Hi Ivan,
On Mon, Mar 7, 2022 at 6:49 AM Ivan V. <[email protected]> wrote:
> Linas,
>
> I'm giving it a thought...
>
> For words similar to other words, or words near other words, directed
> cyclic graphs with a few thousands unique nodes (spanning infinitely
> because they are cyclic) seem the most promising food for thought. If you
> could make CogServer (or analogous app) to output the node-relations set, I
> could try to show them in a browser. Probably, additional search-for-word
> feature would appy along the already seen browsing by dragging nodes around
> to navigate them.
>
I could set this up. It would be the standard cogserver. I'm paranoid about
being hacked, so I would set it up somewhere "secret" where only you can
get to it. The API would be that you give it a word, and get back a list
of edges attached to that word. Each edge has a weight. At the far end of
the edge is another word. I can give you either the "raw API" as I use it,
or I can write some easy-to-use wrappers for it. The raw API looks like
this:
; Some initial setup, "magic incantations"
(define star-obj (foo bar baz))
; Get a list of all "left"-connected atoms
(define atom-list (star-obj 'left-stars (Word "bicycle")))
; The above returns a scheme list. To get the weight, you would say
(define some-atom (list-ref atom-list 42)) ; get the 42nd atom
; get the third floating point number attached at "whizbang"
(cog-value-ref (cog-get-value some-atom (Predicate "key to whiz bang")) 3)
That's it. The atom-list will be either a list of WordNodes or a list or
WordClassNodes, or other atoms, it depends. We can keep it simple, or
general ...
The value system hangs lists of floating point numbers on atoms. The
(Predicate "key to whiz bang") is just the key to some specific list of
numbers. Not all atoms have all keys; this is all very data-set-specific.
As I said, I can wrap up the above in an even simpler API.
>
> For skip-gram disjunct queries, I'm not completely sure how their output
> data is interrelated. But if you can make the queries output
> parent-children data, the rest would be easy.
>
Disjuncts are just atoms.
Consider the sentence "Ivan threw the ball". The disjunct for the verb
"threw" is conceptually: "threw: ivan- & ball+" which says that Ivan is on
the left and ball is on the right. Written in Atomese, it is this a
rather verbose tree:
(Section (Word "throw")
(ConnectorList
(Connector (Word "Ivan") (ConnectorDir "-"))
(Connector (Word "ball") (ConnectorDir "+")) ))
which is painfully verbose ... the "conceptual" version is much easier to
understand. (and read)
Anyway, the above will have various keys on it, holding the count, the log
probability, assorted other numbers that are "interesting".
The query for the above would be the same as before, except that you would
get back a list of these Sections.
If you are interested & ready to do this, let me know. It would take me a
few days to set it up.
>
> - ivan -
>
> ned, 6. ožu 2022. u 20:30 Linas Vepstas <[email protected]> napisao
> je:
>
>>
>>
>> On Sat, Mar 5, 2022 at 3:08 PM Ivan V. <[email protected]> wrote:
>>
>>>
>>> The logical step would be to prepare a CogServer instance filled with
>>> those millions of atoms, keep it always running, and then query only what
>>> is of the current interest to forward it to a browser.
>>>
>>
>> Yes, exactly.
>>
>>
>>> Anyway, who would browse over millions of atoms all at once? One might
>>> only be interested in some subset of it, and if that subset can be measured
>>> in thousands of atoms,
>>>
>>
>> Or even just hundreds. Or dozens.
>>
>>
>>> Do you have any basic glimpse of a kind of visualization you'd like to
>>> have? And what user interactions would pair it to be successful?
>>>
>>
>> That's the hard question. It's hard to find good answers. I need your
>> help finding good answers. Here are some ideas. For example, given one
>> word, find all the other words "related" to it. Order the list by the
>> strength-of-relationship (and maybe show only the top-20). There are
>> various different ways of defining "relatedness". One is to ask for all
>> words that occur nearby, in "typical" text. So, for example, if you ask
>> about "bicycle", you might get back "bicycle wheel", "bicycle seat", "ride
>> bicycle", "own bicycle". Another might be to ask for "similar" words, you
>> might get back "car", "horse", "bus", "motorcycle". A third query would
>> return skip-gram-like "disjuncts", of the form "ride * bicycle to *" or "*
>> was on * bicycle" or "* travelled by bicycle * on foot" -- stuff like
>> that. These are all fairly low-level relationships between words, and are
>> the kind of datasets I have right now, today.
>>
>> My long-term goal, vision is to create a complex sophisticated network of
>> information. Given that network, how can it be visualized, how can it be
>> queried? A classic answer would be a school-child homework assignment:
>> "write 5 sentences about bicycles". This would be a random walk through
>> the knowledge network, converting that random walk into grammatically
>> correct sentences (we're talking about how to do this in link-grammar, in a
>> different email thread. It's hard.)
>>
>> Is there a way of visualizing this kind of random walk? Showing the local
>> graph of things related to bicycles?
>>
>> So the meta-problem is: given a network of knowledge, how does one
>> interact with it? How does one visualize it? How does one make it do
>> things? If I pluck the word "bicycle" like a guitar string, how can I hear,
>> see the vibrations of the network?
>>
>> -- Linas
>>
>> --
>> 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/CAHrUA37FQRCq0Sj_iiJ5bch-D6Fi4VOkrLS_SMcFy3q_B6DMqw%40mail.gmail.com
>> <https://groups.google.com/d/msgid/opencog/CAHrUA37FQRCq0Sj_iiJ5bch-D6Fi4VOkrLS_SMcFy3q_B6DMqw%40mail.gmail.com?utm_medium=email&utm_source=footer>
>> .
>>
> --
> 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/CAB5%3Dj6U4Qs9EqPfnyKvbSr3xgXbXqDesdbjJN7rogO8YxU2ncw%40mail.gmail.com
> <https://groups.google.com/d/msgid/opencog/CAB5%3Dj6U4Qs9EqPfnyKvbSr3xgXbXqDesdbjJN7rogO8YxU2ncw%40mail.gmail.com?utm_medium=email&utm_source=footer>
> .
>
--
Patrick: Are they laughing at us?
Sponge Bob: No, Patrick, they are laughing next to us.
--
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/CAHrUA359jzC%3Dpg7DRVpguQnGdUhAXwda_%3D0thwmEaY%3DuX2s1FQ%40mail.gmail.com.
