=Paper=
{{Paper
|id=Vol-520/paper-6
|storemode=property
|title=Folksonomy Resources as a Data Source for the Social Data in Semantic Web
|pdfUrl=https://ceur-ws.org/Vol-520/paper05.pdf
|volume=Vol-520
}}
==Folksonomy Resources as a Data Source for the Social Data in Semantic Web==
https://ceur-ws.org/Vol-520/paper05.pdf
Folksonomy Resources as a Data Source for the Social
Data in Semantic Web
F. Echarte, J.J. Astrain, A. Córdoba, J. Villadangos
Dpt. de Ingeniería Matemática e Informática
Universidad Pública de Navarra
Campus de Arrosadía. 31006 Pamplona (Spain)
patxi@eslomas.com, {josej.astrain, alberto.cordoba, jesusv}@unavarra.es
Abstract. The increasing popularity of folksonomies has made them interesting
as a tool to bridge the gap between Web 2.0 and the Semantic Web. This paper
describes a modular and generic method (ACoAR) for the automatic
classification of the resources tagged in a folksonomy, using semantic measures
and a reduced number of relevant tags. Generic since it applies to both narrow
and broad folksonomies, and modular since it allows the integration of different
techniques and algorithms. ACoAR updates this classification as the
folksonomy evolves. ACoAR is validated using a del.icio.us sample set to
obtain a set of classification concepts. The accuracy of ACoAR to classify new
resources is analyzed obtaining a well classification rate of 93% of the analyzed
resources.
Keywords: Social data, folksonomies, ontologies, automatic classification
1 Introduction
Early this century, two substantial changes in the evolution of the Web appeared: i)
Tim Berners-Lee advocated his vision of a Semantic Web [1]; and ii) Social Networks
emerged as a useful and widely used tool which enable collaborative relationships
among users (social data). Web goes through a significant evolution in the so-called
Social Sites (Web 2.0), where user roles have been notably modified. Users lived a
relevant evolution: from just information consumers to be content generators, defining
in a great extent the usefulness and evolution of very different kind of webs. Users
have agreed in the use of a great number of different collaborative tools such as
Blogs, Wikis, and social networks like Flickr, Facebook, etc. These tools require a
powerful, simple and easy-to-use mechanism to classify the information managed.
This mechanism must be simpler and more flexible than those based on taxonomies,
and less formal than those mechanisms used on the semantic web that are based on
ontologies. The mechanism more widely used in current webs is the content tagging
based on a no-controlled vocabulary. The term Folksonomy [2] is frequently used for
this kind of collective annotation process in Web 2.0.
Several works focus on bridging the gap between the Semantic Web and Web 2.0.
Semantic information from tags can be obtained using subsumption strategies [3,4],
concentrating in finding groups of highly related tags [5, 6] or relating folksonomies
with ontologies [7, 8]. The semantic information obtained allows the navigation
among the tags of the folksonomy and then, the resource accessibility. Those
proposals build tag classification systems often considering tag co-occurrences, and in
some cases, external sources that allows obtaining either information from tags, either
ontologies in order to describe their semantic. However, those tag classification
systems do not consider fully the resource’s semantics, because they only consider the
tags assigned to a resource, but not the frequency which each tag is associated to the
resource.
Some works [4, 9] deal with narrow folksonomies, where the related frequency is
always the unit. While in a narrow folksonomy (like flickr) only the owner of a
resource can tag it, in a broad folksonomy (like del.icio.us) anyone can tag anything.
In most cases, the process of harvesting the information semantics is not
automatically performed. Moreover, as folksonomies are dynamic systems which
evolve as users introduce new annotations, some mechanisms are required to
accommodate this evolution to the proposed classification systems.
We propose a resource centric method called ACoAR (Automatic Classification of
Annotated Resources) [10] which ensures the automatical classification of resources.
