Content uploaded by Amruth Kumar
Author content
All content in this area was uploaded by Amruth Kumar
Content may be subject to copyright.
The Effect of Student Model on Learning
Adrian MARIES, Amruth KUMAR
Ramapo College of New Jersey
{amaries, amruth}@ramapo.edu
Abstract
Our goal in this study was to compare the
effectiveness of displaying the open student model as a
set of skillometers versus concept maps. The data
suggests that concept maps are significantly more
effective than a set of skillometers when answering
questions that require synthesizing an overview of the
topic.
1. Introduction
The constructivist theory of learning argues that we
gain knowledge by building upon already existing
knowledge, that we learn new concepts by integrating
them with concepts we already know. Constructing
concept maps makes students analyze the structure of
their own knowledge which helps them assimilate the
new information [9]. Several studies have shown that
concept maps help students learn. One study found that
creating concept maps helps students understand and
retain the material presented in class [10,11] Another
study found that studying the already worked-out
concept map was more effective than generating the
concept map from scratch, which was in turn more
effective than generating the concept map from just a
list of concepts or from a list that was already arranged
spatially [12].
Several studies have shown the ben efits of open
student models. Making the student model available to
students can make them aware of their own knowledge,
or lack of it, and, in turn, improve their learning [3,4].
One survey shows that students want to have access to
their learner models [1]. Student model presented in a
table format is difficult to understand [6]. Therefore,
visualization of data is a critical part of the open
student model [7]. Concept maps, with their ability to
reveal the relationships among concepts, have been
proposed as a mechanism to present the learner model
[1,2,5,8].
In our software tutors, we present the open student
model as a concept map. We wanted to find out
whether using the concept map to present the open
student model conferred any benefits over using other
techniques for visualizing the student model. In this
paper, we will describe an experiment that we
conducted to answer this question, analyze the
collected data and present our results.
2. The Tutor and the Protocol
We used a software tutor on arithmetic expressions
for our evaluation (www.problets.org). It presents
problems on evaluating arithmetic expressions in
C/C++/Java/C# (e.g., 5 + 4 % 8), and asks the student to
evaluate them step-by-step, i.e., one operator at a time.
Once the student has entered his/her answer, the tutor
provides delayed feedback – it lists how many steps the
student solved correctly. It displays the correct evaluation
of the expression using under-braces and intermediate
results. In addition, it prints a text explanation for each
step, such as “16 / 5 returns 3. Since both the operands
are integers, integer division is performed. Any fraction
in the result is discarded.” Since the student’s attempt is
displayed in the left panel and the correct evaluation is
displayed in the right panel simultaneously, the student
can compare the two solutions.
In fall 2006 and spring 2007, we used the arithmetic
expressions tutor to evaluate whether using the concept
map as the student model conferred any additional
benefits. The subjects were students in Psychology
courses who used the tutor over the web, on their own
time, as part of the experiential learning requirements
of a Psychology course.
We used a controlled study and the traditional pre-
test-practice-post-test protocol:
1. First, the subjects answered an online pre-test
consisting of 9 multiple-choice questions. 6 of
these were related to arithmetic expressions and 3
were unrelated.
2. The subjects worked with the software tutor for 15
minutes solving expr ession evaluation problems
and reading the feedback. After solving each
problem, the subjects were shown their open
student model. For the test group, the model was
shown as a taxonomic concept map (domain
concepts are nodes, links are is-a and part-of
relations, and the map is an and-or tree), with their
percentage completion of each concept graphically
displayed in the corresponding node (See Figure
1). For the control group, the model was shown
using skillometers (a series of progress bars
graphically showing the completion percentage of
each concept) (See Figure 2).
3. Finally, the subjects answered an online post-test
consisting of the same questions as on the pre-test.
The questions on the pre-test and post-test were:
1. How many arithmetic operators are available in
C++ programming language? 1/2/3/4/5/6. The
correct answer was 5.
2. Pick ALL the operators among the following that
are C++ arithmetic operators (check all that
apply): <, * , !, /, %, ^. The answer was *, / and %.
3. How many of the following numbers are prime: 2,
3, 4, 5, 6, 7, 8, 9, 10. The answer was 4.
4. Which of the following C++ operators have
'integer' and 'real' types (check all that apply)? -
<=, >, &&, +, /, ^. The answer was /.
5. For which of the following C++ operators is
'Dividing by Zero' an issue (check all that apply)?
- >=, ||, %, ^, !. The answer was %.
6. What is the sum of the internal angles of a
triangle? 90/180/270/360/450/600. The answer
was 180.
7. To how many C++ arithmetic operators does the
issue of 'Precedence' apply? – none/only
one/all/only two/only three. The answer was all.
