The revival of structural subsumption in

tableau-based description logic reasoners

V. Haarslev1 , R. Möller2 , S. Wandelt2

1
Concordia University, Montreal, Canada
2
Hamburg University of Technology, Hamburg, Germany?

Abstract. The paper summarizes our experiences with optimization

techniques for well-known tableau-based description logic reasoning sys-
tems, and analyzes the performance of very simple techniques to cope
with Tboxes whose bulk axioms just use a less expressive language such
as ELH, whereas some small parts of the Tbox use a language as expres-
sive as SHIQ. The techniques analyzed in this paper have been tested
with RacerPro, but they can be embedded into other tableau-based rea-
soners such as, e.g., Fact++ or Pellet in a seamless way.

1 Introduction
In practical applications involving ontologies, it should be the goal to formulate
important tasks as reasoning problems, and existing description logic inference
systems should be used to actually solve these problems. Due to our experiences,
for specific tasks of this kind, very often expressive description logics such as ALC
or SHIQ are used. For large parts of the application, however, a description logic
such as EL or maybe ELH might be sufficient. For practical applications it is a
realistic assumption that description logic reasoning systems should be tailored
towards Tboxes consisting of very large sets of (acyclic) concept definitions with
concept descriptions from, for instance, the ELH language on the one hand,
and some smaller set of axioms involving ALC or SHIQ concept descriptions
on the other. Due to our experiences, supporting this kind of Tboxes is of ut-
most importance for the acceptance of description logic reasoners in industrial
contexts.
It has been argued in the literature [1–3] that less expressive languages have
their merits for applications using very large Tboxes because in these less expres-
sive languages, subsumption is a polynomial inference problem. Also, recently,
non-obvious results have been achieved that identify subsumption as a polyno-
mial problem even in the presence of generalized concept inclusions (GCIs) for
the description logic EL++ [4–6].
Subsumption is an important inference problem in many application con-
texts, and it is the predominant inference problem used at development time
when a Tbox is classified. Classification of large Tboxes is a well-investigated,
but still not an easy inference problem from a practical point of view. Exploiting
the results in [6], for this task very promising results have been reported on algo-
rithms implemented in the CEL description logic system [7]. In this work, among
?
This paper has been partially funded by the TONES Project, FP6-7603, 6th EU
Framework Programme.
other knowledge bases, the authors consider a huge Tbox, which is a variant of
SNOMED-CT [8] with approximately 380,000 concept names for which, to a
large extent, corresponding concept definitions exist. If an application is to be
built that uses a Tbox of this size, it might be the case that additional concept
definitions using concept descriptions from ALC or even SHIQ should be added
in order to solve certain application tasks using description logic reasoning. How-
ever, then, description logic inference systems tailored to the EL language (and
its extension EL++ ) can no longer be used for classification purposes at the
current state of the art. The problem is that one often cannot have the cake
(description logics systems ensuring fast classification) and eat it too (i.e., use
the expressive power of description logics for solving application problems).
Due to our experiences, for instance, value restrictions are indeed used in
applications, and are supported by W3C syntaxes for description logic languages
(such as OWL Lite or OWL DL). Thus, following our line of argumentation,
in these kinds of applications, SHIQ reasoners such as FaCT++, Pellet, or
RacerPro seem to be appropriate. Experiments have indicated that Tboxes of
the SNOMED family are hard for these reasoners [5].
In this paper we summarize our experiences with optimization techniques
for well-known tableau-based reasoning systems, and analyze the performance
of very simple techniques to cope with Tboxes whose bulk axioms just use a less
expressive language such as ELH, whereas some small parts of the Tbox use
a language as expressive or more expressive as ALC. The techniques analyzed
in this paper have been tested with RacerPro, but they can be embedded into
other tableau-based reasoners such as, e.g., Fact++ or Pellet in a seamless way.
The paper is structured as follows. In Section 3 we first give a deeper intro-
duction to the problems of tableau-based reasoning techniques in the context
of classifying ELH. Then, in Section 4, a technique is specified that fits well
into tableau-based reasoning systems, and, due to our experiments in Section 5,
can solve the classification problem for important example Tboxes of the class
considered in this paper. Section 6 concludes the paper and analyses the pros
and cons of the approach.

