AGOSTO, 2026 (13-35)Número 29
DISCOURSE ANALYSIS FOR THE
DEVELOPMENT OF A CYBERGROOMING
DETECTION MODEL ON ROBLOX
ANÁLISIS DEL DISCURSO PARA EL
DESARROLLO DE UN MODELO DE
DETECCIÓN DE CIBERGROOMING EN
ROBLOX
DOI: https://doi.org/10.37135/chk.002.29.01
Research Article
Recibido: (07/03/2026)
Aceptado: (27/05/2026)
1Benemérita Universidad Autónoma de Puebla, Facultad de Filosofía y Letras,
Licenciatura en Lingüística y Literatura Hispánica, Puebla, México, email: ana.
castanonm@alumno.buap.mx
2Benemérita Universidad Autónoma de Puebla, Facultad de Ciencias de la
Computación, Ingeniería en Ciencias de la Computación, Puebla, México, email:
brenda.rodriguezco@alumno.buap.mx
3Benemérita Universidad Autónoma de Puebla, Facultad de Ciencias de la
Computación, Ingeniería en Ciencias de la Computación, Puebla, México, email:
andrea.bazand@alumno.buap.mx
4Benemérita Universidad Autónoma de Puebla, Facultad de Ciencias de la
Computación, Coordinador del Laboratorio de Análisis Forense Digital, Puebla,
México, email: enrique.colmenares@correo.buap.mx
Ana Paola Castañón Marroquín1,
Brenda Ailed Rodríguez Colis2,
Andrea Bazán Durán3,
Luis Enrique Colmenares-Guillen4
DISCOURSE ANALYSIS FOR THE DEVELOPMENT OF A
CYBERGROOMING DETECTION MODEL ON ROBLOX
Número 29 / AGOSTO, 2026 (13-35)
14
Cybergrooming represents a growing threat on online gaming platforms such as Roblox,
where anonymity and frequent interaction among child users create conditions conducive
to child abuse and sexual harassment. The objective of the research that led to this article
was to identify linguistic patterns in the discourse of groomers in Spanish-speaking
Roblox communities and incorporate them into a computational model for the automatic
detection of this cybercrime through text. To this end, a mixed-methods approach was
developed, integrating Corpus-Assisted Discourse Studies with the CRISP-DM data mining
methodology. A specialized corpus of 25 conversations was compiled and processed,
then subjected to detailed analysis. As a main result, a pattern of discursive organization
consisting of a sequence of seven conversational modules with specic predictive value
and a set of 21 functional lexicogrammatical patterns with 224 associated collocations were
identied, described, and subsequently incorporated into a text classication model capable
of distinguishing grooming conversations with 93.33% accuracy. In this way, the study
demonstrated the ecacy of discourse analysis as a basis for the development of systems for
the automatic detection of cybercrimes against minors.
KEYWORDS: Discourse analysis, computer crime, child abuse, videogame, computational
linguistics.
El cibergrooming representa una amenaza creciente en plataformas de juego en línea
como Roblox, donde el anonimato y la frecuente interacción de usuarios infantiles generan
condiciones propicias para el acoso sexual y abuso de menores. El objetivo de la investigación
que originó este artículo fue identicar patrones lingüísticos en el discurso de groomers de
comunidades hispanohablantes de Roblox e incorporarlos a un modelo computacional para
la detección automática de este delito informático a través del texto. Para ello se construyó
una metodología de enfoque mixto que integró los Estudios del Discurso Asistidos por
Corpus y la metodología de minería de datos CRISP-DM. Se compiló y procesó un corpus
especializado de 25 conversaciones que fue sometido a un análisis pormenorizado. Como
resultado principal, se delimitó y describió un patrón de organización discursiva constituido
por una secuencia de siete módulos conversacionales con un valor predictivo determinado
y 21 patrones léxico-gramaticales funcionales con 224 colocaciones asociadas, elementos
integrados a un modelo de clasicación textual capaz de distinguir conversaciones de
grooming con un 93.33% de exactitud. De esta forma, el estudio demostró la efectividad
del análisis discursivo como base para el desarrollo de sistemas de detección automática de
delitos informáticos contra menores.
PALABRAS CLAVE: Análisis del discurso, delito informático, abuso de menores,
videojuego, lingüística computacional.
ABSTRACT
RESUMEN
DISCOURSE ANALYSIS FOR THE
DEVELOPMENT OF A CYBERGROOMING
DETECTION MODEL ON ROBLOX
ANÁLISIS DEL DISCURSO PARA EL
DESARROLLO DE UN MODELO DE
DETECCIÓN DE CIBERGROOMING EN
ROBLOX
Ana Paola Castañón Marroquín, Brenda Ailed Rodríguez Colis, Andrea Bazán Durán, Luis Enrique Colmenares-Guillen
CHAKIÑAN. Revista de Ciencias Sociales y Humanidades / ISSN 2550 - 6722
15
INTRODUCTION
The prolonged use of internet-enabled electronic devices among
children and adolescents is a growing phenomenon that has transformed
the social dynamics and access to information. In this regard, Red
Grooming Latam (2024) reported that Latin American minors spend
an average of four hours a day online, split between forums, messaging
applications, social media, and videogames, which are of particular
interest to this study.
