Sentiment-Driven Election Prediction using Multilingual Social Media Analytics

Abstract

Elections influence public policy priorities and social governance. Traditional forecasting methods, such as opinion polling, are subject to sampling biases and delayed feedback. Media coverage still drives most political conversations, but social media has taken precedence, with many users now more comfortable arguing from their phones or other electronic devices [10,18]. We present a cloud-scalable Global Election Prediction System (GEPS) that combines multilingual social media sentiment analysis with machine learning to estimate election trends [16,24]. In this GEPS, natural language processing is employed to normalise posts and classify them into three sentiment groups: positive, negative, and neutral. The three groups summarise sentiment and can be aggregated regionally and over time to track political momentum [2,6]. Logistic Regression, Random Forest, and LSTM models were evaluated, with LSTM being the most accurate forecasting model. In particular, LSTM was the most useful, as it can capture language cues based on context [5,15,23]. We broadly conclude that digital public opinion is useful, often providing early indicators of election outcomes and helping analysts, journalists, and campaigns in the decision-making process [19,20].