a preliminary classification of dialogue genres or correlating properties of activities with...
DESCRIPTION
A preliminary classification of dialogue genres or Correlating properties of activities with properties of dialogue systems. Staffan Larsson Dept. of linguistics Göteborg University. Overview. Introduction Previous classifications of dialogue Dimensions of classification - PowerPoint PPT PresentationTRANSCRIPT
A preliminary classification of dialogue genres
orCorrelating properties of activities
with properties of dialogue systems
Staffan LarssonDept. of linguistics
Göteborg University
Overview• Introduction• Previous classifications of dialogue• Dimensions of classification• Possible additional activity dimensions• Using the classification: decision graphs
and libraries• Summary & future work
Introduction• Goals
– A classification of dialogue genres (types, kinds, …), relevant for development of dialogue systems
– Correlating properties of activities with properties of dialogue systems
– Investigate how this classification can be used in the development of dialogue systems and applications
• Background: GoDiS– An issue-based dialogue system implemented using
TrindiKit (Larsson 2002)– This talk is done with GoDiS in mind, but it the ideas
presented are intended to more general
Dahlbäck (1997)• Modality: spoken/written• Kinds of agents: human/computer• Interaction: dialogue/monologue• Context : spatial, temporal• Number & type of tasks
– Simultaneous?• Dialogue-task distance
– Similarity of dialogue structure – task structure• Kinds of shared knowledge exploited
– Perceptual, linguistic, cultural
Discussion: Dahlbäck• Several dimensions, some relevant but
some not– We currently assume spoken human-
computer dialogue– Dialogue-task distance perhaps too abstract– Context, kinds of shared knowledge used,
and number of tasks relevant, but not yet included in our classification
– Type of task similar to our concept of activity
Allen et. al. (2001)
technique used
example task
task complexity
dialogue phenomena handled
finite-state script
long-distance dialing
least complex
user answers questions
frame-based getting train timetable info
user asks questions, simple clarifications by system
sets of contexts
travel booking agent
shifts between predetermined topics
plan-based models
kitchen design consultant
dynamically generated topic structures, collaborative negotiation subdialogues
agent-based models
disaster relief management
most complex
different modalities (e.g. planned world and actual world)
Discussion:Allen et. al.• Relates properties of system to
properties of activity, BUT• Based on technologies, not properties of
activities– Dialogue phenomena don’t necessarily
come in lumps• Focus on information seeking and
collaborative planning; some types of dialogue not included– Tutorial, Explanatory, Instructional…
Desiderata for a classification of dialogue
• Based on multiple independent properties of (dialogue in) different activities
• Relating properties of activity to properties of system, formulated in the Information State approach
• Covering not only information seeking and collaborative planning dialogue
• Information State (IS)– an abstract data structure (record, DRS, set, stack
etc.)– accessed by dialogue system modules using
conditions and operations• Dialogue Moves
– utterance function (ask, answer, request etc.)• Update rules
– Modify IS based on on observed moves – Select moves to be performed
• IS Approach implemented in TrindiKit
Background: Information State Approach
Dialogue classification & IS approach
• We want to relate our classification to components of the IS approach:– IS type– Dialogue moves– Update rules
• In this talk, rather informally– For GoDiS, we have more formal
descriptions
Some initial dimensions of classification
• Inquiry-oriented vs. Action-oriented dialogue
• Type of result: simple/complex• Type of external process:
active/passive• Distribution of decision rights:
shared/disjoint
Inquiry-oriented vs. action-oriented dialogue
• IOD: raising and addressing issues – E.g. database search
