2 quotes that speak to you with an analysis of why they are significant to development? – 1 Question (with an answer to your

Beyond formalities: The case of language acquisition

MICHAEL TOMASELLO

The Linguistic Review 22 (2005), 183–197 0167–6318/05/022-0183 c©Walter de Gruyter

Abstract

Generative grammar retained from American structural linguistics the ‘formal’ approach, which basically effaces the semantic and pragmatic dimensions of grammar. This creates serious problems for an account of language acquisi- tion, most especially the problem of how to link universal grammar to some particular language (the linking problem). Parameters do not help the situa- tion, as they depend on a prior linking of the lexical and functional categories of a language to universal grammar. In contrast, usage-based accounts of lan- guage acquisition do not posit an innate universal grammar and so have no linking problem. And if children’s cognitive and social skills are conceptual- ized in the right way, there is no poverty of the stimulus in this approach either. In general, the only fully adequate accounts of language acquisition are those that give a prominent role to children’s comprehension of communicative func- tion in everything from words to grammatical morphemes to complex syntactic constructions.

1. Introduction

In the 1960’s Chomskian generative grammar effected a revolution in linguis- tics. But in many ways it was only a technical revolution. American struc- tural linguists were looking for formal distributional patterns in languages, and Chomsky (following Harris) created some especially powerful formal tools for doing this, namely, syntactic transformations. Transformations opened up new avenues of inquiry and enabled the discovery of such things as island con- straints – arguably one of the important advances in 20th-century linguistics.

American structural linguists were not interested in psychological reality, and so they explicitly eschewed any account of how people use language to

Bereitgestellt von | MPI fuer evolutionaere Anthropologie Angemeldet

Heruntergeladen am | 22.12.15 10:20

184 Michael Tomasello

express meaning and communicate – that was considered a problem for psy- chology. In apparent contrast, Chomsky argued that linguistics is a branch of cognitive psychology, and that universal grammar is an innate faculty of the hu- man mind. Curiously, however, in this psychologizing of linguistics, commu- nicative function was still considered irrelevant for characterizing grammatical constructions – with attempts in this direction, for example by generative se- mantics, being explicitly rejected. In this respect, Chomskian linguistics may be seen as the theoretical successor to American structural linguistics, since both ignore for principled reasons the semantic and pragmatic dimensions of grammatical constructions.

The failure to break with structural linguistics on this crucial point has de- termined much of the history of generative grammar (Langacker 1987, Lakoff 1990). Of particular importance is the account of language acquisition. From the beginning, Chomsky stressed that the question of how human beings ac- quire linguistic competence is the key empirical phenomenon – because it is here that we most clearly discern ‘the poverty of the stimulus’ and the way that universal grammar (UG) must fill in the missing information. But a crucial challenge is presented by the fact that children may be born into any one of several thousand different languages – which vary from one another in myriad ways – and so UG must also have a way of ’linking’ itself to particular lan- guages. The initial idea in the early 1980’s was that the linking problem could be solved with a smallish set of innate bi-valued parameters. But specifying exactly what these parameters are has turned out to be extremely difficult, and moreover, it has become apparent in recent years that parameters are of no theoretical help until the linking problem is solved – one cannot set the head direction parameter until one can identify heads in the particular language be- ing acquired. And the linking problem derives, ultimately, from the exclusive focus on the formal dimension – to the neglect of the functional dimension – of grammar.

In this article I argue that generative grammar is not an adequate theory of language acquisition – either with or without parameters – mainly because of the linking problem. I propose an alternative theoretical account that explicitly recognizes the functional dimension of grammatical constructions and the cog- nitive and communicative sources from which constructions emanate histori- cally and ontogenetically. This account does not rely on any form of universal grammar and so the linking problem does not arise.

