Operationalizing Behavior Change Techniques in Conversational Agents

Maria In

es Bastos

1 a

, Ana Paula Cl

audio

1 b

, Isa Brito F

elix

2 c

, Mara Pereira Guerreiro

2 d

Maria Beatriz Carmo

1 e

and Jo

ao Balsa

1 f

LASIGE, Dep. Inform

atica, Faculdade de Ci

encias, Universidade de Lisboa, Portugal

Nursing Research, Innovation and Development Centre of Lisbon (CIDNUR), Nursing School of Lisbon, Lisbon, Portugal

Keywords:

Conversational Agents, Ontologies, Behavior Change Techniques, Dialogue Platforms.

Abstract:

Departing from previous work on the use of well-established behavior change techniques in an mHealth in-

tervention based on a conversational agent (CA), we propose in this contribution a new architecture for the

design of behavior change CAs. This novel approach combines the use of an advanced natural language plat-

form (Dialogﬂow) with the explicit representation, in an ontology, of how behavior change techniques can be

operationalized. The integration of these two components is explained, as well as the most challenging aspect

of using the advanced features of the platform in a way that allowed the agent to lead the dialogue ﬂow, when

needed. A successful proof of concept was built, which can be the basis for the development of advanced

conversational agents, combining natural language tools with ontology-based knowledge representation.

1 INTRODUCTION

We present a novel approach to the development of

conversational agents that combines the design and

development features offered by advanced natural

language tools with the use of knowledge needed to

support agents in the pursuit of their goals.

Besides the importance of building more ad-

vanced and versatile conversational agents, the drive

for this work comes from other two dimensions: the

importance of the role of conversational assistants in

healthcare (Guerreiro et al., 2021) and the incorpora-

tion of mechanisms that allow agents to induce behav-

ior change in their interlocutor.

In this work, we focus on how to provide agents

with the ability of operationalizing speciﬁc behavior

change techniques (BCTs), which are components of

an intervention designed to change behavior (Michie

et al., 2013).

The fact that noncommunicable diseases, like

Type 2 diabetes (T2D), account for a considerable

number of deaths (71% of deaths worldwide in 2016

https://orcid.org/0000-0002-5200-4981

https://orcid.org/0000-0002-4594-8087

https://orcid.org/0000-0001-8186-9506

https://orcid.org/0000-0001-8192-6080

https://orcid.org/0000-0002-4768-9517

https://orcid.org/0000-0001-8896-8152

(World Health Organization, 2021)) is a key indica-

tor of the importance of developing mechanisms that

help people preventing and managing this type of dis-

eases. Besides, the consistent growing availability of

mobile devices led to the development of thousands

of mHealth apps (mobile health applications); in a re-

cent scoping review (Wattanapisit et al., 2020), the

authors concluded that tasks such as ’disease-speciﬁc

care’ and ’health promotion’ can be successfully ful-

ﬁlled with the support of mHealth apps.

As behavior change interventions require multiple

interactions with a patient/user, the idea of a relational

agent (one that is designed to build and maintain long-

term social-emotional relationships) (Bickmore et al.,

2005) became crucial since the development of the

ﬁrst stage of our work (F

elix et al., 2019; Balsa et al.,

2020). In our previous work, we developed a rule-

based prototype of a mobile application with an intel-

ligent virtual assistant to be used in an intervention to

promote the self-care of older people with T2D.

In order to extend this work so that it could be used

in a wider range of situations (other chronic diseases,

multi-morbidity, or targeting diverse types of users,

for instance), we had to overcome two of its limita-

tions: the input from the user had to be chosen from

a limited set of options, and the fact that the BCTs

were to a greater than ideal extent hard coded in the

dialogues deﬁnition.

216

Bastos, M., Cláudio, A., Félix, I., Guerreiro, M., Carmo, M. and Balsa, J.

Operationalizing Behavior Change Techniques in Conversational Agents.

DOI: 10.5220/0010826800003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 1, pages 216-224

ISBN: 978-989-758-547-0; ISSN: 2184-433X

Figure 1: Simpliﬁed system’s architecture.

Continuing this work, we developed, and present

here, a new type of agent that allows us to overcome

the above mentioned limitations, while illustrating it

with BCTs operationalization.

As it is sketched in Figure 1, in which we present

a simpliﬁed view of the architecture, the use of Di-

alogﬂow

allowed us to deal with the ﬁrst limitation.

