WHAT IS THE RELATIONSHIP ABOUT?
Extracting Information about Relationships from Wikipedia
Brigitte Mathiak
1
, V
´
ıctor Manuel Mart
´
ınez Pe
˜
na
2
and Andias Wira-Alam
1
1
GESIS - Leibniz Institute for the Social Sciences, Unter Sachsenhausen 6-8, 50667 K
¨
oln, Germany
2
Institute for Web Science and Technologies, University of Koblenz-Landau, Koblenz-Landau, Germany
Keywords:
Relationship Extraction, Wikipedia.
Abstract:
What is the relationship between terms? Document analysis tells us that ”Crime” is close to ”Victim” and
not so close to ”Banana”. While for common terms like Sun and Light the nature of the relationship is clear,
the measure becomes more fuzzy when dealing with more uncommonly used terms and concepts and partial
information. Semantic relatedness is typically calculated from an encyclopedia like Wikipedia, but Wikipedia
contains a lot of information that is not common knowledge. So, when a computer calculates that Belarus and
Ukraine are closely related, what does it mean to me as a human? In this paper, we take a look at perceived
relationship and qualify it in a human-readable way. The result is a search engine, designed to take two terms
and explain how they relate to each other. We evaluate this through a user study which gauges how useful this
extra information is to humans when making a judgment about relationships.
1 INTRODUCTION
The purpose of semantic relatedness measures is to
allow computers to reason about written text. They
have many applications in natural language process-
ing and artificial intelligence (Budanitsky, 1999), and
have consequently received a lot of attention from the
research community.
However, the pure measure of relatedness in num-
bers is not very helpful to normal users. These peo-
ple are not so much interested in the quantity of relat-
edness, but the quality. We use a modified standard
method to measure relatedness between Wikipedia
entries based on a combination of link analysis and
text analysis, which evaluate comparably to other
similar measures. We leverage information used by
these methods to find text snippets on Wikipedia,
which are significant for the relationship and describe
it in a human-readable way.
The goal of these snippets is to inform the user of
the quality of the relationship, especially in the case
of an information gap. For evaluation, we have made
a user study using Mechanical Turk. We have first
asked the participants what they know about the re-
lation between two concepts, such as Barack Obama
and Chicago and then present them with a number of
snippets extracted with our method. The general feed-
back was very positive, with most participants finding
most of the snippets helpful. The learning effect was
also quite visible, while in one group 58 % of the par-
ticipants felt there was a relationship between both,
only 22 % could specify that relationship precisely,
while the others were either very general (”both in
America”) or plainly wrong (”He was the governor
the State of Illinois”). Yet, they quickly accepted the
connections made by the application as meaningful.
2 RELATED WORK
Current research explores two fundamentally differ-
ent ways to compute semantic relatedness between
two terms. The first is link-based. In a hierarchical
structure, usually a taxonomy, this typically applies to
the shortest path between the two concepts. This is of-
ten modified with other parameters, such as the depth
of the term in the taxonomy, weights derived from
the semantics of the taxonomy and so on (Leacock
et al., 1998; Slimani et al., 2006). This can also be
applied to Wikipedia, through the use of categories,
like is done with WikiRelate (Strube and Ponzetto,
2006). More accuracy is gained by exploiting the link
structure between the articles, such as in (Islam and
Inkpen, 2008), simply because there are many more
links than categories per page. Beyond the very sim-
ple distance of counting the shortest path, in (Milne
625
Mathiak B., Manuel Martínez Peña V. and Wira-Alam A..
WHAT IS THE RELATIONSHIP ABOUT? - Extracting Information about Relationships from Wikipedia.
