TagAssist: Automatic Tag Suggestion for Blog Posts
Sanjay C. Sood
Northwestern University
2133 Sheridan Road
Evanston, IL 60208
(847) 467-4786
Kristian J. Hammond
Northwestern University
2133 Sheridan Road
Evanston, IL 60208
(847) 467-1012
Sara H. Owsley
Northwestern University
2133 Sheridan Road
Evanston, IL 60208
(847) 467-4786
Larry Birnbaum
Northwestern University
2133 Sheridan Road
Evanston, IL 60208
(847) 491-3640
Abstract
In this paper, we describe a system called TagAssist that provides
tag suggestions for new blog posts by utilizing existing tagged
posts. The system is able to increase the quality of suggested tags
by performing lossless compression over existing tag data. In
addition, the system employs a set of metrics to evaluate the
quality of a potential tag suggestion.
Coupled with the ability for users to manually add tags, TagAssist
can ease the burden of tagging and increase the utility of retrieval
and browsing systems built on top of tagging data.
General Terms
Algorithms, Design, Human Factors, Languages
Keywords
Tags, Blogs, Case Base Reasoning, Tag Suggestions, Text
Classification
1. Introduction
The explosion of user-created content on the web has given rise to
tagging systems aimed at annotating this content with meta-
information, usually in the form of keywords to help in
organizing, browsing, and searching. From image tagging,
(Flickr), to web page tagging (Del.ico.us), to social tagging
(Facebook), these system have become popular and are heavily
utilized across the Web.
Different types of tagging systems have emerged for different
types of content. Social/Collaborative tagging systems have
allowed resources to be tagged by multiple people and shared
amongst a group or community of people. Others only allow the
owner of the content to define the set of tags that will be
associated with the content (YouTube). The focus of our system
is the latter type.
Many would argue that the power of tagging lies in the ability for
people to freely determine the appropriate tags for a resource
without having to rely on a predefined lexicon or hierarchy [10].
The dynamism of tagging systems allows the creators of content
to quickly adapt and incorporate novel concepts and changes in
terminology without having to rely on a standardized tag corpus.
Others argue that large user generated tag corpora, or
folksonomies, will converge on a shared vocabulary that can
assist people in finding and browsing information. The power of
the vocabulary is based on the collaborative nature of its creation,
where individual contributors organically learn and extend the
domain language.
Unfortunately, since tagging systems do not enforce fixed or
controlled vocabularies for tag selection, the tag space suffers
from many of the same problems of traditional free text
Information Retrieval systems. Golder et. al., [6] identified three
major problems with current tagging systems: polysemy,
synonymy, and level variation.
Polysemy, in tagging systems, refers to instances where a single
tag can have multiple meanings. For example, a blog post tagged
with “caterpillar” could indicate that the post is about etymology
or could be interpreted as containing information about
construction equipment.
Multiple tags having the same meaning is referred to as
synonymy. Cases of synonymy may be morphological variation
(“blog” versus “blogs”) or semantic similarity (“news” versus
“current events”). In blog post tagging systems, synonymy is
particularly problematic as authors must rely on their own
intuition to pick the appropriate tag to represent the content of
their post, with no guarantee that two users who have posts on the
same topic will chose the same tag to describe their content.
ICWSM’2007 Boulder, Colorado, USA
The third problem identified is level variation. This refers to the
phenomenon of users tagging content at differing levels of
abstraction. Content can be tagged at a “basic level” or at varying
levels of specificity which is often based on a blogger’s expertise
or requirements; a post may be tagged as “car” (basic) but an
enthusiast might find “Volkswagen MKIV 2001 Golf” more
appropriate.
Another factor that complicates current blog tagging systems is
the lack of clear functional pressure to make tagging consistent,
stable, and complete for use in applications dealing with
collaboration/community, clustering, and search. Some authors
tag their posts to make them visible to a larger community, using
general categorical descriptions such as “politics” and
“shopping”. Some bloggers use tags to organize their posts for
their own consumption and interpretation, using non-content
descriptive tags such as “random” or “toRead”. Other may
choose to use very specific or niche tags that are highly
descriptive of the content of the post (i.e., “why I hate best buy”,
“instructions for cooking grandma’s apple pie”).
