An Automated Planning Approach for Generating Argument Dialogue - PowerPoint PPT Presentation

An Automated Planning Approach for Generating Argument Dialogue Strategies Tanja Daub, Elizabeth Black and Amanda Coles King’s College London tanja.daub@kcl.ac.uk Cardiff Argumentation Forum July 7, 2016 1/24

Background Planning Argument Strategies Future Work Overview 1 Background Persuasion Dialogues Classical Planning Planning a Dialogue Policies 2 Planning Argument Strategies Simple Strategies vs Policies Generating a Policy from Simple Strategies 3 Future Work 2/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Persuasion Dialogues Agents have conflicting views on a topic B D Proponent’s goal: convince opponent to accept the topic A Dialogue terminates when the opponent C E accepts the topic or when neither agent asserts any more arguments 3/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Argument Strategies Existing work AI Planning approach for simple persuasion dialogues [Black et al., 2014] Mixed Observable Markov Decision Processes, assumes probabilistic knowledge of opponent strategy [Hadoux et al., 2015] Minimax algorithm [Rienstra et al., 2013] 4/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Classical Planning A classical planning problem consists of a set of state variables a set of actions defined by preconditions and effects a start state a goal state 5/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Persuasion Dialogues as Planning Problems [Black et al., 2014] a set of state variables �→ different dialogue states a set of actions defined by preconditions and effects �→ asserting arguments a start state �→ initial knowledge of proponent and opponent a goal state �→ topic is acceptable to the opponent 6/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Planning a Dialogue Opponent models M 0 = { B } B D M 1 = { C } M 2 = { B , C } A C E Simple strategy { A , D } , { E } 7/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Policies A policy is a set of state-action-pairs that determines which action should be performed in which state 8/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Simple Strategies vs Policies Opponent models A M 0 = { B 0 } : 0 . 3 M 1 = { B 1 } : 0 . 5 M 2 = { B 2 } : 0 . 2 B 0 B 1 B 2 Simple strategy { A , C 1 } , { C 0 } p = 0 . 8 C 0 C 1 C 2 9/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Simple Strategies vs Policies Opponent models A M 0 = { B 0 } : 0 . 3 M 1 = { B 1 } : 0 . 5 M 2 = { B 2 } : 0 . 2 B 0 B 1 B 2 Policy ( s 0 , a A ) ( s B 0 , a C 0 ) ( s B 1 , a C 1 ) C 0 C 1 C 2 ( s B 2 , a C 2 ) p = 1 10/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Generating a Policy from Simple Strategies A B0 B1 B2 B3 B4 C0 C1 C2 C3 C4 11/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Finding a Simple Strategy Opponent models M 0 = { B 0 , B 2 } : 1 / 3 A M 1 = { B 2 , B 4 } : 1 / 3 M 2 = { B 1 , B 3 } : 1 / 3 B0 B1 B2 B3 B4 Simple strategy π 0 C0 C1 C2 C3 C4 { A , C 0 , C 2 } , { C 4 } p = 2 / 3 12/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Generating a Policy ? Opponent models � M 0 = { B 0 , B 2 } π 0 M 1 = { B 2 , B 4 } B 0 , B 2 , B 4 B 1 , B 3 � M 2 = { B 1 , B 3 } ? π 0 ? ? 13/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Replanning for Failed Cases Opponent models M 0 = { B 0 , B 2 } : 1 / 3 M 1 = { B 2 , B 4 } : 1 / 3 A M 2 = { B 1 , B 3 } : 1 B0 B1 B2 B3 B4 Simple strategy π 1 { A , C 1 , C 3 } C0 C1 C2 C3 C4 p = 1 Merge simple strategies into policy 14/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Merging Simple Strategies into a Policy Opponent models � M 0 = { B 0 , B 2 } π 0 = { A , C 0 , C 2 } , { C 4 } M 1 = { B 2 , B 4 } � A M 2 = { B 1 , B 3 } π 1 = { A , C 1 , C 3 } Policy B 0 , B 2 , B 4 B 1 , B 3 ( s 0 , { a A } ) ( s B 0 , π 0 ) ( s B 1 , π 1 ) π 0 π 1 ( s B 2 , π 0 ) ( s B 3 , π 1 ) ( s B 4 , π 0 ) p = 1 15/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Generating a Policy from Simple Strategies A B 0 B 1 B 2 B 3 C 0 C 1 C 2 C 3 16/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Finding a Simple Strategy Opponent models M 0 = { B 0 , B 2 } : 0 . 25 A M 1 = { B 0 , B 1 } : 0 . 25 M 2 = { B 1 , B 2 } : 0 . 25 B 0 B 1 B 2 B 3 M 3 = { B 1 , B 3 } : 0 . 25 Simple strategy π 0 C 0 C 1 C 2 C 3 { A , C 0 , C 2 } , { C 1 } p = 0 . 75 17/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Generating a Policy Opponent models  ? M 0 = { B 0 , B 2 }   M 1 = { B 0 , B 1 } π 0  M 2 = { B 1 , B 2 }  B 0 , B 2 B 1 B 3 � M 3 = { B 1 , B 3 } ? π 0 ? ? ? ? 18/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Finding more Simple Strategies Opponent models M 0 = { B 0 , B 2 } : 0 . 25 A M 1 = { B 0 , B 1 } : 1 / 3 M 2 = { B 1 , B 2 } : 1 / 3 B 0 B 1 B 2 B 3 M 3 = { B 1 , B 3 } : 1 / 3 Simple strategy π 0 C 0 C 1 C 2 C 3 { A , C 0 , C 2 } , { C 1 } p = 2 / 3 19/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Finding more Simple Strategies Opponent models M 0 = { B 0 , B 2 } : 0 . 25 A M 1 = { B 0 , B 1 } : 1 / 3 M 2 = { B 1 , B 2 } : 1 / 3 B 0 B 1 B 2 B 3 M 3 = { B 1 , B 3 } : 1 Simple strategy π 1 C 0 C 1 C 2 C 3 { A , C 1 , C 3 } p = 1 20/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Merging Simple Strategies into a Policy Opponent models  M 0 = { B 0 , B 2 } A   M 1 = { B 0 , B 1 } π 0 = { A , C 0 , C 2 } , { C 1 }  M 2 = { B 1 , B 2 }  B 0 , B 2 B 1 B 3 � M 3 = { B 1 , B 3 } π 1 = { A , C 1 , C 3 } π 0 π 1 ? B 0 B 2 B 3 21/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Merging Simple Strategies into a Policy Opponent models A � M 1 = { B 0 , B 1 } { C 0 , C 2 } � M 2 = { B 1 , B 2 } { C 2 } B 0 , B 2 B 1 B 3 � M 3 = { B 1 , B 3 } { C 1 , C 3 } π 0 π 1 ? B 0 B 2 B 3 22/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies

Background Planning Argument Strategies Future Work Future Work Implement this approach and perform experiments to determine both its scalability and the quality of the policies compared to the optimal How can we identify problems where a policy would perform better than a simple strategy? What is the best simple strategy to start with? How can we deal with more complex dialogue scenarios and opponent strategies? 23/24 Tanja Daub, Elizabeth Black and Amanda Coles KCL An Automated Planning Approach for Generating Argument Dialogue Strategies