Advanced TuTalk Dialogue Agents

1
Advanced TuTalk Dialogue Agents

• Pamela Jordan
• University of Pittsburgh
• Learning Research and Development Center

2
Agenda

• Introducing the TuTalk dialogue system server
• System architecture (briefly)
• Web interface to server for
• Fielding experiments
• Analyzing experiments
• For uploading sc and xml directly to dialogue
  system server
• Introduce additional authoring features
• Enabling and controlling automated feedback
• Optional steps (sc only)
• Looping (sc only)

3
Architecture of TuTalk Dialogue System

• Hub spokes architecture
• Main modules
• Coordinator
• Language recognition
• Language generation
• Dialogue manager
• Dialogue History database

4
Language recognition

• Refers to labelled sets of alternative phrasings
  called concepts
• Inputs from dialogue manager
• normalized sentence,
• set of expected response concepts
• Computes minimum edit distance, number of adds
  and deletes of words needed to match input
  language to a concept
• Returns expected response concept with smallest
  minimum edit distance that falls within a
  threshold

5
Language generation

• Refers to labelled sets of alternative phrasings
  called concepts
• Input from dialogue manager a concept
• If gt 1 alternative phrasing, removes one last
  used according to dialogue history
• Randomly selects from remaining alternative
  phrasings
• Requests output of selected phrasing to student

6
Integrating TuTalk

• Can integrate (embed/wrap) other modules
• ProPL
• Cordillera
• Hub spokes architecture
• Replaceable modules e.g., NLU

7
Replaceable language modules

• Replace
• Use a human (1 experiment)
• Use another approach
• Supplement Human reviews and corrects choices
  made by recognition module (1 experiment)

8
Experiment management/analysis

• Experiment management tools http//pyrenees.lrdc.
  pitt.edu/tutalk/cgi-bin/admin.cgi
• One scenario/script one dialogue agent one
  condition, but can organize in other ways
• One agent per unit
• One agent that tells all knowledge components vs.
  one that elicits vs. one that uses a strategy to
  decide which to do
• Condition management
• Start on server and leave it running
• Designate who is allowed in the condition
• SQL database of information collected during
  interaction
• Can download or query
• Working on producing DataShop format

9
Setting up a scenario
10
SQL database access
11
SQL database access
12
Requesting a transcript
13
Transcript for a student
14
Additional authoring features

• Automated turn transitions/feedback
• Must specify truth-values for responses in
  context of an initiation
• Default truth-value is unknown
• Override automated turn transitions/feedbacks
  with a say
• Optional steps
• Skip if specified condition met in recent
  dialogue history
• Looping
• Repeat the template for a goal until a condition
  is met
• Globally set condition once set

15
Script w/ truth-val in sc
g select-appetizer say enthuse_about_appetizers
say ask_share_appetizer if
agree_to_share_appetizer if skip_appetizer
true do abort ask-soup else do
abort lose-temper do agree-on-appetizer
16
Automatically generated response feedback

• say feature in authoring tool overrides
  automatic feedback
• say following a response in sc overrides
  automatic feedback
• Can globally enable/disable in xml configuration
  section (default is enabled)

17
Script w/ auto feedback override in sc
g select-appetizer say enthuse_about_appetizers
say ask_share_appetizer if
agree_to_share_appetizer if skip_appetizer
true say Okay, in that case I wont get an
appetizer. do abort ask-soup else
do abort lose-temper do agree-on-appetizer
18
Authoring response feedback (1)
19
Authoring response feedback (2)
20
Additional authoring features

• Automated turn transitions/feedback
• Must specify truth-values for responses in
  context of an initiation
• Default truth-value is unknown
• Override automated turn transitions/feedbacks
  with a say
• Optional steps
• Skip if specified condition met in recent
  dialogue history
• Looping
• Repeat the template for a goal until a condition
  is met
• Globally set condition once set

21
Optional Steps
gravitational,
runners
ltanything elsegt
not said
said
22
Example of optional steps
T what horizontal forces are acting on it
while she is pushing it? S Gravity? T In what
direction does gravity act? . T So are
there any other forces on the ball? S no T What
about the runner? . T Okay. After the push
ends, what forces
subdialogue
T what horizontal forces are acting on it
while she is pushing it? S The runners T
Right! So are there any other forces on the
ball? S no T Good. After the push ends, what
forces
23
Script with an optional step and semantic labels
in sc

• g ask-appetizer
• say enthuse-about-appetizers opt sem
  enthuse-about-appetizers
• say ask-appetizer
• if skip-appetizer sem skip-appetizer-order do
  abort soup
• if no do abort soup
• else do lose-temper
• do order-appetizer

24
Additional authoring features

• Automated turn transitions/feedback
• Must specify truth-values for responses in
  context of an initiation
• Default truth-value is unknown
• Override automated turn transitions/feedbacks
  with a say
• Optional steps
• Skip if specified condition met in recent
  dialogue history
• Looping
• Repeat the template for a goal until a condition
  is met
• Globally set condition once set

25
Looping for a dialogue

• Continuous loop on a template repeat template
  until all indicated components covered
• T Okay, great. What should we work on now?
• S Add a loop
• T Thats right. We will have to add a loop.
  Lets figure out why. What made you think of
  using a loop?
• S .
• .
• T Okay. What should we work on now?
• .
• .
• T So, with that were finished with this
  problem.
• (example based on Lane 04)

26
Script with a loop and semantic labels in sc
g start do 76 do 36 do 163 do 58 g 76
loop finished76 say the-first-problem-will-take-
about do prob76 say the-next-problem-will-take
-about-3 sem finished76 if yes else do
logout-msg
27
Configuration section in sc format
config global version "Revision 1.75 "
default-language en We use several proper
nouns in the scenario, so we declare them here,
and provide morphological info so
en-normalizer.py wont spell-correct them. A
different normalizer will probably require (and
hopefully document) a different format. config
normalizer lexicon-supplement "Pythonen spri
ngfield n-springfield-, chicago n-chicago-,
ontario n-ontario-, illinois n-illinois-,
superior n-superior-, huron n-huron-, erie
n-erie-, michigan n-michigan-, ontario
n-ontario- "