Home' Technology Review : January February 2008 Contents Q&A
As director of research at Google,
Peter Norvig is intimately
involved in the attempt to man-
age the world's information. He's a good
match for the job, having spent much of
his life thinking about how computers
think and making them do it more e -
ciently. An expert on artificial intelligence,
he has taught at universities, held research
jobs in the corporate world and at NASA,
and cowritten the influential textbook AI:
A Modern Approach.
Norvig came to Google in 2001 as the
director of search quality; he assumed his
current position four years later. In that
role, he oversees about 100 computer sci-
entists as they work on projects as diverse
as medical records management and
machine translation. An untold number
of Google servers housing the search-
able Web provide them with a test bed. He
says Google is structured to ensure that
researchers are not sequestered from the
rest of the company. "The main allegiance
they have is to the product they're working
on," he says.
When Norvig arrived in Mountain View,
Web search was simply about serving up
the pages most relevant to a given query.
But as the Web has grown, so has people's
need to filter information quickly. Norvig
recently spoke with Technology Review's
information technology editor, Kate
Greene, about what's next for Web search.
TR: Google has many innovative products,
but the look and feel of Web search hasn t
changed much in 10 years. Why?
Norvig: We've hit on something that
people mostly liked. We weren't the first
to do it. Go back to Excite and the search
engines before: you have a box, and you
get a list of 10 results, with a little bit of
information accompanying each result.
We've just stuck with that.
What has changed?
The scale. There's probably a thousand
times more information. It used to be
just Web pages; now it's video, pictures,
blogs, and all sorts of media and formats.
Also, the immediacy has changed. When I
started, we were updating the index once
a month. We thought of it as a library cata-
logue, a long-term thing. Now we're seeing
it more as up-to-the-minute media. When
news breaks, you want to be able to read it
in minutes, not in days, weeks, or months.
You claim that Google s accuracy is pretty
good. How do you know how good it is, and
how do you make it better?
We test it in lots of ways. At the grossest
level, we track what users are clicking on.
If they click on the number-one result, and
then they're done, that probably means
they got what they wanted. If they're
scrolling down, page after page, and refor-
mulating the query, then we know the
results aren't what they wanted. Another
way we do it is to randomly select specific
queries and hire people to say how good
our results are. These are just contractors
that we hire who give their judgment. We
train them on how to identify spam and
other bad sites, and then we record their
judgments and track against that. It's more
of a gold standard because it's someone
giving a real opinion, but of course, since
there's a human in the loop, we can't
a ord to do as much of it. We also invite
people into the labs, or sometimes we go
into homes and observe them as they do
searches. It provides insight into what
people are having di culty with.
Companies such as Ask and Powerset
are betting that the future is in natural-
language search, which lets people use
real, useful sentences instead of poten-
tially ambiguous keywords. What is Google
doing with natural language?
We think what's important about natu-
ral language is the mapping of words onto
the concepts that users are looking for.
But we don't think it's a big advance to be
able to type something as a question as
opposed to keywords. Typing "What is
the capital of France?" won't get you better
results than typing "capital of France." But
understanding how words go together is
important. To give some examples, "New
York" is di erent from "York," but "Vegas"
is the same as "Las Vegas," and "Jersey"
may or may not be the same as "New Jer-
sey." That's a natural-language aspect that
we're focusing on. Most of what we do
is at the word and phrase level; we're not
concentrating on the sentence. We think
it's important to get the right results rather
than change the interface.
How much will Google search become per-
sonalized to individual users?
We're doing some of that in various
places. One good example is in news per-
sonalization, where we give recommenda-
tions for news articles. There, it's easier to
do than in larger Web databases, because
there's a limited number of news stories.
We track what news stories you look at,
and we compare it to other people. And
that seems to work out well. It's harder to
apply it to something as vast as the whole
Web, but we're starting with the easy parts.
Where do you see Google search in two to
You'll see integration of various kinds
of content. We're getting into speech rec-
ognition and all the kinds of interfaces
on phones, where you have a tiny screen
and awkward keyboard. You'll see that
gaining in importance. You'll see integra-
tion of our various properties. We used to
put the onus on the user and ask them if
they wanted Web search or image search
or video search. Now we're trying to solve
that for them and serve up the results in a
way that makes sense.
The evolution of Web search
Photograph by HOWARD CAO
Links Archive March April 2008 November December 2007 Navigation Previous Page Next Page