Part of the allure of classifying things by assigning tags to them is that the user can give free reign to sloppiness. There is no authority —human or computational— passing judgment on the appropriateness or validity of tags, because tags have to make sense first and foremost to the individual who assigns and uses them. And yet, the whole point of distributed classification systems (DCSs) such as del.icio.us and flickr is that the aggregation of inherently private goods (tags and what they describe) has public value: When people use the same tag to point to different resources they are organizing knowledge in a manner, commonly referred to as a folksonomy, that makes sense to them and to others like them. In other words, the tag is the object that brings a resource and a social group together via the shared meaning of a word (although tags also serve to form connections between words and new meanings, as for example when you encounter a link to the Center for Alternative Technology when looking at the tag ‘cat’).

We can say, then, that DCSs function at the intersection of individual choices and the shared linguistic/semantic norms of a social group (the folks in folksonomy). In this paper, I explore two aspects of this intersection. In the first part, I examine some of the open affordances of DCSs in terms of the agency of the code (the program; the computer instructions that make things happen). In other words, I look at how DCSs frame social activity in the process of aggregating individual tagging choices into collective information; in short, how the code shapes social action. At the same time, I also explore the implications of relegating the organization of some social functions to the code.

In the second part, I explore some of the linguistic properties of tags, their role in an attention economy, and outline a set of guidelines for generating tags in ways that maximize the social usefulness of tags. Tag literacy in this sense refers to the ‘etiquette’ of generating tags in a way that increases their social value, balancing individual needs with the needs of the group. Because the code (rightly, I believe) does not enforce normative behaviors when it comes to the creation of tags, I argue that it is up to those users invested in the welfare of the community to develop a normative approach to tagging.

Part I: The Social Agency of Code in DCSs

The greatest strength of distributed classification systems (DCSs) is also their greatest weakness: the way in which the negotiation of meaning during the process of classification is delegated from humans to code. Decisions regarding how to classify things which used to be undertaken by humans in collectivity are now carried out by humans individually, while the code aggregates and represents those decisions. If we see this as a replacement for systems of classification in which one group of people used to impose their classification scheme on the rest, then this might be seen as an improvement. If we see this as a replacement for systems in which equals used to negotiate and collaborate on the definition of a classification scheme (and in the process gave shape to what defined them as a group), then the outcome might not be as welcomed. This is because this process is now conducted by the code, without some of the opportunities for negotiation and collaboration that other systems afford. As is always the case with technology, where the line is drawn between the open affordances of DCSs (what they facilitate and what they constraint) depends on how the technology is applied.

In order to understand how code assumes social agency in DCSs, we must first contextualize the manner of classification that these systems embody. There are two ways in which a classification systems allows for meaning construction. One is in the use of the system to search for resources already in the system. The other is in the contribution of new resources to the system. A traditional classification system, based on a structured taxonomy, guided users in the search for resources by moving from the general to the specific, at each branch presenting clearly defined options. Imagine you wish to find a resource using the Yahoo! Directory. Does the resource have to do with Arts & Humanities, Business & Economy, or one of the other categories? If it’s related to Arts & Humanities, does it have to do with Photography, History, Literature, or one of the other categories? Yahoo! decides what those categories are, and individuals use their familiarity with the classification structure to find things. Now imagine you wish to add a resource to the system. In that case, you would use the same categories to find the appropriate place for the resource. If such a category does not exist, then the administrators of the system must decide whether it needs to be created, and where in the overall scheme it needs to be added.

Folksonomies differ from this structured taxonomy approach in significant ways. The most obvious one is that any user of the system can create tags or categories without permission from any kind of authority. Another important difference is that tags need not be arranged in any particular way. If the tag/category ‘cat’ is close to the tag/category ‘car’ it is probably because of alphabetical reasons, and not because the proximity of ‘cat’ and ‘car’ says something about any of the two signifieds. Because categories do not occupy a specific location in a structure, folksonomies allow for the association of an infinite number of tags to a resource. In other words, a picture of a cat driving a car can be tagged in both categories, as well as any others that the user chooses.

Another difference between folksonomies and structured taxonomies that might not be so obvious is the role of human collaboration in their definition. Structured taxonomies require consensus in the form of at least two collaborating human subjects (whether this consensus is achieved democratically or hegemonically is another topic). If a taxonomy is defined but no one adheres to it, can it be said to exist? Folksonomies, on the other hand, do not require consensus as much as they measure the consensus already established around the use of certain words. In other words, folksonomies assume consensus without involving humans in the process. DCS users have no discussion whatsoever about how categories should be defined, or what they mean, or their relation to each other. Instead, all the code cares about is that if two people used the tag ‘cat,’ it will aggregate and display the resources associated with that tag, regardless of whether one user meant the furry feline and another the Center for Alternative Technology. Of course, if the latter user had employed that tag ‘CAT’ instead of ‘cat,’ the code would react differently (which perhaps means, as Clay Shirky suggests, that there are no such things as synonyms in a folksonomy).

In essence, the code of DCSs removes the need for humans to negotiate meaning around classification. This can be liberating as well as alienating. Liberating because, as I suggested above, there is no governing body dictating what the classification scheme should be. Alienating because, without the mechanisms for deliberation, meaning becomes atomistic, a reflection of what the software has parsed and aggregated from detached individuals, not what has emerged through consensus and deliberation.

By this I do not mean to imply that DCSs do not offer social affordances (they are, after all, ’social software’). I merely want to call attention to this different way in which we are defining and constructing sociality —a sociality that is the result of code doing things to the resources of detached individuals. There are plenty of ’social’ transactions that can be carried out in DCSs, such as being able to see different items classified by different people with the same tag, or the same item classified by different people with different tags, or the resources of a particular individual, etc. But the scope of these affordances is defined by the code, and the community willingly relinquishes a large part of its agency in exchange for individual freedom and the scale of access that only the internet can provide.