Figure 1: The distribution of tags and their frequencies
The lack of a shared or controlled vocabulary has resulted in the
explosion of unique tags in the blogosphere. At last glance,
Technorati [16], one of the leading blog aggregators, was tracking
over 60 million blogs and nearly 11.5 million tags. A sample of
English blog data provided by Technorati from a 16 day period in
late 2006 shows nearly 403,000 unique tags with a mean
frequency of 343.1, median of 8, and mode of 1. The most
frequently occurring tag is “Weblog” with 6,695,762 occurrences.
Nearly 22% (88,212) of tags in the system only occurred once and
only 5.7% of the tags occur more frequently than that average
frequency (343.1). A sample of the distribution of tags and their
frequencies in the sample is illustrated in Figure 1. Because of the
size of the dataset, we ordered the tags by frequency and sampled
every 15 tags (using log distribution to present the data). The
data show that a small percentage of all existing tags are actually
reused by bloggers. The data also show that there is a very large
portion of existing tags that are used rarely, making up a
significant “long tail” [17] distribution. In practice, incredibly
rare tags (that are assigned to posts very infrequently), often
referred to as “meta-noise”, are unlikely to be used for retrieval
due to their idiosyncratic nature. For example, consider the
likelihood of a user utilizing the tags shown in Table 2, a small
random sample of tags that only occur once in our data sample, in
order to browse or search for content.
Content creator-owned tagging systems (those without a
collaborative component), especially suffer from inconsistent and
idiosyncratic tagging. It is not that people are uninterested in
tagging as they often do tag their posts, but when given no insight
into how other bloggers tag, the task of tagging becomes difficult
and the results are less than useful for retrieval and browsing.
For this reason, systems need to be built that can suggest
appropriate tags for content. The goal of such a system is that
users can see how other users are tagging content and choose to
leverage the shared vocabulary or create new tags when
necessary. The overall results would be much more useful
tagging systems without undercutting the prospect or the power of
folksonomies.
Given the state of this tag space, we aimed to build a system that
would assist users in tagging their own blog posts by providing
them with a set of relevant tags. The approach we take is that of a
mediated suggestion system. That is, the tagging system does not
apply the suggested tags automatically, rather it suggests a set of
appropriate tags and allows the user to select tags from the set
they find appropriate. The selected tags are then applied to the
post and incorporated into a larger corpus of post to tag
associations. The system also provides a text box where users can
add additional tags not suggested by the system, allowing new
and emerging tags to be introduced and utilized in the system.
This approach is appealing as it is able to leverage large scale data
processing, while manually checking the suggestions using
minimal human intervention. This type of approach also
fundamentally changes the tagging process from generation to
recognition -- requiring less cognitive effort and time [18][19].
In addition, by providing consistent suggestions to user, we
provide the opportunity for the tag space to organically converge
on a consensus for tag selection. Such a convergence would help
alleviate the issues of synonymy and level variation as users
would have an indication as to the types of tags that other
bloggers are using to describe similar content. Convergence
would also help increase recall by reducing the number of
idiosyncratic tags, reducing the meta-noise in the tag distribution.
offshoreman 'the people', no way!!.., black colored lilies, manila
pride, circuit asia, console customization, eyelash perming,
shadow watcher, miss yah, all female horror, scripture snippets,
insomnia due to quail wailing, streetball china, marriage age,
Wresler, could this possibly be a poem..., Coritsol, goodbye
highbury, 1.2 glossary of terms
Table 2: A sample of tags that only occur once on the blog post
corpus
2. Related Work
There have been several system developed that automatically
generate appropriate tags for a given blog post.
The first of these systems, named TagIt [1], uses Naïve Bayes text
classification methods to determine the appropriate tag to apply to
a new post. While the results of the system were promising, it
was not proven to scale beyond the 330 tags in the training set or
evaluated against blog posts with multiple tags.
Brooks et al, [4] developed another system that automatically
tagged blog posts based on the top three terms extracted from the
post, using TFIDF scoring. While this technique resulted in
closer, more focused clusters of posts it only provides unigram
tags that literally appear in the blog post.
