Table 3 shows mean output activations for the four cases.
There is a strong interaction (
):
output activations are higher for
words in the meta-category corresponding to the linguistic context
than for words in the other meta-category.
In other words, even though the network cannot have generalized
about what constitutes an adjective and what constitutes a noun ---
there is no generalization to be made, after all --- it has made a
distinction between the two meta-categories.
The associations between linguistic inputs (the two linguistic context
units) and linguistic outputs (the 36 word units) are sufficient
to create two classes of words.
We do not believe that the picture is this simple for word learning in
children because there are semantic generalizations to be
made concerning part-of-speech categories.
In a more realistic setting, the straightforward learning demonstrated
in this experiment might serve to bootstrap the learning of
the relatively abstract semantic differences between the
meta-categories.
At any rate, the implication is that the patterns
of errors made by children that implicate distinct noun and
adjective categories could arise only from form-to-form
associations.
Table 3: Experiment 6: Noun and Adjective Response to Noun and
Adjective Linguistic Contexts.
Figures show the
mean activation of noun and adjective output units in response
to 18 object input patterns which belong to neither meta-category and
which are presented together with either noun or adjective linguistic
contexts.