While the benefits of this freedom and scale are obvious, some people rightfully point out the risks of surrendering agency in the process of negotiating how knowledge should be arranged. Representative of the arguments focusing on freedom and scale is the following by Clay Shirky, discussing del.icio.us:

“[A]ggregate self-interest creates shared value… By forcing a less onerous choice between personal and shared vocabularies, del.icio.us shows us a way to get categorization that is low-cost enough to be able to operate at internet scale, while ensuring that the emergent consensus view does not have to be pushed onto any given participant.” (reference)

On the other hand, Matt Locke describes the functions relinquished by the community and how the code assumes those functions in some form or other:

“There are no politics in folksonomies, as there is no meta-level within the system that allows tagging communities to discuss the appropriateness or not of their emergent taxonomies. There is only the act of tagging, and the cumulative, amplified product of those tags.” (reference, my emphasis)

It is in discussing this ‘appropriateness’ that social groups in fact define themselves. Clearly, there are politics in folksonomies, but we need to uncover them by expanding the sort of questions we are accustomed to asking for deconstructing political power (questions such as the ones danah boyd asks here) with questions that take into account the social agency that the code assumes on behalf of people. For example, the question of who benefits and who becomes marginalized in DCSs needs to be reframed to account for the fact that in DCSs anyone with access to the internet can benefit, and no one is more marginalized than anyone else. Instead, it is more important to ask in what ways agency is taken away from users by the code, and what the benefits and risks of this are. Or, put another way: What assumptions (epistemologies, biases, etc.) are embedded in the shared meanings that make up a folksonomy, and that the code makes unnecessary to negotiate?

Of course, the code in DCSs is not static. Improvements and new features are constantly being added: discussion tools through which users can share reasons for tagging things in a particular way, tagging forms which make visible the most common or most popular tags being used by the network, all sorts of network-wide or group-wide tag visualizations, user data, and so on. These new features redraw the line between the agency of the code and that of humans to organize knowledge, and result in entirely new affordances.

Part II: Guidelines for Generating Tags

Before suggesting some specific strategies for increasing the social value of tags, I want to make some observations about their linguistic nature. Tags are text words, used to assign meaning to resources. Consider the following groups of tags, each by a different user (selected from this set), which were used to classify the same resource (in this case an earlier study I wrote on del.icio.us):

• folksonomy Delicious socialnetworks socialsoftware rss syndication
• findability interaction_design to_read
• del.icio.us social bookmark article research classification
• del.icio.us ontology kdt
• annotation classification del.icio.us emergence flickr folksonomies knowledge metadata ontology tags

The list can be analyzed according to the cultural, social, affective and cognitive dimensions of the words that comprise each group. Gunther Kress’ (2003) noted that

“words in combination are not much more than rough outlines waiting for us the readers to colour them in. What the written text provides is words in clear order. Each word asks to be filled with meaning, a meaning that comes from our past experience of that word in our social lives.” (p. 59)

While Kress is referring to written texts with a “clear order” (such as sentences and paragraphs), an unordered group of tags used to describe a resource holds the same potential to be filled with personal and social meaning. Clearly, these groups of words are not devoid of meaning as texts. To their authors, they represent textual associations that will make it easier to find the resource again in the future (one can assume, for example, that the tag to_read is used to classify resources that the user intends to read later). Each group of tags is the user’s framing of the resource according to a personal scheme. At the same time, however, the fact that certain keywords (such as folksonomy, social, del.icio.us, ontology, etc.) are repeated across groups suggests that some of these words have a socially shared meaning apart from their personal meaning. This socially shared meaning is what would allow someone browsing through those tags to find that same resource, or similar ones.

But why would anyone choose to search for resources this way instead of, say, using a search engine? In other words, why do people use del.icio.us instead of (or, probably more accurately, along with) Google? One possible answer is that, in an attention economy, tags represent an allocation of attention. To elaborate: Attention is “the action that turns raw data into something humans can use” (Lanham, in Lankshear and Knobel, 2003, p. 111). Attention economics establishes that what information consumes is “the attention of its recipients. Hence a wealth of information creates a poverty of attention and a need to allocate attention efficiently among the overabundance of information sources that might consume it” (Simon, in Lankshear and Knobel, 2003, p. 109).

Tags are very efficient ways of allocating attention in the face of informational overabundance. It takes very little time to bookmark and tag a resource. Because users are the first ones to benefit from classifying the resources that interest them, there is a very high motivation to tag. Thus, what people are doing in reviewing tags is capitalizing on attention allocated by others, specially on aggregated attention (what happens when large groups of people allocate attention to the same tag or resource, as seen in the ‘Most Popular’ tag or resource feeds in a DCS).

In short, Google yields search results that represent attention allocated by computers, while DCSs yield search results that represent attention allocated by humans. The former method (computer attention) is cheap, and hence ideal for indexing large amounts of information quickly; the latter method (human attention) is not so cheap, and not so quick, but it can yield more socially valuable information because it means a human being has made the association between a resource and a particular tag. Hence, this method is ideal for qualitative indexing. Furthermore, this method can be made cheaper and quicker by distributing the process across large communities and tying it to the individual interest of the user, which is exactly what a DCS does.

Still, the overall quality of a DCS is largely determined by how individuals tag resources.
For the most part, users don’t give much thought to the process of selecting tags, which is what makes tagging so painless (in these systems sloppiness is, by definition, a right). However, the selection of tags plays a very important role in the welfare of the community. As a matter of fact, they are the most important social contribution an individual can make. If tags signify allocation of attention, we want that allocation to be as efficient and useful to the network as possible (while retaining, of course, value to the individual). Therefore, users need to be metacognitively aware of good tag-selection practices. Eventually, these practices of selecting tags that will be useful not only to the individual but to the network as a whole become second nature. This is what is meant by tag literacy.

