Statistical natural language generation for speech-tospeech machine translation systems

This paper presents a statistical natural language generation scheme for trainable speech-to-speech machine translation (MT) systems for limited domain applications using a cascaded approach. The natural language generation scheme in the translation systems is based on a maximum entropy (ME) statistical model fully trained from a corpus, allowing flexible translation outputs. In this pa-per, the system architecture and some of its components, including the parsing, information extraction, and translation etc are briefly overviewed, followed by the descriptions of training and search al-gorithms for ME based sentence level NLG within the MT context. Details of NLG including feature selection and robustness are also addressed. We have implemented the described system for trans-lating between Chinese speech and English speech in an air travel application domain. Encouraging experimental results have been observed and are presented. 1