Automatic Translation, Context, and Supervised Learning in Comparative Politics

Abstract

This paper proves that automatic translation of multilingual newspaper documents deters neither human nor computer classification of political concepts. We show how theory-driven coding of newspaper text can be automated in several languages by monolingual researchers. Supervised machine learning is successfully applied to text in English from British, Spanish and German sources. The paper has three main findings. First, results from human coding directly in a foreign language do not differ from coding computer-translated text. Second, humans can code translated text as well as they can code untranslated prose in their mother tongue. Third, machine learning based on translated Spanish and German training sets can reproduce human coding as accurately as a system learning from English training sets.