Data-driven discourse modeling for semantic interpretation

We propose a task-independent dialog controller that is able to work with different databases and tasks just changing a configuration file. This configuration file specifies a semantic tree for each task. The tree hierarchically organizes the information of the task and database, going from general concepts, towards more specific ones. So the dialogue controller is an inference engine that just follows the strategy specified in the data configuration file. We have defined four types of nodes that imply four different strategies. Moreover all output messages for each node in the tree are defined in the configuration file. So our message generator is task independent too. We have defined six different types of messages and two recovery strategies. Using this approach a system for information retrieval and voice commands has been built at Telefonica I+D.