Professional Documents
Culture Documents
Description Logics of Minimal Knowledge and Negation As Failure
Description Logics of Minimal Knowledge and Negation As Failure
Description Logics of Minimal Knowledge and Negation As Failure
We present description logics of minimal knowledge and negation as failure (MKNF-DLs), which
augment description logics with modal operators interpreted according to Lifschitzs nonmonotonic
logic MKNF. We show the usefulness of MKNF-DLs for a formal characterization of a wide variety
of nonmonotonic features that are both commonly available in frame-based systems, and needed
in the development of practical knowledge-based applications: defaults, integrity constraints, role,
and concept closure. In addition, we provide a correct and terminating calculus for query answering
in a very expressive MKNF-DL.
Categories and Subject Descriptors: I.2.3 [Artificial Intelligence]: Deduction and Theorem
ProvingNonmonotonic reasoning and belief revision; I.2.4 [Artificial Intelligence]: Knowledge
Representations Formalisms and MethodsModal logic; F.4.1 [Mathematical Logic and Formal
Languages]: Mathematical LogicComputational logic
General Terms: Theory
Additional Key Words and Phrases: Description Logics, frame-based systems, nonmonotonic modal
logics, tableau calculi
1. INTRODUCTION
Frame-based systems constitute one of the main tools for the development
of knowledge-based applications [Fikes and Kehler 1985]. The main feature
This research has been funded by Italian MURST, Tecniche di ragionamento non monotono and
Linguaggi per la modellizzazione concetualle dei requisiti, by EC Esprit LTR Project, Foundations
of Data Warehouse Quality, by ASI (Italian Space Agency), and by Italian CNR, Description Logics
per la Modellazione e lAnalisi di Dominio.
Authors addresses: F. M. Donini, Departimento di Elettrotecnica ed Elettronica, Politecnico di
Bari, Via G. Re David 200, I-70125 Bari, Italy; email: donini@poliba.it; D. Nardi and R. Rosati,
Dipartimento di Informatica e Sistemistica, Universita` di Roma La Sapienza, Via Salaria 113,
I-00198 Roma, Italy; email: {nardi,rosati}@dis.uniroma1.it.
Permission to make digital/hard copy of part or all of this work for personal or classroom use is
granted without fee provided that the copies are not made or distributed for profit or commercial
advantage, the copyright notice, the title of the publication, and its date appear, and notice is given
that copying is by permission of the ACM, Inc. To copy otherwise, to republish, to post on servers,
or to redistribute to lists, requires prior specific permission and/or a fee.
178
F. M. Donini et al.
179
180
F. M. Donini et al.
MKNF [Rosati 1999a], their extension to the fragment of first-order logic captured by description logics is necessary.
To this end, we first characterize epistemic interpretations as sets of denumerable DL assertions. This is a new and original characterization for firstorder MKNF. Then, we present a calculus for answering queries posed to
MKNF-DL knowledge bases, which provides a sound and complete reasoning
method. The proposed deduction method proves the decidability of a significant
first-order extension of propositional MKNF.
The paper is organized as follows. We first present DLs, then MKNF-DLs
(obtained by adding to DLs the epistemic operators of MKNF). We then discuss
the representational features of MKNF-DLs by considering several forms of
nonmonotonic reasoning and integrity constraints. Subsequently, we address
reasoning in these logics by providing a calculus for checking satisfiability of
MKNF-DL knowledge bases. To this end we first extend the characterization
of preferred models in propositional MKNF to MKNF-DLs; then we present a
tableau-based calculus for computing preferred models; finally we show termination and correctness of the calculus. We conclude the paper with a discussion
comparing nonmonotonic reasoning in logic programming with MKNF-DLs.
2. SYNTAX AND SEMANTICS OF MKNF-DLS
In this section we first briefly recall description logicsin particular, the basic
DL ALCthen define syntax and semantics of modal extensions with operators
from MKNF.
2.1 Description Logics
In frame-based systems, knowledge is represented by defining frames [Fikes
and Kehler 1985; Karp 1992], which can be regarded as intensional descriptions
of classes of objects. Frames are related to each other through super-classes and
slot specifications, the latter requiring that the values of a slot belong to the
class of objects represented by another frame.
Description logics (DLs) were developed with the aim of capturing the firstorder semantics of frame systems in a piecewise syntax: each possible property
about frames is expressed by a distinguished syntactic construct; then each set
of constructs identifies a different DL.
Using DLs, frames are represented as concept expressions, which are interpreted as unary predicates. A frame structure, containing super-classes and
slot specifications is represented by a conjunctive concept expression, where
there is a conjunct for each superclass and a conjunct for each slot.
For example, the concept expression manager u female describes the class of
individuals that are (i.e. the concept of) female managers, female, manager being its super-classes. Other boolean connectives can be used, as in the concept
mathematician t engineer which expresses a class formed by merging mathematicians and engineers, or as in UScitizen which describes non-US citizens.
Also, the concept expression employee u MANAGES.employee describes the class
of individuals that are employees, and that can manage employees only. The expression MANAGES.employee is a universal role restriction, which corresponds to
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
181
1 \ CI
C1I C2I
C1I C2I
d 1 | d 0 .(d, d 0 ) Ra I and d 0 C I
182
F. M. Donini et al.
manager v employee
managers are employees
manager v MANAGES.employee every manager manages at least one employee
MANAGES.> v manager
only managers can manage
> v MANAGES.employee
only employees are managed
Observe that, intuitively, the last two axioms constrain the two arguments of
the role MANAGES to be in manager and employee, respectively (see Buchheit
et al. [1993]). More precisely, for every model I of the TBox, it holds MANAGESI
managerI employeeI .
Assertions regarding such a TBox could be:
manager(Bob) employee(Paula) MANAGES(Ann,Marc)
manager(Ann) employee(Marc)
Observe that employee(Ann) and employee(Bob) follow from the TBox. Observe
also that the assertions manager(Ann) and employee(Marc) are redundant, since
they follow from the TBox and the assertion MANAGES(Ann, Marc). Moreover, it
follows that there is at least one employee managed by Bobbut we do not
know who; it could also be an individual not mentioned in the above ABox. If
we instead would like to impose that the managed employee appears explicitly
in the ABox, we would need to express constraints about the possible states of
an ABox. These constraints are not expressible in DLs.
We now show how to capture the intuition underlying the last remark by
extending description logics with two epistemic operators.
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
183
=
=
=
=
=
=
1 \ (C) I,M,N
(C1 ) I,M,N (C2 ) I,M,N
(C1 ) I,M,N (C2 ) I,M,N
{d 1 | d 0 .(d, d 0 ) (R) I,M,N and d 0 (C) I,M,N }
1 Recall that Lifschitz [1994] introduced MBNF, slightly extending the MKNF framework. However,
we refer to the original MKNF, since it is sufficient for our goal of representing non-first-order
features of frame systems.
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
184
F. M. Donini et al.
(AC)
I,M,N
(KRa )
I,M,N
(ARa )
I,M,N
(C)J ,M,N
J N
(Ra )J ,M,N
J M
(Ra )J ,M,N
J N
Notice that, if one considers non-epistemic expressions, the semantics is defined as for ordinary DLs on I. Conversely, epistemic expressions are defined
on epistemic interpretations. More precisely, an individual d 1 is an instance
of a concept C in a given interpretation I iff d (C) I,M,N . An individual d 1
is an instance of a concept KC (i.e. d (KC) I,M,N ) iff d C J ,M,N for all interpretations J M. In other words, an individual is known to be an instance
of a concept if it belongs to the concept interpretation of every possible world in
M. Similarly, an individual d 1 is an instance of a concept KR.> iff there
is an individual d 0 1 such that (d, d 0 ) R J ,M,N for all J M.
The interpretation of A is analogous: an individual d 1 is an instance
of a concept AC (i.e. d (AC) I,M,N ) iff d C J ,M,N for all interpretations
J N . At this point the definitions of A and K are indeed symmetrical. Later,
the preference criterion will provide different meanings to the operators, by
introducing minimality for K and default assumption for A.
In principle, for every DL one can define its MKNF extension. In general,
given a particular description logic DL, we denote with DLKN F its extension
with the modal operators of MKNF, and define it analogously to ALCKN F . The
following definitions are valid for every MKNF-DL, and they obviously apply
to ALCKN F .
