Why is language unique to humans?
Jacques Mehler*†
1
, Marina Nespor‡, Mohinish Shukla* and Marcela Peña*
* International School for Advanced Studies, Trieste, Italy, † Ecole des Hautes Etudes en Sciences Sociales,
Paris, France and ‡ Universita di Ferrara, Ferrara, Italy
Abstract. Cognitive neuroscience has focused on language acquisition as one of the main
domains to test the respective roles of statistical vs. rule-like computation. Recent studies
have uncovered that the brain of human neonates displays a typical signature in response
to speech sounds even a few hours after birth. This suggests that neuroscience and lin-
guistics converge on the view that, to a large extent, language acquisition arises due to our
genetic endowment. Our research has also shown how statistical dependencies and the
ability to draw structural generalizations are basic processes that interact intimately. First,
we explore how the rhythmic properties of language bias word segmentation. Second, we
demonstrate that natural speech categories play specific roles during language acquisition:
some categories are optimally suited to compute statistical dependencies while other
categories are optimally suited for the extraction of structural generalizations.
2005 Percept, decision, action: bridging the gaps. Wiley, Chichester (Novartis Foundation Symposium
270) p 251–284
Linguists and psychologists have studied language acquisition; the former have
elaborated the most sophisticated formal theories to account for how this unique
competence arises specifically in humans. For instance, Chomsky (1980) formulated
the Principles and Parameters theory (hereafter, P&P) to account for the acquisi-
tion of language given the poverty of the linguistic data the learner receives. In fact,
infants acquire the grammatical properties of their language of exposure on the
basis of partial and unreliable information. Babies, like adults, are confronted with
incomplete or erroneous sentences.
P&P assumes that infants are born with ‘knowledge’ of Universal Grammar. This
endowment includes genetically specified universal principles, that is, the properties
shared by all natural languages. Moreover, the endowment specifies a number of
binary parameters that capture those grammatical properties that vary systemati-
cally between groups of natural languages. For instance, there are groups of lan-
251
1
This paper was presented at the Symposium by Jacques Mehler, to whom correspondence should
be addressed.
guages that put Heads to the left of Complements while other languages put Com-
plements to the left of Heads. The P&P theory attempts to identify such structural
properties that are basic to natural language distinctions. Parameters can be thought
of as switches that must be set to one of two possible positions to specify the prop-
erties of the language being learned. The linguistic input determines the particular
value of a parameter.
2
P&P has many virtues. First, by exploring the way in which natural languages are
sorted into groups that share coherence for syntactic properties, P&P is one of the
most productive theories ever developed within the linguistic domain, (see Baker
[2001] for an accessible and fascinating account of the P&P proposal). Next, P&P
also addresses the problem of language acquisition without making the simplifica-
tions common to alternative theories. For example, optimists claim that imitation
is the privileged mechanism responsible for the emergence of grammatical com-
petence. The P&P perspective is appealing because it is biologically realistic assum-
ing that infants are equipped with a species-specific mechanism to acquire natural
language that can be explored with the tools available to formal linguistics and to
the explorations of cognitive neuroscience.
While P&P is certainly playing an important role in the domain of language
acquisition, there is a second influential position that asserts that the surface prop-
erties of stimuli can bias the learner towards postulating syntactic properties for the
incoming utterances. While the P&P theory was formulated with the precision nec-
essary to allow us to evaluate it, the general learning device proposal appears to be
somewhat less precise. Criticisms of proposals according to which general learning
mechanisms are sufficient to explain language acquisition have been given by many
theoreticians, see Chomsky (1959), Fodor (1975), Lenneberg (1967) and Pinker
(1994) among many others. We will come back to this point below.
Recently, some attempts were made to show that speech signals contain hitherto
ignored information to allow general learning accounts to explain how language is
acquired. Mostly, these attempts are minor modifications of association, a mecha-
nism that humans and animals share. Within this stream of research, the brain is
252 MEHLER ET AL
2
To illustrate this, consider a child who hears mostly sentences with a Verb–Object order. The
child, supposedly, obtains information automatically from the linguistic input to set the relevant
word order parameter. If this were so, it would constitute a great asset, since fixing this word
order parameter may facilitate the acquisition of grammar and also the acquisition of the lexicon.
