EMBEDD-ER: EMBEDDing Educational Resources Using Linked Open

Data

Aymen A. Bazouzi

1 a

, Micka

el Foursov

1 b

, Ho

el Le Capitaine

2 c

and Zoltan Miklos

1 d

Univ. Rennes, CNRS, IRISA, France

Nantes Universit

e, LS2N, UMR 6004, F-44000 Nantes, France

Keywords:

Educational Resources, Embeddings, Linked Open Data.

Abstract:

There are a lot of educational resources publicly available online. Recommender systems and information re-

trieval engines can help learners and educators navigate in these resources. However, the available educational

resources differ in format, size, type, topics, etc. These differences complicate their use and manipulation

which raised the need for having a common representation for educational resources and texts in general.

Efforts have been made by the research community to create various techniques to homogeneously represent

these resources. Although these representations have achieved incredible results in many tasks, they seem

to be dependent on the writing style and not only on the content. Furthermore, they do not generate repre-

sentations that reﬂect a semantic representation of the content. In this work, we present a new task-agnostic

method (EMBEDD-ER) to generate representations for educational resources based on document annotation

and Linked Open Data (LOD). It creates representations that are focused on the content, compact, and can

be generalized to unseen resources without requiring extra training. The resulting representations encapsulate

the information found in the resources and project similar resources closer to one another than to non-similar

ones. Empirical tests have shown promising results both visually and in a subject classiﬁcation task.

1 INTRODUCTION

The use of technology has become a necessity nowa-

days, whether it is for accomplishing tasks that were

impossible to accomplish before or to facilitate and

automate other tasks. Researchers have tried to make

use of the recent technology breakthroughs and incor-

porate them in every ﬁeld. The ﬁeld of education is

no exception, as it was made clear early on that this

ﬁeld has the potential to yield great results with mean-

ingful impact on our lives. Technology can be used

in many ways to assist educators and learners in en-

hancing their teaching and learning experiences. This

gave birth to the term Technology-Enhanced Learning

(TEL) (Kirkwood and Price, 2014).

TEL can be used in a lot of ways and in close col-

laboration with many ﬁelds. On the one hand, AI and

data mining are among the most popular disciplines

in the research community. On the other hand, the

https://orcid.org/0009-0004-5209-6494

https://orcid.org/0009-0002-1048-8663

https://orcid.org/0000-0002-7399-0012

https://orcid.org/0000-0002-3701-6263

amount of educational content available is bigger than

what humans can go through manually. Therefore,

it quickly became apparent that education can ben-

eﬁt immensely from the use of such techniques and

methods to make the most out of the educational data

such as the different available Educational Resources

(ERs) (Romero and Ventura, 2013).

Usually, ERs are considered to be text documents

whether they contain text such as books, articles, and

presentation slides or they can be transformed into

text which is the case for the use of videos by us-

ing their transcripts. However, there are other differ-

ences that should also be taken into account such as

the differences in sizes, languages, topics, levels, and

concepts covered. This motivates the need for hav-

ing a common representation that allows systems to

process different ERs efﬁciently.

Let us assume that we have a set of ERs that dif-

fer in many ways (Figure 1). In order to be able to

manipulate these ERs, we need to have a common

representation for them. Otherwise, we will need to

treat different ERs in different ways following their

sizes, languages, formats, etc. One way to do that is

to project all the ERs into a latent space so that every

Bazouzi, A., Foursov, M., Le Capitaine, H. and Miklos, Z.

EMBEDD-ER: EMBEDDing Educational Resources Using Linked Open Data.

DOI: 10.5220/0012045300003470

In Proceedings of the 15th International Conference on Computer Supported Education (CSEDU 2023) - Volume 1, pages 439-446

ISBN: 978-989-758-641-5; ISSN: 2184-5026

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

439

ER has a vector representation, this process is known

as embedding. In Figure 1, we use similarity embed-

ding in which similar ERs are projected close to one

another

. Since ERs R

and R

are covering the same

subject, they are projected close to one another de-

spite the differences in size, format and author. As for

, it is projected further than the ﬁrst two as it covers

a completely different subject despite the similarity in

the release year, format and author with R

Figure 1: Embedding of educational resources.

