If everything is syntax, why are words so important? An a-morphous but non-lexicalist approach

5.1 Why concepts and not roots?

At this point it would be reasonable to ask why am I assuming that what is categorized are concepts and not, as has become standard in the literature, roots. Up to now I have been closely following the (more or less divergent) neo-constructionist models (such as DM and XS), but with a crucial difference: instead of the categorization of roots I propose the categorization of concepts. The reason for this discrepancy is clear: because I take seriously the asymmetry described in Section 2, the theory I am outlining is not morphological, but purely syntactic. Perhaps a problem with neo-constructivist models is that they are not sufficiently emancipated from morphology. Recall that my central hypothesis is that there is no link between sound and meaning below the level of the syntactic word, so it makes no sense to introduce purely morphological elements (such as roots and affixes) into syntactic derivations. Anyway, the idea that what is syntactically categorized are conceptual elements is implicit in the discussion about the nature of words in the discourse of anti-lexicalist authors: “Natural language syntax operates on units that are standardly characterized as bundles of features. Such features are lexicalized concepts. Syntax creates ever-larger molecules by combining featural atoms through iterated use of Merge” (Boeckx 2008: 63, [my italics]).

Let us recall now that as early as the thirteenth century speculative grammarians known as Modistae (see Bursill-Hall 1971) claimed that the Latin terms doleo ‘to hurt’ and dolor ‘pain’ had the same meaning but different modi significandi (‘modes of meaning’). At the same time, they said that meanings are not relevant for grammar, which focused on the different modes of meaning. When I argue that a functional category takes a concept as a complement and makes it syntactically computable, I mean that the essential function of syntax is to establish computations between concepts by means of functional categories. In other words, syntax is the only way to mix independent concepts to produce new and more complex concepts (and, ultimately, thought). See Pietroski (2011, 2012) for a development of these ideas.

In fact, Embick and Marantz’s (2008)categorization assumption (CA) establishes conditions for roots very similar to those we are taking now for the concepts that the computational system recruits from the conceptual system:

Roots cannot appear (cannot be pronounced or interpreted) without being categorized; they are categorized by merging syntactically with category-defining functional heads. If all category-defining heads are phase heads in Chomsky’s (2001) sense – that is, if they are heads that initiate spell-out – the categorization assumption would follow from the general architecture of the grammar. (Embick and Marantz 2008: 6)

But note that this condition is much more natural if applied to conceptual elements than if applied to roots, because roots are defined as fragments of words that do not include categorizing morphology, a circular definition. Moreover, as I have suggested, it is in fact the architecture of FL which imposes the CA, and does so in the hierarchy s-word > p-word by virtue of the asymmetry discussed in Section 2.

Allow me to persist with this idea for a moment. If we consider the very reasonable considerations of Panagiotidis (synthetizing the literature on root categorization), it can be seen that they are much more applicable to concepts than to roots: “interpretable categorial features, borne by categorizing heads, define the fundamental interpretive perspective of their complement, thus licensing root material” (Panagiotidis 2015: 78). In fact, later on he says: “grammatical categories, such as ‘noun’ and ‘verb’, are particular interpretive perspectives on concepts” (Panagiotidis 2015: 84), and he suggests that given that “uncategorized roots are FLN-extraneous” they “would not be recognized at the interface between syntax and the Conceptual-Intentional/SEM systems” (Panagiotidis 2015: 94). But again, this restriction seems more reasonable if applied to concepts than to roots. Stating that “roots are essentially ‘imported’ into the syntactic derivation” (Panagiotidis 2015: 94) fits perfectly with the idea that his roots are actually concepts, as indeed he himself seems to suggest: “the ability of FLN to manipulate roots enables it to denote concepts and, ultimately, to be used to ‘refer’” (Panagiotidis 2015: 94).

In the present proposal, an essential lack of concepts is precisely that they are not directly connected to the sensorimotor system. What dolor and doleo share from the semantic point of view, although it is obviously crucial to understand these words, is not a linguistic unit, that is, not a sound/meaning pairing, because the concept only connects with sound when it is within a word. Note that this implies, contrary to the usual models, that the computational system (syntax) is not fed on lexical units (be they conceived as roots, morphemes or words), but directly computes concepts by means of functional categories. Consequently, the assumption I make is that the connection between concepts and sounds (i.e. phonological representations) is necessarily mediated by syntax, and therefore that only categorized concepts (i.e. syntactic words) can have a relation to the phonological component. Borer, although posing a similar model in many respects, produces the opposite reasoning. She defines roots as phonological indexes that have no meaning. ^[10] In Borer’s model roots only receive meaning when realized as phonological forms. But there is something contradictory in such a concept of root, as Borer herself seems to appreciate: “there is also a theoretical claim put forth here which is likewise less tan self-evident: that a root in itself need not, and possibly cannot, serve as an independent domain for phonological spellout” (Borer 2013: 24).

Indeed, in the world’s languages it is common that roots are not pronounceable in isolation, which lends credence to the hypothesis that what enters the syntactic computation is the “concept” and not the “phonology” of the “root”. In fact, considering that roots have only phonological form clearly implies that roots are purely morphological objects, not syntactic ones, which is what our model predicts for roots (and morphemes in general).

On the other hand, if categorization has the effect of imposing an “interpretive perspective” on what is categorized, then it seems more reasonable to think that what is categorized is a concept, and not a root. For a concept it is natural to receive an interpretive perspective, whereas for a root it is not (especially if defined as a phonological index). The reason for this state of affairs, as I have suggested, relates to the asymmetric architecture of the essential components of FL. If a concept is only interpretable if categorized (assuming the arguments about roots of the DM tradition), then the categorization requirement is imposed by the CI system, which is primary with respect to the SM system.

