not Location, which would imply movement of z to y and therefore ‘caused motion’). Since a Possessor is by definition an animate, and the syntactic alternation no longer correlates to a difference in meaning, Green’s (Reference Green1974) ‘animacy constraint’ on the indirect object in both syntactic structures follows.

However, it is not the case that the indirect object is restricted to the expression of animates: (23a–h) are ‘echo’ responses to examples (22a–h):

The test shows clearly that the indirect object referents in (22c, e, f, h) are not animate, and are therefore not Possessors. Consequently, RH&L’s statement that the indirect object of the oblique is constrained by the fact that to ‘only takes animate complements and not inanimate complements that designate places’ (RH&L: 138) is evidently false.Footnote ^[16] This claim is more clearly contradicted by examples not in their data, not just in respect of the oblique:

Contrary to the claim RH&L cite from the literature, the indirect object NPs in bold are not in these cases metonyms for animate referents. They are places: in (24a) London does not mean ‘(a particular set of) the people of or in London’; rather, it refers to the city as an (abstract) whole.Footnote ^[17]

The only independent support given for their assertion that the expressions cited above as (22c–e) ‘involve caused possession’ is that the relation between the indirect and direct objects ‘can be described using the verb have’ (RH&L: 139) (their examples (20a–c)):

This argument turns on the question-begging assumption that have uniformly lexicalises a ‘subevent’ of the schema in (21), i.e. that it always means ‘possess’. But it does not. Whereas John has a house means either that ‘John owns’ or ‘has use of’ a house (‘possesses’, in their sense), neither John has a brother nor John has a bath every day mean that he ‘owns’ or ‘has use of’ a sibling or an act of bathing, and the same is true of all the subjects in (25). Moreover, these subjects can be construed (metaphorically or not) as locations rather than Possessors, and locational PPs can therefore be adjoined, as seen in (26):

This is not possible, however, for the other instances of have just mentioned (including the possessional (27a)), even though all three can also be related to give-sentences:

These uses of have are therefore semantically non-uniform, and do not support the claim that a ‘caused possession’ event underlies (22c–e).Footnote ^[18]

It is reasonable to conclude, then, that the expressions in (22b–i) only have various metaphorical senses of ‘caused possession’, contra RH&L. But this is not an option for their ‘single meaning’ version of the SDH, which entails that the verb imposes uniform semantic constraints on its argument tokens (see Section 2.2 above). RH&L appear to try to pre-empt this difficulty by stating at the outset of their paper that the notion ‘possession’ should be ‘construed broadly’ – broadly enough, apparently, to include the idioms (expressions they define as having ‘a non-literal meaning’) in (22h, i), but not give birth/rise/way to NP, regarding which they surprisingly comment that the verb ‘seems to have been bleached of its possessional meaning’ (RH&L: 155).Footnote ^[19] However, if the concept ‘possessor’ is itself ‘metaphorical’, then it predicts nothing, because any NP at all could qualify as a ‘possessor’ and be licensed as an indirect object, and since according to the SDH, semantics is the sole source of grammatical constraints, it would follow that there are no constraints on grammatical structure at all.Footnote ^[20]

The reductio ad absurdum of this position is that there is no way of distinguishing between ‘caused possession’ and ‘caused motion’. Consider the pairs in (28) and (29):

According to the SDH, the door and Walt are Possessors or Recipients when they appear with give but Locations when they are arguments of put. Yet the paired expressions are clearly synonymous, and have the same entailments (‘the door is covered with paint’; ‘Walt is made responsible’), so there is no way to distinguish one event schema from the other, and no logical riposte to the counterhypothesis that give and all the other ditransitives ‘metaphorically’ instantiate the ‘caused motion’ schema.Footnote ^[21] Any meaningful comparison of hypotheses is therefore impossible.

4.1 A non-determinist hypothesis of lexical entries

The alternative hypothesis concerning the properties of lexical entries is stated informally as follows:

Nothing in (30) precludes a lexical entry from having a single set of semantic arguments but more than one set of syntactic arguments. The hypothesis is neutral as to whether the verb’s meaning itself is ‘structured’ in the sense of e.g. Jackendoff (Reference Jackendoff1990), but semantic arguments are the only part of meaning ‘visible’ to argument structure.

