Aug 25, 2003
And yes, it has to do with RSS, but only indirectly. It rather has more to do with the Semantic Web as a whole. Here's how Phil Windley put it:
When I seehow do I know what it means? As Jon has pointed out, this is where things get tricky. When we use the word "means" we usually think of some rigorous, complete definition. Its fairly easy to see how namespaces might provide us with more metadata and thus increase the information we have available to us about any given XML document. Its much harder to imagine that machines will be able to divine the meaning of the document no matter how much metadata you include.
I have been witness to, either directly or indirectly, a fair number of metadata initiatives recently. Many of these, like CanCore, are intent on fixing the meaning of the tags. They want to tell you, for example, what the range of allowable values is, what it means to select this option rather than another. And so we get to the heart of what we really mean by "typical age range" and other things. Right?
Well, not so. Confusion abounds, ranging from the minor (is it proper to separate the names in the 'creator' field with a semi-colon or a comma?) to the major (does the term 'creator' refer to the author of original content or to the finished product; or does creator denote the institution rather than the author when the work is work for hire) to the really obscure (does the 'URI' point to the resource itself, or can it only ever point to a representation (eg., an HTML version, a WAP version) of a resource?).
Some of the web's heaviest weights have been pondering this sort of issue. Tim Bray writes,
At the end of the day, markup is just a bunch of labels. We should be grateful that XML makes them (somewhat) human-readable and internationalized, and try to write down what we want them to mean as clearly as and cleanly as we can, with a view to the needs of the downstream implementors and users. But we shouldn't try to kid ourselves that meaning is inherent in those pointy brackets, and we really shouldn't pretend that namespaces make a damn bit of difference.He asks, where does metadata come from and, in another post, discusses the deep confusion web architects have when considering some of the fundamental issues.
Everyone agrees that when you get confused about what's being identified, this is a bad thing and makes the Web less useful. As TimBL has said repeatedly: a resource can't be both a person and a picture of a person. Unfortunately, such ambiguity is not a condition that Web software can detect.
Some of the web's problems can be solved through traditional means. Many of the paradoxes that arise, I wrote to him in a letter just prior to my Vancouver trip, are a result of Russell's Paradox - they are the modern-day equivalent of coding the sentence "this sentence is a lie" into metadata, and can therefore be solved by Russell's theory of types. Create a convention, whereby references of the form "about=URI" always refers to metadata, while "
But the problem, though address, is not solved. You cannot enforce such a convention, you cannot make it that everyone will code this way. When faced with a URI, a computer system will always have to dereference it (that is, request the file, open it up, and look inside) in order to determine whether it is metadata, HTML or an image. It's like Wittgenstein says,
I think there cannot be different Types of things! In other words whatever can be symbolized by a simple proper name must belong to one type.The same dispute arises in the field of ontology. A month or two ago I was involved in a dust-up on the Semantic Web mailing list. The question centers around whether the Semantic Web should have one upper level ontology defining basic types and classes. Sounds great... but how would you decide what the entries were?
I wrote,
Disagreements in language can be traced to disagreements in a wide variety of underlying hypothesis, among them including ontology (Quine), causation (Hanson), explanation (van Fraassen), categorization (Lakoff) and meaning (Wittgenstein). These five factors (among others) create a context of discourse, and it is the context of discourse (the 'pragmatics' (Morris)) that changes the meanings of the terms in question, changes the very language of discourse. Because there is no way around this (aside from a regime of ontological authoritarianism) any system of representation must adapt to this, learning, deploying, and teaching new vocabularies, new semantics, as the need arises.
And why do I say this? because if we don't, we get right back to Russell's Paradox again:
By the 'empty node' do you mean 'nothing' or do you mean 'the empty set'? Nothing, as in the absence of a thing, or empty, as in a(n imaginary) container with no members? If we take the empty node, and place it inside a similarly empty node, do we now have the equivalent of one of the original empty nodes, or do we have something that is now a non-empty node?
And I think that the debate over the meaning of metatags, of reference and dereference, of upper level ontologies, resolves to this:
We can have first principles if we wish, but first principles are not matters of discovery, they are matters of agreement, and on some things - even so simple as the characterization of nothing - there may not be agreement, not, at least, if both parties to the discussion understand what follows from such first principles. Experience is like a large jugsaw puzzle, all the pieces scatted on the table in front of us. Some people insist that the only proper way to start is with a corner piece. Others prefer to start at an edge. Others prefer to find pieces of similar colour, or similar shape, and build from within. In jigsaws, this is no problem. In experience, if you start in one place, you get one picture, and if you start in another place, you get another picture. Which picture is 'correct'? There is no way to tell: all we have is experience, and ways of putting it together.