5.2 Why are concepts categorized?

The categorization assumption is indeed a stipulation, but has some interesting (although speculative) theoretical foundation. The basic idea is that categorization makes formally homogeneous concepts that by their very nature are heterogeneous and incompatible. As Boeckx (2008) has suggested, categorization (lexicalization) imposes a common format to all concepts. This would allow lexical items to combine under their new shared format, rather than being limited to their natural affinity (i.e. purely semantic or conceptual). This in turn would mean that concepts that originally reside in different mental modules (and that possibly would be opaque to each other) can be combined to give rise to new concepts:

It is quite possible that what is at first a formal restriction on lexical items is the source of a cross-modular syntax of thought – giving rise to a full-blown language of thought, arguably the source of our Great (mental) Leap Forward at the evolutionary scale. (Boeckx 2008: 78)

Following the same line of thought, Ott (2009) notes a certain consensus in the field of comparative psychology on the fact that humans and other species share much of their conceptual systems. As Hurford has observed, “some (not all) of a human system of common-sense understanding precedes a system of language, both ontogenetically and phylogenetically” (Hurford 2007: 87, cited in Ott 2009), which would imply that such systems are not part of the evolution of human language, but predate such a faculty. It seems that both human neonates and other animals have considerable conceptual abilities, but animals, unlike what happens as humans develop, seem unable to integrate these various “mental languages”. This is what Ott (2009) calls Hauser’s paradox. Ott suggests that the ability to associate concepts with words (lexicalization) is the key to explaining this paradox, in the sense that concepts could be combined beyond their modular constraints, resulting in a productive system that would transcend the limits of core-knowledge domains, such as social relationships or spatial reasoning:

If these speculations are on the right track, the significant cognitive gap between humans and non-linguistic animals is not the result of a profound remodeling of the pre-linguistic mind. Rather, the sudden addition of recursive syntax, paired with a capacity for lexicalization, plausibly led to the explosive emergence of symbolic thought that paved the way for modern human behavior. (Ott 2009: 267 [my italics])

That “capacity for lexicalization”, which is precisely what in this paper I have identified with categorization, is thus a syntactic operation on concepts. In this context, it is not strange that the first phase of computation (categorization) is the minimal phase connected with the phonological component. And it is therefore expected that the first phase is memorized for the purpose of interaction with other speakers. But what is matched is not a “sense” and “sound”, but a morphophonemic structure (a p-word) and a syntactic structure (including an s-word).

The lexicon, thus understood, is not the input of the processes of syntactic derivation, but part of the output. In a certain sense one could say that syntax describes a kind of loop, because it produces derivations from concepts and “returns” them to the CI system creating new concepts (as shown by the vertical arrows in Figure 1). As Boeckx and Ott suggest, syntax decapsulates conceptual systems of various types and puts them at the service of a single computational system. In our specific version, syntactic categories take concepts belonging to different cognitive niches and make them formally homogeneous and computable. That is, s-words “rescue” concepts from “darkness”, multiplying the computational power of our Kind in an unexpected way in the natural realm.

5.3 How are concepts categorized?

Henceforth I will assume Baker’s (2003) theory of syntactic categories, with some modifications adopted from Panagiotidis’ (2015) model. Although these two models are different, they share a crucial trait: in both of them categories (or categorial features) are interpretable, in the sense that they have a specific function within the CI system (to provide “interpretive perspectives on concepts”, according to Panagiotidis’ formulation).

Following Baker (2003) I assume that nouns, verbs and adjectives (N, V and A) are the only lexical categories that exist, but following Panagiotidis (2015) arguments, I do not assume that A is a category of the same level as N and V, or that it can be considered the default lexical category, as proposed by Baker. ^[11] Thus, N and V are basic functional categories that take a concept and turn it into a syntactic word. It could be said, then, that the functional categories N and V are different “flavors” of what Chomsky (2008) calls an Edge-Feature.^[12] In this sense, the syntactic category is the property that makes a concept computable. In the likely event that categories themselves are interpretable, then we could say that the computational system (the Merge function implemented through an Edge-Feature) has the mission of building more complex concepts (such as dolor or doleo) which previously would not be available for the mind, and, of course, for language.

The computational system (as indicated by the vertical arrows in Figure 1) links concepts together and creates more complex concepts without adding anything new, apart from the structure. The rationale for the s-word > p-word hierarchy is based precisely on this fact: if only the concepts included in s-words are relevant for the language of thought (a primary relationship), it is expected that only what is (at least) an s-word will be susceptible of externalization (a secondary relationship). Another way of putting it would be to say that s-words exist prior to p-words and independently of them.

The first operation of syntax is therefore that of converting concepts into syntactic words (N or V). Following Panagiotidis (2015) I will assume that N and V assign to the concept they categorize a particular interpretive perspective (e.g. “extension in space” in the case of N and “extension in time” in the case of V). Syntactic words, under conditions which vary from language to language, are externalized, that is, are paired with phonological representations (phonological words), and these representations will eventually be stored.

The relevant question now is whether this global vision can explain the typically idiosyncratic character of words, that is, if it is the fact that words are the minimum limit for externalization which explains their idiosyncrasy.

6.1 What makes words special?

The externalization of syntactic derivations has the requirement that certain fragments of syntactic derivation are stored in the motor system, i.e. they are integrated into memory linked to phonological forms. Traditional models (and many current ones) tend to identify the stored fragments with the terminal nodes (as shown on the right side of Figure 3), ^[13] while the model I am proposing, along with the NS model, posit a different match, shown on the left side of Figure 3. In this diagram we see that the syntactic “atoms” (A, B, C, D, E) do not match the “phonological atoms” (W1, W2, assuming that the “phonological atoms” accessible to the computational system are not acoustic features or phonemes, but the smallest units of the prosodic hierarchy which are not sensitive to purely phonological information, that is, p-words).

Figure 3:

Two models of the correlation between “syntactic atoms” and “phonological units”.