2 Syntax and semantics of ALC and ELH

For a given application problem one chooses a set of elementary descriptions

for concepts and roles representing unary and binary predicates, respectively.
Elementary descriptions are also called atomic descriptions, or just names for
brevity.
In the following, we use letters A and R for atomic concept and role de-
scriptions, respectively. In ALC (Attributive Language with full Complement),
descriptions for complex concepts C or D can be inductively built using the
following grammar:
We introduce the concept descriptions > and ⊥ as abbreviations for A t ¬A
and A u ¬A, respectively. Concept descriptions may be written in parentheses in
order to avoid scoping ambiguities. A concept description C which is contained
in a concept description D is called a subconcept (of D).
For defining the semantics of concept and role descriptions we consider in-
terpretations I that consist of a non-empty set ∆I , the domain, and an inter-
 C, D −→ A | atomic concept description
C uD | conjunction
C tD | disjunction
¬C | negation
∃R.C | existential restriction
∀R.C | value restriction

pretation function ·I , which assigns to every atomic concept description A a set

AI ⊆ ∆I and to every (atomic) role R a set RI ⊆ ∆I ×∆I . For complex concept
descriptions the interpretation function is extended as follows:

(C u D)I = C I ∩ DI
(C t D)I = C I ∪ DI
(¬C)I = ∆I \C I
(∃R.C)I = {x | ∃y.(x, y) ∈ RI and y ∈ C I }
(∀R.C)I = {x | ∀y. if (x, y) ∈ RI then y ∈ C I }

The semantics of description logics is based on the notion of satisfiability. An

interpretation I = (∆I , ·I ) satisfies a concept description C if C I 6= ∅. In this
case, I is called a model for C.
A Tbox is a set of so-called generalized concept inclusions C v D . A concept
.
definition A = C in a Tbox T is an abbreviation of A v C and C v A, given
there is no other GCI A v D in T and A is not referred to in C, either directly
or indirectly via other GCIs.
An interpretation I satisfies a GCI C v D if C I ⊆ DI . An interpretation
is a model of a Tbox if it satisfies all GCIs in the TBox. A concept description
C is subsumed by a concept description D w.r.t. a Tbox if the GCI C v D is
satisfied in all models of the Tbox. In this case, we also say that D subsumes C,
or D is a subsumer of C.
The set of parents of an atomic concept is the set of concept names which
are most-specific subsumers w.r.t. a Tbox T . The set of children is defined
analogously. The classification problem for T is the problem of computing the
parents for every concept name occurring in T . The resulting lattice is called
taxonomy.

2.1 Decision problems and their reductions

The definitions given in the previous section can be paraphrased as decision

problems. The concept satisfiability problem is to check whether a model for a
concept description exists. The Tbox satisfiability problem is to check whether
there exists a model for the Tbox. The concept subsumption problem is to check
whether C v D holds in all models of the Tbox. The classification problem
is to compute the taxonomy, and is reduced to a quadratic number of concept
subsumption problems in the worst-case.
2.2 The description logics EL, ELH, and others

Given the definition of ALC in the previous section, we introduce description

logic EL by imposing restrictions on the set of possible complex concept descrip-
tions. Only conjunction and existential restrictions are allowed in EL. In ELH
it is possible to additionally specify a set of role inclusions of the form R v S
where R and S are atomic role descriptions. The set of superroles of a role w.r.t.
a set of role inclusions is defined in the obvious way.
A role inclusion is satisfied if there exists an interpretation I such that RI ⊆
I
S . In addition to GCIs, role inclusions can appear in Tboxes in ELH. Since EL
(and ELH) do not support negation, concept descriptions are always satisfiable
(also with respect to Tboxes).
For a definition of the description logic SHIQ see [9]. EL++ is defined in [4].

