Overview of the process for publishing AGROVOC as linked data 

Context in source publication

Context 1

... separate information on the revision of concepts as well as of each label in each language. Thus we are able to state that “ 玉 米 ” has a code, originating from the past pre-RDF versions of AGROVOC, equal to “12332” . This is expressed by the triple 6 : As a further example, we are also able to state that that term was created on December 12, 2002: The Agrontology seen in the above example is a compendium to AGROVOC, providing domain-spe- cific properties for enriching the description of concepts. Agrontology is enriched with VOAF 8 (Vocabulary of a Friend) descriptors, mostly for linking it to AGROVOC (and to other Datasets adopting it, such as the FAO Biotech Glossary 9 ), and to have it men- tioned in the LOV Dataset. Currently, AGROVOC is undergoing a deep analysis in order to make very explicit the modelling style adopted in its various topic areas (see Figure 2). A parallel analysis of the agron- tology is ongoing, with the purpose of reaching a full harmony between domain modelling and the vocabulary currently used for it. All 32,000+ concepts of the AGROVOC thesaurus are hierarchically organized under its 25 top concepts. AGROVOC top concepts are very general and high level concepts , including concepts such as “activities”, “organisms”, “locations”, “products”, “organism”, etc. The fact that 20,000+ concepts fall under the top concept “organism”, confirms how AGROVOC is largely oriented towards the agriculture sector (see Figure 2 for a complete statistic distribution of the dataset concepts under its 25 top concepts). Other important areas of AGROVOC include “substances”, “entities”, “products” and “locations”. Beyond being listed in the AGROVOC website, the list of top concepts can be found in the VoID file for AGROVOC Linked Open Dataset. Moreover, a deep study on the current coverage of AGROVOC is under study, with the purpose of supporting human and machine users alike in their quest for information within the thesaurus and its links. AGROVOC is today published as an Open Linked Dataset with links to thirteen vocabularies, thesauri and ontologies. Five of the linked resources are general in scope: the Library of Congress Subject Head- ings (LCSH) 10 , , RAMEAU Répertoire d'autorité- matière encyclopedique et alphabetique unifie, Eu- rovoc 12 , DBpedia 13 , and an experimental Linked Data version of the Dewey Decimal Classification 14 . The remaining eight resources are specific to various domains: NAL Thesaurus 15 for agriculture, GEMET 16 for environment, STW for Economics, TheSoz 17 for so- cial science and both GeoNames and the FAO Geopolitical Ontology cover countries and political regions. ASFA 18 covers aquatic science and the aptly named Biotechnology glossary covers biotechnology. These linked resources are mostly available as RDF/SKOS resources. The linked resources were considered in their entirety barring RAMEAU, for which only agriculture related concepts were considered (amounting to some 10% of its 150 000 concepts). Candidate mappings were found by applying string similarity matching al- gorithms to pairs of preferred labels [6] and by using the Ontology Alignment API [7] for managing the produced matches. The common analysis language used was English in all cases except the AGROVOC RAMEAU alignment for which French was used. Table 1 shows, for each resource linked to AGROVOC (column 1), its area of coverage (column 2), the language considered for mapping with AGROVOC (column 3), and the number of matches resulting from the evaluation (column 4, see below). Candidate links were presented to a domain expert for evaluation in the form of a spreadsheet. Once validated, the mappings were loaded in the same triple store where the linked data version of AGROVOC is stored. All resulting validated candidate matches were considered to be skos:exactMatch, as in agv:c_12332 skos:exactMatch ("maize" in English). Exact match has the advantage of a looser notion of equivalence than the more formal equivalence found in owl equivalence/identity properties such as owl:sameAs . The objective when linking AGROVOC to other resources was to provide only main anchors, privileging accuracy over recall. This is why we (mostly) rejected skos:closeMatch and relied exclusively on skos:exactMatch , found by means of string-similarity techniques as opposed to more sophisticated context-based approaches. Also, the One Sense per Domain assumption which, in analogy to the “one sense per discourse” [8] assumption, specifies that “t he more specialized a domain is, the less is the influen ce of word sense am- biguity ”, supports our claim that (in our case) similar strings correspond to equivalent meanings. The use of more sophisticated approaches might have contributed to filtering out potential results more than widening their number (thus incrementing precision over recall), however this potential loss of precision was well com- pensated by the manual