A man-machine interface system which produce commands to a machine through interactive speech conversation has been proposed. To improve the accuracy of the speech understanding and to realize natural conversation, it is necessary to make effective use of the task-dependent knowledge. It is easy to achieve the purpose if we adopt the task-dependent algorithm. In this case, however, the system lose the flexibility for the change of target machine. We attempt to construct task-independent algorithm by separating task-dependent knowledge from linguistic knowledge and conversation control strategy. This approach holds down the cost of the system modification which is required when the target machine is changed. This paper describes the method of the description of task-dependent knowledge and the algorithm of the system which utilizes the knowledge.