Fact-in-a-Box: Hiding Educational Facts in Short Stories for Implicit

Learning

Alia El Bolock

1 a

, Caroline Sabty

, Nour Eldin Awad

and Slim Abdennadher

1,2

Informatics and Computer Science, German International University, Cairo, Egypt

Media Engineering and Technology, German University in Cairo, Cairo, Egypt

Keywords:

NLG, Education, Implicit Learning, User Study.

Abstract:

Generating stories on-demand is one of the covered tasks in Natural Language Generation. Stories are being

used in every culture by any age. They have been used for different purposes, such as entertainment and the

education of children. They are an effective way of indirectly providing students with valuable facts that are

easier embedded in their memory. We propose an approach for embedding facts into existing or automatically

generated stories, given a speciﬁc target audience and a story context. As proof of concept, we implemented

a framework called Fact-in-a-Box to hide facts in existing stories or a human-like generated text through a

customized user. The framework is based on a ﬁne-tuned model for children as the target audience and ﬁxes

the story context to animals. Instructors can apply the approach to deliver facts to the learner in an exciting yet

informative way. The framework is composed of two modules, one for selecting the most relevant story and

the other one to embed the fact in it. We tested the proposed approach using an experiment to test the learning

gain of children and a survey for adults to evaluate the language of the resulting stories and the concept itself.

The performance was relatively good in hiding facts inside an existing story where children could correctly

re-convey 50% of the complex facts and 80% of the simpler tasks.

1 INTRODUCTION

Natural Language Generation (NLG) covers a wide

range of diverse tasks, including generating stories

on-demand, providing endless possibilities for enter-

tainment and education. Intelligent tutoring systems

can provide students with instant feedback, increasing

education quality. Stories are an effective way of indi-

rectly providing students with valuable facts that are

easier embedded in their memory. They are used in

classrooms to promote critical thinking and enhance

learning (Alhussain and Azmi, 2021). NLG can be a

more cost-effective tool than hiring authors to write

informative yet interesting tales for kids.

We propose an approach for embedding pre-

selected facts into existing or automatically gener-

ated stories, given a speciﬁc target audience and a

story context. Our main goal was to be able to im-

plicitly hide and deliver knowledge or facts with-

out the conscious awareness of the target audience.

Pre-trained models like Generative Pre-trained Trans-

former (GPT) 2 (Radford et al., 2019), and 3 (Brown

et al., 2020) are used to ﬁne-tune the involved down-

https://orcid.org/0000-0002-5841-1692

stream NLG tasks and obtain accurate results based

on the intended context. We rely on pre-trained

transformer-based models to overcome the shortcom-

ings of Seq2Seq, RNN, and LSTM for maintaining

the coherence and story ﬂow. As proof of concept, we

implement a ﬁne-tuned model for children as the tar-

get audience and ﬁx the story context to animals. Fur-

ther modiﬁcations were done to enhance the model to

make conditional text predictions based on the user’s

selection. As proof of concept, we implement our ap-

proach into a framework called Fact-in-a-Box, to hide

facts in existing stories or a human-like generated text

through a customized user interface based on the tar-

get audience and the speciﬁed context. The frame-

work contains two interaction modes one for the edu-

cator responsible for content creation and one for the

student for receiving implicit learning.

To embed a fact in a story, the framework is di-

vided into two modules. The ﬁrst module is a Base

Story Selector; it contains existing stories and can

generate new stories. The latter task is not the focus

of this presented paper. It takes as input the fact from

the user and selects from the available/generated sto-

ries the most relevant to it. The selection approach is

330

El Bolock, A., Sabty, C., Awad, N. and Abdennadher, S.

Fact-in-a-Box: Hiding Educational Facts in Short Stories for Implicit Learning.

DOI: 10.5220/0011985100003470

In Proceedings of the 15th International Conference on Computer Supported Education (CSEDU 2023) - Volume 1, pages 330-336

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)

applied using TF-ID (Salton, 1983), and Cosine Sim-

ilarity (Pang-Ning et al., 2005) methods. The second

module is the Fact Embedder, which selects the best

position for adding the fact in the selected story. Co-

sine Similarity is used to calculate the similarity be-

tween the story sentences and the fact.

