A Tutoring Rule Selection Method for Case-based e-Learning by
Multi-class Support Vector Machine
Daichi Hisakane, Minami Otsuki, Masaki Samejima and Norihisa Komoda
Graduate School of Information Science and Technology, Osaka University, 2-1, Yamadaoka, Suita, Osaka, Japan
Keywords:
Intelligent Tutoring System, Multi-class SVM(Support Vector Machine).
Abstract:
We develop an intelligent tutoring system on learners’ answers to problems that are dealt with in case-based
e-learning. A facilitator instantiates answers and tutoring advice as a tutoring rule preliminary, and the system
automatically identifies an appropriate instantiated answer which corresponds to the input sentence of an an-
swer from the learner. Although various kinds of tutoring rules are given on a certain problem, the instantiated
answers are very similar to each other among tutoring rules, even if tutoring rules are different. So the input
sentence is similar to the wrong instantiated answer of the tutoring rule, which makes it difficult to select
the tutoring rule correctly. The proposed method selects the tutoring rule for the input sentence by machine
learning of selecting the tutoring rules with the multi-class SVM(Support Vector Machine). The multi-class
SVM, consisting of multiple binary classifiers, can output various tutoring rules identified as corresponding to
one input sentence. In order to identify one correct tutoring rule, the proposed method introduces confidence
on each identification result and integrates the results. The proposed method improves accuracy of selecting
tutoring rules by 17% compared to the similarity-based selection method of tutoring rules.
1 INTRODUCTION
Case-based learning is an educational method
to equip people with knowledge to solve prob-
lems(Hoffmann and Ritchie, 1997). In the case-
based learning, an instructor selects a documented
case that is useful for developing learners’ problem-
solving skill. The learners read the documented case
and think what are problems and how to solve the
problems. When the learners can not answer solu-
tions to the problem correctly, the instructor does not
give the correct answer but gives a hint to derive the
correct answer in order to develop their thinking pro-
cess(Savery, 2006).
However, because the instructor is lacking for the
learner, the learners can not always receive advice
from the instructor, which decreases their learning ef-
fect. If the learner can get appropriate advice with-
out the instructor as well as the instructor gives, the
instructor’s workload can be reduced, and the learner
can learn from the case-based learning by themselves.
So, the intelligent tutoring system that gives advice
automatically is valuable for both the instructor and
the learner. Therefore, we propose an intelligent tu-
toring system that enables learners to be engaged in
case-based learning without the instructor.
Intelligent Tutoring Systems(ITS) have been pro-
posed to give advice to a learner(Corbett et al., 1997).
In order to show advice, ITS have several kinds of tu-
toring rules that are defines as IF-THEN rules. The
tutoring rules represent pairs of learners’ answer and
an instructor’s advice to the answer. ITS need to se-
lect the appropriate tutoring rules by analyzing the an-
swer from the learner based on the domain knowledge
that is related to the learning contents. In this paper,
we address tutoring rule selection for case-based e-
Learning.
The rest of the paper is organized as follows. Sec-
tion 2 describes the literature review of the intelli-
gent tutoring system. Section 3 outlines case-based e-
Learning systems and shows the research issues. Sec-
tion 4 describes the selection method of tutoring rules.
Section 5 shows the experimental results in applying
the proposed method to the real data from learners.
Section 6 deals with the conclusion derived from the
experimental results.
2 LITERATURE REVIEW
Learner support in e-Learning is important because
it influences learning outcomes(Aleven et al., 2003).
119
Hisakane D., Otsuki M., Samejima M. and Komoda N..
A Tutoring Rule Selection Method for Case-based e-Learning by Multi-class Support Vector Machine.
DOI: 10.5220/0005023501190125
In Proceedings of the International Conference on Knowledge Management and Information Sharing (KMIS-2014), pages 119-125
ISBN: 978-989-758-050-5
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
Online individual tutoring with human tutor via Inter-
net is the most effective support. However learners
always cannot receive such tutors’ support unless the
tutor stays for the learners. On the other hand, there
is also an e-Learning system that sends learner’s an-
swers to a tutor in a remote area(Chen and Hsu, 2005).
