=Paper=
{{Paper
|id=None
|storemode=property
|title=Cross-language Semantic Matching for Discovering Links to e-gov Services
in the LOD Cloud
|pdfUrl=https://ceur-ws.org/Vol-992/paper4.pdf
|volume=Vol-992
|dblpUrl=https://dblp.org/rec/conf/esws/NarducciPS13
}}
==Cross-language Semantic Matching for Discovering Links to e-gov Services
in the LOD Cloud==
https://ceur-ws.org/Vol-992/paper4.pdf
Cross-language semantic matching
for discovering links to e-gov services
in the LOD cloud
Fedelucio Narducci1 , Matteo Palmonari1 , and Giovanni Semeraro2
1
Department of Information Science, Systems Theory, and Communication
University of Milano-Bicocca, Italy
surname@disco.unimib.it
2
Department of Computer Science
University of Bari Aldo Moro, Italy
giovanni.semeraro@uniba.it
Abstract. The large diffusion of e-gov initiatives is increasing the at-
tention of public administrations towards the Open Data initiative. The
adoption of open data in the e-gov domain produces different advantages
in terms of more transparent government, development of better public
services, economic growth and social value. However, the process of data
opening should adopt standards and open formats. Only in this way it
is possible to share experiences with other service providers, to exploit
best practices from other cities or countries, and to be easily connected
to the Linked Open Data (LOD) cloud.
In this paper we present CroSeR (Cross-language Service Retriever), a
tool able to match and retrieve cross-language e-gov services stored in
the LOD cloud. The main goal of this work is to help public adminis-
trations to connect their e-gov services to services, provided by other
administrations, already connected to the LOD cloud. We adopted a
Wikipedia-based semantic representation in order to overcome the prob-
lems related to match really short textual descriptions associated to the
services. A preliminary evaluation on an open catalog of e-gov services
showed that the adopted techniques are promising and are more effective
than techniques based only on keyword representation.
1 Introduction and Motivations
The main motivation behind the success of the Linked Open Data (LOD) ini-
tiative is related to well-known advantages coming from the interconnection of
information sources, such as improved discoverability, reusability, and utility of
information [11]. In the last years, many governments decided to make public
their data about spending, service provision, economic indicators, and so on.
These datasets are also known as Open Government Data (OGD). As of Febru-
ary 2013, more than 1,000,000 OGD datasets have been put online by national
and local governments from more than 40 countries in 24 different languages3 .
3
http://logd.tw.rpi.edu/iogds data analytics
As the interest of governments in LOD has grown over the last years, a roadmap
consisting of three data-processing stages, namely the open stage, the link stage,
and the reuse stage, has been proposed to drive the transition from OGD to
Linked Open Government Data (LOGD) [2].
The SmartCities project4 is worth of mentioning in this context. The general
aim of that project is to create an innovation network between governments and
academic partners leading to excellence in the domain of the development and
uptake of e-services, setting a new baseline for e-service delivery in the whole
North Sea region. The project involves seven countries of the North Sea region:
England, Netherlands, Belgium, Germany, Scotland, Sweden, and Norway. One
of the main interesting results of this project is the European Local Government
Service List (LGSL) as part of the Electronic Service Delivery (ESD)-toolkit
website5 . The goal of the LGSL is to build standard lists (i.e, ESD-standards)
which define the semantics of public sector services. Each country involved into
the project is responsible to build and maintain its list of public services delivered
to the citizens, and all of those services are interlinked to the services delivered
by other countries. The ESD-standards are already linked to the LOD cloud6 .
LGSL is a great opportunity for local and national governments all over Eu-
rope. Linking national or local service catalogs to LGSL allows to make local or
national services searchable in several languages, improving also the capability
of EU citizens to access services in a foreign language country, an explicit objec-
tive of the Digital Agenda for Europe (DAE) [1]. Moreover, local and national
governments can learn best practices of service offerings across Europe and com-
pare their service to make their service offering more valuable [13]. Finally, by
linking e-service catalogs to LGSL additional information can be exploited, e.g.,
services in the LGSL are linked to a taxonomy of life events, which is useful to
enrich the service catalogs and support navigation. However, manually linking
e-service catalogs, often consisting of several hundreds - or thousands - of ser-
vices, to LGLS requires a lot of effort, which often prevents administrations from
taking advantage of becoming part of the LOD cloud.
