AN EFFICIENT EDUCATIONAL MATERIAL EXPLORATION

USING EXTRACTED CONCEPTS

Tetsuro Takahashi

1,2

and Richard C. Larson

Fujitsu Laboratories LTD., 4-1-1 Kamikodanaka, Kawasaki 211-88, Japan

Sociotechnical Systems Research Center, Massachusetts Institute of Technology, Cambridge, Massachusetts U.S.A.

Keywords:

Educational Data Mining, Open Educational Resources, Concept Network, Natural Language Processing.

Abstract:

There are huge amount of Open Educational Resources (OER) provided by academic institutes now. Learners

have an advantaged environment in which they can use the educational materials for free in their learning.

General text search techniques, however, are not enough to provide an efﬁcient search function for them. This

drawback makes it difﬁcult for learners to fully exploit the promising OER. We propose to solve the problem

by providing an environment in which learners can search for educational materials efﬁciently using a set of

concept networks derived from the existing OER. The result of experiment with the actual OER shows the

availability of the proposed method.

1 INTRODUCTION

In the decade following the start of sensational edu-

cational project, Open Courseware (OCW) leaded by

Massachusetts Institute of Technology, many colleges

and universities (“institutes” here after) have started

to provide their educational resources for free on the

WWW (Carson, 2009). There are more than 15,000

courses offered in 20 languages through the Open

Courseware consortium

iTunes University is another major platform for

Open Educational Resources (OER). It is an Internet

service offered by Apple, Inc. that allows institutes to

share media ﬁles among students and faculty, as well

as with the general public (Reid, 2008). Currently

more than 350,000 educational materials are provided

by about 800 institutes.

In general, educational materials has the following

hierarchy and each lecture has several materials.

Institute → Subject → Course → Lecture,

Though the educational materials are provided in sev-

eral formats such as video, transcript, lecture note,

slide, exam and so on, we call them all “educational

material” in this paper. For instance, MIT OCW has a

subject “Mathematics” which has about 130 courses

such as “Single Variable Calculus” or “Linear Alge-

bra”. Each course has about 30 lectures in average,

and a lecture has several educational materials.

http://www.ocwconsortium.org

Learners have an advantaged environment in

which they can use the educational materials for free

in their learning. However, there are several difﬁcul-

ties to utilize the OER efﬁciently and effectively. We

discuss the difﬁculties in the next section.

2 DIFFICULTIES IN THE USE OF

OER

2.1 Cross-institute

Every institute has their own subjects and curricula

that the provided OER are following. While the OER

have rigidly structured within an institute, the struc-

tures are not universal to all institutes. For example,

a course “Single Variable Calculus” in an institute

may or may not comparable to a course “Calculus”

in another institute. Moreover it is not obvious which

subject should a course “Calculus for Business” be in

“Mathematics” or “Business”.

Learners need a set of abstracted common concept

rather than instances of existing subjects or curric-

ula for the sake of cross-institute utilization of OER

which are provided from various institutes.

2.2 Inﬂexibleness of Static Hierarchy

Most OER are structured in static hierarchies. A

282

Takahashi T. and C. Larson R..

AN EFFICIENT EDUCATIONAL MATERIAL EXPLORATION USING EXTRACTED CONCEPTS.

DOI: 10.5220/0003958702820287

In Proceedings of the 4th International Conference on Computer Supported Education (CSEDU-2012), pages 282-287

ISBN: 978-989-8565-06-8

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

learner who aim to complete a course may be able to

ﬁnd an appropriate educational material in the static

hierarchies. Learners’ objective, however, is not al-

ways to complete a course. It may take all sorts de-

pending on learners. while a learner needs to learn

whole of elementary calculus, another learner may

need only a part of calculus in order to understand

Gauss’s law in electromagnetism. The existing static

hierarchies are not ﬂexible enough to cope with the

various learning objectives.

