Introduction

Wikipedia contains significant amounts of data relevant for scientometrics, and it has formed the basis for several scientometric studies [5,12,15,16,21,24,27]. Such studies can use the structured references found in Wikipedia articles or use the intrawiki hyperlinks, e.g., to compare citations from Wikipedia to scholarly journals with Thomson Reuters journal citation statistics as in [15] or to rank universities as in [27].

While many Wikipedia pages have numerous references to scientific articles, the current Wikipedias have very few entries about specific scientific articles. This is most evident when browsing the Academic journal articles category on the English Wikipedia. 4 Among the few items in that category are famed papers such as the 1948 physics paper The Origin of Chemical Elements [2] -described in the English Wikipedia article Alpher-Bethe-Gamow paper 5 -as well as the 1953 article Molecular Structure of Nucleic Acids: A Structure for Deoxyribose Nucleic Acid [4] on eight Wikipedias. Another scientific article is Hillary Putnam's Is Semantics Possible? [22] 6 from 1970 on the Estonian Wikipedia.

References in Wikipedia are often formatted in templates, and it takes some effort to extract and match information in these fields. For instance, in a study of journals cited on Wikipedia, a database was built containing journal name variations to match the many different variations that Wikipedia editors used when citing scientific articles [15]. The use of standard identifiers -such as the Digital Object Identifier (DOI) -in citations on Wikipedia can help to some extent to uniquely identify works and journals.

Several other wikis have been set up to describe scientific articles, such as WikiPapers, AcaWiki, Wikilit [19] and Brede Wiki [17]. 7 They are all examples of MediaWiki-based wikis that primarily describe scientific articles. Three of them use the Semantic MediaWiki extension [13], while the fourth uses MediaWiki's template functionality 8 to structure bibliographic information.

Since the launch of Wikidata9 [25], the Wikimedia family includes a platform to better handle structured data such as bibliographic data and to enforce input validation to a greater degree than Wikipedia. Wikidata data can be reified to triples [8], and graph-oriented databases, including SPARQL databases, can represent Wikidata data [9]. The Wikidata Query Service (WDQS) 10 is an extended SPARQL endpoint that exposes the Wikidata data. Apart from offering a SPARQL endpoint, it also features an editor and a variety of frontend result display options. It may render the SPARQL query result as, e.g., bubble charts, line charts, graphs, timelines, list of images, points on a geographical map, or just provide the result as a table. These results can also be embedded on other Web pages via an HTML iframe element. We note that Wikidata is open data published under CC0, and it is available not only through the SPARQL endpoint, but also-like any other project of the Wikimedia family-through an API and dump files. 11In the following sections, we describe how Wikidata has been used for bibliographic information and present Scholia, our website built to expose such information. We furthermore show how Scholia can be used for bibliography generation and discuss limitations and advantages with Wikidata and Scholia.

Bibliographic information on Wikidata

Wikidata editors have begun to systematically add scientific bibliographic data to Wikidata across a broad range of scientific domains -see Table 1 for a summary of Wikidata as a digital library. Individual researchers and scientific articles not described by their own Wikipedia article in any language are routinely added to Wikidata, and we have so far experienced very few deletions of such data in reference to a notability criterion. The current interest in expanding bibliographic information on Wikidata has been boosted by the WikiCite project, which aims at collecting bibliographic information in Wikidata and held its first workshop in 2016 [23].

The bibliographic information collected on Wikidata is about books, articles (including preprints), authors, organizations, journals, and publishers. These items (corresponding to subject in Semantic Web parlance) can be interlinked through Wikidata properties (corresponding to the predicate), such as author (P50), 13 published in (P1433), publisher (P123), series (P179), main theme (P921), educated at (P69), employer (P108), part of (P361), sponsor (P859, can be used for funding), cites (P2860) and several other properties. 14Numerous properties exist on Wikidata for deep linking to external resources, e.g., for DOI, PMID, PMCID, arXiv, ORCID, Google Scholar, VIAF, Crossref funder ID, ZooBank and Twitter. With these many identifiers, Wikidata can act as a hub for scientometrics studies between resources. If no dedicated Wikidata property exists for a resource, one of the URL properties can work as a substitute for creating a deep link to a resource. For instance, P1325 (external data available at) can point to raw or supplementary data associated with a paper. We have used this scheme for scientific articles associated with datasets stored in OpenfMRI [20], an online database with raw brain measurements, mostly from functional magnetic resonance imaging studies. Using WDQS, we query the set of OpenfMRI-linked items using the following query:

