There is increasing attention in the Digital Humanities for information systems to document and analyze scholarly debates, a case for all being that of criticism in literature or performing arts. To make these systems available, there is a call for formal models to represent alternative perspectives about the same entity (text, music composition, artwork, etc.), even when they do not perfectly align or are explicitly incompatible. Building on previous research, we propose the core module for an ontology to model scholarly perspectives as observations. To enable tasks such as data classification, sharing, and reasoning, the ontology takes the advantages of formal languages such as OWL and SWRL. The ontology is developed with a modular architecture, where the presented module is the most general one for reuse across diferent domains. We also present a case study in literary studies to exemplify the ontology and show its potential usage.
There is increasing attention in the Digital Humanities (DH) for the design of tools addressing both
the hypothetical and partial nature of data in the humanities [
To make sense of observational data produced through research investigation, as well as design information systems for their documentation and analysis, we need first of all models suited for their representation. Ideally, once a model in this direction is developed, diferent mechanisms can be used to compare alternative data about the same entities, thus representing debates with competing and conflicting perspectives. In addition, an approach along these lines could be functional in making artificial agents capable of handling plural and contrasting viewpoints.
A model for the representation of observational data in the DH must be able to face several challenges.
First, considering the broad spectrum of research in the community, the model has to be designed at
diferent levels of generality to be reusable across various application contexts. Second, scholars in the
humanities, even within the same research field, can easily adopt diferent conceptual systems, where
the intended meaning of the terms they use often remains implicit or only vaguely defined. Thus, a
model to represent and reason over observational data must be able to deal with such a plurality of
frameworks and concepts, “operationalizing” them for computational processing [
Building on previous research [
The paper is structured as follows. In Section 2 we present some insights on how we conceive observations. The core module of our ontology is discussed in Section 3, where we also present some methodological principles for its use. Section 4 exemplifies how the ontology can be applied to literary studies, these being currently the main application context for our research. Section 5 compares the ontology with similar eforts in the DH. Finally, Section 6 concludes the paper.
In this section, we discuss some foundational features related to the notion of observation, borrowed
from previous work [
First, observations are statements about how things are categorized on the basis of the sorts of arguments one nfids in literary criticism, or technical procedures like empirical measurements with Carbon-14, just to make some examples. Observations therefore have a representational nature, since they make explicit how things are described: they are pieces of information representing how domain entities are understood, classified, and attributed with certain properties. From this perspective, the modeling of observations allows us to take an epistemological stance on property exemplification, since representing that an entity satisfies a property means that it has been attributed with that property. It follows that observations are not necessarily veridical, that is, they may not correspond to reality. For instance, one may express the observation according to which a person was born on a certain date, but the observation could be incorrect, perhaps due to reliance on unreliable historical documents. In the context of this paper, we present a framework to primarily document observations without establishing criteria for their evaluation. This is because such criteria are domain-dependent and often controversial.
Second, observations can be inter-subjectively accessed, i.e. they are not scholars’ introspective beliefs. This allows us to model observations as pieces of information found in texts, without the need to refer to scholars’ mental states. Also, in our framework, it is possible that diferent scholars express the same observation, even by relying on diferent sources. Accordingly, our proposal keeps track of the provenance of observations while abstracting from both observers and observing acts.
Third, we assume that, within specific contexts, scholars share a common vocabulary for observations, namely, they share an observational language. As we will see, our approach employs a finite set of (basic) observation classes that are organized at diferent levels of generality. Each class groups observations corresponding to the attribution of a property (relation) to one (or several) individuals, e.g., an observation class collecting the attributions of the relation being born between persons and dates.2
Fourth, observations about the same entities can be compatible (at diferent abstraction levels), but they can also be incompatible. For instance, a scholar may observe that a person was born in 1920, whereas another may claim that the same person was born in April 1920. Assuming background knowledge about dating and calendar, these two observations are compatible, and what changes is just their timescale. However, a third scholar may claim that that person was born in February 1921, which is incompatible with the previous observations. As said, the modeling of observations has to allow documenting controversies and incompatibilities among the perspectives of domain experts.
