Words and Rules by Steven Pinker, Weidenfeld & Nicolson,
拢14.99, ISBN 0297816470
ARE we born with a brain hardwired with rules for language? Or is the brain a
big sponge, capable of creating rules for language by mopping up what the
growing infant hears?
This has become one of the fiercest battles in the modern study of mind. And
it鈥檚 not hard to see why. Understanding language goes to the heart of some of
science鈥檚 biggest unanswered questions: how does the brain do what it
does鈥攁nd what is intelligence anyway?
Advertisement
The debate has been raging ever since the time of Plato and Aristotle, with
views falling into two main camps. Rationalists argue that intelligence is
rooted in the mind鈥檚 ability to manipulate symbols according to rules, while
empiricists point to our ability to generalise from associations with past
experience to new circumstances.
In Words and Rules, cognitive scientist Steven Pinker of the
Massachusetts Institute of Technology attempts to illustrate this clash of world
views by focusing on a minute part of human language: the way we form the past
tense of verbs in English. And he proposes a resolution of the conflict through
compromise. According to Pinker, the lesson from verbs is that our minds combine
the best of both worlds. The brains we are born with come with an in-built
program for manipulating regular verbs (matched from match,
say) whereas the past tenses of irregular verbs (catch to
caught) have to be individually memorised. Pinker dubs this the 鈥淲ords and
Rules theory鈥�.
Pinker鈥檚 thesis rests on a dazzling range of facts suggesting the brain
handles regular and irregular verbs quite differently. Briefly, he argues thus:
children are much more likely to make mistakes on irregular verbs than regular
ones (鈥渟ee-ed鈥� instead of 鈥渟aw鈥�); the speed of production of an irregular verb
depends on its frequency in the language, while regular verbs seem immune to
this; irregular verbs exhibit 鈥済ang鈥� effects, that is, unfamiliar words are
treated like a member of the gang, where regulars aren鈥檛; irregular verbs are
much more likely to be lost from the language than regular verbs; and people
with localised brain damage sometimes show 鈥渄ouble dissociations鈥� in their
ability to use regular and irregular verbs. Patient A may be able to use regular
inflections but not irregular inflections, while patient B exhibits the opposite
pattern鈥攕uggesting, argues Pinker, the two types of verbs are handled by
distinct brain areas.
In all these situations, Pinker shows how the behaviour of irregular verbs is
symptomatic of an associative memory at work, while rule-governed symbol
manipulation neatly captures the behaviour of the regulars. In his captivating
and elegant style, Pinker convincingly debunks exclusively rule-based accounts,
before turning on his real enemy, the exclusively empiricist school, as
currently represented by the new wave of so-called connectionist modellers.
These researchers put their faith in neural network computer simulations.
These are designed to mimic the learning behaviour of neurons, soaking up
information鈥攁nd detecting patterns in it鈥攂y adjusting the strengths
of the connections in the network. Some connectionists claim that, with a little
training, these simulated networks can create the rules for making, plurals or
past tenses, for example, more or less from scratch. And that observation, they
say, seriously undermines the case for the rules being hardwired in humans.
According to Pinker, though, this enemy has little firepower. Neural network
models lack the necessary computational power to capture the structural features
characterising human language. He dismisses 15 years鈥� of connectionist accounts
of how the past tense is formed in English as too simplistic or implausible.
Readers of Words and Rules have little choice but to recognise the
wisdom underlying Pinker鈥檚 compromise between rationalism and empiricism, or
else dismiss his whole book.
But I must come clean. As a part-time connectionist modeller, I am not
convinced. Many of Pinker鈥檚 dismissals just don鈥檛 square with the facts. Take
verbs derived from nouns. Contrary to Pinker鈥檚 claims, there are connectionist
models that will take a word, usually inflected as a noun, and regularise it in
the past tense, even if it sounds the same as an irregular verb鈥攂rake,
spit and ring, for example.
Likewise, connectionist models exist that show
patterns of mistakes when you train neural networks鈥攎uch as you do
children 鈥攁nd without making implausible assumptions about the training
regime. Even the pattern of development of noun plurals found in young children,
which Pinker assumes beyond the capacity of associationist approaches, has been
captured in a connectionist model.
What Pinker also fails to acknowledge is that neural network models of the
Arabic and German plurals (and the English past tense for that matter) do a
better job at predicting how to inflect a new word than the hybrid model he
advocates. And contrary to a Pinkerian reading of the literature, connectionist
models can mimic the double dissociations between regular and irregular types of
words seen in people with brain damage.
Of course, Pinker would object that these connectionist models offer one-off
fixes to individual problems, rather than a grand theoretical framework. This is
a valid criticism, but one which must be viewed in the context of the great
debate about the brain. First, science usually progresses in small steps,
dividing problems into smaller problems and solving them one at a time. There is
no obvious reason why a grand associationist theory might not unite its
individual successes.
Secondly, the connectionist models that Pinker criticises are all working
models that not only mimic human behaviour but also produce testable behavioural
predictions about, for example, word formation. In contrast, the words and rules
theory has not been implemented in any working model. As any cognitive scientist
will tell you, there is a wide gap between having a theory and making its parts
work together coherently in a computer model.
If you care about language and the mental processes underlying the way we use
it, then Words and Rules is definitely worth reading. It is the most
accessible exposition of the rationalist approach to verb morphology I have
seen. It is not, however, a balanced presentation of the age old controversy
between rationalism and empiricism. For that, readers will have to look
elsewhere or delve into the journals.
You won鈥檛 have to hurry, though. My guess is that the debate still has a
couple of centuries to run.