Case-Based Techniques Used for Dialogue Understanding and Planning in a Human-Robot Dialogue System

Karolina EliassonIJCAI 2007

Introduction

Traditional human-robot dialogue systems often suffer from "brittleness," relying on static, manually authored scripts that fail to account for the vast variability in natural language. In the context of complex tasks, such as commanding an Unmanned Aerial Vehicle (UAV), a dialogue manager must not only parse words but also translate them into actionable sequences while managing sub-dialogues and resolving missing information.

This paper introduces CEDERIC, a system that leverages Case-Based Reasoning (CBR) and Case-Based Planning (CBP) to create a more flexible and adaptive dialogue interface. By treating both linguistic interactions and robot actions as "cases" within a unified memory, CEDERIC can adapt to novel phrasing and automatically resolve implicit commands through a process of continuous learning.

The CEDERIC Architecture

The core of the system is a unified engine that integrates natural language understanding (NLU) with action execution. Unlike systems that separate these functions into isolated modules, CEDERIC uses a shared case-based approach to bridge the gap between "understanding" a command and "knowing how to execute" it.

The CEDERIC system architecture, illustrating the integration between the SGUI, the core engine, and the Robotic Control System.

As shown in Figure 1, the architecture consists of three primary components: the SGUI (Speech and Graphics User Interface) for user interaction, the CEDERIC engine (containing the case-base manager, lexicon, and discourse model), and the Robotic Control System which interfaces with a physical robot or simulator.

Case-Based Reasoning and Planning

CEDERIC employs two primary techniques to manage interactions:

  • Case-Based Reasoning (CBR): Used for interpreting user utterances. Instead of rigid keyword matching, the system compares input to a library of previous interactions. A similarity function evaluates word importance and semantic categories (e.g., treating "car" and "vehicle" as similar), allowing the system to generalize across synonyms and varied sentence structures.
  • Case-Based Planning (CBP): Used to generate and repair action sequences. When a user command implies a sequence of actions, the system checks the preconditions and effects of each step. If a "plan fault" is detected—such as a command to fly to a location while the UAV is still grounded—the system searches its case base for a "repair plan" (e.g., inserting a "take off" action) to fulfill the request.

Experimental Evaluation

The system was evaluated using a custom case base initially seeded with 22 hand-crafted problem cases. Testing involved five users with no prior experience, who were tasked with completing five distinct missions (e.g., navigating to specific buildings or altitudes) using a simulator for a Yamaha RMAX helicopter. CEDERIC was compared against a baseline system that lacked CBR/CBP capabilities, requiring exact linguistic matches and offering no automated planning.

Comparison of dialogue turns required to complete five different missions using the baseline system versus the CEDERIC system.

The results demonstrated significant improvements in both success rates and efficiency:

  • Task Success: All users successfully completed all missions using CEDERIC. In contrast, zero users in the baseline group were able to complete the more complex Mission 4.
  • Efficiency: As illustrated in Figure 4, CEDERIC significantly reduced the number of dialogue turns required to complete tasks. For instance, in Mission 1, the baseline required an average of ~25 turns due to failed attempts, while CEDERIC users finished in roughly 7 turns.
  • Generalization: Manual testing confirmed that the initial 22 cases could resolve 30 additional "similar but not identical" problems through CBR and 8 complex scenarios through CBP.

Conclusion

CEDERIC demonstrates that a unified case-based approach can significantly enhance the flexibility and robustness of human-robot dialogue. By prioritizing semantic similarity over rigid syntax and integrating planning directly into the dialogue management process, the system can handle implicit commands and learn from experience. Future work will address the scalability of the case-base retrieval process and the handling of more complex linguistic structures such as nested conditionals.

Summary updated

BibTeX
@inproceedings{eliasson2007case,
  author = {Karolina Eliasson},
  title = {{Case-Based Techniques Used for Dialogue Understanding and Planning in a Human-Robot Dialogue System}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2007},
}
Figure