Automatic cross-language ontology matching methods can support local and
national administrations in linking their service catalogs to LGSL, and therefore
to the LOD cloud, by reducing the cost of this activity. Although some cross-
language ontology matching methods have been proposed [16], the application of
these methods to the problem of linking local and national service catalogs has to
deal with the poor quality of the descriptions available in the catalogs. Services
are represented by minimal descriptions that often consist of the name of the
service and very few other data. Furthermore, as showed in Figure 1, the labels
associated with services linked in the LGSL are not a mere translation from a
language to another. As an example, the German service (literally translated
as) Acquisition of children daycare contributions and the Dutch service (literally
translated as) Grant Babysitting/Child Services have been manually linked to
4
http://www.smartcities.info/aim
5
http://www.esd.org.uk/esdtoolkit/
6
http://lod-cloud.net/
the English service Nursery education grant by domain experts. Therefore, the
automatic matching of the service text labels is not a trivial task.
Fig. 1: Examples of linked services in the LGSL. Services with the same ID
number are linked by the owl:sameAs relation in the LGSL. The automatic
English translation powered by Bing is reported in brackets.
In this paper we propose Cross-language Service Retriever (CroSeR), a tool
to support the linkage of a source e-service catalog represented in any language
to a target catalog represented in English, where both the source and target cat-
alogs are characterized by minimal descriptions. Ultimately, the aim of CroSeR
is to support human annotators in order to simplify the selection of possibly
matching services. Our tool exploits a cross-language ontology matching tech-
nique that uses an off-the-shelve machine translation tool and annotates the
translated descriptions with Wikipedia concepts in order to extrapolate seman-
tic representations of the services; candidate links are retrieved by evaluating
the similarity between the extracted semantic representations. Our method is
independent from the language adopted in the source catalog and does not as-
sume the availability of further information about the services other than very
short text descriptions used as names for the services.
We conduct an experiment using the English, German and Dutch catalogs
from the LGSL dataset. In the experiment we compare several configurations
of our system that leverage different semantic annotation tools and the Explicit
Semantic Analysis (ESA) [7] method. Our preliminary results show that the
method based on ESA outperforms both the methods based on other annotation
tools and a baseline where no semantic representation is used.
The rest of this paper is organized as follows. Section 2 analyzes the state of
the art. Section 3 describes the general architecture of our system, and Section
4 shows the tools exploited for obtaining a Wikipedia-based representation of
e-gov services. Finally, experimental results are presented in Section 5 and in
Section 6 the conclusion and the future work are summarized.
2 Related Work
Ontology matching, link discovery, and entity linking are tightly related research
areas. In all of these areas, automatic or semi-automatic matching techniques
are applied to discover correspondences among semantically related entities that
appear in a source and a target information source [16]. Different types of cor-
respondences have been addressed (e.g., equivalence, subclass, same as, and so
on), depending on the types of considered entities (e.g., ontology concepts, on-
tology instances, generic RDF resources) and information sources (web ontolo-
gies, linked datasets, semi-structured knowledge bases). Cross-language ontology
matching is the problem of matching a source ontology that uses terms from a
natural language L with a target ontology that uses terms from another nat-
ural language L0 (e.g., L is German and L0 is English) [17]; multi-lingual on-
tology matching is the problem of matching two ontologies that use more than
one language each, where the languages used in each ontology can also over-
lap [17]. These definitions can be easily extended to semantic matching tasks
over other types of information sources (e.g., cross-language and multi-lingual
matching of two document corpuses). In the following we discuss the most rele-
vant approaches to cross-language matching proposed over different information
sources.
The most adopted approach for cross-language ontology matching is based on
transforming a cross-lingual matching problem into a monolingual one by lever-
aging automatic machine translation tools [17, 6, 19]. However, the accuracy of
automatic machine translation tools is limited and several strategies have been
proposed to improve the quality of the final matchings. One of the most recent
approaches uses a Support Vector Machine (SVM) to learn a matching function
for ontologies represented in different languages [17]. This method uses features
defined by combining string-based and structural similarity metrics. A transla-
tion process powered by Microsoft Bing7 is used to build the feature vectors in a
unique reference language (English). A first difference with respect to our work
is that the proposed approach is deeply based on structural information derived
from the ontology; this information is very poor in our scenario and is not used
in our method. Also other translation-based approaches use structural informa-
tion, i.e., neighboring concepts [6] and instances [19], which is not available in
our scenario.
