AUGMENTED TRANSITION NETWORKS (ATNs) FOR DIALOGUE CONTROL: A LONGITUDINAL STUDY
Curry Guinn
Rob Hubal (RTI)
Outline
Architecture Applications Experiments Results
Virtual Human Architecture
Applications
Question-Answer Kiosks (Tradeshow for Tradeshows, Space Congress, John Deere, ASTD, APHA)
Training Applications» Bank Teller Training» Field Interviewer Trainer » Phone Interviewer Trainer» Informed Consent Trainer» Law Enforcement Training for Encounters with the Mentally Ill» Pediatrics Trainer
Profiler» Youth risk-assessment profiler» Prison inmate risk assessment profiler
Applications
(a) John Deere Kiosk(b) Bank Teller
Trainer(c) NIDA Assessment Tool
(d) JUSTTALK Officer Training
(e) Pediatrics Trainer(f) WTC – Informed
Consent
Augmented Transition Network Data Structure
Augmented Transition Network is a Finite State Machine with Variables» Nodes» Transitions» Variables
A B
C
If x < 2
If x >=2
Structure of a NodeName Unique identifier
Type Grammars to load
Grammar Response to say
Response Response to say
VR Command Message to send to VR world
Action Actions to perform (variable settings
Transition Name of node to go to if following conditional is true
Transition Action Actions to perform if transition is taken
Transition Condition Boolean expression
Augmented Transition Network Data Structure
wait_on_input /* Name */Normal /* Type */Grammar: "intro.gram default.gram"Response: "inform(offer_assistance)"VRString: "raise(eyebrows)"Action: "MENTOR = 0.8"Transition: proc_commandTransitionAction: "INPUT = command"Conditional: "command(CONTENT)"Transition: proc_queryTransitionAction: "INPUT = query"Conditional: "query(CONTENT)"
ATN Statistics In Order of Creation of Each System
Application SemanticsNetworkNodes
Normal(Speech)
Nodes
Average Transitions
Per Node
State Variables
Space Congress (Kiosk) 38 107 2 4.401869 35
Bank Teller 56 522 29 4.32567 98
ASTD (Kiosk) 51 216 3 4.430556 36
Door-to-door Survey 71 446 39 3.091928 119
Deere (Kiosk) 34 171 1 4.140351 38
Virtual Asthma Patient 52 355 20 2.309859 100
Telephone Survey 23 109 2 4.229358 58
NIDA StolenGoods 11 39 2 1.948718 56
NIDA Smoking 15 41 2 1.878049 56
JUSTTALK 81 284 2 4.169014 120
Pediatrics Teen 97 586 2 2.1843 120
WTC Informed Consent 66 227 2 4.911894 69
Semantic Complexity vs. Network Complexity
Semantics vs. Nodes
0
100
200
300
400
500
600
700
0 20 40 60 80 100 120
# of Semantics
Node
s
Variables Used
The linear increase is encouraging» Why? » You might expect that as
the number of semantics increase the network expands exponentially
» What’s holding this number down?
The use of variables
Application SemanticsState
Variables
Space Congress (Kiosk) 38 35
Bank Teller 56 98
ASTD (Kiosk) 51 36
Door-to-door Survey 71 119
Deere (Kiosk) 34 38
Virtual Asthma Patient 52 100
Telephone Survey 23 58
NIDA StolenGoods 11 56
NIDA Smoking 15 56
JUSTTALK 81 120
Pediatrics Teen 97 120
WTC Informed Consent 66 69
Decrease in the Number of Normal Nodes
The replication of normal (“speech”) nodes resulted in more repetitive networks
Used variables to store which grammars needed to be loaded next
Greatly decreases the complexity of the network
05
1015202530354045
The number of transitions as a measure of complexity
Over time, we tended to move towards reducing the inter-node connections
Why?» Easier to trace and
debug» Easier to maintain
This at the expense of large number of nodes
Application SemanticsNetworkNodes
Average Transitions Per Node
Space Congress (Kiosk)
38 107 4.401869
Bank Teller 56 522 4.32567
ASTD (Kiosk) 51 216 4.430556
Door-to-door Survey 71 446 3.091928
Deere (Kiosk) 34 171 4.140351
Virtual Asthma Patient
52 355 2.309859
Telephone Survey 23 109 4.229358
NIDA StolenGoods 11 39 1.948718
NIDA Smoking 15 41 1.878049
JUSTTALK 81 284 4.169014
Pediatrics Teen 97 586 2.1843
WTC Informed Consent
66 227 4.911894
Future Work
How do these empirical measurements correspond to “true” costs?
– Time to development– Reliability– Maintenance