Strategical Argumentative Agent for Human Persuasion Ariel Rosenfeld

Strategical Argumentative Agent for Human Persuasion Ariel Rosenfeld and Sarit Kraus Dept. of Computer Science Bar-Ilan University Israel

Time to Graduate? I have 3 conference papers and 2 journal papers. They are great papers, but its not enough. My Advisor I guess you are right… Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016 Me

In Pervious work… Agent Supports Deliberation �Rosenfeld and Kraus, AAAI 2015 (extended version in ACM Tii. S Journal 2016). No Explicit Goal Past deliberations accumulative data Update Capital punishment? Current “Exchanging Opinions” deliberation Trial by jury? Agent Offer arguments =Obtains information

Assumption �The persudee is not STRATIGICAL. �Selects arguments stochastically. Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016

Persuasion �“…is designed to influence others by modifying their beliefs, values, or attitudes. ” (Simons, 1976) �Plenty of works within AI: No explicit argumentation in persuasion technologies (Hunter, 2015) Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016

Our Methodology Human Argumentative Behavior Argumentation Framework Machine Learning Persudee model Optimization Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016 Persuasion Policy

Argumentation Framework � Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016

Example 0. 5 a 0. 2 0. 9 b 0. 9 0. 5 d 0. 8 0. 2 0. 9 c 0. 5 Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016 0. 2 e 1

Reasoning � Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016

Our Methodology Human Argumentative Behavior Argumentation Framework Machine Learning Opponent model Optimization Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016 Persuasion Policy

Two Domains �Computer Science Master’s Degree �Chocolate cake vs. energy bar Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016

Human Argumentative Behavior �Human Dialogs (~50 chats) �Manually annotated for arguments and relations �Human Questioners (~100 subjects) �Arguments Belief �Interaction Strength Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016

Our Methodology Human Argumentative Behavior Argumentation Framework Machine Learning Opponent model Optimization Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016 Persuasion Policy

Machine Learning �Predicting the next argument to be presented given the current state of the dialog. �Probability distribution P(d) �E. g. , P(<a, b, c>)= 0. 5 d, 0. 2 e, 0. 1 f , … �As designed in previous work. �Rosenfeld and Kraus [AAAI 2015]. Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016

Our Methodology Human Argumentative Behavior Argumentation Framework Machine Learning Opponent model Optimization Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016 Persuasion Policy

Persuasive Dialog Optimization �Each persudee may use a different AF. �The persudee uses an unknown argumentation framework. �How can we know? �The persuader presents arguments that are added to AF. �Arguments serve as signals. �Bottom line- Stop when no significant difference is expected. Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016

POMDP �States – Persudee’s AF �Observation – dialog �Actions - arguments �Transition – heavily depends persuadee’s model �Reward – Only when the dialog terminates. �Discounting factor – avoiding long (useless) dialogs �Approximated using Monte-Carlo simulations. Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016

Simulations… Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016

Experimental Setting �Strategic Persuasion Agent (SPA) �Baseline �Human �Chocolate cake vs. energy bar – Binary practical decision. �Computer Science Master’s Degree – Attitude (1 -5 scale). Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016

Evaluation �Computer Science Master’s Degree �Chocolate cake vs. energy bar Agent Subjects Changed opinion SPA 15 26. 6% (4) Baseline 15 6. 6% (1) Baseline 15 0% (0) Human 56 26. 7% (15) Human 28 10. 7% (3) Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016

Humans and Argumentation – ongoing effort �Argumentation Theory in the Field - An Empirical Study of Fundamental Notions, Arg. NLP, 2014. �Providing Arguments in Discussions Based on the Prediction of Human Argumentative Behavior, AAAI-15. �Providing Arguments in Discussions On the Basis of the Prediction of Human Argumentative Behavior, ACM Tii. S 2016. �Strategical Argumentative Agent for Human Persuasion (Submitted). Take home Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016 message

Thanks �Ariel Rosenfeld: arielros 1@gmail. com Ariel Rosenfeld and Sarit Kraus, Dagsthul, Germany - April 2016