2.3 A Limit of Text Search

There are huge amount of OER now. This makes

it difﬁcult for learners to ﬁnd an appropriate educa-

tional material. iTunes University provides a search

function. This function looks for educational mate-

rials which contain character string given as a query

in their subject name or course name. This surface

text search is too rough to look for materials regard-

ing learners’ precise search demand. And the results

tend to be a long list of candidates or empty (so called

zero-hit problem).

For instance, as the result of search with a query

“Algebra” iTunes University returns a long list con-

sisting of 162 courses which makes learners to con-

sume much time to traverse it. On the other hand,

the number of result is zero for a query “Document

Classiﬁcation”. Even if the current material collection

does not have a suitable course for this topic, some re-

lated materials could help learners who want to know

about that.

General text search engines return documents

which contains query string in them. In order to get

suitable result, learners have to know some words

or phrases that are supposed to be in desired educa-

tional materials, however, learners usually do not have

enough knowledge about a new learning object, and

it is a difﬁcult task to choose an appropriate word or

phrase as a query.

3 OER EXPLORATION USING

CONCEPT

We propose to introduce an abstracted expression

layer, “concept” which is common to all institutes

in order to solve difﬁculties described in Section 2.

The concepts enable the system to manage heteroge-

neous OER and to provide learners efﬁcient and ﬂex-

ible functions for OER exploration.

Figure 1 shows how the concepts work in the

structure. There are educational materials in the bot-

tom layer. Over the materials there are several layers

concept

225

concept

184

concept

material 55

material 21

material 32

material 40

material 53

material 29

material 83

material 10

material 2

material 19

step 1

step 2

step 3

step 4

step 5

Figure 1: OER management using concepts.

of concept network. Each material has some links to

relative concepts. Namely, materials have some con-

cepts as background of the materials, and concepts

have some materials which explain the concepts.

Using this structure, system can provide an ap-

proach which leads learners to educational materials

as the following steps.

1. Learners express their objective or concern by

keywords as a search query.

2. System shows (a) a layer of concept network

which contains a most related concept to the in-

put keywords and (b) the related materials to each

concept.

3. System updates the concept network as learners

update the keywords.

4. Learners grasp the target area graphically, and ﬁnd

interesting concept by browsing concept network.

5. System provides appropriate educational materi-

als for the concept.

The concepts are made in a bottom-up manner

by the algorithm described in Section 4 instead of a

top-down manner that the existing subject or courses

are deﬁned in. Because the bottom-up concept does

not depend on any speciﬁc structure (namely sub-

jects, courses or lectures), it can solve the difﬁculty of

cross-institute, and utilize any OER at the same time.

Since the desired granularity level of concept de-

pends on learners’ objectives, we introduced multi-

ple layers for the concept network. Each layer has

different granularity of concept which correspond to

learners’ various levels of objectives. This multiple

layers solves the problem of inﬂexibleness which the

existing hierarchies have, and provides learners an ef-

ﬁcient search environment. Learners can browse ab-

stracted structures of concept avoiding being buried

in a lot of educational materials.

ANEFFICIENTEDUCATIONALMATERIALEXPLORATIONUSINGEXTRACTEDCONCEPTS

283

4 CONCEPT EXTRACTION

4.1 Applying LDA to Educational

Materials

We used an algorithm Latent Dirichlet Allocation

(LDA) proposed by Blei (Blei et al., 2003) to extract

concepts from educational materials. LDA is a gener-

ative probabilistic model of a document set. The ba-

sic idea is that documents are represented as random

mixtures over latent topics, where each topic is char-

acterized by a distribution over words. The model can

be represented as Figure 2. In this model, a topic has

a word distribution φ, and a document has topic dis-

tribution θ. LDA gives a set of topics, θ and φ given a

document set.

topic 2topic 1 topic 3

document

w1 w2 w3 w4 w5 w6 w7 w_v.....

Figure 2: Latent Dirichlet Allocation (LDA).

The structure which LDA is assuming is quite

similar to our proposed one shown in Figure 1. Be-

cause LDA gives enough data to construct the struc-

ture, we chose it for our approach. In the context of

OER, the topic and document in LDA can be con-

sidered as concept and educational materials respec-

tively. Then by applying LDA, we can get the most

likely output described as follows for a given OER.

