Goal-based Generation of Motivational Expressions in a Learning Environment

Ingrid ZukermanAAAI 1987

Presented at AAAI-1987, this research explores the intersection of linguistics, psychology, and artificial intelligence to enhance the human-like qualities of computer tutors. The paper introduces a computational model for generating "Motivational Meta-Technical Utterances" (MTUs)—the encouraging or orienting phrases human tutors use to maintain student engagement.

The Need for Motivational Discourse

In the 1980s, Intelligent Tutoring Systems (ITS) were proficient at explaining technical concepts, such as solving algebraic equations. However, they lacked the "connective tissue" of human discourse. Human tutors do not merely provide formulas; they offer context and encouragement, such as, "This next part is tricky but very important," or "Let's practice this one more time so you don't forget it."

The author identifies a critical gap: while AI tutors focused on Technical Utterances (the subject matter), they neglected Motivational Meta-Technical Utterances (MTUs). These expressions are essential for managing a student's attention, interest, and persistence throughout a lesson.

The FIGMENT System

To address this, the author developed FIGMENT, a generative model based on Goal Theory applied to algebra instruction. FIGMENT tracks student progress through a hierarchical goal model to strategically insert motivational phrases.

Goal Status Modeling

FIGMENT employs a decision-tree process to predict the state of a student's goal. A goal's status determines the type of motivational utterance required. For instance, a student might be "Active" (focused), "Inactive (Remote)" (having forgotten the topic), or "Inactive (Deteriorated)" (tired or bored).

Figure 1: Decision tree for determining the status of a student's goal based on prior encounters and performance.

Hierarchical Propagation

The system represents algebraic knowledge as a hierarchy: Topic → Equation → Solution Strategy → Rules. Motivation is managed through two primary logic flows:

  • Inheritance: If a student is motivated to solve a specific equation, they "inherit" the motivation to learn the underlying rules.
  • Upwards Propagation: If a specific rule is identified as "very important," that importance is propagated upward to motivate the entire equation or topic.
Figure 3: Hierarchical structure of the tutoring domain showing the flow of motivation through inheritance and upwards propagation.

Key Insights and Results

The research treats motivation as essential metadata that helps students prioritize information. A key heuristic used is the "Least Ignorance" Principle: when motivating a student, FIGMENT prefers to imply the least possible ignorance to protect the learner's ego. For example, it might describe a technique as "interesting" rather than telling a student their skills have "deteriorated."

The primary result is the functional architecture of FIGMENT, which successfully generates context-aware phrases:

  • For remote topics: "Let us go over a technique which we haven't seen for a while."
  • For fatigue/repetition: "This type of equation has been practiced a lot, but it still demands some more practice."
  • Combined motivations: "Let us now consider a rather important technique, which enables us to solve some equations of higher degree."

Limitations

The system relies on a "model of a typical student," assuming the tutor can accurately perceive internal states like boredom or diligence—a task that remains difficult to automate. While FIGMENT can select motivational types, the actual phrasing remains somewhat template-based. Future work suggests deeper integration of social and enjoyment goals to better recognize when a student is simply too exhausted to continue.

Summary updated

BibTeX
@inproceedings{zukerman1987goal,
  author = {Ingrid Zukerman},
  title = {{Goal-based Generation of Motivational Expressions in a Learning Environment}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure