Thomas Lukasiewicz. Computing Laboratory, University of Oxford

Transcription

1 Managing and Vagueness in Description Logics for the Semantic Web Scuola di Dottorato in Ingegneria dell Informazione Napoli 2009 Thomas Lukasiewicz Computing Laboratory, University of Oxford

2 Outline 1, Vagueness, and the Web Sources of and Vagueness on the Web vs. Vagueness: A Clarification 2 The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) 3 in DLs 4 Vagueness Vagueness in DLs

3 Outline, Vagueness, and the Web Sources of and Vagueness on the Web vs. Vagueness: A Clarification 1, Vagueness, and the Web Sources of and Vagueness on the Web vs. Vagueness: A Clarification 2 The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) 3 in DLs 4 Vagueness Vagueness in DLs

4 Sources of and Vagueness on the Web vs. Vagueness: A Clarification Sources of and Vagueness on the Web (Multimedia) information retrieval: To which degree is a Web site/page, a text passage, an image region, a video segment,... relevant to my information need? Matchmaking To which degree does an object match my requirements? if I m looking for a car, and my budget is about e, to which degree does a car s price of e match my budget? Semantic annotation To which degree does, e.g., an image object represent a dog?

5 Sources of and Vagueness on the Web vs. Vagueness: A Clarification Information extraction To which degree am I sure that, e.g., SW is an acronym of Semantic Web? Ontology alignment (schema mapping) To which degree do two concepts of two ontologies represent the same, or are disjoint, or are overlapping? Representation of background knowledge To some degree, birds fly. To some degree, Jim is blond and young.

6 Distributed Information Retrieval Sources of and Vagueness on the Web vs. Vagueness: A Clarification Then, the agent has to perform automatically the following steps: 1 The agent has to select a subset of relevant resources S S, as it is not reasonable to assume to access to and query all resources (resource selection/discovery); 2 For every selected source S i S, the agent has to reformulate its information need Q A into the query language L i of the resource (schema mapping/ontology alignment); 3 The results from the selected resources have to be merged (data fusion/rank aggregation).

7 Negotiation Sources of and Vagueness on the Web vs. Vagueness: A Clarification A car seller sells an Audi TT for e, as from the catalog price. A buyer is looking for a sports car, but wants to pay not more than around e Classical DLs: the problem relies on the crisp conditions on price. More fine-grained approach: to consider prices as vague constraints (fuzzy sets) (as usual in negotiation) Seller would sell above e, but can go down to e. The buyer prefers to spend less than e, but can go up to e. Highest degree of matching is The car may be sold at e.

8 Logic-Based Information Retrieval Sources of and Vagueness on the Web vs. Vagueness: A Clarification IsAbout ImageRegion Object ID degree o1 snoopy 0.8 o2 woodstock Find top-k image regions about animals Query(x) ImageRegion(x) isabout(x, y) Animal(y)

9 Top-k Database Querying Sources of and Vagueness on the Web vs. Vagueness: A Clarification HotelID h1 h2. hasloc hl1 hl2. hasloc hasloc distance hl1 cl1 300 hl1 cl2 500 hl2 cl1 750 hl2 cl ConferenceID c1 c2. hasloc cl1 cl2. hasloc hasloc close cheap hl1 cl hl1 cl hl2 cl hl2 cl Find top-k cheapest hotels close to the train station q(h) haslocation(h, hl) haslocation(train, cl) close(hl, cl) cheap(h)

10 Health-Care: Diagnosis of Pneumonia Sources of and Vagueness on the Web vs. Vagueness: A Clarification Given that Temp = 37.5, Pulse = 98, RespiratoryRate = 18, which is the probability that a patient has pneumonia? Temp = 37.5, Pulse = 98, RespiratoryRate = 18 are in the danger zone. These constraints are rather imprecise than crisp.

11 vs. Vagueness: A Clarification Sources of and Vagueness on the Web vs. Vagueness: A Clarification What does the degree mean? There is often a misunderstanding between interpreting a degree as a measure of uncertainty or as a measure of vagueness. The value 0.83 has a different interpretation in Birds fly to degree 0.83 from that in Hotel Verdi is close to the train station to degree 0.83.

12 Sources of and Vagueness on the Web vs. Vagueness: A Clarification : statements are true or false. But, due to lack of knowledge we can only estimate to which probability/possibility/necessity degree they are true or false. For instance, a bird flies or does not fly. The probability/possibility/necessity degree that it flies is Usually, we have a possible world semantics, with a distribution over possible worlds: W ={I classical interpretation}, I (ϕ) {0, 1} µ: W [0, 1], µ(i ) [0, 1] Pr(φ) = X I =φ µ(i ) Poss(φ) = sup µ(i ) I =φ Necc(φ) = inf µ(i ) = 1 Poss( φ) I =φ

13 Vagueness, Vagueness, and the Web Sources of and Vagueness on the Web vs. Vagueness: A Clarification Vagueness: statements involve concepts for which there is no exact definition, such as tall, small, close, far, cheap, expensive, isabout, similarto. Statements are true to some degree, which is taken from a truth space. E.g., Hotel Verdi is close to the train station to degree 0.83 Truth space: set of truth values L and a partial order. Many-valued Interpretation: a function I mapping formulas into L, i.e., I (ϕ) L. Fuzzy Logic: L = [0, 1]. and Vagueness: It is possible/probable to degree 0.83 that it will be hot tomorrow.

14 Outline, Vagueness, and the Web The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) 1, Vagueness, and the Web Sources of and Vagueness on the Web vs. Vagueness: A Clarification 2 The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) 3 in DLs 4 Vagueness Vagueness in DLs

15 The Semantic Web... The Semantic Web (SW) Web Ontology Languages Description Logics (DLs)...aims at an extension of the current WWW by standards and technologies that help machines to understand the information on the Web to support richer discovery, data integration, navigation, and automation of tasks....consists of several hierarchical layers, including the Ontology layer: OWL Web Ontology Language: OWL Lite SHIF(D), OWL DL SHOIN (D), OWL Full; the Rules, Logic, and Proof layers, which should offer sophisticated representation and reasoning capabilities.

16 Layers of the Semantic Web The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) 22 Engineering the Web Fig. 3.2 The Semantic Web Stack c.2006.

