Towards Data-Driven Drama Management: Issues in Data Collection and Annotation

This work argues for a data-driven approach to intelligent drama management, moving away from hand-authored top-down systems toward models learned from real human play. Its central premise is that pen-and-paper role-playing games offer a suitable source of such data, since they embody a pure form of collaborative storytelling governed by a structured decision–action–reaction–processing cycle and an explicit story-facilitating role for the game master. Because the operational strategies of game masters and the mechanics of authorial control remain poorly understood, the authors undertake empirical collection and annotation of play as a foundation for their design.

The empirical basis consists of annotated transcripts from ten sessions involving 51 participants and roughly 60 hours of recorded play, all run by experienced game masters using a D20-based scenario. Player and game master communication is coded along dimensions including narrative progression, character actions, assistance, critique, and dramatic language, and these codes are organized into hierarchical case frames. The authors propose D3M, a case-based reasoning architecture in which such frames can be matched and reused at runtime to guide dramatic decisions.

The findings are explicitly scoped to the classical single-game-master model of pen-and-paper play rather than to the full range of tabletop formats, and the authors are candid that annotation remains in progress. Current work includes automated annotation and the neural or evolutionary prediction of a game master's next action. Taken together, the contribution is best understood as a structured annotation scheme and an architecture for annotated play traces, offered as a first step toward a high-quality interactive drama manager grounded in observed human expertise.