? item wdt : P1325 ? resource . filter strstarts ( str (? resource ) , " https :// openfmri . org / dataset / " )

A similar scheme is used for a few of the scientific articles associated with data in the neuroinformatics databases Neurosynth [26] and NeuroVault [6].

When bibliographic items exist in Wikidata, they can be used as references to support claims (corresponding to triplets with extra qualifiers) in other items of Wikidata, e.g., a biological claim can be linked to the Wikidata item for a scientific journal.

By using these properties systematically according to an emerging data model, 15 editors have extended the bibliographic information in Wikidata. Particularly instrumental in this process was a set of tools built by Magnus Manske, QuickStatements 16 and Source MetaData, 17 including the latter's associated Resolve authors tool. 18 Information can be extracted from, e.g., PubMed, PubMed Central and arXiv and added to Wikidata.

How complete is Wikidata in relation to scientific bibliographic information? Journals and universities are well represented. For instance, 31,895 Wikidata items are linked with the identifier for the Collections of the National Library of Medicine (P1055). Far less covered are individual articles, individual researchers, university departments and citations between scientific articles. Most of the scientific articles in Wikidata are claimed to be an instance of (P31) the Wikidata item scientific article (Q13442814). With a WDQS query, we can count the number of Wikidata items linked to scientific article: select ( count (? work ) as ? count ) where { ? work wdt : P31 wd : Q13442814 . }

As of 12 March 2017, the query returned the result 615,182, see also Table 1.

In comparison, arXiv states having 1,240,585 e-prints and ACM Digital Library states to have 24,110 proceedings. 19 There were 8,617 authors associated with Wikidata items linked through the author property (P50) to items that are instance of scientific article, and the number of citations as counted by triples using the P2860 (cites) property stood at 2,729,164:

select ( count (? citedwork ) as ? count ) where { ? work wdt : P2860 ? citedwork . }

The completeness can be fairly uneven. Articles from PLOS journals are much better represented than articles from the journals of IEEE.

The sponsor property has been used extensively for National Institute for Occupational Safety and Health (NIOSH) with 52,852 works linking to the organization, 18,135 of which are instance of scientific articles, but apart from NIOSH, the use of the property has been very limited for scientific articles. 20 Scholia provides both a Python package and a Web service for presenting and interacting with scientific information from Wikidata. The code is available via https://github. com/fnielsen/scholia, and a first release has been archived in Zenodo [18]. As a Web service, its canonical site runs from the Wikimedia Foundation-provided service Wikimedia Tool Labs at https://tools.wmflabs. org/scholia/, but the Scholia package may be downloaded and run from a local server as well. Scholia uses the Flask Python Web framework [7]. The current Web service relies entirely on Wikidata for all its presented data. The frontend consists mostly of HTML iframe elements for embedding the on-the-flygenerated WDQS results and uses many of the different output formats from this service: tables, bubble charts, bar charts, line charts, graphs and image lists.

Through a JavaScript-based query to the MediaWiki API, an excerpt from the English Wikipedia is shown on the top of each Scholia page if the corresponding Wikidata item is associated with an article in the English Wikipedia. The label for the item is fetched via Wikidata's MediaWiki API. While some other information can be fetched this way, Scholia's many aggregation queries are better handled through SPARQL.

