Computing Horn Rewritings of Description Logics Ontologies

1 Introduction

Reasoning over ontology-enriched datasets is a key requirement in many applications of semantic technologies. Standard reasoning tasks are, however, of high worst-case complexity. Satisfiability checking is 2NEXPTIME-complete for the description logic (DL) SROIQ underpinning the standard ontology language OWL 2 and NEXPTIME-complete for SHOIN, which underpins OWL DL [Kazakov, 2008]. Reasoning is also co-NP-hard with respect to data complexity-a key measure of complexity for applications involving large amounts of instance data [Hustadt et al., 2005].

Tractability in data complexity is typically associated with Horn DLs, where ontologies correspond to first-order Horn clauses [Ortiz et al., 2011; Hustadt et al., 2005]. The more favourable computational properties of Horn DLs make them a natural choice for data-intensive applications, but they also come at the expense of a loss in expressive power. In particular, Horn DLs cannot capture disjunctive axioms, i.e., statements such as "every X is either a Y or a Z". Disjunctive axioms are common in real-world ontologies, like the NCI Thesaurus or the ontologies underpinning the European Bioinformatics Institute (EBI) linked data platform. 1

1 http://www.ebi.ac.uk/rdf/platform

In this paper we are interested in Horn rewritability of description logic ontologies; that is, whether an ontology O1 expressed in a DL L1 can be rewritten into an ontology O2 in a Horn DL L2 such that O1 and O2 are equisatisfiable when extended with an arbitrary dataset. Ontologies that admit such Horn rewritings are amenable to more efficient reasoning techniques that ensure tractability in data complexity.

Horn rewritability of DL ontologies is strongly related to the rewritability of Disjunctive Datalog programs into Datalog, where both the source and target languages for rewriting are function-free. Kaminski et al. [2014b] characterised Datalog rewritability of Disjunctive Datalog programs in terms of linearity: a restriction that requires each rule to contain at most one body atom that is IDB (i.e., whose predicate also occurs in head position in the program). It was shown that every linear Disjunctive Datalog program can be rewritten into plain Datalog (and vice versa) by means of program transposition-a polynomial transformation in which rules are "inverted" by shuffling all IDB atoms between head and body while at the same time replacing their predicates by auxiliary ones. Subsequently, Kaminski et al. [2014a] proposed the class of markable Disjunctive Datalog programs, where the linearity requirement is relaxed so that it applies only to a subset of "marked" atoms. Every markable program can be polynomially rewritten into Datalog by exploiting a variant of transposition where only marked atoms are affected.

Our contributions in this paper are as follows. In Section 3, we show undecidability of Horn rewritability whenever the input ontology is expressed in ALCF. This is in consonance with the related undecidability results by Bienvenu et al. [2014] and Lutz and Wolter [2012] for Datalog rewritability and non-uniform data complexity for ALCF ontologies.

In Section 4, we lift the markability condition and the transposition transformation in [Kaminski et al., 2014a] for Disjunctive Datalog to arbitrary first-order programs with function symbols. We then show that all markable first-order programs admit Horn rewritings of polynomial size. This result is rather general and has potential implications in areas such as theorem proving [Robinson and Voronkov, 2001] and knowledge compilation [Darwiche and Marquis, 2002].

The notion of markability for first-order programs can be seamlessly adapted to ontologies via the standard FOL translation of DLs [Baader et al., 2003]. This is, however, of limited practical value since Horn programs obtained via transposition may not be expressible using standard DL constructors. In Section 5, we introduce an alternative satisfiability-preserving translation from ALCHIF ontologies into first-order programs and show in Section 6 that the corresponding transposed programs can be translated back into Horn-ALCHIF ontologies. Finally, we focus on complexity and show that reasoning over markable L-ontologies is EXPTIME-complete in combined complexity and PTIME-complete w.r.t. data for each DL L between ELU and ALCHIF. All our results immediately extend to DLs with transitive roles (e.g., SHIF) by exploiting standard transitivity elimination techniques [Baader et al., 2003].

