Chinese Room question

[Darth Ruinus]

Ok, I'm taking Introduction to Philosophy right now, and we were going over the mind and different ideas about it. The professor went over the computational theory of the mind, and brought up the chinese room as an idea that refutes the computational theory. Now, we didn't go over this in detail, mainly because it is a simply introductory class and because we only had 30 minutes left, but there is something that I am hoping someone where can help me on.

Now, as I understand it, the argument goes that since the man in the room doesn't understand Chinese then there can't be a mind in the room. But, to me that seems silly, as there is no one part in my brain that understands everything, all my thoughts and actions are the work of various parts of my brain working together. Also, there is no other intelligence (the man) in my brain which receives input and gives out the proper output. Also, I am sure the analogy is not exactly similar, as brains and computers work far faster than it would take a man to look up the chinese words and then look up the correct chinese response, but I suppose if you just speed up the entire room it would look similar. When I brought this up to the professor he said something about syntax and semantics and that the room has no "understanding" of chinese because the man has no understanding. But again, I still think the man is being placed to much importance.

Anyways, as I left all my notes somewhere else, and I am absolutely having no luck understanding the damn argument (but I don't want to give up on it) I ask anyone here who could hopefully explain it.

[Kuroneko]

You're having difficulty understanding the argument primarily because the argument is flawed. The professor's reply was probably something that's equivalent to the following thought experiment: imagining the person in the room as simply memorizing his entire table of syntax rules, thus enabling him to function the same way even outside the room while still being unable to actually translate any Chinese into his or her native language. Thus, there would be a gap between syntax and semantics. However, this simply inverts the situation into a fallacy of division rather than a fallacy of composition.

It doesn't mean that the computational thesis is true, of course, but the argument is not a refutation by any means; in fact, both situations are perfectly consistent--for according to the theory, even the person out of the room would act like is a hardware system simulating another mind.

[Darth Ruinus]

However, this simply inverts the situation into a fallacy of division rather than a fallacy of composition.

Ok. So let me see if I got this:

Fallacy of Division:
- The house is a mind with an understanding of Chinese
- The man is a part of the house
- The man must therefore understand Chinese

Fallacy of Composition
-The man doesn't understand Chinese
-Therefor the entire mind doesn't understand Chinese

So the problem in the first one is that it assumes that one part must be able to accomplish the same thing as the entire system put together, which is ridiculous, otherwise there wouldnt be any other parts besides the part that can accomplish the entire task. The problem in the second one is that it also assumes that the total end product is flawed because one single part is flawed.

This about right?

It doesn't mean that the computational thesis is true, of course, but the argument is not a refutation by any means;

Oh of course. I don't mean to say that simply because this particular argument is wrong then the entire theory is right, but as the professor taught it, and as I understand it, the computational theory of mind is widely accepted, and actually has more scientific backing than other theories of the mind and brain?

in fact, both situations are perfectly consistent--for according to the theory, even the person out of the room would act like is a hardware system simulating another mind.

You sort of lost me here. I understand how the other person would be another mind, but how could both situations (I assume you mean the aforementioned fallacies) be consistent?

[Kuroneko]

So the problem in the first one is that it assumes that one part must be able to accomplish the same thing as the entire system put together, which is ridiculous, ... .

Yes.

The problem in the second one is that it also assumes that the total end product is flawed because one single part is flawed.

In this case, not quite. The fallacy is going from the person's lack of understanding to the claim that there is no part that understands.

I understand how the other person would be another mind, but how could both situations (I assume you mean the aforementioned fallacies) be consistent?

In the second case, for the argument to work, there must be an inference from the person's lack of understanding of Chinese semantics (since he or she consciously operates only at the level of syntax rules) to the claim that the subprocess of the person't thinking does not have such understanding. This is prima facie very reasonable and it may look odd to claim otherwise, as used to identifying minds and brains as we are, but all we need is for the contrary claim to be consistent with the computational thesis, and it is.

