scriptive phrase is stronger: clean, almost spotless, or sparkling. It should
be able to distinguish the positive—
”The room was nice and quiet”—from
the negative—”I was disappointed the
room wasn’t quieter.”
Blog and Twitter Searches
One growing area that poses new challenges for search engines is social
media, such as blogs, Twitter feeds,
and Facebook status updates. “I don’t
think we have really good blog search
yet,” says Marti A. Hearst, a professor
in the University of California, Berkeley School of Information. Along with
Microsoft researchers Susan T. Dumais
and Matthew Hurst, Hearst says blog
search should be able to accomplish
three tasks: find out what people are
thinking about a certain topic over time;
suggest blogs that are good to read for
their style, personality, and other criteria; and find useful information in older
blog posts, along the lines of standard
search of more static documents.
Blog search needs to take into account the differences between blogs
and traditional documents, such as
the former’s use of more informal language, their different link topology, the
importance of timeliness, and the fact
that updates tend to not be full HTML
pages. Blog search must also take into
account that much of the information
on blogs is subjective.
To accomplish these tasks, search
engine designers look for representations of features that might belong
to a particular class of posting, such
as the readability level of a page. Machine learning algorithms can then
figure out that particular distributions
of features may be characteristic of a
“If you have labeled data and examples of things that you think have a
particular attribute, then you can use
that to find something similar,” says
Dumais, principal researcher in Microsoft Research’s Adaptive Systems
and Interaction Group. But rating postings as positive or negative, or figuring
out whether they’re aimed at an older
or younger audience or have a left-leaning, right-leaning, or middle-of-the-road viewpoint, is challenging, she
says. “They do involve a richer understanding of language than most search
engines have,” Dumais notes.
Search can be
improved through a
of a document’s
meaning and a better
grasp of a searcher’s
Twitter use has grown explosively
in recent months, and in October the
company made a deal to open its data to
Microsoft’s search engine. Dumais says
that, with its 140-character limit leading to creative abbreviations of words
and condensed hyperlinks, searching Twitter will pose some interesting
challenges. But once those are tackled,
Twitter users should be able to conduct
more refined searches than the service
currently allows, while the flow of Twitter data provides search designers with
new information that may make search
richer. “The volume of the content [on
the Web] is actually very useful for some
types of algorithms,” Dumais says.
One useful fact is that people with
Twitter feeds and Facebook pages are
making public a lot of information
about themselves that search engines
can use to better understand their
search queries. Just as search can be improved through a deeper understanding
of what documents mean, it can also
improve through a better grasp of the
searcher’s intentions. “The real issue
with a search engine is not just to serve
up results, but to help people accomplish what they’re trying to do,” says Jon
Kleinberg, a professor of computer science at Cornell University.
Search engines trying to provide
the right answer to a query might take
into account what a user has previously
searched for. If a user is looking for a
restaurant or a movie recommendation, the search engine might look at
the user’s friends lists and see what
those presumably trusted sources
liked. And if the user is searching from
a mobile device, that might provide additional clues.
If nothing else, a search from a mobile phone tells the search engine it is
from a phone, so perhaps a search for a
person is really a search for their phone
number. And many mobile devices use
GPS or cell phone towers to determine
their location. A person typing “
Yankees” in Manhattan may be looking
for tickets to tonight’s baseball game,
whereas the same search in Seattle may
represent a desire for last night’s score.
“In a relatively short time frame, we’re
going to think of geolocation as an integral component of a lot of the online
activity we do,” Kleinberg says.
Time is also becoming a characteristic to take into account, Kleinberg says.
One way of judging the importance of a
news story, for instance, is how quickly
it spread and how long interest focused
on it. Dumais points out that many
facts have a time component as well.
The gross national product of Norway,
the population of Brazil, and the prime
minister of Japan—all can have one factual answer in 2000 and a different one
Dumais says future search engines
will have both a better grasp of the intent
of a query and a richer understanding of
Web content. “We’re looking at how we
can support that in ways that go beyond
2.3 words typed into a search box and 10
blue links,” she says.
Etzioni, O., Banko, M., Soderland, S., Weld, D.
Open information extraction from the Web.
Commun. of the ACM 51, 12, 2008.
Hearst, M. A.
Emerging trends in search interfaces.
Search User Interfaces, Cambridge
University Press, New York, NY, 2009.
Backstrom, L., Kleinberg, J., Kumar, R., Novak, J.
Spatial variation in search engine queries.
Proc. 17th Int’l Conf. on World Wide Web,
Hearst, M.A., Hurst, M., Dumais, S. T.
What should blog search look like? Proc. of
the 2008 ACM Workshop on Search in Social
Downey, D., Dumais, S. T., Liebling, D., Horvitz, E.
Understanding the relationship between
searchers’ queries and information goals.
Proc. 17th ACM Conf. on Information and
Knowledge Management, 2008.
Neil Savage is a science and technology writer based in
© 2010 ACM 0001-0782/10/0100 $10.00