HUMANS are quintessentially verbal creatures. Ours is a world of
words—so much so that it is impossible even to imagine life without
language. No wonder, then, that anthropologists are obsessed with finding out
why and how language evolved, and, equally, linguists are obsessed with trying
to understand how each of us comes to acquire this power.
Throughout its history, the study of language acquisition has been dominated
by two opposing schools. One claims that language is simply a social phenomenon,
like culture, and therefore the result of learning, while the other holds that
the grammatical rules on which we build a language are something we are born
with.
This year, both groups have published new findings, fuelling renewed interest
in the field and a brisk trade by all concerned in promoting their own
interpretations. In doing so, Noam Chomsky, for one, has been tidying up his own
theories— theories that founded the innateness movement more than 40 years
ago. This process is seen by some as indicating their weakness, but to others
it’s a sign of strength. The confusion arises partly because the term “innate”
is slightly misleading. We obviously all have something innate in our brain that
permits language acquisition—as Chomsky himself asserts, “everyone has an
innateness hypothesis”. But the question is, is it an innate machinery for
learning language or an innate grammar?
Advertisement
In the 1950s, the learning school dominated linguistics. Chomsky, then a
young researcher at the University of Pennsylvania, was dissatisfied with this
explanation. How is it, he mused, that a child can master the complexity and
abstractness of language at such a precocious age, in such a short time?
Children are not taught language in a systematic manner, but are merely exposed
to a stream of talk that may be grammatical or ungrammatical, with neither
labelled as such. Moreover, it is often incomplete. And yet children converge on
the same grammar.
Given what he called the “poverty of the stimulus”, Chomsky argued that
children must come into the world with the fundamental structure of grammar
hard-wired in their brains. What a child learns is the specific language to
which he or she is exposed, like flesh draped over a skeleton. In 1957, Chomsky
published a book entitled Syntactic Structures, in which he set out his
theory of generative grammar. There is, he argued, a universal grammar that is
just as much a biological endowment of Homo sapiens as the organs of
the body. Thus he established the “innateness school”, at the Massachusetts
Institute of Technology, which dominated linguistic explanations of language
acquisition for three decades, busying itself with trying to find an invariant
core of grammatical rules. This, says David Pesetsky, another key player at MIT,
is the goal of goals.
Meanwhile, the learning school did not disappear, but it was definitely a
minority view. What this school needed was evidence that children are capable of
learning all the complexities of language from the linguistic mess they hear
around them. “Many people assumed that babies are poor learners, and have
insufficient memory and attention span to learn all that there is to learn about
language, given how complex it is,” says Elizabeth Bates, a psychologist at the
University of California, San Diego. “I don’t think that’s true, and there is a
growing body of evidence in support of that view.”
For instance, modern computing power and techniques have recently had an
impact on the field, with neural networks and statistical methods being applied
to the study of language acquisition. Neural networks are computer algorithms
that mimic the way collections of neuron-like processing units learn simple
tasks or identify patterns. “An important property of these algorithms is their
ability to derive structural regularities from relatively noisy input data,”
explains Mark Seidenberg, a neuroscientist at the University of Southern
California in Los Angeles.
No neural network has actually learnt language in a human-like manner, but
they have demonstrated a knack for recognising certain structures in language.
This has encouraged Seidenberg and others to entertain the possibility that the
human brain, with its incredibly extensive neural networks, is capable of
learning language from the kind of noisy input to which children are
exposed.
Elissa Newport, Richard Aslin and Jennifer Saffran, from the University of
Rochester in New York, produced a simple but, to many observers, impressive
study which they argued made this type of language learning not merely possible
but probable (Science, vol 274, p 1926). They showed that infants
exposed to a stream of computer-generated sentences could learn to identify the
boundaries between individual words. Unlike the words on this page, spoken
sentences are a constant flow of sounds. There are gaps in the stream, but these
are as likely to occur in the middle of words as between them. The ability to
distinguish between words and not-words is a crucial skill in language learning
and use—a skill the researchers think can be quantified. The recurring
sound sequences that make up words, they say, crop up much more frequently than
the accidental sound sequences that occur between one word and the next.
Newport and her colleagues tested the idea that by exposing eight-month-old
babies to a two-minute recording of four nonsense words each of three syllables,
repeated in random order and spoken in a monotonic, synthesised voice. Later,
the babies heard a new recording of three-syllable words, but this time with
new syllables or the original ones in a different order. The researchers found
that the babies tended to listen more attentively to the new sequences of
syllables—in other words, they could tell the difference.