Finally, to better understand our notion of observation, it is worth spending some words on the
philosophical debate on correlated notions. Philosophers sometimes speak of facts as entities that
are part of reality and do not depend on the epistemic states of observers [
In this section we present the structure of our ontology. In particular, we present the core module and some methodological principles to extend it (Section 3.1), and show how it can be used to analyze observations (Section 3.2). An exemplification of the ontology to literary studies is presented in Section 4. For the sake of presentation we will use the syntax of description logic; the OWL files are available in a dedicated open-access repository.3
The purpose of this module is to lay down the core modeling elements of the ontology. The core module consists of two main (disjoint) classes, i.e., Text and Observation, including a series of relations used for their formal characterization. We will introduce here three main (disjoint) subclasses of Observation: () basic observation (BasicObs), () argumentation observation (ArgumentObs), and () source observation (SourceObs). The last two subclasses are fundamental to represent debates and possible disagreements about the attribution of some properties (relations) to entities in the domain. To represent the entities to which an observation applies, we use the general relation hasArgument (an OWL object property), which can be further specialized, as we will see.
Texts. Texts are here intended as abstract sequences of words (characters) in one or more languages. Though we mainly refer to texts written in natural language, the proposal also applies to oral utterances or texts written in formal languages (e.g., databases, logical theories, knowledge bases). We also assume that texts difering even for a single word are not identical, although they can be similarly interpreted and can stand in various (e.g., philological or formal) relationships with each other. A text is abstract in the sense that it is not, e.g., the individual pattern of ink on a specific paper that would be destroyed if the paper were destroyed. The same text can have multiple “realizations”; e.g., the physical copies of Orwell’s Animal Farm that one finds in libraries and bookshops. For the sake of generality, we do not commit to the production context of texts, namely, whether each text is produced by someone at a time. This information can be added to further extensions of our ontology, if needed.
As we will see, texts are fundamental to represent the sources of observations, since each observation is always linked to a text reporting the information encoded by the observation itself. Atomic and Compound Observations. Before discussing the three main subclasses of Observation, we introduce a general parthood relation, defined on the whole domain, represented by the (transitive) object property partOf and its (transitive) specialization properPartOf (standing for proper part).4 Such relations allow us to distinguish compound observations, i.e., observations composed by other observations, from atomic ones, i.e., observations that are not mereologically decomposable. The idea is that compound observations group several pieces of information into a single observation. More
4The mereology we would like to adopt is the classical atomic extensional mereology [
We introduce the subclass of atomic observations (AtomicObs) collecting all the observations having no proper parts, see (a1).5 To simplify our framework, basic, argumentation, and source observations are all subclasses of AtomicObs (the option where such subclasses of Observation are closed under mereological sum, e.g., basic observations exclusively composed by basic observations are still basic, is technically more complex). Furthermore, we assume that all the parts of observations are observations themselves, and that observations can be part only of other observations (see (a2) and (a3)). a1 AtomicObs ≡ Observation ⊓ ∀properPartOf− .⊥ a2 Observation ⊑ ∀partOf− .Observation a3 Observation ⊑ ∀partOf.Observation Basic Observations. Basic observations are the most simple observations representing claims about domain entities, e.g., the attribution of authorship to an artwork. Basic observations do not have other observations in their arguments (in the sense of hasArgument); this means that, diferently from argument and source observations (see below), they are not observations about observations. Subclasses of BasicObs can be organized into taxonomies at diferent levels of generalization (specificity or granularity). The core module does not include subclasses of BasicObs, since they are always domaindependent. However, we provide here some methodological principles to specialize BasicObs, and Section 4 discusses an example relative to literary studies. Importantly, it is possible that diferent groups of domain experts come up with alternative subclasses of BasicObs (alternative observation languages). It could be therefore necessary to analyze and possibly align them in scenarios where data exchange and integration are relevant. In principle, one can think in terms of a library of multiple (partially aligned) observation languages, representing diferent scholarly perspectives on domain entities.