Most similar to our system is the collaborative filtering AutoTag
system developed by Gilad Mishne [11]. AutoTag finds similar
tagged posts and suggests some set of the associated tags to a user
for selection. While our system uses a similar technique, we have
improved on AutoTag’s performance by introducing tag
compression and case evaluation to filter and rank tag
suggestions.
3. The System
To help define the task and guide the development of our system,
we instituted a functional framework for blog tags. Functionally,
we wanted tags to help users to retrieve and browse posts based
on a contextual relationship to a tag or set of tags. Although the
tag space is currently used for browsing and retrieval, the lack of
consistency is the tag space leaves a “long tail”, a significant but
rarely seen portion of the tagged blogosphere.
Our solution to this problem takes the form of a recommendation
system that leverages tags previously associated with content to
recommend tags for new content. The system take a new, un-
tagged post, finds other tagged posts that are similar to it,
aggregates the tags associated to those posts, and then
recommends a set of those tags to the end user. In practice, the
system considers several factors when selecting tags to suggest,
including the frequency of occurrence of the tags in previous
posts. To increase the effectiveness of our system we did not treat
every unique tag in the tag space as an atomic symbol, but rather
looked for areas that we could automatically group
morphologically related tags in a lossless compression.
Discovering similar sets of tags allows the system to utilize a
larger portion of the tagged posts in order to provide
recommendations.
To adapt the constant flow of new blog content being created and
to prevent the data from becoming stale, the system also allows
for the continuous update of content in the training system. This
allows the system to react to changes in the blogosphere including
the addition of new tags and the drift of existing tag senses.
3.1 Tag Compression
The tag space compression stage of the system has two primary
phases. The first phase, referred to as the tag normalization phase,
takes each tag in the system and performs a set of operations
aimed at reducing it to its root form. The second phase, called the
compression validation phase, validates the normalization done in
the first phase.
3.1.1 Tag Normalization
Because of the uncontrolled nature of the tag data, token
scrubbing is performed to trim white space and punctuation from
each tag. Each tag is also stemmed to a morphological root using
Porter’s stemmer [12]. In addition, tags that contain more than
one atomic word are tokenized, stemmed, and placed in
alphabetic order. This helps resolve tag variations such as “news
and politics” and “politics and news” which both resolve to “and
new polit”. The resulting buckets of morphologically related
tags (i.e., those with the same root form) are used as the
hypothesis set of final compression. The first round of tag
normalization reduced the overall tag set by 18.581% (402638
unique tags to 327820 unique roots).
3.1.2 Compression Validation
The second phase of the tag space compression, called validation,
aims to confirm each grouping from the normalization phase to
ensure that the system has not grouped tags with different
meanings under the same normalized root. Morphological
normalization is a relatively aggressive technique that poses the
risk of over-stemming, where two terms that share the same root
but not the same meaning are collapsed together. While
“production”, “product”, and “producers” share the same
morphological root “produc”, they each have distinct meanings.
Techniques that validate morphological normalization choices
using dictionaries and thesauri have been developed, but fail to
adapt to novel word senses and lack entries for current
technological and/or blogging terminology [4]. We chose,
instead, to validate our normalization choices by leveraging
relationships between tags as they are used within our blog
corpus.
To perform the validation, the set of related tags is generated for
each tag within the corpus. The related tags for tag t is defined as
the set of tags, rel(t), that co-occur alongside t in posts in the
corpus. Along with the set of related tags, we also collect the
number of times the co-occurrence appeared within the corpus.
The related tag set provides a reasonable set of related or similar
concepts to the usage of tag t in the corpus. Cattuto el al [5]
statistically analyzed collaborative tagging data and determined
the non-trivial nature of co-occurrence relationships amongst tags.
They further demonstrated that the relationship between co-
occurring tags and how the frequent grouping of “generic” tags
with “narrow” tags may encode semantic hierarchical
organization. The use of co-occurring tags has also been used to
some extent in tag clustering [3] and tag visualization systems [7].
