Scholarly citations from one publication to another, expressed as reference lists within academic articles, are core elements of scholarly communication. Unfortunately, they usually can be accessed en masse only by paying significant subscription fees to commercial organizations, while those few services that do made them available for free impose strict limitations on their reuse. In this paper we provide an overview of the OpenCitations Project (http://opencitations.net) undertaken to remedy this situation, and of its main product, the OpenCitations Corpus, which is an open repository of accurate bibliographic citation data harvested from the scholarly literature, made available in RDF under a Creative Commons public domain dedication. RASH version: https://w3id.org/oc/paper/occ-lisc2016.html
Databases of citation data are among the most attractive and used artefacts
in the Scholarly Communication domain. They are one of the main tools
used by researchers for gaining knowledge about a particular topic, and by
scientists in Bibliometrics, Informetrics, and Scientometrics for analysing
the complex relationships that exist within huge networks of citations of
scholarly works. They also serve institutional goals, since they provide one
of the main mechanisms for assessing the quality of research by means
of (sometimes questionable) metrics and indicators calculated from such
citation databases. While some of these resources, e.g. Microsoft Academic
Graph3 and Google Scholar4, are freely accessible (but not downloadable),
those considered the most authoritative by institutions worldwide, namely
Scopus5 and Web of Science6, can be accessed only by paying significant
access fees, which may amount to tens of thousands of pounds annually [
Reference lists within academic articles are core elements of scholarly
communication, since they both permit the attribution of credit and
integrate our independent research endeavours. But the cruel reality is that
these key data are not freely available. In the current age where Open
Access is considered a necessary practice in research, it is a scandal that
reference lists from scholarly publications (conference papers, books,
journal articles, etc.) are not readily and freely available for use by all scholars.
As we have already stated in a previous work [
Citation data now needs to be recognized as a part of the Commons – those works that are freely and legally available for sharing – and placed in an open repository, where they should be stored in appropriate machine-readable formats so as to be easily reused by machines to assist people in producing novel services.
This is the main premise behind the OpenCitations Project [
The rest of the paper is organised as follows. In Section 2 we recall the story of the OpenCitations Project since its beginning in 2010. In Section 3 we describe the revised metadata specification and software tools that have been recently developed within the OpenCitations Project for the creation of a new and improved instantiation of the OCC. In Section 4 we briefly describe other open (and RDF-based) repositories of scholarly document metadata. Finally, in Section 5, we sketch out our future plans. 2
The OpenCitations Project formally started in 2010 as a one-year project
funded by JISC9 (subsequently extended for an additional half year), with
David Shotton as director, who at that time was working in the
Department of Zoology at the University of Oxford. The project’s goal was global
in scope, and was designed to change the face of scientific publishing and
scholarly communication, since it aimed to publish bibliographic citation
information in RDF and to make citation links as easy to traverse as Web
7 https://creativecommons.org/publicdomain/zero/1.0/legalcode
8 http://europepmc.org/
9 http://www.jisc.ac.uk/whatwedo/programmes/inf11/jiscexpo/
jiscopencitation.aspx
links. The main deliverable of the project, among several outcomes10, was
the release of an open repository of scholarly citation data described using
the SPAR (Semantic Publishing and Referencing) Ontologies11 [
In May 2014, OpenCitations was adopted by the Infrastructure Services for Open Access (IS4OA)13 as one of its academic Open Access services. IS4OA is UK-based not-for-profit charitable company that aims to provide benefit to the global community of research information users. It acts as an umbrella organisation that supports openly accessible information and discovery services relating to academic information, research results and scholarly publications, by providing business structure and expertise and a means of channelling financial support to these services.
At the end of 2015, Silvio Peroni joined the OpenCitations Project as co-director, with the aim of setting up a new instantiation of the Corpus based on a new metadata schema and employing several new technologies to automate the ingestion of fresh citation metadata from authoritative sources. The current instantiation of the OCC is hosted by the Department of Computer Science and Engineering (DISI) at the University of Bologna, and since the beginning of July 2016 it has been ingesting, processing and publishing reference lists of scholarly papers available in Europe PubMed Central, as described in the following section. 3
The OpenCitations Project (http://opencitations.net) has recently
created a new instantiation of its open citations database, with an integrated
SPARQL endpoint and a browsing interface to support data consumers.