A DLKN F knowledge base 6 is defined as a pair 6 = hT , Ai, where the TBox
T is a finite set of inclusion statements of the form C v D, where C, D are
DLKN F -concepts, and the ABox A is a finite set of membership assertions of
the form C(a) or R(a, b), where C is a DLKN F -concept, R is a DLKN F -role, and
a, b are names of individuals.
The truth of inclusion statements in an epistemic interpretation (I, M, N )
is defined in terms of set inclusion: C v D is satisfied in (I, M, N ) iff
(C) I,M,N (D) I,M,N . Assertions are interpreted in terms of set membership:
C(a) is satisfied in (I, M, N ) iff a (C) I,M,N and R(a, b) is satisfied in (I, M, N )
iff (a, b) (R) I,M,N .
An inclusion C v D is satisfied by a structure (M, N ) (which is denoted
as (M, N ) |= C v D) iff each interpretation I M is such that (I, M, N )
satisfies C v D. An assertion C(a) (resp. R(a, b)) is satisfied by (M, N ) (denoted
respectively as (M, N ) |= C(a), (M, N ) |= R(a, b)) iff each interpretation I M
is such that (I, M, N ) satisfies C(a) (resp. R(a, b)).
Analogously, a DLKN F knowledge base 6 is satisfied by a structure (M, N )
(denoted as (M, N ) |= 6) iff each interpretation I M is such that every
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
185
186
F. M. Donini et al.
187
188
F. M. Donini et al.
be viewed as KBs with a single epistemic model. In the more general case of
epistemic KBs with multiple models, ICs can be viewed as an a fortiori check
which establishes which of the models are actually allowed. In particular, many
forms of ICs impose properties that must hold for the individuals mentioned
explicitly in the KB. The modal operator K appears as an appropriate way to
distinguish such individuals from the others. Moreover, conditions imposed by
ICs are consistency conditions which cannot change the content of the KB. In
other words, augmenting the KB with ICs can only eliminate the models which
do not satisfy the ICs. As we shall see, the modal operator A turns out to be
well-suited for this purpose.
We now show that the combination of the modalities K and A provides for the
formalization in MKNF-DLs of sophisticated constraints on the KB content.
Example 3.2. Let us consider the IC each known employee must be known
to be either male or female, which is meant to avoid any situation where an
employee has been added to the KB without specifying her/his sex.
The solution to the above problem lies in the use of the autoepistemic belief
operator A. Indeed, if we add to a KB the DLKN F inclusion
I1 = Kemployee v (A male t Afemale)
we formalize the intended meaning of the IC.
In fact, suppose 61 contains only one assertion, employee(Bob). Of course,
61 does not satisfy the IC. A model M for 61 {I1 } should satisfy the assertion
A male t Afemale(Bob),
that is, it should be such that either Bob
T
T
J ,M,M
J ,M,M
male
or
Bob
189
since employee(Bob) holds, the assertion A male t Afemale(Bob) must hold. Now,
since there is no reason to conclude either male(Bob) or female(Bob) from 61 ,
the autoepistemic beliefs A male(Bob) and Afemale(Bob) are not consistent with
the objective knowledge of 61 , therefore A male t Afemale(Bob) is inconsistent
with 61 .
We remark that, for propositional logics, a precise correspondence between
the A operator and Moores L operator of Autoepistemic Logic (AEL) has been
proved [Rosati 1999a, Theorem 3.5]. Namely, given an AEL propositional formula , its models are exactly the models of the MKNF formula K([L/A]),
where [L/A] denotes the substitution of L with A everywhere. From this correspondence, the above assertion A male t Afemale(Bob) can also be understood
as a variation of the classical inconsistent autoepistemic theory {L}. This remark may further clarify our formalization of ICs as believed sentences: if such
a belief is not supported by the objective knowledge, then an inconsistency
arises. We now show a more sophisticated form of IC.
Example 3.3. The IC Each known employee has a known social security
number, which is known to be valid can be correctly formalized by the ALCKN F
inclusion
I2 = Kemployee v ASSN.Avalid
In factas we show belowan ALCKN F ABox 62 does not satisfy the IC iff
62 {I2 } is inconsistent.
Suppose in 62 there is the assertion employee(Bob), but there is no SSN for
Bob. A model M for 62 should satisfy ASSN.Avalid(Bob), that is, there is an
individual x 1 such that M |= ASSN(Bob, x) and M |= Avalid(x). Now con0
sider another interpretation I 0 , such that x 6 validI , and let M0 = M I 0 .
Observe that T
the structure (M0 , M) satisfies 62 {I2 }, since the interpretation
0
of Avalid is J M validJ ,M ,M , same as the interpretation in the structure
(M, M). Therefore, M is not a model for 62 {I2 }, and from the generality of
M we conclude that 62 {I2 } has no models.
A similar reasoning applies in case there is at least one valid x, which is not
known to be Bobs SSN.
To summarize the intuitions of the previous examples, we can state a general
theorem in MKNF-DLs, relating ICs as epistemic queries a` la Reiter and ICs
as epistemic assertions in our setting. We recall that an ICs as epistemic query
is an ALCK-concept CC uses only the modal operator Ksuch that for every
individual a in a KB 6, it must be 6 |= C(a) (otherwise, the IC is violated). In
our setting, we can replace every K in C with A, and use the concept A(C[K/A]),
where C[K/A] is the concept expression obtained from C by replacing each occurrence of K with A, inside the KB, in an inclusion whose antecedent holds for
every individual in O6 . If the IC is violated in a model M of the initial KB, then
M is not a model of the new KB. This ability of MKNF-DLs to embed ICs is formalized in the following theorem. From now on, we call subjective A-assertion
any ALCKN F assertion in which there are no occurrences of the modality K and
all occurrences of atomic concepts lie within the scope of the modality A.
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
190
F. M. Donini et al.
191
I I = I I {a}. Then, since the concept name I does not occur in 60 , it follows
that the structure (M {I 0 }, M) satisfies 6, thus contradicting the hypothesis
that M is a model for 6. Consequently, for each a 1 O6 , (M, M) does not
satisfy KI (a). Thus, (M, M) satisfies the inclusion KI v A(C[K/A]), which
implies that (M, M) satisfies 6 0 . Moreover, by Definition 2.2, for each set of
interpretations M0 such that M0 M, (M0 , M) does not satisfy 6, and since
6 0 contains every inclusion and every assertion in 6, it follows that (M0 , M)
does not satisfy 6 0 , which implies that M is a model for 6 0 .
We note that when the IC must hold only for the individuals that belong
to a particular conceptemployee in the previous examplesone can use this
concept as an antecedent in the inclusions.
3.3 Role and Concept Closure
Finally, we show how two particular forms of closed-world reasoning, namely
role closure and concept closure, can be nicely formalized with a combination of minimal knowledge and autoepistemic belief (that is, bimodal DLKN F concepts). These kinds of closure appear as very useful tools in knowledge
representation.
Closure on roles is available both in CLASSIC [Patel-Schneider et al. 1991]
and in LOOM [MacGregor 1988]. The idea is to state that all fillers for a role
about an individual have been inserted in the knowledge basefor example,
all subparts of a given object, all children of a person. This enables one to
conclude universal role quantifications, restricted to the individuals explicitly
filling the role in the KB. We believe that reasoning enabled by role closure can
be more correctly captured by epistemic concepts. To introduce the problem, we
adapt from Donini et al. [1998b, Sect. 4.2] the following example.
Example 3.5 Role closure, 1.
192
F. M. Donini et al.
When epistemic sentences can also appear inside the knowledge base, role
closure may yield more subtle results, as shown by the following example.
Example 3.6 Role closure, 2. Let 64 = hT , Ai be the following ALC KB:
v programmingSkills
engineer
T = mathematician v programmingSkills
mathematician v
engineer
A = {MANAGES(Ann, Marc), MANAGES(Ann, Paula)}
where the three axioms in T express that both engineers and mathematicians
have programming abilities, and the same employee cannot be both a mathematician and an engineer. Now, suppose we want to formalize the property:
one of the known employees managed by Ann is known to be an engineer,
and another (different) one is known to be a mathematician. One would like
to conclude that all known employees managed by Ann are known to have
programming abilities.