Tag literacy is also about understanding the social life of tags, what code does to tags, how it manipulates them, and what it allows users to do with each other’s tags. What follows are some practical guidelines intended mostly for the new user of DCSs (they are in the format of Frequently Asked Questions):

Tag Literacy v1.0:

• What are distributed classification systems (DCSs)?
• What do tags signify?
• What makes a good tag?
• What is the social value of tags?
• How do I find things using tags?

What are distributed classification systems (DCSs)?

A DCS allows a network of users to contribute resources to the system and to classify those resources by assigning tags to them. While tags serve primarily a personal purpose (facilitating the retrieval of resources by the individual at a later time), the use of the same tag by more than one person engenders a collective classification scheme. Some of the most popular DCSs at the moment are services like del.icio.us (http://del.icio.us/), Flickr (www.flickr.com/), and Furl (www.furl.net).

The most notorious feature of DCSs is that they do not impose a rigid classification scheme. Instead, they allow users to assign whatever classifiers they choose. Although this might sound counter-productive to the ultimate goal of classifying content, in practice it seems to work rather well. While DCSs will probably not replace search engines (which rely on computer algorithms to index resources and match them to queries) or traditional structured taxonomies (in which a group of people defines and controls the classification scheme), they will probably gain more popularity and coexists with them. The shortcomings of one can be offset by the strengths of another, which is why it’s important for people to understand how each can be applied.

back to top

What do tags signify?

In essence, a tag establishes a relation between an online resource and a concept in the user’s mind. This association is expressed in the form of a word. Thus, a picture of a tree might be tagged with the word ‘tree.’ A web page containing a cartoon might be tagged with the word ‘funny’ by one user and ‘not funny’ by another. Resources are often tagged with more than one word, which indicates multiple associations between a resource and various concepts.

A tag also signifies an allocation of attention; it tells us that a resource has been deemed useful by someone, and therefore it might be useful to us as well. But instead of relying on mere serendipity, tags increase our chances of encountering that same resource.

back to top

What makes a good tag?

The beauty of DCSs is that there are no rules regarding what types of words you may use as tags, or the number of tags you may use (this might vary from system to system). Most users opt to use tags that will facilitate their later retrieval. This is perfectly fine, and is largely what makes DCSs valuable. However, in order to maximize the social value of DCSs, users should think about ways of tagging resources that will be useful to other people who use the system. Most likely, people will not spend a lot of time trying to guess all the possible tags that someone might look for. But there are some simple ways that, if adopted widely within the community, can ensure that the value of the system increases:

• Think of tags as personal, but also think of tags as social. In other words, tags can be both for your personal use and for the use of others in the network. Sometimes the same tag can serve both purposes. Sometimes you will need to use one kind of tags for your individual purposes, and another kind of tags for the social purpose. The rest of these guidelines illustrate some of these strategies.
• Use plurals to define categories. When appropriate, instead of blog or tree, use blogs and trees. Tags signify a category which can encompass various resources, so the plural is generally more appropriate. This will avoid having to check both the singular and plural version of a tags (although DCSs will become increasingly smart at aggregating tags). However, sometimes having both a singular and a plural tag is necessary. For example, I would expect to find very different resources under the tags apple(as in the electronics manufacturer) and apples(as in the fruits).
• Avoid capitalization, except when capitalization is the norm. Don’t use Trees or Blogs, but trees or blogs. However, it might be appropriate to capitalize ANTS (Algorithmic Number Theory Symposium) to differentiate it from ants(the insects), for example. Also, per the previous example, you would capitalize Apple if it refers to the company, not the fruit.
• Think specific, but also think general. Select tags that describe the resource in very specific terms, but consider also using tags that describe the resource in general terms. Those terms might be too broad for your benefit, but they might help others find the resource. For example, a user who happens to be a chef tagging a recipe might use specific tags for the main ingredients, but using general tags such as food and recipes would help others find the resource.
• Be idiosyncratic, but also be generic. It’s OK to select tags that have meaning only to you or to a very specialized group of people (such freedom is what makes DCSs so valuable), but try to balance this with the use tags that would also help others using more conventional paths. For example, tagging a picture of panda bears with the tag print_this might be useful to you, but you may also want to consider using the more standard tags panda and bears.
• Group common phrases. Some folks use a period or an underscore to group words in common terms, as in open.source or open_source, instead of the separate words open and source. This avoids the hassle of having to look for the intersection of the two separate tags.
• If you want to be extra-nice, include a couple of synonyms. For example, it may be sufficient for you to tag something with the tag Big_Apple, but you may want to spend a couple of extra seconds to also include the tags NYC and New_York, knowing that those tags have a broad social value. Of course, sometimes synonyms can dilute the associations you wish to make, so if you mean cinema and not film, then you should use whatever word fits your needs.
• Observe the norms of the network. Pay attention to tagging conventions followed by other members of the network, and if they make sense to you, adopt them. Lots of good ideas can come from observing the tagging practices of others.
• Contribute to maintenance efforts. Some DCSs allow you to modify your tags, adding or deleting tags from resources, cleaning up errors, or batch-fixing tags (for example, changing all your blog tags to blogs in one go). Some DCSs even allow you to modify the tags of others. Spending some time doing this takes a minimum amount of effort and increases the value of the system for you and for the network as a whole.

back to top

What is the social value of tags?

As an individual user, a DCS allows you to maintain a set of resources tagged any way you want. But your tags and your resources are shared by the community as well. Similarly, you can benefit from the information classified by others. Here are some examples of the benefits you can derive from the social aspects of a DCS.