17 OWL, Vagueness, and the Web The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) Three sublanguages of OWL: OWL Full is union of OWL syntax and RDF (Undecidable) OWL DL restricted to FOL fragment (decidable in NEXPTIME) OWL Lite is easier to implement subset of OWL DL (decidable in EXPTIME) Semantic layering: OWL DL within Description Logic (DL) fragment OWL DL based on SHOIN (D n ) DL OWL Lite based on SHIF(D n ) DL

18 Description Logics (DLs) The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) The logics behind OWL DL and OWL Lite, Concept/Class: names are equivalent to unary predicates In general, concepts equiv. to formulas with one free variable Role or attribute: names are equivalent to binary predicates In general, roles equiv. to formulas with two free variables Taxonomy: concept and role hierarchies can be expressed Individual: names are equivalent to constants Operators: restricted so that: Language is decidable and, if possible, of low complexity No need for explicit use of variables Restricted form of and Features such as counting can be succinctly expressed

19 The DL Family The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) A given DL is defined by set of concept and role forming operators Basic language: ALC (Attributive Language with Complement) Syntax Semantics Example C, D (x) (x) A A(x) Human C D C(x) D(x) Human Male C D C(x) D(x) Nice Rich C C(x) Meat R.C y.r(x, y) C(y) has_child.blond R.C y.r(x, y) C(y) has_child.human C D x.c(x) D(x) Happy_Father Man has_child.female a:c C(a) John:Happy_Father

20 Example, Vagueness, and the Web The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) Sex = Male Female Male Female Person Human hassex.sex MalePerson Person hassex.male john:person hassex. Female KB = john:maleperson

21 DL Naming The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) For instance, AL: C, D A C D A R. R.C C: Concept negation: C. Thus, ALC = AL + C S: Used for ALC with transitive roles R + U: Concept disjunction: C 1 C 2 E: Existential quantification: R.C H: Role inclusion axioms: R 1 R 2, e.g., is_component_of is_part_of N : Number restrictions: ( n R) and ( n R), e.g., ( 3 has_child) (has at least 3 children) Q: Qualified number restrictions: ( n R.C) and ( n R.C), e.g., ( 2 has_child.adult) (has at most 2 adult children) O: Nominals (singleton class): {a}, e.g., has_child.{mary}. Note: a:c equiv. to {a} C, and (a, b):r equiv. to {a} R.{b} I: Inverse role: R, e.g., ispartof = haspart F: Functional role: f, e.g., functional(hasage) R +: transitive role: e.g., transitive(ispartof ) SHIF = S + H + I + F = ALCR +HIF OWL-Lite (EXPTIME) SHOIN = S + H + O + I + N = ALCR +HOIN OWL-DL (NEXPTIME)

22 Semantics of Additional Constructs The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) H: Role inclusion axioms: I = R 1 R 2 iff R 1 I R 1 I N : Number restrictions: ( n R) I = {x I : {y x, y R I } n}, ( n R) I = {x I : {y x, y R I } n} Q: Qualified number restrictions: ( n R.C) I = {x I {y x, y R I y C I } n}, ( n R.C) I = {x I : {y x, y R I y C I } n} O: Nominals (singleton class): {a} I = {a I } I: Inverse role: (R ) I = { x, y y, x R I } F: Functional role: I = functional(f ) iff z y z if x, y f I and x, z f I then y = z R +: Transitive role: I = transitive(r) iff r I is transitive

23 Concrete Domains The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) Concrete domains: reals, integers, strings,... (tim, 14):hasAge (sf, SoftComputing ):hasacronym (source1, ComputerScience ):isabout (service2, InformationRetrievalTool ):Matches Minor = Person hasage. 18 Semantics: a clean separation between object classes and concrete domains D = D, Φ D D is an interpretation domain Φ D is the set of concrete domain predicates d with a predefined arity n and fixed interpretation d D n D Concrete properties: R I I D Notation: (D). E.g., ALC(D) is ALC + concrete domains

24 OWL DL as Description Logic The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) Concept/Class constructors: Abstract Syntax DL Syntax Example Descriptions (C) A (URI reference) A Conference owl:thing owl:nothing intersectionof(c 1 C 2...) C 1 C 2 Reference Journal unionof(c 1 C 2...) C 1 C 2 Organization Institution complementof(c) C MasterThesis oneof(o 1...) {o 1,...} {"WISE","ISWC",...} restriction(r somevaluesfrom(c)) R.C parts.incollection restriction(r allvaluesfrom(c)) R.C date.date restriction(r hasvalue(o)) R.{o} date.{2005} restriction(r mincardinality(n)) ( n R) ( 1 location) restriction(r maxcardinality(n)) ( n R) ( 1 publisher) restriction(u somevaluesfrom(d)) U.D issue.integer restriction(u allvaluesfrom(d)) U.D name.string restriction(u hasvalue(v)) U. = v } series.= LNCS restriction(u mincardinality(n)) ( n U) ( 1 title) restriction(u maxcardinality(n)) ( n U) ( 1 author) R is an abstract role, while U is a concrete property of arity 2.

25 The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) Axioms: Axioms Abstract Syntax DL Syntax Example Class(A partial C 1... C n) A C 1... C n Human Animal Biped Class(A complete C 1... C n) A = C 1... C n Man = Human Male EnumeratedClass(A o 1... o n) A = {o 1 }... {o n} RGB = {r} {g} {b} SubClassOf(C 1 C 2 ) C 1 C 2 EquivalentClasses(C 1... C n) C 1 =... = C n DisjointClasses(C 1... C n) C i C j =, i j Male Female ObjectProperty(R super (R 1 )... super (R n) R R i HasDaughter haschild domain(c 1 )... domain(c n) ( 1 R) C i ( 1 haschild) Human range(c 1 )... range(c n) R.C i haschild.human [inverseof(p)] R = P haschild = hasparent [symmetric] R = R similar = similar [functional] ( 1 R) ( 1 hasmother) [Inversefunctional] ( 1 R ) [Transitive]) Tr(R) Tr(ancestor) SubPropertyOf(R 1 R 2 ) R 1 R 2 EquivalentProperties(R 1... R n) R 1 =... = R n cost = price AnnotationProperty(S)

