A common tool for improving theperformance quality of natural languageprocessing systems is the use of contextualinformation for disambiguation. Here I describethe use of a finite state machine (FSM) todisambiguate speech acts in a machinetranslation system. The FSM has two layers thatmodel, respectively, the global and localstructures found in naturally-occurringconversations. The FSM has been modeled on acorpus of task-oriented dialogues in a travelplanning situation. In the dialogues, one ofthe interactants is a travel agent or hotelclerk, and the other a client requestinginformation or services. A discourse processorbased on the FSM was implemented in order toprocess contextual information in a machinetranslation system. Evaluation results showthat the discourse processor is able todisambiguate and improve the quality of thedialogue translation. Other applicationsinclude human-computer interaction andcomputer-assisted language learning.