types of learning, these results are only possible given a system which operates on labels
and data at the same time. With an engineered system which we’ve built ourselves it’s no
big deal to add ‘symbolic’ inputs in the form of labels to the inputs already in place for
the ‘raw’ data, and adjust the network architecture so that these two streams interact
optimally. But with us, with people that is, and some non-humans, there’ s a crucial
developmental question: “How do we get to be able to make use of ‘symbols’ in the first
place?” Much of the present paper is concerned with this question, which we call the
‘How question’.
This means that for the purposes of what follows, we will for the most part leave Clark’s
account of the advantages of language once you ’ve ‘got’ it, in place. Another way of
saying what interests us, though is as follows: Clark’s account of language, in common
with much linguistic theorising, emphasises the ‘abstraction amenable’ aspects of
language. That is to say that he focuses on labels, signs, symbols and constructions of
such elements. But if he is broadly correct about the advantages, then an answer to the
question as to how any cognizer can get to count something as a ‘symbol’ at all is
needed, and we maintain that part of the answer to that question is to be found by paying
closer attention to how talk works between people, which is to say drawing on the sorts
of ways Clark looks at robots.
The Poverty of the Stimulus
A fact indubitably in need of some explanation is that human children typically acquire
facility with language within a few years and with little evidence of effort. Debates over
the correct explanation are partly organised around a fault line between empiricists
defending some version of the view that general learning can account for language
acquisition, and nativists insisting that some language-specific innate capacities are
essential. Perhaps the most powerful weapon available to the nativists is the poverty of
the stimulus argument, which can be glossed as follows:
It is clearly the case that a wide range of sets of organising principles are
consistent with the ‘stimulus’ or primary data available to human children, and
further that the sub-set of ‘correct’ principles are not preferable by the standards
of generic criteria for theory choice, such as simplicity. It consequently seems
extraordinarily unlikely that any human child would ever come to behave in
ways counted as grammatical for their mother tongue (or tongues) in the event
that human children were broadly empiricist learners. Since children do come to
be regarded as behaving grammatically with such striking reliability, we can
conclude that they are not empiricist learners, but rather that they have language
specific innate cognitive endowments.3
Debates between empiricists and nativists about language acquisition are not, of course, a
series of confrontations between radical ‘tabula rasa’ empiricists and comprehensive
nativists who see no role for experience or learning at all. Rather, disagreement concerns,
inter alia, questions about the real nature of the ‘stimulus’, what mixture of innate and
learned capacities are required to explain the phenomena, when particular types of
learning start, the extent to which humans and particular non-human animals are
cognitively alike, and the strengths and limitations of different types of learning.
Although the present paper is not directly concerned with grammar, we may as well
stress that we are not Chomskian nativists. That said, with respect to our ontogenetic
concerns we are persuaded that a wide range of innate mechanisms and biases are
required to explain the available data. Our wariness of Chomsky’ s brand of nativism is
fuelled by two major considerations.
4 / 23