Classifying Topics and Detecting Topic Shifts in Political Manifestos

General political topics, like social security and foreign affairs, recur in electoral manifestos across countries. The Comparative Manifesto Project collects and manually codes manifestos of political parties from all around the world, detecting political topics at sentence level. Since manual coding is time-consuming and allows for annotation inconsistencies, in this work we present an automated approach to topical coding of political manifestos. We first train three independent sentence-level classifiers – one for detecting the topic and two for detecting topic shifts – and then globally optimize their predictions using a Markov Logic network. Experimental results show that the proposed global model achieves high classification performance and significantly outperforms the local sentence-level topic classifier.