ICON: An Ontology for Comprehensive Artistic Interpretations

In the context of art interpretations, several attempts have been made to create models that cover some specific elements related to interpretations (i.e., symbolic meanings) or the whole act of interpretation of a cultural heritage object. CIDOC-CRM [5] is a widely used ontology in the context of cultural heritage. It has an event-based structure and covers fundamental aspects of the lifecycle of a cultural heritage object. Carboni and de Luca [8] extended it with the VIR ontology, which explores the concept of visual representations in artworks and associates the portion (called the iconographical atom) of the cultural heritage object to the recognized subject. We use SKOS alignments to refer to parts of CIDOC and VIR in our ontology, and compare the coverage of our ontology and VIR in Section 6.3. Compared to our ontology, VIR focuses only on subjects of level 2, considering iconographies and their attributes, consequently lacking of a clear distinction between levels. The preliminary study conducted in previous work [3] further extends VIR by the addition of an iconological interpretation class linking to the artwork concepts and external cultural phenomena. It is evaluated over 11 real case studies taken from the literature in iconology, which illustrate a wide variety of aspects included in an iconological analysis. In addition, the work is based on a careful theoretical comparison of the main iconological and iconographical interpretation theories. For its comprehensive overview over iconographical and iconological theories, along with the real-base evaluation, it is used as a source for ontology development proposed here, which has to be seen as its development and refinement. We deepen this study by developing aspects not already considered, such as a more detailed description of level 1 and 2 subjects and the integration of multiple interpretations by different art historians. Gartner [27] proposes an ontology to facilitate and automate the identification of subjects (level 2) in works of arts through logical inferences. No alignments were possible to this ontology because it has not been released. ArCo’s ontology [10] was developed to model Italian cultural heritage artifacts by converting information contained in traditional catalogue sheets into linked open data. Among the possible aspects modeled for an artwork, some classes were designed to describe its iconographical apparatus. Apart from this class, the schema does not mention any distinctions between different levels of interpretations. Most of the information about the iconographical and iconological interpretations in ArCo are provided through natural language descriptions with the property dc:description or core:description,⁶ not exploiting the full potential of the Semantic Web [49]. As we mention in Sections 4 and 5, in the development of ICON we designed specific classes and properties to express this information with the necessary granularity. We reuse some parts of ArCo to refer to the concept of artwork and subject. In an earlier work [50], symbolic meanings in cultural heritage in the Simulation Ontology were modeled. Compared to ICON, this ontology does not consider the hermeneutic act of interpretation of associating the symbolic meanings to artworks. Nevertheless, its conceptualization of symbolic meanings using n-ary relationship classes that link a symbol, its symbolic meaning, and the cultural context in which the symbol-symbolic meaning relationship takes place fits well with our modeling of symbols. Therefore, we reuse the Simulation Ontology classes and properties to express the symbolic meanings in our work, inserting it in the context of an interpretation of an artwork.

In the context of knowledge organization, several ontologies addressed the concept of interpretation. CIDOC-CRM models assertions with the class E13_Attribute_Assignment, which relates the assertion made by one agent to the object considered. Since each assertion reflects the agent’s opinion, multiple, contradictory assertions may be represented. The concept of interpretation is applied broadly, including measurements and other types of scientific observations. Similarly, the class Interpretation of Arco is intended to describe every piece of information asserted by an agent about an object on the basis of stated sources.⁷ The CIDOC-CRM extension CRMinf deepens the concept expressed by E13_Attribute_Assignment distinguishing the type of argumentation and if the belief resulting from the argumentation holds true or not. The concept is further explored by the CRMsci, another CIDOC-CRM extension, which integrates CRMinf by formalizing the shared scientific process adopted across different domains and the scientific activities involved. In detail, of great interest is the class crmsci:S4_Observation, subclass of crmsci:I1_Argumentation and of crm:E13, expressing the scientific observation of physical events or reality that is done directly or through measurements. It represents the “transition between reality and propositions” [20]. Furthermore, the VIR ontology adds a domain-definition of crmsci:S4 Observation by declaring its subclass vir:IC12 Representation, which represents an assignment of a solely iconographical status to a physical object.

