Semantics and NLP
Jorn Barger
appears university programs in AI are (still) doing an extremely bad job
of covering *semantics*...
An unabridged dictionary may have half a million entries. Roget tried
to sort these into 1000 categories, in a hierarchical tree:
1.abstract relations
1.existence
2.relation
3.quantity
4.order
5.number
6.time
7.change
8.causation
2.space
1.generally
2.dimensions
3.form
4.motion
3.matter
1.generally
2.inorganic
3.organic
4.intellect
1.formation of ideas
2.communication of ideas
5.volition
1.individual
2.intersocial
6.affections
1.generally
2.personal
3.sympathetic
4.moral
5.religious
Semantic AI has to wrestle with ontologies like Roget's, using a
precisely-defined, *limited* vocabulary to express the very-wide range
of realworld concepts.
And I think so long as one stays in the top three categories of Roget's
tree (abstract relations, space, matter) progress can be-- and is
being-- made.
But when human psychology enters the picture (intellect, volition,
affections) precise definitions and limited vocabularies instantly and
catastrophically *fail*.
And this barrier has apparently traumatized the field of AI so severely
that the topic is practically taboo! It's just excluded from
discussion.
*If* a neat ontology-of-psychology could be created, all the other
problems of NLP might just evaporate. Understanding a sentence would
just require finding the node in the ontology that expresses the exact
same meaning (in a generalised form).
I've been arguing that this ontology _can_ be built if we think of the
nodes not as dictionary-entries, but as the whole 'usual stories'
surrounding a concept or word.
So Roget's 'intellect' node would correspond to the whole usual-story of
human intellect: children are born with limited intellect, they learn,
some learn faster, some learn more, humans use intellect to solve
problems, they teach others, etc. (This could even be encoded in the
form a short, human-readable encyclopedia article, written in simplified
English.)
Particular concepts like 'smarter' can then be linked to specific parts
of this general story as 'specialisations'. (The usual 'smarter' story
involves getting praised in school, becoming annoyingly self-important,
going to college, etc.)
But remarkably, the mental skill required to think in terms of these
usual-stories is much closer to the novelist's art than to anything
taught in comp-sci-- to such a great extent that comp.ai.nat-lang barely
acknowledges the topic as appropriate!
But I think I've found a leverage point, finally: pseudo-XML tagging of
the entries in Web *timelines*.
Because the authors of timelines are trying to limit themselves to the
most significant discrete events (in all of history), timelines do an
excellent job of prioritising human behaviors, and so of identifying the
most-useful limited vocabulary for human history.
Examples:
person1 is born at place on date to mother person2 and father person3
person1 is educated at place by person2
person moves from place1 to place2
person creates creative-work
person founds social-institution
person joins social-institution
person discovers theory
person1 fights person2
person leads group with persons2-3-etc
group fights group
etc
These, then, become the 'root-level' usual-stories in the psychological
ontology.
And my 'newsdiff' suggestion was about the need for a NLP to have a
mega-timeline of history that's continually updated with new
news-events, winnowing just the new developments from the standard
high-redundancy journalistic style.
ai faq: http://www.robotwisdom.com/ai/
timelines project: http://www.robotwisdom.com/science/history.html
--
http://www.robotwisdom.com/ "Relentlessly intelligent
yet playful, polymathic in scope of interests, minimalist
but user-friendly design." --Milwaukee Journal-Sentinel
Elias Ponvert
I have a few comments on this posting, hence I respond in-line below.
On Mon, 5 Nov 2001, Jorn Barger wrote:
> An unabridged dictionary may have half a million entries. Roget tried
> to sort these into 1000 categories, in a hierarchical tree:
Before dealing with the comments below, there are a couple good ontologies
of English. The OED and Webster's (I forget which Webster's -- one of the
big ones) are both good. And then there's Quirk et al. for those with a
lot of time on their hands.
> Semantic AI has to wrestle with ontologies like Roget's, using a
> precisely-defined, *limited* vocabulary to express the very-wide range
> of realworld concepts.
>
> And I think so long as one stays in the top three categories of Roget's
> tree (abstract relations, space, matter) progress can be-- and is
> being-- made.
Perhaps. There's a kind of tradeoff involved here -- the smaller the
ontology wherein you ascribe meaning, the (much) smaller the coverage of
NL expressions you'll be able to handle. I think you make this point, in
so many words, below.