So, concluded the researchers, babies are whizz kid statisticians. In just
two minutes, they could recognise which sound sequences were novel enough to be
the boundaries between words. But what does it all mean? Bates has no doubt.
“This work is compelling evidence that children are capable of extremely
powerful and rapid statistical induction, and there is every reason to think
that the same kind of abilities might extend to other aspects of language
acquisition, such as grammar,” she says.
But when she teamed up with her colleague Jeffrey Elman to voice the same
speculation in an article accompanying the paper in Science, there were
howls of protest about how far the finding could be extended. “Saffran, Bates
and Elman suggest that if children can learn words by recording frequent sound
sequences, they might learn grammar the same way,” wrote Steven Pinker, a
cognitive scientist at MIT and author of The Language Instinct.
“Learning words and learning grammar are…different computational
problems.” Pinker believes that you simply have to memorise words, you cannot
learn rules for putting syllables together to form particular words. Sentences,
by contrast, have grammatical structure for which there are rules.
Mark Hauser, a cognitive scientist at Harvard University, sees the Rochester
results as interesting but he, too, finds them limited in their implications.
“The statistical processing involved wouldn’t help a child very much in a real
language acquisition context because it is extremely simple. What has been shown
so far has nothing to do with language.” He does not deny, however, that
statistical learning may be important in grammar. “An interesting thing about
syntax is that the statistics are long range. You have one word in a certain
position, and then another down the line that has to be there because of the
structure of the sentence.” Does such a long-range statistical learning
mechanism exist in humans? “There’s no answer to that yet,” says Hauser.
But in January this year, researchers at New York University and Amherst
College, Massachusetts, led by Gary Marcus, claimed to have moved closer to
finding that answer (Science, vol 283, p 77). Using experiments similar
to the Rochester group, Marcus and his colleagues showed that seven-month-old
babies could recognise patterns of syllables in artificial words. For example,
if they heard a string of nonsense words such as “gatiti” and “linana”, all of
the general form “A-B-B”, they would recognise that the word “wofefe” would fit
the pattern but not the word “wofewo”.
Because this shows an ability to recognise a pattern in new sounds, the
researchers argue this is a more complex form of learning than that demonstrated
by Newport and her colleagues. “If our position is correct,” they say, “then
infants possess at least two distinct tools for learning language—one
device that tracks statistical relationships and another that manipulates
variables, allowing children to learn rules.”
Rules are the core of grammar, of course. But opinions differ widely about
whether this experiment gives a glimpse of the type of rule learning that might
apply in language acquisition. Bates, for instance, disagrees with Marcus: she
believes that the mechanism for this pattern recognition isn’t so very different
from statistical learning, and that the experiment strengthens the learning
model. But Hauser thinks differently. “It’s a rule nonhuman animals can learn,
too, just as they can do statistical learning. It has nothing to do with
language.” Likewise Pesetsky. “The Marcus work doesn’t look like anything I
would recognise in language acquisition,” he says. Marcus and his colleagues are
somewhat in between. Referring to statistical learning and rule learning, they
wrote in Science: “Even taken together, these tools are unlikely to be
sufficient for language acquisition, but both may be necessary
±č°ů±đ°ů±đ±çłÜľ±˛őľ±łŮ±đ˛ő.”
Just to complicate the debate, Marcus’s paper came out just three weeks
before a symposium organised by Pesetsky where he and others presented new
findings supporting the innate model—findings Pesetsky describes as
“unassailable”. To Chomsky’s problem of how babies learn language despite the
poverty of the stimulus, Pesetsky adds a further question that he believes
points to substantial innateness of language structure. “Why is it that
languages that were not in historical contact nevertheless have the same
structures and properties?” he asks. “There must be a universal, invariant core
of language knowledge that we are born with and [that] is easily triggerable by
environmental stimuli.”
In one of the more Herculean ventures in linguistics, Guglielmo Cinque of the
University of Venice has shown that, among the 500 languages he studied, the
ordering of adverbs in sentences, when sorted out by meaning, is invariant. The
same is true, he found, for other aspects of speech, expressing tense, mood,
aspect, modality and voice. “These results suggest that the core structure of
human sentences is not merely `similar’ from language to language, but is
actually identical in all languages,” said Cinque at Pesetsky’s symposium. “And
it is so for no other obvious reason but the fact that it is inscribed in our
innate predisposition for language.”