The same topic is addressed also by the history domain to represent the frequent case of disagreeing historians’ interpretations of the same events. As reported by Frank [22], several ontologies afford the theme by modeling different views of the same observed events, such as SEM ontology⁸ MIDM [58] and the Ontology Design Pattern (ODP) Event-Model-F,⁹ expanding DOLCE+DnS Ultralite (DUL), in which a distinction between facts and interpretations is already stated [24]. HiCO [18]¹⁰ goes further by adding contextual information to the interpretation, such as interpretation type and criterion.

The concept of interpretation used in our ontology reflects the VIR perspective, narrowing down the field to those interpretations that an observer may have about the visual representation of an artwork, excluding scientific observations about its physical features. Moreover, we further specify the meaning of an interpretation in the context of this work by referring to the definition provided by van Ruymbeke et al. [58] relatively to the archaeological field, stating that the views produced by scholars are “the result of an interpretive reasoning that includes the subjectivity of the author,” due to the uncertainty and incompleteness that often characterizes archaeological data [58].

While modeling an interpretation, it is important to define also what the concept of meaning is. Considering semiotics, there are several aspects of meaning that can be modeled [40]. In the context of this study, meaning is defined according to the semiotic theory [21] in which a signifier (i.e., an icon) signifies the carried signification (i.e., a meaning) [54, pp. 93–94]. Therefore, following Panofsky’s modeling, all the subjects identified at each level of interpretation are considered meanings [36]. This definition corresponds to the notion of meaning as a social object introduced in the work of Picca et al. [40].

Although iconographical and iconological descriptions are not their main focus, some general domain schemas contain information related to art interpretation. Often these schemas rely on the sole subject property (e.g., dc:subject, schema:about) to describe any information about the iconographical and iconological content [2, 15, 33, 44]. One of the outliers in this characteristic is Wikidata and its Wikidata schema [60]. The Wikidata schema contains different properties that link an artwork to its content, such as wdt:P921 (main subject) or wdt:P180 (depicts). Moreover, Wikidata allows adding qualifiers to the statements made with the property wdt:P180 to address specific aspects of elements depicted in the artworks (i.e., their symbolic meaning, qualities).¹¹ Compared to ICON, Wikidata does not explicitly distinguish between levels of interpretations, linking elements of the first and second level to the artwork with the same property. This limits the possibility of art historian driven research questions to be answered with the Wikidata model. Finally, compared to our work, Wikidata does not include potential intrinsic meaning of artworks provided by the third level of interpretation. A more in-depth analysis of the qualifiers and comparison between Wikidata and ICON is presented in Section 6.3.

The scope of iconography has some overlaps with the domain of narratology in the description of the plot for what concerns the represented characters and their actions. A good wealth of studies concerns the semantic modeling of the topic¹² [12, 56], among which some focus on the narrative representation in visual images [14, 63]. Even though they do not describe the iconographical subject with the necessary granularity required by an iconographical study, they provide a solution for organizing the common archetypical knowledge of stories, events, and characters participating in them [14].

Controlled vocabularies and taxonomies of art and culture, despite not being ontologies, are essential for standardizing the reference to elements that belong to the first, second, or third level of interpretation. Iconclass [13] is a classification system that mostly deals with iconographical subjects. The Getty Art & Architecture Thesaurus (AAT) [30] provides permanent identifiers for people, concepts, and places that might be contained in artworks. The two aforementioned taxonomies cover a wide amount of information,¹³ but other exists that were created ad hoc for museums or cultural institutions, such as the Rijksmuseum’s thesaurus [19].

These controlled vocabularies and taxonomies can foster the interoperability between different knowledge graphs that use them, but they do not provide statements regarding the type of element that is depicted in a work of art (whether it belongs to the first, second, or third level of interpretation). For this reason, it is essential for these characteristics to be modeled with a specific ontology.

