Enhancing Post-Incarceration Support: A Custom Chatbot Solution for

the Brazilian Prison System

Geovana Ramos Sousa Silva

, Lurian Correia Lima, Guilherme Pereira Paiva

and Edna Dias Canedo

University of Bras

ılia (UnB), Department of Computer Science, Bras

ılia–DF, Brazil

Keywords:

Chatbot, Conversational Interface, Social Support.

Abstract:

This paper presents the development and implementation of a dedicated chatbot to assist former inmates of the

Brazilian prison system, integrated into the ESVirtual application. This initiative addresses the multifaceted

challenges faced by this vulnerable group by providing a comprehensive and sustainable approach to support

their social reintegration. The proposed architecture leverages Docker Compose, Rasa NLU, Stories, and

Actions, creating a robust and scalable framework capable of understanding natural language, adapting to

diverse interaction scenarios, and executing customized actions. By integrating the chatbot into ESVirtual, we

enhance its utility and accessibility, offering former inmates a reliable and accessible channel for obtaining

information and support. The decision to manually craft the chatbot’s content, rather than using generative AI,

ensures the accuracy, relevance, and reliability of the provided information, allowing for quick adaptation to

changes in policies, legislation, and available services. Socially, this project aims to signiﬁcantly contribute

to the reintegration of former inmates into society, reducing recidivism rates and fostering a more just and

inclusive community. By providing essential support and resources, the chatbot empowers individuals to

overcome the challenges they face after leaving the prison system and to build digniﬁed and productive lives.

This work underscores the potential of technology and innovation to promote social well-being and justice,

marking a signiﬁcant step towards a more humanized approach to former inmates reintegration in Brazil.

1 INTRODUCTION

In Brazil, the issue of reintegrating former inmates is

a multifaceted challenge that requires a holistic and

innovative approach. The transition of individuals

who have served their sentences back into society is

often marked by a series of obstacles, ranging from

stigmatization to lack of access to employment oppor-

tunities, adequate housing, education, and healthcare

(Storck, 2023; Tharshini et al., 2024). These chal-

lenges signiﬁcantly contribute to the high recidivism

rates in the country.

In this context, there is a need to develop effec-

tive solutions that can support former inmates in their

journey of social reintegration. Among these solu-

tions, technology plays a crucial role, providing ac-

cessible and scalable tools that can meet the diverse

needs of this vulnerable group (Dores and Dores,

2019).

https://orcid.org/0000-0002-0304-0804

https://orcid.org/0000-0002-2159-339X

This project proposes the development of a chat-

bot dedicated exclusively to providing information

and support to former inmates of the Brazilian prison

system and introduced through an existing app for

their assistance (Canedo et al., 2024). A chatbot is

an artiﬁcial intelligence application designed to simu-

late human conversations, offering automatic and in-

teractive responses to users’ questions and requests

(Al Husaeni et al., 2024).

By using a chatbot, it is expected to offer former

inmates a reliable source of information and guid-

ance, available at any time and place through mobile

devices or computers connected to the internet. This

is particularly relevant given the difﬁculties former in-

mates face in accessing traditional support and coun-

seling services, often limited by geographical, ﬁnan-

cial, or social stigma barriers (Benard et al., 2023).

The proposed chatbot will provide a variety of

resources and functionalities, including guidance on

their legal rights, information on reintegration pro-

grams, access to employment and education oppor-

tunities, emotional support, and connection to rele-

Silva, G. R. S., Lima, L. C., Paiva, G. P. and Canedo, E. D.

Enhancing Post-Incarceration Support: A Custom Chatbot Solution for the Brazilian Prison System.

DOI: 10.5220/0013086800003929

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 423-433

ISBN: 978-989-758-749-8; ISSN: 2184-4992

423

vant social services. By offering this comprehensive

support, the chatbot aims to empower former inmates,

helping them overcome the challenges they face dur-

ing the reintegration process. It is believed that by

promoting autonomy, conﬁdence, and access to es-

sential resources, the chatbot can help its users to rein-

tegrate themselves into society and reduce recidivism.

2 BACKGROUND

2.1 ESVirtual

In 2021, the CNJ launched the Social Ofﬁce Virtual

(ESVirtual) app (UNDP, 2024), which aims to ex-

pand the reach and enhance the availability of ser-

vices for individuals reintegrating from the prison

system. It complements the in-person assistance pro-

vided at Social Ofﬁces, which are present in 17 states,

and the psychosocial services offered prior to release,

especially during the Covid-19 pandemic (Canc¸ado

et al., 2022). The app uses georeferencing to help

users locate social public service facilities, including

healthcare units, social services, shelters, community

restaurants, legal aid centers, and civil documentation

ofﬁces (Portal CNJ, 2022).

The app is also integrated with the Uniﬁed Elec-

tronic Execution System (SEEU), allowing former in-

mates and their families to track the progress of their

sentences quickly and easily through the app (Portal

CNJ, 2022). Additionally, it offers free courses in

various ﬁelds of knowledge and content speciﬁcally

focused on job and income generation. Beyond its

diverse range of services, the app aims to boost the

implementation of physical Social Ofﬁces, which are

currently established in over 20 municipalities across

the country. The Social Ofﬁce Virtual app is available

for both Android and iOS devices (PlayStore, 2024)

and its interface is shown in Figure 1.