Since ERs are considered to be text documents,

researchers showed great interest in techniques based

on document representation methods when manipu-

lating them. Despite the remarkable results achieved

using these methods, there are still some particular-

ities to be considered when applying them to ERs.

In this work, we are interested in mitigating two

problems that prevent the generic text representation

techniques from being directly used on ERs. First,

these models are mostly used in tasks dependent on

the writing style rather than the content (Mukherjee,

2021). Whereas for ERs, in most cases, the con-

tent is the most important part of the document. The

content can be seen as the core of the ER. We want

two ERs that cover the same concepts and discuss the

same subject to have representations that are similar.

However, for some of the usual methods, if an author

writes two articles about different subjects, the rep-

resentations generated for these two different articles

would still have some sort of similarity because these

representations take into considerations the entirety of

the text, the writing style included. Second, the state

of the art techniques seem to not generate representa-

tions that reﬂect a semantic representation of the text

We opted for a 2D representation for simplicity, the

commonly used embeddings have bigger dimensions.

(Merrill et al., 2021). However, in ERs, we would like

to have representations that are grounded on some sort

of semantic structure.

In this article, we present a novel method to gen-

erate task-agnostic embeddings for ERs using doc-

ument annotations and Linked Open Data. To ex-

plain our contribution, we start by discussing the re-

lated work and some preliminary notions in Section 2.

Then, we discuss the architecture of the model used in

Section 3. In Section 4, we conduct empirical tests on

our method and prove its efﬁciency as it achieves bet-

ter results than the state of the art methods. We con-

clude in Section 5 by summarizing the contribution

and discussing some perspectives for future work.

2 RELATED WORK

2.1 State of the Art

Data mining, and text mining more speciﬁcally, have

been widely studied in education (Ferreira-Mello

et al., 2019). Many techniques were used in educa-

tion such as text classiﬁcation (Lin et al., 2009), clus-

tering (Khribi et al., 2008) and many more. These

techniques were used on different texts related to ed-

ucation, namely ERs, forums, online assignment, etc.

Text mining is used in many tasks related to educa-

tion, it can be automatic evaluation of students (Cross-

ley et al., 2015), recommender systems (Drachsler

et al., 2015), program embedding learning (Cleuziou

and Flouvat, 2021), etc. However, (Ferreira-Mello

et al., 2019) reported that from the articles he ana-

lyzed in his survey, less than 3% used text mining

techniques on documents (or what we have been re-

ferring to as ERs).

Outside of education, document representation

techniques have been evolving at an incredible rate

(Liu et al., 2020). Education can beneﬁt from these

breakthroughs and use these techniques to represent

ERs. From these techniques, we mention the bag-of-

words model, which is a one-hot document encoding

method. TF-IDF is also widely used, it quantiﬁes im-

portance of the terms in a document. Topic Models

such as LDA (Blei et al., 2003) is another method,

but this technique identiﬁes a list of topics that a text

covers which is different from what we want to ac-

complish. Doc2Vec (Le and Mikolov, 2014), which

is a generalization of Word2Vec, is another technique

that is still being used. It is based on the use of the

skip-gram model. Most recently, the use of large lan-

guage models is getting more and more interest, es-

pecially the pretrained ones, such as BERT (Devlin

et al., 2019) which is a model based on a bidirectional

EKM 2023 - 6th Special Session on Educational Knowledge Management

440

deep transformer architecture.

The representations used for ERs and text docu-

ments play a major role in determining the quality of

the results returned for any technique, whether it is

classiﬁcation, clustering, etc. Having an ER repre-

sentation of quality is important when manipulating

it. However, sometimes we do not have an objective

function that can be optimized for the task at hand

to generate suitable representations like some recom-

mender systems. Furthermore, it can be interesting to

have a representation that is task-agnostic and can be

used as a building block for a more complex system.