It uses a set of classification concepts -CCs- (a set of resources with similar
semantics). These concepts are automatically obtained from the resources annotations
taking into account both tags and the frequency with which a resource is tagged. This
resource classification is performed using a relatively small subset of the tags of a
folksonomy. We prove that this subset is valid enough to semantically group the
resources of the folksonomy under classification concepts which represents the main
topics of the folksonomy. Additionally, ACoAR is a dynamic and automatic method
which updates the classification concepts as the folksonomy evolves with new
annotations.
ACoAR is a generic and modular method. Generic, in the sense of it applies to
both narrow and broad folksonomies, and modular since it allows the integration of
different techniques and algorithms.
In order to validate our proposal we have obtained information about web pages
from the broad folksonomy of del.icio.us, since del.icio.us is considered one of the
world's leading social bookmarking service and contains a large set of resources and
annotations.
The rest of the paper is organized as follows: Section 2 is devoted to describe the
ACoAR method, explaining the creation of the classification concepts from an
existing folksonomy and their evolution when new annotations arrive to the
folksonomy; Section 3 describes the experimental results obtained; and finally,
conclusions, acknowledgements and references end the paper.
2 Method Description
ACoAR is intended to provide the automatic classification of annotated resources. As
depicted in Fig. 1 ACoAR has an initial task called CCs Creation, which classifies the
folksonomy resources under a set of CCs that represent the main topics of the
folksonomy. The task uses a Dictionary in order to improve the performance of the
classification. The Dictionary contains the subset of tags which represent the
semantics of the resources. Although folksonomies use to have a very large number of
tags, some works [11, 12] show that annotations follow a power law distribution such
that a reduced subset of tags represents (and preserves) the semantics of the resources.
The information of the folksonomy concerning tags and resources is used to build
the initial set of CCs and to classify the folksonomy resources under them. The CCs
Evolution task updates the classification as the folksonomy evolves with new
annotations (new or existing resources, users and tags). While the initial task runs
only once (at startup), the second one runs whenever a buffer is fulfilled with new
annotations. Both tasks, CCs building and evolution, are performed using
representations of the resources according to the Dictionary tags assigned by users.
Fig. 1. Method Description: it shows the creation of the classification concepts and their
evolution when new annotations are created in the folksonomy
Let F be a folksonomy F:= where T represents the set of tags, R is the
set of resources, U is the set of users and Y U T R is a ternary relation
representing the set of annotations, can be extended to represent the dictionary, the
classification concepts and the relations among resources and concepts, using a tuple
ModelACoAR := , where T, R, U and Y define the folksonomy.
D T is a dictionary which contains the most relevant tags of the folksonomy, C is
the set of classification concepts (CCs) in which folksonomy resources are classified.
And finally, Z R C is a binary relation between a resource R and a classification
concept C, representing that a resource is classified under a concept.
We consider that a resource has converged when its distribution of tags (tags
belonging or not to the dictionary) converges rapidly to a remarkably stable heavy-
tailed distribution. In order to evaluate this convergence we define a threshold, based
on the annotations number. Resources that have converged are encoded using the
vector space model (VSM). The corpus consisting of R converged resources and D
dictionary tags is represented by a matrix A=(aij) RxD. Each row vector ai
corresponds to a resourcei and each column vector corresponds to a tagj of the
dictionary. Each aij represents the number of annotations that relates tagj to resourcei.
Although there exist many other representation methods that could also be
considered, such as TF-IDF, the election of this non normalized method is due to its
simplicity since resources can be directly encoded according to their annotations
without requiring any additional calculus, and also due to the easiness of the CCs’
representation, since it only requires a summatory for each tag of the dictionary over
the resources classified by the concept.
2.1 CCs Creation
In order to create the CCs ACoAR performs several steps, as depicted in Figure 2.
Initially ACoAR builds automatically the dictionary for a given folksonomy. The
behavior of the method is determined by the size of the dictionary, since a large
number of tags in the dictionary implies high computational costs when comparing
resources. The objective is to accurately represent the semantics of the resources
using the minimum number of tags. Dictionary tags can be selected following
different criteria like the tags most frequently used, the tags representing the largest
number of resources, etc. Tags can be also be filtered to discard misspellings, to find
syntactic variations, etc.