Perhaps the following hypothetical might help. Imagine a simulation of an entire world with simulated humans. Since we're upping the person's mental abilities anyway, let's imagine them to the far ludicrous level, capable of running just such a simulation entirely in his or her head. Supposing further that the person fails to realize just what he or she is doing by carrying out the computations, this fact would not take anything away from the simulated world. That's why I called the person in the inverted case "hardware". A lack of understanding on that level is of absolutely no consequence to the computational thesis; it holds or fails independently.

... and as I understand it, the computational theory of mind is widely accepted, and actually has more scientific backing than other theories of the mind and brain?

That's my impression as well, but I really can't say just how wide the acceptance extends.

[Ariphaos]

In my opinion it's far more interesting to explore the seemingly infinitely vast ways in which the Chinese room hypothesis is wrong. The very premise - that programs can't or don't assign semantic meaning of some sort to text is disproved every single time someone fires up their grammar checker.

[Bellator]

I admit, it's been many years since I heard this arugment, but can someone explain to me what precicely is wrong with it?

As I understand it (might mangle it a bit, doing this from memory).
- A computer program (software) simply follows a set basic if A, then B, else C, etc instructions. It contiuously executes simple instructions line by line (depending on input, status, etc).
- This would be analagous to a person inside a room with a big instruction book, simple executing each instruction line by line, depending on input, etc.
- If the particular program is designed to hold a conversation in Chinese, and the instruction book for the man inside the room also contains instructions for holding a conversation in Chinese. The man does not speak Chinese himself, he simply executes the instructions in the book. The instructions (or the programming) is sufficient to convince a "real" chinese speaker (like a Turing test).
- The man inside the room still doesn't understand chinese, he's just following instructions (like a software program). The room "itself" can't very well understand anything, since besides the man inside there is nothing there. There are no other parts. The software program operates in a similar way as the man in the room does, and therefor can't understand chinese either.

To reason your way out of it, I can think of only two possibilities: 1. people don't understand <chinese/english/math/anything> either, or 2. the man + his instruction = understanding. Neither seems too satisfying. Where do I go wrong here?

The very premise - that programs can't or don't assign semantic meaning of some sort to text is disproved every single time someone fires up their grammar checker.

I don't think that is entirely accurate. If you look at the actual lines of code, it still (exagurating) a basic set of if, then, else instructions. The "meaning" assigned only comes from (I'm exagurating a little) matching a list of possible conditions vs what words are used in the text. If conditions X, Y, Z are met, then verb 2 should be past tense. Or something like that. Is that assigning meaning to it similar to the way people do?

[Kuroneko]

According to functionalism, which the computational thesis adopts to at least some degree, "you" are not your brain, but rather the product of your brain operating in a certain manner. In other words, the bearer of understanding is the process of a brain operation, rather than the brain itself.

The person's mind acts as the "brain" for the process of this Chinese exchange, but the fact that he or she does not recognize what's going on is of no import, since understanding is in the process, not the "hardware". Of course, one can simply deny functionalism or CT as pertaining to minds, but the alternatives invariably carry a whole lot more metaphysical baggage.

[Starglider]

The Chinese Room argument is of course the exact type of bullshit that makes me want to dropkick the philosphers who somehow wander into AI forums and act like we should hang on their every word. It is a natural product of completely ridiculous assumptions about what a mind is, derriving from nothing more than the broken-but-cheap-and-workable approximations evolution stumbled upon when creating the human brain. But that's not why I'm here...

The very premise - that programs can't or don't assign semantic meaning of some sort to text is disproved every single time someone fires up their grammar checker.

The word 'semantics' is a bit slippery but 99% of compsci people would say that grammar checkers don't process semantics. Grammar checking is a syntactic operation; verification that the word ordering conforms to certain patterns, without any consideration of what those patterns model or represent. Syntax and semantics are distinct and (for most purposes) exclusive properties of languages.