We have implemented markability checking and evaluated our techniques on a large ontology repository. Our results indicate that many real-world ontologies are markable and thus admit Horn rewritings of polynomial size.

The proofs of all our results are contained in an extended version of the paper (see arXiv:1504.05150).

2 Preliminaries

We assume standard first-order syntax and semantics. We treat the universal truth ⊤ and falsehood ⊥ symbols as well as equality (≈) as ordinary predicates of arity one (⊤ and ⊥) and two (≈), the meaning of which will be axiomatised.

Programs A first-order rule (or just a rule) is a sentence

where variables ⃗x and ⃗z are disjoint, ϕ(⃗x, ⃗z) is a conjunction of distinct atoms over ⃗x ∪ ⃗y, and ψ(⃗x) is a disjunction of distinct atoms over ⃗x. Formula ϕ is the body of r, and ψ is the head. Quantifiers are omitted for brevity, and safety is assumed (all variables in the rule occur in the body). We define the following sets of rules for a finite signature Σ: (i) P ⊤ Σ consists of a rule P(x1, . . . , xn) → ⊤(xi) for each predicate P ∈ Σ and each 1 ≤ i ≤ n and a rule → ⊤(a) for each constant a ∈ Σ; (ii) P ⊥ Σ consists of the rule having ⊥(x) in the body and an empty head; and (iii) P ≈ Σ consists of the standard axiomatisation of ≈ as a congruence over Σ. 2 A program is a finite set of rules P = P0 ∪ P ⊤ Σ ∪ P ⊥ Σ ∪ P ≈ Σ with Σ the signature of P0, where we assume w.l.o.g. that the body of each rule in P0 does not mention ⊥ or ≈, and the head is non-empty and does not mention ⊤. We omit Σ for the components of P and write P ⊤, P ⊥ and P ≈. A rule is Horn if its head consists of at most one atom, and a program is Horn if so are all of its rules. Finally, a fact is a ground, function-free atom, and a dataset is a finite set of facts.

Ontologies We assume familiarity with DLs and ontology languages [Baader et al., 2003]. A DL signature Σ consists of disjoint countable sets of concept names ΣC and role names ΣR. A role is an element of ΣR ∪ {R- | R ∈ ΣR}. The function inv is defined over roles as follows, where R ∈ ΣR: inv(R) = R- and inv(R-) = R. W.l.o.g., we consider normalised axioms as on the left-hand side of Table 1.

An ALCHIF ontology O is a finite set of DL axioms of type T1-T6 in Table 1. An ontology is Horn if it contains no axiom of type T1 satisfying m ≥ 2. Given O, we denote with

2 Reflexivity of ≈ is axiomatised by the safe rule ⊤(x) → x ≈ x.

⊑∗ the minimal reflexive and transitive relation over roles in O such that R1 ⊑∗ R2 and inv(R1) ⊑∗ inv(R2) hold whenever R1 ⊑ R2 is an axiom in O.

We refer to the DL where only axioms of type T1-T3 are available and the use of inverse roles is disallowed as ELU. The logic ALC extends ELU with axioms T4. We then use standard naming conventions for DLs based on the presence of inverse roles (I), axioms T5 (H) and axioms T6 (F). Finally, an ontology is EL if it is both ELU and Horn.

Table 1 also provides the standard translation π from normalised axioms into first-order rules, where at(R, x, y) is defined as R(x, y) if R is named and as S(y, x) if R = S-. We define π(O) as the smallest program containing π(α) for each axiom α in O. Given a dataset D, we say that O ∪ D is satisfiable iff so is π(O) ∪ D in first-order logic.

3 Horn Rewritability

Our focus is on satisfiability-preserving rewritings. Standard reasoning tasks in description logics are reducible to unsatisfiability checking [Baader et al., 2003], which makes our results practically relevant. We start by formulating our notion of rewriting in general terms.