• Track a particular tag. Say you are interested in pictures of birdhouses. You can go to a DCS that indexes pictures (such as flickr) and look up the tag birdhouses. This will show you all the pictures in the system classified with that tag.
• Track a particular user. Suppose you have a friend who also contributes resources to the DCS. Or you may realize that a lot of the resources you are interested in are submitted by a particular user whom you do not know. Or perhaps it has come to your attention that someone you consider an expert in a particular field has an account in the DCS. In all cases, the system allows you to track the resources of those individual people. In del.icio.us, for example, all you have to do is add those users to your Inbox, and every time they add a resource to their collections you will be notified.
• Track social groups. Different DCSs offer different ways to support group collaboration. One of the most basic is to create a tag that defines a social group (e.g.., jones_family), and have members of the group track that tag. Thus, when a user wishes to share a resource with that group, he or she can tag it with the desired tags as well as the group tag (for a study of how this works, see my previous del.icio.us study). More advanced DCSs allow you to form groups, set privacy levels, and access online collaboration tools (such as chat, discussion boards, etc.).
• Track trends. Most DCSs offer ways to track what the most popular tags and resources are. This functions practically as a zeitgeist meter, since it measures which tags are being used the most and which resources are being classified the most by users within the whole network.

back to top

How do I find resources using tags?

There are three main methods of using tags to search for resources that have been entered into a DCS:

• Intentional search. This involves cross-referencing tags to find what you want. If you are interested in how blogs are being applied to education in Asia, you could cross reference, for example, the tags blogs, education and Asia.
• Serendipitous browsing. This works for more general searches. Suppose you are interested in pictures of bears. You could access that tag and see what is available (which may include bears at the zoo, stuffed teddy bears, etc.)
• Subscription to RSS feed. A great feature of DCSs is that they make available RSS feeds for tags, users or even searches. Thus, you can subscribe to the bear picture tag, an individual user’s resources, resources tagged by members of a group, or even to the results of the cross referencing of the tags blogs, education and Asia. This means that you would be notified whenever a new resource is added to any of those lists.

back to top

This has been a short introduction to what distributed classification systems allow you to do with tags, and how to generate tags to maximize the social value of these systems. This guidelines were intended to provide a very cursory view of the power of tags. If there is interest, I can put this text in a wiki so that others may add to it.

Offline References

Kress, G. R. (2003). Literacy in the new media age. London: Routledge.

Lankshear, C., & Knobel, M. (2003). New literacies: Changing knowledge and classroom learning. Buckingham [England]; Philadelphia, Pa.: Society for Research into Higher Education & Open University Press.

[Note: This is a project I did for a class on social and communicative aspects of the internet, taught by Chuck Kinzer. Not a 'real' study, but you might find some of the literature review and listed resources helpful. You may also want to check out a more recent paper I wrote on tagging and del.icio.us: Tag Literacy]

Abstract

Working within the constraints of a very limited data sample, this
study attempts to identify some of the information management and
meaning construction practices of an online distributed classification
(a.k.a. free tagging or ethnoclassification) community. Specifically,
this study seeks to investigate the social and communicative practices
that emerge when users are encouraged to share web links with one
another by using a metadata keyword, or tag, to demark a social group,
apart from using other tags to classify links according to an emergent
taxonomy.

Introduction

We have definitely arrived at a point in the development of human knowledge where the amount of content published online everyday far exceeds the ability of anyone to categorize and index such wealth of information. Even for a hypothetical individual or organization of great skill and capacity, the task of processing all that content would pose an almost insurmountable problem, without even considering the difficulty of developing and maintaining a taxonomy to accommodate the speed at which new knowledge is produced.

But the situation is not hopeless. Solutions to this particular dilemma are emerging that are in accordance to the demands of the Information Age. One approach, made possible by advances in network technologies, is to distribute the task amongst the maximum number of individuals possible. Just as we figured out that scanning outer space for intelligent life signals is a task that can proceed more efficiently by being distributed across many computer processors, we have begun to realize that other tasks that require human involvement can also be distributed across individuals by using the largest human network in history: the internet.

This principle of distribution is at work in socio-technical systems that allow users to collaboratively organize a shared set of resources by assigning classifiers, or tags, to each item. The practice is coming to be known as free tagging, open tagging, ethnoclassification, folksonomy, or faceted hierarchy (henceforth referred to in this study as distributed classification), and is associated with popular online services such as furl (www.furl.net), del.icio.us (http://del.icio.us/), or flickr (www.flickr.com/).

One important feature of systems such as these is that they do not impose a rigid taxonomy. Instead, they allow users to assign whatever classifiers they choose. Although this might sound counter-productive to the ultimate goal of organizing content, in practice it seems to work rather well, although it does present some drawbacks. For example, most people will probably classify pictures of cats by using the tag ‘cats.’ But what happens when some individuals use ‘cat’ or ‘feline’ or ‘meowmeow’ or ‘my.favorite.cat’?

While the present study is obviously interested in such issues, my goal is to focus instead in some of the social dynamics that are emerging through the use of distributed classification systems. My thesis is that a better of understanding of how users perceive these systems, and how they interact with each other through them, can provide us with important insights about individual as well as social processes of knowledge and meaning construction online.

Purpose Statement

To identify, through a series of quantitative as well as ethnographic research methods, some of the social interactions and information management practices exhibited by users of the CCTE Distributed Research (http://ideant.typepad.com/ccte/) system, in order to better understand how distributed classification shapes individual and social processes of knowledge and meaning construction online.

Literature Review and Theory

As the variety of alternatives to describe this phenomenon suggest, distributed classification is still a nascent field, and formal research and theorizing is just beginning. Therefore, very little literature exists on the social and communicative affordances of distributed classification systems. Below, I will summarize some of the main themes in the field and attempt to portray the emerging zeitgeist by quoting extensively from the blogosphere.

To begin, Jon Udell frames the issue in terms of an individual’s motivation for assigning metadata to content:

Conventional wisdom holds that people will never assign metadata tags to content. It just isn’t on the path of least resistance, the story goes, and those few who do step off the path succeed only in creating unwieldy taxonomies… Yet somehow, users of Flickr and del.icio.us do routinely tag content, and those tags open new dimensions of navigation and search. It’s worth pondering how and why this works. (Udell, 2004)

It seems that while most people might not be motivated to contribute to a pre-established system of classification that may not meet their needs, or to devise new and complex taxonomies of their own, they are quite happy to use distributed systems of classification that are quick and able to accommodate their personal (and ever changing) systems of classification.

