Difference between revisions of "Vol-3197/short5"
Jump to navigation
Jump to search
(edited by wikiedit) |
(edited by wikiedit) |
||
Line 9: | Line 9: | ||
|dblpUrl=https://dblp.org/rec/conf/nmr/CasiniS22 | |dblpUrl=https://dblp.org/rec/conf/nmr/CasiniS22 | ||
|wikidataid=Q117341737 | |wikidataid=Q117341737 | ||
+ | |description=scientific paper published in CEUR-WS Volume 3197 | ||
}} | }} | ||
==Defeasible Reasoning in RDFS== | ==Defeasible Reasoning in RDFS== |
Latest revision as of 16:55, 30 March 2023
Paper
Paper | |
---|---|
edit | |
description | scientific paper published in CEUR-WS Volume 3197 |
id | Vol-3197/short5 |
wikidataid | Q117341737→Q117341737 |
title | Defeasible Reasoning in RDFS |
pdfUrl | https://ceur-ws.org/Vol-3197/short5.pdf |
dblpUrl | https://dblp.org/rec/conf/nmr/CasiniS22 |
volume | Vol-3197→Vol-3197 |
session | → |
Defeasible Reasoning in RDFS
Defeasible reasoning in RDFS (Extended Abstract) Giovanni Casini1,2 , Umberto Straccia1 1 ISTI - CNR, Pisa, Italy 2 CAIR, University of Cape Town, South Africa Abstract For non-monotonic logics, the notion of Rational Closure (RC) is acknowledged as one of the main approaches. In this work we present an integration of RC within the triple language RDFS (Resource Description Framework Schema), which together with OWL 2 is a major standard semantic web ontology language. To do so, we start from 𝜌df, an RDFS fragment that covers the essential features of RDFS, and extend it to 𝜌𝑑𝑓⊥ , allowing to state that two entities are incompatible/disjoint with each other. Eventually, we propose defeasible 𝜌𝑑𝑓⊥ via a typical RC construction allowing to state default class/property inclusions. Keywords RDFS, non-monotonic reasoning, defeasible reasoning, rational closure 1. Introduction 20 years [1, 2, 3, 4, 5, 6]. On the other hand, addressing non-monotonicity in the context of RDFS, has attracted RDFS (Resource Description Framework Schema)1 is in comparison little attention so far, and almost all ap- a main standard semantic web ontology language that proaches we are aware of implement non-monotonicity consists of triples (𝑠, 𝑝, 𝑜) (denoting 𝑠 is related via 𝑝 by adding a so-called rule-layer on top of RDFS; see with 𝑜). The introduction of non-monotonic formalisms e.g., [7, 8, 9, 2, 10]. in reasoning with ontologies is useful in particular to deal In the following, our aim is to show how to integrate with situations in which some classes are exceptional Rational Closure (RC), one of the main constructions in and do not satisfy some typical properties of their super non-monotonic reasoning [11], directly within the triple classes, as illustrated with the following example. language RDFS. To to do so, we start from 𝜌df [12, 13], a minimal, but significant RDFS fragment that covers the Example 1.1 (Running example). Consider the following essential features of RDFS, and then extend it to 𝜌𝑑𝑓⊥ , facts (and an intuitive translation into RDFS, where sc is allowing to state that two entities are incompatible/disjoint read as “is a subclass of”). with each other. The results in this paper are presented more in detail in a technical report [14]. - Young people are usually happy; (𝑦𝑃, sc, ℎ𝑃 ) - Drug users are usually unhappy; (𝑑𝑈, sc, 𝑢ℎ𝑃 ) - Drug users are usually young; (𝑑𝑈, sc, 𝑦𝑃 ) 2. 𝜌𝑑𝑓⊥ Graphs - Controlled drug users are usually happy; (𝑐𝐷𝑈, sc, ℎ𝑃 ) - Controlled drug users are drug users; (𝑐𝐷𝑈, sc, 𝑑𝑈 ) We rely on a fragment of RDFS, called minimal 𝜌df [12, Def. 15], that covers all main features of RDFS, and it is We may consider then reasonable to conclude, for example, essentially the formal logic behind RDFS. The vocabulary that controlled young drug users are usually happy. is composed by two pairwise disjoint alphabets U and L denoting, respectively, URI references and literals, where Description Logics provide the logical foundation of a literal may be a plain literal (e.g., a string) or a typed formal ontologies of the semantic Web Ontology Lan- 2 literal (e.g., a boolean value) [15]. With UL, the set of guage (OWL) family and endowing them with non- terms, we will denote the union of these sets. A 𝜌df-triple monotonic features has been a main issue in the past is of the form 𝜏 = (𝑠, 𝑝, 𝑜) ∈ UL × U × UL.3 We NMR 2022: 20th International Workshop on Non-Monotonic Reason- call 𝑠 the subject, 𝑝 the predicate, and 𝑜 the object. A ing, August 07–09, 2022, Haifa, Israel graph 𝐺 is a set of triples. 𝜌df is characterised by the set $ giovanni.casini@isti.cnr.it (G. Casini); of predicates {sp, sc, type, dom, range} ⊆ U, that can umberto.straccia@isti.cnr.it (U. Straccia) � 0000-0002-4267-4447 (G. Casini); 0000-0001-5998-6757 appear only as second elements in the triples. Informally, (U. Straccia) (𝑖) (𝑝, sp, 𝑞) means that property 𝑝 is a subproperty of © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). property 𝑞; (𝑖𝑖) (𝑐, sc, 𝑑) means that class 𝑐 is a subclass CEUR CEUR Workshop Proceedings (CEUR-WS.org) Workshop Proceedings http://ceur-ws.org ISSN 1613-0073 of class 𝑑; (𝑖𝑖𝑖) (𝑎, type, 𝑏) means that 𝑎 is of type 𝑏; 1 http://www.w3.org/TR/rdf-schema/ 2 3 http://www.w3.org/TR/2009/REC-owl2-profiles-20091027 As in [12], we allow literals for 𝑠. 155 �(𝑖𝑣) (𝑝, dom, 𝑐) means that the domain of property 𝑝 is • if (𝑐, 𝑑) ∈ P[[⊥c ℐ ]] then 𝑐, 𝑑 ∈ ΔC ; 𝑐; and (𝑣) (𝑝, range, 𝑐) means that the range of property 𝑝 is 𝑐. We also recall that minimal 𝜌df does not consider • If (𝑐, 𝑑) ∈ P[[⊥c ℐ ]], then (𝑑, 𝑐) ∈ P[[⊥c ℐ ]] (sc- so-called blank nodes [16, 12]. Symmetry); Concerning the semantics of 𝜌df [12], an interpreta- • If (𝑐, 𝑑) ∈ P[[⊥c ℐ ]] and (𝑒, 𝑐) ∈ P[[scℐ ]], then tion is a tuple ℐ = ⟨ΔR , ΔP , ΔC , ΔL , P[[·]], C[[·]], ·ℐ ⟩, (𝑒, 𝑑) ∈ P[[⊥c ℐ ]] (sc-Transitivity); where ΔR , ΔP , ΔC , ΔL are the interpretation domains of ℐ, which are finite non-empty sets, and P[[·]], C[[·]], ·ℐ • If (𝑐, 𝑐) ∈ P[[⊥c ℐ ]] and 𝑑 ∈ ΔC then (𝑐, 𝑑) ∈ are the interpretation functions of ℐ. In particular: (𝑖) P[[⊥c ℐ ]] (c-Exhaustive). ΔR