From a methodological standpoint, we assume that users begin with introducing the main subclasses of BasicObs with a fixed number of arguments, where these subclasses collect observations with the same number of arguments. Accordingly, for each main subclass of BasicObs, users need to introduce () an axiom establishing the number of arguments; () specializations of hasArgument that uniquely identify its arguments, and () axioms that guarantee the existence of these arguments. Then, as said, all the specializations of a main subclass of BasicObs maintain the same number of arguments. For instance, in the case of AuthorshipObs, formulas (e1)–(e3) introduce cardinality restrictions, i.e., (e1) assures that instances of AuthorshipObs have exactly two arguments, (e2) assures that one of these arguments is an author, and (e3) that the other argument is an artwork.
e1 AuthorshipObs ⊑ ∃=2 hasArgument e2 AuthorshipObs ⊑ ∃=1 hasArgument.Author e3 AuthorshipObs ⊑ ∃=1 hasArgument.Artwork
Users can specialize a subclass by imposing additional constraints on the ranges of the arguments. For instance, one could add two subclasses of AuthorshipObs, namely StatueAuthorshipObs and PaintingAuthorshipObs, grounded on the restriction of one argument to Statue and Painting (which are two subclasses of Artwork), respectively. A diferent kind of specialization is conceptually more interesting in the context of observation languages. Consider scholarly debates about artworks’ colors or styles. One may wish to have in the ontology both the main subclass of BasicObs for colors, i.e., ColorObs, and its specializations, e.g., RedObs, BlueObs, etc. These can in turn be further specialized through, say, ScarletObs, CrimsonObs, and MagentaObs for RedObs (without having further subclasses), and similarly for BlueObs. ColorObs and its subclasses have only one argument and, quite plausibly, the range of such argument does not vary across the subclasses.6 Note that the way in which CrimsonObs specializes 5Formulas marked with are the axioms of the ontology, whereas are its SWRL rules. All formulas with are not part of the ontology and are used for examples only. 6It may be the case that certain colors apply only to specific kinds of entities. We do not consider this situation here. RedObs is diferent from the way CarRedObs specializes RedObs: the observation CarRedObs restricts the range of RedObs to Car, whereas CrimsonObs stands for observations at a finer level of granularity, considering that the property of being crimson is more specific than that of being red. One can think, in fact, that the taxonomy introduced above for ColorObs sets the color-values and their granularity: a red-observation is more general and less informative than a scarlet-, crimson-, or magenta-observation. This can be relevant in application contexts where agents have devices with diferent resolutions. For instance, some agents could distinguish only general colors like red, blue, etc., while other agents may be able to distinguish more specific shades of red, blue, etc. In this case one needs to represent the granularity of observations.
These considerations have a technical impact. While for specializations like CarRedObs (similarly for StatueAuthorshipObs and PaintingAuthorshipObs) one can impose identity conditions by means of a SWRL rule like (e4), the same does not apply to RedObs. Suppose to consider (e5). In this case we cannot distinguish, for instance, red-observations obtained through devices that only distinguish the principal colors from red-observations obtained through devices that are able to distinguish diferent shades of red. To distinguish the two cases, one can () allow instances of RedObs that are instances of no subclass, i.e., RedObs is not partitioned by its subclasses; () introduce identity conditions like (e5) only for the leaves of the taxonomy and for the set-theoretically diference between the set of instances of RedObs and the union of the instances of its subclasses.
e4 CarRedObs(?), CarRedObs(?), hasCar(?, ?), hasCar(?, ?) → SameAs(?, ?)
e5 RedObs(?), RedObs(?), hasArgument(?, ?), hasArgument(?, ?) → SameAs(?, ?)