Table 3 shows the top 10 related tags for “Iraq”, illustrating the
effectiveness of related tags to help define a topic space.
related tag count
Politics 462
Bush 410
War 357
Terrorism 275
Iran 230
News 193
Middle East 171
War on Terror 146
Republicans 141
Military 133
Table 3: List of the top 10 related tags and their co-occurrence
frequency for “Iraq”
To perform the validation, each set of tags that share a common
normalized root is placed in a bucket, B
i,
with n total buckets,
where n is the total number of unique roots. The most frequently
occurring tag in each bucket is assigned as the centroid (centroid
i
)
of B
i
and its related tags rel(centroid
i
) are retrieved and
normalized. For each of the remaining candidate tags {t
1
t
2…,
t
k,
}
in B
i
the related tags rel(t
k,
) are retrieved and normalized and an
overlap score P is calculated between rel(centroid
i
) and rel(t
k,
).
The frequency of co-occurrence for each tag is used to weight the
intersection score to favor frequently co-occurring pairs. If P is
above an acceptable a tunable threshold F, the compressed
relationship is maintained. If P is undefined, meaning that t
k
did
not have any related tags, the compressed relationship is also
maintained. If P is less than F, t
k
is labeled an outlier and placed
in a new bucket, B
n+1
. The algorithm is then recursively invoked
until all tags have been placed in an appropriate bucket with
similar tags. At this point, each bucket is assigned a
representative that is the most common variant (mcv) of the
particular morphological root based on its frequency within the
training corpus. The mcv is subsequently used as the actual tag
suggestion to the user, representing the most common use over
the entire corpus. For any tag t, the mcv(t) represent the most
common variant from B
i
which contains t.
The end result of the validation phase was a modest reduction of
overall compression. The reduction of the uncompressed raw tag
data went from 18.581% (327820 unique roots) to 17.001%
(333790 unique roots), but still a large improvement over the total
number of unique raw tags (402638 unique tags).
tag related tag count
apple Mac 333
apple Technology 240
apple iPod 217
apple Software 190
apple Microsoft 143
apple iTunes 135
apples Fruit 60
apples Apple 50
apples Recipes 33
apples Food 31
apples Cooking 26
apples Oranges 20
Table 4: A sample of the tags and their co-occurrence frequency
for “apples” and “apple”
More interesting was this technique’s ability to identify the actual
context in which a tag is used in the corpus, which may be
different than information contained within a standard dictionary
or thesauri. For example, the tags “apple” and “apples” were
combined during the first phase of tag compression, as they share
a common morphological root. A dictionary may very well tell us
that “apples” is a plural form of “apple” or even that “Apple” is
the name of a computer and software manufacturer, but does not
say anything about how the tag “apple” or “apples” is used by
most users in the blogsphere. The related tag set, however,
provides clear evidence that the tag “apple” almost exclusively
refers to the technology firm, while “apples” refers to the fruit.
The differences between these two related tag sets are illustrated
in Table 4. Conversely, this strategy was able to verify many
more compression decisions by proving the semantic relationship
between the two variants. An example of this type of validation is
illustrated in Table 5.
The end result of the compression stage of our system is the
creation of a collapsed tag space that condenses the various
morphological variants and promotes one variant to represent
each set during tag suggestion.
tag related tag count
dogs Pets 364
dogs Dog 108
dogs Puppies 100
dogs Cats 82
dogs Puppy 74
dogs Dog Training 71
dog Dogs 108
dog Pets 93
dog Puppy 83
dog Puppies 76
dog Dog Training 72
dog Dog Clothes 69
Table 5: A sample of the tags and their co-occurrence frequency
for “dogs” and “dog”
3.2 Tag Suggestion Engine
Once the tag space has been normalized and compressed, the
other component of the system, the Tag Suggestion Engine (TSE)
is used to suggest a set of tags to a user. The TSE operates on the
principal of leveraging existing tagged data to provide appropriate
tag suggestions for new content. This approach is very similar to
Case-Based Retrieval Systems [8][13][14] (CBR) where solutions
for new cases are determined by retrieving similar, solved cases
from a large corpus of labeled examples and applying those
solutions (or transformations of those solutions) to the new
problem. Mishne’s AutoTag system takes a very similar approach
to tag recommendation, comparing his system to a recommender
system, a successor to the classic CBR systems.
The TSE contains three main components: the case-base, the case
retriever, and the case evaluator which are all implemented as
web services so they can easily be deployed and integrated with
existing blog post authoring tools.