DOI: 10.5220/0003936506250632
In Proceedings of the 8th International Conference on Web Information Systems and Technologies (WEBIST-2012), pages 625-632
ISBN: 978-989-8565-08-2
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
and Witten, 2008) the anchor texts of links and link
structure itself is used to find. They use link counts
weighted by the probability of the link occurring on
the page (inspired by tf-idf) as a vector representation
of the article while calculating the cosine similarity
on the vectors for the similarity measure. This may
look very similar to our approach, but we use tf-idf
on the terms not the links as well as a directly com-
putable measure for the link structure, so we can cal-
culate our measure online with only two requests to
the Wikipedia API. Thus, we combine a link-based
measure with the second category of text based mea-
sures.
Text based measures take an example corpus of
documents that are known to relate to the two terms
and then calculate the semantic distance between the
two document sets, thereby splitting the problem of
relatedness between terms into two problems: choos-
ing a suitable data set and calculating the semantic
distance between the documents. There are large
numbers of semantic distances to choose from: Lee
distance, tf-idf cosine similarity (Ramos, 2003), Jaro-
Winkler distance, and Approximate string matching
(Navarro, 2001), just to name a few. In (Islam and
Inkpen, 2008), the Semantic Text Similarity (STS)
has been developed as a variety of the Longest Com-
mon Subsequence (LCS) algorithm and a combina-
tion of other methods. It is optimized on very short
texts, such as single sentences and phrases. This
method was evaluated by using definitions from a dic-
tionary.
The Explicit Semantic Analysis ESA (Gabrilovich
and Markovitch, 2007) uses Wikipedia, just like our
approach, and calculates a complete matrix of term to
concept relatedness, which can be further refined by
introducing human judgments. Unlike our approach,
however, it requires the processing of the whole of
Wikipedia in a non-linear process, which is very ex-
pensive and has not been replicated on the scale since.
There are other approaches to mix both link and
text analysis, such as (Nakayama et al., 2008) which
extracts explicit relationships such as Apple is Fruit,
Computer is Machine, Colorado is U.S. state. The
goal of this paper, however, is not to use Wikipedia to
find relationships which conform to established stan-
dards and semantics, but quite the opposite, to pro-
duce explanatory text suited for unusual relationships.
3 RELATIONSHIP EXTRACTION
3.1 Architecture
The RelationWik Extractor was built as a web infor-
mation system. From a user’s point of view, it’s func-
tion it quite simple. The articles for which a relation-
ship is sought and a few parameters are input over a
web site and the system will show the results as both
a score and snippets illustrating the connection from
both sides.
The Wikipedia articles are then downloaded di-
rectly via the Wikipedia API. The text is then scanned
for additional information such as links, templates,
etc. and stripped of its Wikipedia syntax. Both text
and meta-information is stored in a database cache.
The results of the algorithms are visualized with PHP
and the Google Chart API.
3.2 Calculating Relatedness
For the actual calculation of the relationship, two dif-
ferently approaches are used. One is based on the link
structure and the other on the textual closeness of the
texts. A third approach is a mixture of both.
The first algorithm measures the connectedness of
the terms, by studying inlinks and outlinks. When
looking at connections that go over several hops, it
becomes clear that the connection can be quite thin.
For example, Banana and Berlin are connected by an
enzyme that occurs in the Banana and was identified
in Berlin. This gets worse when looking at connec-
tions with even more connectors in between. There-
fore, we decided to ignore all connections involving
more than one intermediary. The connections with
one intermediary that made the most sense occurred
in the scenario where both articles link to the same
article. This occurs, for example, when both of the
given articles A and B are connected to a category or
another larger super-concept by linking to it. Also,
in terms of computation time, it is the fastest possible
link analysis since outlinks are the easiest to extract
from.
Following that argumentation, we only look at ar-
ticles that either have intersecting outcoming links or
have a link from A to B or from B to A. Any other
pairs are given a relatedness of zero. Connected arti-
cles A and B receive a base relatedness b of 0.5 and
are given a boost of 0.1 for additional connections
(e.g. 0.7 when A and B link to each other and link
to at least one third article). This base value is further
modified by the number of backlinks of the linked-to
article in relation to the links from the other article.
WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies
626
Figure 1: The result page of the Relationship Extractor with the terms Bonn and Berlin. The algorithm is set to sentence-sized
snippets. The Score given next to the snippets is a relevance measure based on the terms used in both documents.
And, if applicable, by the ratio between common
outlinks c and total outlinks l
A
respective of l
B
.
Rel
AB
= b
l
AB
+ l
BA
l
total
+
c
2l
A
+
c
2l
B
(1)
It was originally planned to optimize the choice
of b and introduce weighting factors, but the initial
choices performed quite well in the evaluation and so
no further optimization was necessary and might have
introduced overfitting.
For the second algorithm, we use a standard co-
sine similarity between the articles. The articles
are preprocessed by stripping wikisyntax, punctuation
and symbols, removing stop words and unique terms
and using only basic stemming by removing plurals.
The term vectors A and B are calculated with tf idf.
cos(θ) =
n
i=1
A
i
× B
i
p
n
i=1
(A
i
)
2
×
p
n
i=1
(B
i
)
2
(2)
As our third method, we average the results of the
two similarities above. Again weighing was consid-
ered, but since our goal is not to optimize the relation-
ship score, but to present human-readable snippets,
the exact optimal ratio, was of no consequence to us.
4 EVALUATION
For the evaluation of the relatedness score, we use
the WordSimilarity-353 Test Collection (Finkelstein
et al., 2002). It contains 353 English word pairs
with human-assigned similarity judgments. It con-
tains antonyms, synonyms and similar words and con-
nected terms, such as Film and Popcorn. Not all terms
from the dataset can be used directly, e.g. keyboard
was mapped to the Wikipedia article Computer key-
board, Plane to Fixed-wing aircraft, etc. The map-
pings were constructed by using the Wikipedia search
engine and choosing the first entry. A few terms had
to be removed, such as Diego Maradona, because of
errors on the page. The similarity scores from all three
algorithms were tested on the data set by using a Pear-
son linear correlation coefficient (Rodgers and Nice-
wander, 1988). The results are shown in Table 1.
On closer inspection, we can observe that the co-
sine similarity tends to judge too low on somewhat
similar articles, such as Radio, Computer or Inter-
net. The links similarity on the other hand is vulner-
able to over judging relatedness due to singular rogue
links and has problems in general with articles con-
taining only few links. On average, both effects seem
to dampen each other.
The combined score is competitive with other
methods such as described in (Strube and Ponzetto,
2006), (Gabrilovich and Markovitch, 2007) and
WHATISTHERELATIONSHIPABOUT?-ExtractingInformationaboutRelationshipsfromWikipedia
627
Table 1: Pearson-correlation.
Algorithm r-Pearson coefficient
links 0.65
cosine 0.55
combined 0.69
(Strube and Ponzetto,
2006)
0.49
(Gabrilovich and
Markovitch, 2007)
0.75
(Milne and Witten,
2008)
0.69
(Milne and Witten, 2008), but it is very fast, with-
out needing any pre-processing and is able to work
online. The wait time mostly depends on the speed
of downloading both articles. All three other methods
work on a Wikipedia dump, which is more or less ex-
tensively pre-processed. With the pre-processed data-
set however, they achieve much faster response times.
We rarely go beyond 10 seconds for any given pairs
of terms, though this depends on the current traffic on
the Wikipedia server. A caching mechanism has been
implemented to alleviate the effect, lowering response
times to much lower numbers.
We have not addressed some of the serious ques-
tions in the field, such as how to match search terms
from the evaluation set to the Wikipedia articles.
Since we expect user interaction, the disambiguation
can be done with Wikipedia-specific means. Alterna-
tively, methods, as outlined in the above-mentioned
publications can be employed.
