Many language resources are nowadays available in machine readable formats, but still contained in isolated silos. Current Semantic Web-based techniques enable the transformation and linking of those resources to become a navigable graph of linked language resources, which can be directly consumed by third-party applications. The prototype we have developed builds on a web user interface and SPARQL endpoint initially developed to query a single terminological database (Terminesp), now extended to navigate a set of multilingual terminologies. The vocabulary used to represent these terminologies into the linked data format is lemon-ontolex, a de facto standard for representing lexical information relative to ontologies and for linking lexicons and machine-readable dictionaries to the Semantic Web.
The Linguistic Linked Open Data (LLOD) cloud1 is a sub-cloud of linguistic resources provided in an interoperable way (using the Resource Description Framework or RDF data model), freely accessible and linked with each other. In its current state, the LLOD Cloud contains monolingual and multilingual dictionaries, lexicons, thesauri and even corpora. English is the best represented language, and some languages are underrepresented or not present at all.
With Terminesp (a multilingual terminological
database created by the Spanish Association for
Terminology, AETER), we aimed at validating the
lemon-ontolex model as a representation scheme for
1 http://linguistic-lod.org/
lexical resources, specifically, the so-called
vartrans module, a dedicated module that
accounts for terminological variation and translation
relations among entries
In this paper, we focus on the design decisions taken in the transformation and linking steps, and on the impact they have in the search and navigation of the resulting linked terminological data.
In Section 2, we introduce lemon-ontolex and the vartrans module. In section 3, we describe the design decisions taken in the transformation process. In section 4, we refer to the benefits of browsing and navigating linked multilingual terminologies. 2.
lemon-ontolex
The lemon-ontolex model is the resulting work of
the efforts made by the W3C Ontology Lexica
Community Group since 2011 to build a rich model
to represent the lexicon-ontology interface. It is
largely based on the lemon model
Migration and linking of the resources
For the transformation of TERMCAT terminology
repertoires to the LD format and linking to
Terminesp we followed these steps: data exploration,
URI naming strategy, data modeling, RDF
generation and linking
Data exploration. TERMCAT terminology repertoires are divided by domain. Each database consists of a list of entries in Catalan and their translations into Spanish, English, French, etc., along with the term type (full form or abbreviation), references to associated terms, synonyms, and, sometimes, definitions. Data for part-of-speech, gender and number in nouns, and subcategorization of verbs, is also available.
URI naming strategy. Inspired by the work in the Apertium dictionaries4, the term itself, its part of speech and the language of the term are part of the URI of the lexical entry. For lexical senses, the domain is included in the URI.
Modeling. For the modeling process, we regard each term in a set of translations as a specific sense of a lexical entry, a sense that is mapped to a concept in a particular domain. This allows us to have a unique lexical entry red (network), for instance, which occurs both in the lexicon Internet i societat de la informació as in the lexicon Indústria electrònica i dels materials elèctrics, with different senses that we extract from each domain lexicon. This results in a number of RDF lexica that matches the number of languages available in TERMCAT data, and each lexical entry will have a different number of senses depending on its use across domains. In this way, the lexical entry :red-n-es will be mapped to a sense :red-n-es-Internet-sense, as well as to a :red-n-es-Industria-sense, etc. Each of these senses refers to a skos:Concept with a particular definition and domain. Regarding translations, the vartrans module represents them as relations across lexical senses of the entries of each lexicon. Parts of speech, subcategorization, gender and number are accounted for as well.
Generation and linking. For the transformation we used the data cleaning and transformation tool OpenRefine5 with its extension for LD. We linked to lexinfo6 to cover morphosyntactic information, and to Terminesp at the lexical entry level. Linking to DBpedia is also planned as a next step. 4.
Browsing multilingual terminologies
We reuse the Terminesp web user interface (see Fig. 2.) and SPARQL endpoint to browse and query this set of integrated terminologies7. Benefits are related to easy access and reuse of linguistic data by end users (translators, terminologists) and semanticaware software agents.
Acknowledgements. This work is supported by the FP7 EU project LIDER (610782), and the Spanish 4V project (TIN2013-46238-C4-2-R). 5 http://openrefine.org/index.html 6 http://lexinfo.net/ 7 http://linguistic.linkeddata.es/terminesp/