Again I invoke the asymmetric linking of CS with CI and SM to explain this discrepancy: the relation between CS and CI is “optimal” and primary, while the relation between CS and SM is “imperfect” and secondary. The “size” of the smallest units of externalization (p-words) would be determined by inherent properties of SM (e.g. respiration) and are unrelated to the internal language of mind. ^[14]

The main difference between this model and the DM, NS and XS approaches is that I assume that lexical entries have the minimum size of phonological words, that is, I conclude that traditional morphemes lack syntactic and semantic properties. I think it can also be argued that the lexical limit of “idiosyncrasy” is better predicted with this model than with the “first phase model” (e.g. Marantz 1997; Embick and Marantz 2008; Arad 2003 or Panagioditis 2015). In these approaches it is stipulated that the limit of lexical idiosyncrasy is demarcated by categorization, understood as a first phase of syntactic derivation. ^[15] Without prejudging here whether or not categorizing nodes are phases in the Chomskyan sense (see Fábregas 2014 for reasonable objections), it is important to note that the exception to compositionality can not follow solely from the fact that categorization constitutes a phase, since higher phases (v, C) have no such effect, at least systematically. In the present model the limit of idiosyncrasy is justified because the first connection of the computational system with SM is set in the syntactic word, and it is this first connection that involves memorization.

Note that in purely conceptual and computational terms (i.e. in the language of thought) no idiosyncrasy is possible, since all elements (conceptual and functional) are compositionally interpretable. There can only be idiomaticity when the meaning of a linguistic form (e.g. recital or irascible) does not follow from its (presumed) structure; but that can only happen, obviously, when the structure is associated with a phonological form. Idiomaticity does not exist in the inner language of thought, it only exists in the mismatches that occur between the syntactic structure (the meaning, ultimately) and the morphological and phonological structure, part of the SM system.

6.2 Deriving words

Adapting the vision of root categorization of DM to our model, we can say that basic functional categories (F) take a concept from the conceptual system and project through merge in a construction, as shown in Figure 4:

Figure 4:

Syntactic categorization of a Concept.

By definition F can only be V or N, which impose an interpretive perspective on the concept. Following Baker’s (2003) model, it can be assumed that the category V takes a concept as a complement and adds a specifier (a subject). Therefore, if F is V, we have the scheme of Figure 5:

Figure 5:

Verb derivation.

This would be the basic scheme of an unaccusative verb or a fragment of a transitive verb, in the sense that the specifier (X) is the internal argument of the verb. To simplify the argument we assume that V is the category that in some theories is represented as Aspect (Borer 2005b) or Process (Ramchand 2008) and that establishes the relation between the internal argument and the process phase of the event. Although for clarity and convenience I use the traditional labels (V, N), these represent interpretable functional categories and are not only categorial labels. ^[16] The syntactic categorization of the concept has two simultaneous consequences: it converts a concept previously isolated in a computable unit, and categorizes it determining its subsequent behavior in the derivation (and, ultimately, of course, its interpretation).

Let’s assume that the concept of the above scheme is the concept underlying the verb destroy. If the construction of Figure 5 is selected by v (corresponding to the category v of Chomsky 1995, voice of Kratzer 1996 or Init in Ramchand 2008), then we have a new verbalization of the derivation, with the addition of a new specifier (the external argument or initiator), as shown in Figure 6.

Figure 6:

Transitive verb derivation.

So far we have a derivation that roughly means that an argument Y initiates a process that happens to the argument X, a process that is defined by the selected concept (that of ‘destroy’). Note that the terminal nodes of this structure (without considering specifiers) are not words or morphemes, that is, they are neither listemes (the sound/meaning pairings of Borer’s 2005a; Borer 2005b model) nor roots or affixes (as in DM and in Borer 2013). In fact, beyond the inventory of functional categories (which we assume to be universal and limited), there are no lexical entities involved in the derivation. The structure of Figure 6 is purely syntactic and could be externalized as a sentence fragment or as a single word. In other words, on this approach there is still no systematic matching of sounds and meanings; there are only conceptual entities and functional categories.

The hypothesis that the matching between sense and sound is produced only through syntax implies that only once a concept is categorized (is an s-word) is a candidate to be linked to a phonological form (p-word). The structure in Figure 6, once the arguments have been evacuated, could be externalized in English with the p-word destroy, but not, for example, in Spanish. According to Spanish morphology such a sequence is unpronounceable as a p-word, since finite verbs in Spanish express (at least) tense and agreement features, which do not appear in the scheme. Thus, the derivation of a finite verb in Spanish must also reach the functional node T, as in Figure 7:

Figure 7:

Finite verb derivation.

I have suggested the hypothesis that only categorized derivations can have phonological realization. This is a necessary but not sufficient condition. The particular range of materialization of heads in a structure like that of Figure 7 will depend on the morphological structure of p-words (i.e. on the extent of their paradigms). According to this hypothesis, the concept shown in Figure 7 would not be a possible candidate for externalization (in any language), so we assume that it is incorporated into its head (V in this case). ^[17] V would already be a possible candidate, but not in Spanish. Therefore V incorporates to v, but the result remains unsatisfactory in Spanish. Only an incorporation to T will meet in the same continuous span of heads all the features expressed by a Spanish verbal form, such as, for example, destruyó ‘(he/she) destroyed’ (ignoring for the moment singular and third person phi features). Only then can the materialization (and linearization) of the span [T[v[V concept]]] with the p-word destruyó happen in Spanish. ^[18] Note that we can consider the structure in Figure 7 as a schematization of the computational processes that allow interpreting the p-word destruyó. Therefore, we can say that the diagram in Figure 7 is also the “meaning” (the signifié in the Saussurean sign) of the “signifier” destruyó. Thus, the I-lexicon entry of destruyó would include, minimally, the morphological and phonological elements needed to build the exponents (p-words) of the verb paradigm, that is, the “signifier”, connected to the fragment structure shown in Figure 7, that is, its “meaning” (which includes a link to a conceptual element, a perspective of interpretation, an agentive head, and a tense specification). In this model the meaning of lexical units (words) is not stored in the I-lexicon, but it is always “calculated” in the interaction between the computational system and the conceptual-intentional system.

To complete the perspective, let us consider the projection of concepts as nouns, i.e. nominalization. According to Baker’s (2003) theory of lexical categories, nouns are characterized by what he calls the identity criterion, that is, sameness. Baker proposes that only nouns have a semantic component that makes it legitimate to ask whether X is the same as Y. Baker notes that nouns, verbs and adjectives have application criteria, so that by knowing what dog means we can identify which objects are dogs, knowing what blue means we identify what things are blue, and knowing what crying means we recognize who is crying. But, according to him, only nouns “set standards by which one can judge whether two things are the same or not” (Baker 2003: 101).