Given the fact list and their intended context as

input, the framework produces the output story text

with the information embedded. This work demon-

strates how beneﬁcial combining NLG and fact em-

bedding may be in the education domain. The pro-

posed project goes beyond animal-based stories di-

rected at children and expands to educational and

training systems. Instructors can apply the approach

to deliver facts to the learner in an exciting yet infor-

mative way. We tested the proposed approach in two

phases: 1) an experiment to test the learning gain of

5 children and 2) a survey for 40 adults to evaluate

the language of the resulting stories and the concept

itself. The proposed framework performance was rel-

atively good in hiding facts inside an existing story

where children could correctly re-convey 50% of the

complex facts and 80% of the simpler tasks. The pro-

posed approach can be expanded to a broader domain

by using more complex stories than children’s stories

to reach a broader audience.

Several works are done regarding story generation

that can be found in (Alhussain and Azmi, 2021) On

the other hand, there are lots of approaches for uncon-

ventional learning mediums to accompany traditional

face-to-face classes and computer-assisted learning

(Akturk, 2022). These include gamiﬁcation (Nadi-

Ravandi and Batooli, 2022), online learning (Mas-

tan et al., 2022), immersive learning (Bizami et al.,

2022), and indirect learning. Indirect learning is

achieved when knowledge is acquired by watching a

movie, reading a book, or doing a seemingly daily

life activity. While indirect learning can happen un-

intentionally, it can be harnessed by developing ed-

ucational systems that intentionally convey implicit

knowledge through the chosen mediums. Different

use-cases for this implicit learning have been inves-

tigated. For example, movies are used for STEM

education (Kangas et al., 2017), language education

(Obloberdiyevna and Odilkhonovna, 2022), soft skills

(Belda-Medina, 2022), and physical education (Fu

et al., 2022), to name a few. (Faidley, 2021) ex-

plores education through different pop culture medi-

ums, e.g., movies, tv shows, and memes. Storytelling

has always been used as a tool for teaching language

and morals by both caregivers and educators (Abdal-

rahman, 2022; Purnama et al., 2022; Nicolaou, 2023;

Ratih et al., 2022; Hofman-Bergholm, 2023; Quah

and Ng, 2022).

Figure 1: Overview of the Fact-in-a-Box Story Generation.

In this work, we focus on providing a generic tool

for automatically generating indirect learning mate-

rial to intentionally convey information to learners in

an implicit manner in the form of short stories.

2 FACT-IN-A-BOX OVERVIEW

Hiding information inside an existing or generated

text is challenging as many variables should be taken

care of otherwise, the text would not make any

sense. Our proposed framework enables the prob-

lem of embedding words or hiding information within

an already-existing story or a newly generated story.

Fig. 1 gives an overview of the proposed architecture.

Given the chosen input facts to be learned, a story is

selected. The facts are then embedded in an appro-

priate location inside the story based on two scoring

criteria. We will go into the details of each of the two

main modules in the following.

2.1 Base Story Selector

There are two approaches to choosing a story to use

as a basis for embedding the facts. We can either gen-

erate a story from scratch based on the input facts or

choose existing stories to embed the facts. The for-

mer is a pure NLG approach requiring heavy com-

putational power and story tuning. The latter bene-

ﬁts from relying on existing stories adhering to the

Fact-in-a-Box: Hiding Educational Facts in Short Stories for Implicit Learning

331

plot generation rules of stories. When depending on

existing stories, the challenge is choosing a suitable

story to match the facts. While the authors also inves-

tigate generating stories using GPT-2, in the presented

work, we focus on choosing a ﬁtting story from a

database of existing stories. We categorize the global

dataset into smaller topically-clustered and difﬁculty-

clustered datasets. We have different datasets for sto-

ries about animals, crimes, fairytales, etc., and other

datasets for young children’s stories, teenagers, and

adults

. We rely on two metrics to choose a match-

ing story based on the input prompt from the story

database: TF-IDF and cosine similarity.