This is exactly what distributed classification systems such as del.icio.us provide. As far as the actual operation of del.icio.us (the focus of this study), Matt Biddulph describes it as follows:

You submit your links to a website, adding some descriptive text and keywords, and del.icio.us aggregates your post with everyone else’s submissions–letting you slice and dice the information any way you like. Posts with the same keywords are clumped together, and if enough people link to a URL, a loose classification emerges. (Biddulph, 2004)

But distributed classification does not accrue benefits only to the individual. It is a very social endeavor in which the community as a whole can benefit. Jon Udell describes some of the individual and social possibilities of this method of classification:

These systems offer lots of ways to visualize and refine the tag space. It’s easy to know whether a tag you’ve used is unique or, conversely, popular. It’s easy to rename a tag across a set of items. It’s easy to perform queries that combine tags. Armed with such powerful tools, people can collectively enrich shared data. (Udell 2004)

This is indeed one of the most important advantages of using a distributed classification system: the ability to emergently define a taxonomy, or, as it is alternatively known, a folksonomy. The advantage of this ethnoclassification or free tagging or faceted hierarchy process can be described in various ways:

Set this [an imposed taxonomy] against the idea of allowing a user to add tags to any given document in the corpus. Like Del.icio.us, there needn’t be a pre-defined hierarchy or lexicon of terms to use; one can simply lean on the power of ethnoclassification to build that lexicon dynamically. As such, it will dynamically evolve as usages change and shift, even as needs change and shift. (Williams, 2004)

The primary benefit of free tagging is that we know the classification makes sense to users… For a content creator who is uploading information into such a system, being able to freely list subjects, instead of choosing from a pre-approved “pick list,” makes tagging content much easier. This, in turn, makes it more likely that users will take time to classify their contributions. (Merholz, 2004)

Folksonomies work best when a number of users all describe the same piece of information. For instance, on del.icio.us, many people have bookmarked wikipedia (http://del.icio.us/url/bca8b85b54a7e6c01a1bcfaf15be1df5), each with a different set of words to describe it. Among the various tags used, del.icio.us shows that reference, wiki, and encyclopedia are the most popular. (Wikipedia entry for folksonomy, retrieved December 15, 2004 from http://en.wikipedia.org/wiki/Folksonomy)

Of course, this approach is not without its potential problems:

With no one controlling the vocabulary, users develop multiple terms for identical concepts. For example, if you want to find all references to New York City on Del.icio.us, you’ll have to look through “nyc,” “newyork,” and “newyorkcity.” You may also encounter the inverse problem — users employing the same term for disparate concepts. (Merholz, 2004)

As a way to address this issue, many have suggested that synonym control mechanisms be implemented in distributed classification systems. But as Clay Shirky remarks, this solution might diminish some of the benefits that we can derive from folksonomies:

Synonym control is not as wonderful as is often supposed, because synonyms often aren’t. Even closely related terms like movies, films, flicks, and cinema cannot be trivially collapsed into a single word without loss of meaning, and of social context… There is a loss in folksonomies, of course, but also gain, so the question is one of relative value. Given the surprising feedback loop — community creates folksonomy, which helps the community spot its own concerns, which leads them to invest more in folksonomies — I expect the value of communal categorization to continue to grow. (Shirky, 2004)

Lastly, we should also keep in mind that, as Matt Biddulph points out, there are established structures defining just how open these systems are and what kind of knowledge is shared. We might see these as acceptable or even desirable costs, but we should still be aware of the dynamics:

The choice of tags [in the entire del.icio.us system] follows something resembling the Zipf or power law curve often seen in web-related traffic. Just six tags (python, delicious/del.icio.us, programming, hacks, tools, and web) account for 80% of all the tags chosen, and a long tail of 58 other tags make up the remaining 20%, with most occurring just once or twice… In the del.icio.us community, the rich get richer and the poor stay poor via http://del.icio.us/popular. Links noted by enough users within a short space of time get listed here, and many del.icio.us users use it to keep up with the zeitgeist. (Biddulph, 2004)

Having reviewed some of the literature, I will now summarize the concepts that inform this study.

Key Concepts

• Socio-technical system. This study does not frame technological systems as divorced from the people who use them and the context in which they are used. A socio-technical system is conformed of hardware, software, physical surroundings, people, procedures, laws and regulations, and data and data structures (from http://www.computingcases.org/general_tools/sia/socio_tech_system.html, retrieved on November 22, 2004).
• Metadata: Keywords (or ‘tags’) used to describe an object, usually structured according to a taxonomy, or a system of classification.
• Distributed classification (also known as free tagging, open tagging, folksonomy or ethnoclassification): A way for individuals or groups to collectively define their own taxonomy. This is in contrast to taxonomies that are pre-defined and that are imposed in a top-down hierarchy.
• RSS: Real Simple Syndication or Rich Site Summary. An XML format for distributing information about a dynamic web site. Commonly employed by bloggers and news organizations to syndicate new content, allowing users to subscribe to ‘RSS feeds’ which are usually collected with an ‘RSS Aggregator.’ In distributed classification systems (such as del.icio.us), RSS feeds can be created for particular tags or users.

Research Questions

Due to the limited scope and timeframe of this study, none of the results should be considered as definitive answers to my research questions. In fact, my intention is merely to suggest directions for future, more comprehensive studies. These are the questions I have attempted to address in this study:

• How is meaning created in the distributed classification system through the social sharing of bookmarks?
• How is knowledge collectively structured by the use of tags?
• What social conventions emerge through the use of the distributed classification system?

