A Feature Space Expression to Analyze Dependency of Korean Clauses with a Composite Kernel

Abstract

Analyzing of dependency relation among clauses is one of the most critical parts in parsing Korean sentences because it generates severe ambiguities. To get successful results of analyzing dependency relation, this task has been the target of various machine learning methods including SVM. Especially, kernel methods are usually used to analyze dependency relation and it is reported that they show high performance. This paper proposes an expression for dependency analysis of Korean clauses. The proposed expression adopts a composite kernel to obtain the similarity among clauses. The composite kernel consists of a parse tree kernel and a liner kernel. A parse tree kernel is used for treating structure information and a liner kernel is applied for using lexical information. The proposed expression is defined as three types. One is a expression of layers in clause, another is relation expression between clause and the other is an expression of inner clause. The expriment is processed by two steps that first is a relation expression between clauses and the second is a expression of inner clauses. The experimental results show that the proposed expression achieves 82.12% of accuracy.