• AOD: introduces (non-communicative) actions to performed (requests)– E.g. programming a Video Recorder
Dialoguegenre
Moves/rules Information State components
Inquiry-Oriented (IOD)
askanswer
Question stack
Action-Oriented (AOD)
requestconfirm
Action stack
Result type• Is the primary result of the dialogue a simple or a
complex information object?– Simple: proposition, action– Complex: plan, proof, explanation
• Complex results require update rules and information state components (e.g. a tree) enabling incremental construction
• Example: offline planning– U: Get me coffee– R: How do I do that?– U: First, go to the kitchen.– R: OK. And then?– U: Go to the coffee machine.– …
Proactivity of external process
• Passive: database, simple device (e.g. Video Recorder)
• (Pro)active: device, e.g. robot, burglar alarm– May need to interrupt current dialogue, perhaps
even interrupt user utterances• This dimension correlates with
– the way the system is connected to the device • Is the device interface a resource (passive) or a module
(active)?– System intitiative and turntaking mechanisms
Distribution of decision rights• Disjoint: each question directed to a specific DP ; this
DP decides on the answer and does not need to negotiate
• Shared: some question(s) should be answered jointly; negotiation may be needed
• Dialogue system requirements for negotiation:– Dialogue move: propose– Information state component: a stack of pairs of
• issue under negotiation, and• alternative solutions/answers to this issue
• N.B.: we here refer to collaborative negotiation (non-conflicting goals)– E.g. SunDial furniture selection task
activity IOD/AOD
result type external process
decision rights
database search IOD simple: price etc.complex: itinerary
passive (database)
disjoint
ticket booking AOD+ IOD
simple: flight passive (database)
disjoint
simple device control
AOD+ IOD
simple: actions passive or active
disjoint
instructional (sys instructs usr)
AOD+IOD
simple: actions passive (manual)
disjoint
offline planning, incl. itinerary planning, complex device control
AOD complex: plan(s) passive (planner)
shared
online planning, e.g. TRIPS
AOD+ IOD
complex: plan active (device+planner)
shared
explanation IOD complex: proof or explanation
passive (inference engine)
shared
tutorial IOD/ AOD
complex? passive (planner)
disjoint
narration IOD complex: narrative passive disjoint
Possible additional activity-related factors
• Distribution of information– Symmetric: DPs have same kind of information– Asymmetric: DPs have different kinds of information– Relation to distribution of decision rights?
• Shared or conflicting goals– Conflicting goals may lead to non-collaborative
negotiation, which would require argumentation acts, including rhetorical acts
• Number of simultaneous tasks (one or several)– But probably very few activities with just one task
• …
Comments• What we really are classifying are activities
– Table shows a classification of activities according to features of a dialogue system needed to particitpate in dialogues in these activities
• How specific should our activities, or activity types, be?– Action oriented dialogue? Device control? VCR control?
Dialogue with Panasonic VCR 4500? • Is ”genre” still a useful term?
– Could perhaps be reserved for very basic properties, such as IOD/AOD
– Or have genres like ”AOD for active devices and collaborative negotiation and asymmetric distrubution of information”
How can this classification be used?
• Make decision graphs …– … which based on properties of the activity,
including dialogue properties, …– … leads to dialogue genres, or to desired
properties of system.• Based on output of decision graph,
– select the variant of the system closest to the requirements
– E.g. GoDiS for AOD with passive devices and disjoint decision rights
Sample decision graph (partial, and assumes disjoint decision rights)
Does the dialogue involverequests for actions?
Is the goal of the dialogueto control a device?
Is the goal of the dialogueto retrieve information from
a database?
Is the device active?
