Can text structure be incompatible with rhetorical structure

1
Can text structure be incompatible with
rhetorical structure?
Nadjet Bouayad-Agha, Richard Power, Donia
Scott ITRI University of Brighton
2
Rhetorical vs Document

• Rhetorical Structure
• the underlying coherence structure of the text
• hierarchical
• leaves are (unordered) propositions
• nodes are discourse relations (e.g., PURPOSE,
  CONCESSION)
• Text Structure
• the realisation of the rhetorical structure
• hierarchical
• leaves are (ordered) propositions discourse
  markers
• nodes are textual categories (e.g., paragraphs,
  sentences)

3
Elixir contains gestodene. As a result, it
occasionally provokes a mild allergic reaction.
However, Elixir has no serious side-effects.
4
contain(elixir, gestodene)
Elixir contains gestodene as a result, it
occasionally provokes a mild allergic reaction.
However, Elixir has no serious side-effects.
5
The big issue

• mapping from
• rhetorical structure (RhetRep)
• text structure (DocRep)

6
Requirement meaning preservation
(c)
Because it contains gestodene, Elixir
occasionally provokes a mild allergic reaction.
7
Underlying assumptions

• Meaning preservation
• correct content all propositions expressed
• well-formed structure formation rules respected
• structural compatibility organisation of
  propositions in two structures must be compatible
• But what does structural compatibility mean??
• typically isomorphism
• ... exactly corresponding in form and
  relations (OED)

8
Compatible structures the strong definition

• DocRep is compatible with RhetRep if they group
  the elementary propositions in exactly the same
  way.
• i.e.,
• Every set of propositions that is dominated by a
  node in DocRep should be dominated by a node in
  RhetRep, and vice versa.

9
4
10
Compatible structures the weaker definition

• DocRep is compatible with RhetRep if every
  grouping of the elementary propositions in DocRep
  is also found in RhetRep.
• i.e.,
• Every set of propositions that is dominated by a
  node in DocRep should be dominated by a node in
  RhetRep ? but not necessarily vice versa.

11
4
12
4
13
Still too strong?

• If you forget to take your tabletA, take another
  as soon as you rememberB, or wait until it is
  time to take your next doseC. Then go on as
  beforeD.

14
Another example
Consult your doctor immediatelyA if a rash
developsB. It might become seriously infectedC.
15
A key issue extraposition
RhetRep
DocRep
16
Another exception violations of the Nuclearity
Principle
Although Elixir has no serious side-effectsA, it
contains gestodeneC. As a result, it occasionally
provokes a mild allergic reactionB.
17
Implications for NLG

• Structural Compatibility (SC), as we have defined
  it (i.e., the weak version), provides a useful
  constraint for most applications/genres
  therefore text planners should include it.
• But,
• SC is frequently violated in some applications
  (esp. complex instructions), for reasons of
  readability.
• So,
• How can we configure a text planner to be hard
  when it needs to be, and more relaxed when it
  doesnt?

18
Where to relax?
19
A caution

• Transformations must be defined with extreme
  care
• e.g.,
• Linear ordering can make a difference, so if
  following Option 2, the surface rhetorical
  structure (or selected parts of it) will have to
  be ordered.

20

The nuclearity example again
Although Elixir has no serious side-effectsA, it
contains gestodeneC. As a result, it occasionally
provokes a mild allergic reactionB.
DocRep
RhetRep
CONCESSION
s
n
CAUSE
A
C
n
s
B
C
B
as a result
21
Conclusions

• Text planners that assume isomorphism between
  Rhetorical Structures and Text Structures will
  only be able to produce limited ranges of texts.
• Extraposition is a useful device for varying the
  document structure for reasons of
  comprehensibility or style, while still
  preserving the message.
• Segmentation has to take more than the nuclearity
  principle into account.