Definition 1. Let F and F ′ be sets of rules. We say that F ′ is a rewriting of F if it holds that F ∪D is satisfiable iff so is F ′ ∪ D for each dataset D over predicates from F. ⋄

We are especially interested in computing Horn rewritings of ontologies-that is, rewritings where the given ontology O1 is expressed in a DL L1 and the rewritten ontology O2 is in a Horn DL L2 (where preferably L2 ⊆ L1). This is not possible in general: satisfiability checking is co-NP-complete in data complexity even for the basic logic ELU [Krisnadhi and Lutz, 2007], whereas data complexity is tractable even for highly expressive Horn languages such as Horn-SROIQ [Ortiz et al., 2011]. Horn rewritability for DLs can be formulated as a decision problem as follows:

Definition 2. The (L1,L2)-Horn rewritability problem for DLs L1 and L2 is to decide whether a given L1-ontology admits a rewriting expressed in Horn-L2. ⋄

Our first result establishes undecidability whenever the input ontology contains at-most cardinality restrictions and thus equality. This result is in consonance with the related undecidability results by Bienvenu et al. [2014] and Lutz and Wolter [2012] for Datalog rewritability and non-uniform data complexity for ALCF ontologies.

Theorem 3. (L1,L2)-Horn rewritability is undecidable for L1 = ALCF and L2 any DL between ELU and ALCHIF. This result holds under the assumption that PTIME̸=NP.

Intractability results in data complexity rely on the ability of non-Horn DLs to encode co-NP-hard problems, such as non-3-colourability [Krisnadhi and Lutz, 2007; Hustadt et al., 2005]. In practice, however, it can be expected that ontologies do not encode such problems. Thus, our focus from now onwards will be on identifying classes of ontologies that admit (polynomial size) Horn rewritings.

T1. dni=1 Ai ⊑ ⊔mj=1 Cj ∧ni=1 Ai(x) → ∨mj=1 Cj(x) T2. ∃R.A ⊑ C at(R, x, y) ∧ A(y) → C(x) T 3. A ⊑ ∃R.B A(x) → at(R, x, f (x)); A(x) → B(f (x)) T 4. A ⊑ ∀R.C A(x) ∧ at(R, x, y) → C(y) T 5. S ⊑ R S(x, y) → at(R, x, y) T 6. A ⊑ ≤1R.B A(z) ∧ at(R, z, x1) ∧ at(R, z, x2) ∧B(x1) ∧B(x2) → x1 ≈ x2

4 ProgramMarkability and Transposition

In this section, we introduce the class of markable programs and show that every markable program can be rewritten into a Horn program by means of a polynomial transformation, which we refer to as transposition. Roughly speaking, transposition inverts the rules in a program P by moving certain atoms from head to body and vice versa while replacing their corresponding predicates with fresh ones. Markability of P ensures that we can pick a set of predicates (a marking) such that, by shuffling only atoms with a marked predicate, we obtain a Horn rewriting of P. Our results in this section generalise the results by Kaminski et al. [2014a] for Disjunctive Datalog to first-order programs with function symbols.

To illustrate our definitions throughout this section, we use an example program Pex consisting of the following rules:

Markability. The notion of markability involves a partitioning of the program's predicates into Horn and disjunctive. Intuitively, the former are those whose extension for all datasets depends only on the Horn rules in the program, whereas the latter are those whose extension may depend on a disjunctive rule. This intuition can be formalised using the standard notion of a dependency graph in Logic Programming.

Definition 4. The dependency graph GP = (V,E, µ) of a program P is the smallest edge-labeled digraph such that: (i) V contains all predicates in P; (ii) r ∈ µ(P,Q) whenever r ∈ P, P is in the body of r, and Q is in the head of r; and (iii) (P,Q) ∈ E whenever µ(P,Q) ̸= ∅. A predicate Q depends on r ∈ P if GP has a path ending in Q and involving an r-labeled edge. Predicate Q is Horn if it depends only on Horn rules; otherwise, Q is disjunctive. ⋄

For instance, predicates C, D, and ⊥ are disjunctive in our example program Pex, whereas A and B are Horn. We can now introduce the notion of a marking-a subset of the disjunctive predicates in a program P ensuring that the transposition of P where only marked atoms are shuffled between head and body results in a Horn program.