Recommender systems in education have ex-

plored different ways to represent ERs (Urdaneta-

Ponte et al., 2021). Traditionally, recommender sys-

tems are divided into three categories (Adomavicius

and Tuzhilin, 2005). The ﬁrst one is content-based

(CB), the second is collaborative ﬁltering (CF) and

the third and ﬁnal one is hybrid recommender sys-

tems. Since we are interested in the ERs’ content, we

will be focusing in the next paragraph on CB and hy-

brid recommender systems in education as they are

the ones that focus on the ERs’ representations. Even

though we are focusing on recommender systems for

ERs and that our method can be used in such systems,

they are not the topic of this article as we are inter-

ested in ER representations in general.

To represent ERs in a recommender system,

(Broisin et al., 2010) have used WordNet-based key-

word processing combined with matrix decomposi-

tion. (Zhao et al., 2018) used a hybrid similarity mea-

sure based on link and object similarity where objects

(ERs) were similar based on a comparison of their

top-k TF-IDF terms. Other works such as (Lessa and

Brand

ao, 2018) have used pipelines that transform the

text into a vector by doing some preprocessing such

as stop word removal, stemming and other techniques

followed by the use of TF-IDF for weighting the ﬁl-

tered terms found in the text. (Ma et al., 2017) used

a pipeline composed of some of the most used tech-

niques in the text mining community (clean-tokenize-

TFIDF-Word2Vec-Doc2Vec) to create an embedding

for the courses that are used to make recommenda-

tions based on similarity. Other representations of

ERs have been created outside of recommender sys-

tems. For example, (Li et al., 2022) have used BERT

to create an embedding for ER descriptions and used

them in a grade prediction model.

All of these representations suffer from at least

one of the two aforementioned problems. For ex-

ample, BERT is sensitive to the writing style. Other

methods such as TF-IDF are not based on a seman-

tic grounded representation. These problems, make

the representations generated by these methods un-

suitable for some tasks that require a special focus on

the content such as recommender systems.

2.2 Preliminaries

Our contribution is in the form of a pipeline. In or-

der to better present it, we ﬁrst present two key con-

cepts necessary to understand the architecture chosen

for the pipeline.

• Wikiﬁcation: A semantic annotation technique

that uses Wikipedia as a source of possible se-

mantic annotations by disambiguating natural lan-

guage text and mapping mentions into canonical

entities also known as Wikipedia concepts (Hof-

fart et al., 2011).

• Knowledge base entity embedding: Knowledge

base entity embedding consists of training models

to learn embeddings for entities in a knowledge

base such as Wikipedia, DBPedia, Wikidata, etc

(Hu et al., 2015).

3 EMBEDD-ER

Given the speciﬁcities of the ERs, our goal is to cre-

ate a model that generates an embedding representa-

tive of the ER’s content in a more compact format. In

order to go from a text representation of an ER to an

embedding, we create a pipeline composed of three

components.

We start by taking the text of the ER and pass it

to a Wikiﬁcation tool known as Wikiﬁer (Brank et al.,

2017). Wikiﬁer takes a text document as input (Table

1a) and annotates it with links to relevant Wikipedia

concepts. As shown in Table 1b, this process is

achieved through the identiﬁcation of the different

terms found in the input text, disambiguating them

based on the context, linking them to a Wikipedia

page, then attributing an importance score to every

concept. To calculate this importance score, Wiki-

ﬁer

proceeds by creating a bipartite graph called the

mention-concept graph. This graph contains the men-

tions found in the text in one set and the Wikipedia

concepts in the other set. This graph is augmented by

adding links between concepts using a semantic relat-

edness measure. The pageRank algorithm is then ap-

plied to the resulting graph, and the pageRank scores

for the concepts are the importance scores for the con-

cepts returned (Brank et al., 2017). The result of this

step is a list of tuples composed of the concept found,

its Wikipedia page, and its pageRank score. This ﬁrst

step allows us to focus on the content and concepts

https://wikiﬁer.org/info.html

EMBEDD-ER: EMBEDDing Educational Resources Using Linked Open Data

441

Table 1: Generating embeddings with EMBEDD-ER on an ER text example.

(a) Input ER.

