Ideology Classifiers for Political Speech

Abstract

In this paper we discuss the design of ideology classifiers for Congressional speech data. We then examine the ideology classifiers' person-dependency and time-dependency. We found that ideology classifiers trained on 2005 House speeches can be generalized to the Senate speeches of the same year, but not vice versa. The ideology classifiers trained on 2005 House speeches predict recent year Senate speeches better than older speeches, which indicates the classifiers' time-dependency. This dependency may be caused by changes in the issue agenda or the ideological composition of Congress.

Keywords: machine learning, text classification, generalizability, ideology, evaluation