Recall: Semantics of DL
z
z
Semantics defined by interpretations. We consider ALCN once more.
An interpretation I = (4, ·I), where
– Δ is the domain (a non-empty set)
– ·I is an interpretation function such that (for ALCN ):
Reasoning in Description Logic
• AI ⊆ 4
(where A is a class name)
• RI ⊆ 4 × 4 (where R is a Role (property) name)
• iI ∈ 4 (where i is an Individual name)
Jacques Fleuriot
• ⊥I = ∅ and >I = 4
• (¬C)I = 46 CI
• (C u D)I = CI ∩ DI and (C t D)I = CI ∪ DI
• (∀R. C)I = {a ∈ 4 | ∀b. (a, b) ∈ RI → b ∈ CI}
• (∃R. C)I = {a ∈ 4 | ∀b. (a, b) ∈ RI ∧ b ∈ CI}
• (· nR. C)I = {a ∈ 4 | k { b | (a, b) ∈ RI ∧ b ∈ CI } k · n}
• (≥ nR. C)I = {a ∈ 4 | k { b | (a, b) ∈ RI ∧ b ∈ CI } k ≥ n}
MASWS
MASWS
1
2
Example Interpretation for ALCN
Recall: ABox and TBox
(there may be others)
z
4 = {leo, jerome, ralph, lulu, bart, tania, robert}
z
z
GiraffesI = {jerome, lulu, robert}
LionsI = {leo}
z
eatsI = {(leo, jerome), (leo, lulu), (leo, robert), (tania, ralph), (tania, jerome)}
z
A set of DL sentences (making up the KB) about a domain D is divided
into:
– TBox: general knowledge with axioms of the form:
• C v D, C ≡ D, R v S, R ≡ S and R+ v R
Then we have (for instance) that:
where C, D are concepts, R, S are roles, and R+ is a set of transitive
roles.
(¬Giraffes)I = 46 GiraffesI
= {leo, jerome, ralph, lulu, bart, tania, robert}6 { jerome, lulu, robert}
Example: Sibling ≡ Brother t Sister
= {leo, ralph, bart, tania}
– ABox: assertions (specific knowledge):
(∀eats. Giraffes)I = {a ∈ 4 | ∀b. (a, b) ∈ eatsI → b ∈ GiraffesI} = {leo}
• jerome: Giraffe
(Giraffes u Lions)I = GiraffesI ∩ LionsI = ∅ = ⊥I
• (leo, jerome) : eats
(· 2 eats. Giraffes)I = ??
MASWS
MASWS
3
4
Rasoning about TBoxes
z
Satisfiability: A concept C is satisfiable wrt TBox T iff
there is an interpretation I of T such that CI ≠ ∅. Such an
interpretation is known as a model for C.
z
Subsumption: A concept C is subsumed by a concept D
wrt TBox T iff CI⊆ DI for all interpretations I of T. This
can be denoted by T ² C v D.
z
Equivalence: A concept C is equivalent to a concept D wrt
TBox T iff CI = DI for all interpretations I of T. This can
be denoted by T ² C ≡ D.
z
Reduction to Subsumption
z
Subsumption (traditionally) acts as the basic reasoning
mechanism in DL systems.
– This is sufficient in order to implement other inferences since
they can be reduced to subsumption checking (is C v D?) by
the following theorem (provided the DL system contains u and
⊥) :
For concepts C and D, we have
C is unsatisfiable wrt TBox T iff T ² C v ⊥
Disjointness: Two concepts C and D are disjoints wrt TBox
T iff CI ∩ DI = ∅ for all interpretations I of T.
T ² C ≡ D iff T ² C v D and T ² D v C
C and D are disjoints wrt TBox iff T ² C u D v ⊥
Note: In the special case where the TBox is empty, we can write ² C v D and
² C ≡ D, where appropriate.
MASWS
MASWS
5
6
Reduction to Unsatisfiability
z
Reducing Unsatisfiability
If the DL system allows complement (negation), subsumption,
equivalence, and disjointness can be reduced to a satisfiability
problem:
z
For concepts C and D, we have
Let C be a concept. Then the following are equivalent
with respect to a TBox T :
C is subsumed by D wrt TBox T iff C u ¬D is unsatisfiable
z
Unsatisfiability is a special case of each of the following:
subsumption, equivalence, and disjointness:
T ² C ≡ D iff both C u ¬D and ¬C u D are unsatisfiable
1.
