MishMash

Cameron Neylon, Capturing process

Just been at a seminar organised by the Comp Sci people here in Newcastle. Cameron Neylon, currently of the Science and Technology Facilities Council on:

Capturing process: The promise and challenges of connecting experimental records to the semantic web

The promise of the semantic web in its broadest terms to integrate structured data, and in particular scientific data, is breathtaking. Yet our ability to deliver on that promise remains severely limited. On one side, battles are fought over vocabularies and structured description languages, while on the other computer scientists argue over design and architecture. In the middle the experimental scientist is usually left bemused as to what all the fuss is about and how it can have any relevance to their research, which is naturally far too specialised and ground-breaking to fit into any framework that an outsider could ever provide. The challenge lies in finding a pathway that moves towards satisfying the needs and interests of these disparate groups while providing both value and a usable interface to the end user.

We have adopted a blog based approach to capturing the record of research in an experimental laboratory. The use of a free form framework along with templates encourages the user to structure the description of their work in a way that matches their needs. Cycles of template design and optimisation lead to structures with strong similarities, and occasional important differences, with designed ontologies and vocabularies. The challenge for the future lies in determining how best to guide the actions of users towards the use of appropriate controlled vocabularies and structured descriptions, while using their choices to inform the design of these systems. The potential for Google Wave to provide a framework to enable local structuring as well as integration with global structured descriptions will be explored.

Some interesting ideas, as abused by my memory:

• People tell stories, they don't organise stuff the way it needs to be organised for machines to process. A big part of the disconnect between scientists and the folks building "e-science" tools is that these tools don't help people tell stories.
• So we need to start capturing structure through storytelling. That is, allow people to work in a natural way and pick up the structure from what they do.
• You can create a virtuous circle between metadata, which a user would normally not bother filling in or fill in incorrectly, and user interface. Arrange the UI so that the metadata populates the UI. The user wants that UI to make their life easier, and they start to notice that getting the metadata right is necessary.
• (Lab) instruments as bloggers. Have them push results straight into the blog. Take the human out of the loop (I might qualify, as I think was intended: where they add no value).

Cameron has had early access to Google Wave, it seems, and is impressed with the potential. He had some cool examples of Wave "robots" going off and filling in extra information, drawing graphs, creating links, tidying up references.

He was less impressed by the Google client: too much stuff going on in a single UI, too non-standard. He also expressed concern about users understanding the model. It's the eternal problem with geeks developing stuff for "normal people" again. We're geeks because we do abstract with ease.

There was a lot that I recognise, and some new thoughts as well. I'm all about computational analysis in engineering decision making rather than lab-based science, so there are some differences, but there's a lot of overlap too. In theory we should have an easier time here since so much of our "stuff" comes in the form of information artefacts, and our "experiments" take place in silicon. There are instruments, and the process of getting data from them into our analysis is currently extremely painful, but most of that is "upstream" of our analysis.

I liked the picture of an lab process generating artefacts and narrative, and then a separate process of arranging these things into a different narrative for publication. The story we tell when we want to explain our results to others isn't the story of the trials and tribulations of getting to those results, its a story that gets the reader as smoothly as possible from their current knowledge to where we want them to be. That's true of engineering analysis as of science. And in both cases, we'd be in a much better place if, given that latter story, we could easily investigate the artefacts referred to: look at them, do our own visualisations or further processing, and follow the trail back to see exactly how they were generated.

Digital information is different

This is beautifully done, and hits a few nails on the head:

Merlin has a few worthwhile comments on the strengths of paper. Also missing, from my point of view, is the fact that a lot of what we do with computers now (pretty much everything that's off the web, for a start) needlessly recreates the limitations of paper and adds a few new problems for good measure. Now where did I put that file ...

Drawn with a very fine camelhair brush

A commenter to a post of Dave Weinberger's left a link to an essay by Jorge Luis Borges, The analytical language of John Wilkins, which includes the quote which so amused Michel Foucault that it helped inspire his writing of The Order of Things (amazon.co.uk, amazon.com):

These ambiguities, redundancies and deficiencies remind us of those which doctor Franz Kuhn attributes to a certain Chinese encyclopaedia entitled 'Celestial Empire of benevolent Knowledge'. In its remote pages it is written that the animals are divided into: (a) belonging to the emperor, (b) embalmed, (c) tame, (d) sucking pigs, (e) sirens, (f) fabulous, (g) stray dogs, (h) included in the present classification, (i) frenzied, (j) innumerable, (k) drawn with a very fine camelhair brush, (l) et cetera, (m) having just broken the water pitcher, (n) that from a long way off look like flies.

