sarken : writings

Me & Template Tagging

In order for social tagging systems to be most useful the tags that users place on items need to have some sort of global use. Specifically a common problem of tagging systems is that tags that are meaningful for a specific user are not meaningful to a generic user.

I wrote some of my prior thoughts on tagging over at the kozoru blog last year. (Which I’ve snagged from archive.org and placed here.)

First I should say that the Google engine has improved since I wrote that article … Google hadn’t revealed some of their templating solutions yet and they’ve figured out how to tweak page rank and blogs. Both of those improvements have pushed my criticisms into rare edge cases.

But my complaints about tagging remain, although they’ve been refined. With enough users I now believe you can get around most of the problems of disambiguation. For example, I don’t really care that a single user tags a picture of his cat ‘tea’ - with enough data it will be noise to the overwhelming signal of ‘tea’ being about the beverage. This does assume that the majority of users tag to some common sense notion. Let’s assume this is true — cause there’s still another problem.

What happens when a common sense tag is useless to the crowd, but highly useful to an individual. Specifically, let’s deal with the tag me.

Me is a tag that is extremely useful for my personal photo albums. A quick search for me in my photo album returns all pictures that have me in them. But what about a general search for me?

It returns all photos that any user has tagged me. Try it. Clearly that isn’t the semantic me. There are similar problems with tags like you, mom, dad, parents - but solutions to those are more complex.

If I want to search for a specific person I could miss a match due to the me tag.

Take, for example, this photo by tamelyn.

Notice how a search for most recent photos with the keyword “tamelyn” doesn’t return this match.

Tamelyn has taken the time to tag her photo as me which is a big help to her when searching her personal albums, but a quick template could help the crowd.

If every tag of me had another tag auto-generated with the flickr user’s username or true name (notice Tamelyn revealed her name to be “tamelyn feinstein”) the search would return the photo above.

This is a nice example where simply semantic templates can help user generated tagging. There are a number of other tags that could be lumped into this template (self, selfportrait, us). This information could even be displayed to the user in a slightly different color to indicate that it was auto-generated and a user could remove it if it was inaccurate.

It’s my belief that by creating a few template-based tools you could enhance the value of tagging systems without expecting the common user to become an ontologist.

Collaborative Tagging In Social Networks: A Tagging & Scoring Abstraction

Summary

Collaborative scoring sites like Reddit and Digg demonstrate the editorial power of having user-scored data.  Tagging sites like Flickr, delicious and LibraryThing show the power of having user-created categories (aka folksonomies).  Social networking sites like LiveJournal and MySpace show the power of having user-managed relationships. Creating or enhancing a social network with collaboratively scored tagging of user contributions will dramatically increase its value.

Introduction

This paper aims to demonstrate an abstraction that covers both collaborative scoring and tagging implementations, but also allows for a combination of collaboratively scored tagging.  In the pages that follow I’ll demonstrate how this could be used to enhance existing social networking sites.

Base Representation

The key abstraction has been used in graph theory before but hasn’t yet been applied to the collaborative or tagging networks.  There are three components:

1. Thing
2. Relation
3. Score

Relations occur between Things.  This is similar to vertices and edges in graph theory. Two Things could have multiple Relations; in graph theory the vertices could have multiple edges. A Relation could occur between two identical Things;  in graph theory this is known as a loop. A Score is a measure associated with a Relation between Things.  This is similar to weighted edges.

Let’s look at some examples.

Flickr:

MySpace:
Reddit:

In a more general sense, this can be viewed as follows:

Representational Advantages

Currently delicious is the only social site that comes close to using all the features described above.  Delicious has a bookmarked page (Thing) which can be tagged (Identity Relation) and shows the count of the page across its userbase.  The score in the case of delicious is not tied to the tag, but rather the page itself.

By attaching the score to the Relation, rather than the Thing, this method provides the ability to score different aspects of the Thing.

For example, in a search for ‘Toronto’ on Flickr provides the following pages;

In the ‘Most Relevant’ listing we see some canonical photos of Toronto, but a number of other photos that are much less ‘Toronto-ish’ and seem to be better matches for band and purple.