This database, the OpenCitations Corpus (OCC), is an open repository of
scholarly citation data made available under a Creative Commons public
domain dedication (CC0), which provides accurate bibliographic references
harvested from the scholarly literature, described using the SPAR
Ontologies [
The newly revised metadata model used for the data stored in the OCC,
available at [
For convenience, all the terms from the aforementioned ontologies are collected within a new ontology called the OpenCitations Ontology (OCO)20. This is not yet another bibliographic ontology, but rather just a mechanism for grouping existing complementary ontological entities from several other ontologies for the purpose of providing descriptive metadata for the OCC. 3.2
The OCC stores metadata relevant to bibliographic citations in RDF,
encoded as JSON-LD21. In early September 2016, all the ingested data will
be also available as downloadable datasets. In the meantime, two
exemplar dataset, compliant with the OCC metadata model introduced in
Section 3.1, have been made available: the first from article metadata provided
by Springer Nature (available at [
The following six bibliographic entity types occur in the OCC, as well as in the aforementioned exemplar datasets: 14 http://purl.org/spar/fabio 15 http://purl.org/spar/pro 16 http://purl.org/spar/biro 17 http://purl.org/spar/c4o 18 http://purl.org/spar/datacite 19 http://xmlns.com/foaf/spec/ 20 https://w3id.org/oc/ontology 21 http://json-ld.org/ 22 All the resources in the exemplar datasets have URLs that starts with “http: //localhost:8000/corpus/” and do not refer to any existing IRI included in the current version of the corpus.
– bibliographic resources (br), class fabio:Expression – resources that either cite or are cited by other bibliographic resources (e.g. journal articles), or that contain such citing/cited resources (e.g. journals); – resource embodiments (re), class fabio:Manifestation – details of the physical or digital forms in which the bibliographic resources are made available by their publishers; – bibliographic entries (be), class biro:BibliographicReference – the literal textual bibliographic entries occurring in the reference lists within the bibliographic resources, that reference other bibliographic resources; – responsible agents (ra), class foaf:Agent – names of agents having certain roles with respect to the bibliographic resources (i.e. names of authors, editors, publishers, etc.); – agent roles (ar), class pro:RoleInTime – roles held by agents with respect to the bibliographic resources (e.g. author, editor, publisher); – identifiers (id) (class datacite:Identifier) – external identifiers (e.g. DOI, ORCID, PubMedID) associated with the bibliographic entities.
The corpus URL (https://w3id.org/oc/corpus/) identifies the entire OCC, which is composed of several sub-datasets, one for each of the six aforementioned bibliographic entities included in the corpus. Each of these has a URL composed by sufixing the corpus URL with the two-letter short name for the class of entity (e.g. “be” for a bibliographic entry) followed by an oblique slash (e.g. https://w3id.org/oc/corpus/be/). Each dataset is described appropriately by means of the Data Catalog Vocabulary23 and the VoID Vocabulary24, and a SPARQL endpoint25 is made available for all the entities included in the entire OCC.
Upon initial curation into the OCC, a URL is assigned to each entity within each sub-dataset, which can be accessed in diferent formats (HTML, RDF/XML, Turtle, and JSON-LD) via content negotiation. Each entity URL is composed by sufixing the sub-dataset URL with a number assigned to each resource, unique among resources of the same type, which increments for each new entry added to that resource class. For instance, the resource https://w3id.org/oc/corpus/be/537 is the 537th bibliographic entry recorded within the OCC. The final part of such URL, i.e. the two-letter short name for the class of items plus “/” plus the number (“be/537” in the example), is called the internal corpus identifier , since it allows the unique identification of any entity within the OCC.
Each of these entities has associated metadata describing its provenance using the PROV-O26 ontology and its PROV-DC extension27 (e.g. https: //w3id.org/oc/corpus/be/537/prov/se/1). In particular, we keep track of the curatorial activities related to each OCC entity, the curatorial agents involved, and their roles.
