SYSTEM DEMONSTRATION 
CONTENT PLANNING AS THE BASIS FOR AN 
• •INTELLIGENT•TUTORING SYSTEM 
Reva Freedman It, Stefan BrandlJ, Michael Glass z, 
Jung Hee Kim 2, Yujian Zhou 2 and Martha W. Evens 2 
iLRDC #819 
University of Pittsburgh 
3939 O'Hara Street 
Pittsburgh, PA 15260 
2Department of CSAM 
Illinois Institute of Technology 
10 W. 31st Street 236-SB 
Chicago, IL 60616 
freedrk+@pitt.edu, {brandle, glass}@charlie.cns.iit.edu, 
janice@steve.csam.iit.edu, Zhouyuj@charlie.cns.iit.edu, csevens@minna.cns.iit.edu 
http ://www. csam. lit. edu/~circsim 
1. INTRODUCTION 
The negative feedback loop which maintains a steady blood pressure in the human body is one of the 
more difficult topics for first-year medical students to master. CIRCSIM-Tutor v. 3 is the latest in a series 
of dialogue-based intelligent tutoring systems intended to help students master the concepts involved. 
CIRCSIM-Tutor v. 3 differs from many other ITSs in that text planning is an integral part of the system 
rather than part of a natural-language front-end. It contains a global planner whose fundamental goal is 
"generate a conversation resulting in the student knowing <concepts>" rather than "teach <concepts>." 
Constraints on the plan operators can be used to take a variety Of knowledge Sources into account, 
including the tutorial history, the domain knowledge base and a student model. 
To ensure that CIP, CSiM-Tutor is useful in the classroom, we have paid particular attention to broad 
coverage of the domain, maintenance of a coherent conversation and fast response time. We often need to 
model what human tutors do without a deep model of why they say what they say. As a result our content 
planner uses a schema-based.representation which allows us to control the decomposition and sequence 
of goals. Through the use of a reactive planning architecture, we can update our plan based on the 
student's answers. Text realization is accomplished via a high-level template mechanism based on a 
mini-syntax of potential answers. Botl! the content schemata and the text realization templates are based 
on detailed modeling of the dialogu e patterns of expert human tutors. 
2, DESCRIPTION OF DOMAIN AND USER INTERFACE 
CIRCSIM-Tut0r helps students practice the reasoning they have learned in Introduction to Physiology. 
Students are given a simplified qualitative model of the heart, followed by a series of problems which 
utilize the model. In each problem, an incident such as the administration of a drug affects the processing 
This work was supported by the Cognitive Science Program, Office of Naval Research under Grant No. N00014- 
94--1-0338 to Illinois Institute Of Technology. The content does not reflect the position or policy of the government 
and no official endorsement should be inferred. 
*This work was performed while Reva Freedman was at the Illinois Institute of Technology. 
I 
I 
r 
I 
! 
I 
,I 
i I 
I 
I 
280 
of file heart. The student is then asked to predict tile direction of change of seven core variables: 
HR: Heart rate (beats/min) MAP: 
IS: Inotropic state, a measure of the SV: 
heart's contractile ability " CVP: 
TPR: Total peripheral •resistance CO: 
The qualitative causal relationships between the core variables (i.e. an increase in X causes an 
Mean arterial pressure 
Stroke volume (vol/beat) 
Central venous pressure 
Cardiac output (vol/min) 
increase/decrease in Y) are shown on tlie left-hand side of Figure I. In this diagram, NS = nervous 
system and Baro = the baroreceptors in the neck which recognize a change in blood pressure. A finer- 
grained knowledge representation is also available for the tutor to usewhen needed. A section of this 
knowledge base, the leg from the nervous system to TPR, is shown on tlie right-hand side of Figure 1. 
Each variable must be Predicted at three points: the DR or direct response phase immediately after the 
incident, the RR or reflex response phase, which shows the effect of the nervous system, and the SS or 
steady state phase after a new steady state has emerged. After the predictions are made, • the tutor engages 
the student in a dialogue to help the student learn the correct answers and file concepts underlying them. 
The basic user interface is tile screen shown in Figure 2. A description of the current problem is shown at 
the top of the screen. The left-hand Side of the screen contains a table where the student can fill in 
predictions for tile three stages, and the right-hand side contains a window where the dialogue evolves. 