All of which returns me to Wittgenstein.
In the Philosophical Investigations Wittgenstein writes, "For a large class of cases--though not for all--in which we employ the word 'meaning' it can be defined thus: the meaning of a word is its use in the language."
What does this mean?
Knowing the meaning of a word can involve knowing many things: to what objects the word refers (if any), whether it is slang or not, what part of speech it is, whether it carries overtones, and if so what kind they are, and so on. To know all this, or to know enough to get by, is to know the use. And generally knowing the use means knowing the meaning. Philosophical questions about consciousness, for example, then, should be responded to by looking at the various uses we make of the word "consciousness."
To take this a step further,
When a person says something what he or she means depends not only on what is said but also on the context in which it is said. Importance, point, meaning are given by the surroundings. Words, gestures, expressions come alive, as it were, only within a language game, a culture, a form of life. If a picture, say, means something then it means so to somebody. Its meaning is not an objective property of the picture in the way that its size and shape are. The same goes of any mental picture.
This may appear to be abstruse philosophy but it has practical day-to-day implications. I have, for example, been studying French recently. The study of a language, says Don Belliveau (my instructor), quite rightly, is the study of a culture. The clearest indication of this is to look at the French and the English look at the body.
In English, we write, I am sick. But in French, we write, J'ai mal (I have a sickness). In general, in English, the form of expression assumes that the body is something that we are, while in French the assumption is that the body is something we have. Now what would be the 'correct' reading? How do we define the 'self' properly, given the differences in English and in French?There is no correct way. Neither language has a lock on the truth. Nor would an empirical investigation help: we could look at the body all we wanted and advance no further: the body itself does not tell us how to describe it, and the language we use to describe the body is itself permeated with assumptions about the nature of our enquiry. How does one study the 'self'? The same questions arise. When Descartes says, "I think, therefore, I am," he has already acquired a knowledge of I, of think, of am.
Meaning is use. What the term means is determined by how it is used. Do you want to know the meaning of the term
To take all this a step or two further: in other of my writing, i ave been taking about the need to treat complex objects (such as learning objects) as 'words' in a new 'vocabulary' (here and here). When I say this, I mean this quite literally, which means further than when we try to attach meaning to one of these multimedia-words, what I mean is that we need to look at its use.
Right now, people are trying to determine the meaning of, say, a learning object, a priori by means of attaching metatags. The metadata, it is supposed, creates a description of the object, and thereby, creates meaning. How do you know when to use a learning object in your course? You look at what you are trying to say and then find a learning object that matches that description.
Several metadata tags are explicitly about meaning. Tags that have to do with the topic, category or the content of an object, especially, try to fix the object into some sort of semantical space. If you are talking about 'rabbits' then this object will help you. So the theory goes.
But what rule would help us to determine whether a given learning object should or should not be used in the context of a discussion of rabbits? It should be clear that it is not going to be possible to determine this a priori. Should we use all and only those that have the term 'rabbits' in their metadata? This would eliminate those objects depicting 'hares', it would eliminate 'bunnies', and it would eliminate references of rabbits by name (such as 'Peter'). Should we look at the content? Well then this article, because of the preponderance of the word 'rabbit' in its body, will show up in the search.
The only way you can be certain that a word has a given meaning is when the word is used in some context. The only way, therefore, to determine whether a learning object has anything to do with 'rabbits' is to determine that it was actually used in that context. There is no prior determination that will tell you that this object is used correctly, or incorrectly, in that context.
Indeed, this applies to our definition and use of the general class of entities known as 'learning objects'. Rory McGreal characterizes my view as being "anything can be a learning object". It does not, of course, follow that everything is a learning object. rather, my view is that the only way we can know that an object is a learning object is that somebody actually uses it in learning.
In my view, the massive efforts underway to tag, to carefully sort, classify and describe, learning object metadata is misplaced. It is misplaced not because t is wrong or misleading (though that possibility is certainly built in by assumption). It is misplaced because such metadata descriptions can, at best, represent only one point of view of the description and the application of a learning object.
We will not be able to approach the usefulness contained in the promise of learning objects - and of the semantic web more generally - until we get past this idea that we can define it (and in passing, all of human knowledge), a priori. We can't. The very best we can do is establish (through, say, RDF) relations between intended meaning of terms. But at some point, we need to step back and observe how these entities are being used, and to capture that as our definitive metadata.
If you step back, and look at it from a wider view, it is apparent that it could be no other way. The search engines that have been less successful (Yahoo!, DMoz) have tried to categorize and rank a priori, while the most successful (Google) has ranked according to use (as instantiated in actual links).
