Understanding the political opinion of Sri Lankans through deep learning based social media data analysis

Abstract

With the rising popularity of social media usage, the data generated through it has significantly increased. By focusing critically on these data, certain patterns, traits and opinions can be obtained which can be used for the betterment of the society. This research focuses on finding out the possibility of understanding the political opinion of the country based on these social media data. Understanding the impact of these data will prompt the public to use these platforms to a greater effect thus guiding the decision-makers to make better and informed decisions. To achieve the above objective English and Sinhala comments from 60 posts from six prominent politicians in Sri Lanka were collected from November 2019 to December 2021. These data was then annotated as positive, negative or neutral sentiments. 6591 annotated comments were used to fine-tune the XLM-RoBERTa (XLM-R) pretrained model for a text classification task. XLM-R is the new state-of-the-art multilingual masked language model which performs exceptionally well in cross-lingual understanding. To improve the performance of the baseline model a novel approach of adding the context of the comment as a feature in the comment is proposed in the thesis. The XLM-R baseline model achieved an F1 score of 79% while the model using politicians' representation in parliament as a context obtained an F1 score of 91%. The models performed exceptionally well for unseen data as well, when tested with data related to politicians not considered in the training data, the model reported an F1 score of 86%. Predicting the sentiments using the best model for the latest posts of the six main politicians in the study, the current opinion of the people was derived. Based on it, the Government representatives President Gotabaya Rajapaksha, Prime Minster Mahinda Rajapaksha, and Former President Maithripala Sirisena obtained negative sentiments while Opposition MPs Sajith Premadasa and Anura Kumara obtained positive sentiments.