8. Pick all the issues that apply to all the C++
arithmetic operators (check all that apply) –
Coercion, Correct evaluation, Error, Associativity,
Real. The answer was Correct evaluation and
Associativity.
9. Which of the following are types of operators in
C++ (check all that apply)? – Relational,
Abstraction, Repetition, Logical, Selection,
Assignment. The answer was Relational, Logical
and Assignment.
Note that almost all the questions had multiple correct
answering options and the student was asked to select
all those options. Questions 3, 6 and 9 are not related
to arithmetic expressions, and were meant to serve as
control questions for each subject. Answers to
questions 1,2,4,5,7,and 8 cannot be synthesized
without an overview of the domain of arithmetic
operators, since these questions are not about any
particular operator, but rath er about groups of
operators.
Durin g the problem-solving session, the tutor never
explicitly provided the answers to any of these
questions. But, the answers to questions 1,2,4,5,7,and 8
were evident from examination of the open student
models. The spatial organization of the concept map
made these answers more obvious than the list
organization of the skillometers. Take, for example,
question 2: Pick ALL the operators among the
following that are C++ arithmetic operators (check all
that apply): <, * , !, /, %, ^. A quick look at the open
student model displayed as a concept map, Figure 1,
will show that the five C++ operators are the children
of the root node, namely +, -, *, / and %. Even though
this information is available in the student model
displayed as skillometers as well, Figure 2, it is more
difficult to find. Not only do you have to look at most
topics in the skillometers, but the notation is not as
straightforward either.
4. Data Analysis
50 students participated in the evaluation, 32 were
in the test group and 18 in the control group. We used
two grading schemes:
1. In the regular grading scheme, if a problem had n
answering options, m of which were correct, the
student received 1/m points for each correct
answer. E.g, question 8 has 5 options, 2 of which
are correct; if the students’ answer includes one
Figure 1: Open Student Model Displayed as a Concept Map
incorrect option and the two correct ones, they
receive full credit.
2. In the negative grading scheme, students were
penalized for guessing. If a problem had n
answering options, m of which were correct, the
student received 1/m points for each correct
answer and lost 1/(n – m) points for each incorrect
answer. E.g, if the students have the same answer
as before to question 8, they only receive 2*1/2 –
1/3 = 0.66 points.
Aggregate of all the Questions. First, we considered
the aggregate of all the questions for each student. We
conducted a 2 X 2 mixed factor ANOVA analysis of
the aggregate scores with pre-post as the within-
subjects factor and treatment (skillometers versus
concept maps) as the between-subjects factor. Using
regular grading, we found a significant main effect for
time (pre versus post-test) [F(1,48) = 7.728, p = 0.034]
- students scored significantly higher on the post-test
(4.879 points) than on the pre-test (4.158 points).
There was no significant main effect for the treatment,
or significant interaction between pre-post and
treatment. We again found a significant main effect for
time using negative grading [F(1,48) = 17.417, p =
0.000] - students scored significantly more on the post-
test (3.993 points) than on the pre-test (2.822 points).
There was no significant interaction between pre-post
and treatment.
Related and Unrelated Questions. We did a 2 X 2 X
2 mixed factor ANOVA analysis with pre-post scores
and related-unrelated questions as within-subjects
factors and treatment (skillometer versus concept map)
as between-subjects factor. We found that:
• There was a significant main effect for related (ave
2.076) versus unrelated questions (ave 1.410)
[F(1,48) = 16.023, p = 0.000].
• There was a significant main effect for pre-test (ave
1.478) versus post-test (ave 2.008) [F(1,48) =
17.417, p = 0.000] -a clear pre-post increase.
• There was significant interaction between related
versus unrelated questions and pre versus post-test
[F(1,48) = 24.769, p = 0.000]. The average on
related questions increased from 1.525 on pre-test
to 2.627 on post-test for related questions and
decreased from 1.430 to 1.390 on unrelated
questions.
We did not observe any other significant interaction.
Related Questions. Next, we repeated the 2 X 2 mixed
factor ANOVA analysis for the aggregate scores on
related questions only. Once again, there was a
significant main effect for time using regular grading
[F(1,48) = 7.833, p = 0.007] – students scored
significantly more points (3.235) on the post-test than
on the pre-test (2.448). We found the same results with
negative grading [F(1,48) = 23.289, p = 0.000] –
students scored significantly more points (2.654) on
the post-test than on the pre-test (1.417). The
interaction between pre-post and treatment was
significant whether we used regular grading [F(1,48) =
3.925, p = 0.053] or negative grading [F(1,48) = 3.476,
p = 0.068].
Unrelated Questions. We did not find any significant
main effect for time using regular grading [F(1,48) =
0.401, p = 0.53] or negative grading [F(1,48) = 0.252,
p = 0.618].
Easy, Intermediate and Hard Related Questions.