Fig. 2. Classification concepts (CCs) creation, process description
In the second step, ACoAR encodes the folksonomy resources which annotations
have converged, obtaining the representative vector of each resource. The rest of
resources will be encoded when the folksonomy evolves and their annotations
converge. Table 1 shows a folksonomy example. Vector a1≡(40,0,0,0,5,0,0,0,32,0,
0,0,0,0,2,0,0,0,0,0) corresponding to resource r1 (a1 is the representation vector of
resource1) and indicates that r1 has been annotated 40 times with the tag ajax, 5 with
css, 32 with javascript and 2 times with programming.
Table 1. Folksonomy example with dictionary tags and resources
Tag/Resource r1 r2 r3 r4 r5 r6 r7 r8 r9 r10
t1 “ajax” 40
t2 “art” 21 24
t3 “blogs” 13 34
t4 “blogging” 4 21
t5 “css” 5 4
t6 “database” 5
t7 “design” 7 8
t8 “java” 5
t9 “javascript” 32 45
t10 “lisp” 16
t11 “museum” 9
t12 “mysql” 10
t13 “oracle” 8
t14 “php”
t15 “programmin” 2 33 34 25 5 25
t16 “python” 15 18
t17 “socialweb” 5
t18 “sql” 12 12
t19 “twitter” 10
t20 “xml” 5 12
The third step is in charge of the creation of the CCs and the classification of the
resources. The classifier provides a set of clusters (each of them corresponding to a
CC) and the resources are classified under those clusters. Although many interesting
clustering techniques can be considered [13] to create the clusters, k-means technique
provides satisfactory results in our experiments (see Section 3). Each cluster consists
on a centroid and its associated resources. The classifier compares the resources with
the centroids of the clusters in order to obtain the most appropriate cluster for each
resource. There exist many similarity measures that can be used to compare resources.
For example, the cosine similarity (well known on Information Retrieval [14])
measures the angle among their vector representation without requiring any
normalization.
Let illustrate the calculus of the centroids with the aid of Table 2. Using a k-means
algorithm with a k value of four, we obtain the clusters represented by centroids c1, c2,
c3 and c4. Each centroid is obtained, in our case, adding the representation vectors of
the resources (c1=a1+a9). Then, c1≡(40,0,0,0,5,0,0,0,77,0,0,0,0,0,7,0,0,0,0,0) is the
addition of (40,0,0,0,5,0,0,0,32,0,0,0,0,0,2,0,0,0,0,0) and (0,0,0,0,0,0,0,0,45,0,0,0,0,0,
5,0,0,0,0,0), the representation vectors of r1 and r9 (a1 and a9 respectively).
Table 2. Clustering results for the related folksonomy
Cluster centroid Resources Cluster Centroid
c1 r1, r9 (40,0,0,0,5,0,0,0,77,0,0,0,0,0,7,0,0,0,0,0)
c2 r2, r6 (0,45,0,0,4,0,15,0,0,0,9,0,0,0,0,0,0,0,0,0)
c3 r3, r7 (0,0,47,25,0,0,0,5,0,0,0,0,0,0,0,0,5,0,10,0)
c4 r4, r5, r8, r10 (0,0,0,0,0,5,0,0,0,16,0,10,8,0,117,33,0,24,0,17)
ACoAR evaluates the similarity among the clusters in order to merge those clusters
with a high degree of similarity. Although many different measures can be used, in
this case and due to its simplicity, we use the cosine similarity measure to compare
the centroids of the clusters. Table 3 illustrates the semantic similarity among
centroids. In this example, the similarity between c1 and c4 is very low (0.07405) and
then c1 and c4 are not merged in a same cluster. Two clusters having a high similarity
between them (close to the unit) would be merged producing a new cluster if a certain
threshold is exceeded. The centroid of the resulting cluster is the sum of both
centroids (cnew=c1+c4). Then cnew replaces c1 and c4.
