Exploiting Minimal Resources for Subcategorization Frame Acquisition

The d etection of the set of syntactic fram es, i.e. syntactic entities a certain verb su bcategorizes for, is im portant especially for tasks like parsing and gram mar d evelopm ent. Machine-readable d ictionaries listing su bcategorization fram es usually give only expected fram es rather than actual ones and are therefore incom plete, or not available for som e languages, inclu d ing Mod ern Greek (MG). By acquiring fram es au tomatically from corpora, these problem s are overcome altogether. Previou s w ork on learning fram es au tom atically from corpora focu ses mainly on English H eid (1996) w ork on Germ an w hile Basili et al. (1997) d eal w ith Italian and Zem an and Sarkar (2000) focu s on Czech. In m ost of the above approaches, the inpu t corpu s is fully parsed and , if not, only a limited number of frames are learned. As is the case for the m ajority of langu ages, a treebank or a w id e coverage syntactic parser are not yet available for MG. Constru cting a treebank is expensive and tim e-demanding. The au tomatic acquisition of su bcategorization inform ation by exploitation of as lim ited linguistic resources as possible appears to be very challenging. Contrary to English, w hich has a m ore or less fixed-ord er syntactic stru ctu re, in MG the position of the constitu ents of a sentence is a very w eak ind icator of their syntactic role. Morphology, on the other hand, is essential for determining verb-argument structure. Based on the above properties of the language, the environments of the verbs in the corpus are formed and