The Student Model: From Text-Based to Multimedia Tutoring Systems

This work was supported by the Cognitive Science program, Office of Naval Research under Grant No. N00014-94-1-0338 to the Illinois Institute of Technology. The content does not reflect the position or policy of the government and no official endorsement should be inferred.

Abstract

Student modelling has traditionally been used to determine what to tutor. We have studied human-to-human tutoring transcripts to aid in the design of an Intelligent Tutoring System (ITS). We not only identified how a human tutor determines what to tutor, we also identified how a human tutor determines how to tutor. A human tutor, based on the interaction with a student, chooses from a range of tactics. These are arranged along a continuum from the passive (i.e., a didactic explanation) to the active (i.e., an indirect hint). Our tutors' behavior has been codified and simulated in our ITS. Our subjects are first year medical students. The protocols for our human-to-human tutoring experiments and our ITS are nearly identical: the student reads text from the tutor on the screen and replies with natural language via the keyboard.

We wish to extend our study of using student modelling in an ITS to determine how to tutor in a multimedia tutoring system. For example, can recent student performance be used to determine if a topic should be presented textually or graphically? Are there passive and active graphical tutoring tactics? Integrating current ITS technology with multimedia education systems will allow the full potential of multimedia to be exploited in support of learning.

Introduction

Our study of Intelligent Tutoring System (ITSs) has been limited to the design and development of CIRCSIM-Tutor (CST; Kim et al., 1989). CST is designed to tutor first year medical students on the human body's blood pressure regulating system. To understand the dynamics of one-on- one tutoring, we have recorded (Li et al., 1992) and studied human-to-human tutoring sessions. A keyboard-to- keyboard protocol was used to simulate a student-to-ITS interaction.

The most striking phenomenon observed from these transcripts is our tutors' frequent use of hints (Hume et al., 1993, 1996a). We subsequently identified all of our tutors' tactics and positioned then on a passive to active continuum (Hume et al., 1995). Hints engage students in the greatest degree of cognitive activity. A didactic explanation allows the student to be a passive participant. A human tutor constantly monitors the progress of a student. This activity is referred to as student modelling in an ITS. We have identified how a human tutor uses the equivalent of a student model to select tactics (Hume et al., 1996b). Our tutors tend to employ active tactics if they believe that student can make use of them; they resort to passive tactics when the student performs poorly.

We wish to generalize our findings; we are aware that our domain is narrow in many ways. Our students are older and more motivated than most subjects of ITS experiments. The solutions to CST problems require that the student understand qualitative, causal relationships; there is little quantitative manipulation required. The primary form of communication is natural language via the keyboard and screen. We hope to learn (1) how the hinting phenomenon can be implemented in a multimedia environment, (2) how different tactics in a multimedia environment require different degrees of cognitive activity of the student and (3) the role of student modelling in a multimedia tutoring environment.

This paper will (1) illustrate key aspects of our work with CST, (2) introduce a domain that will be used for a new multimedia ITS and (3) outline an argument for why traditional ITS technology should be incorporated in multimedia education.

CST's Human Tutoring Experiments

We used transcripts from human-to-human tutoring experiments to design all aspects of CST. Our most important conclusion is the following definition of a hint (Hume et al., 1993, p. 564):

A rhetorical device that is intended to either: (1) provide the student
with a piece of information that the tutor hopes will stimulate the
student's recall of the facts needed to answer a question, or (2) provide a
piece of information that can facilitate the student's making an inference
that is needed to arrive at an answer to a question or the prediction of
system behavior.

Hints may provide explicit information or they may simply allude to information. Hints may be followed up with explicit questions or the follow up question may be left implied. Hints may be phrased as statements while the intention is to ask a question. Some hints are phrased as questions when the intention is to answer a student question. Hints may summarize segments of preceding dialogue or they may introduce new information. The surface form of hints are varied yet one intention of the tutor is always constant: guide the student towards self discovery.

We have identified two categories of hints: point-to hints (PT-Hints) and convey-information hints (CI-Hints). A CI-Hint explicitly conveys information in the form of an explanation or summary and is followed up with a question. A PT-Hint alludes to information presumed to be available to the student. This information does not contain any part of a desired answer but provides information that should enable the student to proceed.

We have further identified three other tactics: explanations, summaries and directed line of reasoning (DLR). A DLR is an interactive dialogue in which the tutor prompts the student in a stepwise manner to arrive at a solution to a problem. The intention of the tutor is to help the student reason about the problem at hand. Thus, the tactics we have identified, in order of the passive to active participation they require of the student, are: explanation, summary, DLR, CI-Hint and PT- Hint.

