![$tag[$tags] = $tags;
Aaron Straup Cope
Visual Web Meetup July 2009
This is an actual part of the Flickr code base. No one can bring themselves to change it now.](https://image.slidesharecdn.com/visualwebmeetup-090721212756-phpapp01/85/tag-tags-tags-1-320.jpg)
$tag[$tags] = $tags;
- 1. $tag[$tags] = $tags; Aaron Straup Cope Visual Web Meetup July 2009 This is an actual part of the Flickr code base. No one can bring themselves to change it now.
- 2. aaron My name is Aaron. The shortest possible introduction is that once upon a time I was still a painter, and then the Internet happened.
- 3. ?ickr These days I am a senior engineer at Flickr. We are a small (ish) photo sharing website, specializing in pictures of cats. And other things. We are also known for the many tags that our users have added to their photos. We have about 40M unique tags.
- 4. del.icio.us We didn¡¯t invent tagging. We ¡°borrowed¡± the idea from Joshua Schachter¡¯s social bookmarking website del.icio.us.
- 5. keywords facets topic maps categories ontologies It¡¯s important to remember that Joshua didn¡¯t invent tagging either. We¡¯ve been chasing systems and forms of classifying information for as long as we¡¯ve been collecting anything worth calling information.
- 6. ponies It¡¯s been a bit of a wild goose chase really. More than anything, formal ontologies outside of so-called domains of expertise are hard to master and, if we¡¯re being honest about it, boring to use.
- 7. tags (good enough is perfect) But del.icio.us offered tangible proof that if you make the process (for adding tags) simple enough and provide tools for managing those tags then people will participate. Small tools for self-organization. I¡¯ll come back to this idea later on. We added tags because it provided a fast, cheap and easy way for our users to catalog their photos. If that were all tags did, though, they wouldn¡¯t be that interesting. They also double as a kind of foot-bridge between users and meaning; little rabbit-holes of serendipity.
- 8. tag clouds One of the earliest tools for managing the volume of tags was a text-based visualization called a ¡°tag cloud¡± where the size of each tag displayed is relative to the number of photos associated with it. We¡¯re sorry about tag clouds.
- 9. hawt tags Eventually we started to experiment with a variety of algorithms for detecting new and interesting tags. It¡¯s worth noting that there hasn¡¯t been a day since I¡¯ve started working at Flickr when ¡°wedding¡± wasn¡¯t the top tag so that should tell you something about ranked lists.
- 10. tag clusters We also added tag ¡°clusters¡± which are generated nightly by analyzing the entire corpus and feeding them through a variety of hierarchical clustering algorithms. Clusters are good serendipity magnets. As a rule, I ?nd the associations between the different clusters more interesting than the associations between the set of tags in a given cluster. Maybe that¡¯s just me.
- 11. tag maps In 2006 the Yahoo! Research Berkeley (YRB) team released the tag maps project that generated a dynamic, map-driven interface to Flickr photos by analyzing their tags for geographic information. We implemented something like that, incorporating the work we¡¯d done with hot (¡°hawt¡±) tags, for the second iteration of the Flickr map. More recently, researchers have expanded on the work done by YRB and published a really fascinating paper called ¡°Mapping the World¡¯s Photos¡±. I¡¯m not going to talk about it now but the paper is de?nitely worth reading.
- 12. machine tags In 2007, we added formal support for ¡°machine tags¡±. Machine tags are really nothing more than regular tags with a special syntax to denote a faceted relationship: a namespace (or a subject domain); a predicate (or a subject topic); and a value. Our users had already been adding tags using a machine tag like syntax and then parsing out the structure, and the meaning of those tags, themselves using the Flickr API. What we added was the ability for Flickr to recognize and index the different pieces of a machine tag and to allow users to search for them across the entire corpus of photographs accordingly.
- 13. extras! machine tags Machine tag extras are we refer to as the process of using the value of a machine tag to look up data in another service (as de?ned by the namespace) and squirting that information back in to Flickr. For example we recently added machine tag extras support for the Open Library so that when someone tags their photo with an Open Library identi?er we can display the name of that book.
- 14. wildcard tags We also added the ability to query for machine tags as part of a plain old URL. For example, if you want to see all the photos that people have taken at places to eat in the Dopplr Social Atlas you can just go to: http://www.?ickr.com/photos/tags/dopplr:eat= Or all the photos taken at Upcoming or Last.fm events: http://www.?ickr.com/photos/tags/ *:event=
- 15. linked data For anyone familiar with the idea of the Semantic Web machine tags might look a familiar but somewhat causal implementation or a variation on the theme. They are. Machine tags try to provide some of the bridging facilities of the semantic web but without forgetting the original lesson that del.icio.us offered: Keep it simple.