Moreover, Borer (2005a) develops a detailed and sophisticated theory of the functional categories that transform a root into a noun. Borer basically argues that nouns are the listemes dominated by three functional categories: CL(assifer), Q(uantity) and D(eterminer). The closest category to the listeme (the root) in her model is CL, which is related to plural number. According to Borer, plurality has the effect of fragmenting a continuous interpretation in what we might think of as reticles or partitions. The subsequent specification of the Quantity head (with a cardinal or a quantifier) selects a number of cells fragmenting the denotation of the noun. ^[19]

It is interesting to note that Baker’s model, based on the (more diffuse) idea of the identity criterion, is also related to plurality and quantity. Indeed, by definition, plurality is a requirement to count. But it is not the only one. To count dogs, for example, it is not only relevant that there is a plurality of dogs, but that we are sure about whether the next dog is different or equal to the previous one (whether it is the same or not). Meanwhile, mass nouns cannot be counted, but can be measured. And in order to know how much water there is, we must be sure that the water we are measuring now is not the same water that we have already measured. Therefore, it seems reasonable to assume that the basic nominal category is the category that provides the identity criterion or sameness to nouns, and that legitimizes the upper category of plural number.

It does not seem unreasonable to assert that a requirement for plurality is singularity, so I assume that the first nominalizing category is singular number, which will be responsible for the identity criterion (sameness). ^[20] I also assume without further discussion that morphological gender in languages such as Spanish is the expression of singular number, that is, the categorizer. ^[21] Therefore, I will represent nominalization as the merge of a concept to category N (for singular number, which also – e.g. in Spanish – can be masculine or feminine). Plural number (n), equivalent to Borer’s CL head, will therefore be another superior functional category selecting N as a complement. Thus, the derivation of a singular noun such as libro ‘book’ in Spanish (or book in English) will be like the scheme on the left in Figure 8; the derivation of libros (or books) would imply the scheme on the right. As in Figure 7, the concept is incorporated into its head (N), where it gets singular number (i.e. identity criterion, nominality), and the resulting complex is incorporated into n, where it obtains plural number and count reading.

Figure 8:

Noun derivation.

Ignoring for now of the complex semantics of nouns, the key idea is that spell-out is always phrasal, regardless of the morphological complexity of words in a given language. Compare the representation of the words libro and book:

Note that unlike what is assumed in other neo-constructionist models, I am not assuming that morphemes are realizations of syntactic nodes, as would be tempting in the case of Spanish where there is a clear correspondence between N and the suffix -o and between the root libr- and the concept, a correspondence that is expandable to plural (with n being -s):

This is essentially the strategy of neo-constructivist models. However, the proposal I have set out here is based on the fact that syntactic nodes are typically sub-morphemic (Starke 2009). It is well known that the phenomena of fusion and suppletion, typical of many languages, challenge the claim that morphemes correspond to terminal syntactic nodes (see Anderson 1992, Chapter 3, for a full review). Moreover, although I am arguing that words are syntactic structures, it is also known that the claim that the internal structure of words can be explained as the result of the application of syntactic rules to constituent morphemes is inadequate (again, see Anderson 1992, chapter 10 in this case, for a detailed discussion). This morphemic view, characteristic of models such as Selkirk (1982) or DM, implies that morphemes are manipulated by syntax to create words, and that the same syntax then continues to operate with words to create larger structures. No doubt these models are the most theoretically attractive and elegant, but perhaps they are not the most empirically adequate.

I therefore suggest that the schemes in Figure 9 are the correct ones, while the diagram in Figure 10 involves an inappropriate mix of syntactic and morpho- phonological entities. ^[22] The central idea now is that the greater correspondence between both levels in Spanish is a simple historical accident, and therefore it is alien to the computational system. It is certainly true that the morphology of Spanish uses the scheme in Figure 7 to produce the p-word destruyó (and the diagram in Figure 10 to produce the p-word libros) in a different way from how the morphology of English does in producing destroyed (or books), but the syntax is alien to these differences, something that only is clearly reflected in the schemes in Figure 9.

Figure 9:

Parallel derivations of libro and book.

Figure 10:

Derivation of Spanish libros ‘books’.

The hypothesis that I propose is that the correlation between syntactic complexity (compositionality) and morphological complexity (analyzability) is contingent, accidental, and in a certain sense weak, something expectable if we take the asymmetric architecture of FL seriously.

A brief description of so-called derivational morphology will better enable us to show the plausibility and scope of the present model.

6.3 Deriving derived words

I anticipated that the present model, although it is not lexicalist, implies a paradigmatic morphology, and this is true both as regards inflectional morphology and derivational morphology, unlike Borer’s XS model. However, I will assume Borer’s (2013) analysis (common with DM) of derivative affixes as (re-)categorizing functional heads, that is, as more specific versions (more complex in their constituent features) of the basic categories (N, V, and A). But I do not assume that syntax manipulates affixes, because in this model syntax only handles functional categories and concepts, not morphemes. The relation between (re-)categorizing heads and affixes, as stated at the end of the previous section, is not direct or deterministic, although it does often reveal a certain isomorphism and (as discussed in Section 8) it may have an important role in the acquisition, storage and use of p-words.

Thus, the structure that would provide interpretation to the p-word destruction (similarly for Spanish destrucción) would be that in Figure 11, consisting of the merge of the derivation of Figure 7 with the head N, instead of T:

Figure 11:

Syntactic structure of destruction and destrucción.

The representation in Figure 11, inspired by Borer’s (2005b, 2013) proposal for so-called complex event nominals, is intended to show that destruction, as used in Hannibal’s destruction of Rome (or, in the similar Spanish La destrucción de Aníbal de Roma) is not a noun formed from a verb, as a lexicalist theory would argue. Thus, there is no need to postulate a complex system of morphological adjustment rules, or a mechanism of arguments inheritance. The eventive use of destruction implies that the concept merges with V, which creates an event on an argument (Rome), and which is then merged with v, creating an initiation by Hannibal of the event, this then categorized as a noun, with identity criterion, i.e. with singular number (and, in Spanish, feminine gender). And this is exactly what the diagram in Figure 11 represents. Apart from many relevant details of the derivation (such as the system of case assignment to the arguments), the representation in Figure 11 shows that the meaning of destruction is not obtained from the lexical entry of the verb destroy and the subsequent application of rules of word formation, but is obtained, as in “normal” syntax, from the computation of concepts without grammatical properties through the basic operations of syntax and its functional categories. Figure 11 shows the subordination of an event to a noun, as would happen in the construction Hannibal’s action of destroying Rome (or in Spanish La acción de Aníbal de destruir Roma): in fact, the underlying structure of these NPs (in both English and Spanish) would be essentially the same as in Figure 11.

