Twitter data mirrors opinion poll results, study finds

A team from Carnegie Mellon University’s School of Computer Science and Tepper School of Business in Pittsburgh, Pennsylvania, looked at surveys of consumer confidence and presidential approval published by various organisations during 2008 and 2009 and found that they correlated to a ‘sentiment ratio’ based on words and phrases used in tweets in the same period.

“While our results vary across datasets,” they write in a paper published at a social media conference later this month, “in several cases the correlations are as high as 80%, and capture important large-scale trends.”

“The results highlight the potential of text streams as a substitute and supplement for traditional polling,” they concluded.

The team also looked at whether Twitter data was predictive of poll results, but found that results were very different at different times. This could be a result of the changing nature of Twitter, which grew vastly in usage during the period, the researchers said, but it also suggests more work is needed to understand how events and the passage of time influence how words are used.

A 72% correlation was found between Twitter sentiment and job approval ratings in the period after Barack Obama took office in 2009, although there was “no substantial correlation” with results of voting preference polls in the lead-up to the election. It remains unclear, the team said, how the names of election candidates should fit into a model of voting intentions.

“We find that a relatively simple sentiment detector based on Twitter data replicates consumer confidence and presidential job approval polls,” the researchers said. “While the results do not come without caution, it is encouraging that expensive and time-intensive polling can be supplemented or supplanted with the simple-to-gather text data that is generated from online social networking. The results suggest that more advanced NLP (natural language processing) techniques to improve opinion estimation may be very useful.”

They said that sentiment analysis could be useful for modelling traditional survey data and as a stepping stone toward “larger and more sophisticated applications”.

The paper, by Brendan O’Connor, Ramnath Balasubramanyan, Bryan Routledge and Noah Smith, is available online here.