As explained in Section 2, works of art can be analyzed through different layers of interpretations that depend on recognitions. A recognition, in the context of this ontology, is an interpretation act made by an agent (or interpreter, which can be a biological or electronic being) that links works of art to something related to their content. From a conceptual perspective, it is a mental entity reflecting the agent’s subjective point of view. From a technical viewpoint, it is an n-ary predicate that cannot be modeled using OWL due to expressivity limitations; therefore, it was turned into an n-ary relationship class.¹⁷ Coherent recognitions on the same artwork are collected and documented by interpretation descriptions (requirement 8, Section 4).¹⁸ In this iteration, we conceptualize the elements that revolve around recognitions. From the n-ary relationship class icon:Recognition, several properties were designed (or reused from existing ontologies) to link it to its interpreter(s) (or agents), the artwork that is being interpreted, supporting sources for the recognitions. In particular, the aboutWorkOfArt property links the recognition to the artwork (Artwork class). Then, the dul:includesAgent property (from DOLCE [23]) links the recognition to the agent who performed it (requirement 4, Section 4). The class InterpretationDescription is linked to (one or many) Recognition class(es) that comply with it through several properties according to the type of the recognition, namely isCompliantWithPreiconographicalRecognition for pre-iconographical recognitions and formal motif recognitions,¹⁹ isCompliantWithIconographicalRecognition for iconographical recognitions, and isCompliantWithIconologicalRecognition for iconological recognitions. The CiTO [51] properties cito:citesForInformation and cito:citeAsEvidence can be linked to a icon:Recognition class to provide sources or other information that support a recognition (requirement 5, Section 4). Finally, a recognition can also be used to support further recognitions made on the same artwork or another one. For example, Panofsky recognizes that the figure of Chastity sculpted by Giovanni Pisano on the Pulpit of Pisa cathedral is represented with the same appearance of the nude classical iconography of Venus Pudica (formal motif recognition).²⁰ This interpretation provides support to the third-level recognition of the characteristics of the Proto-Renaissance movement in the cultural context of the Medieval Tuscany [37, p. 157]. To express this using our ontology, the property cito:givesSupportTo can link the supporting recognition to another one (requirement 7, Section 4). These elements are also the object of interest of the general CQs (see Table 3, Q0.1–Q0.5).

Depending on the level of interpretation presented in Table 1, four Recognition subclasses have been defined:

Recognitions at each level of interpretation may be based on the results of the recognition at one of the previous levels. Therefore, they can be linked together but ultimately are modeled as independent of one another. This choice is made since (i) the describer may not have available the descriptions of the lower level(s), (ii) the corresponding subjects in the other levels may not be relevant for the recognition, and (iii) it may be possible that a level 3 recognition (i.e., an IconologicalRecognition) is linked to level 1 subjects rather than level 2 ones (e.g., iconological interpretations of a landscape painting, which may not have level 2 subjects [36]).

These classes and their specific usage will be further described in the following sections. Figure 1 shows a rendering of the classes and properties of this iteration.