• a set of concept

• word distribution for each concept (φ)

• concept distribution for each material (θ)

Each concept has a word distribution (φ) which

supports the concept. The distribution enables ﬂexi-

ble matching between learners’ search query and the

concept. Since concept distribution θ reﬂects the

strength of relation between concepts and materials,

we can provide relevant materials for a speciﬁc con-

cept by using the distribution.

4.2 Layered Concept

LDA is a popular algorithm for the sake of clustering

as many researches applied it especially in a research

area of Natural Language Processing. The proposed

algorithm may be regarded as an application of LDA

to a text collection, however, we do not merely aim to

make clusters but the followings.

• to ﬁnd the most appropriate concept for a learners’

query

• to ﬁnd the most relevant educational material for

a concept

This motivated us to have multiple sets of clusters in-

stead of only one reasonable set of clusters. A set

of clusters forms one concept network layer shown

in Figure 1. We proposed to extract multiple layered

concept network that allow us to have granularity and

ﬂexibility in the matching.

We have to note that each layer does not have any

relation with other layers. The layers are made inde-

pendently, thus the layers shown in Figure 1 do not

form a hierarchy.

4.3 Visualization of Concept Network

Since an educational material written in text can be

represented as a word vector, the similarity between

educational materials can be deﬁned as the similarity

of the vectors. On the other hand, a concept have a

word distribution φ which can be used as a word vec-

tor, and the similarity between concepts can be de-

ﬁned as the similarity of the vectors as well.

Once the similarity between elements is deﬁned,

the set of elements can be visualized as a network

on two-dimensional space like Figure 1 using Force-

Directed model proposed by Fruchterman and Rein-

gold (Fruchterman and Reingold, 1991). Force-

Directed model assume a physical spring between two

nodes. And the position of nodes can be calculated by

looking for an equilibrium state in the simulation on

dynamic power modeling. We used the similarity de-

scribed above as strength of the pseudo spring.

The concepts derived by LDA do not have any la-

bel. But they have word distribution (φ) which can be

used to represent a tag cloud.

5 EXPERIMENTAL RESULT

5.1 Experiment

The data we used for our experiment were from two

resources, BLOSSOMS (Larson and Murray, 2008)

CSEDU2012-4thInternationalConferenceonComputerSupportedEducation

284

and MIT OCW. BLOSSOMS had 50 educational

video lectures in a variety of topics. Because all

videos have a transcript, they are suitable for text

processing such as LDA considerably. Though MIT

OCW had more than 2, 000 lectures, we used 1, 019

lectures from which we were able to obtain lecture

notes. The proposed algorithm was applied to the

transcripts for BLOSSOMS, and text data which were

extracted from lecture notes for MIT OCW respec-

tively. While OER are provided in various formats,

the proposed approach is applicable to them as long

as they have text.

After word regularization such as lemmatize and

elimination of stop-words, we applied LDA and de-

rived concepts. We used an algorithm proposed by

Grifﬁths (Grifﬁths and Steyvers, 2004) for the imple-

mentation of LDA. LDA requires several parameters

that we set as the following.

• hyper parameter α : 0.5

• hyper parameter β : 0.5

• number of iteration: 100

• number of concept: 10 to 50 step by 5

We used multiple values for the number of concepts as

proposed in Section 4.2. By this experiment setting,

we obtained nine different layers of concept which

were generated independently.

Figure 3 shows an example of output. The system

uses the output of LDA to provide search function and

to visualize concept network and material network.

The probabilistic word distribution φ works for scor-

ing of the search function. The score of a concept

can be formulated as the summation of φ value(s) for

query word(s).

The ﬁgure “1) Concept Network” in Figure 3

shows the result of the search. For the learner’s query,

”star”, system shows a layer that has 45 concepts in

which the concept number 20 is highlighted as the

most relevant concept to the query. The summary of

the concept number 20 is shown by a tag cloud below

the concept network. The learner can browse among

other concepts by moving mouse cursor over concept

nodes in this network. Browsing concepts, the learner

may ﬁnd a new keyword and add it to the query. “2)

Updated Concept Network” shows the result of up-

dating the query by adding “momentum”.

