G22.2591 - AdvancedTopics in Natural Language Processing
Prof. Grishman
Spring 2009
Thursday 5-7 PM
Understanding natural language requires a lot of prior knowledge --
knowledge of language, and knowledge of the world as reflected in the
language. The challenge for developers of natural language
processing systems is how to capture this knowledge.
Over the past 15 years, the paradigm first shifted from systems based
on hand-coded rules to systems which are trained from
text corpora -- in most cases, from corpora that have been
hand-annotated
with specific linguistic information. In many cases, the result
has
been systems which significantly outperform the earlier systems with
hand-coded
rules.
More recently, with the availability of almost unlimited text on the
Web, the focus has shifted again from supervised methods (which require
annotated corpora) to semi-supervised and unsupervised methods (which
operate on 'raw' text). In effect, we learn about text from text. We will
consider how to create systems that can operate on Web-scale corpora
and offer
the potential of more powerful Web search -- the ability to search for
facts rather than keywords.
In many cases, relatively simple models and learning methods will do
quite
well. For better system performance, however, it is necessary to
understand
the limitations of these models and the linguistic features which can
lead
to better performance. This course will look at several natural
language
processing tasks from this point of view, examining the linguistic
characteristics that support the creation of effective models, and the
learning methods
required to train these models. Among the tasks which may be
considered
are:
- name tagging
- part-of-speech tagging
- chunking
- coreference
- sense disambiguation
- paraphrase
- information extraction
The classes will be a mix of lectures, discussion, and student
presentations.
In addition to preparing two or three presentations for the course
(covering specific articles and the student's project), students
will be expected to run a number of smaller experiments, and one larger
experiment
as a term project.
Students should
have
- some background in natural language processing (if you have not
taken a course in NLP such as G22.2590,
there will be assigned readings from the Jurafsky and Martin text, Speech
and Language Processing, to provide background for this course)
- good programming skills to assemble substantial programs for
class experiments
- sufficient mathematical background to be able to read papers in
machine
learning which include arguments regarding statistics and probability
For further information, contact the instructor at grishman@cs.nyu.edu