Achievement

Dialogue systems used in natural language interfaces are designed to engage in conversation with humans in ways that are flexible and effective. Trainee McMahan (trainee) and Stone designed a system that learned how to generate natural conversations. A key goal was to have the model decide when to perform an utterance and when to wait for the human. The model was developed from data acquired from humans performing a referential language task (object identification) that required a series of utterances. Each utterance was represented by how far it advanced the dialogue to goal completion. Performing cross validation trials (training the model on 23 conversations and testing on the 24th) showed significant improvements in the model’s generation of descriptions and acknowledgments within the dialogue. (Proceedings of the 14th meeting of the SIG on Discourse and Dialogue, Association for Computational Linguistics, Aug 2013).