Wikidata and the bibliography of life

Wikicite

The original scope of Wikidata was to provide structured data to underpin the different Wikipedia projects. Hence, notionally each item in Wikidata had a corresponding entity in at least one of the various Wikipedias. However, as Wikidata has grown the potential of having a single, queryable, community-edited database of structured information has become increasingly clear. Hence many items being added to Wikidata might not themselves have a Wikipedia page but are relevant to the content and goals of Wikipedia. A good example of this are bibliographic citations, which are a key source of support for factual statements made on Wikipedia.

The Wikicite project started out with the goal to provide structured bibliographic data for citations across the different Wikipedia projects. Given that the scope of Wikipedia includes taxonomy, many of the publications cited in Wikipedia (and hence destined to be in Wikidata) are relevant to taxonomy. Furthermore, there is a wiki devoted entirely to taxonomy (Wikispecies), which includes pages for taxa, taxonomists, and taxonomic publications. Many of these pages also have corresponding items in Wikidata. Hence a considerable amount of taxonomic literature has already been added by contributors to the WikiCite project.

Data contributions to Wikidata typically come in two forms, manual edits by individual people or automated edits by software (“bots”). A number of bots add bibliographic metadata sourced from databases such as PubMed and CrossRef. For example, given a CrossRef DOI for an article the CrossRef API can be used to retrieve the metadata for the corresponding article. If one wanted to include only publications cited by Wikipedia, one would then need a list of DOIs cited on Wikipedia pages. Alternatively, one could proactively add articles with DOIs to Wikidata even if they aren’t currently cited on Wikipedia, on the assumption that as Wikipedia grows it is likely that more and more articles will be cited. This means it is a short step to expanding the scope to include most, if not all of the academic corpus in Wikidata. One motivation for this is to have openly accessible bibliographic data which can be used to enable freely accessible measures of the activity and impact of researchers (Nielsen, Mietchen & Willighagen, 2017).

As a consequence of work done by the WikiCite community, and the prominence of taxonomy in Wikipedia and Wikispecies, Wikidata already contains a considerable number of publications relevant to taxonomy. This, coupled with the sophistication of the data model, powerful query language, and the existence of an enthusiastic community of editors makes a strong case for Wikidata being a promising platform for a bibliography of life.

Bibliographic data in Wikidata

The Wikidata model for a publication has evolved over time as the community adds properties and recommendations for their use. Figure 1 shows how a scientific article can be modelled in Wikidata.

Wikidata items are given one or more types using Wikidata property P31 (instance), such as Q13442814 for a scholarly article, and Q571 for a book. There are properties for the typical metadata associated with an article, such as title, journal that contains the article, volume, pagination, and date of publication. Wikidata supports values in multiple languages, so that articles with titles in multiple languages can have all those titles represented. Authorship is handled in two distinct but complementary ways. If an author of a publication is known to have a Wikidata entry then the author property (P50) links the item for the publication to the item for that author. If it is not known whether the author exists in Wikidata their name can be stored as a simple string value (P2093). In Fig. 1 there are examples of both authors. There are tools available to subsequently map those name strings to the corresponding Wikidata items.

External identifiers, such as ones provided by the publishing industry (e.g., DOIs), archiving services (e.g., Handles), and domain-specific databases (such as PubMed, ZooBank, etc.) can also be added to the Wikidata item. Wikidata items are being decorated with an increasing number of diverse identifiers, hence Wikidata is increasingly playing a role as an “identity broker” enabling cross-links between identifiers from different databases (Van Veen, 2019).

Links between publications

Publications rarely exist in isolation from each other, hence we can connect them using a range of properties. The most obvious relationship is citation, where one publication cites another. Adding this information helps flesh out the citation graph, enables us to track the provenance of an idea, and also discover potentially related publications through co-citation (Marshakova-Shaikevich, 1973; Small, 1973).

Other relationships supported by Wikidata include errata where one publication corrects errors or mistakes in a previous publication, and translations, where a publication may exist in more than one language. For example, the article Korotyaev (2018) is an English translation of Коротяев (2018), the corresponding items in Wikidata can be connected by properties reflecting that relationship.

