Neuro Trans

NeuroTransA New Model for Connectionist Machine TranslationXu Luomai, Ph.D.Guangdong University of Foreign StudiesThis thesis aims at developing a connectionist technique for machine translation (MT).This technique treats MT development as seeking the solutions to two sub-problems. The firstproblem concerns obtaining vocabulary translations and is solved by using a distributedneural translation lexicon. The second problem involves adjusting the transliterationsproduced by the lexicon into acceptable target language sentences, which is handled by ahybrid generator.The translation lexicon learns the meaning of words from exampIes and stores theacquired lexical knowledge in a set of lexical networks. During translation, the lexicalnetworks perform lexical disambiguation automatically. With the neural lexicon,programming a disambiguation component for an MT system is no longer necessary. Thetechnique allows a translation lexicon to scale up easily to a full size lexicon for an MTapplication. Although developed initially for English-Chinese translation, the technique canbe used to develop translation lexicons between any language pairs.The hybrid generator consists of a generation network (GN) and a symbolic generator(SG). GN learns a simple jumble of grammar from examples, and SG physically adjuststransliterations to produce turget language sentences. The generator is a bold attempt atlanguage generation without using any formal linguistic theories. This thesis discusses itsstrengths and weaknesses. This new language generation technique is still under development,requiring further research before becoming a practical alternative to those based on thesymbolic approach.