In this iteration, we model the recognitions that happen on a pre-iconographical level. In this level, an interpreter recognizes artistic motifs present in the artwork, and associates to them (i) natural objects (a tree, a man, a sword) without identifying specific individuals from those classes (tree of life, Saint Joseph, Excalibur) that are recognized in level 2 (Section 5.3), (ii) in the form of expressional meanings²¹ (emotions of the depicted elements), (iii) qualities about these elements (size, color, positions), and (iv) performed actions (see Table 1 in Section 2). Assuming that the agent doing the interpretation act might also be a computer, as in the case of the results of object detection through computer vision, we give the possibility to express coordinates of the portion of the image of the artworks where these elements are detected. Furthermore, these coordinates can be expressed using IIIF URIs [53] that point to a specific portion of the work of art. A series of artistic motifs can be grouped together in a composition that can have a compositional structure²² (e.g., pyramidal). Additionally, an interpreter might recognize similarities between artistic motifs present in a work of art with other artistic motifs of another work of art, recognizing a prototypical artistic motif or composition that is reused in another artwork. For example, the level 1 description of Pisano’s figure of Chastity cited earlier is linked through a formal motif recognition to the level 1 description of Venus Pudica, from which its appearance is derived (i.e., a nude woman covering herself with her arms). Artistic motifs and compositions are linked to the class PreiconographicalRecognition respectively through the properties recognizedArtisticMotif and recognizedComposition. Only one artistic motif or composition can be linked to a recognition. Compositions are linked to the artistic motifs that take part in them through the hasPart property. If the artistic motif refers to a natural object or action with a factual meaning, it is linked to the classes NaturalElement or Action through the property hasFactualMeaning. Otherwise, if what is recognized in the artistic motif is an expressional meaning, the property that links it to expressional meanings is hasExpressionalMeaning. If actions, expressional meanings, or natural elements have some specific quality that needs to be highlighted, from the artistic motif the qualities are expressed with the DOLCE hasQuality property. When the pre-iconographical recognition is performed by a computer with an object detection algorithm, or when an IIIF URI is provided, it is possible to associate not only the detected objects but also the coordinates of the image in which they are found. Coordinates of the detected object can be expressed through the data property hasRegionDescription that has the ArtisticMotif or Composition classes as the domain. As mentioned earlier, the use of IIIF URIs for the format of this data property is also welcomed. The FormalMotifRecognition class links the prototypical motif to the copied motif respectively using the hasPrototypicalMotif and hasCopiedMotif properties. Finally, all the coherent formal motif recognitions and pre-iconographical recognitions that take part in an interpretation about a work of art can be linked to an InterpretationDescription class, through the property preiconographicallyCompliesWith. Figure 2 shows a graphical rendering of the classes and properties used in this interpretation level.

In this third iteration, we focus on the Panofsky’s second level of art interpretation: the iconographical interpretation. In this level, the interpreter recognizes images and invenzioni²³ in an artwork. An image represents the subject depicted as a manifestation in the specific artwork taken into account. It is then linked to second-level subjects, which are characters, places, events, named objects,²⁴ symbols, and personifications, identifying iconographies from an abstract and general point of view. This distinction between the general subject level (i.e., characters, symbols) and the artwork-specific one (image) is functional to identify the variants of a subject in relation to the specific context (i.e., Thor as represented in a specific painting may differ from its common one). An invenzione, instead, is the subject matter represented by the combination of general subjects linked to the single images recognized.²⁵ For example, in an artwork, you might recognize three images: the first refers to the general subject of Mary, the second refers to the general subject of Angel Gabriel, and the third refers to the general subject of the Holy Dove. The combination between the general subject of Mary, Angel Gabriel, and the Holy Dove is the Annunciation, which, in our ontology terms, would be considered the invenzione. The same invenzione could be present in multiple artworks, but each artwork maintains its uniqueness by having different images. The classes Story and Allegory are subclasses of the class Invenzione. Stories are more likely to contain characters, named objects, places, and events, whereas allegories are more likely to contain symbols and personifications. We give the possibility to express symbols as just symbolic meanings recognized, or, for a more thorough description, as Simulations (see Section 5.5). The classes Image and Invenzione are linked to the class IconographicalRecognition through the respective properties recognizedImage and recognizedInvenzione (one image or invenzione per recognition). The artistic motif belonging to a pre-iconographical level that refers to the recognition of an image can be linked to it with the property refersToArtisticMotif (i.e., the recognition of the image that represents Mary Magdalene can be linked to the artistic motif that has the factual meaning of woman). This link is important to ensure that the connection between pre-iconographical elements and the respective iconographical subjects is preserved. If the artistic motif is the principal element that enabled a recognition of an image, then it can be linked to that image through the property hasRecAttribute (i.e., the recognition identifying Cupid has recognizing attributes the artistic motifs linked respectively to “wings” and “arrows”). Images are linked to the general subject portrayed through specific properties according to the subject class. The property hasCharacter links an image to the class Character, likewise hasEvent refers to the class Event, hasPlace refers to the class Place, hasNamedObject refers to NamedObject, hasSymbol refers to Symbol, and finally hasPersonification refers to Personification. The cited ICON classes represent second-level subjects represented in the fictional representational space, therefore including both real and fictional, non-existent subjects (e.g., Medusa, the Greek mythological character appearing in various media), in compliance with the modeling of subjects in narratology [12, 14]. An invenzione is linked to the elements that compose it through the property composedOf. Finally, multiple iconographic recognitions that take part in an interpretation of an artwork are linked to the interpretation using the iconographicallyCompliesWith property. Figure 3 shows the classes and properties relative to this level of recognition.

