Using supervised machine learning to code policy issues: Can classifiers generalize across contexts?

Content analysis of political communication usually covers large amounts of material and makes the study of dynamics in issue salience a costly enterprise. In this article, we present a supervised machine learning approach for the automatic coding of policy issues, which we apply to news articles and parliamentary questions. Comparing computer-based annotations with human annotations shows that our method approaches the performance of human coders. Furthermore, we investigate the capability of an automatic coding tool, which is based on supervised machine learning, to generalize across contexts. We conclude by highlighting implications for methodological advances and empirical theory testing.