Natural Language Processing for Improving Transparency in Representative Democracy

Abstract

Language is the natural medium in politics, hence among several types of data, text is the central artifact to capture political behaviour. In this thesis we focus on automating several political analyses using natural language processing techniques, which can improve the transparency of policy-making and thereby the voters' trust in representative democracy. Political scientists have observed that the voters' trust in government is necessary for successful implementation of policies, and in-turn their trust is based on effective implementation of policies and services. In-order to improve the trust, the policy-making process should be more transparent and receptive. The policy-making process typically consists of several stages, and we focus on the two primary stages involving political parties and voters --- policy proposal and its implementation audit.

Specifically, we target three major aspects of policy-making process: (a) analyzing policy proposal during election campaign --- what policy goals are spoken about, and specifically in which context, and what promises are made. (b) Post-election policy implementation audit --- given the pre-election promises, which sets the expectation of voters, does the government make progress towards those promises. (c) Public advocacy for policy changes --- what changes do the voters want. This can be seen as both evaluation of existing policies as well as suggestions for changes. More importantly, active participation of voters in the process reflects their level of trust in the system. We define the individual research targets based on political science literature and automate those using deep learning approaches. We use canonical sources of text for each of the tasks, for example, election manifestos released by political parties (more sources are discussed in Chapter 2).

The challenges involved in this work are multi-fold, starting from defining the task, to dataset creation, to developing suitable models. We hope that this PhD thesis, dealing with political text analysis, will shed light on the available data sources, flavor of tasks at the intersection of both natural language processing and political science, and also the techniques to handle the challenges.