Iconological interpretations (third level) focus on the recognitions of intrinsic meanings.²⁶ An intrinsic meaning links the whole artwork or some parts of it to a cultural phenomenon and a concept that defines it. The IconologicalRecogniton class is linked to the IntrinsicMeaning class²⁷ through the property recognizedIntrinsicMeaning. From there, the n-ary class IntrinsicMeaning can be linked to a specific composition, image, or artistic motif that can be the focus of the intrinsic meaning through the properties hasComposition, hasImage, and hasArtisticMotif. Then, it is linked to the expressed concept through the property recognizedConcept. For the range of this property, we reuse the DOLCE class SocialObject because there was no need to create an ad hoc class for this element.²⁸ Additionally, since an intrinsic meaning can also reflect some cultural phenomena, it is linked to the class CulturalPhenomenon through the property recognizedCulturalPhenomenon. Currently, CulturalPhenomenon has four subclasses, which specify the type of cultural phenomenon, namely Attitude, Belief, CulturalValue, and Tendency. These terms are taken from Panofsky’s vocabulary in the description of the third level of artistic interpretation.²⁹ Finally, all the iconological recognitions that take part in an interpretation made on an artwork are linked to it with the property iconologicallyCompliesWith. A graphical rendering of this fourth iteration, representing the third level of the interpretation, can be found in Figure 4.

To promote ontology interoperability and reusability, we connect to several external ontologies through means of alignments and reuse. We present our alignments and reuse by following guidelines proposed by the state of the art [9, 35]. Our ontology selection for reuse and alignment was guided by different principles: (i) standardization for CIDOC-CRM [5] and FRBRoo [46] because they are considered standard frameworks in the domain, (ii) cognitive and formal analysis for the choice of DOLCE foundational ontology [6, 23] in its OWL version (DOLCE Zero), Simulation Ontology [50], VIR [8], HiCO [18], and CiTO [51], as all of them offer design solutions to the CQs defined from the requirements in Section 4.

Due to the complexity of the field, the number of ontologies to be reused, and the heterogeneous domains from which they come, we adopted a hybrid reuse approach [9], which, depending on the specific cases explained in the following, considers either reusing directly the classes and properties of the aforementioned ontologies (either by importing the whole ontology or parts of it), or (indirect reuse) using them as fully extensional ontology patterns, or just as intensional patterns.

Extensional reuse happens when classes or properties of an ontology O1 are logically aligned to an external ontology O2, which we want to reuse with its full-fledged semantics, because it is compatible, desirable, or necessary. For example, if we extensionally align a O1 class Organisation to a O2 class dul:SocialObject, we intend to inherit the semantics of DOLCE’s social objects (e.g., that they are not physical).

On the contrary, we use parts of an external ontology O3 as purely intensional constructs when we want a limited interoperability, which does not include accepting in O1 all the semantics provided in O3, because it may be partly incompatible. For example, we may intensionally align a O1 class Image to a O3 class crm:E36_Visual_Item because we might not want to inherit the axiom stating that crm:E36_Visual_Item is a subclass of crm:E89_Propositional_Object.

To implement this distinction, indirect reuse is designed using different mapping properties, according to the semantics they provide, and its impact into the resulting reasoning. We have used RDFS (rdfs:subPropertyOf, rdfs:subClassOf) and OWL (owl:equivalentTo) logical properties when we want the alignments to provide first-order extension to ICON schema and data, and we have used SKOS skos:broadMatch, skos:related, and skos:closeMatch for purely intensional mapping, which can be used at query time to integrate data represented with ontologies that may harm the logical integrity of ICON knowledge.

