We shall illustrate and evaluate the LOD by issuing sample SPARQL queries illuminating aspects of it.

1. Simple queries

Query for author. Authors are instances of foaf:Person (except in the rare institutional case, in which they would be foaf:Agent). The SPARQL query in Table 9 answers the question of which persons have been the most prolific authors in the harvested journals.

Query for a scientific name. Biological Latin names are stored in the system as :ScientificName and are mentioned by taxonomic name usages. Table 10 orders scientific names of any rank by the number of unique mentions that they have in articles. It is possible to narrow down the solution to binomial names (species names) by adding the dwc:specificEpithet and dwc:genus properties as shown in Table 11. It is also possible, for example, to determine the most-mentioned scientific name by the number of articles it is mentioned in (Table 12).

Query the article structure. A unique feature of OpenBiodiv-LOD is that articles are broken down to their components (see Table 3) and taxonomic name usages are connected to the specific part of the article and not just to the article in general. Combining this feature with queries from the previous paragraph, we can, for example, look for the most mentioned scientific name in a figure (Table 13) or for the figures present in a particular article (Table 14).

Query for taxonomic concepts. We can create a query uniting information from the GBIF Backbone Taxonomy with semantics coming from the article structure. The query in Table 15 locates taxa that are in the beetle family Curculionidae according to the taxonomic backbone of GBIF and looks for new taxa (:TaxonomicDiscovery) that have been associated with one of its genera.

Fuzzy Queries via Lucene. The SPARQL endpoint of OpenBiodiv-LOD supports fuzzy matching via a Lucene connector (The Apache Software Foundation 2013). This can be a very useful as, due to multiplicity of taxonomic names and the complexities of Latin grammar, one often does not remember the correct spelling of a name. The Lucene query needs to follow the standard Lucene query syntax (The Apache Software Foundation 2013) and is specified as a literal string of the property <http://www.ontotext.com/connectors/lucene#query> of the search variable (Table 16).

Competency question answering via SPARQL

Validity of a name. Of central importance to biological nomenclature is the question of whether a given taxonomic name is valid or not. We shall consider a taxonomic name invalid if and only if at least one of the following invalidation criteria holds:

1. The name has been replaced: i.e. there is a :replacementName property originating in the name and there are no loops (it is impossible to follow the :replacementName edges and come back to the name). This query is illustrated in Table 17.

2. The name has been invalidated: i.e. there is a taxonomic usage of the name with the status :UnavailableName and there is no newer taxonomic name usage revalidating it (:AvailableName). Illustrated in Table 18.

The case of Museu Nacional de Rio de Janeiro (MNRJ). In order to illustrate the capabilities of OpenBiodiv and draw attention to the scientific impact of the tragically lost collection in the fire of the Museu Nacional de Rio de Janeiro (MNRJ), we can ask our system to give us the number of times a specimen from that collection was used in a taxonomic article and in which ones (Table 19). We use MNRJ’s GRSCICOLL (GBIF Registry of Scientific Collections) identifier in addition to its name and collection code.

It turns out that MNRJ has been mentioned 362,062 times in our system in a total of 509 articles. Perhaps more interestingly, we can see specimens of which taxa may have been lost, have declining populations or are threatened by extinction. Examples include the insects (Xestoblatta, Charinus, Lamproclasiopa etc.) which are extinct, Keays's Rice Rats (Nephelomys keaysi) which have declining populations and many others for a total of 1,348 distinct names mentioned in taxonomic articles which reference MNRJ.

Specimen collection. Some of the most important information about biodiversity in an article is within the Materials Examined section. It contains information about the collection of biodiversity samples (specimens), the location where they were found, the taxonomists who identified them, their habitats, the institutions where the specimens are kept and much more. For example, we can query all people who have collected specimens belonging to the insect genus Zelus (Table 20).

Institutional impact. We can use a SPARQL query to understand how collections from different instutions are used to describe taxa. In the example query, we have linked institutional identifiers with treatments which mention them to find out institutional impact per family (Table 21).