Links to facts

A key motivation for including publications in Wikidata is to provide trustworthy sources of references for statements made in Wikidata. For example, statements about the birth and death dates for a person, the exact date of publication of a work, the date at which a journal changed its name, or the publication of a taxonomic name can all be supported by adding references to the relevant source.

As an example, the taxonomic name Euphorbia bicompacta Bruyns was published in Bruyns, Mapaya & Hedderson (2006) as a replacement for the name Synadenium compactum N.E.Br. This publication (Q28960244) is the one discussed above in Fig. 1. The Wikidata item for Euphorbia bicompacta (Q5851419) has a property “taxon name” (P225) with the value “Euphorbia bicompacta” and Wikidata item Q28960244 as a reference for that value (see Fig. 2).

Populating Wikidata

Creating a bibliography of life would only be conceivable if much of the work of populating it could be automated, and if freely accessible sources of data were available. Bibliographic metadata from CrossRef and PubMed are constantly being added by automated tools (“bots”). This means that many publications that have a CrossRef Digital Object Identifier (DOI) or have an entry in PubMed are likely to be already in Wikidata. If they aren’t, then it is straightforward to add them. Data from these sources are typically of high quality, although sometimes the data is limited or incorrect, for example, in not including lists of literature cited, or there may be typographic or character encoding errors in the data. An advantage of a community-editable resource is that these can be found and subsequently corrected by the community.

While much of the biomedical literature, and an increasing fraction of modern taxonomic literature has CrossRef DOIs, much of the taxonomic corpus either lacks a DOI, or may have a DOI issued by a registration agency other than CrossRef. The DOI foundation has several members that issue DOIs, and these differ in the support they provide for resolving DOIs to machine readable data. CrossRef DOIs can return extensive metadata about an article in CiteProc JSON, a default standard for bibliographic metadata (Willighagen, 2019; Bennett, 2021). Some DOI agencies support CiteProc (albeit not as fully populated as CrossRef), however agencies such as ITISC—which is issuing DOIs for many Chinese articles (Wang et al., 2018)—do not support machine readability at all. Hence, not all DOIs are equally easy to work with.

There are also publications with persistent identifiers that are not DOIs (such as Handles), publications which lack persistent identifiers but are online, and publications which may not be online at all. There are various strategies we can use to gather bibliographic data for these publications. Below I describe some of these strategies. Source code for some of these approaches is available at https://github.com/rdmpage/wikidata-bibliographic-data.

Scrape metadata from the web

Web sites for some journals contain embedded machine-readable metadata about publications in their web pages to enhance discoverability by search engines such as Google Scholar. These tags also enable software tools (e.g., reference managers such as Zotero) to easily extract bibliographic metadata to be stored by users of those tools. Although typically there are journal and publisher-specific idiosyncrasies in how the metadata is marked up, it is relatively straightforward to write software to fetch these web pages and extract the metadata.

Lists of literature cited

Most articles will have a list of literature cited, so when taxonomists publish their work they are also continually publishing bibliographic metadata. These lists are becoming increasingly accessible to machines. Furthermore, CrossRef is encouraging publishers to include lists of references cited in their submissions to CrossRef. If both the citing article and the cited article have CrossRef DOIs and the citing article references are submitted to CrossRef, then this citation link will appear in the next update of COCI, the OpenCitations index of open Crossref DOI-to-DOI citations (Heibi, Peroni & Shotton, 2019). While this helps grow the citation network, it overlooks all those publications that lack DOIs (or which lacked them at the time the citing article was published). However, the metadata for references cited which lack DOIs can still be used to help populate Wikidata.

Some publishers provide article text in machine-readable formats such as XML where the references are identified and can be easily extracted. Other publishers may provide lists of references in the web view of an article, sometimes with embedded markup. Hence, we can regard taxonomists as, in effect, “crowd sourcing” the taxonomic literature simply by the act of publishing their research. For example, articles published in the journal Zootaxa together contain over a million references cited (Page, 2020).

