Polarizing discussions on political and social issues are common in mass and user-generated media. However, computer-based understanding of ideological discourse has been considered too difficult to undertake. In this paper we propose a statistical model for ideology discourse. By ideology we mean “a set of general beliefs socially shared by a group of people.” For example, Democratic and Republican are two major political ideologies in the United States. The proposed model captures lexical variations due to an ideological text’s topic and due to an author or speaker’s ideological perspective. To cope with the non-conjugacy of the logistic-normal prior we derive a variational inference algorithm for the model. We evaluate the proposed model on synthetic data as well as a written and a spoken political discourse. Experimental results strongly support that ideological perspectives are reflected in lexical variations.