2.2 Chatbot

A chatbot is an artiﬁcial intelligence (AI) software

designed to simulate human-like conversation with

users through text or voice interactions (Al Husaeni

et al., 2024; Silva and Canedo, 2024; Gonc¸alves et al.,

2024). Chatbots are increasingly popular due to their

ability to provide instant responses, handle multiple

queries simultaneously, and operate 24/7, thereby en-

hancing user experience and operational efﬁciency

(Misischia et al., 2022). There are two primary types

of chatbots: generative and retrieval-based (Pandey

and Sharma, 2023).

Figure 1: ESVirtual app.

Generative chatbots use machine learning algo-

rithms, particularly neural networks, to generate re-

sponses in real-time (Pandey and Sharma, 2023).

These chatbots are capable of producing more ﬂex-

ible and varied responses as they do not rely on pre-

deﬁned scripts. However, they require extensive train-

ing data and computational power, and their responses

can sometimes be unpredictable or less accurate.

Retrieval-based chatbots, on the other hand, oper-

ate by selecting the most appropriate response from a

set of predeﬁned replies (Bachtiar et al., ). They lever-

age machine learning models to match user queries

with the best possible answers from a database. While

they may lack the ﬂexibility of generative models,

retrieval-based chatbots are generally more reliable

and easier to control, making them suitable for many

practical applications.

One of the most widely used frameworks for

building retrieval-based chatbots is Rasa. Rasa is an

open-source framework that provides tools for devel-

oping and managing conversational AI applications

(Joshi and Sharma, 2020). It supports both intent

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

424

recognition and entity extraction, enabling developers

to create sophisticated dialogue systems. The choice

of Rasa can be justiﬁed by its robust architecture, ac-

tive community support, and the ﬂexibility it offers in

customizing and deploying chatbots (Dagkoulis and

Moussiades, 2022). Additionally, Rasa’s ability to in-

tegrate with various messaging platforms and APIs

makes it a versatile choice for building chatbots tai-

lored to speciﬁc business needs.

2.3 Related Work

Brazil has seen a growing interest in the development

of chatbots aimed at addressing various social issues,

particularly in health, education, and community en-

gagement. These initiatives leverage artiﬁcial intel-

ligence and natural language processing to facilitate

communication and provide essential services to di-

verse populations.

In response to the COVID-19 pandemic, a chatbot

was developed by a large telehealth service in Brazil

to assist with screening and monitoring (Chagas et al.,

2023). This chatbot provided users with evidence-

based information regarding COVID-19 symptoms

and facilitated web-based assessments of symptom

severity. Evaluations conducted through a mixed-

methods approach highlighted the usability of the

chatbot, revealing strengths and areas for improve-

ment based on user feedback. This initiative show-

cased the potential of chatbots to deliver timely health

information and support public health efforts during a

crisis. Key challenges included ensuring the chatbot

provided accurate, up-to-date information and creat-

ing a user-friendly interface for a wide range of users.

Adapting the chatbot to the Portuguese language and

cultural context was likely crucial for its success.

Another signiﬁcant project is the Amanda Selﬁe

chatbot (Massa et al., 2023), designed speciﬁcally to

create demand for pre-exposure prophylaxis (PrEP)

among adolescent men who have sex with men

(AMSM) and transgender women in Brazil. Devel-

oped through a participatory process that included in-

put from the target demographic, Amanda Selﬁe func-

tions as a virtual peer educator. The chatbot’s design

reﬂects cultural sensitivity, as it embodies a Black

transgender woman, making it relatable to its audi-

ence. The development process spanned 21 months

and involved rigorous testing to assess acceptabil-

ity, functionality, and usability. The results indicated

that Amanda effectively increased awareness and de-

mand for PrEP services among adolescents, illustrat-

ing the power of targeted digital interventions in pub-

lic health.

The work of (Tueiv and Schmitz, 2023) focused

on optimizing the value delivered by Chatbots, a cru-

cial tool in the digitalization of government services.

By applying the Incremental Funding Method (IFM)

alongside the Analytic Hierarchy Process (AHP)

technique, the study developed a method tailored to

the unique characteristics of Chatbot and Machine

Learning projects. The approach demonstrated that,

using a customized optimization delivery sequencing

method, Chatbot-enabled services, such as those im-

plemented in the Institute of Social Security and Pen-

sion of Brazil (INSS) application, couldcan enhance

the efﬁciency and effectiveness of government ser-

vices. This research highlights the signiﬁcant role of

Chatbots in advancing digital inclusion and accelerat-

ing the transformation of government services to bet-

ter serve citizens.

IAraBot is an architectural model and proactive

chatbot developed to detect misinformation in real-

time on WhatsApp and Telegram in Brazil (Cacabro

et al., 2023). The chatbot aims to provide media

education and help users critically evaluate content.

Key challenges include accurately identifying misin-

formation, providing relevant examples and explana-

tions in Portuguese, and encouraging users to develop

critical thinking skills. Adapting the chatbot’s lan-

guage and approach to different age groups and edu-

cation levels is also important.