- 16. commontag (.org) Recently a project called Common Tag has been launched. It seems to be a short form for addressing authoritative topic descriptors in web pages. I haven¡¯t decided what I think about it.
- 17. tagopedia In 2006 Dave Beckett presented a really great paper called ¡°Semantics Through the Tag¡± at XTech, in Amsterdam. One of the ideas Dave proposed was setting up a Wikipedia-like site for tags, to document their many meanings and uses. He chose Wikipedia speci?cally because that community has developed lots of tools managing con?icts and mechanisms for disambiguating concepts. I think it¡¯s a great idea. Unfortunately, neither Dave nor I want to actually run the site so if someone here wants to take on that responsibility I think it could be a really valuable resource over time.
- 18. equivalencies That would certainly help managing equivalencies in tags, whether it¡¯s equivalencies in concepts or just across languages. We don¡¯t do anything like that right now on Flickr. Personally, I¡¯d like to allow users to de?ne equivalencies between tags but we haven¡¯t been able to think of a way that would be easy enough to warrant doing.
- 19. lexicon But really what we have is this fantastic lexicon of terms and connections that keeps growing every day. We make a point of trying to expose as much of that information as we can via the API and are eager to see someone tease out the shape of language on Flickr. That¡¯s some of what we¡¯ve done with things like hot (¡°hawt¡±) tags and the clustering but there¡¯s still plenty of interesting possibilities to explore.
- 20. ?rst class objects There is also the question of when and why a tag evolves in to being a ?rst class data type and whether that¡¯s actually re?ected in how people use tags. Dates are one example, and geotags another. Each are uniquely indexed in the Flickr database and, still, people continue to add both as tags on their photos. The short answer, of course, is that it¡¯s usually just easier to type 2008 or 2009 than to try and remember a specialized syntax for doing searches.
- 21. magic words That¡¯s the funny thing about language.
- 22. tag as ¡°small horses¡± tag as ponies There¡¯s been a really interesting discussion on the FlickrCommons group around a blog post written by Larry Cebula questioning the limits of user contributed tags, notes and comments citing the volume of conversational additions like ¡°cool¡± and ¡°awesome¡±. Another camp argues that the value of user-contributed data comes from not simply analyzing the photograph itself but analyzing the activity that surrounds photograph a photograph.
- 23. play (social objects) The photograph is a ¡°social object¡± around which people can use tags, and notes and comments, to have a conversation. The different kinds of metadata are devices for shuttling the discussion in a variety of different ways. While the signal to noise ratio can often be higher than researchers are used to the contributions from the ¡°commons¡± have also proven to be valuable and rewarding.
- 24. openlibrary:actionshot= And the value of play as a motivator shouldn¡¯t be underestimated. I mentioned earlier that we added machine tag ¡°extra¡± support for the Open Library. This prompted one user to ask (the Open Library staff) whether they could, or should, tag a photograph of themselves reading a book, rather than the cover itself, with an ¡°openlibrary:id¡± machine tag. The answer was: Of course, why not! It is early days and we can still make our own consensus so let¡¯s see where it goes. Or maybe tag it as an ¡°action shot¡± instead.
- 25. horse=yes It¡¯s not as crazy as it sounds. The Open Street Map project whose mission it is to map the entire world uses just this approach and they¡¯ve been surprisingly successful. In just ?ve or six years they¡¯ve managed to produce a dataset in the UK nearly as good as the Ordinance Survey which has had By using tags like this. No, really.
- 26. time:hour= We¡¯ve used tags as part of the Flickr Clock, a visualization of videos uploaded throughout the day, that was created for us by Stamen Design. Most videos don¡¯t have very much useful metadata, including the day or the time they were created. We were able to use machine tags to give people a way to add structure date/time information to their videos which was then interpreted by the Flickr Clock application. In the process we were able to teach people how to add tags and how to use them and, hopefully, see their value. I¡¯m pretty sure (or at least like to believe) that the moment people understand how something is useful for them is the also the moment they start to think about how to play with it and how to use it for something entirely new.
- 27. discovery This is also called discovery.
- 28. No one said it was easy.
- 29. nubby bits Formal ontologies are useful when you know the boundaries of your domain. I¡¯ve seen people in the public safety sector get very excited about them because it means they can keep track of where are their ambulances are. This is a good thing. I want them to know where the ambulances are. But it¡¯s a pretty brittle approach when applied to something as wide-open and open-ended as the Internet and even more so when it involves communities from all over the world coming together to share and discuss their photos. We try to be mindful of building what a colleague described as ¡°small tools for self- organization¡±. Tags are one such tool because there¡¯s just enough convention (language) for people to have a common ground to operate on but still have a rough enough surface to hang new and wacky ideas off of.
- 30. thank you