3 The problem, an observation and a hypothesis

The problem with contemporary SHIQ classification techniques is that, in the

worst case, subsumption tests for the form C v D with respect to a Tbox T are
translated to a test of the form ¬SATT (C u ¬D), i.e., subsumption is reduced
to non-satisfiability. The introduction of negation might introduce disjunctions.
Since concept descriptions can be deeply nested and usually introduce many
conjuncts, due to the negation, many disjuncts are the result. In tableau-based
reasoners, a technique called lazy evaluation [10] is used, with the effect that
even more disjunction appear implicitly when concept names are “replaced”
with their definition (see [10] for details).
During Tbox classification very many subsumption tests have to be made,
and concerning practical Tboxes it is well-known that most of these tests return
false, i.e, C u ¬D is indeed satisfiable. With the underlying tableau-based prover
technology, the search for a model in the space of disjunctions is slow, however.
The subsumption problem formulated in this way is solved with an exponential
algorithm due to the introduction of negation. It is not the case that in the
experiments discussed below for a specific subsumption test a combinatorial ex-
plosion occurs, causing the test to run hours. But, although a single subsumption
test does not matter much for the whole classification process, at the current
state of the art the tableau prover machinery just seems to be too heavy-weight,
and the enormous amount of subsumption tests provides for a situation in which
classifications times are not tolerable.
An initial experiment with RacerPro (see also [5]) indicated that the above-
mentioned brute-force approach needed days to (correctly) classify a variant of
SNOMED-CT that uses the language ELH [11]. We call this version SNOMED-
ELH. Note that RacerPro implements standard optimization techniques for clas-
sification as described in [12, 13] as well as some others [14]. In addition, as
analyzed by Tsarkov et al. [15], in Tboxes of this kind for quite a substantial
number of concept names the axioms directly indicate the correct place in the
taxonomy (i.e., the parents and children can be identified using static analysis).
The technique is known as the identification of “completely defined concepts”
(CD optimization, [15]). Our experiments with an implementation of this tech-
nique in RacerPro also indicate the effectiveness of this method (see below).
However, still we were not satisfied with classification runtimes (a couple of
hours for SNOMED-ELH).
A backbone optimization technique for classifying Tboxes using tableau-
based reasoners is called “pseudo model merging” [13]. Every concept name
CN satisfiable w.r.t. a Tbox T is associated with a so-called pseudo model data
structure M = (L, ¬L, S ∃ , S ∀ ) where L is a set of positive literals (concept
names), ¬L is a set of negative literals, S ∃ is a set of concept descriptions of
the form ∃R.C, and S ∀ is a set of concept descriptions of the form ∀R.C. The
data structure is computed in the obvious way from the label of the root node
in a tableau test to implement SATT for a concept description C. The idea of
the algorithm is to show that C and ¬D do neither “interact” in terms of posi-
tive and negative literals nor “interact” w.r.t. existential restrictions and value
restrictions (see [13] and Section 4 for details). If there is no interaction, the
conjunction C u ¬D is satisfiable, and C is not subsumed by D.
Although pseudo model merging can be seen as a cheap trick, it is surpris-
ingly effective for a variety of expressive description logics [16]. However, our
tests indicate that this backbone optimization technique for classifying Tboxes
is not effective at all for SNOMED-ELH. Indeed, the technique just introduces
overhead if used in the standard way. Since subsumption tests introduce dis-
junctions (see the arguments above), in many cases, the “wrong” pseudo models
are tried in the merging process (for details of the merging process see [13, 16]),
and the computational efforts do not lead to any results. On the contrary, the
heavy-weight tableau machinery is still needed.
The observation that realistic knowledge bases in use today use inexpressive
languages for most parts of the Tboxes led us to the insight that one might use
the pseudo models computed anyhow in tableau-based reasoners in a different
way. Instead of “merging” the pseudo models computed for C and ¬D (in the
sense of model merging) one could try to “embed” the pseudo model of D into the
pseudo model of C. This technique has been used for many years in description
logic systems and is known as structural subsumption. Deep pseudo models (with
pseudo models computed also for the qualifications in existential restrictions) as
introduced in [16] can be seen as so-called description graphs [3]. An algorithm
that checks whether the description graphs can be embedded can be used to
decide subsumption in case certain conditions are met (for details see below).
There exists an algorithm for this problem that runs in polynomial time (and
it works bottom-up). A simple graph-embedding algorithm, which is the basis of
previous structural subsumption algorithms (back to KL-ONE), works top-down.
It takes two deep pseudo models (or description graphs) as input and checks for
graph embeddings with respect to existential restrictions. In the worst-case it is
exponential, however.
Static analysis has to reveal in which cases the structural subsumption al-
gorithm can be applied, and in which cases the tableau-based techniques are
really required to be sound and complete. It was our research hypothesis that
for Tboxes such as SNOMED-ELH augmented with few SHIQ axioms an “ac-
ceptable” performance for classification can be achieved.
This paper describes our experiments and the results we obtained w.r.t.
whether the hypothesis can be verified or not. To our surprise, structural sub-
sumption, i.e., a prestidigitation, implemented with an algorithm that is expo-
nential in the worst case, can dramatically speed up classification times while
retaining soundness and completeness. SNOMED-ELH with 380,000 concept
names and augmented with some SHIQ axioms can be classified in 12 minutes
on a standard laptop computer.