Scholia uses the Wikidata item identifier as its identifier rather than author name, journal titles, etc. A search field on the front page provides a Scholia user with the ability to search for a name to retrieve the relevant Wikidata identifier. To display items, Scholia sets up a number of what we call "aspects". The currently implemented aspects are author, work, organization, venue, series, publisher, sponsor and topic, see Table 2. The present selection was motivated by the possibilities inherent in the Wikidata items and properties. We plan to extend this to further aspects, e.g., award or determination method. A URL scheme distinguishes the different aspects, so the URL path /scholia/author/Q6365492 will show the author aspect of the statistician Kanti V. Mardia, while /scholia/topic/Q6365492 will show the topic aspect of the person, e.g., articles about Mardia. Likewise, universities can be viewed, for instance, as organizations or as sponsors. Indeed, any Wikidata item can be viewed in any Scholia aspect, but Scholia can show no data if the user selects a "wrong" aspect, i.e. one for which no relevant data is available in Wikidata.

For each aspect, we make multiple WDQS queries based on the Wikidata item for which the results in the panels are displayed, -technically in embedded iframes. For the author aspect, Scholia queries WDQS for the list of publications, showing the result in a table, displaying a bar chart of the number of publications per year, number of pages per year, venue statistics, co-author graph, topics of the published works (based on the "main theme" property), associated images, education and employment history as timelines, academic tree, map with locations associated with the author, and citation statistics -see Fig. 1 for an example of part of an author aspect page. The citation statistics displays the most cited work, citations by year and citing authors. For the academic tree and the citation graph, we make use of Blazegraph's graph analytics RDF GAS API 21 that is available in WDQS. The embedded WDQS results link back to WDQS where a user can modify the query. The interactive editor of WDQS allows users not familiar with SPARQL to make simple modifications without directly editing the SPARQL code.

Related to their work on quantifying conceptual novelty in the biomedical literature [14], Shubhanshu Mishra and Vetle Torvik have set up a website profiling authors in PubMed datasets: LEGOLAS. 22 Among other information, the website shows the number of articles per year, the number of citations per year, the number of self-citations per year, unique collaborations per year and NIH grants per year as bar charts that are color-coded according to, e.g., author role (first, solo, middle or last author). Scholia uses WDQS for LEGOLAS-like plots. Figure 2 displays one such example for the number of published items as a function of year of publication on an author aspect page, where the components of the bars are color-coded according to author role.

For the organization aspect, Scholia uses the employer and affiliated Wikidata properties to identify associated authors, and combines this with the author query for works. Scholia formulates SPARQL queries with property paths to identify suborganizations of the queried organization, such that authors affiliated with a suborganization are associated with the queried organization. bar chart uses the P1104 (number of pages) Wikidata property together with a normalization based on the number of authors on each of the work items. The bars are color-coded according to individual authors associated with the organization. In this case, the plot is heavily biased, as only a very limited subset of publications from the organization is currently present in Wikidata, and even the available publications may not have the P1104 property set. Other panels shown in the organization aspect are a co-author graph, a list of recent publications formatted in a table, a bubble chart with most cited papers with affiliated first author and a bar chart with co-author-normalized citations per year. This last panel counts the number of citations to each work and divides it by the number of authors on the cited work, then groups the publications according to year and color-codes the bars according to author.

For the publisher aspect, Scholia queries all items where the P123 property (publisher) has been set. With these items at hand, Scholia can create lists of venues (journals or proceedings) ordered according to the number of works (papers) published in each of them, as well as lists of works ordered according to citations. Fig. 4 shows an example of a panel on the publisher aspect page with a scatter plot detailing journals from BioMed Central. The position of each journal in the plot reveals impact factor-like information.

For the work aspect, Scholia lists citations and produces a partial citation graph. Fig. 5 shows a screenshot of the citation graph panel from the work aspect for a specific article [3]. For this aspect, we also formulate a special query to return a table with a list of Wikidata items where the given work is used as a source for claims. An example query for a specific work is shown with Listing 1. From the query results, it can be seen, for instance, that the article A novel family of mammalian taste receptors [1] supports a claim about Taste 2 receptor member 16 (Q7669366) being present in the cell component (P681) integral component of membrane (Q14327652). For the topic aspect, Scholia uses a property path SPARQL query to identify subtopics. For a given item where the aspect is not known in advance, Scholia tries to guess the relevant aspect by looking at the instance of property. The Scholia Web service uses that guess for redirecting, so /scholia/Q8219 will redirect to /scholia/author/Q8219, the author aspect for the psychologist Uta Frith. This is achieved by first making a server site query to establish that Uta Frith is a human and then using that information to choose the author aspect as the most relevant aspect to show information about Uta Frith.

