Language and Politics

William Cohen, Noah Smith, and Tae Yano

Most approaches to automatic text analysis and processing treat text as a stream of words or sentences. A typical underlying assumption is that the use of language in the data is literal and that the data represent facts. Many genres, however, do not have these features.

We are exploring automatic methods for analyzing text in the political domain, specifically blog posts on topics pertinent to the 2008 United States Presidential Elections. Political text is often indirect, sarcastic, repetitive, hyperbolic, emotional, biased, manipulative, and riddled with unstated assumptions. Our aim is to automatically separate useful, thoughtful information from redundant "spin," using statistical natural language processing techniques and a data-driven methodology that makes use of the insights of political scientists.

The broader impact of this work will consist of a renewed emphasis exploiting domain knowledge together with text data for more powerful natural language understanding technology, as well as software tools that will promote more informed decision-making among American voters.