A Commonsense Reasoning Framework for Explanatory Emotion Attribution, Generation and Re-classification
Antonio Lieto, Gian Luca Pozzato, Stefano Zoia, Viviana Patti, Rossana Damiano
AA Commonsense Reasoning Framework for ExplanatoryEmotion Attribution, Generation and Re-classification
Antonio Lieto a,b , Gian Luca Pozzato a , Stefano Zoia a , Viviana Patti a , RossanaDamiano a a University of Turin, Department of Computer Science, Turin, Italy b ICAR-CNR, Palermo, Italy
Abstract
In this work we present an explainable system for emotion attribution and recom-mendation (called DEGARI and available here: http://di.unito.it/DEGARI )relying on a recently introduced commonsense reasoning framework (the T CL logic)which is based on a human-like procedure for the automatic generation of novel con-cepts in a Description Logics knowledge base. Starting from an ontological formaliza-tion of emotions based on the Plutchik’s model (known as ArsEmotica and availableat http://130.192.212.225/fuseki/ArsEmotica-core ), the system ex-ploits the logic T CL to automatically generate novel commonsense semantic represen-tations of compound emotions (e.g. Love as derived from the combination of Joy andTrust according to Plutchik’s model). The generated emotions corresponds to proto-types, i.e. commonsense representations of given concepts, and have been used toreclassify emotion-related contents in a variety of artistic domains, ranging from artdatasets to the editorial content available in RaiPlay, the online multimedia platformof RAI Radiotelevisione Italiana (the Italian public broadcasting company). We havetested our system (i) by reclassifying the available contents in the tested dataset withrespect to the new generated compound emotions (ii) with an evaluation, in the formof a controlled user study experiment, of the feasibility of using the obtained reclassifi-cations as recommended emotional content. The obtained results are encouraging andpave the way to many possible further improvements and research directions. Keywords:
Explainable AI, Commonsense reasoning, Knowledge Generation, Concept
Preprint submitted to Elsevier January 12, 2021 a r X i v : . [ c s . A I] J a n ombination
1. Introduction and Background
Emotions have been acknowledged as a key part of the aesthetic experience throughall ages and cultures, as witnessed by terms such as “sublime” [44] and “pathos” [43],associated with the experience of art since the ancient times. The advent of computa-tional tools and methods for investigating the way we respond to objects and situationshas paved the way to a deeper understanding of the intricate relationship between emo-tions and artistic content. For example, [55] have studied how art affects emotionalregulation by measuring the brain response through EEG: their research shows that,in comparison with photographs depicting real events, artworks determine strongerelectro-physiological responses; in parallel, [19] argue that the emotional response toart – measured through facial muscle movements – is attenuated in art critics, andstronger in non-expert, thus showing the universality and spontaneity of this response.The association between art and emotions is even stronger when the artistic expres-sion is conveyed by media, as in music and movies. For example, music has proven tobe an effective tool for emotion regulation: as demonstrated by [54], music can inducespecific emotional states in everyday situations, an effect which is sought for by theusers and can be exploited to create effective affective recommender systems [3]. Fi-nally, emotional engagement is of primary importance in narrative media, such as filmand television, as extensively investigated by a line of research which draws from bothfilm studies and emotion theories [47, 53].As a consequence of the multi-faceted, complex role played by emotions in theexperience of art and media, the investigation of this phenomenon with computationaltools has relied on a variety of models and methodologies, ranging from dimensionalmodels, better suited to investigate physiological, continuous correlate of emotions[45, 57, 29], to categorical models, which lend themselves to inspecting the consciouslevel of emotional experience [39, 12, 4]. Dimensional models typically measure theemotional engagement along the arousal and hedonic axes, and are useful to study howthe emotional response evolves over time. For example, [27] rely on crowdsourced2nnotations of tension, arousal and variance in audio pieces to realize sound-basedaffective interaction in games. Categorical models are useful to collect the audienceexperience as discrete emotional labels, and are easily mapped to textual descriptions ofemotions across languages. As exemplified by [31], discrete emotional labels, mergedfrom different categorical models (from [39] to [35]), can shed light on the receptionof art, letting correlations emerge between attributed emotions, liking and subjects.In many cases, the emotional response of the audience is conveyed through lan-guage, non only in textual media, but also in relation to art and other media (consider,for example, tags and social media comments concerning artworks and exhibitions).Automatically detecting affective states and emotions from text has gained consider-able attention over recent years, leading to the development of several resources - suchas annotated corpora, ontologies and lexicons within the Computational Linguisticscommunity [18, 34, 56]. Affective information expressed in texts is multi-faceted, andthe wide variety of affective linguistic resources developed in the last years community,mainly for English, but also for other languages, basically reflects such richness. Whenwe speak about affective states in the context of natural language communications, wemean to refer to several aspects, which vary in their degree of stability, such as: emo-tion, sentiment, personality, mood, attitudes or interpersonal stance. Given the widevariety of affective states, in recent years research has focused on a finer-grained inves-tigation of the role of emotions, as well as on the importance of other affect dimensionssuch as emotion intensity [32] or activation. Depending on the specific research goalsaddressed, one could be interested in issuing a discrete label describing the affectivestate expressed (frustration, anger, joy, etc.) in accordance to different contexts ofinteraction and tasks. Both basic emotion theories, in the Plutchik-Ekman [12] tradi-tion, and dimensional models of emotions, provided a precious theoretical groundingfor the development of lexical resources [49, 33, 30, 7, 50] and computational modelsfor emotion extraction. However, there is a general tendency to move towards richer,finer-grained models, possibly including complex emotions, especially in the contextof data-driven and task-driven approaches, where restricting the automatic detection toa small set of basic emotions would fall short to achieve the objective. This is also ourperspective. 3rom a computational perspective, the choice of the model of affect to be used inorder to give psychological grounding to the resource or the corpus to be developedis driven from, and highly intertwined with, the specific sentiment analysis task to beaddressed, which, in turn, usually depends on the application domain to be tackledand on the final purpose of mining affective contents in texts. In this sense, evaluat-ing the emotional responses of an audience in front of an artwork, with the purpose ofmonitoring the emotional impact of a cultural heritage exhibition on visitors [5], is dif-ferent from monitoring political sentiment or or mining the levels of anger in commentsthreads of software developers [15]. There are still few works and resources specifi-cally developed to address emotion detection in the art and media domain, includingthe work in [31], where authors described the WikiArt Emotions Dataset, that includesemotions annotations for thousands of pieces of art from the WikiArt.org’s collection,and the work in [5, 38] where the ArsEmotica framework is proposed, which relies onthe combined use of NLP affect resources and an ontology of emotions to enable anemotion-driven exploration of online art collections.The diversity of computational models implied by the analysis of the emotional re-sponse to art and media, and the applications that exploit this response to improve theuser experience – from learning to entertainment – witnesses the complexity of the un-derlying processes (aesthetic, self-regulatory, social and cultural). This diversity, how-ever, can be an obstacle to the development of models which work across domains andformats, preventing techniques from being transferred across similar tasks (e.g., emo-tion annotation and affective recommendation). In particular, the differences in emo-tion annotation between datasets can endanger the development and cross-validation ofnew techniques for analysing and exploiting emotions in art and media. In this sense,techniques for merging and extending emotional categories can be useful to overcomethese limitations. A notable example of such a comprehensive system is SenticNet [7],which relies on the Hourglass model [8]. The Hourglass model, recently revised andextended [52], is inspired by Plutchik’s model of emotions [40]. Such model, formal-ized in the ArsEmotica ontology and described in detail in section 4, can be representedas a wheel of emotions and is formed by: basic or primary emotions; opposite emo-tions; similarity between the emotions; compound emotions (or complex emotions)4enerated by the primary ones.Similarly to the SenticNet framework, our system also relies on Plutchik’s model.The choice of this model is based on the fact that it provides a recipe for the generationof compound emotions that is compliant with the commonsense reasoning frameworkof the T CL logic. As such, we exploited the reasoning mechanisms of T CL to generatethe compound emotions according to the Plutchik’s theory. In this paper, we illustrateand validate this approach by means of the DEGARI system (Dynamic Emotion Gen-erator And ReclassIfier) for emotion attribution and recommendation. In particular,we have exploited the generated compounds to automatically reclassify items in threedatasets in the artistic and media domains. As a result of this reclassification process,an emotional enrichment is obtained and new emotional labels are associated with theitems in the original datasets. Thanks to the properties of T CL framework, the resultsof such reclassication - as will be shown in the paper - are entirely explainable.The paper is organized as follows: after a brief overview of the rationale adoptedby our commonsense reasoning framework (Section 2), we present in Section 3 - forthe sake of self-containedness - a more detailed description of the T CL logic (by refer-ring to [25] for a complete explanation). In Section 4 we present the ontological modelArsEmotica (enriched with an emotional lexicon) formalizing the Plutchik’s theory ofemotions and used as a standard representation to leverage the reasoning capabilitiesof the T CL within the system DEGARI. Sections 5 and 6 present the DEGARI systemthat, starting from the basic emotions represented in ArsEmotica (and according to thePlutchik’s theory), generates compound emotions and uses these novel emotional cate-gories for artistic content reclassification. Finally, Section 7 shows both the outcome ofthe automatic and explainable reclassification obtained with DEGARI and the results ofa user study on 44 people showing the feasibility of using the obtained reclassificationsas recommended contents. Section 8 ends the paper.
