Word selection is a key role of a transfer system, which selects a target word that most closely matches the source word being to be translated. In rule-based approaches, it is difficult to acquire knowledge for word selection. In the previous target language statistics-based approaches, incorrect target words can be selected because senses of source language and structures of target language are not considered.
A word can have many senses, and each sense can be translated to many target words. This thesis presents a two-step word selection method which uses source language-specific transfer rules and target language-specific syntactic cooccurrence statistics. First, a sense of an English word is selected by using English-specific transfer rules. And then, for the selected sense, the most appropriate Korean word is determined by using Korean syntactic cooccurrence statistics.
Our method can reduce the difficulty of knowledge acquisition because Korean syntactic cooccurrence statistics can be automatically acquired from syntactic tagged corpus and it is relatively easy to write source language-specific transfer rules. And, our method can select target words more accurately than the previous statistics-based approaches, because Korean syntactic cooccurrence statistics are used with syntactic relations of the Korean structure which is transformed through a structural transfer.
