Linked Dataset description papers at the Semantic Web journal: A critical assessment

Dataset papers at the Semantic Web journal

On February 29th, 2012, the Semantic Web Journal (SWJ) announced the first “Special Call” for Linked Dataset descriptions.3

When compared with data journals and tracks in other disciplines, the SWJ calls have some subtle but key differences. While in other areas the emphasis is on the potential scientific value of that dataset to that research community, for SWJ datasets the core requirement is that the dataset is modelled appropriately using Semantic Web standards and published correctly using Linked Data principles. In this sense, the topic and domain of the dataset for SWJ is not as important. For example, a Linked Dataset about historical events may not have a direct influence on Semantic Web research in the same way a dataset about genes may have impact in the Bioinformatics community, but can still be accepted by SWJ if deemed of potential impact in another area (or in practice). Thus, the goal for SWJ is in publishing descriptions of high-profile exemplars of what a Linked Dataset can and should be, no matter their topic.

Criteria for review

The criteria for accepting dataset papers differ from those for other tracks. Specifying these criteria in a precise way in the call for papers is crucial to ensure that authors know what requirements they need to fulfil before their paper can be accepted and to ensure that reviewers for different papers can apply a consistent standard. However, given the nature of the track, specifying precise criteria proved challenging and the call has undergone various revisions. The current call for dataset descriptions is as follows:

Linked Dataset Descriptions – short papers (typically up to 10 pages) containing a concise description of a Linked Dataset. The paper shall describe in concise and clear terms key characteristics of the dataset as a guide to its usage for various (possibly unforeseen) purposes. In particular, such a paper shall typically give information, amongst others, on the following aspects of the dataset: name, URL, version date and number, licensing, availability, etc.; topic coverage, source for the data, purpose and method of creation and maintenance, reported usage etc.; metrics and statistics on external and internal connectivity, use of established vocabularies (e.g., RDF, OWL, SKOS, FOAF), language expressivity, growth; examples and critical discussion of typical knowledge modeling patterns used; known shortcomings of the dataset. Papers will be evaluated along the following dimensions: (1) Quality and stability of the dataset – evidence must be provided. (2) Usefulness of the dataset, which should be shown by corresponding third-party uses – evidence must be provided. (3) Clarity and completeness of the descriptions. Papers should usually be written by people involved in the generation or maintenance of the dataset, or with the consent of these people. We strongly encourage authors of dataset description paper to provide details about the used vocabularies, ideally using the 5 star rating [ 20 ].

This captures much of the criteria that the SWJ editors currently feel important for a good dataset description paper, but indeed is always subject to further refinement.

This leads to the second significant challenge. Based on these criteria, it is important in the peer-review process that not only the paper but also the dataset itself is scrutinised. While it is possible to define upfront some high-level criteria that the dataset must meet, deciding that a particular dataset is, e.g., modelled appropriately, free of technical errors, provides high-quality links, uses existing vocabulary in a suitable manner, etc., requires finding reviewers with strong experience in Linked Data publishing, including a variety of technical issues such as HTTP content negotiation, RDF modelling, datatypes, awareness of popular vocabularies that could be reused, datasets that could be linked to, the semantics of RDFS and OWL, the SPARQL protocol and query language, representing taxonomies in SKOS, vocabularies such as VoID used to describe a dataset, services such as Datahub used to publicise a dataset, and so on. Even aside from the current state of the art, many best practices in terms of how Linked Datasets should be published are still in formation.

Furthermore, aside from technical issues, reviewers must also judge the usefulness of a dataset; however, such datasets can be in areas unfamiliar to the reviewer in question – they may be historical datasets, geographic datasets, etc. Thus, within such specialised communities, it may be difficult for reviewers to assess how useful the topic of the dataset is, how complete the data are, whether all of the interesting dimensions of the domain are captured or not, and so forth. For this reason, the call was amended (to the version listed above) to put the burden of proof of quality, stability and usefulness on authors, where, e.g., the paper must now demonstrate evidence of third-party use, such as referencing a third-party paper where the dataset is used: previously the ability to demonstrate the potential usefulness of a dataset was sufficient to be considered for acceptance.

Volume of dataset papers

The first 15 dataset papers were published in the original Special Issue, Volume 4(2), 2013. By the end of 2015, 38 papers had been accepted under the dataset description track at SWJ, of which 33 have been published in print up to, and including Volume 7(1), 2016. To put these figures in perspective, 96 non-dataset papers were accepted for publication in the same time frame (starting with Volume 4(3)), of which 61 have been published in print (up to and including Volume 7(1)).4

Table 1 provides the number of such papers published per year, where by In Press we include those papers that are accepted and published online but have not yet appeared in print. In 2013, a single issue – the first special issue – accounted for all 15 dataset papers; taking this as an equal starting point, only one non-dataset issue with 3 papers was counted thereafter in the same year. One may note the comparatively small number of dataset papers that are in press. We cannot be sure exactly why there has been a drop in these papers, but it may be possible that the stricter version of the call (e.g., requiring concrete evidence of third-party dataset use) has reduced the number of recent, eligible submissions.