26 The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) Abstract Syntax DL Syntax Example DatatypeProperty(U super (U 1 )... super (U n) U U i domain(c 1 )... domain(c n) ( 1 U) C i ( 1 hasage) Human range(d 1 )... range(d n) U.D i hasage.posinteger [functional]) ( 1 U) ( 1 hasage) SubPropertyOf(U 1 U 2 ) U 1 U 2 hasname hasfirstname EquivalentProperties(U 1... U n) U 1 =... = U n Individuals Individual(o type (C 1 )... type (C n)) o:c i tim:human value(r 1 o 1 )... value(r no n) (o, o i ):R i (tim, mary):haschild value(u 1 v 1 )... value(u nv n) (o, v 1 ):U i (tim, 14):hasAge SameIndividual(o 1... o n) o 1 =... = o n president_bush = G.W.Bush DifferentIndividuals(o 1... o n) o i o j, i j john peter Symbols Object Property R (URI reference) R haschild Datatype Property U (URI reference) U hasage Individual o (URI reference) U tim Data Value v (RDF literal) U ESWC07

27 Outline, Vagueness, and the Web in DLs 1, Vagueness, and the Web Sources of and Vagueness on the Web vs. Vagueness: A Clarification 2 The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) 3 in DLs 4 Vagueness Vagueness in DLs

28 and Logic in DLs Any statement is either true or false Due to lack of knowledge, we can only estimate to which probability/possibility/necessity degree they are true or false Ususally, we have a possible world semantics with a distribution over possible worlds Possible world: any classical interpretation I, mapping any statement φ into {0, 1}: W = {I classical interpretation}, I (φ) {0, 1} Distribution: a mapping µ: W [0, 1], µ(i ) [0, 1] satisfying some additional conditions (probability/possibility) µ(i ) is the probability/possibility that the world I is indeed the actual one

29 in DLs The probability of a statement φ is determined as Pr(φ) = I =φ µ(i ) The possibility of a statement φ is determined as Poss(φ) = sup µ(i ) I =φ The necessity of a statement φ is determined as Nec(φ) = 1 Poss( φ) = inf 1 µ(i ) I =φ

30 in DLs Probabilistic setting: φ = sprinkleron wet W sprinkleron wet µ I I I I = I W µ(i ) Pr(φ) = = 0.9

31 in DLs Possibilistic setting: φ = sprinkleron wet W sprinkleron wet µ I I I I = sup µ(i ) I W Poss(φ) = sup(1, 0.8, 1) = 1 Nec(φ) = 1 Poss( φ) = = 0.7

32 Properties of Probabilities in DLs Pr(φ ψ) = Pr(φ) + Pr(ψ) Pr(φ ψ) Pr(φ ψ) min(pr(φ), Pr(ψ)) Pr(φ ψ) max(0, Pr(φ) + Pr(ψ) 1) Pr(φ ψ) = Pr(φ) + Pr(ψ) Pr(φ ψ) Pr(φ ψ) min(1, Pr(φ) + Pr(ψ)) Pr(φ ψ) max(pr(φ), Pr(ψ)) Pr( φ) = 1 Pr(φ) Pr( ) = 0 Pr( ) = 1

33 in DLs Properties of Possibilities/Necessities Poss(φ ψ) min(poss(φ), Poss(ψ)) Poss(φ ψ) = max(poss(φ), Poss(ψ)) Poss( φ) = 1 Nec(φ) Poss( ) = 0 Poss( ) = 1 (in the normalized case) Nec(φ ψ) = min(nec(φ), Nec(ψ)) Nec(φ ψ) max(nec(φ), Nec(ψ)) Nec( φ) = 1 Poss(φ) Nec( ) = 0 (in the normalized case) Nec( ) = 1

34 Probabilistic Logic in DLs Integration of (propositional) logic- and probability-based representation and reasoning formalisms. Reasoning from logical constraints and interval restrictions for conditional probabilities (also called conditional constraints). Reasoning from convex sets of probability distributions. Model-theoretic notion of logical entailment.

35 in DLs Syntax of Probabilistic Knowledge Bases Finite nonempty set of basic events Φ = {p 1,..., p n }. Event φ: Boolean combination of basic events Logical constraint ψ φ: events ψ and φ: φ implies ψ. Conditional constraint (ψ φ)[l, u]: events ψ and φ, and l, u [0, 1]: conditional probability of ψ given φ is in [l, u]. Probabilistic knowledge base KB = (L, P): finite set of logical constraints L, finite set of conditional constraints P.

36 Example, Vagueness, and the Web in DLs Probabilistic knowledge base KB = (L, P): L = {bird eagle}: All eagles are birds. P = {(have_legs bird)[1, 1], (fly bird)[0.95, 1]}: All birds have legs. Birds fly with a probability of at least 0.95.

37 in DLs Semantics of Probabilistic Knowledge Bases World I : truth assignment to all basic events in Φ. I Φ : all worlds for Φ. Probabilistic interpretation Pr: probability function on I Φ. Pr(φ): sum of all Pr(I ) such that I I Φ and I = φ. Pr(ψ φ): if Pr(φ) > 0, then Pr(ψ φ) = Pr(ψ φ) / Pr(φ). Truth under Pr: Pr = ψ φ iff Pr(ψ φ) = Pr(φ) (iff Pr(ψ φ) = 1). Pr = (ψ φ)[l, u] iff Pr(ψ φ) [l, u] Pr(φ) (iff either Pr(φ) = 0 or Pr(ψ φ) [l, u]).

38 Example, Vagueness, and the Web in DLs Set of basic propositions Φ = {bird, fly}. I Φ contains exactly the worlds I 1, I 2, I 3, and I 4 over Φ: fly fly bird I 1 I 2 bird I 3 I 4 Some probabilistic interpretations: Pr 1 fly fly bird 19/40 1/40 bird 10/40 10/40 Pr 2 fly fly bird 0 1/3 bird 1/3 1/3 Pr 1 (fly bird) = 19/40 and Pr 1 (bird) = 20/40. Pr 2 (fly bird) = 0 and Pr 2 (bird) = 1/3. fly bird is false in Pr 1, but true in Pr 2. (fly bird)[.95, 1] is true in Pr 1, but false in Pr 2.

39 in DLs Satisfiability and Logical Entailment Pr is a model of KB = (L, P) iff Pr = F for all F L P. KB is satisfiable iff a model of KB exists. KB = (ψ φ)[l, u]: (ψ φ)[l, u] is a logical consequence of KB iff every model of KB is also a model of (ψ φ)[l, u]. KB = tight (ψ φ)[l, u]: (ψ φ)[l, u] is a tight logical consequence of KB iff l (resp., u) is the infimum (resp., supremum) of Pr(ψ φ) subject to all models Pr of KB with Pr(φ) > 0. Hence, inference tasks in probabilistic logic: deciding satisfiability, deciding logical consequence, computing tight logically entailed intervals for conditional events.

