Looks like he took the extraction approach [1] - copy the information deemed most important by the system to the summary. The second approach, abstraction, involves paraphrasing sections of the source document. It seems much harder.

There are some interesting summarizers accessible on the web : http://search.iiit.ac.in/~jags/summarizer/index.cgi

Machine learning, as always, is uber-cool.

[1] http://en.wikipedia.org/wiki/Automatic_summarization

Somewhat related, see this recent article (6/10/09):

"Extracting Meaning from Millions of Pages: University of Washington software pulls facts from 500 million Web pages."

http://beta.technologyreview.com/computing/22773/