Definition 5. A marking of a program P is a set M of disjunctive predicates in P satisfying the following properties, where we say that an atom is marked if its predicate is in M: (i) each rule in P has at most one marked body atom; (ii) each rule in P has at most one unmarked head atom; and (iii) if Q ∈ M and P is reachable from Q in GP, then P ∈ M. We say that a program is markable if it admits a marking. ⋄

Condition (i) in Def. 5 ensures that at most one atom is moved from body to head during transposition. Condition

(ii) ensures that all but possibly one head atom are moved to the body. Finally, condition (iii) requires that all predicates depending on a marked predicate are also marked. We can observe that our example program Pex admits two markings: M1 = {C, ⊥} and M2 = {C, D, ⊥}.

Markability can be efficiently checked via a 2-SAT reduction, where we assign to each predicate Q in P a propositional variable XQ and encode the constraints in Def. 5 as 2-clauses. For each rule ϕ ∧ ∧ni=1 Pi(⃗si) → ∨mj=1 Qj(⃗tj), with ϕ the conjunction of all Horn atoms in the rule head, we include clauses (i) ¬XPi ∨ ¬XPj for all 1 ≤ i < j ≤ n, which enforce at most one body atom to be marked; (ii) XQi ∨ XQj for 1 ≤ i < j ≤ m, which ensure that at most one head atom is unmarked; and (iii) ¬XPi ∨XQj for 1 ≤ i ≤ n and 1 ≤ j ≤ m, which close markings under rule dependencies. Each model of the resulting clauses yields a marking of P.

Transposition. Before defining transposition, we illustrate the main intuitions using program Pex and marking M1.

The first step to transpose Pex is to introduce fresh unary predicates C and ⊥, which stand for the negation of the marked predicates C and ⊥. To capture the intended meaning of these predicates, we introduce rules X(x) → ⊥(x) for X ∈ {A,B,C,D} and a rule ⊥(x) → ⊥(f(x)) for the unique function symbol f in Pex. The first rules mimick the usual axiomatisation of ⊤ and ensure that an atom ⊥(c) holds in a Herbrand model of the transposed program whenever X(c) also holds. The last rule ensures that ⊥ holds for all terms in the Herbrand universe of the transposed programan additional requirement that is consistent with the intended meaning of ⊥, and critical to the completeness of transposition in the presence of function symbols. Finally, a rule ⊥(z)∧C(x)∧C(x) →⊥(z) ensures that the fresh predicate C behaves like the negation of C (⊥(z) is added for safety).

The key step of transposition is to invert the rules involving the marked predicates by shuffling marked atoms between head and body while replacing their predicate with the corresponding fresh one. In this way, rule B(x) → C(x) ∨D(x) yields B(x) ∧ C(x) → D(x), and C(x) → ⊥(x) yields ⊥(x) →C(x). Additionally, rule D(x) →C(f(x)) is transposed as ⊥(z)∧D(x)∧C(f (x)) → ⊥(z) to ensure safety. Finally, transposition does not affect rules containing only Horn predicates, e.g., rule A(x) →B(x) is included unchanged.

Definition 6. Let M be a marking of a program P. For each disjunctive predicate P in P, let P be a fresh predicate of the same arity. The M-transposition of P is the smallest program ΞM(P) containing every rule in P involving only Horn predicates and all rules 1-6 given next, where ϕ is the conjunction of all Horn atoms in a rule, ϕ⊤ is the least conjunction of ⊥- atoms making a rule safe and all Pi, Qj are disjunctive:

