Election 2012 Roundup: Four Faculty Members on This Year's Campaign

October 18, 2012

Kathleen McKeown
Henry and Gertrude Rothschild Professor of Computer Science

Q. Tell us about your project and how it relates to the election.

Over the course of the last year, we have been collecting data from online political forums, Twitter and different kinds of online discussions. We have also collected news. The program we have created identifies influencers. We look to identify who was able to change people’s minds and how; where sentiment shifted; and what arguments people found persuasive. And we look to correlate shifts in sentiment with events as reported in the news.

Q. How does it work?

The systems detect influence using several programs that perceive components of influence. For example, one program can detect in online discussions whether a participant feels positively, negatively or neutral towards a topic and whether other people on the forum agree or disagree. We use machine-learned algorithms that recognize sentiment based on the way language is used—for instance, negative or positive expressions such as “I like” or “I hate.” The computer uses all of these cues to identify that this person is in favor of a particular claim or this other person is opposed.

We have another algorithm that can detect whether someone is making a claim about a particular topic. And we can detect when someone is attempting to persuade people to go along with their opinion. Then we can look at how those arguments change the sentiment of others or the direction of the conversation.

When we put them all together, we can determine whether somebody is being influential and when they change other people’s opinions. We are still building these systems and plan to look at other kinds of information, such as age and gender.

Q. How would this relate to the national discourse?

We might use it to analyze how people reacted to particular events, which events were important in determining opinions, and how in online social media, influence played a role in determining outcomes. So we might look at whether changes in opinion are correlated to a particular event in the news, like the recent attack in Libya.

There might be a discussion about health care on a forum, and you might have two people taking different sides about whether Obamacare is helpful or not. By looking at what claims are made and who agrees or disagrees, we can identify whether one person was able to influence other people.

Also, we can look at the influence of Obama or people who work for him, what they are reported to have said, how often it was mentioned in the forums and what effect that had on people’s sentiments.

I would note that our programs are not perfect, they have error rates. These are hard problems, and they are not easily solved. Our error rates can go from 10 to 30 percent depending on the component and the data.

Q. What is your overall goal?

Ultimately, we want to be looking across quite a few forums to see what is happening. We can aggregate that information to find out how influence is expressed in different genres, which forums are most influential, how influences changes over time, who within a forum is most influential and how do they get people to agree. Ultimately, we want to be able to detect influence for a variety of purposes such as advertising or online problem solving.

Q. How did you get interested in this field?

I did my undergrad at Brown in comparative literature. So I had an interest in language at that point. I did both French and English, and looked at all time periods. I loved James Joyce and Baudelaire. A lot of my friends were doing computer science and were very enthusiastic about it. So towards my senior year I began taking course in computer science. I learned at the end of undergraduate that there was a field where you could combine computers and language. I heard University of Pennsylvania had a great program in it, and I went there to do a PhD with Aravind Joshi one of the pioneers in the field of natural language processing.

—Interview by Adam Piore