2. The problem of linking

Each animal species eats only certain things, and so they are born knowing how to identify those things perceptually. In some animal species individuals

Bereitgestellt von | MPI fuer evolutionaere Anthropologie Angemeldet

Heruntergeladen am | 22.12.15 10:20

Beyond formalities: The case of language acquisition 185

are born alone and live mostly solitary lives, but when it is time to mate they know how to perceptually identify a conspecific of the opposite sex. There is no problem with organisms having adaptive specializations that contain pre- existing "knowledge" of something they are likely to encounter in their envi- ronments.

The problem is that language cannot work in this way. It cannot work in this way because each of the world’s 6000-plus natural languages is different perceptually, and so any universal that might be a candidate for pre-existing knowledge looks different in different languages. For example, let us say that you have an innate idea of “subject of a sentence” or “clausal head”. You now encounter the following utterance (from Slave, an Athebaskan language of northern Canada; Rice 1989, as cited in Dryer in press):

(1) T’eere li ráreyiht’u.

What is the subject of this sentence? Or the clausal head? Does it help if you are told that the English translation involves ‘girl’, ‘dog’, and ‘hit’? The general problem is formulated by Fodor (2001: 765) in this way:

For UG to be truly helpful, it should supply innate sentence structures and fix their relation to surface words strings . . . As long as there is substantial crosslin- guistic variation with respect to how innately defined structure is overtly lexi- calized, there will be ambiguities of string-to-structure alignment.

The only way out of the impasse is function. Thus, the only explicit theory addressing the problem is Pinker’s (1989) theory of semantic bootstrapping. Pinker proposed the following: (a) a list of key syntactic categories innately given to all human beings, (b) a list of key experiential functions innately given to all human beings, and (c) a set of innate linking rules to connect the two. In the case of ‘subject of a sentence’, as one instance, the process would work in the following way. ‘Subject’ is innately linked to ‘agent of an action’. Thus, if the child saw a girl hit a dog and heard someone say “The girl hit the dog”, she would know through perception and her general causal cognition that the girl is the agent of the action; then, her innate linking rule would automatically con- nect agent to subject. Because she notices the linguistic form associated with the subject, the child can also now recognize future exemplars of ’sentence sub- ject’ on the basis of this form alone (say a particular word order configuration or a particular case marker), even if they are not agents. Thus, the English- speaking child will eventually have to deal with experiential subjects that are not agents (as in John saw Mary) and even passive sentences in which subjects are not agents and agents are not subjects.

However, in the specific case of ‘sentence subject’ it is almost certain that Pinker’s proposal is not correct. On general theoretical grounds it has been

Bereitgestellt von | MPI fuer evolutionaere Anthropologie Angemeldet

Heruntergeladen am | 22.12.15 10:20

186 Michael Tomasello

known for some time that in ergative languages the notion of ‘subject’ does not operate like it does in English and other accusative languages, and so a direct connection to agent is not possible (and some languages probably do not have subjects at all; Croft 2001). Moreover, even if there were some solu- tion to this problem, there are many languages that are so-called split ergative: some of its constructions are ergative while others are accusative based on such things as person or tense (DeLancey 1981; van Valin 1992). In general terms, Slobin (1997) has made a persuasive case that there is much too much vari- ability across languages for any static and innate look-up table to function in the way it would need to solve the problem of linking (see also Braine 1992). Empirically, the problem with Pinker’s proposal is that children do not always begin with agentive subjects. Many early subjects come from such non-causal utterances as I like it, Maria have it, I see it, and It has a hole in which there is no agent of an action at all, and sometimes there are even direct violations in early speech such as Pete hurt by car (patient = subject, agent = oblique; Tomasello 1992). And there are no other syntactic roles or grammatical rela- tions – that is, other than sentence subject – for which there is any detailed analysis or supportive evidence at all.

So what could possibly work? The only realistic possibility is that children construct from their experience with a particular language some kinds of gram- matical categories, based on the function of particular words and phrases in par- ticular utterances – followed by generalizations across these. Then they could look to see if any of these categories corresponds to anything in UG. But then the notion of UG becomes superfluous. If the child is constructing grammatical categories anyway, then what added value is provided by UG?

