Educational Chatbots: A Sustainable Approach for Customizable

Conversations for Education

Donya Rooein

, Paolo Paolini

and Barbara Pernici

Poliotecnico di Milano, Milan, Italy

Keywords:

Educational Chatbot, Adaptive Learning, Conversational Agents, Customizable Conversations.

Abstract:

This paper proposes using chatbots as “tutors” in a learning environment; tutors who are not domain experts

but helpers in guiding students through bodies of learning material. The most original contributions are the

proposal that conversation should be content-independent (although chatbots speak about content); The pro-

duction process should allow non-technical actors to customize chatbots and keep the costs of development

and deployment low. We speciﬁcally discuss conversation customization, which is relevant, especially for

learning applications, where users might have speciﬁc needs or problems. We achieve the features introduced

above via extensive “conﬁguration” (regarding direct programming), making the underlying technology novel

and original. Experiments with teachers and students have shown that chatbots in education can be effective

and that customization of conversations is relevant and valued by users.

1 INTRODUCTION

Chatbots are famous for being a valid alternative

to traditional point-and-click interaction between hu-

mans and computers (Brandtzæg and Følstad, 2017).

There are many studies who have investigated differ-

ent roles of chatbot in education (Hwang and Chang,

2021), and this paper is about customizable (or even

adaptive) conversations for educational chatbots. This

study is an alternative approach to the typical one-

size-ﬁts-all chatbot solution for educational chatbot

conversation design, which is relevant to education,

where psychology and individual needs are of great

importance.

Customization of conversational features may in-

clude the chatbot’s loquacity, the topic of the chat-

bot’s turns, the chatbot’s style and wording, or the

length of the chatbot’s wording in each turn. Our idea

is that customization of the conversation will make

it more effective, i.e., better tuned to the user’s pro-

ﬁle (including students with special needs) and better

tuned to contextual situations (e.g., the user is tired or

in a hurry). Customization means that someone will

make choices (explicit or implicit); in the future, we

envision adaptive conversations, i.e., the chatbot will

https://orcid.org/0000-0002-0368-3084

https://orcid.org/0000-0003-3486-5662

https://orcid.org/0000-0002-2034-9774

interpret data and decide on possible changes to con-

versational features.

There are many different roles for chatbots in ed-

ucation, such as a chatbot to support learning expe-

rience, an assistive tool, or a tutoring and mentor-

ing role (Wollny et al., 2021). We propose to use

chatbots as tutors, helping students through a body

of content. A chatbot acting as a tutor is not a do-

main expert (say History or Computer Science); it

must be able to sustain a good conversation in a learn-

ing experience and help the student move through the

various learning elements

. There are several moti-

vations for this choice: i) there is a vast amount of

digital content already available (e.g., online courses,

MOOCs

, Learning objects, digital resources); rather

than creating new content, chatbots can help at mak-

ing better use of existing ones; ii) with respect to

the traditional (interactive) interface, the conversation

may add friendliness, empathy, and easiness to use all

ingredients relevant for learning; iii) learning, espe-

cially in formal education, has an already established

group of actors (e.g., authors, teachers, publishers,

etc.); a new technology should improve the activities

of these actors, rather than replace them.

Reﬁning the above idea, we elaborated many

The learning element refers to each learning material

unit, e.g., a theoretical description.

Massive Open Online Course

314

Rooein, D., Paolini, P. and Pernici, B.

Educational Chatbots: A Sustainable Approach for Customizable Conversations for Education.

