Author: Berend Driessen, University of Amsterdam, submitted in partial fulfillment for the degree of master of science
The semantic analysis of parliamentary debates is an interesting domain for unveiling political agendas. This paper examines the potential of sentiment information for the analysis of parliamentary corpora. It explores three challenges in this field; the effectiveness of automatic sentiment labelling techniques, the cross-country analysis of parliamentary negativity and the integration of sentiment information in a parliamentary search engine. First, we find that a SVM, MLP and BiLSTM trained on automatic sentiment labels were not able to beat the majority baseline when predicting human labels. However, a BiLSTM can reach accuracy up to 88\% when classifying automatic sentiment labels generated by a domain-specific lexicon. Based on this, we conclude that automatic labelling techniques are not a valid substitute for human labels based on voting behaviour. Second, we find that the parliaments of European countries vary in their general levels of negativity up to 0.42 points and for different queries up to 3.15 points. Additionally, we find that countries vary in the oscillation of their negativity levels. Third, we find indications that a parliamentary search engine which integrates sentiment information is a potentially valuable tool for political analysis by other scientific researchers.
Code - This folder contains all the code and information to run the automatic labelling experiments and install the search engine.
Thesis - Here you can view the final thesis itself.