... This is really what

I mean by race being

a very unexpected un-

dercurrent in The Great

Gatsby. This really is a

completely unnecessary

detail. We’ll never see

that limousine ...

(b) Wikiﬁer results.

Term Wikipedia Link PageRank

Gatsby http://en.wiki.../Gatsby 0.0849

Race http://en.wiki.../Race 0.0328

Limousine http://en.wiki.../Limousine 0.0077

... ... ...

dings.

Embedding PageRank

= [0.1. . . 0.6] 0.0849

= [0.2. . . 0.6] 0.0328

= [0.2. . . 0.5] 0.0077

... ...

covered in the ER instead of the writing style and

other content-irrelevant details.

The result of the previous step is then used as

the input to the following component which is a

pretrained Knowledge base entity embedding model

named Wikipedia2Vec (Yamada et al., 2020). Wiki-

pedia2Vec is based on an extension of the skip-gram

model (Yamada et al., 2016). It is trained on the

paragraphs of the Wikipedia articles as well as a

Wikipedia graph in which the nodes are Wikipedia

entities and the edges represent the presence of hy-

perlinks between these entities, this Wikipedia graph

is a subgraph of DBPedia. The Wikipedia2Vec model

generates embeddings of size 300 for entities in

Wikipedia. The use of Wikipedia2Vec has many ad-

vantages over using a model that is trained on the

current data only, as TF-IDF. For instance, training a

model on a knowledge base, such as Wikipedia, gives

it a larger vision and a better representation of con-

cepts and their relations. If two concepts are syn-

onyms for example, Wikipedia2Vec would attribute

similar embeddings to these two concepts, this means

that if an ER uses more than one term to refer to the

same concept, Wikipedia2Vec will be able to cap-

ture the similarity. Furthermore, being trained on

Wikipedia’s articles and graph gives the model a bet-

ter semantic representation of subjects and concepts

rather than only being trained on paragraph structures

namely the popular models such as BERT that have

been trained on the tasks of tokens masking and next

sentence prediction (Devlin et al., 2019). We use

a Wikipedia2Vec model, pretrained on a 2018 snap-

shot of Wikipedia, on the results returned by Wiki-

ﬁer to generate embeddings for the Wikipedia enti-

ties found in the text. Table 1c shows how for every

Wikipedia concept, an embedding is associated. This

step ensures the embeddings generated are grounded

on Wikipedia, thus they vehicle a semantic meaning

for the text concepts and the document as a whole.

This simulates the notion of understanding the mean-

ing of a concept.

Finally, we know that not every concept men-

tioned has the same importance within an ER. There-

fore, we aggregate the concepts using a normalized

weighted sum for every concept’s embedding and its

pageRank value to give more weight to the more im-

portant entities. The ﬁnal result is an embedding for

the input ER that is representative of its content with

respect to a grounded understanding. To calculate an

embedding e for an ER that has n concepts each hav-

ing an embedding e

and a pageRank p

, we use the

following formula :

e =

∑

i=1

× p

)

∑

i=1

(1)

To illustrate how the aggregation works, we gen-

erated an embedding of an ER from a Yale course

(Figure 2). In this ER, the teacher discusses race in

the novel of The Great Gatsby by F. Scott Fitzgerald.

The embedding generated for the resource following

the equation 1 (in red) is closer to the key concepts of

the ER. However, the embedding generated by com-

puting an unweighted average (in orange) seems less

precise as it does not reﬂect the importance of the key

concepts in the result embedding. T-SNE was used

to reduce the dimensionality of the generated embed-

dings from 300 to 2 in order to plot the results.

Figure 2: Aggregating the entities’ embeddings.

EKM 2023 - 6th Special Session on Educational Knowledge Management

442

4 EXPERIMENTAL EVALUATION

In order to verify if these embeddings are truly repre-

sentative of these ERs, we decided to make two types

of tests. First, we study the similarity of ERs through

embeddings. Second, we use the embeddings gener-

ated in a multi-class subject classiﬁcation

The tests were conducted on a machine equipped

with a 64 bit Fedora Linux 35 OS with 16 GB of RAM

and an 11th Gen Intel i7-11850H @ 2.50GHz x 16

processor.