1. ϕ⊤∧ϕ∧∧mj=1 Qj(⃗tj)∧∧ni=1 P i(⃗si) → Q(⃗t) for each rule in P of the form ϕ ∧ Q(⃗t) ∧ ∧mj=1 Qj (⃗tj ) → ∨ni=1 Pi(⃗si) where Q(⃗t) is the only marked body atom;
2. ⊥(x) ∧ ϕ ∧ ∧mj=1 Qj (⃗tj ) ∧ ∧ni=1 P i(⃗si) → ⊥(x), where x a fresh variable, for each rule in P of the form ϕ ∧ ∧mj=1 Qj (⃗tj ) → ∨ni=1 Pi(⃗si), with no marked body atoms and no unmarked head atoms;
3. ϕ∧ ∧mj=1 Qj(⃗tj) ∧ ∧ni=1 Pi(⃗si) → P(⃗s) for each rule in P of the form ϕ ∧ ∧mj=1 Qj(⃗tj) → P(⃗s) ∨ ∨ni=1 Pi(⃗si) where P(⃗s) is the only unmarked head atom;
4. ⊥(z) ∧ P (⃗x) ∧ P (⃗x) → ⊥(z) for marked predicate P ;
5. P(x1, . . . , xn) → ⊥(xi) for each P in P and 1 ≤ i ≤ n;
6. ⊥(x1) ∧ . . . ∧⊥(xn) → ⊥(f(x1, . . . , xn)) for each n-ary function symbol f in P. ⋄

Note that rules of type 1 in Def. 6 satisfy {P1, . . . , Pn} ⊆ M since Q ∈ M, while for rules of type 3 we have {Q1, . . . , Qm} ∩M = ∅ since P /∈ M.

Clearly, Pex is unsatisfiable when extended with fact A(a). To see that ΞM1 (Pex)∪{A(a)} is also unsatisfiable, note that B(a) is derived by the unchanged rule A(x) → B(x). Fact C(a) is derived using A(x) → ⊥(x) and the transposed rule ⊥(x) → C(x). We derive D(a) using B(x)∧C(x) → D(x). But then, to derive a contradiction we need to apply rule ⊥(z)∧D(x)∧C(f (x)) → ⊥(z), which is not possible unless we derive C(f(a)). For this, we first use ⊥(x) → ⊥(f(x)), which ensures that ⊥ holds for f (a), and then ⊥(x) → C(x).

Transposition yields quadratically many Horn rules. The following theorem establishes its correctness.

Theorem 7. Let M be a marking of a program P. Then ΞM(P) is a polynomial-size Horn rewriting of P.

It follows that every markable set of non-Horn clauses N can be polynomially transformed into a set of Horn clauses N ′ such that N ∪D and N ′ ∪D are equisatisfiable for every set of facts D. This result is rather general and has potential applications in first-order theorem proving, as well as in knowledge compilation, where Horn clauses are especially relevant [Darwiche and Marquis, 2002; Del Val, 2005].

5 Markability of DLOntologies

The notion of markability is applicable to first-order programs and hence can be seamlessly adapted to ontologies via the standard translation π in Table 1. This, however, would be of limited value since the Horn programs resulting from transposition may not be expressible in Horn-ALCHIF.

Consider any ontology with an axiom ∃R.A ⊑ B and any marking M involving R. Rule R(x, y)∧A(y) → B(x) stemming from π would be transposed as B(x)∧A(y) → R(x, y), which cannot be captured in ALCHIF. 3

To address this limitation we introduce an alternative translation ξ from DL axioms into rules, which we illustrate using the example ontology Oex in Table 2. The key idea is to encode existential restrictions in axioms T3 as unary atoms over functional terms. For instance, axiom α2 in Oex would yield B(x) → D(fR,D(x)), where the "successor" relation between an instance b of B and some instance of D in a Herbrand model is encoded as a term fR,D(b), instead of a binary atom of the form R(b, g(b)). This encoding has an immediate impact on markings: by marking B we are only forced to also mark D (rather than both R and D). In this way, we will be able to ensure that markings consist of unary predicates only.

3 Capturing such a rule would require a DL that can express products of concepts [Rudolph et al., 2008].

To compensate for the lack of binary atoms involving functional terms in Herbrand models, we introduce new rules when translating axioms T2, T4, and T6 using ξ. For instance, ξ(α3) yields the following rules in addition to π(α3): a rule D(fR,D(x)) → D(x) to ensure that all objects c with an Rsuccessor fR,D(c) generated by ξ(α2) are instances of D; a rule D(fR,B(x)) → D(x), which makes sure that an object whose R-successor generated by ξ(α4) is an instance of D is also an instance of D. Finally, axioms α1 and α5, which involve no binary predicates, are translated as usual.

