Improving statistical natural concept generation in interlingua-based speech-to-speech translation

Natural concept generation is critical to statistical interlinguabased speech translation performance. To improve maximumentropy-based concept generation, a set of novel features and algorithms are proposed including features enabling model training on parallel corpora, employment of confidence thresholds and multiple sets of features. The concept generation error rate is reduced by 43%-50% in our speech translation corpus within limited domains. Improvements are also achieved in our experiments on speech-to-speech translation.