The work of (Andrade et al., 2020b) presents a

novel approach to enhancing user support for the

EV.G platform through the development of a special-

ized chatbot. This chatbot is designed to handle fre-

quently asked questions and reduce the burden on

human support staff by utilizing an architecture that

adapts to new information and user requests. Lever-

aging open-source tools customized for EV.G’s needs,

the chatbot, named Eva, aims to improve service qual-

ity by providing rapid responses, which is crucial

for customer satisfaction. Notably, EvaTalk is tasked

with understanding and processing Portuguese writ-

ten text, including its variations and informal usage

such as shortcuts, misspellings, and acronyms often

encountered in online communication. These chal-

lenges are underscored by the need for the chatbot

to handle the diverse linguistic expressions used by

users, which deviates from standard Portuguese and

impacts the coherence of interactions. Despite its ca-

pabilities in managing repetitive queries, EvaTalk’s

effectiveness is contingent on addressing these lan-

guage complexities and ensuring that more intricate

issues are escalated to human support.

These examples highlight the innovative use of

chatbots in Brazil for social purposes, demonstrat-

ing their potential to enhance communication, pro-

mote health awareness, and improve access to essen-

Enhancing Post-Incarceration Support: A Custom Chatbot Solution for the Brazilian Prison System

425

tial services. Developing chatbots for Portuguese in-

volves overcoming a range of challenges due to the

language’s diverse dialects, regional variations, and

cultural nuances (D’

Avila, 2018). Portuguese is spo-

ken across multiple countries, each with distinct vo-

cabularies and idiomatic expressions, making it difﬁ-

cult for chatbots to universally understand and gen-

erate appropriate responses. These challenges are

compounded by varying cultural expectations; for in-

stance, among Brazilian users and Portuguese users

(Thom

e-Williams, 2004). Still, language-specialized

models have a tendency to perform worse than multi-

lingual models (Shaitarova et al., 2023).

Additionally, the grammatical complexities of

Portuguese, such as gendered nouns and intricate verb

conjugations (Wilkens et al., 2024), pose signiﬁcant

hurdles for chatbots in accurately interpreting and re-

sponding to user inputs. The scarcity of high-quality,

annotated datasets for Portuguese further hampers the

development of robust natural language processing

models (Trajano et al., 2023). Most existing chatbot

frameworks and AI tools are designed primarily for

English (Liang et al., 2023), requiring developers to

invest extra time and resources to adapt these tools

for Portuguese. This adaptation process can increase

development costs and extend deployment timelines,

complicating the creation of effective chatbot solu-

tions.

3 CHATBOT PROPOSAL

This section presents the chatbot proposal, outlining

the key requirements and the proposed architecture

designed to meet these needs. It details the speciﬁc

functional and technical requirements that the chatbot

must fulﬁll. Following this, it presents the proposed

architecture, which includes a robust framework and

technology stack intended to support these require-

ments efﬁciently. This section will provide a compre-

hensive overview of how the proposed design aligns

with the objectives and operational demands, ensur-

ing a seamless and effective implementation of the

chatbot.

3.1 Requirements

The project proposes an innovative and comprehen-

sive solution to address the challenges faced by for-

mer inmates in Brazil by integrating a dedicated chat-

bot into the ESVirtual application (Canedo et al.,

2024). ESVirtual is already recognized as a robust

platform aimed at providing support and resources for

people in vulnerable situations, and the inclusion of a

specialized chatbot will further expand its utility and

effectiveness in supporting former inmates.

To ensure accessibility and ﬂexibility in imple-

mentation, the chatbot will be developed using a free

open-source framework. This strategic choice will not

only reduce costs but also allow for smoother inte-

gration with the existing ESVirtual app infrastructure.

Moreover, utilizing an open-source framework offers

the advantage of an active developer community and

constantly evolving resources, ensuring the continu-

ous maintenance and improvement of the chatbot over

time.

The chatbot’s scope will be comprehensive, ad-

dressing a variety of crucial topics for former inmates.

This will include information and guidance on shelter,

food, access to healthcare services, ways to contact

support organizations, available beneﬁts (such as so-

cial programs and ﬁnancial assistance), professional

training opportunities, and legal assistance. By offer-

ing support in a wide range of areas, the chatbot aims

to meet the diverse needs of former inmates, provid-

ing them with tangible resources to help them reinte-

grate into society.

A distinctive aspect of this project is the decision

not to use generative artiﬁcial intelligence to gener-

ate responses. Instead, the chatbot’s responses will be

crafted and managed by the project team. This ap-

proach was adopted to ensure the accuracy, relevance,

and consistency of the information provided, as well

as to enable quick responses to changes in policies,

legislation, or services available to former inmates.

Although this requires additional effort from the team,

we believe it is crucial to ensure the quality and relia-

bility of the service offered.

3.2 Architecture

The proposed architecture for the chatbot integrated

with the ESVirtual application is designed to be mod-

ular, scalable, and easy to maintain. To achieve these

goals, we opted to use Docker Compose, a tool that

simpliﬁes and standardizes the deﬁnition and execu-

tion of multi-container Docker applications. There

will be containers for Rasa and the Action Server. The

full architecture is presented in Figure 2.

