=Paper=
{{Paper
|id=None
|storemode=property
|title=A context-based algorithm for annotating educational content with Linked Data
|pdfUrl=https://ceur-ws.org/Vol-685/FIS_2010MLama.pdf
|volume=Vol-685
}}
==A context-based algorithm for annotating educational content with Linked Data==
https://ceur-ws.org/Vol-685/FIS_2010MLama.pdf
A context-based algorithm for annotating
educational content with Linked Data
Estefanı́a Otero-Garcı́a1, Juan C. Vidal1 , Manuel Lama1 , Alberto Bugarı́n1
and José E. Domenech2
1
Depto. de Electrónica e Computación, Universidade de Santiago de Compostela
15782 Santiago de Compostela, Spain
{estefania.otero,juan.vidal,manuel.lama,alberto.bugarin.diz}@usc.es
2
Netex Knowledge Factory, 15172 Oleiros, A Coruña
jose.domenech@netex.es
Abstract. In this paper we present an approach for annotating and
enriching educational contents modeled as Learning Fruits (LFs). LFs
are web books described in XML files and created to make more dynamic
and flexible the learning process. A way to reduce the cost of creating
a LF is to complete its content with information available in the Web.
The solution described in this paper combines syntactic and semantic
analysis techniques to enrich and annotate the LFs with relevant and
reliable data retrieved from the DBpedia repository.
1 Introduction
The education of youth has always been a major concern of society, and a great
deal of effort has been spent with the aim of getting the right tools to facili-
tate the learning process. However, the information society in which we live has
brought about the need to adapt traditional methods of learning to new habits
and requirements. One of the new trends are the Learning Fruits1 .
Learning Fruits (LFs) are learning pills and interactive activities on topics
of the school curriculum. More concretely, LFs are web books that stand out for
their innovative format which provide a friendly and interactive interface to fa-
cilitate the access and the navigation through the course contents. An important
feature of LFs is that they provide links to other content that complement and
extend the information to the student and the teacher. This step is expensive
when implementing a course: (i) it involves identifying which parts need to be
supplemented with external information, that is, it is necessary to determine
the important topics of the LF; and (ii) it involves selecting and analyzing the
external links in order to determine if they contain accurate information.
In this paper we propose a semantic and context-based approach to minimize
the cost of enriching and annotating LFs with external contents. Our solution
This work was supported by the Ministerio de Educación y Ciencia and the Xunta
de Galicia under the projects TSI2007-65677C02-02 and 09SIN065E respectively.
1
http://www.netex.es/santillana/eng/index.html
is based on the application of semantic technologies (i) to identify and annotate
the relevant concepts of the LF; and (ii) to retrieve the corresponding contents
from the web, in our case from Linked Data [1]. With this approach we deal
with the drawbacks of other approaches for annotating learning contents [2, 3],
because we use semantic data represented through standard and large ontologies.
However, since we only want to enrich the LF with relevant data, we need to
discriminate important information from which is not. In this work we called
context of a LF to the set of terms that determine the topics of the course. In
the annotation process, this context will establish the degree of relevance of the
concepts and relations filtered from the Linked Data repository, in our case the
DBpedia [4], and therefore will influence the creation of the the most appropriate
graph to annotate the LF terms. In other words, we will use the context of the
LF to filter the RDF triples that describe its content.
The paper is structured as follows: in Section 2 we describe the sequence
of tasks to obtain the topics that characterize a LF. In section 3 we detail the
algorithm we implemented to retrieve the most appropriate (sub)graph of the
DBpedia for the topics of the course. Finally, in Section 4 we point out some
results and the conclusions.
2 Learning Fruit Context
Figure 1 depicts the sequence of processing that must be executed to obtain
the LF context. The first step is to parse the LF XML document in order to
get its most representative fields, such as title, sections, paragraphs, etc. Once
identified, we weight the content according to the field where it has been found,
because the relevance of a term vary depending on the field in which it appears.
For example, terms in the title or marked in bold are more representative than
those of appearing in a paragraph. The result of this step is a document made
up of fields whose content is classified and weighted according to where it has
been located.
