The original WSD code is no longer used as such. However, I think it
contains a very important -- fundamental, even -- idea within it that has
been long ignored, and needs to be revived and modernized.   And this is
why those subdirectories have not been nuked into oblivion -- they are
kind-of a reminder to "modernize this stuff".

Let me briefly explain the original idea. It was to associate every word
with a vector of word-senses (exactly like Webster's dictionary or any
natural language dictionary).  Given the words in a sentence, get the
associated vector. The claim is that the correct meaning for the sentence
involves only those word-senses that are the most similar.  The canonical
example is "The church bells rang on Sunday" -- the "ring" could be "the
ring on your finger" or "the sound" but since there are no hands or fingers
or other body-parts in that sentence, we conclude it's the sound.   The
"bell" could be the "bell curve" or "the doorbell" or something else. But
bell-curves do not make sounds; door-bells and other bells do (oddly
enough, the kind of sound they make is called "ringing"... Hmm ..
interesting...). A "church" could be the abstract political structure in
Rome, or it could be a reference to a physical, architectural building.
Now, abstract political structures do not contain any "bells" (they do
contain "dog-whistles", but that's something else), so, here, "church"
probably means "the building".  Sunday, well, Sunday could be the 7th day
of the week, it could be the radio advertisement "SUNDAY AT THE US31
SPEEDWAY THE TOP NITRO-FUELED MUNSTERS DRAG-U-LA WILL BE TAKING ON RAY
FARHNER'S BOOTHILL EXPRESS..." (
https://www.museumofamericanspeed.com/boothillexpress.html) but it's
probably more about the religious holiday that comes around once a week. We
conclude this because "churches" have something to do with religion...

The desired generalization of WSD would be the creation of a "vector of
factoids" for each word, and then one of more algorithms to compare
factoids, to discover the most likely ones.  Comparison could be as trivial
as counting the words-in-common in the definitions (e.g. the definition of
"bell" probably has the word "ring" in it) but you could also do more
complex reasoning (e.g. ringing is-a kind-of sound and bells perform
sounds, therefore ...)

Now, natural-language neural-nets do something like the above, but the
actual "reasoning" that they perform is tightly mashed up and compressed
into opaque weight-vectors of floats. We don't know what those
weight-vectors are doing. (it's kind-of like trying to figure out what a
compressed file is, without uncompressing it. It looks like random bytes.
Neural nets are like "homomorphic computing" - they work directly with the
compressed form.) So here, instead, the goal is to "uncompress" those
weight-vectors, and work with a sparse representation, where we have
"human-readable" access to the factoids attached to each word.  The sparse
representation should be more accurate (as, famously, compression loses
image quality/audio quality ....)

I will cut-n-paste this email into the README ...

--Linas


On Tue, Aug 25, 2020 at 1:27 PM [email protected] <[email protected]>
wrote:

> Hi,
>
> I was wondering how much of this Howto still applies.
> https://github.com/opencog/opencog/blob/master/opencog/nlp/wsd/README-howto
> Is nlp/wsd still part of nlp-parse or anything?
>
> Thank you
> Mani
>
> --
> You received this message because you are subscribed to the Google Groups
> "opencog" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to [email protected].
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/opencog/01e159a2-6325-4154-b063-e0869900fcffn%40googlegroups.com
> <https://groups.google.com/d/msgid/opencog/01e159a2-6325-4154-b063-e0869900fcffn%40googlegroups.com?utm_medium=email&utm_source=footer>
> .
>


-- 
Verbogeny is one of the pleasurettes of a creatific thinkerizer.
        --Peter da Silva

-- 
You received this message because you are subscribed to the Google Groups 
"opencog" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To view this discussion on the web visit 
https://groups.google.com/d/msgid/opencog/CAHrUA35OmVZBp7P7uUSs_3Z_JKiAAUVxrgBDAhaDZZ-EaPeCyA%40mail.gmail.com.

Reply via email to