...S='B-NP'¹

Section 3 explains this `subcategorization' of NP into B-NP and I-NP.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... rule²

We call mother rule of a rule r, the rule to which r is an exception.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... features³

This would imply modifications of the system we have not done yet for lack of time.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... nucleus⁴

The partial structures (structures without nucleus) represent 2% of all the structures in the training corpus, and then introduce a little noise.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... often⁵

At least, in the training corpus.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... concern/VBP⁶

Tagging error.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... cover⁷

In the specific application described in [Mooney and Califf(1995)], learning past-tense of English verbs, this minimum threshold is validated, but for the chunking task, exceptions occurring only once in the data degrade results.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.