At this point I would like to make clear a common sense idea that, in light of the proposed model, may be tractable in a formal theory of language: complex words are actually the result of the reanalysis of regular syntactic constructions as “lexical units” (i.e. as syntactic structures associated with single p-words). The compositional meaning of destruction is action of destroying and indeed both expressions have roughly the same underlying syntactic structure (that in Figure 11).

If it is true that language change is limited to the component internalized from the linguistic environment (the I-lexicon, which is the only area in which realignments between meaning and sound can be produced), and if it is true, as our model establishes, that morphology is part of the internalized component (and is therefore the outcome of linguistic change), then it is inevitable to conclude that (inflected and derived) complex words are historical accidents, and therefore that morphemes are purely morphological residues of history. I said before that morphemes are not signs. We might now add that morphemes are actually “ghosts” (or “fossils”) of ancient words.

I will return to this in Section 9, but first we must address the crucial question of how a neo-constructivist model can account for “lexical integrity” effects and for the typical (and unexpected) lack of compositionality (and of productivity) of derivative morphology.

7.1 The lexical integrity hypothesis

The apparent lack of derivational continuity between the internal and the external syntax of words (i.e. on the left and right sides of P-RaD in the scheme of Figure 2) is one of the most robust objections to a syntactic view of the internal structure of words. Ackerman and Neeleman (2007: 332), for example, argue that if the lexical and phrasal syntax were the same, then it would be expected that a noun that is incorporated into a higher affix could strand its complements or modifiers. Indeed, this is often impossible. Consider the Spanish example of (1):

If the internal structure of this derived word (zapatero ‘cobbler’) is built in syntax, then there is no way to explain why the example of (1b) is ungrammatical (which indeed it is, if intended to represent the sense ‘Cobbler that mends leather shoes’). The objection is relevant, but it only affects “morphemic” theories postulating that morphological word formation is performed in the syntax; that is, in our example, that derive zapatero ‘cobbler’ from zapato ‘shoe’, and that derive zapato from the root zapat-, according to the scheme in Figure 12.

Figure 12:

A morphemic syntactic derivation of zapatero.

Indeed, if the scheme in Figure 12 is a legitimate syntactic object, nothing should prevent the incorporation of N (zapato) into the superior N (-ero), producing (1b). ^[23] Any stipulation to prevent this process would add complexity to the system, eliminating the advantage over a lexicalist interpretation, or making the former indistinguishable from the latter. As Bermúdez-Otero (2012: 50) points out, this is like turning a component into the waste bin of another one. ^[24] However, the system I describe here, based on the hypothesis that word formation, even if it is done in the syntax, does not work with words, roots or affixes, can avoid such an objection while maintaining the null hypothesis that there is only one generative component in human language. Consider the derivation in Figure 13 as an alternative to the “morphemic” derivation of Figure 12:

Figure 13:

A syntactic derivation of zapatero.

The selected concept (the same concept associated with the Spanish word zapato ‘shoe’) is incorporated into V forming a verb (which would explain that the meaning of zapatero ‘person who mends shoes’ involves the activity of repairing shoes). V, in turn, is selected by v, introducing an initiator of the event (represented by an empty pronoun PRO in the scheme). Following Fábregas’ (2012) analysis of agentive derivatives with the Spanish suffix -dor (‘-er’), we could stipulate that the affix -ero corresponds to a set of features that include a D feature (which enables it to saturate an argument position, in this case the event initiator) and an N feature (which makes it a nominalizer). Following the derivation proposed by Fábregas, N is “re-projected” taking v as a complement and nomializing it. ^[25]

According to the left diagram in Figure 3, above, I postulated that a structure like that in Figure 13 is associated in memory with the phonemic form zapatero, which allows us to say that this structure is the meaning of the p-word. Note, crucially, that a similar structure would underlie the noun phrase Persona que repara zapatos ‘Person who mends shoes’ (which includes a relative clause). What matters now is that the speaker links such a fragment of syntactic structure to a single p-word (zapatero). Or to put it in another way, that such fragment of derivation corresponds to a p-word in the speaker’s I-lexicon. Syntax is thus a uniform computational system that allows us to “calculate” the meaning of linguistic expressions combining functional categories and concepts, while ignoring the issue of whether or not there are p-words available for particular structure fragments, or whether they are morphologically complex or not. The Spanish word zapatero is a nominal element with a “relative clause” which is materialized as a single p-word. We might also point out here that there is no place in Figure 13 for a possible prepositional complement of zapato, precisely because zapato does not appear as a noun in the derivation. And this fact explains the impossibility of Example (1b) without throwing weeds to the neighbor’s yard.

The most important implication is that, contrary to intuition (that of both speakers and linguists), I’m assuming that zapatero is not derived from zapato, but that the two words share the same concept, part of their structure, and also part of their phonological form (i.e. they form part of a paradigm), while they differ in the functional categories involved in their internal structure.

The prediction of this model is, therefore, that the interpretation of words, beyond Saussurean arbitrariness, is always compositional. The explanation for the abundant cases in which this expectation is not met is that in such examples the structure associated with the phonological form is different, regardless of morphological structure. The present approach then, following the DM framework, adheres to the principle of Full Decomposition: “No complex objects are stored in memory; i.e. every complex object must be derived by the grammar” (Embick 2015: 17). Nevertheless, I disagree with the notion of “complex object” used in this framework. Embick argues that complex objects are those objects “that consist of more than one morpheme” (Embick 2015: 17), assuming that they are derived syntactically “every time they are employed” (Embick 2015: 17). Under the present approach, the principle of Full Decomposition applies to the internal elements of words, but assumes that they are purely syntactic (functional categories and categorized concepts) and not morphological. Thus, I will show that the hypothesis that “every word and every phrase is built every time it is used” (Embick 2015: 17), as well as being theoretically interesting, is in fact more empirically adequate once we adopt the (“a-morphous”) alternative notion of “complex object”.