Checking the answers, the tutor gives the appropri-
ate advice for the answer. The system has functions
of displaying Q&A and pop up hint, and exchanging
messages with other learners for developing learning
ability.
However, a learner does not still obtain advice in
a moment and this system does not reduce a tutor’s
burden, because the human tutor must think and send
the advice for a learner’s answer. The system has not
solved the problems of the limitation of the tutors and
obtaining advice to the learner in a moment.
Currently many studies on Intelligent Tutoring
System (ITS) have been conducted. In ITS, a pro-
grammed tutor agent teaches learners instead of a hu-
man tutor. The tutor agent provides advice to learners
based on the learner model in order to provide ap-
propriate advice to each learner. The learner model
consists of learner properties, degree of understand-
ing and the degree of guesswork. ITS estimates the
learner model based on time taken to solve and which
answer is correct and so on(Sobue et al., 2004; San-
tos and Jorge, 2013). The learner property and the
degree of understanding are estimated by analyzing
the history of learning statistically. For example, the
degree of understanding can be calculated as a rate
of the correct answers to all the answers in case of
that the problem has clear and deterministic answers
as well as the multiple choice questions. In fact, e-
Learning systems target not only the problem having
deterministic answers but also the problems having
non-deterministic answers:
• Mathematics
ITS for mathematics have been developed in the
past(Heffernan and Koedinger, 2000; Virvou and
Sidiropoulos, 2013; Pholo and Ngwira, 2013)
which enables the dialogue as well as a human tu-
tor. By inputting questions with natural language,
the learner can understand mathematical problems
through dialogue with the system.
• Circuit design
There has been the system that accepts inputs of
free descriptions of circuits and gives advice on
the circuit design(Dzikovska et al., 2010). The
system simulates circuit action and displays hints
according to student’s input in circuit education.
As the above systems deal with non-deterministic
answers, the case-based e-Learning system also deals
with non-deterministic answers that are inputted as
natural language by the learners. However, generally
in case-based e-Learning, the learners input answers
from various aspects. So, our case-based e-Learning
system must equip the function to give advice for var-
ious answers.
3 CASE-BASED E-LEARNING
SYSTEM
3.1 Outline of the Case-based
E-Learning System
The GUI(Graphical User Interface) of this e-learning
system is shown in Figure 1. Case description is
shown on the upper-left text area in this window, and
the learner thinks what are the problems and how to
solve the problems from this case description. Then
the learner inputs the problems and the solutions into
the answer from at the bottom of the window. The ad-
vice corresponding to the answers is indicated in the
upper-right text area in the window. The instructor
of the case-based learning empirically judges advice
appropriately corresponding to the learner’s answer.
So we propose a rule-based system to give ap-
propriate advice based on the instructor’s experience.
The instructor instantiates answers and appropriate
advice to the answers. Based on the pair of the instan-
tiated answers and the appropriate advice, we make a
tutoring rule as the following IF-THEN rule: IF the
input sentence corresponds to the instantiated answer
THEN the system outputs the advice in the pair. For
an example in Figure 1, IF the input sentence corre-
sponds to “I don’t know.” THEN the system outputs
“What happens when Project Manager(PM) makes
members work late?”
!"#$%&$#'()*+,-!
.-#/$(%0,(1"!
"#$%&!'()!*+,-"!
,)($%&./!#+0!/(%-1(+/!
2$"(-$(!
3)"4,56$"
%&#)+&)2/!*+,-"!
#+0!#03*4&!
7$4$'8%"%'$44%"-&%
)-*68%"-#/$(!
.&9)'$:%;<"8%<"**$-#%
/<$-%=>%1"?$#%%
1$1@$(#%/,(?%4"8$A!
B-*68:%B%&,-C8%?-,/%/<"8%
#,46+,-#%=>%#<,64&%&,D!
Figure 1: GUI of case-based e-learning system.