2. Commonsense Concept Invention via Dynamic Knowledge Combination
The overall rationale assumed in the T CL reasoning framework is that the process ofautomatic generation of novel concepts within a knowledge base (also known as knowl- dge invention ) can be obtained, as happens in humans [25], by exploiting a process ofcommonsense conceptual combination. This generative phenomenon highlights somecrucial aspects of the knowledge processing capabilities in human cognition. Such abil-ity, in fact, concerns high-level capacities associated to creative thinking and problemsolving. Still, however, it represents an open challenge in the field of Artificial Intelli-gence (AI) [6]. Dealing with this problem, indeed, requires, from an AI perspective, theharmonization of two conflicting requirements that are hardly accommodated in sym-bolic systems [14]: the need of a syntactic and semantic compositionality (typical oflogical systems) and the one concerning the exhibition of typicality effects. Accordingto a well-known argument [36], in fact, prototypes (i.e. commonsense conceptual rep-resentations based on typical properties) are not compositional. The argument runs asfollows: consider a concept like pet fish . It results from the composition of the concept pet and of the concept fish . However, the prototype of pet fish cannot result from thecomposition of the prototypes of a pet and a fish: e.g. a typical pet is furry and warm, atypical fish is grayish, but a typical pet fish is neither furry and warm nor grayish (typi-cally, it is red). The pet fish phenomenon is a paradigmatic example of the difficulty toaddress when building formalisms and systems trying to imitate this combinatorial hu-man ability. In this paper, we exploit the recently introduced nonmonotonic extensionof Description Logics T CL (typicality-based compositional logic, introduced in [25]),that is able to account for this type of human-like concept combination . More specif-ically, we show how it can be used as a tool for the generation of novel compoundemotions and, as a consequence, for the suggestion of novel emotion-related contents.In this logic, “typical” properties can be directly specified by means of a “typi-cality” operator T enriching the underlying Description Logic (from now on, DL forshort), and a TBox can contain inclusions of the form T ( C ) (cid:118) D to represent that Other works have already shown how such logic can be used to model complex cognitive phenomena[25], creative problem solving [26, 22] and to build intelligent applications in the field of computationalcreativity [24]. Alternative approaches to the problem of commonsense conceptual combination have beenrecently discussed in [13], [21], [10]. The main advantages of T CL with respect to such approaches aredetailed in [25]. C s are also Ds ”. As a difference with standard DLs, in the logic T CL onecan consistently express exceptions and reason about defeasible inheritance as well.Typicality inclusions are also equipped by a real number p ∈ (0 . , representing theprobability/degree of belief in such a typical property: this allows us to define a seman-tics inspired to the DISPONTE semantics [41] characterizing probabilistic extensionsof DLs, which in turn is used in order to describe different scenarios where only sometypicality properties are considered. Given a KB containing the description of twoconcepts C H and C M occurring in it, we then consider only some scenarios in orderto define a revised knowledge base, enriched by typical properties of the combinedconcept C (cid:118) C H (cid:117) C M by also implementing a heuristics coming from the cognitivesemantics.By relying on T CL , this work introduce the system DEGARI which, first, automati-cally builds prototypes of existing compound emotions by extracting information about concepts or properties by relying on the ArsEmotica ontology enriched with the NRCEmotion Intensity Lexicon [32] (associating, in descending order of frequency, wordsto emotional concepts). In this setting, words with the highest frequencies of associa-tion to emotional concepts have been used as typical features of the basic emotions inthe Plutchik model. Such prototypes of basic emotions have been formalized by meansof a T CL knowledge base, whose TBox contains both rigid inclusions of the form BasicEmotion (cid:118)
Concept , in order to express essential desiderata but also constraints, for instance Joy (cid:118)
PositiveEmotion as well as prototypical properties of the form p :: T ( BasicEmotion ) (cid:118) TypicalConcept , representing typical concepts of a given emotion, where p is a real number in the range (0 . , , expressing the frequency of such a concept in items belonging to that emo-tion: for instance, .
72 :: T ( Surprise ) (cid:118) Delight is used to express that the typicalfeature of being surprised contains/refers to the emotional concept Delight with a fre-quency/probability/degree of belief of the .Given the ArsEmotica knowledge base (see Section 4) with the prototypical de-7criptions of basic emotions, DEGARI exploits the reasoning capabilities of the logic T CL in order to generate new derived emotions as the result of the creative combinationof two (or even more) basic or derived ones. DEGARI also reclassifies the artistic andmultimedia contents taking the new, derived emotions into account. Intuitively, an itemof the tested dataset belongs to the new generated emotion if its metadata (name, de-scription, title) contain all the rigid properties as well as at least the of the typicalproperties of such a derived emotion. In this respect, DEGARI can be seen as a “whitebox” recommender system, able to suggest to its users artistic contents belonging tonew emotions by providing an explanation of such a recommendation.We have tested DEGARI by performing two different kinds of evaluation that arereported and discussed in Section 7, namely an automatic evaluation, and an evaluationof the satisfaction of users, showing promising results. In the following, to make thepaper self-contained we recall in more detail the main features of the above described T CL logic.
3. THE DESCRIPTION LOGIC T CL FOR CONCEPT COMBINATION
The logic T CL [25], used by the system DEGARI as the basis for the generation ofnew compound emotions combines three main ingredients. The first one relies on theDL of typicality ALC + T R introduced in [16], which allows to describe the protoype of a concept. In this logic, “typical” properties can be directly specified by meansof a “typicality” operator T enriching the underlying DL, and a TBox can containinclusions of the form T ( C ) (cid:118) D to represent that “typical C s are also Ds ”. As adifference with standard DLs, in the logic ALC + T R one can consistently expressexceptions and reason about defeasible inheritance as well. For instance, a knowledgebase can consistently express that “normally, athletes are fit”, whereas “sumo wrestlersusually are not fit” by T ( Athlete ) (cid:118) Fit and T ( SumoWrestler ) (cid:118) ¬ Fit , given that
SumoWrestler (cid:118)
Athlete . The semantics of the T operator is characterized by theproperties of rational logic [20], recognized as the core properties of nonmonotonicreasoning. ALC + T R is characterized by a minimal model semantics corresponding toan extension to DLs of a notion of rational closure as defined in [20] for propositional8ogic: the idea is to adopt a preference relation among ALC + T R models, whereintuitively a model is preferred to another one if it contains less exceptional elements,as well as a notion of minimal entailment restricted to models that are minimal withrespect to such preference relation. As a consequence, T inherits well-establishedproperties like specificity and irrelevance : in the example, the logic ALC + T R allowsus to infer T ( Athlete (cid:117)
Bald ) (cid:118) Fit (being bald is irrelevant with respect to being fit)and, if one knows that Hiroyuki is a typical sumo wrestler, to infer that he is not fit,giving preference to the most specific information.As a second ingredient, we consider a distributed semantics similar to the one ofprobabilistic DLs known as DISPONTE [42], allowing to label inclusions T ( C ) (cid:118) D with a real number between 0.5 and 1, representing its degree of belief/probability, as-suming that each axiom is independent from each others. Degrees of belief in typicalityinclusions allow to define a probability distribution over scenarios : roughly speaking, ascenario is obtained by choosing, for each typicality inclusion, whether it is consideredas true or false. In a slight extension of the above example, we could have the need ofrepresenting that both the typicality inclusions about athletes and sumo wrestlers havea degree of belief of , whereas we also believe that athletes are usually young witha higher degree of , with the following KB:(1) SumoWrestler (cid:118)
Athlete (2) . T ( Athlete ) (cid:118) Fit (3) . T ( SumoWrestler ) (cid:118) ¬ Fit (4) .