3.2.1 The Case Base
In order to leverage previously tagged blog posts, they had to be
available for retrieval from a repository. For this purpose, we use
the off-the-shelf Lucene search engine. We have had success in
the past [15] using Lucene, as it was an easy-to-use and configure
repository for our previous text classification system. We used
Lucene’s default content analyzers to index each tagged post in
our corpus along with a unique post identifier so we can retrieve
the associated tags for the post. Once indexing has been
completed, Lucene is able to take a text-based query and provide
a relevance ranked list of posts that contain one or more of the
specified query terms using a simple vector space comparison
model.
3.2.2 The Case Retriever
The second component of the TSE is the case retriever. The main
purpose of this component is to take a new post (target post) to be
tagged and to retrieve other similar posts from the case-base. To
generate a compressed representation of the target post, we
employ a simple TFIDF unigram scoring schema using the corpus
to determine the document frequency component of each term’s
score. In addition, we set minimum and maximum selection
thresholds (St
min
and St
max
) for term inclusion in the query. The
use of St
max
helps in filtering out common corpus-wide unigrams
from the query and St
min
aides in identifying cases where
unigrams are misspelled or non-English. In addition to the
unigram-based term vector, we also identify salient bi-grams
(using TFIDF scoring) from the target post and include those in
our final query. To be included, the bigram must not contain a
term with score below the minimum scoring threshold and must
occur at least twice in the post. To prevent favoring the
vocabulary of any one blogger, we only process the first two posts
from any particular blogger. We experienced the best
performance by using lengthy queries (a maximum length of 30)
and retrieving the top 35 results from Lucene for evaluation.
3.2.3 The Case Evaluator
Once similar posts have been retrieved from the case base, the
unique post identifier is used to retrieve information about the
blog the post was contained in as well as the tags that have been
assigned to that post. For each tag that is retrieved, the most-
common variant, mcv(t), is found, utilizing the tag compression.
Each tag is then scored and/or filtered using five metrics that
evaluate the relative usefulness of the tag t. The five different
scoring/filtering parameters for tag evaluation are as follows:
Frequency – freq(t) is the number of times that t appears as an
associated tag in the top 35 results returned by the case evaluator.
A tag is discarded if freq(t) < 2. This is effective under the
assumption that the stronger the consensus of the tag across
different bloggers, the higher the potential utility of the tag.
Text Occurrence – whether the raw tag t or the mcv(t) appears in
the actual target post. The appearance of the tag in a post may be
an indicator of relevance.
Tag Count – count(t) is the number of times tag t (and all of its
variants) have been used in the training corpus. The tag is
discarded if count(t) < 2. Tags that are utilized more in the
blogosphere have a higher potential of being useful to a user.
Rank – the relative rank of the blog that contained the post that
was assigned tag t. The rank of a blog is analogous to its overall
popularity in the blogosphere as determined by the number of
inbound links.
Cluster – using the same clustering technique that we use to
validate the tag compression, we determine whether any of the
candidate tags are members of topically related clusters by
comparing the pair-wise similarity of each tag’s related tag set.
This allows us to find the semantic relationships between tags that
are not morphological variants. Topical agreement amongst
disparate tags in the results set is an indication of their potential
utility.
After each tag has been processed and scored, the individual
scores are weighted and combined to form an aggregate tag score.
The tags are ordered by score and filtered once again by score.
The goal is to provide only the best tag suggestions to the user. To
this end, we only return tags that have an aggregate score greater
than the mean score for all the tag candidates.
4. Evaluation
To evaluate TagAssist, we used data provided to use by
Technorati, a leading authority in blog search and aggregation.
Technorati provided us a slice of their data from a sixteen day
period in late 2006. The data contains only English content with
8.1M blog posts from 2.7M unique blogs. Out of these posts,
1.9M posts are tagged with an average of 1.75 tags per post.
To gauge the effectiveness of our system compared to other
similar systems, we developed a version of our tagging suggestion
engine that was integrated with the raw, uncompressed tag data
and did not use the case-evaluator for scoring, aside from
counting frequency of occurrence in the result set. This baseline
system returned the top 10 tags ordered by frequency. In addition
to comparing our system’s performance against the baseline, we
were also interested in examining how our system compared to
the original tags that were assigned to the post in our training
corpus.