Docker Compose is a tool that allows for the

easy and consistent deﬁnition and execution of multi-

container Docker applications. It uses a YAML ﬁle

to conﬁgure the application’s services and dependen-

cies, ensuring that the development environment is

replicable and consistent across different machines.

This simpliﬁes the installation, update, and mainte-

nance of the chatbot, while also allowing for more

efﬁcient scalability by adding or removing containers

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

426

Figure 2: Architecture of the proposed chatbot.

as needed.

Rasa is an open-source framework for building

chatbots based on machine learning and natural lan-

guage processing (NLP). It enables the development

of highly customized chatbots capable of understand-

ing and responding to questions contextually. The

Rasa chatbot will be the core of our solution, pro-

viding interaction and response capabilities to users.

Rasa allows for the creation of machine learning mod-

els speciﬁcally trained for the chatbot’s context, en-

suring that responses are relevant and accurate. Ad-

ditionally, Rasa supports the implementation of com-

plex conversation ﬂows, allowing for a more natural

and intuitive interaction with users.

The Action Server is a component of Rasa that

allows for the integration of custom actions into the

chatbot. It will be responsible for executing speciﬁc

actions, such as performing complex operations that

cannot be directly handled by the chatbot. The Ac-

tion Server will be crucial for extending the chatbot’s

capabilities and providing more sophisticated and per-

sonalized responses to users.

These components, when combined using Docker

Compose, form a cohesive and scalable architecture

for the chatbot integrated with the ESVirtual applica-

tion. The use of Docker containers ensures that the

development environment is consistent and easily re-

producible across different machines, while the use

of Rasa and the Action Server provides the ﬂexibil-

ity and customization capabilities needed to meet the

speciﬁc demands of users. This modular approach

facilitates the update and maintenance of individual

components without affecting the entire system, pro-

moting greater operational efﬁciency and the ability

to adapt the chatbot to new needs and challenges.

Although there was a choice of not adding addi-

tional components for data tracking, as seen in (An-

drade et al., 2020a), the architecture is ﬂexible enough

to be incremented in the future. This choice was due

to privacy concerns of storing user data, which may

be sensitive since users are free to send any message.

Therefore, data is only in memory during the course

of conversation, and it is not persisted after the con-

versation is over.

4 RESULTS

This section presents the results of the chatbot devel-

opment, emphasizing both its content and interface. It

details how the chatbot’s responses have been reﬁned

for relevance and engagement, and how the interface

has been designed to ensure user-friendliness. Addi-

tionally, we will discuss the challenges encountered

in evolving the chatbot, including maintaining perfor-

mance while integrating new features and adapting to

content changes.

4.1 Knowledge Base

To ensure that the chatbot provides accurate and rel-

evant responses to users, it is essential to supply it

with appropriate content and train it to understand and

adequately respond to user queries. This is achieved

through the use of Rasa NLU (Natural Language Un-

derstanding), Stories, and Actions, which together

form the core intelligence of the chatbot. Below, we

detail each of these components and how they con-

tribute to the chatbot content feeding process.

Rasa NLU is responsible for understanding the in-

tents and entities in user messages. It utilizes ad-

vanced natural language processing (NLP) techniques

to extract relevant information from messages and

map them to the intents and entities deﬁned in the

chatbot model. To feed Rasa NLU with content, it is

necessary to provide examples of user messages along

with the associated intents and entities. These exam-

ples are used to train the Rasa NLU model, enabling

it to recognize similar patterns and contexts in future

messages.

Stories are representations of dialogues between

the user and the chatbot. They describe sequences of

interactions between the user and the chatbot, includ-

ing user messages, actions performed by the chatbot

in response to these messages, and the conditions for

transitioning between different dialogue states. Sto-

ries are used to train the chatbot’s conversation model,

allowing it to learn to follow speciﬁc conversation

ﬂows and make appropriate decisions based on the

context of the interaction. To feed the chatbot with

content, it is necessary to create and provide a variety

of stories that cover different conversation scenarios

and relevant use cases for users.

Actions are the tasks that the chatbot can perform

in response to user messages. These may include

Enhancing Post-Incarceration Support: A Custom Chatbot Solution for the Brazilian Prison System

427

sending a text message, asking a follow-up question,

calling an external API to retrieve additional infor-

mation, among others. Actions are deﬁned and im-

plemented by the developer and are triggered based

on the decisions made by the chatbot’s conversation

model during user interaction. To feed the chatbot

with content, it is necessary to deﬁne and implement

the actions required to respond to the different user

intents and achieve the chatbot’s objectives.

By feeding the chatbot content using Rasa NLU,

Stories, and Actions, we ensure that the chatbot can

understand and adequately respond to user queries,

providing a natural and effective conversational ex-

perience. The chatbot’s design ensures it covers a

comprehensive range of topics relevant to the needs

of former inmates, providing them with a reliable and

accessible source of information and support. Below,

we outline the main topics covered by the chatbot:

1. General Interactions

(a) Greeting and Farewell: The chatbot can ini-

tiate conversations, greet users, and bid them

farewell.

(b) Afﬁrmations and Negations: It handles user af-

ﬁrmations and negations, ensuring smooth in-

teraction ﬂow.

identity and capabilities.

