Fine-tuning GPT-2 on annotated RPG quests for NPC dialogue generation

Abstract

GPT-2, a neural language model trained on a large dataset of English web text, has been used in a variety of natural language generation tasks because of the language quality and coherence of its outputs. In order to investigate the usability of GPT-2 for text generation for video games, we fine-tuned GPT-2 on a corpus of video game quests and used this model to generate dialogue lines for questgiver NPCs in a role-playing game. We show that the model learned the structure of quests and NPC dialogue, and investigate how the temperature parameter influences the language quality and creativity of the output artifacts. We evaluated our approach with a crowdsource experiment in which human judges were asked to rate hand-written and generated quest texts on language quality, coherence and creativity. CCS CONCEPTS• Applied computing → Computer games; • Computing methodologies → Natural language generation.

Generating Role-Playing Game Quests With GPT Language Models

“…Our results suggest that even the largest variant of GPT-2, fine-tuned on our well-curated data set, cannot be used to autonomously generate high-quality quest descriptions reliably. This confirms findings in related work [22]. We especially found that Quest-GPT-2 lacks the ability (i) to distinguish between multiple entities, and (ii) to "glue" quest ingredients well together while not relaying illogical information.…”

“…Based on a small user study with 75 participants, they found that the GPT-2 quests were experienced as more valuable and coherent, but less surprising and novel than quests produced by random assignment or Markov chains. Most closely related to our work, van Stegeren and Myśliwiec [22] have recently fine-tuned GPT-2 for the generation of quest descriptions told from the perspective of an NPC. Crucially though, they solely use data from the Massively Multiplayer Online Role-Playing Game (MMORPG) World of Warcraft [23].…”

“…We adopt the hypothesis from related work [22] that the data used to pre-train GPT-2 does not contain a sufficient amount of quest examples to facilitate high-quality quest generation without additional fine-tuning based on a separate, specialized data set. We confirmed this hypothesis by investigating the output of the vanilla GPT-2 model with 744M parameters, if presented with different quests (cf.…”

TropeTwist: Trope-based Narrative Structure Generation

“…Furthermore, the generation of narratives, stories, and quests using a variety of techniques such as planning algorithms [47,61], grammars [7,27], or machine learning [53,58], is a growing and important field within games research and narrative research in general [18,22,32,62]. One typical approach for the generation of content and stories is the use of patterns representing different elements such as level design patterns [4,56], quest patterns and common quests in games [16,57], or identifying fundamental units and assembling them based on various pre-conditions [20,31].…”

TropeTwist: Trope-based Narrative Structure Generation

“…Furthermore, the generation of narratives, stories, and quests using a variety of techniques such as planning algorithms [21]- [23], grammars [24], [25], or machine learning [26], [27], is a growing and important field within games research and narrative research in general [8], [20], [28], [29]. One typical approach for the generation of content and stories is the use of patterns representing different elements such as level design patterns [9], [30], quest patterns and common quests in games [25], [31]- [33], or identifying fundamental units and assembling them based on various pre-conditions [34], [35].…”