To provide a Roman character-KANA (Japanese syllabary)/KANJI (Chinese character) direct conversion device capable of automatically and directly converting a Roman character string continuously inputted without executing divided writing or shift between capitals and small letters into a KANA/KANJI mixed sentence by using candidates having high conversion definition.

Words are roughly sorted into adjuncts, conjugating independent words and non-conjugating independent words, dictionaries 311, 331, 341 for respective sorts are prepared and the headings of the conjugating independent word ditionary 331 are especially expressed by Raman character stems. Occasional conversion start and constant conversion start are executed in parallel through a direct conversion loop consisting of respective word sort searching parts 31, 33, 34 applying left search, connection relation among the word sorts and the search priority of the word sorts, a word sort searching part selection means 32, a regression judging means 36, a converted Roman character string reducing means 37 for reducing one alphabet from the final end of a Roman character string to be converted in each regression, and an adjunct connecting means 35 to an independent word and direct conversion from a short Roman character string to be converted to a head coincident clause or word sort and its output are repeated.