95 :: T ( Athlete ) (cid:118) YoungPerson
We consider eight different scenarios, representing all possible combinations of typi-cality inclusion: as an example, { ((2) , , ((3) , , ((4) , } represents the scenario inwhich (2) and (4) hold, whereas (3) does not. Obviously, (1) holds in every scenario,since it represents a rigid property, not admitting exceptions. We equip each scenariowith a probability depending on those of the involved inclusions: the scenario of theexample has probability . × . (since 2 and 4 are involved) × (1 − . (since 3is not involved) = 0 .
152 = 15 . . Such probabilities are then taken into account inorder to choose the most adequate scenario describing the prototype of the combined9oncept.As a third element of the proposed formalization we employ a method inspiredby cognitive semantics [17] for the identification of a dominance effect between theconcepts to be combined: for every combination, we distinguish a HEAD, representingthe stronger element of the combination, and a MODIFIER. The basic idea is: given aKB and two concepts C H (HEAD) and C M (MODIFIER) occurring in it, we consideronly some scenarios in order to define a revised knowledge base, enriched by typicalproperties of the combined concept C (cid:118) C H (cid:117) C M .Let us now present the logic T CL more precisely. The language of T CL extends thebasic DL ALC by typicality inclusions of the form T ( C ) (cid:118) D equipped by a realnumber p ∈ (0 . , – observe that the extreme . is not included – representing itsdegree of belief, whose meaning is that “we believe with degree/probability p that,normally, C s are also D s” Definition 3.1 (Language of T CL ). We consider an alphabet of concept names C , ofrole names R , and of individual constants O . Given A ∈ C and R ∈ R , we define: C, D := A | (cid:62) | ⊥ | ¬ C | C (cid:117) C | C (cid:116) C | ∀ R.C | ∃
R.C
We define a knowledge base K = (cid:104)R , T , A(cid:105) where: • R is a finite set of rigid properties of the form C (cid:118) D ; • T is a finite set of typicality properties of the form p :: T ( C ) (cid:118) D where p ∈ (0 . , ⊆ R is the degree of belief of the typicality inclusion; • A is the ABox, i.e. a finite set of formulas of the form either C ( a ) or R ( a, b ) , where a, b ∈ O and R ∈ R . The reason why we only allow typicality inclusions equipped with probabilities p > . is due to oureffort of integrating two different semantics: typicality based logic and DISPONTE. In particular, as detailedin [25] this choice seems to be the only one compliant with both the formalisms. On the contrary, it wouldbe misleading to also allow low degrees of belief for typicality inclusions, since typical knowledge is knownto come with a low degree of uncertainty.
10 model M in the logic T CL extends standard ALC models by a preference relationamong domain elements as in the logic of typicality [16]. In this respect, x < y meansthat x is “more normal” than y , and that the typical members of a concept C are theminimal elements of C with respect to this relation . An element x ∈ ∆ I is a typicalinstance of some concept C if x ∈ C I and there is no C -element in ∆ I more normal than x . Formally: Definition 3.2 (Model of T CL ). A model M is any structure (cid:104) ∆ I , <, . I (cid:105) where: • ∆ I is a non empty set of items called the domain; • < is an irreflexive, transitive, well-founded and modular (for all x, y, z in ∆ I , if x < y then either x < z or z < y ) relation over ∆ I ; • . I is the extension function that maps each atomic concept C to C I ⊆ ∆ I , andeach role R to R I ⊆ ∆ I × ∆ I , and is extended to complex concepts as follows: – ( ¬ C ) I = ∆ I \ C I – ( C (cid:117) D ) I = C I ∩ D I – ( C (cid:116) D ) I = C I ∪ D I – ( ∃ R.C ) I = { x ∈ ∆ I | ∃ ( x, y ) ∈ R I such that y ∈ C I } – ( ∀ R.C ) I = { x ∈ ∆ I | ∀ ( x, y ) ∈ R I we have y ∈ C I } – ( T ( C )) I = M in < ( C I ) , where M in < ( C I ) = { x ∈ C I | (cid:64) y ∈ C I s.t. y ALC + T R has the same complexityof the standard ALC ), whereas adopting multiple preference relations could lead to higher complexities. 11 model M can be equivalently defined by postulating the existence of a function k M : ∆ I (cid:55)−→ N , where k M assigns a finite rank to each domain element [16]: therank of x is the length of the longest chain x < . . . < x from x to a minimal x , i.e.such that there is no x (cid:48) such that x (cid:48) < x . The rank function k M and < can be definedfrom each other by letting x < y if and only if k M ( x ) < k M ( y ) . Definition 3.3 (Model satisfying a knowledge base in T CL ). Let K = (cid:104)R , T , A(cid:105) bea KB. Given a model M = (cid:104) ∆ I , <, . I (cid:105) , we assume that . I is extended to assign adomain element a I of ∆ I to each individual constant a of O . We say that: • M satisfies R if, for all C (cid:118) D ∈ R , we have C I ⊆ D I ; • M satisfies T if, for all q :: T ( C ) (cid:118) D ∈ T , we have that T ( C ) I ⊆ D I , i.e. M in < ( C I ) ⊆ D I ; • M satisfies A if, for each assertion F ∈ A , if F = C ( a ) then a I ∈ C I ,otherwise if F = R ( a, b ) then ( a I , b I ) ∈ R I . Even if the typicality operator T itself is nonmonotonic (i.e. T ( C ) (cid:118) E does notimply T ( C (cid:117) D ) (cid:118) E ), what is inferred from a KB can still be inferred from any KB’with KB ⊆ KB’, i.e. the resulting logic is monotonic. As already mentioned, in orderto perform useful nonmonotonic inferences, in [16] the authors have strengthened theabove semantics by restricting entailment to a class of minimal models. Intuitively,the idea is to restrict entailment to models that minimize the atypical instances of aconcept . The resulting logic corresponds to a notion of rational closure on top of ALC + T R . Such a notion is a natural extension of the rational closure constructionprovided in [20] for the propositional logic. This nonmonotonic semantics relies onminimal rational models that minimize the rank of domain elements . Informally, given It is worth noticing that here the degree q does not play any role. Indeed, a typicality inclusion T ( C ) (cid:118) D holds in a model only if it satisfies the semantic condition of the underlying DL of typical-ity, i.e. minimal (typical) elements of C are elements of D . The degree of belief q will have a crucial role inthe application of the distributed semantics, allowing the definition of scenarios as well as the computationof their probabilities. x has rank 2 (because forinstance z < y < x ) , and another in which it has rank 1 (because only y < x ), weprefer the latter, as in this model the element x is assumed to be “more typical” thanin the former. Query entailment is then restricted to minimal canonical models . Theintuition is that a canonical model contains all the individuals that enjoy properties thatare consistent with KB. This is needed when reasoning about the rank of the concepts:it is important to have them all represented.Given a KB K = (cid:104)R , T , A(cid:105) and given two concepts C H and C M occurring in K ,the logic T CL allows defining a prototype of the combined concept C as the combinationof the HEAD C H and the MODIFIER C M , where the typical properties of the form T ( C ) (cid:118) D (or, equivalently, T ( C H (cid:117) C M ) (cid:118) D ) to ascribe to the concept C areobtained by considering blocks of scenarios with the same probability, in decreasingorder starting from the highest one. We first discard all the inconsistent scenarios, then:• we discard those scenarios considered as trivial , consistently inheriting all theproperties from the HEAD from the starting concepts to be combined. Thischoice is motivated by the challenges provided by task of commonsense concep-tual combination itself: in order to generate plausible and creative compounds itis necessary to maintain a level of surprise in the combination. Thus both sce-narios inheriting all the properties of the two concepts and all the properties ofthe HEAD are discarded since they prevent this surprise;• among the remaining ones, we discard those inheriting properties from the MOD-IFIER in conflict with properties that could be consistently inherited from theHEAD;• if the set of scenarios of the current block is empty, i.e. all the scenarios havebeen discarded either because trivial or because preferring the MODIFIER, werepeat the procedure by considering the block of scenarios, having the immedi-ately lower probability.Remaining scenarios are those selected by the logic T CL . The ultimate output of ourmechanism is a knowledge base in the logic T CL whose set of typicality properties is13nriched by those of the compound concept C . Given a scenario w