As part of the project, the researchers intend to use artificial intelligence to create an innovative theatrical performance, which is expected to premiere early next year. Rudolph Rose, head of the research team, said:
“The main idea of our research came from Tomáš Studenik, an innovator who noticed that the centenary of the play“ RUR ”is approaching. This was a key moment for robotics, as the idea of ”Robot”, including the word “robot” itself, was invented by Karl Czapek and his brother Josef, who wrote this play.
“Tomasz believed that this should be properly celebrated and formulated the idea to turn the story around: 100 years ago, a man wrote a play about robots; what if robots write a theatrical play about humans today? ”
Before starting the project, the researchers reviewed previous literature that explored the potential of artificial intelligence techniques to create other arts.
While there are now many works on the art that AI creates, including those in which computational techniques have been used to create dialogues or ideas for plays, the automatic creation of an entire theatrical performance is an extremely difficult task never before undertaken.
Rose’s team decided to split the production of their play into several parts. Their plan is to use an approach called hierarchical generation, where the script is broken down into smaller manageable chunks.
While other research groups have already used this approach to create dialogue, few have tried to create an entire piece.
“With the upcoming anniversary, our main goal is clear and defined: by January 2021, we need to prepare the play for the premiere. Since it will be prepared by a professional theater group, we need to prepare a script in September so that there is enough time for dramatization and rehearsals.
“So even though we are developing a common script generation tool in theater, we really only need to create one script at this point.”
Researchers began experimenting with a pretrained language model called GPT-2. This is an open source model developed by the OpenAI consortium organized by Elon Musk, trained by a huge amount of online texts in English.
Once properly trained, a generative language model is capable of completing unfinished texts using a similar language and covering related topics.
For example, if the first paragraph of a news article is served in GPT-2, the model will try to generate several additional paragraphs on the same topic, using existing text for inspiration, while generating sentences in new concepts.
“When we filed a GPT-2 scene and some lines of drama script, it generated additional lines in the same style and focused on the theme of the input script fragment. So we didn’t have to train anything (at the moment) as we limited the generator a bit to control the task. ”
“So we can use the large GPT-2 model, which has been trained for a very long time on very large texts that we could not afford to implement in training.”
Contact us: [email protected]