PARADIME Parallel Agentbased Dialogue Management Engine

1
PARADIMEParallel Agent-based Dialogue Management
Engine

• Harry Bunt Simon Keizer
• (Tilburg University, Dept. of Language and
  Information Science)

2
Overview

• Multidimensionality and multifunctionality in
  dialogue
• Dynamic Interpretation Theory
• Multidimensional Dialogue Management in IMIX
• The ISA metaphor
• Multi-Agent Dialogue Manager
• Architecture of Paradime Act Generator
• Information State Demo
• Summary
• Future Plans

2
3
Multidimensionality in dialogue

• In natural dialogue, speakers use their
  utterances to address several aspects
  (dimensions) of the communication.
• Example
• A Wat is RSI?
• B RSI staat voor Repetitive Strain Injury.
• A Ja maar wat is het?
• Utterance 3 is multifunctional
• - ja positive auto-feedback
• - maar negative allo-feedback
• - wat is het task-oriented WH-Question
  turn-giving

3
4
Dialogue Management

• Central task of a Dialogue Manager module to
  decide how to continue the dialogue.
• Common approach DM decides which dialogue act to
  perform (when the system has the turn).
• Traditional (one-dimensional) design choice of
  dialogue acts determined by the task, plus error
  handling in case of problems.

4
5
IMIX Question Answering

• One-dimensional treatment of example dialogue
• U Wat is RSI?
• S RSI staat voor Repetitive Strain Injury.
• U Ja maar wat is het?

5
6
IMIX Question Answering

• One-dimensional treatment of example dialogue
• U Wat is RSI?
• S RSI staat voor Repetitive Strain Injury.
• U Ja maar wat is het?
• S het gt RSI

6
7
IMIX Question Answering

• One-dimensional treatment of example dialogue
• U Wat is RSI?
• S RSI staat voor Repetitive Strain Injury.
• U Ja maar wat is het?
• S het gt RSI
• RSI staat voor Repetitive Strain Injury.

7
8
IMIX Question Answering

• One-dimensional treatment of example dialogue
• U Wat is RSI?
• S RSI staat voor Repetitive Strain Injury.
• U Ja maar wat is het?
• S het gt RSI
• RSI staat voor Repetitive Strain
  Injury.
• The multifunctionality of the utterance should be
  taken into account.

8
9
Dynamic Interpretation Theory

• Dialogue utterances are in general
  multifunctional
• Dialogue acts are defined semantically as update
  operations on the information states (or
  contexts) of the participants.
• Dialogue acts consist of two components
• Semantic Content what they are about
• Communicative Function how the semantic content
  should be used to update the information states
• Communicative functions form a multi-dimensional
  taxonomy,
• dimensions representing aspects of communication
  that can be addressed independently
• every utterance having at most one function per
  dimension

9
10
Dialogue Act Taxonomy (1)

• Task/domain
• Dialogue control
• Feedback
• Auto-feedback
• Allo-feedback
• Interaction management
• Contact management
• Time management
• Turn management
• Topic management
• Own communication management
• Partner communication management
• Dialogue structuring
• Social obligations management

10
11
Dialogue Act Taxonomy (2)

• Task/domain
• Dialogue control
• Feedback
• Auto-feedback ( /- perception, interpretation,
  evaluation, execution)
• Allo-feedback
• Feedback-giving ( /- perc, int, eval, exec)
• Feedback-elicitation (perc, int, eval, exec)
• Interaction management
• Social obligations management
• Salutation (init-greeting, react-greeting)
• Self introduction (init-self-introduction,
  react-self-introduction)
• Gratitude (thanking-init, thanking-downplay)
• Apologising (apology-init, apology-downplay)
• Valediction (init-goodbye, react-goodbye)

11
12
Dialogue Act Taxonomy (3)

