Natural Language Understanding for Interaction with Digital Characters
Flavia Cavaliere
Master's Thesis, September 2024
Supervisors: Henning Metzmacher, Börge Scheel, Dr. Fabio Zünd, Prof. Dr. Bob Sumner
Abstract
Non-Player Characters (NPCs) play a crucial role in video games by adding depth to the experience and guiding players through the story, but their interactions are usually limited to preset dialogue trees. This approach has been the industry standard for many years without significant developments, despite the major advances in Artificial Intelligence (AI) and their application to other aspects of video game design. However, following recent breakthroughs in the field of Natural Language Processing (NLP), small research projects, tech demos, and short video games have begun experimenting with integrating Large Language Models (LLMs) into dialogue generation for NPCs, offering new methods of interaction. This work explores the use of multiple independent LLM instances to guide and create a story-driven conversation between a player and multiple NPCs. A game demo, based on the video game Fac ̧ade, was developed as a proof of concept, and a user study was conducted to explore how players perceive the conversations generated within the game. The results show that the large majority of participants found the dialogue from our LLM version of the game more consistent, realistic, engaging, interesting and interactive compared to the pre-scripted game version imitating the original Fac ̧ade game. Overall, the findings of this work show that player-NPC interactions could be greatly enhanced by the use of LLMs and a prompting strategy that splits up the responsibilities of story progression and dialogue generation. The contributions of this thesis are a novel prompting strategy that facilitates interactive and responsive conversations with game characters, a working game demo that implements the strategy, and a preliminary user study into the perception of LLM-generated dialogues in video games.