Our Mind the Taxonomist

So finally after the many pages of blogs and thoughts, I think my main points finally came into focus after the last blog, after Clay Shirky’s clear and direct claim which I reiterate here:Here’s what’s radical about what del.icio.us protends: My vocabulary on del.icio.us folksonomy is personal, not vernacular — no one knows or needs to know which class I’m talking about when I tag something ‘class’, or that I use LOC to mean Library of Congress. This isn’t the same as, say, the dictionary of thieves slang from the mid-18th c. because no one else needs to know my bookmark system, and I don’t need to know anyone else’s,

So here is what’s radical about what Shirky protends: if I want to, I can tag my bookmarks with any vocabulary I chose. So here goes, a sample of my tags: “hdfjkfb”, “orjfkido”, “hjfoå”, “krlofpke”. So this certainly ensures that no one else knows my system, and if everyone else does the same, I won’t know theirs. But how useful would this be to anyone?? So at this rather radical extreme the Shirky claim is without content.
But maybe the example is too radical because the notion of vocabulary precludes the use of random tags. Fine. Let us think of a different example which is perhaps closer to the spirit of Clay’s idea. So I make up a system that only I know, where all interesting sites end with “xyz”, technical sites begin with “krp”, and so on. Everyone else can make up their own system, and no one knows anybody else’s. But an immediate problem with this is that it is cheating … it smuggles in the more natural English vocabulary via the back door by simply equating each English term with an expression in the new “vocabulary system”. Still, this is private knowledge so maybe that is O.K. for the example. But this leads to the interesting hypothesis: suppose you made up vocabularies like “krp…”, “…xyz”, and so on, and got users to adopt them. These can be mapped onto their “parent” vocabularies that they are derived from (remember “krp…” = “technical site”, etc.). To the individual user, each system would be equivalent (more or less, as Katie Melua would say). But which vocabularies would make a better del.icio.us?

I think it is pretty clear that we DO need to know SOMETHING about everyone else’s bookmark system! The success of a “social bookmarking system” depends on the fact that we do understand each others vocabularies (to some extent), and can extract value from that shared understanding.

What I have been arguing is that the aggregate data shows us something about what we know about everyone else’s bookmark system. Why are some terms so “natural”? Do all “natural” terms become popular? Are all “natural” terms natural in the same way, or are there different categories according to which terms can naturally relate to their categories (a bit like facets in taxonomies).
So my position is that folksonomies can provide interesting data which can give a clue about the way humans organize knowledge, AND about the ways in which we share each other’s organizational systems. My Ph.D. was originally in Cognitive Psychology, and during my studies I came to believe in the hypothesis that mental architecture fundamentally shapes our perceptions and organization of the world in which we live. Further, essential aspects of the mental architecture are fixed and therefore shared by all humans, which is what makes communication and shared understanding possible. The overlap is not perfect. I say Library of Congress, but Clay wants to say LOC. Good for him. But pity the poor soul who calls it “the square root of negative 2”!
The mind creates categories, because that is what minds do (there is an awful lot more to say about that!). The mental architecture enforces the range of possible ontologies and taxonomies that we can bring to bear on the understanding of our universe. All humans share fundamental aspects of mental architecture and therefore properties of possible taxonomies. Folksonomies provide a fantastic window into the worshop of the mental taxonomist!


More Differences

Just when I finished my previous post on the problems with focusing on individual differences, I discovered this post , once again by Clay Shirky. He makes a good point about why it might be important to think about individual differences when it comes to tag use. He says:

Here’s what’s radical about what del.icio.us protends: My vocabulary on del.icio.us folksonomy is personal, not vernacular — no one knows or needs to know which class I’m talking about when I tag something ‘class’, or that I use LOC to mean Library of Congress. This isn’t the same as, say, the dictionary of thieves slang from the mid-18th c. because no one else needs to know my bookmark system, and I don’t need to know anyone else’s …

So pretty clearly the view is that an individual can do whatever s/he wants, in complete isolation from every other user, and the “radical” system will accommodate each individual equally. Thus, while many people might agree, the system is “ensuring that the emergent consensus view does not have to be pushed onto any given participant“.

So this is why individual differences matter, because they are allowed to exist. Taxonomies are bad because they force everyone to be the same, folksonomies are good because they allow everyone to be individuals. This is probably good ideology, but is it good science? Is it even interesting? Is it useful?

These are the questions I have been asking all along. Where, pray tell, is the evidence that highly unusual tags that differ from the “norm” are even useful? Here are some of my less frequent tags: adhoc, bar, and controlled. I have NO IDEA what any of them stand for! Maybe I am a bad tagger? So ignore me.

Here is an observation based on a single example (so take with grain of salt .. the same grain you use for all other single examples!): a few weeks ago the most popular tag for the New York Times was “news” with 2093 instances. The next few are “newspaper” with 550, “daily” with 370, “nyc” and “media” with 229. These tags kinda make sense .. but after this you have “english” with 15 instances, “business” with 17, and “noticias” with 16. These last few are a real mix. I wouldn’t be surprised if “business” turned out to be useless for most users: it sounds like a tag you add when there is pressure to improve on just “news”, but probably never used as a retrieval aid! (Clearly an empirical question, but would someone so interested in business news ever bookmark news sources that DIDN’T have business news??). “English” and “noticas” are interesting in that they appear to cater to internationalization needs. This is definitely an important requirement for category systems, which many don’t address adequately.

But the question is, is there a significant benefit after “news”, “newspaper”, and say, “daily”? How many users would have so many links that they would not locate the New York Times without also including “business”???