Methodology and Data Analysis

Setting

The study took place mostly online. Two web locations were the main sites of data gathering:

• del.icio.us (http://del.icio.us): As discussed above, del.icio.us is an online service that lets users collect and categorize web links (URLs). According to its author, “del.icio.us is a social bookmarks manager. It allows you to easily add sites you like to your personal collection of links, to categorize those sites with keywords, and to share your collection not only between your own browsers and machines, but also with others” (http://del.icio.us/doc/about, retrieved on December 4, 2004).
• CCTE Distributed Research (http://ideant.typepad.com/ccte/): CCTE DR is a portal created by me that provides instructions for using del.icio.us to collect bookmarks and, by including the special tag ‘ccte,’ share those bookmark with others. Users can visit the portal just listed or subscribe to an RSS feed of the ‘ccte’ del.icio.us feed. In my own personal blog, I described the motivation for creating CCTE DR as follows: “[A]t the graduate program where I am studying (Communication, Computing and Technology in Education, or CCTE), we usually share links by email or through classroom discussion boards. This means things don’t get archived collectively, and only some people benefit from such knowledge. Some of us have blogs, but we don’t really use them to share bookmarks. What if the CCTE community was encouraged to use a ‘ccte’ tag when bookmarking stuff on del.icio.us? And what if I created a little portal to display the RSS feed of that tag (as well as provide some instructions)?” The CCTE DR portal was launched on September 25, 2004. As of December 4, 2004, the site had received 424 hits, or an average of about 6 visits per day (interestingly, someone is Spain also started using del.icio.us tags in the same way at around the same time; cf. http://www.eibar.org/blogak/luistxo/en/166, retrieved on December 4, 2004).

Additionally, I gathered informal interview (“ethnographic”) data through email and, in some cases, face to face conversations.

Actors

There were six individuals who voluntarily participated in this study by choosing to use the system. All are graduate students at Teachers College, Columbia University, and all but one of them are enrolled in the Communication, Computing and Technology in Education program. It should also be acknowledged that one of the subjects or users is me, the author of this study.

Additionally, an analysis of the server logs shows that there were a number of visitors to the CCTE DR portal (people who simply browsed the links, but never posted a bookmark), although the exact number and affiliation is hard to determine.

Events and Processes

After creating and troubleshooting the system, I invited a number of people to use it. In order to post, all users had to create a del.icio.us account first (there is no cost involved in doing this, and all that is required is to create a username, a password, and enter an email address). I included in the CCTE DR portal detailed instructions for how to post a bookmark once the user acquired a del.icio.us account. The following describes the process in brief.

Upon encountering a web location that the user wished to bookmark, classify and share, the subject would click on a bookmarklet located on their browser bar (this bookmarklet was installed as part of the del.icio.us account creation process). This would cause a del.icio.us pop-up form to appear. The form would have the following fields: url, description, extended, and tags. There would also be a Save button. All fields would be editable, and some were prefilled. The url field would be automatically filled with the corresponding web address of the page that was being viewed. The description field would be automatically filled with the title of the page. The extended field would be empty; here the user could enter additional information about the web page in question. The tag field would also be empty. Here, users could enter whatever metadata keywords they wished to associate with the particular page in questions. It was emphatically stressed in the CCTE DR instructions that if users wanted to share the link with the rest of the CCTE DR community, they should include as one of their tags the keyword ‘ccte.’ This allowed me to use the corresponding RSS feed generated by del.icio.us for this tag to automatically publish all items that contained ‘ccte’ to the CCTE DR portal. To do this, I used a service called RSS Digest (http://www.bigbold.com/rssdigest/) that automatically queries the specified RSS feed for new items every 30 minutes and generates a formatted view of the information that can be published on a web page.

After the required 30 minutes or so, the link bookmarked by the user would appear on the CCTE DR page. Other users would see the link either the next time they visited the CCTE DR web page, or if they used an RSS Aggregator, the next time they checked their subscriptions (I myself subscribed to the ‘ccte’ RSS feed and would check it daily with my Shrook RSS aggregator, which meant that I didn’t have to visit the CCTE DR page except to make sure things were working fine).

Specific communication and knowledge-building practices amongst users are discussed in the Findings section of this study.

Data Collection Strategies

Because this study was not intended to be a comprehensive application of a research methodology, I tried to combine aspects of both quantitative and qualitative methods to provide as interesting a picture as possible of the social practices that emerged through the use of the CCTE DR socio-technical system. It should be acknowledged from the start that the amount of data used in this study is not meant to be considered a sufficient sample.

Quantitative strategies:

• Analysis of logs: I had access to usage reports made available by the server that hosted the CCTE DR page, the service that monitored RSS subscriptions, and whatever information del.icio.us provides (such as who posted each item, etc.).
• Analysis of tag use: Using mostly information from del.icio.us, I put together the tables in the Appendix that summarize usage, including most popular themes by tag, tag ranking, and individual user tag use. It is worthwhile to note that there are tools, such as tag.alicio.us (http://frenchfragfactory.net/ozh/archives/2004/10/05/tagalicious-a-way-to-integrate-delicious/) and extisp.icio.us (http://kevan.org/extispicious), that aid in querying and visualizing the use of tags in del.icio.us. I only had opportunity to play with the latter, however.

Qualitative strategies:

• Informal interviews with users: Since the launch of the system, I maintained informal and irregular communication with the users of the system. This included email exchanges as well as face to face conversations. I did not follow a particular template for these interviews. Sometimes I would engage in extemporaneous exchanges about the CCTE DR system with users. Other times, I would prepare brief questionnaires that I would send by email, and to which users replied voluntarily. For purposes of this study, interview quotes do not include any statements made by me as a user (I felt that my bias as researcher would be reflected in these statements).