Argumentation Observations. Relying on research work in argumentation theory, in particular on
the bipolar argumentation framework presented in [
Compound observations are important when it is a whole conjunction of observations that supports an observation and not each single conjunct by its own, i.e., without any part of the supporting compound observation being suficient to support that observation. The same goes for defeat observations. Vice versa, the support (defeat) of a compound observation can always be represented (for countable scenarios) by the logical conjunction of the supports (defeats) of all its atomic parts.
Axiom (a4) characterizes the arguments of SupportObs (the axiomatization technique is the same as the one for the subclasses of BasicObs).7 Accordingly, each instance of SupportObs is related via hasArgument to exactly two entities; one of these arguments is the supported observation, the other one is the supporting observation. Relations hasSupportedObs and hasSupportingObs are restrictions of hasArgument which apply only to observations and are intended to explicitly distinguish the supporting observation from the supported one (following the previous discussion, the range of hasSupportedObs is AtomicObs). Defeat observations have a similar axiomatization to distinguish between defeating (via hasDefeatingObs) and defeated (via hasDefeatedObs) observations.
a4 SupportObs ⊑ ∃=2hasArgument,
SupportObs ⊑ ∃=1hasSupportedObs, SupportObs ⊑ ∃=1hasSupportingObs
As said in Section 2, all observations are pieces of information representing how certain entities have properties or are related. In particular, argumentation observations represent when an observation supports or defeats another observation. In this view, it is not possible to have diferent support 7For the sake of brevity, we introduce only the core aspects concerning the axiomatization of the introduced classes and object properties; readers can refer to the OWL file for a complete overview of the axioms. observations with the same supporting and supported observations. We impose this constraint via the SWRL rule (r1).8 Similarly for defeat observations.
r1 SupportObs(?), hasSupportedObs(?, ?), hasSupportingObs(?, ?),
SupportObs(?), hasSupportedObs(?, ?), hasSupportingObs(?, ?) → SameAs(?, ?) Source Observations. Source observations play a fundamental role, because they allow us to model some aspects pertaining to the provenance of observations, in particular, what is the source asserting (AssertObs) or rejecting (RejectObs) an observation. Following the idea that sources must be accessible in an intersubjective manner, we assume that the source of an observation is always a text. This is common in the humanities, where scholars refer to texts to point out that someone has expressed an observation. In our approach, texts also include databases and knowledge bases as sources for observations, which is more common in science. Furthermore, it is important to stress that assertions and rejections are observations themselves. This allows us to have further argumentation or source observations (making explicit the provenance of the source observations themselves) about them. In particular, source observations can be stacked up, i.e., one can recursively have texts asserting or rejecting observations about what other texts assert or reject. We will come back to this point. Here, we just want to note that our observational framework allows for the partial documentation of the subjective interpretation of a text by making explicit the observations that each interpreter associates to a text. It is possible to explicitly represent several characteristics of the contents of texts and to study the diferent points of view of interpreters. In this sense, source and argumentation observations, which account for the evidence the interpreters explicitly provide to argue in favor or against a given claim, allow us to partially capture the debates among scholars.
Source observations hold between two arguments, i.e., a textual source and the asserted or rejected observation, see (a5) (analougously for RejectObs). The relations hasSrc (standing for has source) and hasObs (has observation) specialize hasArgument to uniquely individuate the two arguments of source observations. Note that the assertion of a compound observation can be represented by the assertions of its atomic parts, and the rejection of a compound observation by the disjunction of the rejections of its atomic parts. Thus, we assume that hasObs has range AtomicObs.
a5 AssertObs ⊑ ∃=2hasArgument, AssertObs ⊑ ∃=1hasSrc, AssertObs ⊑ ∃=1hasObs Similarly to the case of argumentation observations, source observations with the same source and observation arguments are identical: (r2) is the SWRL rule for AssertObs, analogously for RejectObs. r2 AssertObs(?), hasObs(?, ?), hasSrc(?, ?),
AssertObs(?), hasObs(?, ?), hasSrc(?, ?) → SameAs(?, ?)
