Supervised Sentiment Analysis of Parliamentary Speeches and News Reports

Abstract. In this paper, we use several supervised machine learning approaches and compare their success in predicting the sentiment of Austrian parliamentary speeches and news reports (German language). Prediction results in learning- based sentiment analysis vary strongly. They depend on the choice of algorithm and its parameterization, the quality and quantity of available training data as well as the selection of appropriate input feature representations. Our training data contains human-annotated sentiment scores at the phrase and sentence level. Going beyond the dominant bag-of-words modeling approach in traditional natural language processing, we also test sentiment analysis for neural network-based distributed representations of words. The latter reflect syntactic as well as semantic relatedness, but require huge amounts of training examples. We test both approaches with heterogeneous textual data, compare their success rates and provide conclusions on how to improve the sentiment analysis of political communication.