A way to integrate context processing in the MT component of spoken, task-oriented translation systems

Handling the context, be it global, dialogic or linguistic, seems necessary to increase translation quality in spoken dialogue translation systems aiming at casual users speaking spontaneously in task-oriented situations. Traditionally, the unit of translation is a text produced by speech recognition in analysis and some intermediate form in generation (as the IF in some CSTAR systems). We propose to enrich it by a structured textual representation of the context and sketch a way to use the context during analysis and generation, at the same time producing a modified context for handling subsequent utterances.