It turns out that the correct formalization is provided by the use of both
modalities A and K. Formally: let
I4 = AMANAGES.Kengineer u AMANAGES.Kmathematician(Ann)
Then, hT , A {I4 }i |= KMANAGES.KprogrammingSkills(Ann).
Epistemic operators make it natural to extend the notion of closure to concept
expressions.
Example 3.7 Concept closure. Let 65 be the following ALC KB:
UScitizen(Paula)
UScitizen(Marc)
Observe that, following the previous examples for role closure, closing the role
MANAGES for Ann in 65 leads to the non-monotonic conclusion that All employees
that are known to be managed by Ann are known to be US citizens. In formulae,
65 |= KMANAGES.KUScitizen(Ann)
(2)
Suppose now we add to 65 the information I5 , informally stated as: Ann manages one of the already known non-US-citizens of the company. (Here already
means that you should not infer from I5 that there can be another non-UScitizen you dont know about.) Now, since Paula and Carl are the only non-UScitizens in 65 , we should conclude that either Paula is managed by Ann, or Carl
is. Now what about the non-monotonic conclusion (2)? Intuitively, it should be
false, either because of Paula, or because of Carl.
This intuitive behavior can be formally obtained by letting
I5 = KMANAGES.AUScitizen(Ann)
sinceas we now show65 {I5 } has only two epistemic models, one in which
Paula is one of Anns employees, and the other in which Carl is.
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
193
We first prove that in every model either Paula or Carl is managed by Ann.
In fact, suppose there is a model N such that in N neither Paula nor Carl are
managed by Ann, i.e. N 6|= MANAGES(Ann, Paula) and N 6|= MANAGES(Ann, Carl).
Since N models I5 , there should be another individual x 1 such that N |=
KMANAGES(Ann, x) and N |= AUScitizen(x), with x 6= Paula and x 6= Carl.
However, N |= AUScitizen(x) means that, for every interpretation I N ,
it is x (UScitizen)I . Now consider the interpretation I 0 in which x
0
(UScitizen)I , and let M = N {I 0 }. Observe that (M, N ) still satisfies 65 {I5 },
since for every interpretation I M, it is
\
(UScitizen)J ,M,N
x (AUScitizen)I,M,N =
J N
194
F. M. Donini et al.
4. REASONING IN MKNF-DLS
In this section we study reasoning in MKNF-DL knowledge bases. Specifically,
we address satisfiability of ALCKN F knowledge bases, containing both a TBox
and an ABox. We then show that instance checking can be reduced to satisfiability, and discuss the problem of subsumption.
We develop our reasoning method for a subset of ALCKN F , which allows one
to express all the nonmonotonic features described in the previous section
defaults, rules, integrity constraints, epistemic queries, and role and concept
closure. The resulting language amounts to a large first-order fragment of
MKNF, and it represents a significant extension over previous results on reasoning in nonmonotonic modal logics. In fact, previous work on reasoning
in first-order autoepistemic logic [Niemela 1993] does not allow any form of
quantifying-in. Another proposal for reasoning on a first-order fragment of nonmonotonic modal logic is Lakemeyer [1996], which is based on a formal model
of limited reasoning which makes it difficult to compare with our work.
Our strategy, which we inherit from several approaches to reasoning in
propositional modal nonmonotonic logics (e.g., Gottlob [1992]; Marek and
Truszczynski
[1993]; Niemela [1992]), aims at reducing the reasoning problem
in ALCKN F to a number of reasoning problems in the underlying non-modal
logic ALC. The basis for such a technique is the characterization of each model
of an ALCKN F KB 6 by means of an ALC KB: in this way, the set of ALC
KBs representing all the models of 6 constitutes a non-modal representation
of 6, and allows for using classical reasoning techniques in order to solve the
reasoning problem in ALCKN F .
Formally, we say that a set of interpretations M is first-order representable
(resp. ALC representable) if there exists a first-order theory (resp. an ALC KB)
6M such that
M = {I : I satisfies 6M }
In other words, we want to characterize the interpretations belonging to M by
means of a first-order theory (or an ALC KB). Of course, since ALC corresponds
to a fragment of first-order logic, if M is ALC representable then M is also
first-order representable. Notice that the theory 6M may be either finite or
infinite.
In the following, we study both first-order representability and ALC representability of the models of ALCKN F KBs. We first prove that, in the general
case, the models of an ALCKN F KB are not first-order representable. Then,
we define the subset of subjectively quantified ALCKN F KBs, and show that
such KBs constitute a maximal subset of ALCKN F KBs whose models are both
first-order and ALC representable. While an ALC-based characterization of
subjectively quantified KBs is an interesting result in itself, it is not sufficient for obtaining an effective procedure for ALCKN F based on an ALC reasoner. Therefore, we define the class of simple ALCKN F KBs, a subset of subjectively quantified ALCKN F KBs whose models admit a characterization in
terms of finite ALC KBs, and prove decidability of reasoning in such a class of
ALCKN F KBs.
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
195
In order to prove the thesis, we show the existence of a model M for 6 such
that there exists no first-order theory 9 such that
M = {I : I satisfies 9}
The proof is organized in the following way. First, we define a set of interpretations M, and prove that M is a model for 6. Then, we show that M cannot
be characterized in terms of a first-order theory.
We start by defining M. First, since there are no occurrences of the operator
A in 6, from the semantic definition of ALCKN F in Section 2.2, the evaluation
of 6 in a structure (M, M0 ) does not depend on M0 . Therefore, in the following
we say that a set of interpretations M satisfies 6 (or any assertion without
occurrences of the operator A) iff every interpretation I M is such that
(I, M, M0 ) satisfies 6 (for any choice of the set M0 ). Let d 1 , d 2 , . . . be an ordering
of the elements in 1, let Odd be the set of odd-positioned elements in such an
ordering, and let S be the infinite set of formulas
^
S=
R(d 1 , d ) C(d )
d Odd
We now show that M is a model for 6. From Definition 2.2, we have to prove
that M is a model for 6 and that there is no set of interpretations M0 such that
M0 M and M0 satisfies 6. First, we prove that M satisfies 6. From the above
definition of M it follows that M satisfies an assertion of the form KR(d 1 , d ) iff
d Odd, and M satisfies an assertion of the form KC(d ) iff d Odd. Moreover,
each I M is such that I satisfies D(d ) and P (d 1 , d 0 ) for some d , d 0 Odd.
Consequently, for each I M, (I, M) satisfies the assertions KR.D(d 1 ) and
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
196
F. M. Donini et al.
P .KC(d 1 ). Finally, for each I M there exists d 1 Odd such that I satisfies
R(d 1 , d ) and I satisfies C(d ). Since, for such a d , M does not satisfy KC(d ), it
follows that, for each I M, (I, M) satisfies the assertion R.(C u KC)(d 1 ).
Therefore, M satisfies 6.
Then we prove that there is no set of interpretations M0 such that M0 M
and M0 satisfies 6. Let M0 be a set of interpretations, and suppose that M0 M
and M0 satisfies 6. Furthermore, let I be an interpretation such that I
M0 M. From the above definition of M it follows that either I does not
satisfy S or I does not satisfy any formula of the form f (i, j, l ) such that i 0,
j Odd, l Odd, which implies that at least one of the following conditions
holds:
(1)
(2)
(3)
(4)
(5)
First, suppose Condition 1 holds. Then, there exists d k Odd such that M0
does not satisfy KR(d 1 , d k ). Now, all interpretations J contained in M and
satisfying one of the formulas f (i, k , l ) such that i > 0 and l Odd are such
that (J , M0 ) does not satisfy KR.D(d 1 ). Consequently, M0 does not satisfy 6.
Contradiction. Hence, Condition 1 does not hold.
Now suppose Condition 2 holds. Then there exists d k Odd such that M0 does
not satisfy KC(d k ). Now, all interpretations J contained in M and satisfying
one of the formulas f (i, j, k ) such that i > 0 and j Odd are such that (J , M0 )
does not satisfy P .KC(d 1 ). Consequently, M0 does not satisfy 6. Contradiction.
