Kees van Deemter

1
Formal IssuesinNatural Language Generation

• Lecture 4
• Shieber 1993 van Deemter 2002

2
Semantics

• Formal semantics concentrates on information
  content and its representation.
• To what extent does good NLG depend on the right
  information? To what extent does good NLG depend
  on the right representation?
• Note GRE, but also more general.

3
Information in NLG
Logical space all the ways things could turn out
to be
4
Information in NLG
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the cake (C).
John ate BC.
John ate the banana (B).
John ate A, BC.
Logical space all the ways things could turn out
to be
5
A proposition - information
Identifies particular cases as real possibilities
6
For example
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the cake (C).
John ate BC.
John ate the banana (B).
John ate A, BC.
Here is a particular proposition.
7
A wrinkle
Computer systems get their knowledge of logical
space, common ground, etc. from statements in
formal logic.
Lots of formulas can carry the same information.
8
For example
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the cake (C).
John ate BC.
John ate the banana (B).
John ate A, BC.
ABC ? ABC ? ABC ? ABC
9
For example
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the cake (C).
John ate BC.
John ate the banana (B).
John ate A, BC.
AB ? AB
10
For example
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the cake (C).
John ate BC.
John ate the banana (B).
John ate A, BC.
(A ? B) ? (A ? B)
11
For example
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the cake (C).
John ate BC.
John ate the banana (B).
John ate A, BC.
F ? (A ? B)
12
Shieber 1993

• The problem of logical form equivalence is about
  how you get this representation.
• In general, an algorithm can choose this
  representation in one of two ways
• In a reasoner that does general, non-grammatical
  inference.
• Using at least some grammatical knowledge.

13
Shieber 1993

• If it is chosen without access to the grammar
  (modularly) then the surface realizer has to know
  what logical formulas mean the same.
• This is intractable,
• philosophically, because the notion is
  impossible to pin down and
• computationally, because our best attempts are
  not computable.

14
What about GRE?

• Arguably, GRE uses a grammar.
• Parameters such as the preference order on
  properties reflect knowledge of how to
  communicate effectively.
• Decisions about usefulness or completeness of a
  referring expression reflect beliefs about
  utterance interpretation.
• Maybe this is a good idea for NLG generally.

15
Letting grammar fix representation

• Choice of alternatives
• reflects linguistic notions discourse
  coherence, information structure, function.

ABC ? ABC ? ABC ? ABC
AB ? AB
(A ? B) ? (A ? B)
F ? (A ? B)
16
Now theres a new question

• If grammar is responsible for how information is
  represented, where does the information itself
  come from?
• To answer, lets consider information and
  communication in more detail.

17
Information in NLG
Logical space all the ways things could turn out
to be
18
Information in NLG
Common ground the possibilities mutual
knowledge still leaves open.
19
Information in NLG
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the cake (C).
John ate BC.
John ate the banana (B).
John ate A, BC.
Common ground the possibilities mutual
knowledge still leaves open.
20
Information in NLG
Private knowledge the things you take as
possible.
21
Information in NLG
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the banana (B).
John ate the cake (C).
John ate BC.
John ate A, BC.
Private knowledge the things you take as
possible.
22
Information in NLG
Communicative Goal an important distinction that
should go on the common ground.
23
Information in NLG
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the banana (B).
John ate the cake (C).
John ate BC.
John ate A, BC.
Communicative Goal an important distinction that
should go on the common ground.
24
Formal question

• What information satisfies what communicative
  goals?
• Objective modularity
• general reasoning gives communicative goals,
• grammar determines information.
• Another meaty issue.

25
Information in NLG
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the banana (B).
John ate the cake (C).
John ate BC.
John ate A, BC.
Communicative Goal an important distinction that
should go on the common ground.
26
For example
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the banana (B).
John ate the cake (C).
John ate BC.
John ate A, BC.
What John ate was a piece of fruit.
27
For example
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the banana (B).
John ate the cake (C).
John ate BC.
John ate A, BC.
John didnt eat the cake.
28
For example
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the banana (B).
John ate the cake (C).
John ate BC.
John ate A, BC.
John ate one thing.
29
For example
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the cake (C).
John ate the banana (B).
John ate BC.
John ate A, BC.
John ate at most one thing.
30
For example
John ate nothing.
John ate AC.
John ate the apple (A).
John ate AB.
John ate the banana (B).
John ate the cake (C).
John ate BC.
John ate A, BC.
What John ate was the apple.
31
Formal questions

• What information satisfies what communicative
  goals?
• Let u be the info. in the utterance.
• Let g be goal info.
• Let c, p be info. in common ground, private
  info.
• u g?
• p ? u ? g?
• c ? u c ? g?
• p ? c ? u ? c ? g?

32
Logical form equivalence

• An inference problem is inevitable
• u g?
• p ? u ? g?
• c ? u c ? g?
• p ? c ? u ? c ? g?
• But the problems are very different
• not always as precise (entailment vs.
  equivalence)
• not always as abstract (assumptions, context,
  etc.)
• Consequences for philosophical computational
  tractability.

33
GRE, again

• We can use GRE to illustrate, assuming
• c domain (context set)
• g set of individuals to identify
• represented as set of discourse refs
• u identifying description
• represented as a conjunction of properties
• solution criterion
• c ? u c ? g

34
GRE

• How does the algorithm choose representation of
  u?
• The algorithm finds a canonical representation of
  u, based on incremental selection of properties.
• And how does the representation and choice of u
  relate to the representation and choice of an
  actual utterance to say?
• The representation of u works as a sentence plan.