Nesstags
A Tagging Proposal, May 2005
Oveirview of Tagging
The internet is packed with content of many forms including images, text, music, animations, videos etc. Modern search engines have taken up the enormous task of organizing this content and they have made it surprisingly accessible to the internet community. Although existing search tools are nowhere near perfect, they are remarkable technologies. At its core, a search engine (SE) works by parsing web pages and running the content through a algorithm. It is this black box that decides how to rank the billions of pages on the internet.
Most ranking algorithms are designed to mimic what a human viewer would want, but it has to perform this task without actually understanding the content. Before artificial intelligence becomes available, only humans are well-suited to fully understand some bit of content. We can know what a particular page is about, but Googlebot (the spider that Google sends out to crawl the pages on the internet), can only make 'assumptions.' These systems analyze factors like incoming links, keyword density and page titles. Despite this limitation, the algorithms do an amazing job of approximating what it would be like to understand some content. With this approximation, the algorithm decides which results to present first and which to suppress. From the millions of pages about digital cameras, for example, it has to determine which to present to the user at the top of the results page.
While present day search engines are adept with text, they do not perform well with other types of media, like pictures. Needless to say, however, Google, Yahoo and MSN are hard at work at figuring out how to get a computer to be able to mimic the interpretation of pictures, audio and video. For the time being, however, humans are far superior at this task than any computer algorithm.
This is where tags come in. In one sense, tags are a competing schema to search algorithms. Tagging is not a new idea; in fact the father of the internet, Tim Berners-Lee, has been evangelizing tags, in the form of the semantic web, for many years. In its most general form, tagging refers to the association of some meaningful snippets to some underlying content. flickr, for example, allows users to tag the photographs that they post online. These photographs are made publicly available and you can search through the tags to help find related photos.
A user that posts a scene of the golden gate bridge in rain, may provide the following tags: bridge, San Fransisco, rain, red, clouds. anyone that subsequently visited the site, and searched for "bridge," would be given this picture as one of the results. The tag, therefore, attaches some meaning to the photograph that a computer algorithm couldn't. As more and more people post their pictures, you start to get results for a search for "red bridge". You can start to see the power of this, because with a very simple step of adding tags, a huge number of photographs become very easy to catagorize and search through.
del.icio.us, a website that allows people to save their favorite website addresses, uses a similar idea. instead of organizing ones bookmarks into a folder, they are added to their free account associated with one or more tags. these bookmarks are shared publicly. based on the belief that people will only bookmark valuable links, you can begin to see the power of such a bookmarking folksonomy. A search for "programming" will yield a list of valuable programming sites.
Can flickr be better than Google Images? Maybe. Can del.icio.us be better than Google's search engine? Probably not, but it provides some benefits that Google cannot offer.
The Problem with Tags
It is clear that the potential of tagging is vast. There is, however, at least one significant failing. Although the assignment of a tag informs the system about the nature of the content, the system has no way of knowing how blue the picture is. When you perform a flickr search for "blue", you will find some very blue pictures listed alongside some not-so-blue pictures. The system is not provided with a method of sorting the results of the search - everything with the same tag is treated as an equal. del.icio.us has the ability to rank bookmarks based on the number of people who tagged them, but because many people bookmarked a particular site, means that it is a valuable site, not that it is particularly blue.
When there are only a few items in a database, the lack of ranking is not very relevant. If there are 5 pictures associated with a tag that you have searched for, who really cares which is presented first and which is last. What happens when there are 3,000 pictures, or 30,000 pictures? Enter nesstags.
The Nesstags Proposal
When we search for "blue" pictures, we naturally expect the pictures with the most blue-ness to be presented first; as we scroll down the results pages we should be presented with pictures of decreasing blueness. Since current computer systems have difficulty assigning characteristics such as rainy, smile or depressing, the human tagger would do it. A picture with a small bridge in the distance would would be given low bridge-ness score, but a direct view of the Golden Gate Bridge should be tagged as overflowing with bridge-ness.
I have a texture site that uses such nesstags so you can observe the results. Notice that a search for rust is sorted from the most rusty to the least rusty - you may have to navigate to the subsequent results pages to arrive at the slightly-rusty and not-so-rusty textures. In the interest of full disclosure, the ness scores for each texture are provided by a closed set of editors, so the quality of the sorting is particularly high; there is no reason to believe, however, that nesstags in an open folksonomy wouldn't drastically improve the system.
But doesn't this type of categorization defeat the simplicity which is what has made these folksonomies so successful? In a word: no. Nesstags don't have to be complex. How about a one digit number attached to the end of each tag. A "red9" tag would be assigned to something that is extremely red, and "bridge1" would be given to a picture of a bridge in the distance. Don't forget, that multiple tags can be given to a single item. A close-up picture of a zebra might be given the following tags: striped7 furry6 zebra9
Of course, this sort of scoring is open for abuse, but since most folksonomies are based on people's own collections as well as the collective collection, they have something to lose themselves. If a user rates something not-so-red with a red9, then searches through their own private collections, in which they have a particular interest, would also be flawed. In other words, everyone is incentivized to buy into the system. Appending every tag with a one-digit score is only negligibly more difficult; the long term benefits of nesstags, especially when the databases grow exponentially larger, could be substantial.