To make a simple example, let us assume to have an instance 1 of the basic authorship observation AutorshipObs. Formula (e6) represents that there is an assert observation 2 having observation 1 and source txt . In simpler terms, (e6) says that txt asserts 1.9 One could add another assert observation 3 saying that txt ′ asserts 2, as well as a reject observation 4 saying that txt ′′ rejects 3, and so on. One can therefore have chains of source observations starting from basic or argumentation observations or, more generally, from any kind of observation.
e6 AuthorshipObs(1), AssertObs(2), hasObs(2, 1), hasSrc(2, txt )
In our approach, the general intuition is that all atomic observations, including source observations, have a source. To represent this constraint by excluding at the same time infinite chains of source observations, we introduce the notions of sourceChainRootIn, assertChainRootIn, and kb (where assertChainRootIn is a restriction of sourceChainRootIn to assertion chains rather than assertion/rejection chains). In principle, the source of a compound observation could be reduced to the compound of all the sources of its atomic parts. To simplify our framework we do not consider this aspect here.
From an application perspective, users of the ontology do not need to care about populating sourceChainRootIn and assertChainRootIn since they are automatically built using the recursive SWRL rules (r3)-(r6) from the observations present in the knowledge base. Rule (r3) says that when source observation has observation (i.e., asserts or rejects) , then is a source chain rooted in (starting from) . Recursively, (r4) says that when source observation has observation (i.e., asserts or rejects) , which has a source chain rooted in , then is a (larger) source chain rooted in . In the previous example, 2, 3 and 4 are all source chains rooted in 1; 3 and 4 are source chains rooted in 2; and 4 is a source chain rooted in 3. Similarly for (r5) and (r6) in the case of assertChainRootIn. r3 SourceObs(?), hasObs(?, ?) → sourceChainRootIn(?, ?) r4 SourceObs(?), hasObs(?, ?), sourceChainRootIn(?, ?) → sourceChainRootIn(?, ?) r5 AssertObs(?), hasObs(?, ?) → assertChainRootIn(?, ?) r6 AssertObs(?), hasObs(?, ?), assertChainRootIn(?, ?) → assertChainRootIn(?, ?) The individual kb is an instance of Text representing the knowledge base of the system itself. It is intended to document all the existing positions expressed in the analyzed texts, rather than to express its own viewpoint on the texts. The interpretative dimension of kb is therefore unavoidable and is explicitly represented by assertion and rejection observations with source kb.
We assume that all source chains end in kb; thus avoiding infinite source chains. This constraint is formally represented by () introducing the subclass KBAssertObs of AssertObs and the subclass KBRejectObs of RejectObs characterized via (a6) and (a7), respectively, i.e., these classes identify the source observations with source kb; () introducing the subproperty kbSourceChainRootIn of sourceChainRootIn that is automatically populated starting from KBAssertObs, KBRejectObs, and sourceChainRootIn by means of the SWRL rules (r7)-(r10); and () adding the axiom (a8) claiming that every atomic observation is the root of a source chain ending in the kb. Note that, by assuring that kb-observations are rooted in themselves, (r9) and (r10) are “technical artifices” that make (a8) operative. Finally, we assume that kb is “coherent” (see (r11)): given the disjointness of AssertObs and RejectObs, (r11) generates an inconsistency when kb asserts and rejects the same observation. a6 KBAssertObs ≡ AssertObs ⊓ ∃hasSrc.{kb} a7 KBRejectObs ≡ RejectObs ⊓ ∃hasSrc.{kb} r7 KBAssertObs(?) → sourceChainRootIn(?, ?) r8 KBRejectObs(?) → sourceChainRootIn(?, ?) r9 KBAssertObs(?), sourceChainRootIn(?, ?) → kbSourceChainRootIn(?, ?) r10 KBRejectObs(?), sourceChainRootIn(?, ?) → kbSourceChainRootIn(?, ?) a8 AtomicObs ⊑ ∃kbSourceChainRootIn− .(KBAssertObs ⊔ KBRejectObs) r11 hasObs(?, ?), hasObs(?, ?), hasSrc(?, kb), hasSrc(?, kb) → SameAs(?, ?) In practice, in order to block infinite regressions, for every observation introduced under Observation, users need to provide a chain of observations ending at kb. It is important to stress that in our approach kb is not intended as a direct source of information; it is rather a collector of observations coming from other sources. Formally, this means that there should be no observation having kb as a direct source, hence, at least a chain of length 2 is assumed (see examples in Section 4).