Hence, Condition 2 does not hold.
Then suppose Condition 3 holds. Then (I, M0 ) does not satisfy the assertion
KR.D(d 1 ), since for each d 0 1, M0 satisfies KR(d 1 , d 0 ) only if M satisfies
KR(d 1 , d 0 ), and M satisfies KR(d 1 , d 0 ) only if d 0 Odd. Consequently, M0 does
not satisfy 6. Contradiction. Hence, Condition 3 does not hold.
Suppose now that Condition 4 holds. Then (I, M0 ) does not satisfy the assertion P .KC(d 1 ), since for each d 0 1, M0 satisfies KC(d 0 ) only if M satisfies
KC(d 0 ), and M satisfies KC(d 0 ) only if d 0 Odd. Consequently, M0 does not
satisfy 6. Contradiction. Hence, Condition 4 does not hold.
Finally, suppose Condition 5 holds. Since Condition 2 does not hold, it follows
that each I 0 M0 M satisfies C(d ) for each d Odd. Hence, M0 satisfies KC(d )
iff d Odd. Now suppose (I, M0 ) satisfies the assertion R.(C u KC)(d 1 ). Then
there exists d 1 such that I satisfies both R(d 1 , d ) and C(d ), and M0 does
not satisfy KC(d ). There are two possible cases:
(1) d Odd: in this case, M0 does not satisfy KC(d ), thus contradicting the
hypothesis. Hence, Condition 5 does not hold;
(2) d 1 Odd: in this case, since Condition 5 holds, either R(d 1 , d ) or C(d )
are not satisfied by I. Contradiction. Hence, Condition 5 does not hold.
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
197
Therefore, I does not satisfy any of the 5 conditions above, which implies that
I M, which contradicts the hypothesis that M0 M. Consequently, there
is no set of interpretations M0 such that M0 M and M0 satisfies 6, which
implies that M is a model for 6.
Then, we prove that the above defined set of interpretations M cannot be
characterized in terms of a first-order theory. Let I be the first-order interpretation such that
RI
CI
PI
DI
=
=
=
=
{(x, y) : x = d 1 y Odd}
{x : x Odd}
{(d 1 , d 1 )}
{d 1 }
198
F. M. Donini et al.
199
200
F. M. Donini et al.
201
202
F. M. Donini et al.
(P 0 , N 0 ) is consistent with 6 0 ;
ObK (P, N ) |= ObK (P 0 , N 0 );
ObK (P 0 , N 0 ) 6|= ObK (P, N );
ObK (P, N ) |= ObA (P 0 , N 0 ).
PROOF. (): Suppose (P, N ) is the partition of MA1 (6) induced by M and
satisfying the above conditions, and suppose that M is not a model for 6. Then
there exists a set M0 M such that the pair (M0 , M) satisfies 6. Now consider the KB 6 0 as in the above Condition 5, and let (P 0 , N 0 ) be the partition of MA1 (6 0 ) induced by (M0 , M). Since (M0 , M) satisfies 6 and M0 M,
from Lemma 4.8 it follows that (P 0 , N 0 ) is consistent with 6; moreover, let
AC(x) (resp. AR(x, y)) be a modal atom such that C(x) ObK (P, N ) (resp.
R(x, y) ObK (P, N )): since (M, M) |= C(x) (resp. (M, M) |= R(x, y)), it follows that (M0 , M) |= AC(x) (resp. (M, M) |= AR(x, y)), and since (P 0 , N 0 ) is
the partition of MA1 (6) induced by (M0 , M), AC(x) (resp. AR(x, y)) belongs to
P 0 . Consequently, (P 0 , N 0 ) is consistent with 6 0 , hence Condition (a) holds.
Then, since M0 M, M0 = {I : I |= ObK (P, N )}, and since by Definition 4.5
and Definition 4.3 M0 {I : I |= ObK (P 0 , N 0 )}, it follows that ObK (P, N ) |=
ObK (P 0 , N 0 ), hence Condition (b) holds.
Moreover, by hypothesis (Condition 2) M contains exactly the interpretations satisfying ObK (P, N ), hence there is at least one assertion in ObK (P, N )
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
203
which is not satisfied by (M0 , M), that is, which is not in ObK (P 0 , N 0 ): therefore,
ObK (P 0 , N 0 ) 6|= ObK (P, N ), hence Condition (c) holds.
Finally, ObK (P, N ) |= ObA (P 0 , N 0 ), since any assertion of the form
AC(x)(P 0 , N 0 ) or AR(x, y) satisfied in (M0 , M) is also satisfied in (M, M) (since
the second element of the two pairs is the same): hence, Condition (d) holds.
In conclusion, (P 0 , N 0 ) satisfies Conditions (a), (b), (c), (d), which implies
that (P, N ) does not satisfy Condition 5, thus contradicting the hypothesis.
Consequently, M is a model for 6.
(): Suppose M is a model for 6. Let (P, N ) be the partition of MA1 (6)
induced by (M, M). We prove that (P, N ) satisfies each one of the conditions
15 of the claim.
Since M satisfies 6, from Lemma 4.8 it follows that (P, N ) is consistent with
6, hence Condition 1 holds.
Now suppose M 6= {I : I |= ObK (P, N )}: since, by Definition 4.3 and
Definition 4.5, each interpretation in M satisfies ObK (P, N ), it follows that
there exists at least one interpretation J which does not belong to M and such
that J |= ObK (P, N ). It can be immediately seen that the pair (M {J }, M)
induces the same partition (P, N ), hence (M {J }, M) satisfies 6, thus contradicting the hypothesis that M is a model for 6. Therefore, M = {I : I |=
ObK (P, N )}, hence Condition 2 holds.
Moreover, since (P, N ) is the partition induced by (M, M), by Definition 4.5 it
follows that, if AC(x) N (resp. AR(x, y) N ), then ObK (P, N ) 6|= C(x)(P, N )
(resp. ObK (P, N ) 6|= R(x, y)), hence Condition 3 holds.
Now suppose that ObK (P, N ) 6|= ObA (P, N ); then, there exists an assertion AC(x) (or AR(x, y)) in P such that ObK (P, N ) 6|= C(x)(P, N ) (resp.
ObK (P, N ) 6|= R(x, y)). As shown above, M = {I : I |= ObK (P, N )}, hence
M 6|= C(x)(P, N ) (resp. M 6|= R(x, y)): on the other hand, since (P, N ) is the partition induced by (M, M), it follows that M |= C(x)(P, N ) (resp. M |= R(x, y)).
Contradiction. Therefore, ObK (P, N ) |= ObA (P, N ), hence Condition 4 holds.
Finally, suppose there exists a partition (P 0 , N 0 ) of MA1 (6 0 ) which satisfies
conditions (a), (b), (c), (d). Consider the pair (M0 , M), where M0 = {I : I |=
ObK (P 0 , N 0 )}. Since ObK (P, N ) |= ObA (P 0 , N 0 ) and (P 0 , N 0 ) is consistent with
6 0 , from Definition 4.4 (Condition 1 and Conditions 5, 6) it follows that P 0 contains the set {AC(x) : C(x) ObK (P, N )} {AR(x, y) : R(x, y) ObK (P, N )},
hence ObK (P, N ) is equivalent to ObA (P 0 , N 0 ): therefore, (M0 , M) induces the
partition (P 0 , N 0 ) of 6 0 , and since (P 0 , N 0 ) is consistent with 6 0 , it follows
that (M0 , M) satisfies 6 0 , which implies that (M0 , M) satisfies 6. Moreover,
since ObK (P, N ) |= ObK (P 0 , N 0 ) and ObK (P 0 , N 0 ) 6|= ObK (P, N ), it follows that
M0 M. Consequently, M is not a model for 6, thus contradicting the hypothesis. Therefore, Condition 5 holds, which proves the thesis.
As a corollary of the above theorem, we obtain that the models of any subjectively quantified ALCKN F KB are ALC representable. In fact, let M be a
model of the subjectively quantified ALCKN F KB 6, and let (P, N ) be the
partition of MA1 (6) induced by (M, M). By the above theorem, each such
model M is constituted by the set of interpretations satisfying the ALC KB
ObK (P, N ).