satisfying the aboveproperties, we define the properties of C as the set of inclusions p :: T ( C ) (cid:118) D , forall T ( C ) (cid:118) D that are entailed from w in the logic T CL . The probability p is such that:• if T ( C H ) (cid:118) D is entailed from w , that is to say D is a property inheritedeither from the HEAD (or from both the HEAD and the MODIFIER), then p corresponds to the degree of belief of such inclusion of the HEAD in the initialknowledge base, i.e. p : T ( C H ) (cid:118) D ∈ T ;• otherwise, i.e. T ( C M ) (cid:118) D is entailed from w , then p corresponds to the degreeof belief of such inclusion of a MODIFIER in the initial knowledge base, i.e. p : T ( C M ) (cid:118) D ∈ T .The knowledge base obtained as the result of combining concepts C H and C M into the compound concept C is called C - revised knowledge base, and it is defined asfollows: K C = (cid:104)R , T ∪ { p : T ( C ) (cid:118) D } , A(cid:105) , for all D such that either T ( C H ) (cid:118) D is entailed in w or T ( C M ) (cid:118) D is entailed in w , and p is defined as above.As an example, consider the following version of the above mentioned Pet-Fish problem. Let KB contains the following inclusions: Fish (cid:118) LivesInWater (1) . T ( Fish ) (cid:118) Greyish (2) . T ( Fish ) (cid:118) Scaly (3) . T ( Fish ) (cid:118) ¬ Affectionate (4) . T ( Pet ) (cid:118) ¬ LivesInWater (5) . T ( Pet ) (cid:118) LovedByKids (6) . T ( Pet ) (cid:118) Affectionate (7)representing that a typical fish is greyish (2) , scaly (3) and not affectionate (4) , whereasa typical pet does not live in water (5) , is loved by kids (6) and is affectionate (7) .Concerning rigid properties, we have that all fishes live in water (1) . The logic T CL Pet and Fish , by using the latter as the HEAD and the formeras the MODIFIER. The prototypical Pet-Fish inherits from the prototypical fish thefact that it is scaly and not affectionate, the last one by giving preference to the HEADsince such a property conflicts with the opposite one in the modifier (a typical pet isaffectionate). The scenarios in which all the three typical properties of a typical fish areinherited by the combined concept are considered as trivial and, therefore, discarded,as a consequence the property having the lowest degree ( Greyish with degree . ) isnot inherited. The prototypical Pet-Fish inherits from the prototypical pet only prop-erty (6) , since (5) conflicts with the rigid property (1) , stating that all fishes (then, alsopet fishes) live in water, whereas (7) is blocked, as already mentioned, by the HEAD/-MODIFIER heuristics. Formally, the Pet (cid:117) Fish -revised knowledge base contains, inaddition to the above inclusions, the following ones: . T ( Pet (cid:117) Fish ) (cid:118) Scaly (3’) . T ( Pet (cid:117) Fish ) (cid:118) ¬ Affectionate (4’) . T ( Pet (cid:117) Fish ) (cid:118) LovedByKids (6’)In [25] it has been also shown that reasoning in T CL remains in the same complexityclass of standard ALC Description Logics.For the purposes of this paper it is worth-noticing that, as mentioned, the T CL rea-soning framework presented in this section has been applied, via the DEGARI system,to the generation of new compound emotions by starting from the affective ontologicalknowledge base named ArsEmotica. Such ontological model is described in the nextsection. 4. The ArsEmotica Ontological Model enriched with the NRC Emotion IntensityLexicon The affective knowledge leveraged by the T CL logic via the DEGARI system isencoded in an ontology of emotional categories based on Plutchik’s psychological cir-15umplex model [39] called ArsEmotica and includes also concepts from the Hour-glass model [8]. The ontology structures emotional categories in a taxonomy, whichcurrently includes 32 emotional concepts. The design of the emotional categories taxo-nomic structure, of the disjunction axioms and of the object and data properties mirrorsthe main features of Plutchik’s circumplex model. As already mentioned, such modelcan be represented as a wheel of emotions (see figure 1) and encodes the followingelements: Figure 1: The Wheel of Emotion of the Plutchik’s Model The ArsEmotica ontology is available here: http://130.192.212.225/fuseki/ArsEmotica-core and queryable via SPARQL endpoint at: http://130.192.212.225/fuseki/dataset.html?tab=query&ds=/ArsEmotica-core 16 Basic or primary emotions: joy, trust, fear, surprise, sadness, disgust, anger,anticipation; in the color wheel this is represented by differently colored sectors.• Opposites: basic emotions can be conceptualized in terms of polar opposites: joyversus sadness, anger versus fear, trust versus disgust, surprise versus anticipa-tion.• Intensity: each emotion can exist in varying degrees of intensity; in the wheelthis is represented by the vertical dimension.• Similarity: emotions vary in their degree of similarity to one another; in thewheel this is represented by the radial dimension.• Complex emotions: complex emotions are a mixtures of the primary emotions;looking at the Plutchik’s wheel, the height emotions in the blank spaces are com-positions of basic emotions called primary dyads.We have chosen to encode the Plutchik’s model in the ontology for several rea-sons. First, it is well-grounded in psychology and general enough to guarantee a widecoverage of emotions. This is important for implementing successful strategies aimedat mapping tags to the emotional concepts of the ontology. Second, as already men-tioned, the Plutchik’s “wheel of emotions” is perfectly compliant with the generativemodel underlying the T CL logic. Finally, it encodes interesting notions, e.g. emotionalpolar opposites, which can be exploited for finding new relations among artworks.Within the ArsEmotica ontology, the class Emotion is the root for all the emotionalconcepts. The Emotion’s hierarchy includes all the 32 emotional categories presentedas distinguished labels in the model. In particular, the Emotion class has two disjointsubclasses: BasicEmotion and ComplexEmotion. Basic emotions of the Plutchik’smodel are direct sub-classes of BasicEmotion. Each of them is specialized again intotwo subclasses representing the same emotion with weaker or the stronger intensity(e.g. the basic emotion Joy has Ecstasy and Serenity as sub-classes). Therefore, wehave 24 emotional concepts subsumed by the BasicEmotion concept. Instead, the classCompositeEmotion has 8 subclasses, corresponding to the primary dyads. Other rela-tions in the Plutchik’s model have been expressed in the ontology by means of object17roperties: the hasOpposite property encodes the notion of polar opposites; the hasSib-ling property encodes the notion of similarity and the isComposedOf property encodesthe notion of composition of basic emotions. Moreover, a data type property hasS-core was introduced to link each emotion with an intensity value i mapped into theabove mentioned Hourglass model. Due to the need of modeling the link among wordsin a language and the emotions they refer to, the ArsEmotica Ontology is also inte-grated with the ontology framework LExicon Model for ONtologies (LEMON) [28].In particular, such integration allows to explicitly differentiate between the languagelevel (lexicon based) and the conceptual one representing the emotional concepts [38].Within this enriched framework, it is possible to associate a plethora of emotionalwords, with the encoding of language information, to the corresponding emotionalconcepts. Apart from the already available linking with the lexical resources suchas WordNet-Affect, MultiWordNet, we have now equipped the ArsEmotica emotionalconcepts with the NRC Emotion Intensity Lexicon mentioned above [32]. Such lexi-con provides a list of English words, each with real-values representing intensity scoresfor the eight basic emotions of Plutchik’s theory. The lexicon includes close to 10,000words including terms already known to be associated with emotions as well as termsthat co-occur in Twitter posts that convey emotions. The intensity scores were obtainedvia crowdsourcing, using best-worst scaling annotation scheme. For our purposes, weconsidered the most frequent terms available in such lexicon (and associated to the ba-sic emotions of the Plutchik’s wheel) as typical features of such emotions. In this way,once the prototypes of the basic emotional concepts were formed, the T CL reasoningframework was used to generate the compound emotions. 