Figure 2 is an outline of our proposed sys-
tem. The input of this system is an answer from a
learner(hereafter called input sentence), and the out-
put of this system is advice corresponding to the input
KMIS2014-InternationalConferenceonKnowledgeManagementandInformationSharing
120
sentence. The instructor preliminary registers tutor-
ing rules in a data base(hereafter called tutoring rule
DB). Due to various kinds of input sentences, the in-
structor sets multiple instantiated answers on each tu-
toring rule. For an example in Figure 2, the Rule B
has instantiated answers of “Delivery delay”, “Stag-
nant work” and so on. In addition, the instructor can
not instantiate answers that cover all the input sen-
tences. In order to give advice for the input sentences
that do not correspond to any instantiated answers, the
system collects such input sentences from history of
the learners’ inputs. And the collected past input sen-
tences are registered into the tutoring rule of ‘Others’
in order to identify the input sentences that do not cor-
respond to any tutoring rules.
The intelligent tutoring system judges whether the
input sentence corresponds to the instantiated answer
in each tutoring rule DB or not. Then the system se-
lects a tutoring rule which includes the instantiated
answer corresponding to the input sentence and shows
advice corresponding to the tutoring rule. If the input
sentence corresponds to the tutoring rule of ‘Others’,
the system shows a message to induce the learner to
change the input sentence by rephrasing the expres-
sion. The learner reads the shown advice and inputs
the answer repeatedly until the learner inputs a com-
pletely correct solution to the problem. The system
has to select the correct tutoring rule corresponding
to the input sentence to show appropriate advice to
learners.
!"#$"%!
!"#$%"&%'
("')%*+,-./0!
&'()%'*'$#+",)-%%"(.-$/"$)",+-,
,"0#*.1",#$(2"%,-3,%41"(,
!")+&-1%'-),+&%'(2'
3%*%&1%)'.$*%0!
4%5$%31'
.%6+"7$10!
!")+&-1%'8),+&%'
94$*%':";%*+,%./')%*-/'+3<=!
>%*%&1''
&(..%37(")%)'
'.$*%!
#$%&'(!
5$.4+,,
("$+"$)"!
5$+"11'6"$+,74+-%'$6,89(+"*!
:41",;)!?)&(..%37(")'1('+"31-"@-1%)'-"3A%.''
''''''''''''''''''''''BC%DE%.F3'E$.)%"F0''
'''''''''''''''GHIJ)!")+&-1%'-),+&%'B!2'1K%'E$.)%"<F'
:41",<)!?)&(..%37(")'1('+"31-"@-1%)'-"3A%.'
'''''''''''''''''''''''B;%*+,%./')%*-/L'>1-M"-"1'A(.NF0'
'''''''''''''''GHIJ)!")+&-1%'-),+&%'B;%*+,%./')%*-/<F'
=+>"%(
,,!?)&(..%37(")'1('BO1K%.3F'+"'1K%'7-31'3%"1%"&%'
''''''''''''''''GHIJ).%5$%31'.%6+"7$10!
G$1(.+"M'4$*%'
;:!
5$(+%4)+-%,
?-%%"(.-$/!
@'(+-%9,-3,
+>",.#(+,
1"#%$"%(!
:"."#+,'$.4+,
#$/,#/A'("B!
P(..%37(")'
'1('BO1K%.3F!
?#("!
Figure 2: Case-based e-learning system.
3.2 Research Issue
Similar input sentences tend to correspond to the
same advice in the tutoring rule. So a typical ap-
proach to judge the correspondence between 2 sen-
tences is to measure the similarity between the sen-
tences. Even if the sentences are the same in meaning,
different similar words are often used in the input sen-
tence and the instantiated answers. In order to judge
the similarity between the sentences, it is necessary
to judge the different similar words. For this prob-
lem, recognizing textual entailment (RTE) has been
developed(Mark et al., 2011). RTE-based method se-
lects the rule by the similarity between the input sen-
tence and each instantiated answer by identifying the
similar words with the conceptual dictionary Word-
Net(Jung and VanLehn, 2010).
