Kein Folientitel

1
Linguistic interaction as ontological mediation
John Bateman
AG5 Ontologisches Wissen und sprachliche
Kodierung 25. Jahrestagung Deutsche Gesellschaft
für Sprachwissenschaft Freitag 28. Februar 2003
2
Structure of Talk

• Design, Motivation and Use of a
  Linguistically-motivated Ontology for Natural
  Language Generation the Penman Upper Model
• Problems and Open Questions the issue of
  variation
• New Questions and Perspectives the roles of
  negotiation and function

3
Development of the Penman Upper Model

• a linguistically motivated ontology for driving
  natural language generation (USC/ISI, 1985-89)
• Similar ontologies adopted in variety of NLG
  systems
• event-based
• induced from grammatical patterns, not lexical
• language-specific unless similar grammatical
  patterns occur in differing languages

4
Relations between ontology and client knowledge
representation
Simplest relationship straight inheritance
Client concepts are subordinated to linguistic
ontology concepts and thereby inherit the
linguistic expressions possible for
them... e.g., a robot movement is subordinated
to the linguistic ontology concept movement
Linguistic Ontology
Client Concepts
5
Relations between ontology and client knowledge
representation
Simplest relationship straight inheritance
methods for linguistic expression
Linguistic Ontology
Client Concepts
inherited methods for linguistic expression
6
Example of use semantic specifications
The robot moves to the red block
(x / nondirected-action lex move actor (r /
object lex robot) destination (p / object
lex block color-property-ascription (c
/ color-quality lex red))))
7
Example of use semantic specifications
The robot moves to the red block
(x / nondirected-action lex move actor (r /
object lex robot) destination (p / object
lex block color-property-ascription (c
/ color-quality lex red))))
Upper model concepts
8
Example of use semantic specifications
The robot moves to the red block
(x / nondirected-action lex move actor (r /
object lex robot) destination (p / object
lex block color-property-ascription (c
/ color-quality lex red))))
Upper model relations
Upper model concepts
9
Relation between upper model and domain models
(1988-1990)
10
Relation between upper model and domain models
(1988-1990)
11
Example of use semantic specifications
The robot moves to the red block
(x / move actor (r / robot) destination (p
/ block color-property-ascription (c /
red))))
12
Example of use semantic specifications
The robot moves to the red block
(x / move actor (r / robot) destination (p
/ block color-property-ascription (c /
red))))
domain concepts subordinated to upper model
concepts
13
Summary of properties of the semantic
specification

• anchored in the concepts and relations of the
  upper model
• event-centered semantic roles only defined in
  terms of the event-types they participate in
• very little lexical information
• grammatical information factored out into
  flexible semantics-grammar mappings
• convenient interface between linguistic
  processing components and non-linguistic
  components

14
Purpose of the upper ontologies
SUMO DOLCE OntoLingua/KIF
UOarea
UOarea
UOarea
supporting more powerful and reusable reasoning
in a domain re-use of abstract organization
UOarea
DM
15
Purpose of the upper ontologies
Upper Model
stating possible alternatives for linguistic
expression in a semantically transparent fashion
Domain Model
16
UOarea
UM
relationship
UOarea
?
UOarea
UOarea
DM
DM
17
A problem of expression
Generating explanations for a digital circuit
design expert system
a component C is faulty if
? I i input (i,C) s.t. ?i value (i)
expected-value (i) ?o output (o,C)
value (o) ? expected-value (o)
Bateman Paris (1989) Phrasing a text in terms
a user can understand. IJCAI89.
18
Alternatives
? I i input (i,C) s.t. ?i value (i)
expected-value (i) ?o output (o,C)
value (o) ? expected-value (o)
A component is faulty if there exists a set I of
members i which are inputs of C such that for all
i it is the case that the value of i equals its
expected value and there exists o an output of C
such that the value of o does not equal its
expected value
19
Alternatives
? I i input (i,C) s.t. ?i value (i)
expected-value (i) ?o output (o,C)
value (o) ? expected-value (o)
A component is faulty if the values of all the
inputs of the component are as expected and the
value of an output is not its expected value.
20
Alternatives
? I i input (i,C) s.t. ?i value (i)
expected-value (i) ?o output (o,C)
value (o) ? expected-value (o)
A component is faulty if its inputs are fine and
its output is wrong.
21
Alternatives
? I i input (i,C) s.t. ?i value (i)
expected-value (i) ?o output (o,C)
value (o) ? expected-value (o)
A component is faulty...
if there exists a set I of members i which are
inputs of C such that for all i it is the case
that the value of i equals its expected value and
there exists o an output of C such that the value
of o does not equal its expected value
if the values of all the inputs of the component
are as expected and the value of an output is not
its expected value.
if its inputs are fine and its output is wrong.
22
Relation between logical form and linguistic
expression
23
Result...