5. DEGARI: GENERATING NOVEL EMOTIONS from ArsEmotica In this section we describe DEGARI, the system exploiting the logic T CL on the Ar-sEmotica knowledge base in order to generate and suggest novel emotion related con-tents and tested on the RaiPlay catalog , as well as on two artwork datasets: WikiArt and ArsMeteo [1, 5]. DEGARI is implemented in Python and it makes useof the library owlready2 for relying on the services of efficient DL reasoners (likeHermiT).DEGARI’s prototypes generation proceeds in two steps: in the first one, it buildsa prototypical description of basic emotions in the language of the logic T CL , in orderto describe their typical properties; as a second step, it exploits the above describedreasoning mechanism of such a Description Logic in order to combine the prototypi-cal descriptions of pairs of basic emotions, generating the prototypical description ofcompound emotions. As mentioned above, the obtained ontology is then tested byre-classifying the items belonging to RaiPlay, Wiki Art and ArsMeteo keeping the gen-erated compound emotions into account: this allows us to describe a novel and com-pletely explainable recommending system, which is able to suggest items belongingalso to compound emotions.Concerning the first step, DEGARI builds a knowledge base in the logic T CL char-acterized by typicality inclusions of the form p :: T ( BasicEmotion ) (cid:118) Property where BasicEmotion is one of the eight basic emotions of Plutchik’s theory: Joy,Trust, Fear, Surprise, Sadness, Disgust, Anger, and Anticipation. Typical propertiesare selected from the list of words characterizing each basic emotion in the NRC Emo-tion Intensity Lexicon where, as already mentioned, the probability p represents theintensity scores for the emotion. In detail, for each basic emotion, we consider the sixproperties/words having the highest scores.As an example, consider the basic emotion Joy . The words having the highestscores are happiness ( . ), bliss ( . ), to celebrate ( . ), jubilant ( . ), ecstatic( . ), and euphoria ( . ). Therefore, the knowledge base generated by DEGARIwill contain, among others, the following inclusions: Joy (cid:118) ¬ Holocaust http://saifmohammad.com/WebPages/wikiartemotions.html https://pythonhosted.org/Owlready2/ . 98 :: T ( Joy ) (cid:118) Happiness . 97 :: T ( Joy ) (cid:118) Bliss . 97 :: T ( Joy ) (cid:118) Celebrating . 97 :: T ( Joy ) (cid:118) Jubilant . 95 :: T ( Joy ) (cid:118) Ecstatic . 94 :: T ( Joy ) (cid:118) Elation DEGARI then computes novel compound emotions by combining existing ones (byusing the same logical procedure of the pet-fish problem). As an example, let us con-sider the combination of the above basic emotion Joy with Fear , whose prototypicaldescription is as follows: . 96 :: T ( Fear ) (cid:118) Kill . 95 :: T ( Fear ) (cid:118) Annihilate . 95 :: T ( Fear ) (cid:118) Terror . 98 :: T ( Fear ) (cid:118) Torture . 97 :: T ( Fear ) (cid:118) Terrorist . 97 :: T ( Fear ) (cid:118) Horrific In order to obtain a description of the compound emotion Guilt as the result ofthe combination of the two basic emotions ( Joy (cid:117) Fear ) in the logic T CL , DEGARIcombines the two basic emotions by implementing a variant of CoCoS [23], a Pythonimplementation of reasoning services for the logic T CL in order to exploit efficientDLs reasoners for checking both the consistency of each generated scenario and theexistence of conflicts among properties, following the line of the system DENOTER[9]. More in detail, DEGARI considers both the available choices for the HEAD andthe MODIFIER, and it allows to restrict its concern to a given and fixed number ofinherited properties. The combined emotion Guilt has the following T CL description(concept Joy (cid:117) Fear ): . 98 :: T ( Joy (cid:117) Fear ) (cid:118) Happiness . 97 :: T ( Joy (cid:117) Fear ) (cid:118) Celebrating . 97 :: T ( Joy (cid:117) Fear ) (cid:118) Bliss . 98 :: T ( Joy (cid:117) Fear ) (cid:118) Torture . 97 :: T ( Joy (cid:117) Fear ) (cid:118) Terrorist . 97 :: T ( Joy (cid:117) Fear ) (cid:118) Horrific Obviously, rigid properties of basic emotions (if any) are inherited by the compoundemotion (in the example, Joy (cid:117) Fear (cid:118) ¬ Holocaust ), and this avoids the system toconsider any inconsistent typical property even if it has the highest probability.It is worth noticing that the properties of the derived emotion are still expressedin the language of the logic T CL , therefore the combined emotion, Guilt in the exam-ple, can be used to be further combined with another emotion, in order to iterate theprocedure. 6. Reclassification of Emotion-related Content based on DEGARI By starting from the generated prototypes of the compound emotions in ArsEmot-ica, DEGARI is also able to perform an emotion-oriented reclassification of the itemsof the consideted datasets.In particular, DEGARI employs two different strategies to extract metadata fromthe items to reclassify. In a first case (e.g. for the datasets of ArsMeteo and WikiArt)the metadata are either stored in the provided resource (e.g. in WikiArt) or are theresult of a social tagging activty based on the artistic community. In the second case(e.g. in the case of the RaiPlay dataset) the metadata associated to every and each item(title, name of the program/episode, description of the program/series, description ofthe episode) are extracted from a crawler. Such metadata are then used to generatethe typical description of the items via the computation of the most frequent termsretrieved in their textual description (the assumption is that the most frequently usedterms to describe an item are also the ones that are more typically associated to them).The frequencies are computed as the proportion of each property with respect to the setof all properties characterizing the item, in order to compare them with the propertiesof the derived emotion. If the item contains all the rigid properties and at least the of the typical properties of the compound emotion under consideration, then the itemis classified as belonging to it. Last, DEGARI suggests the set of classified contents,21n a descending order of compatibility, where a rank of compatibility of a single itemwith respect to an emotion is intuitively obtained as the sum of the frequencies of“compatible” concepts, i.e. concepts belonging to both the item and the prototypicaldescription of the genre. Formally: Definition 6.1. Given an item m , let DerivedEmotion be a compound emotion gener-ated form the ArsEmotica mode as defined in Section 5 and let S m be the set of wordsoccurring in m . Given a knowledge base KB of compound emotions built by DEGARI,we say that m is compatible with DerivedEmotion if the following conditions hold: m contains all rigid properties of DerivedEmotion , i.e. { C | DerivedEmotion (cid:118) C ∈ KB } ⊆ S m m contains at least the of typical properties of DerivedEmotion , i.e. | S m ∩ S DerivedEmotion || S DerivedEmotion | ≥ . , where S DerivedEmotion is the set of typical properties of DerivedEmotion . As another example, consider the derived emotion Joy (cid:117) Surprise , which in Plutchik’swheel corresponds to the combined emotion “delight”. The knowledge base in the logic T CL describing such a compound emotion is as follows: . 98 :: T ( Joy (cid:117) Surprise ) (cid:118) Happiness . 97 :: T ( Joy (cid:117) Surprise ) (cid:118) Bliss . 97 :: T ( Joy (cid:117) Surprise ) (cid:118) Celebrating . 97 :: T ( Joy (cid:117) Surprise ) (cid:118) Jubilant . 95 :: T ( Joy (cid:117) Surprise ) (cid:118) EstaticJoy (cid:117) Surprise (cid:118) ¬ HolocaustJoy (cid:117) Surprise (cid:118) ¬ Anticipation For instance, the multimedia item “ `E arrivata la felicit`a” (“Happiness has come”)( ) from theRaiPlay dataset is reclassified in the novel, generated emotion Joy (cid:117) Surprise , since:• all rigid properties of both basic emotions are satisfied, that is to say neither Holocaust nor Anticipation belong to the properties extracted for the item;22 more than the of the typical properties of the compound emotion are sat-isfied; in particular, “ `E arrivata la felicit`a” has Happiness ( . ) and Surprise ( . ).This item will be then recommended by DEGARI as shown in Figure 2.Figure 2 also shows how DEGARI can be considered as an explainable AI sys-tem: indeed, an explanation of the reasons why the multimedia item has been reclas-sified in the compound concept is provided, in order to let the user be aware of theprocedure of the system. As a difference with “black box” approaches, DEGARI ex-plicitly reports that the “instance’s description has the following word(s) in commonwith category prototype”, followed by the two above matching properties surprise andhappiness. Moreover, the whole procedure is completely known and could be usedto further expand the feedback provided by system: from the axiom system and thecorresponding semantics of the Description Logic with typicality to the DISPONTEsemantics adopted by the logic T CL in order to compute the prototypical description ofthe compound emotion. 