This paper examines the impact of investor sentiment on stock market volatility using a natural language processing classification method applied to a large-scale dataset of social network data. We also apply numerous forecasting techniques not only including conventional linear models, but also different machine learning models and compare its results. Among various economic and sentiment features, we employ the least absolute shrinkage and selection operator (Lasso) for linear models and a tree-based nonlinear variable selection method to demonstrate the critical role of sentiment measures in market volatility. The results show that sentiment variables are identified to be one of the most important variables in relationship with stock market volatility and improve the future prediction of volatility when considered.