Two ontology matching methods have been recently proposed, which use
the concepts’ names, labels, and comments to build search keywords and query
web data. A first approach queries a web search engine and uses the results
to compute the similarity between the ontology concepts [14]. The system sup-
ports also cross-language alignment leveraging the Bing API to translate the
keywords. A second approach submit queries to the Wikipedia search engine [8].
The similarity between a source and target concept is based on the similarity
of the Wikipedia articles retrieved for the concepts. Cross-language matching is
supported by using the links between the articles written in different languages,
7
http://www.bing.com/translator
which are available in Wikipedia, and by comparing the articles in a common
language. The authors observe that their approach has problems when it tries to
match equivalent ontology elements that use a different vocabulary and lead to
very different translations (e.g., Autor von(de) and has written(en)). Despite we
do also leverage Wikipedia, our matching process uses semantic annotation tools
and ESA. We can therefore incorporate light-weight disambiguation techniques
(provided by the semantic annotation tools) and match entities that, when trans-
lated, are represented with significantly different terms (in particular when the
system use the ESA model).
Another interesting work presented in literature applies the Explicit Seman-
tic Analysis (ESA) for cross-language link discovery [9]. The goal of that paper
is to investigate how to automatically generate cross-language links between re-
sources in large document collections. The authors show that the semantic sim-
ilarity based on ESA is able to produce results comparable to those achieved by
graph-based methods. However, in this specific domain, algorithms can leverage
a significant amount of text that is not available in our case.
Finally, the impact of the translation quality on the quality matching in
cross-lingual scenarios is investigated in [5]. From that research it emerges that
good translation quality is a prerequisite for achieving good quality cross-lingual
matches. However, this is likely true only compared to accuracy of monolingual
matching [17].
3 CroSeR: Cross-language Service Retriever
Fig. 2: CroSeR general architecture
We assume that each service is labeled by a short textual description (called
service label in the following), e.g., see examples in Figure 1, and represents
an abstract service, i.e., a high-level description of concrete services offered by
one or more providers8 [13]. The intuitive semantics of a link between a source
8
http://www.smartcities.info/files/Smart Cities Brief What is a service list.pdf
and a target service description is that the two descriptions refer to the same
abstract service. Although service descriptions conceptually represent categories
of concrete services, which are offered by actual providers, these descriptions are
represented as ontology instances and, consistently with the ESD approach, we
represent the links with owl:sameAs relations. In order to discover links between
a source list of services described in an arbitrary language L, and a target list
of services described in English (LGSL), we design a matching algorithm for
retrieving the top-k services that are best candidate to be linked to a given
source service. We implemented the matching algorithm in a system called Cross-
Language Service Retriever (CroSeR).
Given a source service described in an arbitrary language L, and a target list
of services described in English (LGSL), the CroSeR system returns a ranked
list of k services in LGSL that are candidate to be owl:sameAs-related to the
source service. The user will use the service list returned by CroSeR to validate
the link between the source and the target service.
The architecture of CroSeR is depicted in Figure 2. The CroSeR system con-
sists of two components: the Content Analyzer analyzes the service descriptions
in the source and target lists and builds a semantic annotation for each ser-
vice; the Retriever discovers the links between the source and target services by
computing the similarity between the semantic annotations.
Content Analyzer. The input to the Content Analyzer are the service
catalogs in different languages to be linked to the LGSL. The first step performed
by this component is to translate the service labels in English by leveraging the
Bing API9 . Next, the translated labels are exploited for obtaining the Wikipedia-
based annotation of the services. For each service s, a set of Wikipedia concepts
Ws semantically related to the service label is generated; we call Wikipedia-based
annotation of s the set Ws . The set Ws is built by processing the short text in
the label of the service with semantic annotation techniques. The Wikipedia
concepts are generated for all the services (English ones, as well), since we need
to adopt an unified representation.
The Wikipedia-based annotation aim to capture the main topics (represented
by the corresponding Wikipedia concepts) related to a service. In this context,
a Wikipedia concept is defined as the title of a Wikipedia article. This solu-
tion is also able to perform a sort of word sense disambiguation of a natural
language text without the application of elaborate algorithms based on lexical
ontologies such as Wordnet10 . Furthermore, the annotation of a service with a
set of Wikipedia concepts represents an additional link between the service and
the LOD cloud (by using DBpedia as input).