Online gaming platforms have established a strong presence as
signicant spaces for social interaction within youth culture (Morreale
& Rosa, 2024), a phenomenon related both to the integration of chat
and voice channels into their interfaces (Excelin et al., 2024), and to the
formation of user communities on external applications (Pyslar, 2025).
A paradigmatic example is Roblox, a platform that enables the creation
of gaming experiences shared by millions of users (Kou et al., 2025;
Red Grooming Latam, 2024).
However, interacting with strangers in virtual environments has led to
the emergence of harmful behaviors and cybercrimes (Rozgonjuk et
al., 2023), such as cybergrooming, a practice in which an adult contacts
minors through digital platforms with the aim of involving them in
sexual activity or obtaining sexual content (Lorenzo-Dus, 2023; Ortiz,
2024). In particular, cases of grooming associated with the use of
Roblox have been reported and have led to legal complaints (Carville
& D’Anastasio, 2024; Dorsey, 2024).
As a form of child abuse, online grooming has been extensively studied
from a behavioral science perspective, although research adopting a
discursive approach is scarce (Lorenzo-Dus, 2023). Nevertheless, the
practice of grooming materializes by means of a series of communicative
exchanges between adults and minors.
This study performed the rst analysis of the conversational discourse
of groomers in Spanish-speaking Roblox virtual communities. To
this end, in line with previous research (Broome et al., 2025; Joleby
et al., 2021; Lorenzo-Dus et al., 2020; Lorenzo-Dus & Izura, 2017;
Pienczykowski & Madella, 2026), the procedures of Corpus-Assisted
Discourse Studies, as dened in the work of Gillings et al. (2024),
were adopted. The aim was to identify linguistic patterns at the level
of conversational organization and lexico-grammatical resources with
functional value (Halliday & Matthiessen, 2004; Vaamode & González,
2008) associated with pragmatic-discursive strategies.
DISCOURSE ANALYSIS FOR THE DEVELOPMENT OF A
CYBERGROOMING DETECTION MODEL ON ROBLOX
Número 29 / AGOSTO, 2026 (13-35)
16
Finally, a computational model was developed using the CRISP-
DM methodology, in which the detected patterns were organized
into structures suitable for use in automated text analysis (Manning
et al., 2009): a sequence of conversational modules with a specic
predictability percentage and collocations or collocation pairs (Daohuan
& Xuri, 2023; Schweinberger, 2024) associated with the patterns of
each module.
METHODOLOGY
This study adopted a mixed-methods approach (Molina et al., 2024)
that combined two distinct methodologies: Corpus-Assisted Discourse
Studies or CADS for discursive analysis, and the CRISP-DM data
analysis methodology for the development of the computational model.
The result of this integration is presented in Figure 1.
Figure 1: Diagram of the integrated CADS ad CRISP-DM
methodology and its implementation in the study
Ana Paola Castañón Marroquín, Brenda Ailed Rodríguez Colis, Andrea Bazán Durán, Luis Enrique Colmenares-Guillen
CHAKIÑAN. Revista de Ciencias Sociales y Humanidades / ISSN 2550 - 6722
17
CADS is a mixed-methods paradigm that employs corpus linguistics
tools in discourse analysis, namely the study of language in use
(Gillings et al., 2024; Partington et al., 2013). In the rst phase, this
methodology requires the identication of high-frequency linguistic
patterns, and in the second phase, the qualitative interpretation of
discourse characteristics (Gillings et al., 2024).
In conceptualizing cybergrooming as a distinct discursive practice, two
theoretical postulates were adopted. First, the denition proposed by
Van Dijk (1997) of discourse as a complex and bounded set of linguistic
constructions with eects on context, conceived as “the structure of all
properties of the social situation that are relevant for the production or
the reception of discourse” (p. 19). Second, the characterization oered
by Lorenzo-Dus (2023) of digital grooming as a form of manipulative
discourse with specic resources and strategies.
For pattern identication, Sketch Engine was employed, a software
program specialized in the management and analysis of corpus or large
collections of natural language data (Lutzuky & Kehoe, 2022), which
enabled the generation of statistically representative results (Isti’anah et
al., 2023) and the determination of their relative frequency.
In this study, a linguistic pattern as a recurring choice of form and
meaning (Hunston, 2010), was examined at two levels: as a unit that
reects a regular discursive organization, and as a set of valency
structures and preferred lexical collocations associated with a particular
meaning or implicature (Hunston, 2025).
Regarding the rst level, drawing on the concept of sequential
organization proposed by Scheglo (2007) and considering that
discourse may be analyzed “in terms of a number of typical formal
categories and their specic order and function” (Van Dijk, 1997, p.
37), it was posited that a series of communicative acts or interventions
—referred to here as conversational modules— bearing a specic
discursive or thematic function, constitute a pattern.
With respect to the second level, categories from Systemic Functional
Linguistics were employed to detail the valency structure that
congures the pattern, following the semantic role annotation methods
for patterns proposed by Hunston (2025). This theoretical framework is
conceived as a tool for the socio-semiotic study of language (Halliday
& Matthiessen, 2004), thereby rendering it compatible with concepts
from discourse analysis and pragmatics drawn upon in the second phase
of CADS.