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
204
F. M. Donini et al.
205
206
F. M. Donini et al.
207
208
F. M. Donini et al.
u-rule
if hw : C u D(x)i B, C u D contains at least one occurrence of a modality, and either
h1 : C(x)i 6 B or h1 : D(x)i 6 B
then add h1 : C(x)i and h1 : D(x)i to B.
t-rule
if hw : C t D(x)i B, C t D contains at least one occurrence of a modality, and neither
h1 : C(x)i B nor h1 : D(x)i B
then add h1 : C(x)i or h1 : D(x)i to B.
M-rule-1
if hw : MC(x)i B and h1 : MC(x)i 6 B
then add h1 : MC(x)i to B.
M-rule-2
if h1 : MC(x)i B and C contains at least one occurrence of a modality,
then add hw : C(x)i to B, for each w present in the branch.
M-rule-1
if hw : MC(x)i B and h1 : MC(x)i 6 B
then add h1 : MC(x)i to B.
M-rule-2
if h1 : MC(x)i B and there exists no w such that hw : NNF (C(x))i B
then add hw0 : NNF (C(x))i to B, where w0 is new in the branch.
-rule
if hw : MR.NC(x)i B
then for each h1 : MR(x, y)i B add h1 : NC( y)i to B.
-rule
if hw : MR.NC(x)i B and there is no y such that both h1 : MR(x, y)i and
h1 : NC( y)i B
then add h1 : MR(x, z)i and h1 : NC(z)i to B, for some z such that z OB {} and
6 OB .
KR-rule
if h1 : KR(x, y)i B and h1 : AR(x, y)i 6 B,
then add h1 : AR(x, y)i to B.
trigger-rule
if B is weakly completed, KC v D 0, x OB , h1 : KD(x)i 6 B, and ObK (B) |= C(x)
then add h1 : KD(x)i to B.
mcut-rule
if h1 : C(x)i B, D M S(C), and neither h1 : D(x)i B nor h1 : NNF (D(x))i B
then add h1 : D(x)i or h1 : NNF (D(x))i to B.
Fig. 1. ALCKN F tableau rules.
209
(2) the M-rules and M-rules are analogous to the usual rules for handling
modalities in tableau methods for the logic S5 [Fitting 1983];
(3) the -rule and -rule are analogous to the rules for handling quantified
concept expressions in tableau methods for epistemic Description Logics
(see Donini et al. [1998b]). In particular, the -rule either chooses, in a
nondeterministic way, an individual among the ones already appearing in
the branch, or generates a new individual;
(4) the KR-rule is a special rule which enforces the presence of the assertion
A R(x, y) for each assertion of the form KR(x, y) appearing in the branch;
roughly speaking, such a constraint is needed due to the preference semantics of ALCKN F ;
(5) the mcut-rule is applied by generating assertions on individuals based on
the set of modal subconcepts and modal role expressions appearing in 6.
However, if the modal atoms were applied to all individuals in 1, the method
would not terminate. Moreover, termination would not be guaranteed even
if such an application were restricted to the individuals already present in
the branch, due to the possibility of creating (as a consequence of an application of the mcut-rule) a new world, which in turn (due to the presence of
a modal atom of the form MR.NC) causes the creation of a new individual
to which the mcut-rule can again be applied, thus giving rise to an infinite
branch. Consequently, the mcut-rule is applied to a certain individual only
for the modal atoms which are within the scope of the existing assertions
about such an individual, and which do not occur inside a quantifier. As
we shall see, this restriction on the application of the mcut-rule is correct
under the assumption that the KB is simple. In particular, the semantics
of ALCKN F guarantees that it is safe to ignore all the occurrences of modal
atoms that do not appear in a completed branchsuch modal atoms are
assumed to be implicitly decided according to a minimal knowledge criterion, that will be explained later;
(6) the trigger-rule is only applied when B is weakly completed, that is, no other
rule is applicable in B (see below). Such a rule is needed in order to take
into account the effect of simple epistemic inclusions: specifically, the rule
applies each simple inclusion in 6 to each individual which appears in the
tableau branch and such that the left-hand side of the inclusion holds for
that individual. Notice that this last property is verified by means of the
ALC KB ObK (B) associated with the weakly completed branch B, which is
formally defined below. As already mentioned, we consider this part of the
reasoning process (which can be done by an ALC reasoner) for granted, and
do not develop it through an auxiliary tableau.
We now define the notions of closure and completeness of a branch. To
this aim, we give the following preliminary definitions. We say that a branch
is weakly completed if no rule different from the trigger-rule is applicable
to it, and say that a branch is completed if no rule is applicable to it. Of
course, if a branch is a completed branch then it is also weakly completed.
We define the partition (PB , NB ) associated with a weakly completed branch B
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
210
F. M. Donini et al.
as follows:
PB = {MC(x) : h1 : MC(x)i B} {MR(x, y) : h1 : MR(x, y)i B}
{MR.ND(x) : hw : MR.ND(x)i B for some prefix w}
NB
211
212
F. M. Donini et al.
N i+1
PB1
213
[
=
Pi
NB1 =
i=0
Ni
i=0
214
F. M. Donini et al.
215
PROOF. Let C(x) A and let B0 be the initial branch of the tableau for 6,
that is, the initial, unexpanded set of formulas for the tableau for 6. Then,
KC(x) PB0 and, since (P, N ) is consistent with 6 and KC(x) MA1 (6),
KC(x) P . Analogously, if R(x, y) A (or KR(x, y) A), then KR(x, y) PB0
and, since (P, N ) is consistent with 6 and KR(x, y) MA1 (6), KR(x, y) P ,
and if AR(x, y) A, then AR(x, y) PB0 and, since (P, N ) is consistent with
6 and AR(x, y) MA1 (6), KR(x, y) P .
Now we prove that, given a branch B0 and an injective function f 0: OB0
O6 1 O6 such that, for each modal atom 9 MA1 (6), if 9 Pf 0 (B0 ) then
9 P and if 9 N f 0 (B0 ) then 9 N , for each application of a tableau rule to
B 0 there exists an expansion B of B0 , obtained by such a rule, and an injective
function f: OB O6 1 O6 such that, for each modal atom 9 MA1 (6),
if 9 P f (B) , then 9 P and if 9 N f (B) , then 9 N . The proof of such a
property is immediate for most rules, due to the fact that (P, N ) is consistent
with 6. The only non-trivial cases concern the KR-rule, the -rule, the -rule
and the trigger-rule.
As for the KR-rule, suppose KR(x, y) PB0 . Then, KR(x, y) P and PB =
PB0 {AR(x, y)}. Since by hypothesis, if KR(x, y) P , then AR(x, y) P , it
follows that the above property holds.
As for the -rule, suppose M1 R.M2 D(x) PB0 . Then, M1 R.M2 D(x) P
(or equivalently M1 R.M2 D(x) N ), and since (P, N ) is consistent with 6,
there exists y 1 such that both M1 R(x, y) P and M2 D( y) P . Now, if B is
an expansion of B 0 obtained by the application of such a rule, then PB = PB0
{M1 R(x, z), M2 D(z)} where z is either an individual name already appearing
in B or a new individual name i. Consequently, if y = f (t) for some t already
present in B 0 , then the expansion B of B0 that chooses z = t satisfies the above
property, while if there is no t present in B0 such that y = f (t), then, the function
f such that f (i) = y and f (x) = f 0 (x) for each x OB0 , and the expansion B
of B 0 that chooses z = i, satisfy the above property. In the same way, it follows
that the above property also holds when the -rule is applied to a modal atom
of the form M1 R.M2 D(x) PB0 .
As for the -rule, suppose M1 R.M2 D(x) PB0 . Then, M1 R.M2 D(x) P
(or equivalently M1 R.M2 D(x) N ), and since (P, N ) is consistent with
6, for each y 1 such that M1 R(x, y) P , M2 D( y) P . Now observe
that the only rules that introduce modal atoms of the form M1 R(x, y) in
a branch are the KR-rule and the -rule: therefore, as shown above, if
M1 R( f 0 (x), f 0 ( y)) Pf 0 (B0 ) , then M1 R(x, y) P . Consequently, for each modal
atom M2 D( f 0 ( y)) introduced by the application of the -rule, the modal atom
M2 D( y) belongs to P . Therefore, the expansion B obtained by B0 satisfies the
above property. In the same way, it follows that the above property also holds
when the -rule is applied to a modal atom of the form M1 R.M2 D(x) PB0 .
