>However, it is not clear to me how the competence/performance
>distinction has "actually disappeared" in the framework you describe.
>Crucial to your approach is the "large corpus of analyzed utterances,"
>and implicit in that phrase is some set of rules -- discrete,
>nonprobabilistic rules -- for what constitutes an allowable analysis.
>Admittedly those rules may be incredibly simple, for example, you can
>allow all possible binary trees covering a sentence as the set of
>possible analyses for that sentence....
>
>If your theory is intended to be a cognitive model, then you have no
>choice but to interpret those underlying principles as something that
>people know about the possible structures of their language, as
>distinguished from the actual structures that get observed or
>produced. And that's competence.
Philip,
Sure you're right with respect to the technical sense of the word
competence. But a competence model which allows you to generate the set
of all labeled trees with words as terminals can hardly be seen as
something that people *know* about their *language*. Moreover, in my
knowledge of the word "competence model" it is also required that a
person's competence be described by a compact, preferably non-redundant
set of rules that summarize linguistic regularities. And this is what the
framework I have described actually skips.
Still, also Steve Finch is right in saying that there is a huge
bootstrapping problem here. In fact, I can only *effectively* put forward
my framework by saying: take you're favorite linguistic theory (yes, a
competence model..) and annotate a large language corpus with it; let the
result be the basis of your performance model, by means of which new
utterances can be parsed and produced. Of course, in order to do so, one
still needs to define (1) the corpus-fragments that may be used as units
in constructing new utterances, (2) a definition of the operations to
combine fragments, and (3) the way probabilities are computed on the
basis of occurrence-frequencies of the fragments in the corpus. And many
of you will certainly claim that at least (2) and maybe (1) belong to the
realm of a competence model.
To go short: I still hold that once you have a large language corpus in
which each utterance is provided with a syntactic/semantic/pragmatic etc.
analysis that was appropriate for understanding that utterance, you can
(probabilistically) parse and produce new utterances without explicitly
or implicitly invoking the underlying principles that summarize the
regularities of your language.
--Rens