Table 3. Semantic similarity among centroids
Similarities c1 c2 c3 c4
c1 1.00000 0.00473 0.00000 0.07405
c2 0.00473 1.00000 0.00000 0.00000
c3 0.00000 0.00000 1.00000 0.00000
c4 0.07405 0.00000 0.00000 1.00000
Finally, the method creates a classification concept for each cluster, and assigns
them a name. For such purpose, the method considers those tags having the greater
values in the cluster centroid vector. Table 4 shows the CC assigned to each cluster.
When the frequency of a certain tag is greater than the frequency of the rest of tags
(this tag has a high relevance), then this tag becomes the classification concept
associated to the cluster. When the most significant tags have similar frequencies and
their frequency is amply high in comparison to other tags, the CC name is obtained by
merging those tags. Other alternatives like looking for syntactic variations, prefixes
and suffixes, or even looking up external resources like Wordnet or Wikipedia, or
even ontologies, could be used to assign more adequate names to the CCs.
Table 4. Classification concepts created from clusters
Cluster Cluster main tags Classification Concept
c1 javascript:77, ajax:40, programming:7, css:5 Javascript & Ajax
c2 programming:117, python:33, sql:24, xml:17, lisp:16, Programming
mysql:10, oracle:8, database:5
c3 art:45, design:15, museum:9, css:4 Art
c4 blogs:47, blogging:25, twitter:10, java:5, socialweb:5 Blogs
2.3 CCs Evolution
Once the CCs have been created and the resources have been classified under them, is
necessary to define how to update these CCs when the folksonomy evolve with new
annotations. ACoAR also considers that users can make changes in their existing
annotations, for example deleting a tag assigned to a resource. New folksonomy
annotations are accumulated into a buffer. When this buffer is filled ACoAR
processes the annotations like Fig. 3 depicts. A four steps process is applied,
processing the annotations and updating the classification information.
A folksonomy resource may belong to one of these three sets: i) pending when the
resource has not yet received enough annotations to have converged, ii) converged
when the resource has converged but it is not yet classified, and iii) classified when
the resources has converged and it is classified under a CC. After the CCs Creation
phase, all the folksonomy resources belong to pending or classified sets.
Fig. 3. CCs evolution based in the new annotations on the folksonomy
The first step is intended to classify new resources under the existing CCs. It
process annotations and encodes pending resources that have reached the minimum
annotations threshold, assigning them to converged set. It also updates the
representation vectors of converged resources that have received new annotations.
After that, converged resources are processed to find the most similar CC,
comparing their representation vectors. When the similarity between a resource and a
CC reaches a minimum threshold, the resource is classified under CC and it is moved
to classified set. When the similarity doesn’t reach the threshold the resource remains
in the converged set. Once a resource is assigned to classified set, ACoAR doesn’t
allow moving it to other set.
The second step checks the annotations and the criteria used to create the
Dictionary and update it when necessary. If the Dictionary is modified, ACoAR
updates the representation vectors of resources and CCs according to the new
situation.
The third step updates the resources representation vectors that belong to the
classified set with the new annotations they have received.
The fourth step is equivalent to CCs Creation – Step 3. Its objective is to update
the classification of the classified resources taking into account the new information
obtained from the annotations. However in this case, the classification is much
simpler than in the CCs Creation, because the clustering algorithm can initialize the
clusters with the previously obtained classification. When using k-means the
convergence is faster because the initial centroids are not randomly created, they are
obtained from the CCs representation vectors.