Retriever. The Retriever adopts the Vector Space Model (VSM) for repre-
senting services in a multidimensional space where each dimension is a Wikipedia
concept. Therefore, each service is represented as a point in that space. In this
first implementation we weighing each concept in Ws by adopting the simplest
schema represented by the number of occurrences of the concept in Ws . For-
9
http://www.microsoft.com/en-us/translator/
10
http://wordnet.princeton.edu/
mally, each service is represented as a vector s =< w1 , . . . , wn > where wk is
the occurrence of the the Wikipedia concept k in Ws . We guess that more so-
phisticated weighing measures such as TF-IDF and BM25 [15] do not improve
the performance of our system at this step.
Finally, the similarity between two services (vectors) is computed in terms
of cosine similarity. Therefore, given a service in one of the supported languages
(the query), the retriever is able to return a ranked list of the most similar
English services from the LGSL.
Please, note that we adopt a variety of techniques (described in details in Sec-
tion 4) for annotating the services, each of which represents a different CroSeR
configuration. In addition to those Wikipedia-based configurations, we evaluated
our system also by setting hybrid configurations obtained by merging the key-
words extracted from the label associated to s with the corresponding Wikipedia
concepts in Ws . In that case, all the above-mentioned definitions are still valid,
with the slight difference that the vector space is built both on keywords and
Wikipedia concepts (instead of Wikipedia concepts alone).
4 Semantic annotation of e-gov services
We exploited different techniques for semantically annotate service labels with
a set of Wikipedia concepts. In particular, we adopted three well-known on-line
services that perform semantic annotation, namely Wikipedia Miner, Tagme,
DBpedia Spotlight, and we implemented a semantic feature generation tool based
on the Explicit Semantic Analysis (ESA) technique. The on-line services take
as input a text description (the service label), and return a set of Wikipedia
concepts that emerge from the input text. Also ESA generates a set of Wikipedia
concepts as output, but the insight behind it is quite different. All those services
allow to configure some parameters in order to favor recall or precision. Given
the conciseness of the input text in our domain, we set those parameters for
improving the recall instead of precision.
Wikipedia Miner. Wikipedia Miner is a tool for automatically cross-referen-
cing documents with Wikipedia [12]. The software is trained on Wikipedia ar-
ticles, and thus learns to disambiguate and detect links in the same way as
Wikipedia editors [3]. The first step is the disambiguation of terms within the
text by means of a classifier. Several features are exploited for learning the clas-
sifier. The two main features are commonness and relatedness [10]. The com-
monness of a Wikipedia article is defined by the number of times it is used
as destination from some anchor text. For example, the anchor text tree has
a higher commonness value for the article Tree (plant) than the article Tree
(data structure). However, this feature is not sufficient to disambiguate a term.
Therefore, the algorithm compares each possible sense (Wikipedia article) for a
given term with its surrounding context by computing the relatedness value. This
measure computes the similarity of two articles by comparing their incoming and
outgoing links.
Tagme. Tagme is a system that performs an accurate and on-the-fly seman-
tic annotation of short texts via Wikipedia as knowledge base [4]. The annota-
tion process is composed of two main phases: the anchor disambiguation and
the anchor pruning. The disambiguation is based on a process called ”collective
agreement”. Given an anchor text a associated with a set of candidate Wikipedia
pages Pa , each other anchor in the text casts a vote for each candidate annota-
tion in Pa . As in Wikipedia Miner, the vote is based on the commonness and
the relatedness between the candidate page pi ∈ Pa and the candidate pages
associated to the other anchors in the text. Subsequently, an anchor pruning is
performed by deleting the candidate pages in Pa considered to be not mean-
ingful. This process takes into account the probability of the anchor text to be
used as link in Wikipedia and the coherence between the candidate page and
the candidate pages of other anchors in the text.
DBpedia Spotlight. DBpedia Spotlight [11] was designed with the explicit
goal of connecting unstructured text to the LOD cloud by using DBpedia as
hub. Also in this case the output is a set of Wikipedia articles related to a text
retrieved by following the URI of the DBpedia instances. The annotation process
works in four-stages. First, the text is analyzed in order to select the phrases that
may indicate a mention of a DBpedia resource. In this step any spots that are
only composed of verbs, adjectives, adverbs and prepositions are disregarded.