2. Legal and Documentation Support

(a) Fine Penalties: Information on penalties and

ﬁnes.

(b) Document Regularization: Assistance with reg-

ularizing various personal documents, includ-

ing birth certiﬁcates, voter registration, CPF,

CTPS, and military service certiﬁcates.

support, including assistance from public de-

fenders and legal assistants.

3. Employment and Professional Development

(a) Job Opportunities: Information on ﬁnding job

opportunities.

(b) Professional Training: Details on obtaining

scholarships for courses, ﬁnding vocational

courses, and the requirements and documenta-

tion needed for these courses.

4. Housing and Shelter

(a) Temporary Housing: Information on temporary

shelters, including locations, availability, oper-

ating hours, and acceptance policies.

(b) Social Housing Programs: Assistance with so-

cial housing programs, rental subsidies, and re-

lated services.

5. Health and Welfare

(a) Public Health System: Access to information

about the public health system, including how

to obtain a SUS card, schedule appointments,

and ﬁnd medical establishments.

(b) Psychological Support: Access to psycholog-

ical support services for former inmates and

their families.

6. Food and Nutrition

(a) Food Assistance: Information on ﬁnding food,

community restaurants, and food cards.

(b) Basic Food Supplies: Guidance on obtaining

basic food supplies.

7. Community and Social Support

(a) Support Services: Information about CRAS,

CREAS, and Centro Pop services.

(b) Connecting with Other Former Inmates: Facili-

tating contact and support networks among for-

mer inmates.

8. Children and Family

(a) Childcare and School Enrollment: Assistance

with enrolling children in school and accessing

childcare services.

(b) Parental Support: Support for parenting and

custody issues.

9. Social Programs and Beneﬁts

(a) Social Beneﬁts: Information on various social

programs like BPC, PBF, social electricity tar-

iffs, and reclusion assistance.

(b) Eligibility and Registration: Guidance on eligi-

bility and registration for social programs and

subsidies.

10. Miscellaneous Support

(a) Transportation and Mobility: Information on

transportation options and mobility assistance.

(b) Leisure and Recreation: Details on accessing

free leisure and recreational activities.

cies, including drug addiction help and domes-

tic violence reporting.

A signiﬁcant number of intents needed to be inte-

grated into the chatbot to meet the requirements of so-

cial ofﬁces. We examined the progression of training

examples and their impact on intent prediction conﬁ-

dence across multiple releases of the model. Figure

3 illustrates the number of training examples used in

each release. This visual representation highlights the

increasing volume of data incorporated into the train-

ing process, driven by the high volume of intents re-

quired.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

428

Figure 3: Number of training examples per release.

To assess the model’s performance, we analyzed

the distribution of intent prediction conﬁdence at dif-

ferent stages of development. Figure 4 shows the con-

ﬁdence distribution for the ﬁrst release. The initial

distribution reveals insights into the model’s baseline

performance and the areas needing further reﬁnement.

As the model evolved, the intent prediction conﬁ-

dence distribution showed some stability, despite the

subtle increase in data volume depicted in Figure 3.

Maintaining this stability was a challenge. It required

constant updates to the pipeline to accommodate the

incoming new data. This stability was not coinciden-

tal but rather the result of continuous adjustments to

the pipeline, including modiﬁcations to the base train-

ing model and its hyperparameters.

Furthermore, with the large volume of intents, it

became necessary to progressively reduce the fall-

back threshold. This threshold determines when the

chatbot will return an “I didn’t understand” response.

Lowering it helped improve the chatbot’s ability to

handle the increasing number of intents accurately.

Figure 5 presents the conﬁdence distribution for the

last release, showcasing the improvements achieved

through these ongoing adjustments. Although the ad-

justments helped in the keeping the conﬁdence ratings

acceptable, the high volume of intents to predict still

resulted in some decrease in performance, which was

expected.

The evolution of the pipeline for the chatbot re-

ﬂects the growing sophistication in handling the nu-

ances of the Portuguese language and improving over-

all performance. Each iteration of the pipeline has

introduced advancements aimed at enhancing both

language understanding and user interaction quality.

Here’s a closer look at the progression and the rea-

sons behind these changes:

The initial pipeline was quite basic, relying heav-

ily on traditional NLP techniques, with Rasa’s de-

fault tokenizers, featurizers, classiﬁers and response

selector. The second pipeline made signiﬁcant strides

Figure 4: Intent prediction conﬁdence distribution for the

ﬁrst release.

Figure 5: Intent prediction conﬁdence distribution for the

last release.

in handling the Portuguese language more effectively

by integrating specialized tools and feature extraction

techniques using spaCy’s robust NLP tools, such as

SpacyNLP and SpacyTokenizer.

Introducing spaCy’s portuguese NLP model

(pt core news sm) and tokenizer enhanced the

pipeline’s ability to handle Portuguese text more

effectively, thanks to spaCy’s pretrained language

models and more sophisticated tokenization. Later,

the spaCy model was changed to the medium-sized

model (pt core news md) for better performance but

longer training duration. Additionaly, adding the

RegexEntityExtractor allowed for the extraction of

Enhancing Post-Incarceration Support: A Custom Chatbot Solution for the Brazilian Prison System

429

named entities based on regular expressions, improv-

ing the model’s ability to recognize speciﬁc terms

and entities in Portuguese through regex patterns and

the list of works provided in the custom lookup table.