To tackle the second limitation, we added a module

that allows the explicit representation of knowledge

on BCTs operationalization.

The core sections of this paper will describe in de-

tail these components and how they interconnect. But

ﬁrst, in section 2, we highlight and discuss some re-

lated work. In section 3, we present the knowledge

component of our agent, namely the ontology that al-

lows us to better characterize BCTs operationaliza-

tion. In section 4, we describe the dialogue engine and

how Dialogﬂow was used. In section 5, we present

the detailed architecture and explain how the agent

works, illustrating it with a demo in section 6. Fi-

nally, in section 7, we present some conclusions and

directions for future work.

2 RELATED WORK

Our work derives from contributions in several di-

verse scientiﬁc areas. In Figure 2 we schematize how

those different areas directly or indirectly contribute

to our work.

The widespread use of mobile equipment across

people from all generations has been responsible for

the dissemination of mobile applications with diverse

purposes. Among these, a great number of develop-

ments were made in mHealth applications, the ones

that somehow “support the achievement of health ob-

jectives” (World Health Organization, 2011). A re-

A platform for the development of natural lan-

guage conversation interfaces (https://cloud.google.com/

dialogﬂow/)

Figure 2: Context of our contribution.

search carried out by Bhuyan et al. concluded that

60% of the adults who were mHealth apps users con-

sidered them useful in achieving health behavior goals

(Bhuyan et al., 2016). Moreover, mHealth applica-

tions are often used by people with little experience

of technology, as noted by Zapata et al. (Zapata et al.,

2015). So, mHealth offers a vast terrain of opportu-

nity to design digital solutions as tools to help im-

prove self-care, which can have a positive effect in

individual health outcomes and reduce the burden on

health systems. However, the conception of such ap-

plications entails high responsibility and the process

must be grounded on valid scientiﬁc studies.

The use of embodied conversational agents (ECA)

playing the role of virtual assistants has proved to be

well accepted by users in several health applications

(Baptista et al., 2020; Gong et al., 2020). In particu-

lar, Baptista et al. conducted a study to evaluate the

acceptability of an ECA (called Laura) to deliver self-

management education and support for patients with

T2D. The users are prompted to complete weekly in-

teractive sessions with Laura. This virtual assistant

provides education, feedback and motivational sup-

port for glucose level monitoring, taking medication,

physical activity, healthy eating and foot care. The

study had 66 respondents (mean age 55 years) and

the results identiﬁed positive reaction of the majority

of users for having a friendly, nonjudgmental, emo-

tional and motivational support provided by a human-

like character. Just around a third of them consid-

ered Laura not real, boring and annoying (Baptista

et al., 2020). Users respond to Laura by choosing

or by speaking out one of the options displayed in

the screen. The authors refer a “sophisticated script

logic” enabling the user to interact with Laura in sev-

eral predetermined variations. The solution was im-

plemented by a company that provides chatbots and

no details are given about these scripts. A more recent

12 months trial with 187 adults with T2D (age around

57) study resorting to the same solution, performed

during 12 months involving 187 adults suffering from

T2D (mean age 57), found a successful adoption of

Operationalizing Behavior Change Techniques in Conversational Agents

217

the program and a signiﬁcant improvement in partic-

ipants’ health-related quality of life (HRQOL) (Gong

et al., 2020).

A recent systematic review about Artiﬁcial

Intelligence-Based Conversational Agents for

Chronic Conditions (Schachner et al., 2020) revealed

the immaturity of the ﬁeld, using the authors’ own

words. They concluded that there is a lack of

evidence-based evaluation of the solutions, most

of them quasi-experimental studies. These authors

suggest that important future research would be

the deﬁnition of the AI architecture that should be

adopted and the adequate assessment process for the

overall solution.

As mentioned before, our research started some

time ago (Buinhas et al., 2019) with the development

of a ﬁrst prototype. Key project features were re-

sorting to an anthropomorphic assistant with a rel-

evant role in the interaction with the user, and to

support the intervention design in a well-deﬁned

theoretical framework, the Behavior Change Wheel

(Michie et al., 2014). This theoretical framework pro-

vides guidance to choose the more adequate behavior

change techniques in speciﬁc contexts.

