Sentimental analysis on political speeches

Abstract

Politics is an essential part of human society. From the start of human civilization, politics has controlled every human society. Political speeches have had one of the most influential roles in shaping the world. Speeches of the written variety have been etched in history. These sorts of speeches have a great effect on the general people and their actions in the coming few days. With advancing technologies, people from all across the world get to listen to these speeches hence the impact on the listener is increasing on a global scale. We analyzed the performance of different models on our corpus of speeches using sentiment and context analysis and then we compared the results of those models to see the difficulty in analyzing sentiment and context of speeches of country leaders. In our research we have focused on the presidents/prime ministers of the five permanent members of the United Nations Security Council which are France, China, Russia, United Kingdom and United States. Moreover, if left unchecked, a political personnel or party may cause major problems. In many cases there may be a warning sign that the government needs to change their policies and also listen to the people. By classifying the speeches into positive, negative or neutral categories in terms of sentiment and five context categories international, nationalism, development, extremism and others and evaluated the accuracy of our models. By using approaches such as Longformer (RoBERTa based model), TF-IDF with ensemble learning models and LDA topic modeling along with ensemble learning models, we were able to achieve some satisfactory results. We have used a modified Bidirectional Encoder Representations from Transformers (BERT) algorithm which is Longformer and TF-IDF with ensemble learning models for sentiment analysis and an LDA based topic model implemented on ensemble learning models to analyze our speeches for context analysis. We have achieved a 0.67 score on the accuracy of Sentiment and we also achieved a 0.67 accuracy on contexts.