Answer three of the following. Try not to spend more than two
hours on the exam altogether.
Submit your answers to me,
preferably
by email, by 11:59 pm, Tuesday,
Dec. 16. Please include "645" somewhere in the subject of your
message.
If you have questions about the exam, send me a message, again
including "645" in the subject.
For each of the following, say how statistical approaches would handle or fail to handle
it:
subject-verb agreement
the similarity in meaning between the words election
and party
the tendency for the object of the verb meet to be
human
Using what we have learned about computational speech act theory,
show how a speaker would plan the sentence is it raining?.
You will need the operator Informif and a version of the operator Request for yes-no
questions.
You have the task of developing a system which can assign
one of a set of 10,000 classifications (e.g.,
medieval European history,
cell biology, knowledge representation) to the journal articles
received by a library.
The system should work without the benefit of a grammar or
a dictionary but simply make use of the statistical
properties of the
texts.
Describe the system you design for this task and how you would
train it.
What connectionist
options are there for representing
(or learning to represent) the structure
of phrases such as the following?
a green square
a green square to the right of a circle
a green square to the right of a circle containing two
blue triangles
a green square to the right of a circle containing two
blue triangles on either side of a green square
Why are symbolic approaches still preferred for practical systems that
have to deal with such phrases?
Assume that learning a language is
learning an HPSG-style
grammar. Speculate on how learning could be facilitated by
having certain aspects of the grammar innate
and on how HPSG-style parsing and learning would interact.