C is unsatisfiable;
C and D are disjoints wrt TBox T iff C u D is unsatisfiable
2.
T ²Cv⊥;
3.
T ²C≡⊥;
4.
C and > are disjoint.
So, to obtain decision procedures for the above, we just need
algoritms that decide the satisfiability of concepts
– DL needs to support u and negation (¬) for arbitrary concepts
MASWS
MASWS
7
8
Reasoning Tasks for ABoxes
z
z
z
Structural Subsumption Algorithm(s)
ABox contains two types of assertions: concept assertions of
the form a: C and role assertions of the form (a,b): R.
z
Algorithms usually have two phases:
– normalize the concepts to be tested for subsumption
– compare the syntactic structure of the normal forms
An ABox A is consistent wrt a TBox T iff there is an
interpretation that is a model for both A and T.
z
Normal Forms:
– An atomic concept A is in normal form
Instance checking: An assertion α is entailed by A, written A ²
α, iff every interpretation that satisfies, that is, every model of
A, also satisfies α.
– The negation of an atomic ¬A concept is in normal form
– ⊥ is in normal form
– A concept C, in a DL with only u, ∀, ⊥, ¬, is in normal form if it has
the following form:
– If is α of the form a: C then, we have A ² a: C iff A ∪ {¬ a: C} is
inconsistent.
A1 u … u Am u ∀R1. C1 u … u ∀Rn. Cn
– A concept is C satisfiable iff {a: C} is consistent wrt to ABox A for
an arbitrary a.
where A1, … , Am are distinct atomic concept names different
from ⊥ and R1, … , Rn are distinct role names and C1, … , Cn
are concepts in normal form.
Note: a: C and (a,b): R are often written as C(a) and R(a,b) respectively.
MASWS
MASWS
9
10
Subsumption Check Algoritm:
Comparison
Converting to Normal Form
z
The second phase of the algorithm for ALC :
The conversion to normal form for previous DL relies on a number of
identities:
– ⊥ is subsumed by every concept;
– C is subsumed by D (C v D) if:
– AuA≡A
– AuB≡BuA
– A u (B u C) ≡ (A u B) u C
– ∀R.(A u B) ≡ (∀R. A) u (∀R. B)
•
B1 u … u Bk u ∀S1. D1 u … u ∀Sr. Dr is the normal form of D
1. for all i, 1 · i · k there exists j, 1 · j · m such that Bi = Aj
– Au⊥≡⊥
2. for all i, 1 · i · r there exists j, 1 · j · n such that Si = Rj and
Cj v D i
To deal with number restrictions, we need additional rules such as:
– m > n → (· nR. C) u (≥ mR. C) ≡ ⊥
– n ≥ 1 → (≥ nR. C) u (∀R. ⊥) ≡ ⊥
z
A1 u … u Am u ∀R1. C1 u … u ∀Rn. Cn is the normal form of C
and
– A u ¬A ≡ ⊥
z
•
z
Exercise: Work out the normal form of: (∀R.¬A) u A u (∀R. A u (∀R. B))
To deal with number restrictions (for ALCN), we need additional
rules such as:
(≥ nR. C) v (≥ mR. C) iff n ≥ m
MASWS
MASWS
11
12
Examples and Remark
z
CvC
z
CuDvC
z
∀R. (C u D) v ∀R.C
z
A u B u ∀R.(C u D u ∀S. E) u ∀T. ⊥ v B u ∀T. F u ∀ R.(C u ∀ S. E)
On Tableau Algorithms
z
These algorithms do not do subsumption testing directly
z
Negation is used to reduce subsumption to (un)satisfiability of
concepts using the fact that:
C v D iff C u ¬D is unsatisfiable
z
They act as decision procedures for unsatisfiability
– if a formula is decidable, they can exhibit a model
z
For DLs with full negation (i.e. not just applied to concept names) and
full existential restriction ∃R. C, structural subsumption cannot be used.
Instead, one can use a tableau-based algorithm.
Idea: Incrementally build model by decomposing negation
normal form (nnf) formula in a top-down fashion. All
possibilities are looked by applying a number of
‘transformation rules’ exhaustively until no more rules apply.
