A Hybrid Theory and Data-driven Approach to Persuasion Detection with Large Language Models