3 Experimental results
We have used data retrieved from del.icio.us to evaluate the proposed method. We
have obtained, with the aid of two page scrapers, information about web pages
annotated by users using the Recent Bookmarks page1. The first scraper processes the
recent bookmarks page looking for resources bookmarked by a minimum number of
100 users, and stores the url of the resources. We consider this value because it has
been empirically proved in [12] that the frequency of each tag in a resource stabilizes
after the first 100 annotations. The election of this value as the convergence criteria
for the annotations of a resource implies the consideration of the resources annotated
by a minimum of 100 users (maybe including more than 100 annotations, if users
assign two or more tags to a resource). The selected bookmarks are used by the
second scraper to retrieve their information from the “Everyone’s Bookmarks” pages
in del.icio.us, that consist of a 40 pages listing with the most recent bookmarks and a
summary with the most 30 frequently assigned tag to the resource. These scrapers
were used in April 2009 obtaining a total of 25,251 resources, with 93,247,161
annotations, 1,039,796 users and 584,722 tags.
We have built two datasets2 from the information retrieved in order to evaluate the
CCs Creation and the accuracy of the method when classifying converged resources.
The first dataset (ds1) consists of 24,251 resources and their associated annotations,
and it is used to create the classification concepts using k-means. The second dataset
(ds2) consists of 1,021 new resources and their annotations, and it is used to validate
the accuracy of the classification, since these resources are classified under the CCs
created from ds1. This validation has been performed manually by ten experts (distinct
from authors) checking if the proposed CCs are appropriate to the resources.
We have created the dictionary considering those tags with a minimum number of
500 annotations, obtaining a cardinality of 4,085 tags. It is interesting to note that
although these tags represent less than the 0.1% of the retrieved tags, they are used in
more than the 95% of the retrieved annotations. In such way, we consider they
represent the semantic of the folksonomy resources in a great extent. Once defined the
dictionary, the first experiment has been performed in order to evaluate the creation of
the classification concepts using the k-means algorithm. The value of k has been
1 http://delicious.com/recent/?setcount=100&min=100
2 http://www.eslomas.com/index.php/publicaciones/sdow09/
determined using the expression k n 2 , where n=24,251 and then k = 110.
Nevertheless, many other techniques like [13] can be applied in order to optimize this
value. The initial centroids have been defined randomly, since any a-priori
knowledge is considered. The clustering has been performed using several parallel
processes. A total amount of 8 processes have been used over a four Intel Core 2 Duo
processors and the algorithm has converged in 34 iterations.
The 110 resulting clusters have been analyzed to detect possible equivalent
clusters, comparing them and merging those clusters with a similarity greater than
0.75. As result, we obtain 104 clusters.
Table 5. Example of Classification Concepts created based in their most frequent tags
Tag 1 Tag 2 Tag 3 Tag 4 Classification Concept
astronomy space science nasa Astronomy, Space and
(32,683) (27,018) (24,950) (10,651) Science
photography photos photo images
Photography and Photos
(401,387) (220,623) (199,908) (126,910)
programming development reference code
Programming
(132,858) (38,452) (34,160) (24,917)
Social, Web 2.0,
social web2.0 community socialnetworking
Community and
(132,977) (104,364) (88,841) (65,240)
Socialnetworking
Finally, a name has been assigned to each cluster to create the CCs. Names have
been created using the most relevant tag and concatenating the next ones when their
weight is greater than the 50% of the most relevant tag weight. So, a cluster which the
two most relevant tags php (weight 127,427) and programming (weight 39,743), has
been named as “Php”. Table 5 shows some examples of the classification concepts
created.
50 40
43 36
45
35 32
40
33 30
35
30 25
25 20 17
20 16
15
15
9 9
10 10
6
5 3 5
1 1 0 0 1
0 0 0 1 0
0
0
0-150
150-300
300-450
450-600
600-750
750-900
900-1050
1050-1200
1200-1350
1350-1500
[0,0.1]
[0.1,0.2]
[0.2,0.3]
[0.3,0.4]
[0.4,0.5]
[0.5,0.6]
[0.6,0.7]
[0.7,0.8]
[0.8,0.9]
[0.9,1]
Fig. 4. Number of concepts attending to the number of resources they group (left) and
to the average similarity among concepts and resources classified under (right).
Fig. 4 (left) shows the number of classification concepts obtained attending to the
number of resources they group. Most of concepts group less than 300 resources, and
there exists a concept with a large number of resources (Art and Design). Fig. 4
(right) shows the average similarity between resources and concepts. Most of the
concepts have an average similarity with their resources greater than 0.5.