Jorge Luis Borges

The rest of the article makes it clear why the quote is significant: it looks ridiculous, so ridiculous that if it catches you at the right angle (it did me) it makes you laugh, hard and out loud, but it is ridiculous in <em>just the same way</em> as a lot of the classification schemes that we use all the time.

It is clear that there is no classification of the Universe not being arbitrary and full of conjectures

Jorge Luis Borges

Reticulation

I've not been paying attention. My earlier note in response to Sean McGrath's comments on hierarchy obviously hit the spot. I've only just noticed that Sean picked it up.

Jon Udell on del.icio.us

Gary King points out a screencast on the "social bookmarking" service del.icio.us by Jon Udell. Great work, as usual, from Jon; his earlier effort on the "Heavy Metal Umlaut" article at Wikipedia was perhaps even more powerful. Both draw out some of the power of Internet-scale social/community tools.

It's particularly amusing to see myself on the small screen: I'm one of the people who posted an article on domain specific languages. I posted it under dsl, because domainspecificlanguage(s) seemed a bit long. As Jon points out, dsl is a bit ambiguous. I could change it, but then I'd move my posts away from everyone elses on domain specific languages.

The problems with tagging remain clear. Supporting these social processes is vital, but the web will not be "semantic" until we, somehow, get beyond slapping simple sets of text strings against items. Ambiguity of acronyms is only the tiniest part of it. Still, they are indeed a powerful tool.

Tag stemming

Well, this won't solve the folksonomy argument, but it is a useful tool. Here's to the power of web service APIs.

Matt Biddulph's tag stemming tool for del.icio.us.

Darwinian metadata

Ben Lund suggests that tags and folksonomies can be seen in a Darwinian light. A nice idea.

The four pre-requisites for a Darwinian process are Reproduction, Variation, Inheritance and Selection. Let's say our individuals are tags and our species is a folksonomy:

• Reproduction: Certainly. New tags are created with almost every entry, be that bookmark, article, or photo.
• Variation: Absolutely.
• Inheritance: Probably. When I copy a bookmark, I often copy the tags too.
• Selection: Yep. 'Edit tag' was the first feature requested after we launched Connotea.

I think there's some mileage in this idea, but I'm not sure that individual=tag, species=folksonomy works perfectly. Then again, it's an analogy, let's not get too worked up about it. More importantly, "edit tag" isn't enough on the selection front. Tags need to die on a large scale, and that isn't happening. Nor is it obvious how it can happen without things that were tagged no longer being tagged, or old tags being ignored, which seems to lose too much value.

Taxonomy, folksonomy, and value

Let's try again.

Shelley Powers has written a good summary of some arguments about the relative merits of tagging and more formal metadata. Nice pics, too.

Shelley is worried that the rise of "cheap" metadata, in particular tags, will inhibit development of tools to make "real" metadata available to the masses.

Theory: the value of terms in any metadata system used in an uncontrolled way tends to zero with time.

That's just as true for rich, formal so-called “ontologies” as it is for tag soup. There is a difference, however. Tags don't lie about their status. We know, right off, that they're just terms which people have attached to things, for some reason. We know that we don't know what sense the term is meant in, of the undoubtedly many possible. We know that we don't know the intention of the tagger in tagging. The spam problem arises from this: spammers are applying tags in order that pictures show up in certain places, not because they think those tags are relevant. The business of tagging flickr photos with the "offensive" tag is another case in point, where the intention is now to stop items showing up in certain places, and really has nothing to do with offensiveness.

Problematic, perhaps, but not unique to tagging. Exactly the same is true of uncontrolled use of any classification system, however good the tools are which provide access to it. Things will be tagged in a way which is inconsistent with the intended use of any system, either through ignorance or intent.

The rel="nofollow" tag which Google has announced it will use to control the inclusion of links in its PageRank calculations is indeed a nice demonstrator of an important point here. It's meaning is very well defined: it is grounded not in human understanding, but in machine behaviour. Nofollow can't be spammed: either Google honours it, or they don't. The web now has a mechanism for systematically witholding page rank. What it does with it is out of Google's control.

Point is, that some terms have meaning in this, mechanical way, like those which make up a programming language. Others have meaning in a fluid, human way. The latter will always be as fluid as they are human; no fancy tools will change that.

Fancy tools to make use of formal taxonomies easier would be great, of course. Very useful for those who care enough to put in the effort to understand, and make sure they stick to, the formal intended meanings of terms. But however easy that is made, it will always be hard, and people will always get it wrong. And as people get it wrong, the extensional definition of the term becomes less precise, and the value of the term tends to zero. You might slow the process down, but you can't, simply can't, stop it.

