Genuine semantic publishing

1. Introduction

Many scholars have pointed out that the classical way of publishing scientific articles is ill-suited to deal with the rapid growth of both, volume and complexity, of scientific contributions [4,39]. To overcome these problems, next generation scientific publishing [15] has to respond to the increasing importance of datasets and software, and needs to provide methods to automatically organize reported scientific findings. Perhaps the most important shortcoming of the current publication system is that scientific papers do not come with formal semantics that could be processed, aggregated, and interpreted in an automated fashion.

Semantic publishing [18,44,45] is a general approach to tackle this problem of scholarly communication by using the concepts and tools of the Semantic Web and related fields. This idea was basically born together with the idea of the Semantic Web itself. In 2001, Tim Berners-Lee and James Hendler sketched how they expect researchers in the future to produce machine-readable descriptions of their experiments and findings, in the form of mark-up of their research papers or as independent representations made public on the web [7]. Unfortunately, subsequent work has deviated from this general proposal.

The topic of semantic publishing has received considerable attention during the last few years, most prominently in events that carry the term in their names, specifically the workshop series on Semantic Publishing (SePublica)1

Semantic publishing has been defined as “anything that enhances the meaning of a published journal article, facilitates its automated discovery, enables its linking to semantically related articles, provides access to data within the article in actionable form, or facilitates integration of data between papers” [44], and this definition accurately reflects how the term “semantic publishing” has been used in the recent literature. We argue here, however, that this definition is in one way too restrictive and in another way too inclusive if we want to be faithful to the literal and intuitive meaning of the term and if we aim to follow the spirit of the Semantic Web vision.

In our view, the definition above is too restrictive because semantic publications according to this definition are required to accompany a “journal article” or a “paper.” An entity that only contains a semantic representation of a scientific result, without an accompanying narrative article, could not be considered a semantic publication. On the other hand the definition is too inclusive, in our view, because it covers very shallow approaches that add little – if anything – to established approaches of publishing. For example, letting authors choose keywords from standardized vocabularies for their paper – as many journals do – in fact “enables its linking to semantically related articles,” and therefore by the definition above makes it a semantic publication. As another example, a semantic annotation performed by a third party on an article “enhances the meaning of a published journal article” and therefore would have to be called a semantic publication, even if the semantic annotation is not even made public. In general, the existing literature seems to interpret the term “semantic publishing” as “adding semantics to something that is published” instead of the more intuitive readings of “publishing something that is semantic” or “publishing in a semantic manner.” (We are using the word semantic in its narrow technical sense of carrying a formal logic-based representation of the content’s meaning.)

We argue here for a more intuitive definition of semantic publishing that is broader in the sense that no narrative article needs to be present, and that is at the same time narrower in the sense that the semantic representation has to be a first-class object created and published by the authors. We propose the definition that genuine semantic publishing occurs when somebody publishes a work that includes authentic and fine-grained representations of its content in a shared and interoperable semantic notation, where these semantic representations have essential coverage and are a primary component of the published entity. We explain in more detail below what we mean by authentic, fine-grained, essential coverage, and primary component.

OPEN IN VIEWER

Fig. 1.

The concept of genuine semantic publishing compared to what has been called semantic publishing, explained by an analogy where scientific papers are represented by boxes and formal semantics by flowers.

Figure 1 illustrates our point with a simple analogy. Classical papers are shown on the left hand side as boxes that are closed and hard to access for automated techniques. Existing approaches to what has been called semantic publishing merely adorn this box with formal semantics – represented by flowers in the picture – but leave it closed. This adornment is very useful, to be sure, but it does not reach to the main content of the box. By only looking at the formal semantics, one can possibly find out the topic of the paper but not its main message. Moreover, the adornment is often attached at a later point, after the box has been shipped so to say, and is therefore not a proper part of it. Speaking in terms of this metaphorical image, we argue that we should open the box and let semantics bloom right from the inside. We should represent the paper’s main message with formal semantics. As we see on the right hand side of the figure, this metaphorically turns the box into a flowerpot. Now semantics is the main content, and the scientific paper has become a container for semantics instead of a closed box with a secondary usage as a pin board for semantic annotations.