7.2 Compositionality

Indeed, it may be objected that Full Decomposition only works for compositional derivatives (or for simple words), which does not avoid the fact that we have to postulate a specific lexical entry when the meaning of a derivative includes features not present in the structure or when parts of the (alleged) structure are not interpreted. This, as we know, is extremely common. Indeed, the Spanish word zapatero typically means ‘person who mends shoes’, but not things like ‘person who buys shoes’, ‘person who draws shoes’ or ‘person who falsifies shoes’. However, a scheme like that in Figure 13 predicts that zapatero should also have these meanings (and anything else that involves a transitive verb). ^[26]

It is important at this point to return to the essential function of syntax according to the diagram in Figure 1: to create more complex concepts from simpler concepts. The specific meaning of zapatero ‘person who mends shoes’ (as is the case with the meaning of most derived words) is a complex mixture of compositionality and idiosyncracy. Therefore, if (as we have assumed) interpretation is always compositional, we must stipulate that the structure in Figure 13 has additional information, for example, assuming that the verbalizer V is associated with a certain conceptual element and not others, as in the scheme in Figure 14:

Figure 14:

An alternative derivation of zapatero.

Figure 14 is intended to show that the structure associated with the p-word zapatero in Spanish (i.e. its meaning) specifies that the event involved in the verbalization (the Concept-2 in the scheme) is related to the concept ‘repair’ (or ‘make’) but not to ‘falsify’ (or ‘destroy’, etc.). Note that no complexity is added to the system, as Figure 14 also would be the underlying structure of the analytical version Persona que repara zapatos ‘Person who mends shoes’. Syntax is capable of building very specific and refined concepts (which is no surprise, in that this is basically its raison d’être), with the peculiarity that only a very small number of these complex and specific concepts, contingently, are associated with memorized phonological forms. Thus, in Spanish, the complex concept ‘person who repairs shoes’ (whose derivation is that in Figure 14) has an associated p-word (zapatero), while the complex concept ‘person who destroys shoes’ does not, although both complex concepts are perfectly legitimate and compositionally generated by syntax. As I noted above, the diagram in Figure 14 is actually the non-lexicalist version of the Saussurean linguistic sign: the structure (including the basic concepts) is the signifié and the p-word (zapatero) is the signifiant. The potential signifié ‘person who destroys shoes’ has no specific signifiant, and therefore it is not usually considered a sign.

7.3 Morphological compositionality?

The problems of lack of compositionality begin when the interpretation of a morphologically complex word does not correspond to its supposed structure. But here the emphasized predicate is crucial. In a purely syntactic approach (as is the case here) the interpretation of words is always compositional (i.e. fully decomposable) because the morphological structure is irrelevant (invisible) for syntax. In fact, a morphologically complex word may have a relatively simple syntactic structure. For example, transmission with the meaning ‘gearbox’ would have a syntactic structure similar to that in Figure 8 (left side), while its compositional version would have a structure similar to that in Figure 11. Of course, the relation between morphological structure and syntactic structure is not arbitrary, but the fact that the relation is not arbitrary does not necessarily imply that it is isomorphic or that one structure determines the other.

In fact, they are structures of a different nature. Syntactic structure is hierarchical (the result of iterated merge), while the structure of derivative morphology is linear (i.e. concatenative). ^[27] The claim that morphological structure is the result of the historical reanalysis of syntactic constructions predicts such a difference, since the syntactic structure “belongs” to the computational system, while the morphological structure “belongs” to the sensorimotor system. And this is so because the morphological structure is the result of the externalization (linearization) of syntactic structure (as reflected in the left side of Figure 3, in which a flat sequence, W, corresponds to a fragment of hierarchical structure). If we consider the formulation made by Everaert et al. (2015) about the asymmetry discussed in Section 2, that is, “what reaches the mind is unordered, what reaches the ear is ordered” (Everaert et al. 2015: 740), we can conclude that morphological structure is part of the linearization of the hierarchical structures generated by syntax, and therefore it is not itself hierarchical, but linear.

This model predicts that there is a (non-deterministic, but psychologically real) correlation between morphological complexity and syntactic complexity, so that a greater degree of morphological complexity corresponds to greater syntactic complexity, and vice versa. This is clearly seen if we contrast the representation of book (Figure 9) and that of destruction (Figure 11). These representations clearly illustrate the correlation between the complexity of syntactic and morphological structures, and also allow us to understand the temptation of correlating these structures deterministically, a common feature of morpheme-based theories. However, it is easy to find, in all languages that have derivative morphology, morphologically complex expressions with unexpectedly simple underling syntactic structures. Thus, we can compare the structure associated with transmission in its reading of a complex event nominal (left side of Figure 15) with the structure associated with the same p-word in its idiomatic reading ‘gearbox’ (right side of Figure 15, in which it is assumed that the concept involved is different):

Figure 15:

Two representations of transmission.

Borer (2013: 500) proposes an explanation for the phonological form of the complex event nominal through an incremental derivation by phase from a hierarchical morphological structure, but it then remains unexplained why the non-compositional versions (or those with smaller internal structure, as in the case of R-nominals) have exactly the same morphological and phonological structure. The phrasal spell-out model that we assume, along with a paradigmatic conception of derivational morphology, explains more naturally this unexpected identity. On the one hand, the conception of derivatives as the externalization of phrases explains the tendency towards the correlation between syntactic complexity and morphological complexity (both in inflectional and derivative morphology). On the other hand, the fact that this externalization is the result of a historical reanalysis (see Section 9) explains the possible disparity of complexity in pairs such as those in Figure 15. Thus, the morphologically complex structure of transmission is explained as a consequence of the fact that it is the materialization of a complex derivation (‘the action of transmitting’), whereas the same morphological structure found in transmission ‘gear box’ is explained as the result of the impoverishment of the original syntactic structure, again the result of a reanalysis. ^[28] It is important to note that the impoverishment of the original syntactic structure involves the association of the derivation of the new interpretation to a different concept (in this case, the same one that would underlie the compound gear box). ^[29]

The vision of the lexicon that emerges from this model coincides with the notion of I-lexicon of Figure 1: a “passive” vocabulary of formants (p-words) associated with syntactic derivations produced by the computational system. ^[30] But if there are no word formation rules in the lexicon, and if syntax does not manipulate morphemes to create words, what is the source of complex words? What is the source of the intuition that some words are created from others, and that different words share a smaller number of constituents? A possible answer to these questions comes from two notions of traditional grammar: paradigm and analogy.