Some evidence for semantic arguments as types, existing independently of syntactic realisation, is supplied by the following:

Devoid of context, the pronominal subject arguments of die and shine in (31) are likely to be construed as animate and inanimate respectively, since it is animates that die and inanimates that shine. Similarly, in (32), the subject and indirect object arguments of give and owe have their ‘default’ animate interpretation. However, as we have seen, argument tokens need not conform to these properties:

Nonetheless, as was shown in the echo-question test (23), the non-prototypical tokens retain their inherent semantic properties:

In other words, the semantic properties of the argument types and the argument tokens are disjoint, a phenomenon which will be returned to below.

In this light, we now look at give in more detail. The Oxford English Dictionary gives the ‘general sense’ of the verb in modern English as ‘to make another the recipient of (something that is in the possession, or at the disposal, of the subject)’.Footnote ^[22] The definition intuitively characterises a physical act of transference involving three participants, which can again intuitively be distinguished: (i) an agentive, animate (prototypically human) agent; (ii) an animate recipient; (iii) an inanimate physical entity capable of being passed from (i) to (ii).Footnote ^[23] These semantic properties characterise and individuate the verb’s semantic arguments.

Note that it is not sufficient for the direct object argument to be inanimate; it must also possess certain concrete properties (e.g. solidity and mobility) to be physically transferable. When an argument token lacks one or more of these properties, there is a concomitant effect on the construal of the act of transfer itself. Compare the following:

While a laboratory is inanimate, it has properties (size, mass, immobility) which make the inference of physical transfer fail; only the inference of transfer of ownership or use remains. The NPs a punch and an important task have (possible) physical manifestations, but lack concrete properties; the nouns hour and reason denote entirely abstract entities. This mismatch between the properties of the semantic argument and those of the tokens results in the ‘transfer’ meaning denoted by the verb being only metaphorically interpretable; in all apart from (37a), even the notion of ‘possession’ becomes abstract.

As already borne out by the data in Section 3.2, the indirect object can be realised by non-animate tokens (see also (38b–d) below), but the same is true of the subject, where inanimate tokens ‘conflict’ with the agentive, animate argument type associated with this position:

Again, the subject NPs retain their inherently inanimate reference (as a what? test will confirm) and therefore are not inherently ‘agentive’. But their syntactic position, where an animate argument type is interpreted, triggers an interpretation whereby they cause ‘transfer’ of some entity to a metaphorical recipient.

What of the relation between the verb’s semantic arguments and its syntactic argument structure? As already established, give has two syntactic structures, but only one meaning, and therefore only one set of semantic arguments, so the same arguments must be mapped to both structures. Furthermore, since the verb has no sense of ‘caused motion’, the preposition to of the oblique cannot denote a Path, and as the oblique is synonymous with the double object, it follows that to is meaningless, and must be simply a stipulated form in the lexical entry.

The role of syntactic arguments in determining constraints, plus that of semantic arguments in the triggering of metaphor, are supported by a brief initial consideration of give-idioms, of which a small sample is provided in (39):

Four things are striking: first, that give-idioms, in principle at least, occur in both syntactic variants. Second, as noted by Hudson (Reference Hudson1992: 262) and Bruening (Reference Bruening2010: 536), the ‘fixed’ part of give-idioms (in boldface) is always the direct object, never the indirect object alone. Third, as can be seen in (39b–e), the ‘open’ indirect object position can be occupied by an inanimate. Fourth, the fixed direct object, though inanimate in its literal reference, does not otherwise conform to the physical properties of the argument type, so we do not find idioms like give NP a book. Idioms therefore share properties with metaphorical uses in that (i) they obey the verb’s syntactic constraints; (ii) they do not observe putative semantic constraints; (iii) at least one argument token (the direct object in the case of idioms) must be semantically disjoint with the verb’s corresponding argument types.Footnote ^[25]

We turn now to send. Given the premise that this verb is polysemous, it has two sets of semantic argument types. One set, entailed from the ‘caused possession’ sense, will be assumed to be identical to that of give above, and these arguments will be mapped to the same alternation, repeated below:

It is then predicted that send in this sense should behave in principle just like give in allowing subjects and indirect objects in either variant to be realised by inanimates as well as animates, and direct objects to be non-concrete:

As (41b) shows, the hypothesis that the argument type linked to the indirect object is a Recipient is supported by the fact that only to, and not a true Path preposition, is licensed. It is also notable that, as with give, there are no idioms with a fixed indirect object and open direct object. On the other hand, the ability of an inanimate indirect object to have a ‘metaphorical recipient’ interpretation seems considerably more limited than give:

