Calculating Selectional Preferences of Transitive Verbs in Korean

Copyright 2012 by Sanghoun Song and Jae-Woong Choe 26th Pacific Asia Conference on Language,Information and Computation pages 89–98 Calculating Selectional Preferences of Transitive Verbs in Korean Sanghoun Song Department of Linguistics University of Washington Box 354340 Seattle, WA 98195-4340, USA sanghoun@uw.edu Jae-Woong Choe Department of Linguistics Korea University 145 Anam-ro Seongbuk-gu Seoul, 136-701 Korea jchoe@korea.ac.kr Abstract This study calculates the selectional prefer- ence strength between transitive verbs and their co-occurring objects, and thereby in- vestigates how much they are co-related to each other in Korean. The selectional pref- erence strength is automatically measured in a bottom-up way, and the outcomes are eval- uated in comparison with a manually con- structed resource that indicates which verb takes which class(es) of nouns as its depen- dents. The measurement offered by this study not only can be used to improve NLP appli- cations, but also has a theoretic significance in that it can play a role as distributional evidence in the study of argument structure. 1 Introduction Selectional Preference Strength (henceforth, SPS) refers to the degree of correlation between two co- occurring linguistic items. This study, exploiting some Korean language resources and employing the Kullback-Leibler Divergence model formulated by Resnik (1996), aims to calculate SPS between tran- sitive verbs and the classes of co-occurring nouns that function as objects. As far as we know, there has been no previous study to calculate SPS in Korean. Now that sev- eral Korean resources constructed on a comprehen- sive scale are currently available, it would be very interesting to conduct a systematic analysis of SPS in Korean and to see what kind of significant pat- terns and results can be found through such analysis. This research is an endeavour in that direction, and reports some results of our analysis of SPS between predicates and their object argument, which is based on language resources like treebanks, wordnets, and electronic dictionaries. We also expect that our anal- ysis would make a meaningful contribution to our understanding of the semantic interaction between verbal items and argument structure in Korean. This paper is structured as follows. Section 2 discusses why it is necessary to look into SPS in NLP, and offers a brief explanation of the back- ground knowledge. Section 3 covers the computa- tional model that this study employs, and Section 4 measures SPS using a Korean wordnet (i.e. KorLex) and a development corpus (i.e. the Sejong Korean Treebank). The results are evaluated quantitatively as well as qualitatively in Section 5. This paper closes in Section 6 with a brief look at our further work to help NLP systems perform better. 2 Background The Korean language, as is well-known, is an agglu- tinative language with a large number of grammat- ical function morphemes. It also has features like the right-headness, scrambling, and virtually free deletion of any element from a sentence. On the more semantic side, Korean shows the usual restric- tion between a predicate and its selection of argu- ments. The sentence pair in (1) exemplifies the syn- …

Content truncated to 500 words