Natural human conversations are complex and hard for AI to mimic convincingly. Most conversational AI agents are trained to implement dialogue for goal-directed tasks such as answering questions (eg chatbots), querying information (eg Siri), setting appointments, texting people, etc.
Live improvised acting takes conversation a step further through more abstract contextualisation, dynamic role-play, audience-suggested scenes, and rapid-fire decision-making.
To improvise well, performers must think and react quickly, collaborate fluidly, generate tension and conflict effectively, and build on each other’s offers while progressing the narrative, story, and character development.
AI Improvisers
Mathewson et al produced Pyggy and ALEx, Artificial Improvisers, to perform live comedy with. The former uses classic machine learning and deterministic rules, and the latter uses deep neural networks and advanced natural language processing (NLP) methods.
Each of these systems was tested live in front of audiences between 5 and 100 people, for a total of 25, 7-60 minute performances between 8 April 2016 and 1 June 2017. The dialogue system was iteratively improved with audience feedback.
An Artificial Improvisor dialog system is composed of three major building blocks: 1) speech recognition, 2) speech generation, and 3) a dialogue management system.
Pyggy, built using the speech recognition functions of the Google Cloud Speech API, was trained on 220k+ versatile conversational exchanges from 617 films. However, Pyggy had no means to understand topics and was creating dialogues through inferences from the training data.
ALEx (Artificial Language Experiment) used recurrent neural network-based language models to generate sentences word-by-word. Using a Long-Short Term Memory (LSTM) implementation helped the AI retain context and capture themes. ALEx’s language model was trained on subtitles from 102,916 movies released from 1902 to 2016.
ALEx somewhat keeps track of the general theme of the conversation– it creates dramatic dialogue with words related to navigation and combat. The creators noticed limitations in speech recognition and perception of non-verbal cues.
AI Improvisers have potential for improvement. Future iterations could include improvisational rules such as creating a status contrast, and accepting and expanding on offers through ‘Yes, and…’ Systems can be taught how to generate dialogue with a cohesive story and a narrative arc. Information gleaned through reading humans’ non-verbal cues can be encoded to inform the dialogue-production.
Very very unique topic!!
Rarely written topic! AI and humor is possibly a challenging idea! As humor is very contextual and some time personal. To stretch a lil, it may be sensitive. Improvisation may lead by individual to help AI to come up with improvised dialogue.