Balisage Paper: Seeing things whole

My topic this afternoon is seeing things whole.^[1]

The German philosopher Leibniz says several times that if you really correctly know any thing, you know all of its relations to other things, you know all of the predicates which apply to it in much the same way that knowing a number entails knowing all of its prime factors. (This is more than just a vague analogy. In some sketches for what he called a characteristica universalis or universal writing system, Leibniz proposed to use prime numbers to encode properties and the product of prime numbers to encode concepts formed by the conjunction of those properties.) In any case, according to Leibniz, perfect knowledge of anything entails knowledge of all of the predicates true of that thing.^[2]

He suggested that it might be possible to exploit this fact by developing a universal writing system in which signs for complex concepts are built up by conjoining the signs for the primitive concepts that go together to make up the complex concept and that with a writing system like that, logical inference becomes much more straightforward, almost trivial, like arithmatic calculation — just as checking the divisibility of one number by another is straightforward, even trivial, if you have identified the prime factors of the two numbers.

This sounds hopelessly naive, over-optimistic, rather old-fashioned. (But let’s remember we are talking about Leibniz here. He also had an idea that if you could just get a plausible notation for concepts of infinitesimal change, you might be able to solve problems of motion and acceleration which had thus far — for millenia — eluded solution. And you know what? On that one, he was right, and we use his notation to this day in differential and interal calculus, which you might recall he invented at the same time as and independently of Newton.)

And as for being old-fashioned, well, Ludwig Wittgenstein, often regarded as one of the preeminent philosophers of the 20th century, argued something very similar in his Tractatus where he writes:

This line of argument ties in with Wittgenstein’s argument that the relation of logical consequence cannot in principle provide us with new information and thus cannot in principle surprise us. And for similar reasons, he argues, logic has no need of an identity symbol since the identity of any thing we have sucessfully identified must be trivial and visible on its face.

His argument is perhaps not assailable, but it is difficult to assail because he was a pretty smart guy and he didn’t leave many angles of attack. But it does seem pretty clear that for most of us, logical inference does have the capacity to surprise us and therefore that we live in a rather different psychological world from the one suggested by Wichtenstein, if we had any doubt that he lived in a rather different world from the one most of us inhabit.

On the other hand, it does seem reasonable to believe that if we are surprised by a logical inference from some thing or state of affairs, then after we’ve made that inference, we understand that thing or that state of affairs better than we did before, so perfect knowledge — even though we don’t have it yet — we may infer would involve no longer being able to be surprised because we understand the thing or the state of affairs so thoroughly that we grasp all of its logical implications and are beyond surprise. If we can as a rule be surprised by discoveries, it is because as a rule our knowledge of things is incomplete and imperfect.^[3]

This is a bit like saying that a perfect understanding of an elephant involves understanding its rope nature and its snake nature and its wall nature and so forth, or equivalently, that perfect vision of a forest involves seeing and knowing not just the forest but all of the trees and all of the leaves on all of the trees. That’s an awful lot for us to keep in mind, but that’s what perfect knowledge would be like.

But since in fact we don’t have perfect knowledge, we don’t always immediately see all the implications of a given proposition or set of propositions. It’s very helpful under these circumstances to have tools to work through some set of premises or some given data and calculate some of the implications we’re interested in. And this is particularly so when we’re dealing with complex systems made from many, many small parts intricately interrelated.

An outstanding example that we’ve seen here this week are the streamability rules of XSLT 3.0, which are such a complex system of many, many things intricately interrelated. The streamability tool described on Tuesday by John Lumley is useful in part because it does calculate out the implications of the constructs of the data that we give it Lumley. The rules for writing streamable functions outlined by Abel Braaksma in his talk on Wednesday also provide a valuable thread through that particular labyrinth Braaksma.

XSLT streamability is, of course, not the only complex system we have seen here this week. The National Information Exchange Model (NIEM) described by Priscilla Walmsley on Wednesday Walmsley and again this morning by Betty Harvey Harvey is another very complex system with many moving parts. I think the lessons drawn in Priscilla’s talk from implementation experience should be valuable to people working in very different projects, and the techniques that Betty described for describing extracts from NIEM and making them accessible and understandable to the people who have to work with them should also be very widely applicable.