Figure 3 shows the tutoring rule selection by simi-
larity and its issue. The similarity is calculated by Jac-
crad coefficient(Broder, 1997). In comparing a sen-
tence A to a sentence B, let a, b, c denote the number
of words in sentence A, the number of words in sen-
tence B, and the number of common words in both
sentences. The Jaccard coefficient is decided by the
following formula:
Jaccard coe f ficient =
c
a+ b− c
(1)
After the similarities to all the instantiated an-
swers are calculated, the tutoring rule that has most
similar instantiated answer is selected for advice.
However, the instantiated answers in different tu-
toring rules for the same problem are similar to each
other. As shown in Figure 3, the input sentence is sim-
ilar to not only the instantiated answer in the tutoring
rule corresponding to the input sentence but also one
incorresponding to the input sentence. Therefore, the
conventional method often selects the wrong tutoring
rule.
!"#$%"&%'("')*%'+(,-'
./'0"&,%120"3'
'
4%4.%,52'.$,6%"7'
89(,,%&)'1"2+%,!:$;%'<="
:$;%'>'
!"2)1"?1)%6'>"2+%,''
#
$<$,6%"'./'1660?("1;'+(,-7'
:$;%'<'
@A14B;%'>"2+%,'
$#C%;0D%,/'
6%;1/'./'2)13"1")'+(,-7'
E040;1,0)/'
F7G"
E040;1,0)/'
F7H"
%E%;%&)'
:$;%'>'
0"&(,,%&);/'
!"#"$%&'"()*%(+%*,-'%().,&"
I$)(,0"3'
:$;%'
C<"
/,%&(,&"
Figure 3: Selection of tutoring rule by similarity and an is-
sue.
4 SELECTION METHOD OF
TUTORING RULE
4.1 Outline of the Method
We focus on that the instantiated answers correctly
correspond to the correct advice in the tutoring rule.
So our proposed method learns how to select the tu-
toring rule based on the similarity between the instan-
tiated answers as supervised data. As a learning clas-
sifier, we use 1 vs 1 multi-class SVM(Support Vector
Machine) which indicates good performance of dis-
criminating various input sentences for the correct tu-
toring rule(Brunner et al., 2012). 1 vs 1 multi-class
SVM consists of 2-class SVM that are learned with
ATutoringRuleSelectionMethodforCase-basede-LearningbyMulti-classSupportVectorMachine
121
supervised data on 2 classes, and integrates the tu-
toring rules identified as corresponding to the input
sentence by each SVM.
In integrating the identified rules by all SVM, the
proposed method emphasizes on the identified tutor-
ing rule that is considered to correspond to the input
sentence more correctly. Furthermore, the method se-
lects the tutoring rule using the instantiated answer
which has similar confidence to the confidence of the
input sentence.
Figure 4 shows the outline of the tutoring rule
selection method. This method calculates the con-
fidence of the identified rule by discriminating in-
stantiated answers to each tutoring rule by applying
multi-class SVM to the input sentence and the instan-
tiated answer in the tutoring rule DB. When different
tutoring rules have similar instantiated answers each
other, the confidence of identified input sentence also
increases on both of tutoring rules. So our method
selects the tutoring rule corresponding to the instan-
tiated answer by identifying the instantiated answer
which has similar confidence to the confidence of the
input sentence.
!"#$%&
'("%(")(!
*+'),+-+".%(&,$/('&."0&
)./)$/.%(&)1"20(")(&
34&-$/56)/.''&789!
7(/()%&,$/(&1:&
(;.-#/(&."'<(,&<+%=&
'+-+/.,&)1"20(")(!
7(/()%(0&
,$/(!
>?@!
A?B!
A?>!
>?C!
>?D!
A?E!
F&
G&
H&
>?C!
>?D!
A?A!
I$/(&G! I$/(&F!
!"#"$%!
H1"20(")(&
J!"#$%&'("%(")(K&
H1"20(")(&
JL;.-#/(&."'<(,K&
"!
"!
#!
G!
F!
#!
#!
M$%#$%&#1''+3+/+%4&1:&)1,,()%&,$/(&
0+'),+-+".%(0&34&(.)=&789&.'&)1"20(")(?&
I$/(&G!
"!
"!
#!
G!
F!
#!
#!
"!
"!
$!
G!
F!
#!
#!
>?C!