However, this restriction can be attributed to the fact that send, even in the above sense, has a more specific lexical meaning than give (something like ‘transfer by causing to go’) and is therefore less readily available to metaphorical usage per se.Footnote ^[26]

In its ‘caused motion’ sense, send has a Path argument instead of a Recipient. This argument maps to a PP, which consists of a preposition specifying the Path, and an NP complement specifying its endpoint. Unlike give, there is no lexical restriction on the preposition:

In fact, the head of the PP is clearly semantic, so in this context to is ‘allative’, i.e. it has its meaning of ‘up to an endpoint’, and is not a meaningless grammatical marker. The preposition is therefore lexical, and itself selects its NP complement (its ‘subject’ being the direct object of the verb). Therefore, the internal properties of the PP are not relevant to the syntactic argument structure of the verb, which is as follows:

Further evidence that the Path argument is mapped to the whole PP (and not to the NP inside it) is provided by a comparison between the ‘caused possession’ sense of (45a) and the ‘caused motion’ one of (46a):

In (45a), the Mexicans realises a Recipient argument and is animate, as shown by the use of who? in (45b). But in (46), where to is a (modifiable) Path preposition, the PP is questioned with where, and so is a Path argument and interpreted as one, even though the answer is ‘to the Mexicans’ (meaning that the Mexicans is a metaphorical ‘endpoint’).

The polysemy view of send solves an apparent ‘idiom anomaly’ illustrated by the following expressions, originally cited in Larson (Reference Larson1988: 340) as evidence that (some) ‘goal’ indirect objects of ditransitives can be idiomatic:

However, as Larson (Reference Larson2017: 423) concludes, these idioms are in fact not true ditransitives, but based on ‘caused motion’ verbs. If these verbs have Path arguments interpreted on their PP, then the bolded NPs are metaphorical endpoints like the Mexicans in (46a) (in the idioms they are actually metaphorical endpoints of metaphorical Paths).Footnote ^[28]

The third verb to be considered is put, which superficially seems similar to ‘caused motion’ send, subcategorising for a PP, but permitting a ‘free’ choice of preposition:

Although Jackendoff (Reference Jackendoff1987: 391) remarks that the verb has a ‘curious selectional restriction’, disallowing a preposition of source (from), direction (to, towards) or route (through), this is not an absolute restriction, as the following metaphorical/idiomatic expressions show:

Thus subcategorisation for a lexically unspecified PP must be correct; the apparent ‘selectional restriction’ says something about the argument type mapped to this PP.

Comparison with send and put in a prototypical context shows that their Path argument is not of the same type:

This distribution of PPs suggests that this argument of put describes a trajectory, but also a spatial relationship between the complement of the preposition and the direct object of the verb (concepts that to and towards cannot denote), so the verb does not really lexicalise ‘caused motion’, but something more like ‘caused location’. In contrast, as we have seen, the Path of send includes an endpoint (concepts denoted by to and towards).

In fact, the meaning of put is arguably at least as close to give as to send. As expected, metaphorical and idiomatic usage reveals the lack of semantic constraints on argument realisation, but put manifests both metaphorical Themes (like give), and metaphorical Paths (like send) (non-prototypical forms in bold):

Nonetheless all these examples meticulously respect the syntactic argument structure represented in (44).

If there are no semantic constraints on argument realisation, there is a further prediction. Although semantically ‘mismatching’ argument tokens might lead to coherent metaphorical interpretations, they need not. Thus expressions with ‘fictional’ senses are licensed by the grammar (52a), as well as surreal (52b) and nonsensical ones (52c):

If argument tokens are subject only to syntactic constraints, then expressions such as those above are entirely non-deviant as far as the grammar is concerned.

In conclusion, the hypothesis of (30) correctly predicts that the only grammatical constraint on argument tokens of these verbs is provided by their syntactic argument structure, with a verb’s semantic argument types only determining their interpretation. Hence metaphorical and idiomatic expressions are licensed in principle. However, this does not in itself account for how metaphorical expressions ‘work’, or how they are related to ‘idiomatic’ expressions. The next two sections turn to these questions.