DOI: 10.5220/0011083200003182

In Proceedings of the 14th International Conference on Computer Supported Education (CSEDU 2022) - Volume 1, pages 314-321

ISBN: 978-989-758-562-3; ISSN: 2184-5026

high-level requirements as follows: i) the conversa-

tion afforded by the chatbot should have as much

as possible a human ﬂavor, providing functional and

nonfunctional aspects; ii) the chatbot should conduct

mainly a proactive conversations taking the lead in or-

der to take the learner across the material effectively;

iii) the content should be customizable according to

the proﬁle of the learner, her speciﬁc needs and also

contextual situation; iv) the conversation itself should

be customizable, in order to be accepted and effec-

tive for the learner; v) actors in educations should

be empowered to control customization both of con-

tent and conversation since they are in charge of or-

ganizing learning; vi) learners should be able to fur-

ther control and customize their learning experience.

vii) the overall process for creating chatbots should

be streamlined, effective, low cost, and with little IT

personnel involvement. Consequently, a speciﬁc type

of chatbot named by TalkyTutor was developed.

In this paper, we focus speciﬁcally upon the fea-

ture of the chatbot that allows the conversation to be

customized. The basic idea is that the conversation

is controlled by several conﬁguration data that can

be easily modiﬁed without programming. Initial ex-

periments (with teachers and students) show that cus-

tomizing conversation is very promising and appreci-

ated by users. In Section 2, we discuss related work;

in Section 3, we discuss how conversations are de-

scribed via a conﬁguration-driven approach and can

be customized; in Section 4, we describe the initial

experimentation, the related qualitative and quantita-

tive assessment; in Section 5, we draw the conclu-

sions and discuss future work.

2 RELATED WORK

2.1 Chatbots in Education

The concept of educational chatbots has its origins

from intelligent tutoring systems, which address the

idea of building a learning tool that is intelligent

enough to understand learners’ needs and proceed ac-

cordingly (Song et al., 2017). Authors in (Burkhard

et al., 2021) presented a theoretical basis for the use

of smart machines like chatbots in education by con-

sidering the role of teachers and focusing on the ne-

cessity for teachers to play an active role in the digital

transformation.

From the broad application of conversational

agents in education over time, we mention the Grasser

et al. a study from 20 years ago who introduced tu-

toring systems as a conversational agent to help col-

lege students learn about computer literacy (Graesser

et al., 2001). Heller et al. developed a chatbot de-

signed by open source architecture of AIML

to im-

prove student-content interaction in distance learning

(Heller et al., 2005). In 2008, Kerry et al. worked

on using conversational agents for self-assessment

in e-learning (Kerry et al., 2008). Chatbots are

often applied for organizational support to perform

speciﬁc tasks, e.g. automated FAQ (Han and Lee,

2022). An intelligent teaching assistant (iTA) intro-

duced by (Duggirala et al., 2021) to help students by

providing detailed answers to their questions by us-

ing a generative model, extracted the relevant content

from the top-ranked paragraph to generate the answer.

2.2 Chatbots and Conversation

According to McTear, despite all the progress in

speech recognition and natural language understand-

ing, but still chatbots suffer from the lack of the con-

versational abilities that make the interaction with

them unnatural (McTear, 2018). Conversation by it-

self is a complex system, but the aspect of the con-

versation design plays a crucial role on the chat-

bot’s effectiveness. Natural Conversation Framework

(NCF) presented a new approach for implementation

of the multi-turn conversation between chatbot and

human (Moore and Arar, 2019). This approach uses

a library of predeﬁned UX patterns driven by natu-

ral human conversation. The conversation structure

in chatbots are mainly hand-crafted, which means the

whole conversation is embodied by content and devel-

oped by IT experts (Paek and Pieraccini, 2008). This

approach makes chatbot development and mainte-

nance expensive and not easily generalizable to other

domains. On the other hand, statistical methods use

machine learning algorithms to learn about an optimal

dialogue strategy from the interaction with users. In

addition, there are end-to-end dialogue systems that

use deep neural networks to generate responses from

the large corpora of dialogues (Serban et al., 2016).

2.3 Chatbots and Customization

If we look deeper into the past three years of research

on educational chatbots, we can consider Tegos et al.’s

work on a conﬁgurable design of chatbots for syn-

chronous collaborative activities in MOOCs in a uni-

versity setting relevant to our research (Tegos et al.,

2019). They targeted MOOCs and integrated them us-

ing chatbots to support simple tasks such as collecting

feedback from learners. A recent study(Rooein et al.,

2020) has shown the use of chatbots also for train-

ing new employees in a factory by adaptive approach.

