Software sorts out subjectivity
Technology Research News
of the fundamental challenges in getting computers to sort and analyze text
is finding ways to automatically classify information.
Applications like search engines that group similar documents do
so using topic-based categories. Sentiment analysis techniques add another
dimension by determining the author's attitude about a topic rather than
just identifying a topic.
Existing techniques tend to concentrate on finding words, phrases
and patterns that indicate sentiment. This has proven difficult, however.
"This laptop is a great deal", for instance, shows strong sentiment, but
contains the same words as the neutral sentence "The release of this new
laptop drew a great deal of media attention."
In this example, it's not just the presence of a cue word like "great"
that matters, but also its meaning in context.
People can easily tell the difference between the phrases because
they understand the meaning of the words. Enabling computers to deal with
meaning is an extremely difficult challenge, however.
Researchers from Cornell University have devised a way to improve
sentiment classification that sidesteps having to deal with meaning by instead
concentrating on context. Their method weeds out neutral sentences. "Getting
rid of neutral sentences like 'The release of this new laptop drew a great
deal of media attention' [makes] the overall sentiment more obvious," said
Lillian Lee, an associate professor of computer science at Cornell University.
The method improved sentiment classification performance from 82.8
to 86.4 percent, which is statistically very significant, according to Lee.
The method could eventually be used to maintain review-aggregator Web sites,
to filter search results by viewpoint, and to track attitudes toward a given
topic, she said.
It is not readily apparent that classifying text as subjective or
objective is any easier than classifying text as positive or negative, said
Lee. But it turned out to be easier simply because people tend to switch
between objective and subjective statements less often than they switch
between positive and negative phrases.
A movie reviewer, for instance, may begin with several sentences
of objective text concerning a movie's plot before switching to a subjective
statement about how good the movie was, said Lee. "If the sentence appears
in the context of a block of other obviously objective sentences, there's
a good chance that it is also objective," she said.
To take advantage of this clustering, the researchers represented
text as a network, or graph. "Imagine that each sentence is represented
by a network point, or node," said Lee. To model contextual information
between each pair of sentence nodes, the researchers added a link whose
strength represented how much the two sentences deserved the same label
-- objective or subjective -- based on criteria including how close the
sentences are to the text, and whether they are separated by a paragraph
The model also took into consideration the evidence within a sentence
that the sentence is subjective or objective. Possible evidence that a sentence
is subjective, for example, includes the presence of a word like 'wonderful',
or 'terrible', said Lee.
Each sentence was linked strongly or weakly to a special subjective
and objective nodes depending on the amount of evidence there was within
the sentence that it was subjective or objective.
The sentences are then clustered into subjective and objective camps
based on the strength of the links. This is a graph partitioning problem
known as finding the minimum cut, and it can be solved exactly by a quick,
efficient algorithm, said Lee.
One way to visualize how this works is to picture someone taking
the special subjective node in one hand and the special objective node in
the other hand, and pulling them in opposite directions so that the weaker
links snap until the network is broken into two pieces, said Lee. "Two sentences
that prefer to be in the same class will tend to be in the same piece because
they had a strong link between them, but they could still be separated if
they have very strong links to opposite special nodes," she said.
Once the subjective vs. objective classification is done, the researchers
use standard pattern recognition techniques to classify each document as
positive or negative based just on the portions identified as subjective.
The researchers found that seemingly empty words and phrases can
turn out to be unexpectedly informative when it comes to sentiment classification.
In the context of movie reviews, for example, the word "good" provides less
evidence for positive sentiment than the word "still" followed by a comma,
said Lee. "This makes sense in retrospect -- a typical use would be something
like 'still, this film is worth seeing' -- but illustrates how subtle the
sentiment problem can be."
The researchers are working on improving their method for estimating
the affinity sentences have for being classified the same way, said Lee.
"We used very simple cues like distance... but more sophisticated information
ought to be incorporated," she said.
Longer-term, they are aiming to develop methods that can handle
variations in language, said Lee. "This is very important in dealing with
on-line text, since Internet sources can very widely in form, tenor and
even grammaticality," she said. "One can get reviews from the highly-edited
New York Times or from a stream-of-consciousness personal Web log."
The ultimate aim is to be able to handle rhetorical devices like
irony and sarcasm, said Lee. "Given that even humans are occasionally misled
by such rhetorical devices, this is going to be very challenging," she said.
People are incredibly creative at expressing negative opinions,
said Lee. For example, this sentence not only contains no obviously negative
words, but has a lot of potentially positive words: "If you think this laptop
is a great deal, I've got a nice bridge you might be interested in."
The system could be deployed now for domains that have fairly consistent
language and training data that the system can use to learn what cues work
in that domain, said Lee.
It will take at least a decade before the system can readily handle
unrestricted texts containing arbitrary rhetorical devices, she said.
The method could be used to automate the maintenance of review-aggregation
sites, said Lee. "A system could crawl on-line information sources and automatically
extract ratings, even from documents like New York Times book reviews that
don't include explicit scores," she said.
It could be used by search engines to sort or filter results by
viewpoint to, for instance, help users distinguish between objective and
biased Web sites, said Lee.
It could also be used to track changes in attitudes toward a given
topic by, for instance, analyzing press articles, she said. "An analyst
might desire a summary of the international press's reaction to a particular
act of political violence, as well as a list of which countries approve
of the act and which condemn it," she said.
And companies could use the system to gather business intelligence
such as finding out what people think of their products or the products
of their competitors. "A computer company might crawl blogs to find out
whether or not people like its latest laptop model," said Lee.
Lee's research colleague was Bo Pang. The research is published
in the Proceedings of the 42nd Annual Meeting of the Association for Computational
Linguistics, held July 21 to 26, 2004 in Barcelona, Spain.
The research was funded by the National Science Foundation (NSF),
the Alfred P. Sloan Research Foundation, and the Cornell cognitive studies
Timeline: 10 years
Funding: Government, Private, University
TRN Categories: Natural Language Processing; Databases and
Story Type: News
Related Elements: Technical paper, "A Sentimental Education:
Sentiment Analysis Using Subjectivity Summarization Based on Minimum Cuts,"
published in the Proceedings of the 42nd Annual Meeting of the Association
for Computational Linguistics, held July 21 to 26, 2004 in Barcelona, Spain
and posted at arxiv.org/abs/cs.CL/0409058
November 17/24, 2004
Fibers mix light
Software sorts out
Nanotubes tune in light
material holds hydrogen
artificial hip wear
make 3D color display
Lasers drive nano
plastic holds up
Atom flip energy measured
Research News Roundup
Research Watch blog
View from the High Ground Q&A
How It Works
News | Blog
Buy an ad link