Cinque can see no reason why the structure is as it is. It is just a frozen
accident of evolutionary history, he says, “much as our genetic endowment
imposes on all humans one pair of eyes instead of two, five fingers instead of
seven.” Asked why it has taken so long for what he calls “the secret skeleton of
language” to come out of the closet, he says: “As the differences among
languages are so much more obvious than the similarities, it is no wonder that
it took many years of careful work on a wide variety of languages to discover
that they all share quite a substantial core of common properties.”
Mark Baker of Rutgers University, a participant in Pesetsky’s symposium,
agrees. “The more you look at languages, the more similar they seem,” he says.
At the symposium, he described his work comparing Mohawk, a language spoken by
Native American people from the northeast, with English. “Superficially, Mohawk
looks about as different from English as you can imagine,” he said. “However, in
a list of six parameters that govern how sentences are assembled in the two
languages, they differ in only one.” Because languages are so complex, with
infinite combinations, he says, any small differences become magnified, so that
they look more different than they actually are.
Pesetsky described his own study of the structure of questions. There are
some nuances that differ, he said, but the strongest signal is that all
languages follow the same rule: that the word that asks the question is
associated with the periphery of the sentence. “This work, and the work of my
colleagues here at MIT, and Baker’s and Cinque’s work…all represents a
redeeming of a promissory note of the Sixties.” In other words, the innately
programmed universal grammar that Chomsky posited is beginning to be
identified.
Meanwhile, Chomsky has been revising his theory of universal grammar with
what he terms the Minimalist Program. This is an attempt to resolve the tension
between innate structure being sufficiently detailed to shape any construction
in any human language, and simple enough to reflect the small set of innate
mechanisms that allow humans to acquire and use that language. In other words,
there has to be a balance between the huge task of acquiring the many
complexities of languages and the limited set of mechanisms that are present to
do it. The bottom line is that despite Chomsky’s admitting that “many of the
standard assumptions (including my own) were just wrong,” he still believes that
we are all born with some form of language acquisition device into which is
wired elements of grammatical structure.
Radical departure
Although some linguists view the programme as a radical departure in
Chomsky’s thinking, Pesetsky says it is more like cleaning up the theory of
universal grammar. “As Chomsky modified his theory over the years, new pieces
were added that made others redundant,” he says. “The redundancies built up,
making the theory untidy. So he did some house cleaning. The theory isn’t weaker
than he proposed previously. If anything, it’s stronger.”
Not surprisingly, proponents of the learning model see things differently.
Elman suggests that what the innate school calls universals are in fact the
result of constraints in the systems that generate language. “Think of a hundred
different neural networks, with different architectures, and think of these as
English and Chinese and Italian, and so on. Now have them solve the problem of
assembling sentences. What would happen is that they would all converge on the
same solution, the only good solution. The similarities you see across languages
are the result of the cognitive processes that produce them, not something
that’s built in from the start.” Bates takes a similar view, but she also
suggests that when you look deeply at what is being claimed to be universal, it
turns out not to be that universal.
Pinker places himself in between, suggesting that there may be an innate
machinery for learning rules. “If there is anything that is innate in language,
it has to be the machinery that does the learning,” he says. “The machinery
looks for word boundaries and categories, and tries to formulate rules for
recombination, maps sounds onto meaning, and so on. If this machinery is
optimised for learning language, as opposed to being a generic mechanism for
everything we learn about the world, then it isn’t easy to distinguish that from
innate knowledge. You could say that the knowledge is implicitly embodied in
these special-purpose learning mechanisms.”
If the difference between innate knowledge and innate mechanisms will be hard
to spot, will the puzzle of how language is acquired ever really be solved to
either school’s satisfaction? As if to keep the two groups talking past each
other, they turn the debate to the subject of whether, if there are any innate
mechanisms, these mechanisms are unique to language—and so to our own
species. “I believe there are unique mechanisms in the human brain, but they are
not specific to language,” says Elman. “Instead, they have been
opportunistically used for language, and this results in the fact that all human
languages look alike.”
Similarly, Seidenberg argues that “brain organization…constrains how
language is learned, but the principles that govern the acquisition,
representation, and use of language are not specific to this type of knowledge”
(Science, vol 275, p 1599). Pesetsky’s counter is that “it’s just a
statement of belief”.
As if to echo Chomsky’s problem, it seems that linguists themselves are
attempting to solve the highly complex problem of how we acquire language on the
basis of scanty, conflicting and incomplete information. A wealth of opinions
against a paucity of evidence. As Pesetsky comments on what remains to be
achieved: “Yes, linguistics is a science, but it is very much like a
17th-century science. We still have a long way to go on these questions.”