Definition 8. Let O be an ontology. For each concept ∃R.B in an axiom of type T3, let fR,B be a unary function symbol, and Φ the set of all such symbols. We define ξ(O) as the smallest program containing π(α) for each axiom α in O of type T1-T2 and T4-T6, as well as the following rules:

• A(x) → B(fR,B(x)) for each axiom T3;
• A(fR′ ,Y (x)) → C(x) for each axiom T2 and R′ and Y s.t. fR′ ,Y ∈ Φ and R′ ⊑∗ R.
• A(finv(R′ ),Y (x)) → C(x) for each axiom T4 and R′ and Y s.t. finv(R′ ),Y ∈ Φ and R′ ⊑∗ R.
• A(x) ∧ Y (finv(R′ ),Y (x)) → C(finv(R′ ),Y (x)) for each axiom T2 and R′ and Y s.t. finv(R′ ),Y ∈ Φ and R′ ⊑∗ R.
• A(x) ∧ Y (fR′ ,Y (x)) → C(fR′ ,Y (x)) for each axiom T4 and R′ and Y s.t. fR′ ,Y ∈ Φ and R′ ⊑∗ R.
• A(z)∧B(fR′ ,Y (z))∧at(R, z, x)∧B(x) → fR′ ,Y (z) ≈ x for each ax. T6 and R′ , Y s.t. fR′ ,Y ∈ Φ and R′ ⊑∗ R.
• A(finv(R′ ),Y (x))∧B(x)∧at(R, finv(R′ ),Y (x), y)∧B(y) → x ≈ y for each axiom T6 and R′ and Y s.t. finv(R′ ),Y ∈ Φ and R′ ⊑∗ R.
• A(z) ∧ B(fR′1 ,Y1 (z)) ∧ B(fR′2 ,Y2 (z)) → fR′1 ,Y1 (z) ≈ fR′2 ,Y2 (z) for each axiom T6 and fR′i ,Yi ∈ Φ s.t. R′i ⊑∗ R.
• A(finv(R′1 ),Y1 (x)) ∧ B(x) ∧ B(fR′2 ,Y2 (finv(R′1 ),Y1 (x))) → x ≈ fR′2 ,Y2 (finv(R′1 ),Y1 (x)) for each axiom T6 and each R′i and Yi s.t. {finv(R′1 ),Y1 , fR′2 ,Y2 } ⊆ Φ and R′i ⊑∗ R. ⋄

Note that, in contrast to the standard translation π, which introduces at most two rules per DL axiom, ξ can introduce linearly many rules in the size of the role hierarchy induced by axioms of type T5.

The translation ξ(Oex) of our example ontology Oex is given in the second column of Table 2. Clearly, Oex is unsatisfiable when extended with A(a) and E(a). We can check that ξ(Oex) ∪ {A(a), E(a)} is also unsatisfiable. The following theorem establishes the correctness of ξ.

Theorem 9. For every ontology O and dataset D over predicates in O we have that O∪D is satisfiable iff so is ξ(O)∪D.

This translation has a clear benefit for markability checking: in contrast to π(O), binary predicates in ξ(O) do not belong to any minimal marking. In particular, Mex = {B, D, ⊥} is the only minimal marking of ξ(Oex).

Proposition 10. (i) If ≈ is Horn in ξ(O) then so are all binary predicates in ξ(O). (ii) If ξ(O) is markable, then it has a marking containing only unary predicates.

Thus, we define markability of ontologies in terms of ξ rather than in terms of π. We can check that π(Oex) is not markable, whereas ξ(Oex) admits the marking Mex.

We conclude this section with the observation that markability of an ontology O can be efficiently checked by first computing the program ξ(O) and then exploiting the 2-SAT encoding sketched in Section 4.

