Data and metadata once hidden in dusty card cabinets of thousands of galleries, libraries, archives and museums worldwide are now available online in digital formats. An incredible explosion of metadata has been expanding in the quantity of digitized data, the richness and sophistication of data models, the number of institutions and private citizens that contribute and their interconnection. A fundamental issue, however, limits this undeniable success: current data models force the expression of a single point of view. For example, the eld author is either set to a value or to another one. Any disagreement about the content of a eld is resolved before the publication of the data and forever lost. Yet, we argue, the expression of di erent and contrasting points of views is a keystone of scholarship, as well as one of the most engaging aspects for everyone. Bowdlerized, sterile, con ict-free data records fails to capture the core of important scholarly debates and thus fails to attract the interest of the general public. The root cause of this issue is technical rather than cultural: current standards for data models (e.g. RDF, OWL) do not simply support the expression of contrasting statements. In this paper we propose both a methodological approach to address this problem, and a proof of concept of how this problem could be fully and cleanly overcome with modest extensions to the existing standards. We name this approach \contexts and conjectures".
Why do Galleries, Libraries, Archives and Museums digitize collections? We can list a few reasons: preservation of knowledge, i.e., associating to physical or cultural artefacts what is known about them, should they get destroyed or forgotten about; interlinking, i.e., historically situating these artefact by showing the connections between them; circulation of knowledge, i.e., allowing larger audiences to gain access to information about cultural artefacts that otherwise (because of fragility, physical distance, obscurity) would be hard or impossible to reach.
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Yet, we notice, we are failing on all counts. We fail at preserving knowledge because an extremely limited quantity of information ends up being associated with our physical and cultural artefacts; we fail at interlinking collections because technological constraints force us to choose between consistent or connected datasets; and, most importantly, we fail at circulating knowledge because we cannot interest a substantially large audience to our digital collections.
We believe that all these failures stem from a single cause: a (doomed ) search for objectivity, forced by digital models that mostly ignore the derivative and strati ed nature of cultural objects, and allow for only one point of view to be expressed, leading to unbalanced and incomplete descriptions of our artefacts.
This eliminates debates and disagreements, exactly what has a chance to create and maintain interest in scholars and general audiences alike.
We rather believe that we should explicitly aim at representing competing points of view and opinions, and make sure that we fully document their existence, their strengths and the ideas behind them, so that our audiences can
nally perceive representations that are truer and more interesting than the sterilized and boring renditions forced by so-called objectivity.
In many disciplines of the humanities, truths are hard to come by and facts are rare. In most cases, we use words such as facts and truths just to mean \statements for which there is an acceptable trail of supporting sources", or \statements that are more or less accepted by the majority of the relevant scholars" or even \statements that so far have not been disproven".
This is not a surprise, as precise knowledge in many cases is impossible or outside our reach. Yet, scholars can easily deal with incomplete or conjectural knowledge, and even when they are personally convinced of the truthfulness of some information, they are aware and able to accept that di erent viewpoints, dissent and speculations may exist about them.
Unfortunately, the ability of contemplating and handling di erent or opposing interpretations over the same piece of information is not embodied in the software and in the digital data structures that we use to represent them: single points of view are what we expect in our digital collections. The inevitable con
icts and disagreements cannot be expressed explicitly, and need to be simpli ed, resolved and eliminated before committing information to the digital realm.
Yet, unknowns, disputes and dissenting opinions are often what in the rst place attracted, fascinated and keeps fascinating scholars. Neutered, indisputable and unequivocal digital data have the twin problems of a) being a poor representation of what we know and think and b) being fundamentally boring and unable to attract anyone, especially laypeople and younger students.
Our approach to making datasets able to express the richness of the nondigital scholarly debates is composed of two parts: contexts (Section 3) to allow the expression of con icting opinions on an artefact, and conjectural graphs (Section 4) to encode scholarly statements as quali ed arguments, rather than as absolute truths, and a mechanism, collapse to reality, to turn some conjectural graphs into plain RDF statements that can be interlinked and get them to interoperate with other RDF-based datasets.
Consider the example in Fig. 1, taken from a record in Europeana, coming from
the Bildarchiv Foto Marburg [
The existing record3 is accompanied by metadata stating that the item being described is a 53,9 x 41,3 cm print by G.B. Piranesi, dated 1756 (and/or 1787), titled \Veduta dell'An teatro Flavio detto il Colosseo", with subject James Caul eld and King Gustaf III of Sweden. Another (now deleted) record described the (very same) item as the Colosseum, a 70-80 A.D. building by Vespasianus, and, in a plain text description, represented via a 1960/70 photo by Konrad Helbig of a print by G.B. Piranesi (see Fig. 2).