A few redirects for external identifiers are also implemented. For instance, with Uta Frith's Twitter name 'utafrith', /scholia/twitter/utafrith will redirect Listing 1. SPARQL query on the work aspect page for claims supported by a work, -in this case Q22253877 [1].

SELECT distinct ? item ? itemLabel ? property ? propertyLabel ? value ? valueLabel WHERE { ? item ? p ? statement . ? property wikibase : claim ? p . ? statement ? a ? value . ? item ? b ? value . ? statement prov : wasDerivedFrom / < http :// www . wikidata . org / prop / reference / P248 > wd : Q22253877 . SERVICE wikibase : label { bd : serviceParam wikibase : language " en " } } ORDER BY ? itemLabel to /scholia/Q8219, which in turn will redirect to /scholia/author/Q8219. Scholia implements similar functionality for DOI, ORCID and GitHub user identifier.

Using Wikidata as a bibliographic resource

As a command-line tool, Scholia provides a prototype tool that uses Wikidata and its bibliographic data in a L A T E X and BibT E X environment. The current implementation looks up citations in the latex-generated .aux file and queries Wikidata's MediaWiki API to get cited Wikidata items. The retrieved items are formatted and written to a .bib that bibtex can use to format the bibliographic items for inclusion in the L A T E X document. The workflow for a L A T E X document with the filename example.tex is latex example python -m scholia . tex write -bib -from -aux example . aux bibtex example latex example latex example

Here, the example document could read

\ documentclass { article } \ usepackage [ utf8 ]{ inputenc } \ begin { document } \ cite { Q18507561 } \ bibliographystyle { plain } \ bibliography { example } \ end { document }

In this case, the \cite command cites Q18507561 (Wikidata: a free collaborative knowledgebase [25]). A DOI can also be used in the \cite command: instead of writing \cite{Q18507561}, one may write \cite{10.1145/2629489} to get the same citation. Scholia matches on the "10." DOI prefix and makes a SPARQL query to get the relevant Wikidata item.

The scheme presented above can take advantage of the many available style files of BibT E X to format the bibliographic items in the various ways requested by publishers. We have used Scholia for reference management in this paper.

Discussion

WDQS and Scholia can provide many different scientometrics views of the data available in Wikidata. The bibliographic data in Wikidata are still quite limited, but the number of scientometrically relevant items will likely continue to grow considerably in the coming months and years.

The continued growth of science data on Wikidata can have negative impact on Scholia, making the on-the-fly queries too resource demanding. In the current version, there are already a few queries that run into WDQS's time out, e.g., it happens for the view of co-author-normalized citations per year for Harvard University. If this becomes a general problem, we will need to redefine the queries. Indeed, the WDQS time out will be a general problem if we want to perform large-scale scientometrics studies. An alternative to using live queries would be using dumps, which are available in several formats on a weekly basis, with daily increments in between. 23 The problem is not a limitation of SPARQL, but a limitation set by the server resources. Some queries may be optimized, especially around the item labeling.

Working with Scholia has made us aware of several issues. Some of these are minor limitations in the Wikidata and WDQS systems. The Wikidata label length is limited to 250 characters, whereas the 'monolingual text' datatype used for the 'title' property (P1476) is limited to 400 characters. There are scholarly articles with titles longer than those limits.

Wikidata fields cannot directly handle subscripts and superscripts, which commonly appear in titles of articles about chemical compounds, elementary particles or mathematical formulas. Other formatting in titles cannot directly be handled in Wikidata's title property, 24 and recording a date such as "Summer 2011" is difficult.