>?D!
A?E!
F&
G&
H&
H1"20(")(&
J!"#$%&'("%(")(K&
N+O=&)1"20(")(&
+"&'1-(&,$/('!
&!
'!
(!
P$%1,+"O&I$/(&*G!
!"'%."5.%(0&
F"'<(,!
Figure 4: Outline of the method of selecting tutoring rule.
4.2 Confidence Estimation in
Discriminating Input Sentence
For the sake of calculating the confidence of the dis-
crimination with a criterion of the similarity, the pro-
posed method learns how to select the tutoring rule by
the similarity between words included in instantiated
answers.
In learning from instantiated answers, the similar-
ity between words in 2 instantiated answers is calcu-
lated with the conceptual dictionary WordNet(Bond
et al., 2012; Isahara et al., 2008). As Figure 5
shows, the WordNet consists of associated conceptual
groups ‘synset’. Each synset represents a concept,
and has higher and lower rank groups(Leacock and
Chodorow, 1998).The similarity between the word
w
m
and w
n
is calculated by the following formula:
sim(w
m
, w
n
) =
L
m
L
1
+
L
n
L
2
L
m
+ L
n
(2)
where L
1
and L
2
are the route length from the synset
of w
m
and w
n
to the upper synset where links from
synsets of w
m
and w
n
join together. And L
m
and L
n
are the route length from the synset of w
m
and w
n
to
the uppermost synset linked from w
m
and w
n
.
!"#$%#&&'!
"
(!
)*+&,-./0-1%$,!1%'!
#
(!
&2.2#!
3
4!
3
5!
3
%!
3
6!
#%!2+!
2#6)1"."+00
&2.2#
!
!
!
!
!
-1%$,!1%!
7&+%&#2!
Figure 5: Conceptual groups in WordNet.
The feature vector is defined as similarities be-
tween the words. Let F denote the feature vector of
the similarities s
ij
between words w
i
in the input sen-
tence and w
j
in instantiated answers in a tutoring rule.
F =
s
ij
∀i ∈ I, j ∈ J (3)
where I and J are a set of words in the input sen-
tence and a set of words in all the instantiated answers
in a tutoring rule. The dimensionality is reduced by
principal component analysis in order to use only the
characteristic similarities(Ian, 2005).
The distance from hyperplane on the multi-class
SVM is generally used as the confidence in discrim-
inating tutoring rules. Figure 6 shows the confidence
calculation of discriminating learners answer. SVMs
as binary classifiers learn with instantiated answers
for 2 kinds of tutoring rules that are selected from
the tutoring rule DB. This learning process is applied
to all combination of the tutoring rules, and 1 vs 1
multi-class SVM is created. Then the confidence is
estimated by classifying input sentences and instanti-
ated answers.
The bottom of Figure 6 shows the flow of dis-
criminating input sentence to each tutoring rule. Each
SVM is applied to the feature vector of the similarity
whose dimensionality is reduced by principal compo-
nent analysis. SVM has high confidence of discrim-
inated input sentence when the identified result is far
KMIS2014-InternationalConferenceonKnowledgeManagementandInformationSharing
122
away from the hyperplane. So the confidence is de-
fined as the distance from the hyperplane and the fea-
ture vector of the confidence is defined as the total
distance on SVMs that discriminates the input sen-
tence to the tutoring rule. For example, a classifier for
tutoring rules of A and B, and a classifier for tutor-
ing rules of A and C discriminate the input sentence
to A with confidence of 0.2 and 0.3. The total confi-
dence 0.5 is used to select tutoring rule A. The feature
vector of the confidence of all instantiated answers is
also calculated by the same way of calculating feature
vector of confidence of the input sentence.
!"#$%
&%'(%)!
!"#$%
&%'(%*!
!"#$%
)%'(%*!
!"#$%&%
!"#$%)%
!"#$%*%
!"#$%
&%'(%)!
!"#$%
&%'(%*!
!"#$%
)%'(%*!
+,-!
+,.!
/,0!
!"#$%&%
!"#$%)%
!"#$%*%
*1#2"#13$%
2'456$42$%
'7%$8$(9%("#$!
