International organizations like FAO are intrinsically multilingual. FAO is currently experimenting with semantic-oriented technologies based on ontologies, with the purpose of integrating data across various information systems and providing better services to end users. However, in order for these technologies to be used in real-life scenarios, models and tools for accommodating and managing multilingual data are needed. This paper analyzes the requirements for the treatment of multilinguality as resulting from the experience we gained at FAO.
Currently, a number of pilot studies are being carried on within FAO to improve the level of interoperability among FAO information systems (IS) by means of semantically oriented technologies. In order for these pilot studies to be successful the multilingual issues need to be carefully addressed: in terms of models, software, support to editing and interfaces to editors.
In the rest of this paper we illustrate in some detail the multilingual resources available at FAO, and the future direction currently under study (Section 2). In Section 3 we provide our requirements for the treatment of multilinguality, and in Section 4 we discuss them.
The reference tables used to store and access statistical data are an example of
multilingual resources where the one-to-one correspondence between languages is
given. For example, the reference tables on fisheries (currently managed by a
Reference Table Management System (RTMS) [
AGROVOC [
Semantically oriented technologies promise to improve the level of interoperability
between IS. Therefore, FAO is experimenting with the possibility of using
interconnected and networked ontologies [
These relationships include synonymy (e.g. ‘vacca’ and ‘mucca’ are
synonyms and both translate the Chinese ‘ ’), acronyms (‘BSE’ stands
for the English ‘Bovine spongiform encephalopathy’, but ‘ESB’ stands for
the French name of the same disease, ‘Encéphalopathie spongiforme
bovine’), spelling variances, different types of names (e.g. ‘Gabonese
Republic’ is the official long name, and ‘Gabon’ is the official short name of
the same country);
4. impose constraints at the lexical level, such as the number of allowed
translations and acronyms;
5. add additional information at the lexical level, such as comments, definitions,
and images, in order to better specify the intended meaning of a term;
6. account for the use of the same terms in different context, that are translated
by different terms in other languages (e.g. ‘benzene’ is considered a pollutant
in a specific area, and a fuel in another sphere [
Requirement 1 and (partially) Requirement 2 can be met by using the xml:lang attribute that allows one to specify the language in which a given element is expressed. However, the simple use of this attribute may not be enough to satisfy Requirements 3, 4 and 5. For instance we are not able to set relationships between terms in different languages that may be used to represent the same object, and we may not be able to assign specific properties to those objects. Requirements 6 and 7 are both related to the concerns expressed in Requirements 1 and 3, i.e. that the linguistic (terminological) level is distinguished by the “factual” level. They also somewhat imply that languages have specificities and that ontologies (and thesauri, for the matter) built in one language and then translated can lead to counterintuitive results.
Requirement 8 is relevant both to the exploitation of large multilingual ontologies and to their maintenance. In fact, for efficiency reasons it may be very useful to be able to select and extract only one or few languages instead of manipulating the entire data set. Also, during the editorial process it may be very useful to be able to add a new language layer for the entire ontology (as opposed to adding a new element, say a property ‘title’, individually for each element).
Requirements 9 and 10 address the lowest level features of a multilingual ontology, namely the possibility of properly encoding and visualizing a wider variety of languages than just the western ones.
Most of our requirements can be addressed by ad hoc modeling solutions (e.g. representing the lexical items as specific entities, classes or instances, characterized by an xml:lang property, and using object-type relationships to link terms). These ad hoc models can then be accessed and exploited by software components that are bound to be hardly reusable and non-generalizable. We raise the concern that some more general treatments of multilinguality should be provided: be that at the level of ontology language, modeling, or applications for editing and exploiting ontologies. Finally, one issue to take into consideration for the implementation of the requirements at which we arrived is the existence of massive legacy data (e.g. relational and terminological databases in various formats) that will continue to be used and to which new information systems will need to refer or connect.