3. Do parameters help?

On first blush, one might imagine that the notion of parameters would be help- ful in solving the linking problem. But actually the opposite is the case: pa- rameters are of no help until the linking problem is solved since parameters presuppose the existence of certain lexical and functional categories. The point is made quite nicely by Mazuka (1996: 24–25):

Setting a Head Direction parameter by analyzing the syntactic structure of the input involves a paradox. The Head Direction parameter is supposed to deter- mine the order in which the head and complement should appear in the language the child is acquiring. But, for a child to set this parameter, she must first find out which units are the heads and the complements in the sentence she hears. If her linguistic skills are sophisticated enough to know which are heads and com- plements, she will also know which order they came in. If she already knows

Bereitgestellt von | MPI fuer evolutionaere Anthropologie Angemeldet

Heruntergeladen am | 22.12.15 10:20

Beyond formalities: The case of language acquisition 187

which order the head and the complements come in a sentence, there is no need to set the parameter.

The hard part is thus recognizing ‘heads’ and ‘complements’ in a particular language, and this difficulty is logically prior to any act of parameter setting. That this recognition is far from trivial is highlighted by Fodor (2001: 761):

It would have been more convenient for children if natural language parameters were all concerned with surface facts . . . Instead . . . the relation between word strings and their parametric generators is opaque at best.

It is also telling that there is no agreement among generative grammarians about what kinds of parameters there might actually be, that is, what kinds of things might be parameterized. Again, Fodor (2001: 734) is especially helpful:

It might turn out that there are 20 parameters or 30 or 100 and more . . . On one recent estimate . . . there would be at least 32 parameters controlling the landing site for verb movement, perhaps multiplied by the number of possible verb forms (finite/infinitive/past participle, etc.).

And in point of fact the two major proponents of a theory of parameters, Mark Baker (2001) and Janet Fodor (2003) provide lists with very little overlap (see Table 1).

Empirically, there are only a few truly cross-linguistic tests of parameters, and they lead to uniformly negative conclusions. For example, perhaps the best- studied parameter is the null subject (pro-drop) parameter. It specifies that the following features should go together in a language (i.e., because they all re- sult from the setting of the parameter): the possibility of null thematic subjects in tensed clauses, null nonthematic (expletive) subjects, subject inversion, and that-trace violations. But as Newmeyer (1998: 357–359) shows, once we get away from European languages on the basis of which this parameter was for- mulated, the correlations simply do not hold. The other widely cited parameter is the head-direction parameter, and it fares no better in cross-linguistic per- spective. In a typological survey of a sample of 100 languages, Dryer (in press: 57) demonstrates that there are many languages where, again, the hypothesized correlations (in this case concerning the placement of heads in various types of phrases and clauses) simply do not hold. And to the extent that these corre- lations do hold, they are best explained by general processing considerations (Hawkins 1994).

It should also be pointed out that, a priori, the proposal of innately pre- specified parameters of variation in a cognitive skill is an extremely implau- sible biological mechanism. To my knowledge, no one has proposed anything remotely similar for any other cognitive or social skills that human beings pos- sess. For example, such skills as music and mathematics are, like language,

Bereitgestellt von | MPI fuer evolutionaere Anthropologie Angemeldet

Heruntergeladen am | 22.12.15 10:20

188 Michael Tomasello

Table 1. Lists of hypothesized parameters from the two most well-known theorists of parameters, with possible overlaps listed on the same line

Baker (2001) Fodor (2003)

Polysynthesis Adjective Neutralize Head Directionality Ergative Case Topic Prominent Obligatory Topic Subject Side Subject Initial Verb Attraction Subject Placement Serial Verbs Null Subject Null Subject