After interviews with our tutors, we have concluded that they use a very coarse scheme to evaluate their students. Specifically, they maintain a local and global assessment. A local assessment is the tutor's assessment of how the student is responding during a short segment of the tutoring session. This is based on qualitative characteristics of the dialogue. As an example, a student who understands that an utterance is a hint can be evaluated positively by the tutor even if the student's answer is incorrect. The tutor's global assessment of the student is a measure of the student's behavior throughout the entire tutoring session.

We have observed that our tutors use the following rules to determine when to hint (Hume et al., 1996c):

  1. Initially try hinting when errors are made. The exception is when the global assessment is very low.
  2. If the global assessment is sufficiently high, try a second hint if the first hint is not successful.
  3. Continue to provide hints on a topic as long as:
    1. The global and local assessment are sufficiently high, and
    2. The number of hints while tutoring one topic is sufficiently low.
  4. If a follow up hint is to be provided then:
    1. Use a PT-Hint when the local assessment is high, and
    2. Use a CI-Hint when the local assessment is low.

A Tutor for a Pseudo Machine Code Interpreter

The domain of our proposed multimedia tutoring system is a pseudo computer machine code (Brookshear, 1994; Appendix B). The target students are college undergraduates with little or no background in computer science. The domain topic is one component in Valparaiso University's Computers and Computation course. This course satisfies a science requirement in the College of Arts and Sciences general education requirement. No computer science majors enroll in this course; occasionally a student becomes a computer science major after completion of this course. Most enrolled students do not have a strong background in quantitative manipulation.

Several class periods are spent on mathematical and computer related topics. Eventually, students are expected to trace the execution of a pseudo machine code. After completion of such traces, students are asked to state what values reside in various registers and memory locations. Most students need to be tutored to reach these goals. This is an ideal domain for multimedia tutoring because:

  1. There is a variety of required, isolated operations.
  2. Most operations require that students understand some background concepts.
  3. Many higher level tasks require that students have mastered multiple operations.
  4. Students make errors because:
    1. They are have not mastered an isolated operation,
    2. They have incorrectly applied an isolated operation, or
    3. They do not understand the sequence of operations required to carry out a higher level task.
  5. Isolated skills and higher level tasks can be presented as text or in a graphical format.
A human tutor in this environment can observe errors and sometimes, with a follow up interaction, determine the cause of errors. This is how human tutors, and many ITSs, decide what topic to tutor. We suspect that multimedia tutoring will provide greater range of tactics to be employed (more choices for how to tutor).

An Example

Machine code instructions are placed in memory. Memory and register addresses, instructions and data are all presented in hexadecimal (hex) format. Figure 1 is an example of a program.

Machine Address          Code
         00              1A
         01              A0
         02              1B
         03              A1
         04              5C
         05              AB
         06              3C
         07              A2
         08              C0
         09              00
         (later in memory)
         A0              FA
         A1              37
Figure 1. A sample pseudo language program.

Students are asked to trace such a program by emulating the fetch-decode-execute cycle of a computer chip. Every fetch brings four hex digits into the instruction register (IR): the op code, and fields F1, F2 and F3. The program counter (PC) points to the next instruction; the PC is incremented, by two, after the fetch. The op codes relevant for the program in Figure 1 are:

1 - Place value found in memory location F1F2 into register F3
3 - Place value found in register F1 into memory location F1F2.
5 - Add, in two's complement, the contents of registers F2 and F3. 
    Then, place result in register F1.
C - Halt execution.
The sequence of digits placed in the IR and the ensuing action from the program in Figure 1 is:

1AA0      Places FA (from location A0) into Register A.
1BA1      Places 37 (from location A1) into Register B
5CAB      Add (two's complement) FA and 37 
          (from Registers A, B). Places result, 31, in
          Register C.
3CA2      Places 31 (from Register C) into location A2
C000      Halt.
This small program illustrates the relationship between the underlying concepts, isolated operations and higher level tasks required of the student. For example, an often used isolated skill is the conversion of numbers between hex and binary format. The student also must understand the underlying concepts of place numbering systems (i.e., places represent the base raised to successive powers, the base determines the number of digits, any number can be represented in any base). Animation or traditional means (i.e., the scrolling of text) can be employed to tutor either the isolated operation or the underlying concepts.

An example of a higher level task is the '5' op code from the above example. Values are retrieved from registers A and B. The two numbers, interpreted as two's complement numbers, are added. The result, converted back to hex, is placed in a register C. This one higher level task requires that the student has mastered many isolated operations and understands many underlying concepts.

We are beginning to design a protocol for tutor-student interaction. It is important to derive a system where errors can be identified and assumptions about the cause of errors can be verified. We then hope, through pictures and animation, to simulate the hinting activity that is natural with human tutors. One possibility is an interactive multimedia environment where the student is prompted to complete pictures that illustrate a concept.