Links between holotype descriptions (literature), institutions holding the holotypes and genomics. OpenBiodiv integrates information about examined specimens and taxonomic descriptions from literature with external identifiers of institutions holding the specimens, as well as records identifiers pertaining to the genomic sequences of the specimens. We can retrieve this information with the SPARQL query in Table 22.

Fulfilment of the Principles of Linked Open Data

Linked Open Data (Heath and Bizer 2011) is an idea of the Semantic Web (Berners-Lee et al. 2001) aimed at ensuring that data published on the Web are reusable, discoverable and, most importantly, that pieces of data published by different entities can work together. The principles of LOD are the following (Heath and Bizer 2011):

1. Use URIs as names for things.

2. Use HTTP URIs so people can look up these things.

3. When someone looks up a URI, provide useful information, using the standards (RDF, SPARQL).

4. Include links to other URIs so they can discover more things.

We have followed these guidelines when creating the OpenBiodiv-LOD. We will now discuss each of these points separately.

Usage of URIs as resource identifiers. Every instance in OpenBiodiv-LOD is uniquely identifiable by a HTTP URI of the following form: http://openbiodiv.net/uuid-(suffix). All instance identifiers in OpenBiodiv LOD follow this schema. The optional suffix field is assigned only to resources extracted from GBIF.

Identifiers for resources from Pensoft and Plazi. During the RDF-isation of the sources Pensoft and Plazi, when a new concept is discovered (e.g. a person, a scientific name etc.), a UUID is generated. Then the resource is always referred to in the database by this UUID in the OpenBiodiv namespace, http://openbiodiv.net/. Pensoft and Plazi furthermore share the UUID part of the identifier in the semi-structured representation of treatments. For example, Lyubomir Penev is a resource identified by http://openbiodiv.net/416FDF84-1029-4115-B43F-E9E734004489.

Identifiers for GBIF taxonomic concepts. GBIF offers its taxonomic backbone as a DarwinCore (Wieczorek et al. 2012) tab separated file (TSV). Each row in the TSV corresponds to a taxonomic concept published by GBIF. GBIF does not offer a globally-unique ID of its concepts, but only a local ID (e.g. 4239 is the GBIF ID for their concept for weevils, Curculionidae sec (GBIF Secretariat 2019). This is why we generated a UUID (here: F1DD0CF0-217D-422B-BAA4-58901976D7B4-4239) for each row of the data table published from GBIF. Each taxonomic concept is linked to a taxonomic name and to a taxonomic concept label (name + “sec." + reference to the literature). It was impractical for programmatic reasons to generate a new UUID for these linked entities. This is why their unique identifiers are suffixed. We use the suffix -scName to denote scientific names and -label to denote taxonomic concept labels (e.g. http://openbiodiv.net/F1DD0CF0-217D-422B-BAA4-58901976D7B4-4239-label)

Usage of HTTP URIs and dereferencing. As per the Linked Data principles, we use dereferenceable HTTP URIs for our resources. For example, if a web browser opens http://openbiodiv.net/416FDF84-1029-4115-B43F-E9E734004489, a webpage is displayed (Fig. 4) providing useful information for the person Lyubomir Penev, such as his affiliations, research output and co-authors. In addition, it is possible to request OpenBiodiv resources via Curl with the header Content-Type application/rdf+xml and an RDF representation of the resources is returned.

a

b

Figure 4.

Visualisation of a semantic resource via a template on the OpenBiodiv website. The figure shows information related to a Person resource which is displayed when the resource is resolved.

Linking to other resources. All resources in OpenBiodiv form a graph (there are no disconnected parts), following a data model discussed in the next subsection. Second, resources are linked to external databases via properties like datacite:hasIdentifier and openbiodiv:mentionsIdentifier. These identifiers can be: GBIF IDs, ZooBank IDs, Zenodo IDs, ORCIDs, BOLD BINs, BOLD Records or GenBank accession numbers. We have created links between people and their ORCID records, publications and their GBIF dataset records, as well as Zoobank records and genomic records within BOLD and GenBank. See Table 21 for a SPARQL query which demonstrates the linking of taxonomic name data with external institutional identifiers from the GBIF Registry of Scientific Collections (GRSciColl) and Table 22 for a query which retrieves taxonomic circumscriptions of all holotypes, which have genomic identifier(s) and associated institutional records, which signify where the examined materials are stored.