But apart from the usefulness issue, the really interesting observation is still the remarkable overall agreement. Surely the New York Times is “daily”, and it includes a “business” section .. so these are all equally valid features by which to reference it. But why do (roughly) 10 to 200 times the people prefer “news” to the other two?

And then there is my thought experiment. If the evil millionaire convinced 100000 people to tag the New York Times as “finglewick”, how long would that survive? What about “really super cool site”? What about “should subscribe”? Why are some of these tags good but others not? Why not “finglewick”? Why would some good tags, that would probably generally be judged as appropriate (like “should subscribe”) not survive (in my opinion) while some others, like “news”, definitely do??

Del.icio.us lets us label the New York Times any which way we like. But in spite of infinite freedom we all label it “news”. Now THAT is news to me!

Individual Differences and the trouble with the Standard Social Science Model

I have just realized that in my thoughts about all this tagging, I have really been focusing on the historical and other forms of aggregate data, and not really been worried about individual users. This is really in contrast to the analyses carried out by the die hard tags vs. categories types. A prototypical example of this is again Clay Shirky’s (classic?) post. Here are the titles of his figures illustrating del.icio.us: “Tags per user”, “A single user’s tags”, “Different tag ‘signatures’ for different URLs”. O.K., so he only has three graphs. But a full 66% of them refer to single users.
This reflects differences in bias …. no, not really bias …. but attitudes toward the appropriate level of description at which phenomena are best understood. I remember the moment my attitude suddenly made sense to me, when I came across the brilliant book containing an introductory chapter by the psychologists John Tooby and Leda Cosmides (the last of whom I met at a lunch once, to my great pleasure. Leda is a brilliant and really nice lady). In presenting their Adaptationist view of cognition, they argue that the traditionally popular Standard Social Science Model in which human minds are presented as “blank slates” to be written on by experience and culture, is simply wrong. If I remember correctly (this is reaching back a few years!) they illustrate their argument with the failures of Anthropolgical research in much of the 20th century. The problem is that researchers were so keen on identifying the vast power of cultural differences on shaping human behavior, they over focused on finding those differences at the expense of ignoring the similarities. So in one culture men marry one woman, in others they marry 100. There is clearly a difference here, but is that difference more important than the striking similarity that both cultures have some concept of “marriage”? The truly important observation made by Tooby and Cosmides is that there is an absolutely remarkable series of commonalities between ALL cultures on the face of the planet; similarities to do with affiliation, punishment, and so on. How can this be? It is the “aggregated” similarities rather than the specific differences that pose the really interesting puzzles.
This observation is in many ways similar to the works of the great linguist Noam Chomsky, who changed the face of linguistics when he postulated that the really important part of understanding language is to understand the nature of the mental faculties that must exist in order to learn and to use human language in the way that humans do. The study of these faculties has shown that the knowledge of language in speakers of all human languages probably overlaps a great deal, and that the apparently vast differences between languages are due to relatively unimportant factors like parametric variations and different vocabularies. I am being somewhat cavalier in this, I know, but let me illustrate the basic point with my previous linguistic example: it is true that different individual English speakers might chose to say either “Who do you want to go with” or “Who do you wanna go with”. We could of course have all sorts of theories about the sorts of people would want to say one or the other. Or we could say “look .. I really (really) couldn’t care less about why you would chose to say one or the other”. What is really interesting is that both speakers will admit (unless they have some sort of non-linguistic psychological problems!) that either way of saying it is fine. Furthermore, they will also both agree that “Who do you want to help you” is good, but the superficially similar variation “Who do you wanna help you” is not good. All of the sudden questions about the superficial choice of different surface forms is superseded by a much deeper question about the nature of (shared) knowledge that allows the expressions in the first place. Of course the differences between languages is less trivial, and we can learn a great deal about language by the ways in which different languages can differ. But again it is the generalizations and categorizations of the kinds of possible differences that is important.
So maybe the point about all this to tags is obvious? Well, for what its worth … I have a feeling that the aggregate information is telling us something interesting about the way people (in general) use tags, that will subsequently also help us understand why a person (individually) chooses to use them in their own unique (to a point?) way.
So in a way this is just an attitude (bias?) that I have. But the wonderful thing about science is that our respective biases are eventually tested and weeded out. Does one level of analysis give us deeper understanding than the other? Which one has more predictive power, in the sense that we can uncover more new facts and explain more old ones?? In some cases like linguistics the answer is fairly clear, in psychology less so, and with respect to tags … we have a long way to go.
But I am optimistic about my attitude. Here is a prediction I can make. In an earlier post I suggested an experiment in which an evil millionaire who wants the whole world of computers to bend to his will (hmm .. sounds familiar!), pays a lot of people to insert his favorite tags into popular web bookmarks. So he pays people to insert tags like really_super_cool, must_have_for_my_birthday, blue, and so on. What would happen? My feeling is that, once he stopped paying, the popularity of these tags would quickly drop off. How could we explain that if it were true? Again my feeling is that every possible explanation would eventually boil down to the claim that people just don’t like tags like that to describe resources, in the same way that they like tags like news. Why not? Well, we are back to the linguistics-style argument again.
I have other predictions of course, which Joshua could easily confirm or dis confirm. I bet people mostly click on the big words in tag clouds. But I am not too deeply committed to this feeling. On the other hand I bet more that when people search with combinations of tags, they tend to use relatively few and high frequency tags. And I bet the aggregate data would show this!