The need for interoperability among geospatial resources in different natural languages evidences the difficulties to cope with domain representations highly dependent of the culture in which they have been conceived. In this paper we characterize the problem of representing cultural discrepancies in ontologies. We argue that such differences can be accounted for at the ontology terminological layer by means of external elaborated models of linguistic information associated to ontologies. With the aim of showing how external models can cater for cultural discrepancies, we compare two versions of an ontology of the hydrographical domain: hydrOntology. The first version makes use of the labeling system supported by RDF(S) and OWL to include multilingual linguistic information in the ontology. The second version relies on the Linguistic Information Repository model (LIR) to associate structured multilingual information to ontology concepts. In this paper we propose an extension to the LIR to better capture linguistic and cultural specificities within and across languages.
The symbiosis between ontologies and natural language has
proven more and more relevant on the light of the growing
interest and use of Semantic Web technologies. Ontologies that
are well-documented in a natural language not only provide
humans with a better understanding of the world model they
represent, but also a better exploitation by the systems that may
use them. This “grounding in natural language” is believed to
provide improvements in tasks such as ontology-based
information extraction, ontology learning and population from
text, or ontology verbalization, as pointed out in [
Nowadays, there is a growing demand for ontology-based applications that need to interact with information in different natural languages, i.e., with multilingual information. This is the case of numerous international organizations currently introducing semantic technologies in their information systems, such as the Food and Agriculture Organization or the World Health Organization, to mention just a few. Such organizations have to manage information and resources available in more than a dozen of different natural languages, and have to customize the information they produce to a similar number of linguistic communities.
In the present research, we are concerned with a further use case: the geospatial information. The importance of multilingualism in this domain lies in the need for interoperability among multiple geospatial resources in different languages, and a flexible human interaction with multilingual information. For many years, geospatial information producers have focused on the collection of data in one or different languages without considering the interoperability level among them. If we take as example the case of Spain, data have been collected from multiple producers at different levels (national, regional and local), and in the different languages that are official in the country (Spanish, Catalan, Basque and Galician), but there have been no efforts to make these data interoperable.
Today, the widespread use of geospatial information and the globalization phenomenon have brought about a radical shift in the conception of this information. In this context, multilingualism has reached a pre-eminent position in the international scene. The rapid emergence of international projects such as EuroGeoNames 1 confirms this trend. The main goal of EuroGeoNames is to implement an interoperable internet service that will provide access to the official, multilingual geographical name data held at national level and make them available at European label.
Ideally, the “meaning” expressed by ontologies would provide the
“glue” between geospatial communities [
Other international projects in a similar line of research are:
eSDI-NET+ (http://www.esdinetplus.eu) or GIS4EU
(http://www.gis4eu.org/)
exponent of what has been called culturally-dependant domains
[
The rest of the paper is structured as follows. In section 2 we present the state of the art on formalisms and models to represent linguistic information in ontologies. Then, in section 3 we try to characterize the problem of conceptual mismatches or discrepancies among conceptualizations in multilingual knowledge resources. Section 3 is devoted to a brief description of hydrOntology. The inclusion of linguistic information in the ontology by means of the RDF(S) labels is described in section 4. Then, the LIR model is presented in section 5, and its instantiation with the linguistic information related to hydrOntology is detailed in section 6. By describing the two versions of hydrOntology, we aim at showing the main benefits and drawbacks of each modelling modality. Finally, we conclude the paper in section 7.
Most of the ontologies available nowadays in the Web are documented in English, i.e., the human-readable information associated to ontology classes and properties consists of terms and glosses in English. Most of these ontologies, not to say all of them, make use of the rdfs:label and rdfs:comment properties of the RDF Schema vocabulary, a recommendation of the W3C 2 In [3] the terminological layer in an ontology is defined as the terms or labels selected to name ontology elements.
Consortium to provide “a human-readable version of a resource’s name”3.
It is also specified that labels can be annotated using the “language tagging” facility of RDF literals 4, which permits to indicate the natural language used in a certain information object. The RDF(S) properties can be complemented by Dublin Core metadata 5 that have been created to describe resources of information systems. Examples of the Dublin Core Metadata elements are: title, creator, subject or description. Since it is possible to attach as many metadata as wished, this has been used to associate the same metadata in different natural languages to obtain an ontology documented in different natural languages, in other words, to obtain a multilingual ontology. This is precisely one of the main advantages of this representation modality, namely, associating as much information in different languages as wished.