40 Example, Vagueness, and the Web in DLs Probabilistic knowledge base: KB = ({bird eagle}, {(have_legs bird)[1, 1], (fly bird)[0.95, 1]}). KB is satisfiable, since Pr with Pr(bird eagle have_legs fly) = 1 is a model. Some conclusions under logical entailment: KB = (have_legs bird)[0.3, 1], KB = (fly bird)[0.6, 1]. Tight conclusions under logical entailment: KB = tight (have_legs bird)[1, 1], KB = tight (fly bird)[0.95, 1], KB = tight (have_legs eagle)[1, 1], KB = tight (fly eagle)[0, 1].

41 in DLs Deciding Model Existence / Satisfiability Theorem: The probabilistic knowledge base KB = (L, P) has a model Pr with Pr(α) > 0 iff the following system of linear constraints over the variables y r (r R), where R = {I I Φ I = L}, is solvable: r R, r = ψ φ r R, r = ψ φ l y r + u y r + r R, r =ψ φ r R, r =ψ φ (1 l) y r 0 ( (ψ φ)[l, u] P) (u 1) y r 0 ( (ψ φ)[l, u] P) y r = 1 r R, r =α y r 0 (for all r R)

42 in DLs Computing Tight Logical Consequences Theorem: Suppose KB = (L, P) has a model Pr such that Pr(α) > 0. Then, l (resp., u) such that KB = tight (β α)[l, u] is given by the optimal value of the following linear program over the variables y r (r R), where R = {I I Φ I = L}: minimize (resp., maximize) l y r + r R, r = ψ φ r R, r = ψ φ u y r + r R, r =ψ φ r R, r =ψ φ y r r R, r = β α subject to (1 l) y r 0 ( (ψ φ)[l, u] P) (u 1) y r 0 ( (ψ φ)[l, u] P) y r = 1 r R, r =α y r 0 (for all r R)

43 in DLs Towards Stronger Notions of Entailment Problem: Inferential weakness of logical entailment. Solutions: Probability selection techniques: Perform inference from a representative distribution of the encoded convex set of distributions rather than the whole set, e.g., distribution of maximum entropy, distribution in the center of mass. Probabilistic default reasoning: Perform constraining rather than conditioning and apply techniques from default reasoning to resolve local inconsistencies. Probabilistic independencies: Further constrain the convex set of distributions by probabilistic independencies. ( adds nonlinear equations to linear constraints)

44 in DLs Entailment under Maximum Entropy Entropy of a probabilistic interpretation Pr, denoted H(Pr): H(Pr) = I I Φ Pr(I ) log Pr(I ). The ME model of a satisfiable probabilistic knowledge base KB is the unique probabilistic interpretation Pr that is a model of KB and that has the greatest entropy among all the models of KB. KB = me (ψ φ)[l, u]: (ψ φ)[l, u] is a ME consequence of KB iff the ME model of KB is also a model of (ψ φ)[l, u]. KB = me tight (ψ φ)[l, u]: (ψ φ)[l, u] is a tight ME consequence of KB iff for the ME model Pr of KB, it holds either (a) Pr(φ) = 0, l = 1, and u = 0, or (b) Pr(φ) > 0 and Pr(ψ φ) = l = u.

45 in DLs Logical vs. Maximum Entropy Entailment Probabilistic knowledge base: KB = ({bird eagle}, {(have_legs bird)[1, 1], (fly bird)[0.95, 1]}). Tight conclusions under logical entailment: KB = tight (have_legs bird)[1, 1], KB = tight (fly bird)[0.95, 1], KB = tight (have_legs eagle)[1, 1], KB = tight (fly eagle)[0, 1]. Tight conclusions under maximum entropy entailment: KB me tight (have_legs bird)[1, 1], me KB tight (fly bird)[0.95, 0.95], KB me tight (have_legs eagle)[1, 1], me KB tight (fly eagle)[0.95, 0.95].

46 Lexicographic Entailment in DLs Pr verifies (ψ φ)[l, u] iff Pr(φ) = 1 and Pr = (ψ φ)[l, u]. P tolerates (ψ φ)[l, u] under L iff L P has a model that verifies (ψ φ)[l, u]. KB = (L, P) is consistent iff there exists an ordered partition (P 0,..., P k ) of P such that each P i is the set of all C P \ i 1 j=0 P j tolerated under L by P \ i 1 j=0 P j. This (unique) partition is called the z-partition of KB.

47 in DLs Let KB = (L, P) be consistent, and (P 0,..., P k ) be its z-partition. Pr is lex-preferable to Pr iff some i {0,..., k} exists such that {C P i Pr = C} > {C P i Pr = C} and {C P j Pr = C} = {C P j Pr = C} for all i<j k. A model Pr of F is a lex-minimal model of F iff no model of F is lex-preferable to Pr. KB lex (ψ φ)[l, u]: (ψ φ)[l, u] is a lex-consequence of KB iff every lex-minimal model Pr of L with Pr(φ)=1 satisfies (ψ φ)[l, u]. KB lex tight (ψ φ)[l, u]: (ψ φ)[l, u] is a tight lex-consequence of KB iff l (resp., u) is the infimum (resp., supremum) of Pr(ψ) subject to all lex-minimal models Pr of L with Pr(φ) = 1.

48 in DLs Logical vs. Lexicographic Entailment Probabilistic knowledge base: KB = ({bird eagle}, {(have_legs bird)[1, 1], (fly bird)[0.95, 1]}). Tight conclusions under logical entailment: KB = tight (have_legs bird)[1, 1], KB = tight (fly bird)[0.95, 1], KB = tight (have_legs eagle)[1, 1], KB = tight (fly eagle)[0, 1]. Tight conclusions under probabilistic lexicographic entailment: KB lex tight (have_legs bird)[1, 1], lex KB tight (fly bird)[0.95, 1], KB lex tight (have_legs eagle)[1, 1], lex KB tight (fly eagle)[0.95, 1].

