The content of grammar
3.5 The phase as the smallest unit of grammatical organization
The above account entails that the smallest unit of grammar is a unit of reference; that any such unit is based on a lexical concept that is as such non-referential but has a descriptive content that will be the basis of its referential use; and that the possible forms of reference for a given lexical concept depend on how it is grammaticalized: what grammatical role it plays and which head takes its lexical
14
Relegating predication, reference, and truth, to‘semantics’ or C-I, which we have argued against above, interestingly unites Minimalist grammar and its rivals. Thus, for Jackendoff (2011), the inertness of grammar in relation to the genesis of a propositional semantics is even more pronounced, as is the form– content distinction, with syntax in charge of the former and semantics in charge of the latter. In line with this, he stresses the existence of relational cognition in non-human primates, as Boeckx (2010) does, who suggests regarding Merge as completely unconstrained, with relevant constraints on grammar coming from the way in which derivations based on Merge hook up with extra-linguistic concepts such as thing or location at phase-boundaries. Our scepticism with regard to calling these ‘concepts’ is clear from the above.
projection as the complement. Putting these observations together, we obtain (26), where the‘interior’ houses the lexical concept and its lexical projection (e.g. Noun Phrase (NP), as in (27)), and the edge houses grammatical elements that govern how it is used to refer (e.g. determiners):
(26) [EDGE [INTERIOR]]
(27) [the [man]]
This unit of grammaticality cannot be reduced further. If there was no edge governing reference, we would only have a unit of conceptualization, not reference. If there is an edge but it is‘weak’, in a sense that we will make precise in Chapter 4, we get a unit of grammatical organization and hence reference, but reference will be established via the descriptive content of the lexical material in the interior only, as in I like kings of France, where the referent of the internal argument is no objects in particular, but objects only insofar as they instantiate the generic description‘being a king of France’. If we only had an edge, we might have functional reference, not an act of human grammatical reference, which is always based on a lexical concept with a descriptive content used predicatively.
Note that there is no recursion in this unit of grammar: nothing recurs. There is a single edge and a single interior. In line with this, there is a single referent, established at the edge, whatever it might be; and there is a single lexical concept. If another substantive lexical concept is added, it thus has the status of a modifier, which as such is not required and goes beyond the minimality exhibited in (26):
(28) [DPthe [NPgood [NPman]]]
As we would predict from this, modifiers exhibit recursivity in a much clearer way than units of reference: they can be stacked unboundedly (subject to conceptual constraints, though, as studied in Cinque,1999):
(29) [the [good old wise [man]]]
Crucially, the recursion, insofar as there is one, is part of the interior, not the edge (but we return to determiner and clitic doubling below and in Chapter4). Moreover, none of it disturbs the essential generalization above: that there will be a single referent at the edge—even if described by multiple predicates—and a single lexical head that the unit in question turns into a referential expression. By not exhibiting recursion, the smallest unit of grammar exactly resembles, as it should, an act of (e.g. index-finger) pointing: no such act is‘recursive’ either, in the sense that it would ever contain two referents, with one referent embedded in the other. The question thus arises what the status of recursion in grammar is, which we address in the next subsection.
