Multimedia content service and delivery have long been plagued by the difficulty in obtaining feedback on users' true quality of experience. Existing estimation methods do not adequately cover all relevant factors, whereas questionnaires are costly, time-consuming, and impossible to scale. In this work, we present a framework for estimating a viewer's reactions toward on-screen content in real time by capturing and analyzing his/her facial video, thus allowing up-to-date learning of the viewer's preferences to occur, enabling the content provider to serve the most desirable and relevant contents and advertisements. Experiments have shown that the proposed sentiment analysis method can predict the viewer's preferences with good accuracy.