However, we identify several drawbacks for an appropriate
exploitation of the resulting multilingual ontologies:
(1) All annotations are referred to the ontology element they are
attached to, but it is not possible to define any relation among the
linguistic annotations themselves. This results in a bunch of
unrelated data whose motivation is difficult to understand even
for a human user.
(2) When different labels in the same language are attached to the
same ontology element, absolute synonym or exact equivalence is
assumed among the labels. As reported in [
On the light of the drawbacks outlined, additional approaches
have been proposed to connect linguistic and ontological
information. In this sense, we will first refer to the Linguistic
Watermark initiative [
using their content to “enrich and document ontological objects”. The authors already propose a description for WordNet and a set of dictionaries called DICT. Their idea would be to directly import the linguistic information contained in those resources and integrate it in the ontology. However, it seems as if the reused information is included in the ontology by making use of the RDF(S) properties, and this shows the same disadvantages presented above. This approach is technologically supported by the OntoLing Protégé plugin6.
A further effort to associate linguistic information to ontologies is
represented by the LexInfo [
Finally, we will briefly mention the Simple Knowledge
Organization System (SKOS) [
The reconciliation of different representations (within the same
natural language) can be solved by establishing mappings among
those representations. When facing representations in different
languages, the mapping process results in a multilingual system.
A collection of mapping approaches with monolingual and
multilingual resources can be found in [
modifying the conceptualization (for a detailed analysis of modeling modalities to represent multilinguality in knowledgebased systems, see [1]).
To the best of our knowledge, the most recurrent conceptual discrepancies could be systematically classified as follows: (a) 1:1 or exact equivalence (as illustrated in Figure 1) (b) n:1 subsumption relation (isSubsumedBy) (illustrated in Figure 2) (c) 1:n subsumption relation (subsumes) (represented by Figure 3) In case (a) both conceptualizations or world views share the same structure and the same granularity level. This is normally reflected in the language by means of a word or term that designates that concept. In the situation represented by (b), the original conceptualization (the one belonging to the English language) makes a more fine grained distinction of a certain reality that does not correlate with the granularity level in the target representation of the same reality. In that case, the target concept is slightly more general, and it could be understood as encompassing the n concepts in the original conceptualization. This results in two terms in the English language, for instance, to designate those two concepts, whereas in the target culture, only one term is available. The last case (c) depicts the same situation as in (b) but exactly the other way round.
Watercourse
Curso de agua River
Río Ditch
Dique
Acequia However, if our objective is to rely on one ontology to “glue” the different conceptualizations of reality that cultures make, we will need to assume a trade-off between what is represented in the ontology and what is left out, so that every culture can feel that conceptualization as its own, and can meet its representation needs.
Coming back to case (c), if we agree on representing the view of
the English culture in the ontology, we will be missing the
granularity level of the Spanish world view. We think that those
cultural discrepancies could still be reported at the terminological
layer of the ontology. A further option would be to integrate the
granularity level of the target culture in the common ontology,
but, here again, a certain compromise would be necessary.
However, if there are more than two or three cultures and
languages involved in the multilingual ontology the suggested
option will not be an optimal one. In that case, one possible
solution could be to include specific language modules in the
ontology, and support different linearizations or visualizations of
the same ontology according to the language selected. To the best
of our knowledge, currently there is no system to support this
latter option. Therefore, our proposal is to account for those
categorization mismatches in an elaborated model of lexical and
terminological information, separated from the ontology.
