RE: CG: SUO: RE: Link Grammar and Parser
John Velman wrote:
> On Tue, Dec 02, 2003 at 09:40:11AM +0100, John Bateman wrote:
> > >
> > >
> > >One reason I like RRG is that it postulates a surface
> > >>structure based on
> > >>(roughly speaking) predication, postulates a lexicon that
> includes the
> > >>'logical structure of the verb' that forms the basis for
> a 'semantic
> > >>representation', and postulates a bi-dirctional linking
> > >>algorithm between
> > >>the semantic representation of a clause and the surface structure.
> > >>I could go on, but won't unless there is some interest.
> > >
> > >
> > >
> > >I'm very interested. If the group objects, we can go offline,
> > >though I believe some of the group is interested in this topic.
> > >
> > RRG is just one, rather old, approach to linguistic
> description tht has
> > these
> > properties; its computational implementations are, since
> some rather early
> > work, more or less non-existent. I am surprised that it is being
> > 'reactivated'
> > in this way given that the current state of the art and practice in
> > linguistic
> > engineering is not being kept a big secret somewhere as far
> as I can tell.
> > Perhaps this is another case of different communities not
> communicating
>
> There is sure a lot of that going around!
>
> > (although we are, via the web, all now at least linkwise
> neighbors?).
> >
> > Is there some particular reason that RRG is being raised
> and not LFG,
> > HPSG, TAG and half a dozen others all with solid histories of
> > computational implementation? Add to that the sizeable work on
> > broad coverage analysis and there is quite a literature to work
> > through.
>
> John -
>
> Not quite sure what you mean by RRG "being raised." I
> mentioned it because
> that's where I've gone since my link grammar/dependency
> grammar days. By
> the way, some place I read that RRG is a kind of dependency
> grammar, but I
> don't see it that way. If you have any references along
> these lines I'd be
> happy to see them.
>
> I think RRG has -- as presented in Van Valin and LaPolla "Syntax:
> Structure, meaning, and function" -- some aspects that have a
> bearing on
> generating a knowledge base from text. These include the
> fact that the
> central concepts turn on predication, the nature of
> predicates, and the
> relationships between arguments and predicates. The semantic level
> described by VV&LP is a lot closer to what I would like to
> see than the
> phrase structures other folks end up with.
I can't seem to google up the URL now, but Van Valin also
posted a paper on Intro to RRG (forget the title) which went
over the predication concepts. It seems nicely done, and
well suited to translation into language understanding
systems. From having read that paper last weekend, it seemed
to me that RRG can, at least in principle, be put into a
link grammar form. If someone undertook that task already,
then the CMU link grammar parser should be, in principle,
able to work with RRG specifications.
But I'm still concerned that RRG, a la Van Valin, seems to be
getting organized to polish off the Chomsky Universal Grammar
(CHUG) theories. Van Valin uses dozens of languages in the
intro paper to exmplify the RRG concept in that framework.
If so, then RRG research is oriented more toward theoretical
universality than toward just English language parsing. My
own bent toward the practical has me concerned about the
practical usage of RRG if that is typical of the motivations
behind the field.
> Another thing I like about RRG is the separate operator
> projection, and the
> guidelines as to the scope of various operators (operators:
> markers for
> tense, aspect, negation, evidentials, illocutionary force,
> modality,...).
> How this all works out computationally remains to be seen.
>
> By the way -- what do you mean by 'old'? As far as I can tell RRG
> developed during the 80s more or less, and would appear to be
> about the
> same age as the other things you mentioned -- but I'm no
> expert, and would
> appreciate any guidance.
Van Valin's papers seem to be written in the early 90's. If RRG
dates from the 80's, there must be a more original work that might
be interesting to study.
> One feature of RRG may be positive or negative, depending on your
> perspective. This is it's explicit attempt to be a more or
> less universal
> grammar. This put me off a little at first, but I find that
> it doesn't
> keep me from understanding how it applies to English. VV&LP
> support the
> theory with examples from about 100 languages (based on a
> quick count of
> lines in the language index). Van Valin has said "RRG grew out of an
> attempt to answer the question 'what would linguistic theory
> look like if
> it were based on an analysis of Lakhota, Tagalog, and
> Dyirbal, rather than
> an analysis of English?'". My experience in applying math
> has been that
> sometimes it is easier to solve a problem by looking at it in a larger
> space than seems to bee required, so my gut feel is that this analysis
> based on many languages should be helpful.
