Genetic Programming for Task Selection in Dialogue Systems

Natural language is too complex and ambiguous to be understood by a computer using currently known methods. However, in some cases natural language interfaces are possible because interaction is limited by the set of tasks the system can perform. In this context, when a user starts a dialog, the system tries to identify the intended task, which determines the course of the dialog. Modeling tasks in order to allow selecting one is labor intensive and may cause conflicts if the system performs many tasks. We propose using ripple down rules as a task selection mechanism, and genetic programming for automatic generation of such rules. Advantages of this approach are ease of generation and possibility to learn from user interaction. We tested the approach in a multi-agent system named OMAS, where agents interact with users using natural language.