In the existing literature, we often encounter the implicit assumption that the semantic representation of knowledge has to start from a textual representation, and therefore writing a statement down in natural language always needs to be the first step. For example, we can read in a paper on semantic publishing that “learning how the brain creates and decodes meaning from text is essential if we are to provide better tools for scientific inquiry” and that we need to “train computers to help us read scientific text” [18]. While these are certainly interesting problems, it is not obvious why they are essential if we take the approach of semantic publishing literally, i.e. if we ensure that the published artifacts come with semantic representations from the start. There is no law of nature that research findings can only be formalized after they have been expressed in a narrative text. It can very well be the other way round, such as a researcher writing a narrative text verbalizing existing formal statements she has come up with. More likely, these two will go hand in hand in an iterative process, much like manuscripts and their content typically being shaped through several rounds of revisions. It has in fact been argued – convincingly in our opinion – that this iterative process of scientific writing contributes in an important way to scientific understanding and discovery [26], and therefore it seems beneficial for the semantic representations to participate in these iterations from the start, and not to come into play only at the point where the text is already finalized. However, many articles in the area of semantic publishing seem to make this implicit text-first assumption, as exemplified by papers presented at semantic publishing workshops claiming that “annotations on all levels pave the way for shared knowledge understanding” [46] and that “semantic publishing [...] can be defined as the activity of enhancing a document” [41], among many others (e.g. [17,22,32]). The entire approach of semantic annotation is based on this text-first paradigm, which is a perfectly valid approach but is surely not the only possibility.

We get a similar picture if we look at the Semantic Publishing Challenge held at the Semantic Web conferences ESWC from 2014 until 2017 [20,30,49]. There were three tasks defined for each of these three challenges, but none of them actually deals with publishing. Instead they are about automatically extracting and interlinking semantic data from existing publications. Only the “in-use” task of the first challenge was general enough to not exclude publishing (“showcase the potential of Semantic Web technology for enhancing and assessing the quality of scientific production”), but it did not specifically mention the publishing process either. Unsurprisingly then, the approaches presented at these challenges deal with extraction from and annotation of articles that are already written and published (with the only exception being a paper introducing a publishing platform for Research Objects [36]).

To be clear, we do not mean to deny the value or importance of this body of existing work. To the contrary, these approaches are highly valuable to deal with the wealth of existing publications and those that will become available in the near future. Besides this important work, however, we also need clear and bold visions for the future on how we can improve the form in which such publications are created in the first place.

Despite the prevalence of approaches that deal exclusively with semantic annotations and semantic interlinking of already published articles, there are a number of existing approaches where the artifacts to be published include from the start semantic representations that originate from the researchers themselves. They cover different aspects that we consider important for genuine semantic publishing. These approaches include Research Objects, executable papers, scholarly HTML, Structured Digital Abstracts, Micropublications, and Nanopublications.

Research Objects2

From a different angle, a number of approaches have been proposed for what has been called executable papers (e.g. [27,35,50]), in particular in response to the Executable Paper Grand Challenge3

Yet another angle, with a focus on scientific writing instead of source code and datasets, is taken by various approaches on scholarly HTML, including the work of a W3C community group with that name5

Structured Digital Abstracts [11,43], having been first proposed ten years ago, are probably the oldest approach of this kind. Their basic idea is to require for articles to come with a machine-interpretable summary of the main claims, besides the classical abstract for human readers. They proposed to let authors themselves capture the claims of their own scientific contributions, such as a newly discovered protein-protein interaction, in a notation with formal semantics. Even though these abstracts are attached to narrative articles, the formally represented findings can be processed and interpreted independently from the narrative text. We proposed a similar approach in previous work with abstracts in controlled natural language [28].

Micropublications [16] are a further approach, which puts the emphasis on the structure and interrelation of scientific arguments and their underlying pieces of evidence. The authors stress that the network of arguments is an essential part of science, of which claims and hypotheses are necessary but not sufficient ingredients. They argue that formal representations of scientific claims are often not practically feasible, whereas the structure among them can be captured more easily and is moreover more important and more valuable to help scientists with computer-aided knowledge management.

