Understanding machines and the mind.
Dr. Aravind Joshi, GEE’58, GrE’60, has won the International Joint Conference on Artificial Intelligence Research Excellence Award for 1997, the highest citation in its field. It is given every two years to a scientist who has carried out a research program of consistently high quality, yielding substantial results over a long period of time. “It’s really an honor,” says Joshi, the Henry Salvatori
Professor of Computer and Cognitive Science and Professor of Linguistics
at Penn. “It’s a recognition by your peers both here and abroad.” He
will accept his award and address the conference August 23-29
in Japan.
In
1958 and 1959 he developed the world’s first parser for natural
language at Penn, in a project directed by the late Dr. Zellig Harris, G’30, Gr’34, a linguistics professor. (Dr. Lila Gleitman, Gr’67, psychology
professor and currently co-director with Joshi of Penn’s Institute for
Research in Cognitive Science, also worked on the team.) Joshi also
invented Tree-Adjoining Grammar, or T.A.G.
A
parser, in simple terms, breaks down sentences into their grammatical
parts (Think of a sophisticated version of the sentence diagramming that
schoolchildren do). T.A.G., which is used by some of today’s parsers,
operates much like a dictionary, telling what grammatical structures are
associated with which word. For example, with the verb “hit,” it will
give a structure for using that word, with open slots for a subject
noun-phrase and an object noun-phrase.
T.A.G.
has only two ways of combining structures: substitution — In the
sentence “John hit Bill angrily,” for example, “John” is substituted in
the open slot for the subject — and adjunction, attaching one structure
to another. “Angrily,” an adverb, is attached to the end of the phrase
“John hit Bill.” All constructions in English and other languages can be
made with these two operations, Joshi says.
“Parsing
is really a way of figuring out what a sentence really means,” he says.
“Roughly speaking, it tells you who did what to whom and how … The
idea is that when we understand [a] sentence, we are actually doing some
computation like this internally.”
Applications
for parsing include language translation, information extraction, and
summarization. “You read a newspaper story and you want information only
about Bill Gates, for example,” Joshi says. “…It’s not that easy,
because in one sentence, for example, it might be ‘he.'” To know if ‘he’
refers to Gates requires analyzing each sentence.
“If
you want a machine to do it,” Joshi says, “the tasks that appear quite
easy to us are actually not so easy at all.”
Joshi grew up in India and earned his bachelor’s degree in electrical
and mechanical engineering at Poona University. When he arrived at Penn
as a graduate student in 1954, there was no computer science department;
that field grew out of electrical engineering. He gradually shifted his
concentration from electrical engineering to computer science, and also
did work in the linguistics department.
Joshi co-directs the IRCS at Penn, which brings together researchers in the pursuit of understanding the human mind.
By Susan Lonkevich