4. hydrOntology: an ontology of the
hydrographical domain
hydrOntology [
In order to develop this ontology following a top-down approach,
different knowledge models (feature catalogues of the IGN-E, the
Water Framework European Directive, the Alexandria Digital
Library, the UNESCO Thesaurus, Getty Thesaurus, GeoNames,
FACC codes, EuroGlobalMap, EuroRegionalMap,
EuroGeonames, several Spanish Gazetteers and many others)
have been consulted; additionally, some integration issues related
to geographic information and several structuring criteria [
Currently, the hydrOntology ontology is available in two versions. The first one in Prótége makes use of the RDF(S) labeling model to document the ontology in natural language. In a subsequent stage, the ontology was associated to the Linguistic Information Repository (LIR) model, currently supported in the NeOn Toolkit. The first version of the ontology is available in Spanish and English, whereas in the second version two more languages were added: French and Catalan, as will be reported in section 6. Regarding the first version of the ontology, hydrOntology was originally developed in Spanish, and therefore, the labels given to the concepts in the original ontology were in Spanish. Later on, English labels were also related to ontology concepts, and the language of those labels was specified by means of language tags. Definitions or glosses describing the concepts were also included in Spanish and English, if available, by making use of the comment property. Finally, one metadata element of Dublin Core (source) and one additional annotation (provenance) were used to report about the resources from which the different definitions (comments) and labels had been obtained, respectively. It must be noted that the process of documentation was not systematically carried out for different reasons, and not all types of annotations are available for every concept.
A snapshot of the class hierarchy of hydrOntology in the Protégé ontology editor can be seen in Figure 4. The concept Río (River) has been chosen for illustration. It has nine annotations related to it: three provenance annotations, two comment annotations, three label annotations, and one source annotation. As already reported, the provenance annotation gives information about the linguistic resources (glossaries, thesauri, dictionaries, etc.) labels have been obtained from. Since there are no mechanisms for relating the label (e.g. River) with its source of provenance (e.g. Water Framework Directive), the authors have decided to include the label in the provenance text for the sake of clarity (e.g.; “River – Water Framework Directive. European Union”@en).
Two comments are included, one in Spanish, and one in English, though no relation to any of the labels is given. Finally, three label annotations are given: two in Spanish (in addition to the one given in the URI, i.e., Río) and one in English. The two additional labels are Curso de agua principal (Main Watercourse), and Curso fluvial (Watercourse). According to the authors, the main difference among the three synonyms is the discourse register. The label Río is the general word, and would appear in general documents, whereas the other two additional labels would only come up in technical documentation managed by experts in the domain. It is worth noting that such fine-grained aspects could be relevant for certain indexing or information extraction tasks, but cannot be made explicit in the RDF(S) labeling model. Regarding the English translation, River, it is not possible to know to which of the Spanish labels is related to or is translation of. River is considered to be in a complete equivalence relation with Río, which would be appropriate in this case, but it is rarely the case, as explained in section 2. However, the RDF(S) labeling model does not offer any means to report about those cultural differences that, more often than not, occur between two languages.
Because of these deficiencies in the representation of multilinguality in ontologies in OWL, and with the aim of giving response to the increasing demand for multilingual ontologies, the Linguistic Information Repository (LIR) model was developed. In the next section, we present the LIR model and how it aims at solving some of the representation problems identified so far.
The Linguistic Information Repository or LIR is a proprietary
model expected to be published and used with domain ontologies.
In itself, it has also been implemented as an ontology in OWL. Its
main purpose is not to provide a model for a lexicon of a
language, but to cover a subset of linguistic description elements
that account for the linguistic realization of a domain ontology in
different natural languages. A complete description of the current
version of the LIR can be found in [
The lexical and terminological information captured in the LIR is
organized around the Lexical Entry class. Lexical Entry is
considered a union of word form (Lexicalization) and meaning
(Sense). This ground structure has been inspired by the Lexical
Markup Framework (LMF). The compliance with this standard is
important for two main reasons: (a) links to lexicons modeled
according to this standard can be established, and (b) the LIR can
be flexibly extended with modular extensions of the LMF (or
standard-compliant) modelling specific linguistic aspects, such as
deep morphology or syntax, not dealt by LIR in its present stage.