Within this theory, the unit of analysis is the clause, a construction with a
conjugated verb in which three metafunctions are realized at the lexico-
grammatical level (Halliday & Matthiessen, 2004). In particular, the
experiential component of the ideational metafunction was employed
DISCOURSE ANALYSIS FOR THE DEVELOPMENT OF A
CYBERGROOMING DETECTION MODEL ON ROBLOX
Número 29 / AGOSTO, 2026 (13-35)
18
within the transitivity subsystem, where six types of verbal processes
and their participants are identied, as dened in Table 1 based on
Halliday and Matthiessen (2004) and the ADESSE database project
reviewed by Vaamode and González (2008).
Table 1: Types of verbal processes and their participants in the
systemic-functional approach
Furthermore, the interpersonal metafunction was addressed through
the mode subsystem, which distinguishes three discursive realizations:
declarative, imperative, and interrogative (Halliday & Matthiessen,
2004). In addition to collocations, high-frequency lexical pairs that occur
within the same construction (Daohuan & Xuri, 2023; Schweinberger,
2024), these elements completed the structure of the patterns.
The corpus analyzed was composed of 25 conversations collected by
means of a convenience sampling technique (Molina et al., 2024),
conrming that the conversations complied the following selection
criteria: relation to Roblox, explicit request for sexual content, and
participation of a ctional or real minor.
Corpus compilation involved the systematic search for public reports
on social media documenting, by means of conversations, cases of child
abuse in Roblox communities, as well as the creation of ctional minor
proles to interact with potential groomers in public and private social
media groups related to the videogame, registering the conversations
that met the established criteria.
Ana Paola Castañón Marroquín, Brenda Ailed Rodríguez Colis, Andrea Bazán Durán, Luis Enrique Colmenares-Guillen
CHAKIÑAN. Revista de Ciencias Sociales y Humanidades / ISSN 2550 - 6722
19
The conversations were transcribed, anonymized by removing direct
identiers of the interlocutors, and normalized through orthographic
correction, translation of anglicisms, and completion of contextually
relevant apocopated forms. Subsequently, the complete corpus was
compiled in Sketch Engine to facilitate both the qualitative analysis
aimed at delimiting the conversational modules and the identication
of linguistic patterns by means of the specic tools of the software in
the next stage.
Conversational modules, clauses, the predictability percentage —a
value established for the automatic detection of each module, where a
zero percentage conrms that constitutes a neutral scheme not directly
associated with cybergrooming —, relevant conversational modules —
those exceeding 1% predictability—, and the interpretation of discourse
through the identication of pragmatic-discursive strategies (Calsamiglia
& Tuson, 2012), were delimited and developed qualitatively. In contrast,
the calculation of the relative frequencies of these elements, as well as
the keyword list automatically generated by means of Sketch Engine’s
Wordlist, Word Sketch, and Concordances tools —consulted for the
delimitation of modules, patterns, and collocations — correspond to
the quantitative dimension of the approach.
For the development of the computational model, the CRISP-DM data
mining methodology was adopted. According to Mariscal et al. (2010),
this methodology consists of six interrelated phases applied in the
research as detailed in Figure 1.
In the data preparation phase of CRISP-DM, a dataset was constructed.
This is a collection of data with observations in a structured format
that facilitates storage, retrieval, and automatic analysis (Badman &
Kosinski, 2024). This collection was created by integrating a sample of
50 observations from Spanish-Tweets (Pérez et al., 2022), a large, open-
source, public dataset hosted on Hugging Face by the Pysentimiento
Project (McDonough, 2023) which compiles Spanish-language tweets
characterized by their neutrality and diversity of regional and thematic
variants. Finally, the dataset was supplemented with the conversational
corpus.
Two elds were also incorporated for each unit of analysis in grooming
conversations: an identier and the associated text. This structure,
visualized in Table 2, retained the original text sequence, and facilitated
the subsequent stages of data cleaning, computational analysis, and
modeling.
DISCOURSE ANALYSIS FOR THE DEVELOPMENT OF A
CYBERGROOMING DETECTION MODEL ON ROBLOX
Número 29 / AGOSTO, 2026 (13-35)
20
Table 2: Structure of the conversational units processed for the
cybergrooming dataset
Regarding the dataset cleaning process, this included converting text
to lowercase, removing punctuation marks and grammatical categories
without lexical value, as well as lemmatization, a widely used text
preprocessing technique in natural language processing —an area of
computational linguistics— that performs automatic morphological
analysis to convert a word to the base form or lemma, reducing textual
complexity (Manning et al., 2009). Lemmatization was performed using
the udpipe library (Straka & Straková, 2020), which removes prexes
and suxes from words.
The following techniques and metrics were applied during the model
evaluation and implementation phases: confusion matrix, to examine
correct and incorrect classications by class; accuracy, as an overall
measure of performance; recall and specicity, to evaluate performance
dierences between positive and negative classes (IBM, 2025; Steward,
2023; Ting, 2010); as well as the visualization of a classication test.