A BCT is an observable, replicable, and irre-

ducible component of an intervention designed to

change behavior. They have been organized in a tax-

onomy (BCTTv1) (Michie et al., 2013; Cane et al.,

2015), which provides a deﬁnition for each of the

93 techniques. An example of a BCT is “problem

solving”; it requires users to pinpoint factors inﬂuenc-

ing the behavior and subsequently select strategies to

achieve it. After surveying 166 medication adherence

apps to ascertain whether they incorporated BCTs,

Morrissey et al. concluded that, from the 93 possi-

ble techniques only a dozen were found in the evalu-

ated apps (Morrissey et al., 2016). This result clearly

shows that more work is needed in incorporating evi-

dence on BCTs in available applications.

Although conversational agents (CA) have been

researched for decades, namely since the seminal

ELIZA (Weizenbaum, 1966), the shift towards be-

havior change is more recent (for instance, to pro-

mote a healthier lifestyle – typically regarding physi-

cal activity and type of diet). As reported by (Kramer

et al., 2020), CAs were found to have an important

value regarding the use of persuasive communication

in the health domain (Kramer et al., 2020), namely

when targeting coaching tasks. As Zhang et al. point

out, there is a “lack of understanding around the-

oretical guidance and practical recommendations on

designing AI chatbots for lifestyle modiﬁcation pro-

grams” (Zhang et al., 2020). In order to overcome

this, these authors developed an AI chatbot behav-

ior change model that has persuasive conversational

capacity as a central component. Our work has this

same feature, although the architecture we present

goes beyond the coverage of just two speciﬁc topics,

as Zhang and his colleagues do.

Regarding the use of ontologies, it is worth men-

tioning two recent works that, in a different way,

stress the importance of its use in the context of our

work.

Within the Human Behaviour-Change Project

(Michie et al., 2017), an ontology is being developed

for representing behavior change interventions (BCIs)

and their evaluation (Michie et al., 2021). Although

it focuses on the more general aspects that character-

ize BCIs, it clearly opens the possibility of a link to

our work, via the common BCT concept that, having

a different perspective, has in our work the grounds

for the representation of its operationalization.

Also relevant to us is the recent work on the de-

velopment of dialogue managers combining ontolo-

gies and planning. Teixeira et al. combine a conver-

sational ontology and Artiﬁcial Intelligence planning

to generate dialogue managers capable of performing

goal-oriented dialogues in the health domain (Teixeira

et al., 2021). As the authors note, in the health domain

it is critical to have predictable and reliable systems,

making a knowledge component crucial to guarantee

that, even if complemented, as it is in our case, with

other resources, namely for natural language under-

standing tasks.

2.1 Dialogue Engines

In order to create our conversational agent, we chose

to adopt an existing tool, since many excellent tools

are currently available. The two main features we

wanted were: the possibility of having natural lan-

guage interaction and some control of the dialogue

ﬂow, in order to address situations where the agent is

the one responsible for leading the dialogue (and not

the opposite, like what happens in question/answering

contexts, for instance).

Several dialogue tools were analyzed and tested

before choosing the most adequate. Due to space lim-

itation, we will just enumerate the tools we consid-

ered: Twine

, Yarn Spinner

, WOOL

, Watson As-

sistant

, FAtiMA Toolkit

, and Dialogﬂow.

As stated before, in this work, one of our main

interests was to make the dialogue more dynamic.

https://twinery.org/

https://yarnspinner.dev/

https://www.woolplatform.eu/

https://www.ibm.com/cloud/watson-assistant

https://fatima-toolkit.eu/

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

218

Dialogﬂow, being a natural language understanding

platform, has proven to be quite useful in that aspect,

making it easier to build and deliver the agent mes-

sages and capturing the user’s response. In previous

work, the dialogue portions corresponding to BCTs

operationalization were initially created in YARN

(Balsa et al., 2020). Besides the fact that deﬁning di-

alogs with that type of tool is a time-consuming task,

the deﬁnition of the whole logic of the interaction was

also harder than what we have with Dialogﬂow. Be-

sides, with this tool, the only user dialogue we have

to provide is the training phrases, and the agent will

learn from over time and usage.

3 AN ONTOLOGY OF BEHAVIOR

CHANGE CONCEPTS

One of the main limitations of the antecedent agent

was that BCT operationalization was too rigid, i.e. the

dialogue ﬂow was too dependent on the way the cor-

responding dialogues were deﬁned. In order to make

dialogues more natural and ﬂexible, we decided to

characterize BCT operationalization in a more gen-

eral way. For that, we needed to identify the main

concepts involved and the relations between them.