Before applying the TF-IDF technique, we

cleaned the text by removing extra white spaces, con-

verting all the text into lowercase, removing digits,

and removing stop words. Lastly, we added the stem-

ming of the words, which is the process of converting

similar words into a single word, e.g., running and

runs would be converted to run. This improves the

calculations of the TF-IDF function to focus on the

essential parts only rather than calculating misleading

information and preventing the function from diverg-

ing from the critical information. Then after pass-

ing the list of inputs, each input is passed into the

TF-IDF function with each story, and the maximum

probability score is chosen. This score resembles the

most likely ﬁtting story for the input fact. This pro-

cess is repeated for the rest of the inputs. Also, the

highest-scoring story appearing for each input is cho-

sen as the story to embed these facts in. The other

approach is similar to TF-IDF but applies cosine sim-

ilarity instead. We are trying to ﬁnd a matching story

that would make our input ﬁt in the sentence with-

out being placed incorrectly and without affecting the

story’s ﬂow and the text’s coherence. For this reason,

we propose another way to ﬁnd the matching story by

calculating the cosine similarity scores for each input

with every sentence in the story. We ﬁrst split our

story into a list of sentences. The function is then cal-

culated by encoding each sentence to have them as

vectors so that the function can be applied. Accord-

ingly, the score is calculated for each input n times

where n is the number of sentences in a story. We

calculate a list of probabilities that we sum for each

story. We pick the highest probability scores to ﬁnd

the most ﬁtting stories in our dataset. Then, we cal-

culate the total score for each fact from our input and

store it for each story, respectively. The ﬁnal stage for

picking the base story is choosing a story randomly

from the list of highest probabilities. This ensures

different outputs every time the story generator is run

and avoids producing repetitive, boring results.

https://libguides.stcc.edu/c.php?g=886516&p=6370592

2.2 Fact Embedder

The Fact Embedder is responsible for choosing ap-

propriate locations within the base story to embed the

different facts. This is done by relying on the cosine

similarity lists already calculated by the Base Story

Selector. Each sentence in the story is encoded, and

we generate a combined list of cosine similarities be-

tween every sentence in the story and each input fact.

This yields a list of probability scores for each input

fact. We then randomly choose one of the maximum

three scores for each fact. Again, this is to ensure

non-determinism and enable different generated sto-

ries. The same fact-embedding approach is applied

when inserting facts into auto-generated new stories.

3 FACT-IN-A-BOX DESIGN

Fact-in-a-Box in a box was designed in a simple man-

ner. The aim of the initial prototype was not to in-

troduce any design elements that might affect the re-

sults of measuring the learning gain from the embed-

ded facts.

3.1 User Interface and Features

The developed user interface is a web application with

two different interaction modes: an educator mode for

content creation and a student mode for indirect or

implicit learning. In the following, we will discuss

each mode’s different features/views.

Educator Mode - Content Creation

The Educator Mode consists of three main views:

Story Selection, Fact Input, and Story Browsing. Story

Selection is an optional feature. Educators can use it

to choose a speciﬁc story from the dataset or upload

their own story as the base story for the fact embed-

ding. This step can be skipped if a story should be au-

tomatically selected. The Fact Input of Fact-in-a-Box

is shown in Fig 2. Educators can input the four facts

they want to teach learners. The educator should also

specify the protagonist animal to improve the base

story selection process. After conﬁrming, the facts are

used as a basis for selecting the base story or the ed-

ucator’s input story. The Fact Embedder then inserts

the facts in the suitable locations in the story before

displaying the generated story in the Story Browsing

view. In this view, the facts are highlighted for the

educator to evaluate their location. The educator can

then conﬁrm the generated story or rerun the process.

CSEDU 2023 - 15th International Conference on Computer Supported Education

332

Figure 2: The Fact Input View of Fact-in-a-Box.

Student Mode - Indirect Learning

The Student Mode currently consists of the Story

Reading view, where learners are presented with dif-

ferent stories the educator assigns. Students can

switch between the different stories.

3.2 Implementation Details

The main aim was to implement a rapid prototype as

a proof-of-concept to the proposed idea of fact hid-

ing for implicit learning. Thus, we implemented a

basic architecture for the proposed approach relying

on readily available and easy-to-use tools. We used

Python and Google Colaboratory to build the back-

end of Fact-in-a-Box, and Flask API

to build the

web app. Flask is a Python-based microweb frame-

work. Flask is classiﬁed as a micro-framework since

it does not require using speciﬁc tools or libraries. It

lacks a database abstraction layer, form validation, or

other components relying on pre-existing third-party

libraries to perform common operations. Thus, Flask

is used for building the UI that links with the back-end

code rather than using other frameworks that might

slow down the story generation process in Google Co-

lab’s environment. We are using it for both our gener-

ation and running the server. We used Ngrok

to host

the local webserver to the internet.

