Situation Comprehension and Emotion Bob Marinier University of
Situation Comprehension and Emotion Bob Marinier University of Michigan June 2005 1
Emotions: Appraisal theory l Evaluate situation along several dimensions Relate situation to current goals l Compare situation to expectations l Determine causality l Determine social acceptability l l Evaluations l map onto emotions Example: Highly relevant, predictable, goalconducive situation Joy 2
Evaluating from scratch is hard l In general, how do we… l l Relate the situation to current goals? Make predictions? Determine causality? In general, how does one understand what is going on? l “Soar interprets the environment by applying comprehension operators to specific aspects of an environment it wants to comprehend…. What gets produced by executing a comprehension operator is a data structure in the current state that is the comprehension, by virtue of its being interpretable by other parts of Soar in doing other tasks. ”* *Newell, UTC, 1990 3
Properties of a general comprehension theory l Incremental l l Happens over time l l Situations unfold as a series of ordered events Immediate comprehension* l l We perceive situations in pieces, so we need to work with those pieces Don’t waste time – fully utilize this moment May want to respond now Early commitment to an interpretation prevents combinatorial blowup of possibilities Ambiguity resolution l l Commit to an interpretation even if not sure which is correct Repair interpretation if new information indicates previous ambiguity was resolved incorrectly *Just & Carpenter 1987 4
Previous work: NL-Soar* l NL-Soar has these properties l l l NL-Soar is organized around models l l l Incremental: Process one word at a time Happens over time: Words come in sequence Immediate comprehension: Commits to a parse structure Ambiguity resolution: “Snips” previous commitments when current parse fails Utterance model Situation model NL-Soar is about understanding language – we want to understand situations *Lewis 1993 5
Schema theory l l Schema = model A schema is knowledge about a concept l l Data structure representing interpretation Long-term knowledge about the relationships between the concept parts l l Default values Constraints Interpretation = instantiated schema Situation schemata are sequences of lowerlevel events which compose higher-level abstract events 6
Schema example: Structure l Water l Throw l l balloon toss Body language (happy/angry) Travel Trajectory (near/far) l Speed (slow/fast) l l Catch Success (true/false) l Causal agent (catcher/thrower) l 7
Schema example: Predictions l Water balloon toss l Throw l l Travel l Body language: (happy/angry) Trajectory: (near/far) Speed: (slow/fast) Catch l l Success: (true/false) Causal agent: (thrower/catcher) If the thrower looks happy, then I expect the balloon’s travel to be near and slow. l If the balloon’s travel is near and slow, then I expect the catch to succeed. l 8
Schema example: Causality l Water balloon toss l Throw l l Travel l Trajectory: (near/far) Speed: (slow/fast) Catch l l l Body language: (happy/angry) Success: (true/false) Causal agent: (thrower/catcher) If the balloon’s travel is near and slow and the catcher fails to catch it, then it’s the catcher’s fault. 9
Comprehension system Event perceived Categorize event Add to event set NO Find schema that fits current event sequence Fill in schema with default values for predicted events Current schema? YES Does event fit? Swap event in for prediction NO Reject current schema 10
Comprehension system l Incremental l Handle one event at a time l Happens l over time Events occur in sequence l Immediate l Early commitment to an interpretation l Ambiguity l comprehension resolution Reinterprets if previous commitments fail 11
Comprehension example l Event Set l Throw l Body language: happy l Current interpretation: Water balloon toss l Throw l l Travel l Body language: happy Trajectory: near Speed: slow Catch l l Success: true Causal agent: thrower 12
Comprehension example l Event Set l Throw l l Body language: happy l Current interpretation: Water balloon toss l Travel l l Trajectory: near Speed: slow Throw l l Travel l Body language: happy Trajectory: near Speed: slow Catch l l Success: true Causal agent: thrower 13
Appraisals and interpretation l Situation interpretation provides us with a framework for generating appraisals l Relate interpretation to goals l l Evaluate the comprehension process l l Example: Do I want the outcome I’m predicting? Example: Are my predictions accurate? Determine causality l Example: What does my schema knowledge say about who’s at fault? 14
Appraisal example l Event Set l Throw l Body language: happy l Current interpretation: Water balloon toss l Throw l l Travel l Body language: happy Trajectory: near Speed: slow Catch l l Success: true Causal agent: thrower 15
Appraisal example l Event Set l Throw l l l Body language: happy Current interpretation: Water balloon toss l Throw Travel l Trajectory: far Speed: fast l Body language: happy Travel Trajectory: near Discrepancy from l Speed: slow l expectation! l Catch l l Success: true Causal agent: thrower 16
Appraisal example l Event Set l Throw l l Body language: happy l l Throw l Trajectory: far Speed: fast Appraisals l Current interpretation: Water balloon toss Travel l Discrepancy from expectation Surprise! l Travel l Update predictions l l Body language: happy Trajectory: far Speed: fast Catch l l Success: false Causal agent: thrower 17
Other appraisals l Causality l is stored in the schema Get for free if understand what’s happening l Relating long-term goals to situation may require a goal representation beyond Soar’s goal stack 18
Nuggets & Coal l Nuggets l Situation comprehension provides a framework for building agent functions l l Appraisal generation Goal generation via coping and predictions l Coal l Learning schemata requires generating appraisals from scratch l l Event grammar? Implementation in very early stages 19
- Slides: 19