Subsequently, a set of candidate DBpedia resources is built by mapping the
spotted phrase to resources that are candidate disambiguations for that phrase.
As in the abovementioned tools, the disambiguation process uses the context
around the spotted phrase to decide for the best choice amongst the candidates.
Finally, there is a configuration step whose goal is to set the best parameter
values for the text to be disambiguated.
Explicit Semantic Analysis. Explicit Semantic Analysis (ESA) is a tech-
nique proposed by Gabrilovich and Markovitch [7], that uses Wikipedia as a
space of concepts explicitly defined and described by humans. The idea is that
the meaning of a generic term (e.g. ball ) can be described by a list of concepts
it refers to (e.g. the Wikipedia articles: volleyball, soccer, football,...). Formally,
given the space of Wikipedia concepts C = {c1 , c2 , ..., cn }, a term ti can be rep-
resented by its semantic interpretation vector vi =< wi1 , wi2 , ..., win >, where
wij represents the strength of the association between ti and cj . Weights are
obtained from a matrix T , called esa-matrix, in which each of the n columns
corresponds to a concept, and each row corresponds to a term of the Wikipedia
vocabulary, i.e. the set of distinct terms in the corpus of all Wikipedia articles.
Cell T [i, j] contains wij , the tf-idf value of term ti in the article (concept) cj .
The semantic interpretation vector for a text fragment f (i.e. a sentence, a doc-
ument, a service label) is obtained by computing the centroid of the semantic
interpretation vectors associated with terms occurring in f .
We can observe that while the intuition behind Wikipedia Miner, Tagme,
and DBpedia Spotlight is quite similar, ESA implements a different approach.
Indeed, the first three tools identify Wikipedia concepts already present in the
text, conversely ESA generates new articles related to a given text by using
Wikipedia as knowledge base. As an example, let us suppose that we want to
annotate the service label Home Schooling. Wikipedia Miner, Tagme and DBpe-
dia Spotlight link it to the Wikipedia article Homeschooling, while ESA generates
(as centroid vector) the Wikipedia articles Home, School, Education, Family, ....
Hence, we can state that the three first tools perform a sort of topic identification
of a given text, while ESA performs a feature generation process by adding new
knowledge to the input text. Another example enforces the motivation behind
the need of producing a sematic annotation of the service labels. Let’s consider
the English service label Licences - entertainment and the corresponding Dutch
service in LGSL Vergunning voor Festiviteiten (translated as: Permit for Festiv-
ities). A keyword-based approach never matches these two services. Conversely,
the Tagme annotation generates for the English Service the Wikipedia concepts
License, Entertainment, and for the translated Dutch label the concepts License,
Festival.
5 Experimental Evaluation
In this experiment we evaluated the different configurations in terms of:
(1) effectiveness in retrieving the correct service in a list of k service to be
presented to the user, (2) capability in boosting the correct service in the first
positions of the ranked list.
Design and Dataset. We adopted two different metrics: Accuracy@n (a@n)
and Mean Reciprocal Rank (MRR) [18]. The a@n is calculated considering only
the first n retrieved services. If the correct service occurs in the top-n items,
the service is marked as correctly retrieved. We considered different values of
n = 1, 3, 5, 10, 20, 30. The second metric (MRR) considers the rank of the correct
retrieved service and is defined as follows:
PN 1
i=1 ranki
M RR = , (1)
N
where ranki is the rank of the correctly retrieved servicei in the ranked list, and
N is the number of the services correctly retrieved. The higher is the position of
the services correctly retrieved in the list, the higher is the MRR value for a given
configuration. We decided to adopt this normalization instead of considering
N as the total number of the services in the catalogue in order to evaluate
the ranking of each configuration independently from its coverage. Hence, a
configuration with a good ranking, but a small coverage will obtain a higher
MRR value than a configuration with a better coverage, but a worse ranking.
The dataset is extracted from the esd-toolkit catalogue freely available on-
line11 . We indexed English, Dutch, and German services. The number of Dutch
services is 225, and the number German services is 190. For each service we
extracted and represented its textual label in terms of Wikipedia concepts by
exploiting the methods described in Section 4. The labels have an average length
of about three words.
11
http://standards.esd-toolkit.eu/EuOverview.aspx
Results. The baseline of our experiment is the keyword-based configuration.