4.1 Data

In these past years, a lot of efforts were made to

make educational resources easily accessible and free

to use for both students and educators. Resources of

this type were named Open Educational Resources

(OERs) by UNESCO. In this work, we are interested

in using this type of ERs as they allow educators

and learners to retain, reuse, revise, remix, and redis-

tribute them.

We have used a dataset extracted from Open Yale

Courseware (OYC)

(Connes et al., 2021). This

dataset is formed in a hierarchical manner. It is com-

posed of series, within which we ﬁnd episodes, which

are divided into chapters themselves. A chapter is the

atomic building block of this dataset, it is represented

as a text that covers a speciﬁc subject. We consider

these chapters to be ERs. This dataset has in total

2550 chapters (ERs). However, we will not be using

all of them due to the class imbalance between sub-

jects for these ERs.

To measure similarity, we randomly pick four sub-

jects : African-American studies, Biomedical Engi-

neering, Economics, as well as Geology and geo-

physics. Then, from these subjects we chose 3 ERs

per subject. This is an arbitrary choice and has been

made to have a compromise between visual simplic-

ity and representativeness. We use similarity mea-

sures to compare these 12 ERs. We also use the ERs

belonging to the 3 most represented subjects in the

dataset and use them both to test similarity in a vi-

sual manner and in classiﬁcation tasks. The 3 subjects

are : Ecology and Evolutionary Biology, African-

American studies, and History. They have 407 ERs,

224 ERs, and 203 ERs respectively.

https://gitlab.inria.fr/abazouzi/embedd-er

https://oyc.yale.edu/

Figure 3: Similarity between embeddings generated by

EMBEDD-ER and BERT for 12 ERs using the l2-norm and

cosine similarity.

4.2 Results

4.2.1 Similarity

For testing the similarity, we use two measures that

are widely used in embeddings comparison : the l2-

norm and cosine similarity. We apply these mea-

sures to embeddings generated with EMBEDD-ER

and BERT for 12 ERs from the OYC dataset, every

3 ERs belong to a different subject. The results are

reported using matrix heat-maps.

In the similarity measures computed on the

EMBEDD-ER embeddings (Figure 3), we notice that

the resources belonging to the same subject are more

similar to one another than to ERs of different sub-

jects. This can be noticed by observing the 3 × 3

tiles formed along the diagonal axis of the similar-

ity matrices. However, for the embeddings generated

by BERT, these similarities are not as visible. This

means that the embeddings generated by our method,

project similar ERs close to one another and relatively

far from different ERs. This can be explained by the

fact that our method takes only the concepts which is

not the case for BERT.

To further investigate similarity, we performed di-

mensionality reduction on the ERs belonging to the

3 most represented subjects. We reduced the dimen-

sionality using T-SNE from d=300 for EMBEDD-ER

and d=768 for BERT to d=2.

From the results presented in Figure 4, we

can observe that the generated 2D coordinates for

EMBEDD-ER have formed homogeneous clusters, in

EMBEDD-ER: EMBEDDing Educational Resources Using Linked Open Data

443

Figure 4: T-SNE performed on the embeddings generated by BERT and EMBEDD-ER for ERs covering 3 different subjects.

which ERs covering the same subjects (in the same

color) are closer to one another than to other ERs.

However, the coordinates generated for the BERT em-

beddings are more dispersed with different ERs pro-

jected close to one another than to similar ERs in

some cases.

4.2.2 Classiﬁcation

To test our method, we used embeddings of the ERs

belonging to the three most present subjects in the

OYC dataset mentioned in Section 4.1.

For the baselines, we used the following models

that have been mentioned in the related work section:

• BERT: We used a pretrained BERT model to gen-

erate embeddings of size 768.

• TF-IDF: We applied TF-IDF on the ERs of the

three classes which generated embeddings of size

18508. An embedding of this size is too big.

Therefore, we applied T-SNE to reduce the size

of the embeddings to 700 so that it is close to the

size of the BERT embeddings.

