Re: [ontolog-forum] Axiomatic ontology

[Pat Hayes <phayes@xxxxxxx>]

At 3:45 PM -0500 2/1/08, John F. Sowa wrote:

> Pat and Ed,
>
> Nobody "extracts" common sense.  They just use
> it.

"just use"? Surely in order to use it, one must possess it in a usable form. This applies to our artifacts as well as to us. So how are we to get it into these artifacts?

> PH> ... but because [knowledge in text
> is] 'implicit'
> > doesn't mean it can be extracted by any algorithm.
>
> As an example of an alternative approach to natural language,
> note that machine translation systems that attempt to map the
> source language to a "logical form" and then to a target NL
> have
> failed.

The first generation did, yes. But the statistical methods which replaced them also have severe limitations. The newest work in this area combines insights from both schools of thought. BUt in any case, the task of translation from one NL to another is not the same as that of extracting ontological information from NL text.

>  The most widely used MT system is
> SYSTRAN, which began
> as the Georgetown Automatic Translator (GAT), for which research
> was *terminated* in 1963.  The GAT-SYSTRAN method is based on
> long
> lists of word and phrase pairs from the source to the target.
>
> Many recent MT systems apply statistical techniques to bilingual
> corpora to find patterns in the source language that match
> patterns
> in the target.  Although statistical methods are very different
> from
> GAT, they both succeed for similar reasons:  the knowledge
> needed
> for language understanding

For translation, not for comprehension. Purely statistical translation systems have severe limitations on their usefulness, and have to be checked and corrected by human translators in order to be useable.

>  is encoded in the surface patterns,
> and
> much can be done without an intermediate form that resembles
> logic.
>
> EB> ... the very good and effective work on knowledge
> acquisition
>  > from unstructured text that is "guided" by a
> reference ontology.
>  > That approach provides the search engine with a
> "starter ontology"
>  > that defines the principal concepts and relationships in
> the domain
>  > and then extends that knowledge base (by Bayesian analysis)
> by
>  > extracting and interpreting natural language from a broader
> corpus.
>
> I agree that some kind of ontology is important, and the methods I
> have been proposing could benefit from whatever ontology is
> provided.
> Some serious research issues:
>
>   1. How big, how detailed, and how formal should that starter
> be?
>
>   2. What kind of internal representation would be useful?
>
>   3. Is a closed-form "definition" of the symbols
> necessary?

Obviusly not.

>   4. What methods of learning,
> statistics, analogy, pattern matching,
>      *and* logic are appropriate?
>
>   5. What intermediate goals would promote progress, and what
> goals
>      would be more distracting or misleading than
> helpful?
>
> > Find me, anywhere on the Web other than in Cyc, an account
> > of the different senses of 'cover' used in...
>
> > Or talk to a linguist about the many senses of "in"
> used in
> > English (approximately 30, though it is hard to be exact),
> > which require an ontology to be used in order to disambiguate
> them.
>
> What I am questioning is the need for an a priori list of all
> possible interpretations.  That is not the assumption that
> underlies GAT or the statistical methods of MT.
>
> Lexicographers dutifully prepare such lists in unabridged
> dictionaries such as the OED or the MW
> 3rd.

No, they catalog word senses. Related, but not the same task. Cyc is obliged to make many more distinctions than are found in any dictionary, but also ignores many dictionary distinctions based on for example on historical or lexicographical distinctions. No dictionary will make the continuant/occurrent contrast.

>  Those lists are
> useful for some purposes, but 99.44% of the world's population
> get along quite well without them.

I wonder if this is true: but in any case it is irrelevant. People gain knowledge from many sources, and use it in many ways to help them understand language. Most of the common sense knowledge we all have of the everyday physical world is learned without the use of language at all, during the first 5 or 6 years of life; yet it is used during language comprehension in ways we have probably not fully understood yet.

>  People who use dictionaries
> to learn a foreign language would probably be happy with word
> and phrase pairs derived automatically by statistical
> methods.

Im sure, but people are already NL comprehenders of their native tongue.

> > This [data from Google] tells one nothing more than that the
> > words are associated. That is not enough to state a coherent
> > proposition, let alone a coherent piece of ontological
> content.
>
> What the data from Google shows is that the information is
> there.

No, it does not.