[Ariphaos]

The word 'semantics' is a bit slippery but 99% of compsci people would say that grammar checkers don't process semantics. Grammar checking is a syntactic operation; verification that the word ordering conforms to certain patterns, without any consideration of what those patterns model or represent. Syntax and semantics are distinct and (for most purposes) exclusive properties of languages.

I should explain better, the point is better stated: a grammar checker assigns more to a given word than its pure 'translation' from English to binary. It uses an annotated dictionary, looks for subject-verb order, tests for voice, and so on, which is a great deal more than just checking that the individual words are valid.

You could actually syntactically go quite a ways with this, to the point where the checker observes the gender of given nouns and enforces the necessary structure later on in a document. Here then we have a situation where it may pick out a person, and it uses identifiable pronouns to track what references this person, making sure that gender references are specific, flagging ambiguities, tracking plurals, and so on. I don't see any absolute need to use a black box AI like a neural net for this.

Once you have a suitably perfect understanding of the grammar, any translation automatically conveys meaning.

[Starglider]

I should explain better, the point is better stated: a grammar checker assigns more to a given word than its pure 'translation' from English to binary. It uses an annotated dictionary, looks for subject-verb order, tests for voice, and so on, which is a great deal more than just checking that the individual words are valid.

Structure is not meaning. Meaning associates the symbolic representation with some external refferent (real or hypothetical). AI programs that attempt to process the semantics of text usually create an object-relation model in memory, that serves as the basis for logical inference that can draw meaningful conclusions. A grammar checker doesn't do any of that, it will cheerfully pass meaningless text as long as it is structurally correct.

You could actually syntactically go quite a ways with this, to the point where the checker observes the gender of given nouns and enforces the necessary structure later on in a document.

This still isn't semantics (if you don't believe me, just look at how arbitrary noun gender assignment is in German ).

I don't see any absolute need to use a black box AI like a neural net for this.

Why on earth would you? As you say, it is (relatively) easily implemented by a set of logical rules.

Once you have a suitably perfect understanding of the grammar, any translation automatically conveys meaning.

I don't know what you're trying to say here, but note that it is possible to do even machine translation using stastical methods with no modelling of semantics at all.

[Ariphaos]

I wasn't saying it was semantics in that post, however, it is more than the symbols, which is what the CR hypothesis assumes to be true from the outset.

[petesampras]

Structure is not meaning. Meaning associates the symbolic representation with some external refferent (real or hypothetical). AI programs that attempt to process the semantics of text usually create an object-relation model in memory, that serves as the basis for logical inference that can draw meaningful conclusions. A grammar checker doesn't do any of that, it will cheerfully pass meaningless text as long as it is structurally correct.

I've never been full convinced that the distinction between syntax and semantics is as clear cut as this. A grammar checker might reject statements like "The food cat ate the fast", but except statements like "The fast cat ate the food". That is telling you something about the world. You could imagine a simpler set of English grammar rules which allow both statements, and then rely on semantic analysis to reject the former statement. One could also imagine a more complex set of grammar rules, than we currently have for English, which reject via syntax statements which currently require semantic analysis. The implementation of checking syntax or semantic consistancy will often fundamentally work out to be the same kind of task. It is somewhat arbitrary where you place the boundary.

[Wyrm]

I've never been full convinced that the distinction between syntax and semantics is as clear cut as this. A grammar checker might reject statements like "The food cat ate the fast", but except statements like "The fast cat ate the food". That is telling you something about the world.

You seem to think a utility called a "grammar checker" will always accept or reject a sentence based solely on whether it is grammatically well-formed. Not necessarily. It can reject or accept a sentence for reasons other than simple grammar. It turns out that both of the sentences you cite are equally well-formed in English. I won't bore you with a breakdown of the syntax trees, but there is nothing about the first sentence ("The food cat ate the fast.") that violates the syntax of English.

