Case analysis is a process to find semantic relations between components of a sentence. Case analysis has been treated in many approaches including rule-based approaches and example-based approaches. These approaches, however, have many problems in that it is difficult in keeping consistency of the rules and in extensibility of the system. Because of these difficulties, several approaches tried to use neural networks. But, in turn, they require input to be a fixed sized sentence, which is too strong restriction. To resolve this problem, we propose a case analysis system that treats the triples of constituents, i.e., noun, verb, and postpositional particle, rather than dealing with whole sentence as a unit.
The proposed case analysis system consists of three parts: preprocessing part, word sense disambiguation, and case role assignment. In preprocessing, each data is prepared automatically through a mapping table for each Korean word into a feature vector. network nodes. The word sense disambiguation and the case role assignment parts are implemented with neural networks.
In this thesis, the catagorization of cases which is suitable for Korean sentences is designed, and the distributed representation of word senses which is efficient for case analysis is proposed. To enhance the performance, word sense disambiguation method using neural network is introduced, and the case assignment neural network is implemented.
Experiments show that the performance on the trained data is 95.8%, and that of untrained one is 89.6%, which are promising.