4 EVALUATION AND

EXPERIMENT

Our goal was to hide some facts regarding a speciﬁc

topic in a story to help a speciﬁc group of individu-

https://ﬂask.palletsprojects.com/en/2.2.x/

https://ngrok.com/

als better retain the information. We wanted to eval-

uate whether our approaches yield the intended re-

sults of giving information to individuals without ex-

plicitly mentioning them and without giving the users

the feeling that they are on a learning endeavor. As

proof of concept, we chose to test our Fact-in-a-Box

approach on the target group of young children, i.e.,

generating stories with easy difﬁculty. Children were

chosen as they are the most obvious target group for

such a tool, especially in its initial phases. Children

have an amazing capacity for retaining knowledge

and, at the same time, are the most adverse age group

to sitting and receiving it (Chau, 2008). As we are

dealing with children, we decided to constrain the tar-

get information to be learned to four facts. We restrict

the global story dataset to stories topically related to

animals. The already existing stories were scraped

”Folklore and Mythology Electronic Texts” (Ashli-

man, 1996) that were all related to animals to test the

model for only one domain. All scraped books were

cleaned, ﬁltered, and processed.

4.1 Pilot Experiment

We experimented with measuring the children’s abil-

ity to recall the information they read in the story and

how much information they could retain. We also

tested their ability to explicitly state the facts we em-

bedded inside the story.

4.1.1 Experiment Design and Setup

The experiment aimed to give clear insights into chil-

dren’s perceptions of the application. Accordingly,

we choose a small sample size for this initial evalua-

tion. We invited 9 children ages seven and nine. We

needed to evaluate their reading ability before start-

ing the experiment, and this was crucial to ensure they

could extract and understand knowledge from reading

texts to conduct our experiment. Four children did not

meet this criterion and were thus excluded from the

experiment.

After that, each child was given the same three

generated stories to read. Each of the stories had four

facts, but they had variable lengths. The order of the

three stories was counterbalanced in a Latin square

manner. The experiment consisted of four question

groups that serve as evaluation metrics. (1) A regu-

lar post-test to measure the learning gain of the chil-

dren from the story with respect to the intended learn-

ing facts. Each child is given the same two questions

about the facts embedded in the three stories. (2) Chil-

dren were asked for three facts they learned from the

story about the protagonist’s animal to evaluate how

well they recognized the included facts. (3) Mislead-

Fact-in-a-Box: Hiding Educational Facts in Short Stories for Implicit Learning

333

Figure 3: S

with the input facts highlighted.

ing information contradicting the story’s context was

embedded in one of the cases to see whether the chil-

dren would retain it or not. This helped us understand

whether the proposed model can inﬂuence the chil-

dren’s judgment of the story.

4.1.2 Results

We will present the three generated stories, S

, S

and S

, with their learning facts, the post-test ques-

tions, and the results of the experiment on this story.

The input facts are highlighted in the generated stories

to display where the model predicted and embedded

each fact. The model outputs a story that revolves

around a speciﬁc protagonist animal: elephants, kan-

garoos, and lions, respectively.

1. The generated story S

is shown in Fig. 3. The

following post-test questions were asked:

• ” Do you ﬁnd elephants gentle creatures?”: The

reason behind asking this question is that the

context of the story implies that elephants were

causing harm to rabbits. However, we men-

tioned in our facts that elephants were kind, so

we wanted to see if the effect of our input as

it was placed by the model early at the begin-

ning of the story would inﬂuence the children’s

understanding.

• ” Do Elephants feed on grass?”: This question

was clear as we needed to measure how well the

children retained the relatively easy question or

will they encounter difﬁculty.

• ” Mention three facts about the elephant from

the story you have just read.”: Here, we are try-

ing to estimate how much information the child

can understand and receive from our input.

Figure 4: S

with the input facts highlighted.

To analyze the results of the answers to the sur-

vey questions of the S

, we investigated the chil-

dren who said” No” to speciﬁc questions. We

found that the reason for answering” No” in the

ﬁrst story was the story’s context, describing ele-

phants in a way that they harmed another animal

(rabbit). Hence, the answer was” No”, as the ele-

phant hurts another animal, so it cannot be gen-

tle. The second question was easy, and all the

kids scored correctly except for one. When we

investigated the results of the ﬁrst question, we

found that the story had already mentioned that”

elephants eat leaves” before we stated that” ele-

phants eat grass”. Almost every child stated that”

elephants are huge” and that” the elephants drink

a lot of water”. Still, although four facts were em-

bedded in the story, all the children could not state

three facts and only recalled two.

2. The generated story S

is shown in Fig. 4. The

following post-test questions were asked:

• ”Are Kangaroos adapted to hot environment?”:

This question was fairly indirect to ask the chil-

dren; however, we needed to see how well they

answered this question to measure their ability

to capture the information if it has been stated

at a higher linguistic level.

• ”What do Kangaroos feed on?”: This question

was clear as we needed to measure how well the

children retained the relatively easy question or

will they encounter difﬁculty.