Likewise, the child exposed to a language that can have sentences without an overt subject e.g.
Italian (‘piove’, ‘mangiano arance’ etc.), or to a language whose sentences require overt mention
of subjects e.g. English (‘it is raining’, ‘they eat oranges’), supposedly gets information from the
linguistic input to set the relevant pro-drop parameter.
regarded as a huge network that works in a Hebbian fashion (see Hebb 1949).
3
This
may explain why many psychologists and neuroscientists have adopted a viewpoint
that ignores the complexity of syntax and assumes that by focusing exclusively on
speech perception and production, a functional account of how language is
processed will follow. Undeniably, behavioural scientists have made great strides in
the study of perception and production. Some of them believe that it is sufficient to
study how language production and perception unfold during development to
understand how syntax (or semantics) arises in the infants’ mind. Of course it is
easier to study speech perception in babies or animals than trying to figure out how
the human brain computes syntax, semantics and pragmatics of utterances, some-
thing that animal brains cannot do. Psychologists who adopt a general learning
device framework often assume that the mystery of syntax acquisition will disap-
pear once we understand how infants just learn to extract the distributional prop-
erties of language (see Seidenberg & MacDonald 1999 among many others).
At this point we would like to point out that although there is a huge contrast
between the two stances presented above there are many points of agreement as
well. Investigators working in both the P&P tradition and in the general learning
framework agree that some parts of grammar must be learned. Indeed, no one is
born knowing Chinese, Malay or any other natural language. Each learner has to
acquire the language spoken in his or her surrounds. What distinguishes the differ-
ent positions is the scope and nature of learning they are willing to posit. P&P
assumes an initial state characterized by ‘knowledge’ specific to a putative language
module, i.e. Universal Grammar (UG). In contrast, general learning theoreticians
assume that the initial state is characterized by learning principles that apply across
the different domains in which organisms acquire knowledge. The general learning
device is undeniably a powerful account to explain how organisms will use the
surrounds to acquire behaviours that satisfy the organism’s needs. Thus, it is an em-
pirical issue whether just one of these theories is sufficient to explain how the
human mind acquires its capacities including natural language.
Do we have accounts of how syntax can be acquired after the child has learned
the lexicon of the language of exposure applying general learning mechanisms? Is
there any evidence that the acquisition of syntax does not start until at least part of
the lexicon is acquired? Depending on the answers we give to these questions, either
the P&P model or the general learning model should be abandoned for syntax
acquisition.
HUMAN LANGUAGE 253
3
Hebbian networks are a set of formal neurons synaptically connected but with connectivity
values that change with functioning. If two neurons are active at the same time, the value of their
connection increases. Otherwise, the value of the connection stays identical to what it was or
decays.
So far, we have tried to highlight the positive aspects of both P&P and general
learning mechanisms. However, problems arise with both frameworks. While the
first tries to cope with language acquisition in a realistic sense the second focuses
on the acquisition of, at best, toy-languages. P&P is problematic because of the
many implicit assumptions that investigators make when trying to explain the acqui-
sition of grammar. P&P was formulated with syntax acquisition in mind and
researchers generally take for granted that infants, in one way or another, have
already acquired the lexicon, before setting syntactic parameters. Presupposing that
infants begin processing speech signals only when they start learning the lexicon
justifies neglecting the study of language acquisition during the first year of life and
explains why P&P investigators have mostly reported data from language produc-
tion studies.
