Relations that are seldom or never signaled

[Alistair Knott]

Here are some followups to issues raised by John Bateman, Daniel Marcu,
Fiorella de Rossis and Bill Mann relating to my heretical comments about
ELABORATION..

1) I started by suggesting that the linguistic resources used for
signalling relations should be used as a source of evidence about how
those relations should be defined, and I illustrated the idea by
suggesting that the fact that `but' can serve to mark several different
relations should count as evidence that these relations have something in
common in their representation. Daniel and Bill both commented that
ambiguous words don't always have anything to do with each other
semantically (e.g. the two senses of `bank' are unrelated). Daniel says:

> ... The word "bank" means different things in different contexts.
> And semanticists do not suggest that the different senses should have
> anything in common.

That's certainly true for the word `bank', where we need to postulate two
separate lexical entries that happen to share the same orthography.
However, the different senses of the word `but' (e.g. in signalling a
violated expectation or a contrast) also show up cross-linguistically, as
do the different senses of the word `because' in signalling a causal
relation and an argumentative relation. In addition, there's an argument
from parsimony that says that if we can find a generalisation between the
two senses, this is preferable to simply postulating two distinct lexical
items. I'd argue that the case is similar for `logically polysemous' words
like `enjoy', which can mean `enjoy eating', `enjoy reading' etc depending
on the argument it takes: again, we don't want to multiply lexical entries
unnecessarily, and there's plenty of cross-linguistic evidence that the
different senses aren't related together accidentally.

2) Now to some specific analyses given. John Bateman's example is a prima
facie case where a BACKGROUND relation is signalled using a connective
(which I was claiming never happens):

> (1)     Look, you'll just have to go with him to the store.
> (2)     Since his foot is broken, he can't go by himself.

> Background? (2a,2b) with 2b as nucleus.

As John says, it's hard to distinguish between Background, Justify, and
Evidence here.

Daniel Marcu gave two texts in which he suggested a background or
justification relation holds between large text spans, and in which the
notion of focus wouldn't be helpful in explicating the relationship
between the spans. (The texts are repeated below.)

In all three texts, it's interesting that there are difficulties in
distinguishing between the RST relations of background and justification.
So there are two separate questions: (i) Is the relation in question
background or not? (ii) If it is background, (a) is it signalled by a
connective? and (b) can a notion of focus continuity be used to explain
the coherence between the two spans?  I must admit that I find it
extremely difficult in practice to use the definitions given in classical
RST to distinguish between background and justification relations. On the
other hand, I think the question of whether or not a particular connective
would be suitable to signal a particular relation is much easier to
determine: at least some of the time the text already contains a
connective, and in the absence of a connective, the question of whether
one would be appropriate draws on the kind of judgement we make all the
time as writers/speakers. For instance, it seems quite clear that we could
use a connective like `because' to signal the relation in Daniel's `Flu
stopper' text:

> [Running nose. Raging fever. Aching joints. Splitting headache. Are
> there any poor souls suffering from the flu this winter who haven't
> longed for a pill to make it all go away? Relief may be in
> sight] , BECAUSE
> [Researchers at Gilead Sciences, a pharmaceutical company in
> Foster City, California, reported last week in the Journal of the
> American Chemical Society that they have discovered a compound that
> can stop the influenza virus from spreading in animals. Tests on
> humans are set for later this year.]

And I really think it's difficult to use `because', or any other
connective, in Daniel's other example text.

>         [The lore of ancient Greece recalls an Olympic athlete who was
> determined to become the strongest person in the world. Every day
> Milon of Croton would pick up a calf, raise it above his head and
> carry it around a stable. As the calf grew, so did Milon's strength,
> until eventually he was able to lift the full-grown cow.
>
>         Milon, who won the wrestling contest five times,
> intuitively grasped one of the basic tenets of contemporary sports
> science.] [Progressive resistance training - the stressing of muscles
> with steadily increasing loads - is something well understood by the more
> than 10,000 athletes from 197 countries who will go to Atlanta, Ga.,
> next month for the centennial of the modern Olympic Games. (...)

This is not just because the spans are large: as Bill Mann noted, we can
signal relations using phrases or even clauses, such as `This is because'
or `This is for the following reason', but no phrases of this sort seem to
fit either. I'd argue that this points up an interesting difference
between the Olympic text and the other two. There has got to be something
that explains why connectives are suitable in two cases and not in the
third. This is true quite independently of any discussion about which RST
relation should be used in analysing the texts. However, given the
difficulty of distinguishing between background and justification going by
the definitions in classical RST, I think it might make sense to take the
question of whether or not a connective can be used as a diagnostic for
whether the relation is background or not. Of course, this means it would
no longer be an empirical result to say that background relations cannot
be signalled with connectives: it'd be true by definition. The interesting
research question would then be to look for a new definition for
background and justification which (a) makes the two easier to
distinguish, and (b) accounts for the data about the places where a
connective can and cannot be used. Again, this would be a significant
departure from RST, whose relations are explicitly defined with no
reference to surface linguistic features.

If I adopt this position, then John Bateman's text, and Daniel Marcu's
`Flu stopper' text would be taken by definition not to involve background,
while Daniel's Oympic text would be taken to involve background. So I'd be
left with Daniel's question about how focus can be used to account for the
coherence of the Olympic text. I guess I'd have to say that `progressive
resistance training' can be thought of as an entity that's mentioned in
the first span (it's referred to as `one of the basic tenets of
contemporary sports science'), and that becomes the focus of the second
span. However, I expect that there are other `connectiveless' texts where
I'd be forced to pick a focussed entity that was either abstract or not
explicitly mentioned or both; I think that these are what Daniel has in
mind, and I agree that these would be problematic for my focus-based
account.

Incidentally, you have to sympathise with the cow in Daniel's Olympic
story. It must have wondered what the world was coming to being lifted up
and carried around the stables all day.

3) Finally, Fiorella de Rossis writes in defence of ELABORATION-OBJ-ATTRIB
that it has been used successfully in several existing text planning
systems.  That's certainly true, but I don't think that this counts as
very strong evidence in favour of its inclusion in a set of relations. For
one thing, my main argument against this relation is on grounds of
parsimony: I was arguing that it does the same job as focus, and that
there's no point in having two constructs in the model that do the same
job. But of course there's no reason why you couldn't nonetheless
implement a planner that included both ELAB-OBJ-ATTR as well as focussing
constraints.

In any case, as I said in my original message, I'm wary of using the
success of a generation system implementing a certain theory of discourse
as strong support for the theory. I think that the fact that RST has been
so widely taken up by people building generation systems can be taken as
good support for its general concept of discourse relations. But at a
finer level of granularity, I feel it's hard to use the performance of
systems implementing alternative sets of relations, or using relations in
different ways, as a way of deciding between these alternatives.  It's
hard to compare the performance of existing systems in any case, because
they tend to operate in different domains and with different goals. But
even if we could compare the outputs of different systems, there's so much
in a generation system besides a declarative theory of discourse (in
particular the algorithms for content selection and text planning) that
would have to be controlled for before we could use these comparisons to
distinguish between the discourse theories implemented.

Ali