AOD-PassiveIOD
Libraries?• Disadvantages of ”system variants” approach
– Large number of system variants– Same code respresented in several system variants
• Ideally,– system properties should correlate with modular
libraries of moves, rules, and IS components;– These libraries can be combined into a system
suitable for dialogue in the activity.• Libraries e.g. for
– AOD, IOD– Simple results, complex results– Negotiation
Independent ”decision graphs” for libraries: examples
• Does the dialogue involve questions and answers?– Yes -> use ”IOD” library
• Does the dialogue involve requests for actions?– Yes -> use ”AOD” library
• Does the dialogue involve an active external process?– Yes -> use ”ActiveDevice” library– No -> use the ”PassiveDevice” library
• Are there issues with shared decision rights?– Yes -> use ”Negotiation” library
Libraries, cont’d• Libraries would also simplify
implementation:– Enables upgrading a library without having
to change anything else– E.g. plug in a new analysis of grounding– Allows reuse of the same rules etc. in
multiple genres • However, it may be difficult to achieve
the required degree of modularity
Summary• By
– relating properties of (dialogue in) activities to properties of dialogue systems,
• we can – determine which variant of a system (or
which combination of libraries) to use for a system in a given activity
• We provided a first attempt at such a classification, – and discussed how it could be used
Future work• Extend the number of dimensions of
classification– More activity-related factors– Add modality-related factors?
• Explore the idea of libraries– May be difficult to implement
• (Extend capabilities of GoDiS– Currently, IOD and AOD for passive devices,
disjoint decision rights, asymmetric distribution of information, shared goals, multiple simultaneous tasks)
?
More thoughts• Rule libraries come with infostate
extensions/requirements, and with additional moves– Requirements not only on structure,
but also on how it’s to be used, e.g. What does the order of a queue mean?
Interactive Communication Management
• The presence of ICM may be independent of activity– … but not the form of ICM– Have different ICM grammars for different kinds of
activity– Which factors determine genre-specific ICM?
• Written/spoken• Noisiness• Available modalities• How important to be right? AOD->higher requirements on
recognition, more checks? • Negotiation (in ”alternatives” sense) not really
directly correlated with shared decision rights
Modality-related properties• Written• Spoken
– Not noisy– Noisy
• Determines choice of feedback mechanisms
• To some extent activity-related
Allwood’s activity-based pragmatics
• Levels of activity/context– Physical: artifacts etc.– Biological– Psychological: beliefs, desires, intentions, …– Social: incl. rights & obligations, communicative and task-
related• How do these fit with the proposed activity-related
factors?– Distribution of decision rights: social– Proactivity of external process: Physical (Biological?
Psychological?)– Result type: Psychological?– Information state components: Psychological and social
Cutouts…
GoDiS: an issue-based dialogue system
• Built using TrindiKit– Toolkit for implementing and experimenting with
dialogue systems based on the information state approach
• Explores and implements Issue-based Dialogue Management (IBDM)
• Extends theory to more flexible dialogue– Multiple tasks, information sharing between tasks – Interactive Communication Management (ICM),
including feedback, and grounding– Question accommodation– Negotiation of alternatives– Menu based action oriented dialogue
input inter-pret
TIS
DATABASE LEXICON DOMAIN
data-base
control
update select gene-rate output
lexicon domainknowledge
DME
TrindiKit
GoDiS
GoDiS-IOD GoDiS-AOD
TravelAgency
Auto-route
Xeroxmanual
VCRmanager
IBDM
homedevice
manager
ISapproach
genre-specific
activity-specific
General dialogue phenomena- may appear in any activity
• We assume grounding & accommodation probably present in all spoken H-H dialogue– However, grounding works very differently in noisy
environments, and of course in written dialogue• We don’t use these factors to distinguish activities
Feature Moves/rules Infostatecomponents
ICM & grounding
ICM moves Temporary storage, grounding issues
Question accommodation
Accommodation rules
-
Added 030531• AOD/IOD- complicated cases
– Web search:IOD/AOD; what is a non-communicative action?
– Offline planning (should be IOD, unless DP requested to carry out the plan)
• Distinguish different kinds of computer DPs– Robots vs. Stationary devices, etc.
• Additional dimension– Pronoun resolution needed? Or can it
be ignored? How determine this y looking at dialogue?
• Turntaking related to– Grounding (modality)– Passive/active device– …?
• Why not use all libraries (maximal variants)?– Because more work adapting to new
domains