7. Evaluation DEGARI evaluation has been carried out on three different datasets and evaluatedin a twofold way. The datasets considered (described in detail below) are: ArsMe-teo, the RaiPlay catalog and Wiki Art Emotion. The evaluation has concerned a first,completely automatic, test consisting in calculating the percentage of the reclassifieditems within the novel hybrid emotion classes generated by the system via T CL . In thiscase, as a positive indicator has been considered the spread of the reclassified itemsalong most of the concepts of the wheel of emotion. This aspect, indeed, shows howthe created prototypes of the compound emotions are mostly meaningful and able toreclassify the artistic content available in three original datasets.A second evaluation, aimed at measuring the satisfaction of the potential users ofthe system when exposed to the contents of the novel categories suggested by DE-23 igure 2: An Example of the Laconic Explanation provided by DEGARI about the reasonsdetermining the reclassification of an item in the RaiPlay dataset. involving 44 subjects, who evaluated a total of 30recommendations generated by the system. All the participants were recruited onlineusing an availability sampling strategy. Participants were all naive to the experimentalprocedure and to the aims of the study.In the following we briefly describe the adopted datasets: two of them are art-related ones (ArsMeteo and WikiArt Emotion), while the RaiPlay dataset contains allthe multimedia items (e.g. movies, tv series, tv shows, documentaries etc.) availableon the online multimedia platform or RAI, Radiotelevisione Italiana. ArsMeteo [1] is an art portal for sharing artworks and their emerging, connectivemeanings. Its development is leaded by a non-profit cultural organization called As-sociazione Culturale ArsMeteo (AMA), based in Turin, Italy. It enables the collectionof digital (or digitalized) artworks and performances, belonging to a variety of artis-tic forms including poems, videos, pictures and music. Meanings are given by thetagging activity of the community. All contents are accessible as “digital commons”.Currently, the portal has collected over 350,000 visits and gathers a collection of over9,000 artifacts produced by 307 artists; it has collected almost 38,000 tags. The RaiPlay dataset is composed by 4,612 multimedia items extracted from RaiPlay : the online platform of RAI. Such dataset containsdifferent types of multimedia content grouped in six main narrative categories: Movies,Fiction, Kids, TV Series, Drama, Comedy.As mentioned, each multimedia item/episode is equipped by some information,namely: title, name of the program/episode, description of the program/series, descrip-tion of the episode. Such descriptions are used by DEGARI to extract the relevant This is one of the most commonly used methodology for the evaluation of recommender systems basedon controlled small groups analysis, see [46]. WikiArt Emotions is a dataset of 4,105 artworks with annotations for the emotionsevoked in the observer [31]. The artworks were selected from the online visual artencyclopedia WikiArt.org. Each piece of art is annotated for one or more of 20 emotioncategories (including neutral). Annotations were obtained via crowdsourcing, askingannotators all emotions evoked by the title of the artwork, the image of the artworkor the artwork as a whole. The annotators were also asked to point out if the artworkdepicted a face, or a human body but not a face: these additional information is includedin the dataset (if an artwork didn’t depict a face nor a body, it was marked as ”none”).In order to decide if an emotion applies or not to an artwork, the authors specified anaggregation threshold of : if at least of the responses indicated that a certainemotion applied, then the label was chosen. Other distributions of the dataset withdifferent aggregation thresholds ( and ) are available, but we chose to use the threshold version, as recommended by the authors of the dataset [31]. The obtained results for what concerns the automatic evaluation are presented in thetable below. Overall the figure shows that for two of the three datasets (ArsMeteo andRaiPlay) DEGARI is able to reclassify and spread the original items along the entirewheel of emotion assumed in the Plutchik’s model, thus allowing a more fine grainedcharacterization.In these two cases, the percentage of the reclassified items is of , and , respectively. On the other hand, the WkiArt Emotions dataset, contains orthogonalresults since, in this case, 16 out of the 31 generated compound emotions are filledwith reclassified items. In this case, however, a large part of the dataset ( , ) isinvolved in such a reclassification. 26 igure 3: Overall Table The main reason for these orthogonal results is in the kind of input considered byDEGARI. Indeed, while for ArsMeteo and RaiPlay the metadata associated to the itemsare either the result of a social tagging activity by a community of artists (like in Ar-sMeteo ) or the result of an information extraction pipeline (in RaiPlay); in WikiArt The ArsMeteo dataset has the additional difficulty of being an heterogeneous dataset, touching different igure 4: Top Reclassified Content of ArsMeteo in the Compound Emotional Classes generatedby DEGARI Emotions, all the metadata associated to the items are the result of a controlled crowd-sourcing activity based on predefined emotion tags.While this fact, on one hand, creates - for the WikiArt Emotions dataset - moreclean metadata and allows the reclassifications of most of the items available in thedataset, on the other hand forces the user to use a predefined vocabulary for annotationthat, as such, inhibits more free association that could have led to a wider reclassifica-tion and redistribution of the items along the entire Plutchik’s wheel. In all the 3 cases,however, most of the compound emotions generated by DEGARI are filled with new artistic genres (from poetry to literature to paintings). igure 5: Top Reclassified Content of RaiPlay in the Compound Emotional Classes generatedby DEGARI items.In the figures 4, 5, 6, finally, are also reported, for each dataset, the generated com-pound emotions that have received more reclassifications (the horizontal histogramsindicate the number of the reclassified items for each compound emotion). The goal of the user study was to assess the acceptance of the emotion categoriessuggested by DEGARI, with the ultimate goal of using the reclassifications producedby the system to improve the annotation of artworks and media, and, consequently, theapplications which depend on it, such as personalization and recommendation.29 igure 6: Top Reclassified Content of WikiArt Emotions in the Compound Emotional Classesgenerated by DEGARI Methods and material . The user study consisted in an online questionnaire (in Ital-ian). The questionnaire contained items, each represented by an image, and for themultimedia items, by the film poster accompanied by the link to the online player forwatching the content. For each item, the users received two questions: the first question(Question ) asked them to rate the association of the item with the emotional categoryprovided by DEGARI on 10-point scale; the second question (Question ) asked themto associate the item to additional emotion categories, taken from Plutchik’s model.Users were divided in 3 groups, each corresponding to a different set of items. Foreach dataset, the selected items were the ones ranked higher by DEGARI for each30 able 1: User ratings of the emotions proposed by DEGARI ArsMeteo WikiArt RaiPlay AllAverage rating . 47 6 . 12 6 . 42 6 . Standard deviation . 48 2 . 