Nanopublications, finally, are an approach to use the RDF language to represent “the smallest unit of publication” [34] or “core scientific statements with associated context” [25]. This statement-level approach is therefore at a more granular level than most other approaches whose unit of publication is at the article level. Nanopublications consist of three parts, each represented in RDF: an assertion containing the actual content in the form of an atomic small piece of knowledge (e.g. a scientific claim, or a data entry), a provenance part containing metadata about the origin and context of the assertion (e.g. how it was measured), and a publication information part with metadata about the nanopublication as a whole (e.g. when and by whom it was created). Even though the details of their integration into the scientific publishing workflow have remained largely unspecified, nanopublications have received considerable attention during the last few years, with several large dataset having been published in this format [2,12,40].

Table 1 summarizes the different types of approaches introduced above, showing that they cover different types of semantic representations. We also observe from the table that none of the existing approaches covers all aspects, but a complete coverage could be achieved by a combination of them. Executable papers, Structured Digital Abstracts, and Micropublications stick to the article as their unit of publication, whereas Research Objects operate at a higher level (at what we call the project level) and nanopublications at a lower one (statement level). All these approaches, except some flavors of scholarly HTML, mandate that formal semantic data are part of the published entity, but three out of the five approaches also require or assume that a narrative text accompanies the data. Not coincidentally, these are also the approaches that work on the article level, and therefore stick to the classical unit of publication. While none of the presented approaches has yet managed to find widespread acceptance, small practical and less intrusive steps have already been successfully implemented, such as the use of unambiguous references to biomedical resources in the form of Research Resource Identifiers (RRIDs) [3].

To summarize, there are a number of existing approaches that cover important aspects of what we think deserves the term semantic publishing under its intuitive interpretation, but they are normally overshadowed by the more conventional approaches of semantic annotation and semantic interlinking introduced above. These conventional approaches, in fact, advocate the term semantic publishing more visibly, whereas approaches like Research Objects often do not use the term at all. This situation can lead to a general impression that the conventional approaches are the best we can do to apply the ideas of the Semantic Web to scientific publishing. It is our goal to convince the readers of this paper otherwise. While none of the existing approaches covers all aspects that we consider important for genuine semantic publishing, they provide – as we will show below – important and highly valuable building blocks towards that endeavor.

To conclude our discussion of related work, we would like to point out that a large number of general technologies have been developed in the last years that can serve as the basis for approaches on genuine semantic publishing. They come in the form of data formats, ontologies, and software tools. The Semantic Publishing and Referencing Ontologies (SPAR)6

As we have shown above, most approaches that go under the label semantic publishing are not actually about publishing, and the approaches that do target the publication of semantic representations cover different aspects thereof that only partly overlap. We therefore think that there is a need for clear criteria of genuine semantic publishing that include concerns about the authoritativeness and essential coverage of semantic representations, as well as their status in relation to narrative articles and their granularity level.

The first aspect we would like to discuss here is what we call essential coverage of semantic representations with respect to the entity to be published. A representation has essential coverage with respect to a work if it covers (at least) the essence of the work. The essence of a work is its main message, which for scientific articles normally consists of the main claims, findings, and arguments. A semantic representation may not cover all aspects discussed and described throughout a scientific work, but for it to have essential coverage it has to cover the main points: If you had to summarize a paper in one sentence, the content of this sentence has to be present in the semantic representation too. One can also see it as a kind of democratization process of making automated agents first-class citizens: English-speaking agents (e.g. human researchers) get the main content of the work in their English-based representation (i.e. the narrative text); so RDF-speaking agents (e.g. Linked Data aware software) should also get the main content of the work in their RDF-based representation. As we will see below, this perspective aligns very well with the well-established web technique of content negotiation.