are the resources (the domain or universe of ℐ); (𝑖𝑖) ΔP are property names (not necessarily disjoint from These new constraints are such to model relevant prop- ΔR ); (𝑖𝑖𝑖) ΔC ⊆ ΔR are the classes; (𝑖𝑣) ΔL ⊆ ΔR erties of disjointedness, and allow the definition of an are the literal values and contains L ∩ 𝑉 ; (𝑣) P[[·]] is a entailment relation ⊨𝜌𝑑𝑓⊥ . An important feature of 𝜌𝑑𝑓⊥ function P[[·]] : ΔP → 2ΔR ×ΔR ; (𝑣𝑖) C[[·]] is a function is also that it preserves the 𝜌df property that a graph is C[[·]] : ΔC → 2ΔR ; (𝑣𝑖𝑖) ·ℐ maps each 𝑡 ∈ UL ∩ 𝑉 into always satisfiable, avoiding the possibility of unsatisfia- a value 𝑡ℐ ∈ ΔR ∪ ΔP , and such that ·ℐ is the identity bility and the ex falso quodlibet principle. This is in line for plain literals and assigns an element in ΔR to each with the 𝜌df semantics [12, 19]. From an inference system element in L. point of view, new derivation rules are added to the 𝜌df An interpretation ℐ satisfies a graph 𝐺 if for each derivation system [14, Sect. 2.3]. The following are just a (𝑠, 𝑝, 𝑜) ∈ 𝐺, 𝑝ℐ ∈ ΔP and (𝑠ℐ , 𝑜ℐ ) ∈ P[[𝑝ℐ ]], and few examples: moreover ℐ satisfies a series of constraints related to the (𝐴,⊥c ,𝐵) ; (𝐴,⊥c ,𝐵),(𝐶,sc,𝐴) ; (𝐴,⊥c ,𝐴) . (𝐵,⊥c ,𝐴) (𝐶,⊥c ,𝐵) (𝐴,⊥c ,𝐵) 𝜌df-predicates. For example, a constraint imposing that The new derivation relation ⊢𝜌𝑑𝑓⊥ that we have defined is P[[scℐ ]] is transitive over ΔP indicates that the subclass correct and complete w.r.t. the entailment relation ⊨𝜌𝑑𝑓⊥ relation sc must be transitive. We refer to [12, Def. 15] [14, Th. 2.1]. Eventually, we say that a graph 𝐺 has a for the full definition of the satisfaction relation, and of conflict if, for some term 𝑡, either 𝐺𝑠 ⊢𝜌𝑑𝑓⊥ (𝑡, ⊥c , 𝑡) or the correspondent entailment relation. 𝐺𝑠 ⊢𝜌𝑑𝑓⊥ (𝑡, ⊥p , 𝑡) holds. The intuitive meaning is that Definition 2.1 (Entailment ⊨𝜌𝑑𝑓⊥ ). Given two graphs 𝐺 𝐺 has a conflict if we can derive for some term 𝑡 that it and 𝐻, we say that 𝐺 entails 𝐻, denoted 𝐺 ⊨𝜌𝑑𝑓 𝐻, if is either an empty class, (𝑡, ⊥c , 𝑡), or an empty predicate, and only if every model of 𝐺 is also a model of 𝐻. (𝑡, ⊥p , 𝑡). In [12] the reader can find also a deduction system, con- Example 2.1 (Running example cont.). In Exam- sistent and complete w.r.t. the 𝜌df entailment relation, that ple 1.1 we could add the triple (𝑢ℎ𝑃, ⊥c , ℎ𝑃 ) to is based on rules, such as indicate that ‘being happy’ and ‘being unhappy’ are incompatible. Notice that from (𝑢ℎ𝑃, ⊥c , ℎ𝑃 ), (𝐴, sc, 𝐵), (𝐵, sc, 𝐶) (𝑐𝐷𝑈, sc, ℎ𝑃 ), (𝑐𝐷𝑈, sc, 𝑑𝑈 ) and (𝑑𝑈, sc, 𝑢ℎ𝑃 ) we (𝐴, sc, 𝐶) conclude (𝑐𝐷𝑈, ⊥c , 𝑐𝐷𝑈 ), that is, that being a con- encoding the transitivity of sc. trolled drug user is incompatible with being a controlled Defeasible reasoning can be built only when faced with drug user (that is, 𝑐𝐷𝑈 should be an empty class). Analo- a conflict between the properties of a class and of a sub- gously, from (𝑢ℎ𝑃, ⊥c , ℎ𝑃 ), (𝑑𝑈, sc, 𝑦𝑃 ), (𝑦𝑃, sc, ℎ𝑃 ) class. e.g., in Example 1.1,“Drug users are usually un- and (𝑑𝑈, sc, 𝑢ℎ𝑃 ) we conclude (𝑑𝑈, ⊥c , 𝑑𝑈 ). happy” appears in conflict with “Controlled drug users are usually happy”. 