Publishing Partner: Cambridge University Press CUP Extra Publisher Login

Discussion Details

Title: Reply to disc. posting 16.1156
Submitter: Peter Hallman
Description: In issue 16.1156, Richard Sproat and Shalom Lappin challenge the
Minimalist community to ''to produce, by May of 2008, a working P&P
[Principals and Parameters Framework] parser that can be trained in a
supervised fashion on a standard treebank, such as the Penn
Treebank, and perform in a range comparable to state-of-the-art
statistical parsers,'' between 80% and 90% accuracy under certain

The goals of the P&P approach to language acquisition are
dramatically different from those of statistical approaches, which
makes a comparison in terms of accuracy alone uninformative. The
P&P framework seeks to connect typological universals to the
mechanism of language learning, in effect explaining those universals
as properties of the initial state of the trainable parser. A statistical
parser can, within physical limitations, recognize and learn any
statistically significant pattern, not merely those patterns that occur in
human languages. The P&P approach finds this disadvantageous,
because the P&P framework seeks to answer the question ''What is a
possible human language (type)?'' The P&P parser that Sproat and
Lappin envision would answer this question; comparable statistical
parsers do not.

A successful P&P parser would not only acquire the target language
accurately, it would behave like a language learner in its acquisition
timeline and would fail to acquire languages that violate language
universals. It would have to display these properties in order to
successfully learn the target language, because these properties
ought to be inherent in the parameters underlying the system.

So the P&P parser that Sproat and Lappin envision would accomplish
much more than comparable statistical parsers, which makes the
proposed accuracy metric a poor yardstick for comparison, and
furthermore, I suspect, it makes the three-year timeline unrealistic,
especially since there is no reason to believe that the discovery of
parameters and implicational relations among them is finished at the
present time and ready to form the basis of a trainable parser.

Nonetheless, I hope someone takes up the challenge (it's not my
field), since the attempt can only benefit the P&P framework. Perhaps
there should be a prize.

Peter Hallman
Department of Linguistics
McGill University
Montreal, Quebec, Canada
Date Posted: 20-Apr-2005
Linguistic Field(s): Computational Linguistics
Discipline of Linguistics
LL Issue: 16.1251
Posted: 20-Apr-2005

Search Again

Back to Discussions Index