If the learner selects the concept number 30 by

clicking the node in this network, the system shows

a network of educational materials which related to

the selected concept as shown in the ﬁgure “3) Mate-

rial Network”. The size of nodes in this network re-

ﬂects the degree of relevance to the selected concept.

The learner can grasp a summary of an educational

material with a tag cloud by moving mouse cursor

over a node in the network as well as the concept net-

work. The system can also show a similarity network

in where a selected node and its similar nodes are rep-

resented in a network as shown in “4) Similarity Net-

work”. Browsing concept network, material network

and similarity network, the learner can grasp the tar-

get area and select a appropriate material to where the

learner can jump, and then start learning with a mate-

rial as shown in “5) Material Page”.

5.2 Evaluation

The result shows that materials from both BLOS-

SOMS and MIT OCW are used all together. The

ﬁgure “3) Material Network” in Figure 3 shows the

two lectures from different resources (“0315”: MIT

OCW lecture “Extrasolar Planets: Physics and Detec-

tion Techniques”, “22”: BLOSSOMS video “Galax-

ies and Dark Matter”) are represented in one network.

The proposed approach can incorporate any OER pro-

vided by different institutes and make concept net-

works which does not depend on any existing struc-

ture. On the concept networks, learners can look for

materials which are appropriate to their educational

demand. This result shows the algorithm works to

solve the ﬁrst and second problem described in Sec-

tion 2.

The proposed algorithm gives the interactive

search with abstracted concept level in which learn-

ers can grasp the target area and related ones. This

helps learners to avoid facing a long list of search

result. Learners can view OER gradationally from

bird’s-eye-view to the detail. In the interactive search,

the tag clouds over the concept network help learners

to ﬁnd unexpected keywords.

Theoretically, concept networks give extensional

expression which explains a target concept by the re-

lated materials and the neighbor concepts. The ex-

tensional expression works especially in a case that

learners are looking for materials to solve their own

problem. Because learners usually lack knowledge

about the target itself. The experimental result shows

that related materials and neighbor concepts can help

learners to understand the target. From this result,

the proposed approach solved the third difﬁculty de-

scribed in Section 2, limit of text search.

6 RELATED WORK

In this work, we aim to provide learners suitable edu-

cational materials that relevant to their demand. This

objective is the same as what many researches are

tackling in an area of information retrieval (Manning

ANEFFICIENTEDUCATIONALMATERIALEXPLORATIONUSINGEXTRACTEDCONCEPTS

285

1) Concept Network

3) Material Network

4) Similarity Network

2) Updated Concept Network

Figure 3: Exploration Steps.

et al., 2008). In most cases of information retrieval,

users select a desired content from a list of contents

which system generated for a given query. This work

ﬂow assumes that users have ability to select desired

one from the list. In the context of educational ma-

terial search, however, learners usually do not have

enough knowledge about the target area. Hence it is

difﬁcult to select suitable one for the learners. The

layered concept network which we proposed helps

learners to have a bird’s-eye view and to understand

the target area. By the interactive interface, learners

can ﬁnd not only what kind of concepts are there for

the given keywords, but also what are there as similar

or related concept, and what do they need to learn the

target concept.

Tudhope (Tudhope et al., 1995) and many works

tried to represent search results by visual interface

such as two-dimensional network instead of a simple

list. These techniques correspond to the representa-

tion of the bottom layer of Figure 1 in our work. In

addition to the representation, the proposed algorithm

has upper layers on the bottom one, and utilize them

for the search.

Brusilovsky (Brusilovsky and Rizzo, 2002) and

Simko (Simko and Bielikova, 2009) proposed to

make a concept network (they called this network

“concept map”) in order to help learners to understand

a target area and to ﬁnd educational materials. The

objective is the same to ours. While Simko made a

concept network by using a set of concepts deﬁned

by hand, we proposed to derive not only relations

but also a set of concepts itself from a bottom-up au-

tomatically. Brusilovsky derived both concepts and

relations by applying Self Organizing Map to OER.