Combining the conversational corpus with the sample obtained from
social media creates a dataset containing dierent types of language and
contexts, which could introduce biases in the evaluation of the model.
Therefore, this validation strategy was established to measure the
performance of the classier and the ability to adapt to the dierences
between the two data sources within the test set.
RESULTS AND DISCUSSION
A sequence of seven conversational modules was identied in the
conversations of the corpus, described and exemplied in Table
3, which also species their relative frequency and predictability
percentage. This sequence is presented as a pattern of discursive
organization characteristic of digital groomers in Spanish-speaking
Roblox communities. Moreover, three to ve patterns were obtained
for the relevant modules, contained in Table 4, which also presents a
Ana Paola Castañón Marroquín, Brenda Ailed Rodríguez Colis, Andrea Bazán Durán, Luis Enrique Colmenares-Guillen
CHAKIÑAN. Revista de Ciencias Sociales y Humanidades / ISSN 2550 - 6722
21
sample of 10 of the 224 total collocations associated with the modules.
Table 3: Conversational modules of cybergrooming on Roblox
The rst module corresponds to the salutation, presented in 76% of
conversations. This speech act is also referred to as an opening frame
sequence, which includes interrogative formulas or exclamatory
expressions that serve a basic interpersonal function: opening and
acknowledging the existing relationship between the two dialogue
participants (González-Sanz, 2024). Within the corpus, this module
evidenced the denition of the interactional relationship through the
calibration of parameters such as social position and degree of intimacy
(Calsamiglia & Tuson, 2012; González-Sanz, 2024), as corroborated
by the use of vocatives or adjectivized appellatives as rapport-building
strategies.
Additionally, in some cases the salutation was accompanied by personal
introductions aimed at projecting an apparently legitimate identity of
DISCOURSE ANALYSIS FOR THE DEVELOPMENT OF A
CYBERGROOMING DETECTION MODEL ON ROBLOX
Número 29 / AGOSTO, 2026 (13-35)
22
the perpetrator, appealing to the strategic conguration of a face or
image in the terms of Goman (Walsh, 2022; Calsamiglia & Tuson,
2012). Together, these resources foster the development of a trust bond
with the minor and initiate manipulation (Sorlin, 2017) in the access
phase of the model of Lorenzo-Dus (2023).
Table 4: Exemplication of linguistic patterns and collocations from
the cybergrooming corpus on Roblox
The second conversational module presented a frequency of 76% and
was designated as the personal information request. In line with the two
subsequent modules, it was named in terms of the illocutionary force of
Ana Paola Castañón Marroquín, Brenda Ailed Rodríguez Colis, Andrea Bazán Durán, Luis Enrique Colmenares-Guillen
CHAKIÑAN. Revista de Ciencias Sociales y Humanidades / ISSN 2550 - 6722
23
a speech act with a greater or lesser degree of indirectness that threatens
the groomer’s face (Calsamiglia & Tuson, 2012; Walsh, 2022); in this
case, a request whose primary function is the solicitation of personal
data from the victim. Five patterns were identied:
− Pattern A, with a relative frequency of 3%, in which the groomer
employs a verbal process that positions the minor as the sayer of
personal data or information, with the groomer as the receiver.
− Pattern B, with a frequency of 68%, congures the minors as
entities that assign themselves attributes or personal traits, both
psychological and physical, by means of a relational process.
− Pattern C, which inquires into the habits and family context of the
minor by means of a behavioral process, related to risk assessment
for isolation (Lorenzo-Dus et al., 2020), and with a frequency of
3%.
− Pattern D, with a frequency of 12%, explores the preferences of the
minor as a stimulus regularly experienced through a mental process
of sensation or perception.
− Pattern E, identied in constructions with a greater degree of
indirectness, whereby the groomer discloses potentially false
personal information (Pienczykowski & Madella, 2026) by means
of a relational process, which presented a frequency of 6%.
The most frequent mode of discursive realization was the interrogative,
with 82%, as visualized in Figure 2 alongside the relative frequency
data of the three modes analyzed in relevant modules.
Figure 2: Graph of the frequency of the modes of discursive
realization in relevant conversational modules of cybergrooming on
Roblox
DISCOURSE ANALYSIS FOR THE DEVELOPMENT OF A
CYBERGROOMING DETECTION MODEL ON ROBLOX
Número 29 / AGOSTO, 2026 (13-35)
24
The third module, Robux —the virtual currency of the videogame— or
rewards oer, presented a frequency of 96%. This constitutes the most
relevant contribution of this study and particularizes the cybergrooming
discourse on Roblox. Contrary to the implicit oer of aective goods
—such as validation and recognition— in this context, the incentives
are material and related to the dynamics of the videogame. The oer
of these rewards is “conditional and thereby put pressure on the child”
(Joleby et al., 2021, p.10). Four patterns were identied:
- Pattern A, presented in 34% of the conversations, entails a
mental process of sensation that congures the minor as the
experiences of diverse, though generally positive, internal states
in response to a possible reward.
- Pattern B, with a frequency of 11%, presents an attributive
relational process that posits a positive state related to the
possession of Robux or rewards.