<snip>
> *If* a neat ontology-of-psychology could be created, all the other
> problems of NLP might just evaporate.
First of all, that's a really big 'if'. Frankly, I feel tha no such 'neat'
ontology exists. If by 'ontology' I assume you mean semantic network,
things become very tricky. Take notions of belief. While you can imagine
where BELIEF would fit into such a network (e.g. a BELIEF is a kind of
EPISTEMIC_OBJECT -- or, a BELIEF is a component of an EPISTEMIC_STATE)
beliefs also wrap around, in some sense, propositions. Namely that,
propositions, or something like them, are often taken to be the *objects
of beliefs*. On the other hand, other EPISTEMIC_OBJECTs would consist of
PERCEPTs. Here now, we find the objects of PERCEPTs to be, among other
things, cows and trees. So generalizing BELIEFs and PERCEPTs to
EPISTEMIC_OBJECTs we question what are ther objects of EPISTEMIC_OBJECT
itself? A union of COW, TREE and PROPOSITION? TOP or THING? Undefined? At
this point already one runs into difficulty discussing coherently a class
PROPOSITION that could bear any relation to COW and TREE. I feel that
that such an ontology would, a priori, fail any human-readable criteria of
'neat'-ness.
Secondly, all that aside, even if such an ontology were possible, it is
hardly the case that any of the significant NLP problems would evaporate
at all. Three such problems come to mind:
1 Ambiguity -- the problem of ambiguity resolution, which is, to my
understanding, THE problem of NLP, would remain effectively untouched
regardless of how big or sophisticated your semantic ontology gets
2 Discourse -- the resolution of discourse referents, traditionally
pronomial and temporal referents, as well as "the" terms, is very
difficult to achieve, again, regardless of your ontology.
3 Oblique contexts -- perhaps the oldest problem of modern linguistics,
resolving modal and/or belief contexts with finite memory and processing
power (i.e. Kripke structures, as traditionally understood, are simply
TOO BIG to be handled in an NLP setting).
And these are just a few. While, yes, semantics should be a core part of
NLP -- more important, perhaps, than syntax itself -- semantics cannot be
reduced to an ontology or semantic network.
I had concerns, along the lines of my points above, with the rest of the
posting, but I think if I continued a running commentary I'd do little
more than repeat myself ad naseum, so I stop here.
Elias Ponvert
Jorn Barger
Elias Ponvert <e...@uts.cc.utexas.edu> wrote:
> I feel that that such an ontology would, a priori, fail any human-readable
> criteria of 'neat'-ness.
You just need hierarchies that can be orthogonal with self-similar
copies: http://www.robotwisdom.com/ai/thicketfaq.html
Joshua Tauberer
of your points that I don't completely follow...
> 1 Ambiguity
Wouldn't an ontology help to rule out many cases of ambiguity that don't
make sense, within the ontology? Or are you saying that there is just a lot
of irony, symbolism (is this what you mean? what else?) out there?
> 3 Oblique contexts -- perhaps the oldest problem of modern linguistics,
> resolving modal and/or belief contexts with finite memory and processing
> power (i.e. Kripke structures, as traditionally understood, are simply
> TOO BIG to be handled in an NLP setting).
I looked up what oblique contexts are.... Really interesting. I never
thought of that before. What are Kripke structures, how are they related,
and just how big are they?
Thanks.
- Joshua Tauberer
**Nothing Unreal Exists**
Holger Schauer
> Before dealing with the comments below, there are a couple good
> ontologies of English. The OED and Webster's (I forget which
> Webster's -- one of the big ones) are both good.
I don't think these count as ontologies, at least not in my
opionion. I would like to see some formal principles to hold in an
ontology and I guess none of these would (even Wordnet doesn't
AFAIK). Second, I don't know what "onotlogy of English" is
supposed to mean, quite frankly. Ontology is the study of /beings/
(and perhaps also of non-beings), not of (a particular) language.
>On Mon, 5 Nov 2001, Jorn Barger wrote:
>> *If* a neat ontology-of-psychology could be created, all the other
>> problems of NLP might just evaporate.
Another one of Jorns typical of-the-shelf constructions: an ontology
of psychology. What's that supposed to mean? An ontology of terms used
in psychology? An ontology of things rendered valid by psychologists?