Tag Set Accuracy
Original Tags 48.85%
TagAssist 42.10%
Baseline 30.05%
Table 6: Accuracy values for human evaluation of the three tag
sets
Our study used human judges to evaluate the appropriateness of
tags for a post. For testing data, we randomly selected posts, with
2 or more originally assigned tags, from our blog corpus and
presented them to a human judge along with a list of tags
generated by our system, a list of tags generated by a baseline
system, and the tags originally assigned to the post in our corpus.
The post was presented in a web page along with a list of tags and
corresponding checkboxes. The judges were asked to read each
post and then check the boxes next to tags they thought were
appropriate for the post.
In all, we collected and analyzed 225 responses from a total of 10
different judges. Table 6 lists the precision values for each of the
tag sets, that is, the average percentage of tags in each set that the
judges found appropriate. As the results show, 48.85% of tags
originally assigned to a post were determined to be relevant by
our judges. Our method resulted in a precision of 42.10% and the
baseline came in third with a precision of 30.05%. While
TagAssist did not outperform the original tag set, the performance
is significantly better than the baseline system without tag
compression and case evaluation.
Given that less than 50% of tags originally assigned to a post are
not deemed as relevant by third party judges, we found it less
useful to perform automatic evaluation of our system by
calculating precision/recall values for our system against the
original tags. It also makes it difficult to automatically tune the
system when there is not reliable data to gauge the system’s
performance. For the sake of comparison to other systems, we
performed the evaluation by processing 1000 posts through our
system and the baseline system and then comparing the suggested
tags against those originally assigned. We did not use string
distance to compare tags, but chose instead to use exact string
equality for comparison. As a result, the precision/recall values
are much lower than the results of human evaluation. Table 7
shows the results of automated evaluation for both our method
and the baseline. The results show almost identical recall values
between both systems with our system outperforming the baseline
in precision.
Suggestion Method Precision Recall
TagAssist 13.11% 22.83%
Baseline 7.66% 23.14%
Table 7: Precision and recall values for automated testing over
1000 posts using exact tag matching
5. Discussion
Our evaluation shows that TagAssist is able to provide relevant
tag suggestions for new blog posts. The novel tag compression
algorithm and case evaluation component helped increase the
precision of the system without reducing recall. A system that
can effectively propose relevant tags has many benefits to offer
the blogging community.
Firstly, shifting the tagging process from a purely generative
process to one that require users to recognize appropriate tags
significantly reduces the cognitive burden and increases
performance of blog post tagging. If we work to make the
tagging process easier, we are more likely to increase the number
of bloggers who tag their posts. If more users tag posts, we are
likely to increase the richness of the folksonomy and provide
more content to tag search interfaces.
Secondly, providing users with tag choices based on their actual
usage in the blogosphere can accelerate the convergence of tag
vocabulary to be more consistent and useful for retrieval and
browsing.
6. Future Work
One of the interesting results from our human evaluation is the
relevance score for the original tags assigned to a blog post. On
average, less that 50% of these original tags were deemed as
relevant by third-party judges. Given that our system is trained
off this data, we believe we can drastically improve the
performance of our system by identifying the blog posts have
been effectively tagged, meaning that the tags associated with the
post are likely to be considered relevant by other users. We are
currently investigating techniques to identify these effectively
tagged blog posts and hope to incorporate it into future versions
of TagAssist.
Automatic evaluation of tag suggestion engines is also critical to
building effective systems. Given issues with the perceived
relevance of user-generated tags, it is important that we have a
ground truth testing corpus to evaluate system performance. But
with no functional constraints on the use of tags, it is difficult to
build a gold standard that everyone can agree on.
User feedback is another component that we would like to add to
TagAssist. Given that we are providing a list of tags to a user and
having them select the most appropriate ones give a blog post, we
can use their feedback to help tune our system. A more involved
scenario might ask users to evaluate each tag, providing the
system with explicit feedback on the utility of each tag. In
addition, given that we allow users to freely enter additional tags,
we can use that information to improve TagAssist.
We are also exploring novel way of presenting the suggestion list,
besides using plain text. We would like to explore ways of
presenting the tags that reflect their usage in the larger
blogosphere. The interface to our system will be critical in
making it effective and usable by various bloggers.