It’s not just specifications or vocabularies that can consist of many small parts intricately related. If so, we could say, oh, well, the Working Group just didn’t succeed in boiling it down. Sometimes it’s our own documents. If you take almost any conventional document and you try to tease out the declarative or, for that matter, the procedural meaning of its markup by transcribing it in RDF or First-Order Predicate Calculus (FOPC), we are apt to see a huge explosion in volume. Sometimes, as in the EARMARK system described by Francesco Poggi yesterday Peroni, Poggi, and Vitali, or in the use of First-Order Predicate Calculus described by Yves Marcoux and myself Sperberg-McQueen, Marcoux, and Huitfeldt, this explosion can allow us the use of new tools for working with the information, but it does make for an awful lot of distinct objects to focus on. You will recall that John Lumley observed that in the streamability analyzer, one line of XSLT not infrequently turns into about a kilobyte of information in the browser Lumley. And as he said, You can look at all this in any XML editor, but then you can’t see the wood for the trees.

A desire to see the wood, and not just the trees, motivates what has historically been a strong interest in tools for controlling complexity and for hiding complexity from ourselves. The application of styling is one such tool, illustrated again by John Lumley in the streamability tool and his encore presentation on dynamic documents. Josh Lubell’s XForms interface for small arcane non-trivial data sets is a way to make things easier to grasp, easier to grok Lubell. Betty Harvey’s documentation of profiles of complex vocabularies falls into the same category Harvey.

Another tool that we use and cultivate is declarative programming, illustrated here by Anne Brüggemann-Klein’s work with her students on a computer game using almost exclusively declarative XML-based technologies Sayih, Kuhn, and Brüggemann-Klein, Josh Lubell’s work again Lubell, and Han-Jurgen Rennau’s work on the declarative description of tools and, from those declarative descriptions, the generation of consistent user interfaces and so forth Rennau.

A third tool that we use to control complexity is validation. Validation controls complexity in a slightly different way: it simplifies the population of inputs that we have to think about. If we know that the input conforms to a schema, we may know, for example, that every chapter is going to have a title, and that means our code for handling chapters or chapter titles does not need to say Oh, wait, first of all, does my input look as I expect it to look? Given a chapter, do I actually have a title? The stripping out of those now unnecessary conditionals amounts to a huge reduction in volume and a huge simplification of the programmer’s task.

Validation is not by any means a solved problem. Anne Brüggemann-Klein’s work on identity constraints Brüggemann-Klein, Maalej, and Sayih and Eliot Kimber’s continuing quest for good ways to validate DITA Kimber and Bina provide evidence of that.

We also try to tame complexity by bullding more powerful tools, or more convenient tools. The work reported on by Benito van der Zander is a good example of this: simple patterns that allow us to do very complicated things in a very simple way van der Zander. Or consider Pat Case’s work on XQuery full-text Case; Pat observed that one of the big advantages of XQuery full-text is: You do not have to define search fields; you can search on all your data. I think of that as another way of saying you can view your data whole; you can see things whole. One compelling argument for more powerful query tools or extraction tools was formulated here by David Lee: The bigger the data, the more you need less of it. Lee And of course, broader dissemination of existing query tools falls into this category too, as described by Cliff Anderson just a few minutes ago Anderson.

A fifth approach to controlling complexity lies in achieving greater uniformity — this is less of a tool and more of a practice. It is illustrated by the topic tools described by Hans-Jürgen Rennau Rennau and by Steve DeRose’s work on providing consistency and uniformity of interface between the XML accessors in a given language and everything else in that language DeRose. It’s a question of context: Is it more important to have all of the XML accessors look the same across languages, or is it more important to have them look like the rest of the language? Steve suggests, I think quite rightly, that programmers are more apt to care about consistency within a program and across their experience with a given programming language than they are about consistency between programming languages. One of the reasons we have different programming languages, as Wendell [Piez] suggested in a discussion a little while ago, is that they’re good for slightly different things. We don’t want perfect consistency across programming languages. That would be boring.

David Lee’s programmatic desire to generalize and extend XPath is also seeking a kind of uniformity Lee. They say that if all you have is a hammer, then everything looks like a nail. What I understand David to be saying is, when all I want to use is XPath, then I want everything to look like a node. And he’s right. Everything that we can make look like a node is something we can work with conveniently in XPath — and if we really want world domination (I’m not sure we do), those of us who are interested in world domination should take that to heart.

Kurt Cagle’s suggestion that devices can describe themselves and their statuses and their capabilities by using a uniform RDF interface is another attempt to reduce complexity by achieving uniformity Cagle.

There’s another way, if not to control complexity, then to reduce its impact. It’s not exactly a tool; maybe it’s a technical fact; maybe it’s just a psychological fact. But we can reduce the apparent importance or difficulty of complexity by zooming out, by considering things in a larger context.

