Year: 2022
Type: dissertation
Abstract: Recent advances in artificial intelligence research have brought forth text-generating language models with promising computational storytelling capabilities. This thesis leveraged one of the most successful Transformer models, GPT-2, to automatically generate video game quest descriptions. We aimed to replace traditional procedural content generation methods for quests in games, as these often produce uninteresting, mechanical descriptions. We have gathered and processed a novel quest data set from a selection of popular 3D role-playing games. We have fine-tuned GPT-2 on the data set using various learning optimizations. Most notably, we replaced proper nouns within raw quest data with generic placeholder tokens to reduce unnecessary variance. We validated the resulting \questgpt{} model via an online user study performed by role-playing game players. Our results indicate that one in five quest descriptions would be deemed acceptable by a human critic, yet the variation in quality across individual quests is large. Further work utilizing next-generation language models and our quest data set is expected to result in improved quest description quality.
Source: Aaltodoc (Aalto University)
Author: Värtinen, Susanna
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7117220604