Finally, as for the trigger-rule, suppose B 0 is a weakly completed branch
and suppose there exists x such that ObK (B 0 ) |= KC(x). Then, since P Pf 0 (B0 ) ,
N N f 0 (B0 ) , it follows that ObK (P, N ) |= ObK ( f 0 (B 0 )), therefore ObK (P, N ) |=
C( f 0 (x)). Since (P, N ) is consistent with 6 and KC( f 0 (x)) MA1 (6), it follows
that KC( f 0 (x)) P , and by Definition 4.4, KD( f 0 (x)) P . Therefore, the expansion B obtained by B 0 through the trigger-rule satisfies the above property.
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
216
F. M. Donini et al.
217
218
F. M. Donini et al.
and is satisfied by (M0 , M). Now, consider a modal atom of the form KC(x) or
AC(x). If such a modal atom does not contain occurrences of subjectively quantified concept expressions, then it follows immediately that if such a modal atom
belongs to Pf (B) , then it is satisfied by (M0 , M), and if it belongs to N f (B) , then
it is not satisfied by (M0 , M). Then, since in MA1 (6) each quantified concept
subexpression of the form AR.ND, AR.ND, is such that in D there are no
occurrences of role expressions of the form KR, and since P and P 0 only differ
for modal atoms of the form KR(x, y), it follows that, since for each y 1
AR(x, y) belongs to P iff (M0 , M) satisfies AR(x, y) and MD( y) belongs to P
iff (M0 , M) satisfies MD( y), AR.ND(x) (respectively AR.ND(x)) belongs to
P iff (M0 , M) satisfies AR.ND(x) (respectively AR.ND(x)). And since (P, N )
agrees with f (B), it follows that, if AR.ND(x) (respectively AR.ND(x)) belongs to Pf (B) then (M0 , M) satisfies AR.ND(x) (respectively AR.ND(x)) and
if AR.ND(x) (respectively AR.ND(x)) belongs to N f (B) then (M0 , M) does
not satisfy AR.ND(x) (respectively AR.ND(x)). Moreover, since by definition of M0 , if KR(x, y) belongs to Pf (B) then (M0 , M) satisfies KR(x, y) and if
KR(x, y) belongs to N f (B) then (M0 , M) does not satisfy KR(x, y), and since
(P, N ) agrees with f (B) it follows that if KR.ND(x) (respectively KR.ND(x))
belongs to Pf (B) then (M0 , M) satisfies KR.ND(x) (respectively KR.ND(x))
and if KR.ND(x) (respectively KR.ND(x)) belongs to N f (B) then (M0 , M) does
not satisfy KR.ND(x). Therefore, if a modal atom belongs to Pf (B) then such a
modal atom is satisfied by (M0 , M), and if a modal atom belongs to N f (B) , then
such a modal atom is not satisfied by (M0 , M). Since Pf (B) contains all assertions
in A, it follows that each assertion in A is satisfied by (M0 , M). Moreover, by
definition, each interpretation in M and M0 satisfies each inclusion in T , hence
(M0 , M) satisfies T . Finally, since there are no occurrences of role expressions
of the form KR in 0, for each inclusion in 0 of the form KC v D and for each
x 1, KC(x) P iff KC(x) is satisfied by (M0 , M) KD(x) P iff KD(x) is
satisfied by (M0 , M). Since (P, N ) is consistent with 6, it follows that for each
inclusion in 0 of the form KC v D and for each x 1, if KC(x) is satisfied
by (M0 , M), then D(x) is satisfied by (M0 , M). Consequently, (M0 , M) satisfies
each inclusion in 0, which implies that (M0 , M) satisfies 6.
Now, since (M0 , M) satisfies 6 and M0 M, it follows that M is not a model
for 6, thus contradicting the hypothesis. Therefore, there is no modal atom of
the form KR(x, y) such that KR(x, y) P and KR(x, y) 6 Pf (B) . Since by hypothesis ObK (P, N ) |= ObA (P, N ), if AR(x, y) P , then KR(x, y) P , thus
KR(x, y) Pf (B) , and by the KR-rule of the tableau, AR(x, y) Pf (B) . Conversely, since (P, N ) agrees with f (B), if KR(x, y) Pf (B) , then KR(x, y) P ,
and if AR(x, y) Pf (B) then AR(x, y) P . Therefore, for each modal atom of
the form KR(x, y), KR(x, y) P iff KR(x, y) Pf (B) and for each modal
atom of the form AR(x, y), AR(x, y) P iff AR(x, y) Pf (B) . Now, let
M00 = {I : I |= ObK ( f (B))}. By the last property, in a way analogous to the
case of (M0 , M), it follows that, if a modal atom belongs to Pf (B) , then such
a modal atom is satisfied by (M00 , M), and if a modal atom belongs to N f (B) ,
then such a modal atom is not satisfied by (M00 , M). Since Pf (B) contains all
assertions in A, it follows that each assertion in A is satisfied by (M00 , M).
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
219
(P 0 , N 0 ) is consistent with 6 0 ;
ObK (P, N ) |= ObK (P 0 , N 0 );
ObK (P 0 , N 0 ) 6|= ObK (P, N );
ObK (P, N ) |= ObA (P 0 , N 0 ).
Suppose B is not a preferred branch. Then, by Definition 4.14 and by the properties shown above, it follows that there exists an open and completed branch
B 0 of the tableau for 6 00 such that the conditions (a), (b) and (c) in Definition 4.14
do not hold, which implies that the following conditions hold for f (B) and f (B0 ):
(1) ObK ( f (B)) |= ObK ( f (B 0 ));
(2) ObK ( f (B 0 )) 6|= ObK ( f (B));
(3) ObK ( f (B)) |= ObA ( f (B 0 )).
Since (P, N ) = (Pf1(B) , N f1(B) ) and, by Definition 4.15,
1
0
1
ObK ( f (B 0 ) = ObK P 1
f 0 (B 0 ) , N f 0 (B 0 ) and ObA ( f (B )) = ObA P f 0 (B 0 ) , N f 0 (B 0 ) ,
by the above conditions it follows that the partition (Pf1(B0 ) , N f1(B0 ) ) is such that
1
(1) ObK (P, N ) |= ObK (P 1
f 0 (B 0 ) , N f 0 (B 0 ) );
1
(2) ObK (P 1
f 0 (B 0 ) , N f 0 (B 0 ) ) 6|= ObK (P, N );
1
(3) ObK (P, N ) |= ObA (P 1
f 0 (B 0 ) , N f 0 (B 0 ) ).
220
F. M. Donini et al.
The above theorem states correctness of the tableau method with respect to
satisfiability of a simple ALCKN F KB. This fact and termination of the tableau
method imply the following property.
THEOREM 4.23. Let 6 be a simple ALCKN F KB. Then, the problem of establishing satisfiability of 6 is decidable.
With regards to instance checking of an epistemic assertion C(a) with respect
to a simple ALCKN F KB 6, we remark that, due to the preference semantics of
ALCKN F , the deduction theorem does not hold for this logic, as in the vast majority of nonmonotonic logics. Hence, it is not possible to directly reduce logical
implication to unsatisfiability. However, in Rosati [1998] it has been shown that
in propositional MKNF, such a reduction is possible by means of a simple transformation of the query. We now extend such a result to the case of ALCKN F . As
in the previous section, we denote as C[K/A] the concept expression obtained
from C by replacing each occurrence of K with A.
THEOREM 4.24. Let 6 = hT , Ai be an arbitrary ALCKN F KB, and let C be
an ALCK concept. Then, 6 |= C(a) iff the ALCKN F KB
6 0 = hT , A {A(C)[K/A](a)}i
is unsatisfiable.