Data Analysis Procedures

These are some of the measures I looked at, both quantitative and qualitative:

Quantitative:

• Total number of users in the system (up to a certain date)
• Total number of items submitted by user (up to a certain date)
• Most popular tags in the system (referred to as ‘themes’)
• Tag use by individual user (i.e., what keywords they used, and how frequently)
• Items archived by the user that did NOT include the ‘ccte’ tag (i.e., personal bookmarks)
• Number of visits to the CCTE DR page
• Number of RSS subscriptions

Qualitative:

• Understanding of the function and potential benefits of the system
• Difficulty in using the system
• Additional desired features not currently found in the system, in particular features that would enhance social interaction

Findings

In this section, I will try to summarize some of trends that emerged in the analysis of the quantitative and qualitative data. These findings are by no means conclusive, specially considering the small study sample. Thus, they should be taken more as possible directions for further exploration.

How is knowledge collectively structured by the use of tags?

In total, there were 6 users of the CCTE DR system who contributed 156 items between September 25th and December 12th of 2004 (this means that 156 items included the tag ‘ccte’). The most active user contributed 117 items using 156 different tags, while the least active user contributed 2 items using only the ‘ccte’ tag (a summary of the data is provided in the Appendix). In the interest of disclosure, I should acknowledge that I was the most active user of the system.

What kinds of resources did users share? Given the focus of the Computers, Communication and
Technology in Education program, it is not surprising that users (mostly students in this program) contributed items mostly having to do with the use of technology in social and educational settings. By analyzing tag use and grouping similar tags (e.g., blog and blogs), I found that the most common themes were: blogs, games, social, collaboration, academia, and virtuality. As expected, users did employ many of the same tags. For example, the single most used tag by more than one user (excluding ‘ccte,’ which was used by all users) was ‘collaboration,’ used 11 times by two users. ‘Blog’ and ‘blogs’ were used 11 times by three users. ‘Games’ and ‘gaming’ were used 12 times by 2 users.

The sharing of resources related to these topics mirrors the interests of the larger internet community. Although it was not possible to conduct searches to see how many times a particular tag was used by all del.icio.us users (del.icio.us does not provide this information as part of its search functions), a search of the use of these terms in the blogosphere (conducted using www.technorati.com) reveals that terms such as ‘blogs’ and ‘gaming’ are indeed used more than terms such as ‘identity’ or ‘globalization’ (for a comparison of this ranking, see the Appendix).

Through the interviews, I was able to get a glimpse of how the community of users decided what types of resources should be shared. This took the form of discussions about ‘who the audience for this is’ and how this information influenced the kinds of bookmarks users believed it was pertinent to share, and the kinds of bookmarks they would collect for their own research purposes without sharing them with the rest of the community. The following passage by a user summarizes this kind of decision making:

I thought of the CCTE DR blog as being for a certain audience… I’d only choose to put links that were of a general enough interest to appeal to a variety of people. At the same time, I saw that audience as having certain interests (socio-cultural approaches to understanding tech, social software, games, new literacies…) that pretty much reflect the interests of the people who I know post to the blog. So, I decided to start my own set of research bookmarks that were more specific to my own interests, including stuff that others might not have a broad appeal (technology and liberal arts colleges, for example).

Another user remarked: “I’ve only bookmarked a few things and all of them have the ccte tag. However, I would only use this tag when I feel I want to share with the ccte audience.” The use of del.icio.us to classify personal bookmarks without sharing them with the rest of the CCTE community is measured by the number of items submitted without the ‘ccte’ tag. For example, one user saved 9 items out of her total of 22 without the ‘ccte’ tag. Other subjects used the ‘ccte’ tag almost for everything; I employed the ‘ccte’ tag for 117 out of 118 items.

What social conventions emerge through the use of the distributed classification system?

Some of the most interesting social conventions emerged through the use of the extended field in the del.icio.us form (the pop-up form used to submit an item). Since all other fields (url, description, and tags) served specific purposes, the extended field was adopted for more informal means of communication within the community. For example, comments entered in the extended field such as “Did anyone attend this event?” are most likely addressed to the other members of the CCTE DR community, and not to the del.icio.us community at large. I make this inference based on the fact that since there is no way for other del.icio.us members to reply, the author of this comment expected that if one of the CCTE DR members had indeed attended the event in question, he or she would make the author aware of this through another means of communication available only to them (email or face to face conversation). Also, one user sometimes included a note in the extended field to signify to whom the bookmark in question might be particularly relevant. For example: “[for David].” This information obviously was not intended for (and could not be made sense of by) other del.icio.us users.

Equally important are the conventions that users felt could not emerge due to the lack of features in the system. For example, one user said: “[C]omments for posts! I’d like it if we can discuss the things people post,” referring to the lack of a feature that would allow users to annotate the bookmarks already submitted by others. Another user remarked: “No ‘people knowledge’: No information on the participants other than the collection of bookmarks they post” and “[N]o way to know if others find posts useful. I feel if I got feedback on my participation I would post more often and with more relevance,” which suggests this user felt a user profile and a rating system would increase the usefulness of the system.

It is also clear from usage patterns and interview data that some users did not find the system useful or easy to operate. This might have been related to the level of interest, the availability of personal time, or the clarity of the instructions on how to use the system and what its potential benefits are. For example, one user commented: “I only tried twice briefly and it looks like I messed it up both times. I didn’t explore the other features either (mostly due to my time constraints)” and “It’s also not easy to see the big picture. I cannot easily find the main page, where the other links and resources are located…” This suggests, among other things, that the connection between CCTE DR and deli.cio.us was not explained clearly in the instructions, at least for this particular user. However, it is not surprising to note that the users who experienced more difficulties in using the system, like the one just quoted, are also the ones who contributed less items (i.e., who spent less time experimenting with the system).

Finally, it is interesting to note that, as far as social conventions go, people seem to find more value in reviewing links than in submitting them. Although the degree of ‘lurking’ is difficult to quantify, I quote the following comment by a user: “I don’t post that often. I do find things daily that I feel are worth posting but I don’t. I do, however, check the CCTE DR page daily for new stuff.”

How is meaning created in the distributed classification system through the social sharing of bookmarks?