> 'neat' ontology exists. If by 'ontology' I assume you mean semantic
> network, things become very tricky. Take notions of belief. While
> you can imagine where BELIEF would fit into such a network (e.g. a
> BELIEF is a kind of EPISTEMIC_OBJECT -- or, a BELIEF is a component
> of an EPISTEMIC_STATE) beliefs also wrap around, in some sense,
> propositions.
Yup. Epistemics are tricky in such systems, but not impossible (well,
at least if you don't restrict the notion of ontology to 1980-style
frame systems). One approach would be to partition your A-box
(holding the assertions), so that you have some partition BU1 (read:
belief of some user U1) which holds assertions about the beliefs of
U1. I'm quite sure that there are both more elegant and more
sophisticated methods.
> Secondly, all that aside, even if such an ontology were possible, it
> is hardly the case that any of the significant NLP problems would
> evaporate at all. Three such problems come to mind:
>
> 1 Ambiguity -- the problem of ambiguity resolution, which is, to my
> understanding, THE problem of NLP, would remain effectively
> untouched regardless of how big or sophisticated your semantic
> ontology gets
I don't buy that argument nor do I buy this one:
> 2 Discourse -- the resolution of discourse referents, traditionally
> pronomial and temporal referents, as well as "the" terms, is very
> difficult to achieve, again, regardless of your ontology.
In fact, our research group as well as several others (the SRI group
headed by Jerry Hobbs comes to mind) have quite succesfully shown how
to tackle both problems with an ontology. Of course, a lot of tough
problems and unsolvable examples remain, but having an ontology may
help a lot.
> 3 Oblique contexts -- perhaps the oldest problem of modern
> linguistics, resolving modal and/or belief contexts with finite
> memory and processing power (i.e. Kripke structures, as
> traditionally understood, are simply TOO BIG to be handled in an NLP
> setting).
Well, of course, if your try to represent /all/ possible worlds, this
is certainly going to pose quite a problem. However, it may turn out
that there are possible workarounds, although doing proofs may become
quite hard (read: impossible). But I'm not an expert in modal logics,
so I'll keep my mouth shut.
> And these are just a few. While, yes, semantics should be a core
> part of NLP -- more important, perhaps, than syntax itself --
> semantics cannot be reduced to an ontology or semantic network.
Nah, but having an ontology might actually help a lot to a) represent
the semantics and b) to derive the semantic representation.
Holger
Jorn Barger
> 1 Ambiguity
> 2 Discourse
> 3 Oblique contexts
By the way, I have a standing challenge/offer on this newsgroup, to
represent _any_ sentence/concept/etc that anyone proposes, using my
notation.
Ray Dillinger
Elias Ponvert wrote:
> On Mon, 5 Nov 2001, Jorn Barger wrote:
> > *If* a neat ontology-of-psychology could be created, all the other
> > problems of NLP might just evaporate.
> First of all, that's a really big 'if'. Frankly, I feel tha no such 'neat'
> ontology exists. If by 'ontology' I assume you mean semantic network,
I contend that the ontology-of-psychology probably exists, but
we will need a broad class of mathematical and theoretic tools
which we do not now have before we can address it in a way where
it would seem "neat."
"neatness", in this context, is just a word for "a set of ideas
that follows mathematical models we understand." Psychology
really isn't that, at the moment. We chase it with a bunch of
ad-hoc and poorly understood models and heuristics. If somebody
finds a model that works well, it will probably be horribly "messy"
by current standards -- but then it will be studied, picked at,
and perhaps eventually understood mathematically -- whereupon,
in the time of our grandchildren, it will become "neat." Until
that time it will be enhanced in a groping way, empirically, by
doing experiments and seeing what works better.
> Secondly, all that aside, even if such an ontology were possible, it is
> hardly the case that any of the significant NLP problems would evaporate
> at all. Three such problems come to mind:
>
> 1 Ambiguity -- the problem of ambiguity resolution, which is, to my
> understanding, THE problem of NLP, would remain effectively untouched
> regardless of how big or sophisticated your semantic ontology gets
You get into seriously murky areas here, because the way humans resolve
ambiguity (this is my opinion only) is by framing their sentences in the
first place in light of their assumptions about how their fellow humans
will disambiguate them. So it involves belief -- my belief that you
understand that the subject we've been talking about is "a herd of
cows" for example informs me that I can use a word like "they" and
that you will understand what it means. My belief that you are aware
of the inability of cows to fly and the rarity of airborne cows in any
circumstances informs me that a word like "over" will be understood
as "beyond" rather than "above". If I'm right in all my assumptions
about *YOU*, then when I say "they're over that ridge," you don't get
confused. Well, assuming that there's only one ridge in sight, or I
point to say which I mean.