Taxonomic databases

The numerous taxonomic databases being developed by the community, often focussed on a particular taxonomic group, are yet another source of bibliographic data. Regrettably, in many cases taxonomic databases do not treat the taxonomic literature as a first-class citizen, and hence the data may be stored in an abbreviated form (such as the micro-citations mentioned above). But some databases do provide high-quality curated literature which can be used to help populate Wikidata.

Databases of researchers

Yet another potential source of data are the collections of articles created by researchers as part of an online profile or identity, such as ORCID (https://orcid.org/) or ResearchGate (https://www.researchgate.net/). Using a combination of manual input and web services, ORCID assembles a list of publications (and other outputs) linked to a researcher’s unique identifier (their ORCID id). This data is openly available via an API. In contrast, ResearchGate is a commercial website where members can upload lists of their publications, and provide access to the publications themselves (on the understanding that their members have the legal right to do so). Although ResearchGate is “closed” in that it lacks a publicly available API, they do embed structured markup in their web pages which links authors to their publications using terms from the https://schema.org/ vocabulary.

Wikis

The sources which perhaps most closely match the notion of “crowd sourcing” are Wikidata itself, and other wikis of the Wikipedia Foundation, such as Wikipedia and Wikispecies. Indeed, in much the same way that we can regard Wikipedia as an Encyclopaedia of Life (Page, 2010), Wikispecies can be regarded as a crowd sourced “bibliography of life” where volunteers are assembling a wiki with one page per taxon, often including extensive lists of references cited. However, these references are often entered as simple text strings with little or no structured markup, making it challenging to extract structured metadata, and hence limiting the utility of Wikispecies.

Full-text

Wikidata stores metadata rather than full-text content, that is, it stores information about a publication, not the contents of the publication itself. A growing proportion of the taxonomic literature is being digitised, such that articles may be available in formats such as PDF or sets of images (e.g., scans of printed works). Given the alarming ease with which links to online content can break (Laakso, Matthias & Jahn, 2020) a convention on Wikidata is to include not only a link to a freely available PDF but also a link to an archived version, e.g. on the Internet’s Wayback Machine (https://web.archive.org/). Another strategy (one that I have regularly used) is to store a copy of the PDF on Internet Archive itself and include the Internet Archive identifier as a property of the publication on Wikidata.

Other ways to access content include tools that take a DOI and return a PDF if one is available online, either freely available, e.g. Unpaywall (http://unpaywall.org/) or “pirated” (Bohannon, 2016). Some publishers such as the China National Knowledge Infrastructure (CNKI) have mobile phone apps that provide access to their content through that app.

Being able to access the content of the articles themselves not only means that we can read the article, but it also provides a way to augment existing metadata. In my own experience key data such as page numbers were often not recorded in the available metadata for an article. This can make it harder to link publications to taxonomic names using “microcitations”, where the only information we have is a journal, a volume, and a page number. However, if we have access to a digital version of the article we can extract the page numbers. This need not be a manual process, for instance the Internet Archive generates a file for each PDF that contains a best-guess of the page numbers in the PDF. We can use those to add missing pagination values to the corresponding Wikidata items.

A community of editors

One of the challenges in community-based editing of scientific data is assembling that community. We could create a domain-specific database and hope a community coalesces around that database. Alternatively, we take the data to where an active community already exists. This is the approach taken by projects such as Gene Wiki (Good et al., 2012). If Wikidata is going to be the place to assemble the bibliography of life, a natural question is: “Does the community actually edit taxonomic publications?” To assess this, I looked at the edit history of the random sample of 10,000 Wikidata items generated above. For each of these items I retrieved the number of edits made since the Wikidata item was created, when those edits were made, and what properties were edited.

