A HYBRIDNEURAL NETWORK FOR LANGUAGE IDENTIFICATION FROM TEXT Eniko' Beatrice Bilcu, Jaakko Astola
Korkeakoulunkatu · 2006
Inthis paper, theproblem oflanguage identification from awritten textisaddressed. Theproblem isconnected to text-to-phoneme mapping application wheretheletters of awritten textmustbetranslated into their corresponding phonemes. Whenthewordsofthewritten text belong only toasingle knownlanguage, transcription ofthewritten letters into their phonemes canbedonedirectly (without the needoflanguage identification). Whenthelanguage ofthe input textisnotknown,usually thefirst stepistoidentify thelanguage and,based onthis information, thetextto-phoneme mapping isdone. Inthis paper, weintroduce a novelhybrid approach composed ofa neural network equipped withadecision rule forlanguage recognition from written text. Comparative results showing theimproved performance ofournovel approach compared withanexisting similar method arealso shown.