In the ‘Most Interesting’ listing we see few canonical photos of Toronto, but are much lower in the display ranking.  A number of the photos returned have no inherent Toronto-ness.

By allowing users to score the tags on the images (or Score the Relations on Things) canonical results for a search on Toronto could appear.
Some pictures may exhibit high Toronto-ness and high Blue-ness, while others that currently exhibit Toronto-ness could be downgraded and exhibit high Bono-ness, and medium Cowboy Hat-ness.

Similarly, by implementing the ability to score relationships in a social networking site like MySpace users can choose degrees of relations, as well as add new relations beyond simple friendship.  In the social arena, raw numbers may not be as user-friendly, but degrees of friend relations between two people could become scored like long-term friend, close friend, new friend; or friendship relations expanded with family, co-worker, college chum.

Sites like Reddit could do away with distinct sub-reddits (reddits devoted to a single topic - like programming and web 2.0) and instead auto-generate sub-reddits as pages that have highly-scored tags (Things with highly-Scored Relations).  Delicious could implement their version of sub-reddits by using the scores on the tags as opposed to the count of the bookmarks.

This implementation could also be used to enhance Netflix, Amazon, iTunes and AllMusic genre listings.  A movie wouldn’t be simply in film noir and sci-fi but could be extremely sci-fi, moderately film noir and poor family-friendliness.  A song might be 75% rock / 25% pop or 3-stars electronic, 5-stars ambient.

Conclusion

The Thing-Relation-Score representation model represents the next logical step in social networking tools.  It leverages techniques that have proven successful in other social networks and provides no less functionality that what any of the sites currently implement.  When presented to the user in a non-intrusive way this functionality could help maximize the value of folksonomies.

Notes

1. One aspect of graph theory that is not addressed is directed vs undirected edges.  It’s feasible that Relations between Things could be directed, but it is my belief that this would be too confusing for a collaborative site.
2. In terms of UI design care must be taken to not overwhelm the user with data.  Just as the majority of users don’t tag items, the majority of taggers may not score.  However, taggers that do score can increase usefulness of the social site for all users.
3. See http://www.sarken.org/bettr/bettrflickr.html for a UI mock up of this abstraction applied to Flickr.

Acknowledgments

There are more questions than answers? Debatable!

Johnny Nash wrote a song ‘There are more questions than answers’ but, while I love Johnny, spending the last 10 years thinking about AI and Search makes me think it’s just not so.

After college I wanted to ‘do AI’ and went to Texas to work on the CYC project. During my time there the internet took off - nearly overnight.  This explosion created more information than a handful of people could encode and translate into a rule-based predicate calculus. My belief in a hybrid approach of statistical and rule-based modeling wasn’t shared at the time and so I moved on and spent time writing Linux device drivers and middle-ware in the telecom sector.

But that didn’t stop me from continuing to think about AI and search. Eventually I settled into working on more semantic web applications (which is a whole ‘nother post) but was wooed away to try a different approach.

During the course of these events I became a proponent of Brute Force AI — doing simplistic things that leverage massive amounts of power (like modern processors, clusters and networks) or people (like Wikipedia and the web).

I saw how people actually ask questions and perform search. This has made me believe that there simply aren’t that many questions to be answered. Old-school AI, linguistics and philosophy may be able to come up with numerous ways of phrasing questions and speak of the infinite combinations of language - but in my experience there’s not that many ways that real people ask real questions.

A while back I had this idea of how to approach the questions and answers field - a 90% solution. I sat on the idea too long and now major parts of it have surfaced, scattered across the web.

My inspiration came from a page at Wikipedia - Why is the sky blue?. This is a redirect page, which takes a page title and sends it to another page - Diffuse sky radiation.

Why not have all questions redirect to an answer?

A few months ago, someone beat me to an implementation over at AnswerWiki. It’s not quite what I’d have done - but they have the right domain name and they’re first-to-market. Let me describe what I had planned out in hopes that one of these answer sites¹ will implement my ideas.