These records are both incomplete and wrong. They contain, at the same time, objective and factual information (e.g., the size of 53,9 x 41,3 cm), some 3 https://www.europeana.eu/en/item/2064137/Museu ProvidedCHO Bildarchiv Foto Marburg obj20089555 T 001 T 071 The image was present in 2012 in two separate and dissimilar records in Europeana, and is now (2020) present in only one, the second having been removed probably due to being a duplicate and because of the number of issues in its description. reasonable interpretations (e.g., the author) and many factual errors (e.g., James Caul eld and King Gustav III are clearly not the subject of the image, this item is not a building and its creation date is not 70-80 A.D.).
So, the facts are few and actually constitute the most boring and uninspiring part of the data items we are shown, and many details, although wrongly characterized, paint a much more complex and interesting story, once we take it upon ourselves to understand and explain them.
We have a better story: this is a black/white photograph from a retrospective of the works by Konrad Helbig (1917-86), a famous German art photographer, known for his groundbreaking photographs of Italian classical works of art, and of tanned, underage, naked Sicilian boys. The photograph depicts a print of the Colosseum by Giovan Battista Piranesi (1720-78) taken from \Antichita romane", a four-volume book of high quality etchings of various sceneries of Rome and its countryside, a bestseller of the XVIII Century, often brought back home by the young nobles and high society bourgeois upon their return from their \Grand Tour" in Italy.
The rst edition of the book (1756) was dedicated to James Caulfeild, Lord Charlemont (1728{1799). In 1756 Gustav III (1746{1792) was 10 years old, not a king yet, and likely uninterested in the arts. He was King, on the other hand, after Piranesi's death, when his son Francesco (1758-1810) made Gustav III the new patron of the publishing and dedicated the new edition (1787) to him. This explains both the two dates and the double dedication of the volume.
This may be a better story, but can we now expect the Europeana record
to be xed accordingly? Uhm, well. First, ours is just one of many possible
explanations: it should not be promoted to truth only because it is richer or more
plausible than the current one. Second, and most importantly, this explanation
deals with many di erent levels of abstraction of the artefact (the chain of
reproductions, the physicality of these reproductions, the people associated with
them, etc.), and the data model employed by Europeana [
The example in the previous section has shown various typical problems in representing metadata (digitally or otherwise) about our cultural heritage. We like to classify these issues in a few broad categories: reticence: information that was well known was not recorded, due to haste, lack of skill, or, more probably, lack of policies and software support. The role of Helbig, the relationship between the table and whole book are missing, etc. attening: the fact that an artefact depicts another artefact is never made explicit, and here it actually occurs twice (a photograph of an etching of a building). A single set of metadata describes the image, the physical support, and the chain of entities depicted, as an inconsistent and meaningless record. coercion: important information for which no appropriate eld was found, was forced into inappropriate elds (e.g., the dedication of the whole book to James Caulfeild and Gustav III ends up as the subject of a single table), leaving to the reader the job of making things straight, and forever ba ing any automatic tool tasked with indexing and searching collections by subject. dumping: important information for which no appropriate eld was found, was forced as plain text inside a descriptive eld, easy for humans to read but forever lost to any automatic tool. For instance, the existence of Helbig, the date of the photograph and the exact placement of the table within a fourvolume book are text-only in the Description eld, and never formalized.
In addition to these, there are a few remaining problems we should also consider: not all statements of our story have the same reliability; our statements are clearly authorial (we are expressing our own personal interpretations); and we ourselves have fairly diverse levels of con dence about the events described.
In the following section we propose an approach called contexts. Contexts cohesively address these issues, and their adoption enables data models to express not only ne details about the artefacts, as they do now, but also a wide variety of opinions, points of views and conjectures that constitute the largest part of our knowledge about cultural heritage. 3
Plenty of data models exist in the digital humanities to represent metadata, facts
and information about our cultural heritage. They encode ne details about our
artefacts but cannot easily handle more complex situations like the one presented
above and cannot describe the circumstances within which these details nd their
place, their role, their correctness, their plausibility. The approach we suggest
to overcome these limits is the use of contexts, presented in more details in [
In general, we de ne contexts as sets of statements meant to characterize the metadata about an artefact, rather than the artefact itself. For instance, determining the provenance of the creation date of an ancient vase is not a statement about the vase, but a statement about a statement about the vase.