For that configuration, only stemming and stopword elimination are performed
on the text. We compared the baseline with the above-mentioned four different
Wikipedia-based configurations (i.e., Wikipedia Miner, Tagme, DBpedia Spot-
light, ESA) as well as with a combination of keywords and Wikipedia configu-
rations. The latter configurations were obtained by adding to the keywords the
corresponding Wikipedia concepts generated by the different methods.
Results for the Dutch language are reported in Table 1. We can observe that
the best configuration in terms of a@n is ESA. The configuration becomes more
effective by returning several services (n > 5), since the matching is particularly
difficult in this domain. The worst configuration is that based on Wikipedia
Miner. This is likely due to a low effectiveness of Wikipedia Miner in identifying
topics from very short text. Most configurations seem to improve their accuracy
by merging the Wikipedia concepts with keywords; however they do not generally
overcome the baseline. The only tool for which the keywords do not generally
improve the performance is ESA. However, by analyzing the MRR values we can
observe that ESA produces the worst ranking of the retrieved list of services.
Conversely, the method with the best ranking is Wikipedia Miner, but it has a
very small coverage (only 24 services). Hence, we can state that ESA is able to
identify the correct correspondence for the largest number of services (∼ 82%
of services) but under the condition to extend the list of retrieved service. Very
similar results are shown for the German services (see Table 2). Also in this
experiment ESA is the configuration with the best accuracy, while Wikipedia
Miner achieves the best ranking. These results are very promising since the
service labels in LGSL are written and matched by human experts and they are
not always mere translations of the English labels.
Table 1: Accuracy@n and MRR for the Dutch language.
The highest values are reported in bold (total services = 225).
Configuration a@1 a@3 a@5 a@10 a@20 a@30 MRR N
keyword 0.333 0.458 0.502 0.538 0.542 0.547 0.610 123
tagme 0.120 0.164 0.178 0.182 0.187 0.187 0.643 42
tagme+keyword 0.316 0.453 0.484 0.551 0.560 0.569 0.555 128
wikiminer 0.080 0.093 0.107 0.107 0.107 0.107 0.750 24
wikiminer+keyword 0.324 0.440 0.484 0.529 0.542 0.547 0.593 123
esa 0.311 0.480 0.538 0.622 0.689 0.716 0.378 185
esa+keyword 0.311 0.476 0.542 0.622 0.689 0.716 0.378 185
dbpedia 0.182 0.236 0.244 0.249 0.258 0.258 0.707 58
dbpedia+keyword 0.329 0.449 0.498 0.556 0.569 0.573 0.574 129
Table 2: Accuracy@n and MRR for the German language.
The highest values are reported in bold (total services = 190).
Configuration a@1 a@3 a@5 a@10 a@20 a@30 MRR N
keyword 0.204 0.338 0.396 0.413 0.418 0.418 0.489 94
tagme 0.124 0.147 0.151 0.151 0.151 0.151 0.824 34
tagme+keyword 0.218 0.342 0.400 0.427 0.431 0.431 0.505 97
wikiminer 0.098 0.124 0.124 0.129 0.129 0.129 0.759 29
wikiminer+keyword 0.218 0.342 0.396 0.422 0.427 0.427 0.510 96
esa 0.244 0.360 0.431 0.484 0.556 0.600 0.350 157
keyword+esa 0.244 0.360 0.431 0.484 0.556 0.600 0.350 157
dbpedia 0.138 0.169 0.182 0.182 0.182 0.182 0.756 41
dbpedia+keyword 0.231 0.360 0.413 0.440 0.440 0.440 0.525 99
6 Conclusions and Future Work
In this paper we proposed a tool called CroSeR that is able to perform a cross-
language matching of e-gov services. Four different semantic Wikipedia-based
representations were investigated. The most accurate representation turned out
to be ESA that, given a Dutch or German service, is able to retrieve the cor-
responding English service for most of services. Hence, adding new external
knowledge for representing a very short textual description is an effective solu-
tion in this specific domain. However, the correct service is generally not boosted
in the first positions of the ranked list. Therefore, as a future work we want to
combine different representations to generate the ranked list. For example, we
can start by adopting the representation with the highest MRR value, and then
shifting to the representations with a worse ranking but a better accuracy. We
want also to extend the experiment to the other languages in the esd-toolkit
catalogue (Belgian, Norwegian, Swedish). Another idea is to exploit also the
other relations stored into the catalogue (life event, interactions) for improving
the service matching. Finally, we will carry out a user study where users can
directly formulate their information need instead of using the service label as
query.