:$13"($%8$23'(%
'7%;<=<#1(<39!
:$13"($%8$23'(%
'7%2'456$42$!
*'42$>3"1#%
?<2@'41(9%
A$1(4%7('=%
<4;314@13$6%14;B$(!
C"#@D2#1;;%EFC%
G+%8;%+H%
*#1;;<79%<4>"3%;$43$42$%
146%<4;314@13$6%14;B$(!
*'42$>3"1#%
?$2@'41(9%
I"3'(<4J%!"#$%?)!
K4>"3%
;$43$42$!
?<;2(<=<413$%
&%7('=%)!
?<;2(<=<413$%
&%7('=%*!
?<;2(<=<413$%
)%7('=%*!
/,L!
/,0!
!
!
!
!
/,M!
*1#2"#13$%2'456$42$%'7%
<4;314@13$6%14;B$(%
<4%3N$%;1=$%B19,!
/,0!
/,-!
!
!
!
!
/,O!
+,L!
/,M!
/,+!
+,-!
+,.!
/,/!
"!
#!
#!
"!
$!
&!
)!
%!
#!
#!
%!
$!
&!
*!
"!
%!
%!
"!
$!
*!
)!
I"3'(<4J%!"#$%?)!
K4;314@13$6%
&4;B$(!
Figure 6: Confidence calculation of discriminating learner’s
answer.
4.3 Tutoring Rule Selection by
Classifying Confidence
The tutoring rule is selected by the similarity of fea-
ture vectors of the confidence between the input sen-
tence and the confidence of the instantiated answer.
Figure 7 shows the outline of selecting tutoring rules.
As Figure 7 shows, distance is small between the
feature vectors of confidences on the instantiated an-
swers that correspond to the same tutoring rule. The
feature vectors of instantiated answers consist clusters
that correspond to the tutoring rule.
Therefore, we can formulate the selection prob-
lem of tutoring rules as classifying the feature vec-
tor of input sentence to the cluster of the instantiated
answers. So our method selects the tutoring rule as
a major instantiated answer neighbor the input sen-
tence by k-NN method on the confidence(Duda and
Heart, 1973). K-NN method needs to set the number
of the neighborhood k. The number of the instantiated
answers is set based on the number of instantiated an-
swers r. The value of k is 1 at a minimum, and the
number of the instantiated answers r at a maximum.
Therefore, the range of k is set as follows:
1 ≤ k ≤ r (4)
Because it is possible to calculate the rate of the
number of instantiated answers assigned to the cor-
rect cluster of tutoring rule as precision, our method
selects k that indicates maximum precision rate within
1 ≤ k ≤ r. If some k have the maximum precision
rate, the largest k is used for selecting tutoring rules.
In Figure 7, the value of k is set as k = 3. Then the
tutoring rule is selected as a major instantiated answer
in k instantiated answers.
Figure 7: Tutoring rule selection by classifying confidence.
5 EXPERIMENT
5.1 Outline of the Experiment
In this experiment, we have developed the case-based
e-learning system and collected answers for a docu-
mented case about project management from 20 stu-
dents majoring in Information Science. The num-
ber of collected answers were 82 sentences. The
document of this case describes project manager’s
decision-making on the project delay due to the addi-
tional development required by the customer. The in-
structor preliminary set 4 problems shown in the fol-
lowing. The number of tutoring rules, instantiated an-
swers, and input sentences in each problem are shown
in Table 1.
• Problem 1. How to solve the problem of deal-
ing with change request from the customer with-
out manager’s authorization.
• Problem 2. How to solve the problem of re-
maining a difficult work later during the project
duration.
• Problem 3. How to solve the problem in dealing
with additional work.
• Problem 4. How to solve the problem to speed-
up developmentby working overtimeand on a day
off.
In the proposed method, the principal component
analysis for reducing dimension is implemented by
Weka (Hall et al., 2009). And multi-class SVM for
learning instantiated answer is implemented by SVM-
Light (Kanungo and Joachims, 1999).
