Machine Learning Techniques for Persuasion Detection in Conversation

摘要

We determined that it is possible to automatically detect persuasion in conversations using three traditional machine learning techniques, naive bayes, maximum entropy, and support vector machine. These results are the first of their kind and serve as a baseline for all future work in this field. The three techniques consistently outperformed the baseline F-score, but not at a level that would be useful for real world applications. The corpus of data was comprised of four types of negotiation transcripts, labeled according to a persuasion model developed by James Cialdini. We discovered that the transcripts from the Davidian standoff in Waco, Texas were significantly different from the rest of the corpus. We have included suggestions for future work in the areas of data set improvements, feature set improvements, and additional research. Advancements in this field will contribute to the Global War on Terror by alerting intelligence analysts to enemy persuasion attempts and by enabling U.S. forces to conduct more effective information and psychological operations using local persuasion models.