February 17, 2006
Folksonomies, taxonomies and population thinking
In preparation for a talk on “Science as social practice” I came across a quote by Ernst Mayr (via Three Toed Sloth):
The assumptions of population thinking are diametrically opposed to those of the typologist. …. For the typologist, the type (eidos) is real and the variation an illusion, while for the populationist the type… is an abstraction and only the variation is real.
This seems to me to perfectly capture the issues that divide advocates of folksonomies and taxonomies. Taxonomies attempt (and, as advocates acknowledge, always fail) to find a complete and sharp set of “types” with which to classify their domain. Folksonomies rely on the population of users to create a population of tags that can adequately organize a population of content items.
Perhaps the biggest differences come in managing change. Folksonomies are inherently squishy and messy, and accomodate change much more comfortably — at the cost of never providing the benefits of crisp classification. They are much more successful online because they leverage the power and fluidity of digital environments.
Taxonomies are inherently sharp edged and therefore brittle. They tend to break rather than change gracefully. Furthermore, because they typically require a substantial investment of human effort in construction, training, labeling, etc. change is expensive and painful. On the other hand, sometimes we find these prices worth paying. Searching a physical library where the books were informally tagged by users (perhaps using post-its) would be a nightmare.
We are working on the population side of this dichotomy. Our difference from existing folksonomies (e.g. del.icio.us) is that current sites provide only an extensional definition of their tags, while our technology will create an intensional definition (though squishier than the standard philosophical idea of intension). This lets us provide significantly more complete support for the real diversity and complexity of the user population than existing sites. Let me explain.
On existing sites, the meaning of a tag is simply the set of things that have been given that tag. You want to know whether a tag is appropriate in a given case? Look at what else has been given that tag, and decide whether the thing you were planning to tag is “like” the others. This is a classic case of extensional definition.
Our initial tagging system will create intensional definitions of tags, based on a given user’s choices. Specifically, it will summarize a population of tagging decisions by a classifier. Initially we’ll use a simple bayesian filter, like today’s spam filters. If this doesn’t work well enough, we’ll use other technologies — there are a lot to choose from (Support Vector Machines, Latent Semantic Indexing, etc.). But we think simple methods will work well enough.
So what does this do for our users? There are two benefits that we think are important. First, since our system has an intensional definition of what a given user means by a tag, it can classify new items for that user. If it is wrong the user can correct it and thus improve that definition.
Second, since the system has these intensional definitions for each user, it can compare them and find similar tag definitions. Note that it makes no difference how a given user spelled their tag. If I say “SF”, you say “Science Fiction”, and someone else says “Space opera”, the system judges the similarity of the intensional definitions and ignores the way their names are spelled. In our environment, similarity is a relation on the intensional definitions, not on the tag spellings.
So how does all this relate to population thinking? Systems with shared extensional definitions of tag meaning or shared moderation values implicitly push their users toward a single, shared, “ideal” perspective. Tags and moderation values tend to be viewed in terms of “consensus meaning” — and sites like Slashdot have explicit meta-moderation to enforce this concensus!
In contrast, by generating intensional definitions we explicitly accept (and celebrate!) that the users of this technology will be a diverse population, each with different interests and ways of classifying the world. Our goal is to accomodate the real diversity of the user population, while also giving people ways to adopt each others’ perspectives, and to collaborate when they want to.
Update: Tim Spalding of LibraryThing has some interesting meditations on how to compute intensional meanings from an extensional tag pool. (He doesn’t use those terms, but he needs some sort of intensional definition to make the similarity judgements he’s aiming for.) He specifically comments on the messy nature of tagging and the need to use statistical methods to extract cleaner information from the tags.