Data Model

When creating the RDF graph, we have conformed to the OpenBiodiv-O ontology described in Senderov et al. 2018, Dimitrova et al. 2019b and well-established community ontologies (Fig. 5). In particular, (1) we use the Semantic Publishing and Referencing Ontologies (SPAR, (Peroni and Shotton 2018) to model entities from published documents such as Journal, Article, Section, Figure, Table and so on; and (2) we use the DarwinCore (DwC) (Wieczorek et al. 2012) community standard to model entities in the biodiversity domain.

Figure 5.  

OpenBiodiv-O is an ontology that links the publishing domain with the biodiversity domain. Major resource types covered by each of the ontology families are given in the box below the Venn diagram. Each of them is present in the OpenBiodiv-O ontology as a class. Important resources from the publishing domain are listed in the left-most column and from biodiversity informatics in the right-most column. The middle one covers important OpenBiodiv-O resources.

SPAR provides facilities to deal with the dichotomy between the abstract representation of knowledge through the class Work and its concrete representation through the class Expression. For example, a fabio:JournalArticle can be the realisation of a fabio:ResearchPaper. On the other hand, the DwC community standard gives a standard way to express properties from taxonomy and biodiversity science.

In the most recent verision of OpenBiodiv-LOD (Dimitrova et al. 2019b), we have represented occurrences, taxonomic identifications, locations and events via the respective DwC classes dwc:Occurrence, dwc:Identification, dcterms:Location and dwc:Event, as well as various other DwC properties which model information related to these four resource types (e.g. coordinates, type statuses, life stages). We have also integrated classes and properties from the DataCite ontology, part of the SPAR ontologies (Peroni and Shotton 2018), to model external resources via their identifiers. In particular, we have implemented the class datacite:ResourceIdentifier to model Zoobank, Zenodo, BOLD and GenBank identifiers and datacite:PersonalIdentifier to model ORCIDs. The property datacite:hasIdentifier has also been implemented to help establish links between existing resources and their external identifiers.

Performance

The current iteration of the database holds over 360 Mln triples. The expansion ratio under the RDFS-Plus (Optimised) ruleset is 1.20, i.e. for each asserted statement, we materialise, on average, 1.20 implicit statements. In a previous release of the dataset, which was under the OWL2-RL rule-set, the most complex rule-set supported by GraphDB, the expansion ratio was about 3.7; however, we encountered significant performance issues using it. Even with the lighter rule-set (RDFS-Plus Optimised) (Ontotext 2021), we still see performance degradation with increasing database size.

We observed a steady increase of implicit (inferred) statements during the upload of new triples. An example of such an inferred statement is :A po:contains :B, generated from the statement :B po:isContainedBy :A because po:isContainedBy is an inverse property of po:contains. Upon closer inspection, it turned out that the import of external ontologies, in addition to OpenBiodiv-O, leads to the generation of superfluous inferred statements. For instance, in the SKOS ontology, the property skos:exactMatch is transitive and is also a sub-property of skos:closeMatch. The same ontology defines that skos:closeMatch is a sub-property of skos:mappingRelation. Therefore, after importing the SKOS ontology, GraphDB infers that all treatments which have the property skos:exactMatch (these are only Plazi treatments for which we have information about their Plazi treatment id, for example, openbiodiv:03894A65-5824-FFE9-571B-B65D2F47F95E skos:exactMatch plazi_treatment:03894A65-5824-FFE9-571B-B65D2F47F95E), also have an additional statement with property skos:mappingRelation. This inferred statement does not actually bring any new semantic information to the knowledge graph, hence we consider it superfluous.