The created classification concepts have been used to evaluate the accuracy of the
method when the folksonomy evolves and new resources must be classified
(converged resources). The 1,021 resources in ds2 have been encoded and the
classifier has been applied, obtaining the most similar concept for each resource, the
similarity between them, and the similarity with the rest of concepts in order to
evaluate the results. It is interesting to note that resources are classified according to
the users of the folksonomy knowledge. There exist resources that may correspond to
marketing and publishing companies that have been classified under “Art and
Design”, since del.icio.us users bookmark these pages due to their impressive design,
or because they may use them as an inspiration for other designs.
Table 6 shows a summary of the results. The number of well classified resources
by ACoAR is 951 (93.14%) versus 70 (6.86%) misclassifications. It also shows the
average similarity among CCs and the resources classified under them, and the
average difference between the similarities of the two most similar CCs (delta) to
each resource. Considering for example four CCs and a resource ri to be classified,
with a similarity of 0.05 with c1, 0.50 with c2, 0.80 with c3 and 0.00 with c4, the value
of delta is 0.30 (0.80-0.50). High delta values indicate that the concept suggested by
ACoAR has a high similarity with the resource and a low similarity with the next
most similar concept. Low delta values indicate that the resource has a similarity
value very similar with the two closest classification concepts, so the classification
could be erroneous or may be there is not enough information to classify the resource.
Table 6. Classifications results, including average values for similarity and delta
Results Average Similarity Average Delta
Well classified 951 (93.14%) 0.621868 0.242415
Misclassified 70 (6.86%) 0.483300 0.100786
Table 6 shows that the average similarity of the well classified resources is greater
than the obtained for the misclassified resources. The same happens with the delta
value. ACoAR is able to provide classification concepts with a great similarity value
for well classified resources and furthermore, the second classification concept has an
average similarity of 0.38 (0.62–0.24), that is relatively low.
As described in the method description, we can define a threshold function to
adjust the classification and to reduce the classification errors. This function allows
defining some minimum conditions that must be fulfilled to classify the resource
under the suggested classification concept, like a minimum similarity value, a
minimum delta, or a combination of both of them.
Fig. 5 (left) depicts the number of resources well classified and misclassified
according to the similarity measure between resources and the suggested CCs. Fig. 5
(right) shows the same information based in a threshold function using delta values.
Both figures show that the election of a threshold considering the delta value is a
good choice since most of misclassifications belong to the interval [0.0,0.1], while
the misclassifications considering a threshold based on the similarity value are more
homogenously distributed. Consequently, the selection of a threshold function based
on a minimum delta value of 0.10 would reduce the classification errors from 70 to
23. As a counterpart, the 238 resources well classified with a delta value under 0.1
would be assigned to converged set until they could be classified more accurately.
Therefore, the results of the classification would be 285 resources not classified
(converged set), 713 resources well classified and 23 resources misclassified,
representing a correct ratio of 96.88% (713 out of 736).
250 300
200 250
200
150
150
100
100
50
50
0
0
[0,0.1]
[0.1,0.2]
[0.2,0.3]
[0.3,0.4]
[0.4,0.5]
[0.5,0.6]
[0.6,0.7]
[0.7,0.8]
[0.8,0.9]
[0.9,1]
[0,0.1]
[0.1,0.2]
[0.2,0.3]
[0.3,0.4]
[0.4,0.5]
[0.5,0.6]
[0.6,0.7]
[0.7,0.8]
[0.8,0.9]
[0.9,1]
Correct Misclassified
Correct Misclassified
Fig. 5. Results obtained attending to the similarity value between resource and
suggested CC (left) and delta values (right).