From this document, we analyse the morphology, similarity, and frequency of
the terms in order to determine the most relevant ones, and so characterize the
context of the LF:
– Morphological analysis. The morphological analysis is carried out with the
GATE tool [5], and it is used to determine the grammatical category of each
word in the document. This analysis affects the LF context creation in two
points:
XML XML Document Morphology Similarity Frequency Context
Translation
parsing analysis analysis analysis
Fig. 1. Sequence of tasks to obtain the context of a Learning Fruit
• Terms that are not representative to characterize the document content
or are not included in the DBpedia will be ruled out. For example verbs,
conjunctions, prepositions or determiners are never included in the con-
text.
• Identification of composite terms. For example, terms like Ancient Egypt
or Ramesses II cannot be considered separately, and they are detected
as regular expressions through the JAPE rules of the GATE tool.
– Similarity analysis. Since a term may appear in different forms, we create
clusters of terminological similarity in order to increase the frequency of
occurrence of a word, and also to avoid words that share the same meaning or
arise from the same word. In our approach, we use the metrics Monge-Elkan
[6] and Jaro-Winkler [7] to calculate the similarity among the document
words, because these return adequate values for words with a common root.
Thus, two given strings s and t are divided into substrings s = a1 . . . aK and
t = b1 . . . bL , and the similarity between those words is calculated as:
K
1 L
sim(s, t) = max sim (Ai , Bj ) (1)
K i=1 j=1
where sim is a secondary distance function, in our case the two metrics
mentioned above.
– Frequency analysis. The frequency is a quantitative measure which provides
the number of occurrences of a term in the LF document. It indicates the
relevance of a term within the document and, therefore, it must be com-
bined with the field weights to obtain a relative weighted frequency for each
document term:
K
j=0 (nij ∗ pj ) + s
fi = (2)
N
where nij is the number of occurrences of the term i in the field j of the
document; pj is the weight of the field; N is the sample size; and s is the
number of similar terms of the term under analysis. It is important to remark
that the similarity analysis is used to calculate the relative frequency of each
document term.
As a first approximation, we consider terms whose relative weighted fre-
quency is between 4% and 15%: terms are discarded if this frecuency is less
than 4%, because they are not representatives, or greater than 15%, because
we consider them too general.
The last step for obtaining the LF context is the translation of terms to the
English language. Although some of the properties of concepts in the DBpedia
are multi-language, most of them are only in English. Thus, if the LF is in a
different language, a translation should be performed to obtain better results.
Table 1 shows the context of a LF written in Spanish, whose subject is the
Ancient Egypt. Note that although all the terms of the Table 1 are included in
the context, some of them are too general. For example, the term land is not
relevant in the domain of the Ancient Egypt.
Table 1. Context of the LF about Ancient Egypt
Term Translation Relevance
Osiris Osiris 0.054814816
Horus Horus 0.057777777
templo temple 0.05925926
tumba tomb 0.062222224
Cleopatra VII Cleopatra VII 0.06518518
Ra Ra 0.06962963
sacerdote priest 0.07111111
dios god 0.072592594
tierras lands 0.07703704
faraón Pharaoh 0.07851852
Ramsés II Ramses II 0.08592593
Pirámides de Gizeh Pyramids of Giza 0.08888889
Nilo Nile 0.093333334
Egipto Egypt 0.11111111
Antiguo Egipto Ancient Egypt 0.14222223
3 Semantic Filtering of Linked Data
The objective of the semantic filtering is to select the DBpedia nodes that are
relevant to annotate the terms of the context that characterizes the LF doc-
ument. As it is depicted in Figure 2, to get this objective the first step is to
identify the DBpedia URI (resource) that match each context term. Once this
step is executed through the DBpedia lookup service we need to deal with two
issues: (i) the lookup service may retrieve many URIs for a given keyword, but a
term can only be paired with a single URI; and (ii) not all the relationships that
describe the URI are relevant to annotate the LF context term. For example, if
the LF is about the Ancient Egypt, we are not interested in relationships with
URIs that describe contemporary facts or persons.
To solve these two issues each URI is expanded to asses whether the node
deserves to be considered. This expansion process is an iterative deepening depth-
first algorithm [8], which carries out a detailed search through the semantic DB-
pedia graph until a depth limit. For each URI we perform a SPARQL query and
retrieve all its relationships; that is, all its related RDF triples. Thus, according
to the type of the object of those RDF triples, we take the following actions:
– If the object is a literal, we analyze the literal to check the relevance of the
relationship. We consider that a literal is related to the LF if it contains any
of the context terms, and assess its relevance with the similarity measures
described in Section 2. This analysis also takes into account the relative
frequency of the terms of the context and uses a threshold to consider which
relations are relevant or not.