6 RewritingMarkable Ontologies

It follows from the correctness of transposition in Theorem 7 and ξ in Theorem 9 that every ALCHIF ontology O admitting a marking M has a Horn rewriting of polynomial size given as the program ΞM(ξ(O)). In what follows, we show that this rewriting can be expressed within Horn-ALCHIF.

Let us consider the transposition of ξ(Oex) via the marking Mex, which is given in the third column of Table 2. The transposition of α1 and α5 corresponds directly to DL axioms via the standard translation in Table 1. In contrast, the transposition of all other axioms leads to rules that have no direct correspondence in DLs. The following lemma establishes that the latter rules are restricted to the types T7-T20 specified on the left-hand side of Table 3.

Lemma 12. Let O be an ontology andM a minimal marking of ξ(O). Then ΞM(ξ(O)) contains only Horn rules of type T1-T2 and T4-T6 in Table 1 as well as type T7-T20 in Table 3.

We can now specify a transformation Ψ that allows us to translate rules T7-T20 in Table 3 back into DL axioms.

Definition 13. We define Ψ as the transformation mapping (i) each Horn rule r of types T1-T2 and T4-T6 in Table 1 to the DL axiom π-1 (r) (ii) each rule T7-T20 on the left-hand side of Table 3 to the DL axioms on the right-hand side. 4 ⋄

4 For succinctness, axioms resulting from T7, T8, T12, T13, T14, T16 and T18 are not given in normal form.

Intuitively, Ψworks as follows: (i) Function-free rules are "rolled up" as usual into DL axioms (see e.g., T7). (ii) Unary atoms A(fR,Y (x)) with A ̸= ⊥ involving a functional term are translated as either existentially or universally quantified concepts depending on whether they occur in the body or in the head (e.g., T10, T11); in contrast, atoms ⊥(fR,Y (x)) in rules ⊥(x) → ⊥(fR,Y (x)) are translated as ∃RY .⊥, instead of ∀RY .⊥ (see T9). (iii) Rules T15-T18, which involve ≈ in the head and roles R′ and R in the body, are rolled back into axioms of type T6 over the "union" of R and R′ , which is captured using fresh roles and role inclusions.

The ontology obtained by applying Ψ to our running example is given in the last column of Table 2. Correctness of Ψ and its implications for the computation of Horn rewritings are summarised in the following lemma.

Lemma 14. Let O be a markable ALCHIF ontology and let M be a marking of O. Then the ontology Ψ(ΞM(ξ(O))) is a Horn rewriting of O.

A closer look at our transformations reveals that our rewritings do not introduce constructs such as inverse roles and cardinality restrictions if these were not already present in the input ontology. In contrast, fresh role inclusions may originate from cardinality restrictions in the input ontology. As a result, our approach is language-preserving: if the input O1 is in a DL L1 between ALC and ALCHI, then its rewriting O2 stays in the Horn fragment of L1; furthermore, if L1 is between ALCF and ALCIF, then O2 may contain fresh role inclusions (H). A notable exception is when O1 is an ELU ontology, in which case axioms T2 and T3 in O1 may yield axioms of type T4 in O2. The following theorem follows from these observations and Lemma 14.

Theorem 15. Let L be a DL between ALC and ALCHI. Then every markable L ontology is polynomially rewritable into a Horn-L ontology. If L is between ALCF and ALCHIF, then every markable L ontology is polynomially rewritable into Horn-LH. Finally, every markable ELU ontology is polynomially rewritable into Horn-ALC.

7 Complexity Results

We next establish the complexity of satisfiability checking over markable ontologies.

We first show that satisfiability checking over markable ELU ontologies is EXPTIME-hard. This implies that it is not possible to polynomially rewrite every markable ELU ontology into EL. Consequently, our rewriting approach is optimal for ELU in the sense that introducing universal restrictions (or equivalently inverse roles) in the rewriting is unavoidable.

Lemma 16. Satisfiability checking over markable ELU ontologies is EXPTIME-hard.

