Use of maximum entropy in natural word generation for statistical concept-based speech-to-speech translation

Abstract

Our statistical concept-based spoken language translation method consists of three cascaded components: natural language understanding, natural concept generation and natural word generation. In the previous approaches, statistical models are used only in the first two components. In this paper, a novel maximum-entropy-based statistical natural word generation algorithm is proposed that takes into account both the word level and concept level context information in the source and the target language. A recursive generation scheme is further devised to integrate this statistical generation algorithm with the previously proposed maximum-entropy-based natural concept generation algorithm. The translation error rate is reduced by 14%-20% in our speech-to-speech translation experiments.