𝜌df is not expressive enough to model such conflicts. So, we need to introduce at least a no- 3. Defeasible 𝜌𝑑𝑓⊥ tion of incompatibility, of disjunctiveness [17]. Hence we Next we show how to model defeasible information. Here enrich the 𝜌df vocabulary with two new predicates, ⊥c we consider defeasibility w.r.t. the predicates sc and sp and ⊥p , representing incompatible information: (𝑐, ⊥c , 𝑑) only, and introduce the notion of defeasible triple: (resp., (𝑝, ⊥p , 𝑞)) indicates that the classes 𝑐 and 𝑑 (resp., the properties 𝑝 and 𝑞) are disjoint. Of course we can 𝛿 = ⟨𝑠, 𝑝, 𝑜⟩ ∈ UL × {sc, sp} × UL , further enrich the language allowing for logically stronger notions such as negation [18], but it is not necessary for where 𝑠, 𝑜 ̸∈ 𝜌𝑑𝑓⊥ . The intended meaning of the purpose of the present paper. e.g., ⟨𝑐, sc, 𝑑⟩ is “Typically, an instance of 𝑐 is also an We call the new formalism, obtained by adding ⊥c and instance of 𝑏”. Analogously, ⟨𝑝, sp, 𝑞⟩ is read as “Typi- ⊥p to 𝜌df, 𝜌𝑑𝑓⊥ . Some new constraints are added to the cally, a pair related by 𝑝 is also related by 𝑞”. semantics of 𝜌df [14, Sect. 2.2]. Here are a few examples: 156 �Example 3.1 (Running example cont.). In Exam- Example 3.3 (Running example cont.). We wonder ple 1.1 the statements containing ‘usually’ can whether ⟨𝑐𝐷𝑈, sc, 𝑢ℎ𝑃 ⟩ is in the RC of our graph. This more correctly be modelled using defeasible triples, triple is interesting because it would be derivable in that is, ⟨𝑦𝑃, sc, ℎ𝑃 ⟩, ⟨𝑑𝑈, sc, 𝑢ℎ𝑃 ⟩, ⟨𝑑𝑈, sc, 𝑦𝑃 ⟩ and the monotonic 𝜌𝑑𝑓⊥ -graph we have considered up to ⟨𝑐𝐷𝑈, sc, ℎ𝑃 ⟩. Exmple 2.1, but it is undesirable since we are aware that ⟨𝑐𝐷𝑈, sc, ℎ𝑃 ⟩ and that ‘Drug users are usually There are various ways of reasoning in a defeasible frame- happy’, that is a defeasible statement. If we consider work. Here we take under consideration RC [11], since, our entire graph, we already know (Example 3.2) that despite having some limits from the inferential point of 𝑐𝐷𝑈 is exceptional, that is, substituting the defeasi- view [20], it is a main inference relation in conditional ble triples with their 𝜌𝑑𝑓⊥ counterparts, we obtain reasoning on top of which we can define other interesting (𝑐𝐷𝑈, ⊥c , 𝑐𝐷𝑈 ). The same if we consider the graph forms of entailment [20, 21, 22]. obtained eliminating all the defeasible triples of rank We give here only a short overview of the reasoning 0. Only once we eliminate also the triples of rank procedure, inviting the reader to check [14] for a compre- 1, and we consider only the graph {⟨𝑐𝐷𝑈, sc, ℎ𝑃 ⟩} ∪ hensive presentation. Given a defeasible graph 𝐺 and a {(𝑐𝐷𝑈, sc, 𝑑𝑈 ), (ℎ𝑃, ⊥c , 𝑢ℎ𝑃 )}, we are not able to query ⟨𝑠, 𝑝, 𝑜⟩, we decide whether ⟨𝑠, 𝑝, 𝑜⟩ is in the RC derive (𝑐𝐷𝑈, ⊥c , 𝑐𝐷𝑈 ) anymore. That is, we do of 𝐺 through a two-step procedure: not have a conflict anymore on 𝑐𝐷𝑈 . Our query 1. We rank all the defeasible triples in 𝐺, considering ⟨𝑐𝐷𝑈, sc, 𝑢ℎ𝑃 ⟩ will be decided considering only this the potential conflicts and the relative logical specificity portion of the original graph: {⟨𝑐𝐷𝑈, sc, ℎ𝑃 ⟩} ∪ of the first elements of the triples. We give priority (that is, {(𝑐𝐷𝑈, sc, 𝑑𝑈 ), (ℎ𝑃, ⊥c , 𝑢ℎ𝑃 )}. In order to decide a higher rank) to more specific triples. To check the pres- whether ⟨𝑐𝐷𝑈, sc, 𝑢ℎ𝑃 ⟩, we check whether its 𝜌𝑑𝑓⊥ - ence of potential conflicts in a graph, we translate all the counterpart, (𝑐𝐷𝑈, sc, 𝑢ℎ𝑃 ), is derivable from the 𝜌𝑑𝑓⊥ - defeasible triples into the correspondent 𝜌𝑑𝑓⊥ triples, that counterpart of the portion of the graph we consider, that is, we create a new 𝜌𝑑𝑓⊥ graph in which every defeasible is, {(𝑐𝐷𝑈, sc, ℎ𝑃 ), (𝑐𝐷𝑈, sc, 𝑑𝑈 ), (ℎ𝑃, ⊥c , 𝑢ℎ𝑃 )}. It ⟨𝑠, 𝑝, 𝑜⟩ is substituted by (𝑠, 𝑝, 𝑜). is easy to check that there is no way of deriving (𝑐𝐷𝑈, sc, 𝑢ℎ𝑃 ) from this graph. Example 3.2 (Running example cont.). In Example 2.1 The semantics for defeasible 𝜌𝑑𝑓⊥ are defined with a rank- we have seen that from the 𝜌𝑑𝑓⊥ version of our graph ing of 𝜌𝑑𝑓⊥ -models: the lowest the rank of the model, the we obtain (𝑐𝐷𝑈, ⊥c , 𝑐𝐷𝑈 ) and (𝑑𝑈, ⊥c , 𝑑𝑈 ). From more expected the situation it describes is considered. As this we conclude that all the defeasible triples with for the propositional and DL case [23], given a defeasi- 𝑐𝐷𝑈 or 𝑑𝑈 as first element (e.g., ⟨𝑐𝐷𝑈, sc, ℎ𝑃 ⟩ and ble graph 𝐺 its RC is determined by its minimal ranked ⟨𝑑𝑈, sc, 𝑢ℎ𝑃 ⟩) have priority (a higher rank) w.r.t. the model, that is, the model of 𝐺 in which every 𝜌𝑑𝑓⊥ -model other defeasible triples. That is, ⟨𝑦𝑃, sc, ℎ𝑃 ⟩ has is ranked as low as possible. The technical details can be rank 0, while the other defeasible triples are excep- found in [14, Sect. 3]. tional. We then reiterate the procedure consider- ing only the exceptional triples and the 𝜌𝑑𝑓⊥ -triples, 4. Conclusions that is, {⟨𝑑𝑈, sc, 𝑢ℎ𝑃 ⟩, ⟨𝑑𝑈, sc, 𝑦𝑃 ⟩, ⟨𝑐𝐷𝑈, sc, ℎ𝑃 ⟩} ∪ The main features of our approach are: (i) the defeasible {(𝑐𝐷𝑈, sc, 𝑑𝑈 ), (ℎ𝑃, ⊥c , 𝑢ℎ𝑃 )}. Translating the de- 𝜌𝑑𝑓⊥ we propose remains syntactically a triple language feasible triples into 𝜌𝑑𝑓⊥ -triples, the only conflict we by extending it with new predicate symbols with specific can still derive is (𝑐𝐷𝑈, ⊥c , 𝑐𝐷𝑈 ), hence we have semantics; (ii) the logic is defined in such a way that any that ⟨𝑑𝑈, sc, 𝑢ℎ𝑃 ⟩, ⟨𝑑𝑈, sc, 𝑦𝑃 ⟩ have rank 1, while RDFS reasoner/store may handle the new predicates as ⟨𝑐𝐷𝑈, sc, ℎ𝑃 ⟩ is exceptional. From {⟨𝑐𝐷𝑈, sc, ℎ𝑃 ⟩}∪ ordinary terms if it does not want to take into account of {(𝑐𝐷𝑈, sc, 𝑑𝑈 ), (ℎ𝑃, ⊥c , 𝑢ℎ𝑃 )} we cannot derive any- the extra non-monotonic capabilities; (iii) the defeasible more (𝑐𝐷𝑈, ⊥c , 𝑐𝐷𝑈 ), hence ⟨𝑐𝐷𝑈, sc, ℎ𝑃 ⟩ has rank 2 entailment decision procedure is built on top of the 𝜌𝑑𝑓⊥ and we have finished the ranking of the graph. entailment decision procedure, which in turn is an exten- Note that, given a graph 𝐺, the ranking procedure needs sion of the one for 𝜌df via some additional inference rules, to be done