If no obvious contradiction or clash is found, then the formula
is satisfiable, otherwise, it is unsatisfiable.
MASWS
MASWS
13
14
Tableau Algorithm: Informal View
z
List all assertions in ABox A.
z
Apply specific completion rules and add resulting assertions to A
yielding A’.
Negation Normal Form
z
• ¬(C u D) ≡ ¬C t ¬D
– Completion rules can be:
De Morgan’s laws
• ¬(C t D) ≡ ¬C u ¬D
• deterministic yielding a uniquely determined set of assertions
• ¬∀R.C ≡ ∃R. ¬C
• non-deterministic yielding several alternative sets of assertions,
each set corresponding to a branch of the tableau
z
An ALC formula can be transformed into negation normal
form (nnf) using the following identities:
• ¬∃R. C ≡ ∀R. ¬C
Apply the completion rules until either a contradiction or clash
occurs in every branch, or there is a completed branch where no
more rule is applicable.
z
In nnf, the negation is pushed as far inward as possible
until it is applied to a concept name.
– A completed branch of the tableau gives a model for the completed
system.
MASWS
MASWS
15
16
Transformation Rules for
Satisfiability Algorithm
Clash
branching
z
A clash corresponds to a logical contradiction.
z
It occurs when a set of assertions contains both
C(a) and ¬C(a).
z
A clash is an unsatisfiable set of assertions. The
tableau algorithm closes a branch as soon as a
clash is encountered.
Recall:
C(a) is a: C
R(a,b) is (a,b): R
In algorithm:
(C1 u C2)(x) stands for
x : (C1 u C2)
MASWS
MASWS
17
18
Example
Example
Does ∀R. (C u D) v ∀R. C u ∀R. D ?
Does ∃R. C u ∃R. D v ∃R.(C u D)?
Check whether ∀R. (C u D) u ¬(∀R. C u ∀R. D) is unsatisfiable using the tablean
algorithm:
Tableau:
1.
( ∀R. (C u D) u ¬(∀R. C u ∀R. D))(x)
2.
(∀R. (C u D) u (¬∀R. C t ¬∀R. D))(x)
by De Morgan’s law
3.
(∀R. (C u D) u (∃R. ¬C t ∃R. ¬D))(x)
by nnf
4.
(∀R. (C u D))(x)
5.
(∃R. ¬C t ∃R. ¬D)(x)
(∃R. ¬C)(x)
6.
by
→∀ –rule to 4:
by
by
→∃ –rule
by
→∃ -rule
(C u D)(z)
by
→∀ –rule to 4
9b’.
C(z)
by
→u -rule
9b.
D(z)
R(x, y)
7b.
R(x, z)
7a’.
¬C(y)
7b’.
¬C(z)
(C u D)(y)
8.
9a.
C(y)
9b.
D(y)
Tableau:
(∃R. ¬D)(x)
7a.
8.
→u -rule
by
→u –rule
X
X
MASWS
MASWS
19
20
Expressive Roles
z
Reading
The DL(s) considered only allow for atomic roles. The expressivity of the
languages can be extended by allowing role constructors that build
complex roles from atomic ones:
z
Inverse role: (R−)I = {(b, a) ∈ 4 × 4 | (a, b) ∈ RI}
z
Role composition:
(R ◦ S )I = {(a, c) ∈ 4 × 4 | ∃b. (a, b) ∈ RI ∧ (b, c) ∈ SI }
z
Transitive closure: (R+)I =
∪n≥1 (RI)n
(RI)1 = RI
z
F. Baader and W. Nutt. Basic Description Logics. In Baader,
Calvanese, McGuinnes Nardi and Patel-Schneider, (eds). The
Description Logics Handbook, chapter 2, pages 43-95. Cambridge
University Press, 2003. Available from MASWS web-page.
z
F. Baader and U. Sattler. Tableau Algorithms for Description
Logics. Proceedings of the International Conference on
Automated Reasoning with Tableaux and Related Methods
(Tableaux 2000), Vol. 1847, Springer-Verlag. Available at
http://citeseer.ist.psu.edu/baader00tableau.html
(RI)n+1 = (RI)n ◦ RI
z
etc.
z
In general, there’s a price to pay in terms of reasoning/deciding a formula.
MASWS
MASWS
21
22
© Copyright 2026 Paperzz