The large di usion of e-gov initiatives is increasing the attention of public administrations towards the Open Data initiative. The adoption of open data in the e-gov domain produces di erent 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 administrations 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 problems 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 e ective than techniques based only on keyword representation.
The main motivation behind the success of the Linked Open Data (LOD)
initiative is related to well-known advantages coming from the interconnection of
information sources, such as improved discoverability, reusability, and utility of
information [
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) [
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 de ne 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
Europe. 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
objective of the Digital Agenda for Europe (DAE) [
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
crosslanguage ontology matching methods have been proposed [
the English service Nursery education grant by domain experts. Therefore, the automatic matching of the service text labels is not a trivial task.
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 catalogs 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 technique that uses an o -the-shelve machine translation tool and annotates the translated descriptions with Wikipedia concepts in order to extrapolate semantic 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 assume 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 con gurations
of our system that leverage di erent semantic annotation tools and the Explicit
Semantic Analysis (ESA) [
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.
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 [
The most adopted approach for cross-language ontology matching is based on
transforming a cross-lingual matching problem into a monolingual one by
leveraging automatic machine translation tools [
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 rst approach queries a web search engine and uses the results
to compute the similarity between the ontology concepts [
Another interesting work presented in literature applies the Explicit
Semantic Analysis (ESA) for cross-language link discovery [
Finally, the impact of the translation quality on the quality matching in
cross-lingual scenarios is investigated in [
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 o ered by
one or more providers8 [
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 consists of two components: the Content Analyzer analyzes the service descriptions in the source and target lists and builds a semantic annotation for each service; 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 di erent languages to be linked to the LGSL. The rst 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 Wikipediabased 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 uni ed 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 de ned as the title of a Wikipedia article. This solution 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 representing 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 rst implementation we weighing each concept in Ws by adopting the simplest schema represented by the number of occurrences of the concept in Ws. For
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
sophisticated weighing measures such as TF-IDF and BM25 [
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 Section 4) for annotating the services, each of which represents a di erent CroSeR con guration. In addition to those Wikipedia-based con gurations, we evaluated our system also by setting hybrid con gurations obtained by merging the keywords extracted from the label associated to s with the corresponding Wikipedia concepts in Ws. In that case, all the above-mentioned de nitions are still valid, with the slight di erence that the vector space is built both on keywords and Wikipedia concepts (instead of Wikipedia concepts alone). 4
We exploited di erent 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 di erent. All those services allow to con gure 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-referencing documents with Wikipedia [
Tagme. Tagme is a system that performs an accurate and on-the- y
semantic annotation of short texts via Wikipedia as knowledge base [
DBpedia Spotlight. DBpedia Spotlight [
Explicit Semantic Analysis. Explicit Semantic Analysis (ESA) is a
technique proposed by Gabrilovich and Markovitch [
We can observe that while the intuition behind Wikipedia Miner, Tagme, and DBpedia Spotlight is quite similar, ESA implements a di erent approach. Indeed, the rst 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 DBpedia 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 rst tools perform a sort of topic identi cation 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 Festivities). 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
In this experiment we evaluated the di erent con gurations in terms of: (1) e ectiveness 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 rst positions of the ranked list.
Design and Dataset. We adopted two di erent metrics: Accuracy@n (a@n)
and Mean Reciprocal Rank (MRR) [
M RR =
PN 1 i=1 ranki ;
N (1) 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 con guration. 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 con guration independently from its coverage. Hence, a con guration with a good ranking, but a small coverage will obtain a higher MRR value than a con guration with a better coverage, but a worse ranking.
The dataset is extracted from the esd-toolkit catalogue freely available online11. 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 con guration. For that con guration, only stemming and stopword elimination are performed on the text. We compared the baseline with the above-mentioned four di erent Wikipedia-based con gurations (i.e., Wikipedia Miner, Tagme, DBpedia Spotlight, ESA) as well as with a combination of keywords and Wikipedia con gurations. The latter con gurations were obtained by adding to the keywords the corresponding Wikipedia concepts generated by the di erent methods.
Results for the Dutch language are reported in Table 1. We can observe that the best con guration in terms of a@n is ESA. The con guration becomes more e ective by returning several services (n > 5), since the matching is particularly di cult in this domain. The worst con guration is that based on Wikipedia Miner. This is likely due to a low e ectiveness of Wikipedia Miner in identifying topics from very short text. Most con gurations 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 con guration 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. In this paper we proposed a tool called CroSeR that is able to perform a crosslanguage matching of e-gov services. Four di erent 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 corresponding English service for most of services. Hence, adding new external knowledge for representing a very short textual description is an e ective solution in this speci c domain. However, the correct service is generally not boosted in the rst positions of the ranked list. Therefore, as a future work we want to combine di erent 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
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).