Sentiment Analysis on Current Government using Tweets

Abstract

In the age of the internet, social media connects us all at the tip of our fingers. People are linked through different social media. The social network, Twitter, allows people to tweet their thoughts on any particular event or a specific political body, which provides us with a diverse range of political insights. This paper serves the purpose of (a) Natural Language Processing (NLP) of a multilingual dataset (Urdu, English, and Roman Urdu), (b) exploring machine learning solutions for sentiment analysis and training models, (c) collecting data on government from Twitter and applying sentiment analysis, and (d) providing a python library that classifies input texts as positive or negative. The training data contained tweets in three languages: English: 200000, Urdu: 200000, and Roman Urdu: 11099. Five different classification models are applied to determine sentiments, and eventually, the use of ensemble technique to move forward with the acquired results is explored. The Logistic Regression model performed best with an accuracy of 75%, followed by the Linear Support Vector classifier and the Stochastic Gradient Descent models, both having 74% accuracy. Lastly, the Multinomial Naïve Bayes and Complement Naïve Bayes models both had 73% accuracy.