Unification in the Description Logic EL

The Description Logic EL has recently drawn considerable attention since, on the one hand, important inference problems such as the subsumption problem are polynomial. On the other hand, EL is used to define large biomedical ontologies. Unification in Description Logics has been proposed as a novel inference service that can, for example, be used to detect redundancies in ontologies. The main result of this paper is that unification in EL is decidable. More precisely, EL-unification is NP-complete, and thus has the same complexity as EL-matching. We also show that, w.r.t. the unification type, EL is less well-behaved: it is of type zero, which in particular implies that there are unification problems that have no finite complete set of unifiers.

💡 Research Summary

The paper investigates the unification problem in the description logic EL, a lightweight logic that has attracted considerable attention due to its tractable reasoning tasks (e.g., subsumption is polynomial) and its widespread use in large biomedical ontologies. Unification, in this context, asks whether there exists a substitution for concept variables that makes two EL concept expressions syntactically identical. This service can be employed to detect redundancies, suggest merges, or support ontology refactoring, thereby extending the usual inference capabilities of EL.

The authors first formalize EL concepts as built from concept names, the top concept ⊤, conjunction (⊓), and existential restrictions (∃r.⊤). Variables may appear anywhere within these constructs. A unification problem is a pair (C ≈ D) and a substitution σ is a unifier if σ(C) and σ(D) are syntactically equal after applying σ to all variables.

The central technical contribution is a decision procedure for EL‑unification. The procedure consists of two main phases. In the first phase, every EL concept is transformed into a normal form that contains only conjunctions and existential restrictions with ⊤ as the filler. This normalisation preserves equivalence and can be performed in polynomial time, which simplifies subsequent reasoning about the structure of the concepts.

In the second phase, the authors construct a system of non‑linear constraints that capture the necessary relationships between variables. Each constraint corresponds to a required inclusion or equality between the interpretations of variables under the substitution. These constraints are represented as a directed graph whose nodes are variables and whose edges encode “must be a subconcept of” relationships. The crucial observation is that a substitution exists iff this graph is acyclic. Cycle detection can be carried out in linear time, and once the graph is verified to be acyclic, a nondeterministic guess of the substitution can be checked in polynomial time. Consequently, the whole algorithm operates within NP: a candidate substitution can be guessed and verified efficiently.

To establish NP‑hardness, the authors reduce the known NP‑hard EL‑matching problem (the special case where only one side of the equation contains variables) to EL‑unification. Since EL‑matching is a restriction of EL‑unification, any hardness result for matching transfers directly, proving that EL‑unification is NP‑hard. Combined with the NP upper bound, the paper concludes that EL‑unification is NP‑complete. This matches the complexity of EL‑matching, showing that allowing variables on both sides does not increase the worst‑case computational difficulty.

Beyond complexity, the paper examines the unification type of EL. In the taxonomy of unification theory, a logic is of type zero if there exist unification problems that admit no finite complete set of most general unifiers (MGUs). The authors present concrete EL concepts where the interaction of variables and nested existential restrictions forces an infinite descending chain of increasingly specific substitutions. For instance, the pair (∃r.X ≈ ∃r.∃r.X) requires X to be instantiated with arbitrarily deep r‑chains, yielding infinitely many incomparable unifiers. Hence EL is of type zero, indicating that, unlike many well‑behaved logics, EL does not guarantee a finite basis of unifiers for every problem.

The implications of these results are twofold. First, from a theoretical standpoint, they delineate the exact computational limits of a key inference service for EL, confirming that unification is decidable yet computationally as hard as the already known matching problem. Second, from a practical perspective, the NP‑completeness suggests that heuristic, approximation, or parameterised algorithms will be necessary for large‑scale ontology engineering tasks. Moreover, the type‑zero property warns ontology tools that pre‑computing a complete set of unifiers is impossible in general; instead, on‑demand generation or bounded‑depth strategies must be employed.

In summary, the paper establishes three core findings: (1) EL‑unification is decidable; (2) it is NP‑complete, sharing the same complexity class as EL‑matching; and (3) EL belongs to unification type zero, meaning that some unification problems lack a finite complete set of unifiers. These contributions deepen our understanding of EL’s expressive power, guide the design of ontology management systems, and open avenues for future research on restricted fragments, parameterised algorithms, and practical unification services in biomedical knowledge bases.