4 Conclusions
Social tagging systems are nowadays the preferred way to classify information in the
Web 2.0 sites. As their popularity increases, many works try to solve some of their
intrinsic problems derived from their uncontrolled vocabulary. In this paper we have
proposed a modular and generic method, called ACoAR, that automatically: i) creates
a classification based in the semantics of the resources of a folksonomy using a
relatively small subset of the existing tags, and ii) allows the evolution of this
classification when the associated folksonomy evolves with new annotations. Due to
its modularity ACoAR allows the use of different clustering algorithms, as well as
different similarity measures. ACoAR3 allows browsing folksonomies by means of
the semantics of its resources increasing the chance of finding interesting results.
ACoAR can minimize the gap between Web 2.0 and the Semantic Web defining an
ontology from the CCs created, and evolving this ontology as the folksonomy does.
3 http://acoar.eslomas.com/demo
We have performed an evaluation of the proposed method using del.icio.us.
Experimental results consider an initial set of classification concepts that classify
folksonomy resources, and provide a 93.14% well classification rate when the
folksonomy evolves and the method classifies new folksonomy resources without
using any threshold. This rate may increase using an adequate threshold based on the
similarity values between resources and concepts, or based on the difference between
the similarities of the CCs (delta). ACoAR is a modular method which allows the
integration of different techniques and algorithms. Some of these modules are:
dictionary generation, convergence with different criteria, cluster algorithms,
similarity measures, etc.
Acknowledgements
Research partially supported by the Spanish Research Council under research grants
TIN2006-14738-C02-02 and TIN2008-03687.
5 References
1. Berners-Lee, T., Hendler, J., Lassila, O.: The Semantic Web. Scientific American (2001)
2. Vander Wal, T.: Folksonomy, http://vanderwal.net/folksonomy.html
3. Zhou, M., Bao, S., Wu, X., Yu, Y.: An Unsupervised Model for Exploring Hierarchical
Semantics from Social Annotations. In: Sixth International Semantic Web Conference
(ISWC 2007), LNCS, vol. 4825, pp. 680-693, Springer Verlag (2007)
4. Schmitz P.: Inducing ontology from flickr tags. In: Collaborative Web Tagging Workshop,
WWW. Edinburgh, Scotland (2006)
5. Heymann, P., García-Molina, H.: Collaborative Creation of Communal Hierarchical,
Taxonomies in Social Tagging Systems. TR Stanford InfoLab 2006-10 (2006)
6 Begelman G., Keller P., Smadja F.: Automated tag clustering: Improving search and
exploration in the tag space. In: Collaborative Web Tagging Workshop, WWW (2006)
7. Specia, L., Motta, E.: Integrating Folksonomies with the Semantic Web. In: ESWC 2007.
LNCS, vol. 4519, pp. 503-517, Springer Verlag (2007)
8. Passant, A., Laublet, P.: Meaning Of A Tag: A collaborative approach to bridge the gap
between tagging and Linked Data. In: Workshop on Linked Data on the Web, WWW.
CEUR-WS, vol. 369 (2008)
9. R. Abbasi, S. Staab, Cimiano, P.: Organizing resources on tagging systems using t-org. In
4th European Semantic Web Conferences, pp. 97-110 (2007)
10. Echarte, F., Astrain, J. J., Córdoba, A., Villadangos, J., Labat, A.: ACoAR: a method for
the automatic classification of annotated resources. In Fifth International Conference on
Knowledge Capture (K-CAP '09), pp. 181-182, ACM, New York (2009)
11. Michlmayr, E.: A Case Study on Emergent Semantics in Communities. In: Workshop on
Social Network Analysis, International Semantic Web Conference. Galway, Ireland (2005)
12. Golder, S.A., Huberman, B.A.: The Structure of Collaborative Tagging Systems. Journal of
Information Science, vol. 32, no. 2, pp. 198-208 (2005)
13. Maitra, R.: Initializing Partition-Optimization Algorithms. Computational Biology and
Bioinformatics, IEEE/ACM Transactions vol.6, no.1, 144-157 (2009)
14. Salton, G.: Automatic Text Processing: the transformation, analysis, and retrieval of
information by computer. Addison-Wesley Longman Publishing, Boston, USA (1989)