I have not written in a while, because I have been busy developing a theory about the cognition of tags … and probably more importantly, implementing a system that actually does something useful with what I discover. I will write about the first part here, the psychology part. By the way, there is a popular blog which is also about roughly this topic, but is very different in its scope. Rashmi is mainly concerned with why tagging seems cognitively less effortful than other forms of more structured categorization. Her analysis is therefore not very informative about what the tags themselves are like. Is there something interesting to say about the sorts of categories people use in their tags? If so, can this tell us about interesting things to with those tags by means of post processing? This is what I am going to write about.

First, why I think tags are, with few exceptions, real categories. What sort of categories? Ad hoc ones, of course.

The analysis of ad hoc categories comes from the work of the cognitive psychologist Lawrence Barsalou. He was particularly interested in categories like “things to sell at a garage sale” and “things to take on a camping trip”, which are spontaneously generated categories that group entities in goal directed ways. By comparing these to “natural” categories he hoped to discover some interesting differences. The thing is, he didn’t find too many differences! Both types reveal strong typicality (prototype) effects, which are stable across time and people. As a result he proposed a more general theory of categorization which subsumes both common and ad hoc categories. Here is a quick summary his general model.

Ad hoc categories are made up to fulfil some goal. The critical role of ad hoc categories is to provide an interface with a person’s world model, in a way that can help achieve a goal. A world model is ” … a person’s knowledge of locations in the environment, together with knowledge of the entities and activities that exist currently in these locations”. The world model is not the general knowledge one has about the world, but an instantiation involving “… specific knowledge and beliefs about the current state of the world”, which might include culturally shared information and the like. The primary building blocks of the world model are the common taxonomic categories like bird, flower, chair, and so on. Finally, whenever people wish to achieve any goal, they instantiate an event frame which describes the necessary components for achieving the goal. The successful realization of the goal described by the frame depends on a satisfactory interface between the event frame and the individual world model. For example if one wishes to buy groceries then the relevant frame will include things like locations to find groceries, times the store is likely to be open, forms of payment, and so on. But to achieve this goal we need to know specific locations, times and forms of payment. This is where ad hoc categories provide the mapping, by establishing specific categories like places to buy groceries, and so on. Crucially, “mapping different event frames into the same world model defines different partitions on entities in the world model”, requiring flexible ad hoc categorization. Taxonomic and goal directed categories are two complementary ways to categorize the world: taxonomies describe the relatively stable kinds of things in the world whereas goal directed categories are ad hoc collections of different taxonomic kinds that are created to map particular event frames to particular world models.

One potential problem in applying this theory to the folksonomy data is that Barsalou’s ad hoc, goal derived categories tend to be expressed as multi word phrases whereas the majority of tags are, well, single word tags. Part of the reason for this, on del.icio.us at least, is artefactual since the user interface specifically prevents the use of compound words as tags. Marieke reports that 10% of tags recovered from delicious showed evidence that people were trying to form compounds by using some sort of punctuation symbol to represent a space. For example there are examples like “Devel/C++”, “Devel/perl”. But this number does not include the items in which words are simply concatenated, so the prevalence of complex tags may be quite high, opening up the possibility that complex ad hoc categories are used in folksonomies. But an equally important point is that goal derived categories may become lexicalized with common usage. Thus “buyer”, “payment”, “donor” and “gift” are lexicalized concepts that have an important role in many commonly used event frames. But Barsalou gives no indication of how many lexical items are of this sort, or how to identify them.

This is where my work has been focused, since I believe that very many words are lexicalized ad hoc categories. What remains to be done is to discover what goal they serve .. are there high level abstract descriptions of these goals which might allow us to lump words into semantically coherent groups? Are these the sorts of words people are using as tags?

I think the answer to all these questions is “yes”. Next time I will show why.