4 Structural subsumption as model embedding

Model merging is a standard optimization technique for description logic classi-

fiers, and we define the procedure here for the sake of completeness by considering
the case for the description logic ALC. The approach can easily be extended to
more expressive description logics such as SHIQ.
The idea of merging pseudo models (pmodels) for a subsumption test C v D
is to evaluate pmodels mergable(pmodel(C), pmodel(¬D)) whose definition is
given in Algorithm 1.

Algorithm 1 pmodels mergable?((L1 , ¬L1 , S1∃ , S1∀ ), (L2 , ¬L2 , S2∃ , S2∀ ))
return (L1 ∩ ¬L2 = ∅) ∧ (¬L1 ∩ L2 = ∅)
∧ not exists (∃R.C) ∈ S1∃ such that
there exists (∀S.D) ∈ S2∀ with S ∈ ancestors(R)
∧ not exists (∃R.C) ∈ S2∃ such that
there exists (∀S.D) ∈ S1∀ with S ∈ ancestors(R)

The function ancestors applied to a role R returns the set of superroles of R

including R.
Since the pseudo models are computed anyway during classification in a
standard tableau-based classification algorithm, the idea is to exploit these data
structures also in a different way, namely in a structural subsumption test for
the language ELH.
Every ELH concept C is associated with a pseudo model data structure M =
(L, {}, S ∃ , {}), i.e., there are no negated literals and no value restrictions. We
assume that the data structure pmodel(C) being associated with C is computed
on demand (with respect to a given Tbox). As has been mentioned before, an
ELH concept is always satisfiable.
If now during classification it is to be checked whether A1 subsumes A2 this
can be done with a call to Algorithm 2 in the form of

pmodel embeddable?(pmodel(A1 ), pmodel(A2 ))

Algorithm 2 pmodel embeddable?((L1 , {}, S1∃ , {}), (L2 , {}, S2∃ , {}))
return f ilter(L1 ) ⊆ L2
∧ for all (∃R.C) ∈ S1∃
there exists (∃S.D) ∈ S2∃ such that
R ∈ ancestors(S)
∧ pmodel embeddable?(pmodel(C), pmodel(D))
Note that we do not claim that the idea behind Algorithm 2 is new. The al-
gorithm is presented here in order to discuss details in the context of pseudo
models.
The function f ilter is used for the following purpose. Given a Tbox
. .
{A = ∃R.C, A1 v B u ∃R.C, A2 = B u A}

it is obvious that A1 is subsumed by A2 , and it holds that3

pmodel(A1 ) = ({A1 , B}, {}, {∃R.C}, {})

pmodel(A2 ) = ({A2 , B, A}, {}, {∃R.C}, {}).