Artiﬁcial Intelligence Markup Language

Educational Chatbots: A Sustainable Approach for Customizable Conversations for Education

315

Authors in (von Wolff et al., 2019) and (Winkler et al.,

2020) focused on the educational actors (students and

teachers) to extract important features and investi-

gated the requirements implementing a chatbot in uni-

versity setting and also conﬁrm the acceptance of the

chatbots in academic environments.

All the mentioned approaches are highly involved

by IT experts to develop and maintain the conversa-

tion module of the chatbots. Here, we start with hand

crafted approach to design the conversation compo-

nent of chatbots, and later by a conﬁgure driven ap-

proach make it more sustainable in the production and

maintenance point of view and also delivering a more

ﬂexible conversation.

3 DESIGNING AND DEPLOYING

ADAPTIVE CONVERSATIONS

In this section, we describe our approach for design-

ing and implementing conversation machine in edu-

cational chatbots. The key aspect of this approach is

that it does not support a speciﬁc conversation, but it

allows designers, authors and teachers to model the

conversations that they wish. And, not less important,

it allows learners to customize the conversation of the

chatbot during the learning experience. There are two

very original aspects: i) conversation design is totally

independent of the content that the chatbot will de-

liver; ii) a great deal of the conversation features is

controlled via conﬁguration rather than by program-

ming; changing conﬁguration data allows modifying

the conversation at low (nearly zero) cost.

Following what we had said in the Section 1, the

conversation that we envision for this chatbot should

exhibit a number of relevant features:

• As close as possible to natural conversa-

tions (Moore and Arar, 2019).

• Proactive, in the sense that the chatbot leads the

conversation.

• Responsive, in the sense that the chatbot properly

reacts to user turns, that could be solicited (e.g.

asking for feedback) or unsolicited.

• Customizable (and adaptive) in a number of ways.

• Designed by nontechnical actors. (i.e. authors,

teachers, conversation experts, . . . )

• Be sustainable, both in terms of costs and time

required.

In the following of this section we describe some

of the components of TalkyTutor that allow to imple-

ment the above features: Conversation Machine (the

strategy of the conversation of the chatbot), Dialouge

Categories, Turns of the chatbot (what the chatbot

says and when), Wording, Templates, and variables

(the speciﬁc utterances of the chatbot and its formula-

tion), Intents (what users may say and how it is inter-

preted), and Loquacity control (controlling how often

the chatbots speaks and the length of its utterances).

3.1 Conversation Machine

Conversation Machine is the most complex piece of

the machinery to generate a conversation between

chatbot and human. This component in chatbot’s ar-

chitectures performs these tasks: i) it generates the

conversation turns of the chatbot; ii) it understands the

user turns of conversation, iii) it organizes the ﬂow of

turns, and iv) it calls other “engines” of the chatbot

when needed (e.g. to fetch a new item of content)

The conversation is modeled as a state machine. The

turns of the chatbot corresponds to entering or leaving

a state. All turns of the chatbots are made of messages

from different categories.

The conversation machine, coupled with corre-

sponding turns, deﬁnes the overall strategy of the con-

versation; a conversation machine deﬁnes a family of

chatbots, in the sense that all the members of the fam-

ily adopt a similar strategy of conversation (indepen-

dently of the content being delivered).

A state machine is relatively complex; therefore,

it needs a conversation designer to shape it. In addi-

tion, each state machine requires a new programming

if they need more states than the default state ma-

chine, since proper hooking to various actions needs

to be established (e.g., when to ask for a new piece

of content). Once a state machine is created (deﬁning

a family of chatbots), several speciﬁc chatbots can be

created, just via conﬁguration with no programming.

3.2 Conversation Categories

As it was said above, the turns of the chatbot are clas-

siﬁed into categories. Figure 2 shows an example of

categories. Categories are not built-in by the technol-

ogy; they are part of the conﬁguration: They can be

modiﬁed without programming, and are clearly de-

pendent upon the design of the conversation machine.

