Modeling generalization and specialization with Extended Conceptual Graphs

Abstract

AbstractThe final goal of our research is to show that the performance of statistical rule induction can be improved by augmenting training data with semantic information. In order to prove this hypothesis, a statistical grammar induction system is to be created the knowledge base of which is represented by Extended Conceptual Graphs (ECGs). Since generalization and specialization are the basic operations of induction, they are of great significance in machine learning. As a consequence, the paper aims at investigating the least common generalization and the greatest common specialization of two ECG graphs. These operations should be traced back to the examination of ECG graph element instances. For this reason, a domain-specific ECG element instance type lattice (T′,≺) has been generated for the given test environment. Our final conclusion is that the least common generalization and the greatest common specialization of two ECG graphs always exist and can be computed. Therefore, the definition of the ≺ relation on element instances can be extended to a partial relation ⪯ on ECG diagram graphs, according to which F 1 ⪯ F 2 if graph Γ 1 is more specialized than Γ 2.