Next, we repeated the above analysis for easy (1 and
2), intermediate (4 and 5) and hard (7 and 8) questions
considered together. The criterion we used to divide
the problems into the three categories was the
likelihood of finding the answer by taking a quick look
at the open student model. Question 1 (How many
arithmetic operators are available in C++ programming
language?), for example, can be answered fairly easily.
A quick look at the concept map will allow us to see
that there are 5 second-level nodes that have operator
names. To answer the hard questions, we need to
carefully inspect the student model. In order to find the
answer to question 7 (To how many C++ arithmetic
operators does the issue of ‘Precedence’ apply?), we
have to look at all third-level nodes and scan for nodes
with the specified name, a task that requires more than
just a quick glance at the concept map. Intermediate
questions are somewhere in between, harder than what
we call easy questions, but easier to answer than the
hard ones.
Figure 2: Open Student Model Displayed
as Skillometers
For easy questions, there was no main effect for
time or treatment and no significant interaction
between th e two.
For intermediate questions, there was a significant
main effect for time [F(1, 48) = 17.936, p = 0.000] –
the average score improved from 0.66 on the pre-test to
1.22 on the post-test. The effect for treatment was
marginally significant [F(1,48) = 3.0, p = 0.09] – the
test group scored significantly lower than the control
group on the pre-test (0.515 versus 0.941). The
interaction between time and treatment was not
significant.
For hard questions, there was no significant main
effect for time or treatment, but the interaction between
the two was significant [F(1,48) = 4.147, p = 0.047]:
whereas the control group score decreased from pre-
test to post-test (from 0.794 to 0.588), the test group
score increased from pre-test to post-test (0.646 to
0.894). We have summarized the pre-post change in
scores on the three types of questions, and the
statistical significance of the control-test group
difference in table 1.
Table 1: Analysis of Easy, Intermediate and
Hard Questions
Treatment Pre-post
Change Easy Inter
m Hard
Non-negative grading
Avg. 0.041 0.353
-
0.206
Skillomet
ers St. Dev. 0.644 0.786 0.751
Avg. 0.183 0.667 0.248 Concept
Map St. Dev. 0.774 0.816 0.743
Between-subjects p-value 0.495 0.196 0.051
Negative grading
Avg. 0.471 0.459
-
0.254
Skillomet
ers St. Dev. 0.736 0.783 0.523
Avg. 0.576 0.776 0.175 Concept
Map St. Dev. 0.868 0.717 0.672
Between-subjects p-value 0.657 0.174 0.017
Hard Questions. The analysis on questions 7 and 8,
taken separately, revealed the following results. For
questions 7, using regular grading, we did not find a
significant main effect for either time or treatment.
Similarly, there was no main effect for time and
treatment for questions 8, but the interaction between
the two was marginally significant [F(1, 48) = 2.988, p
= 0.090]. While the average for the control group on
this question dropped from 0.215 to 0.079, the average
for the test group increased from 0.147 to 0.171.
However, using negative grading, we found a
marginally significant difference between the control
and test groups on question 8, as shown in Table 2.
Table 2: Analysis of Questions 7 and 8
Treatment Pre-post
Change
Question
7
Question
8
Non-negative Grading
Avg. -0.118 -0.088
Skillometers St. Dev. 0.6 0.476
Avg. 0.152 0.096 Concept
Map St. Dev. 0.566 0.411
Between-subjects p-value 0.136 0.185
Negative Grading
Avg. -0.118 -0.136
Skillometers St. Dev. 0.6 0.27
Avg. 0.152 0.024 Concept
Map St. Dev. 0.566 0.329
Between-subjects p-value 0.136 0.073
Question 8 is one of the five multiple answer
questions the students had to answer. They are the only
type of questions on which guessing has a reasonable
chance of improving the score. The reason why there is
no significant difference between the two groups using
the grading scheme that doesn’t use negative grading is
due to the fact that guessing helps them find the correct
answer and, since it is a hard questions, many students
resorted to guessing. Choosing two correct options and
two incorrect ones (out of a total of two correct and
three incorrect options), for example, will give students
full credit using regular grading while only giving
them 0.33 points using the negative grading scheme.
5. Discussion
We found a significant improvement from pre-test
to post-test on questions that are related to arithmetic
expressions, whether regular grading or negative
grading was used. On the other hand, the scores on the
questions unrelated to arithmetic operators did not
improve with time. This means that the improvement
on the related questions did not occur by chance. It
shows that going through the tutor helped students
better answer the questions related to arithmetic
expressions.
We argue that the pre/post-test improvement is not
due to the problems the students solved working with
the tutor, but due to the open student model. Recall that
on most questions, students were asked to select all the
applicable options, e.g., all the arithmetic operators.