Among the reused ontologies, we have used an intensional (or “terminological”) mapping for CIDOC, VIR, and FRBRoo, because we have noticed potential problems when reasoning is jointly made with both the axioms from ICON, and from those ontologies. For example, a full extensional alignment of the class icon:Image as rdfs:subclassOf crm:E36_Visual_Item would make an automated reasoner infer that icon:Image rdfs:subclassOf crm:E89_PropositionalObject, which is not defendable, since propositional entities typically exclude visual, musical, or other information modalities. In other words, CIDOC contains here a debatable assumption, which should be ignored when reusing data that use CIDOC as their schema. Now, if we use a purely intensional mapping: icon:Image skos:broadMatch crm:E36_Visual_Item, we make a commitment that can be discussed, and the triple can be used to make SPARQL-based data integration, but we will not get the inference that images are propositions.

In this section, we give a thematic overview of the classes and relations reused for satisfying a specific task, and we refer to the documentation (see Section 7) for further details on the single alignments. Table 4 shows the direct reuse of external classes and properties in ICON, and Table 5 shows the indirect alignments.

Artistic interpretations are, by definition, subjective. Nevertheless, the characteristics and relationships that surround recognized elements can be subject to logical constraints. Each recognition is made on exactly one artwork, involves exactly one agent, and different recognitions require adequate elements. A pre-iconographical recognition targets as recognized elements, either artistic motifs or compositions. A formal motif recognition instead deals with a prototypical motif and a copied motif, and both of them can be either an artistic motif or a composition. Then, an iconographical recognition recognizes exactly one image or one invenzione. Finally, an iconological recognition refers to an intrinsic meaning. For what it concerns the elements that are recognized, the recognition of an artistic motif, in Panofsky terms (and thus in our ontology) implies that the interpreter associates either a natural or expressional meaning to a portion of the artwork. At the same time, the recognition of an image implies the presence of either a character, event, named object, symbol, personification, or specific place in the artwork. Furthermore, images, artistic motifs, and intrinsic meanings cannot be instantiated without having a recognition that addresses them. To ensure this, we added several restrictions. For example, images must be linked to exactly 1 recognition through the property isIconographicallyRecognizedBy. The difference between stories and allegories is that the former generally includes characters, places, events, and named objects, and the latter is more focused on symbols and personifications. Nevertheless, a story might contain symbols and an allegory may contain characters, so the logical restrictions in these cases are not very strict. An intrinsic meaning can refer to either a cultural phenomenon or a conceptual object.

An exemplification of some restrictions on main classes through OWL axioms follows, formalized in Manchester Syntax³⁶:

This list contains only the directly created axioms. The restrictions inherited by the alignment to external ontologies are available in the documentation.³⁷

To the best of our knowledge, existing linked open data hubs do not include detailed descriptions of iconological aspects needed to evaluate the ontology. Therefore, we tested ICON on manually created data extracted from Panofsky’s Studies in Iconology interpretations³⁸ [37]. This text was chosen for its historical importance and authoritativeness in the domain at hand. Since iconological interpretations do not have a fixed structure, no automatic recognitions were implemented to extract the knowledge. The entire process is based on the author’s qualitative reading and interpretation of the text. Therefore, it has to be considered that the statements obtained depend on the author’s subjective comprehension of Panofsky’s work. For this reason, the author is indicated as responsible for the graph created, whereas Panofsky’s text is always cited as the source of each statement. The data creation was conducted as follows. First, we created a tabular data structure reflecting the ontology structure, including information on each level of interpretation. Second, we interpreted Panofsky’s claims and described the features of interest according to the data structure. The quality of the data was assessed through values validation through controlled lists. Finally, the data were converted into RDF³⁹ according to the modeling of the ontology described in Section 5 and reusing CIDOC-CRM for the description of artworks’ metadata. As a result, the dataset contains a total of 28,864 triples about 152 artworks, 1,980 interpretations, and 928 subjects. Additional statistics about the dataset can be found in Table 7.