once and for all. favouring a potential implementation; (iv) the computa- tional complexity of deciding entailment in 𝜌df and 𝜌𝑑𝑓⊥ 2. Given a query ⟨𝑠, sc, 𝑜⟩ (resp., ⟨𝑠, sp, 𝑜⟩), we check are the same; and (v) defeasible entailment can be decided the rank of 𝑠, i.e., we check which is the lowest rank in via a polynomial number of calls to an oracle deciding which we do not derive (𝑠, ⊥c , 𝑠) (resp., (𝑠, ⊥p , 𝑠)), and ground triple entailment in 𝜌𝑑𝑓⊥ and, in particular, decid- then we check whether we can derive (𝑠, sc, 𝑜) (resp., ing defeasible entailment can be done in polynomial time. (𝑠, sp, 𝑜)) considering only the defeasible triples with at While an extended version of the paper is under review at least such a rank. the moment, a technical report is online [14]. 157 �Acknowledgments Semantic Web Conference (ESWC-2009), 2009, pp. 857– 862. This research was partially supported by TAILOR (Foun- [11] D. Lehmann, M. Magidor, What does a conditional dations of Trustworthy AI – Integrating Reasoning, Learn- knowledge base entail?, Artif. Intell. 55 (1992) 1–60. ing and Optimization), a project funded by EU Horizon [12] S. Muñoz, J. Pérez, C. Gutierrez, Simple and Efficient 2020 research and innovation programme under GA No Minimal RDFS, Web Semantics: Science, Services and Agents on the World Wide Web 7 (2009) 220–234. URL: 952215. http://dx.doi.org/10.1016/j.websem.2009.07.003. doi:10. 1016/j.websem.2009.07.003. References [13] C. Gutiérrez, C. A. Hurtado, A. O. Mendelzon, J. Pérez, Foundations of semantic web databases, J. Comput. Syst. [1] P. A. Bonatti, M. Faella, I. Petrova, L. Sauro, A new Sci. 77 (2011) 520–541. URL: https://doi.org/10.1016/ semantics for overriding in description logics, Artificial j.jcss.2010.04.009. doi:10.1016/j.jcss.2010.04. Intelligence Journal 222 (2015) 1–48. URL: http://dx. 009. doi.org/10.1016/j.artint.2014.12.010. doi:10.1016/j. [14] G. Casini, U. Straccia, Defeasible RDFS via rational artint.2014.12.010. closure, CoRR abs/2007.07573 (2020). URL: https:// [2] T. Eiter, G. Ianni, T. Lukasiewicz, R. Schindlauer, H. Tom- arxiv.org/abs/2007.07573. arXiv:2007.07573. pits, Combining answer set programming with descrip- [15] RDF, http://www.w3.org/RDF/, W3C, 2004. tion logics for the semantic web, Artificial Intelligence [16] A. Hogan, M. Arenas, A. Mallea, A. Polleres, Every- 172 (2008) 1495–1539. URL: http://dx.doi.org/10.1016/ thing you always wanted to know about blank nodes, j.artint.2008.04.002. doi:10.1016/j.artint.2008. J. Web Semant. 27-28 (2014) 42–69. URL: https:// 04.002. doi.org/10.1016/j.websem.2014.06.004. doi:10.1016/ [3] T. Eiter, G. Ianni, T. Lukasiewicz, R. Schindlauer, Well- j.websem.2014.06.004. founded semantics for description logic programs in the [17] G. Casini, U. Straccia, T. Meyer, A polynomial time semantic web, ACM Transaction on Computational subsumption algorithm for nominal safe ℰℒ𝒪⊥ un- Logic 12 (2011) 11:1–11:41. URL: http://doi.acm.org/ der rational closure, Inf. Sci. 501 (2019) 588–620. 