Contexts provide boundaries to opinions and make them comparable to facts. Identifying and expressing the context of all statements is fundamental for their correct interpretation and use. Without them, we could draw false conclusions. The following are a few contexts that we have identi ed: Temporal relationships: facts and assertions are often constrained inside a temporal interval. For instance, Gustav III was the King of Sweden only between 1771 and 1792; he was alive in 1756, but he was not King then, and hardly the addressee of a dedication from a Roman printer.
evolving concepts and statements referring to them must be quali ed. For example, was Caulfeild Irish? James Caulfeild, 1st Earl of Charlemont (17261799), was a noble of the Kingdom of Ireland, then under the rule of the Crown of England, with little or no direct connection with the current Republic of Ireland. His nationality is a handy simpli cation for a much more nuanced characterization of his true origins.
Part-whole relationships: the item being described in Helbig's photograph is only one page in a four-volume book, itself published in at least two editions.
We should be able to distinguish facts regarding the individual page (the heading of the image, the subject, etc.) from those regarding the volume as a whole (the author, the publication date(s), the dedications, etc.) Object-subject relationships: Object-subject chains can be particularly deep, intricate and fascinating. In our example what we are describing is in fact a JPEG, derived from a high-resolution TIFF scan, of a photograph by Konrad Helbig dated 1956, of an etching dated 1756 by G.B. Piranesi, of a 70-80 a.D. building in Rome. Each of these items deserves descriptive metadata that must be associated with the correct entity.
Provenance: Statements of the metadata come from sources that should be identi ed: an individual, a text, the direct observation of the item, etc. Provenance not only gives backing and traceability to the stated facts, but most importantly it allows multiple di erent and competing statements to coexist, even if in contrast with each other.
Con dence: the sources themselves expose varying degrees of con dence in expressing this or that fact. Recording a con dence level allows for conjectural, hypothetical and even wacky statements to correctly coexist with established and settled information.
The use of contexts provides the objectivity and truthfulness that software needs and computer scientists crave for, without giving up on the richness of information favored by domain scholars. E.g., the statement \the Marburg Foto Archiv organized a retrospective about Konrad Helbig" is clearly speculative, non-objective and conjectural. Adding a little context, the statement \Barabucci et al. (2020) speculate that the Marburg Foto Archiv organized a retrospective about Konrad Helbig" becomes objective and easily veri able. The addition of provenance created a statement that is undeniable and usable in a scienti c discourse.
Contexts allow expressing more than just settled and boring data: their ability to enunciate con icting views allows the creation of a much more interesting and nuanced representation of our cultural past. 3.1
The natural habitat for contexts as presented in the previous sections is in datasets expressed with Semantic Web standards such as RDF4 and OWL5. Expressing metadata using RDF is easy, and most of the existing models used in cultural heritage already have an RDF representation.
RDF is used to express statements about entities. In the RDF model each statement is expressed using a triple composed of a subject, a predicate and an object, the subject being the entity being described. For example: \Antichita Romane" (subject ) has author (predicate) \G.B. Piranesi" (object ).
Modelling contexts means expressing statements whose subject is not the artefact being described, but another statement in RDF that expresses some quality about the artefact. The simplest approach is called rei cation, which 4 Resource Description Format: http://www.w3.org/TR/rdf11-concepts/ 5 OWL 2 Web Ontology Language: http://www.w3.org/TR/owl2-overview/ requires treating the statement itself as an entity, which becomes the subject of the contextual statement. For example, in Fig. 3, :S is a statement that asserts that the book \Antichita Romane" was authored by G.B. Piranesi (1). The statement at the end (2) a rms that statement :S was attributed to GBarabucci, FTomasi and FVitali, thus assigning a clear and unambiguous provenance to it. : S rdf : type rdf : Statement ; rdf : subject :AntichitaRomane ; rdf : predicate dc : author ; rdf : object : GB Pir an es i .
¬ : S prov : w a s A t t r i b u t e d T o [ : GBarabucci , : FTomasi , : FVitali ] .
We cannot, however, simply add these RDF statements to existing RDF collections. These statements would be understood by RDF processors and reasoners as expressing at the same time two related but independent sentences: 1) that \G. B. Piranesi is the author of the book `Antichita Romane'", and 2) that \Barabucci, Tomasi and Vitali assert this". The truth of the rst statement is not under discussion: it is atly asserted. The provenance of this statement is provided in addition to it and is at the same level. Contexts expressed in this way do not restrains the statement of authorship, as we were hoping to obtain.
