Farrar on Ontologies for NLP

1
Farrar on Ontologies for NLP

• Fourth w/s on multimodal semantic representation,
  Tilburg 10-11 Jan 2005

2
How I understand the paper

• Activity 4 sometimes titled
• Logico-semantic relations / entities
• Semantic Data Categories
• Farrars paper discusses
• Ontologies for NLP (cf. referential descriptors)
• Defense of Bateman 1992, The theoretical status
  of ontologies in NLP
• Main claim ontologies for NLP need to be shaped
  by linguistic concerns

3
The ontology debate (Hobbs 1985, quoted by
Bateman)

• Semantics is the attempted specification of
  the relation between language and the world. (..)
  There is a spectrum of choices one can make (..).
  At one end of the spectrum (..), one can adopt
  the correct theory of the world, the one given
  by quantum mechanics and the other sciences. (..)
  At the (other) end, one can assume a theory of
  the world that is isomorphic to the way we talk
  about it.

4
The standard response to Hobbs question

• Have your cake and eat it multilevel semantics
  (e.g., various systems at Philips and BBNearly
  theories of underspecification)
• a deep semantics thats as close to the
  scientific facts as you like
• a shallow semantics thats as close to
  linguistic structure as you like
• any finite number of levels in between
• a computable mapping between adjacent levels

5
Farrar Bateman

• Defend the Generalised Upper Model (Bateman et
  al.)
• Appear to use multilevel semantics separation
  between linguistic and nonlinguistic levels of
  knowledge. Conceptual/semantic distinction
• Appear to argue against classic denotational
  (e.g.,Montague-style) semantics
• (e.g., this seemed to be the gist of example
  school?xpurpose(x)learning,where x is
  either institution or building)

6

• To find out how the Upper Model compares with
  denotational multilevel semantics, lets look at
  some examples

7
Example 1 American in TENDUM(Bunt et al.,
Philips/IPO)

• Shallow semantics
• American company ?x(company(x) AM(x))
• American passenger ?x(passenger(x) AM(x))
• American airplane ?x(airplane(x) AM(x))
• Deep semantics replace AM by one of
  ?x(country(headq(x))USA
  company?x(nationality(x)USA
  passenger?x(country(headq(carrier(x)))USA
  airplane?x(country(headq(builder(x)))USA
  airplane

8
The idea behind this trick

• Shallow semantics one shallow constant AM
  used for many different properties, matching
  English usage
• Deep semantics AM replaced by an expression that
  has straightforward denotation in the world

9
Generalised Upper Model (GMU)

• GMU appears to opt for shallow constants
• Often, these appear to cover cases that are
  semantically very different
• It is not always clear how they are linked with
  deep (denotational) expressions(cf., Bunt
  Romary 2004 Do concepts have modeltheoretic
  semantics?)
• This may result in formulas whose meaning we
  dont really understand

10
Example 2 Generalised possession (after Bateman
1992)

• The handle of the door, the doors handle, ..
• part-of relation
• The desk of John, Johns desk, ..
• ownership relation
• The son of Abraham, Abrahams son, ..
• son of relation
• The father of Isaac, Isaacs father, ..
• father of relation

11
Shallow relation constant POS(part of, owned by,
)

• The handle of the door?x(POS(x,door)
  handle(x))
• The desk of John?x(POS(x,John) desk(x)) But
• Abrahams son?x(POS(x,Abraham) son(x)) ???
• Isaacs father?x(POS(x,Isaac) father(x))
  ???

12
The issue illustrated by possession

• Is a shared form sufficient for postulating a
  shared (shallow) representation (e.g., POS in the
  case of generalised possession)?
• This issue can also be illustrated by focussing
  on the definite determiner

13
Example 3 Identifiability

• Generalised Upper Model works with abstract
  notions like definiteness (identifiability)
  The pope lives in Italy He is the son of a rich
  banker A man collapsed. (..) The man died He
  is the best left-footer in Scotland
• All these could be generated from something like
  .. IDENT pope.., ..IDENT son .., etc.

14
Identifiability (ctd.)

• Using one constant IDENT does not tell us what
  the different usages of the definite article have
  in common.
• In NLG, it could put an unreasonable burden on
  previous modules (which decide whether to
  generate IDENT or not)
• Perhaps the right question is not How close to
  NL should an ontology be?, but How do we link
  the different levels of meaning?

15
Example 4 The weather

• The deepest question in this area How deep
  should deep representations be?(Quantum
  mechanics, cf. Hobbs??)
• The wind blows (fiercely), and its snowing
  too
• What we want
• Generate this from numerical weather data
• Interpret it and draw inferences
• Note it ltgt the wind

16
The wind blows (fiercely)

• Suppose shallow rep. blow(w) deep rep.
  speed(w)gt50mph
• Mapping from shallow to deepblow
  ?x(speed(x)gt50mph)
• The wind blows shallow blow(w)deep
  ?x(speed(x)gt50mph)(w) speed(w)gt50mph

17
It is snowing

• Whats a suitable shallow representation?
• it does not refer
• Maybe just an atomic proposition Snow
• Possible mapping from shallow to deep
• Snow ?x(precipitation(x)
  type(x)snow
  quantity(x)gt10mm p/h)

18
Questions

• Are these the kinds of mismatches between NL and
  reality that you see as the main challenges for
  building ontologies that are useful to NLP?
• Does the proposed (classical multilevel
  semantics) approach look reasonable to you?
• How does this approach compare to the
  Generalised Upper Model?