Object Final Complementizer Initial V to I Movement I to C Movment Question Inversion Affix Hopping Topic Marking Wh- Movement Obligatory Pied Piping Null Topic

unique to humans and universal among human groups, with some variations. But no one has to date proposed anything like Universal Music or Universal Mathematics, and no one has as yet proposed any parameters of these abili- ties to explain cross-cultural diversity (e.g., +/−mathematical variables, which some but not all cultures use in such things as algebra – or certain tonal pat- terns in music). Perhaps these empirical and logical problems with parameters are the reason that many contemporary theories (e.g., Minimalism; Chomsky 1995) no longer make use of them.

So parameters as theoretical constructs have not worked empirically, and the immediate prospects for progress do not seem promising. Parameters as- sume a range of formally specified lexical and functional categories, whose cross-linguistic status is problematic at best (Croft 2001), and whose linking to particular languages in the cases where they do apply seems to be pretty much impossible (unless we want to posit an absolutely incredible number of innate linking rules). One possibility, then, is that the search for formal param- eters is a mistake, and we would be much better served by trying to figure out how children construct the particular grammatical categories of their particu-

Bereitgestellt von | MPI fuer evolutionaere Anthropologie Angemeldet

Heruntergeladen am | 22.12.15 10:20

Beyond formalities: The case of language acquisition 189

lar language, with the question of universality being an empirical one – with what language universals there are resulting from universal processes of human cognition, communication, and vocal-auditory processing.

4. Is there really a poverty of the stimulus?

The fundamental argument for the existence of an innate universal grammar is the argument from the poverty of the stimulus. Chomsky has made this clear in a number of places, and it has recently been reiterated by Crain and Pietroski (2001). The problem is that the argument is formulated in terms of a formal generative grammar as adult endpoint and a child who has available only be- havioristic learning theory – which enables her only to string words together in a Markov chain (with no understanding of phrasal organization or any other structure-function correlations), making blind associations and inductive infer- ences in the process (with no conceptual understanding of linguistic function at all).1 But, as Tomasello (2003) argues, there is no poverty of the stimulus if (1) language is conceived as a set of symbolic instruments for directing the intentional and mental states of others (with complex linguistic constructions being one kind of form-function pairing for doing this; Goldberg 1995); and (2) children are given credit for possessing the cognitive and pragmatic skills necessary for learning such meaningful linguistic symbols and constructions, including such things as categorization, analogy, statistical learning, competi- tion among structures, and so forth.

An excellent example of the problem appears in arguments for so-called structure dependency. Chomsky (1980) gives the following example of ques- tion formation in English.

(2) a. The man is tall. b. Is the man __ tall?

(3) a. The man who is smoking is tall. b. *Is the man who __ smoking is tall? c. Is the man who is smoking __ tall?

The idea is that forming simple questions such as “Is the man tall?” could be done on the basis of either of two hypotheses: move the first-occurring auxiliary to the front or move the auxiliary from the main predicate to the front. To differentiate between these two hypotheses children supposedly need to see

1. In their recent textbook, Crain and Lillo-Martin (1999) have section headings on classical and instrumental conditioning (which have basically been dead in the study of language acquisi- tion for close to half a century), whose inadequacies in accounting for language acquisition they then go on to demonstrate.

Bereitgestellt von | MPI fuer evolutionaere Anthropologie Angemeldet

Heruntergeladen am | 22.12.15 10:20

190 Michael Tomasello

examples like 3c in which the subject NP contains a relative clause with an auxiliary (which did not move to the front). Chomsky (1980: 40) has famously claimed that children almost never hear such sentences. But in an analysis of some written corpora and corpora of child directed speech, Pullum and Scholz (2002) find many of just the right kind of examples that children need, such things as:

(4) Can those who are leaving early __ sit near the door?

(5) Is the boy who was crying __ still here?

(6) Could those who are coming __ raise their hands?

