Document Type

Conference Document



Embargo Period



Machine learning, feature selection, sentiment classification, text categorization, e-Rulemaking




Library and Information Science


Recently there has been increasing interest in constructing general-purpose political opinion classifiers for applications in e-Rulemaking. This problem is generally modeled as a sentiment classification task in a new domain. However, the classification accuracy is not as good as that in other domains such as customer reviews. In this paper, we report the results of a series of experiments designed to explore the characteristics of political opinion expression which might affect the sentiment classification performance. We found that the average sentiment level of Congressional debate is higher than that of neutral news articles, but lower than that of movie reviews. Also unlike the adjectivecentered sentiment expression in movie reviews, the choice of topics, as reflected in nouns, serves as an important mode of political opinion expression. Manual annotation results demonstrate that a significant number of political opinions are expressed in neutral tones. These characteristics suggest that recognizing the sentiment is not enough for political opinion classification. Instead, what seems to be needed is a more finegrained model of individuals' ideological positions and the different ways in which those positions manifest themselves in political discourse.