The student part of tile dialogue is free text. Instead of restricting what students can say, we attempt to 
guide them toward understandable responses through the way the questions are framed, e.g. by asking 
short-answer questions instead of open-ended ones. Input processing is based on finite-state transducers. 
Eachquestion the system can ask is associated with a transducer which can categorize and extract a 
variety of student answers. Often if the student just uses some of the right words, the sentence is accepted 
since the input processor uses very little syntax. Spelling correction, an essential in any system which 
allows free-text typing, is based on algorithms developed by Elmi \[1998\]. Tile transducers were coded by 
hand using as input both dialogues with human tutors and logs from earlier versions of CIRCSIM-Tutor. 
F 
1- ¸ 
F 
+ 
q 
•+ 
q 
I Nervous system I 
I Arteriolar muscle tone I 
\[ Arteriolar diameter ,I 
,L- 
Arteriolar resistance l 
++ 
I I 
Figure 1: Two aspects of the CIRCSIM-Tutor domain model 
281 
3. DIALOGUES GENERATED BYTHE SYSTEM 
At the top levels, the conversation generated by the system is hierarchical. Within each stage, the text is 
divided into segments, one for each •incorrect core variable. The variables are discussed in a partially 
ordered sequence which corresponds to the solution trace of the problem. 
Each variable is tutored using one of a number of tutoring methods which we have isolated from studies 
of human tutoring transcripts. The tutoring methods are implemented using an extended form of schema 
which allows full unification, static and dynamic preconditions, and recursion. The following schema is 
typical (schemata are implemented in Lisp): 
To correct student's ideas about any variable ?v controlled by the nervous system 
• Teach about mechanism of control of ?v 
Teach about when this mechanism is activated 
Check to find out whether student knows the correct answer now 
Circsim.Tutorv3.0 Help Quit Debug 
Problem: Pacemaker malfunctions, increasing to 120 beats/min. 
DR RR SS 
Central Venous Pressure 
Inotropic State 0 
Stroke Volume 
Heart Rate + 
Cardiac Output + 
Total Peripheral Resistance + 
Mean Arterial Pressure + 
T> Can you tell me what controls TPR? 
S> Maybe vasoconstriction? 
T> And what causes that? 
S> The nervous system. 
T> Right. And what stage are we in now? 
• Student notes go here. 
! 
i| 
Figure 2: User interface screen 
Each tutoring method is composed of a number of topics. Unless it includes a recursive call to another 
schema, each topic is instantiated using standard text generation primitives like elicit and inform. In 
addition to arguments specifying the content, the primitives can be modified with arguments specifying 
where the primitive falls on Halliday's interpersonal and narrative axes. Thus, for example, a sentence 
like Remember that we're in the pre-neural period could be generated from a form like <T-informs 
info=DR-info attitude=remind>. Optional arguments are also provided for generating several kinds of 
discourse markers and temporal clauses. 
Instead of planning the complete text as in a monologue, we interleave planning and execution, planning 
only as much as necessary to generate the next turn. When the student gives an unexpected response, 
which includes various kinds of "near-misses" as well as wrong answers, we can choose between retrying 
the current goal, adding a new g0al at the top of the agenda, or dropping the current schema and 
replacing it by another one. In this way We can reply flexibly to the student while still maintaining a long- 
282 
(1) Can you tell me how TPR is controlled? / What is the primary mechanism which controls TPR? 
Nervous system 
(2) 
Sympathetic Radius of I have 
vasoconstriction arterioles no idea 
TPR is 
(3) 
Right Right And what Which is 
~ \[ control;that? c22tUr2alllelYd c2entUr2alllelYd 
Nervous / / 4 
And we're in the pre-neural period now / 
Remember that we're in the pre-neural period 
I 
So what do you think about TPR now? 
• Figur e 3: Sample dialogues 
range plan. Each path through Figure 3 shows one piece of conversation which can occur as a result of 
the schema shown above. From left to right, the paths show a right answer, a couple of near-misses which 
require the use of the more detailed knowledge base, and a wrong answer. 

REFERENCES 
\[Elmi 1998\] Elmi, M. A. and M. W. Evens. 
1998. Spelling Correction Using Context. Proceedings of 
COLING-ACL '98, Montreal. 