4.2 The relation between verbal semantics and metaphor

Lakoff (Reference Lakoff1987) observes that metaphor is universal and pervasive in natural languages, but his contention that metaphor moulds the grammatical system itself is unpersuasive. Sullivan’s (Reference Sullivan2013) study of metaphoric language emphasises that the reverse is true: grammatical structures have an essential role in constraining metaphorical interpretations. Metaphors are cognitive phenomena that involve a process of mapping a ‘source’ concept (usually of a concrete kind) to a ‘target’ concept (usually abstract). Crucial to the understanding of verbal metaphors is the fact that, in Sullivan’s terms, verbs are conceptually ‘dependent’ elements (in the sense of usually requiring other ‘autonomous’ elements – arguments, in fact – for their interpretation): verbs therefore make good ‘source domains’ for metaphors involving ‘target domain’ arguments. For this to be possible, metaphors must belong to the conceptual system, which is independent of the grammatical system (one of the key principles of ‘representational modularity’ argued for in Jackendoff Reference Jackendoff1997). In this light, we can now understand that the empirical problems which metaphors pose for the SDH derive directly from the lack of any syntactic argument structure representation, which makes it impossible for ‘source’ and ‘target’ domains to be distinguished.

The principle that source and target domains must be conceptually distinct allows us to be more precise about the role of semantic arguments in terms of the present hypothesis. It is the meaning of a verb and its semantic argument types that provide the source of a metaphorical interpretation, and argument tokens that provide the target.Footnote ^[30] A metaphorical interpretation can then be triggered if at least one argument token is semantically disjoint with the corresponding semantic argument type. Compare again:

The lexical meaning of give and its semantic argument types provide the source: the basic sense of transfer by an animate Agent of a concrete Theme to an animate Recipient. In (53a) the semantic properties of every token are ‘subsumed’ by the properties of the corresponding type, so the ‘target’ NPs are understood as instantiations of the source types – hence the ‘prototypical’ sense. In (53b), the metaphorical interpretation – ‘Walt punched Jesse’ – is triggered by the target a punch, whose semantic properties are now disjoint with the properties of the source type linked to the direct object position.Footnote ^[31] Combined with the (source) sense of ‘transfer’ from the verb and the source types of the other two arguments, this generates the metaphor (that punching somebody is like transferring an object to them).

4.3 The relation between verbal metaphor and idioms

The notion of an ‘idiom’, as addressed in the literature, encompasses a rather miscellaneous collection of ‘multi-word’ expressions, ranging from highly irregular (including grammatically fossilised) forms at one extreme (see van Gestel Reference van Gestel, Everaert and van der Linden1995 for Dutch data) to ‘conventionalised’ phrases such as center divider (Nunberg, Sag & Wasow Reference Nunberg, Sag and Wasow1994) at the other. Another property that has been held to be criterial, particularly for verbal idioms, is semantic opacity or ‘non-compositionality’, as evidenced by kick the bucket and shoot the breeze, although it has been observed that it does not, in fact, appear to be very typical of expressions which are intuitively idiomatic.Footnote ^[32] Having established that metaphorical usage fully respects the syntactic argument structure of the verb, and that the only difference between prototypical and metaphorical usage lies in the properties of the argument tokens themselves, it is possible that there is no basis for positing a separate category of ‘idioms’ for any of the expressions studied here. Indeed, Larson (Reference Larson2017) argues that there are no true ditransitive idioms at all. Consider the following three sets of expressions:

Larson first draws attention to expressions of type (54a), whose direct objects denote communicable illnesses: these seem ‘completely productive’, and the direct objects all have their usual dictionary meanings. The same can be said for the ‘denominal’ direct objects in (54b): as the meaning of the direct object is transparent, the meaning of its combination with the verb can be derived. Larson goes on to argue that expressions of the type in (54c) (where the regularity of the direct objects is less convincing) are no different: for example, if the creeps means ‘a feeling of apprehension or horror’, then the expression The Count gave me the creeps can be generated without special stipulation.

