Speech as data for the politics and partisanship of legislators

Thesis

How can we empirically prove complex theories about legislative behaviour? Parliamentary speeches provide large quantities of data about legislators’ behaviour, daily. Until recent developments in natural language processing (NLP), however, researchers could not make full use of the richness of this data. Introducing new methods, I automatically code millions of statements made in parliaments into data about legislators’ partisan loyalties and political positions. I use these methods to yield new evidence about old questions in the study of legislatures, which so far have had long-standing theoretical debate and mixed empirical evidence. These include the effects of electoral system (change), of constituency-level opinion and electoral threat, and of legislators’ ethnicity and gender. Firstly, I use NLP methods to identify the similarity of legislators’ speeches to their party’s relevant manifesto chapters based on the topic they are discussing. This similarity is used to measure strength of partisanship, and determine the effect of electoral system change in New Zealand on party discipline. The results show greater partisanship after introducing closed party lists. Secondly, I use a large language model (LLM) as an ‘expert coder’ to label over 2 million speeches in the UK parliament for their left-right political leaning on two axes. I compare these positions to estimates of constituency opinion and voting intention polls. I show that MPs’ positions are not significantly correlated with those of their constituents, but where UKIP, the Brexit Party, or Reform UK gained support in their constituency, MPs took more right-wing positions, especially on non-economic issues. Finally, I add to this dataset LLM-coded topics identifying relevance to ethnic minorities’ and women’s interests, and measure positions on these topics separately. I show that ethnic minority and female MPs’ speeches are generally more left-wing (especially on ethnic and gender topics), except for ethnic minority MPs in the Conservative Party.

Authors: Blayney, M

• DOI: 10.5287/ora-44xz9eeb2
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: electoral systems; large language models; legislative behaviour; New Zealand; LLM; descriptive representation; natural language processing; NLP; United Kingdom; legislatures; UK politics; parliaments
• Subjects: Legislative Behaviour; Politics and government