CSEDU2012-4thInternationalConferenceonComputerSupportedEducation

286

While they used only one concept network, we used

multiple layered concept networks. Because the con-

cepts and educational materials have complex rela-

tion each other, a two-dimensional visualization is not

enough to represent the complete relation consistently

, especially, in a case that new educational materi-

als are being added incrementally. We, however, do

not have to take care the difﬁculty, because we use a

concept network only to guide learners to appropri-

ate educational materials. This means that we do not

have to draw a perfect concept network in our work.

We just have to represent subminimal relative rela-

tions for guiding learners. For this sake, the statistical

distribution produced by LDA works effectively and

robustly for various kinds of OER.

7 CONCLUSIONS

We proposed a method to guide learners to appropri-

ate educational materials efﬁciently using a layered

concept networks that is derived from existing OER.

Using the layered concept networks, we achieved the

objectives of this research, namely to provide learners

an environment in which they can understand the tar-

get area and explore educational materials efﬁciently.

The experimental result shows the proposed approach

works to solve difﬁculties in utilization of existing

OER. The advantages of the approach can help learn-

ers, however, the result did not show any quantitative

efﬁciency in searching that is one of important future

work.

The proposed algorithm classiﬁes materials into

multiple concepts. This classiﬁcation is based on

the types of concepts, however, it says nothing about

quality or difﬁculty of materials. Those information

must help learners to chose materials which are ap-

propriate for their demand. This should be added by

an algorithm to compliment the concept extraction us-

ing LDA.

We propose one directions which is worth for

future study, dependence extraction. The proposed

method uses only similarity relation between con-

cepts and materials deﬁned as the similarity between

vectors. The concepts and materials can have depen-

dence as the meaning of pre-condition. For example,

learners have to learn “addition” before “multiplica-

tion”. We call the relation dependence, and we can

say that multiplication depends on addition for the

above example. The dependence will help learners

We cannot represent even a small graph such as a com-

plete graph which consist of four vertex connected with

equivalent strength in two-dimension.

to ﬁnd their necessary concepts or educational mate-

rials. Because the number of combination between

concepts and materials is very huge, we have to make

an algorithm which works automatically.

REFERENCES

Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). La-

tent dirichlet allocation. J. Mach. Learn. Res., 3:993–

1022.

Brusilovsky, P. and Rizzo, R. (2002). Map-based horizontal

navigation in educational hypertext. In Proceedings of

the thirteenth ACM conference on Hypertext and hy-

permedia, HYPERTEXT ’02, pages 1–10, New York,

NY, USA. ACM.

Carson, S. (2009). The unwalled garden: growth of the

opencourseware consortium, 20012008. Open Learn-

ing: The Journal of Open and Distance Learning,

24(1):23 – 29.

Fruchterman, T. and Reingold, E. (1991). Graph drawing

by force directed placement. Software Practise and

Experience, 21(11):1129 – 1164.

Grifﬁths, T. L. and Steyvers, M. (2004). Finding scientiﬁc

topics. PNAS.

Larson, R. and Murray, M. (2008). Open educational re-

sources for blended learning in high schools: Over-

coming impediments in developing countries. Asyn-

chronous Learning Networks, 12(1).

Manning, C. D., Raghavan, P., and Schuetze, H. (2008). In-

troduction to Information Retrieval. Cambridge Uni-

versity Press.

Reid, A. (2008). Portable composition: itunes university

and networked pedagogies. Computers and Composi-

tion, 25(1):61 – 78.

Simko, M. and Bielikova, M. (2009). Automatic concept

relationships discovery for an adaptive e-course. In

Educational Data Mining, pages 171–178.

Tudhope, D., Taylor, C., and Benyon-Davies, P. (1995).

Navigation via similarity in hypermedia and informa-

tion retrieval. In HIM’95, pages 203–218.

ANEFFICIENTEDUCATIONALMATERIALEXPLORATIONUSINGEXTRACTEDCONCEPTS

287