PROOF. (): Let 6 0 be unsatisfiable, and suppose 6 6|= C(a). Then, there exists a model M for 6 such that (M, M) 6|= C(a), which implies that (M, M) |=
KC(a). Then, since (M, M) interprets each concept expression of the form
AD in the same way it interprets the concept expression KD, it follows that
(M, M) |= AC[K/A](a), and since (M, M) |= 6, (M, M) |= 6 0 . Since by hypothesis M is a model for 6, it follows that, for each M0 such that M0 M,
(M0 , M) 6|= 6, therefore, for each M0 such that M0 M, (M0 , M) 6|= 6 0 , which
implies that M is a model for 6 0 , thus contradicting the hypothesis that 6 0 is
unsatisfiable. Therefore, 6 |= C(a).
(): Let 6 |= C(a), and suppose 6 0 is satisfiable. Then, there exists a cluster M0 which is a model for 6 0 . Hence, (M0 , M0 ) |= AC[K/A](a), and since
AC[K/A](a) is a subjective A-assertion and for every cluster M00 , (M0 , M0 )
and (M00 , M0 ) interpret subjective A-assertions in the same way, it follows that
(M00 , M0 ) 6|= C(a) for every cluster M00 . Since M0 is a model for 6 0 , it follows
that there is no cluster M00 such that both M00 M0 and (M00 , M0 ) |= 6 0 . As
shown above, for every cluster M00 , (M00 , M0 ) |= AC[K/A](a), hence, for each
M00 such that M00 M0 , (M00 , M0 ) 6|= 6. Consequently, from Definition 2.2 it
follows that M0 is a model for 6, and since (M0 , M0 ) 6|= C(a), it follows that
6 6|= C(a), which contradicts the hypothesis. Therefore, 6 0 is unsatisfiable.
The above theorem allows for using our tableau method for computing instance checking of subjectively quantified assertions against simple
ALCKN F KBs.
THEOREM 4.25. Let 6 = hT , Ai be a simple ALCKN F KB, and let C be a
subjectively quantified ALCK concept. Then, 6 |= C(a) if, and only if, there is
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
221
222
F. M. Donini et al.
number of epistemic instance checking problems in 6i , and since the set of ALC
KBs representing all the models of 6 is finite, it follows that subsumption in
a simple ALCKN F KB can be reduced to a finite number of instance checking
problems in simple ALCKN F KBs.
Finally, let us remark the fact that, without quantifying-in, it is possible to
define much simpler reasoning methods for MKNF-DLs. Like in the case of subjectively quantified ALCKN F , the key property is the possibility of identifying
a finite subset of the set of modal atoms MA1 (6) of a KB 6. Furthermore, in
this case it is possible to completely decouple the monotonic and nonmonotonic aspects of reasoning, which allows for defining a method which is independent of the particular non-modal DL used in the formalization of the KB.
Specifically, in Donini et al. [1997a] a general deductive method for MKNFDLs without quantifying-in is defined, which is parametric with respect to
the DL used, since it uses as an auxiliary procedure a deduction method for
the non-modal DL adopted. Such a method also allows for proving that the
MKNF extension without quantifying-in of DLs preserves decidability of reasoning. The interested reader is also referred to Rosati [1999b] for a more general discussion of reasoning in modal first-order fragments of MKNF without
quantifying-in.
5. SUMMARY AND CONCLUSIONS
We proposed a formalism called description logics of minimal knowledge and
negation as failure (MKNF-DLs) with the aim of filling the gap existing between
description logics (DLs), which characterize the fragment of monotonic firstorder logic of frame-based systems, and implemented systems, which provide
in practice several nonmonotonic features that cannot be formally captured
by DLs.
We proved the usefulness of MKNF-DLs as a formal tool, by showing some of
the representational features of the proposed formalism: defaults, procedural
rules, integrity constraints, role and concept closure. The enhanced expressiveness of MKNF-DLs required the extension of reasoning techniques developed
for propositional MKNF to an interesting class of assertions that we have characterized as subjectively quantified. This extension, which is based on a fixed,
infinite domain of interpretation, has been achieved by presenting a correct and
terminating tableau calculus for query answering in MKNF-DLs.
It is worth noticing that the expressiveness of MKNF in representing forms of
nonmonotonic reasoning can be exploited by extending frame-based systems
formalized through MKNF-DLswith other nonmonotonic reasoning features,
as well as forms of abductive reasoning, due to the known correspondences
between abduction and autoepistemic logic (see e.g. Kakas and Mancarella
[1990]; Niemela [1993]).
More interesting, however, is the fact that MKNF has proved to be adequate
for characterizing the features of both frame-based systems and of logic programming languages. In fact, it is well-known that the logic MKNF is able to
formally reconstruct logic programs with negation as failure, interpreted under
the stable model (or answer set) semantics [Lifschitz 1991]. This indicates that
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
223
the same formal tools used in the logical reconstruction of logic programs and
deductive databases (i.e., the notions of minimal knowledge and autoepistemic
belief) can also be used to capture not simply defaults, but several nonmonotonic features of frame-based systems that do not have a counterpart in logic
programming. Hence, our proposal for MKNF-DLs shows that MKNF provides
a common ground for giving a precise semantics of two different families of KR
systems: logic programming/deductive database systems, and systems based
on structured representation of information.
However, Datalog (and its extensionssee e.g., Eiter et al. [1997]) and subjectively quantified ALCKN F , represent two rather different frameworks, since
they are obtained from first-order MKNF by imposing different kinds of restrictions. Roughly speaking, the restrictions imposed by Datalog are the form
of function-free, definite Horn clauses and the safeness of the clauseswhich
leads to the interpretation of a Datalog program over a finite interpretation domain (the Herbrand universe of the program, namely the constants appearing
in the program). On the other hand, MKNF-DLs impose a variable-free syntax (as every DL), which is constrained by the language constructs provided in
the DL, not admitting n-ary relationships, but are interpreted on an infinite
interpretation domain. Such differences make it difficult to provide a precise
comparison of the expressiveness of the two formalisms.
Languages integrating description logics and Datalog have already been proposed: notably, AL-log [Donini et al. 1998a] and CARIN [Levy and Rousset
1998]. However, both proposals keep the extensional, open-world semantics of
DLs and the fixpoint, closed-world semantics of Datalog separated. A more thorough integration of frame-like structures and logic programming has been proposed in F-Logic [Kifer et al. 1995]. F-Logic is a general-purpose logic programming language, with a frame-like language for the types of objects. A unified
proof-theoretic semantics is given in terms of inference rules. However, F-logic
treats the nonmonotonic aspects of frames in an admittedly procedural way
[Kifer et al. 1995, p.806]: nonmonotonic inheritance is given a fixpoint semantics, much in the spirit of overriding mechanisms of object-oriented languages,
and negation in clauses is given a perfect model semantics.
Instead, MKNF-DLs open the possibility to combine logic programming and
frame-based languages in one expressive, decidable fragment of first-order
MKNF, with a single, model-theoretic semantics. This approach becomes feasible after recent proposals that overcome the restriction imposed by classical
DLs of dealing with unary and binary predicates only [Calvanese et al. 1988].
ACKNOWLEDGMENTS
224
F. M. Donini et al.
BAADER, F. AND HOLLUNDER, B. 1992. Embedding defaults into terminological knowledge representation formalisms. In Proceedings of the 3rd International Conference on the Principles of Knowledge
Representation and Reasoning (KR92). Morgan Kaufmann, Los Altos, 306317.
BAADER, F. AND HOLLUNDER, B. 1995. Embedding defaults into terminological knowledge representation formalisms. J. Auto. Reason. 14, 149180.
BERGAMASCHI, S. AND SARTORI, C. 1992. On taxonomic reasoning in conceptual design. ACM Trans.
Data. Syst. 17, 3, 385422.
BORGIDA, A. 1995. Description logics for data management. IEEE Trans. Knowl. Data Eng. 7, 5,
671682.
BRACHMAN, R. J. 1992. Reducing CLASSIC to practice: Knowledge representation meets reality.
In Proceedings of the 3rd International Conference on the Principles of Knowledge Representation
and Reasoning (KR92). Morgan Kaufmann, Los Altos, 247258.
BRACHMAN, R. J. AND LEVESQUE, H. J., Eds. 1985. Readings in Knowledge Representation. Morgan
Kaufmann, Los Altos.
BUCHHEIT, M., DONINI, F. M., AND SCHAERF, A. 1993. Decidable reasoning in terminological knowledge representation systems. J. Artif. Intell. Res. 1, 109138.