• ”Mention three facts about the Kangaroos from

the story you have just read.”: Here, we are try-

ing to estimate how much information the child

can understand and receive from our input.

As expected in the answers to the survey ques-

tions of the second story, no child could answer

the ﬁrst question, but all children easily managed

to answer the second.

3. The generated story S

is shown in Fig. 4. The

following post-test questions were asked:

• ” Is the lion the king of the forest”: We wanted

to investigate the children’s ability to receive in-

CSEDU 2023 - 15th International Conference on Computer Supported Education

334

Figure 5: S

with the input facts highlighted.

formation from our model when they received

a relatively straightforward fact.

• ” Is the lion powerful?”: This question contra-

dicts the story’s context, where the lion was de-

feated by another animal. Still, we wanted to

check if the child could see the lion as powerful

and if it would affect his judgment.

• ”Mention three facts about the lion from the

story you have just read.”: Here, we are trying

to estimate how much information the child can

understand and receive from our input.

The ﬁrst question got three ” No” answers. The

reasoning behind Child 2’s answer was that the

lion was defeated in the story by the Gnat; hence

it is not powerful. Same for Child 4 and 5 men-

tioned that the lion could not stop the Gnat. This

left only 2 children with ”Yes”, supporting our

claim that implicitly affecting the children’s de-

cisions, lets them obtain information that is not

accurate. All the children could identify that” the

lion is the king of the forest” in the second ques-

tion we asked them. Finally, we received an aver-

age of two facts out of the 3 for all the children.

The pilot experiment results showed that all the

children could identify at least two out of four facts

in each story. And they received clear and mentioned

facts with a rate of up to 80 percent. Accordingly,

we found that we could involve false information to

the children, and around 50% were able to verify its

authenticity even though it conﬂicted with the actual

storyline. Finally, The linguistic difﬁculty of the sen-

tences greatly affected the children’s ability to answer

the questions. This was yielded from the open-ended

discussion conducted with the children after the ex-

periment, where they pinpointed some sentences as

too complicated.

4.2 Human Evaluation of Stories

Following the proposed approaches presented in (Sai

et al., 2022), we also evaluated the generated stories

using human evaluation. To test the quality of the gen-

erated stories, we surveyed a group of 40 adults. They

Table 1: Summary of the results of the subjective evalua-

tion.

(Q1) (Q2) (Q3) (Q4)

40% very good (5)

avg = 3.9

yes = 75% yes = 85% yes = 75%

42.5% very good (5)

avg = 4.1

yes = 75% yes = 85% yes = 75%

42.5% very good (5)

avg = 4.1

yes = 75% yes = 85% yes = 87.5%

were mainly asked to (Q1) evaluate the story’s gram-

mar (Likert scale from 1 (lowest) to 5), (Q2) whether

there are repetitions in some places (yes/no), (Q3),

whether a child will be able to identify the facts, and

(Q4) whether they would read it to a child as it’s a

good source for learning facts. The summary of the

results can be found in Table 1. S

surprisingly scored

better in (Q4). The results strengthen the position of

our approach and claim in being able to convey the

hidden facts within the text, making it unnoticeable

to the reader that there is something off or misplaced

with the text he is reading without being to identify

that a certain part of the text was hidden explicitly.

5 CONCLUSION

We have shown that NLG and story generation may

be beneﬁcial in domains like education and fact em-

bedding. We implemented an approach to embed

facts in related stories. We proposed a framework

called Fact-in-a-Box as a proof of concept that con-

tains two different interaction modes one for educa-

tors and one for students. It contains animal-based

stories directed to children. The process starts by tak-

ing the fact as input from the educator and goes to the

story selector module to select the most relevant story

to it. The selection is done using two methods TF-

IDF and Cosine Similarity. After the story is selected,

the fact embedder module that adds the fact between

the available sentences by selecting the best position.

In the end, the generated story could be shown to the

student. We tested our proposed approach by evaluat-

ing the learning gain of some students and conducted

a survey for adults to evaluate the language of the re-

sulting stories and the concept itself.

In the future, we want to apply more testing for the

different modules of the application. In addition, we

want to enhance the performance of the story gener-

ation module and include different domains. We will

also add a quizzing module, where educators can de-

ﬁne quizzes in the Educator Mode. In the Student

Mode, students will then be able to take a quiz after

reading a speciﬁc story, if a quiz is deﬁned for it.

