Parliamentary Debates in Norway A Computational Social Science Approach Martin Søyland Thesis submitted for the degree of Philosophiæ Doctor © Martin Søyland, 2020 Series of dissertations submitted to the Faculty of Social Sciences, University of Oslo No. 822 ISSN 1564-3991 All rights reserved. No part of this publication may be reproduced or transmitted, in any form or by any means, without permission. Cover: Hanne Baadsgaard Utigard. Print production: Reprosentralen, University of Oslo. Summary Legislative debates are an understudied institution in the political science literature. Modern tools for automatic analysis of speech content has, fortu- nately, given us an opportunity to make inferences based on an ever increas- ing accessibility of vast corpora of texts that has been analyzed only in parts previously. In this thesis, I study Norwegian parliamentary debates with the overar- ching hypothesis that parliamentary speech can be used to assess the effects of institutional and external shocks on MP behavior. In order to test this ar- gument, we need the data to do so, the methods suited for doing so, and the contextual knowledge of how to interpret the results. The thesis has three main contributions. First, it provides a new data set on Norwegian parliamentary debates (1998-2016); a large corpus of automatically annotated speeches accompanied with a wide set of meta data. These are both the first openly accessible structured data on parliamentary debates in Norway and, to my knowledge, the first open access linguistically annotated parliamentary speech data in the world. Second, the thesis has a more general contribution in that I show how data structuring and contextual knowledge is an integral part of the text analysis process. On the one hand, I provide an analysis on which MPs get to take the plenary floor in the Norwegian parliament. This builds an important foundation for understanding the content of the speeches in parliament by, for example, showing that committee membership is essential for MP’s floor access. On the other hand, I show that the language features fed to our text models are important for subsequent inference. Even small scale language tweaks are shown to have strong impact on possible inferences made from these analyses. Finally, I provide two examples of inferential analyses on MP behavior based on parliamentary debates. On the one hand, research on the effect of electoral reform on different political institutions are numerous, but this the- sis provides the first analysis on the effect of electoral reform on the content of parliamentary debates. This confirms the theory stating that going from SMD systems to PR systems re-alters the vote-seeking incentives of MPs; iii going from more personal to more party based. On the other hand, even though the parties are very unified in the Norwegian parliament, I also show that disproportional external shocks across electoral districts can alter the behavior of MPs based on their constituency, and not necessarily based on their party affiliation. In sum, the thesis makes three distinct contributions to the literature on parliamentary debates: 1) new and innovative data on Norway, 2) in depth analysis of pre-processing consequences, and 3) how institutional design and external shocks affect the content of parliamentary debates. iv Acknowledgements This thesis concludes my ten year long stay with the Department of political science at the University of Oslo. I am grateful for all my fellow students for the first half and my colleagues for the later half of that period. Coming from a non-academic line of work, I had a weak knowledge foun- dation compared to my fellow students when I first started studying political science. I want to thank Einar, Magnus, Lars, Ole, Marie, and Malin espe- cially for helping me patch the holes in my foundation, laughing at me when I’m unreasonable, and being the invaluable friends you are. I also want to thank Haakon and Peter, the next generation of Norwegian political science stars, for teaching me how to be academic and for inspiring me with their great work ethic and interesting research. With a unique drive to help others, being able to see everyone she meets as interesting individuals, and taking all challenges put in front of her by the horns, I want to thank Ingebjørg. You are a true inspiration for everyone fortunate enough to have you in their lives. Of all the leaders I have encountered, none come close to Bjørn Høyland. From inspiring me to shift my attention to Norwegian politics during our first meeting to supervising my Master’s thesis and PhD thesis, Bjørn has made my career in