We did this by deﬁning an ontology representing the

relevant behavior change intervention concepts. Since

the operationalization of a BCT depends also on user

speciﬁc information, in order to characterize a spe-

ciﬁc way of operationalizing a BCT we had to incor-

porate concepts related to the user’s characteristics.

The resulting ontology can be divided in two main

class entities: Behavior Change Intervention (BCI)

and User. The ﬁrst one, BCI, includes several classes

representing both general concepts (like BCT or Be-

havior determinant – see below) and speciﬁc ones

(like Food Topic, representing the relevant topics re-

lated to a healthy diet). As a starting point, our illus-

trative domain is healthy nutrition, as it should be an

universal concern, independent of age or health con-

dition. The User class includes concepts critical to

the operationalization, like the age category (adult, se-

nior, . . . ) or the identiﬁcation of some risk condition.

Some of these entities are related by six object

properties: has active goals, has active topic, re-

lated to, targeted during, triggered by, and triggers.

For further characterization, there are also eight data

properties: BCT order, has age, has BMI (Body Mass

Index), has competence score, has genre, has height,

has risk level, and has weight.

The classes that the current work focuses on and

are: Behavior Change Intervention (an intervention

that has the aim of inﬂuencing human behavior);

Behavior Change Technique (an observable, replica-

ble, and irreducible component of an intervention de-

signed to alter or redirect causal processes that regu-

late behavior — an “active ingredient” on an inter-

vention (Michie et al., 2013); Behavior Determinant

(a factor that inﬂuences positively or negatively a be-

havior). For instance, a reason of non-adherence to

the user’s agreed goal – lack of motivation, forget-

fulness, . . . ); Behavior Goal (a goal deﬁned in terms

of the behavior to be achieved); Behavior Topic (a

topic that is targeted during the intervention, aimed to

help achieve/maintain the desired behaviors); Opera-

tionalization (the act of delivering one or more BCTs,

based on several conditions); User (a person who’s

being subjected to the Behavior Change Intervention).

3.1 How the Ontology Is Used

This ontology can be used for several purposes within

the development of a dynamic Behavior Change In-

tervention, making it easier to connect the concepts

involved and shape the structure of the intervention.

When creating an application aimed at inﬂuencing

the human behavior, the use of an ontology can be

useful to detach the logic of the application from the

speciﬁc data (the individuals and their relationships).

In order to understand better, the following para-

graphs describe a scenario where the ontology is used

during the interaction with the user. Figure 3 shows

the classes and relationships that are relevant to the

scenario described.

During Review Tasks (one of the dialogue stages,

as explained in the next section), the user is asked

whether he/she completed or not a previously agreed

goal. When the user answers negatively, the agent

will ask about the reason for not meeting the goal, and

the reason the user gives is denominated determinant.

The determinants of non-adherence can be as simple

as lack of motivation, or reluctance in changing their

habits, so the individuals of the class Behavior Deter-

minant are key words of such motives (so far, the de-

terminants included are motivation, habits, difﬁculty,

appetite, and loneliness). These determinants are ex-

amples and are not related with a particular behavior.

Figure 3: Ontology entities relevant to BCTs operational-

ization.

Operationalizing Behavior Change Techniques in Conversational Agents

219

After the user gives an answer (providing a deter-

minant), the agent queries the ontology to see which

operationalization related to the active goal is trig-

gered by that determinant. When the right opera-

tionalization is identiﬁed, the agent chooses the right

BCT to trigger. If it is the case that the operationaliza-

tion is sequential, the agent also determines the trig-

gering order. After that, it searches the dialogue ﬁle

and fetches the dialogue that is supposed to be deliv-

ered to complete the execution of a speciﬁc BCT.

4 DIALOGUE ENGINE

As stated before, we chose to use Dialogﬂow CX as

the dialogue engine. In this section, we describe how

our Dialogﬂow agent is built and used, along with the

details regarding the training process and the way we

deal with the situations where the agent takes control

of the dialogue.