• Doc2Vec: We created a pipeline that removes stop

words, does lemmatization then trains a Doc2Vec

model on the processed text. The generated em-

beddings have the same size as TF-IDF (700).

We used a 4-fold cross-validation grid search to

determine the best hyper-parameters for 4 classiﬁ-

cation models : Decision trees, SVM, Naive Bayes

(GNB), and KNN. The results reported represent the

mean accuracy of a 4-fold cross validation for these

models with the best hyper-parameters found with

grid search.

Table 2: Comparison of multi-class classiﬁcation accuracy

between TF-IDF, Doc2Vec, BERT, and EMBEDD-ER.

Method Decision

tree

SVM GNB KNN

TF-IDF 76.97%

±6.32%

65.81%

±8.38%

78.27%

±11.3%

85.23%

±9.18%

Doc2Vec 48.56%

±0.51%

48.8%

±0.17%

44.12%

±1.79%

44.61%

±2.91%

BERT 78.41%

±1.73%

94.12%

±3.13%

90.52%

±5.48%

92.44%

±5.61%

EMBEDD-

91.24%

±3.55%

98.2%

±0.71%

95.8%

±4.87%

97.36%

±1.1%

In multi-class classiﬁcation, EMBEDD-ER is the

best model followed by BERT (Table 2). These re-

sults highlight the fact that despite not having trained

EMBEDD-ER on classiﬁcation tasks, it still performs

better than the state of the art models used for docu-

ment representations while generating more compact

representations.

Table 3: Comparison of the time consumed (in seconds)

for the different steps of TF-IDF, Doc2Vec, BERT, and

EMBEDD-ER.

Method Pre-

processing

Model

Embedd-

ing

TF-IDF × × 635.80

Doc2Vec 4.15 × 705.28

BERT × 2.15 217.72

EMBEDD-

5695.89 517.19 142.32

Table 3 summarizes for each model the different

EKM 2023 - 6th Special Session on Educational Knowledge Management

444

steps and the time necessary. For EMBEDD-ER, it

takes a lot of time annotating the input ERs. However,

we notice that EMBEDD-ER and BERT generate

the embeddings faster than the TF-IDF and Docvec

with a slight advantage for EMBEDD-ER. Further-

emore, BERT is a pretrained model and EMBEDD-

ER uses Wikipedia2Vec which is also a pretrained

model which means that they need time to be loaded

into memory. Since EMBEDD-ER takes more time to

be loaded but less time to generate embeddings com-

pared to BERT, it means that for big preprocessed (an-

notated with Wikiﬁer) datasets EMBEDD-ER can be

faster than BERT.

4.2.3 Ablation study

Given that EMBEDD-ER is made by combining three

components, it can be interesting to alter its architec-

ture and check the performance of the different gen-

erated versions and study the impact of altering the

components. For that purpose, we have created three

more versions of EMBEDD-ER:

• BASE: The model presented in Section 3.

• BASE UW: Same as BASE, but for the aggrega-

tion is computed using an unweighted average :

e =

∑

i=1

(2)

• BERT-ER: Same as BASE, but we use BERT in-

stead of Wikipedia2Vec for generating the con-

cepts’ embeddings.

• BERT-ER UW: Same as the previous model

(BERT-ER) but we aggregate the embeddings us-

ing Equation 2’s formula.

We use the different architectures in a multi-class

classiﬁcation task. Using the same experimental setup

presented in the previous section. The results are re-

ported below:

Table 4: Ablation study on multi-class classiﬁcation accu-

racy.

Method Decision

tree

SVM GNB KNN

BASE 91.24%

±3.55%

98.20%

±0.71%

95.8%

±4.87%

97.36%

±1.1%

BASE

89.68%

±3.65%

98.44%

±0.71%

94.23%

±6.43%

97.12%

±0.59%

BERT-ER 81.05%

±5.06%

94.96%

±1.25%

74.1%

±1.51%

84.77%

±2.58%

BERT-ER

76.50%

±1.62%

96.29%

±1.48%

74.59%

±3.74%

79.62%

±2.44%

By observing the results of the multi-class clas-

siﬁcation reported in Table 4, we notice that BASE

outperforms all the models when used with Decision

trees, GNB and KNNs while BASE UW performs

better with SVM and gets the second best results with

the remaining classiﬁers.

