An efficient approach to identifying anti-government sentiment on Twitter during Michigan protests

Trust in the government is an important dimension of happiness according to the World Happiness Report (Skelton, 2022). Recently, social media platforms have been exploited to erode this trust by spreading hate-filled, violent, anti-government sentiment. This trend was amplified during the COVID-19 pandemic to protest the government-imposed, unpopular public health and safety measures to curb the spread of the coronavirus. Detection and demotion of anti-government rhetoric, especially during turbulent times such as the COVID-19 pandemic, can prevent the escalation of such sentiment into social unrest, physical violence, and turmoil. This article presents a classification framework to identify anti-government sentiment on Twitter during politically motivated, anti-lockdown protests that occurred in the capital of Michigan. From the tweets collected and labeled during the pair of protests, a rich set of features was computed from both structured and unstructured data. Employing feature engineering grounded in statistical, importance, and principal components analysis, subsets of these features are selected to train popular machine learning classifiers. The classifiers can efficiently detect tweets that promote an anti-government view with around 85% accuracy. With an F1-score of 0.82, the classifiers balance precision against recall, optimizing between false positives and false negatives. The classifiers thus demonstrate the feasibility of separating anti-government content from social media dialogue in a chaotic, emotionally charged real-life situation, and open opportunities for future research.