49 in DLs Probabilistic knowledge base: KB = ({bird penguin}, {(have_legs bird)[1, 1], (fly bird)[1, 1], (fly penguin)[0, 0.05]}). Tight conclusions under logical entailment: KB = tight (have_legs bird)[1, 1], KB = tight (fly bird)[1, 1], KB = tight (have_legs penguin)[1, 0], KB = tight (fly penguin)[1, 0]. Tight conclusions under probabilistic lexicographic entailment: KB lex lex tight (have_legs bird)[1, 1], KB tight (fly bird)[1, 1], KB lex tight (have_legs penguin)[1, 1], KB lex tight (fly penguin)[0, 0.05].

50 in DLs Probabilistic knowledge base: KB = ({bird penguin}, {(have_legs bird)[0.99, 1], (fly bird)[0.95, 1], (fly penguin)[0, 0.05]}). Tight conclusions under logical entailment: KB = tight (have_legs bird)[0.99, 1], KB = tight (fly bird)[0.95, 1], KB = tight (have_legs penguin)[0, 1], KB = tight (fly penguin)[0, 0.05]. Tight conclusions under probabilistic lexicographic entailment: KB lex lex tight (have_legs bird)[0.99, 1], KB tight (fly bird)[0.95, 1], KB lex tight (have_legs penguin)[0.99, 1], KB lex tight (fly penguin)[0, 0.05].

51 Bayesian Networks in DLs Well-structured, exact conditional constraints plus conditional independencies specify exactly one joint probability distribution. Joint probability distributions can answer any queries, but can be very large and are often hard to specify. Bayesian network (BN): compact specification of a joint distribution, based on a graphical notation for conditional independencies: a set of nodes; each node represents a random variable a directed, acyclic graph (link directly influences ) a conditional distribution for each node given its parents: P(X i Parents(X i )) Any joint distribution can be represented as a BN.

52 Example, Vagueness, and the Web in DLs I m at work, neighbor John calls to say my alarm is ringing, but neighbor Mary doesn t call. Sometimes it s set off by minor earthquakes. Is there a burglar? Variables: Burglary, Earthquake, Alarm, JohnCalls, MaryCalls Network topology reflects causal knowledge: a burglar can set the alarm off an earthquake can set the alarm off the alarm can cause Mary to call the alarm can cause John to call John sometimes confuses the telephone ringing with the alarm. Mary likes rather loud music and sometimes misses the alarm.

53 in DLs Burglary P(B) P(E) Earthquake Alarm B E t t t f f t f f P(A) JohnCalls A f t P(J) MaryCalls A t f P(M).70.01

54 Global Semantics in DLs Global semantics defines the full joint distribution as the product of the local conditional distributions: P(X 1,..., X n ) =Π n i = 1P(X i Parents(X i )) e.g., B E A J M P(j m a b e) = P(j a)p(m a)p(a b, e)p( b)p( e) = =

55 Inference Tasks in DLs Computing unconditional and conditional probabilities Pr(φ) and Pr(φ ψ) from a Bayesian network: Simple queries: compute posterior marginal P(X i E = e), e.g., P(Burglary Alarm = true, John = true, Mary = false). Conjunctive queries: P(X i, X j E = e) = P(X i E = e)p(x j X i, E = e). Many algorithms: probability propagation (runs in linear time for polytrees), clustering, variable elimination, sampling,...

56 Possibilistic Knowledge Bases in DLs Possibilistic formulas have the form P φ l or N φ l with l [0, 1] Encode to what extent φ is possibly resp. necessarily true Possibilistic interpretation: π : I Φ [0, 1] π(i ) is the degree to which the world I is possible π is normalized: π(i ) = 1 for some I I Φ Possibility/Necessity of an event φ under π: Poss(φ) = sup{π(i ) I I Φ, I = φ} (where max = 0) Necc(φ) = 1 Poss( φ) Truth under π: π = P φ l iff Poss(φ) l π = N φ l iff Nec(φ) l

57 Deciding Entailment in DLs Reduction to classical entailment Let KB l = {φ N φ l KB, l l} KB l = {φ N φ l KB, l > l} Then: KB = N φ l iff KB l = φ KB = P φ l iff KB 0 = φ or there is P ψ l KB with l l and KB 1 l {ψ} = φ

58 First-Order Logics of Probability in DLs Rough classification of first-order logics of probability: types of probability structures underlying probabilistic formalism Types of Probability Structures Probabilities on a set of possible worlds: a probability structure is a tuple (D, S, π, µ), where D is a domain, S is a set of possible worlds, π associates with every s S an interpretation of the function and predicate symbols over D, and µ is a discrete probability function on S. The probability that Tweety (a particular bird) flies is 0.9

59 in DLs Probabilities on the domain: a probability structure is a tuple (D, π, µ), where D is a domain, π is an interpretation of the function and predicate symbols over D, and µ is a discrete probability function on D, which is extended to D n by µ(d 1,..., d n ) = Π n i=1 µ(d i). The probability that a randomly chosen bird flies is at 0.9 Probabilities on both a set of possible worlds and the domain: a probability structure is a tuple (D, S, π, µ D, µ S ), where D is a domain, S is a set of possible worlds, π maps every s S to an interpretation of the function and predicate symbols over D, and µ D (resp., µ S ) is a discrete prob. function on S (resp., D).

60 Underlying Probabilistic Formalism in DLs By far the most first-order probabilistic formalisms have a semantics in probabilities on a set of possible worlds. Such a semantics can be easily based on one of the many existing semantics of propositional probabilistic formalisms: model-theoretic probabilistic logic probabilistic logic under maximum entropy strong nonmonotonic probabilistic logics Bayesian networks A query on such a first-order probabilistic knowledge base KB is then evaluated by (i) constructing a relevant propositional prob. knowledge base KB, and (ii) evaluating the query on KB. Very few first-order formalisms with probabilities on the domain and with probabilities on both a set of possible worlds and the domain.

61 Probabilistic Ontologies in DLs Main types of encoded probabilistic knowledge: Terminological probabilistic knowledge about concepts and roles: Birds fly with a probability of at least Assertional probabilistic knowledge about instances of concepts and roles: Tweety is a bird with a probability of at least 0.9. Main types of reasoning problems: Satisfiability of the terminological probabilistic knowledge. Tight conclusions about generic objects (from the terminological probabilistic knowledge). Satisfiability of the assertional probabilistic knowledge. Tight conclusions about concrete objects (from both the terminological and the assertional probabilistic knowledge).