4.1 How Dialogﬂow Is Used

Dialogﬂow CX has a visual builder in its console,

where conversation paths are graphed as a state ma-

chine model, making it easier to design, enhance, and

maintain. Conversation states and state transitions

are ﬁrst-class types that provide explicit and powerful

control over conversation paths. We can clearly de-

ﬁne a series of steps that we want the end-user to go

through. Following the work of Bickmore’s relational

agents group (Bickmore et al., 2005), dialogues are

organized in eight steps: opening, social talk, review

tasks, assess, counseling, assign tasks, pre-closing

and closing. The visual state machine created for our

agent is represented in Figure 4.

Each state is represented by a page, that can be

conﬁgured to collect information from the end-user

that is relevant for the conversational state represented

by the page. For simpler dialogue steps, such as

opening, social talk, assess, pre-closing and closing,

we conﬁgured the respective pages to give static text

responses, since there are no parameters or external

conditions that they depend on. For more complex

dialogue steps, such as review tasks, counseling and

assign tasks, we conﬁgured their respective pages to

enable webhook, in order to provide a dynamic re-

sponse based on external conditions.

To each page can be added state handlers, that

are used to control the conversation by creating re-

sponses for the end-users and/or by transitioning the

current page. There are two types of state handlers:

routes and event handlers. Routes are called when

an end-user input matches an intent (categorizes an

Figure 4: Conversation path of the Dialogﬂow agent.

end-user’s intention for one conversation turn) and/or

some condition.

Figure 5 shows the routes in our Counselling page.

The ﬁrst three routes (with the blue left border) are

intent routes, which means that either one of those

routes are called when the user input matches one of

those intents. The last two routes (with the orange

left border) are condition routes, that are called if the

respective condition is satisﬁed. At the end of a route,

there’s a Transition ﬁeld, which deﬁnes the next page

in the conversation.

In other resources, and in Dialogﬂow ES, an intent

usually contains a ﬁeld with the training phrases and

Figure 5: Counselling page.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

220

a ﬁeld with the agent response, that is sent back to the

user if their input matches the training phrases for that

intent. During an interaction, a user can say “yes” or

“no” multiple times, which means that, with this type

of intent, every time the user can say “yes” or “no”

during the conversation, there would have to be an in-

tent with the same training phrases but different agent

response, depending on the context. In Dialogﬂow

CX, an intent only contains the logic to detect what

the user says (the training phrases), which means an

intent can be reused in multiple places of a conversa-

tion.

4.2 Training Process

When an agent is trained, Dialogﬂow uses the train-

ing data to build machine learning models speciﬁcally

for that agent. This training data primarily consists of

intents, intent training phrases, and entities referenced

in a ﬂow, which are effectively used as machine learn-

ing data labels. However, agent models are built using

parameter prompt responses, state handlers, agent set-

tings, and many other pieces of data associated with

the agent.

In the agent settings, it can be chosen to train

the agent automatically or manually. By default, the

training is executed automatically, showing a popup

dialogue in the console every time there’s an update

of the ﬂow.

4.3 Leading the Dialogue

Technologies like Dialogﬂow are generally used in

virtual agents meant to handle questions from the

user. In our case, it is important that the agent is

the one taking control of the conversation, asking the

questions, and delivering the content in a “doctor-

patient” type of way, i.e. with the doctor leading the

interaction.

On previous work, after the agent’s response, the

user interface had buttons for the user to choose their

answer, and during longer explanations or a change

in the dialogue phase, it would present a “Continue”

button (Balsa et al., 2020). When the user has total

control over the input, we can not expect them to ca-

sually answer that. The agent must seek the user’s

attention and interest, keeping the user engaged in the

conversation.

In Dialogﬂow CX, all the agent responses are han-

dled in page, and a page can have an entry fulﬁllment,

and a static response for each route. The entry fulﬁll-

ment is optional, and it is what the agent will respond

to the end-user when a page initially becomes active.

For each route added to a page, there is a fulﬁllment

ﬁeld, where it is possible to add several types of re-

sponse messages, although, in this work we only used

text response messages.

The entry fulﬁllment feature is a great advantage

since it makes it possible for the agent to say some-

thing to the user without the need to have a previous

input. This gives the agent more initiative and con-

trol over the conversation, being extremely useful be-

tween dialogue phases.

5 AGENT ARCHITECTURE

The architecture of our agent has comprises three

main components: the Core, the Dialogﬂow Engine,

and the Ontology (Figure 6). The Core controls the

interface and, along with the Dialogﬂow Engine, con-

trols the ﬂow of execution and the speech of the agent.