Multimedia Tutoring

An intelligent multimedia tutor for assisting students to solve the problem described here would have several features:

  1. It would incorporate a student model to be used in determining what to tutor and how to tutor;
  2. It would be able to select from a variety of tutoring tactics; and
  3. It would be able to interact with the student using information presented in a variety of ways.
Hypertext is likely to remain the principal means by which information to assist the student will be presented, although the links that are embedded may take the user to graphical or pictorial formats as well as text. Pictures, whether from real life (photographs) or drawings, have great potential to illustrate many of the important concepts that need to be understood if these problems are to be solved successfully. Animation offers a particularly powerful aid to the students' understanding of the data flow that occurs in the program being analyzed. Sound can be used to provide information to the user and may be particularly useful in conjunction with any kind of visual image.

How would a sophisticated tutor (human or ITS) use such multimedia resources? How would the tutor promote the kind of active learning that we are attempting to get CST to encourage? What is a visual hint? We can only suggest possible answers to these questions at this time. Much research will be needed to confirm the validity of such answers.

Both PT-Hints and CI-Hints convey some information to the user without actually providing the answer being sought. A visual hint might consist of an image (still or animated) delivered without accompanying labels or other textual information (whether printed to the screen or delivered by voice). The degree of cognitive activity required of the student can then be varied by providing labels and/or text in a graded fashion. A fully labeled and annotated figure would then represent the least active tutoring tactic (the equivalent of an explanation in CST).

An ITS able to tutor students in this manner could be used to address a number of significant issues. Does the simultaneous use of two sensory modalities assist in the acquisition of information? For example, does an auditory explanation accompanying a graphical image (auditory and vision) work better than a written explanation accompanying a graphical image (vision only)? Can iconic representations of processes or data flow paths serve to trigger the desired cognitive activity? Can an icon be a PT-Hint?

Conclusion

We are starting research in a multimedia ITS for teaching a pseudo machine language code because:

  1. We have found no literature on student modelling in multimedia education.
  2. Our attempts at determining how to tutor is limited to tactics available with text scrolling on a computer screen. Multimedia allows choices to be made between written text, voice and visual images.
  3. We wish to continue our study of hinting and extend this phenomenon to a multimedia tutoring environment. How does one hint in such an environment and is it effective in advancing student learning?
  4. We wish to compare our current findings with a different student population and a different subject domain. What effect does student age have on hinting? Does tutoring about qualitative reasoning differ from tutoring about quantitative manipulation?
  5. The pseudo machine code problems are simple enough to implement, yet rich enough to provide interesting tutoring situations. Concepts related to the pseudo machine code can be presented textually and graphically.

References

Brookshear, J. Glenn. (1994). Computer science: an overview. Benjamin/Cummings Publishing Co., Inc. Redwood City, CA.

Hume, G., Michael, J., Rovick, A., & Evens, M. (1993). The use of hints as a tutorial tactic. Proceedings of the 15th Annual Conference of the Cognitive Science Society (pp. 563-568). Boulder, CO. Hillsdale, NJ: Lawrence Erlbaum Associates, Inc.

Hume, G., Michael, J., Rovick, A., & Evens, M. (1995). Controlling active learning: how tutors decide when to generate hints. Proceedings of the 8th Florida Artificial Intelligence Research Symposium (pp. 157-161). Melbourne Beach, FL.

Hume, G., Michael, J., Rovick, A., & Evens, M. (1996a). Hinting as a tactic in one-on-one tutoring. The Journal of Learning Sciences (5(1), 23-47). Hillsdale, NJ.: Lawrence Erlbaum Associates, Inc.

Hume, G., Michael, J., Rovick, A., & Evens, M. (1996b). Student responses and follow up tactics in an ITS. To appear in the Proceedings of the 9th Florida Artificial Intelligence Research Symposium. Key West, FL.

Hume, G., Michael, J., Rovick, A., & Evens, M. (1996c) The use of hints by human and computer tutors: the consequences of the tutoring protocol. To appear in the Proceedings of the 2nd International Conference on the Learning Sciences. Evanston, IL.

Kim, N., Evens, M., Michael, J., & Rovick, A. (1989). CIRCSIM-TUTOR: An intelligent tutoring system for circulatory physiology. In H. Maurer (Ed), Computer Assisted Learning: Proceedings of the International Conference on Computer-Assisted Learning (pp. 254-266). Dallas, TX. Berlin: Springer-Verlag.

Li, J., Seu, J., Evens, M., Michael, J., & Rovick, A. (1992). Computer dialogue system (CDS): a system for capturing computer-mediated dialogue. Behavior Research Methods, Instruments, & Computers, 24, 535-540.