Taking things in a larger context, taking a wider view, taking a longer view has a number of advantages. It can help us to look past superficial differences of opinion and see strengths, advantages, utility, good work in other people, other technologies, other groups, even when they refer to things we’ve been working on for a long time as pet projects. (There was much more than this one point in Robin Berjon’s talk on Monday, but that was an important take home from it Berjon.)

Taking a broader view can help us see there can be multiple paths to the same goal. Alex Miłowski showed us on Monday how some features of HTML5 can help us to achieve some things in the browser that some people had hoped to achieve using XML support in the browser, but which never quite materialized out of the XML support in the browser Miłowski. But if we can achieve them now, does it matter whether the features are called XML support or something else?

Phil Fearon showed us how to work both sides of the street by processing XML as HTML5, and HTML5 as XML Fearon. (I did notice that the list of technologies he deployed was quite long, and I began to think that this might be another case of many, many things intricately inter-related.) But sometimes that’s what you have to do to make things work. And we should all take to heart his memorable remark: There is no reason to choose one or another. We can choose both. That’s nice to be able to do. Alex Miłowski (there he is again!) and Norm Walsh showed more of the same direction on Wednesday Miłowski and Walsh.

And, as I said in the discussion yesterday, it looks as though those of us who are interested in overlap had better start learning SPARQL and the other technologies used by the EARMARK team Peroni, Poggi, and Vitali, because the work being done in Bologna is changing the landscape of that discussion and it’s going to get harder and harder to take any consideration of overlap seriously that doesn’t engage with that work.

Another excellent example of taking a broad view and an example of the kind of consideration of a wide variety of factors was given by the paper on the Recommended Formats of the Library of Congress as presented Monday by Ardie Bausenbach and Kate Zwaard Bausenbach and Zwaard.

Taking a broad view helps us understand that no single point of view serves everyone. If you needed another example of that, I thought the MathML panel on Wednesday showed that very well Dineen et al..

And if you take a sufficiently broad view, you end up with sort of panoramic views of whole industries and vocabulary developments, and a real focus on the bigger picture, as illustrated by David Webber’s talk this morning Webber.

There’s another way to zoom out, and that’s to take a historical view. Steve DeRose’s talk about hypertext and the history of hypertext theory and practice on Monday DeRose (symp.), and Liam Quin’s talk about markup systems on Wednesday Quin, both illustrate to us the general principle that seeing things in historical depth can sometimes help us to get them in focus. For many of the things that we would like to do, it helps to remember that a lot of people before us may have tried to do them. And (as Lauren Wood remarked, while putting up a copy of the old DOM Ranges spec as a poster after listening to some of the early talks), even if we don’t see ourselves able to adopt their solutions wholesale, there may be things we can learn from them. The past, said an English novelist, is like a foreign country; they do things differently there.^[4] That means that even if they were trying to do something that we are also trying to do, they may have had a slightly different point of view, and we can learn from that difference in point of view.

If we take a broad view of the possible applications of our data, we may find ourselves led to the kind of rich markup and intelligent use of it for a broad variety of applications that were illustrated this morning by Joe Wicentowski’s talk about the use of XML in the State Department’s Office of the Historian Wicentowski. Rich views of data and the reuse of information for different purposes have always been important in the use of descriptive markup; they are in large part why we have the concept of descriptive markup.

One particularly difficult aspect, technically speaking, of the quest for a really rich and informed view of any complex data is the requirement, or desire, to capture and expose something about its development over time, a perennial topic of concern. I think we saw some progress here this week in the proposal discussed by Robin La Fontaine La Rontaine, and in Ari Nordström’s account of a custom version management system Nordström, and also in a slightly different vein in Josh Lubell’s discussion of ways to display overlays over complex technical documents Lubell.

But perhaps the most striking example of a proposal for a richer account of our information was the proposal made by Domenic Denicola during the symposium on HTML5 and XML on Monday, with his discussion of the development of web components Denicola. He reminded us in what can be a salutory way that even when we provide an elaborate vocabulary definition document, be it a DTD, or an XSD schema, or a RELAX NG schema, or a big wad of documentation, XML data doesn’t in itself do anything. I like things, he said, that do things.

And it’s clear, as he argued on Monday and as Alex Miłowski and Norm Walsh showed on Wednesday Miłowski and Walsh, that web components will make it far easier to do interesting things with user-defined markup. But I have to confess that, I think that if we want to see things whole, and particularly if we want to see things in a historical light, we may want to take a slightly different view of what is happening here.

