Tagging -2.0?

I was at first impressed with the power of folksonomy-based tagging (in the sense of allowing users to invent their own taxonomies and metadata for information objects). But now I’m not so sure.

I just attended a SXSW panel called “Taxonomy 2.0”, with Tom Vander Wal, Prentiss Riddle, Rashmi Sinha, Adina Levin and moderated by Don Turnbull. They had a lot of really interesting and thought-provoking ideas about folksonomies and free-tagging, and I learned a lot.

But something stuck in my craw about the concept itself: The whole point behind the technological development of search engines, especially their artificial intelligence and natural language capabilities, is that humans shouldn’t be the ones indexing information. Folksonomies seem, to me, like a step backwards.

Put Your Audience to Work

The first thing to note about folksonomies is that they already work. Millions of people already spend a little bit of time every day manually adding metadata to their own and, more importantly, other people’s information. In addition to adding tags to my own data, such as this blog post, I can also add tags to a web site URL in delicious, a photo in Flickr, a news article on CNN or Newsvine, etc.

It’s that latter part that’s amazing to me: People are actually doing (free!) work for other people, adding metadata to information where the information’s “owner” could have done that work. The brilliant thing about folksonomies is that internet users have shown themselves time and time again to be remarkably willing to do their part to help the greater good, even if it means doing labor that happens to bring financial benefit to someone else. I wont go into any specfic examples of this (just think open source, freeware, file-sharing networks, etc), but I think you can see this effect in full force every day on the web. Nobody ever went broke underestimating web users’ willingness to geek out and do free technology work.

I think this community participation effect is a good thing, and it is one of the best strengths of folksonomies.

What are Taxonomies for?

I think my problem lies, instead, with the basic purpose of taxonomy and the role of taxonomies in the contemporary digital environment. So first here’s a little bit of my understanding of how we got here:

Taxonomies were originally intended to allow human beings to easily take a peice of information and mark it up in such a way that it will be easily retreiveable by another human being later on, usually as part of a larger hierarchy-based classification scheme. A controlled vocabulary of keywords could be used to empower, say, a librarian to know where and how to classify a book, and where and how to find it again later on.

Later, in the digital land of databases and web sites, this system became keyword-based metadata, permitting machines to do the same sort of finding of information: by adding a limited number of keywords to the meta tag of your HTML page or database entry, you could allow an old-fashioned search engine (one which ONLY read the metadata) to find your information object based only on a keyword search. Later, when search engines could parse and index a whole page’s contents, the metadata would act as “hints” to help identify the content a little more specifically.

But Taxonomies Are Obsolete

This is what we do when we add folksonomy tags to information: we’re adding “hints” to help other users and search engines find our information. My problem, however, is that tagging itself is already old-fashioned. Google long ago stopped looking at keyword metadata because their search engine has for a long time been using far, far more powerful automated methods of indexing your HTML page. Are there any tag-based search methodologies that are superior to Google’s full-text (and more) search engine?

So my mistrust of folksonomies belies an inherent distrust not in the “folks” part, but in the “onomy” — that is, that I’m starting to think that the whole model of defining any taxonomy at all to assist in the findability of information may be a historical relic already well on the way towards being replaced by complex and powerful tools, and that we shouldn’t still have a real need to bother to create any sort of human-readable classification schemes at all.

Machines to Replace Human Labor

There are plenty of great technologies already in existence, and even more powerful technologies right around the corner, that make taxonomy- and keyword-based tagging look almost midieval to me.

• Credibility-Based Indexing is the original heart and soul of Google, where information sources are valued more as search results when they are cited more often by other information sources (for example, a site with lots of people linking to it is more useful than a site with only a few people linking to it).
• Context-Sensitive Indexing allows a search engine to recognize that an article about courage and perseverence in the sports section of the newspaper might be a good result for someone looking for articles about athletic coaching tips.
• Error correction, exemplified by Google’s “Do you mean xyz?” feature, makes dumb obstacles like misspellings a thing of the past.
• Natural Language Processing, or NLP, where a search engine can read a text and deduce the most important keywords automatically, recognizing which words best represent the text’s meaning using complex language-recognition algorithms.
• Concept-Based Indexing, which is like NLP on steroids, is when a search engine posesses a knowledge engine able to recognize when search terms are related to one another according to a limited, or even unlimited, body of relational concepts. For example, a book search for “web design” might return a book entitled “HTML for Dummies” because the engine knows that “HTML” is conceptually related to “web design”.
• Artificial Intelligence, which is like NLP on PCP, can allow even more sophisticated matches between a human beings query and the appropriate results. AI will, hopefully, make the search interface user experience itself simpler, saving us the trouble of having to “think like a machine” when we craft our search queries and think of which keywords to enter into that search box.