Once we collect basic observations linked to specific observation languages, together with source and
argumentation observations concerning them, we can analyze the status of some claims or sources
(e.g., coherence, ambiguity, conflict) on the basis of explicit criteria. In the following, we mainly adapt
principles considered in assertion logics [
Let us begin with the notion of disputability of observations: an observation is strongly disputable when there are two sources directly accessible by kb, one asserting and one rejecting . This notion can be captured through the class StronglyDisputableObs and the rule (r12), ensuring that when there are two kb assertions, one asserting that a source asserts , and one asserting that a source rejects , then is strongly disputable. According to (r12), kb has direct access to the observations that is asserted and rejected by some source. Strong incoherence can be weakened by allowing kb to have indirect access to such observations, i.e., they are reported to kb via a chain of assertions. To represent this, one can rely on kbAssertChainRootIn (a single rejection in a chain would prevent the transfer of information from the original source to kb) to introduce the rule (r13) for the new class DisputableObs. r12 KBAssertObs(?), KBAssertObs(?), hasObs(?, ?), hasObs(?, ?),
AssertObs(?), RejectObs(?), hasObs(?, ?), hasObs(?, ?) → StronglyDisputableObs(?) r13 kbAssertChainRootIn(?, ?), kbAssertChainRootIn(?, ?),
AssertObs(?), RejectObs(?), hasObs(?, ?), hasObs(?, ?) → DisputableObs(?) A similar analysis can be done focusing on the incoherence of sources, i.e., when a source both asserts and rejects an observation. See the OWL file in the repository for the SWRL rules and the new classes necessary for the two versions of incoherence.
Following argumentation theory, we can also study some properties of the “network” of support and defeat observations. In particular, we can analyze compound observations collecting all the observations relevant for a given context by classifying them under some classes.10 For example, a (compound) observation is conflictual when it has two parts one defeating the other. A strong notion of conflict is introduced through the rule (r14). Again, such notion can be weakened by considering assertions chains.
r14 partOf(?, ?), partOf(?, ?), KBAssertObs(?), hasObs(?, ?), DefeatObs(?), hasDefeatingObs(?, ?), hasDefeatedObs(?, ?) → StronglyConflictualObs(?)
In this section, we exemplify how the core module of our framework can be applied to literary studies,
and notably how it can be used to model specific kinds of scholarly observations about the contents of
literary texts. We partially reuse the case study presented in a first-order setting of our framework [
The case study consists of the representation of observations by diferent scholars concerning the final
novella (tale) of Boccaccio’s Decameron, namely the tale (X,10) of Griselda and Gualtieri. In short, the tale
narrates the story of the nobleman Gualtieri who marries the poor Griselda. To test his wife’s obedience,
Gualtieri submits Griselda to increasingly cruel trials, such as letting her believe that their children need
to be murdered. Griselda accepts everything he puts her through. In the end, Gualtieri reveals the truth
and they live happily ever after. Due to its controversial plot, the critical debate around the tale is still
open. We focus on how Vittore Branca and Michelangelo Picone analyzed the tale’s (literary) characters,
as well as how they conceived the relationships between the tale and other texts. For shortness, we
shall report only some observations; the reader can refer to the available OWL file, as well as to [
Among his main theses, Branca characterizes Griselda by linking Boccaccio’s work to biblical and hagiographic texts. In particular, he argues for the similarity between Griselda and the Virgin Mary, which is in turn supported by the observation that they are both patient figures. The theses by Picone follow a diferent direction, connecting Boccaccio’s tale to the Medieval Lai de Fresne. In doing so, Picone investigates the relationships between the characters of Griselda and Fresne, rejecting the similarity between Griselda and Mary. 10Argumentation theory usually refers to sets of arguments rather than compound arguments.