Extending current data models with the ability to assert statements as true
only within and depending on the truth of a given context is problematic due to
shortcomings of the underlying RDF meta-model. The same issues arise not only
in the Semantic Web, where RDF and OWL are used, but also in traditional
databases, where the ER model [
We could, of course, overcome this limit by introducing ad-hoc interpretation rules in our own metadata processors, e.g., by having them ignore assertions outside of explicitly activated contexts. This approach would work in practice, but would be project-speci c and would prevent the sharing of our data with the rest of the world, e.g. as Linked Open Data: we cannot expect all LoD participants to use our modi ed rules instead of standard RDF reasoners to process this and other RDF datasets.
Or, we could adopt a RDF extension called nested named graphs [
Adopting contexts would make it possible to express di erent opinions while providing the consistency and the truthfulness needed to support the coexistence of heterogeneous point of views. 4
Contexts provide a clear and systematic way to record possibly contrasting statements about the same object from multiple points of views, but they do not address how contrasting statements may be processed, combined and presented by a piece of software.
Our approach to make contextual metadata coexist and work with other RDF-based datasets goes under the name of conjectures. Conjectures are scholarly statements that are formally expressed without being asserted, i.e., without an associated truth value. Statements expressed in conjectures are as in a limbo: their truth state is unknown, and they can, if desired, be subjected to conditions, called guards. Now and then, for example when evaluating a what-if scenario, conjectures can be programmatically turned into usual assertions through a speci c operation called collapse to reality (or collapse to truth).
Multiple statements can be collected in conjectural graphs, a set of statements that may or may not be asserted collectively and that must be either all asserted or none asserted at the same time. When a conjecture is collapsed to reality, the corresponding conjectural graph becomes equivalent to a plain RDF graph.
In the following sections we will illustrate through a running example how conjectures, conjectural sets and contexts t together, discussing basic semantics, notation and compatibility with RDF. The basis of our example are two incompatible conjectures: \Francesca Tomasi states that Giovanni Battista Piranesi is the author of Antichita Romane" and \Fabio Vitali states that Elvis Presley is the author of Antichita romane". 4.1
The core semantics of a conjecture, in contrast to classic RDF statements, is that, by entering a conjecture, we do not assert the corresponding statements, but only assert that 1) these conjectures exist, 2) they are available for examination, and 3) they can be collapsed to reality.
We use the a simple extension to the TriG [
S1 P1 O1 .
S2 P2 O2 . } Cj prov : w a s A t t r i b u t e d T o A .
Conjectural graphs cannot be linearized as RDF graphs without losing their semantic expressivity. Nonetheless, there exists a simple and straightforward representation of conjectural graphs in RDF as named graphs, and it can be generated algorithmically with a very limited procedural e ort.
By construction, the conjecture \Cj" is de ned as equivalent to a new graph \CTXj", called the weak context graph of Cj, in which the predicates in the statements have been replaced by weakened predicates that are only conjecturally equivalent to their original counterparts: CTXj {
S1 PG1 O1 .
S2 PG2 O2 . PG1 isConjecturallyEquivalentTo P1 .
PG2 isConjecturallyEquivalentTo P2 . } CTXj prov : wasAttributedTo A.
The graph is a weak form of the conjecture because instead of using the expected predicates P1 and P2, we use ad hoc predicates PG1 and PG2, which are di erent and independent from P1 and P2, but are actually asserted, and connect the same subjects to the same objects. Weak predicates are associated to their respective strong predicates through the new predicate isConjecturallyEquivalent which provides the semantic closeness and structural weakness that we need to establish the connection between them.
In this way we can assert multiple clashing statements at the same time: conjectures can be expressed side by side and their assertion of truth is absent or postponed.
Our running example can now be expressed as follows: CONJECTURE :T { :AntichitaRomane dc : creator : GBPiranesi . } :T prov : wasAttributedTo : FTomasi .
CONJECTURE :V { :AntichitaRomane dc : creator : EPresley . } :V prov : wasAttributedTo : FVitali .
As said, these conjectural graphs are equivalent to the following weak contextual graphs in plain RDF: :T { :AntichitaRomane T: creator : GBPiranesi .
T: creator conj : isConjecturallyEquivalent dc : creator . } :T prov : wasAttributedTo : FTomasi . :V { } :V prov : wasAttributedTo : FVitali .
:AntichitaRomane V: creator : EPresley .
V: creator conj : isConjecturallyEquivalent dc : creator . where the syntax T:creator and V:creator denotes two weak predicates that live in separate namespaces, di er from each other and di er also from the original Dublin Core dc:creator, but are explicitly represented as conjecturally equivalent to it. 4.2
The process one can use to actually assert the content of conjectural graphs is that of collapse to reality, in which the statements belonging to a conjecture are reasserted in their full and strong form, while leaving other conjectures in their weak, non-asserted form. The collapse does not modify the nature of the conjecture, but re-empowers it. The term refers to the reduction of the superposition of multiple independent (and, possibly, incompatible) conjectures to a single one which is then asserted as true, and it is inspired by a similar concept in quantum mechanics, where the wave function collapse is the e ect of an observation from the external world that reduces multiple states of a quantum system to a single one.