The Ontology holds external data that can be queried

by the Core whenever necessary.

Figure 6 shows the steps that take place for one

conversational turn of a session:

1. The end-user types something, known as user in-

put.

2. The user interface (UI), responsible for the view

provided to the user, receives the input and for-

wards it to the Dialogﬂow API in a detect intent

request (handled by the Application Controller).

3. The Dialogﬂow API receives the detect intent re-

quest. It matches the user input to an intent or

form parameter, sets parameters as needed, and

updates the session state. In case it needs to call

a webhook-enabled fulﬁllment, it sends a web-

hook request to the Webhook Service, otherwise

it jumps straight to step 6.

4. The Webhook Service receives the webhook re-

quest and it takes any actions necessary, such as

querying the ontology and/or fetching dialogue

from external sources (JSON ﬁles).

5. The Webhook Service builds a response and sends

a webhook response back to Dialogﬂow.

6. Dialogﬂow creates a detect intent response. If

a webhook was called, it uses the response pro-

vided in the webhook response. If no webhook

was called, it uses the static response deﬁned in

the Dialogﬂow Agent. The detect intent response

is send to the user interface.

7. The user interface receives the detect intent re-

sponse and forwards the text response to the end-

user.

8. The end-user sees the response.

Operationalizing Behavior Change Techniques in Conversational Agents

221

Figure 6: Agent architecture.

In short, when a user submits a message, it is sent

to Dialogﬂow to detect the intent of the user. Di-

alogﬂow will process the text, then send back a ful-

ﬁllment response (either static, or dynamic, by means

of the Webhook Service).

6 DEMO

During a few dialogue steps, there are speciﬁc BCTs

that are always operationalized in the same way,

therefore being simpler to execute. For example, dur-

ing Review Tasks, the BCT Self-monitoring of behav-

ior (2.3)

is always executed and delivered in the same

way: by collecting data on the user’s behavior. To

do that, our system takes advantage of the following

ontology classes: User and Behavior Goal. Linking

those classes is an object property, labeled has active

goals, that connects one or more behavior goals to a

speciﬁc user. That way, the system can easily get ac-

cess to the goal that the user agreed on, and through

its label, ask if they completed it or not.

During Assess and Counseling, the procedure is

a bit more complex, since the BCTs operationalized

on those dialogue steps are executed in several differ-

ent ways, depending on more than just one condition.

The numbers next to the names of the BCTs refer to the

codes used in the taxonomy (Michie et al., 2013).

Following the case mentioned in the previous para-

graph, after Review Tasks, comes Assess, and during

that step, the system takes advantage of the follow-

ing classes: Operationalization, User, Behavior Goal,

Behavior Determinant and Behavior Change Tech-

nique. As was mentioned in Section 3.1, when the

system gets the non-adherence determinant, it queries

the ontology in order to ﬁnd which operationalization,

related to the active goal, is triggered by that given de-

terminant. After that, the system accesses the object

property triggers, to see which BCTs are triggered in

that operationalization. In case the operationalization

is of the Sequential type (a complex operationaliza-

tion where the BCTs are delivered in sequential or-

der), the system accesses the data property BCT order,

to be able to deliver the BCTs in the suitable order.

Figure 7 illustrates an output demonstration of the

examples given in the previous paragraphs. The agent

asks the user if they completed the agreed goal ( “have

at least three main meals”, highlighted in green on the

ﬁrst chat message on the left upper corner), to which

the user answers negatively. After that, there is the

operationalization of another simpler BCT, Feedback

on behavior (2.2), statically delivered by Dialogﬂow

since it only depends on the intent (“Yes” or “No”).

The user answers saying “I don’t have motivation”, to

which the system catches the “motivation” behavior

determinant. The operationalization related to the ac-

tive goal, and triggered by motivation, is the individ-

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

222

Figure 7: An example dialogue.

ual OP Alim Det1 2, which triggers the BCTs Prob-

lem solving (1.2) and Information of health conse-

quences (5.1), having “1.2, 5.1” as BCT order. Given

those conditions, the system accesses the dialogue ﬁle

and extracts the dialogue parts related to those BCTs,

ordering them accordingly and sending them to the

user. In the last chat message on the lower left corner

of Figure 7, highlighted in dark blue is the dialogue

delivering Problem solving (1.2), and higlighted in

purple is the dialogue part that delivers Information

of health consequences (5.1).