To make sense of these observations in a modular extension of our ontology, we need to introduce at least the following individuals in the domain of quantification: the figures of Griselda ( griselda) and Virgin Mary (mary ), Branca’s text Boccaccio medievale (bmd ), Picone’s text Boccaccio e la codificazione della novella (bcn), a collection of hagiographic texts (hag ),11 and the Decameron’s tale X,10 (tlx ). At the class level, we introduce two classes of basic observations: () PatientObs for the observation of being a patient figure (a9); and () ResemblanceObs for the observation of being resemblant to other domain entities (a10). As in previous examples, in the case of (a9), the axiom makes explicit the range of hasArgument to a specific class of the ontology, which is Figure in this case. We introduce the latter class only in a preliminary way for the sake of the example to model domain entities like Griselda and Virgin Mary; the modeling of (literary) characters deserves much more robust work to be done in a principled manner. In the case of (a10), hasResemblantArg and hasResembledArg are subrelations of hasArgument meant to distinguish between the two arguments of the observation.
a9 PatientObs ⊑ ∃=1hasArgument.Figure a10 ResemblanceObs ⊑ ∃=2hasArgument,
ResemblanceObs ⊑ ∃=1hasResemblantArg, ResemblanceObs ⊑ ∃=1hasResembledArg Let us assume that it is not controversial that Griselda is a figure; hence, we can introduce Figure(griselda) (analogously for the Virgin Mary). (e7) represents that, according to Branca’s text bmd , Boccaccio’s tale tlx asserts that Griselda is patient: 1 is the observation of Griselda being patient; 2 is the observation of tlx asserting 1; 3 is the observation of bmd asserting 2 (i.e., 3 is the assertion done in bmd that tlx asserts that Griselda is patient). Finally, 4 is the assertion of 3 in kb which is required to close the source chains. Similar observations are introduced for the characterization of the Virgin Mary with respect to the text hag (see the OWL file).
e7 PatientObs(1), hasArgument(1, griselda), AssertObs(2), hasObs(2, 1), hasSrc(2, tlx ),
AssertObs(3), hasObs(3, 2), hasSrc(3, bmd ), KBAssertObs(4), hasObs(4, 3) According to (e8), 9 is the observation of Griselda resembling the Virgin Mary; 10 is the assertion of 9 expressed in bmd ; and 11 is the required assertion of 10 in kb to close the source chain. e8 ResemblanceObs(9), hasResemblantArg(9, ), hasResembledArg(9, ),
AssertObs(10), hasObs(10, 9), hasSrc(10, ), KBAssertObs(11), hasObs(11, 10) (e9) represents the argument put forward by Branca in favour of the resemblance between Griselda and the Virgin Mary: 13 is the supporting of 9 (the resemblance of Griselda with the Virgin Mary) by 12, where 12 is the compound of 2 (tlx asserting Griselda being patient) with 6 (hag asserting Virgin Mary being patient, see the OWL file); 12 14 is the assertion of the supporting observation 13 in bmd (we omit here 15 as assertion of 14 in kb).