However Tim is still effectively throwing away a lot of information. A given user might have a very focused meaning for some of the tags he mentions — e.g. “WWI” or “memoir”. But since LibraryThing is looking only at the total pool of tags spelled the same, rather than the meaning of those tags within each user’s library, that focused meaning gets washed out. This is exactly the effect we are trying to avoid. The key differences are that Tim is computing intensional definitions over the whole tag pool, while we are computing them for each user separately, and Tim is using only the tags and the identities of the tagged items, while we are also using the “fine structure” of the tagged items (the tokens they contain). Of course for books this fine structure is often not available.
Filed by Jed Harris at 9:57 pm under Populations, Tagging
Thanks for this stimulating post. I will read it a few times. One note of correction: LibraryThing does not compute tags based on what’s spelled the same. Rather, in keeping with its many other user-driven features (eg., user-driven edition and author disambiguation) it allows tags to be “merged” by users on a global level. So, users have merged “wwii,” “ww2,” “world war two,” “world war II” etc. Users who go to merge something are treated to a sermon about not combining things that merely *seem* similar. (The differences between a tag like LGBT and GLBT are fascinationg.) And there’s a changelog, which helpful users—LibraryThing has a lot of librarians, after all—police.In some ways, this is a “dirty hack.” In others, a user-driven solution to a problem arising from user-driven data. Anyway, it does work pretty well.
Part of a longer response to email from Tim:
I think if our ideas can generate any benefit for you, it is as much due to a shift of perspective as to better technology. We’re trying to look at the individual correlations — in your example, when a particular user tags The Seven Pillars of Wisdom with “memoir”, what other books does *that user* tag with “memoir”? This is likely to be a lot more relevant than someone else who uses the “memoir” tag entirely on late 20th century female writers. I think if you look at the patters user by user, you are likely to get more interesting results — but like I said, your empirical approach is right!
I don’t know how a more individual approach would play out in your code (or if it even makes sense to think about it). But that is probably where the main benefits would be.
I agree that the LibraryThing approach, which ” allows tags to be “merged” by users on a global level”, e.g. “wwii,” “ww2,” “world war two,” “world war II”, based on their own judgement, and providing defaults that can be overridden per user, may be the only or most universal way. It deals nicely with users who speak multiple languages, too, and don’t care to segment their notes to themselves. Two users, one of whom speaks German and French, and another who speaks German and Russian, would need a way at least to distinguish pages in German (which they can both read and write) from those in French or in Russian only. But similar considerations apply for people who speak only a limited range of English (Basic English or so-called Simple English, or some specific grade level). The individual user’s idea of “wwii” is less interesting than the correlations between what various users with similar backgrounds or interests tag that way, e.g. those who speak German and French may put the Rhineland and Ruhr annexations and disputes, even the 1870 Franco-Prussian war, into the “wwii” category, while those who speak French and Russian may be more inclined to make note of the Molotov-Ribbentrop pact, the Second Internationale, etc.. Both of them might link the Paris 1919 conference though for different reasons. All speaking statistically, of course.
And none of this deals with the question of the “don’t know” and “don’t care” relationships. Not everyone sees every page and not everyone cares enough about every page they see to tag it. So “tag X applied”, “tag X considered and refused”, “no tag considered” typically because the page was not seen or had no means to accept it, “tag X removed” after others applied it, are all different. Psychology and semantics differ drastically between these four states. Most web pages are in a permanent “don’t know” state because the publishers “don’t care” to have the users tag them enough to enable any technology to do that - unless one trusts alexa.com or can dig up the crit.org database or pile through everyone’s bookmarks to see where they put that page. So introducing “don’t know” and “don’t care” introduces a point of view (POV) that need not be universal or individual.
More typically, it’s shared among some social network or other, including a small creative network (a “team”) doing some task in common, or a large power network (like a political party), that have agreed to handle a wide range of tasks in a common way, and to mechanisms (like voting systems) that decide what their common correlations are. These mechanisms should also be available to those deciding how to aggregate tags. They may, for instance, want their tags to match those used by the peers they work with on specific tasks. That should be encouraged…