Induction and revision of terminologies
Floriana Esposito, Nicola Fanizzi, Luigi Iannone, Ignazio Palmisano, Giovanni Maria Semeraro · CINECA IRIS Institutional Research Information System (University of Bari Aldo Moro) · 2004
Description Logics (DLs) and derived markup languages are a standard for representing ontological knowledge bases which can be a powerful tool for supporting other services, such as reasoning and retrieval. Such languages are generally endowed with well-founded semantics and reasoning services investigated in the DLs field [1]. In this context, we examine the problem of the induction and refinement of definitions for the concepts and their properties forming the T-box, on the ground of basic information (assertions) made on individuals that may be available in a knowledge base, i.e. the A-box, representing the world state. The induction and refinement of structural knowledge is not new in Machine Learning. Recently, in Inductive Logic Programming, attempts have been made to extend the classic relational learning techniques toward hybrid representations [5]. In order to cope with the problem complexity, former methods are based on a heuristic search and generally implement bottom-up algorithms, such as the least common subsumer (lcs) [3], that tend to induce overly specific definitions which may suffer for poor predictiveness (overfitting). Hence, in some approaches maximal generalizations are preferred [4]. Moreover, like the least general generalizations for clausal representations, lcs’s sizes tend to increase exponentially. Other approaches have shown that also a top-down search is feasible [2], yet the refinement is less operational: it is intended to show the properties of the related search space rather than specifying how to exploit the heuristics based on the assertions in the A-box. Intending to give operational instruments for performing the inference and refinement of conceptual descriptions, we introduce an algorithm based on multilevel counterfactuals [7] for operating with an ALC representation.