As a general conclusion for the ablation study, the

most important part of the architecture is the use of

Wikipedia2Vec as the top two performing models use

it. The use of Wikipedia2Vec allows the embeddings

to vehicle semantic meaning and the use of a weighted

average adds more precision.

5 CONCLUSIONS

In this work, we presented a new method for

generating homogeneous, compact, writing-style-

independent, and content-focused representation for

ERs. This method is based on the use of a pipeline

that generates an embedding from an ER’s text by us-

ing document annotations and a knowledge base em-

bedding model pretrained on Wikipedia.

The embeddings generated by this method have

been tested in two different ways. The ﬁrst was a sim-

ilarity test, it showed that embeddings of ERs cover-

ing the same subjects are projected close to one an-

other in the embedding space. The second was a sub-

ject classiﬁcation task in which EMBEDD-ER per-

formed better than the state of the art text represen-

tations. This is due to EMBEDD-ER generating em-

beddings using the content while taking into account

the importance scores of the different concepts cov-

ered by the ER.

As for future works, we want to test EMBEDD-

ER on other datasets of different sizes, languages, and

subjects to test its generalization capabilities. It can

also be interesting to have a local implementation of

Wikiﬁer to avoid using the web API or even test other

tools for annotating the input text. This might reduce

the computation time for our method. We would also

like to use this method in real world tasks such as in-

corporating it in a recommender system, especially

content-based ones, or in an automatic curriculum de-

sign task.

ACKNOWLEDGEMENTS

This work has received a French government support

granted to the Labex Cominlabs excellence laboratory

and managed by the National Research Agency in the

“Investing for the Future” program under reference

ANR-10-LABX-07-01.

EMBEDD-ER: EMBEDDing Educational Resources Using Linked Open Data

445

REFERENCES

Adomavicius, G. and Tuzhilin, A. (2005). Toward the

next generation of recommender systems: A sur-

vey of the state-of-the-art and possible extensions.

IEEE transactions on knowledge and data engineer-

ing, 17(6):734–749.

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

Dirichlet Allocation. Journal of machine Learning re-

search, 3(Jan):993–1022.

Brank, J., Leban, G., and Grobelnik, M. (2017). Annotat-

ing documents with relevant Wikipedia concepts. Pro-

ceedings of SiKDD, 472.

Broisin, J., Brut, M., Butoianu, V., Sedes, F., and Vidal, P.

(2010). A personalized recommendation framework

based on cam and document annotations. Procedia

Computer Science, 1(2):2839–2848.

Cleuziou, G. and Flouvat, F. (2021). Learning student pro-

gram embeddings using abstract execution traces. In

14th International Conference on Educational Data

Mining, page 252–262.

Connes, V., de la Higuera, C., and Le Capitaine, H. (2021).

What should I learn next? ranking educational re-

sources. In 2021 IEEE 45th Annual Computers,

Software, and Applications Conference (COMPSAC),

page 109–114.

Crossley, S., Allen, L. K., Snow, E. L., and McNamara,

D. S. (2015). Pssst... textual features... there is more to

automatic essay scoring than just you! In Proceedings

of the Fifth International Conference on Learning An-

alytics And Knowledge, LAK ’15, page 203–207, New

York, NY, USA. Association for Computing Machin-

ery.

Devlin, J., Chang, M.-W., Lee, K., and Toutanova,

K. (2019). BERT: Pre-training of deep bidi-

rectional transformers for language understanding.

(arXiv:1810.04805). arXiv:1810.04805 [cs].

Drachsler, H., Verbert, K., Santos, O. C., and Manouselis,

N. (2015). Panorama of recommender systems to sup-

port learning, page 421–451. Springer.

Ferreira-Mello, R., Andr

e, M., Pinheiro, A., Costa, E., and

Romero, C. (2019). Text mining in education. WIREs

Data Mining and Knowledge Discovery, 9(6):e1332.