If the grammar checker marks the first as invalid for any reason, then it is not due to any grammar considerations. The dictionary may have 'fast' incorrectly marked (it may also be a noun or a verb), or it may not recognize as valid the production NP -> 'The' 'food' 'cat', even though it is a valid production. These are errors in design, not anything to do with semantics being part of grammar.

You could imagine a simpler set of English grammar rules which allow both statements,

English grammar rules do allow both sentences. Period.

and then rely on semantic analysis to reject the former statement. One could also imagine a more complex set of grammar rules, than we currently have for English, which reject via syntax statements which currently require semantic analysis.

The former statement can only be rejected on semantics grounds: we don't know what a 'food cat' is, and eating a state of an extended period of not eating is absurd. The statement is syntactically well-formed, because I know that it states an absurdity, and is not simply gibberish. "Green the furiously ideas colorless sleep," is an example of a syntactically ill-formed sentence, and is rejected on different grounds than "The colorless green ideas sleep furiously." The latter states an absurdity of an abstract concept that is both a certain color and devoid of all color, and performing an action furiously, an action that is by its nature sedate. The former is just a bunch of random words unconnected by any English syntactic structure. Grammar only decides the well-formedness of a sentence, not whether it makes sense. Indeed, we can only decide the sense of a well-formed sentence.

The implementation of checking syntax or semantic consistancy will often fundamentally work out to be the same kind of task. It is somewhat arbitrary where you place the boundary.

Funny, I've never had problems drawing a line between the two. Then again, I've taken linguistics. Although the specific form of the syntax tree you construct with the grammar rules affects they way you analyze the semantics, the two steps are quite distinct and easily separated from each other.

[petesampras]

You seem to think a utility called a "grammar checker" will always accept or reject a sentence based solely on whether it is grammatically well-formed. Not necessarily. It can reject or accept a sentence for reasons other than simple grammar. It turns out that both of the sentences you cite are equally well-formed in English. I won't bore you with a breakdown of the syntax trees, but there is nothing about the first sentence ("The food cat ate the fast.") that violates the syntax of English.

I would have thought that the sentence "the food cat ate the fast" was grammatically incorrect due to the fact that 'fast' is an adjective, and it thus breaks the rules of constructing sentences? It's been a long time since I had to study English grammar and I've never studied linguistics. I did have to take compiler design which involved a limited study of context free grammars. My understanding of a general grammar is that you have a set of symbol types and a set if rules for how those symbol types can follow each other. Is that incorrect?

The former statement can only be rejected on semantics grounds: we don't know what a 'food cat' is, and eating a state of an extended period of not eating is absurd. The statement is syntactically well-formed, because I know that it states an absurdity, and is not simply gibberish. "Green the furiously ideas colorless sleep," is an example of a syntactically ill-formed sentence, and is rejected on different grounds than "The colorless green ideas sleep furiously." The latter states an absurdity of an abstract concept that is both a certain color and devoid of all color, and performing an action furiously, an action that is by its nature sedate. The former is just a bunch of random words unconnected by any English syntactic structure. Grammar only decides the well-formedness of a sentence, not whether it makes sense. Indeed, we can only decide the sense of a well-formed sentence.

Isn't this a matter of degree though? The sentence "Green the furiously ideas colorless sleep" violates a great many rules of grammar. It thus reads as utterly incomprehensible. The sentence "the food cat ate the fast" has an adjective in the wrong place but a more sound overall structure. Hence we see it as describing an absurdity. It is still breaking rules, however, and hence in the strictist sense is not a well-formed sentence. I am intepreting a well-formed sentence to mean one in which all the symbols/words are following rules. Is there not a rule that forbids having an adjective hanging like that?

Funny, I've never had problems drawing a line between the two. Then again, I've taken linguistics. Although the specific form of the syntax tree you construct with the grammar rules affects they way you analyze the semantics, the two steps are quite distinct and easily separated from each other.

