TL;DR
This paper introduces "Usnea," a novel authorship tool for creating interactive fiction (IF) that bridges the gap between traditional IF storytelling and modern neural semantic parsing for dialog systems. By integrating retrieval-based semantic parsing, Usnea allows authors to define branching narratives with more flexible player input than previous systems, while providing a user-friendly interface for non-programmers.
What Problem Does This Paper Address?
Interactive Fiction (IF) has a rich history, evolving from early text adventures like Zork to modern narrative games. However, authoring IF often requires specialized programming knowledge or adheres to strict input formats, limiting creative freedom. Existing tools, while powerful, can have steep learning curves. This paper addresses the challenge of creating an IF authoring tool that is both powerful and accessible, allowing authors to leverage the flexibility of modern natural language understanding (NLU) techniques without needing deep technical expertise. They aim to relax the rigid lexical constraints of traditional IF parsers by employing retrieval-based semantic parsing, inspired by virtual assistant dialog systems.
Methods
The core of Usnea's approach lies in its integration of retrieval-based semantic parsing for Language Understanding (LU). Instead of relying on complex, hand-crafted grammars or deep learning models requiring extensive training data, Usnea uses a nearest-neighbor classification approach. Authors define a set of "exemplar strings" (positive examples) and "anti-exemplar strings" (negative examples) for each semantic intent. The system then uses a semantic similarity function (a "semantic kernel") to match user input to these exemplars. This method is inspired by recent work in dialog systems and allows for a more freeform interaction style, as long as the author can anticipate and guide the reader's inputs into defined "neighborhoods" of exemplars.
The system's Dialogue Management (DM) is based on a directed graph, where nodes represent story states and edges represent transitions. These transitions are triggered by the semantic intents identified by the LU system. The DM also incorporates a "World" state, a string-to-string map that can store and mutate game state (e.g., player inventory), enabling more complex narrative logic. Additionally, an "AutoAdvance" property on nodes allows for automatic transitions, streamlining certain narrative flows.
The Authoring Tool itself is designed with two main components: a graph editor for visualizing and manipulating the story's topology, and a graph inspector for editing node and edge properties (prompts, mutations, conditions, match candidates). The tool is implemented as a web application using Angular and Firebase for persistence.
Datasets
The paper does not explicitly mention specific datasets used for training the core semantic similarity model. However, it references the Universal Sentence Encoder (Cer et al., 2018) as the implementation for the semantic similarity function D. The training data for this encoder is drawn from the SNLI corpus and additional unsupervised multi-task training data from sources like Wikipedia and web news. The authors note that while this general domain data provides reasonable results, domain-specific fine-tuning could further improve performance for IF.
Key Insights and Ideas
- Bridging IF and Dialog Systems: The paper highlights the surprising parallels between IF and modern dialog systems, both involving turn-taking interactions and requiring semantic understanding. This connection allows for the application of advanced NLU techniques to IF authoring.
- Democratizing IF Authoring: By using retrieval-based semantic parsing with author-defined exemplars, Usnea significantly lowers the technical barrier to entry for IF creation. Authors don't need to be programmers or machine learning experts to define sophisticated interactions.
- Flexible Input Handling: The retrieval-based approach allows for more natural and varied player input compared to traditional IF parsers that rely on strict verb-object commands. As long as the author provides relevant exemplars, players have more freedom in how they express their actions.
- Iterative Authoring Workflow: The inclusion of a "Node Tester" and "Preview Mode" within the authoring tool emphasizes an iterative development process. Authors can easily test their semantic parsing definitions and story flow, making it easier to refine content based on anticipated reader interactions.
- The "World" as a Narrative Engine: The concept of a mutable "World" state, managed by the Dialogue Manager, is crucial for enabling complex IF narratives with persistent state changes, akin to inventory management or character status.
Key Results
The paper focuses on the model specification and the authoring tool rather than presenting quantitative results from a large-scale user study or benchmark. The key "results" are the successful integration of retrieval-based semantic parsing into an IF authoring framework and the development of a user-friendly tool that facilitates this process. The authors demonstrate the functionality through examples and figures, showcasing how authors can define story branches, handle player input, and manage game state. The open-sourcing of the tool is also a significant outcome.
Limitations and Future Work
- Semantic Similarity Model: The authors acknowledge that while the Universal Sentence Encoder provides good general-purpose similarity, domain-specific fine-tuning could significantly improve the accuracy and nuance of semantic matching for IF. They suggest that recent advancements in pre-training (e.g., BERT, XLNet) could be leveraged.
- Authoring Effort: While Usnea simplifies authoring, defining comprehensive sets of exemplars and anti-examples for complex IF can still be a significant undertaking. The effectiveness of the system relies on the author's ability to anticipate and cover a wide range of potential player inputs.
- Scalability of Exemplars: For very large or complex IF projects, managing a vast number of exemplars might become challenging.
- Future Work: The authors express hope that authors will find the tool useful for creating engaging IF. They also implicitly suggest future work in exploring more advanced semantic similarity models and potentially integrating more sophisticated NLU techniques as they become available. The flexibility of their architecture, with injected services for persistence and semantic similarity, is designed to accommodate future improvements.
Summary updated