e9 properPartOf(2, 12), properPartOf(6, 12), SupportObs(13), hasSupportingObs(13, 12), hasSupportedObs(13, 9), AssertObs(14), hasObs(14, 13), hasSrc(14, bmd ) For lack of space, (e10) reports only the source observation of Picone’s text bcn rejecting 9, i.e., the resemblance of Griselda with the Virgin Mary (as for 14 in (e9), we omitted the assertion of 16 in kb). e10 RejectObs(16), hasObs(16, 9), hasSrc(16, bcn)
Reasoning over the ontology and the data, by means of (r12) and (r13), 9 is classified by both StronglyDisputableObs and DisputableObs, i.e. there is a disagreement about the observation that 11The introduction of hag is a simplification with respect to Branca, who talks of hagiographies without reference to specific texts. This might lead to the modeling of “vague” entities, which we leave to future work. 12The symmetry of resemblance is not assumed, which motivates the introduction of both hasResemblant and hasResembled as specializations of hasArgument. This is because, first of all, there are open discussions in cognitive science about the symmetry of general similarity relations; second, we do not commit to the fact that the resemblance Branca has in mind is symmetric, even though it is supported by the sharing of a common trait (being patient) among the two figures. Furthermore, symmetric relations raise a problem concerning observations, e.g.: is it necessary to distinguish ‘Griselda resembling the Virgin Mary’ from ‘the Virgin Mary resembling Griselda’, or is the single observation ‘the resembling of Griselda and the Virgin Mary’ enough? We leave this open to future work.
Griselda resembles the Virgin Mary. Depending on the application context and the represented data, our framework can support the automatic classification of individuals in various ways, e.g., to spot the presence of incoherent sources, conflictual observations, and so on.
The representation of observations in the Digital Humanities has been attracting the interest of the community for a while. We now wish to compare our approach with some existing works.
The CRM Argumentation Model (CRMinf) [
The Historical Context Ontology (HiCO) [
To model claims’ contents, some of the authors of HiCO have been developing an approach based on
named graphs, see [
To make a distinction between factual and hypothetical data in cultural heritage, Carriero et al. [
Stressing a bit further some key diferences between our approach and the ones just mentioned,
from a conceptual standpoint, we do not take observations as mental entities, unlike CRMinf. As said
in Section 2, they are pieces of information expressed through observational languages. This is an
important departure point: our aim is to model the claims that scholars express in accessible sources,
rather than their mental ideas. We also depart from pure technical approaches like the one in [
Another key aspect of our framework is its axiomatization, which makes it applicable in contexts where both query-answering and reasoning can be applied. From this perspective, we have relied on existing work in argumentation theory to represent some elements of the argumentation put forward by scholars to support or defeat claims. Moreover, we have introduced various SWRL rules to automatically analyze claims and their relations on the basis of transparently designed criteria, e.g., to check whether observations express conflictual claims.
In this paper we presented the core module of a Semantic Web ontology to represent and reason over observational data, collected through the design of (domain-specific) controlled observational languages. These latter need to be developed in close collaboration with domain experts to reflect as much as possible their critical stance on the observed entities. It is important to stress that to make sense of the variety of conceptual systems in the humanities, our approach allows for the development of alternative observational languages, possibly linked in what we call a library of observational languages, to represent diferent sorts of claims and their mutual relationships. We also introduced SWRL rules for the analysis of both observations and their sources, e.g., to check whether a single observation makes conflictual claims. Further rules can be added by extending our ontology and making it suitable to specific application contexts. Conceiving the framework with a modular architecture means to make it available for diferent domains. By collaborating with domain experts to test the ontology against their knowledge, we have also shown a preliminary exemplification of the use of our ontology in literary studies, giving a sample of its potential uses to represent scholarly claims. Further work is needed to refine our ontology with respect to a larger quantity and variety of data.
As a final remark, it is important to stress that our approach is compatible with research and
application eforts in Computational Literary Studies (CLS) [
Sanfilippo and Tomazzoli’s work is supported by the project MITE – Make it explicit: Documenting interpretations of literary fictions with conceptual formal models (2023-2025) funded by the European Union – Next Generation EU, Mission 4, Component 1, CUP B53D23028830001. Mosca’s work is supported by the PNRR MUR project PE0000013-FAIR project Future Artificial Intelligence Research funded by the European Union – Next Generation EU.