62 in DLs Probabilities about Generic and Concrete Objects Combining generic and concrete probability distributions: Conditioning: Generic distributions are conditioned on the (classical) information about concrete distributions. Probabilistic default reasoning: Generic distributions are constrained by the (not necessarily classical) information about the concrete distributions, and techniques from default reasoning resolve local inconsistencies. Minimum cross entropy: Generic and concrete distributions are combined via cross entropy minimization.

63 Probabilistic DLs in DLs R. Giugno, T. Lukasiewicz. P-SHOQ(D): A probabilistic extension of SHOQ(D) for probabilistic ontologies in the SW. In Proc. JELIA probabilistic generalization of the description logic SHOQ(D) (recently also extended to SHIF(D) and SHOIN (D)) terminological probabilistic knowledge about concepts and roles assertional probabilistic knowledge about instances of concepts and roles terminological probabilistic inference based on lexicographic entailment in probabilistic logic (stronger than logical entailment) assertional probabilistic inference based on lexicographic entailment in probabilistic logic (for combining assertional and terminological probabilistic knowledge) terminological and assertional probabilistic inference problems reduced to sequences of linear optimization problems

64 in DLs M. Jaeger. Probabilistic reasoning in terminological logics. In Proceedings KR probabilistic generalization of the description logic ALC terminological probabilistic knowledge about concepts and roles assertional probabilistic knowledge about concept instances, but no assertional probabilistic knowledge about role instances terminological probabilistic inference based on logical entailment in probabilistic logic (by solving linear optimization problems) assertional probabilistic inference based on cross entropy minimization relative to terminological probabilistic knowledge (by an approximation algorithm; no exact algorithm known so far)

65 in DLs D. Koller, A. Levy, and A. Pfeffer. P-Classic: A tractable probabilistic description logic. In Proceedings AAAI probabilistic generalization (of a variant) of the description logic Classic so-called p-classes express terminological probabilistic knowledge about concepts, roles, and attributes but assertional classical and probabilistic knowledge about instances of concepts and roles is not supported probabilistic semantics based on Bayesian networks determines exact probabilities for conditionals between concept expressions in canonical form probabilistic inference can be done in polynomial time, when the underlying Bayesian network is a polytree

66 Probabilistic OWL in DLs P. C. G. da Costa. Bayesian Semantics for the Semantic Web. PhD thesis, George Mason University, Fairfax, VA, USA, P. C. G. da Costa and K. B. Laskey. PR-OWL: A framework for probabilistic ontologies. In Proceedings FOIS probabilistic extension of OWL probabilistic semantics based on multi-entity Bayesian networks (MEBNs), which are a Bayesian logic that combines first-order logic with Bayesian probabilities: represents knowledge as parameterized fragments of Bayesian networks expresses repeated structure represents probability distribution on interpretations of associated first-order theory

67 Use of Probabilistic Ontologies in DLs Representation of terminological and assertional probabilistic knowledge (e.g., in the medical domain or at the stock exchange market). Information retrieval, for an increased recall (e.g., Udrea et al.: Probabilistic ontologies and relational databases. In Proc. CoopIS/DOA/ODBASE-2005). Ontology matching (e.g., Mitra et al.: OMEN: A probabilistic ontology mapping tool. In Proc. ISWC-2005). Probabilistic data integration, especially for handling ambiguous and controversial pieces of information.

68 Possibilistic DLs in DLs Generalization of DLs by possibilistic uncertainty, which is based on possibilistic interpretations rather than probabilistic interpretations. Directly extends propositional possibilistic logic Expressions: P α l or N α l, where α is a classical description logic axiom and l [0, 1] Example: N ( owns.porsche richperson carfanatic) 0.8 P (richperson golfer) 0.7 N ((Tom, 911): owns) 1 N (911: Porsche) 1, N (Tom : carfanatic) 0.7 logically implies P (Tom : golfer) 0.7.

69 Outline, Vagueness, and the Web Vagueness Vagueness in DLs 1, Vagueness, and the Web Sources of and Vagueness on the Web vs. Vagueness: A Clarification 2 The Semantic Web (SW) Web Ontology Languages Description Logics (DLs) 3 in DLs 4 Vagueness Vagueness in DLs

70 Vagueness and Logic Vagueness Vagueness in DLs Statements involve concepts with no exact definition, e.g., tall, small, close, far, cheap, expensive, is about, similar to,... A statements is true to some degree, which is taken from a truth space, e.g., Hotel Verdi is close to the train station to degree 0.83 The image is about a sun set to degree 0.75 Truth space: set of truth values L and a partial order ; usually: L = [0, 1], but also { 0 n, 1 n,..., n n } for an integer n 1. Interpretation: a function I mapping atoms A into L, i.e., I (A) L.

71 Vagueness Vagueness in DLs Problem: what is the interpretation of, e.g., φ ψ? E.g., given I (φ) = 0.83 and I (ψ) = 0.2, what is the result of I (φ ψ) = ? E.g., in multimedia retrieval: if an image region is white to degree 0.8 and the object is about a dog to degree 0.4, to which degree is the image about a white dog? That is, what is ? More generally, what is the result of n m, for n, m [0, 1]? The choice cannot be any arbitrary computable function, but has to reflect some basic properties that one expects to hold for a conjunction. Norms: functions that are used to interpret connectives such as,,, and t-norm: interprets conjunction s-norm: interprets disjunction Norms are compatible with classical two-valued logic.

72 Axioms for Norms Vagueness Vagueness in DLs Axioms for t-norms and s-norms: Axiom Name Conjunction Strategy Disjunction Strategy Tautology / Contradiction a 0 = 0 a 1 = 1 Identity a 1 = a a 0 = a Commutativity a b = b a a b = b a Associativity (a b) c = a (b c) (a b) c = a (b c) Monotonicity if b c, then a b a c if b c, then a b a c Axioms for implication and negation functions: Axiom Name Implication Strategy Negation Strategy Tautology / Contradiction 0 b = 1, a 1 = 1, 1 0 = 0 0 = 1, 1 = 0 Antitonicity if a b, then a c b c if a b, then a b Monotonicity if b c, then a b a c Usually, a b = sup{c : a c b} is used. It is called r-implication and depends on the t-norm only.