• General-purpose Communicative Functions can be
  applied in any dimension. Examples
• Information Transfer
• Information seeking YN-Question, WH-Question,
  ...
• Information providing Inform, WH-Answer,
  YN-Answer, ...
• Action Discussion
• Commissives Offer, Accept/Decline-Request, ...
• Directives Request, Accept/Decline-Offer, ...
• General-purpose vs. dimension-specific functions
• What is RSI? WH-Question task/domain
  dimension
• What did you say? NEG-AUTO-FEEDBACK-PERCEPTION
• Did you say five? YN-Question
  auto-feedback-perc dimension

12
13
Context Model

• represents the Information State of the system
• contains all information considered relevant for
• interpreting user utterances, and
• generating system utterances
• Structure (dimensions) in context models
• Linguistic Context dialogue history,
  conversational/topical structure, dialogue
  future
• Semantic Context information about the
  underlying task
• Physical and Perceptual Context availability of
  communicative channels, participants presence
  and attention
• Cognitive Context participants state of
  processing
• Social Context interactive and reactive
  pressures

13
14
Information Search Assistant

• Human-Computer Interaction (HCI) metaphor used to
  present the dialogue system to the user as an
  Information Search Assistant (ISA), i.e., the
  ISA
• is not a (medical) domain expert itself
• helps users formulate answerable questions for
  the QA modules
• deals with problems in finding the required
  information
• deals with problems in the communication process

14
15
The ISA in the IMIX demonstrator

• ISA reflected in the way information is
  presented
• Answers get an isolated status within the system
  utterance
• ISA may reflect some attitude towards the
  answers
• S I have found the following answers ...

15
16
Agent-based Dialogue Management

• Dialogue act generation supporting
  multifunctionality
• Design of dialogue act agents each agent
  associated with a specific dimension
• Each agent monitors the context, and is triggered
  on the basis of either goal conditions or
  interactive or reactive pressures.
• After triggering, the agent tries to satisfy its
  enabling conditions, thereby generating a
  candidate dialogue act.
• Design of additional evaluation agent constructs
  the semantic- pragmatic content of a
  multifunctional system utterance from available
  candidate dialogue acts

16
17
Architecture
17
18
Feature Structure representation of Context Model
LinguisticContext user_utts
lt...gt system_utts lt...gt topic_struct
... conv_state openingbodyclosing candi
date_dialacts lt...gt combined_dialacts
lt...gt SemanticContext user_info_needs
lt..., question ... satisfied - ,
...gt task_progress ...
CognitiveContext own_proc_state
proc_problem percintevalexecnone h
ypotheses lt...gt partner_proc_state
proc_problem percintevalexecnone h
ypotheses lt...gt SocialContext int
eractive_press nonegrtvaledapothk react_pr
ess nonegrtvaledapothk
18
19
Information State Monitor
DEMO... Example dialogue with simulated user
dialogue acts
19
20
Summary

• First version of a DIT-based multi-agent dialogue
  manager
• Rule-based dialogue act recogniser
  (POS/word-patterns)
• Paradime dialogue manager and dialogue act
  recogniser integrated in (DAM) module of IMIX
  demonstrator, supporting
• Multifunctionality in both user and system
  utterances
• Evaluation and selection of QA results
• Signalling various processing problems
• Limited reactive behavior based on social
  conventions

20
21
Future Plans (1)

• Identify mechanisms to support more coherent and
  natural interaction
• Context-awareness
• Giving occasional positive feedback
• Displaying more refined social behaviour
• Modelling mechanisms underlying the generation of
  articulate feedback
• Example
• U1 I would like to know more about the flu
• S2 I found too many answers. Do you want
  information about symptoms of the flu?

21
22
Future Plans (2)

• Corpus Development (in collaboration with Vidiam)
  in support of
• improved dialogue act recognition
• contextually appropriate dialogue act generation
• Implementation of belief update mechanisms in DIT
• Creation, cancelling, adoption and strengthening
  of beliefs in the context model
• Research tool for simulating dialogues and
  automatic update of beliefs in the model
• Integration of implemented belief update
  mechanisms in the Paradime framework

22