Modeless Japanese Input Method Using Multiple Character Sequence Features

Abstract

Recently, the rapid growth of globalization requires writing a large number of multilingual texts. However, Japanese PC users need to switch the input mode between Japanese and the Latin alphabet on conventional Japanese input method. That is cumbersome. Meanwhile, the solution system using a dictionary is hard to maintain because new words are created every year with high frequency. This paper proposes a modeless Japanese input method which automatically switches the input mode without using a dictionary. Using the model called "multiple character sequence features", this method discriminates whether to convert alphabet into Kana or not. There are multiple character sequence features, namely, character surface features and character type features both based on n-gram. These model features are learned by a Support Vector Machine from corpora especially from those of a large number of living words on Web. The evaluation of this method showed that the statistical accuracy by F-measure for both chatting texts and news texts was over 90% (mostly over 99%).