A category deﬁnes the overall semantic of a turn of

the chatbot; “GREETINGS” for example means ex-

changing pleasantries (e.g. “nice to see you again”);

“FORECAST” means to anticipate what will happen

for completing the current learning path, etc. Clas-

sifying the turns of the chatbot into categories has

a double purpose: helping designers to identify the

need for a turn; and also helping them to put turns in

CSEDU 2022 - 14th International Conference on Computer Supported Education

316

Figure 1: Structure of chatbot’s wording and template example.

Figure 2: An example categories for the turns of the chatbot.

the proper order. At the start of a session, for example,

“GREETINGS” are at the beginning of a sequence;

at the end of a session they should come at the end.

These categories are attached to the states and arcs

of state machine to describe the chatbot dialogues in

the conversation with user. Also, these categories are

used in the unsolicited turns to understand users and

provide them the response. The chatbot for example

could say “did you like the item?” (“FEEDBACK”

category), and wait for the reply. Alternatively, the

user (in a different turn), may say “I did not like it”,

which is interpreted as an unsolicited feedback.

3.3 Wording, Templates and Variables

An original and relevant feature of TalkyTutor is the

possibility of easily controlling style and wording.

Style deﬁne an overall way of speaking (for the chat-

bot); examples could be friendly, professional or em-

pathic. The speciﬁc turns of the chatbot , as used by

the conversation machine are deﬁned via a “master ta-

ble”, that is used to debug and tune a speciﬁc family

of chatbots. Designers, however, may specify what

the chatbot actually says, by associating each turn in

the master table with a speciﬁc sentence. Wording

conﬁguration tables allow designers to introduce their

own wording for each speciﬁc turn, controlling both

style and length of the turn. A relevant aspect of the

state machine is that it is not dependent on the con-

tent; however, it can speak about individual items and

the learning pathway currently being used. This is

obtained by encoding templates (describing complex

information) and variables (describing simple infor-

mation). In the other word, each chatbot’s category

is consisted by one to many sentences, and each sen-

tence contains several ﬁxed wordings, templates and

variables as shown in Figure 1. Templates and vari-

ables can be created by authors, without interfering

with conversation design. Various versions of tem-

plates can be created (e.g., a standard, short or long

length version). Figure 1 also shows an example of

template for an item of content; it embeds some ﬁxed

words and simple variables derived from metadata as-

sociated to the content item. All chatbot’s wording in

the format of the ﬁxed words, templates and variables

are represented in tables.

We must remark again that style, wording and

templates are controlled via conﬁguration tables and

therefore may be part of the customization by the var-

ious actors, without the need of programming.

3.4 Intent Recognition

The modular architecture brings the opportunity to

use different intent detection services in the chatbot.

Currently, TalkyTutor chatbots can select either an in-

ternal module for intent detection trained based on

the classic BERT model developed by Google (De-

vlin et al., 2018) or inspecting external services such

Educational Chatbots: A Sustainable Approach for Customizable Conversations for Education

317

as IBM Watson. In both cases, the chatbot only un-

derstands the range of deﬁned intents in conﬁguration

data such as ”ask for help” or ”ask for summary”.

3.5 Loquacity Model

An additional issue is the loquacity of the chatbot:

how many turns does it take? Should they be more

or less? How verbose are these turns? It is clear that

there is no predeﬁned answer: in some situations, and

for some users, very few and short turns can be suit-

able. In other situations, instead, more (longer) turns

may be better. There is also another issue: some of the

turns of chatbot are mandatory, in the sense that they

can’t be skipped in the conversation; other turns (e.g.

