Unveiling political power in textual parliamentary debate using machine learning and deep learning techniques

Abstract

A parliamentary debate can be described in general terms as a formal discussion on a particular topic that is strictly controlled by an institutional set of rules and presided over by the Speaker of the House. Debates in national parliaments do not only affect the fundamental aspects of citizens' life, but often a broader area, or even the whole world. As a form of political debate, however, parliamentary speeches are often indirect and present a number of challenges to computational analyses. In this task, we focus on identifying whether the speaker in the parliamentary debate belong to a governing party or a party in opposition. This task is formulated as a binary classification tasks. This report comprises of two parts. In the first part, we have applied Sentiment Analysis on a labelled dataset of Danish parliamentary debate, by using five Machine learning and four Deep learning models, and have done a comparative study of the performance of each of these models. For Danish parliamentary debate dataset, the highest score that we got among the five Machine Learning models is that of Support Vector Machine (accuracy = 74%) and lowest score is of Gaussian Naïve Bayes (accuracy = 53%), and the highest score that we got among the four Deep Learning models is that of Bi-LSTM (accuracy = 67%) and lowest score is of LSTM (accuracy = 39%). Although European parliamentary debate datasets are available on the Internet, there are no such dataset available on Indian parliamentary debate that’s why we have decided to prepare one. In the second part, we have prepared a dataset of Indian parliamentary debate from the unstructured pdf files of Lok Sabha debate, available in the Parliament Digital Library website, keeping all the fields of this dataset similar to that of Danish parliamentary debate dataset and then we have applied Sentiment Analysis on our prepared dataset by using five Machine learning models, and have done a comparative study of the performance of each of these models. But unlike Danish parliamentary debate dataset, here we have to identify whether the speaker in the parliamentary debate belong to a governing party or a party in opposition or neither of them. This task is formulated as a three class classification tasks. For Indian parliamentary debate dataset, the highest score that we got among the five Machine Learning models is that of K-Nearest Neighbour (accuracy = 64%) and lowest score is of Gaussian Naïve Bayes and K-Nearest Neighbour with K=1 (accuracy = 41%).