10.1145/1877714.1877717. doi:10.1145/1877714. URL: https://doi.org/10.1016/j.ins.2018.09.037. doi:10. 1877717. 1016/j.ins.2018.09.037. [4] L. Giordano, V. Gliozzi, N. Olivetti, G. L. Pozzato, A [18] U. Straccia, G. Casini, A minimal deductive system for non-monotonic description logic for reasoning about typ- rdfs with negative statements, in: Principles of Knowl- icality, Artificial Intelligence Journal 195 (2013) 165– edge Representation and Reasoning: Proceedings of the 202. URL: http://dx.doi.org/10.1016/j.artint.2012.10.004. 19th International Conference, KR 2022, Haifa, Israel, doi:10.1016/j.artint.2012.10.004. July 31 - August 5, 2022, forth. [5] B. Motik, R. Rosati, Reconciling description logics and [19] S. Muñoz, J. Pérez, C. Gutiérrez, Minimal deductive rules, Journal of the ACM 57 (2010). systems for RDF, in: 4th European Semantic Web Con- [6] K. Britz, G. Casini, T. Meyer, K. Moodley, U. Sattler, ference (ESWC-07), volume 4519 of Lecture Notes in I. Varzinczak, Principles of klm-style defeasible de- Computer Science, Springer Verlag, 2007, pp. 53–67. scription logics, ACM Trans. Comput. Logic 22 (2020). [20] D. Lehmann, Another perspective on default reason- URL: https://doi.org/10.1145/3420258. doi:10.1145/ ing, Annals of Mathematics and Artificial Intelligence 15 3420258. (1995) 61–82. [7] A. Analyti, C. V. Damásio, G. Antoniou, Extended [21] G. Casini, U. Straccia, Defeasible inheritance-based de- RDF: computability and complexity issues, Annals of scription logics, Journal of Artificial Intelligence Re- Mathematics and Artificial Intelligence 75 (2015) 267– search 48 (2013) 415–473. 334. URL: https://doi.org/10.1007/s10472-015-9451-0. [22] G. Casini, T. Meyer, K. Moodley, R. Nortjé, Relevant doi:10.1007/s10472-015-9451-0. closure: A new form of defeasible reasoning for descrip- [8] G. Antoniou, Nonmonotonic rule systems on top of on- tion logics, in: E. Fermé, J. Leite (Eds.), Proceedings tology layers, in: Proceedings of the 1st International of the 14th European Conference on Logics in Artificial Semantic Web Conference (ISWC-02), volume 2663 of Intelligence (JELIA-14), volume 8761 of LNCS, Springer, Lecture Notes in Computer Science, Springer Verlag, Sar- 2014, pp. 92–106. dinia, Italia, 2002, pp. 394–398. [23] L. Giordano, V. Gliozzi, N. Olivetti, G. Pozzato, Semantic [9] T. Eiter, G. Ianni, R. Schindlauer, H. Tompits, Effective characterization of rational closure: From propositional integration of declarative rules with external evaluations logic to description logics, Artificial Intelligence Journal for semantic-web reasoning, in: The Semantic Web: 226 (2015) 1–33. Research and Applications, 3rd European Semantic Web Conference (ESWC-06), volume 4011 of Lecture Notes in Computer Science, Springer Verlag, 2006, pp. 273–287. [10] G. Ianni, T. Krennwallner, A. Martello, A. Polleres, A rule system for querying persistent RDFS data, in: The Semantic Web: Research and Applications, 6th European 158 