The terms, we are most interested in, are not
generally found in evaluation data sets. And with
good reason: they are terms, which have hidden or
not commonly known relationships. Human judges
would give varying degrees of relationship, depend-
ing on whether they happen to know the details about
the relationship or not. Terms, such as Belarus and
Ukraine
1
are blindly judged as completely unrelated
by 28.8 % of the participants, yet our algorithms judge
the relationship as high. We believe the algorithm is
right and that the information gap is something to be
closed.
5 SNIPPETS
We assume that a common user is not so much inter-
ested in how much two terms are related, but rather
how they are related. As we have shown above,
1
Two neighboring countries that used to be part of the
USSR.
the connections that we find are positively correlated
to the human-perceived relation between two terms.
This offers us a connection point between the terms.
For the linking algorithm it is rather simple. The links
themselves serve as a direct connection. For intersect-
ing outlinks the links to the intersecting articles are a
connection point. The cosine algorithm also gives us
a measure of which terms are most highly relevant for
the similarity and we can use those a connector. Those
connection points are then transformed into snippets
and shown to the user. All three can be mapped to one
or more specific text positions inside of Wikipedia ar-
ticles. The corresponding snippet is generated from
there, choosing either paragraph, sentence or a fixed-
size window. Since there can be more than one link
on a page and terms can be mentioned several times,
we receive a large number of snippets. The remaining
questions are: What is the optimal window? What
is best way to rank? And most importantly, is the
method beneficial in the first place?
5.1 Methodology of the User Study
We offered .25 US$ at Amazon Mechanical Turk to
40 participants to answer a short survey on our snip-
pets. We chose 10 pairs of terms (ref. 2). The last
4 term pairs are taken from the WordSimilarity-353
test set as a control group. The first 6 were cho-
sen based on a lesser known connection. We were
striving to take into account a variety of connections,
such as biographical events, historical similarity, re-
cent news, spatial closeness, same super-category and
part-of. Also, we were trying to find term pairs in
which at least one term should be known by the par-
ticipants, and preferably both.
For the study, we first asked the participants to
rate the relationship between the terms between 0 and
10, without using any secondary information sources,
such as the Internet. We then presented up to five
snippets for each pair and asked for an evaluation of
each individual snippet on a scale ranging from not
good to very good (description of the relationship).
Then we asked again for a rating of the relationship
between the two terms.
As a control mechanism, we asked the participants
to give us a catchword description of the relationship
from their point of view, in order to understand why
they would rate in a certain way. This was checked
both on the initial rating and after the snippets had
been given.
Final methodological note: As one of the partici-
pants pointed out to us (a self-proclaimed MS in Re-
source Economics), we did not offer an ”opt-out” but-
ton on our rating scales. This introduces bias towards
WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies
628
Table 2: Term pairs used for the user study.
Term 1 Term 2 connection
Barack Obama Chicago Where he went to law school
Bonn Berlin Both were capital of Germany at some point
Google Apple Recent law suit concerning Motorola
Belarus Ukraine Neighboring countries; ex-USSR
German language English language Both indogermanic languages
Dave Mustaine Metallica Founding member; guitarist
Radio Television
Cat Tiger
Sex Love
Student Professor
stating an opinion even if there is no informational ba-
sis for this opinion. The participants have to answer
the question to gain the monetary incentive, even if
they do not in fact know anything about the subject
matter, thus (in economical theory) they are prone to
answer randomly
2
. We have pondered this issue, but
since our aim is to measure how much they know in
the first place, giving them an easy way out seemed
like losing too much information and thus introducing
a bias against the uninformed. We assume they will
not answer randomly, but choose something on the
low end of the scale. The data seems to corroborate
our assumption. There is a strong gap between the
relationship rating in the first 6 pairs between before
the snippets and after, although with a large spread,
which could be a result of the random choices.
