>On the other hand, if one views a person's knowledge of a language as an
>extensive record of previous language experiences, represented as a large
>corpus of analyzed utterances, by means of which new analyses (and their
>probabilities) can be constructed out of parts of analyses that occur in
>the corpus, one obtains a "probabilistic grammar" where the
>competence/performance distinction has actually disappeared.
How so? I'm sure many traditional linguists would agree that a "large
corpus of analysed utterances" (e.g. the ones found in Radford and a few
others), encapsulating an "extensive record of previous language
experiences" would be very useful in setting parameters within a
performance model which still leaves plenty of room for a competence
model. The point is that an algorithm such as you describe can still be
thought of as implementing a competence theory; indeed the Penn Tree Bank
can be thought of as a set of examples analysed according to a particular
competence theory, and as such in some sense partially encapsulating that
theory. The parsing model you propose can be thought of as exploiting a
partial encapsulation of a competence model.
To put it another way, while such an algorithm might constitute a
performance model, it does not by itself do away with a competence model.
There is a strong bootstrapping problem here; one needs to be able to
analyse utterances before one has a large corpus of them which, by
hypothesis, one needs before new sentenses can be analysed. I believe
this bootstrapping problem can be overcome without heavy innate knowledge
since the statistical regularities in language (evidenced in a large
corpus of trivially analysed utterances) are so strong, and if it can be
overcome I agree with the conclusion that a competence theory becomes
unnecessary (but possibly still useful), but there is a long long way to
go to show that it is possible.
Steve.