Optimization Strategies for Instance Retrieval.
Volker Haarslev, Ralf Möller · 2002
In this paper new techniques for optimizing instance retrieval in DLsyMxL) are described. The algorithms are evaluated with application examples from a natural language processing application. 1 Motivation Many applications of description logic (DL)sy)2Tk use the instance retrieval inference problem [5] in order to properly formalize subtasks. For instance, in [1, 2] acase-study with the application of DL inference services in a natural language (NL) interpretation syerp is presented. In particular, the inference retrieval service of the Racersy2M) [4] is investigated for various application-specific subtasks (e.g., resolution of referring expressions, content determination, and content realization). In this application many ABoxes are generated on thefly (see [1, 2] for details) and for each ABox a specific instance retrievalquery is computed. In order to achieve good performance in the NL application, the performance of the instance retrieval procedure provided by the DLsyz(] is crucial. Furthermore, since ABoxes change quite frequently , standard techniques for optimizing instance retrieval using indexing techniques (see below for an explanation) canhardly be employ ed in order to improve performance because of the overhead of computing index structures in beforehand. In this paper new techniques for optimizing instance retrieval in DLsyMx)) are described. The algorithms are evaluated with application examples from the natural language processing application described above (Racer 1-6-2, 1GHz Pentium). The TBox consists of 165 possibly cyibl "definitions" and "primitive definitions" for concepts as well as domain and range restrictions for 18 roles. In the ABox around 250 individuals are mentioned in concept and role assertions. The DL used in the knowledge base is...