Grounding and Repair Joe Tepperman CS 599 – Dialogue Modeling Fall 2005
Grounding Establishing mutual belief Collaborative –More than one active participant Acknowledgement Necessary for: –Dialogue flow, theorem proving, etc. –User modeling –Repairing dialogue & ASR errors
Clark and Schaefer’s Contribution Model (1989) Influential, but not practical Contributions in two parts: Presentation Phase 1.Contributor presents content, Partners try to understand it Acceptance Phase 2. Contributor & Partners move towards a grounding criterion: mutual belief that the contributor was understood sufficiently
Presentation Phase –A assumes that B has understood u if B demonstrates some minimum evidence e or stronger Acceptance Phase –B assumes A will believe he has understood u if A registers that B has demonstrated evidence e’ Assumptions Requires acceptance of acceptance?
Display: B repeats A’s presentation verbatim Demonstration: B demonstrates what he has understood Acknowledgement: B makes some sign that he has understood Initiate Next Contribution: B makes a relevant contribution Continued Attention: B shows he is satisfied with A’s presentation Types of Evidence strongest weakest Strongest? Oblivious?
Main Problem with the Model How to tell the current state for each utterance: presentation or acceptance phase? A: Move the boxcar to Corning. A: And load it with oranges. B: Okay. A: Move the boxcar to Corning. B: Okay. A: And load it with oranges. B: Okay. Where does the presentation end?
The Grounding Acts Model (Traum 1992) Collapses all different types of acceptance Single-utterance level grounding units Allows automatic recognition of a within- utterance grounding act –No need to wait for the next phase to start before identifying completion of current one
Grounding Acts Initiate: Begin new content Continue: Add related content Acknowledge: Demonstrate or claim understanding Repair: Correct a perceived misunderstanding Request Repair Request Acknowledgment Cancel: Leave unit ungrounded Includes all C&S “evidence”
State Transition Matrix I: initiator R: responder S: start F: grounded D: “dead” state 1: ack needed for grounding 2: repair I needed 3: ack I needed 4: repair R needed
Previous Example DU1 1: initiate I 1 2: continue I 1 3: acknowledge R F DU1DU2 1: initiate I 1 2: acknowledge R F 3: initiate I F1 4: acknowledge R FF DU: Discourse Unit A: Move the boxcar to Corning. A: And load it with oranges. B: Okay. A: Move the boxcar to Corning. B: Okay. A: And load it with oranges. B: Okay.
Open Problems with this Model Binary grounded/ungrounded decision –No levels of “groundedness” Leaves the unit size unspecified Confusability of grounding acts –e.g. repetition = acknowledgment, repair, or request for repair? Only well-suited for spoken language grounding
A More Complete Psychological Model How is a particular grounding act realized? How important is the grounding? –How useful will it be to the system? What criteria are needed? How well will a particular act ground its intended content? And what is the opportunity cost of performing this act? –Is it worth it?
Levels of Analysis: Quartet, Paek & Horvitz 2000 Channel Level: attempt to open communication channel with some behavior Signal Level: behavior is intended as a signal Intention Level: understanding of semantic content occurs Conversation Level: a joint activity is proposed and responded to lowest highest *All levels require coordination between speaker and listener
System Design Two modules: –maintenance –intention Conversation Control –exchanges info between the modules –determines grounding state –weighs costs and benefits –evaluates module performance & reliability Signal & Channel level Intention level Conversation level
Benefits of this Design ASR can model probabilistic dependencies among levels Easier to pinpoint and fix problems in system understanding Models psychological strategies for grounding on lower levels first Flexibility in multiple domains: simply changing the intention module
Grounding Strategies
Signal Failure
Detecting Miscommunication: Dybkjaer et. al. 1996
GP6: Avoid obscurity of expression
Detecting & Verifying ASR Errors: Krahmer et. al. 2001
Utterance Features System –Implicit/Explicit question –Number of verified slots –Default assumptions: true? –Number, type, and recurrence of errors User –Length (in words) –Answer to verification question? –Ordinary word order? –Confirmation/Disconfirmation markers –Number of repeated, new, and corrected slots When do you want to travel to Amsterdam? So you want to travel to Amsterdam? Date, time, destination, etc. e.g. travel today Human-labeled I want to go to Amsterdam Where I want to go is Amsterdam Yes, no, yeah, nope, etc.
Nonverbal Grounding: Nakano et. al Speaker/Listener gP: gaze at partner gM: gaze at map gMwN: gaze at map & nod UU: utterance unit (intonational)
Grounding Model for MACK