7 CONCLUSIONS

The main goal of this work was achieved with the de-

sign and implementation of a novel architecture for

the development of conversational agents for behav-

ior change interventions. This architecture allows the

combination of an advanced dialogue engine with a

learning ability (DialogFlow) with the representation

of knowledge on the operationalization of behavior

change techniques, by means of deﬁning an ontology

of behavior change intervention concepts.

The design was explained and the functioning of

the agent illustrated.

Regarding future work, two immediate steps fol-

low: the incorporation of additional knowledge so

that the agent capacity can be enlarged; the additional

training of the system with a set of dialogues that were

developed in a previous work.

Additionally, we intend to incorporate in the de-

sign mechanisms that will allow the consideration of

an ethical dimension in this type of agents. As rec-

ognized by Zhang and her colleagues (Zhang et al.,

2020), the ethical dimensions regarding conversa-

tional agents development has been completely ab-

sent. But, in contexts where the goal is to induce be-

havior change in humans, the incorporation of ethical

principles and the insurance of responsibility in the

systems’ designers is of paramount importance.

ACKNOWLEDGEMENTS

This work was supported by FCT through the

LASIGE Research Unit, ref. UIDB/00408/2020 and

ref. UIDP/00408/2020.

REFERENCES

Balsa, J., F

elix, I., Cl

audio, A. P., Carmo, M. B., Silva, I.

C. e., Guerreiro, A., Guedes, M., Henriques, A., and

Guerreiro, M. P. (2020). Usability of an Intelligent

Virtual Assistant for Promoting Behavior Change and

Self-Care in Older People with Type 2 Diabetes. Jour-

nal of Medical Systems, 44(7):130.

Baptista, S., Wadley, G., Bird, D., Oldenburg, B., Speight,

J., and Group, T. M. D. C. R. (2020). Acceptabil-

ity of an Embodied Conversational Agent for Type

2 Diabetes Self-Management Education and Support

via a Smartphone App: Mixed Methods Study. JMIR

mHealth and uHealth, 8(7):e17038.

Bhuyan, S. S., Lu, N., Chandak, A., Kim, H., Wyant, D.,

Bhatt, J., Kedia, S., and Chang, C. F. (2016). Use of

Mobile Health Applications for Health-Seeking Be-

havior Among US Adults. Journal of Medical Sys-

tems, 40(6):153.

Bickmore, T. W., Caruso, L., Clough-Gorr, K., and Heeren,

T. (2005). ‘It’s just like you talk to a friend’ relational

agents for older adults. Interacting with Computers,

17(6):711–735.

Buinhas, S., Cl

audio, A. P., Carmo, M. B., Balsa, J.,

Cavaco, A., Mendes, A., F

elix, I., Pimenta, N., and

Guerreiro, M. P. (2019). Virtual Assistant to Improve

Self-care of Older People with Type 2 Diabetes: First

Prototype. In Garc

ıa-Alonso, J. and Fonseca, C., edi-

tors, Gerontechnology, Communications in Computer

and Information Science, pages 236–248. Springer In-

ternational Publishing.

Cane, J., Richardson, M., Johnston, M., Ladha, R., and

Michie, S. (2015). From lists of behaviour change

techniques (BCTs) to structured hierarchies: Com-

parison of two methods of developing a hierarchy

of BCTs. British Journal of Health Psychology,

20(1):130–150.

elix, I. B., Guerreiro, M. P., Cavaco, A., Cl

audio, A. P.,

Mendes, A., Balsa, J., Carmo, M. B., Pimenta, N.,

and Henriques, A. (2019). Development of a Com-

plex Intervention to Improve Adherence to Antidia-

betic Medication in Older People Using an Anthro-

pomorphic Virtual Assistant Software. Frontiers in

Pharmacology, 10:680.

Operationalizing Behavior Change Techniques in Conversational Agents

223

Gong, E., Baptista, S., Russell, A., Scuffham, P., Riddell,

M., Speight, J., Bird, D., Williams, E., Lotfaliany, M.,

and Oldenburg, B. (2020). My Diabetes Coach, a Mo-

bile App–Based Interactive Conversational Agent to

Support Type 2 Diabetes Self-Management: Random-

ized Effectiveness-Implementation Trial. Journal of

Medical Internet Research, 22(11):e20322.

