[Don Patrick]

It looks like Abiword's source code is right there: https://www.abisource.com/downloads/abiword/3.0.2/source/

Parsey doesn't handle ambiguity at human level though. It only does well on "well formed text", and the score is not quite what it seems: individual dependencies between words includes all the times it gets "and" and "the" right. Its score will be drastically lower when you look at entire sentences: 95% accuracy means 1 out of 20 words will be wrong, and with that it will get one out of every two sentences wrong. It uses statistical probability to handle ambiguity.

No sole programmer needs to label nouns manually, they can just use existing dictionaries. Parsers produce parse trees, and dependency parsers not only tell you which words are nouns, but also which group of words belong together as a noun phrase like "the black cat". Simply put it looks for valid combinations of e.g. determiner-adjective-noun, predetermined sequences. It also marks the main verb so you know to keep that separate. However, which parts of the parse tree you select to output is up to you to program.

there are hints to solve such problems, like `Last night` and `my pajams` parts of the sentence that hint that the shooter wore a pajama

I'm curious, In what way do you consider "last night" a hint? I handle the elephant/shooter in pajamas ambiguity using knowledge of their sizes, among other things. The top score in the Winograd Schema Challenge was 58%, though people have come to claim up to 70% success since, but that is of course while already knowing the questions.

[ivan.moony]

there are hints to solve such problems, like `Last night` and `my pajams` parts of the sentence that hint that the shooter wore a pajama

I'm curious, In what way do you consider "last night" a hint? I handle the elephant/shooter in pajamas ambiguity using knowledge of their sizes, among other things.

Last night? People usually wear pajamas when they sleep at night (besides that they generally wear pajamas, sometimes during day), while elephants are not wearing pajamas neither during day, neither during night, unless someone is trying to make a fun out of a poor elephant, which doesn't depend on day time. So, human wearing pajamas depending on night versus elephant wearing pajamas not depending on night is what may increase a probability that it is a man who wears pajama. At least, that is set in my NN fuzzy mind, I'm not sure about the real probability proof. I could be wrong, of course.

As for NN parsing, I'd bet on NN picking the right combination, considering the whole context, part of which is knowledge of our world. But I'm curious of a crucial thing: is NN capable of hosting what is considered to be loops in algorithms? I believe everything could be described in terms of semantic tables, and I believe NN handles these well, but what if there is a loop inside a semantic table, a recursive reference in a cell that points to a parent semantic table? I wonder if NN is able to describe such a thing. If it can, I'd dare to say that NN language is a Turing complete language, and I'd raise my bets in NN handling correct parsing.

[Edit]
Fixed-point combinators make me believe that NN is Turing complete. Y combinator in particular can be used to represent a recursive functions in languages that don't support recursion at general level.

[Edit2]
But I think that NN is a brute force thing, there could pass centuries until it clues up the right combination. It would be interesting to see how the thing behaves after years or decades of unsupervised learning. A learning curve may raise exponentially, considering our species.

[LOCKSUIT]

Does this do what abiword's link grammar does in showing you which is incorrect? http://www.link.cs.cmu.edu/cgi-bin/link/construct-page-4.cgi#submit
see attachements FIRST...it looks like it may idk !


Else > Does anyone have a Unix system here? I looked at building AbiWord tar file on Windows and it's a long list of steps, hence I'd rather like it if someone could screen-capture a test in AbiWord for me.

I want you to take several sentences I'd write and see what it underlines in green underlining with link grammar turned on.

I seen an image on Wiki showing 2 short sentences but I need to see more examples!

I need    abiword    XD.............the thing in it......

[LOCKSUIT]

Ok now I need to know why, see attachment,
1) why does it return 20 possible linkages for this sentence and not 1 correct linkage? That defeats the goal, I could do that too lol!
2) why does, in my screenshot, the bottom linkage combine so so so many words on the left as 1 big long piece and only at the right end does it BUILD bigger? While the top linkage actually is building with the little atomic parts of the sentence? Is it in actuality building at EVERY + symbol sign? Or only when it goes up a level with the + sign ABOVE the below? That would fatally wound my goal if it thought those first 10 words were 1 part/utility/tool/fact at the smallest level and not "he said" or "he" !

[Don Patrick]

It seems to me you do not understand the basic concept of hierarchy in parse trees. Do some reading.
https://en.wikipedia.org/wiki/Parse_tree

Ivan: I would think that if an elephant were to wear pajamas, it would also do so at night because that's what pajamas are for. If a program already has the knowledge that humans wear pajamas, or that elephants don't, then the time of wearing doesn't add anything in my view. Still, it's interesting that a specification of time could be a hint in other scenarios.