academia. I really appreciate his traits of being able to trust and listen to a junior researcher, giving clear and useful feedback, while at the same time making sure to always taking my complaints seriously. This thesis would not exist without Bjørn. I was also fortunate enough, as a political scientist interested in infor- matics, to find a partner in crime that is an informatics PhD with interest in political science. Emanuele’s influence is woven into every sentence, word, part of speech and morpheme of this thesis. I also want to thank Daniel M. Smith for kindly inviting me to stay at his department for the last half of 2016 – an experience I would not trade for anything; Zoltán Fazekas for showing me how teaching should be done and giving me invaluable feedback on all parts of this thesis; and Jon H. Fiva for taking interest and giving feedback on my work. Outside of academia, I want to thank my little sister, Guro, for being able to knock sense into me when I lose touch with reality and for making the v coolest kids in the world. I am also in debt to my mom and dad for always letting me do what I want and supporting some of the toughest decisions I have made. I want to thank Joar, Joakim, and <Cruel> for taking my mind off things during my off-hours. Having close friends that have no clue as to what my work is has been essential for me staying sane these last couple of years. Finally, thanks to my greatest inspiration in life. Sandra, you are every- thing. vi Contents 1 Introduction 1 1.1 The case of Norway . .4 1.1.1 Political system . .4 1.2 Theoretical framework . .7 1.2.1 Political institutions . .7 1.2.2 Parliamentary debates . .8 1.2.3 Defining parliamentary debates in Norway . .9 1.3 Text analysis in political science . 11 1.4 Methodological approach . 13 1.4.1 Data structuring . 14 1.4.2 Language pre-processing . 18 1.4.3 Modeling . 22 1.5 Thesis structure . 24 1.6 Implications . 26 1.7 Further studies . 27 I Cotext, data, and language 29 2 Parliamentary Debates in the Norwegian Parliament 31 2.1 Introduction . 32 2.2 Institutional and party system background . 33 2.3 The institutional setting of legislative debate . 36 2.4 What is the role of intra- and inter-party politics in legislative debates . 39 2.4.1 Descriptive analysis . 39 2.4.2 Multivariate analysis . 41 2.5 Debate Participation . 47 2.6 Conclusions . 51 vii 3 Talk of Norway: An Open Resource for the Computational Social Sciences 53 3.1 Introduction . 54 3.2 Related Work . 54 3.2.1 Datasets . 55 3.3 The Talk of Norway Dataset . 56 3.3.1 Data format . 59 3.4 Preliminary experiments . 59 3.5 Discussion . 61 3.6 Future Work . 65 3.7 Conclusion . 66 4 Multi-party Classification of Parliamentary Debates 69 4.1 Introduction . 70 4.2 Classifying parties . 71 4.3 Data and methods . 76 4.4 Classifier performance . 82 4.5 Intra-party cohesion . 84 4.6 Inter-party relations . 87 4.7 Classification as a measure in practice . 91 4.8 Discussion . 92 II Effects on parliamentary debates 95 5 Electoral Reform and Parliamentary Debates 97 5.1 Introduction . 98 5.2 Electoral Rules, Reforms, and Legislative Behavior . 99 5.3 Electoral reform in Norway . 101 5.4 Methods and Data . 103 5.5 Results . 106 5.5.1 Reform change . 107 5.5.2 Robustness . 110 5.5.3 Placebo reforms . 111 5.6 Conclusion . 113 viii 6 Climate Politics in Hard Times: How Local Economic Shocks Influence MPs Attention to Climate Change 117 6.1 Introduction . 118 6.2 Climate politics in hard times . 118 6.3 Research design . 121 6.4 Results . 122 6.5 Conclusion . 125 III Appendices 129 ix List of Tables 1.1 Norwegian electoral districts and number of seats in the 2013 election. .5 1.2 Cabinet attributes (1998-2016) . .6 1.3 Examples of other data sets. 17 2.1 Descriptive statistics on # speeches and # words . 40 2.2 Regression Models: # Speeches (Negative Binominal) and # Words / # Speeches (OLS) . 46 2.3 Regression results: Participant in debate (Logit) . 50 3.1 Speaker classes in the ToN data . 58 3.2 Party-wise classification results for the best performing classifier. 60 4.1 Descriptive statistics for selected variables in the ToN dataset. 76 4.2 List of feature sets accompanied by the macro F1 score and accuracy. 83 5.1 Descriptive comparison between subset of MPs with seat be- fore and after reform and full data. 105 5.2 Number of speeches for parties in different sub-samples of the data. 106 5.3 Topics focused on in the analysis, with expected effect direc- tion and short description. 106 6.1 Estimated topic proportions before and after the oil price shock (2014). 126 A4.1 List of feature sets accompanied by the macro F1 score and accuracy for random forest estimation.