Introduction
Published in the 1984 AAAI proceedings, the Meno-tutor system represents a significant milestone in the evolution of Intelligent Tutoring Systems (ITS). Developed during an era when computer-aided instruction was largely restricted to binary "right/wrong" feedback, Meno-tutor introduced a sophisticated, hierarchical approach to dialogue management. By utilizing a Discourse Management Network (DMN), the system moved beyond simple data retrieval to emulate the nuanced, context-aware pedagogical strategies employed by human tutors.
The Problem: Context-Blind Tutoring
In the early 1980s, machine tutors typically lacked the ability to reason about the broader context of a conversation. If a student provided an incorrect answer, the system would simply output the correct fact. This approach failed to distinguish between a minor slip and a fundamental misconception. Human teachers, by contrast, adapt their rhetorical strategy based on the student's history, the importance of the concept, and the current flow of the discourse. Meno-tutor was designed to address this gap by planning its output through a multi-layered decision-making process.
The Meno-tutor Architecture
The core innovation of Meno-tutor is the Discourse Management Network (DMN). This framework organizes tutoring decisions into three distinct levels, allowing the system to separate high-level pedagogical goals from low-level linguistic choices.
- Pedagogic Level: This top layer determines the overall approach, such as whether to introduce a new topic or continue tutoring the current one.
- Strategic Level: This level refines the pedagogical goal into a specific plan. For instance, the tutor might decide to "explore competency" through questioning or "describe the domain" through direct exposition.
- Tactical Level: The final layer selects the specific "move" or utterance type, such as an emphatic correction, a hint, or a simple acknowledgment.
The system operates similarly to an Augmented Transition Network (ATN) but incorporates meta-rules. These rules act as interrupts to default paths. For example, if a student repeatedly fails to answer questions correctly, a meta-rule might trigger a shift from a questioning strategy to a "misconception-repair" strategy, effectively abandoning a failing line of inquiry to address the root cause of the student's confusion.
Key Insights: Context Over Correctness
Meno-tutor posits that providing the correct answer is not always the most effective pedagogical move. The system's flexibility allows it to adjust its "conservatism" or "aggressiveness" in correcting errors. In some contexts, the tutor might choose to "barely acknowledge" a wrong answer and pivot to providing more foundational data, encouraging the student to self-correct.
This decoupling of planning from generation allows the pedagogical logic to remain domain-independent. The system's knowledge base, represented as a KL-ONE network, is annotated with pedagogical importance, enabling the tutor to prioritize foundational concepts over supplemental details.
Results and Portability
The researchers demonstrated the robustness of the Meno-tutor framework by applying it to two distinct domains: Geography (rainfall patterns) and Computer Science (PASCAL looping constructs). The system successfully generated varied discourses for the same topic, proving it could "deliberately plan" rhetorical structures. For example, it could detect a specific misconception—such as the "northern states are dry" fallacy—and immediately shift its strategy to address it, rather than continuing with a pre-scripted sequence of questions.
Limitations and Future Directions
Despite its pioneering nature, Meno-tutor had several limitations. The system's domain knowledge was relatively shallow, and it lacked a live natural language parser; student inputs were hand-coded into conceptual equivalents. Furthermore, the integration with the surface language generator (MUMBLE) was incomplete at the time of publication. Future work aimed to develop a more autonomous system with a deeper knowledge base and integrated natural language processing capabilities.
Summary updated