However, the pseudo model of A2 cannot be embedded into the pseudo model
of A1 due to the defined concepts A2 and A in the literal list. The function f ilter
eliminates the concept names for which there exists a concept definition (and
keeps those for which there exists no definitions or only GCIs with the name on
the left-hand side).
The information encoded in the pseudo model is usually exploited for the
model merging process. For the implementation of model embedding, however,
defined concepts such as A1 should be transparent. The function f ilter used
in Algorithm 2 eliminates all concept names from L1 for which there exists a
concept definition and keeps the names for which there exists a GCI with the
name on the left-hand side.
It can be easily seen that Algorithm 2 can exhibit exponential behavior in
the worst case. Our conjecture was that for realistic knowledge bases, the com-
binatorial behavior just does not occur because, usually, there are only very few
existential restrictions for a particular role R used in a concept definition. In
order to verify this hypothesis we wanted to conduct several experiments with
SNOMED-ELH. The goal was however, not to build a dedicated ELH prover
but to integrate this technique into a standard SHIQ prover such as Racer-
Pro. The above-mentioned model embedding technique should hardly introduce
any overhead since pseudo models are computed anyway for the standard model
merging optimization techniques (see above). Nevertheless, a static analysis is
required in order to identify when the model embedding technique is actually
applicable for two concepts A1 , A2 to be checked for subsumption.
The applicability conditions of pmodel embeddable? are defined as specified
in Algorithm 3.

Algorithm 3 pmodel embedding applicable?(A1 , A2 , T )

return ¬meta constraints?(T )
∧ elh concept?(A1 , T ) ∧ elh concept?(A2 , T )
∧ ¬self referencing?(A1 , T ) ∧ ¬self referencing?(A2 , T )

The function meta constraint? checks if GCI absorption was successful, i.e.,
the function tests whether there are GCIS left after GCI absorption. The function
3
We assume that taxonomic encoding [13] is not used.
elh concept? exploits results from a static analysis of the axioms in the Tbox.
For a concept name A it is determined whether all subconcepts “reachable” via
concept definitions or GCIs are ELH concepts. The function self referencing?
determines whether a concept name A is cyclic with respect to the Tbox axioms.
In Algorithm 4 the subsumption test is defined. The function shiq subsumes?
applies the standard subsumption test with optimization techniques as described
in [12–14].

Algorithm 4 subsumes?(A1 , A2 , T )
if pmodel embedding applicable?(A1 , A2 , T ) then
return pmodel embeddable?(pmodel(A1 ), pmodel(A2 ))
else
return shiq subsumes?(A1 , A2 , T )

5 Evaluation
We conducted several experiments to verify the hypothesis that Algorithm 2,
which is worst-case exponential, dramatically speeds up classification times for
SNOMED-ELH even in the case where there are some GCIs that use ALC
concept descriptions.
Table 1 shows the results that we obtained for classifying SNOMED-ELH
with approximately 380,000 concept names using RacerPro 1.9.2 beta (Intel
2.4GHz, Core 2 Duo, Mac OS X, Leopard, 64bit).

CD optimization pseudo model embedding ELH ? runtime

disabled disabled yes several days
enabled disabled yes several hours
disabled enabled yes minutes (15 min)
enabled enabled yes minutes (12 min)
Table 1. Effects of optimization techniques.