IOW, the ambiguity of english sentences reflects the speaker or writer's
beliefs about the reader or listener's beliefs. And to disambiguate
something, sometimes the reader or listener needs to infer the proper
meaning from the reader or listeners beliefs about what the speaker
or writer believes about the reader or listener's beliefs....
So, offhand, I'd tend to argue that an ontological understanding of
psychology -- beliefs-about-beliefs -- is a necessary and vital thing
for purposes of disambiguation of ordinary sentences.
Bear
Charles Moeller
Barger) writes:
>By the way, I have a standing challenge/offer on this newsgroup, to
>represent _any_ sentence/concept/etc that anyone proposes, using my
>notation.
>
Here are two:
1. The single-shot gun is loaded, then at some later point it is fired, that
action causing the gun to become unloaded.
2. Fred is alive, then at some later point he is shot, after which action he
subsequently dies.
Good luck,
Charlie
Dave Davies
Jorn Barger wrote:
> By the way, I have a standing challenge/offer on this newsgroup, to
> represent _any_ sentence/concept/etc that anyone proposes, using my
> notation.
I'd like to see your representation of:
"By the way, I have a standing challenge/offer on this newsgroup, to
represent _any_ sentence/concept/etc that anyone proposes, using my
notation."
dave
Jorn Barger
but politely please.]
Dave Davies <da...@ozemail.com.au> wrote:
> I'd like to see your representation of:
> "By the way, I have a standing challenge/offer on this newsgroup, to
> represent _any_ sentence/concept/etc that anyone proposes, using my
> notation."
I expect it was just accidental on your part, but this challenge really
interests me.
The generalised 'story' that we want to start from might be:
- person conceives theory
- others not-accept theory
- person challenges others to challenge theory
- (other e-quotes challenge, and somewhere Goedel weeps!)
In the 'usual science story', these others would challenge it
immediately, I think... but this depends on a distinction between
'credible theory-proposer' and 'not-yet-credible theory-proposer'...
with the PhD system being the conventional filter...
(A common alternate might be:
- non-PhD conceives theory
- non-PhD communicates theory to open-minded PhD
- o-m PhD brings theory to attention of others
but sometimes instead, alas:
- non-o-m PhD blackballs non-PhD)
So imagine that our eventual core-knowledgebase includes a mega-summary
of the 'usual science story' made up of bits like these, collating all
the usual steps in the evolution of scientific theories. 'Non-PhD
challenges others' will be a sidebar off the main path.
The non-PhD's _specific_ challenge here concerns knowledge
representation in NLP-AI. There's something unavoidably self-defeating
about trying to find a symbolic representation for any particular
_unsolved_ question in science, because the symbols you choose can't
possibly be as appropriate as the ones you'll be able to replace it
with... _once the problem is solved._
And similarly, any representation of a theory ought to take into
consideration that the theory may later be proved wrong...?
But we might at least treat it at the level of 'string of ascii
characters claiming to be insightful'. Or even 'string of ascii
characters claiming to be pseudo-code'...?
I also wonder about the notion of 'standing challenge'. Roget's
classifications of 'challenge' are a big mess:
http://machaut.uchicago.edu/cgi-bin/ROGET.sh?word=challenge
There's an element of 'contest' where anyone who accepts the challenge
may 'win' or 'lose'. A government is in effect a standing challenge--
'I dare you to overthrow me'...
Elias Ponvert
Thanks for the response!
Joshua Tauberer <taub...@for.net> wrote:
> I'm really interested in what you've said and I hope you can clarify a few
> of your points that I don't completely follow...
Be happy to, if I can.
>> 1 Ambiguity
> Wouldn't an ontology help to rule out many cases of ambiguity that don't
> make sense, within the ontology? Or are you saying that there is just a lot
> of irony, symbolism (is this what you mean? what else?) out there?