In the 1960s and the 1970s, those involved in the initial efforts that led to the concept of generic markup took the view — or so at least we might more of less fairly infer from the work they did and the documents they produced — that an external stylesheet language that would allow us to specify the font and the layout and other typographic characteristics of any unknown element that we might provide to a processor, or that might appear in the input to a processor, would constitute an adequate provision for the semantics of any user-defined markup in a document. If we could provide that typographic information, then user-defined elements could have the same capabilities as those supported by the intrinsic semantics of the application.

Now in the 1990s when CSS was being developed, that earlier view began to look a little naive, because there are properties that we need in order to specify the display of an element in a browser that had not been foreseen in the work of the GenCode committee or Working Group 8 of Subcommittee 18 of Joint Technical Committee 1 of ISO and IEC (ISO/IEC JTC 1/SC 18/WG 8). And also in an interactive application, the CSS developers reminded everyone, it’s very helpful to keep the selectors in a stylesheet rule (what some people speak of as the left-hand side of the rule) as simple as possible because the more complex the selectors are, the slower the application of the stylesheet to the data becomes. And that’s really important in an interactive application like a web browser. So both the right-hand side and the left-hand side of the rules needed to look different from what the earlier work had expected. An adequate specification of the semantics of an element required a rule or set or rules in a CSS stylesheet or set of CSS stylesheets, not just the kind of semantics that had been foreseen earlier.

The developers of CSS, on the other hand, did not foresee Web 2.0 or later developments. So web components now are being designed in part around the insight that more is needed to specify the semantics of an element than a set of CSS properties, and to fill that gap by providing a way to specify an interface between the element and the rest of the browser, thus finally enabling the markup designer to describe the full semantics of non-HTML elements in a way that puts them on the same footing as elements intrinsically understood by the browser.

That, at least, is the goal. And this time, surely, we’ve got it right. Surely we’ll never be around this track again. There is no danger at all surely that the primitive semantic notions of web components might strike future observers as naive, or that they might fail to foresee the future in the same way that the failure of CSS1 to include hyperlinking behavior, for example, might strike some observers (both now and then) as naive.

This is the last time around this track; once we’ve gotten this round of development done, we’ll be finished and go on to other things. Right? Maybe.

Now, some of you know that the account I gave a moment ago of the thinking behind CSS and the thinking behind earlier efforts toward generic markup was not at all accurate. (And I’d like to thank you now for not shouting out to interrupt me while I was telling the story.)

Because the developers of the notion of generic markup, as far as I know, never actually held any of the views attributed to them in the story I just told.

It is true, at least according to what I’ve been told (I wasn’t there) that many of the pioneers of generic markup initially came together with the goal of defining a single markup language useful for all documents, so that publishers could code their documents in this generic coding, or GenCode, and be done with it. But by all accounts, they saw very rapidly that no single tagset would suffice or could possibly suffice for the applications they could easily foresee, and they knew quite well that there would be applications that they could not foresee.

So they didn’t spend a lot of time arguing over the merits of including this or that element and specifying its typography this way or that way. They didn’t spend any time at all on a universal tagset because, to borrow a phrase from Anne Brüggemann-Klein, they realized that it didn’t really matter, because the whole concept was doomed anyway.

I have gotten the impression from conversations with some members of the relevant working groups that in fact they used to remind each other from time to time during meetings that the markup languages, or meta-languages, they were developing must not assume that the only use of markup would be to mark up manuscripts for typesetting, that people must be able to use them for arbitrarily unusual, even arcane, topics like security controls (of the kind that Josh Lubell was talking about yesterday Lubell), or interactive games (of the kind Anne Brüggemann-Klein showed us an implementation of Sayih, Kuhn, and Brüggemann-Klein), or even — ne plus ultra of arcane and unusual information — the kind of linguistic of literary and linguistic analysis that Wendell Piez showed us in his visual tour de force the other day Piez.

When the Text Encoding Initiative came along and actually did define an SGML vocabulary for literary and linguistic study among other things, the reaction was unforgeable. They clearly were elated, even moved, by the thought that they had prophesied such an application of descriptive markup, and that their prophecy had come true. It must be a remarkable feeling to try, more or less blindly, to ensure that the system you are building is usable for applications you are not in a position to foresee, and then to have someone say, What you were trying to make possible, you did make possible.