(26) resembles the recent notion of a phase as conceptualized in Chomsky (2007a and b; 2008; 2013), and we will thus suggest that the phase is the smallest unit of
grammatical organization. From the non-recursivity of this unit, at any given point of the computation, it then follows that there will only ever be a single phase in a derivation, resulting in a ‘single-phase’ architecture. At any one moment in the derivation, that is, only one referent is ever computed. This is true of units such as (31) as well, where ‘that’ is the edge element governing reference and the TP, which corresponds to NP in the nominal case (30), is the interior specifying the descriptive content of a particular state of affairs:
(30) [the [NPkings of France]]
(31) [that [TPthere are kings of France]]
The same template repeats again in (32), where the embedded CP functions as a descriptive predicate by means of which we pick out a referent at the edge of v*, in this case a mental event of believing:
(32) [v*believes [CPthat [TPthere are kings of France]]]
We see it onefinal time in (33):
(33) [CPC [TPBill T [v*believes [that there are kings of France]]]]
Here C is again the edge—(33) as a whole refers to a truth value—and TP provides the descriptive content, as in (31). In each of (30–3), there is a single referent and a single identifying description, illustrating the presence of the same, single, non-recursive template in (26). No sentence has two subjects or two predicates; no sentence has two truth-value referring expressions, as illustrated in (33), whose meaning is such that no truth value is assigned to the embedded clause‘that there are kings of France’. Instead there is only one truth value and it is assigned to (33) as a whole. The embedded clause is therefore not evaluated for reference/truth, in the way it would be if it was not embedded as an argument. This is the exact prediction that our template makes, and the intensionality of clauses that are embedded arguments is thereby deduced: for given the constraints of the template, the embedded clause must function as a predicate within the higher phase in which it enters, rather than carrying a truth value into the process of semantic composition. Far from being the anomaly that intensionality has been, causing puzzlement and consternation from a viewpoint of semantic compositionality for more than a century, it now follows from the archi- tecture of grammar, being a core aspect of grammatical semantics (Hinzen, Sheehan, and Reichard,2013).
The referents of the phases are formal-ontologically distinct: if we wished to use ontological language, the nominal phase is mapped onto an‘object’, the verbal to an ‘event’, the clause to a ‘proposition’. Crucially, these are asymmetrically ordered with respect to one another, with propositions always (or architecturally) containing events as part of their structure, forming their thematic core, and events in turn
containing objects. We can thus think of these three phases as the‘first’, ‘second’, and ‘third’ phases. The non-recursivity of the phasal template entails that none of these formal-ontologically distinct objects of reference will ever embed another object, within the same phase. Instead, when an event embeds an object, say, as all events do, we expect that the object in question mustfirst be mapped from a predicate, in order to be licensed to the event. This is what we find: arguments are relations (Bowers, 2011). An object-referential expression such as ‘John’ can only function as an object within an event if it is interpreted as bearing a grammatical relation to it, which is what the term ‘thematic role’ denotes. The form of a verb–noun embedding is therefore not of the form (34) but (35):
(34) [VP [DP]] (35) [VP [Ł DP]]
That is, the object is an Agent, Patient, Goal, etc. Crucially, neither of these reflect properties intrinsic to the referents in question, in which case they would be lexical- semantic, rather than grammatical relations. With theta-roles, therefore, we are firmly in the domain of grammatical meaning. As Bowers (2011: 8) stresses, ‘Agent’ viewed as a theta-role is a grammatical category, not a (lexical-)semantic one. As he explains, following Emonds (1976), a pseudo-passive as in (37) is possible just in case the subject in (36) bears the argument-relation of Agent rather than Theme: (36) The bird/the book flew across the room
(37) The room was flown across by the bird/*the book
The issue here is not animacy, a semantic notion, since even the animacy of the subject‘the bird’ does not prevent it from being construed as a Theme-argument in (36), in which case the bird is merely involved as an inert projectile and the pseudo- passive is now also impossible for it. Similarly, a fairy tale context could allow the construal of ‘the book’ as an Agent-phrase, showing that nothing in its intrinsic lexical feature specifications rules the construal of a referring expression as exhibiting a particular grammatical semantics. The structural nature of the ambiguity in question is illustrated in (38), where ‘John’ can again have the grammatical meaning of either Agent or of Theme:
(38) John appealed to Mary.
In the second case, (38) means that John was attractive to Mary, in which case the pseudo-passive is again impossible:
(39) *Mary was appealed to by John.
Non-grammatical notions such as ‘intentional agency’ do not seem to succeed in explaining these facts. As Bowers notes, (40), with ‘unintentionally’, only has one interpretation:
(40) The room was deliberately/unintentionally flown across by the bird.
It uniquely means that the bird crossed the room intending tofly somewhere else; it does not mean that the bird was sent across the room in the manner of an inert projectile. In other words, the grammar of the argument as an Agent requires constru- ing the world in particular terms, namely with the bird as an intentional agent, no matter whether lack of intentionality is specified lexically through the adjunct. A world seen through the lenses of grammatical meaning becomes different indeed.