The ontology was evaluated on Panofsky’s dataset through the CQs listed in Table 3. Each SAMOD iteration corresponds to a group of questions at levels 0, 1, 2, or 3, and expected results for each CQ are described. To test the correctness of the single classes, the CQs were further subdivided into more detailed ones closer to the ontology structure. All unit tests that query the test dataset are available through GitHub in the form of Jupyter notebooks.⁴⁰ For each level, we describe one or two CQs. Table 8 contains an overview of the metadata about the artworks included in the queries.

To comply with RQ2, we propose a qualitative comparison between ICON and existing ontologies. We do so by describing an authoritative example from Panofsky’s bibliography, in which there are iconographical and iconological interpretations. For this purpose, we selected the most complete ontologies for iconographical descriptions, namely Visual Representation Ontology and Wikidata.

We selected the frontispiece of François Perrier’s “Segmenta nobilium signorum statuarum...” (1638 edition) depicting the iconography of Father Time, a subject that emerges from the Renaissance onward but, despite it originates from classical sources, was never visually represented in previous times. Panofsky reconstructed its genesis, claiming that it originated by the fusion of a medieval French iconography of Time represented with wings (Temps) and the sinister characteristics of Saturn (e.g., old age, scythe, the act of devouring his children) [37]. Late antique writers had already enriched the figure of the god Saturn with attributes referring to time (i.e., a dragon or snake biting its tail) or through a reinterpretation of the traditional Saturn’s attributes, such as the sickle, associated with the times recurring,⁴¹ or the act of devouring the children, reinterpreted as Time devouring “whatever he has created” [37, p. 74]. Panofsky claims that such evolution of the iconography of Time is evidence of the phenomenon of pseudomorphosis, according to which figures with a classical appearance did not exist in the classical visual arts, albeit they were described in classical literature (level 3). In addition to that, the example here described faithfully shows Cesare Ripa’s description of time the destroyer as a demon with iron teeth standing among ruins, a symbol of the fact that time ruins everything without any effort.⁴² Table 13 resumes the understanding of this artwork at each different level, whereas Figures 6 through 9 shows how this example can be modeled with the ontologies considered.

Table 14 gives an overview of the comparison results. It is apparent from these schemas that neither VIR nor Wikidata include properties or classes that represent a third-level meaning. Nevertheless, they do express important aspects of the domain. On the first level of description, VIR offers only the limited expressivity given by the class Iconographical Atom, which is intended to describe the physical portion of the artwork to which a subject is bound [8]. In contrast, Wikidata offers several level 1 specifications of the subjects through the qualifiers of the property wdt:P180 (e.g., “nudity” and “old” referred to the subject “man,” through the qualifier wdt:P1354 “shown with features”). Nevertheless, none of the ontologies distinguishes the types of objects described, failing to express immaterial items properly, such as actions or emotions (VIR) and the connections existing between these immaterial aspects and the subjects doing them. Therefore, ICON aims at solving this issue by introducing Panofsky’s concept of the Composition of Artistic Motifs. As a result, level 1 objects as depicted in the specific artwork cannot only be described in detail but also gathered in meaningful groups. This structure allows specifying the actors involved in the action, as shown in Figure 6 (the subjects “man,” “gnawing away,” and “statue” are part of the same Composition ART1282-COMP2), or the actors feeling emotions. At the second level of interpretation, with VIR it is possible to express important characteristics of the representation, such as attributes, personifications, symbols, places, and characters (Figure 8), but not events. Wikidata, on the contrary, tends not to specify the type of object depicted, even if a symbolical meaning can always be expressed through the qualifier wdt:P4878 “symbolizes” (Figure 9). The distinction between stories and allegories is not included in both of them, and, for different reasons, the contextual appearance of the subjects cannot be carefully described. This is due to the fact that in VIR, it is not possible to properly describe the subjects at the first level, whereas in Wikidata, the depicted subjects cannot be related to each other. Therefore, as shown by the example, the act of a person gnawing away at a statue cannot be related to the allegory of Time devouring the past in VIR, and cannot be recognized as an allegory in Wikidata. To solve that, the ICON Iconographical Recognition allows relating the second-level subjects to its level 1 representation, and the subjects can be part of Stories or Allegories. In addition, the n-ary class icon:Image allows separating the general description of subjects from the contextual one. In this way, subjects can be described carefully including characteristics that would be inappropriate to include in the vocabulary-level description of the subject considered, highlighting variations in their contextual representation. Whereas the VIR and Wikidata features described previously allow a description of the first two levels of interpretation, none of them represents the domain of knowledge of iconology by considering the third level of interpretation. Therefore, ICON introduces icon:Iconological Recognition relating a third-level meaning (concept or cultural phenomenon) to the whole artwork or to its specific parts.