ATutoringRuleSelectionMethodforCase-basede-LearningbyMulti-classSupportVectorMachine
123
Table 1: The number of tutoring rules, instantiated answers,
and input sentences.
Problem 1 Problem 2 Problem 3 Problem 4
Tutoring rules 4 5 5 7
instantiated answers 21 26 43 82
Input sentences 15 20 20 27
5.2 Evaluation of the Proposed Method
For evaluatingthe proposed method, we compared the
accuracy rate of the tutoring rule selections, by the
maximum similarity between input sentence and in-
stantiated answer, by the maximum confidence cal-
culated by multi-class SVM, and by the proposed
method. The accuracy rate is the rate of selecting
correct tutoring rules from input sentences. Together
with the learners, we judge whether the selected tu-
toring rule is correct or not. In applying the proposed
method to give advice to a certain learner, the pro-
posed method uses the answer from the other learner
as instantiated answers. Figure 8 shows the accuracy
rate of the whole problems and each problem by each
method.
The accuracy rate of tutoring rule selection by
maximum similarity was 63%(the number of correct
selection was 52), by confidence calculated by multi-
class SVM was 71%(the number of correct selection
was 58), and by the proposed method was 80%(the
number of correct selection was 66). The proposed
method improves the accuracy rate by 17%(the num-
ber of correct selection was 14) compared to the
method by the similarity.
!"#
$!"#
%!"#
&!"#
'!"#
(!"#
)!"#
*!"#
+!"#
,!"#
$!!"#
-./01# 23.4156#$# 23.4156#%# 23.4156#&# 23.4156#'#
708#9:6:103:/;# 708#<.=>?5=<5#4;#9@7# 23.A.B5?#65/C.?#
)&"!
(!"!
)*"!
*("!
)&"!
*$"!
)("!
')"!
+("!
*+"!
+!"!
*&"!
*("!
,!"!
+$"!
Accuracy rate(%)!
Figure 8: Accuracy rate.
According to Figure 8, the proposed method par-
ticularly improves the accuracy rate or Problem1 and
4 compared to the method of selecting tutoring rule by
similarity. In Problem 1 and 4, the proposed method
can select the correct advice from the input sentence
from which the selection method by similarity selects
the wrong tutoring rule because many input sentences
have high confidence in some tutoring rules. For ex-
ample, Problem 4 has 7 tutoring rules, and 2 of the 7
tutoring rules include the following words:
• Tutoring rule 1Stagnant work
• Tutoring rule 2Delivery delay
Although the input sentence ‘Stagnant work
caused by tired people fatigue leads to the delivery
delay.’ corresponds to Tutoring rule 2, it is judged to
correspond to Tutoring rule 1 by the similarity-based
selection method. The input sentence has a meaning
of the Tutoring rule 1 and 2, but the proposed method
correctly selects Tutoring rule 2.
6 CONCLUSION
In this paper, we proposed the intelligent tutoring sys-
tem that gives an appropriate advice corresponding to
the answer from a learner. This system needs to se-
lect the correct tutoring rule corresponding to the in-
put sentence from the tutoring rules which consists of
the instantiated rules and advices registered by the in-
structor.
Then we proposed the selection method of tutor-
ing rules by the confidence of discriminating learner’s
answer. The proposed method selects the tutoring rule
of the input sentence by multi-class SVM learning the
tutoring rule selection based on the similarity of in-
stantiated answer. The tutoring rule is selected by the
confidence as distance from the hyperplane of each
SVM. And, the proposed method integrated the dis-
criminant results by the confidence.
In the experiment, we adopt the case of project
management and got 82 sentences inputed by about
20 person who have experienced case-based Learn-
ing. The proposed method improves the accuracy rate
by 17% compared to the selecting method by the max-
imum similarity and by 9% compared to the selection
method by maximum confidence calculated by multi-
class SVM.
REFERENCES
Aleven, V., Stahl, E., Schworm, S., Fischer, F., and Wal-
lace, R. (2003). Help seeking and help design in inter-
active learning environments. Review of Educational
Research, 73(3):277–320.
Bond, F., Baldwin, T., Fothergill, R., and Uchimoto, K.