And, of course, because the effort involved in understanding and carefully applying the terms of a formal taxonomy is relatively high — independently of tools, which can reduce the accidental difficulties but cannot attack the essence — it is just not useful in the sort of "fire and forget" systems like del.icio.us, where any effort is too much.

In the comments to Shelley's post, Nick Sweeney writes:

I’m a bit of a Saussurean about this, in that I think that taxonomy (or ontology, depending upon your disciplinary point of origin) is crystallised/calcified folksonomy. Authorised folksonomy, if you like.

Right on! Those two terms, “crystallised” and “calcified” are a really useful duo. The one has connotations of order, beauty, and value. The other of dull rigidity. Well designed formal taxonomies have the first of those in abundance. Both have the second, and you can't avoid it. A taxonomy can only ever be an encoding of one incomplete and imperfect view of the world.

Shelley:

I agree with Clay that the semantic web is going to be built ‘by the people’, but it won’t be built on chaos. In other words, 100 monkeys typing long enough will NOT write Shakespeare; nor will a 100 million people randomly forming associations create the semantic web.

100 monkeys indeed will not write Shakespeare, but 100 Shakespeares didn't write Shakespeare either. Shakespeare did. The situation we have is that of 100 million people (or, at any rate, some large and growing number) who need tools to organise stuff, and to help them find stuff. 100 million people who will go right on randomly forming associations, whether the results pretend to be anything other than randomly formed associations or not.

We're dealing here not with the formal space of so-called ontologies and logic, but with the messy, human space of language and meaning. The processes involved are, essentially social. Tools which deny this will fail.

Shelley:

Clay believes that ultimately ontologies will fall to folkonomies, as the latter gain rapid acceptance because of their low cost and ease of use; I believe that ultimately interest in folksonomies will go the way of most memes, in that they’re fun to play with, but eventually we want something that won’t splinter, crack, and stumble the very first day it’s released.

When you design a building to ride out earthquakes with minimal damage, you build in flexibility. Of course, the clever part is in deciding where and how to build in that flexibility. Tagging provids an abundance of flexibility, so much that the building can barely stand up.

But in some ways it works. Delicious is already very useful, on an individual basis, and has some use at a social scale. I follow what other people post about Lisp, Scheme, and Smalltalk for example. These are well defined terms, referring to programming languages rather than highly ambiguous and mis- or differently interpretable concepts. With less well defined concepts, the problems lie as much in the nature of the concepts as in the tagging approach.

Simply aggregating tags a la Technorati as it stands doesn't show a great deal of promise, but it is possible to envisage tools which show tag clouds and, by presenting these to the user, encourage convergence around particular tags. The results will never be complete, never perfect, but always changing, and sometimes useful.

When you have a Web-load of people, things happen from the bottom up. That's the way the Web works.

Losing time

Damn. Firefox ate my homework. I'd just spent some (too much) time commenting on the debate about taxonomies versus folksonomies, particularly as summarised by Shelley Powers, when it crashed, taking my efforts with it.

It crashes quite a lot, really, often after freezing up for a while. If I switch to a different application then switch back, particularly to a page with text fields in a form, it often takes a while to update and come back to life, and sometimes dies before it does so.

I guess I'd better just get used to using a desktop client rather than the TypePad web form to post. Another app to fiddle with proxy settings for depending on whether I'm at home or at work. Rats.

On Hierarchy

A nice line:

Hierarchies. Cannot think within them. Cannot think without them.

Sean McGrath

Arthur Koestler in "The Ghost in the Machine" (amazon.com, amazon.co.uk) talks about arborisation and reticulation as complementary. Hierarchical structuring is powerful, but a closed hierarchy is restrictive. The leaves of the hierarchy (at least) must connect with those of other hierarchies to form powerful structures.

I think Koestler's insight provides quite a conceptual leg-up when thinking about the composition of system representations ("models") with which to drive simulations. Considered as gross components, subsystem representations are assembled in a composition hierarchy which is a straightforward tree. The leaves of this tree are atomic components which are defined in some way other than by composition. Composites do something useful because they define connections between the inputs and outputs of their component parts, and ultimately this alsways means connecting the inputs and outputs of the atomic leaf components. Thus the modeller is simultaneously defining a tree and a network.

When a single-purpose model is being constructed, this will result in a single tree. We can imagine, however, that long-running systems expose the data flows at their leaves in some way (publish and subscribe). In this case it should be possible to define new systems which connect to these flows, and then we start to define a forest of individual trees.