Data from animal experiments suggests that the vertebrate auditory system is
optimally suited to process some of the linguistically relevant cues that speech
affords. Thus, at least some properties of language could be acquired precociously
from speech signals. Indeed, animals with auditory systems similar to our own tend
to respond to speech patterns much like infants younger than eight months (see
Kuhl 1987 and Ramus et al 2000 among many others). Apes, but also dogs, have
‘lexicons’ that can attain a few dozen words (Premack 1971, 1986). However, such
abilities are insufficient to enable non-human animals to construct a grammar com-
parable to that of humans. Nonetheless, together with other pieces of evidence that
we lay out below, we assume that the sensory capacity of many vertebrates licenses
the processing of speech from the first year of life and, consequently, we should
not neglect the aquisition that humans make during their first year. We show below
that language acquisition begins with the onset of life. Indeed, several investigators,
regardless of the position they defend, have found empirical evidence suggesting
that the sound pattern of language are identified by very young infants and that
some properties can be attested even in neonates. The sound pattern of speech
contains cues that might bias language acquisition at different stages. As is becom-
ing obvious, the viewpoints we presented above are complementary. Indeed, while
rationalists and empiricists acknowledge the role of learning in language
acquisition, the nature of learning conceived by each of the viewpoints is
radically different. In the pages below we will try to show that it is desirable to keep
in mind that only human infants use the acoustic properties of speech to acquire
grammar. In order to explain how such uniqueness comes about, the theory
that will eventually be preferred will be the one that fits best with biological
processes.
We know that the uniqueness of syntax must be explored formally and explained
with models that are biologically realistic. Indeed, we are confronting a human apti-
tude that will bloom under several types of impoverished learning environments.
The linguistic input comes usually in the form of speech signals or, less often, in
254 MEHLER ET AL
the form of hand gestures as produced by deaf humans. Whether the learner is
hearing and seeing, deaf or even blind, s/he will attain a grammar that is as rich
and complex as we expect it in humans without sensory filters, see Klima & Bellugi
(1979) and Landau & Gleitman (1985) amongst others.
Thus, not only do we have to account for the uniqueness of the human language
ability but we also have to account for how language arises despite all the described
impoverished conditions. The best way to attain such an aim is to use the specifi-
cations given in P&P to explain what needs to be learned and what may be mas-
tered throught general learning procedures.
Chomsky (1980, 1986) and others have argued that conceiving acquisition of lan-
guage from a P&P perspective will bring clarity to the field. However, the mecha-
nisms for the setting of parameters in the P&P theory were seriously underspecified
so as to make it hard to judge. In fact, Mazuka (1996) argues that, in its usual for-
mulation, P&P contains a paradox (see below). Morgan et al (1987), Cutler (1994)
and Nespor et al (1996) among others, have proposed some putative solutions to
some of the problems arising within the P&P proposal. However, few proposals
have explored how the infant evaluates and computes the triggering signals. Some
recent results suggest that two-month-olds are sensitive to the prosodic correlates
of the different values of the head-complement parameter (Christophe et al 1997,
2003).
In the early 1980s, some scholars like Wanner & Gleitman (1982) already foresaw
some of the difficulties in the existing theories of grammar acquisition and pro-
posed that phonological bootstrapping might help the infant out of this quandary. They
held that some properties of the phonological system learnt by the child may help
him/her to uncover lexical and syntactic properties. Some years later, Morgan &
Demuth (1996) specifically added that prosody contains signals that can act as trig-
gers and thus help the child learn syntax. Indeed, these authors conclude, as we do
above, that the study of speech signals that can act as triggers is essential if we are
to understand the first steps into language.
To overcome the poverty of the stimulus argument, innate dispositions were pos-
tulated. However, as pointed out above, the proposal for language acquisition is not
sufficiently specific. Indeed, if an important part of the infant’s endowment comes
as binary parameters, we still need to understand how these are set to values that
are adequate to the surrounding language. The general assumption was that by
understanding a few words or simple sentences like ‘drink the juice’or‘eat the soup’
the child would generalize that in her/his language, objects follow verbs. As Mazuka
(1996) pointed out, this assumption is unwarranted. Indeed, how does the child
know that soup means soup (Noun) rather than eat (Verb)? Even if the mother always
says eat in front of different foods, the child may understand that what she means
is simply food! If the signals were to inform the child about lexical categories or
word order, one could find a way out of this paradox. Before we know if this is a
HUMAN LANGUAGE 255