- Pattern C, with a frequency of 17%, involves a behavioral
process performed by the groomer and employed for indirect
speech acts that trigger conversational implicatures or situated
inferential processes that speakers activate to understand
utterances (Calsamiglia & Tuson, 2012); in this case, to interpret
the clause as an oer.
- Pattern D, presented in 29% of the conversations, consists of a
relational process of transference with the groomer as the entity
or possessor and the reward as the attribute.
The interrogative mode presented a relative frequency of 57% and the
declarative mode of 43%, the latter corresponding to indirect oers.
The fourth conversational module, sexual content request, was present
in 100% of the corpus and is the dening scheme of the discursive
practice of cybergrooming.
Five patterns were identied in this module:
- Pattern A, identied in 40% of the conversations, consists of a
material process of creation that positions the minor as the actor
that produces sexual content with the groomer as the beneciary.
- Pattern B, with a relative frequency of 21%, in which a
relational process of transference or possession congures the
minor as the possessor or entity capable of providing sexual
content.
- Pattern C, with a frequency of 19%, also entails a relational
process of transference or possession, but one in which the
groomer is presented as the entity related to positive states
derived from the possession of sexual content.
- Pattern D, with a frequency of 7%, constitutes a mental process
of sensation in which the groomer is presented as the experiences
of positive states derived from the possession of sexual content.
Ana Paola Castañón Marroquín, Brenda Ailed Rodríguez Colis, Andrea Bazán Durán, Luis Enrique Colmenares-Guillen
CHAKIÑAN. Revista de Ciencias Sociales y Humanidades / ISSN 2550 - 6722
25
- Pattern E, with a frequency of 7%, comprises a relational
process in which a challenge created by the groomer is dened
as the action of sending sexual content in the form of an indirect
request.
In this module of request, the imperative and declarative modes were
employed with greater frequency. Furthermore, this was characterized
by the inclusion of both explicit and vague sexual language: explicit
terms coincide with the onset of the sexual gratication process (Joleby
et al., 2021), while vague terms serve the function of “minimizing the
oence (...) and desensitizing the child to sexual communications”
(Pienczykowski & Madella, 2026, p. 11), thereby preserving the face
of the groomer in face-threatening acts (Calsamiglia & Tusón, 2012).
The penultimate module, with a relative frequency of 92%, is that of
resistance management. This module exhibited greater variability in
clause types and constructions, as well as dierent speech acts and an
overlap of cyber-grooming phases (Joleby et al., 2021; Lorenzo-Dus,
2023).
This module employs both overt persuasion and coercion strategies
(Chiang & Grant, 2019), as well as politeness strategies to improve the
image of the victims and attribute to them a sense of responsibility for
their actions comparable to that of the aggressor, or what Pienczykowski
& Madella (2026) refer to as “role and responsibility reversal” (p. 10).
Moreover, this was presented in the form of hypothetical statements,
softened or approximate rephrasings, conditional statements, requests
for specic information, or new oers, with the aim of managing or
defusing the conict that arises during the interaction in response to
the resistance of the victim (Evans & Lorenzo-Dus, 2025). The four
patterns identied were:
- Pattern A with a mental process of sensation or cognition that
motivates the child and provides them with a false sense of
security by appealing to positive internal states, with a frequency
of 22%.
- Pattern B, with a 4% frequency, comprises a verbal process
that inquires information regarding the sexual content and the
refusal to comply with the request of this content from the minor.
- Pattern C, with a frequency of 22%, entails a relational process
that assigns a positive evaluation with an attenuating eect to
the action of sending content.
- Pattern D, with a frequency of 19%, presents a relational
process of transference in which the groomer oers additional
rewards.
The declarative mode was the most common in the resistance
management module, accounting for 59% of the clauses in the corpus.
DISCOURSE ANALYSIS FOR THE DEVELOPMENT OF A
CYBERGROOMING DETECTION MODEL ON ROBLOX
Número 29 / AGOSTO, 2026 (13-35)
26
The nal relevant conversational module identied, with a relative
frequency of 40%, is that of compliments. The function of this
module is to build trust by explicitly praising the appearance, abilities,
possessions, or personality traits of the victim (Lorenzo-Dus & Izura,
2017).
The degree of directness of this module changed depending on the prole
of the aggressor, whether these were hypersexual, with more direct
compliments, or intimacy-focused and adaptive, with more indirect
compliments (Broome et al., 2025). Three patterns were identied:
- Pattern A, with a frequency of 56%, constitutes a relational
process in which positive attributes are assigned to the minor or
to some of their physical or psychological traits in the form of
a direct compliment which emphasizes maturity or docility to
equalize the power relationship.
- Pattern B, with a frequency of 13%, involves a relational
process that positively denes a personal trait of the minor in
the form of an indirect compliment.
- Pattern C presents a mental process of sensation in which the
groomer openly declares their reaction to the victim or to the
content as a stimulus.
The most common discursive realization mode in the compliment
module was the declarative, which appeared in 94% of the clauses.
