Bringing Stories Alive: Generating Interactive Fiction Worlds (2001.10161v1)

Abstract: World building forms the foundation of any task that requires narrative intelligence. In this work, we focus on procedurally generating interactive fiction worlds---text-based worlds that players "see" and "talk to" using natural language. Generating these worlds requires referencing everyday and thematic commonsense priors in addition to being semantically consistent, interesting, and coherent throughout. Using existing story plots as inspiration, we present a method that first extracts a partial knowledge graph encoding basic information regarding world structure such as locations and objects. This knowledge graph is then automatically completed utilizing thematic knowledge and used to guide a neural language generation model that fleshes out the rest of the world. We perform human participant-based evaluations, testing our neural model's ability to extract and fill-in a knowledge graph and to generate language conditioned on it against rule-based and human-made baselines. Our code is available at https://github.com/rajammanabrolu/WorldGeneration.

Procedural Generation of Interactive Fiction Worlds

The paper "Bringing Stories Alive: Generating Interactive Fiction Worlds" by Ammanabrolu et al. addresses the challenge of procedurally generating interactive fiction worlds using narrative intelligence. The authors present a methodology that leverages existing story plots to create rich, coherent, and text-based worlds akin to those found in text-adventure games. These text-driven environments test agents' abilities to understand and manipulate virtual worlds purely through language.

Methods for World Building

Central to the authors' approach is the generation of interactive worlds guided by neural LLMs and thematic commonsense knowledge. The process involves constructing a structured knowledge graph from stories and using this graph to guide the generation of flavorful descriptions of environments, characters, and objects, thus creating an engaging game world.

1. Knowledge Graph Construction: The authors introduce AskBERT, a neural question-answering framework that extracts relevant entities from existing narratives and establishes relationships between them to develop a coherent world structure. Contrasted with OpenIE5, a rule-based approach, AskBERT utilizes ALBERT for vertex and relation extraction based on thematic questions, enabling the model to capture sophisticated semantic relations from the source text.
2. Description Generation: Once the knowledge graph is constructed, a conditional transformer model, namely GPT-2, finetuned on genre-specific story corpora, generates rich and flavorful textual descriptions. This process involves creating prompts based on narrative context and exploiting thematic evidence from similar genre stories for more authentic descriptions.

Results of Human Evaluation

The paper incorporates two human participant studies to evaluate the generated worlds—the first focused on the knowledge graph and the second on the complete game experience. Both studies include genre-specific evaluations, with mystery and fairy-tale plots serving as test cases.

• Knowledge Graph Evaluation: The researchers assessed coherence and genre adherence of extracted graphs. The neural model, in general, produced graphs that were perceived as more coherent, especially in genres where thematic commonsense significantly diverged from everyday knowledge, as was observed with fairy-tales.
• Full Game Evaluation: Comparing entire game experiences generated by different methodologies, the human-authored counterpart generally excelled, but the neural model was preferred over rule-based approaches for fairy-tales in terms of interestingness and coherence. This suggests that thematic commonsense plays a critical role in engaging user experience, particularly in genres with complex thematic elements.

Implications and Future Directions

The successful integration of thematic knowledge into procedural generation methods demonstrates potential for creating more engaging and semantically rich interactive environments. This could impact areas such as automated storytelling, AI-driven game design, and virtual world creation. The paper suggests that future research should explore more nuanced thematic extraction techniques and advanced models for generating interactive content, with increasingly diverse genres and contexts offering expansive opportunities for innovation. Additionally, further refinements to neural model strategies may yield even more coherent and genre-adherent virtual experiences.

The findings also imply productive avenues in AI narrative systems, where a deep understanding of thematic contexts can elevate the fidelity and attractiveness of generated narratives. As the methodological foundations of interactive fiction generation become more robust, the expansion of narrative intelligence applications across AI research domains can grow, pushing boundaries in fields intersecting narrative, cognition, and artificial intelligence.