SUO: RE: Monosemy in a controlled language, disambiguating word senses [was: Informal definitions of ... ]
Frederick B. Kintanar wrote:
> Added a URI to Copestake and Briscoe on polysemy
> and TELIC qualia below.
>
>
> Richard Cooper wrote:
> > Hi Fred,
> >
> > Frederick B. Kintanar wrote:
> >
> >>Richard Cooper wrote:
> >>
> >>>The notion of monosemy in a controlled language
> >>>bothers me.
> >>
> >>Please try to explain what bothers you.
> >>What intuitions are unsatisfied with
> >>matching one vocabulary term with one
> >>unit of meaning?
> >
> >
> >
> > Monosemy is just too rigid to feel like
> > normal language. If I use anything even
> > approaching the variation in daily conversations,
> > it violates the monosemy rules in John's CLCE.
> >
> > Consider "a pretty little girl's school", sort
> > of a benchmark phrase in that it can be interpreted
> > in many different ways, depending on where you
> > place the accent.
> >
> > Or what about Chomsky's counterexample: "Colorless
> > green ideas sleep furiously." Clearly not a
> > natural sentence, yet grammatically correct from
> > a purely syntactic point of view.
> >
> > My rough vision is that most daily language
> > is repetitive, with moderate variation. Published
> > language is more variable in fiction and much
> > more descriptive. Journal language is precise,
> > with long sentences, and lots of conditioning
> > of cases. None of these fit well into a monosemic
> > framework.
> >
> >
> >
> >>>I'm looking for ways to disambiguate
> >>>polysemous words, now that I have a relational
> >>>database of WordNet synsets.
> >>>
> >>>One thing that suggests itself is to use
> >>>paraphrasing of words in a synset. That led
> >>>me to think about tools to automatically
> >>>generate paraphrases from a corpus.
> >>
> >>I would be interested in characterizing
> >>the different senses of a *verb* in terms
> >>of type constraints on "participant roles"
> >>as evidenced in sample sentences that
> >>exemplify that sense.
> >
> >
> > Yes, and if the sample sentences are generated
> > from a single initial seed sentence, an observer
> > can mark which ones don't make sense, and which
> > ones do.
> >
> > Normal conversation uses only a few thousand
> > words (other than names). If we could get
> > those few thousand words into a data structure
> > of sentences that can be efficiently searched,
> > we could make a better syntactic parser with
> > the first glimmer of semantic filtering.
I've been reading the CoreLex dissertation at
http://www.cs.brandeis.edu/~paulb/CoreLex/corelex.html
by PaulBuitelaar, who did his work under the influence
of James Pustejovsky, who also has some good stuff
on the web. The size of CoreLex looks manageable,
and since it was generated from WordNet, it should be
consistent with that dictionary, though it was made
using version 1.5.
The URL above lets you browse his ontology, but I
don't think he has any inference rules. He uses
what appears to be standard linguistics classes, and
has a good set of relations showing the vocabulary
and how it maps into his class definitions. Also,
he seems to be using a disambiguation mechanism
that is similar to the link grammar specs for words.
It looks like it might be managed by a constraint
solver, with the core classes being the first level
of implementation.
> >>I think we would
> >>need to constrain both the thematic
> >>roles themselves (and John S's proposal
> >>in http://www.jfsowa.com/ontology/thematic.htm
> >>seems an excellent starting point) and
> >>the types of the participants. I'm not
> >>sure how much the FrameNet work already
> >>addresses this idea.
I've read John's paper on thematic roles, but I still
don't understand the intricacies of these roles yet.
They seem to be well known by linguists, so I guess
its just a question of time before I understand them
well enough to use them.
One thing I do have trouble with is the subjectivity
of how roles are assigned. For example, Telic roles,
as you mentioned, refer to "purpose" while Agentive
roles refer to an active agent. So tools would be
Telic, but what constitutes a tool isn't so clear
cut. If I ask someone to do something, is that
someone a tool I use (Telic), but an agent who performs
the task (Agentive)?
The CoreLex classes are combinations of basic types,
but the combining processes aren't clear to me yet.
It seems somewhat like the link grammar concepts, but
the dotted notation (which seems similar to And/Or
representations) still throws me. I guess its just
a question of reading more.
> > I haven't seen much published from FrameNet,
> > though someday maybe a good tutorial will
> > come out of that group on how FrameNet is
> > organized.
> >
> > John's thematic roles sound like a good
> > starting point, since they seem to be role
> > names often used by linguists. But as a
> > nonlinguist, I would like to see a clear
> > definition of each word; John's treatment
> > assumes you already understand each of
> > the role words. As I study the cases
> > more, maybe I'll become familiar with
> > them.
> >
> >
> >
> >>This would produce something like a
> >>semantic treebank of exemplifying
> >>phrases (perhaps based on simplified
> >>sentences with just the head words
> >>of the noun phrases, but adding
> >>some implicit information about
> >>the situation being mentioned, enough
> >>to distinguish this word sense from
> >>others).
> >
> >
> > Yes, phrases that are indexed by WordNet
> > synsets could make a searchable data
> > structure of example phrases, including
> > sentence forms with variables that can
> > represent the noun phrase, verb phrase,
> > and other variations in a corpus.
