Selectional preferences of Korean verbal items

초록

The present study computes the selectional preference the verbal items show in relation to their co-occurring subjects and objects in Korean. The selectional preference indicates how significant is the semantic corelation between a verbal item and the class of nouns that appear in its argument position. The selectional preference measurements presented in the current work is automatically measured in a bottom-up way, using two types of language resources: One is a parsed corpus (the Sejong Korean Treebank), and the other is lexical hierarchies or wordnets. The three Korean wordnets that are made use of are CoreNet, KorLex, and U-WIN. The acquisition model is grounded upon the Lowest Common Subsumer that represents the closest common ancestor node for the given two nodes within the hierarchy. As a way to address the issue of exponential growth in calculation time and others, this study makes use of a programming technique called hill climbing. The selectional preferences, specifically Selectional Preference Strengths and Selectional Associations proposed in Resnik (1996), are defined by Kullback-Leibler Divergence, and their values are derived based on the collection of LCSs. The results are evaluated with reference to the Sejong Electronic Dictionary, with human evaluation, with parsing evaluation using a stochastic parser, and finally in a qualitative manner.

키워드