Given this overview nature of the questions, one needs
an overview of the topic to glean the correct answer.
Such an overview is provided by the open student
model, whether it is presented as a set of skillometers
or as a concept map. Concept map version of open
student model has the advantage over skillometers in
that it clarifies the is-a/part-of relationships among the
concepts. Alternatively, students may have constructed
an overview answer by choosing only the operators on
which they solved problems, but this explanation
contravenes Occam’s razor.
The difference between control and test groups
grows statistically more significant as we go from easy
to intermediate to hard questions for both grading
schemes (Table 1). Clearly, as the difficulty of the
questions increases, the gap between the ease of
answering the questions using the two different means
of displaying the student model widens. While it is
relatively straightforward to answer the easy questions
using either of the two ways of displaying the open
student model, it is harder to answer the hard questions
using skillometers than the concept map. In other
words, students are less likely to implicitly learn the
relationships among concepts using a set of
skillometers than using concept maps.
The difference between control and test groups on
question 8 is more significant with negative grading
than with regular grading (Table 2). Clearly, at least
some of the students guessed at least some of the
answers; some of these guesses were correct, and
others were incorrect. By penalizing guessing, negative
grading brought the differences between control and
test groups into sharper focus.
Our goal in this study was to compare the
effectiveness of displaying the open student model as a
set of skillometers versus concept maps. The data
suggests that concept maps are significantly more
effective than a set of skillometers when answering
questions that require synthesizing an overview of the
topic.
6. References
[1] Bull S., 2004. Supporting learning with open learner
models. Proceedings of fourth Hellenic Conference on
Information and Communication Technologies in Education,
Athens, Greece, pp. 47-61.
[2] Bull S., Mangat M., Mabbott A., Abu Issa A.S., Marsh J.,
2005. Reactions to inspectable learner models: seven year
olds to University students. Proceedings of Workshop on
Learner Modelling for Reflection, International Conference
on Artificial Intelligence in Education, pp. 1-10.
[3] Dimitrova, V., Brna, P., and Self, J. A. 2002. The Design
and Implementation of a Graphical Communication Medium
for Interactive Open Learner Modelling. In Proceedings of
the 6th International Conference on Intelligent Tutoring
Systems (June 02). S. A. Cerri, G. Gouardères, and F.
Paraguaçu, Eds. LNCS vol. 2363. Springer-Verlag, London,
pp. 432-441.
[4] Hauser, S., Nückles, M., and Renkl, A. 2006. Supporting
concept mapping for learning from text. In Proceedings of
the 7th international Conference on Learning Sciences
(Bloomington, Indiana, June 27 - July 01, 2006).
International Conference on Learning Sciences, pp. 243-249.
[5] Kay, J. 1997. Learner Know Thyself: Student Models to
Give Learner Control and Responsibility, in Z. Halim, T.
Ottomann & Z. Razak (eds), Proceedings of International
Conference on Computers in Education (AACE), pp.17-24.
[6] Mabbott A., Bull S., 2004. Alternative views on
knowledge: presentation of open learner models. Seventh
International Conference of Intelligent Tutoring Systems.
Springer, Berlin, Heidelberg, pp. 689-698.
[7] Marshall, B. B., Chen, H., Shen, R., and Fox, E. A. 2006.
Moving digital libraries into the student learning space: The
GetSmart experience. J. Educ. Resour. Comput. 6, 1 (Mar.
2006), 2.
[8] Marshall, B., Zhang, Y., Chen, H., Lally, A., Shen, R.,
Fox, E., and Cassel, L. N. 2003. Convergence of knowledge
management and E-learning: the GetSmart experience. In
Proceedings of the 3rd ACM/IEEE-CS Joint Conference on
Digital Libraries (Houston, Texas, May 03). pp. 135-146.
[9] Mazza, R. and Dimitrova, V. 2004. Visualising student
tracking data to support instructors in web-based distance
education. In Proceedings of the 13thIinternational World
Wide Web Conference, (New York, NY, USA, May 04).
WWW Alt. '04. ACM Press, New York, NY, pp. 154-161.
[10] Mazza R. and Milani C. Exploring Usage Analysis in
Learning Systems: Gaining Insights From Visualisations. In:
Workshop on Usage analysis in learning systems. 12th
International Conference on Artificial Intelligence in
Education (AIED 2005). Amsterdam, July 05. pp. 65--72.
[11] Willis, C. L. and Miertschin, S. L. 2005. Mind tools for
enhancing thinking and learning skills. In Proceedings of the
6th Conference on information Technology Education
(Newark, NJ, USA, Oct. 05). SIGITE '05. pp. 249-254.
[12] Zapata-Rivera and Greer, 2004. Interacting with
inspectable bayesian student models. International Journal of
AI in Education. vo1 4 (2). pp. 127-163.