Introduction
Geographic literacy among young adults in the United States has seen a significant decline, with many struggling to locate major countries on a world map despite their prominence in global news. To address this, researchers have developed a "Serious Dialogue Game" called RDG-Map. This approach combines the engagement of gaming with the pedagogical structure of Intelligent Tutoring Systems (ITS), while avoiding the high costs and time constraints typically associated with expert-authored content.
The RDG-Map Framework
RDG-Map is a collaborative game involving two roles: a Director and a Matcher. In this study, the human participant typically takes the role of the Director, who is shown a target country highlighted on a map. Their task is to describe its location to the Matcher (the agent), who views an unlabeled map. This role reversal turns the human into an "implicit tutor," forcing them to formulate spatial descriptions and resolve references through natural conversation.
As seen in the dialogue above, participants often use "anchoring points"—well-known landmarks or countries—to guide the agent. For instance, using Egypt as a reference point to locate South Sudan allows for a structured, pedagogical interaction that reinforces the human's own geographic knowledge.
Methodology: Nellie and Neil
The study explored two implementations of the RDG-Map agent:
- Nellie (Web-Based): A browser-based version used for large-scale data collection via Amazon Mechanical Turk, resulting in 8 hours of recorded audio.
- Neil (Embodied): A version utilizing the Furhat social robot, which features a back-projected face and shared attention capabilities. This version resulted in a larger corpus of 26.8 hours of audio.
Both systems were operated using a Wizard-of-Oz (WoZ) framework. In this setup, a human "wizard" monitors the conversation and selects the agent's responses and map actions from a predefined interface. This allows researchers to simulate a highly capable agent to study human-AI interaction patterns before investing in fully autonomous systems.
Key Results and Insights
The study yielded several significant findings regarding engagement and learning:
- Subjective Learning Gains: In the web-based study, 64.6% of participants reported learning something new, and 70.8% found the game to be an enjoyable educational tool.
- Performance Improvement: In the embodied study, participants' scores increased from an average of 22.24 points in the first round to 29.0 points by the third, indicating rapid skill acquisition.
- Sustained Engagement: Unlike many social robot interactions that suffer from a "novelty effect," engagement with the RDG-Map remained stable over repeated sessions, suggesting the game mechanics were the primary driver of interest.
- Naturalness: The agent's conversational style was rated highly (3.92/5), suggesting that the WoZ-mediated dialogue felt natural to users.
Conclusion and Future Work
The RDG-Map project demonstrates that collaborative dialogue games can effectively foster geographic learning while maintaining high user engagement. By positioning the human as the tutor, the system bypasses the need for expensive expert-authored content. The ultimate goal is to transition from the Wizard-of-Oz setup to a fully autonomous Spoken Dialogue System (SDS). Researchers are currently utilizing the extensive corpora collected during these studies to train agents capable of understanding and generating complex spatial references autonomously.
Summary updated