LEVEL 1 2 3 4
CONTEXT
Tebas S
Frequency
Necrópolis E
Similarity T
Faraones M
... E
... A
R
N
M
T
I
A
C
N
ANOTATION N
F
O
I
T
Luxor forma parte de la antigua ciudad L
A
T
llamada Uaset (en egipcio antiguo ), o también T
E
conocida como Tebas (en griego), I
R
denominada por Homero "La ciudad de las O
I
cien puertas ", por las numerosas puertas N
N
Es la ciudad de los grandes templos del G
antiguo Egipto (Luxor y Karnak), y de las
célebres necrópolis de la ribera ...
Fig. 2. Filtering process to annotate LF documents
– If the object is an URI and we have not reached the depth limit, we con-
tinue the exploration through this URI. At this point we have distinguished
between two types of URIs:
• Those that describe a concept. For example, the term Pharaoh is repre-
sented with the URI http://dbpedia.org/resource/Pharaoh.
• Those that defines a category that allows to classify the resource. For ex-
ample, http://dbpedia.org/resource/Category:Ancient Egypt titles speci-
fies the category Ancient Egypt titles in which the resource identified by
http://dbpedia.org/resource/Pharaoh is classified.
In the case of categories, the algorithm adds a new behavior: if the search
process retrieves a resource that shares one of the categories of the resource
from which the expansion was realized, we consider this category relevant.
Therefore, the category is expanded, which means that URIs with this cat-
egory will also be processed in our filtering process.
Figure 3 shows the result obtained when this algorithm is applied to the LF
about the Ancient Egypt. In this example, we have retrieved 1579 RDF triples
for the 15 terms that compose the LF context.
4 Conclusions
In this paper we described two of the key processes for enriching the contents of
LFs with information extracted from the DBPedia. The first process identifies
the main topics of the LF, by means of the combination of frequency, similarity
and morphology analysis. From the result of this first process, a filtering process
retrieves from the DBpedia the most suitable (sub)graphs to annotate the terms
of the LF.
Fig. 3. Screenshot of the application to annotate LF documents
References
1. Bizer, C., Heath, T., Berners-Lee, T.: Linked data: The story so far. International
Journal on Semantic Web and Information Systems 5(3) (2009) 1–22
2. Jovanovic, J., Gasevic, D., Devedzic, V.: Ontology-based automatic annotation of
learning content. International Journal on Semantic Web and Information Systems
2(2) (2006) 91–119
3. Simov, K., Osenova, P.: Applying ontology-based lexicons to the semantic anno-
tation of learning objects. In: Proceedings of the RANLP-Workshop on Natural
Language Processing and Knowledge Representation for eLearning Environments,
Borovets, Bulgaria (2006) 49–55
4. Bizer, C., Lehmann, J., andSören Auer, G.K., Becker, C., Cyganiak, R., Hellmann,
S.: Dbpedia: A crystallization point for the web of data. Journal of Web Semantics
7(3) (2009) 154–165
5. Cunningham, H., Maynard, D., Bontcheva, K., Tablan, V.: A framework and graph-
ical development environment for robust nlp tools and applications. In: Proceed-
ings of the 40th Annual Meeting of the Association for Computational Linguistic
(ACL’02), Philadephia, USA (2002) 168–175
6. Monge, A., Elkan, C.: An efficient domain-independent algorithm for detecting ap-
proximately duplicate database records. In: Proceedings of the SIGMOD-Workshop
on Research Issues on Data Mining and Knowledge Discovery, Tucson, USA (2003)
7. Cohen, W.W., Ravikumar, P.D., Fienberg, S.E.: A comparison of string distance
metrics for name-matching tasks. In: Proceedings of the IJCAI-Workshop on Infor-
mation Integration on the Web (IIWeb’03), Acapulco, Mexico (2003) 73–78
8. Russell, S.J., Norvig, P.: Artificial Intelligence: A modern approach. 3rd edn. Pren-
tice Hall (2009)