Inference Fusion: A Hybrid Approach to Taxonomic Reasoning

We present a hybrid way to extend taxonomic reasoning using inference fusion, i.e. the dynamic combination of inferences from distributed heterogeneous reasoners. Our approach integrates results from a DL-based taxonomic reasoner with results from a constraint solver. Inference fusion is carried out by (i) parsing heterogeneous input knowledge, producing suitable homogeneous subset of the input knowledge for each specialised reasoner; (ii) processing the homogeneous knowledge, collecting the reasoning results and passing them to the other reasoner if appropriate; (iii) combining the results of the two reasoners. We discuss the benefits of our approach to the ontological reasoning and demonstrate our ideas by proposing a hybrid modelling languages,DL(D)/S, and illustrating its use by means of examples. Motivation and background Current approaches to ontology reasoning during the knowledge lifecycle management are based on a wide variety of structured knowledge models, each enabling different automated capabilities. Different from the Object-oriented and Frame-based approaches, models based on Description Logics (DLs) like OIL/DAML+OIL (Fensel et al. 2001) are equipped with a whole set of specialised deductions based on taxonomic reasoning. Such deductive services include, among others, semantic consistency check and contradiction detection, explicitation of hidden knowledge, subsumption, and concept classification (Donini et al. 1996). Therefore, DL-based approaches are particularly appealing for applications such as ontology reasoning in the Semantic Web, where taxonomic reasoning has been recognised as one of the core inferences (Fensel et al. 2001). Moreover, DLs use the notions of concept (i.e. unary predicate) and role (i.e. binary relation) to model declarative knowledge in a structured way. Using different constructors defined with a uniform syntax and unambiguous semantics, complex concept definitions and axioms can be built from simple components. Therefore, DLs are particularly appealing both to represent ontological knowledge and to reason with it. Unfortunately, because the expressive power needed to model complex real-world ontology is quite high, ontology reasoning was initially ruled out of the list of services to Copyright c © 2003, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. be provided by ontology management tools (Fikes & Farquhar 1999). Nevertheless, it as been re-introduced by the OIL/DAML+OIL effort as a first-class issue, providing a solution within the framework of a DL-based, frame-centred approach (Fensel et al. 2001). However, despite its expressivity, the OIL/DAML+OIL approach does not yet provide practical support to reasoning with concrete domains or local constraints (i.e. role-value maps). This is because the knowledge model of the iFaCT DL engine (Horrocks 1999), which provides the deductive services for the ontology inference layer, does not currently include concrete domains or role-value maps. Some approaches have been proposed to include concrete domains in DL-based concept definitions which are normally restricted to abstract domains. Despite the diversity of their representations, most of them have based on ALC (Schmidt-Schauß & Smolka 1991) or its expressive successor SHIQ (Horrocks, Sattler, & Tobies 2000). They concentrated on extending the original tableau-based algorithm (Schmidt-Schauß & Smolka 1991), i.e. create a tableaux containing both concept constructors and constraint predicates, during which process, the complex intervention of abstract and concrete knowledge is inevitable. It has been proved that adding concrete domains (e.g. numeric constraints) directly to expressive DL-based systems may result in undecidable inferential problems (Lutz 2001). The dilemma faced by DL-community brings up a new question: although single-purposed reasoning systems have improved substantially, their homogeneous approaches are limited in two ways: (i) the expressive power of their representation is restricted in order to ensure computational tractability, completeness and decidability; (ii) the specialist nature of their reasoning means that they are only successful at carrying out particular inferential tasks. We believe that if a knowledge model is too expensive to be analysed by a reasoner (e.g. DLs) alone, other representation and reasoning paradigms must be jointly used. Therefore, it’s reasonable to consider that a hybrid approach to heterogeneous knowledge management may provide, among other things, a wider and better support to ontology reasoning. In this paper, we thus present a generic hybrid schema to extend existing DL-based systems with the ability of representing and reasoning with numeric constraints. Our