Finally, the dened sequence concludes with a closing module that
appears in 32% of the conversations. In this section, the groomer engages
in actions such as expressing gratitude, making threats, and saying
goodbye; or, once the conversation has ended —whether voluntarily
or involuntarily— the groomer ceases to interact, thereby violating the
cooperative principle proposed by Grice (Calsamiglia & Tuson, 2012)
On the other hand, a computational model was developed capable of
classifying conversational text as either cybergrooming or neutral, as
well as identifying the relevant thematic module in each instance of
grooming text by detecting collocations of the predened linguistic
patterns.
For the performance of the model, cross-validation was employed
(Allgaier & Pryss, 2024), which consisted of partitioning the data into k
distinct groups where k − 1 constituted the training sets. The remaining
group contained the data used to evaluate the accuracy; this process was
repeated k times until each group had been used as a validation set.
The automatic analysis performed by the developed model was based
on the detection of collocations associated with each conversational
module. Each time a collocation was identied, the corresponding count
was incremented; subsequently, the intensity was calculated by dividing
the total number of collocations found by the absolute frequency of the
Ana Paola Castañón Marroquín, Brenda Ailed Rodríguez Colis, Andrea Bazán Durán, Luis Enrique Colmenares-Guillen
CHAKIÑAN. Revista de Ciencias Sociales y Humanidades / ISSN 2550 - 6722
27
collocation pairs within the module.
Based on these intensities, a prediction function was proposed, illustrated
in Figure 3. This function generated a risk score by combining the
intensities through a weighted scoring scheme, following the principle
described by Manning et al. (2009) for classication models based on
linguistic features.
With the score calculated for each conversation, an optimal threshold
was established (Lyu & Ishwaran, 2023), the point at which specicity
and sensitivity are optimized for detection and classication in each
case.
Additionally, based on the patterns identied by means of the linguistic
analysis, a conditional decision was implemented that classied a
conversation as positive when it included the sexual content request
module or at least three relevant conversational modules. These criteria
were determined empirically during the model evaluation process.
Overall, a conversation was classied as grooming if the score was equal
to or higher than the optimal threshold, or if the conditional described
above was met.
Figure 3: Prediction function for the cybergrooming detection model
on Roblox
Regarding the rst method for evaluating the performance of the
model, the confusion matrix (Ting, 2010) determines the total number
of correct and incorrect predictions. Figure 4 presents the confusion
matrix corresponding to the test set and indicates zero false positives,
which means that no neutral conversations were classied as grooming.
Figure 4: Confusion matrix of the test set in the cybergrooming
DISCOURSE ANALYSIS FOR THE DEVELOPMENT OF A
CYBERGROOMING DETECTION MODEL ON ROBLOX
Número 29 / AGOSTO, 2026 (13-35)
28
detection model on Roblox
Subsequently, accuracy was calculated, which indicates how frequently
the model makes a correct prediction (IBM, 2025). The model achieved
an accuracy of 0.9333, meaning that the system correctly classied
93.33% of all evaluated conversations. Similarly, as shown in Figure
5, other evaluation metrics were calculated: sensitivity or recall, which
measures the correct detection of positive cases, and specicity, which
measures the detection of negative cases (Steward, 2023).
Figure 5: Results of the evaluation metrics of the cybergrooming
detection model on Roblox
A sensitivity of 0.8333 indicates that the model correctly identied
83.33% of grooming cases, although it implies that the remaining
16.67% went undetected. A specicity of 1.0 indicates that the
model correctly classied all cases that were not grooming, without
generating any false positives. Finally, this system was tested with
new conversations. The results were organized in an output table that
includes the estimated probability interpreted as the risk level, the
number of modules identied, and the nal binary classication with
grooming=yes and neutral=no.
Ana Paola Castañón Marroquín, Brenda Ailed Rodríguez Colis, Andrea Bazán Durán, Luis Enrique Colmenares-Guillen
CHAKIÑAN. Revista de Ciencias Sociales y Humanidades / ISSN 2550 - 6722
29
Table 5: Visualization of a prediction test of the cybergrooming
detection model on Roblox
Ultimately, the results in Table 5 reveal a directly proportional
relationship in which the estimated probability of grooming increases
as the density of identied modules augments, suggesting that the
accumulation and aggregation of linguistic indicators positively
inuence the statistical reliability of the classication. This trend is
evident in the detected modules column of Table 5, where the absence
of these modules results in a negative prediction; however, this behavior
is conditioned by the established decision threshold. This threshold
regulates the sensitivity to ambiguous cases, acting as a methodological
lter designed to reduce the false positive rate. Accordingly, instances
that record high probabilities but are ultimately categorized as negative
(No) reect a precise adjustment in the decision criteria of the model.
DISCOURSE ANALYSIS FOR THE DEVELOPMENT OF A
CYBERGROOMING DETECTION MODEL ON ROBLOX
Número 29 / AGOSTO, 2026 (13-35)
30
CONCLUSIONS
This study performed the rst discourse analysis of groomers in Spanish-
speaking Roblox communities by identifying a series of specic
linguistic patterns. After analyzing a corpus of 25 conversations, a pattern
of discursive organization was delimited in the form of a sequence of
seven conversational modules with specic relative frequencies and
predictability percentages that reect a consistent strategic progression:
salutation, personal information request, robux or reward oer, sexual
content request, resistance management, compliments, and closing.