a reinforcement message) could be skipped without

spoiling the functionality. In order to cope with this,

issue we have devised a way to control how many

turns the chatbot takes. The turns deﬁned in the con-

versation machine are the maximum: using trimming

algorithms we can cut them down. The problem is

that trimming has to be appropriate: e.g. i) never skip-

ping mandatory turns; ii) making sure that no cate-

gory is over represented; iii) making sure that no cat-

egory is skipped forever. For the time being we de-

signed a preliminary trimmed algorithm which is di-

rectly controllable via conﬁguration; what can be con-

trolled is the level of loquacity is directly related to the

metadata of each wording’s category ( e.g, mandatory

categories are not skippable in a different loquacity,

but optional once may be skip on the different turn’s

of chatbot). The user decides if more or less turns are

desired, and the algorithms do the job. Experimental

testing (discussed in the next section) shows that users

like to control the loquacity of their chatbot.

3.6 Customizing the Chatbot

The combination of the above can create very adap-

tive conversations.

The conﬁguration of the conversation above described

may require some conversation expertise but no pro-

gramming at all. Redeﬁning a state machine and cor-

responding Master Table (including the deﬁnition of

transitions and rules to ﬁre them) might be not easy.

The amount of time could vary depending on the level

of the modiﬁcations and it could take a few hours or

few days. Modifying Style Tables and Alternatives

is almost trivial. The current implementation of the

above machinery has shown that we can control the

conversational features of the chatbot by greatly re-

ducing reprogramming time.

There is a ﬁnal overall issue: who does what, in

terms of design and conﬁguration? Since most of the

features discussed in this section have not direct coun-

terpart in the literature, we did experiment with vari-

ous possibilities, and the end we came up with rules

shown in Figure 1, that seem reasonable for an edu-

cation environment. These rules could be modiﬁed in

the future, for different application realms.

In the next section we discuss as the above fea-

tures were used in an empirical testing involving 12

teachers and 80 students (of school and higher educa-

tion).

4 EXPERIMENTING WITH

ADAPTIVE CONVERSATIONS

The approach and the technology described in the pre-

vious sections have been tested in this paper with two

different bodies of content: “Advanced Computer Ar-

chitecture” (ACA) in English and “La curtis” (me-

dieval history) in Italian.

4.1 Experiment

The experimentation of evaluating the TalkyTutor

runs over several months. Firstlt, we asked teachers

were to use chatbots with two points of view: i) were

they willing to adopt them for their students? ii) what

could have been the reaction of their students? Over-

all, 12 teachers were involved, 7 from Higher educa-

tion and 5 from schools. In terms of disciplines, 5

teachers were from humanities and 7 from STEM. A

qualitative investigation was performed through struc-

tured interviews. Several changes were introduced

based on this preliminary assessment (especially to

the interface and adaptivity features).

Secondly, new experimentation was conducted

with a focus group. The same group of teachers was

asked to repeat their experience. Students (selected

by their teachers in high school and invited by email

for higher education) were involved with using the

chatbot in their own environment (school or univer-

sity classes). Overall, 81 students were asked to use

the chatbot to simulate a learning session (30 minutes

at least) and ﬁll up a survey. A few students were also

involved in a focus group for interviews. 33 students

were from Higher education and 48 students from ju-

nior high school. Students were guided by the teach-

ers that also provided local instructions.

For the experimentation a speciﬁc family of chat-

bot was created. Two different content were adapted

from existing material: ACA, developed by a profes-

sor in higher education, and “La Curtis”, a course

about Medieval History. Both courses were real, in

CSEDU 2022 - 14th International Conference on Computer Supported Education

318

Table 1: Current rules for chatbot conﬁguration.

Conversation machine:

This is created by conversation designers with the support of IT specialists.

It is done only once for a family of chatbots.

Categories and turns:

They are created by experts in the conversation for education.

They are deﬁned only once for a family of chatbots.

Stule, wording and templates:

They are initially deﬁned for a family of chatbots. They can be (optionally)

modiﬁed by publishers, authors, and teachers.

User intents:

What users may say and how it is interpreted. They are initially created

by designers but can easily expanded.