For more details on the interoperability of the LIR with further
standards see [
The rest of the classes that make up the LIR are Language, Definition, Source, Note and Usage Context (see Figure 5). These can be linked to the Lexicalization and Sense classes. Each Lexicalization is associated to one Sense. The Sense class represents the meaning of the ontology concept in a given language. It has been modelled as an empty class because its purpose is to guarantee interoperability with other standards. The meaning of the concept in a certain language (which may not completely overlap with the formal description of the concept in the ontology) is “materialized” in the Definition class, i.e., is expressed in natural language. The Usage Context gives us information about how a word behaves syntactically in a certain language by means of examples. Source information can be attached to any class in the model (Lexicalization, Definition, etc.), and, finally, the Note class has been meant to include any information about language specificities, connotations, style, register, etc., and can be related to any class. By determining the Language of a Lexical Entry, we can ask the system to display only the linguistic information associated to the ontology belonging to a given language. Thanks to this set of linguistic descriptions, the LIR is capable of managing lexicalizations within one language, and their translations to other languages. Relations of synonymy can be expressed among lexicalizations in the same language, and the preferred lexicalization can be determined (main Entry), as well as other term variant relations (such as Acronym, Multi Word Expression or Scientific Name). Finally, relations of translation equivalence can be established among lexicalizations in different languages.
However, as we stated previously, more often than not lexicalizations in different languages are not exact equivalents, because the senses they represent do not completely overlap in their intensional and/or extensional descriptions. In order to account for cultural and linguistic specificities of languages, we propose an extension of the LIR to allow declaring semantic relations among the senses (Sense) of lexicalizations within and across languages. The semantic relations identified with this purpose are: equivalence (isEquivalentTo), subsumption (subsumes or isSubsumedBy), or disjointness (isDisjointWith). So, the relation isRelatedTo that currently links senses (Sense) in the model is further specified.
The current version of the LIR is supported by the LabelTranslator system7, a plug-in of the NeOn Toolkit8.As soon as an ontology is imported in the NeOn Toolkit, the whole set of classes captured in the LIR is automatically associated to each Ontology Element, specifically, to ontology classes and properties, by means of the relation “has Lexical Entry”. In this way, the rest of linguistic classes organized around the Lexical Entry class are linked to an ontology element.
LabelTranslator [
As already mentioned, in the second version of hydrOntology, our purposes were to enrich the ontology in French and Catalan. With this aim, we imported the ontology originally documented in Spanish in the NeOn Toolkit, and automatically, all the linguistic classes of the LIR were associated to the concepts and properties in the ontology. The linguistic information associated to the URI of ontology concepts and properties in the original ontology automatically instantiated the LIR classes, i.e., a Lexical Entry was created for each ontology element, with its corresponding identifier (e.g., LexicalEntry-1), the Language of the label was identified and instantiated (e.g., Spanish), and a Lexicalization related to the Lexical Entry was also instantiated with the label in Spanish (e.g., Río). The rest of the linguistic information contained in the original ontology was not imported by the tool, and this fact was reported to the developers.
The next step was to manually introduce the labels in English,
already available in the Protégé version of hydrOntology,. Since
not all the concepts had been originally translated into English,
we decided to make use of the LabelTranslator system to
semiautomatically obtain translations for the original labels. The
process was carried out in a semi-automatic way, and the
translation candidates returned by LabelTranslator were evaluated
by a domain expert. Since the purpose of this paper is not to
evaluate the LabelTranslation plug-in, we will only refer to some
of the results by way of example. To obtain more information
about the experimental evaluation conducted with this tool, we
refer to [
9 The reason for not obtaining correct translations for some ontology terms may be due to the fact that the resources currently accessed by the system are quite general.
Then, the following step was the enrichment of the ontology with information in French and Catalan. Since these languages are not supported by LabelTranslator, we resorted to authoritative terminological resources in the domain 10 , and manually introduced the information in the LIR by means of the LIR API (see Figure 6). For the sake of comparison, we will illustrate the results by taking the concept River as example, as in the case of the Protégé version of hydrOntology.
As shown in Figure 6, seven Lexical Entries with Part of Speech noun were associated to the concept Río: three in Spanish, one in English, one in Catalan and two in French. By clicking on each Lexical Entry we are able to visualize the rest of linguistic information associated to it: Lexicalizations, Senses, Usage Contexts, Sources and Notes.
Figure 6. Linguistic Information associated to the concept Río in the LIR API (supported by LabelTranslator) The three Lexical Entries in Spanish (Río, Curso de agua principal, and Curso fluvial) are related by means of the hasSynonym relation (see Figure 7 for Lexical Entry Relationships). The differences in use depending on register (formal vs. informal) are explained in the Note class. With the new extension to the LIR that we propose in this paper, the Senses of these Lexical Entries could additionally be related by an equivalence relation (isEquivalentTo).