Perhaps, in some cases. But I feel that an ontology alone cannot root
out all of the really tricky kinds of ambiguity that exist.
First of all there's lexical ambiguity. This alone falls into two kinds:
Lexical semantic ambiguity and lexical grammatical ambiguity. (I make no claims
that these are the terms agreed upon by the AI / Linguistics community.
They're just what came to mind now -- undoubtedly there are some technical
terms whose names I'm forgetting. Anyway ...)
Semantic ambiguity is relatively simple -- a word, has two (or more) distinct
meanings. Or to be more precise, two words sound (or are spelled) precisely
the same. Take 'bank'. My parser has to recognize "John was at the bank".
To resolve this sentence correctly, the parser has to figure out, either
from discourse context or some knowledge about John, whether John was
at a financial institution or the side of a river or stream. Or, take
a sentence that is not ambiguous, to human English speakers: 'John
depositted his money at the bank'. You're traversing your search tree
looking for an appropriate meaning for this sentence. This tree will
branch at both 'deposit' and 'bank' (deposit can mean to hand money over
to a financial institution or to dump a lot of soil or soot or whatever
somewhere). In order to resolve the meaning of these words coherently,
you need to find and evaluate 'money' correctly.
Grammatical ambiguity occurs when a word (or word-sign) is encountered
that could be of more than one grammatical type. So, a classic
example is 'The horse ran past the barn fell'. Here 'ran' must be understood
as a past-participle, not a past-tense indicative verb -- although
everything in its context (except the final 'fell') would indicate the latter.
(I suppose you could say, don't worry about context, just generate the
whole search tree and do a blind search. Or something. This may work,
but in general such search tree tend to be huge, and its enormously helpful
to cut them down at any possible point. Using contextual information is
one such technique.)
Then, there is what's known as structural ambiguity -- in so far as this is
beyond the scope of just words. E.g. 'John can sing and dance the tango'
(suppose one could sing the tango. Maybe one can, I don't know.) Here,
we can get two distinct readings -- 'John can sing (basically whatever
he wants) and can dance the tango' v. 'John can sing the tango and dance
the tango'. Also, there are semantic ambiguities that evade reduction
to syntax -- or the lexicon. The de re/de dicto distinction is one
such. So, for example, 'Every man loves a woman' can either mean that
for every man x, there is a corresponding woman y, say, girlfriend(x), such
that x loves y (or x loves girlfriend(x) ); or, it can mean there exists a
woman y, say, Princess Di, such that for every man x, x loves y (or
x loves Princess Di). Most syntactic theories I've seen would assign
this sentence a single structure. (Montague grammar distinguished them
syntactically in terms of the ORDER of application of syntactic rules).
Anyway, not to get too philisophical in my discussion of NLP. I don't
personally feel that these problems are intractable. I only meant that
ontologies alone can't solve them.
>> 3 Oblique contexts -- perhaps the oldest problem of modern linguistics,
>> resolving modal and/or belief contexts with finite memory and processing
>> power (i.e. Kripke structures, as traditionally understood, are simply
>> TOO BIG to be handled in an NLP setting).
> I looked up what oblique contexts are.... Really interesting. I never
> thought of that before. What are Kripke structures, how are they related,
> and just how big are they?
Well, Kripke structures are a kind of model developed by Saul Kripke
(or Ruth Barcan Marcus, as I've been led to believe) to handle the semantics
of the modal operators 'necessarily' and 'possibly', which are intimately
related to oblique contexts in general. So, in short, take your toy
logical language L = { p, q, r, ... } closed under 'not', 'and' and
'implies'. A model, or an interpretation for this language consisted
of and interpretation function I() on the letters p, q, r ... (so, like,
I(p) = true, I(q) = false etc.) and was extended to the rest of the language
as follows:
I(P and Q) = true iff I(P) = true and I(Q) = true
I(not P) = true iff I(P) = false
I(P implies Q) = tuee iff I(Q) = true or I(P) = false
and so forth. If you extend L to include expressions like nec(P) and pos(P),
the interpretation you need looks like the following: instead of a
single interpretation function I you have a set of functions S and a
relation R: S x S, i.e. on these interpretation functions. (These functions
are often called "Worlds".) A full Kripke structure is the tuple of this
set of worlds with this relation : (S, R), often together with a world
assumed to be 'this world': I. So, the operators and, not and implies are
interpreted just as the were above, in I. However,
nec(P) = true iff for all f such that I R f, f(P) = true
i.e. in every accessible world, P is true
pos(P) = true iff there exists an f such that I R f and f(P) = true
i.e. in some accessible world, P is true
First of all, interpretations along these lines capture logicians intuitions
neatly as to 'what possibility and necessity mean'. Secondly, there is a
cool correspondance between the axiom schemes of modal logic and properties
of the accessibility relation R. Thirdly, this is elegant and cool. Does
it help processing the semantics of modal concepts as they arise in NL?