Hoffart, J., Yosef, M. A., Bordino, I., F

urstenau, H., Pinkal,

M., Spaniol, M., Taneva, B., Thater, S., and Weikum,

G. (2011). Robust disambiguation of named entities

in text. In Proceedings of the 2011 conference on em-

pirical methods in natural language processing, page

782–792.

Hu, Z., Huang, P., Deng, Y., Gao, Y., and Xing, E. (2015).

Entity hierarchy embedding. In Proceedings of the

53rd Annual Meeting of the Association for Computa-

tional Linguistics and the 7th International Joint Con-

ference on Natural Language Processing (Volume 1:

Long Papers), page 1292–1300.

Khribi, M. K., Jemni, M., and Nasraoui, O. (2008). Au-

tomatic recommendations for e-learning personaliza-

tion based on web usage mining techniques and in-

formation retrieval. In 2008 Eighth IEEE Interna-

tional Conference on Advanced Learning Technolo-

gies, page 241–245.

Kirkwood, A. and Price, L. (2014). Technology-enhanced

learning and teaching in higher education: what is ‘en-

hanced’ and how do we know? a critical literature re-

view. Learning, media and technology, 39(1):6–36.

Le, Q. and Mikolov, T. (2014). Distributed representations

of sentences and documents. In Proceedings of the

31st International Conference on Machine Learning,

page 1188–1196. PMLR.

Lessa, L. F. and Brand

ao, W. C. (2018). Filtering graduate

courses based on LinkedIn proﬁles. In Proceedings of

the 24th Brazilian Symposium on Multimedia and the

Web, page 141–147.

Li, J., Supraja, S., Qiu, W., and Khong, A. W. (2022). Grade

prediction via prior grades and text mining on course

descriptions: Course outlines and intended learning

outcomes. International Educational Data Mining So-

ciety.

Lin, F.-R., Hsieh, L.-S., and Chuang, F.-T. (2009). Dis-

covering genres of online discussion threads via text

mining. Computers & Education, 52(2):481–495.

Liu, Z., Lin, Y., and Sun, M. (2020). Representation learn-

ing for natural language processing. Springer Nature.

Ma, H., Wang, X., Hou, J., and Lu, Y. (2017). Course rec-

ommendation based on semantic similarity analysis.

In 2017 3rd IEEE International Conference on Con-

trol Science and Systems Engineering (ICCSSE), page

638–641. IEEE.

Merrill, W., Goldberg, Y., Schwartz, R., and Smith, N. A.

(2021). Provable limitations of acquiring meaning

from ungrounded form: What will future language

models understand? Transactions of the Association

for Computational Linguistics, 9:1047–1060.

Mukherjee, S. (2021). Sentiment Analysis, page 113–127.

Apress, Berkeley, CA.

Romero, C. and Ventura, S. (2013). Data mining in educa-

tion. WIREs Data Mining and Knowledge Discovery,

3(1):12–27.

Urdaneta-Ponte, M. C., Mendez-Zorrilla, A., and

Oleagordia-Ruiz, I. (2021). Recommendation

systems for education: Systematic review. Electron-

ics, 10(1414):1611.

Yamada, I., Asai, A., Sakuma, J., Shindo, H., Takeda,

H., Takefuji, Y., and Matsumoto, Y. (2020).

Wikipedia2Vec: An Efﬁcient Toolkit for Learn-

ing and Visualizing the Embeddings of Words

and Entities from Wikipedia. (arXiv:1812.06280).

arXiv:1812.06280 [cs].

Yamada, I., Shindo, H., Takeda, H., and Takefuji, Y. (2016).

Joint learning of the embedding of words and enti-

ties for named entity disambiguation. arXiv preprint

arXiv:1601.01343.

Zhao, Q., Wang, C., Wang, P., Zhou, M., and Jiang, C.

(2018). A novel method on information recommenda-

tion via hybrid similarity. IEEE Transactions on Sys-

tems, Man, and Cybernetics: Systems, 48(3):448–459.

EKM 2023 - 6th Special Session on Educational Knowledge Management

446