5.2 Experiences from the User Study
There were no technical difficulties with the Ama-
zon Mechanical Turk platform. However, setting up a
suitable survey was a bit tricky for non-psychologists
(see above for some pitfalls). We were forced to
change the procedures quite a bit, before finding a
method to adequately measure what we were inter-
ested in. Still, some participants simply did not play
by the rules, e.g. one participant wrote in the gen-
eral comments ”(...)we can complete this survey eas-
ily using search engines like Google(...)”, although
we stated twice and in large letters that the Internet
was not to be used. Overall the general comments
2
The literature on this is in fact extensive and not as
clear-cut as that. While the general opinion, such as (Dill-
man and Bowker, 2001) seems to be that it is better to avoid
forced-choice answer sets as it puts extra strain on the par-
ticipants, they are common practices for special purposes
like memory tasks (Martin et al., 1993). In (Smyth et al.,
2006), it has been shown that forced-choice increases both
the time spent on answering the question and the quality of
the data in a web scenario similar to ours.
were very helpful in designing better versions of the
survey.
There were some complaints concerning for ex-
ample money or a lack of understanding about the
purpose of the survey. However, we decided it was
not wise to explain what we were looking for in or-
der to avoid the interviewer-compliance bias as much
as possible. Some treated it as a game and wondered
whether they had won.
Quite a lot (37%) of the participants did not com-
plete the survey, for unknown reasons. We did not
raise the incentive to test for lack of incentive. Some
of the drop-outs can probably be explained by partic-
ipants being annoyed with the forced-choice answers.
Many of the participants that eventually dropped-out
gave mocking answers to the open-ended questions.
We did award the incentive to everyone who an-
swered more than a few questions and claimed to be
finished and used answers from unfinished question-
naires for the analysis.
5.3 Learning Effect
When looking at the median or mean differences in
the relationship rating the difference seems slight,
comparable to the variance in the control group (cf.
table 3). Now, when looking at the ratio of zero rela-
tionship votes (cf. table 4), we can see a pronounced
change between before and after. What is surprising,
though, is that the values also drop significantly for
the control group.
One part of this effect is that the snippets show
new information between well known concepts such
as Cat and Tiger, two concepts that both belong to the
same animal class. A look at the catchwords that were
provided by the participants to explain their relation-
ship rating confirms this view. New information from
the snippets was incorporated there allowing the par-
ticipants to waver from their belief that there was no
connection at all. However, while some new informa-
WHATISTHERELATIONSHIPABOUT?-ExtractingInformationaboutRelationshipsfromWikipedia
629
Table 3: Ratings before the snippets were shown and after. Numbers are (in order): median, simple mean.
Term 1 Term 2 before after
Barack Obama Chicago 7, 7.0 8, 7.2
Bonn Berlin 5, 5.4 7, 6.7
Google Apple 6, 5.3 6, 5.2
Belarus Ukraine 6, 5.9 7, 6.1
German language English language 5, 5.0 6, 5.7
Dave Mustaine Metallica 5, 5.5 7, 6.7
Radio Television 6, 6.2 7, 6.8
Cat Tiger 7, 6.7 8, 7.2
Sex Love 7, 6.7 7, 6.5
Student Professor 8, 7.1 7, 6.9
Table 4: Ratio of relationships rated as zero.
Term 1 Term 2 before after
Barack Obama Chicago 18.2 3
Bonn Berlin 28.8 1.5
Google Apple 18.2 1.5
Belarus Ukraine 28.8 0.0
German language English language 27.3 0.0
Dave Mustaine Metallica 36.4 7,8
Radio Television 13.6 0.0
Cat Tiger 9.1 0.0
Sex Love 13.6 1.5
Student Professor 13.6 1.5
tion led to upgrades in the relationship rating, it often
led to downgrades as well.
One reason for this was misinformation. A num-
ber of participants wrongly stated that Barack Obama
used to be the governor/senator of Illinois. After they
read the snippets, they revised this opinion and ac-
cordingly downgraded the relationship level. On the
other hand, most of the participants rating the rela-
tionship between the President and Chicago as zero,
gave only general catchwords, such as ”America” and
later upgraded their rating when they learned more.
