Semantic analysis of prepositional phrases in English-to-Korean machine translation

Abstract

The correct analysis of a prepositional phrase(PP) has been an important issue in natural language processing. In English-to-Korean machine translation(EKMT), the analysis of prepositional phrases in English has critical effect on the generation of the corresponding Korean expressions. This dissertation proposes a semantic analysis system to resolve PPs in EKMT. The system consists of two major parts; a PP-attachment resolution part and a semantic role selection part. The PP-attachment resolution part is a hybrid system with a neural network and a rule-based controller. The neural network takes two candidates from an input sentence and chooses the right PP-attachment by comparing the two attachment candidates. The rule-based controller manages the execution of the neural network and chooses the best attachment candidate. The controller makes it possible to process a variable sized input sentence that is difficult to manipulate with simple neural networks. The semantic role selection system consists of three neural networks. Two of them are for disambiguating the senses of the attachment candidate chosen by the PP-attachment resolution part and the noun phrase of the PP. The remaining network takes the outputs of the two sense disambiguation networks into the input layer and determines a proper Korean-style case role of the PP. The role represents the Korean-style case relation designed with the consideration of the characteristics of Korean expressions on various postpositions to generate a correct Korean expression for a PP. As inputs to the neural networks, the system uses the semantic features of each word. The semantic features are defined based on several researches. The semantic features of an input word are automatically extracted from the WordNet using a mapping table which maps the senses of WordNet into the semantic features of this system. Automatic acquisition of the semantic features guarantees objectiveness in expanding the system. We tested the p...