Although more data would be needed to make substantive claims in this area, some observations can be made about the processes of meaning making through the use of del.icio.us and the CCTE DR portal. One observable trend is the difficulty to make the conceptual switch from using fixed to using flexible taxonomies. All of the users who voluntarily supplied interview data asked about the inclusion of categories at one point or another: “[I]s there a way to organize the links into categories, if not on the main page, then perhaps a list of categories in the side bar?”, “[T]he added resources and links are not necessarily categorized, which make it difficult to locate things quickly” and “Is it going to be possible to have posts grouped by category?” These comments suggest that it’s hard to let go of established modes of classification. Even though del.icio.us gives users the power to build their own taxonomies, most still felt that some pre-arranged order would make the system more useful. One user, while expressing some dissatisfaction with a flexible taxonomy system, also suggested ways in which this problem could be addressed: “The free tagging feature is too free. I feel that it might be better to pick from a list and to add new tags only when the list doesn’t contain the tag you need” (incidentally, some del.icio.us plug-ins such as nutr.icio.us, http://supergreg.hopto.org/nutritious/, are starting to head in that direction).

On the other hand, it seems that those users who spent more time with the system, and explored more in depth the features of del.icio.us, began to perceive the potential of the system. The following remark unveils a user’s thought process as she discovers an additional dimension of using del.icio.us tags that she had not thought of before:

I thought of it [del.icio.us, as opposed to CCTE DR] more as a way of having a set of bookmarks accessible on the web, instead of in a menu in my browser, so that I could refer to them even when I didn’t have my computer with me. Also, I thought of using the tags as a way to file the bookmarks by subject, which is a problem with the CCTE links. Of course, as I’m thinking about this, I realize that I could just as well give a link two tags, which would send it to CCTE and file it under my own bookmarks using my own filing system.

Here’s another comment by a user who suddenly realizes the multiplicities of meaning that can be realized through using tags in a flexible taxonomy:

I don’t think anyone in the ccte program does this sort of research [referring to a specific bookmark’s topic] but I did add the tag because it was directly related to my own research. I’m thinking now that perhaps the tag serves a dual purpose. One, to tag resources that I want to share with others. Two, to tag resources that share my interests with others such that I’m sharing “people knowledge” about myself.

Additionally, users of CCTE DR also began to realize how the system can be used to share and collectively construct meaning within the CCTE community (“I’m trying to get all the video games folks to post their links there instead of to our mailing list”) and with the external online community at large (“What’s also cool about this is that people who are interested in what’s going on at CCTE, can visit this site to get a sense of what people are interested in”).

Conclusions

As the findings of this preliminary and limited study show, it is hard for people to make the initial conceptual shift from traditional forms of classification (using fixed taxonomies) to distributed classification schemes (using flexible taxonomies). The freedom to define individual and social structures of classification emergently can be perceived as chaotic, lacking rigor and utility. However, the more comfortable users become with a system’s features, the more aware they become of the benefits of distributed classification, and the more aware they also become of working within its limitations.
It can be argued that distributed classification systems such as del.icio.us do not exhibit some of the features commonly thought of as necessary to support online communities (features such as the ability to access ‘profile’ knowledge about individual users, the ability to communicate directly with other users, and the ability to rate the quality of submissions). It seems that del.icio.us did not set out to become that kind of community tool, so those features might never become part of its toolset. However, one question to explore further is to what degree such features would enhance the sense of community, or if there are other ways in which del.icio.us accomplishing that.

One thing that did not seem to be very clear in the minds of the people participating in this study is how CCTE DR users are contributing not only to that particular community, but to the larger del.icio.us community and their efforts. Perhaps the creation of the CCTE DR portal itself served as a distraction, confusing users about its purpose and the role that del.icio.us plays in fulfilling that purpose. At the same time, I would argue that at least the CCTE DR portal allowed users to realize the social benefits of the distributed classification system more quickly, and in ways more relevant to their own interests. However, more thinking needs to be done on how to get inexperienced users to understand the benefits of distributed classification, and if indeed creating subcommunities within these systems is the best way to do that.

As the designer of the CCTE DR portal, it became clear to me that more needs to be done to get people to understand that CCTE DR is just a place where reviewing links submitted by others can be done quickly; thus, questions of whether it is appropriate to submit a particular link should not be central. Since it is very easy to quickly scan the bookmarks submitted, CCTE DR is intended to be a real-time snapshot of what the community is researching, and nothing more. Any further exploration, archiving and classification should be handled through the del.icio.us interface by each user.

Accordingly, more needs to be done also to investigate if people would find accessing the RSS feed of CCTE DR more useful than visiting the web page every day. For this to happen, more needs to be done to explain to users that the essence of CCTE DR is really the del.icio.us tag, and that RSS or the CCTE DR web page are simply ways to view items associated with that tag (it would be nice to be able to see as part of an item which other tags were used in classifying it).

Lastly, it will be interesting to monitor how community dynamics and meaning/knowledge construction processes change if the number of CCTE DR users increases sharply. My hope is to continue to monitor usage and be able to expand this study.

References

Biddulph, M. (2004, November 10). Introducing del.icio.us. Retrieved on December 15, 2004 from http://www.xml.com/lpt/a/2004/11/10/delicious.html

Merholz, P. (2004, October 19). Metadata for the masses. Retrieved on December 15, 2004 from http://www.adaptivepath.com/publications/essays/archives/000361.php

Shirky, C. (2004, August 25). Folksonomy. Retrieved on December 15, 2004 from http://www.corante.com/many/archives/2004/08/25/folksonomy.php

Udell, J. (2004, August 20). Collaborative knowledge gardening. Retrieved on December 15, 2004 from http://www.infoworld.com/article/04/08/20/34OPstrategic_1.html

Williams, A. (2004, November 19). Terms of the night: Folksonomy and ethnoclassification. Retrieved on December 15, 2004 from http://www.livejournal.com/users/zamiel/831808.html

Appendix: Summary of Data

Click image below to open full-size version in a pop-up window.