73 Typical Norms Vagueness Vagueness in DLs Łukasiewicz Logic Gödel Logic Product Logic Zadeh Logic a b max(a + b 1, 0) min(a, b) a b min(a, b) a b min(a + b, 1) ( max(a, b) a + b a b max(a, b) a b min(1 a + b, 1) 1 if a b min(1, b/a) max(1 a, b) a 1 a b otherwise ( 1 if a = 0 0 otherwise ( 1 if a = 0 0 otherwise 1 a Any other t-norm can be obtained as a combination of Lukasiewicz, Gödel, and Product t-norm. Zadeh Logic: not interesting for fuzzy logicians: it is a sublogic of Łukasiewicz: Z φ = Ł φ φ Z ψ = φ Ł (φ Ł ψ) φ Z ψ = Ł φ Ł ψ

74 Vagueness Vagueness in DLs Some additional properties of t-norms, s-norms, implication functions, and negation functions of various fuzzy logics. Property Łukasiewicz Logic Gödel Logic Product Logic Zadeh Logic a a = 0 a a = 1 a a = a a a = a a = a a b = a b (a b) = a b (a b) = a b (a b) = a b a (b c) = (a b) (a c) a (b c) = (a b) (a c) If all conditions in the upper part of a column are satisfied, then we collapse to classical two-valued logic, i.e., L = {0, 1}.

75 Propositional Fuzzy Logics Vagueness Vagueness in DLs Formulas: propositional formulas Truth space is [0, 1] Formulas have a degree of truth in [0, 1] Interpretation: is a mapping I : Atoms [0, 1] Interpretations are extended to formulas using norms to interpret connectives,,, I (φ ψ) = I (φ) I (ψ) ; I (φ ψ) = I (φ) I (ψ) ; I (φ ψ) = I (φ) I (ψ) ; I ( φ) = I (φ), r [0, 1] may appear as atom in formula, where I (r) = r.

76 Example, Vagueness, and the Web Vagueness Vagueness in DLs In Lukasiewicz Logic: φ = Cold Cloudy I Cold Cloudy I (φ) I max(0, ) = 0 I max(0, ) = 0 I max(0, ) = 0.6 I max(0, ) = 1....

77 Vagueness Vagueness in DLs Note: I (r φ) = 1 iff I (φ) r I (φ r) = 1 iff I (φ) r Semantics: I = φ iff I (φ) = 1 I = KB iff I = φ for all φ KB KB = φ iff for all I : if I = KB then I = φ

78 Vagueness Vagueness in DLs Let: φ KB = inf{i (φ) I = KB} (truth degree) φ KB = sup{r KB = r φ} (provability degree) then φ KB = φ KB. Also, in Lukasiewicz Logic: φ KB = 1 φ KB Proposition φ KB = min x such that KB {φ x} satisfiable.

79 Vagueness Vagueness in DLs Decision Algorithm (for Lukasiewicz Logic) We use MILP (Mixed Integer Linear Programming) to compute φ KB Let r [0, 1], variable or expresson 1 r (r variable), admitting solution in [0, 1], r = 1 r, r = r For each propositional letter p, let x p be a variable denoting the degree of truth of p Apply inference rules: r p x p r, x p [0, 1] p r x p r, x p [0, 1] r φ φ r φ r r φ r (φ ψ) x 1 φ, x 2 ψ, y 1 r, x i 1 y, x 1 + x 2 = r + 1 y, x i [0, 1], y {0, 1} (φ ψ) r x 1 φ, x 2 ψ, x 1 + x 2 = 1 r, x i [0, 1] r (φ ψ) φ x 1, x 2 ψ, r + x 1 x 2 = 1, x i [0, 1] (φ ψ) r x 1 φ, ψ x 2, y r 0, y + x 1 1, y x 2, y + r + x 1 x 2 = 1, x i [0, 1], y {0, 1} Now we have to solve a MILP problem of the form min c x s.t. Ax + By h where a ij, b ij, c l, h k [0, 1], x i admits solutions in [0, 1], while y j admits solutions in {0, 1}

80 Example, Vagueness, and the Web Vagueness Vagueness in DLs Consider KB = {0.6 p, 0.7 (p q)} Let us show that q KB = = max(1, ) = 0.3 Recall that q KB = min x. such that KB {q x} satisfiable: KB {q x} = {0.6 p, 0.7 (p q), q x, x [0, 1]} {x p 0.6, x q x, 0.7 (p q), {x, x p} [0, 1]} {x p 0.6, x q x, p x 1, x 2 q, x 1 x 2 = 1, {x, x p, x i } [0, 1]} {x p 0.6, x q x, x p x 1, x q x 2, x 1 x 2 = 1, {x, x p, x i } [0, 1]} = S It follows that 0.3 = min x such that Sat(S) satisfiable.

81 Predicate Fuzzy Logics Vagueness Vagueness in DLs Formulae: First-Order Logic formulae, terms are either variables or constants we may introduce functions symbols as well, with crisp semantics (but uninteresting), or we need to discuss also fuzzy equality (which we leave out here) Truth space is [0, 1] Formulae have a a degree of truth in [0, 1] Interpretation: is a mapping I : Atoms [0, 1] Interpretations are extended to formulae as follows: I( φ) = I(φ) 0 I(φ ψ) = I(φ) I(ψ) I(φ ψ) = I(φ) I(ψ) I( xφ) = sup c I I c x (φ) I( xφ) = inf c I Ic x (φ) where Ix c is as I, except that variable x is mapped into individual c Definitions of I = φ, I = T, = φ, T = φ, φ T and φ T are as for the propositional case

82 Fuzzy DLs, Vagueness, and the Web Vagueness Vagueness in DLs In classical DLs, a concept C is interpreted by an interpretation I as a set of individuals In fuzzy DLs, a concept C is interpreted by I as a fuzzy set of individuals Each individual is instance of a concept to a degree in [0, 1] Each pair of individuals is instance of a role to a degree in [0, 1]

83 Fuzzy ALC Vagueness Vagueness in DLs The semantics is an immediate consequence of the First-Order-Logic translation of DLs expressions Interpretation: Concepts: Assertions: Inclusion axioms: I = I C I : I [0, 1] R I : I I [0, 1] = t-norm = s-norm = negation = implication Syntax Semantics C, D I (x) = 1 I (x) = 0 A A I (x) [0, 1] C D (C 1 C 2 ) I (x) = I C 1 (x) I C2 (x) C D (C 1 C 2 ) I (x) = I C 1 (x) I C2 (x) C ( C) I (x) = C I (x) R.C ( R.C) I (x) = sup y I R I (x, y) C I (y) R.C ( R.C) I (u) = inf y I R I (x, y) C I (y)} a:c, r, I = a:c, r iff C I (a I ) r (similarly for roles) C D, r, individual a is instance of concept C at least to degree r, r [0, 1] Q I = C D, r iff inf x I (C I (x) D I (x)) r.