Not really because you then have to traverse a, likely infinite,
Kripke structure in search of an appropriate world. Moreover, checking
the truth of a moderately complex sentence (above is just a propositional
toy language -- things become more tricky still when you allow sets
or, equivalently, predicates) is a pretty time heavy operation in a
non-trivial knowledge base. Having to check a sentence the number of times
it would take to run a Kripke structure would be very very costly.
This, of course, is just a summary. Again, I think this is a fasinating
problem, and not an intractable one. Still, a good ontology alone can't solve
it.
> Thanks.
> - Joshua Tauberer
Hope this all helps.
Elias Ponvert
Elias Ponvert
Holger Schauer <Holger....@gmx.de> wrote:
> On Sun, 11 Nov 2001, Elias Ponvert wrote:
>> Before dealing with the comments below, there are a couple good
>> ontologies of English. The OED and Webster's (I forget which
>> Webster's -- one of the big ones) are both good.
> I don't think these count as ontologies, at least not in my
> opionion. I would like to see some formal principles to hold in an
> ontology and I guess none of these would (even Wordnet doesn't
> AFAIK). Second, I don't know what "onotlogy of English" is
> supposed to mean, quite frankly. Ontology is the study of /beings/
> (and perhaps also of non-beings), not of (a particular) language.
That's all fair. I just meant that, in my experience, the good-old
fashioned dictionaries have a lot of useful generalizations about
English syntax and semantics formal linguists tend to overlook.
And, yes, 'ontology of English' is a bad term.
> I don't buy that argument nor do I buy this one:
>
>> 2 Discourse -- the resolution of discourse referents, traditionally
>> pronomial and temporal referents, as well as "the" terms, is very
>> difficult to achieve, again, regardless of your ontology.
> In fact, our research group as well as several others (the SRI group
> headed by Jerry Hobbs comes to mind) have quite succesfully shown how
> to tackle both problems with an ontology. Of course, a lot of tough
> problems and unsolvable examples remain, but having an ontology may
> help a lot.
<snipping my further comments on possible worlds>
> Well, of course, if your try to represent /all/ possible worlds, this
> is certainly going to pose quite a problem. However, it may turn out
> that there are possible workarounds, although doing proofs may become
> quite hard (read: impossible). But I'm not an expert in modal logics,
> so I'll keep my mouth shut.
>
>> And these are just a few. While, yes, semantics should be a core
>> part of NLP -- more important, perhaps, than syntax itself --
>> semantics cannot be reduced to an ontology or semantic network.
> Nah, but having an ontology might actually help a lot to a) represent
> the semantics and b) to derive the semantic representation.
>
> Holger
Absolutely. I simply meant to say, yes, a solid ontology is a probably
necessary part of a good interpretation system. But a semantic network
alone can't do all your work for you. For example, an inference engine
shoould be present as well.
So: I fully agree with Holger on (a) and (b). I object merely to the
proposition that ontologies alone can make the problems I mentioned
disappear.
--
elias ponvert | cit | diia | ut austin
t 512.475.6341
f 512.475.6353
e e...@uts.cc.utexas.edu
w www.utexas.edu/academic/cit
Jorn Barger
> 1. The single-shot gun is loaded, then at some later point it is fired, that
> action causing the gun to become unloaded.
The 'usual weapon story' is a specialisation of the 'usual tool story':
- person makes tool/weapon
- person uses tool/weapon
- person repairs tool/weapon
A gun is a composite weapon with a container-of-fuel-plus-slug as one
part, that must be 'repaired' after each use.
> 2. Fred is alive, then at some later point he is shot, after which action he
> subsequently dies.