Acknowledgements
Our thanks to Technorati for providing us with the datasets used
to create TagAssist. Particularly, we’d like to thank Peter
Hirshberg, Adam Hertz, and Ian Kallen for their support on this
project. Also, thanks to Shannon Bradshaw for his insight on this
project.
References
[1] Al-Shurafa, H. Blog Post Labeling using Naïve Bayes Text
Classification Methods.
[2] Bateman, S., Brooks, C. and McCalla, G. Collaborative
Tagging Approaches for Ontological Metadata in Adaptive
ELearning Systems. In Proceedings of the Fourth
International Workshop on Applications of Semantic Web
Technologies for E-Learning (SW-EL 2006) in conjunction
with 2006 International Conference on Adaptive Hypermedia
and Adaptive Web-Based Systems (AH2006) (June 20, 2006).
Dublin, Ireland, 3-12.
[3] Begelman, G., Keller, P., and Smadja, F. Automated Tag
Clustering: Improving search and exploration in the tag
space. In Proceedings of the Fifteenth International World
Wide Web Conference (WWW2006) (Edinburgh, Scotland,
May 22-26, 2006). ACM Press, New York, NY, 2006.
[4] Brooks, C., and Montanez, N.
Improved Annotation of the
Blogopshere via Autotagging and Hierarchical Clustering
.
In Proceedings of the Fifteenth International World Wide
Web Conference (WWW2006) (Edinburgh, Scotland, May
22-26, 2006). ACM Press, New York, NY, 2006.
[5] Cattuto, C., Loreto, V., and Pietronero, L. Semiotic
Dynamics and Collaborative Tagging.
http://www.isrl.uiuc.edu/~amag/langev/paper/cattuto06semio
ticDynamics.html, 2006.
[6] Golder, S.. Huberman, B. Usage Patterns of Collaborative
Tagging Systems. In Journal of Information Science (2006),
32(2). 198-208.
[7] Hassan Montero, Yusef, and Herrero-Solana, Víctor.
Improving Tag-Clouds as Visual Information Retrieval
Interfaces. In Proceedings of the International Conference
on Multidisciplinary Information Sciences and Technologies
(InSciT ‘06) October 25-28, 2006, Mérida, Spain.
[8] Leake, D. CBR in Context: The Present and Future. In Case-
Based Reasoning: Experiences, Lessons, and Future
Directions. AAAI Press/MIT Press, 1996, 1-30.
[9] Marlow, C., Naaman, M., Davis, M., and Boyd, D. HT06,
Tagging Paper, Taxonomy, Flickr, Academic Article,
ToRead. 2006. In Proceedings of Hypertext 2006. ACM
Press, New York, NY, 2006.
[10] Mathes, A. Folksonomies: Cooperative classification and
communication through shared metadata. In Computer
Mediated Communication, 2004.
[11] Mishne, G. AutoTag: A Collaborative Approach to
Automated Tag Assignment for Weblog Posts. In
Proceedings of the Fifteenth International World Wide Web
Conference (WWW2006) (Edinburgh, Scotland, May 22-26,
2006). ACM Press, New York, NY, 2006.
[12] Porter, M. F. An Algorithm for Suffix Stripping. In
Readings in Information Retrieval. Morgan Kaufmann
Publishers Inc., San Francisco, CA, 1997, 313-316.
[13] Riesbeck, C., and Schank, R. Inside Case-based Reasoning.
Erlbaum, Northvale, NJ, 1989.
[14] Schank, R. Dynamic Memory: A Theory of Learning in
Computers and People. Cambridge University Press, New
York, NY, 1982.
[15] Sood, S., Owsley, S., Hammond, K. and Birnbaum, L.
Reasoning Through Search: A Novel Approach to Sentiment
Classification. In Submission.
[16] Technorati, www.technorati.com, 2006.
[17] Anderson, S. "The Long Tail", Wired, Oct. 2004
[18] Goldstein, D. G., and Gigerenzer, G. The Recognition
Heuristic: How ignorance makes us smart. In Simple
heuristics that make us smart. Oxford University Press,
Oxford, England, 1999.
[19] Tulving, E. Elements of Episodic Memory. Oxford
University Press, New York, New York, 1983.