Another important aspect is the authoritativeness of the source of the semantic representations, which determines their authenticity. Semantic representations can only be considered authentic if they originate from an agent that is authoritative in the given situation. In the case of the publication of a scientific result, the only authoritative source are the researchers (who are called authors in this context). Semantic representations of scientific results are only authentic if they are provided by the researchers themselves, and this relation can be made explicit with a precise provenance representation. It has long been known in the area of knowledge engineering that the process of formalizing expert knowledge is not merely a process of “transferring” or “converting” knowledge from existing representations inside the heads of experts to formal representations of a form that can be stored in a knowledge base. Rather it has to be seen as a creative modeling process [14,48] where formal structures are generated that existed only in an incomplete, implicit, and unconscious form in the experts’ heads. Explaining a result in a narrative is simpler than formally modeling it, in the sense that natural language allows the writer to remain vague and even ambiguous. Accurately modeling knowledge only from such a narrative text with its inherent vagueness and ambiguity is therefore in general not just difficult but strictly impossible without a further connection to the authoritative source. Genuine semantic publishing requires the authors of scientific results to perform the modeling task themselves, because they are – by definition – the only authoritative source. We claim that – contrary to many existing approaches – we should not try to relieve the authors of this burden (though we should of course try to help them). Otherwise, the semantic representations cannot be considered authoritative and should therefore not be considered part of the publication’s content (unless the person who produces the formal representations becomes a co-author). As Tim Berners-Lee and James Hendler made clear when the Semantic Web was just about to come into existence as a research field, it “involves asking people to make some extra effort” [7]. In the case of scientific publishing, it involves asking authors to make the extra effort of providing formal semantic representations of their findings.

To make the semantic representations first-class citizens, they furthermore need to have an existence in their own right. We cannot call something a genuine semantic publication if the semantic representations are attached to an already published article at a later point, or if they can only be interpreted in the context of the narrative article. Neither should these semantic representations be considered just another type of supplementary material, listed somewhere at the very end of the article as a noncommittal extra file. In fact, one of the defining properties and one of the big advantages of declarative and monotonic semantic notations like RDF is that statements are in an important sense self-explanatory and independent. Such a formal statement can be taken out of its context and stripped from natural language explanations attached to it, and it still means exactly the same thing, as far as the formal semantics are concerned.

In turn, this self-explanatory and independent nature allows for publications of semantic representation to be very light-weight and fine-grained. More so than narrative texts, formal representations with declarative and monotonic semantics can be easily broken down into independent pieces, and therefore we should allow people to exploit this nice property. Such light-weight semantic publications might consists of just a single statement (like “X is related to Y”), and for larger chunks of semantic representations we should make it possible to refer to such individual statements in a fine-grained way (e.g. refer explicitly to the statement “A causes B” within a larger set of statements).

Based on these arguments, we define that genuine semantic publishing needs to comply with the following criteria:

A scientific work needs to come with formal representations that are semantic, in the sense that they are not just machine processable but machine interpretable, and that are linked so they add to the existing formal body of knowledge.

Here, we should briefly discuss an aspect that we deliberately left out of our criteria. Several of the related approaches introduced above (in particular executable papers and scholarly HTML) have a specific focus on how semantic representations can enhance the user experience in the form of interactivity. While we think such interactivity can be highly valuable, we argue for a clear distinction between publication and use, where interactivity belongs to the latter. It is precisely the benefit of formal semantic representations that they facilitate all kinds of subsequent (interactive) use but are agnostic about the precise circumstances and technology. Genuine semantic publications may therefore come with specific interactive features, but it is not appropriate to make that a strict requirement.