All Horn DLs from ALC to ALCHIF are EXPTIMEcomplete in combined complexity and PTIME-complete in data complexity [Kr¨otzsch et al., 2013]. By Theorem 15, the same result holds for markable ontologies in DLs from ALC to ALCHIF. Finally, Lemma 16 shows that these complexity results also extend to markable ELU ontologies.

Theorem 17. Let L be in-between ELU and ALCHIF. Satisfiability checking over markable L-ontologies is EXPTIMEcomplete and PTIME-complete w.r.t. data.

8 RelatedWork

Horn logics are common target languages for knowledge compilation [Darwiche and Marquis, 2002]. Selman and Kautz [1996] proposed an algorithm for compiling a set of propositional clauses into a set of Horn clauses s.t. their Horn consequences coincide. This approach was generalised to FOL by Del Val [2005], without termination guarantees.

Bienvenu et al. [2014] showed undecidability of Datalog rewritability for ALCF and decidability in NEXPTIME for SHI. Cuenca Grau et al. [2013] and Kaminski et al. [2014a]

proposed practical techniques for computing Datalog rewritings of SHI ontologies based on a two-step process. First, O is rewritten using a resolution calculus Ω into a Disjunctive Datalog program Ω(O) of exponential size [Hustadt et al., 2007]. Second, Ω(O) is rewritten into a Datalog program P. For the second step, Kaminski et al. [2014a] propose the notion of markability of a Disjunctive Datalog program and show that P can be polynomially computed from Ω(O) using transposition whenever Ω(O) is markable. In contrast to our work, Kaminski et al. [2014a] focus on Datalog as target language for rewriting (rather than Horn DLs). Furthermore, their Datalog rewritings may be exponential w.r.t. the input ontology and cannot generally be represented in DLs.

Gottlob et al. [2012] showed tractability in data complexity of fact entailment for the class of first-order rules with singleatom bodies, which is sufficient to capture most DLs in the DL-Litebool family [Artale et al., 2009].

Lutz and Wolter [2012] investigated (non-uniform) data complexity of query answering w.r.t. fixed ontologies. They studied the boundary of PTIME and co-NP-hardness and established a connection with constraint satisfaction problems. Finally, Lutz et al. [2011] studied model-theoretic rewritability of ontologies in a DL L1 into a fragment L2 of L1. These rewritings preserve models rather than just satisfiability, which severely restricts the class of rewritable ontologies; in particular, only ontologies that are "semantically Horn" can be rewritten. For instance, O = {A ⊑ B ⊔ C}, which is rewritable by our approach, is not Horn-rewritable according to Lutz et al. [2011].

9 Proof of Concept

To assess the practical implications of our results, we have evaluated whether real-world ontologies are markable (and hence also polynomially Horn rewritable). We analysed 120 non-Horn ontologies extracted from the Protege Ontology Library, BioPortal (http://bioportal.bioontology.org/), the corpus by Gardiner et al. [2006], and the EBI linked data platform (http://www.ebi.ac.uk/rdf/platform). To check markability, we have implemented the 2-SAT reduction in Section 4 and a simple 2-SAT solver.

We found that a total of 32 ontologies were markable and thus rewritable into a Horn ontology, including some ontologies commonly used in applications, such as ChEMBL (see http://www.ebi.ac.uk/rdf/services/chembl/) and BioPAX Reactome (http://www.ebi.ac.uk/rdf/services/reactome/). When using π as first-order logic translation, we obtained 30 markable ontologies-a strict subset of the ontologies markable using ξ. However, only 27 ontologies were rewritable to a Horn DL since in three cases the marking contained a role.

10 Conclusion and FutureWork

We have presented the first practical technique for rewriting non-Horn ontologies into a Horn DL. Our rewritings are polynomial, and our experiments suggest that they are applicable to widely-used ontologies. We anticipate several directions for future work. First, we would like to conduct an extensive evaluation to assess whether the use of our rewritings can significantly speed up satisfiability checking in practice. Second, we will investigate relaxations of markability that would allow us to capture a wider range of ontologies.

Acknowledgments

Work supported by the Royal Society, the EPSRC projects Score!, MaSI3 and DBOnto, and the FP7 project Optique.