47 2 . . Median generated compound emotional category. Participants and procedure . The study involved users ( female, male). Con-cerning the age groups, users were below ; users were in the - age range; in the - range; in the - range; were older than . Users were randomlyassigned to the questionnaires. The first questionnaire was filled out by users; thesecond questionnaire was filled out by users; the third questionnaire was assignedto users. As a result, ratings and emotion categories were collected. Results and analysis . Concerning Question , the average rating assigned by the usersto the emotion category proposed by Degari was . , with only slight differences be-tween the datasets (see Table 1). The average rating was . for ArsMeteo, . forRaiPlay, and . for WikiArt. The standard deviation was . ( . for ArsMeteo, . for WikiArt, and . for RaiPlay), suggesting that the differences in ratings werelimited. Also, the median rating is for all data sets, with only proposed emotioncategories ( from Arsmeteo and from WikiArt) rated below .Concerning Question (namely, the additional emotions attached by the users to theitems), were attached to the items in ArsMeteo, to the items in WikiArt, and to the items in RaiPlay, yielding user emotion categories. The average num-ber of emotion categories per users was . . In order to investigate the overlappingbetween the set of emotion categories proposed by DEGARI for each item (apart forthe top ranked category tested through Question ), we compared the emotion cate-gories selected by the users with the ones proposed by DEGARI (Table 2). Data showthat 22.05% of the emotion categories additionally proposed by DEGARI for each item31 able 2: Overlapping of user tags with DEGARI emotions ArsMeteo WikiArt RaiPlay AllUser tags 308 308 449 1065 Proposed emotions 36 50 41 127 Overlapping . % % . % . %matched those selected by the users, with a higher value for ArsMeteo ( . %), anda lower value for WikiArt ( %) and RaiPlay ( . %). This datum is a positive onesince is concerns the non-top ranked emotional categories suggested by the system (forwhich the degree of acceptability by the users was always above 5 out of a 10-pointscale except for items, and with a median of 7 for every considered dataset).To conclude, the collected data suggest that the emotion categories proposed byDEGARI as a result of the reclassification process are generally accepted by the users,with few exceptions that deserve further investigation. The acceptance is clear for thetop ranked emotion, but an acceptable degree of acceptance can be inferred for theremaining suggested categories, for which an overlapping of % and more with theuser tags has been found for all datasets. 8. Discussion and Conclusion In this paper, we presented DEGARI: an explainable AI system relying on the T CL Description Logics and on the ArsEmotica knowledge base to generate, according tothe Plutchik’s theory of emotion, compound emotional concepts starting from the basicones. Such newly created categories, characterized by lexicon-based typical features,are then used in DEGARI to reclassify, in an emotional settings, the items of threedifferent datasets.Overall, DEGARI propose a transparent approach for emotionally-driven contentreclassification and its combinatorial reasoning engine could be useful for addressingthe very well known filter bubble effect [37] in recommender systems, by introducing32eeds of serendipity in content discovery by users. One fundamental discussion aboutthe applicability of DEGARI in practice is whether or not it represents a truly innova-tive technical solution for an emotion-based recommender system. According to [48]recommender systems “try to identify the need and preferences of users, filter the hugecollection of data accordingly and present the best suited option before the users byusing some well-defined mechanism”. Despite the huge amount of proposals, the mainfamilies of recommender systems can be identified as based on: i) collaborative filter-ing; ii) content-based filtering; iii) hybrid filtering. At their core of functioning, col-laborative filtering exploits similarities of usage patterns among mutually affine users,while content-based filtering exploits content similarity. DEGARI by definition fallsinto the latter category since in its current form it uses content description (obtained indifferent ways) as the input. From the technical point of view, however, it differs fromthe current mainstream approaches that are mostly based on the comparison and match-ing of visual and perceptual features of the content [51, 11]. In practice, our approachadds a logic layer capable of mapping and representing - in a commonsense and cog-nitively compliant fashion - new emotional categories which can be used to affect userpreferences and content consumption in a way that cannot be derived from the purestatistical analysis of content and/or the comparison of similar users. Moreover, theproposed approach has been applied to a well-known model, the Plutchik’s circumplexmodel of emotions [39], but could in principle applied to other models which organizeemotions by similarity, opposition and composition, such as for example the extendedversion of the Hourglass model used in SenticNet [52]. Being independent from thespecific application model and type of expression, this approach can work effectively indifferent domains, as shown by its use on the datasets of artworks and media illustratedin this paper. In this sense, it can promote the interoperability of affective annotationsand the cross-domain reuse of techniques and methods.In the future work, we plan to extend the evaluation currently conducted in the formof a user study to a large scale one to further validate the effectiveness of the proposedapproach. We also plan to extend the applications of this system to different domains.A first extension will be in the field of the emotional-oriented recommendation of art-works within Museums and cultural heritage sites (this is a work currently under devel-33pment within the H2020 European SPICE project https://spice-h2020.eu/ .In addition, also the field of music recommendation represents a current area of in-vestigation. From a technical perspective, in future research, we aim at studying theapplication of optimization techniques in [2] in order to improve the efficiency of theDEGARI knowledge generation system. Secondly we aim at considering more accu-rate and multimodal descriptions of artistic and media items, by exploiting AutomaticSpeech Recognition data and semantic visual categories extracted from video and au-dio channels of the content. Finally, as mentioned, we plan to improve the providedrecommendations by justifying the provided content reclassification (and recommen-dation) also referring to the probabilistic ranks assigned to the shared features betweenthe generated emotion and the items to reclassify. References [1] Acotto, E., Baldoni, M., Baroglio, C., Patti, V., Portis, F., and Vaccarino, G. (2009).Arsmeteo: artworks and tags floating over the planet art. In Proceedings of the 20thACM conference on Hypertext and hypermedia , pages 331–332.[2] Alberti, M., Bellodi, E., Cota, G., Riguzzi, F., and Zese, R. (2017). cplint onSWISH: probabilistic logical inference with a web browser. Intelligenza Artificiale ,11(1):47–64.[3] Andjelkovic, I., Parra, D., and O’Donovan, J. (2019). Moodplay: Interactive musicrecommendation based on artists’ mood similarity. International Journal of Human-Computer Studies , 121:142–159.[4] B¨anziger, T. and Scherer, K. R. (2010). Introducing the geneva multimodal emo-tion portrayal (gemep) corpus. Blueprint for affective computing: A sourcebook ,2010:271–94.[5] Bertola, F. and Patti, V. (2016). Ontology-based affective models to organizeartworks in the social semantic web. Information Processing & Management ,52(1):139–162. 346] Boden, M. A. (1998). Creativity and artificial intelligence. Artificial Intelligence ,103(1-2):347–356.[7] Cambria, E., Li, Y., Xing, F. Z., Poria, S., and Kwok, K. (2020). Senticnet 6:Ensemble application of symbolic and subsymbolic ai for sentiment analysis. In Proceedings of the 29th ACM International Conference on Information & Knowl-edge Management , pages 105–114.[8] Cambria, E., Livingstone, A., and Hussain, A. (2012). The hourglass of emotions.In Cognitive behavioural systems , pages 144–157. Springer.[9] Chiodino, E., Luccio, D. D., Lieto, A., Messina, A., Pozzato, G. L., and Rubinetti,D. (2020). A knowledge-based system for the dynamic generation and classifica-tion of novel contents in multimedia broadcasting. In Giacomo, G. D., Catal´a, A.,Dilkina, B., Milano, M., Barro, S., Bugar´ın, A., and Lang, J., editors, ECAI 2020 -24th European Conference on Artificial Intelligence, 29 August-8 September 2020,Santiago de Compostela, Spain, August 29 - September 8, 2020 - Including 10thConference on Prestigious Applications of Artificial Intelligence (PAIS 2020) , vol-ume 325 of Frontiers in Artificial Intelligence and Applications , pages 680–687.IOS Press.[10] Confalonieri, R., Schorlemmer, M., Kutz, O., Pe˜naloza, R., Plaza, E., and Eppe,M. (2016). Conceptual blending in EL++. In Lenzerini, M. and Pe˜naloza, R., edi-tors, Proceedings of the 29th International Workshop on Description Logics, CapeTown, South Africa, April 22-25, 2016. , volume 1577 of CEUR Workshop Proceed-ings . CEUR-WS.org.