Open questions

Given that we have had 2–3 years of collecting and publishing dataset papers, and given the relative novelty of such a call, here we wish to offer a retrospective on this track. In particular, we wish to perform some analyses to try to answer two key questions about these papers and, more importantly, the datasets they describe:

Based on these results, in terms of the future of the track, we discuss some lessons learnt in Section 4, before concluding in Section 5.

Papers with inadequate links

In some cases, it was difficult to find links to a Datahub entry, a Linked Data IRI, or a SPARQL endpoint (where available) in the papers. In other cases, the links provided were out-of-date although resources were available elsewhere. In other cases still, the links provided were dead. In terms of links, the current call simply requires a “URL”.

One possibility to counteract the lack of links is to make a link to a Datahub entry mandatory, and links to at least 2 or 3 Linked Data IRIs mandatory. Likewise any other resources described in the paper, such as a SPARQL endpoint, a VoID description, a vocabulary, a dump, and so forth., must be explicitly linked from the paper. All such links should be clearly listed in a single table.

We could also ask that all corresponding links be provided on the Datahub entry; this would allow locations to be updated after the paper is published (if they change) and allow the public to discover the dataset more easily. If accepted, we could also encourage authors to add a link from Datahub to the description paper.

Datasets with technical issues

We found a number of datasets with core technical issues in terms of how the Linked Dataset is published: how entities are named, how IRIs are dereferenced, how the SPARQL endpoint is hosted, and so on. The current call lists the quality of the dataset as an important criterion for acceptance, but as mentioned in the introduction, finding reviewers able and willing to review a dataset for basic technical errors – let alone more subjective notions of quality, such as succinctness of the model, re-use of vocabulary, etc. – is difficult. Currently the call requires authors to provide evidence of quality; however, it is not clear what sort of evidence reviewers could or should expect.

One possibility is to put in place a checklist of tests that a dataset should pass. For example, in Footnote 7, we mentioned two validators for Linked Data IRIs that we used in these papers to help find various technical errors in how the datasets surveyed were published; unfortunately however, both of these tools were themselves sometimes problematic, throwing uninformative bugs, not supporting non-RDF/XML formats, and so on. An interesting (and perhaps more general) exercise, then, would be to consider what are the basic checks that a Linked Dataset should meet to be considered a “Linked Dataset” and to design systems that help authors and reviewers to quickly evaluate these datasets.

Datasets with poor potential impact

Some datasets are perhaps too narrow, too small, of too low quality, etc., to ever have notable impact, but it may be difficult for reviewers to assess the potential impact of a dataset when its topic falls outside their area of expertise.

Recently, the call was amended such that authors need to provide evidence of third-party use of the dataset before it can be accepted. As such, this sets a non-trivial threshold for the impact of the dataset: it must have at least one documented use by a third-party. Judging by the relatively fewer dataset papers in press at the moment, it is possible that this criterion has reduced the number of submissions we have received; in fact, many dataset papers have recently been rejected in a pre-screening phase. Thus, for the moment, we feel that this modification for the call suffices to address this issue (for now at least).

Short-lived or unstable datasets

We encountered a notable number of papers describing datasets where nothing is available online any longer, not even the website. Other papers offer links to services, such as SPARQL endpoints, that are only available intermittently. Of course, there are a number of factors making the stable hosting of a dataset difficult: many such datasets are hosted by universities on a non-profit basis, researchers may move away, projects may end, a dataset may go offline and not be noticed for some time, etc. Sometimes, nobody, including the dataset maintainers, can anticipate downtimes or periods of instability. At the same time, we wish to avoid the Linked Dataset description track being a cheap way to publish a journal paper: we wish to avoid the (possibly hypothetical) situation of authors putting some data online in RDF, writing a quick journal paper, getting it accepted, then forgetting about the dataset.

Now that papers in this track need to demonstrate third-party usage, we would expect this to naturally increase the threshold for acceptance and to encourage publications describing datasets where the authors are serious about the dataset and about seeing it adopted in practice. Likewise the call requires authors to provide evidence about the stability of their dataset; though what sort of evidence is not specified, it could, e.g., include statistics from SPARQLES about the historical availability of a relevant SPARQL endpoint, etc.

Summary

On the one hand, we need to keep the burden on reviewers manageable and allow them to apply their own intuition and judgement to individual cases; providing them long lists of mandatory criteria to check would likely frustrate reviewers and make it unlikely for them to volunteer. Likewise, we wish to keep some flexibility to ensure that we do not create criteria that rule out otherwise interesting datasets for potentially pedantic reasons.

On the other hand, to avoid accepting papers describing datasets with the aforementioned issues, we may need to (further) tighten the call and provide more concrete details on the sorts of contributions we wish to see. The results from this paper have certainly helped us gain some experiences that we will use to refine the call along the lines previously discussed.