There is a whole class of description - what I have been calling third party metadata - which aids us in this determination. The full and sole purpose of third party metadata is to capture the use of a learning object. And it is my contention that, in the long run, such metadata will be far more useful for the organizing and retrieval of learning objects than any a priori system will ever be.
PostScript
Here is the email I sent Tim Bray. I'm not completely happy with it. For example, what i would now want to say is something like: the best evidence that something is an image, say, is that it is referred to in an image tag. But this should suffice as background.
Hiya Tim,
I've been thinking about the problem posed in your 'On Resources' bit since you posted it a few days ago. I've run through the documentation, the posts, and more. I believe I have a handle on the problem, but maybe not - groups like this tend to start talking with their own language after a short while and what I think it might be might not be what it actually is.
My vast over simplification: it is not possible to tell, from a URI, whether you are pointing to the resource itself, or whether you are pointing to a document about the resource (which, perversely, may also be a resource). "There is no way for the Web software to distinguish between an 'Information Resource' and any other kind."
It seems to me that this is an instance of Russell's Paradox. The paradox occurs when you create a class of entities, with a fuction f that defines what entities fall into that class, in such a way that the class you just created is mapped by f to be a member of the class you just created. In the philosophy of language, this occurs when you make a statement about an element of language, where the element of language you are describing is contained in the statement you just made. Thus, a sentence like 'this sentence is false' is an instance of the paradox.
Russell's response is to introduce a theory of types. In a simplistic nutshell, instances of linguistic resources (ie., sentences) are divided into 'language' and 'metalanguage'. Now I know that in your response you reject the idea of dividing the universe of resources into two buckets. "Let?s please not pretend that this distinction is fundamental to, or even noticeable in, the architecture of today?s Web." But I think the other side is right here, in a sense. The distinction, true, is not a property of the web architecture, but it is a property of language in general. Language can be self-referential, and where self-reference can occur, paradox arises. Failing to address this issue at this juncture may introduce into the web architecture an incongruity that is being sensed by the members of TAG.
But what the theory of types introduces is not an ontological distinction, that is, it makes no attempt to divide the world into 'sentences that are about things' and 'sentences that are about sentences'. Rather, it proposes a stratification of discourse, with no assumption about the type of entity contained at a given level of discourse. In the contenxt of our present discussion, this amounts to being able, somehow, to distinguish between URIs that point to resources (level 0) and URIs that point to types of resources (level 1), where (critically) a 'type' is a description of a resource (that may or may not exist). In other words, we distinguish between pointing to the thing itself, and pointing to something *about* (ie., a description of, and therefore a characterization of a type) the thing itself.
From my (possibly naive) point of view, everything falls out of this. If we distinguish, in our discouse, between whether we are referring to a type (ie., a description) or a token (ie., a resource), then we need not make any claims about the names used to identify what we are referring to (just as, in language, there is no rule governing how we name things).
In language (English, at least), we usually employ a single quote to make the distinction. This, for example, I distinguish between saying "It appears that the Kremlin is red" (where I am talking about the building) and (It appears that 'the Kremlin' is red" (where I am talking about the words that make up the description of the building). Notice that there is nothing syntactically distinct between 'the Kremlin' as it is used in the first case, and 'the Kremlin' as it is used in the second; the set of letters is identical. It is how the words are *used* within the sentence (once inside single quotes, once not) that makes the difference. If we can establish this syntactical nuance in the world of the web, we are away to the races.
My naive understanding of XML and RDF has already drawn that distinction. In my mind (and possibly in nobody else's, because I don't claim by any means to be an expert in the syntactical niceties of XML and RDF) I have distinguished between two types of uses of URIs, one that refers to a type (ie., to a description), and one that refers to a token (ie. to a resource). Specifically:
In a tag, any tag, if you place the URI *between* the tags, you are talking about at token:
This, thus, would denote the location of the resource itself.
But if you place the URI *within* the opening tag, you are talking about the type:
Notice that no ontological distinctions are being drawn here. All I am doing is separating between the two levels of description, between types and tokens. In another context, the very same entity denoted in
In HTML it's a bit trickier, because no type/token distinction was coded into the original specs. But the same logic can be applied:
Now there is no rule saying that the URIs in the HREF tag and the ABOUT tag must be different; a document (say) can contain its own metadata. And all this formulation does is to allow our dereferencer to shift between levels. I can still link to my RSS file with but if I wanted, I could place metadata about that file (meta-meta-data) in a separate file
This, thus, would denote the location of metadata *about* whatever is denoted by
and