The third pipeline represents a more advanced and

modern approach to building a Portuguese chatbot.

The RegexEntityExtractor was updated to be case-

insensitive, improving its ﬂexibility and accuracy in

identifying entities regardless of case variations. The

third pipeline’s use of advanced models like BERT

highlights a shift towards leveraging deep learning for

more nuanced language understanding and contextu-

ally appropriate responses. This version also reﬂects a

greater emphasis on improving entity recognition and

feature extraction capabilities, which are crucial for

handling the complexities of Portuguese.

4.2 Custom Actions

In the realm of conversational AI, custom actions

are pivotal in extending the capabilities of chatbots

beyond predeﬁned responses and static dialogues.

These actions enable chatbots to perform complex

tasks, interact with external APIs, and provide per-

sonalized user experiences. By leveraging custom ac-

tions, developers can tailor chatbot functionalities to

meet speciﬁc user needs and integrate seamlessly with

existing workﬂows and systems.

In the context of the ESVirtual application, cus-

tom actions play a crucial role in enhancing user inter-

action, ensuring that the chatbot can deliver accurate

information and handle nuanced queries effectively.

This section delves into the development of two such

custom actions, highlighting their design and imple-

mentation to bolster the ESVirtual chatbot’s perfor-

mance and user satisfaction.

The ActionPhoneNumber custom action was de-

veloped to enhance the ESVirtual chatbot’s ability to

provide users with relevant contact information for

various essential services. This action is triggered

when a user requests a phone number for a speciﬁc

entity, such as emergency services or public institu-

tions. The ActionPhoneNumber class, extending the

Action class from rasa sdk, is designed to retrieve and

deliver phone numbers based on user queries. The ac-

tion operates as follows:

1. Entity Extraction: It extracts the entity (e.g.,

”SAMU”, ”Pol

ıcia Militar”) from the user’s

message using the tracker.get slot( ”enti-

dade telefone”) method.

2. Phone Number Retrieval: It calls the

obter telefone method, which contains a prede-

ﬁned dictionary of service phone numbers, to

fetch the appropriate contact number for the given

entity.

3. Response Generation: If a valid phone number

is found, the action constructs a response mes-

sage with the phone number. For certain entities

like ”SAMU” and ”Disque Sa

ude,” the response

includes additional information about health ser-

vices. For ”Defensoria P

ublica,” the message pro-

vides further details about legal assistance ser-

vices. If the phone number is not found, the ac-

tion prompts the user to try again with a different

keyword.

4. Message Dispatch: The response is

then sent to the user using the dis-

patcher.utter message(text=response) method.

This design ensures that users receive precise and

helpful information promptly, improving the overall

utility and responsiveness of the chatbot.

The ActionFallbackWithSuggestions custom ac-

tion addresses scenarios where the chatbot fails to

understand the user’s input. By offering alternative

suggestions, this action aims to guide the user to-

wards more recognizable queries, enhancing user ex-

perience and interaction efﬁciency. The ActionFall-

backWithSuggestions class, also extending the Action

class from rasa sdk, operates as follows:

1. Intent Ranking Retrieval: It accesses the

ranking of the latest message’s intents using

tracker.latest message[’intent ranking’].

2. Top Intent Selection: It selects the top three in-

tents (excluding the fallback intent itself) for sug-

gesting to the user.

3. Intent Mapping: It loads a mapping of intent

names to their display names from a CSV ﬁle us-

ing the load intent mapping method. This ensures

that the suggestions are user-friendly and easily

understandable.

4. Button Generation: For each of the top intents, it

constructs a button with a title (display name) and

a payload (intent name).

5. Fallback Message Dispatch: The action

sends a fallback message to the user, in-

cluding the generated buttons, using the

dispatcher.utter message(text=message, but-

tons=buttons) method. This message encourages

the user to rephrase their query or select one of

the suggested options.

6. Utterance Reversion: It reverts the user’s utter-

ance to maintain the conversation ﬂow, ensuring

that the chatbot remains responsive and helpful.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

430

Figure 6: Initial presentation of the chatbot section in the

ESVirtual app.

By incorporating these mechanisms, the Action-

FallbackWithSuggestions custom action signiﬁcantly

enhances the chatbot’s ability to handle misunder-

standings and guide users effectively, thereby improv-

ing the overall conversational experience.

4.3 User Interface

As shown in Figure 2, the chatbot was developed as a

standalone service to facilitate easy maintenance and

evolution. It was made accessible to the ESVirtual

app via a REST API, ensuring that the chatbot in-

terface remained entirely separate from the question-

answer processing. Figures 6 and 7 illustrate the sec-

tions of the app where the chatbot is available.

Although the service and interface were modular-

ized, certain details needed to be synchronized for a

successful integration. For example, some keywords

in the chatbot responses were used to redirect users to

the corresponding help pages for additional informa-

tion. This was necessary because some topics were

Figure 7: Message exchange with the chatbot.

too extensive to be fully addressed in a chatbot mes-

sage.