Now, it’s true that what most of those early adopters and promoters of descriptive markup wanted was applications whose operational semantics were typographical. But they did not confuse the semantics of markup with the operational semantics of that markup for one particular application in one particular environment. At most, a stylesheet rule — or a Web component interface specification — specifies the operational semantics for one application of the tagset in one environment. It is not the creators of SGML who were naive in not specifying operational semantics, or even in not specifying (at the meta-level) a language for specifying operational semantics. It is anyone who believes that any concrete specification of operational semantics can ever constitute a full specification of the meaning of markup, especially user-defined markup.

There is a reason that SGML has given rise to applications that have lasted for decades and has been applied in areas that its original developers foresaw only vaguely or not at all, and TeX (for example) has not. The reason is that TeX came with a perfectly well-specified semantics. What TeX markup means is ink on paper here); it’s very good of its kind, but it does not countenance application to sonnet structure. It does not countenance application to markup to distinguish a Leibnizian view of a monad and distinguish it from other markup that identifies a Russellian view of the same phenomena.

Domenic Denicola was not wrong to like things that do things. Denicola We all do. But the designers of SGML stayed away from operational semantics for a very good reason: they liked things that do more than one thing. They liked things that do things today and will continue to do things tomorrow in a different environment, a decade from now, a century from now. They took the view that has been expressed here several times in discussion by Steve DeRose and Eliot Kimber and others: once you have named something, then you can do things with it.

Antoine de Saint-Exupéry is often quoted as saying that the designer has achieved perfection not when there is nothing left to add, but when there is nothing left to take away. Those who gave us the concepts of descriptive markup invented many things, but one of the most important virtues in their design lies not in what they put into it, but what they kept out of it. Descriptive markup does not do anything, and that is its strength.

But seeing things whole also means looking past the surface a bit. I disagree with Domenic Denicola’s view of what it means to define the semantics of user-supplied markup Denicola. But on Monday Robin Berjon described a choice which he said had been faced at some point by the SVG Working Group Berjon: Which do you want? To be pushing a particular syntax? Or to have cool graphics on the Web?

We face perhaps a similar choice: What do we want?

Do we want to be recognized as having had, at some point in the past, ideas similar to those being pursued by others today? To get credit for the goals and aspirations of the members of the SGML on the Web working group? Or do we want to have web browsers that support user-defined markup in a system with consistent separation of content from the specification of processing semantics?

What does it matter, in the long run, whether those who are specifying web components and building them into the infrastructure of the Web believe that in doing so they are fighting with and overcoming the baleful side-effects of XML; or recognize (as I would see it) that they’re helping carry on and carry out the program and achieve the goals that motivated the creators of XML and related technologies ten and fifteen years ago, and before that, the creators of SGML and its sister specifications thirty and more years ago, and before that, the propagandists for GenCode of forty and fifty years ago?

Can we take yes for an answer? Even if it involves looking past the fact that those providing the answer believe (wrongly, as we may think) that they’re answering no?

One key to understanding, Leibniz told us long ago, relies in understanding the true nature of things and not letting ourselves be misled by superficial properties of language or by the names that we or other people have chosen to use for things.

Leibniz hoped that it would be possible to reduce complex notions unambiguously to particular conjunctions of primitive notions, or basic notions, in just the same way that each number can be reduced unambiguously to a unique list of prime factors. Three hundred years later we may not find that proposition so obvious or even plausible. We may not believe that all interested parties are likely to agree on how to reduce complex notions to more basic ones. We may not even be confident that a complex concept has a unique decomposition into primitive concepts, analogous to the fundamental theorem of arithmetic.^[5]

A simple example is given by the definition of trees in graph theory. In treatments of graph theory, we can find a variety of definitions of tree. There are eight listed in the textbook on graph theory I consult most often, and it is an exercise for first-semester students of graph theory to prove they are equivalent, i.e. that they denote the same set of mathematical objects. But none of them is reducible to the other.^[6]

Leibniz’s dream would have made it possible for us to learn all about elephants by careful study of the character used in his philosophical language to denote elephants. We could have learned all about the salient characteristics of forests by studying the character used to denote forests.

We no longer share Leibniz’s hope.

We will not learn about elephants or forests by studying the words elephant and forest. Instead, we need to get together in physical proximity to other people who have seen different parts of the elephant, or different parts of the forest, exchange information, and discuss how to fit all that information together.

Indiviudally, we see parts of things.

When we come together, we can come closer to seeing things whole: forests, trees, leaves, and all.

That’s what conferences like Balisage are for. Thank you for coming to Balisage. See you all next year.