CALVANESE, D., DE GIACOMO, G., LENZERINI, M., NARDI, D., AND ROSATI, R. 1998. Description logic
framework for information integration. In Proceedings of the 6th International Conference on
Principles of Knowledge Representation and Reasoning (KR98). 213.
CALVANESE, D., LENZERINI, M., AND NARDI, D. 1994. A unified framework for class based representation formalisms. In Proceedings of the 4th International Conference on the Principles of
Knowledge Representation and Reasoning (KR94), J. Doyle, E. Sandewall, and P. Torasso, Eds.
Morgan Kaufmann, Los Altos, Bonn (Germany), 109120.
DONINI, F. M., LENZERINI, M., NARDI, D., NUTT, W., AND SCHAERF, A. 1992. Adding epistemic operators to concept languages. In Proceedings of the 3rd International Conference on the Principles of Knowledge Representation and Reasoning (KR92). Morgan Kaufmann, Los Altos,
342353.
DONINI, F. M., LENZERINI, M., NARDI, D., AND SCHAERF, A. 1996. Reasoning in description logics. In
Principles of Knowledge Representation, G. Brewka, Ed. Studies in Logic, Language and Information. CSLI Publications, 193238.
DONINI, F. M., LENZERINI, M., NARDI, D., AND SCHAERF, A. 1998a. AL-log: integrating datalog and
description logics. J. Intell. Inform. Syst. 10, 227252.
DONINI, F. M., LENZERINI, M., NARDI, D., SCHAERF, A., AND NUTT, W. 1998b. An epistemic operator
for description logics. Artificial Intelligence 100, 12, 225274.
DONINI, F. M., NARDI, D., AND ROSATI, R. 1995. Non-first-order features in concept languages. In
Proceedings of the 4th Conference of the Italian Association for Artificial Intelligence (AI*IA95).
Number 992 in Lecture Notes in Artificial Intelligence. Springer-Verlag, 91102.
DONINI, F. M., NARDI, D., AND ROSATI, R. 1997a. Autoepistemic description logics. In Proceedings
of the 15th International Joint Conference on Artificial Intelligence (IJCAI97). 136141.
DONINI, F. M., NARDI, D., AND ROSATI, R. 1997b. Ground nonmonotonic modal logics. J. Logic
Comput. 7, 4 (Aug.), 523548.
EITER, T., GOTTLOB, G., AND MANNILLA, H. 1997. Disjunctive Datalog. ACM Trans. Data. Syst. 22, 3,
364418.
ETHERINGTON, D. AND REITER, R. 1983. On inheritance networks with exceptions. In Proceedings
of the 3th National Conference on Artificial Intelligence (AAAI83). 104108.
FIKES, R. AND KEHLER, T. 1985. The role of frame-based representation in reasoning. Commun.
ACM 28, 9, 904920.
FITTING, M. 1983. Proof Methods for Modal and Intuitionistic Logics. Reidel.
GOTTLOB, G. 1992. Complexity results for nonmonotonic logics. J. Logic Comput. 2, 397425.
GOTTLOB, G. 1995. NP trees and Carnaps modal logic. J. ACM 42, 2, 421457.
HAYES, P. J. 1979. The logic of frames. In Frame Conceptions and Text Understanding, D. Metzing,
Ed. Walter de Gruyter and Co., 4661. Republished in [Brachman and Levesque 1985].
KAKAS, A. AND MANCARELLA, P. 1990. Generalized stable models: a semantics for abduction. In
Proceedings of the 9th European Conference on Artificial Intelligence (ECAI90). 385391.
KARP, P. D. 1992. The design space of knowledge representation systems. Tech. Rep. SRI AI
Technical Note 520, SRI International, Menlo Park (CA, USA).
ACM Transactions on Computational Logic, Vol. 3, No. 2, April 2002.
225
KIFER, M., LAUSEN, G., AND WU, J. 1995. Logical foundations of Object-Oriented and frame-based
languages. J. ACM 42, 4, 741843.
LAKEMEYER, G. 1996. Limited reasoning in first-order knowledge bases with full introspection.
Artificial Intelligence 84, 209255.
LEVESQUE, H. J. 1984. Foundations of a functional approach to knowledge representation. Artificial Intelligence 23, 155212.
LEVESQUE, H. J. 1990. All I know: a study in autoepistemic logic. Artificial Intelligence 42, 263
310.
LEVY, A. Y. AND ROUSSET, M.-C. 1998. Verification of knowledge bases based on containment checking. Artificial Intelligence 101(12), 227250.
LIFSCHITZ, V. 1991. Nonmonotonic databases and epistemic queries. In Proceedings of the 12th
International Joint Conference on Artificial Intelligence (IJCAI91). 381386.
LIFSCHITZ, V. 1994. Minimal belief and negation as failure. Artificial Intelligence 70, 5372.
MACGREGOR, R. 1988. A deductive pattern matcher. In Proceedings of the 7th National Conference
on Artificial Intelligence (AAAI88). 403408.
, M. 1993. Nonmonotonic LogicsContext-Dependent Reasoning.
MAREK, W. AND TRUSZCZYNSKI
Springer-Verlag.
NADO, R. AND FIKES, R. 1987. Semantically sound inheritance for a formally defined frame language with defaults. In Proceedings of the 6th National Conference on Artificial Intelligence
(AAAI87). 443448.
NEBEL, B. 1990. Reasoning and Revision in Hybrid Representation Systems. Lecture Notes in
Artificial Intelligence, vol. 422. Springer-Verlag.
NIEMELA , I. 1992. On the decidability and complexity of autoepistemic reasoning. Fundamenta
Informaticae 17, 1,2, 117156.
NIEMELA , I. 1993. Autoepistemic logic as a unified basis for nonmonotonic reasoning. Ph.D.
dissertation Department of Computer Science, Helsinki University of Technology, Finland.
PADGHAM, L. AND ZHANG, T. 1993. A terminological logic with defaults. In Proceedings of the 13th
International Joint Conference on Artificial Intelligence (IJCAI93). 662668.
PATEL-SCHNEIDER, P. F., MCGUINESS, D. L., BRACHMAN, R. J., RESNICK, L. A., AND BORGIDA, A. 1991.
The CLASSIC knowledge representation system: Guiding principles and implementation rational. SIGART Bull. 2, 3, 108113.
QUANTZ, J. AND ROYER, V. 1992. A preference semantics for defaults in terminological logics. In
Proceedings of the 3rd International Conference on the Principles of Knowledge Representation
and Reasoning (KR92). Morgan Kaufmann, Los Altos, 294305.
REITER, R. 1990. What should a database know? J. Logic Program. 14, 127153.
REITER, R. AND CRISCUOLO, G. 1981. On interacting defaults. In Proceedings of the 7th International
Joint Conference on Artificial Intelligence (IJCAI81). 270276.
ROSATI, R. 1998. Reducing query answering to satisfiability in nonmonotonic logics. In Proceedings of the 15th National Conference on Artificial Intelligence (AAAI98). AAAI Press / The MIT
Press, 853858.
ROSATI, R. 1999a. Reasoning about minimal belief and negation as failure. J. Artif. Intell. Res. 11,
277300.
ROSATI, R. 1999b. Towards first-order nonmonotonic reasoning. In Proceedings of the 5th International Conference on Logic Programming and Nonmonotonic Reasoning (LPNMR99). Lecture
Notes in Artificial Intelligence, vol. 1730. Springer-Verlag, 332346.
SCHMIDT-SCHAUSS, M. AND SMOLKA, G. 1991. Attributive concept descriptions with complements.
Artificial Intelligence 48, 1, 126.
STRACCIA, U. 1993. Default inheritance reasoning in hybrid KL-ONE-style logics. In Proceedings of the 13th International Joint Conference on Artificial Intelligence (IJCAI93). Morgan
Kaufmann, Los Altos, Chambery (France), 676681.
WOODS, W. A. AND SCHMOLZE, J. G. 1992. The KL-ONE family. In Semantic Networks in Artificial Intelligence, F. W. Lehmann, Ed. Pergamon Press, 133178. Published as a special issue of
Computers & Mathematics with Applications, Volume 23, Number 29.
Received June 1999; revised October 2000; accepted January 2001