Instead of the highly complex grammar and semantics of English I'll use a simpler example for what I am trying to say.

Imagine a language used by a simple AI to describe a very simple world.

This language can have two types of words, actions and objects.

actions are {push, pull, read, memorise}
objects are {book, screen, brick, pebble, stone}

The only sentences allowed are of the form -> 'action word'.

the sentence "screen brick" is grammatically invalid, the sentence "read brick" is semantically invalid. You would need semantic analysis in the AI to recognise that the latter is nonsense.


Now imagine a more complex grammar where you have four types of words: alphaactions, betaactions, alphaobjects and betaobjects.

alphaactions are {push, pull} betaactions are {read, memorise}
alphaobjects are {book, screen} betaobjects are {brick, pebble, stone}

allowable sentances are now of the form -> 'alphaaction alphaword' or 'betaaction betaword'

Now the sentence "read brick" gets rejected by grammar checking. Thus, in this very simple example, you can turn aspects of the semantics of a simpler grammar into the structure of a more complex grammar.

[petesampras]

I misunderstood part of your response before replying. The word 'fast' of course can also be a noun, hence a grammar checker would clearly miss it. My example was thus a poor one, but it would have been valid if 'fast' was only an adjective to do something quickly?

The simplified language example I gave hopefully gets across what I am trying to say here.

[Kuroneko]

We often use analyze the meaning of words to guess at the right grammatical function, but it need not be the case for every language--you're right in that one can get rid of objections of that nature by requiring that each word takes only one grammatical role. This isn't true for natural languages, but there is no reason for an artificial language constructed with this requirement to be incapable of conveying meaning if its speakers were conscientious enough.

However, in your example, semantics would at minimum include the interrelations between the words, e.g., that a noun adjunct 'bachelor' implies adjective 'unmarried', and this type of processing is not part of grammar; this is not to say it cannot be done computationally at least in principle, only that it would be a distinct type of computation. I say "at minimum" because depending on how narrowly one wishes to define semantics, one could also require a higher level of representation of the referent for that word, this also being one way to accept the lack of understanding in the Chinese room thought experiment while still denying that it holds any substantial relevance to the computational theory of mind (although not too interesting, since the hypothetical can always be adjusted for that).

[petesampras]

I think that I agree with most of what you are saying, but am not sure about this line -

and this type of processing is not part of grammar

Would you agree that the two example 'languages' I gave could be turned into formal grammars?

If so, is it not the case that the structure of the grammar in the more complex example contains information that was only present in the semantics of the first system?

Therefore semantics can be encoded in grammar?

I wasn't trying to argue that the information ceases to be semantics if it is encoded in the structure of the grammar. Merely that grammar can contain information about semantics and semantic processing can be incorporated into grammar checking if you use a more complex grammar structure.

[Wyrm]

You can write down a grammar that could only generate and parse semantically sensible sentences. However, such a grammar would be of limited use. If a structure is ill-formed, you literally cannot relate one part of the ill-formed structure to another. In your alpha/beta grammar, you cannot tell why "read brick" makes no semantical sense, because you cannot relate the meanings of the alpha_action and the beta_word. In such a situation, the only response you can make is "What?"

On the other hand, "read brick" parses into (action word), so we know how the meanings of the two tokens are related: "read" takes a word that has property [+readable], which "brick" doesn't have (normally), so you can say, "I cannot read a brick."

Semantic-coded syntax can only tell when a sentence is both well-formed and semantically correct, but cannot distinguish between the two if it's not both. A separate semantic analysis can tell when a well-formed sentence is absurd, and why it is absurd—because it can relate meanings through a syntax tree.

