Relations that are seldom or never signale

Bill Mann bill_mann at SIL.ORG
Tue Feb 8 23:26:09 UTC 2000


Fellow RST list folk:

The topic of unsignaled relations in text has been quiet for a while, but I
think that there are some issues that remain.

To summarize prevailing rates of signaling: several people have looked at
signalling in text.  For monologue English text, Maite Taboada reported a set of
analyses in which about 31% of the RST relations were signaled by conjunctions.
Claudia Soria reported 37.3% signaled by a connective in English written texts.
(Several people report that the percentages are moderately higher for spoken
text.)

I have modified the web site so that for the 6 published and 9 unpublished
analyses there, percentages of signaling are shown. For the published texts, 27%
are signaled, and for the unpublished texts, 28% are signaled.  These figures
reflect a very open definition of a signal, including not only conjunctions and
other common connectives but phrases, syntactic patterns (e.g. question syntax
for solutionhood), and even page formatting (e.g. title formats signaling
Preparation.)  (The numbers for conjunctions alone would be lower.)

Doing the counting, I was very much aware that the notion of a signal had not
been formalized and made precise.  However, the imprecise methods are strongly
indicative of what text is actually doing.

Everyone seems to agree that some relations are signaled most of the time, some
are seldom signaled, and that for each relation it is possible to find (or
construct) examples in which that relations is not signalled. (I suspect that
nearly everyone would agree that each of the relations can be signaled. One of
them, Elaboration, can be signaled by "One of them.")

To fully understand my point in this paragraph, you need know that signaling of
Background has been part of the previous discussion.  You might also be
interested in knowing that the RST Background relation can be signalled by "To
fully understand my point in this paragraph, you need to know that ...".

The relation signalling measures cluster around 60-70% unsignalled, and thus
30-40% signaled.  And the 30-40% signaled are not made definite by the signal;
many of the signals are used for several relations.

The dominant case seems to be that the observers are able to assign a relation
without reliance on signals, even the collective effects of a wide diversity of
signals.

If  discourse structure, the kind found by RST, is conveying its own part of
what the text conveys, frequently distinct from what the clauses convey, then
there is an important question about how this happens.  It seems to call for at
least an extension of what is usually called semantics.  But it also appears
that semantics as we know it is tied tightly to the requirement for explicit
signals.

Even if only 10% of the "meaning" of a text is conveyed in an unsignalled way,
that would be a major semantic challenge.  But the percentage is possibly much
higher.  This is a serious problem for linguistic theory.  The percentages are
not important, but the qualitative understandings of how this works are
extremely important.  They may absolutely require a revision in how semantics
works.

I suspect that any explanation of this will involve processed that are also
found within the sentence (or utterance, for those who use that term.)

This is not a simple problem.  A credible solution will involve a great deal of
complexity.  We should not be afraid to posit complex solutions.

Bill Mann



p.s. For your expository comfort: the following spellings are all recognized as
correct: signaled signalled unsignaled unsignalled signaling signalling

Also, I can provide the details for the signaling numbers found on the website
if anyone wants to have them.



More information about the Rstlist mailing list