Crucially, there is no recursion in the theta-domain either: theta-roles, like argu- ment-structure itself, form a strictlyfinite, and indeed very limited, template. The grammar stretches to generating structures of three syntactic arguments, but not more: all other dependents must be adjuncts. An‘Agent-of ’ phrase embeds no other Agent-of phrase recursively, but a Theme-phrase. So the‘intransitive’ structure in (35) can be transitivized, in which case the additional head v introduces another argument, and we get two (necessarily different) theta-roles:
(41) [[Ł2DP] v [VP [Ł1DP]]]
We then obtain, formal-ontologically, two events, but crucially not two separate ones, and not two of the same kind: one will be embedded or contained in the other as a proper part: an event of killing Bill, say, will involve an event of ‘making dead’ performed by an Agent, which is followed by its causal consequence, the state of the Theme’s being dead. So event embedding comes at the price of the embedded event becoming a proper part of the embedding one. The V-phrase that is embedded in the v-phrase is again not referentially independent, but integrated with the higher head via a grammatical relation, described mereologically as that of‘part’ (see Section 2.5 above). And again, the process is sharply limited: after transitivizing a phrase [Bill [dead]] so as to obtain The bride made Bill dead, which can be spelled out as The bride killed Bill, we cannot transitivize again, with (44) meaning that Elle made the bride kill Bill:
(42) [Bill [dead]]
(43) The bride made [Bill [dead]] = The bride killed Bill (44) *Elle killed the bride Bill.
A general principle, then, seems to veto the embedding of the same in the same. Embedding in grammar is constrained by having the embedded element play a rela- tional role, specifically becoming a part in whatever the embedding phrase refers to.15
15 Our claim here goes counter to Ramchand (2008), who argues that argument structure in syntax is
‘constructed systematically on the basis of primitive recursive syntactic relationships’ (p. 16), which on her view involve three primitive ingredients in the human conceptualization of event-structure: a causing or initiating sub-event, a process-denoting sub-event, and a sub-event corresponding to a resulting state.
In this sense, embedding has an inherent content, a grammatical meaning, and is not ‘blind’ or ‘formal’. The same is true in clausal embedding, where the embedded clause necessarily plays the role of a content of a mental event referred to by the matrix verb, or in NP–NP embedding, where the embedded NP plays a role as a Possessed, a Theme, or a location:
(45) John’s mother (46) The bride’s purse16
(47) The vase on the table
All of these instantiate our template, and they are single referential expressions, with (47) clearly referring to one thing, a vase, which is identified through a descriptor encoding the predicate‘being on the table’, which is related to the higher DP via the quasi-thematic role of location. This is also clear from Agreement facts:
(48) The vase on the table is/*are broken
What is not possible is a juxtaposition of two referential DPs, without one of these playing a role as part in relation to the other:
(49) *John mother (50) *the bride purse (51) *the vase the table
This is so unless the two DPs are part of an enumeration, as in (52), where the grammar computes two referents, and the Agreement facts change:
(52) The vase, the table are/*is broken.
This architecture may explain why the basic pattern is that V embeds N and not the other way around. Suppose that‘V’ and ‘N’ are lexical ‘features’, which are intrinsic to the specifications of a lexical item. Then they are atomic and independent, and we predict nothing about how they will relate. That is, we will have to stipulate how they relate: through relations of semantic or categorial ‘selection’. Relatedly, the same problem arises if we take a lexical and intrinsicalist view of events and objects: if these are what is denoted by V and N, respectively, they will again be independent of one another, and their relation, if any, will have to be stipulated. On the present model,
Although these components can be combined (‘merged’, on her model), however, and the result has compositionally determined readings, the combinatorics is strictly templatic rather than unboundedly recursive: these primitives only compose in a highly limited and predetermined fashion. Similar remarks apply to the notion of‘composition’ of the basic elements of argument structure in Hale and Keyser (2002: ch. 1).
the mystery of why V embeds N and not the other way around dissipates and in fact becomes unformulable. The explanation of the generalization is that event-denota- tion is not lexical: it is grammatically derived. Unlike object-reference, which is mono-phasal, event-reference is necessarily bi-phasal. So any object-phase will be embedded in an event-phase, and it is a general constraint on embedding that embedding-as-argument in grammar creates a part-whole relation between embed- der and embedded. So objects will be parts of events, not the other way around. Put differently, V contains N, necessarily, exactly as events contain objects.
