Software Open Access
A conversational agent is a computer program resulting from advances in artificial intelligence (AI), which aims to mimic human dialogue interactions. Theatre has made little use of this tool, even though, historically, it has been a practice centred on uttering text. On the basis of this observation, the research team directed by Nicolas Zlatoff at La Manufacture (Haute École des Arts de la Scène) is dedicated precisely to the design, development and staging of a conversational agent on the model of an actor or an actress who would improvise from a text, as he or she usually does in rehearsals.
The team uses a tried-and-tested theatrical technique, originating in the method of one of the most important pedagogues in the history of 20th century theatre, Constantin Stanislavsky, known as action analysis. It allows the actor or actress to improvise and recompose a theatrical text that he or she does not yet know by heart. Using his or her partial knowledge of the text, he or she performs it in fragments, sometimes in his or her own words. A group of acting graduates from La Manufacture will work on this practice with computer scientists who will in turn model it and gradually develop a conversational agent-actor able to engage in a dialogue in writing as well as orally with a human partner on stage.
This transdisciplinary collaboration intends to shift the perception paradigm of an AI, using the acting tools specific to theatre practitioners, and will explore the tension created by the coexistence on stage of a living and an artificial presence.
This is the release of six neural models trained for the project (117, 355 and 774 million parameters GPT-2 models, with regular and dialogue-oriented finetuned versions, all in French).
The codebase to use these is available on Zenodo/GitHub: 10.5281/zenodo.5143706 (here).