> >
> > Then there's still the issue of how
> > to choose one interpretation when
> > multiple interpretations are possible.
> > I've glossed over that issue because
> > I don't yet have the data structure to
> > experiment with, but that will be a
> > tricky problem.
In principle, I think that's a constraint
solving problem. In practice, it will be
very important to discover heuristics to
limit the number of combinations that are
searched. I think that's what people learn
to do in daily conversation, which is why
much daily conversation is quite repetitive.
That repetitiveness seems to be what makes
simple Eliza-like chatbots seem slightly
to understand statements, but with little
flexibility.
> >>For nouns, qualia like TELIC (which
> >>constrains the meaning of nouns that
> >>refer to artifacts in relation to some
> >>purpose, typically characterized by
> >>a verb that fulfills that purpose) might
> >>provide a comparable degree of resolution.
> >
> >
> > WordNet has no definition of TELIC, and
> > a google search brings up few good sounding
> > hits. Do you have a URL that might be
> > good for explaining TELIC to me?
> >
> http://www.cl.cam.ac.uk/~aac10/papers/jsem.pdf
>
> Ann Copestake and Ted Briscoe, 1995. Semi-productive
> Polysemy and Sense Extension (PDF) (Revised version
> of ACQUILEX II WP NO. 23) Journal of Semantics, 12, 15-67.
>
> This was also published in _Lexical Semantics:
> the Problem of Polysemy_ edited by James Pustejovsky
> and Branimir Bofuraev, Clarendon, 1996.
>
> > I did find a URL that seems to contradict
> > what I've said above:
> > http://www.cogsci.ed.ac.uk/~kversp/ftp_html/node165.html
> > This author claims that
> >
> > "... an adequate computational lexicon can
> > only be established on the basis of
> > top-down design derived from a linguistic
> > theory in combination with bottom-up
> > information derived from corpora about
> > specific usage of language."
> >
> > I don't find this a compelling argument,
> > except that a simple parse (e.g. link parser)
> > could help organize the samples into phrases.
> > People don't learn a theory of language
> > and then fill it in. They learn meanings
> > and ways to combine atomic meanings into
> > molecular sentences and compound paragraphs.
> > There has been little success in NLP
> > over the last five decades while following
> > a syntax-first approach.
Although that seems to be changing slowly with
semantics pushing ahead.
> > It seems to me that the way people learn
> > language, the steps we go through, should
> > be emulated by a program to see if this
> > approach could provide improvement in the
> > results.
> >
> > But then the author goes on to describe
> > how corpora could be used to deal with
> > specific linguistic representation problems,
> > and I like his suggestions in that vein.
> >
> >
> >
> >>I'm getting this from the article by
> >>Ann Copestake (and somebody else? I
> >>don't have the book handy) in Lexical
> >>Semantics edited by Pustejovsky and
> >>Briscoe (?))
> >
> >
> >
> > Googling "Ann Copestake", I found:
> > http://www.cl.cam.ac.uk/~aac10/papers/gslt-slides.pdf
> > On slide 13, she claims that Terry Windograd's
> > SHRDLU program doesn't scale up well.
> > However, I think the NLP problem is really
> > one of engineering effective ways to accomplish
> > this scaling. SHRDLU's limited context has
> > to be augmented with more complex contexts,
> > and with a library of context frames that
> > covers many common experiences.
> >
> > She does discuss question-answering systems
> > in several parts of this paper, and that is
> > the kind of thing I would like to experiment
> > with. Once a SHRDLU2 is developed that can
> > handle a small number of contexts that relate
> > to linguistics, and a library of common object
> > simulations referenced in daily speech, I think
> > the system can start to fill in its own tables
> > by searching corpora, accepting corrections
> > from observers, and otherwise acting like
> > children do as they learn language.
> >
> >
> >
> >>I'm afraid I still don't
> >>understand qualia in lexical semantics,
> >>and haven't had time to study Corelex.
> >
> >
> >
> > Qualia are experiences of sensation, so they
> > represent world knowledge, right? Lexical
> > descriptions of qualia are dependent on the
> > experiencing subject, and probably do vary
> > more than most domains.
> >
> >
> >
> >>I think this approach can be tested
> >>in small domains like the text of
> >>*definitions* for small standard
> >>vocabularies.
> >
> >
> > I agree. One step up from SHRDLU would
> > be a useful experiment. If well instrumented,
> > it could provide a basis for the next step.
> >
> >
> >
> >>For example, I have
> >>been playing around with the verbs
> >>and nouns in the definitions of the
> >>15 terms in simple Dublin Core. This
> >>could be helpful in ensuring that
> >>terms introduced in qualified DC
> >>have appropriate relationships with
> >>the terms they are refining.
> >
> >
> >
> > Dublin Core is very small, so yes I think
> > that might be a good one for you to work
> > with. I'm still required to keep my
> > experiments work related, and I can't
> > justify studying DC for that reason.
> >
> >>Fred
> >
> >
> > Thanks for your email. Hopefully we can
> > help each other with our work.
> >
> > Rich
Again,
Rich