NNs typically do not loop, they just branch from input to output. However, Google set one up to feed the output back to the input in their Deep Dream project and this led to "hallucinations", where initially faint patterns are enhanced. http://www.clowndotpy.com/deepdream
This is not recommendable however, as it leads the machine to see things where there aren't. It's like you have a statistical method that's 90% correct, and then you pass that result through the same NN, then you get 90% of 90% correct, which is 81%, etc, etc. NN loops become less accurate as far as I understand.

[LOCKSUIT]

So parse tree hierarchies works by rules, which evolve around/on the given sentence and doesn't apply for all sentences that contain "we were" in them? For example, two very structurally different sentences contain "we are", and as the parse rules say ah Ss ah ts ah those two make Mv and so, it might see the other sentence very differently and won't always single out "we are" as a smaller atomic part in both sentences? In sentence 1 it may have the smallest part comes out to be "and we are here" while in sentence 2 (separate sentence) the smallest part may come out as "we are next"?

[spydaz]

It seems to me you do not understand the basic concept of hierarchy in parse trees. Do some reading.
https://en.wikipedia.org/wiki/Parse_tree

Ivan: I would think that if an elephant were to wear pajamas, it would also do so at night because that's what pajamas are for. If a program already has the knowledge that humans wear pajamas, or that elephants don't, then the time of wearing doesn't add anything in my view. Still, it's interesting that a specification of time could be a hint in other scenarios.

NNs typically do not loop, they just branch from input to output. However, Google set one up to feed the output back to the input in their Deep Dream project and this led to "hallucinations", where initially faint patterns are enhanced. http://www.clowndotpy.com/deepdream
This is not recommendable however, as it leads the machine to see things where there aren't. It's like you have a statistical method that's 90% correct, and then you pass that result through the same NN, then you get 90% of 90% correct, which is 81%, etc, etc. NN loops become less accurate as far as I understand.

I try to think of it as the "Subject/Object" >> is "Action" << "Subject/Object" in this way it matters not if the subject is an elephant and the object is pyjamas.... the first aim is to collect the components and structure.... If an Elephant can or cant wear pyjamas can be handled by a later component which deals with the logic of "Can an elephant wear pyjamas"  as unless both of the "noun" objects/Subject have entanglement information collected....it must be TRUE, until a contradictory entanglement presents itself for probablisitc reasoning to begin... or even another confirmation of the FACT elephant is wearing pyjamas ?....

The important factor is the Truths which can be determined from the sentence ; and the extracted subjects/objects/actions Which enable higher level scripts to process the collected data.... as Parsing a sentence etc ... if all components are not discovered... it may not even be "Saved" as its incomplete record. but it may still be processed to return a response.. or extract some meaning.

[LOCKSUIT]

Post #65 - am I right?

Yes I realized myself that before a sentence can match a 'node', it must first cleans the data, like "alice whom was dieting drove a car and saw bob" or "Last night I shot an elephant in pajamas." needs re-arranging or known facts like elephants/animals don't wear pajamas, they could! but don't, last night and humans wear pajamas matches humans wear pajamas at night. The aim though next is to match a memory like the very sentence itself, the understanding part is already done.

[8pla.net]

Here is what I was thinking...

http://aihax.com/analyze/

[spydaz]

A cat didn't enter a box because it was too big.


Conditional Sentence(Propositional Logic)

Implied IF / BECAUSE

A cat didn't enter a box < Premise> 
(a cat<Subject>) Did not enter (Predicate) (a box<Object>)
      because
           it was too big. <Conclusion>
It)- Subject (was too big) (Predicate)

What was too big, a cat or a box?
(Subject = A cat,It)


Potentially !

The important factor is in the normalising of the sentence as we know, the First named subject, Enables for the Identification of the following Identification as "She/He/They/It" .... these factors in a joined sentence...(complex sentence rules).... for grammatically correct sentences.
When saving the data collected from the sentence ... the data base should not have references to he or she which should be replaced by the named object/subject... usually confirm able as a noun/ noun phrase etc

[Don Patrick]

So parse tree hierarchies works by rules, which evolve around/on the given sentence and doesn't apply for all sentences that contain "we were" in them?

Yes. Sentences with different structures will produce parse trees with different parts. The smallest parts are always single words. You'd probably want the medium parts like main verb (VP) and noun phrases (NP), but I don't believe you've yet mentioned what you need any of this for, at least not in language that I understand.

