Philip Resnik's research in computational political science

Note: This page needs an update. The work described here may not be current, but it does provide a good sense of my research approach and may therefore still be useful.

Computational political science is emerging as an incredibly interesting research area, one that shares the momentum that's building in computational social science more generally. I've been finding myself involved in a number of public conversations about this topic, and in my work I've been focusing on the following:

Outside the academic context, I'm also an advisor to Converseon, a leading social media consultancy, and I was founder of React Labs, which commercialized my work on mobile real-time response technology.

For more information about me or my research, see my home page.

Syntactic Framing

One of my very favorite linguistic examples: when he was about four years old, why did my son say "My toy broke" instead of "I broke my toy"? He was using syntax to package up the statement about what happened in a way that de-emphasizes semantic properties such as causation, volition, and change-of-state. This is an example of using syntax to frame the issue, an example of "spin" no different from what Ronald Reagan did in 1987 when he sidestepped attributing responsibility for the Iran-contra scandal; remember "Mistakes were made"? (Precocious child.) My student Stephan Greene did a fascinating dissertation on this topic, and for a conference-paper-length description see our 2009 NAACL paper. Some more recent work includes replicating some of his basic results in Arabic in collaboration with Aya Zirikly.

React Labs: real-time polling

One of the big challenges facing data-driven work in political science is gathering good data. Analyzing Twitter is all well and good, but it has significant disadvantages, too. What you'd really like is something that combines the control you get in focus groups or polls (particularly the ability to focus on particular questions) with the large-scale, instantaneous reactions of social media populations.

I started the React Labs project to create just that.

The project developed an app that lets people watch live events like political debates and speeches and react to them using their smartphones or other mobile devices in real time. During the event, live charting is updated second by second to provide a dynamic picture of people's reactions, and it's also possible to push survey questions to them dynamically. The ambition is essentially to put a more sophisticated kind of Nielsen dial in the hands of every smartphone user in the country. It can also be used in more controlled settings -- for example, enabling the audience to provide real-time feedback to a speaker. A journal article about the project appeared in the Spring 2014 issue of Public Opinion Quarterly.

I am working with folks at the Maryland Technology Enterprise Institute on commercializing this technology. The platform debuted commercially in a partership with ABC7/WJLA-DC during the 2012 Obama-Romney debates and, in the process, helped a team comprising NewsChannel8, Politico, and ABC7/WJLA-TV, win a 2013 Walter Cronkite Award for Excellence in Television Political Journalism for their election coverage (see this announcement).

React Labs smartphone app January 7 debate example Feb 22 debate example

Computational modeling of sentiment, perspective, and framing

In their important 2008 book on the topic, Bo Pang and Lillian Lee pointed out that there was a "land rush" starting in the area of sentiment analysis. Today it's easy to find companies tracking sentiment about brands and products in social media for business purposes. The political world has caught on, as well, with lots of people tracking mentions of political candidates and some even suggesting that automatic analysis of opinions on Twitter could be better than traditional polling. I have talked about the use of sentiment analysis in political contexts, and some of the issues and hype surrounding it, in some recent (and by now also not-so-recent) public discussions.

In my current academic research, I'm less focused on straight positive-versus-negative sentiment analysis (though sentiment analysis for other languages remains a topic of interest). What's really got me excited is the idea of applying the same sorts of computational ideas to the related topic of framing -- that is, the way that language can be used to emphasize or de-emphasize different aspects of a topic or issue, often in service of a particular agenda or ideological perspective. (When this is done deliberately in political discourse, particularly by someone we disagree with, we often call it spin.)

Until pretty recently, framing was a subject that had received relatively little attention from a computational perspective. As an example of some relatively early work, my student Stephan Greene did a nice dissertation laying the foundations for a linguistically well founded approach to framing, focusing on the way that grammatical choices encourage particular interpretations, which we summarized in a 2009 paper. (When my youngest son says "Daddy, my toy broke" to de-emphasize his own role in what happened, he's exploiting the same linguistic strategy Ronald Reagan did when he famously said "Mistakes were made". Notice that this isn't just about passive voice; he is exercising the causative/inchoative alternation, not the active/passive distinction.)

SHLDA model of agenda-setting and framing

In more recent work, colleagues and I have been developing computational models of this syntactic framing phenomenon, as well as lexical framing (e.g. the choice of whether to say death tax or estate tax), and hierarchical models for framing that are inspired by theories of framing as second level agenda-setting. For example, in the figure above, a hierarchical supervised analysis of Congressional floor debates has automatically identified agenda issues including taxation (B), and it distinguishes the Republican-oriented framing of taxation in terms of business interests (C) from the Democratic framing in terms of the programs that taxation supports (children, education, health care veterans, etc. Colleagues and I have also recently looked at intra-Republican framing in Viet-An Nguyen, Jordan Boyd-Graber, Philip Resnik, and Kristin Miler, Tea Party in the House: A Hierarchical Ideal Point Topic Model and Its Application to Republican Legislators in the 112th Congress" (Association for Computational Linguistics Conference, Beijing, July, 2015).

Here are links to some other recent publications on supervised hierarchical topic models of agenda-setting and framing, political ideology detection using recursive neural networks, learning concept hierarchies from multi-labeled documents, identifying media frames and frame dynamics within and across policy issues.

Computational modeling of influence

Who has control of the agenda in a political conversation, and what do they do with it? Clearly if someone has the role of moderator, they exercise some control over what gets talked about. But how strong is that control? And can we detect and measure objectively when somebody takes control over the agenda in a conversation? Jordan Boyd-Graber, grad student Viet-An Nguyen, and I have developed computational models that analyze a conversation like a political debate or an episode of Crossfire, and automatically identify both the topics under discussion and the degree of topic control that participants are exercising.

As an example, the automatic analysis of the Biden-Palin debate in 2008 identifies that Gwen Ifill, the moderator, was the person controlling the topic most of the time (long gray bars in the "Shift" column), but it also automatically uncovers situations like turn 48 of in debate transcript, where Palin responded to a question about the mortgage crisis by talking about energy policy. (This is a particularly clean example, but you'll certainly find that Biden did the same sort of thing, too!)

Topic control example

Interestingly, if you look at the degree of topic control by all participants in the four 2008 presidential debates, you find that Palin had the strongest propensity among all the candidates to exert control over the topic being discussed, and you also find objective empirical support for the widespread perception that Ifill was a weak moderator. (Compare Biden and Palin to Ifill, below, then compare McCain and Obama to their moderators.)

Topic control example

For a conference length description see Viet-An Nguyen, Jordan Boyd-Graber, and Philip Resnik, "SITS: A Hierarchical Nonparametric Model using Speaker Identity for Topic Segmentation in Multiparty Conversations", in Proceedings of the 50th Meeting of the Association for Computational Linguistics, Jeju, South Korea, July 8-14, 2012. Full details appear in our 2013 article in Machine Learning Journal, and Viet-An's doctoral dissertation.

Conversations about computational social science