Fact-in-a-Box: Hiding Educational Facts in Short Stories for Implicit Learning

335

REFERENCES

Abdalrahman, K. K. (2022). Teaching and learning vocab-

ulary through short stories. Canadian Journal of Lan-

guage and Literature Studies, 2(2):7–15.

Akturk, A. O. (2022). Thirty-ﬁve years of the journal of

computer assisted learning: A bibliometric overview.

Journal of Computer Assisted Learning, 38(5):1220–

1253.

Alhussain, A. I. and Azmi, A. M. (2021). Automatic story

generation: a survey of approaches. ACM Computing

Surveys (CSUR), 54(5):1–38.

Ashliman, D. (1996). Folklore and mythology electronic

texts. DL Ashliman.

Belda-Medina, J. (2022). Promoting inclusiveness, creativ-

ity and critical thinking through digital storytelling

among eﬂ teacher candidates. International Journal

of Inclusive Education, 26(2):109–123.

Bizami, N. A., Tasir, Z., and Kew, S. N. (2022). Innovative

pedagogical principles and technological tools capa-

bilities for immersive blended learning: a systematic

literature review. Education and Information Tech-

nologies, pages 1–53.

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,

Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,

Askell, A., et al. (2020). Language models are few-

shot learners. Advances in neural information pro-

cessing systems, 33:1877–1901.

Chau, M. (2008). The effects of electronic books designed

for children in education.

Faidley, E. W. (2021). “movies, tv shows, and memes... oh

my!”: An honors education through popular culture

and critical pedagogy.

Fu, H. S., Silva, P. H. B. d., Silva, A. P. d., Souza Junior,

M. B. M. d., and Melo, M. S. T. d. (2022). Movies

as strategies for physical education classes at school.

Movimento, 28.

Hofman-Bergholm, M. (2023). Storytelling: The ancient

tool of using stories to communicate knowledge for

a sustainable future. In Integrated Education and

Learning, pages 237–253. Springer.

Kangas, T. C., Cook, M., and Rule, A. C. (2017). Cine-

matherapy in gifted education identity development:

Integrating the arts through stem-themed movies.

Journal of STEM Arts, Crafts, and Constructions,

2(2):3.

Mastan, I. A., Sensuse, D. I., Suryono, R. R., and Kautsa-

rina, K. (2022). Evaluation of distance learning sys-

tem (e-learning): a systematic literature review. Jurnal

Teknoinfo, 16(1):132–137.

Nadi-Ravandi, S. and Batooli, Z. (2022). Gamiﬁca-

tion in education: A scientometric, content and co-

occurrence analysis of systematic review and meta-

analysis articles. Education and Information Tech-

nologies, 27(7):10207–10238.

Nicolaou, C. (2023). The secret power of digital story-

telling methodology: Technology-enhanced learning

utilizing audiovisual educational content. In Enhanc-

ing Education Through Multidisciplinary Film Teach-

ing Methodologies, pages 235–246. IGI Global.

Obloberdiyevna, D. S. and Odilkhonovna, K. U. (2022).

Teaching languages using modern educational meth-

ods. International Journal of Intellectual Cultural

Heritage, 2(3):105–111.

Pang-Ning, T., Steinbach, M., and Kumar, V. (2005). Intro-

duction to data mining addison-wesley.

Purnama, S., Ulfah, M., Ramadani, L., Rahmatullah, B.,

and Ahmad, I. F. (2022). Digital storytelling trends

in early childhood education in indonesia: A system-

atic literature review. Jurnal Pendidikan Usia Dini,

16(1):17–31.

Quah, C. Y. and Ng, K. H. (2022). A systematic literature

review on digital storytelling authoring tool in edu-

cation: January 2010 to january 2020. International

Journal of Human–Computer Interaction, 38(9):851–

867.

Radford, A., Wu, J., Child, R., Luan, D., Amodei, D.,

Sutskever, I., et al. (2019). Language models are un-

supervised multitask learners. OpenAI blog, 1(8):9.

Ratih, G. K., Iriani, A., and Dwikurnaningsih, Y. (2022).

Kindergarten teachers training in integrating anti-

corruption education through storytelling and game.

Jurnal Obsesi: Jurnal Pendidikan Anak Usia Dini,

6(03):1628–1639.

Sai, A. B., Mohankumar, A. K., and Khapra, M. M. (2022).

A survey of evaluation metrics used for nlg systems.

ACM Computing Surveys (CSUR), 55(2):1–39.

Salton, G. (1983). Introduction to modern information re-

trieval. McGraw-Hill.

CSEDU 2023 - 15th International Conference on Computer Supported Education

336