I know my ontology needs a health-status attribute, but I don't have any
neat place for it yet. (The usual-weapon-story involves a person
suffering a decline in health.)
Charles Moeller
>Charles Moeller <cmoe...@aol.comnulspam> wrote:
>> 1. The single-shot gun is loaded, then at some later point it is fired,
>that
>> action causing the gun to become unloaded.
>
>The 'usual weapon story' is a specialisation of the 'usual tool story':
>
>- person makes tool/weapon
>- person uses tool/weapon
>- person repairs tool/weapon
>
>A gun is a composite weapon with a container-of-fuel-plus-slug as one
>part, that must be 'repaired' after each use.
>
>> 2. Fred is alive, then at some later point he is shot, after which action
>he
>> subsequently dies.
>
>I know my ontology needs a health-status attribute, but I don't have any
>neat place for it yet. (The usual-weapon-story involves a person
>suffering a decline in health.)
Thanks for your effort. I appreciate the fact that you
...continuing your method:
-person from x is born is alive
-person by y is mortally wounded
-person by (bleeding/destruction of vital organ) is not alive (dead)
Your "makes/uses/repairs" and the "alive/wounded/dead" list, I take it,
are visited one at a time as in frames (standard programming method).
There is no evidence of communication between frames, so the "uses"
can not know if "makes" is complete. The "repairs" can not know the
"uses" is complete. Etc.
Your ontology appears to suffer the frame
problem, as does all linear-sequential logic.
Regards,
Charlie (cmoe...@aol.com)
Jorn Barger
> Your "makes/uses/repairs" and the "alive/wounded/dead" list, I take it,
> are visited one at a time as in frames (standard programming method).
Standard design-artifact. Fractal-thickets are way beyond frames, but I
don't think we have to get into that for these examples.
> There is no evidence of communication between frames, so the "uses"
> can not know if "makes" is complete. The "repairs" can not know the
> "uses" is complete. Etc.
In the ontology itself, tries-to-use-before-make-is-done has its own
node. The 'usual' interpretation is that the make will precede the use.
If you're parsing real text, and you find an instance of 'use', you
assume it's a usual-use until the facts bring you up short, and then you
explore non-usual paths.
> Your ontology appears to suffer the frame
> problem, as does all linear-sequential logic.
This is story-logic.
Charles Moeller
>
>Charles Moeller <cmoe...@aol.comnulspam> wrote:
I meant to say that I appreciate the fact that
you assign cause in your ontology, a necessary
act if we are going to get any more progress.
Regards,
Charlie
Joshua Tauberer
After looking at "Re: What's the difference between a "SEMANTIC NETWORK" and
an "ONTOLOGY"?..." it seems that in terms of what we're talking about, the
two are interchangable. ("a type hierarchy, specifying classes and their
subsumption relationships") Is this true?
Anyway, I think I disagree that an ontology wouldn't help much. :)
"Elias Ponvert" <e...@uts.cc.utexas.edu> wrote in message
news:9srrbr$nr6$1...@geraldo.cc.utexas.edu...
>> 1 Ambiguity
> Semantic ambiguity is relatively simple -- a word, has two (or more)
distinct
> meanings.
I think that an ontology would be able to rule out in many (some?) cases
what words fit into what places..
> "John was at the bank".
This brings up the need to be able to understand context. If there's no
context, then the sentence can't be parsed by anyone/anything.
> 'John depositted his money at the bank'
I'm not sure that money can be deposited in the "dump a lot of soil" sense,
so an ontology might rule that sense out. Assuming that you can deposit
money in a river bank, then.... From a human point of view, unless someone
was writing ironically, no one would say that unless they meant in the
normal sense. "deposit money [at a bank]" is a frequent usage, so maybe
that should be encoded in the ontology (a more likely state of existence),
and that more likely interpretation would be prefered over the other.
> Grammatical ambiguity
> 'The horse ran past the barn fell'
I don't understand it myself (even with your explanation), so I won't hold
an NLP to that standard. :)
> structural ambiguity
> 'John can sing and dance the tango'
The interpretation of singing the tango would be ruled out because a tango
can't be sung. Rather, one sings a song, and since a tango is not a song,
it can't be sung. But, in cases where there is complete ambiguity, then
without context nothing could possibly choose which is correct. So, I think
this reduces to understanding context.... or asking the speaker to rephrase.
