Uses a natural language processing approach (convolutional neural networks) to measure content and synactic complexity and predict persuasiveness of messaging. Dataset includes 134 debates (with 129,480 sentences on different topics), attitude polls from the audience before and after debates, and a follow-up experiment.