validation of candidate links by a domain expert. Figure 3 (next page) provides a high-level view of the entire AGROVOC ’s maintenance process and its publication as linked data. The figure emphasises the three levels of data maintenance (bottom layer), data storage (middle layer), and data publication (top layer). The relational database is still necessary as many existing applications interface with this legacy model via SQL. Such conversions are thus needed to syn- chronize the data accessed by editors using legacy tools. This duplication of data repositories and the consequent data conversions is obviously not ideal and in principle should be limited. On the other hand, AGROVOC has supported a worldwide community of users (people and institutions) for decades who have developed a number of applications relying on the legacy relational model: these conversion steps are thus currently unavoidable and give an idea of the com- plexity inherent to historic, distributed collaboration scenarios. Elaborate procedures are rendered necessary, and the conversion effort, modelling issues and information needs are just the tip of the iceberg com- pared to the real effort spent in content and services maintenance. Several conversion steps are then present in the AGROVOC lifecycle. Note that this data flow is not always monotonic. Although the main authoring tool is VocBench, contributions to AGROVOC may also occasionally come from legacy formats such as spreadsheets and SQL files. This updated content is thus contributed separately (through different modali- ties) and then merged to produce a new copy. When a VocBench version is finalized with contributions coming from different sources and formats, it is then converted back to the relational DB for legacy applications. At the same time, a SKOS-XL version is produced and enriched with information, such as metadata descriptors from the VoID vocabulary to feed the LOD endpoint with updated data. Currently, no versioning info for the dataset as a whole (i.e. which AGROVOC release a client is accessing), is ex- plicitly reported inside its triples, while editorial notes provide fine-grained details about its content, with creation and modification dates for all concepts and labels present in the dataset. The linked data version of AGROVOC is now available online thanks to a collaboration between FAO and MIMOS Berhad. Data is stored in an RDF triple store (Allegrograph 19 ) hosted on a server in Kuala Lumpur, Malaysia. A SPARQL endpoint, combined with http resolution of AGROVOC entities, allows for publication as linked data. The Pubby service men- tioned in section 2 is also hosted by MIMOS. Both RDF and HTML access are resolved through content negotiation on FAO servers and redirected to the proper MIMOS service. During the more than 30 years of its existence, AGROVOC has seen a growing community of users exploiting its content for a progressively wider set of uses. In this section we report the more important uses of which we are aware. In some cases they exploit AGROVOC to give further explicit contribution to the LOD cloud itself, while in others, the availability of AGROVOC data as LOD will foster wider access pos- sibilities and probably see an increase in use as the number of potential users augment due to their inter- action with the entirety of the LOD cloud and to AGROVOC ’s position within it. In 2011, following a wave of enthusiasm caused by Linked Open Data initiatives and benefiting from the successful experiences of the AIMS group working on AGROVOC and other concept schemes and vocabularies ported to the Semantic Web, FAO ’s Information Technology Division (CIO) chose to add a taste of Semantic Web to their ambitious data integration project data.fao.org. The project which launches publicly in December 2012, bring s much of FAO’s statistical, textual and geographical data under one roof, fostering data integration and harmonization first within FAO itself, and later publicly via LOD. The models which are being exploited are many, mainly covering domain representation (OWL and SKOS as core modelling vocabularies), flanked by “standard” vocabularies such as FOAF [9]) and statistical data reporting (Data Cube Vocabulary [10]). AGROVOC, the first FAO resource to embrace the Semantic Web and to be published on the LOD cloud, was chosen as a common, controlled vocabulary for tagging the information resources (documents, media etc..) in data.fao.org. AGROVOC will also act as an interlingua to easily match RDF resources from different datasets, which still maintain a certain independ- ence and which thus expose potential overlaps with other datasets. A new, potentially wider, set of linksets on a star configuration with AGROVOC in the centre will be elaborated for establishing a global intercon- nected network of resources within FAO. AGROVOC and other vocabularies hosted on VocBench (e.g. Journal Authority Descriptions) have for some years been supported by an extensive set of SOAP web services 20 that ...