The results shown in Table 1 reveal that the CD optimization technique

(see Section 3) has a dramatic effect (line 2) but model embedding has an even
larger effect (line 3). Using both techniques in combination yields a slightly
better performance (line 4).
We have also run CEL on SNOMED-ELH (Intel Pentium IV 2.8 GHz, Linux,
32bit). RacerPro and CEL need approximately the same time (half an hour on
the Intel P4 machine under Linux).
RacerPro can still be used if we “spoil” the Tbox with an ALC axiom of the
form A1 v ∀R.A2 for arbitrary SNOMED-ELH concept names A1 and A2 and
role names R. Further experiments have been carried out to investigate in which
situations the performance will degrade. Figure 1 indicates that the effect is
Fig. 1. A SNOMED classification test (Intel Pentium IV 2.8 GHz, Linux, 32bit) with
value restrictions added to the definition of 50 randomly selected parents of ⊥. An
increasing fragment of the Tbox is used (10%, 20%, . . . , 100%). The line with boxes
indicates the runtimes (in minutes) without value restrictions, the line with diamonds
shows the runtimes with “spoiling” value restrictions.

Fig. 2. A SNOMED classification test (Intel Pentium IV 2.8 GHz, Linux, 32bit) with
an increasing number of value restrictions added to the definition of randomly selected
parents of ⊥. Only the first 10 % of SNOMED are used for this test (37900 concept
names).
dramatic. A few value restrictions can cause long classification times. The effect
is less dramatic if there are “few” concept names in the Tbox (see Figure 2).
Note that we are aware of the fact that RacerPro cannot even be used in cases
where CEL could be expected to still be quite fast, i.e., if the extensions that
EL++ includes were used in the Tbox. In addition, the optimization techniques
presented in this paper cannot be used in the presence of cycles. The cases in
which the techniques can indeed be used are automatically detected, however.
CEL cannot be run in the spoiled case.

6 Conclusion
The techniques investigated in this paper are attractive for tableau-based de-
scription logic systems because they are very easy to implement. Data structures
being computed anyway (pseudo models) are now also used for model embed-
ding as a kind of structural subsumption test. The proposed technique requires a
static analysis for implementing the applicability check defined in Algorithm 3.
We conjecture that the information needed in Algorithm 3 is already computed
in all optimized tableau-based reasoners existing today.
The evaluation with SNOMED-ELH shows encouraging results. Neverthe-
less, the tests reveal that even a relatively small number of value restrictions
can cause problems with large Tboxes. Nevertheless, we can have the bulk ELH
Tboxes that meet sporadic SHIQ requirements for small parts.
Furthermore, we found it surprising that an exponential algorithm, namely
a (naive) structural subsumption test, can be used to classify Tboxes such as
SNOMED-ELH. With pmodel embeddable? the “right structures” are made
available. The potential problem that there are many non-effective recursive
calls of pmodel embeddable? does not occur in practical Tboxes, and SNOMED-
ELH seems to be a good example for this effect. An implementation of a poly-
nomial algorithm for pmodel embeddable? and its evaluation in the context of
SNOMED-ELH is left for further studies.
The experiments shed some light on the real problem with tableau-based
subsumption tests. The problem is not the exponential worst-case behavior per
se. The worst-case-exponential algorithm for pmodel embeddable? works fine for
SNOMED-ELH. In our opinion, the problem is that in standard tableau reason-
ers the pseudo models are retained, but the tableau structures used for deciding
the problem subsumes?(D, C) = ¬SAT (C u ¬D) are always discarded after the
test is completed. This could be avoided if also tableau structures for C and ¬D
were retained and somehow manipulated in an effective way to get the answer.
We conjecture that it would be possible to also deal with, for instance, value re-
strictions effectively in a tableau-based algorithm if the “right structures” were
kept between multiple calls to subsumes. This is subject to further research as
well.

