Descripción del proyecto
Our ability to communicate using language in conversation is considered the hallmark of human intelligence. Yet, while holding a dialogue is effortless for most of us, modelling this basic human skill by computational means has proven extremely difficult. In DREAM, I address this challenge by establishing a new computational model of a dialogue agent that can learn to take part in conversation directly from data about language use. DREAM stands at the crossroads of the symbolic and the sub-symbolic traditions regarding the nature of human cognitive processing and, by extension, its computational modelling. My model is grounded in linguistic theories of dialogue, rooted in the symbolic tradition, but exploits recent advances in computational learning that allow the agent to learn the representations that it manipulates, which are distributed and sub-symbolic, directly from experience. This is an original approach that constitutes a paradigm shift in dialogue modelling --- from predefined symbolic representations to automatic representation learning --- that will break new scientific ground in Computational Linguistics, Linguistics, and Artificial Intelligence. The DREAM agent will be implemented as an artificial neural network system and trained with task-oriented conversations where the participants have a well-defined end goal. The agent will be able to integrate linguistic and perceptual information and will be endowed with the capability to dynamically track both speaker commitments and partner-specific conventions, leading to more human-like and effective communication. Besides providing a breakthrough in our capacity to build sophisticated conversational agents, DREAM will have substantial impact on our scientific understanding of human language use, thanks to its emphasis on theory-driven hypotheses and model analysis.