Loquacity of the chatbot:

It can be controlled dynamically by chatbot designers

(teachers say they that they should also control it)

the sense that were created using text and slides of ex-

isting courses. “ACA” was delivered in English (with

the chatbot speaking in English) to higher education

students only; “La Curtis” was delivered in Italian

(with chatbot speaking in Italian) to Higher Education

students and Junior High school students. In the fol-

lowing of this section we brieﬂy discuss i) the quali-

tative analysis of teacher reactions; ii) the quantitative

analysis of the surveys ﬁlled by students; iii) a quali-

tative analysis of comments made by students.

Let us examine ﬁrst the reactions of teachers. Sev-

eral issues were discussed in the interviews, but we

only present overall reactions and opinions about con-

versation adaptivity. Teachers, in general, were highly

positive about having Talky tutor chatbots support-

ing learning by their students. Most teachers enjoyed

their experience. Some teachers (higher education)

found the conversation with the chatbot a little slow

in pace. “In the beginning, I was suspicious; then I

realized that it could work very well, especially for

less motivated students” (school teachers).

Teachers did not sufﬁciently understand the need

for conversational adaptivity at ﬁrst. After a while,

they started to like the idea. There is an interesting

divide: i) school teachers think they should control

style and loquacity for their students (adaptivity for

individual needs); ii) higher education teachers think

learners should be empowered.

To our surprise, wording customization is felt as

signiﬁcantly important by teachers; almost all of them

declared that they would spend 2-3 hours to substitute

standard sentences of the chatbot with their wording.

There several motivations:

• Learners will recognize their teacher, improving

their motivations to learn.

• The effect of presence (by the teacher) would be

enhanced.

• The psychological impact will make the chatbot

more persuasive.

“It will provide less feeling of speaking to a

robot” (school teacher)

Figure 3: The of experiment with students.

Let us now discuss the analysis of the surveys

ﬁlled by the students. Figure 3.a shows the overall

appreciation of the experience with TalkyTutor chat-

bots. The majority (score 5 or 4) liked it; a minority

did not like it (score 0 or 1). School students were

slightly more positive than higher education students

(by less than 2%).

Figure 3.b shows how pleasant was the use of

the chatbot for the students. Again the vast majority

found the experience pleasant (score 4 or 5) while a

minority had a negative opinion (score 0 or 1). Again

school students were slightly more positive. Fig-

ure 3.c shows how the students found the experience

useful for learning. Figure 3.c demonstrates the role

of adaptativity between students.

While compiling the surveys students could add

comments. Table 2 shows a few comments by stu-

dents. Some of them are from the (tiny) minority that

do not appreciate chatbots.

Table 3 shows a few comments by students about

the adaptivity of the conversation. A few comments

were misplaced since they were referring to other

Educational Chatbots: A Sustainable Approach for Customizable Conversations for Education

319

Table 2: Students about the idea of using TalkyTutor. H:

Higher education; S: Junior High School.

No. Level Comment

1 H It’s cool to have a guide that helps you

keeping track of learning and that can

customize your learning experience,

but in the worst case it’s just a com-

plex table of content, not that useful.

2 H Being praised while learning it’s a

thing to not be underestimated in my

opinion because it pushes the stu-

dent to be more active and focused

on the videos with respect to just

watch them without any kind of inter-

action like in simply playing a video

playlist.

3 H It was fun to talk with the chatbot.

4 S 1It will help students with problems

to how study a course in different

ways.

5 S It is an important alternative to tradi-

tional learning; I hope they will be in-

troduced very soon at school.

6 S Besides content, the chatbot should

propose quizzes.

7 H I prefer human relationships.

8 H I don’t like chatbots in general.

problems. Overall, we may say that teachers and stu-

dents liked learning with chatbots and found the ex-

perience pleasant and (potentially) useful for learn-

ing. We are aware that the perception after 30 minutes

could be different from actual usage over an extended

period of time.