All these RDF data are stored in BibJSON28 encoded as JSON-LD, defined through an appropriate JSON-LD context 29 which hides the complexity of the model (shown in Fig. 1) behind natural language keywords. For instance, the following excerpt is the JSON-LD linearisation of the aforementioned “be/537” entity: { } "iri": "gbe:537", "a": "entry", "label": "bibliographic entry 537 [be/537]", "content": "Svahn, HA, Berg, A. Single cells or large populations , Lab Chip , 2007, 7, 544, 546, DOI: 10.1039/b704632b , PMID: 17476370", "crossref": "gbr:1601"
In this excerpt, “iri” defines the URL of the resource in consideration (where “gbe:” is a prefix for “https://w3id.org/oc/corpus/be/”), while “a”, “entry”, “label”, “content” and “crossref” stand for rdf:type, biro:Biblio graphicReference, rdfs:label, c4o:hasContent and biro:references respectively (where “gbr:” is a prefix for “https://w3id.org/oc/corpus/br/”).
Additional information about OCC’s handling of citation data, and the
way they are represented in RDF, are detailed in the oficial OCC Metadata
Document [
The ingestion of citation data into the OCC is handled by two Python scripts called Bibliographic Entries Extractor (BEE) and the SPAR Citation Indexer (SPACIN), available in the OCC’s GitHub software repository30.
responsible for the creation of JSON files containing information about the articles in the OA subset of PubMed Central (retrieved by using the Europe PubMed Central API31). Each of these JSON files is created by asking Europe PubMed Central about all the metadata of the articles it stores that have available the source XML file. Once identified, BEE processes all the XML sources so as to extract the complete reference list of each paper under consideration, and includes all the data in the final JSON file. An excerpt of one of those JSON files is as follows: { "doi": "10.1007/s10544 -016-0081-z", "pmid": "27299468", "pmcid": "PMC4908161", "localid": "MED -27299468", "curator": "BEE EuropeanPubMedCentralProcessor", "source": "http://www.ebi.ac.uk/europepmc/webservices/rest/PMC4908161/ fullTextXML", "source_provider": "Europe PubMed Central", "references": [ ... { "bibentry": "Svahn , HA, Berg , A. Single cells or large populations , Lab
Chip , 2007, 7, 544, 546, DOI: 10.1039/b704632b , PMID: 17476370", "pmid": "17476370", "doi": "10.1039/ b704632b", "process_entry": "True" } ]
In particular, for each articles retrieved by means of the Europe PubMed Central API, BEE stores all the available bibliographic identifiers (in the example, “doi”, “pmid”, “pmcid”, and “localid”) and all the textual references, enriched by their own related bibliographic identifiers if those are available. In addition, the JSON file also includes provenance information about the source, its provider and the OCC curator (i.e. the particular BEE Python class responsible for the extraction of these metadata from the source). The created JSON files are then processed, independently, by the tool presented in the next section.
We have undertaken some tests to determine the performances of BEE in generating these JSON files. In particular, we queried Europe PubMed Central for the metadata of articles while running BEE for 30 minutes on a MacBook Pro, with 2 GHz Intel Core i7 processor, 8 GB DDR3 1600 MHz, OS X 10.11.3. During that time, we were able to create 185 JSON ifles containing all the aforementioned metadata, giving a rate of about 6 new JSON files per minute. 30 https://github.com/essepuntato/opencitations 31 https://europepmc.org/RestfulWebService
JSON file created by BEE, retrieving additional metadata information about all the citing/cited articles described in it by querying the Crossref API32 and the ORCID API33. These API are also used to disambiguate bibliographic resources and agents by means of the identifiers retrieved (e.g., DOI, ISSN, ISBN, ORCID, URL, and Crossref member URL). Once SPACIN has retrieved all these metadata, appropriate RDF resources are created (or reused, if they have been already added in the past) and stored in the file system in JSON-LD format (as shown in Section 3.2) and, additionally, within the OCC triplestore. It is worth noting that, for space and performance reasons, the triplestore includes all the data about the curated entities, but does not store their provenance data nor the descriptions of the datasets themselves – these are accessible only via HTTP, not via the SPARQL endpoint.
The SPACIN workflow, described in Fig. 2, is a process that runs until no more JSON files are available from BEE. Thus, the current instance of the OCC is evolving dynamically in time, and can be easily extended beyond Europe PubMed Central by reconfiguring it to interact with additional REST APIs from diferent sources, so as to gather new article metadata and their related references.