[Modax]

The problem with the chinese room argument is that it relies on a falty understanding of how the brain works. The brain is made up of hundreds of billions of neurons and trillions of synapses. A modern CPU is made up of hundreds of millions of transistors, so there is still several orders of magnitude diference in complexity, but the metaphor is still completely valid. Neither a neuron nor a transistor can do anything is isolation, but many neurons together form a network, just as many transistors are needed to form a logic gate. Neurons and transistors even work in roughly the same way. Both are simple on/off switches. The difference is that neurons run at around 100 Hz, whereas transistors switch on and off billions of times per second.

[Kuroneko]

Ah. I think I see the fundamental reason for the apparent disagreement, and it turns out to be just a simple difference between 'could' and 'should'. That's essentially what Wyrm touched on.
---
If you're saying that given some language L, it can be transformed into some other (perhaps derivative) language L' in which at least some of the information available only at the semantic level in L is available at the grammatical level in L', then I've no reason to disagree. We can even observe this sort of thing between different natural languages, where multiple lexemes in one language can become one lexeme in another, with the semantic variation indicated by the structure of the word instead.

But that doesn't mean that semantics and grammar are not distinguishable; just that the line doesn't quite cut the same way across different languages. One can check that "the slithy toves did gyre and gimble in wabe" as grammatically correct even if one is left to wonder about its meaning. More formally, one can define a qualitative measure of "semantic mixing" of grammar according to how the grammatical categories and rules are tied to particular lexemes. In other words, there is genuinely something much less fundamental about making a category for "readable things" rather than the more abstract "specific things" (proper nouns) or "nominalizations of verbs" or whatnot.

Therefore, if one takes into account that languages evolve, particularly in the addition of new lexemes (grammar rules are slower to change), it becomes very natural, and also most useful, to draw the "line" between grammar in semantics in such a way as to give the least possible "semantic mixing".

P.S. The "read brick" example is somewhat droll in Russian, where "brick" (кирпич) is a common slang term for a book that's practically unreadable (нечитабельная).

[petesampras]

Yes, I think that I was mainly arguing from a 'could' rather than 'should' perspective. Regardless of if it is possible, combining grammar and semantic checking into the same structures and functions would only be of any interest for curiosities sake.

At a purely philosophical level, I wonder if even the purely structural/conjunctive words offer some, allbeit very loose, information about the world that they describe. In other words, the phrase "the slithy toves did gyre and gimble in wabe" is telling us something about the world, it can only refer to a subset of possible situations by the nature of its structure.

From any practical perspective, however, I would agree the arguments presented - that you can draw a meaningful distinction between structural grammar and semantics.

[petesampras]

At a purely philosophical level, I wonder if even the purely structural/conjunctive words offer some, allbeit very loose, information about the world that they describe. In other words, the phrase "the slithy toves did gyre and gimble in wabe" is telling us something about the world, it can only refer to a subset of possible situations by the nature of its structure.

Actually, that is a very poorly thought out paragraph from me. I'm treating the semantic meanings of conjunctive words as if it was part of the grammar. Or something like that. Anyway. What I should have said is I wonder if phrases of the form - "verb pronoun conjunctive pronoun noun" may tell us something about the world, with only the grammatical structure.

Of course, as I said before, this would be pure mental masturbation as a topic, rather than anything of any practical use.

[Wyrm]

Actually, that is a very poorly thought out paragraph from me. I'm treating the semantic meanings of conjunctive words as if it was part of the grammar. Or something like that. Anyway. What I should have said is I wonder if phrases of the form - "verb pronoun conjunctive pronoun noun" may tell us something about the world, with only the grammatical structure.

You can tie semantics to structure, but only at the cost of decreasing the flexibility of the language while increasing the complexity. This is important, because complicated grammars require more brainpower to process, and a less flexible grammar is harder to amend.

