dAIrector: Automatic Story Beat Generation through Knowledge Synthesis

• dAIrector generates semi-automated story beats through knowledge-graph synthesis for human improvisers.The system is designed for live performance and writing assistance, with quantitative and qualitative evaluation.Source: §1 · p. 1

2.1 Automated Story Generation

• Automated story-generation systems operate across abstraction levels, balancing narrative coherence against specific detail.Prior approaches include symbolic planning, hand-engineering, machine learning, and event-based recurrent models.Source: §2.1 · pp. 1–2
• For improvisational theatre, situation descriptions are preferable to prescribed dialogue or actions because vague prompts preserve actors’ freedom to explore.This design differs from systems that generate lines, character choices, or stage directions directly.Source: §2.1 · p. 2

2.2 Digital Storytelling

• Digital storytelling research has used improvisational concepts in virtual worlds, agents, scene introduction, affect detection, and robotic actors.Generative plot systems also aim to inspire human creators to explore diverse and unfamiliar narrative possibilities.Source: §2.2 · p. 2

2.3 Improvised Theatre

• Improv has no script or rehearsal: performers write and perform simultaneously while integrating remembered details with new prompts, dialogue, and actions.External suggestions add entropy that actors must justify within the current scene.Source: §2.3 · p. 2
• dAIrector models improv scenes as platform, tilt, and finding a new normal, generating platforms and tilts for actors to resolve through performance.This structure supports evaluation of cohesive plots and context-rich narrative prompts.Source: §2.3 · p. 2

3 Approach

• dAIrector is a human-machine co-creation system for live improvisation and writing assistance, using linked Plotto clauses augmented with TV Tropes information.Human artists connect the prompts through dialogue and actions to evolve narratives collaboratively.Source: §3 · pp. 2–3

3.1 Plotto

• Plotto represents plot fragments as graph nodes and connectable successors as edges, whose labels prescribe character-symbol modifications for consistency.The system parses roughly 3000 notes into JSON and generates stories by walking from a random node.Source: §3.1 · p. 3

3.2 TV Tropes

• Plotto’s abstract, ambiguous, symbol-based fragments suit improvisation because they invite interpretation, while TV Tropes supplies related high-level plot information for scene tilts.Platforms derive from Plotto fragments, whereas tilts refine the current scene using TV Tropes tropes.Source: §3.1 · p. 3, §3.2 · p. 3

3.3 Paragraph Vectors

• Paragraph vectors encode Plotto fragments and trope descriptions, enabling context-dependent matching while displaying only concise trope names to actors.The system returns five tilt options with minimum cosine distance and keeps full trope descriptions hidden from performers.Source: §3.3 · p. 4

3.4 Plot Tree Generation

• The system expands a random Plotto start node into a configurable-depth tree, then lets actors select child platforms using prompts aligned with their interpretation of the scene.Successors provide alternatives without requiring actors to follow the system’s guidance.Source: §3.4 · p. 4

3.5 Contextual Tilts

• Contextual tilts select semantically related TV Tropes entries, filter redundant matches, and randomly sample among the five closest candidates.Actors may use a tilt for additional direction or follow their own impulse.Source: §3.5 · pp. 4–5

3.6 Stage Presence

• During performance, actors request platforms or tilts through console or speech interaction, and the system matches prompts against plot-tree children or all TV Tropes candidates.Character codes are replaced with consistent names, but pronouns are not automatically changed.Source: §3.6 · p. 5

4.1 Evaluation of Tilts

• Top-5 error was 40%, compared with a 66% top-1 error rate, reflecting ambiguity in human trope assignments.Annotators often faced multiple subtly different tropes without a clear best choice.Source: §4 · pp. 5–6, §4.1 · p. 6
• The system’s roughly 700-trope evaluation set contained alternatives that could also fit human labels, but it treated broad and specific tropes as equally applicable.For example, Clear Their Name, Mystery Literature, No Mere Windmill, Lipstick Mark 5, and Prison Riot were all plausible alternatives to Get Into Jail Free.Source: §4.1 · p. 6
• System-generated tilts were graph-near human annotations, with a median distance of 2 versus 3 across the full TV Tropes set.The distance counts links between tropes; excluding exact matches, generated tilts averaged 2.5 links from human annotations.Source: §4.1 · p. 6

4.2 Sample Stories

• Sample outputs formed coherent beginning-to-end narratives, and selected tilts could refine scenes by adding context or unexpected connections.Examples include using It Belongs in a Museum to elaborate a vague secret danger and connecting clothing details to Tailor Made Prison.Source: §4.2 · pp. 6–7

4.3 Qualitative Evaluation by Professional Improvisor

• A professional improvisor described prompting dAIrector during difficult moments as fun and viewed tilts as optional story expansion rather than plot complications.The performer preferred contextual tilts despite their occasional risk of introducing difficult curveballs.Source: §4.3 · pp. 7–8
• The performer identified a limitation: dAIrector does not know what the actor wants the scene to concern or which decisions the actor makes.Future work was proposed to incorporate actors’ dialogue and actions as system prompts.Source: §4.3 · p. 8

• Treating all graph nodes as equal prevents dAIrector from controlling whether generated tilts are broad or overly specific.The authors suggest using structural information in datasets to produce tilts better suited to each application.Source: §5 · p. 8
• Using specialized narrative databases or structured television-plot data could support plots spanning seasons, character arcs, episodes, and scenes.Source: §5 · p. 8
• The graph-based story-generation walk could extend beyond narratives to directed exploration and knowledge synthesis over large graphs such as Wikipedia.Source: §5 · p. 8