Extraction of Simple Sentences from Mixed Sentences for Building Korean Case Frames

EXTRACTION OF SIMPLE SENTENCES FROM MIXED SENTENCES FOR BUILDING KOREAN CASE FRAMES- Dan-Hee Y ang*, Ik-Hwan Lee**, Mansuk Song* Department of Computer Science, Department of English, Yonsei University, Seoul 120-749, Korea. E-mail: {dhyang, mssong}@december.yonsei.ac.kr, ihlee@bubble.yonsei.ac.kr ABSTRACT A large number of simple sentences are needed to construct practical Case frames automatically. Until now, most studies have assumed that there are already extensive training data (especially here simple sentences) and linguistic information for their work. However, this is not true at least of Korean. Furthermore, Korean syntactic structures are significantly different from those of English. So, this paper first of all, compares Korean with English in relation to extracting simple sentences from mixed ones. Second, we suggest fundamental and detailed principles. For convenience and practicality, however, we deliberately exclude some linguistic phenomena. Finally, we attempt to develop a reliable algorithm to extract simple sentences with the ultimate goal of building Case frames. 1. INTRODUCTION In NLP, the Case frames of a language are very important for a correct syntactic and semantic analysis of the language. The term Case frames is originated from the Case grammar of Fillmore. However, the term may currently refer to the syntactic part of a lexical entry in grammars such as HPSG, LFG, and the like, or in other place it sometimes includes the semantic part, too. To confirm the common deficiency of the recent approaches to the acquisition of Case frames, let us review some of the related works. Chae-Deug Park studied on learning the Case frames of English without any consideration of preparing sufficient simple sentences as training data [7]. Chae-Kwan Song tried to automatically extract sentence patterns and the information of semantic attributes from the corpus manually tagged with parts-of-speech [8]. Tanaka used sentences analyzed by means of a full parser as training data [9]. Most of such researches so far have used the corpus tagged either by hand or by a parser. Experimental studies in a small scale manage to prepare training data manually. Such a manual arrangement, however, always results in a barrier to doing practical researches on the entire language. On the other hand, the use of any full parser, as it is without any additional processing, brings about a contradiction because the training data are obtained from the unreliable parser. Notice that Case frames are the very information for a further reliable syntactic analysis. Furthermore, currently available parsers for Korean are not even as good as those for English. The training data needed to construct the Case frames for the entire Korean verbs are nothing but a large amount of simple sentences. Unfortunately, however ordinary sentences are not in the form of simple sentences but mixed ones. If we might extract only originally simple sentences from a given corpus, hence a corpus of tremendous size would be required, which is not expected to be available in the near future [11]. This implies that we have to extract simple sentences from mixed ones. Concerning this, assuming that the Case structures and argument …

Content truncated to 500 words