For other term pairs (Apple, Google), they digested
the new information, but did not see any reason to ad-
just their rating. For some pairings, such as Bonn,
Berlin and Dave Mustaine, Metallica, there was quite
a shift in the mean rating, but mostly, it seemed, to
account for the fact that many did not know of Bonn
or Dave Mustaine beforehand.
What is interesting, though, is that the knowledge
of the participants concerning the numerical relation-
ship rating was somewhat stable, regardless of addi-
tional information. This is a good sign that relation-
ship ratings tend to become clearer with more infor-
mation; the variance gets lower, and especially the ex-
treme statement of ”not related” gets rarer. This trend
for humans to rate more gradual with more informa-
tion is especially visible when looking at which rating
category received the most ”votes” (cf. table 5). For
the pairs with hidden information this jumped from 0
to the mean, while for the control group it remained
stable.
5.4 Sentences vs. Paragraphs
Apart from the general learning effect gained from the
snippets, we also investigated which type of snippets
(paragraph or sentence) were favored and why. For
each set of snippets we asked the participants, which
they found most useful. They were split into two con-
trol groups, each alternating sentence and paragraph.
Overall, 40 participants chose 112 sentences and 89
paragraphs as the best description, almost an equal
number. The slight bias does not allow a conclu-
sive choice. We therefore decided to integrate a user
choice into the web interface.
Still, distribution was not equal. For ”Barack
Obama” and ”Chicago”, the top choices were a para-
graph with 25% and a sentence with 17.5% of the
votes. However, both were from the same connec-
tion point, telling the story about the career of Barack
Obama with some timeline. For ”Dave Mustaine”
and ”Metallica”, the participants again chose the same
WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies
630
Table 5: Rating that most participants agreed upon.
Term 1 Term 2 before after
Barack Obama Chicago 0 10
Bonn Berlin 0 8
Google Apple 0 6
Belarus Ukraine 0 7
German language English language 0 7
Dave Mustaine Metallica 0 8 &10
Radio Television 8 8
Cat Tiger 7 8
Sex Love 10 10
Student Professor 10 7
connection point about the career history, regardless
of paragraphs or sentences. The connection itself was
the criteria for voting the best snippets. In the para-
graph scenario, they just happened to be more inter-
esting than the alternatives. For the other pairs, we
had a consistent bias towards sentences.
Curiously, we found that the characteristics of a
”best” snippet is not so much tied to length, but to
containing interesting bits of knowledge, regardless
of the information content to the relationship. For ex-
ample, in the snippets for ”German language” ”En-
glish language”, the snippet ”English replaced Ger-
man as the dominant language of science Nobel Prize
laureates during the second half of the 20th century.
was chosen as best by the majority. However, the
snippet does not so much add to the knowledge of
the relationship, as it is simply an interesting piece
of trivia. In a similar vain, the sentence ”Television
sends the picture as AM and the sound as AM or FM,
with the sound carrier a fixed frequency (4.5 MHz in
the NTSC system) away from the video carrier.” won
the majority vote for ”Radio” and ”Television”.
We conclude that users prefer interesting snippets
over relevant ones and that length does not seem to
be important. Therefore, the vote of which snippet
is best cannot be used as a direct measure to which
snippet educated the user best about the relationship.
This can only be determined indirectly (e.g. through
the analysis of catchwords).
6 CONCLUSIONS AND FUTURE
WORK
By showing text snippets around the connec-
tors, the user gets a good overview on the
nature of the relationship between any two
given terms, especially in those cases in which
the relationship is usually not well known
(e.g. Berlin and Bonn). Please try yourself on
http://multiweb.gesis.org/RelationShipExtractor/.
