Combining Lexicon-based and Deep Learning-based methods for automated emotion analysis of newspaper articles in Dutch

Abstract

The increasing use of emotional opinion in media and politics made necessary the development of tools able to monitor these phenomena, i.e. identify in a newspaper article when an emotion is being expressed and who is the target and the expresser of such emotion. In this thesis project a number of techniques were used to develop a system able to derive such information starting from a rule-based system that was able to classify a large set of unlabeled sentences, and then use these to train a deep network. In this study is shown that the deep network is not only able to learn the behavior of the rule-based classifier, it is also able to generalize over the original lexicon and increase the recall of the classifier.