7 Acknowledgments
The work presented in this paper has been partially supported by the Italian
PON project PON01 00861 SMART (Services and Meta-services for smART
eGovernment).
References
1. European Commission. A digital agenda for europe. COM(2010) 245 final/2, 2010.
2. L. Ding, V. Peristeras, and M. Hausenblas. Linked Open Government Data. IEEE
Intelligent Systems, 27(3):11–15, 2012.
3. Samuel Fernando, Mark Hall, Eneko Agirre, Aitor Soroa, Paul Clough, and Mark
Stevenson. Comparing taxonomies for organising collections of documents. In
Proceedings of COLING ’12, pages 879–894. The COLING 2012 Organizing Com-
mittee, 2012.
4. P. Ferragina and U. Scaiella. Tagme: on-the-fly annotation of short text fragments
(by Wikipedia entities). In Proceedings of CIKM ’10, pages 1625–1628. ACM,
2010.
5. B. Fu, R. Brennan, and D. O’Sullivan. Cross-lingual ontology mapping — an
investigation of the impact of machine translation. In Proceedings of ASWC ’09,
pages 1–15. Springer-Verlag, 2009.
6. B. Fu, R. Brennan, and D. O’Sullivan. Using pseudo feedback to improve cross-
lingual ontology mapping. In Proceedings of ESWC ’11, pages 336–351. Springer-
Verlag, 2011.
7. E. Gabrilovich and S. Markovitch. Wikipedia-based semantic interpretation for
natural language processing. Journal of Artificial Intelligence Research (JAIR),
34:443–498, 2009.
8. S. Hertling and H. Paulheim. WikiMatch - Using Wikipedia for Ontology Match-
ing. In Proceedings of the 7th International Workshop on Ontology Matching (OM
2012). CEUR, 2012.
9. P. Knoth, L. Zilka, and Z. Zdrahal. Using explicit semantic analysis for cross-
lingual link discovery. In Proceedings of 5th International Workshop on Cross
Lingual Information Access: Computational Linguistics and the Information Need
of Multilingual Societies, 2011.
10. O. Medelyan, I.H. Witten, and D. Milne. Topic indexing with Wikipedia. In Pro-
ceedings of AAAI Workshop on Wikipedia and Artificial Intelligence: an Evolving
Synergy, pages 19–24. AAAI Press, 2008.
11. P. N. Mendes, M. Jakob, A. Garcı́a-Silva, and C. Bizer. DBpedia Spotlight: Shed-
ding light on the web of documents. In Proceedings of I-SEMANTICS ’10, pages
1–8. ACM, 2011.
12. D. Milne and I. H. Witten. Learning to link with Wikipedia. In Proceedings of
CIKM ’08, pages 509–518. ACM, 2008.
13. M. Palmonari, G. Viscusi, and C. Batini. A semantic repository approach to
improve the government to business relationship. Data Knowl. Eng., 65(3):485–
511, 2008.
14. H. Paulheim. WeSeE-Match results for OEAI 2012. In Proceedings of the 7th
International Workshop on Ontology Matching (OM 2012), 2012.
15. Stephen Robertson and Hugo Zaragoza. The probabilistic relevance framework:
BM25 and beyond. Found. Trends Inf. Retr., 3(4):333–389, April 2009.
16. P. Shvaiko and J. Euzenat. Ontology matching: State of the art and future chal-
lenges. IEEE Trans. Knowl. Data Eng., 25(1):158–176, 2013.
17. D. Spohr, L. Hollink, and P. Cimiano. A machine learning approach to multilingual
and cross-lingual ontology matching. In Proceedings of ISWC 2011, pages 665–680.
Springer-Verlag, 2011.
18. E. M. Voorhees. TREC-8 question answering track report. In Proceedings of
TREC-8, pages 77–82. NIST Special Publication 500-246, 1999.
19. S. Wang, A. Isaac, B. Schopman, S. Schlobach, and L. Van Der Meij. Matching
multi-lingual subject vocabularies. In Proceedings of ECDL ’09, pages 125–137,
Berlin, Heidelberg, 2009. Springer-Verlag.