The robux or rewards oer module was particularly signicant,
presented in 96% of the conversations analyzed. This demonstrates
that the material incentives inherent to the videogame replace the
aective goods oer, reported in previous research on digital grooming
discourse. Similarly, the module of sexual content request was identied
in 100% of the conversations and recognized as the dening criterion of
cybergrooming discourse on Roblox.
With reference to the lexico-grammatical patterns, 21 specic patterns
were identied, with 224 associated collocations, which collectively
demonstrated their eectiveness for integration into automated detection
systems. An analysis based on concepts from Systemic Functional
Linguistics and the identied pragmatic-discursive strategies revealed
that the groomer systematically employs constructions involving
verbal, relational, mental, material, and behavioral processes, as well as
declarative and interrogative modes of expression, most frequently to
manipulate minors and obtain sexual content while projecting an image
of legitimacy and closeness.
Furthermore, the computational model developed demonstrated to be
an eective tool for the automatic detection of cyber-grooming, due
to the integration of contributions from discourse analysis as the basis
for development and the data processing techniques employed. This is
reected in the evaluation of performance metrics, which indicates a
correct classication with 93.33% accuracy.
DECLARATION OF CONFLICTS OF INTEREST: The authors
state that they have no conicts of interest.
AUTHOR CONTRIBUTIONS STATEMENT AND
ACKNOWLEDGMENTS: The contribution of each author is listed
using the CRediT Taxonomy below:
Ana Paola Castañón Marroquín, Brenda Ailed Rodríguez Colis, Andrea Bazán Durán, Luis Enrique Colmenares-Guillen
CHAKIÑAN. Revista de Ciencias Sociales y Humanidades / ISSN 2550 - 6722
31
− Ana Paola Castañón Marroquín: Lead autor, Conceptualization,
Formal Analysis, Research, Methodology, Validation, Visualization,
Writing-revision, and editing.
− Brenda Ailed Rodríguez Colis: Conceptualization, Formal Analysis,
Research, Methodology, Validation, Visualization, Writing-original
draft, Writing-revision, and editing.
− Andrea Bazán Durán: Conceptualization, Formal Analysis,
Research, Methodology, Project Management, Resources,
Validation, Visualization, Writing-original draft, Writing-revision,
− Luis Enrique Colmenares-Guillen: Conceptualization, Research,
Methodology, Project Management, Resources, Supervision,
Validation, Writing-original draft, Writing-revision, and editing.
The authors would like to thank the Benemérita Universidad Autónoma
de Puebla for its support, and in particular to students Karla Hernández
and Pedro Vera of the Bachelor’s Program in Forensic Science, and
Jesús Semita and María Sarmiento of the Bachelor’s Program in
Criminology, who participated in the collection of the corpus over an
eight-month period at the Digital Forensic Analysis Laboratory of the
School of Computer Science at the Benemérita Universidad Autónoma
de Puebla, although they are not responsible for the content of this
article.
STATEMENT OF DATA AVAILABILITY: The authors declare that
the data used in this research are available and freely accessible for
analysis by interested parties, in the repository: https://doi.org/10.5281/
zenodo.19241233
BIBLIOGRAPHIC REFERENCES
Allgaier, J., & Pryss, R. (2024). Cross-Validation Visualized: A
Narrative Guide to Advanced Methods. Machine Learning
and Knowledge Extraction, 6(2), 1378–1388. https://doi.
org/10.3390/make6020065
Badman, A., & Kosinski, M. (2024). ¿Qué es un conjunto de datos?
IBM. https://www.ibm.com/mx-es/think/topics/dataset
Broome, L., Izura, C., & Davies, J. (2025). An investigation of the
linguistic and deceptive characteristics of online grooming
types. Journal of Sexual Aggression, 31(3), 378–395. https://
doi.org/10.1080/13552600.2023.2300470
Calsamiglia, H., & Tuson, A. (2012). Las cosas del decir. Ariel.
Carville, O., & D’Anastasio, C. (2024, July 22). Roblox Predator
Problem Potentially Exposes Kids to Pedophiles. Bloomberg.
DISCOURSE ANALYSIS FOR THE DEVELOPMENT OF A
CYBERGROOMING DETECTION MODEL ON ROBLOX
Número 29 / AGOSTO, 2026 (13-35)
32
https://www.bloomberg.com/features/2024-roblox-pedophile-
problem/
Chiang, E., & Grant, T. (2019). Deceptive Identity Performance:
Oender Moves and Multiple Identities in Online Child Abuse
Conversations. Applied Linguistics, 40(4), 675–698. https://doi.
org/10.1093/applin/amy007
Daohuan, L., & Xuri, T. (2023). The Contextualized Representation of
Collocation. En M. Sun, B. Qin, X. Qiu, J. Jiang, & X. Han
(Eds.) The 22nd Chinese National Conference on Computational
Linguistics (1st ed., pp. 836–846). ACL Anthology. https://
aclanthology.org/2023.ccl-1.71/
Dorsey, E. (2024). Problems at Roblox (RBLX) #4. The Bear Cave.
