Engineering Conversation: Understanding the requirements of control in language generation.

This event has ended.

In recent years, control theoretic principles have been at the heart of neurobiologically-plausible models of brain function and behaviour and, increasingly, are also informing the design of biologically-inspired robots. But while these control-based approaches have led to great advancements in the context of motor control, they have not yet been applied to the challenge of language generation.

State-of-the-art social robots generate language either on the basis of inflexible rule-based systems, or of more flexible data-driven systems that are however dependent on the availability of large amounts of annotated data. Both of these approaches therefore have limitations, and neither is guided by our understanding of the psychology and neuroscience of language production. While computational models of speech motor control are well developed (e.g., the DIVA model), they do not currently extend beyond production of single syllables.

In this workshop, we bring together an interdisciplinary group of experts to discuss how control theoretic principles could apply to language generation from message to sound, and in the context of conversation.