Many of the general, reference questions that people ask can be answered in a few sentences or less. This is the idea behind Wikipedia’s lead section.

The lead section is the section before the first headline. It is shown above the table of contents (for pages with more than three headlines). It should establish significances, large implications and why we should care.

The lead section contains the short answer to questions that can be reduced to the form of “Tell me about X”.

More specific questions can be answered through basic templating. “When was X born” is a template already being used by Google (try when was Aristotle born).² Wikipedia has a user-generated category system which would allow you to write templates that could produce answers for some the most popular categories:

• Towers by Country could be used for “How tall is X”.
• Category:People for person related questions like “[Where | When] was X born?”
• Category:Landforms for geographical questions like “How [high | long] is X?”
• even Category:Universities for questions like “What is the motto of X?”

It is my belief that relatively few of these template solutions could pre-populate an answer site (wiki or otherwise) with the vast majority of questions that people actually ask. For example the following questions would all redirect to “Tell me about James Dean”.³

Tell me about James Dean James Byron Dean (February 8, 1931 – September 30, 1955) was an American film actor who epitomized youthful angst. Dean’s mainstream status as a cultural icon is best embodied in the title of his most cited role in Rebel Without a Cause. As with Buddy Holly, Bruce Lee, and Marilyn Monroe his death at a young age helped guarantee a legendary status. Who is James Dean? When did James Dean die? James Dean died when? When was James Dean born?

This would only answer the factual set of questions, but the same techniques could be used to mine the more subjective questions on answer sites, expert sites and even blogs.

While most questions that are actually asked can be answered this way, some reference questions don’t have good answers available via template search solutions. My favorite one to ask, being a history buff, is “How long did it take to sail from England to India?”. The answer is in the wiki page on India Pale Ale with a more detailed answer in Passage East.⁴

This alternative question and answer population could occur organically, like Wikipedia and the answer sites, or it could occur in a more structured way - by using Amazon’s mechanical turk, or something analogous,⁵ for as yet unanswered questions.

Finally, certain questions that aren’t reference based also have great templates - see Google’s SMS Demo for examples of mapping, yellow pages and currency conversions.^6 7

Now there may be no answer to some of the philosophical questions that Johnny Nash raises;

• Why is there so little love among men?⁸
• What is life?
• How do we live?
• What should we take and how much should we give?

Yet a few clever templates, combined with leveraging the knowledge of the massses, could answer the majority of the questions people expect to be able to find answers for.

Notes

1. Answer sites like Google Answers, Yahoo! Answers, MSN QnA or About.
2. Google mines both wikipedia and who2.
   • They also distinguish between when and where.
   • It appears they have some threshold before the template triggers since When was Jason Arnott born doesn’t use the template but the same query with Wayne Gretzky does.
3. Notice that the summary on Wikipedia doesn’t have how James Dean died. That answer would be edited in my idealized version.
4. India Pale Ale gets it’s hoppy flavor because hops were a way of preserving the beer over the six month voyage from England to India — which became two months with the opening of the Suez Canal and eventually got down to three weeks.
5. Cellphedia, AskMeNow and Mozes likely use this human-swarm plus a cache of questions and answers.
6. Some of this reference templating is available via Google Define which is visible through Google SMS. I haven’t found it as useful - there’s some disconnect between the Define group and the web search templating group. Example: define: wayne gretzky gives me a hold-em poker hand — (with no option of more results on SMS) — whereas define: michael jordan gives me an expected reference-based answer.
7. I think mobile devices are the preferred showcase for these short answers to questions. Standard web search is perfect for research, where a page of many results helps to alleviate questions of authority and its feasible to return 10 or more results. Mobile search has limited screen real-estate and, in my experience, is used for more recall-based questions - where you know something vaguely but require short specifics or there’s a strict algorithm for correctness (like currency exchange, weather and the yellow pages).
8. The lyric sites online list this as “Why is there so little of a moment”.  That’s not what Johnny says - give it a listen.