conclusion relation

[Manfred Stede]

Hi all,

>John Bateman wrote:
>
>Some important basic issues of description coming up in
>this interesting discussion: that hasn't happened for a while
>in RST! :-)

interesting discussion indeed!

(btw, if you are an RST fan interested in new horizons and haven't clicked on
http://www.rst.com.tw/ yet, you should really do so.)

It might help to distinguish between (a) theory and (b) annotation practice
here. The latter is the job to provide clear enough guidelines for
annotators to decide on labelling the discourse phenomena we're
interested in (for example, as a basis to devise the theory). This, I
think, is what Gisela has in mind when she points out that focusing on
writer's purposes often leads to quite good results when instructing
people to come up with RST analyses. It seems to me this is highly
dependent on text type, though; according to our own experiences in
Potsdam, it does help for relatively short argumentative text, where
people are often able to identify central statements and supporting
material, and use RSTs nuclearity to represent this.

As everybody in the discussion seems to agree, there are other things
going on in text besides constructing the argument / realizing the
discourse purpose. One important thing is the thematic development,
and I'd argue (and maybe not everybody agrees...) that once we decide
to use RST trees to capture writer's purposes, we can't simultaneously
capture thematic development with those same trees very well. cf. the
discussion of ELABORATION by Knott et al. 2001, and others.

For our corpus of argumentative text, I was thus lead to represent
writer's purposes with a "deeper" structure than RST - capturing just
the support relationships between arguments and counter-arguments.
(In RST, usually EVIDENCE).
Importantly, this is usually incomplete: it does not span the entire
text, as for instance descriptive background material will not be
part of the argument-as-such. Also, and as a consequence, support
relations may hold between non-adjacent material, contrary to what RST
prescribes (unless one tries to exploit the "nuclearity principle" for
this, but that leads to trouble, I think)

Then, thematic development forms a separate level of analysis/annotation,
independent of "deep" writer's purposes, just recording what things
are "about". This is closely related to co-reference relations in the
text. I'm using a hierarchical representation much like that of
Grosz/Sidner 86. This level does span the complete text.

Which leaves the remaining relationships that encode "more" than just
thematic relatedness but "less" than argumentative moves. This probably
largely corresponds to RSTs subject-matter relations. For annotation
practice in our corpus, we decided not to create a dedicated level for
them (again, this is text-type dependent, I think), and instead went
to Martin's conjunctive relations as a surface-oriented level of description,
as discussed by John in his mail below. (It's also quite similar to
what is being annotated in the Penn Discourse Treebank).

Taking all these levels together, as annotated separately, we end up
with a pretty good picture of what is going on in (our type of) text.
Technically, we're annotating all levels with distinct, specific
annotation tools and then fuse everything into a single database for
evaluation/retrieval.
The extended version of the proposal is here:

@InCollection{stede-rst-nuclearity,
   Author="{Stede, Manfred}",
   Title="{RST Revisited: Disentangling nuclearity}",
   BookTitle="'Subordination' versus 'Coordination' in Sentence and Text",
   Editor="Fabricius-Hansen, Cathrine and Ramm, Wiebke",
   Publisher="John Benjamins",
   Address="Amsterdam and Philadelphia",
   Year=2008}

But what about "theory"? The systemic metafunctions discussed by Ed and
John are probably the most thought-through attempt to provide a uniform
and well-defined framework for any kind of multi-level idea. It does not
map perfectly to the annotation practice I described above, though.
Rather than going into that (this mail is long enough already, sorry!!),
just two remarks: The interpersonal metafunction (as I understood it from
Halliday) is not quite intended to capture claim-argument relationships, or
is it? Similarly, I'm not so sure that the interpersonal/ideational
distinction corresponds to the presentational/subject-matter distinction in
RST (which is widely accepted in the community, and rightly so, I think).
Finally, I'm not sure where a level of thematic structure would belong -
there is overlap with the idea of "textual" but again it does not seem to
be the
same. Thus, to my mind, metafunctions are in principle on the right track but
haven't exactly made it.
Then, working with annotated data might help to come up with improvements
to the theory...

cheers,
Manfred

---
John Bateman wrote:

Some important basic issues of description coming up in
this interesting discussion: that hasn't happened for a while
in RST! :-)

A comment:

> Now my question: When one works with a set of labels that are
> hybrids, then how can one describe the situation when only one
> metafunction alone obtains, if there is no 'pure' label for the
> relation found there?  I.e., all relations one would tend to place
> there also bring in connotations from the other metafunction(s)?  (In
>  other words, how does one in practice bring out one of the
> concurrent relations?)

technically this cannot occur . For 'one metafunction alone' to apply
would violate the sense of metafunctions: they must all apply.
*Alternatively* one could choose not to work with
the hybrid relations, but with the individual
metafunctional contributions instead. The latter would then be free
to provide a description that is specific to particular metafunctions
leaving others unspecified.

This is then much closer to Ed's:

> It seems to me the analyst has more freedom to describe the interplay
> of levels (= metafunctions) when he/she can describe each relation
> separately, using an 'exhaustive' set.

The drawback is that many of the constraints that come from
RST (nuclearity, adjacency) begin to be weakened and one is
more or back with something that looks like a conjunctive
relation analysis, which is significantly underconstrained
structurally.

Both cases are were explored in Bateman/Rondhuis, Discourse
Processes,  1997, using both hybrid labels composed out of contributions
from different metafunctions and explicit specification of individual
metafunctions. Pretty schematic though.

John B.