(2012). Japanese semcor: A sense-tagged corpus of
japanese. In Global WordNet Association.
Broder, A. Z. (1997). On the resemblance and containment
of documents. In Proccedings of Compression and
Complexity of Sequences, pages 21–29.
KMIS2014-InternationalConferenceonKnowledgeManagementandInformationSharing
124
Brunner, C., Fischer, A., Luig, K., and Thies, T. (2012).
Pairwise support vector machines and their applica-
tion to large scale problems. Journal of Machine
Learning Research, 13:2279–2292.
Chen, C.-M. and Hsu, S.-H. (2005). Applying interac-
tive mobile teaching agent to support e-learning plat-
form for learning performance promotion. In Proceed-
ings of the Fifth IEEE International Conference on
Advanced Learning Technologies, ICALT ’05, pages
205–206, Washington, DC, USA. IEEE Computer So-
ciety.
Corbett, A. T., Koedinger, K. R., and Anderson, J. R.
(1997). Handbook of humancomputer interaction.
North Holland.
Duda, R. O. and Heart, P. E. (1973). Pattern classification
and scene analysis. John Wiley & Sons.
Dzikovska, M. O., Bental, D., Moore, J. D., Steinhauser,
N. B., Campbell, G. E., Farrow, E., and Callaway,
C. B. (2010). Intelligent tutoring with natural lan-
guage support in the beetle ii system. Sustaining TEL:
From Innovation to Learning and Practice, pages
620–625.
Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann,
P., and Witten, I. H. (2009). The weka data min-
ing software: an update. ACM SIGKDD Explorations
Newsletter, 11(1):10–18.
Heffernan, N. T. and Koedinger, K. R. (2000). Intelligent
tutoring systems are missing the tutor: Building a
more strategic dialog-based tutor. In Proceedings of
the 2000 AAAI Fall Symposium on Building Dialogue
Systems for Tutorial Applications, pages 14–19.
Hoffmann, B. and Ritchie, D. (1997). Using multimedia to
overcome the problems with problem based learning.
Instructional Science, 25(2):97–115.
Ian, J. (2005). Principal component analysis. John Wiley
& Sons.
Isahara, H., Bond, F., Uchimoto, K., Utiyama, M., and
Kanzaki, K. (2008). Development of the japanese
wordnet. In Proceedings of the Sixth International
Conference on Language Resources and Evaluation
(LREC’08).
Jung, S. Y. and VanLehn, K. (2010). Developing an in-
telligent tutoring system using natural language for
knowledge representation. Lecture Notes in Computer
Science, 6095:355–358.
Kanungo, T. and Joachims, T. (1999). Advances in Kernel
Methods. MIT Press.
Leacock, C. and Chodorow, M. (1998). Combining local
context and wordnet similarity for word sense iden-
tification. WordNet: An electronic lexical database,
49(2):265–283.
Mark, S., Vydiswaran, V. V., and Roth, D. (2011). Multilin-
gual Natural Language Applications: From Theory to
Practice. IBM Press.
Pholo, D. and Ngwira, S. (2013). Integrating explicit
problem-solving teaching into activemath, an intelli-
gent tutoring system. In International Conference on
Adaptive Science and Technology (ICAST), pages 1–8.
Santos, G. S. and Jorge, J. (2013). Interoperable intelli-
gent tutoring systems as open educational resources.
The IEEE Transactions on Learning Technologies,
6(3):271–282.
Savery, J. R. (2006). Overview of problem-based learning:
Definitions and distinctions. Problem-based Learn-
ing, 1(1).
Sobue, Y., Matsumoto, T., and Onishi, N. (2004). Construc-
tion of the statistical student model using learning his-
tory on web-based learning system. IEICE Techni-
cal Report (Institute of Electronics, Information and
Communication Engineers), 103(697):89–94.
Virvou, M. and Sidiropoulos, S.-C. (2013). An intelligent
tutoring system over a social network for mathematics
learning. In Information, Intelligence, Systems and
Applications (IISA), pages 1–4.
ATutoringRuleSelectionMethodforCase-basede-LearningbyMulti-classSupportVectorMachine
125