10.1 Why are words so important?

Non-lexicalist theories often conclude that words are epiphenomena. Julien (2007), for example, states that “the discussion of whether complex words are formed in the syntax or prior to syntax is futile, because words as such are not formed in the grammar at all. They are not grammatical entities” (Julien 2007: 210), which leads her to conclude that “the concept of ‘word’ has no theoretical significance in the grammar at all” (Julien 2007: 212). It seems reasonable to claim that there are no word formation procedures, but this does not mean that words are not created. Words are created in syntax when concepts are categorized (s-word), and they are created in morphology when syntactic structures are materialized (p-word). The two types of words do not necessarily coincide, as there is a hierarchy s-word > p-word and the matching range depends on the morphological structure (i.e. on the uninterrupted history) of each language. In the model I have outlined in this contribution, syntax determines the internal structure of words, but crucially it does not determine their morphological structure, a task for morphology and history. ^[34]

What people learn, use, recognize, utter, forget, recall, and miss are words, not morphemes. Julien herself (2007) notes that words have a greater appearance of psychological reality than morphemes, which is surprising if it is true that syntax operates with morphemes and that words do not exist. The explanation Julien offers is that the reason why words are more accessible to the speaker has to do with their distributional properties: “since words are the minimal morpheme strings that can be used as utterances and that may be permuted more or less freely, words are the minimal linguistic units that speakers can manipulate consciously” (2007: 234). In light of this, “word-internal morphemes, by contrast, cannot be consciously manipulated in the same way, and consequently, word internal morphemes are less salient than words in the awareness of speakers” (Julien 2007: 234). Such an explanation is clever and interesting, but fails to explain why words have the distributional privilege that morphemes lack. The most natural explanation is to state quite clearly that words acquire their distributional independence by virtue of being the smallest units of syntax that connect to the SM component. Morphemes are just that, pure forms that more or less accurately “recapitulate” the internal structure of words. The morphological rules of a language do not determine the internal syntactic structure of a word, but determine its shape. In thus sense, morphology is morphology.

The model I have presented is actually a variant of DM, on which it is based. The essential difference is that DM is a morphemic model, with the difficulty inherent in all models that mix syntactic and morphological entities. In the present approach syntax does not work with morphemes, but only with syntactic categories, incorporating concepts through merge (categorization). However, the morphological theory involved is realizational (late insertion) in the sense that morphology operates with fragments of derivations to produce or select forms within paradigms. Thus, the words cloud and cloudy are related, but not derivatively. Both are constructions on the same concept, though certainly they do not express the same meaning, in accordance with the syntactic derivation of each one.

The terminal nodes in the syntactic trees of lexicalist models are usually words, while they are morphemes in non-lexicalist accounts, so that sentences are made directly with morphemes without the intervention of the notion of word. The model outlined in these pages vindicates the relevance of words, not as lexical units, but as fragments of syntactic derivation linked to phonology. The central question is to what extent this conception of the word (i) really explains those aspects that support the lexicalist hypothesis, and at the same time (ii) permit us to dispense with it.

Williams (2007: 356) summarizes these issues as follows:

Williams argues that either the facts of (2) are not so, or “we need something like the Lexical Hypothesis” (Williams 2007: 356). The non-lexicalist model I have outlined suggests that (2a) and (2c) are not really indisputable facts, showing that the internal structure of words is just regular syntax (although synthetically externalized, so to speak). And both (2b) and (2d), to the extent that they are different, are explained through the hypothesis that words are fragments of derivation sensitive to phonological spell-out. The early materialization (spell-out) makes these derivations opaque to other syntactic processes, predicting lexical atomicity effects.

Lexicalist theories, like traditional theories based on Indo-European morphology, start from words and project them in syntax. Syntactic theories unmake the word, pushing it aside as an epiphenomenon, and focus on the syntactic construction of sentences by means of morphemes. The present model aims to combine the best of both traditions: on the one hand, the word is constructed syntactically, it is syntax, but on the other hand, the word (as a phonological/morphological form) exists independently as a system of materialization and linearization of syntactic structures. Words therefore do not exist as stored lexical units prior to syntax, but exist as phonological forms stored and organized in (more or less extensive) paradigms. Syntax forms derivations using functional categories and concepts; an essential part of the derivation is the categorization (lexicalization) of concepts, at which point the phonological component is accessed and a word (that is, a fragment of derivation associated with a phonological form or a paradigm) is selected or created.

For all these reasons, it is perhaps not surprising that, as Anderson (2015) notes, Saussure did not use the term morpheme once in his posthumous and celebrated Cours. He did not perceive the complex syntactic structure underlying words, but he did understand that morphemes were not an indispensable part of the essence of languages, as suggested by the fact that bound morphemes, unlike phonemes, words or sentences, are not universal in languages (although remarkably common).

10.2 Where is morphology?

The model developed over the course of these pages, although it vindicates the word, is radically anti-lexicalist in the sense that morphology is located in the sensorimotor system, that is, outside the FLN. Moreover, as reflected in Figure 1, in this model the lexicon (defined as I-lexicon) is also “external” and belongs to the sensorimotor system. Following the important insight of DM, I conclude that what is distributed is not the morphology, but the lexicon. The famous Saussurean signifiant belongs to SM, and the no less famous (and elusive) signifié belongs to CI. According to the present framework only syntax (CS) establishes a link between them. The relation between syntax and CI is strictly compositional, while the relation with SM is, although sometimes isomorphic, accidental and variable.