Indexing Non-Text Objects

It’s no coincidence that Flickr is probably the most successful example of a folksonomy. Images cannot be indexed by any of these fancy search engine technologies described above, so for the time being humans must do this work ourselves. Likewise, the same situation applies to indexing music, video, interactive applications, physical locations, and other data points. We simply cannot effectively index these things without human intervention.

But this is changing, too.

As I’ve previously noted, we are rapidly seeing advances in image recognition technologies, in particular in face recoginition. I’ll guess that a third of the image tags used in Flickr are people’s names, referring to who is depicted in a given image. With new commercial tools like Riya, we can have software that automatically identifies the personages in each image on our Flickr account, saving us lots of tagging time and effort.

For music, the MusicBrainz project allows software to do an audio waveform signature scan of a music file, in any format or quality, and detect what song it is in much the same way that CDDB works. No need for any end-user to manually tag any widely-released music file, at least not if it was done initially by the music’s creator.

And for video, closed captioning is a ready-made text index we can start using today. Admittedly this is also human-generated work, but if it already exists for a lot of video, a search engine can parse that text for meaning and derive a better index than keyword-based tagging can.

Tom Vander Wal’s diagram (photo by Andrew Huff)

Power to the People

Several people said to me that the real benefit of folksonomies is that they are “me”-centric. That is, that you can idiosyncratically invent tags that may not make sense to a computer, or even to other people, but that would make perfect sense to you. Tom Vander Wal had a diagram that tried to capture this aspect of the system, too, but it’s admittedly a slippery thing to describe. De.licio.us is a great example of this, where one is generally tagging a URL for the purpose of reminding oneself about it later. For example, if I’m doing research for a project for Client X, I might tag a site “Client X” to help me find it later, even though this tag probably wouldn’t help anyone else find the item (in fact, it might hurt their chances of finding it!).

[This is what Tom has called “refindability”, i.e., the common problem of finding something you’ve already seen before. I think this term is awkward, however, because re-finding an object should be, and is, the same as finding. I think what Tom means is “finding something you merely stumbled upon the last time you saw it”. Either way, a search engine must be able to find something based on your search query, whether you’ve seen it before or not.]

Adam Greenfield points out that folksonomys allow us to create “meta messages” through folksonomic tagging, for example giving “shout outs” to people via tags. To me, when these tags are appropriate to the information, this is no different from normal metadata, and as such is as useless as normal metadata. When these tags are inappropriate, for example tagging a photo of Adam with “ninja”, it only serves to muddy the waters of findability. It’s cool and interesting, but not helpful.

Two Solutions

When I thought about how we can still leverage the community’s willingness to participate in helping the findability of objects, which again I think is the whole point of folksonomies, two things came to mind:

• Commentary Tagging: If you think about it, tags are kind of like one-word comments. And if you think about it more, comments are metadata! So why limit folksonomies to single words, or why even “ghettoize” community participation to metadata? By permitting verbose commentary, we add context and meta-information to a data object. Every blog post that has extensive comments attached to it is exponentially more indexable, and thus findable, than a blog post that stands on its own or is accompanied by a handful of tags.
• Star-Based Tagging: When we tag an item with 5 stars, indicating that we really appreciate the information, we are adding a measurable credibility rating to a link. This is something that only a human being can do. This credibility rating is useful both to the individual who made the rating, to help them find something they found interesting in the past, and to the community as a whole, to help them find information that other people found credible.
  (This also solves Tom’s refindability problem: Any site I’ve even looked at before could by default get one star, and more stars if I manually say so. When doing a normal search for the information I want, sites with one or more stars would rise to the top of that list of results.)

Linking these two ideas is also possible: rating comments with stars to help a search engine determine how much to weight one comment over another when building an index for the data object.

UPDATE: Apparently the Flock browser has several features that do precisely this. The “Star” feature in Flock is their version of bookmarking, and it seems very cleverly designed. It quickly marks a page as interesting, and drops it directly into your delicious or shadows account. Also, Flock indexes your browser history, which in combination with the Star feature, seems to be a “killer app” of refindability. Flock does tagging, also, but I think they were wise to make tagging an optional part of marking a page for future reference, realizing that the combination of an indexed search engine with the star feature ought to do the job without needing to resort to referencing metadata tags.