Furthermore, it is probably helpful to briefly discuss and illustrate what types of claims a scientific work can make. A large part of the body of scientific work deals with what has been called “normal” or “puzzle-solving” science [29]. In this type of science, known kinds of relations and properties are discovered for objects of known kinds, such as a statement that a given mutation of a given gene can be the cause for a given disease. Such types of statements are relatively straightforward to formalize, for example by connecting a concept identifier for the given gene mutation with the concept identifier for the given disease by the use of a relation denoting the causal relationship, possibly augmented with the needed qualifications and contexts (such as the species to which it applies). In a next step, such a statement as a whole can be formally linked to its authors and to the study from which they derived it (such as a clinical trial and its properties). If the authors represent these formula in a specific language like RDF (assuming existing established vocabularies cover all needed terms), save them in a file, and share and archive them on the web, then we have perfect case of a genuine semantic publication. The authors may want to add a narrative to it, but they do not need to, as the semantic representation speaks for itself. More disruptive and more abstract kinds of scientific contributions involve the criticism of existing concepts or arguments, and the advocation of new ones. In the most extreme case, this can consist of proposing a paradigm shift that can lead to a scientific revolution [29]. By their nature, these types of contributions are harder to formalize, but it is always possible to at least make the action of criticizing or advocating explicit and to position the objects in the space of related concepts, arguments, or paradigms.

Finally, before we move on to demonstrate in detail how advocating a new concept can be achieved with a genuine semantic publication, let us reflect for a moment on the potential impact of such a proposal. The machine-interpretability of publications’ main claims entails that software could automatically connect, aggregate, and reason about the body of published scientific work. For example, we could automatically answer complex questions or produce interactive science maps, not only at the meta-level of papers, authors and their relations, but also on the domain level of tangible and abstract concepts and objects of study. This will allow scientists (and others) to acquire a more accurate and more complete picture of the current state of science with much less effort, which in turn can accelerate scientific work and improve its quality. The support for small fine-grained publications can further speed up scientific discovery, as researchers no longer need to wait for a larger body of work to assemble, but can publish smaller findings as they come in. Results from such software solutions will never be error-free, but due to our authenticity requirement we can find out which authors are to blame for mistakes we find in the semantic representations, instead of some anonymous software component or human annotator. This in turn can put strong incentives on authors to provide good formal representations for their works. It is hard to foresee how all the involved technical – let alone social and institutional – aspects would unfold, but it is not hard to imagine that such technology could have a profound positive impact on the communication of science.

5. Genuine semantic publishing in action

It turns out that all the technologies needed for applying genuine semantic publishing are already available and most of them are very mature and reliable. There are no technical obstacles preventing us from releasing our results from today on as genuine semantic publications, even though more work is needed on ontologies that cover all relevant aspects and areas and on nice and intuitive end-user interfaces to make this process as easy as possible.

The paper that you are reading is in fact a genuine semantic publication. It has different representations for different types of usage. You might be reading these lines while sitting on a beach and reading from a sheet of paper printed from the article’s PDF version, or you might be reading it in your office from a web page in HTML format within your browser window. In either case, these representations contain the narrative text, which we carefully wrote to explain and motivate our ideas to human readers. But we also make our work available to software agents, for which we have different representations that consist of formal RDF statements instead of narrative text. Importantly, these RDF statements convey the same main message as the narrative text: They are different representations of the same work.

To formally represent the main content of the paper, we can make use of existing ontologies and vocabularies, such as CiTO [37] and SKOS [33]. Specifically, our paper’s main message is the advocacy of the new concept of genuine semantic publishing, which can be expressed as follows in the Turtle RDF notation [6]:

There is to our knowledge no existing ontology that would exactly capture the relation of a publication advocating a given concept, but the combination of the two relations describes and supports from the CiTO ontology comes close. We as authors should of course say a bit more about this new concept, most importantly that it is related to the existing concept of semantic publishing:

And we can express our critical position on that concept:

Next we can formally represent the five criteria based on which we define our new concept:

We can try to capture part of the content of these criteria in RDF as well, but at some point we have to stop and be content with an informal description in natural language (at the latest when we hit the symbol grounding problem). However, we believe that it is always possible to build a formal representation of the main content at the highest level, such as introducing and advocating a new concept, even though we will mostly not be able to provide a complete formal definition. In this sense, such a representation is underspecified but has essential coverage.

We would like to note here that – while we are confident in declaring that our own representation complies with our criteria – we do not intend to claim that it achieves them to the highest degree possible. It is, to the contrary, still a quite crude representation that leaves many details and aspects of our main claims and arguments untouched. For example, we state that our paper critiques the concept of semantic publishing, but we do not say why and in what way, namely that we claim its interpretation to be not intuitive and not visionary. We are not aware of any ontology that would allow us to express this, and we restricted ourselves for this demonstration to existing resources. More work will be needed on establishing such ontologies and best practices to facilitate more precise and more inclusive formal models of scientific findings and arguments, but the currently existing vocabularies already allow – at least in our case – to achieve a basic level of genuine semantic publishing.