Guerreiro, M. P., Angelini, L., Rafael Henriques, H., El Ka-

mali, M., Baixinho, C., Balsa, J., F

elix, I. B., Khaled,

O. A., Carmo, M. B., Cl

audio, A. P., Caon, M., Da-

her, K., Alexandre, B., Padinha, M., and Mugellini, E.

(2021). Conversational Agents for Health and Well-

being Across the Life Course: Protocol for an Evi-

dence Map. JMIR Research Protocols, 10(9):e26680.

Kramer, L. L., ter Stal, S., Mulder, B. C., de Vet, E., and

van Velsen, L. (2020). Developing Embodied Con-

versational Agents for Coaching People in a Healthy

Lifestyle: Scoping Review. Journal of Medical Inter-

net Research, 22(2):e14058.

Michie, S., Atkins, L., and West, R. (2014). The behaviour

change wheel : a guide to designing interventions. Sil-

verback.

Michie, S., Richardson, M., Johnston, M., Abraham, C.,

Francis, J., Hardeman, W., Eccles, M. P., Cane, J., and

Wood, C. E. (2013). The Behavior Change Technique

Taxonomy (v1) of 93 Hierarchically Clustered Tech-

niques: Building an International Consensus for the

Reporting of Behavior Change Interventions. Annals

of Behavioral Medicine, 46(1):81–95.

Michie, S., Thomas, J., Johnston, M., Aonghusa, P. M.,

Shawe-Taylor, J., Kelly, M. P., Deleris, L. A.,

Finnerty, A. N., Marques, M. M., Norris, E., O’Mara-

Eves, A., and West, R. (2017). The Human Behaviour-

Change Project: harnessing the power of artiﬁcial

intelligence and machine learning for evidence syn-

thesis and interpretation. Implementation Science,

12(1):121.

Michie, S., West, R., Finnerty, A., Norris, E., Wright, A.,

Marques, M., Johnston, M., Kelly, M., Thomas, J.,

and Hastings, J. (2021). Representation of behaviour

change interventions and their evaluation: Develop-

ment of the Upper Level of the Behaviour Change

Intervention Ontology. Wellcome Open Research,

5(123).

Morrissey, E. C., Corbett, T. K., Walsh, J. C., and Molloy,

G. J. (2016). Behavior Change Techniques in Apps

for Medication Adherence. American Journal of Pre-

ventive Medicine, 50(5):e143–e146.

Schachner, T., Keller, R., and v Wangenheim, F. (2020). Ar-

tiﬁcial Intelligence-Based Conversational Agents for

Chronic Conditions: Systematic Literature Review.

Journal of Medical Internet Research, 22(9):e20701.

Teixeira, M. S., Maran, V., and Dragoni, M. (2021). The in-

terplay of a conversational ontology and AI planning

for health dialogue management. In Proceedings of

the 36th Annual ACM Symposium on Applied Comput-

ing, pages 611–619, Virtual Event Republic of Korea.

ACM.

Wattanapisit, A., Teo, C. H., Wattanapisit, S., Teoh, E.,

Woo, W. J., and Ng, C. J. (2020). Can mobile health

apps replace GPs? A scoping review of comparisons

between mobile apps and GP tasks. BMC Medical In-

formatics and Decision Making, 20(1):5.

Weizenbaum, J. (1966). ELIZA—a computer program for

the study of natural language communication between

man and machine. Communications of the ACM,

9(1):36–45.

World Health Organization (2011). mHealth : new horizons

for health through mobile technologies. World Health

Organization, Geneva.

World Health Organization (2021). Noncom-

municable diseases: Mortality. [online].

https://www.who.int/data/gho/data/themes/topics/

topic-details/GHO/ncd-mortality [accessed 2021-09-

28].

Zapata, B. C., Fern

andez-Alem

an, J. L., Idri, A., and Toval,

A. (2015). Empirical Studies on Usability of mHealth

Apps: A Systematic Literature Review. Journal of

Medical Systems, 39(2):1.

Zhang, J., Oh, Y. J., Lange, P., Yu, Z., and Fukuoka,

Y. (2020). Artiﬁcial Intelligence Chatbot Behavior

Change Model for Designing Artiﬁcial Intelligence

Chatbots to Promote Physical Activity and a Healthy

Diet: Viewpoint. Journal of Medical Internet Re-

search, 22(9):e22845.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

224