Statistical Input Method based on a Phrase Class n-gram Model

We propose a method to construct a phrase class n-gram model for Kana-Kanji Conversion by combining phrase and class methods. We use a word-pronunciation pair as the basic prediction unit of the language model. We compared the conversion accuracy and model size of a phrase class bi-gram model constructed by our method to a tri-gram model. The conversion accuracy was measured by F measure and model size was measured by the vocabulary size and the number of non-zero frequency entries. The F measure of our phrase class bi-gram model was 90.41%, while that of a word-pronunciation pair tri-gram model was 90.21%. In addition, the vocabulary size and the number of non-zero frequency entries in the phrase class bi-gram model were 5,550 and 206,978 respectively, while those of the tri-gram model were 22,801 and 645,996 respectively. Thus our method makes a smaller, more accurate language model.