In this talk we describe our continuing research on speech-to-speech machine translation. Our current JANUS system is designed to handle spontaneous conversational spoken language in limited, but fairly broad domains. In collaboration with six partner labs across the globe, we have designed an appropriate interlingua that ballances expressiveness and simplicity and has proven to be highly effective for task-oriented dialogue. In order to expand our system to broad and new domains, The MT components of our system have been engineered to build and manipulate multi-domain parse lattices that are based on modular grammars for multiple semantic domains. We are recently also exploring how to apply machine learning techniques to reduce the required level of human grammar development in our system. Our new parsing approach relies on phrase-level semantic grammars combined with trained classifiers for mapping into interlingua representations. We have also devoted significant attention to evaluating system performance and usability. In addition to user studies, we conduct two types of system evaluation - a sentence-by-sentence evaluation of translation accuracy and an overall task completion evaluation.
For the colloquium series schedule, see the UMD Computational Linguistics Colloquium Series web page at http://umiacs.umd.edu/~resnik/cl_colloquium/. If you are interested in meeting with the speaker, please contact Philip Resnik (resnik@umiacs.umd.edu).