• The different possibilities were not arbitrary,
  but patterned together.
• The differences allowed text representations to
  be generated according to the level of expertise
  of the user.
• Different registers (text styles) were being
  achieved by different mappings between domain
  (the logical form) and the upper model.

24
Further variation...

• Take a street is it
• three dimensional in the street
• two dimensional on the street
• zero dimensional at the street
• ?

( ... spatial-locating (s /
three-d-location ...) ...)
25
Further variation...

• Take a desk is it
• three dimensional in her desk
• two dimensional on her desk
• zero dimensional at her desk
• ?

26
?
on the bridge on the town along the
bridge along the town across the bridge
across the town at the bridge at the
town in the bridge in the town
27
Issues
Just how much variability is there? What are
the criteria for choice?
28
Sources and locations of variability

• register selection
• selection of speaker perspective
• selection of speaker communicative strategy
• selection of speaker goals and purposes for action

29
University of Bielefeld SFB360
30
I Redest du jetzt von hinten oder
vorne? K ltlaut lachtgt Also ich habe jetzt das
Gerät vor mir, ja. Und dieses ltpausegt dieses,
dieser rote Würfel mit der roten Schraube, das
ist jetzt bei mir vorne. Da gucke ich jetzt so
drauf. I Das ist eigentlich hinten. K Ja oder
hinten, ist doch egal. I Okay, ja.
SFB360. Konzeptdynamik. taken from Rieser (1997,
1999)
31
Sanford, Garrod, Anderson the maze game (1982-)
32
Garrod, Sanford, Anderson maze game data

• Path description
• See the bottom left, go along one and up one,
  thats where I am.
• Line description
• Im on the second level, second from the left.
• Co-ordinate description
• Im at E two
• Figural description
• See the rectangle at the bottom left, well Im
  in the middle box on the bottom of it.

33
Locally stable description styles

• the bottom one
• the 3rd
• 2nd column bottom one
• 5th column 3 boxes up
• 4th column 3 up
• the top box
• 5th column along bottom
• 2 up

• the bottom row
• the 4th block
• the 3rd one
• 4 along 3 up
• 2 1, 3 1, 2 3, etc.
• the bottom row 3 along
• top row 5th along
• the 2nd box

34

• X-th column Y boxes up

3rd one 3rd column

• X-th block/box along

3rd one 3rd block
35
Basically, a mental model of a situation
consists of two things (1) some set of
autonomous objects which map onto the situation
and give it an ontology... and (2) a tight knit
set of relations between the objects in the
domain which capture the topology of the
situation... Garrod Sanford (1988)
Discourse models as interfaces between language
and the spatial world Journal of Semantics. 6
36
And...

• This ontology in Garrod and Sanfords sense is
  NEGOTIATED...
• during the interaction
• usually implicitly

37
New Project Achieving interaction between
spatial and linguistic modules in complex robotic
assistance systems
38
Back to basics...
An experimental approach where aspects of the
communicative situation can be systematically
controlled Focus the choices language users
make, their interpretations and the constraining
role of their dialogic interaction
39
Robot types
Bremer Rolland
Krieg-Brückner and group
40
Speech Input
Augmenting the Map
Drive Instructions
The fridge is on the right, and left of the stove
Drive tothe kitchen
Drive in frontof the fridge
41
The fridge is on the right, and left of the stove
route graph
42
Drive in frontof the fridge
On the left!
Is the fridgeon the left oron the right?
43
Expressions in context

• in front of the wardrobe
• in front of the TV
• in front of the fridge
• to the TV / to the fridge
• put the cup on the table
• on the fire
• in the street / on the street / at the
  street

44
Hypotheses...

• many terms come with functional wrappers which
  could be glossed as lets agree to talk about
  this as if it were an X in some respects
• these are much more susceptible to negotiation in
  context than commonly assumed
• in many areas functional control is going to be
  the important choice criterion for linguistic
  usage over and above the spatial-temporal
  modeling

45
relationship
UOarea
classifications by functional control simple
spatial-temporal properties
hard spatial-temporal properties
UOarea
UM
UOarea
UOarea
DM