Of Minds and Language (60 page)

Fig. 20.1
illustrates what CP is. This graph (made after Liberman et al. 1957), shows how native speakers of English perceive two distinct phonemes in an acoustic continuum: stimuli 1–4 are perceived as /b/, whereas stimuli 5–8 are perceived as /d/, and the perceptual change is sharp, as the different lines show. To the person's ear, the sound “changes” to another sound at one point in the acoustic continuum, so that the line goes down sharply.

Language was central in the discovery of this perceptual mechanism, which was originally explained by Liberman et al. (1957), and was taken as evidence that speech is perceived differently from other types of auditory stimuli. At that time it was thought that CP was acquired and language-specific. Later, Eimas et al. (1971) found CP in babies (1–4 months), which meant it was an innate mechanism. A few years later, Kuhl and Miller (1975) successfully trained chinchillas to perceive the voicing contrast between /da/ and /ta/ categorically. In short, CP is innate, but it is not restricted to speech or speech-like stimuli and occurs with stimuli that bear no resemblance to speech sounds (Harnad 1987). In fact, even crickets have been reported to show signs of CP (Wyttenbach et al. 1996).

So here is a perceptual mechanism that is probably essential to understanding and explaining certain architectural properties of language categories such
as discreteness, a fundamental property of phonemes and words. But this perceptual mechanism is not specific to language or to our species, though it is innate and involved in language development and perception. This does not render it irrelevantor uninteresting for a language researcher, of course, but it clearly makes it a poor candidate for UG because it operates in a broader domain.

Another example of an innate mechanism that appears very significant for language is found in the study of the perceptual salience of rhythmic/prosodic properties of speech. Interestingly, the history of its discovery raises a similar point to the one in the previous example. It was originally discovered that newborns are very good at discriminating language groups based on rhythmic information: they can discriminate their mother's language-type using this information (Mehler et al. 1988, Cutler and Mehler 1993, Nazzi et al. 1998, Ramus and Mehler 1999). This capacity is already functioning at the time of birth, and it makes a good candidate for a language-specific mechanism. Recently, however, it was learned that tamarin monkeys (Ramus et al. 2000) and rats (Toro et al. 2003, Toro 2005) can detect rhythmic contrasts too, though not as well as humans. Again, here is a mechanism that appears to be a prerequisite for language, which is not specific to humans; it is a perceptual capacity that nonlinguistic beings can display.

Accordingly, when we try to determine the fundamental underlying properties of human language, we must distinguish between prerequisites that we share with other species, and those properties, if any, that are specific to language (and therefore to humans). In the words of Hauser et al. (2002: 1570), “The empirical challenge is to determine what was inherited unchanged from this common ancestor, what has been subjected to minor modifications, and what (if anything) is qualitatively new.”

There are undoubtedly important discoveries to be made regarding innate, phylogenetically ancient mechanisms that our species might be using in slightly different ways, in general or in some particular domain. Usually, the debate about specificity in language is framed as a yes/no question, whereas what I would like to stress, as Marc Hauser did in his talk (see
Chapter 5
), is that perhaps we will increasingly find that some inherited, prelinguistic mechanisms have become specialized in humans for tasks that our biological relatives do not engage in. Both in the case of categorical perception and in the case of rhythm detection, humans appear to be particularly good at these capacities and apply them to a novel function (language) in order to select categories that are not merely perceptual, such as phonemes or words. Our task is to find out how this happens, how we push these mechanisms to take a path that other creatures do not tread – the path of language, with categories further and further removed from perception.

Let us review the acquisition of phonemes in more detail as an illustration of what I mean. What crickets and chinchillas are trained to do is acoustic discrimination, but not phoneme perception. It is relatively well known that very young toddlers are capable of fine-grained phonetic discrimination, so that a child born in a Japanese-speaking community will be able to discriminate between /r/ and /l/ even though this distinction is not phonologically relevant in Japanese, and even though the adults surrounding this baby cannot perceive the distinction. Werker and Tees (1984) showed that at about 10 months of age, children “specialize” for those contrasts that are phonologically distinctive in the language they are acquiring, and become like their parents, in that they no longer discriminate contrasts that are not phonologically relevant in their language.

From what we know, this specialization process only happens in humans. Apparently, what we humans do is build a second, higher level of representation on top of a basic, common auditory capacity.
³
This higher-order category is the phoneme, a language-specific category. We take a mechanism for auditory perceptual discrimination, perhaps refine it, and build a language category using it, in ways that are still not completely understood. In this regard, the peculiar thing about human babies is that they are able very quickly to construct something new, something different, using largely an old perceptual mechanism. If we go back to
Syntactic Structures
again, one of the central claims made there was that if you want to understand human language, the first thing you must understand is that language involves different levels of representation. Phonology and syntax, for example, have their own separate primitives and rules. This is widely agreed in linguistics today, but it was not an agreed property of language in the late fifties. In this light, what human babies do is build a repertoire for a new type of category, the phoneme, apparently using the same perceptual mechanism for acoustic perception, and presumably employing other types of cues for category membership. Accordingly, we are now talking about something that might be “qualitatively new” in human language.