But actually, if one thinks about it for a bit, the child does not really need to encounter such sentences at all. If the child understands NPs with relative clauses – if she understands that the whole phrase is used to make one act of reference – then there would never be any temptation to extract an auxiliary from it; she would simply understand that that unit stays together as one func- tional unit. It may be said that this is simply another way of stating that the child understands structure dependence. True. And that is the point. If we al- low the child to have some notion of meaning or function, then she understands the structure of the sentence to the extent needed to form a conventional En- glish yes-no question. Modern functionalist theorists are not behaviorists who believe the child works with unstructured linear strings, but rather they are cognitivists who believe in structure – just not of the purely formal kind.

A related argument may be made about island constraints. For example, the so-called subjacency constraint posits, among other things, the impossibility of questioning an element embedded too far down and in the wrong way in a sentence structure. For example, 7b seems to be a question form of 7a, but the parallel pair in 8 does not work in the same way.

(7) a. Bob believes that Jill married Sam. b. Who does Bob believe that Jill married __?

(8) a. Bob believes the rumor that Jill married Sam. b. *Who does Bob believe the rumor that Jill married __?

Subjacency precludes the possibility of moving a WH-word out of an embed- ded clause which is part of a complex NP (the rumor + clause), as in (8b). Subjacency is supposed to demonstrate the argument from the poverty of the stimulus because there is no relevant evidence available to the child in the input (Chomsky 1986).

But van Valin (1998) recasts the constraint by focusing on the role of infor- mation structure in question formation: T he element questioned must be in a clause within the potential focus domain of the sentence. The potential focus

Bereitgestellt von | MPI fuer evolutionaere Anthropologie Angemeldet

Heruntergeladen am | 22.12.15 10:20

Beyond formalities: The case of language acquisition 191

domain may be determined by various kinds of tests including a negation test. Thus, with respect to (7), if we ask “Does Bob believe that Jill married Sam?”, we may say “No, Jeff" (i.e., he believes she married Jeff), which shows that the final NP (Sam) is in the potential focus domain. We may not do the same with (8). If we ask “Did Bob spread the rumor that Jill married Sam?”, and we answer “No, Bill”, this can only mean that Bill is the one who spread the rumor. In general, van Valin (1998: 232) explains the restriction in this way:

Questions are requests for information, and the focus of the question signals the information desired by the speaker. It makes no sense, then, for the speaker to place the focus of the question in a part of a sentence which is presupposed, i.e. which contains information which the speaker knows and assumes that the hearer knows or can deduce easily [e.g., in an object complement, as in (8) – MT]. The content of adverbial clauses and restrictive relative clauses is nor- mally presupposed, and consequently constructing questions with the focus in one of these structures generates a pragmatic contradiction.

We do not need to go into details here. It is enough to simply note that if one takes function into account – in this case pragmatic notions of topic and fo- cus – there is no poverty of the stimulus since even the most abstract syntactic principles from generative grammar may be understood as emanating from the language users’ sensitivity to the communicative function of the pieces of lan- guage she is using.

Other examples could be adduced. For example, the binding principles are supposed to require the purely formal syntactic operation c-command – be- cause again the child is supposed to have no relevant linguistic experience – but a much more natural account based on the notion of cognitive domains in discourse is provided by Van Hoek (1997, 2003) – an account which explains not only intra-sentential constraints, as the c-command account, but also inter- sentential constraints in discourse as well. Again, the details are not important for current purposes, but what is important is that there is no poverty of the stimulus if we assume that children are cognitive beings seeking to make sense of the linguistic communication addressed to them.

