sardonick

It hit me after installing tomboy a little while ago:
the current state of tag-based information representation theory, as painted by flickr, gmail, delicious, etc., is that of tags as very limited-scope metadata. What I mean is this: Given a document, say this blog entry, I have different scopes of metadata. The “date” metadata item (2004/11/08 @ 18:40 PST), the subject metadata, and this “category” metadata (“meta”, “thought”) that tries to implement the same sort of tag-based categorization that flickr and gmail and all our organization-by-tagging friends have.

And I’ve said before that I can’t wait for tags to be better-implemented in blogs and wikis and (filesystems too!).

But should we look at tags this way?

What about looking at tags as entities just as solid as the things that they tag. This would shift the paradigm from:

document has-date-metadata $date
document has-tag-metadata tag1
document has-tag-metadata tag2
document has-tag-metadata tag3
document links-to document2
document links-to document3

to

document links-to tag1
document links-to tag2
document links-to tag3
document links-to document2
document links-to document3


The difference is subtle. In effect, it makes it so that tags themselves become documents, and documents that a page links to (and that a page is linked from) become tags. Tags become a sort of generalized URL, and documents become a sort of specialized tag (This makes a certain sense when you think about it. Everything that external page that I link to on my www home page provides some information and perspective on what I want my home page to represent. And every page I link to on my blog defines, by slow, exact, painstaking enumeration, the big-picture content and aim of my blog). Do you see what this all does? It, in effect, blurs line between context and content.

So, how can we make this work for us, practically?

Assuming unidirectional links:

1. “view all things that this document links to”: this returns the current tagset of the document, where normal tags function as they did before, and where linked document URLs are a sort of very specific tag.
2. Because tags are documents as well, tags of tags become a structure to our ontology. “Windows” tag and “Linux” tag could both link to “OS” tag, “computers” tag, etc. (Does this work? Directed, binary relationships between tags could possibly be a solution to creating an ontology that is both successfuly emergent and hierarchical).
3. “view all things that link to this document”: this view is the opposite that we are accustomed to when browsing the www. But in the case that “this document” is a tag, it returns all objects that had been given a certain tag.
4. It doesn’t have to stop at user-created tags. Links could be created automatically, say by document similarity clustering, which could further add structure to our emergent tag ontology.

(I might be inspired by Latent Semantic Indexing, which attempts to represent the meaning of a word using a high-dimensional vectors in n-space. These meaning-vectors are derived by looking at what words co-occur in different documents. We know that, over hundreds of thousands of documents, “pomegranate” appears in the company of the words “tree”, “ripe”, and is less likely to co-occur with “linux” or “mousepad” or “SUV” or “Manchester”. This says something about the meaning “pomegranate”. Likewise, Bush has been able to form some link between “Al Qaeda” and “Iraq” by using them together a lot ;). Perform a little bit of dimensional reduction and we got ourselves rough vectors that correspond to word meanings. But more talk on Latent Semantic Indexing later. The inspirational bit was that the words surrounding the word you’re specifically looking at are not only content but also “tagging” in a rough sort of way. The line between context and content is blurred.)

Would love to hear any reader thoughts on this! Please!