BERT for discourse analysis
: a pragmatist approach to governmentality

  • Ayan Gupta

Student thesis: Doctoral ThesisDoctor of Philosophy (PhD)

Abstract

This thesis is an investigation of the potential utility of using large language models – specifically BERT – to address the concerns of approaches to discourse analysis within the radical sociological traditions prominent since the 1970s, e.g., neo-Marxist theories of ideology, Critical Discourse Analysis, Foucauldian discourse analysis. Though using natural language processing (NLP) techniques within sociology is now a widespread practice, work on integrating these computational approaches within discourse analysis is sparse – thus there is no work which uses NLP methods within a Foucauldian framework. In addition, though the importance of large language models as the current state of the art
within NLP is well recognised, sociological work which uses large language models is also sparse. These two gaps are the focus of this thesis. To address these gaps, I first develop a reading of Foucault’s work on governmentality that incorporates pragmatist views on language. This reading has two purposes. First, it is used to highlight and address some of the blind spots of current approaches to discourse analysis which prevent such approaches from addressing questions about the large-scale behaviour of discourses, e.g. the question of how exactly discourses spread across and within organisations. Second, it is used to provide the theoretical tools needed to interpret the results of using
BERT to conduct text analysis. Existing Foucauldian approaches were not developed with the use of NLP techniques in mind, meaning they do not come already equipped with the tools needed to interpret NLP results. I demonstrate the utility of my pragmatist-governmentality framework for text analysis via BERT by using it in an empirical examination of the British state’s use of crisis neoliberal discourses about ‘resilience’, ‘sustainability’ and ‘wellbeing’. Using a dataset of 92 million tokens sampled from approximately 170000 documents of legislation and 12 central government departments produced between 2000 – 2020, I use BERT as part of an analysis of the British state’s vocabulary and the state’s use of ideological vocabulary. I use the patterns in these words’ uses revealed through BERT to discuss the sociolinguistic mechanics through which crisis neoliberal discourses spread across and within the organisations of the British state, and to discuss the pragmatic logic of the linguistic agency which mediates the spread of crisis neoliberal discourses. I argue that an important part of understanding the discursive aspect of governmental power is understanding the sociolinguistic mechanics which emerge from the pragmatic logic of the linguistic agency underlying the ‘microphysical’/‘capillary’ practices through which governmental power is exerted.
Date of Award21 Mar 2023
Original languageEnglish
Awarding Institution
  • University of Bristol
SupervisorThomas S D Osborne (Supervisor), Leonidas Tsilipakos (Supervisor), Gregor McLennan (Supervisor) & Lee K R Marshall (Supervisor)

Cite this

'