> A student who was learning English could
> read that information
> to learn enough about the use of those words to translate any
> of that information to another language, natural or
> artificial.

A human student, even illiterate and with Downs syndrome, already has a huge wealth of common-sense knowledge of the world, far more than is in Cyc. (The example that sparked my own work on naive physics was what one needs to know in order to spread a folded sheet over a bed, an example which is still beyond Cyc.) What humans can do is irrelevant to the discussion, therefore.

> > How does one extract information about relationships from
> free
> > text or word associations? Associations, remember, are
> symmetrical.
>
> "Extracting" is not the goal.  The goal is to read a
> text in a
> natural language in order to solve some
> problem

What is the proposed connection between (1) reading an NL text and (2) solving a problem, if it does not involve somehow learning something from the reading that provides new functionality with the problem-solving? As problem-solving is not itself conducted in natural language, this intermediate knowledge must be somehow represented in a formalism which can be input to the problem-solving machinery. The passage from the NL text to this knowledge is what I mean by 'extraction'.

> that one would
> normally ask an intelligent human being to do.  Many of those
> problems could be solved by variations of pattern matching
> as described above.

What problems can be solved by pattern matching? Cite ANY AI work which justifies this claim.

> > As Ive already pointed out, most common sense is never said
> or
> > written to other people.
>
> Precisely!  And there is no reason why it should be said,
> written,
> or typed into a computer system that processes language.

You entirely miss my point. It is not said because there is no need to say it, because it is presumed. Not because it is of no use. In fact, failures of comprehension can often be directly traced to the failures of such assumptions, and we have all kinds of conversational strategies for correcting such failures by reverting to an explanation of the missing information.

> > I don't know of ANY methods or projects (including analogical
> > structure matching, by the way, which is being used actively
> by
> > dozens of people at NorthWestern, where it was invented)
> which
> > can be said to reliably extract a single nontrivial
> ontological
> > proposition from the entire Web.
>
> Their algorithms take polynomial time.

The original did, but they are using the technique in many Web-based applications and getting interesting results (in, for example, retrieval of 'relevant' news items for a given topic, or 'similar' web pages; but not in anything like NL understanding.)

> It is impossible to process
> the WWW (or even the data on your laptop) with algorithms that
> take polynomial time.
>
> I'm not claiming that analogical methods applied to the WWW will
> solve all the problems of NL understanding tomorrow.

I fail to see how they will solve any problems of NL understanding at any time. After all, the Web is only a large corpus of free text. Such corpora have been available for many years, and have not by themselves produced any NL understanding.

>  But I do
> claim that we will *never* do that if we must first construct
> a knowledge base or ontology that remotely resembles Cyc.

Well, you are entitled to your opinion, but I see no evidence for it. And after all, we have now first constructed such a knowledge base, so apparently it can be done.

Pat

> John
>
>
>  
> _________________________________________________________________
> Message Archives:
> http://ontolog.cim3.net/forum/ontolog-forum/ 
> Subscribe/Config:
> http://ontolog.cim3.net/mailman/listinfo/ontolog-forum/ 
> Unsubscribe: mailto:ontolog-forum-leave@xxxxxxxxxxxxxxxx
> Shared Files: http://ontolog.cim3.net/file/
> Community Wiki: http://ontolog.cim3.net/wiki/
> To Post: mailto:ontolog-forum@xxxxxxxxxxxxxxxx
>

-- 

---------------------------------------------------------------------
IHMC               (850)434 8903 or (650)494 3973   home
40 South Alcaniz St.       (850)202 4416   office
Pensacola                 (850)202 4440   fax
FL 32502                     (850)291 0667    cell

http://www.ihmc.us/users/phayes      phayesAT-SIGNihmc.us

http://www.flickr.com/pathayes/collections

_________________________________________________________________
Message Archives: http://ontolog.cim3.net/forum/ontolog-forum/  
Subscribe/Config: http://ontolog.cim3.net/mailman/listinfo/ontolog-forum/  
Unsubscribe: mailto:ontolog-forum-leave@xxxxxxxxxxxxxxxx
Shared Files: http://ontolog.cim3.net/file/
Community Wiki: http://ontolog.cim3.net/wiki/ 
To Post: mailto:ontolog-forum@xxxxxxxxxxxxxxxx    (01)