Concerning the interpretation’s attribution of responsibility, both VIR and Wikidata allow registering the person responsible for the statement in different ways. In VIR, the person responsible and possible sources can be related to the class IC12 Iconographical Recognition, on which the whole recognition of the artwork’s content without a specification for every subject recognized depends. In Wikidata, every statement can be attributed to a person responsible, but information about sources is not always provided. A consistent difference lies in the fact that whereas VIR expresses the person responsible for the claim, Wikidata considers the person responsible for the data inserted, avoiding a possible interesting comparison of authoritative art historian claims. As a result, ICON introduces an interpretation for each statement, giving subjectivity and authoritativeness for each subject recognized. In this way, more agreeing claims can be expressed in the same recognition description, to better represent the realistic case in which an art historian agrees with the claim of others, quoting them and adding further interpretations. In addition, this structure fosters the interoperability of online sources and data integration.

A part from the evaluations made to verify the explicit goals declared in our research questions, we also evaluate some more technical aspects of our ontology using automatic tools and services. We validated our ontology syntax by the WRC RDF Validation Service.⁴³ No syntax problems were highlighted by this tool. Then, we evaluated the logical consistency of our ontology through the OOPS [41] tool⁴⁴ that provides feedback on the ontology in form of highlighted pitfalls of different levels of importance. Most of the issues raised by this tool do not come from our modeling; instead, they are linked to the reused ontologies that might have missing information (e.g., no ranges and domain in properties or inconsistent labeling). The only highlighted pitfall that was directly linked with classes developed by the ontology is “P30: Equivalent classes not explicitly declared.” This issue suggests the possibility that classes such as icon:Character and dul:Quality, dul:Role, and dul:Reference should be equivalent. At the same time, it suggests that icon:Image should be equivalent to sim:Simulacrum and that icon:Story should be equivalent to dul:Narrative. The first equivalence would be fundamentally wrong because the classes themselves do not represent the same concept. The same can be said for Image and Simulacrum, as in the Simulation Ontology a simulacrum is said to be a general symbol that could have different representations, which is why it is equivalent to our class icon:Symbol. An image is considered instead as a specific representation of a symbol in an artwork; therefore, it is not considered to be the general concept of the symbol itself. Furthermore, icon:Story and dul:Narrative might have similarities, but the latter class has no description, so we refrained from making ambiguous equivalences. Apart from these three cases, the ontology was evaluated pitfall-free in all the other aspects (considering issues that dealt with originally created classes only). Finally, we analyzed our ontology with the FOOPS [26] tool, which evaluates how much an ontology complies with the FAIR principles. Our ontology scored 90%. In particular, it received a score of 8.5 out of 9 on reusability, a score of 8 out of 9 on findability, a score of 3 out of 3 on interoperability, and a score of 2 out of 3 on accessibility. The main problem highlighted by this tool is that the ontology is not yet inserted in the linked open data vocabulary (lov).⁴⁵ This issue will be addressed in the future. Finally, we verified that the logical axioms of all the external imported classes and properties that were aligned to our ontology did not cause any inconsistency in ICON by running the Hermit Reasoner [28].