세종 구문분석 말뭉치를 기반으로 한 확률 문맥자유문법 규칙

Abstract

The Sejong Korean Treebank (SKT) was built as part of 10 year government-sponsored Sejong project, and more than 80 million graphic-word Korean parsed corpus has been released to the public at the end of 2007. The purpose of this paper is to extract Context-Free Grammar (CFG) rules from SKT and to draw some linguistic generalizations based on the CFG rules. We introduce an extraction algorithm that was used in this study and prove that it meets the minimal requirements as an objective extraction method in terms of its precision and recall rates. Then our discussion of the extracted CFG rules proceed in terms of the minimal tree structure containing a mother node (MN) and its two daughter nodes (Left DN, Right DN). We arrive at various linguistic or stochastic generalizations restricting the distribution of the categories in the minimal tree structure for Korean, for example, one that states 'In more than 95% of the cases that involve S, VP, NP, VNP, and AP, MN and RDN share the same category.' We provide most of the detailed statistical information regarding the basic properties of SKT and the CFG rules derived from it.