In a similar way, T contains v/V as part of its internal architecture; non-finite T will be part offinite T; and finite T will be a part of C. None of these, then, quite simply, will be lexical notions. Interpreting them as‘features’ fails to recognize their nature as grammatical and derived. Similarly, putative relations of ‘selection’ holding between one feature and others are a misleading way of stating what is in effect a grammatical dependency: intuitively, an event has whatever parts it does, but it doesn’t ‘select’ its parts, any more than a human body selects an arm. Nor does it make sense that within a phasal template, D selects N, or C selects T. Rather, every referential expression is based on the lexical predicate it is based on, and the two pairs necessarily form a unit (though the edges can be missing, as discussed in more detail in Chapter4). Where edges are differentiated in a more fine-grained way, as in Rizzi (1997), a selection relation between functional projec- tions again makes no sense, as the relevant hierarchy of projections is hierarchically ordered.
In line with our stress on the finiteness of the process of creating part-whole relations in the formal ontology, the process as a wholefinds an ultimate limit in the assignment of the truth value, which contains literally everything else as a part and hence cannot itself be a part. More precisely, there is nothing, at the point of reaching the truth value, that the grammar can do so as to change the formal ontology. Anything that can happen, such as‘slifting’ (53), assertions of veridicality (54), adding tags (55), or evidentials (56), leaves the formal ontology and the truth value itself unchanged:
(53) The bride was killed, I believe. (54) It is true that the bride was killed. (55) The bride killed Bill, didn’t she? (56) The bride killed Bill—I saw it.
The essential non-recursivity of grammar as depicted above, and the reinterpretation of the recursivity of grammar in terms of a generation of an ontology structured by a part–whole relation, now predicts that there will be no recursivity in the elements added in (53–6) after the truth value is assigned: and indeed, there is no such recursivity:
(57) *The bride was killed, I believe I hope.17
(58) It is true that it is true that the bride was killed.18 (59) *The bride killed Bill, didn’t she did she?
And where languages encode evidentiality morphemically, recursive stacking of such morphemes is unavailable. Why then is it that the grammatical process should come to an end with the truth value assigned? Why is there no further ontological category that grammaticalizes? We return to this question in Chapter9.
Let us now come back to the operation Merge and to the standard claim that grammar is recursive, with Merge being its one most essential element. One could point out that our phasal template contains a combination of elements, and hence presupposes something like Merge. Indeed this is the standard justification of Merge as the basic operation of grammar: where grammar is not said to be based on it, the operation is somewhere hidden in whatever operations are assumed. Now, we agree of course that the template contains a combination of elements—but then, saying this about our template adds no insight to it. Indeed, one might call it an abstraction from grammar, given that the symmetries that Merge builds precisely arise by abstracting from the fact that no symmetric structures ever arise when two syntactic objects merge. Merge thus throws no light on the observation that symmetric structures in grammar are banned (Kayne,1994; Moro, 2000; diSciullo, 2005).
If recursion in grammar is modeled through Merge, moreover, embedding will typically be interpreted‘formally’, in line with the standard assumption that there is no grammatical meaning and form and content are distinct. This means that semantic effects of embedding as described above, relating to the emergence of a formal ontology structured by part–whole relations, have to come from something other than grammar. But where should we go to derive the effects? The lexicon, we have argued, as much as the putative non-linguistic C-I systems, are the wrong place to look (even if they were to support some initial stages of the formal-ontological hierarchy, before we come to the part that depends on grammar). Nor would it help,