"A cat didn't enter a box because it was too big."
My program presumes that the cat did not enter because it was intimidated by the large size of the box. It would be a different matter if the cat could not enter the box, as opposed to did not. Anyway, no need to explain to me why syntax parsing is an important preprocessing phase, but sometimes the syntax is ambiguous itself and requires disambiguation during parsing.

[LOCKSUIT]

Thanks for confirming that Don! So wait, it shows the structural BIG parts, the smallest parts (single words), but not some of them like "blah blah"? Ex. it outputs "blah blah blah blah blah" and of course we know the single words "blah" but doesn't show the second next most smallest parts "blah blah"?

I need this because I need my AI to know the smallesssst parts to even know what a "blah blah" means, then it needs to know what "blah blah" and "blah blah" mean to know what "blah blah blah blah" means, it can't just understand one big 16 word long sentence! And the bigger stuff is made out of the smaller stuff and doesn't store "I am" twice. It has to know the fundamental features of the real world, like water drops, gravity pulls, I am, not "swimming I" or "but water" that can be found in bigger sentences! Ex. After swimming I left the room. I could enter them in manually but that's laborious. Right now I have 4  valid ways to do the same thing: manual, link grammar (yep still 'works'), a backprop method, or a universal discovery method.

Yeah I'm using Parsey McParseFace hopefully before it reads sentences. Once text is re-arranged or clarified that I was wearing pajamas it'll answer back knowingly, it can then parse, and find the real parts of the sentences like "I am" but not "swimming I".

[spydaz]

"A cat didn't enter a box because it was too big."
My program presumes that the cat did not enter because it was intimidated by the large size of the box. It would be a different matter if the cat could not enter the box, as opposed to did not. Anyway, no need to explain to me why syntax parsing is an important preprocessing phase, but sometimes the syntax is ambiguous itself and requires disambiguation during parsing.

I am currently researching the usages of the Could/Would/Should Relationships for gaining greater sentence understanding/meaning.... (modals)... i'm glad the phrase did not contain that added complexity...as of yet i have no handler for those sentences implemented yet either...

[8pla.net]

Here is what I was thinking...

http://aihax.com/analyze/

Results of analysis:

A_cat_didn't_enter_a_box_because_it = too_big.

What would be the next logical step in the algorithm?

Results of analysis:

What_would = the_next_logical_step_in_the_algorithm?

[spydaz]

Here is what I was thinking...

http://aihax.com/analyze/

Results of analysis:

A_cat_didn't_enter_a_box_because_it = too_big.

What would be the next logical step in the algorithm?

Results of analysis:

What_would = the_next_logical_step_in_the_algorithm?

(Not all of the algorithms return perfect information (hence the need for multiple algorithms for different functions)) ... Incomplete..... or not valid for current interrogation? (a Check would be required to test the captured results to validate the algorithm is capturing valid data)

I was thinking about this also ; as there is still information not taken from the sentence; Probably as relational Subject / Object inferences.....

A_cat_didn't_enter_a_box_because_it = too_big.

A cat <is bigger than> A box....................

The problem with information extraction is to understand if the knowledge is temporal knowledge or long-term knowledge also another step.... because the predicate that the cat is "bigger" than the box is temporal knowledge and Should not be saved as a universal truth or long-term knowledge. Its hard to identify that this is a temporal statement. and yet by the Loose nature of the statement IE: He said / She said / It said clearly denotes that this knowledge is probably "conversational in nature" Specific to the HE/SHE... OR Event/Scenario also temporal.... (after reading i would hope that a sub function captured the fact(long Term) The Cat is Bigger than "Some" boxes) <<<< always the case (we notice even more entanglements) <<< Isnt English so complicated?....Very musical (great for composition)

Information becomes universal if it holds true for all possibilities for the even/Scenario Also hard to recognise. (from the sentence or paragraph structure)... yet again it also goes towards recalling the event or scenario? it becomes a case based reasoning argument. now sessions become "cases"to be stored as atomic argument related to their temporal moment  in contrary to a factual statement to be incorporated in to the Data warehouse/ knowledge base. (other functions may store Full uniformed data, whereas another set of functions serve as temporal functions, such as temporary knowledge trees or Queries)

Also the posed question ; is regarding the "case" and not stored factual information ; as cases can be "fictitious scenarios"....

There is often always more information to be retrieved from a sentence... Now when executing a script the time it takes repeats each new function execute speed. now its time for some Parallel Programming. the same sentence may be parsed many times by many different functions each retriving different structures of information (although some items may be the same or repeated)