Take your alpha/beta grammar:

S -> alpha | beta
alpha -> alpha_action alpha_object
beta -> beta_action beta_object
alpha_action -> 'push' | 'pull'
beta_action -> 'read' | 'memorize'
alpha_object -> 'brick' | 'pebble' | 'stone'
beta_object -> 'book' | 'screen'

Now, is it possible to push a screen, to (say) get it out of the way? Certainly! But that's not a production of our grammar, so we cannot form or analyze these sentences. On the other hand, it is possible for there to be writing on a brick, pebble or stone, so reading and memorization is not confined to beta_objects. Finally, what if there's a new class of words that may be learned, such as maleables? 'squeezing' or 'stretching' of 'play do' or 'silly putty'. These require additional production rules to cover. But we cannot gain any leverage from the fact that alpha's and beta's follow the same form, as they are distinct production rules. We are forced to add, by hand, new production rules such as:

alpha_object -> 'screen'

alpha -> alpha_action beta_object

S -> gamma
gamma -> gamma_action gamma_object
gamma_action -> 'stretch' | 'squeeze'
gamma_object -> 'play do' | 'silly putty'

as well as the the rules used to resolve their meanings. Furthermore, the semantics are realized in grammar rules, so if something goes wrong, you can say that 'pull' does not go with 'silly putty', but you cannot explain why they do not go together; all you know is that the statement is incorrect in some way, failing somewhere here "alpha_action gamma_object".

On the other hand, let's take the simple grammar:

S -> verb noun
: $action is assigned $verb; $theme is assigned $noun.

with the following dictionary entries:

'push' : verb, $theme ~requires~ [+alpha]
'pull' : verb, $theme ~requires~ [+alpha]
'read' : verb, $theme ~requires~ [+beta]
'memorize' : verb, $theme ~requires~ [+beta]

'brick' : noun, [+alpha]
'pebble' : noun, [+alpha]
'stone' : noun, [+alpha]
'book' : noun, [+beta]
'screen' : noun, [+beta]

That way, when you get hit with "push book", you know why the sentence doesn't make sense: the noun 'book' did not have [+alpha], which 'push' requires. To make screens pushable, just add [+alpha] to 'screen'. To add new verbs and nouns, just add the appropriate entries:

'stretch' : verb, $theme ~requires~ [+gamma]
'squeeze' : verb, $theme ~requires~ [+gamma]
'play do' : noun, [+gamma]
'silly putty' : noun, [+gamma]

What you're trying to do is getting a process that is more about how the sentence is structured and use that structure to decide how the semantics fit together. To hamstring it with semantics simply complicates your grammar, cripples it, makes it inflexible, and makes it sap more brainpower. Just about nothing good can come from mixing levels like this.

[petesampras]

I was conceding that argument to you and Kuroneko, though probably didn't make it very clear. Your post did help clarify the issue to me though .

The other point I was trying to make related to the fact that the phrase Kuroneko used "the slithy toves did gyre and gimble in wabe", seems to carry some meaning despite using nonsense words. Hence structure must carry some form of meaning, just by being structure. I was stupidly ignoring the fact that the conjunctive words themselves have semantics.

In the second reply I was trying to allude to the idea that the patterns of words in sentences will contain some information about what they are trying to describe, even if no semantics are known. So - "retic gluhj retic gluhj" is likely to describe something different from "hijo asdig thue jako" because of the repeating pattern. In retrospect, I think this idea is complete bollocks.

[Wyrm]

The other point I was trying to make related to the fact that the phrase Kuroneko used "the slithy toves did gyre and gimble in wabe", seems to carry some meaning despite using nonsense words. Hence structure must carry some form of meaning, just by being structure. I was stupidly ignoring the fact that the conjunctive words themselves have semantics.

Yes, and you were also ignoring the fact that a semantics-in-syntax grammar would be completely boggled by such a sentence; it would refuse to parse because knowing the meanings of all the words is vital for sorting out the syntax in such a grammar. Even if you recognized 'in' and 'the', you would not recognize 'wabe' and could not tell what kind of preposition phrase production to use. Because you do realize that 'wabe' is a noun, even though you do not know its meaning, is proof that the syntax does not encode semantics.

Hope we helped.