Agreed; at some point in time, it might also be useful for
an implementation, but is it ready for prime time now?
> VV&LP, while getting to the semantics, clearly are
> linguistically oriented
> and their semantics, while implicitly based on some metaphysical
> assumptions of course, is elaborated in its details based on
> linguistic
> features. For example, while devoting a good deal of space
> to thematic
> roles, they note that it is important to distinguish between
> the role a
> participant plays in a state of affairs, and the semantic
> interpretation of
> an argument in the logical structure of a sentence.
Yes, and from my (quick) reading of the Intro paper, they find
that English is an "... unusual language ..." in this regard.
Apparently English is more flexible in role order because it is
less flexible in word sequence order than many other languages.
So RRG might be better at solving a problem in universality
of representation, but not providing a significant advantage
for English translation to logic. But then I just read that one
paper all the way through.
> (For those who have read this far and are curious, RRG doesn't use the
> traditional grammatical functions and phrase structures. It
> is organized
> around a clause nucleus (which contains the predicating element), the
> clause core (which contains the nucleus and the arguments),
> and a clause
> periphery (which contains modifying pp and adverbials).
> There is of course
> a lot more to it than this, but this might give some idea of
> how the flavor
> is different from traditional phrase structure grammars.)
The English examples are actually easier to understand than most others.
It might actually turn out to be intuitively more appealing, but I keep
getting lost in the dozens of linguistic examples. I'm still looking
for an overview that treats English primarily, but haven't found
one yet.
> I've managed to avoid any familiarity with HPSG so far :-),
> but have had a
> reasonably good look at LFG. I kind of like LFG and its use
> of unification,
> but don't see how it gets to the semantics. In LFG one has
> the Constituent
> Structure, the F-Structure (feature structure), and other structures
> including a Semantic Structure. I've looked through
> Bresnan's latest book
> and Falk's recent book, and a good many papers on LFG, and I don't see
> things that I want to see. I see a lot of XBar theory, and a
> lot of how to
> get from Constituent structure to F structure. I haven't
> completely given
> up on LFG, but ...
>
> Since I have no deadlines to meet, I'm slowly trying to
> develop a parser
> that is RRG compliant (more or less) that uses Attribute
> Value Structures
> and unification.
I'm surprised you're not working on an LG implementation, given
the speed and quality of the CMU LG parser and the WordNet extensions
to the CMU English link grammar. Am I missing something important?
> Well, I'm not trying to sell RRG to anyone, but I like it so
> far. I may
> like other theories of grammar better eventually, but for now
> RRG is my
> favorite grammatical theory. Parsers are, perhaps, a separate issue.
> Also, it may well be that some of the things I like about RRG
> (as presented
> by VV&LP) can be ported to other theories or to existing parsers.
It appears that RRG can be implemented in LG, but I don't know of
any implementation yet.
>I'm
> willing to discuss pros and cons of RRG with anyone who wants to --
> probably off list unless we come up with some conclusions particularly
> relevant to SUO or CG.
I still think that the automatic translation of broad domain English
corpora, along with specialized functional domain corpora, could
be the best answer to a SUO. Some kind of automated approach might
be able to keep up with the dynamic nature of language and culture,
but a static SUO certainly won't be able to. So the principle aim
of this group, IMHO, should be toward automating a generator that
can be treated with standard performance improvement analyses to
provide that tool.
> >
> > Link grammar is not quite in the same boat, as it is at least
> > computationally
> > "there" and the link to Wordnet is interesting; EVCA is
> also just one of
> > such approaches of its kind. Links between EVCA and WordNet can be
> > found on my ontology portal page under the Common
> Comparison/Combination
> > tasks section (http://www.purl.org/net/ontoportal); if
> anyone has any more
> > I'd like to add them.
>
> Thanks for the references.
>
>
> >
> > John B.
> >
> >
> >
> >
>
>
> Best,
>
> John Velman
>
Glad to get your inputs John S and John V.
Rich Cooper