WITP 5(01).book(WITP_A_315580.fm)

Many research questions in political communication can be answered by representing text as a network of positive or negative relations between actors and issues such as conducted by semantic network analysis. This article presents a system for automatically determining the polarity (positivity/negativity) of these relations by using techniques from sentiment analysis. We used a machine learning model trained on the manually annotated news coverage of the Dutch 2006 elections, collecting lexical, syntactic, and word-similarity based features, and using the syntactic analysis to focus on the relevant part of the sentence. The performance of the full system is significantly better than the baseline with an F1 score of .63. Additionally, we replicate four studies from an earlier analysis of these elections, attaining correlations of greater than .8 in three out of four cases. This shows that the presented system can be immediately used for a number of analyses.