Larson’s analysis is plausible in principle, but there are a number of problems with it. If the expressions of the (54c) type are indeed compositional, it is necessary that their direct objects combine in that form, since they are generally invariant:

Therefore there must be in the lexicon an idiomatic NP ‘the creeps’ meaning ‘feeling of unease’, an NP a hand meaning ‘help’, an NP ‘the boot’ meaning ‘rude dismissal’, and so on, in order for the expressions to be computed as Larson claims. Furthermore, these NPs and their relationship to the verb are no different from those in generally uncontested idioms such as keep tabson ‘monitor closely’, spill the beans ‘reveal a secret’ or pull strings ‘exert influence’, in two respects. First, their direct objects have the specially idiomatic property that they supply metaphorical sources, not targets, together with the verb (as noted by Sullivan Reference Sullivan2013: 96). For example, both spill and the beans evoke ‘reveal’ and ‘a secret’. Equally, give evokes ‘transfer’ and a hand evokes ‘help’. Second, Larson’s contention that the apparent give-idioms are merely ‘collocations’ (a collocation being a ‘conventionalized co-occurrence’) is problematic, because for him, ‘conventionalized’ appears to mean only that words are ‘used together frequently’ (Larson Reference Larson2017: 399), like rancid butter and dead serious. But while the collocation of rancid and butter can be predicted from the meaning of rancid, the same cannot be said of give and the direct objects in (54a–c). The meanings of lend NP a cold, put a kiss on NP, and throw NP the boot are compositional/computable, but they are not collocations because they are not conventionally associated with the meanings that the expressions in (54) are. Rather, the direct objects of (54) collocate with give in the sense of being ‘conventionally associated in the lexicon’.Footnote ^[33]

A further problem is that Larson does not succeed in showing that all ditransitive idioms are syntactically productive. Since give syntactically alternates in both prototypical and metaphorical use, we would expect all candidate idioms to do the same. But this is by no means clearly the case:

For these reasons, we must conclude that ditransitive idioms are genuine, and fit into the hypothesis (30) as ‘partly lexicalised’ metaphors, in the sense that (i) the idiomatic constituent is represented phonologically; (ii) the metaphorical meaning, like that of kick the bucket, is therefore stored with the whole expression. In Section 5 it will be shown how a theoretical description can account for both ‘compositionality’ (of metaphorical and prototypical uses) and (partial) ‘non-compositionality’ (of idioms).Footnote ^[35]

5.1 Theoretical background and assumptions

Although no particular theory of syntax will be adopted here, a constraint-based, Construction Grammar (CxG)-type framework is assumed. CxG takes the lexicon of the grammar to consist of constructions: in the most general terms, constructions are associations of ‘form’ with ‘function’. The main theoretical value of constructions for the present study is that they are non-derived and non-uniform: they may be as small as morphemes, or be syntactic units (possibly semi- or fully-lexicalised). Hence CxG is readily amenable to the representation of idioms and other idiosyncratic expressions (see e.g. Kay & Fillmore Reference Kay and Fillmore1999).Footnote ^[36] But if the argument properties (syntactic and semantic) of a lexical head are taken to be non-derived, then a lexical head and its dependants are a candidate construction also.

This notion of ‘construction’ is combined with the compatible principles of lexical structure proposed in Jackendoff (Reference Jackendoff1997, Reference Jackendoff2002). Jackendoff’s model for lexical items conforms to the properties of his ‘parallel architecture’: a tripartite linguistic system composed of strictly autonomous generative subsystems (phonology, semantics, syntax) linked by ‘interface rules’. On his view, a lexical item is a form of interface constraint, since in the typical case it registers information from all three systems. As Jackendoff (Reference Jackendoff2002: Section 6.5) argues, the properties of idioms require these too to be stored (as phrasal ‘lexical items’), but in his model these properties can easily be handled by the interface rules. In the present proposal, this form of idiom representation is generalised to an argument-taking lexical head itself, so that the head appears in a syntactic configuration with its argument types. The major advantage is that the structural principles of a verbal idiom are then identical to those of its (non-idiomatic) head verb, the major difference being in what is represented.

5.2 The Lexical Argument Construction

A Lexical Argument Construction (henceforth LAC) is a subtype of construction that represents the phonological, semantic, and syntactic properties of a lexeme (e.g. a lexical verb). It takes the form of an attribute-value matrix, where the attributes are phonology (phon), semantics (sem), and syntax (syn), and the values are the specified corresponding properties of the LAC. The phon value is that of the verb and any other stipulated lexemes; the sem value is the meaning of the LAC (usually just the verb) plus a representation of the semantic arguments (sem-args) that it entails; the syn value specifies the category of the head (X) plus those of its syntactic arguments (syn-args) in phrasal (XP) form (plus any other relevant morphological properties). A syn-arg is therefore a purely formal type, e.g. ‘NP’.

