ADS04 20040615

1
Design and First Tests of a Chatter Hans
Dybkjær SpeechLogic, Prolog Development Center
A/SLaila DybkjærNISLab, University of
Southern Denmark
2
Chatting

• Dialogue type not common in state-of-the-art
• Eliza, chatbots written interaction
• New kinds of application
• edutainment
• chat with character from commercials series
• small-talk while waiting instead of music
• Test-bed for new conversational techniques
• express feelings
• understand feelings
• non-task oriented dialogue
• other new features

How far can we push current technology towards
free conversation?
3
Kurt

• Entertain users through chat (in Danish)
• Limited vocabulary (350 words)
• Phone-based
• Preferences of food, notably fruit and vegetables
• Kurt, e.g. his name, his age, and where he works
• Personality
• childish
• affective
• self-centred
• defensive with an underlying uncertainty
• evasive

Personality designed to hide shortcomings of
understanding level
4
Features for emotion modelling

• Available
• (Phonetic) lexicon
• Grammar
• Recognition scores
• Phrasing
• Dialogue flow
• Available, but not used
• n-best ambiguity
• barge-in event handling
• complex task domain

• Not available (input)
• Glottal stop
• Stress
• Prosody
• Non-linguistic vocal phenomena, e.g. laughter
• Mood (anger, joy, ...)
• Aware sites
• Overlapping speech(back-channelling)
• ...

Platform allows limited emotion modelling features
5
Interaction model
You are stupid
Fool yourself,
Linguistic personality
Compute affect
Generate output
s t a t e
Flow model
Manage dialogue
Standard dialogue model extended with affective
state and handling
6
Linguistic personality

• Lexicon tagged with
• Face value
• Preference
• Embarrassment
• Used for
• Input interpretation

• Face value
• Kurt sensitive to losing face
• Negative face value e.g. corrections and insults
• Positive face value e.g. praise
• Preference
• Words are liked, disliked or neutral
• Embarrassment
• Certain words embarrassing
• All other words neutral

stupid
Fool,
perslity
affect
output
s t a t e
flow
manage
Context-independent assumption
7
Negation

• Changes face value and preference
• Does not affect embarrassment
• Syntactic negation
• you are not stupid
• Semantic negation
• you hate apples
• Implication of negation may depend on question or
  statement
• you hate apples dont you hate apples
• you are not stupid ? arent you stupid
• Though and ? are not fully semantically
  correct, they hold with respect to face value
  and preference

More complex logic negation not useful for spoken
language
8
Affect computation

• Self-confidence
• Recognition scores
• Changed by accept/reject
• Embarrassment
• Means topic change

• Face value
• Complex, simplify
• if any negative input, take minimum
• otherwise take maximum
• Preference
• Positive/negative face value gt knock-on effect
• Not a function of single words
• But
• if any negative input, take minimum
• otherwise take maximum

stupid
Fool,
perslity
output
affect
s t a t e
flow
manage
Simplified but transparent
9
Affective state

• Self-confidence
• Influences
• magnitude of satisfaction changes
• flow

• Satisfaction
• Main personality control
• scale from angry (low) to exalted (high)
• Overflow at both ends
• Initial level is neutral
• Changes computed from
• input preference
• input face value
• self-confidence level

stupid
Fool,
perslity
affect
output
Hangup
Get Angry
s t a t e
flow
Angry
Exalted
Current
manage
Two-parameter model
10
Dialogue management

• Flow model
• Questions
• Answers
• Statements
• Jokes
• Feedback
• implicit, explicit

• Embarrassment
• Joke and change topic
• Satisfaction
• Underflow leads to hangup
• No other flow effect
• Self-confidence

0
low
high
1
medium
Implicit
Explicit
Feedback
None
stupid
Fool,
perslity
At accept
Joke
Joke
None
affect
output
s t a t e
flow
manage
Simple task solving plus some more chat-like
interaction
11
Generate output

• Phrases
• Canned
• Composed of
• Change marker
• Insults and jokes
• Answers and feedback
• Prompts

• Change marker
• Notifies user of systems emotional state
• Function of satisfaction state and satisfaction
  change
• High, high Happy
• Low, low Angry
• High, low Forbearing
• Low, High Distrustful
• Random phrases
• Variation, less rigid

stupid
Fool,
perslity
affect
output
s t a t e
flow
manage
A simple scheme with large variability
12
Example dialogue
13
Data collection

• No controlled experiments
• Dialogues collected from demo-line
• 86 dialogues transcribed from 3 system iterations
• Many dialogues performed by children
• First output voice by 40 years old male
• Second output voice by 14 years old boy

Small but sufficient to give impression
14
Learned from dialogues (1)

• Start
• identity
• age
• location
• knows about
• how are you
• During call
• mostly questions concerning Kurt
• maybe search for common ground
• little volunteered information
• dialogue on the conversation

Dinner party conversation with a twist
15
Learned from dialogues (2)

• Topics asked about by users
• personal (where he works, where he lives,
  childhood, wife, children, health, hair,
  eye-colour, glasses, smokes, ) (parents, )
• adjective descriptions (stupid, clever, handsome,
  )
• likes and dislikes (alcohol, food, football,
  music, work, sex, )
• utterances related to what the system says
  (insults, long input, )

Topics depends on modelled person
16
Next steps

• Extend grammar coverage
• Extend Kurts knowledge about himself
• Provide him with interests
• Let Kurt ask questions about the user
• Experiment with addition of new parameters
  (patience, balance, self-esteem,
  pessimism/optimism)
• Weighting of parameters depends on personality
• New kinds of interaction patterns (hand over
  phone, detection of repeated calls from same
  number)

Extended conversational and emotional coverage
17
Conclusion

• Clearly too small vocabulary and grammar for
  longer interactions
• Entertaining despite all shortcomings
• In particular
• repetition of what was understood
• reactions to insults

Simple but entertaining aspects