References
1. Brachman, R.J., Levesque, H.J.: The tractability of subsumption in frame-based
description languages. In: Proc. of the 4th Nat. Conf. on Artificial Intelligence
(AAAI’84). (1984) 34–37
 2. Borgida, A., Patel-Schneider, P.F.: A semantics and complete algorithm for sub-
sumption in the CLASSIC description logic. J. of Artificial Intelligence Research
1 (1994) 277–308
3. Baader, F.: Terminological cycles in a description logic with existential restric-
tions. In Gottlob, G., Walsh, T., eds.: Proceedings of the 18th International Joint
Conference on Artificial Intelligence, Morgan Kaufmann (2003) 325–330
4. Brandt, S.: Polynomial time reasoning in a description logic with existential re-
strictions, GCI axioms, and—what else? In de Mantáras, R.L., Saitta, L., eds.:
Proceedings of the 16th European Conference on Artificial Intelligence (ECAI-
2004), IOS Press (2004) 298–302
5. Baader, F., Lutz, C., Suntisrivaraporn, B.: Is tractable reasoning in extensions of
the description logic EL useful in practice? In: Proceedings of the Methods for
Modalities Workshop (M4M-05), Berlin, Germany (2005)
6. Baader, F., Brandt, S., Lutz, C.: Pushing the EL envelope. In: Proc. of the 19th
Int. Joint Conf. on Artificial Intelligence (IJCAI 2005). (2005) 364–369
7. Baader, F., Lutz, C., Suntisrivaraporn, B.: CEL—a polynomial-time reasoner for
life science ontologies. In Furbach, U., Shankar, N., eds.: Proceedings of the 3rd
International Joint Conference on Automated Reasoning (IJCAR’06). Volume 4130
of Lecture Notes in Artificial Intelligence., Springer-Verlag (2006) 287–291
8. Spackman, K.A., Campbell, K.E., Cote, R.A.: SNOMED RT: A reference termi-
nology for health care. J. of the American Medical Informatics Association (1997)
640–644 Fall Symposium Supplement.
9. Horrocks, I., Sattler, U., Tobies, S.: Reasoning with individuals for the description
logic SHIQ. In McAllester, D., ed.: Proc. of the 17th Int. Conf. on Automated
Deduction (CADE 2000). Volume 1831 of Lecture Notes in Computer Science.,
Springer (2000) 482–496
10. Horrocks, I., Tobies, S.: Reasoning with axioms: Theory and practice. In: Proc. of
the 7th Int. Conf. on the Principles of Knowledge Representation and Reasoning
(KR 2000). (2000) 285–296
11. Spackman, K.A.: Managing clinical terminology hierarchies using algorithmic cal-
culation of subsumption: Experience with snomed-rt. J. of the American Medical
Informatics Association (2000) Fall Symposium Special Issue.
12. Baader, F., Franconi, E., Hollunder, B., Nebel, B., Profitlich, H.J.: An empirical
analysis of optimization techniques for terminological representation systems or:
Making KRIS get a move on. Applied Artificial Intelligence. Special Issue on
Knowledge Base Management 4 (1994) 109–132
13. Horrocks, I.: Optimisation techniques for expressive description logics. Technical
Report UMCS-97-2-1, University of Manchester, Department of Computer Science
(1997)
14. Haarslev, V., Möller, R.: High performance reasoning with very large knowledge
bases: A practical case study. In: Proc. of the 17th Int. Joint Conf. on Artificial
Intelligence (IJCAI 2001). (2001) 161–168
15. Tsarkov, D., Horrocks, I., Patel-Schneider, P.F.: Optimizing terminological rea-
soning for expressive description logics. J. of Automated Reasoning 39 (2007)
277–316
16. Haarslev, V., Möller, R., Turhan, A.Y.: Exploiting pseudo models for TBox and
ABox reasoning in expressive description logics. In: Proc. of the Int. Joint Conf. on
Automated Reasoning (IJCAR 2001). Volume 2083 of Lecture Notes in Artificial
Intelligence., Springer (2001)