Autodiscovery of hierarchies through tagging

There’s been a growing amount of thought recently on how to use the data from folksonomies to organize information in other ways. One blogger suggests data mining folksonomies to populate a hierarchical taxonomy. Through David Weinberger I find that Library Thing notes that tags standing alone don’t always provide as much context as a taxonomy.

The ordered structure of subject headings gives added meaning. History > Philosophy is very different from Philosophy > History — a distinction that isn’t necessarily apparent when searching history or philosophy separately as tags.

Expanding on this concept a bit, I’d note that with a detailed taxonomy it will be difficult to rely soly on a document’s tags to determine where within that taxonomy the document should go. Software > Design is a very different concept than Design > Software (the act of designing a piece of software as opposed to a piece of software used for design) and if your document simply includes the tags design and software how would the classification engine decide where it goes? With a larger tag set it should be possible to infer additional meaning. The additional tags could provide clues as to whether you’re talking about a software product or the process of creating software.

It’s possible for instance to look at the tags mustang, ford, horsepower, specs and determine that this is about the Ford Mustang automobile instead of a horse and that mustang, running, animal are tags that don’t describe a car. But building a taxonomy based on these tags will require that a person with knowledge about automobile brands build a dictionary of tags that map to nodes in their taxonomy. Once you have a corpus of tags that do and don’t place a document into your taxonomy you can start using machine learning — perhaps through statistical analysis — to learn what other tags would define a document as either a horse or a car.

While part of this problem can be solved by using sufficiently large and detailed training sets, but it’s going to be a lot of work. You can’t assume that training a learning algorithm where tags belong in one taxonomy will map well to other similar taxonomies.

Taxonomies are structured for a particular knowledge domain by someone with detailed understanding of that domain. An algotrithm isn’t able to understand nuances in how I want my content structured. It is unable to adjust for my personal preferences and biases. Tagyu tends to classify pages about podcasts as Entertainment but you might feel that everything podcast-related is better suited for the Technology category. With this sort of problem in a generic flat taxonomy, imagine how much more complicated it would be to place items into a detailed hierarchy with lots of similar nodes.

With Tagyu, I’m working on some solutions to this, and I’m excited to see what others are coming up with.