Relations that are seldom or never signale

[Bill Mann]

Dear RST-list folk:

I am encouraged by the discussion on unsignaled relations, partly because we are
getting at some issues that -- in the company of other developments -- can be
very significant.

Alistair's message of 1/20/2000 is long and has many aspects that I would like
to comment on.  I will reply by editing his message and inserting WM: marks on
my comments and AK: marks on his.   Apologies to those who require short
messages.


     ...

AK: Anyway, to recap: the multiple levels issue came up as a result of a
discussion we were having about the relationship between connectives and
relations. The idea that I'm pushing is that we should posit a strong
relationship between connectives and relations, so that tests to do with
inserting connectives in particular contexts can be used as concrete
diagnostics for the presence of particular relations. Bill Mann said he was
happy to use connectives as part of the decision procedure about which
relation applies, but still doesn't want to use them as a way of defining
which relation applies.

WM: Multiple levels, as an issue, has to do with what sort of abstract shape
of theory we would prefer.  It is part of much larger issues of what the
scientific status of RST is.  Is it really a theory?  How does evidence work?
 -- and so forth.

It happens that recently I have done some new work in this area, and I have
proposed a paper to one of the conferences coming up this summer.  The
proposed title is:

The Scientific Status of RST -- Two Views

So, it fits right in here.  I'll sketch it.

As a conceptual orientation, think about a distinction between phenomena and
theories of them.  Informal perceptions of what is in the world can be
identified, and a few isolated and described.  Phenomena can be clumped into
apparently interdependent groups.  Patterns of association can be identified.
None of this is necessarily causal, although the phenomena might be found by
assuming a causal framework beforehand.

For RST the phenomena include the widespread occurrence of texts, writers,
readers and media that deliver texts.  For readers, the impression that certain
texts are coherent is a further set of phenomena.

On the theory side, there are various names of things: theory, model,
explanation, and also validation, proof, demonstration, account.  These terms
are here simply to evoke the phenomena/theory distinction that I am using.

View One:   One view would be that the phenomena include  texts, impressions
that they are structured, and perhaps impressions that they are coherent.  The
theory would include RST; imposing RST on a text would be a way of explaining
the impression of structure and the impression of coherence.  As many have
pointed out, RST has some strengths and abundant weaknesses in this role.
Processes are missing, for example.


View Two:   The second view starts out the same as the first did: the phenomena
include  texts, impressions that they are structured, and perhaps impressions
that they are coherent.  It goes on to say that these are fragments of a larger
field of phenomena.  Written monologues are one kind of attempt by humans to
communicate.  It is closely comparable to other phenomena, such as oral
monologues, dialogues, group interactions using language, persuasion,
remembering texts, errors in remembering ... a much wider field.  A fragmentary
theory is of little value unless it can be related to a suitable larger field of
phenomena, on one hand, and theories on the other hand.

In the larger view, RST is providing more precise data that supplement the
impressionistic statements about coherence, for example.  So observers who find
analyses for texts also regard them as coherent, and they can use their analyses
to explain their finding coherence.  View two is not really making claims about
how reading works or how impressions of coherence arise.  Instead, it has a role
somewhat like building a bigger telescope.  It increases the opportunity to
build explanations or theories, to test them and refine them.  But it is seen as
data gathering rather than theory.

One of the unfortunate outcomes if View Two is taken is that "Rhetorical
Structure Theory" is no longer seen as a theory.

That is a sketch of part of that presentation.

I think that Alistair is working with View One, and I am working with View Two.

AK: We can imagine placing theories of discourse structure on a continuum
of `specificity'. At one end of the continuum, we have a theory that says
that a coherent text can be analysed by a single tree-like structure of
relations, where relations can be defined by concrete tests that relate to
surface signals.  At the other end of the continuum, we have a theory that
says (a) that on top of the tree-like structure we have to allow that
several relations may apply simultaneously between two given text spans,
(b) that relations may hold between non-adjacent text spans, and (c) that
there are no simple surface-based tests to determine the presence or
absence of relations. The theory at the `specific' end of the continuum is
quite readily testable, and probably `more simple than possible' (to quote
John Bateman quoting Einstein). The theory at the `nonspecific' end of the
continuum is very hard to test in the first place.  It seems to me that
the way to progress is to start off with a very specific theory, find out
where it goes wrong, and then look for the simplest and most tightly
specified alternative theory that remedies the problem. My complaint about
positing multiple levels of relations is that seems like too big a jump
towards the nonspecific/unfalsifiable end of the continuum.

WM: In other words, lets make a set of testable models, process theories,
that have various degrees of practical testability.  Certainly.  Why not?

Even the simplest such model will have to have some provision for a
lexicon, assumptions about the analysis process, and many other details.
Its processes might resemble those of the PISA system (Sanders et al) or
the LDM (Polanyi and Scha) or any of several text generation systems.  Or
the processes might be modeled more directly on RST analysis.  There are
many options, and there is a lot of room for diverse model building
efforts.   When these are built, they can be tested to see how they
conform or fail to conform to human judgments about particular texts,
including RST data.

AK: For what it's worth, here are two jumps away from the most specific end
of the spectrum that I would advocate. Judge for yourselves how far I'm
jumping!

1) As Bill noted in his original posting in this thread, there are some
   relations that are seldom if ever signalled by a connective. I suggested
   that ELABORATION and BACKGROUND were the two principal offenders. Rather
   than abandoning the mapping between relations and connectives, I suggested
   that ELABORATION and BACKGROUND should be treated as exceptions, which are
   better modelled using the entity and NP-based metaphor of focus, which is
   needed for independent reasons in a theory of coherence.

   WM: Those two indeed can be expressed without a signal.  But so can the
   other 10 on the list, and the other 16 or so that were not on the list.  We
   need a more comprehensive approach, and when we have it, maybe it will do E
   and B as well.

AK: 2) The set of connectives is itself hierarchically structured: there are
some
   very general connectives, like `and', `so', `then' and so on, which have
   different meanings in different contexts, Rather than treating such
   connectives as lexically ambiguous words with unrelated senses, I argued
   that all relations that can be signalled by a given general connective have
   some semantic component in common. The idea is that connectives don't
   necessarily signal whole relations, but rather signal components of
   relations. The model that results is one where relations are thought
   of as composite constructs, defined in terms of the values of a
   number of orthogonal parameters. This model has something in common
   with a model that  allows multiple simultaneous relations. But it
   still makes quite concrete predictions about the relationship between
   connectives and relations.

   WM:  What little we know about RST in a cross-language context
   suggests that CRST or the slightly extended version on the web site
   move to other languages and work well with minimal adjustment. That's
   anecdotal, of course.

   On the other hand, the signals that indicate relations (partially or
   uniquely), including connectives and many other sorts of linguistic
   objects, vary wildly from one language to another, and are
   diachronically one of the most unstable areas of the (morphophrasal?)
   lexicon.  That's anecdotal too, at least for me.

   The picture of a very stable collection of relations being
   represented by a wildly irregular set of signals needs strong defense
   to be taken as correct.

Anyway, there are a couple more worms for your perusal!

Ali

     WM: Thank you, Ali, for much food for thought.    Bill