Due to the limitations of mobile screen size, we

faced the challenge of balancing the amount of in-

formation in the chatbot responses to avoid excessive

scrolling, which could confuse users or cause them to

overlook previous interactions. This required careful

content curation and UX writing to ensure that infor-

mation was presented clearly and concisely.

5 CONCLUSION

The development and implementation of a chatbot

dedicated to assisting former inmates of the Brazilian

prison system, integrated with the ESVirtual applica-

tion, represents a crucial initiative in the search for in-

novative solutions to the challenges faced by this vul-

nerable group. Throughout this project, we explored a

comprehensive and sustainable approach to providing

support and resources to former inmates, considering

technical, functional, and ethical aspects.

Enhancing Post-Incarceration Support: A Custom Chatbot Solution for the Brazilian Prison System

431

The proposed architecture, based on the use of

Docker Compose, Rasa NLU, Stories, and Actions,

offers a solid and scalable framework for the chatbot,

ensuring its ability to understand natural language,

adapt to different interaction scenarios, and execute

customized actions. Integrating the chatbot with the

ESVirtual application further enhances its utility and

accessibility, providing former inmates with a reliable

and accessible channel to obtain information and sup-

port on their journey to social reintegration.

The decision not to use generative artiﬁcial intel-

ligence to generate responses, opting instead to man-

ually craft the chatbot’s content, demonstrates our

commitment to the accuracy, relevance, and reliability

of the information provided to users. This allows for

a quick response to changes in policies, legislation, or

services available to former inmates, ensuring that the

chatbot remains effective and useful over time.

On a social level, the implementation of this

project has the potential to signiﬁcantly contribute to

the reintegration of former inmates into society, re-

ducing recidivism rates and promoting the building of

a more just and inclusive society. By providing essen-

tial support and resources to former inmates, we em-

power these individuals to overcome the challenges

they face after leaving the prison system and to build

a digniﬁed and productive life.

In summary, this work represents an important

step towards a more humanized and effective ap-

proach to dealing with the reintegration of former in-

mates of the Brazilian prison system, demonstrating

the power of technology and innovation to promote

social well-being and justice.

ACKNOWLEDGMENTS

This study was ﬁnanced in part by the FUNDAC¸

DE APOIO

A PESQUISA DO DISTRITO FED-

ERAL (FAPDF), Brasil – Finance Code 01/2023.

REFERENCES

Al Husaeni, D. F., Haristiani, N., Wahyudin, W., and Rasim,

R. (2024). Chatbot Artiﬁcial Intelligence as Educa-

tional Tools in Science and Engineering Education: A

Literature Review and Bibliometric Mapping Analy-

sis with Its Advantages and Disadvantages. ASEAN

Journal of Science and Engineering, 4(1):93–118.

Andrade, G. G., Silva, G., J

unior, F. D., Santos, G.,

de Mendonc¸a, F. L., and J

unior, R. S. (2020a).

EvaTalk: A Chatbot System for the Brazilian Govern-

ment Virtual School, volume 2. SCITEPRESS, Portu-

gal.

Andrade, G. G., Silva, G., J

unior, F. D., Santos, G.,

de Mendonc¸a, F. L., and J

unior, R. S. (2020b).

Evatalk: A chatbot system for the brazilian govern-

ment virtual school. In Proceedings of the 22nd Inter-

national Conference on Enterprise Information Sys-

tems - Volume 1: ICEIS,, pages 556–562. INSTICC,

SciTePress.

Bachtiar, F. A., Dysham, A. A., Fauzulhaq, A. D., Az-

zam, J. S., and Aryadita, H. FLUENT: Factoid Re-

trieval Based Chatbot Using LSTM. In 2023 1st In-

ternational Conference on Advanced Engineering and

Technologies (ICONNIC), page 14. IEEE.

Benard, A. K., Msomba, G., and Pesha, J. (2023). Social

Challenges Facing Ex-Prisoners During their Integra-

tion into Societies in Iringa Municipality, Tanzania.

Asian Res. J. Arts Soc. Sci., 21(3):78–89.

Cacabro, M., Franco, W., Monteiro, J. M., and Machado,

J. (2023). Iara - an architectural model to assist the

development of advising bots for misinformation de-

tection. In Proceedings of the 29th Brazilian Sym-

posium on Multimedia and the Web, WebMedia ’23,

page 168–176, New York, NY, USA. Association for

Computing Machinery.

Canc¸ado, E. C. R., Bandeira, I. N., Costa, P. H. T., Fortes,

J., Moreira, D. C., Amaral, L. H. V., and Canedo,

E. D. (2022). Exploring user-centered requirements

validation and veriﬁcation techniques in a social inclu-

sion context. In Filipe, J., Smialek, M., Brodsky, A.,

and Hammoudi, S., editors, Proceedings of the 24th

International Conference on Enterprise Information

Systems, ICEIS 2022, Online Streaming, April 25-27,

2022, Volume 1, pages 85–92. SCITEPRESS.

Canedo, E. D., Canc¸ado, E. C. R., Mota, A. P. B., Ban-

deira, I. N., Costa, P. H. T., Lima, F., Amaral, L., and

Bonif

acio, R. (2024). Using Design Thinking to break

social barriers: An experience report with former in-

mates. J. Software: Evol. Process, 36(7):e2648.