Each new resource recorded within the OCC by SPACIN occupies between 0.3 and 4 KB, plus an additional 32 KB dedicated to storage of its provenance data. Each day, the workflow adds about 2 million triples to the corpus, describing more than 20,000 new citing/cited bibliographic resources and about 100,000 new authors, 5% of whom are disambiguated through their ORCID ids. 32 http://api.crossref.org/ 33 http://members.orcid.org/api/
We have tested the performances of SPACIN in processing the JSON ifles generated by BEE and produce new RDF resources for the OCC. In particular, we run SPACIN on two subsets of JSON file: 67 JSON files describing all 67 papers included in the Proceedings of ISWC 2015, and the first 67 JSON files produced by BEE from the Open Access subset of PubMed Central as the outcome of the experiment described in Section 3.3. We use the same configuration as before, i.e. a MacBook Pro, with 2 GHz Intel Core i7 processor, 8 GB DDR3 1600 MHz, OS X 10.11.3.
ISWC 2015 dataset. SPACIN took 45 minutes to process all 67
papers in the ISWC 2015 Proceedings, and the outcomes have been published
in [
Europe PubMed Central dataset. SPACIN took 210 minutes to
process 67 papers from Europe PubMed Central, and the outcomes have
been published in [
In Table 1 we summarise some metrics related to the resources included in the aforementioned exemplar datasets. While these data are far from having a full coverage, they provide interesting snapshots of these two communities. On the one hand, the community of ISWC 2015 is composed by a relatively small number of people. Even if the average number of references per paper is quite small (average 23, the paper with the most references having 47 citation links), there are several papers that were cited more than one time (the most cited one received 7 citations). Many of the citations are to resources for which Crossref was not able to return any metadata. This is understandable, since many citations in these Semantic Web papers are to Web documents (e.g. W3C Recommendations) and to workshop papers not indexed by Crossref (e.g. CEUR Workshop Series34). Some of these non-Crossref-indexed publications are well known and well cited within this community (the most cited one has 4 citations within these 67 ISWC papers). 34 http://ceur-ws.org/
Property Max. number of bibliographic references within a paper
Max. number of citations received by a paper within this
sample Percentage of cited resources for which Crossref did not
return any metadata
Max. number of citations received by a cited resource for which Crossref did not return any metadata 47 7 44% 4 320 3 13% 1
On the other hand, we see a quite diferent citation behaviour in the Europe PubMed Central papers. In this case, as expected, the number of average references is higher (average 50, with one review paper having 320 citation links). The paper within this small sample that has been cited most received only 3 citations from the other 66 papers, and this is clearly due to the dimension of the citation graph of the Biomedical and Life Science community to which PubMed Central relates, which is clearly bigger and more sparsely linked than the ISWC one. Additionally, these papers usually cite others published in journals to which proper identifiers (e.g. DOIs) have been assigned, explaining the lower percentage of citations to resources that are not indexed by Crossref. 4
In recent years we have seen a growing interest within the Semantic Web community for in creating and making available RDF datasets concerning bibliographic metadata of scholarly documents. While the list of such works is quite extensive, inIn this section we describe four of the most important contributions in the area.
The Semantic Lancet35 Project [
Springer LOD37 [
OpenAIRE38 [
Finally, Scholarly Data39 [
In this paper we have introduced the OpenCitations Project, which has created an open repository of accurate bibliographic references harvested from the scholarly literature: the OpenCitations Corpus (OCC). The new instance of the OCC has recently been established, and is already populated with data describing 595,222 citation links (as of August 24, 2016) – a number that will grow quickly over the coming months as the continuous workflow adds new data dynamically from Europe PubMed Central and other authoritative sources. The OCC SPARQL endpoint is presently available for use, and distributions of the OCC datasets will shortly be made openly available for bulk download – the first of these by early September 2016, with subsequent incremental additions.
We are currently working on two diferent aspects. First of all, we are developing tools for linking the resources within the OCC with those included in other datasets, e.g. Scholarly Data and Springer LOD. In addition, we are experimenting with the use of multiple parallel instantiations of SPACIN, so as to increase the amount of new information that can be processed daily into OCC.
Acknowledgements. All the scripts used in OpenCitations have been
developed as outcomes of several personal communications with people
responsible for the external services that OCC uses. We would like to thank
leading people in Europe PubMed Central (in particular Johanna McEntyre
and Vid Vartak), Crossref (in particular Ed Pentz, Geofrey Bilder, and
Karl Ward) and ORCID (in particular Josh Brown and Laurel Hack) for
their help. We would also like to thank Alfred Hofmann and Aliaksandr
Birukou (Springer Nature) for allowing us to publish in Figshare the OCC
metadata concerning the Proceedings of ISWC 2015 [