Figure 3 visualises the edit history for the sample of 10,000 publications as a scatter plot of creation timestamp against edit timestamp. If an item was only edited at the time it was created then all points would fall along the diagonal and the lower right triangle in Fig. 3 would be empty. This diagonal continues to go up and to the right as time goes on. Any edit to an item appears as a dot to the right of the dot on the diagonal that represent the item’s creation. If there are no dots to the right of the diagonal, then an item has not been edited since its creation. Figure 3 shows that many items undergo a series of sporadic edits over time. Some of these edits occur shortly after item creation. For example, there are Wikidata bots whose function is to add a description for a new item in a specific language. Other edits may happen later in the life cycle of an item, for example if a user associates a publication with its author or links a publication to its main subject. Or there may be a bulk update of many items by a bot that edits a specific property.

The most common edits observed in the publications involved the authors of those publications (properties P2093 and P50), as well as adding values for P921 “main subject” (a form of tagging an item) (Fig. 4). Edits in Wikidata can be made by people, either directly by editing a record in Wikidata, or using bulk tools such as Quickstatements (https://quickstatements.toolforge.org/#/). Edits can also be made by automated programs (“bots”). Of the top ten editors of publications, half are bots (Fig. 5).

This approach to measuring edit activity assumes that only edits made to an item itself are relevant. However edits may be made to other items that link those items to the current item. For example, adding a “cites work” statement to an item does not result in any changes to the item being cited (i.e., the target of the “cites work” statement).

Knowledge graph density

Conceptually a knowledge graph comprises entities (nodes) that are connected by facts (edges). The number of facts for an entity is a measure of the knowledge graph’s density, which for many graphs is low, often averaging less than two facts per entity (Hegde & Talukdar, 2015). Note that this definition of “facts” ignores simple statements associated with an entity (e.g., the number of pages in an article). These are also facts in the sense of being statements about an entity, but we don’t need a knowledge graph to store them. The true power of a knowledge graph comes from the density of the connections between entities.

To assess the connection density of bibliographic entities in Wikidata, I counted the number of links between bibliographic items and other Wikidata entities in the sample of 10,000 bibliographic items. In counting these connections some entities, such as those for language, were not counted to avoid inflating the estimate of knowledge density based on what are essentially administrative metadata. The properties that were counted are shown in Table 2.

The average link density for the sample of publications was 4.17, with the modal number of connections being one. Hence this part of the knowledge graph is relatively sparse, with most publications having just the connection to a parent publication (typically a journal). Some publication items are connected to other items via citation relationships, either as the source or the target of that relationship (i.e., citing or cited by).

Author coverage

The bulk of publications added to Wikidata treat authors as “strings” not “things”, that is, most authors are listed as names using the P2093 “authors name string” property, rather than as Wikidata items using the P50 “author” property (see Fig. 1). Ideally all authors of publications would be Wikidata items, not simply text strings, and indeed making that conversion is among the most commonly made edits (Fig. 4). Realising this goal requires that all authors of taxonomic publications have items in Wikidata, which in turn is part of a broader goal of having a Wikidata item for everyone involved in taxonomic research (Groom et al., 2020).

There are several databases of taxonomists that have representation in Wikidata, although their coverage in Wikidata is variable. For example, the International Plant Names Index (IPNI) contained approximately 43,000 authors in 2013 (Lindon et al., 2015), and currently some 53,000 Wikidata items have IPNI author IDs. At the time of writing (2021) ZooBank (Pyle & Michel, 2008) contains some 87,000 authors, of which 17,000 are in Wikidata. The Biodiversity Heritage Library has 28,500 authors in Wikidata, while Wikispecies contributes 61,000 authors to Wikidata. There is overlap among these sources. For example, almost all of the ZooBank authors that are in Wikidata are also in Wikispecies, whereas the majority of authors sourced from IPNI are unique to IPNI (Fig. 6). What is unclear is how much of the lack of overlap between authors in the different sources databases is real (do they represent different sets of authors?), vs. a lack of mapping between identifiers (how many records are for the same people, just using different identifiers?). There is considerable scope for reconciling authors between these databases, as well as other sources of information on people, such as ORCID and ResearchGate. It is not enough to merely have authors represented in Wikidata, we also need to link them to their publications. The source databases (BHL, IPNI, Wikispecies, and ZooBank) all contain links between authors and their publications, and much more use could be made of these sources to add P50 author links (Page, 2019).