We came to the conclusion that all necessary RDF logic is stored in OpenBiodiv-O and does not require the import of other ontologies, since OpenBiodiv-O already includes the essential relations from these ontologies. Therefore, in the latest release of the repository, we have only imported the OpenBiodiv-O ontology.

Another important aspect of performance is the RDF-isation time or the time it takes to convert a single XML into RDF in trig serialisation and to upload it to the database. Our observations show that the most time-consuming part of this process are the MongoDB requests used to get and set resource identifiers. Even though they provided an improvement to the previous model, which used queries to GraphDB to obtain and set identifiers, MongoDB requests can add up to a significant amount of time per XML document. We noted that adding a MongoDB index and using it to search for text content does not improve the speed at all. As an alternative solution, we now use sha256 hashing to compact value strings associated with identifiers to a fixed-length hash string. This method is explained in detail in the Methods section.

The generated dataset OpenBiodiv-LOD, similar to the expanded ontology OpenBiodiv-O, is already a solid resource for biologists, as it includes information from most articles published by Pensoft and Plazi and counts over 360 million RDF triples.

An important conclusion that can be drawn from the work is that it is possible to use a semantic graph for the integration of a large volume of data on biodiversity. We were unexpectedly given the opportunity to illustrate the power of the knowledge graph by analysing the damage from the tragic fire at the Museu Nacional in Rio de Janeiro. In addition, we have illustrated that it is possible to write relatively simple logical conclusions to check the validity of a taxonomic name.

Due to the large amount of data, we found that, although the use of a semantic graph was possible, some of the initially-chosen technologies proved to be inapplicable or difficult to apply. We have observed that the practical application of the full logical OWL model is difficult due to performance problems. Instead in the end, we utilised RDFs which are less powerful, but faster. In addition, we found that triple stores are not a universal solution to all data integration problems, but can be used in combination with other database technologies (e.g. MongoDB) to efficiently store and query semantic resources.

A great difficulty was the disambiguation of resources, such as author names or taxonomic names. In the functional design of the RDF4R package, we have input modules that allow us to insert a list of functions/rules for disambiguation when searching for an identifier for a given resource. However, we had only limited success with the rule-based disambiguation and, for this reason in the production system, it was discontinued for the moment.

Considering these and other "lessons", the future development of the OpenBiodiv-LOD project can be outlined in the following way:

1. As an immediate goal, to optimise the workflow for processing XMLs. This would be achieved by distributing the work across multiple workers operating on different machines.
2. Improve the search functionalities of the OpenBiodiv website to enable more user-friendly querying of the knowledge graph without requiring any undertanding of SPARQL from the users. We have identified user groups (biologists, taxonomists, institutions etc.) and respective competency questions which the website would aim to answer via different "apps". These apps will provide an interface in which users will fill in text fields and use a simplified faceted search functionality to answer different biodiversity questions and narrow down their answers. We could potentially augment the search functionality of the OpenBiodiv website by implementing additional semantic entities, extracted from articles via the Pensoft Annotator (Dimitrova et al. 2020), a text-to-ontology mapping tool developed by Pensoft.
3. Integrate OpenBiodiv with a nanopublication publishing service. Nanopublications (Groth et al. 2010) are a type of publication aiming to store and attribute even the smallest bit of information which can be expressed as a semantic triple. They follow certain guidelines (Gray et al. 2020) defining their constituent graphs (Assertion, Provenance, Publication Info) and the links between them. OpenBiodiv is a potential source for nanopublications since it already contains semantic statements about biodiversity (assertions), as well as provenance information, encapsulated in the publication metadata. As such, it can be used to generate nanopublications which could become part of the existing decentralised network of nanopublications (Kuhn et al. 2016). Existing nanopublication infrastructure and publishing formats can also help to establish a system for commenting, supporting and contradicting biodiversity statements extracted from literature.

We envision OpenBiodiv-LOD as an integral part of the existing semantic network of biodiversity knowledge, based on HTTP identifiers and controlled vocabularies. By semantically enhancing and linking knowlege in OpenBiodiv to existing machine-readable data, we augmment biodiversity data quality and increase the potential for its reuse.