Chagas, B. A., Pagano, A. S., Prates, R. O., Praes, E. C.,

Ferreguetti, K., Vaz, H., Reis, Z. S. N., Ribeiro,

L. B., Ribeiro, A. L. P., Pedroso, T. M., Beleigoli, A.,

Oliveira, C. R. A., and Marcolino, M. S. (2023). Eval-

uating User Experience With a Chatbot Designed as a

Public Health Response to the COVID-19 Pandemic

in Brazil: Mixed Methods Study. JMIR Human Fac-

tors, 10.

D’

Avila, T. C. (2018). KINO: an approach for rule-based

chatbot development, monitoring and evaluation. Uni-

versidade Federal de Minas Gerais.

Dagkoulis, I. and Moussiades, L. (2022). A Comparative

Evaluation of Chatbot Development Platforms. In

ACM Other conferences, pages 322–328. Association

for Computing Machinery, New York, NY, USA.

Dores, A. P. and Dores, A. P. (2019). The Role of Infor-

mation and Communications Technology in the Social

Reintegration of Ex-Prisoners. IGI Global.

Gonc¸alves, L. P., Silva, G. R. S., and Canedo, E. D. (2024).

Documentation artifacts for conversation-related re-

quirements speciﬁcation in chatbots. In Lucena, M.,

Lencastre, M., and Ballejos, L. C., editors, Anais do

WER24 - Workshop em Engenharia de Requisitos,

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

432

Buenos Aires, Argentina, August 7-9, 2024. Even3,

Brasil.

Joshi, M. and Sharma, R. K. (2020). An Analytical

Study and Review of open Source Chatbot frame-

work, RASA. International Journal of Engineering

Research & Technology, 9(6).

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., and Zou, J.

(2023). GPT detectors are biased against non-native

English writers. PATTER, 4(7).

Massa, P., de Souza Ferraz, D. A., Magno, L., Silva, A. P.,

Greco, M., Dourado, I., and Grangeiro, A. (2023). A

Transgender Chatbot (Amanda Selﬁe) to Create Pre-

exposure Prophylaxis Demand Among Adolescents

in Brazil: Assessment of Acceptability, Functionality,

Usability, and Results. J. Med. Internet Res., 25.

Misischia, C. V., Poecze, F., and Strauss, C. (2022). Chat-

bots in customer service: Their relevance and impact

on service quality. Procedia Comput. Sci., 201:421–

428.

Pandey, S. and Sharma, S. (2023). A comparative study

of retrieval-based and generative-based chatbots us-

ing Deep Learning and Machine Learning. Healthcare

Analytics, 3:100198.

PlayStore (2024). Escrit

orio Social Virtual ESV – Apps no

Google Play. [Online; accessed 23. Jul. 2024].

Portal CNJ (2022). Escrit

orio Social Virtual - Portal CNJ.

[Online; accessed 23. Jul. 2024].

Shaitarova, A., Zaghir, J., Lavelli, A., Krauthammer, M.,

and Rinaldi, F. (2023). Exploring the Latest High-

lights in Medical Natural Language Processing across

Multiple Languages: A Survey. Yearbook of Medical

Informatics, 32(1):230.

Silva, G. R. S. and Canedo, E. D. (2024). Towards user-

centric guidelines for chatbot conversational design.

Int. J. Hum. Comput. Interact., 40(2):98–120.

Storck, W. (2023). More Education of Juvenile Offenders

In Sentences Of Imprisonment: A Reform And Justi-

ﬁcation Approach As A Consequence Of Niklas Luh-

mann’s Systems Theory. Issue 3/2023, 8(0).

Tharshini, N. K., Wong, S. K., Mas’ud, F. H., Ali, K., and

Ahmad, N. A. (2024). Readiness for re-entry among

pre-release prisoners in Sarawak, Malaysia. Journal

of Forensic Practice, ahead-of-print(ahead-of-print).

Thom

e-Williams, A. C. (2004). Sociolinguistic aspects of

forms of address in portugal and brazil: Tu or voce.

Intercultural Communication Studies, 13(3):85–99.

Trajano, D., Bordini, R. H., and Vieira, R. (2023).

OLID-BR: offensive language identiﬁcation dataset

for Brazilian Portuguese. Lang. Resources &. Eval-

uation, pages 1–27.

Tueiv, M. and Schmitz, E. (2023). Maximizing the value

delivered of chatbots in e-gov using the incremental

funding method. In Proceedings of the 16th Inter-

national Conference on Theory and Practice of Elec-

tronic Governance, ICEGOV ’23, page 242–246, New

York, NY, USA. Association for Computing Machin-

ery.

UNDP (2024). CNJ e parceiros lanc¸am aplicativo para

apoio a pessoas egressas do sistema prisional. [On-

line; accessed 23. Jul. 2024].

Wilkens, R., Zilio, L., and Villavicencio, A. (2024). As-

sessing linguistic generalisation in language models:

a dataset for Brazilian Portuguese. Lang. Resources

&. Evaluation, 58(1):175–201.

Enhancing Post-Incarceration Support: A Custom Chatbot Solution for the Brazilian Prison System

433