In any case, the benefits of such a representation of the main message of a paper might not seem obvious at this point. One of the main advantages comes when subsequent work starts referring to these formal representations. As a fictitious example, a subsequent paper might propose the concept of “advanced semantic publishing” that includes our criteria 1 to 4, but criticizes number 5 and suggests to replace it with a different one:

This example shows how we can formally capture the high-level relation of papers’ content, and thereby place them in the wider context of the literature on the respective topic.

The above RDF representations are interpretable by machines, and thereby automated software agents of all sorts can read and process them. Human readers, of course, normally prefer a natural text representation of a paper’s content. To account for such different demands, resources on the web can in general have different equivalent representations for different types of agents. Content negotiation can then be used in the background to find a suitable representation based on the agent’s request (mediated by the browser) and the available representation formats on the server side. Alternatively, we can use special kinds of hyperlinks on a landing page to achieve the same effect within HTML. We will use this landing page approach here for demonstration purposes because it makes the different representations more explicit, but the presence of a landing page is not required.

Such a landing page links to the different (classical and semantic) representations of the work. With just a few lines of HTML code, we can define a canonical URL and some minimal metadata, such as title and authors of the work (more metadata is available in the actual representations):

And then we can link to different representations of the content of the given work:

Specifically, we link to the PDF version of this work, two flavors of HTML (Dokieli and RASH), and RDF representations in Turtle (without provenance information and metadata) and TriG (with provenance information and metadata in the form of nanopublications), thereby also showcasing how existing technologies can contribute to achieve genuine semantic publishing.

OPEN IN VIEWER

Fig. 2.

The landing page pointing to different versions of the work.

Figure 2 shows what such a minimal landing page looks like in a browser, and the respective data can be found online9

To illustrate the last criterion of being fine-grained and light-weight, let us assume that somebody wanted to add at a later point just a single triple to assert the connection between our first criterion and the concept of Linked Data:

We can save this triple in a file and create a bare minimum landing page that could look as follows:

Together, these two files, containing fewer than 500 bytes, form a complete publication according to our criteria. This demonstrates that fine-grained contributions down to single triples can be published in a very light-weight manner with an overhead of just a few hundred bytes.

The downsides and limitations of the current scientific publishing paradigm have become apparent in many ways, from the researchers unable to deal with the avalanche of new papers published in their fields to the struggles of elevating scientific datasets to the level of appreciation they deserve. We argue that we need both, grand visions and small practical steps, to move forward and advance science communication, to make sure that the benefits of future breakthroughs are not offset by our inefficiency in communicating them.

We have to make sure, however, that we do not confuse our grand vision with the small practical steps towards it. Semantic publishing was once a grand vision but the term was then hijacked by approaches implementing small practical steps. These small steps are certainly important, but they also made us lose sight of the longer-term vision.

In this position paper, we aimed to focus again on the grand vision, which we propose to call genuine semantic publishing. We argued that genuine semantic publications should not only come with representations that are machine interpretable, but that these representations also need to have essential coverage of the work’s main claims, that they need to be authentic and approved by the authors, that they should form a primary component of the work, and that they should allow for fine-grained and light-weight contributions.

By explaining how this very paper was written as a genuine semantic publication, we demonstrated that – as far as technology is concerned – the vision is not that grand after all. Technically, genuine semantic publications are at a basic level already feasible nowadays with established and mature technologies. But many grand challenges remain, including the development and deployment of stable overarching formal models that include aspects such as evidence and arguments, reliable domain ontologies for the various still under-resourced fields, intuitive user interfaces, data publishing infrastructures, methods for attribution and recognition of scientific efforts, and effective incentive structures. All these challenges can only be addressed, however, with a clear vision of how scientific publishing should develop in the future.