Creating Unique Gameplay Scenarios Using Natural Language Generation

Abstract

Language and text generation is a complicated task for computers, there is a lot of nuances and context required to understand a sentence. For decades natural lan-guage processing has been a field of study, for computers to process, understand and generated texts. Games such as Dungeons & Dragons are heavily based in language, the players talk during the game, and the dungeon master leads the game by describing what the players can do. However, these games also base themselves on rules, limiting the freedom that a storyteller such as an author has. This thesis investigates creating a machine learning model which can create unique gameplay scenarios, called encounters, which fit within the rules of Dungeons & Dragons. The GPT-2 text generation models are based on the transformer architecture created by Google and can generate high-quality text without additional training. A fine-tuning process allows the model to train on specific types of text, teaching it how to create similar texts. Using gathered data this fine-tuning process can create a model which is able to generate Dungeons & Dragons encounters quickly.