Relations that are seldom or never signaled

[Alistair Knott]

I'd like to follow up Bill Mann's note about RST relations that are seldom
or never signalled explicitly in text. As Bill says, this seems to happen
very frequently, and seems particularly common for a certain subset of RST
relations. At least, there is a certain subset of relations which are
seldom signalled _using connectives_; there may be other cohesive methods
of signalling them in surface text, as Simon Corston pointed out in an
earlier message.

I've been interested for several years in a methodological approach
whereby the set of linguistic mechanisms available for signalling
discourse relations in a language is used as a source of evidence about
how those relations should be defined. To take a simple example, there are
several RST relations that can be signalled using the connective `but';
this suggests that these relations have something in common, and we would
like this commonality to be reflected in the formal definitions of the
relations concerned. (A similar story can be told for many other
connectives: `and', `because', `so' and so on).  The fact that there are
some relations which are seldom or never signalled using connectives is
particularly significant in the light of this methodology:  if we are
using a taxonomy of linguistic devices as a source of evidence about how
to define relations, we might want to argue that those relations which
aren't signalled using connectives constitute a completely separate class,
to be defined using quite different semantic primitives.  (Alternatively,
if it's found that the definitions of these relations don't seem radically
different from those of relations signalled by connectives, this would be
prima facie evidence against using a taxonomy of linguistic devices as a
source of evidence about how to define relations.)

Before going further, I think that one important distinction to make is
between relations which are _not often_ signalled by connectives, and
relations which _cannot_ be signalled by connectives. Some of the
relations in Bill's list (e.g. EVIDENCE, JUSTIFY, INTERPRETATION,
RESTATEMENT, LIST) are sometimes signalled by connectives. My feeling is
that there are only a couple of relations that are never signallable by
connectives: one is ELABORATION (or rather, the subtype of elaboration
called OBJECT-ATTRIBUTE ELABORATION), and the other is BACKGROUND. [I'd
be interested in other RST analysts' opinions on this, though.] So a
prediction made by the methodology outlined above is that these two
relations fall into a different class than the others, and that
this will be reflected in their definitions.

I think there is a good case to be made for these relations being
different from the others. My suggestion is that these relations are in
fact better modelled using the metaphor of `focus', rather than the
metaphor of `relations between propositions'. The idea is that what's
described as an OBJECT-ATTRIBUTE ELABORATION could equally well be
described as the maintenance of a given focus entity across two text
spans, or a legal shift of focus from one entity to another. In fact, I
would say that the focus-based description seems more appropriate. For
instance, consider the text `I have a brother called Bill. He's 20 years
old'. We could model the coherence of this text by noting that they are
linked by an ELABORATION relation. But what we have here is only a
`relation between two propositions' in a very derivative sense: the
relationship only holds in virtue of the fact that both propositions
include reference to a particular entity (Bill). In other words, an
account of which entities are in focus seems to be primary in this case.
(I think that a similar point can often be made for the BACKGROUND
relation, although the case is not as clearcut.)

Both `entities-in-focus' and `relations-between-propositions' have been
used extensively to model the phenomenon of discourse coherence, and I
think most theorists would agree that both metaphors are needed for a
complete account of coherence. Often, relations-between-propositions and
focus are construed as multiple simultaneous constraints on coherence, as
for instance in Grosz and Sidner's model, or Hovy and McCoy's `Focussing
your RST' paper.  However, if we decide that OBJECT-ATTRIBUTE ELABORATION
and BACKGROUND are better thought of as reflections of focussing
constraints, and we accept that an account of the focus structure of a
text is something which is needed on independent grounds (for instance to
model the pattern of anaphora in a text), then it seems redundant to
continue to treat them as ordinary RST relations: the work they are doing
is already being done by a different, and more appropriate metaphor.

Of course, leaving OBJECT-ATTRIBUTE ELABORATION and BACKGROUND out of the
set of RST relations has serious consequences: it means that it's no
longer possible to build a complete tree of relations for every coherent
text. In fact, since these relations are amongst the most common in RST
analyses, it means that it will seldom be possible to build a complete
tree. The picture that remains is one where a coherent text is described
by one or more RST trees, with RST trees being linked by legal focussing
moves in the case where there is more than one.

Comments welcome on this subversive view! I should say that this model of
text coherence is one that has been used with reasonable success in a text
planning system (the ILEX system---for more details, see
http://cirrus.dai.ed.ac.uk:8000/ilex/ ), though I should also say that I
don't think that this kind of demonstration-by-implementation counts for
very much.

Cheers -

Alistair Knott.