Procedural Generation of Interactive Stories using Language Models

In this paper we introduce an architecture, an implementation and an evaluation of a system for the automatic creation of interactive stories for games. Our goal is to algorithmically create a branched story for the entire game; in each game run a different variant is generated. The architecture uses natural language processing (NLP) to generate meaningful stories. For NLP we use a statistical language model based on a neural network (Generative Pretrained Transformer, GPT-2). The basic architecture generates stories with too many characters which tend to get incoherent for longer texts, so we add a component restricting the number of persons and improving the consistency. The system is initialized with a hand-written game introduction that defines the main characters and the inventory; it also sets the goals for the game. From that text the remainder of the game story is generated algorithmically. We have fully implemented our system, and we report initial, encouraging experimental results.