In practical terms, collapsing a graph means programmatically adding extra statements that short-circuit the guarded predicates. They re-represent the guarded predicates as actual facts. This is done either via the introduction of predicates such as owl:equivalentProperty to connect weak and strong predicates, or by re-asserting in their full form the statements expressed in weak form. Since the two approaches are redundant but not reciprocally incompatible, we can do both so as not to depend on OWL tools to obtain the desired result.
Collapses to reality are, themselves, graphs that can be contextualized, for instance with provenance. For example, if we want to express the fact that Gioele Barabucci believes in the attribution by Francesca Tomasi, we would augment \T" with a collapsing predicate and attribute it to him, as follows: : c o l l a p s e O f T prov : w a s A t t r i b u t e d T o : G B a r a b u c c i . : c o l l a p s e O f T conj : co ll ap se s : T . : c o l l a p s e O f T { : G B P i r a n e s i dc : creator :AntichitaRomane .
T : creator owl : e q u i v a l e n t P r o p e r t y dc : creator . }
The system sees and processes the new statement as any other statement, without needing speci c support for conjectural graphs. The collapse to reality is well contextualized and can be made temporary, since we can cleanly undo it by removing the collapsed graph as a switch in what-if scenarios. At the same time, all con icting conjectures are still in place and available for further processing by conjecture-aware tools.
To summarize, the collapsing process is powerful but simple to implement, requiring only the use of well-de ned and contained pre- and post-processing procedures during the ingestion of conjecture-based datasets. This allows conjecturebased datasets to be used together with more classic RDF- and OWL-based datasets currently published as Linked Open Data. 5
Much of the past and current data modelling e orts towards the description of
cultural objects are focused on being able to represent an increasing amount of
details. In [
Although these models expose a complex, multidimensional and interconnected landscape, none of them addresses the issues discussed here, because of their implicit assumption that data description should be neutral and objective.
Challenges to the notion of neutrality of data models can be seen in HiCO6,
where interpretation acts are used to express the provenance of interpretations.
The Mauth tool [
Similarly, ontologies such as PROV-O7, for describing the provenance, those
in the SPAR family8 (FaBiO for publishing data, PRO for managing roles, CiTO
for representing citations) or GAF9, for linking mentions, can all be used to add
pieces of context to data descriptions. Further, the CMV+P document model [
Furthermore, the recent RiC-O10 model uses the similarly named notion of
\contexts" to indicate a plurality of paratextual information that translate the
classical siloed approach into a graph of connections between vocabularies and
ontologies. Additionally, contexts and conjectures can be seen as a generalization
of the Factoid model [
The inability of current technologies such as RDF to express multiple contrasting points of views over the same cultural artefact is causing an impoverishment of the datasets published by GLAMs. In addition, the fact that each collection can state only a single point of view is making it harder (and often impossible) to interlink datasets that use, voluntarily or involuntarily, di erent points of view.
To address this issue, we propose a combination of two approaches. First, contexts: a method to allow the expression of con icting opinions on all aspects of an artefact. Second, conjectural graphs : a way to encode scholarly statements as quali ed arguments, and not as absolute universal truths, as required by RDF.
In addition, we present a mechanism, named collapse to reality, that makes it possible, in a limited and controlled way, to turn speci c conjectural graphs into plain RDF statements that can be linked to and made interoperable with other RDF-based datasets.
In contrast to other techniques, our approach does not require the publishers of datasets to settle on a single point of view upfront. Instead, institutions can publish rich datasets containing multiple interpretations of the same data, leaving to the users or to the scholars the ability to accept certain conjectures while refuting others.
The ability to express con icting opinions, di erent hypotheses and, in general, all that is part of a lively and meaningful scholarly debate, should represent a key requirement for all data formats used to describe cultural artefacts. 7 W3C PROVenance Interchange Ontology: https://www.w3.org/TR/prov-o 8 Semantic Publishing and Referencing Ontologies: http://purl.org/spar 9 Grounded Annotation Framework: http://groundedannotationframework.org 10 ICA Records in Contexts Ontology: https://www.ica.org/standards/RiC/ontology
This is not currently the case. We hope that the proposed contexts and conjectural graphs will provide a productive start point for a theoretical, practical and methodological re ection on how to beyond the current state of the art and address what is probably the biggest challenge that cultural institutions face when it comes to reaching and engaging scholars and the citizens alike.