5 CONCLUSIONS AND FUTURE

WORK

Let us summarize the most relevant contributions of

this paper; ﬁrst of all, we put forward the idea of mak-

ing chatbot customizable in order to be tuned to the

user proﬁle and the context; this is speciﬁcally rel-

evant when chatbots are used for a learning experi-

ence, where personalization is a great relevance (Cai

et al., 2021). Next, we propose a speciﬁc role for

chatbots in education: tutors leading users across con-

tent. Finally, we advocate the need to streamline the

production process with two main goals: i) reducing

costs and effort for deployment; ii) empowering non-

technical actors to direct control the features of the

chatbot.

Table 3: Students about the Conversation Adaptativity. H:

Higher education; S: Junior High School.

No. Level Comment

1 H I tried different modes and levels

of loquacity and I appreciated a lot

that we can personalize it based on

our tastes.

2 S It helps to make the chatbot an al-

ternative to traditional learning.

3 S It is important, otherwise, the

chatbot could become boring and

repetitive.

4 S It can help to make chatbot speaks

as human beings speak.

5 H It is important and it should be fur-

ther developed.

6 H It is important to be able to speed

up the interaction.

7 S It is important to help to under-

stand what the chatbot says.

8 H Differences among styles could be

stronger. Empathic style should be

improved.

In order to make the above real, we have de-

veloped an original technology, where chatbots are

shaped via extensive use of conﬁguration data; some-

how, we have developed a chatbot generator. We en-

vision the production of a chatbot into the following

steps:

(a) Create a family of chatbots, sharing a common

conversation strategy.

(b) Instantiate a speciﬁc chatbot and plugging the

content (adequately organized and with proper

metadata).

data, controlling the turns of the chatbot, the style

and the wording, the loquacity, etc.

It should be noted that steps “b” and “c” could

be interchanged, and repeated several times. In addi-

tion, experimentation with teachers and students has

shown that chatbots as tutors can be effective and that

customization of conversation is perceived as impor-

tant.

We currently developed a platform for building

TalkyTutor chatbots and attaching the learning ma-

terials. The platform is standalone for now, but it

should be delivered as a service to be more scalable

in the future. From an application point of view, it is

also essential to deﬁne actors’ role in customization:

who does control of what? Authors, publishers, teach-

ers, and students ﬁnd customization (of content and

conversation) valuable, but they have different ideas

CSEDU 2022 - 14th International Conference on Computer Supported Education

320

about who does what. Teachers at school, for exam-

ple, would like to constrain conversation customiza-

tion possibilities for their students tightly; they want

to take basic choices, while students might have a dif-

ferent idea.

From the technical side, two issues are preemi-

nent: i) to add wider choices of interfaces, includ-

ing vocal ones and integration with Alexa style frame-

works; ii) to further improve the production process,

making it easier for non-technical actors to work on

customization.

As a research issue, we are investigating how

to move from customization to adaptation. There

is a component in our architecture responsible for

decision-making along with the conversation; at the

moment, it takes the simple decisions about suspend-

ing or stopping the session rather than keep going.

As a real tutor, it should take more important deci-

sions: the content should decide if an item needs to

be repeated or if the current pathway is appropriate or

should be changed; for the conversation.

ACKNOWLEDGEMENTS

This work has been partially supported by a grant of

EIT Digital and IBM Italy.

REFERENCES

Brandtzæg, P. B. and Følstad, A. (2017). Why people

use chatbots. In Kompatsiaris, I., Cave, J., Satsiou,

A., Carle, G., Passani, A., Kontopoulos, E., Diplaris,

S., and McMillan, D., editors, Internet Science - 4th

International Conference, INSCI 2017, Thessaloniki,

Greece, November 22-24, 2017, Proceedings, volume

10673 of Lecture Notes in Computer Science, pages

377–392. Springer.

Burkhard, M., Seufert, S., and Guggemos, J. (2021). Rel-

ative strengths of teachers and smart machines: To-

wards an augmented task sharing. In CSEDU (1),

pages 73–83.

