Program analyzes tweets to predict Oscar winners

With the Academy Awards coming up Sunday, the Annenberg Innovation Lab hopes its new tool will help predict Oscar winners.

The Oscar Senti-meter, developed in partnership with the Annenberg Innovation Lab, the Viterbi Signal Analysis and Interpretation Lab, and the Los Angeles Times, analyzes opinions using thousands of tweets about the Academy Awards to chart popular opinions.

The device uses language-recognition software Streams to find tweets about best actress, best actor and best picture nominees and plots the results on a graph.

Annenberg Innovation Lab director Professor Jonathan Taplin said the lab used the IBM-donated software for several other projects before the Senti-meter.

The lab first used the software to analyze political tweets, then moved to track the success of the box office.

The team first became inspired with the idea for the Senti-meter by the software’s accuracy at predicting the success of movies, Taplin said.

“We saw that Cowboys and Aliens was going to bomb in the box office when the studios thought it was going to be a big movie,” Taplin said. “We saw The Help was going to do well when Disney wasn’t too sure about it.”

Taplin said the Senti-meter measures anywhere from 300,000 to 400,000 tweets each day.

Shrikanth Narayanan, director of Viterbi SAIL and a professor of psychology, linguistics and computer science, said the lab had previously used the software for understanding language in settings such as marriage counseling, but Twitter presented a new challenge.

“The power of Twitter is in using only a few characters to pack a lot of your emotions and feelings into the message,” Narayanan said. “What we try to do is see the kind of words people use and how they use them to map it to their sentiment.”

The shortness of tweets, however, can also make them more difficult for the software to read. But now the Senti-meter can understand complex features of language such as emoticons and sarcasm, though Taplin said achieving that result took hard work.

“What we do is we have students look at 5,000 or 7,000 tweets and compare their meaning to what the computer thought they meant,” Taplin said.

Nuances in meaning that humans naturally analyze in speech patterns can be difficult for the computer to read as well.

“Sentiment expressions are very complex to quantify. People might say one thing but mean the opposite, such as when they are being sarcastic. But if you look in context, you can try to figure out what the words really mean,” Narayanan said.

Narayanan said he hopes to put the technology to use in other areas.

“One of the big problems that interest us a lot is to watch for patterns over time, see how opinions change over time and see how long opinions last,” Narayanan said.

The Oscar Senti-meter will be updated on the Los Angeles Times website.