SLANG (2017)

Unpictured
  • Followers
About this project

A System for Learned Artificial Narrative Generation (SLANG): We plan to develop an algorithm that would be able to write syntactically and contextually consistent stories of any arbitrary length, with minimal user input beyond a random seed and a select few parameters. This system would learn from a large collection of stories written by humans, and would be able to generate a story of its own without attempting to mimic any particular story or story structure. Our current direction involves using a Variational Autoencoder to encode sentences into latent variable distributions, and train a Generative Adversarial Network that would operate on the encoded space to learn correlated sentence sequences. Beyond correlated sentence sequences, we are considering a hierarchical recursive model that encodes an higher-level and more abstract ideas from which to derive sentences, as well as a representation model for characters to properly derive correct actions for each character.

Project Members