Playing with Data: Procedural Generation of Adventures from Open Data

Key insights

6 takeaways

01
The system generates playable adventure games by mining relationship paths between two real people from DBpedia (Wikipedia structured data), then turning those paths into NPCs, locations, and items enriched with OpenStreetMap geodata and Wikimedia Commons imagery.
02
In testing across 851 runs using Time's 100 Most Important People of the 20th Century, the generator produced playable games in 96% of cases (827 runs), with failures mostly due to parsing errors and missing DBpedia paths.
03
Generated games averaged about 8.3 nodes per path, and 74% of NPCs corresponded to real people; the brute-force crawler enumerates 2^(n-1) relation-direction combinations per n-node path and scores candidates by length and uniqueness.
04
Games have shallow dead ends and can be completed simply by visiting every location and NPC, indicating limited strategic depth.
05
The system produces incongruous associations, such as politicians linked only by the shared honorific "The Right Honourable," a South African locale mislabeled "New York City," and mismatched imagery like Aung San Suu Kyi standing in for Margaret Thatcher.
06
The authors argue that this absurdity is an unavoidable and partly desirable consequence of drawing on freely edited web data, and the work is a horizontal slice centered on a manhunt mechanic, with inventory, puzzles, narrative, and user studies deferred to future work.

Summary

This paper presents a system that procedurally generates simple adventure games from open data, situating its contribution within procedural content generation and "data games" in the tradition of "Where in the World is Carmen Sandiego?". The core idea is to mine structured Wikipedia data, via DBpedia, for relationship paths connecting two real people, then convert those paths into playable game elements: non-player characters, locations, and items, enriched with OpenStreetMap geodata and Wikimedia Commons imagery and linked by clues that include false leads and dead ends. The authors frame the work as a functional horizontal slice centered on a manhunt mechanic involving travel, conversation, and item consultation, with inventory and puzzles deferred to future development.

The generation pipeline uses a brute-force crawler that enumerates relation-direction combinations for a path of n nodes, yielding 2^(n-1) possibilities, and scores candidate paths by length and uniqueness. In testing across 851 runs using Time's 100 Most Important People of the 20th Century, the system produced playable games in 96% of cases (827 runs). Failures were mainly attributed to parsing errors and missing DBpedia paths. Generated games averaged roughly 8.3 nodes per path, and 74% of NPCs corresponded to real people.

The evaluation also surfaces clear limitations. Games exhibit shallow dead ends and can be completed by visiting every location and NPC, indicating limited strategic depth. More notably, the system produces incongruous associations, such as two politicians linked only by sharing the honorific "The Right Honourable," or a South African locale mistakenly labeled "New York City," along with mismatched imagery, such as a photograph of Aung San Suu Kyi standing in for Margaret Thatcher. The authors argue that such absurdity is an unavoidable and even partly desirable consequence of drawing on freely edited web data. Planned future work targets more diverse mechanics, a coherent narrative and player motive, and user studies to evaluate the generator's effectiveness.