Apart from the obvious use of educating the user
about the relationship between two terms, the snip-
pets are also an adhoc way to produce natural lan-
guage about the relationship, as can be leveraged by
Text Mining systems. In ontologies, such as DBPe-
dia, there are often only a limited number of different
properties defined between two concepts, making it
difficult to properly understand the semantics of the
property, as sometimes only a word is given. There
are too many properties defined to encode the seman-
tics manually, yet many relationships go unnoticed,
because they do not fit the scheme. The snippets of-
fer a way of solving both problems. The semantics
of an existing property can be described by using ex-
amples from the database and generating snippets for
them. Unknown relationships can be found by using
the scores and indirectly defined over the snippets.
REFERENCES
Budanitsky, A. (1999). Lexical semantic relatedness and its
application in natural language processing. Technical
report, Department of Computer Science, University
of Toronto.
Dillman, D. A. and Bowker, D. K. (2001). The Web Ques-
tionnaire Challenge to Survey Methodologists, pages
159–178. Pabst Science Publishers.
Finkelstein, L., Gabrilovich, E., Matias, Y., Rivlin, E.,
Solan, Z., Wolfman, G., and Ruppin, E. (2002). Plac-
ing search in context: the concept revisited. ACM
Trans. Inf. Syst., 20:116–131.
Gabrilovich, E. and Markovitch, S. (2007). Computing se-
mantic relatedness using wikipedia-based explicit se-
mantic analysis. In Proceedings of the 20th inter-
national joint conference on Artifical intelligence, IJ-
CAI’07, pages 1606–1611, San Francisco, CA, USA.
Morgan Kaufmann Publishers Inc.
Islam, A. and Inkpen, D. (2008). Semantic text similarity
WHATISTHERELATIONSHIPABOUT?-ExtractingInformationaboutRelationshipsfromWikipedia
631
using corpus-based word similarity and string similar-
ity. ACM Trans. Knowl. Discov. Data, 2:10:1–10:25.
Leacock, C., Miller, G. A., and Chodorow, M. (1998). Us-
ing corpus statistics and wordnet relations for sense
identification. Comput. Linguist., 24:147–165.
Martin, R. C., Bolter, J. F., Todd, M. E., Gouvier, W. D.,
and Niccolls, R. (1993). Effects of sophistication and
motivation on the detection of malingered memory
performance using a computerized forced-choice task.
Journal of Clinical and Experimental Neuropsychol-
ogy, 15(6):867–880.
Milne, D. and Witten, I. H. (2008). An effective, low-
cost measure of semantic relatedness obtained from
wikipedia links. Proceedings of the first AAAI
Workshop on Wikipedia and Artificial Intelligence
WIKIAI08, pages 25–30.
Nakayama, K., Hara, T., and Nishio, S. (2008). Wikipedia
link structure and text mining for semantic relation ex-
traction. Workshop on Semantic Search (SemSearch
2008) at the 5th European Semantic Web Conference
(ESWC 2008), pages 59–73.
Navarro, G. (2001). A guided tour to approximate string
matching. ACM Comput. Surv., 33:31–88.
Ramos, J. (2003). Using tf-idf to determine word rele-
vance in document queries. Department of Computer
Science, Rutgers University, 23515 BPO Way, Piscat-
away, NJ, 08855.
Rodgers, J. L. and Nicewander, W. A. (1988). Thirteen
ways to look at the correlation coefficient. The Amer-
ican Statistician, 42(1):59–66.
Slimani, T., Ben Yaghlane, B., and Mellouli, K. (2006).
A new similarity measure based on edge counting.
World Academy of Science, Engineering and Technol-
ogy, 23(8):34–38.
Smyth, J. D., Dillman, D. A., Christian, L. M., and Stern,
M. J. (Spring 2006). Comparing check-all and forced-
choice question formats in web surveys. Public Opin-
ion Quarterly, 70(1):66–77.
Strube, M. and Ponzetto, S. P. (2006). Wikirelate! comput-
ing semantic relatedness using wikipedia. In proceed-
ings of the 21st national conference on Artificial intel-
ligence - Volume 2, pages 1419–1424. AAAI Press.
WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies
632