Conversation as Action Under Uncertainty Tim Paek Eric Horvitz
Abstract Conversation uncertainties of various kinds Quartet architecture Inference/Decision making under uncertainty 4 levels of analysis Representations Inference procedures Decision strategies Interaction between a user and dialog system
1. Introduction Spoken dialog system robust, unified architecture for sharing key uncertainties Quartet
2. Toward an Architecture for Conversation General conversation To make themselves understood Uncertainty: communication failure Collaborate/establish mutual belief “feedback” = coordination Head nod, “uh huh” Clear up uncertainties, take other measures Conversation = collaborative effort or joint activity
Grounding 상호 이해를 위해 대화의 presentation 또는 acceptance 를 coordinate 하는 과정 Consideration of how key uncertainties depend on each other and influence mutual understanding What decisions to make in light of these uncertainties
Quartet Treat grounding as decision making under uncertainty Key uncertainty Bayesian Network Local expected utility/VOI Identify action maximize mutual understanding Focus on handling uncertainties Enable dialog system to adapt strategies Allow components to share uncertainties
2.1 Four Levels of Analysis Evaluate grouding at 4 levels of analysis 4 Levels Channel level Signal level Intention level: signal 의 semantics 이해 Conversation level All levels need coordination/collaboration Unified architecture Model probabilistic dependencies between levels
2.2 Models, Inference, Decisions atemporal/temporal Bayesian networks compute likelihood of variables of interest Ex) overall grounding status
2.2.1 Modularity Maintenance module handle uncertainties about channel and signal levels Intention module About intention level Conversation control Keep track of the grounding status Where to focus on resolving uncertainties What grounding actions to take in light of their likely costs/benefits
Reasons for distinct modules Probabilistic dependencies abundant with the modules sparse between them Enhancement of the flexibility of Quartet To be applied to variety of task domains Belief, decision, degree of detailed necessary for each level varied significantly depending on the communication medium or modality
2.2.2 Representation for Decision Making Grouding status Diagnoses the overall mutual understanding
2.2.3 Refinement through Value of Information Need to refine actions by sharing uncertainties Utterance design improve understanding Conversation decision making Capitalize on VOI analysis to ask questions, make recommendations, and seek out information
VOI analysis find best evidence in light of the inferred probabilities Compute VOI 모든 observation 에 대해 expected utility of the best decision 을 계 산 VOI recommend a piece of evidence to observe: system 이 recommendation 제공 또는 phrase a question
Exact computation of VOI all possible sequences of observations Greedy VOI: reasonable approximation
3. Runtime Behavior Dialog system ASR NLPWin parser Face-pose tracker
3.1 Providing Services 완전 이해후의 highest expected utility action User 와 requested action 을 연관시키는 것
3.2 Exploiting Multiple Levels Goal 이 not so clear at the intention level Consider uncertainties at lower levels of analysis
3.3 Adapting Strategies over Time Mutual understanding 이 fluctuate adapt strategy over time Do requested service / resort to only three repair strategies ASK REPEAT, CONFIRM, TROUBLESHOOT (use VOI analyses to recommend procedures to improve mutual understanding)
4. Conclusions Approach to continuous spoken dialog centering on an architecture called Quartet