To continue our search for such phenomena, let us turn from phonology to syntax, a component of language further removed from perception. Again, we start by remembering one of the main arguments in
Syntactic Structures
– namely, that phrase structure, or constituency, is an essential property of human languages that models of language must capture. The combinatorial
and recursive nature of grammar that
Syntactic Structures
argued for is also common ground in linguistics today, as we can see for instance in this quote from O'Donnell et al. (2005: 285):

Fig. 20.2. Example of a finite state grammar.

There are other universals, which are so basic that they are implicit in every linguistic theory and become most obvious when we compare language with other animal communication systems. These include the fact that language is built up from a set of reusable units, that these units combine hierarchically and recursively, and that there is systematic correspondence between how units combine and what the combination means.

Let us remind ourselves of the argument in
Syntactic Structures
: language cannot be captured by a model with no phrase structure. For in stance, language cannot be captured by a finite state grammar (FSG). In an FSG you generate a piece of language by going from one point/state/word to the next along whichever path you choose among the ones available, until you reach the final state, at the end of the path (
Fig.20.2
).

This grammar does not give you any kind of constituency, an important problem if you want to understand and explain how human language is organized.
Syntactic Structures
shows that certain aspects of English cannot be accounted for by a grammar like the one in
Fig. 20.2
. The reason why this is so is that the syntactic structures of human languages can resemble matryoshkas, those Russian wooden dolls you open to find smaller but identical dolls nested inside. Consider for instance the English sentence:

(1) The girl the boy saw thinks the parrot likes cherries

Here, we find sentences nested inside sentences, and there is no grammatical limit to the number of times I can make a bigger doll, a longer sentence. Of course, this is not only a property of English, but a property of language, and the fact that all human grammars can build these matryoshka-structures tells us that this is a very essential aspect of human language. This property receives the name of recursion, and it has also been brought up in other talks and discussions
in this conference by Randy Gallistel, Rochel Gelman, and Juan Uriagereka, for instance. Here, I will focus on three recent studies that have asked whether phrase structure is qualitatively new and specific to humans and language.

Fig. 20.3. FSG versus PSG learning results in humans and tamarin monkeys (from Fitch and Hauser 2004).

For example, Fitch and Hauser (2004) have asked this very question regarding species-specificity. They taught two artificial languages to two groups of tamarin monkeys, where the difference between the two languages was precisely phrase structure. Whereas one language could be accounted for by a FSG, the other one had to be accounted for by a phrase structure grammar (PSG), so the FSG could not capture it. Fitch and Hauser found that tamarins, given time, did quite all right distinguishing grammatical versus ungrammatical sequences for the FSG, but interestingly, they could not manage to learn the PSG.

In
Fig. 20.3
, taken from Fitch and Hauser (2004), we can see that whereas the human group could discriminate grammatical vs. ungrammatical sequences for both grammars (results on the left), the monkeys (on the right) seemed to grasp this contrast for the FSG (top right) but not for the PSG (bottom right), where they failed to discriminate between grammatical vs. ungrammatical sequences. Does this mean that we have found a specific property of human cognition? Have we found a specific property of human language? In order to be able to
answer this question, we still need to know more. For instance, we need to know whether it is only we humans who can grasp constituent structure, the unbounded combination of symbols that yields recursion in human language (Chomsky 1995b). Recently, Gentner et al. (2006) reported that starlings do in fact grasp recursion. I think the jury is still out on this claim, mainly because it is not sufficiently clear whether what the starlings do is recursion or counting, but in any event, songbirds are a good species to investigate, because their songs are long, structured, and in some species acquisition and cortical representation parallels humans in intriguing respects (Bolhuis and Gahr 2006).

Another way of determining whether phrase structure is a good candidate for UG membership is to try to determine whether our own human brain processes phrase structure in a special way. Two recent neuro-imaging studies indicate that this might be so. Musso et al. (2003) and Friederici et al. (2006a) taught human subjects human-like, and non-human-like grammars (a similar idea to the previous animal study) to see how the brain reacted to each.
⁴
The aim was of course to determine whether there is a property of human language that only human language has (specificity in the strongest sense). If this were the case, we could expect to find some evidence of that in the brain.

Musso and co-workers (2003) taught native German speakers three rules/ constructions of true Italian and true Japanese, and three unnatural rules of a fake Italian-like language and a fake Japanese-like language. I say Italian-like and Japanese-like because the words employed in these unnatural languages were the same as in the corresponding natural language. For example, one such unnatural rule placed negation always after the third word of the sentence. The rule is trivial, but no human language does this, because a rule that counts words necessarily ignores phrase structure. The rules are easy and consistent, but they pay no attention whatsoever to phrase structure. What the authors found is that detection of violations of natural rules triggers an activation of Broca's area that is not found when subjects detect violations of unnatural rules.