Title and names of items can change. Authors can change names, e.g. due to marriage, and journals can change titles, e.g. due to a change of scope or transfer of ownership. For instance, the Journal of the Association for Information Science and Technology has changed names several times over the years. 25 Wikidata can handle multiple titles in a single Wikidata item and with qualifiers describe the dates of changes in title. For scientometrics, this ability is an advantage in principle, but multiple titles can make it cumbersome to handle when Wikidata is used as a bibliographic resource in document preparation, particularly for articles published near the time when the journal changed its name. One way to alleviate this problem would be to split the journal's Wikidata item into several, but this is not current practice.

In Wikidata, papers are usually not described to be affiliated with organizations. Scholia's ability to make statistics on scientific articles published by organization is facilitated by the fact that items about scientific articles can link to items about authors, which can link to items about organizations. It is possible to link scientific articles to organization by using Wikidata qualifiers in connection with the author property. However, this scheme is currently in limited use.

This scarcity of direct affiliation annotation on Wikidata items about articles means that scientometrics on the organizational level are unlikely to be precise at present. In the current version, Scholia even ignores any temporal qualifier for the affiliation and employer property, meaning that a researcher moving between several organization gets his/her articles counted under multiple organizations.

Data modeling on Wikidata gives rise to reflections on what precisely a "publisher" and a "work" is. A user can set the publisher Wikidata property of a work to a corporate group, a subsidiary or possibly an imprint. For instance, how should we handle Springer Nature, BioMed Central and Humana Press?

Functional Requirements for Bibliographic Records (FRBR) [11] suggests a scheme for works, expressions, manifestations and "items". In Wikipedia, most items are described on the work level as opposed to the manifestation level (e.g., book edition), while citations should usually go to the manifestation level. How should one deal with scientific articles that have slightly different "manifestations", such as preprint, electronic journal edition, paper edition and postprint, or editorials that were co-published in multiple journals with identical texts? An electronic and a paper edition may differ in their dates of publication, but otherwise have the same bibliographic data, while a preprint and its journal edition usually have different identifiers and may also differ in content. From a scientometrics point of view, these difference in manifestation may not matter in some cases, but could be the focus of others. Splitting a scientific article as a work (in the FRBR sense) over multiple Wikidata items seems only to complicate matters.

The initial idea for Scholia was to create a researcher profile based on Wikidata data with list of publications, picture and CV-like information. The inspiration came from a blog post by Lambert Heller: What will the scholarly profile page of the future look like? Provision of metadata is enabling experimentation. 26 In this blog post, he discussed the different features of several scholarly Web services: ORCID, ResearchGate, Mendeley, Pure, VIVO, Google Scholar and ImpactStory. In Table 3, we have set up a table listing Heller's features for the Wikidata-Scholia combination. Wikidata-Scholia performs well in most aspects, but in the current version, Scholia has no backend for storing user data, and user features such as forum, Q&A and followers are not available.

Beyond the features listed by Heller, which features set Wikidata-Scholia apart from other scholarly Web services? The collaborative nature of Wikidata means that Wikidata users can create items for authors that do not have an account on Wikidata. In most other systems, the researcher as a user of the system has control over his/her scholarly profile and other researchers/users cannot make amendment or corrections. Likewise, when one user changes an existing item, this change will be reflected in subsequent live queries of that item, and it may still be in future dumps if not reverted or otherwise modified before the dump creation.

With WDQS queries, Scholia can combine data from different types of items in Wikidata in a way that is not usually possible with other scholarly profile Web services. For instance, Scholia generates lists of publications for an organization by combining items for works and authors and can show co-author graphs re- stricted by affiliation. Similarly, the co-author graph can be restricted to authors publishing works annotated with a specific main theme. Authors are typically annotated with gender in Wikidata, so Scholia can show gender color-coding of co-author graphs. On the topic aspect page, the Scholia panel that shows the most cited works that are cited from works around the topic can point to an important paper for a topic -even if the paper has not been annotated with the topic -by combining the citations data and topic annotation. References for claims are an important part of Wikidata and also singles Wikidata out among other scholarly profile Web service, and it acts as an extra scientometrics dimension. The current version of Scholia has only a single panel where the query uses references: the "Supports the following statement(s)" on the work aspect page, but it is possible to extend the use of this scientometrics dimension.