Following Jackendoff (Reference Jackendoff2002), the three components of an LAC are linked in various ways by ‘correspondence rules’ in the form of co-indexing. The strict separation of the components means that (i) ‘lexicalisation’ of any part of an LAC means representation as a phon value; (ii) the properties of the syn-args and sem-args are independent of each other and therefore non-redundantly stateable; (iii) there need be no fully isomorphic mapping between the elements of the three components.

To make things concrete, the LAC for the lexeme give is illustrated below:Footnote ^[37]

It will be seen that the phon attribute specifies the phonological content of the LAC, which here consists of the lexical verb and the preposition to (which must be stipulated). These items are coindexed with heads in the syn attribute, thereby associating each phon form with its respective lexical category.

The sem part of the LAC represents the meaning of give (‘give’) and also identifies its sem-args arguments, distinguished by the symbols shown.Footnote ^[38] Each is co-indexed with NPs represented in the syn attribute. There is no coindexing between to or P and any part of sem, ensuring that the preposition is semantically ‘void’. However, it will be noticed that whereas the sem component contains only one representation, the syn part contains two, and the sem-args are linked differently. This describes the fact that although there is only one meaning of give, its sem-args arguments can be realised in two syntactic variants (with no effect on meaning).

The importance of the co-indexing of the sem-args and syn-args is that it indirectly links semantic sources with semantic targets (argument tokens) via the syntactic licensing of the latter.Footnote ^[39] The syn value represents a constraint on argument realisation: only tokens which are NPs will be licensed in each position. However, the coindexing ensures that a licensed NP will be interpreted by the corresponding sem-arg. For example, any NP token licensed in indirect object position ($\text{NP}_{\text{y}}$) will be linked to the sem-arg <give-rct > and interpreted as an ‘instance’ of this type, if, as explained in Section 4.2, the semantics of the token is subsumed by the semantics of the type. If one or more tokens is semantically disjoint with the corresponding type(s), then a metaphorical interpretation may be triggered. Other than preventing the licensing of any item which is not of the stipulated category, the syn representation exercises no further constraints on argument realisation. The corollary is that an unbounded set of phonologically filled NPs may unify with the phonologically empty (i.e. unlinked) argument positions in the LAC.

On this basis, we now look at how the LAC represents verbal idioms. The LAC for give a hand (to) is illustrated below:

Comparing (58) with (57), we see that the phon string a hand is stipulated, which is now coindexed with the $\text{NP}_{\text{j}}$ (direct object) positions in syn. This has the effect of ‘blocking’ the licensing of any other NP in that position. The sem-arg linked with that position in the LAC of give is now absent, since the meaning of the verb (glossed as ‘help’) can now be thought of as associated with the whole string give a hand (to). Only two sem-args remain, which are coindexed with the subject and indirect object NPs unlinked to phonological material. The NPs that they are linked to will license NP tokens in the same way as with give above, but of course they will now be interpreted differently, since the sem-args are now of different types, in line with the special meaning of the idiom.Footnote ^[40] This representation assumes, of course, that this idiom is truly alternating. If the string give a hand to cannot be licensed, then the phon representation of the preposition and the oblique syn representation are simply omitted.

We now illustrate the LACs for send.

As two senses are postulated for this verb, accordingly two LACs are represented: (59) illustrates the ‘caused possession’ sense. In its formal properties it is identical to give. However, (60) has a somewhat different representation: (i) there is only one syn structure; (ii) no internal structure is represented for the PP, and there is no phon material linked to it; (iii) the PP is coindexed with a different sem-arg <send-path >. This entails, as desired, that any kind of PP token will be licensed. But if its meaning is disjoint with the sem-arg <send-path > to which it is linked, it will be interpreted as either metaphorical or deviant: as with give, a prototypical interpretation will only result if the semantics of all argument tokens are subsumed by the verb’s sem-args.

Finally, we show the LAC for put:

As the reader can verify, this LAC is syntactically identical to that of (60), but it is the properties of the sem representation that account for the differences identified in Section 4.1: for example, the linking of the PP to the sem-arg <put-loc > entails a different type of interpretation on a PP token. Idioms of both verbs would follow the same principles as illustrated above for give a hand to, representing a different meaning, different and fewer sem-args, and more phon material (e.g. for put NP at risk, the PP would be coindexed with at risk).