[11] Deldjoo, Y., Constantin, M. G., Eghbal-Zadeh, H., Ionescu, B., Schedl, M., andCremonesi, P. (2018). Audio-visual encoding of multimedia content for enhancingmovie recommendations. In Proceedings of the 12th ACM Conference on Recom-mender Systems , pages 455–459. ACM.[12] Ekman, P. (1999). Basic emotions. Handbook of cognition and emotion , 98(45-60):16. 3513] Eppe, M., Maclean, E., Confalonieri, R., Kutz, O., Schorlemmer, M., Plaza, E.,and K¨uhnberger, K.-U. (2018). A computational framework for conceptual blend-ing. Artificial Intelligence , 256:105–129.[14] Frixione, M. and Lieto, A. (2011). Representing and reasoning on typicality informal ontologies. In Proceedings of the 7th International Conference on SemanticSystems , pages 119–125. ACM.[15] Gachechiladze, D., Lanubile, F., Novielli, N., and Serebrenik, A. (2017). Angerand its direction in collaborative software development. In Proceedings of the 39thInternational Conference on Software Engineering: New Ideas and Emerging Re-sults Track , ICSE-NIER ’17, pages 11–14, Piscataway, NJ, USA. IEEE Press.[16] Giordano, L., Gliozzi, V., Olivetti, N., and Pozzato, G. L. (2015). Semantic char-acterization of Rational Closure: from Propositional Logic to Description Logics. Artificial Intelligence , 226:1–33.[17] Hampton, J. A. (1987). Inheritance of attributes in natural concept conjunctions. Memory & Cognition , 15(1):55–71.[18] Jurafsky, D. and Martin, J. H. (2019). Lexicons for sentiment, affect, and con-notation. In Speech and language processing: an introduction to natural languageprocessing, computational linguistics, and speech recognition, 3nd Edition , PrenticeHall series in artificial intelligence, chapter 21. Prentice Hall, Pearson Education In-ternational. Draft, available online: https://web.stanford.edu/ jurafsky/slp3/21.pdf.Accessed on December 2020.[19] Leder, H., Gerger, G., Brieber, D., and Schwarz, N. (2014). What makes anart expert? emotion and evaluation in art appreciation. Cognition and Emotion ,28(6):1137–1147.[20] Lehmann, D. and Magidor, M. (1992). What does a conditional knowledge baseentail? Artificial Intelligence , 55(1):1–60.[21] Lewis, M. and Lawry, J. (2016). Hierarchical conceptual spaces for concept com-bination. Artificial Intelligence , 237:204–227.3622] Lieto, A., Perrone, F., Pozzato, G. L., and Chiodino, E. (2019a). Beyond sub-goaling: A dynamic knowledge generation framework for creative problem solvingin cognitive architectures. Cognitive Systems Research , 58:305–316.[23] Lieto, A., Pozzato, G., and Valese, A. (2018). COCOS: a typicality based COn-cept COmbination System . In Montali, M. and Felli, P., editors, Proceedings of the33rd Italian Conference on Computational Logic (CILC 2018) , CEUR WorkshopProceedings, pages 55–59, Bozen, Italy.[24] Lieto, A. and Pozzato, G. L. (2019). Applying a description logic of typicality asa generative tool for concept combination in computational creativity. IntelligenzaArtificiale , 13(1):93–106.[25] Lieto, A. and Pozzato, G. L. (2020). A description logic framework for com-monsense conceptual combination integrating typicality, probabilities and cognitiveheuristics. Journal of Experimental & Theoretical Artificial Intelligence , 32(5):769–804.[26] Lieto, A., Pozzato, G. L., Perrone, F., and Chiodino, E. (2019b). Knowledge cap-turing via conceptual reframing: A goal-oriented framework for knowledge inven-tion. In Proceedings of the 10th ACM Conference on Knowledge Capture, K-CAP2019, Marina del Rey , pages 109–114. ACM.[27] Lopes, P., Liapis, A., and Yannakakis, G. N. (2017). Modelling affect for horrorsoundscapes. IEEE Transactions on Affective Computing , 10(2):209–222.[28] McCrae, J., Spohr, D., and Cimiano, P. (2011). Linking lexical resources andontologies on the semantic web with lemon. In Extended Semantic Web Conference ,pages 245–259. Springer.[29] Mehrabian, A. (1996). Pleasure-arousal-dominance: A general framework for de-scribing and measuring individual differences in temperament. Current Psychology ,14(4):261–292.[30] Mohammad, S. (2018a). Obtaining reliable human ratings of valence, arousal,and dominance for 20,000 English words. In Proceedings of the 56th Annual Meet- ng of the Association for Computational Linguistics (Volume 1: Long Papers) ,pages 174–184, Melbourne, Australia. Association for Computational Linguistics.[31] Mohammad, S. and Kiritchenko, S. (2018). WikiArt emotions: An annotateddataset of emotions evoked by art. In Proceedings of the Eleventh InternationalConference on Language Resources and Evaluation (LREC 2018) , Miyazaki, Japan.European Language Resources Association (ELRA).[32] Mohammad, S. M. (2018b). Word affect intensities. Proceedings of the 11thEdition of the Language Resources and Evaluation Conference (LREC) .[33] Mohammad, S. M. and Turney, P. D. (2013). Crowdsourcing a word-emotionassociation lexicon. Computational Intelligence , 29(3):436–465.[34] Nissim, M. and Patti, V. (2017). Chapter 3 - semantic aspects in sentiment analy-sis. In Pozzi, F. A., Fersini, E., Messina, E., and Liu, B., editors, Sentiment Analysisin Social Networks , pages 31 – 48. Morgan Kaufmann, Boston.[35] Noy, P. and Noy-Sharav, D. (2013). Art and emotions. International journal ofapplied psychoanalytic studies , 10(2):100–107.[36] Osherson, D. N. and Smith, E. E. (1981). On the adequacy of prototype theory asa theory of concepts. Cognition , 9(1):35–58.[37] Parisier, E. (2012). The Filter Bubble: What the Internet Is Hiding from You .Penguin Books Limited.[38] Patti, V., Bertola, F., and Lieto, A. (2015). Arsemotica for arsmeteo. org:Emotion-driven exploration of online art collections. In The Twenty-Eighth Interna-tional Florida Artificial Intelligence Research Society Conference (FLAIRS 2015) ,pages 288–293. Association for the Advancement of Artificial Intelligence.[39] Plutchik, R. (1980). A general psychoevolutionary theory of emotion. In Theoriesof emotion , pages 3–33. Elsevier.[40] Plutchik, R. (2001). The nature of emotions. American scientist , 89(4):344–350.3841] Riguzzi, F., Bellodi, E., Lamma, E., and Zese, R. (2015a). Probabilistic descrip-tion logics under the distribution semantics. Semantic Web , 6(5):477–501.[42] Riguzzi, F., Bellodi, E., Lamma, E., and Zese, R. (2015b). Reasoning with prob-abilistic ontologies. In Yang, Q. and Wooldridge, M., editors, Proceedings of theTwenty-Fourth International Joint Conference on Artificial Intelligence, IJCAI 2015,Buenos Aires, Argentina, July 25-31, 2015 , pages 4310–4316. AAAI Press.[43] Ross, W. (1959). Ars Rhetorica. Aristotle . Clarendon Press, Oxford.[44] Russell, D. A. et al. (1964). On the sublime . Clarendon.[45] Russell, J. A. (1980). A circumplex model of affect. Journal of personality andsocial psychology , 39(6):1161.[46] Shani, G. and Gunawardana, A. (2011). Evaluating recommendation systems. In Recommender systems handbook , pages 257–297. Springer.[47] Smith, G. M. (2003). Film structure and the emotion system . Cambridge Univer-sity Press.[48] Sohail, S. S., Siddiqui, J., and Ali, R. (2017). Classifications of recommendersystems: A review. Engineering Science and Technology Review , 10(4):132–153.[49] Strapparava, C. and Mihalcea, R. (2007). Semeval-2007 task 14: Affective text.In Proceedings of the 4th International Workshop on Semantic Evaluations , Se-mEval ’07, pages 70–74, Stroudsburg, PA, USA. Association for ComputationalLinguistics.[50] Strapparava, C. and Valitutti, A. (2004). WordNet affect: an affective extensionof WordNet. In Proceedings of the Fourth International Conference on LanguageResources and Evaluation (LREC’04) , Lisbon, Portugal. European Language Re-sources Association (ELRA).[51] Sun, M. S., Li, F., and Zhang, J. (2018). A multi-modality deep network forcold-start recommendation. Big Data and Cognitive Computing , 2(7).3952] Susanto, Y., Livingstone, A. G., Ng, B. C., and Cambria, E. (2020). The hourglassmodel revisited. IEEE Intelligent Systems , 35(5):96–102.[53] Tan, E. S. (2013). Emotion and the structure of narrative film: Film as an emotionmachine . Routledge.[54] Thoma, M. V., Ryf, S., Mohiyeddini, C., Ehlert, U., and Nater, U. M. (2012).Emotion regulation through listening to music in everyday situations. Cognition &emotion , 26(3):550–560.[55] Van Dongen, N. N., Van Strien, J. W., and Dijkstra, K. (2016). Implicit emotionregulation in the context of viewing artworks: Erp evidence in response to pleasantand unpleasant pictures. Brain and Cognition , 107:48–54.[56] Wang, Z., Ho, S., and Cambria, E. (2020). A review of emotion sensing: catego-rization models and algorithms. Multimedia Tools and Applications , 79.[57] Watson, D., Clark, L. A., and Tellegen, A. (1988). Development and validationof brief measures of positive and negative affect: the panas scales.