Project Name: P-Interact
Project Manager (Class, Year, Major): Andy Liu (HMC 23)
Preferred Method of Contact (Facebook, email): Messenger (@uilydna), Email ([email protected])
Released in beta by OpenAI in 2020, the language model GPT-3 has become a go-to for generating human-sounding English text. However, while GPT-3 outputs can often be indistinguishable from written text, its working memory of only a few hundred words strongly limits how much context it can maintain. For these reasons, GPT-3 has largely proven unsuccessful at maintaining consistency or coherency over longer works such as stories. One way to naturally bridge this gap may be using GPT-3 to generate works of interactive fiction (fiction where the reader can select from a list of continued actions to influence the course of the story) similar to those generated by Twine or published to "Choose Your Own Adventure" websites.
Having gained access to the beta API of GPT-3, we plan to start by playing around with its parameters and inputs to generate "prompts" for potential actions given human-defined introductions, then using GPT-3 on these prompts to generate next steps in the interactive fiction. We will then experiment with different methods of seeding GPT-3 prompts or "encoding" the state of the story (e.g. characters and their status, plot direction, setting details, using existing summarizer models to remember past sections) in order to promote coherence. We can test the success and accuracy of these encoding methods by manually labelling choose your own adventure fiction according to whatever state encoding we use to feed into GPT-3 and seeing if our model can accurately keep track of this encoding without human intervention. If there is time, we can consider scraping existing fiction and attempting to retrain smaller language models, or further experimentation with different encoding variants.
At a minimum, we will deliver a functional web app that takes in user prompts and generates full interactive fiction works according to user input, and use manual labelling of outputs based on existing CYOC stories to tune hyperparameters for coherency. If time permits or if there is greater interest from the team, we can try to modify our GPT-3 usage to generate full text adventures where users can type in any actions to influence the plotline (as a stretch goal). Other potential extensions during the semester might include experimenting with different GPT-3 inputs to adjust the style of the text output, or anything else proposed by the team.
Assuming ~ten weeks (starting early October and ending mid December) for project execution:
Copy of Rough timeline of project