This is another post inspired by sitting in a 3.5 hour lecture. This time, it’s about tagging — you know, that thing that people do with Flickr photos, Google images and the likes. (As an aside, in case you’re about to ask, the answer is “yes, indeed. This is what $16,000 in tuition every academic year buys you at a first-class university: hours of talk about things that people $16,000 richer do without nary a thought.”)
There are a lot of problems with tagging, ranging from the obvious — what forms of the words are used — to the subtle — what motivations do people have in creating a tag. As an example of the obvious, suppose what I want to flag that this post as being about tagging, should I use the tag “tag”, “tags”, or “tagging”? What if I made a spelling error and wrote “tagg”? What if I had written a phrase, “tagging post”? What should a tagging management system do to conflate those dissimilarly worded terms into one concept?
The diametrically opposing problem from that described above goes by the name of “homography”, or the same words used to indicate different concepts. Suppose I tagged something as “play”. What could I mean with that tag? Perhaps I mean a theatrical performance. Perhaps I’m referring somebody relaxing at the beach, strumming an instrument, or engaging with a video game. Although context can clarify these ambiguities, for example if the tag were “play Grand Theft Auto“, you would know that I was talking a video game, the problem persists because many tagging management systems automatically splits tag terms into individual words. That is, “play Grand Theft Auto” would be split into “play”, “grand”, “theft”, “auto”. In order to have your tag indexed as you intend, you would need to know the rules of the management system. For example, on Flickr, I believe you would need write the above tag with underscores, “play_Grand_Theft_Auto”.
Then there’s the problem with specificity. If I tagged something as “Beethoven’s Ninth Symphony” this something is not going to be one of the first returned hits if you search for “classical music”. This is just as well when you’re not looking for Beethoven’s 9th, but is not as desirable when you are in fact looking for that particular composition without knowing the composer nor the name of the piece. How much should a tagging system do to assure that tag terms are associated with other meta tags that may make the item more easily located? WordPress suggests tags for posts, but these tags come from the posts themselves and are not relational tags. In other words, WordPress may suggest “Grand Theft Auto” and “Beethoven” as tags for this post, but it’s not going to suggest something I haven’t written here but reflects the relationship this post bears to other concepts.
The most intricate problem in tagging is that of motivation. Any tagging system assumes that a person who tags wants others to use his tags to find the item. However, this may not be the case. A person may mean his tags to carry private meanings, the equivalent of an inside joke. “Beethoven’s 9th Symphony” may be tagged as “dead dude’s notes”, which may make sense for members of a specific social group but not for anybody else. This problem is impossible to solve unless you have many different people engaged in tagging the same item, in which case you can throw out the outlying, idiosyncratic tags.
So this is what you would’ve learned if you sat in class with me for those 3+ hours. Interesting? Certainly (although I bet you’re happy it didn’t take you three hours to read this post), but would we have expected tagging to be without any of these problems? After all, if the adage “a picture is worth a thousand words” is true, then no tag is going to capture the entirety of a picture. Even 1,000 tags aren’t guaranteed to get that job done. (It’s not for nothing that rich, smart Google offers a “search by image” option.)
Furthermore, I would argue that this problem is not confined to images. One would run into the same issues if one were asked to describe Anna Karenina to somebody who’s never heard of the book or its creator. What would one say? “This is a book about a married woman who fell in love with a handsome man who happened not to be her husband and eventually jumped under a train”? Do you feel like you can uniquely pick out Anna Karenina from that description if you didn’t know it already? This is still true if instead of a sentence, I had written a 300-page literary criticism on Anna Karenina without mentioning its name.
The fundamental problem, therefore, isn’t specificity, or homography, or any other “ee”. The problem is that knowledge transmission is difficult. On the way from not knowing to knowing, from searching to finding, from identifying a problem to solving it, is a million dead ends, man holes, and pitfalls. Tread carefully, and best with the help of immense resources like Google algorithms.