Then, the three Lexical Entries in Spanish are related to the Lexical Entry in English (River), the one in Catalan (Riu), and the last two in French (Rivière and Fleuve) by means of the hasTranslation relation (see Figure 7). The Lexical Entry in English and the Lexical Entries in Spanish are considered equivalents in meaning, and the same happens with the Catalan equivalent. Therefore, their senses could also be related by the equivalence relation (isEquivalentTo). 10 For instance, the Diccionari de l’Enciclopèdia Catalana for the Catalan language (http://www.enciclopedia.cat), and the Dictionnaire français d'hydrologie for the French language (http://www.cig.ensmp.fr/~hubert/glu/indexdic.htm) However, the two French Lexical Entries represent two more specific concepts that would stay in a relation of subsumption with the Spanish Río, the Catalan Riu, and the English River. This is an example of conceptual mismatch. The French understanding of river has a higher granularity level and identifies two concepts which are intensionally more specific, and extensionally do not share instances. These concepts are Rivière and Fleuve. According to the specialized resources accessed, Rivière is defined as a stream of water of considerable volume that flows into the sea or into another stream, and Fleuve is defined as a stream of water of considerable volume and length that flows into the sea. Therefore, in order to make explicit those differences in meaning, we relate them to two different Senses, and provide a definition in natural language for each of them (see Figure 6 for the Definition of Rivière in French). Then, with the new functionality of the LIR, we would establish a relation of subsumption between these two senses and the Spanish, English, and Catalan senses for Río, River, and Riu (isSubsumedBy). This further specification of the isRelatedTo relation among Senses allows accounting for categorization discrepancies among languages, which are not simply motivated by the fact that there are more lexicalizations in one language than in another, but by the different granularity levels that cultures make of the same world phenomenon. One could argue that these language specificities are only captured in the terminological layer of the ontology, but not in the conceptual model. However, this may suffice for certain ontology-based tasks such as information extraction or verbalization, whereas it may be insufficient for others. In that sense, a modification of the conceptualization to adapt the specificities of a certain language could be directly carried out by considering the lexical and terminological information contained in the LIR.
The aim of this paper has been twofold. On the one hand, we have discussed the difficulties involved in the interoperability of resources in the same domain created in different cultural settings, specifically because of the different granularity levels in which world phenomena are dealt with. We have described and illustrated the problematic issues of so called cultural dependent domains taking as example concepts of the hydrographical domain. On the other hand, our objective has been to compare two modalities for the representation of multilingual information in ontologies, with the aim of emphasizing the benefits of associating complex and sound lexical models to ontological knowledge. To achieve this we have presented in detail two versions of a multilingual ontology of the hydrographical domain, hydrOntology. The first version shows the representation possibilities offered by the OWL formalism to account for the multilingual information associated to ontology concepts in two languages: English and Spanish. The second version describes the representation possibilities of the Linguistic Information Repository (LIR), a proprietary model designed to associate lexical and terminological information in different languages to domain ontologies. Thanks to such a portable model, the lexical information can be structured for better exploitation purposes of ontology-based applications, and can account for linguistic and cultural discrepancies among languages.
This work is supported by the Spanish R&D Project Geobuddies (TSI2007-65677C02) and the European Project Monnet (FP7248458). We would also like to thank Óscar Corcho for valuable comments on a draft version of the paper. 9. REFERENCES [1] Aguado de Cea, G., Montiel-Ponsoda, E., Ramos, J. A. (2007) Multilingualidad en una aplicación basada en el conocimiento TIMM, Monográfico para la revista SEPLN [2] Arpírez JC, Gómez-Pérez A, Lozano A, Pinto HS (ONTO)2Agent: An ontology-based WWW broker to select ontologies. In: Gómez-Pérez A, Benjamins RV (eds) ECAI’98 Workshop on Applications of Ontologies and Problem-Solving Methods. Brighton, (UK), 1998, pp 16–24. [3] Barrasa, J. Modelo para la definición automática de correspondencias semánticas entre ontologías y modelos relacionales. PhD Thesis, UPM, Madrid, Spain. 2007.