Cai, W., Grossman, J., Lin, Z. J., Sheng, H., Wei, J. T.-Z.,

Williams, J. J., and Goel, S. (2021). Bandit algorithms

to personalize educational chatbots. Machine Learn-

ing, pages 1–30.

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

(2018). Bert: Pre-training of deep bidirectional trans-

formers for language understanding. arXiv preprint

arXiv:1810.04805.

Duggirala, V. D., Butler, R. S., and Kashani, F. B. (2021).

ita: A digital teaching assistant. In CSEDU (2), pages

274–281.

Graesser, A. C., VanLehn, K., Rose, C. P., Jordan, P. W., and

Harter, D. (2001). Intelligent tutoring systems with

conversational dialogue. AI Magazine, 22(4):39.

Han, S. and Lee, M. K. (2022). Faq chatbot and inclusive

learning in massive open online courses. Computers

& Education, 179:104395.

Heller, B., Proctor, M., Mah, D., Jewell, L., and Cheung, B.

(2005). Freudbot: An investigation of chatbot tech-

nology in distance education. In EdMedia+ Innovate

Learning, pages 3913–3918. Association for the Ad-

vancement of Computing in Education (AACE).

Hwang, G.-J. and Chang, C.-Y. (2021). A review of oppor-

tunities and challenges of chatbots in education. In-

teractive Learning Environments, 0(0):1–14.

Kerry, A., Ellis, R., and Bull, S. (2008). Conversational

agents in e-learning. In International Conference on

Innovative Techniques and Applications of Artiﬁcial

Intelligence, pages 169–182. Springer.

McTear, M. (2018). Conversation modelling for chat-

bots: current approaches and future directions. In

Berton, A., Haiber, U., and Minker, W., editors,

Studientexte zur Sprachkommunikation: Elektronis-

che Sprachsignalverarbeitung 2018, pages 175–185.

TUDpress, Dresden.

Moore, R. J. and Arar, R. (2019). Conversational UX De-

sign: A Practitioner’s Guide to the Natural Conversa-

tion Framework. Association for Computing Machin-

ery, New York, NY, USA.

Paek, T. and Pieraccini, R. (2008). Automating spoken

dialogue management design using machine learn-

ing: An industry perspective. Speech communication,

50(8-9):716–729.

Rooein, D., Bianchini, D., Leotta, F., Mecella, M., Paolini,

P., and Pernici, B. (2020). Chatting about processes

in digital factories: A model-based approach. In En-

terprise, Business-Process and Information Systems

Modeling, pages 70–84. Springer.

Serban, I., Sordoni, A., Bengio, Y., Courville, A., and

Pineau, J. (2016). Building end-to-end dialogue

systems using generative hierarchical neural network

models. Proceedings of the AAAI Conference on Arti-

ﬁcial Intelligence, 30(1).

Song, D., Oh, E. Y., and Rice, M. (2017). Interacting with a

conversational agent system for educational purposes

in online courses. In 2017 10th international confer-

ence on human system interactions (HSI), pages 78–

82. IEEE.

Tegos, S., Demetriadis, S., Psathas, G., and Tsiatsos, T.

(2019). A conﬁgurable agent to advance peers’ pro-

ductive dialogue in moocs. In International Work-

shop on Chatbot Research and Design, pages 245–

259. Springer.

von Wolff, R. M., N

ortemann, J., Hobert, S., and Schu-

mann, M. (2019). Chatbots for the information ac-

quisition at universities–a student’s view on the ap-

plication area. In International Workshop on Chatbot

Research and Design, pages 231–244. Springer.

Winkler, R., Hobert, S., and Appius, M. (2020). Empower-

ing educators to build their own conversational agents

in online education.

Wollny, S., Schneider, J., Di Mitri, D., Weidlich, J., Rit-

tberger, M., and Drachsler, H. (2021). Are we there

yet?-a systematic literature review on chatbots in edu-

cation. Frontiers in Artiﬁcial Intelligence, 4.

Educational Chatbots: A Sustainable Approach for Customizable Conversations for Education

321