Machine learning and natural language processing - The Computing Universe: A Journey through a Revolution - page 293

Fig. 14.10. The Haiti earthquake in 2010
reduced much of the capital, Port au
Prince, to rubble. Using machine learn-
ing on the available “parallel” corpus of
Creole-English texts - such as the Bible -
it was possible to build a translator in
less than five days. The National Palace
in Port au Prince (a) before and (b) after
the earthquake.

(a)

(b)

IBM's Watson and Jeopardy!

After IBM's success with Deep Blue and computer chess in 1997, the com-
pany had been looking for an equally audacious challenge with which to cap-
ture the public's imagination ( Fig. 14.12 ). In 2005, after a suggestion from IBM
manager Charles Lickel, the director of IBM Research, Paul Horn, tried to inter-
est his researchers in building a machine to beat humans on Jeopardy! The game
attracted many millions of viewers, and as Horn said, “People associated it with
intelligence.” 9
The American TV quiz show Jeopardy! debuted on the NBC network in 1964.
It is a quiz game with contestants competing to match questions to answers on
a wide variety of topics. The U.S. television host Merv Griffin devised the game.
He credits its strange reverse answering style, in which contestants receive
clues in the form of answers and must frame their responses as questions, to a
conversation with his wife:

100

80

Switchboard test

60

40

20

0

1992 1994 1996 2000 2009 2010 2011 2012

Year

My wife Julann just came up with the idea one day when we were in a plane
bringing us back to New York from Duluth. I was mulling over game show
ideas, when she noted that there had not been a successful “question and
answer” game on the air since the quiz show scandals. Why not do a switch,
and give the answers to the contestant and let them come up with the
question? She fired a couple of answers to me: “5,280” - and the question of

Fig. 14.11. The change in the WER with
time for the U.S. National Institute of
Standards “Switchboard” test. This
shows the dramatic improvement made
in the last few years using Deep Neural
Network techniques.