5. What is the alternative?

So if humans are not born with an innate universal grammar, then how do they construct the abstract dimensions of grammatical structure and why are there universals in this process? The key in usage-based approaches is to explic- itly recognize the semantic and pragmatic functions of grammatical construc- tions, that is, to recognize that linguistic constructions are nothing other than complex linguistic symbols with a form and function (Langacker 1987, 1991;

Bereitgestellt von | MPI fuer evolutionaere Anthropologie Angemeldet

Heruntergeladen am | 22.12.15 10:20

192 Michael Tomasello

Goldberg 1995). They are constructed by communities of people historically, via processes of grammaticalization, and re-constructed by individual children ontogenetically.2

In this view, there is no need to posit a specific genetic adaptation for gram- mar because processes of grammaticalization can actually create grammatical constructions out of concrete utterances – and grammaticalization is a cultural- historical process not a biological one. Thus, it is a historical fact that the spe- cific items and constructions of a given language are not invented all at once, but rather they emerge, evolve, and accumulate modifications over historical time as human beings use them with one another and adapt them to changing communicative circumstances (Croft 2000). Through various discourse pro- cesses (involving various kinds of pragmatic inferencing, analogy making, etc.) loose and redundantly organized discourse structures congeal into more tightly and less redundantly organized constructions (see Traugott and Heine 1991, Hopper and Traugott 1993, for some recent research). Universals of language structure emerge not from an innate universal grammar, but rather from the si- multaneous interaction of universals of human cognition, communication, and vocal-auditory processing in the process of grammaticalization.

In acquisition, children are confronted not with abstract grammatical princi- ples but with particular, language-specific utterances. For example, the English passive construction conveys an abstract meaning packaged in a certain infor- mation structure: an entity has something done to it (with the agent either not expressed or expressed obliquely). Because semantic and pragmatic function are part of the definition of the construction, children are presented with no linking problem and there is no poverty of the stimulus. Indeed, children can learn to use this or any construction appropriately and productively only by witnessing events in the world and attempting to discern the speakers’ commu- nicative intention in referring to these events with a certain pattern of linguistic symbols – in much the same way they pair a simple word with some type of environmental scene(s). Given that each language has its own lexical and gram- matical conventions, there is really no alternative to this basic account of how children acquire the productive use of the constructions of their language.

2. Some people may doubt that cultural-historical processes can create abstract structures such as those embodied in the grammatical constructions of modern-day languages. But, although the analogy is clearly not perfect, there are many highly abstract structures in modern math- ematics that could only have been created by cultural-historical processes since they are not universal among cultures (e.g., those of algebra and calculus). Again, there are many dis- analogies between language and mathematics (which is more closely related, both logically and historically, to written language). The only point is that abstract symbolic systems can be created by groups of human beings working together over historical time in the domain of mathematics, and so perhaps they can also be created in similar yet different ways in the domain of language.

Bereitgestellt von | MPI fuer evolutionaere Anthropologie Angemeldet

Heruntergeladen am | 22.12.15 10:20

Beyond formalities: The case of language acquisition 193

Children then find generalizations that apply across some subset of gram- matical constructions and end up with a structured inventory of grammatical constructions, organized in an inheritance hierarchy similar to that used to de- scribe knowledge representation in other branches of cognitive science (Gold- berg 1995). Importantly, there are many subsets of constructional patterns in every natural language, including ones that are highly idiosyncratic as com- pared with the rest of the language (e.g., the ‘let alone’ construction, Fillmore et al. 1988; the ‘nominal extraposition’ construction, Michaelis and Lambrecht 1996; the ‘WXDY’ construction, Kay and Fillmore 1999). The constructional approach attempts to account for all different kinds of constructions, from sim- ple words to abstract constructions that are canonical or idiosyncratic. And indeed, the existence of idiosyncratic constructions – and also ‘mixed’ con- structions that are partially canonical and partially idiosyncratic – create insur- mountable problems for generative theories that posit two sets of acquisition processes: one for the lexicon and other unruly linguistic structures (words) and another for those that conform with the core requirements of universal gram- mar (rules) (cf. Pinker 1999). There are many constructions that do not fall neatly into either of these categories. One can also make the argument that if the irregular …