84 Basic Inference Problems Vagueness Vagueness in DLs Consistency: Check if knowledge is meaningful Is KB consistent, i.e., satisfiable? Graded instantiation: Check if individual a instance of class C to degree at least r KB = a:c, r? BTVB: Best Truth Value Bound problem a:c KB = sup{r KB = a:c, r }? Top-k retrieval: Retrieve the top-k individuals that instantiate C w.r.t. best truth value bound ans top k (KB, C) = Top k { a, v v = a:c KB }

85 Vagueness Vagueness in DLs Towards Fuzzy OWL Lite and OWL DL Recall that OWL Lite and OWL DL relate to SHIF(D) and SHOIN (D), respectively We need to extend the semantics of fuzzy ALC to fuzzy SHOIN (D) = ALCHOIN R + (D) Additionally, we add modifiers (e.g., very) concrete fuzzy concepts (e.g., Young) both additions have explicit membership functions

86 Concrete Fuzzy Concepts Vagueness Vagueness in DLs E.g., Small, Young, High, etc. with explicit membership function Use the idea of concrete domains: D = D, Φ D D is an interpretation domain Φ D is the set of concrete fuzzy domain predicates d with a predefined arity n = 1, 2 and fixed interpretation d D : n D [0, 1] For instance, Minor = Person hasage. 18 YoungPerson = Person hasage.young functional(hasage)

87 Modifiers, Vagueness, and the Web Vagueness Vagueness in DLs Very, moreorless, slightly, etc. Apply to fuzzy sets to change their membership function very(x) = x 2 For instance, slightly(x) = x SportsCar = Car speed.very(high)

88 Fuzzy SHOIN (D) Vagueness Vagueness in DLs Concepts: Assertions: Axioms: Syntax Semantics C, D (x) (x) A A(x) (C D) C 1 (x) C 2 (x) (C D) C 1 (x) C 2 (x) ( C) C(x) ( R.C) x R(x, y) C(y) ( R.C) x R(x, y) C(y) {a} x = a ( n R) y 1,..., y V n n. i =1 R(x, y i ) V 1 i <j n y i y j ( n R) y 1,..., y n+1. V n+1 i =1 R(x, y i ) W 1 i <j n+1 y i = y j FCC µ FCC (x) M(C) µ M (C(x)) R P P(x, y) P P(y, x) Syntax Semantics α a:c, r r C(a) (a, b):r, r r R(a, b) Syntax Semantics τ C D, r x r (C(x) D(x)), where is r-implication fun(r) x y z R(x, y) R(x, z) y = z trans(r) ( z R(x, z) R(z, y)) R(x, y)

89 , Vagueness, and the Web Vagueness Vagueness in DLs Example (Graded Entailment) audi _tt mg Car audi _tt mg ferrari _enzo SportsCar KB KB KB = = = = ferrari _enzo speed Car u hasspeed.very (High) hferrari _enzo:sportscar, 1i haudi _tt:sportscar, 0.92i hmg : SportsCar, 0.72i Managing and Napoli 2009 Thomas Lukasiewicz

90 Example (Graded Subsumption) Vagueness Vagueness in DLs Minor = Person hasage. 18 YoungPerson = Person hasage.young KB = Minor YoungPerson, 0.2 Note: without an explicit membership function of Young, this inference cannot be drawn

91 Example (Simplified Negotiation) Vagueness Vagueness in DLs a car seller sells an Audi TT for e, as from the catalog price. a buyer is looking for a sports-car, but wants to to pay not more than around e classical DLs: the problem relies on the crisp conditions on price more fine grained approach: to consider prices as fuzzy sets (as usual in negotiation) seller may consider optimal to sell above e, but can go down to e the buyer prefers to spend less than e, but can go up to e AudiTT = SportsCar hasprice.r(x; 30500, 31500) Query = SportsCar hasprice.l(x; 30000, 32000) highest degree to which the concept C = AudiTT Query is satisfiable is 0.75 (the possibility that the Audi TT and the query matches is 0.75) the car may be sold at e

92 Reasoning, Vagueness, and the Web Vagueness Vagueness in DLs Depends on the semantics and reasoning method (tableau-based or MILP-based) Tableaux method: MILP based method: under Zadeh semantics a tableau exists for fuzzy SHIN, solving the satisfiability problem classical blocking methods apply similarly in the fuzzy variant the management of General concept inclusions (GCI s) is more complicated compared to the crisp case a translation of fuzzy SHOIN to crisp SHOIN also exists (not addressed here) the tableaux method is not suitable to deal with fuzzy concrete concepts and modifiers the BTVB can be solved, but not efficiently under Zadeh semantics, Łukasiewicz semantics, and classical semantics exists for fuzzy ALC + linear modifiers + fuzzy concrete concepts exists for fuzzy SHIF + linear modifiers + fuzzy concrete concepts solves the BTVB as primary problem MIQP based method: using Mixed Integer Quadratically Constrained Programming optimization problem (MICQP) for product T-norm exists for fuzzy SHIF + linear modifiers + fuzzy concrete concepts. Important as it simulates probabilistic reasoning under independent event assumption. solves the BTVB as primary problem the fuzzydl solver also allows to mix all three semantics

93 Summary and Outlook Summary: probabilistic, possibilistic, and fuzzy description logics for the Semantic Web: syntax, semantics, computation problems, solution techniques, and applications in the (Semantic) Web. Current/Future Research: Implementation of systems for probabilistic, possibilistic, and fuzzy description logics. (Recent systems for all of them). Development of highly scalable probabilistic, possibilistic, and fuzzy description logics. Use of probabilistic, possibilistic, and fuzzy description logics in Web and Semantic Web applications. Learning of probabilistic, possibilistic, and fuzzy description logics from data.

94 Literature Thomas Lukasiewicz and Umberto Straccia. Managing and Vagueness in Description Logics for the Semantic Web. Journal of Web Semantics, 6(4), , November 2008.