The Turing Quest: Can Transformers Make Good NPCs?
Qi Gao, Ali Reza Emami · 2023
In this paper, we investigate the potential of using large pre-trained language models to generate non-playable character (NPC) scripts in video games.We introduce a novel pipeline that automatically constructs believable NPC scripts for various game genres and specifications using Transformer-based models.Moreover, we develop a self-diagnosis method, inspired by prior research, that is tailored to essential NPC characteristics such as coherence, believability, and variety in dialogue.To evaluate our approach, we propose a new benchmark, The Turing Quest, which demonstrates that our pipeline, when applied to GPT-3, generates NPC scripts across diverse game genres and contexts that can successfully deceive judges into believing they were written by humans.Our findings hold significant implications for the gaming industry and its global community, as the current reliance on manually-curated scripts is resource-intensive and can limit the immersiveness and enjoyment of players.