> 'Every man loves a woman'
Well... this is a stretch, but.... In order for one to love someone, you
have to know someone. So, this statement is true only if every man knows a
woman. It's possible to encode (don't know how) that it is not possible for
every man to know the same woman, so 'a woman' must be interpreted to mean
'there is a corresponding woman y.' I think a sufficiently "good" ontology
could handle that. (I think this is why I myself understand that sentence
with that meaning.)
So, I'm sure there are better examples, but I think they would reduce to the
above cases of needing context or a rephrase.
> Hope this all helps.
Well, I haven't gotten to reading your explanation of the Kripke structures
yet since my first skim of it left me dazzled. But, thanks, I appreciate
your insight.
My conclusion here is that all of the tractable problems (ie not structural
ambiguity which I think can be impossible to solve, and not a few instances
of lexical ambiguity) can be reduced to needing a sufficiently large
ontology and needing contextual clues.
Looking forward to seeing where I went wrong. :-)
Holger Schauer
> Holger Schauer <Holger....@gmx.de> wrote:
>> On Sun, 11 Nov 2001, Elias Ponvert wrote:
>>> Before dealing with the comments below, there are a couple good
>>> ontologies of English. The OED and Webster's (I forget which
>>> Webster's -- one of the big ones) are both good.
>> I don't think these count as ontologies, at least not in my
>> opionion. I would like to see some formal principles to hold
> fashioned dictionaries have a lot of useful generalizations about
> English syntax and semantics formal linguists tend to overlook.
I am not even convinced of this. Could you perhaps give an example to
show what you're making reference to? I could imagine that you're
thinking of representing the semantics of the english noun "bank" via
some symbol 'house (Montague grammar, anybody?), which is of course
ridiculous. But still people have tried to incorporate such problems
into pretty rigid and formal approaches to semantics (see e.g. the
work by Kohlhaase etal. in the proceedings of ACL'2000 -- I don't know
whether it's available on the net somewhere).
> Absolutely. I simply meant to say, yes, a solid ontology is a
> probably necessary part of a good interpretation system. But a
> semantic network alone can't do all your work for you. For example,
> an inference engine shoould be present as well.
Yes, sure. Otherwise your ontology won't be of much use, albeit people
nowadays seem to employ WordNet (which does not provide any inference
engine) for quite a lot of stuff.
> So: I fully agree with Holger on (a) and (b). I object merely to the
> proposition that ontologies alone can make the problems I mentioned
> disappear.
Thanks for the clarification,
Holger
--
--- http://www.coling.uni-freiburg.de/~schauer/ ---
"Linux stands for the 'Linguistic Institute of Nocturnal
Undertakers on XEmacs'."
-- Anonzymous Coward on slashdot
Holger Schauer
> Holger Schauer <Holger....@gmx.de> wrote:
>> On Sun, 11 Nov 2001, Elias Ponvert wrote:
>>> Before dealing with the comments below, there are a couple good
>>> ontologies of English. The OED and Webster's (I forget which
>>> Webster's -- one of the big ones) are both good.
>> I don't think these count as ontologies, at least not in my
>> opionion. I would like to see some formal principles to hold
> fashioned dictionaries have a lot of useful generalizations about
> English syntax and semantics formal linguists tend to overlook.
I am not even convinced of this. Could you perhaps give an example to
show what you're making reference to? I could imagine that you're
thinking of representing the semantics of the english noun "bank" via
some symbol 'bank (Montague grammar, anybody?), which is of course
ridiculous. But still people have tried to incorporate such problems
into pretty rigid and formal approaches to semantics (see e.g. the
work by Kohlhaase etal. in the proceedings of ACL'2000 -- I don't know
whether it's available on the net somewhere).
> Absolutely. I simply meant to say, yes, a solid ontology is a
> probably necessary part of a good interpretation system. But a
> semantic network alone can't do all your work for you. For example,
> an inference engine shoould be present as well.
Yes, sure. Otherwise your ontology won't be of much use, albeit people
nowadays seem to employ WordNet (which does not provide any inference
engine) for quite a lot of stuff.
> So: I fully agree with Holger on (a) and (b). I object merely to the
> proposition that ontologies alone can make the problems I mentioned
> disappear.
Thanks for the clarification,