The conclusion that morphology is part of SM is consistent with recent theoretical developments in the NS framework: “Morphology, we argue, may be reduced entirely to the function that spells out the syntactic tree by choosing and inserting phonologically contentful lexical items” (Bye and Svenonius 2012: 428). But not surprisingly it is also consistent with findings from avowedly lexicalist perspectives:

Modern Hebrew verb roots and their alternation classes, like those of English and Latin, thus furnish yet another example of morphology by itself, leaving lexical meaning to reside where it belongs, not in roots, which are purely grammatical objects, but in lexemes, where language meets the world. (Aronoff 2007: 828)

The possibility that traditionally opposing positions can be considered (to a certain degree) complementary, capturing the empirical advantages of both, is a direct consequence of the hypothesis that the syntactic word is the minimal connection between CI and SM.

I have assumed a paradigmatic morphology without entering into its formal details. In general the model described here is consistent with Bermúdez-Otero’s vision according to which “morphology connects the output representations generated by the syntax with the underlying forms in the input to the phonology” (Bermúdez-Otero 2012: 50). According to his strictly modular model of the relationship between syntax, morphology and phonology, the morpheme (the morph in his terms) constitutes “the representational currency of morphology”. In this approach, morphological operations are strictly modular only “if they treat morphemes as inalterable units, and never change their syntactic specifications or their phonological content”. As the author notes, “in this overall conception of grammar, only syntax manipulates syntactic features, and only phonology manipulates phonological features” (Bermúdez-Otero 2012: 50). ^[35]

The image Bermúdez-Otero proposes of morph(eme)s as insects trapped in amber is indeed an elegant metaphor, and provides a satisfying reflection of the vision I have proposed of morphemes as a result of history (and as tools of memory):

[A] morph is like a translucent droplet of amber encasing a fossilized insect, the phonological content of the morph being like the body of the insect itself, and the morphology like the laboratory of an entomologist working with a collection of such specimens. (Bermúdez-Otero 2012: 51)

Beard, a representative of another influential lexicalist account, concludes that “the simplest and most consistent universal theory of morphology will represent grammatical morphemes as purely phonological operations on lexemes or phrasal positions, not as listable objects” (Beard 1995: 378), a position that is also compatible with the vision of morphology set out in my proposal.

I have suggested that the lexicon of languages is a paradigmatic collection of p-words stored in the sensorimotor system (but see footnote 18). The idea that the lexicon is distributed implies that the meaningful part of the traditional linguistic sign belongs to the conceptual-intentional system. Note that on this view, the traditional central components of human languages (meaning and sound) actually belong to the “external” components of the FL in Hauser et al.’s (2002) model, and not to its (supposedly) specifically human and specifically linguistic core, which may be surprising. The hypothesis that words are always syntactic constructions may allow us to better understand why the human computational system creates, in its interaction with the “external” systems, a specifically human system of knowledge and communication. The key idea, then, is that the meaning of words (like the meaning of phrases) is not encoded in the lexicon (which is purely morphological and phonological), but is the result of the interaction between the computational system and the conceptual-intentional system.

10.3 What is the meaning of words?

According to the present model, words do not copy or translate extra-linguistic concepts, but simply compute them syntactically. When a person sees a book, we assume that he/she accesses his/her concept of book, not the meaning of the word book. Therefore, when instead of seeing a book a person hears the word book, we should continue to assume that he/she accesses the book concept, not the meaning of the word. But when we use language we do not access the concept directly, but through the syntactic structure of the word book. It could be said then that syntax contributes to the creation of a substitute for the perceptual representation of the book itself. Language actually builds substitutes of perceptions and emotions. The sentence You saw that man provides sophisticated and complex instructions to replace in our interlocutor’s brain his/her past perception of such a man. Language allows the brain to experience perceptions and states that are not real, and what is interesting is that it does so through operations that emulate perceptual operations. Such operations are modulated by functional categories, including the basic ones (V, N, A), profiling concepts as events, objects or properties, and others, such as tense, aspect, definiteness, etc.

People accumulate concepts because it is the only thing that their brains understand, and when they think, what they do is to manipulate concepts, to relate them, to compare them, to position themselves with respect to them. But to do this, concepts are not enough; functional categories and syntax are needed. People do not normally seek to communicate concepts but want others to reproduce in their brains those operations that we make with concepts: we want to communicate the relations between concepts, how we perceive them, how we understand them, how concepts move us. And for this reason words do not correspond to concepts, but to syntactic computations. When we tell someone I’ve seen the book we do not want to talk about the concept of book or about the concept of seeing (unless you’re a philosopher): we want to reproduce in his/her mind the scene of ourselves seeing the book as if our interlocutor had been there. The only strategy to do so is to compute concepts with functional categories and translate such computations phonologically (or visually). We select the concepts most related to the objects and events involved, and we build structures using functional categories that provide an emulation of the perceptual stimuli that the listener would have had if she/he had been there. Perhaps this explains why when we understand the word walk the brain circuits that we actually use to walk are activated (Pulvermüller et al. 2005): because the word walk, besides its syntax, includes the “motor concept” walk itself, not a meaning or a sign of it.

A language that is not used to speak does not require words (p-words): concepts and syntax would do the job. But such a language allow us to access only our own concepts and our own experiences. It is conceivable that we might in this way have an even rich inner life, but it would almost certainly be different. When connecting the CI system with the SM system we associate a computation (which includes concepts and instructions for interpreting them) to a phonological form. This can be interpreted as an additional system of memory (and, if Jackendoff 2012 is right, as an enhanced access of thought to the conscious mind). In the process of syntactic derivation we associate a piece of computation to a phonological form (e.g. book). But that is not a concept, it is, as it were, a point of view on a concept, it is a set of instructions for computing a concept (see Pietroski 2011). The minimal computation (the syntactic word) enters a new memory system: the phonological form. The phonological form (morphology and phonology) is a “translation” of a syntactic computation into the motor system. This allows the storage and use of computations, and also the learning of computations from others. When we learn a language, we learn how to do computations on concepts from what others have done before. And that makes the difference, too.