OPOQ: a Bayesian framework for querying political substance within natural language in support of large scale participatory decision-making

点击翻译按钮获取中文摘要

My research will answer this question via the development of a theoretical and computational framework that closes the gaps between three core aspects of democracy: opinion, policy, and outcomes. The framework created will use political text as data, utilising natural language processing and multiple advanced quantitative methods. It will be comprised of a) a flexible Python data scraping and processing pipeline, b) policy position versus argument classification, c) multi-dimensional opinion polarity analysis e.g. economically left, right / eco-friendly, neutral / socially traditional, liberal, d) granular policy breakdown by type, e) an end to end democratic data schema, f) econometric modelling beyond GDP to wellbeing metrics, and g) a democratic decision support system bringing these elements together. This methodology is explained further below. Overall, my research will take a normative, vocational tack towards supporting the advance of the democratic project: demonstrating how hybrid democratic governance, comprising complementary representative and direct elements, may be supported by an ensemble of advanced quantitative methods and machine learning techniques wrapped into a participatory decision support system. My methodological approach will draw from data science, econometrics and Human Computer Interaction best practice, whereby the methods used will be predominantly quantitative, but also incorporate qualitative field tests with target users. The framework will combine these methods as follows:a) The creation of a flexible data processing pipeline in Python: that scrapes and processes political text from multiple sources, such as opinions in media and social media, party manifestos, parliamentary debates and committee reports, parliamentary votes, and transforms it into 'text as data' for analysis.b) Policy position versus argument classification: via 'parts of speech' (POS) natural language processing (NLP), building upon the work of Lawrence et al. (2017) and Burnap and Williams (2015). This will support the automated separating and keyword tagging of policies and opinion.c) The development of a multi-dimensional polarity analysis algorithm: to automatically code policy positions by axis, for example economic left-right, socially liberal-traditional, environmentally friendly-neutral, with higher semantic precision than achieved by focusing on the left-right cleavage alone.d) Granular automated policy type breakdown: beyond that applied by the Comparative Manifesto Project, and Lowe and Benoit (c.f. Lowe et. al., 2011), to identify and distinguish policy, legislative actions, budgets, goals and performance. e) The development of an end to end democratic data schema connecting opinion, policy and outcomes together with generalisable coding processes.f) Econometric outcome modelling: of economic metrics such as GDP and income equality, and wellbeing metrics such as suffering and life expectancy, as a function of the data features (variables) developed in a) to d). g) The development of a democratic decision support system: comprising the above, that shows relationships between opinions, policy, and outcomes, and uses Bayesian methods to deliver probabilistic predictions. This system will 'machine learn' over time from new data, including from participant decision-makers in field tests.While the objective for this research is generalisable democratic decision support, development will first be carried out around one substantive topic, for example, austerity. This will ensure that tangible advances are made early on, and built upon during the project, as further topics and tuning are iteratively applied.