https://thebearcave.substack.com/p/problems-at-roblox-rblx-4
Evans, C., & Lorenzo-Dus, N. (2025). A corpus-assisted discourse
analysis of children’s and groomers’ talk in online grooming
interactions. Applied Corpus Linguistics, 5(3), 100147. https://
doi.org/10.1016/j.acorp.2025.100147
Excelin, G., Dinansyah, F., & Anugrah, C. (2024). Communication
Patterns in the Use of Communication Features in Online Games
Case Study: Valorant. Jurnal Vokasi Indonesia, 11(2), 45–56.
https://doi.org/10.7454/jvi.v11i2.1203
Gillings, M., Learmonth, M., & Mautner, G. (2024). Taking the Road
Less Travelled: How Corpus‐Assisted Discourse Studies Can
Enrich Qualitative Explorations of Large Textual Datasets.
British Journal of Management, 35(4), 1667–1679. https://doi.
org/10.1111/1467-8551.12816
González-Sanz, M. (2024). Análisis pragmalingüístico de las
secuencias de apertura en conversación de WhatsApp. Rilce
Revista de Filología Hispánica, 40(3), 995–1023. https://doi.
org/10.15581/008.40.3.995-1023
Halliday, M., & Matthiessen, C. (2004). An Introduction to Functional
Grammar. Routledge.
Hunston, S. (2010). How can a corpus be used to explore patterns? In
A. O’Keee & M. McCarthy (Eds.), The Routledge Handbook
of Corpus Linguistics (1st ed., pp. 152–166). Routledge. https://
doi.org/10.4324/978003856949
Hunston, S. (2025). Pattern, Construction, System. Cambridge
University Press. https://doi.org/10.1017/9781009629065
IBM. (2025). Métricas de evaluación de precisión. IBM. https://www.
ibm.com/docs/es/ws-and-kc?topic=metrics-accuracy
Ana Paola Castañón Marroquín, Brenda Ailed Rodríguez Colis, Andrea Bazán Durán, Luis Enrique Colmenares-Guillen
CHAKIÑAN. Revista de Ciencias Sociales y Humanidades / ISSN 2550 - 6722
33
Isti’anah, A., Febrina, R., Suhandano, S., Winarti, D., Sutrisno, A., &
Jumanto, J. (2023). Big Data, Computer, and Technology in
Language Studies: The Potentials of Sketch Engine in Indonesia’s
Research. 2023 International Seminar on Application for
Technology of Information and Communication (ISemantic),
46–51. https://doi.org/10.1109/iSemantic59612.2023.10295357
Joleby, M., Lunde, C., Landström, S., & Jonsson, L. (2021). Oender
strategies for engaging children in online sexual activity. Child
Abuse & Neglect, 120, 105214. https://doi.org/10.1016/j.
chiabu.2021.105214
Kou, Y., Hernández, R., & Gui, X. (2025). “The System is Made
to Inherently Push Child Gambling in my Opinion”: Child
Safety, Monetization, and Moderation on Roblox. In N.
Yamashita, V. Evers, K. Yatani, X. Ding, B. Lee, M. Chetty, &
P. Toups-Dugas (Eds.) CHI ‘25: Proceedings of the 2025 CHI
Conference on Human Factors in Computing Systems (1st ed.,
pp. 1-18). Association for Computing Machinery. https://doi.
org/10.1145/3706598.3713170
Lorenzo-Dus, N. (2023). Digital Grooming Discourses of Manipulation
in Cyber-crime. Oxford University Press. https://doi.
org/10.1093/oso/9780190845193.001.0001
Lorenzo-Dus, N., & Izura, C. (2017). “cause ur special”: Understanding
trust and complimenting behaviour in online grooming discourse.
Journal of Pragmatics, 112, 68–82. https://doi.org/10.1016/j.
pragma.2017.01.004
Lorenzo-Dus, N., Kinzel, A., & Di Cristofaro, M. (2020). The
communicative modus operandi of online child sexual groomers:
Recurring patterns in their language use. Journal of Pragmatics,
155, 15–27. https://doi.org/10.1016/j.pragma.2019.09.010
Lutzky, U., & Kehoe, A. (2022). Using corpus linguistics to study online
data. In C. Vasquez (Ed.), Research methods for digital discourse
analysis (2nd ed., pp. 219–236). Bloomsbury Academic.
Lyu, J., & Ishwaran, H. (2023). Commentary: To classify means to choose
a threshold. Journal of Thoracic and Cardiovascular Surgery,
165(4), 1443–1445. https://doi.org/10.1016/j.jtcvs.2021.08.009
Manning, C., Raghavan P., & Schütze H. (2009). Introduction to
Information Retrieval. Cambridge University Press. https://nlp.
stanford.edu/IR-book/pdf/irbookonlinereading.pdf
Mariscal, G., Marbán, Ó., & Fernández, C. (2010). A survey of
data mining and knowledge discovery process models and
methodologies. The Knowledge Engineering Review, 25(2),
DISCOURSE ANALYSIS FOR THE DEVELOPMENT OF A
CYBERGROOMING DETECTION MODEL ON ROBLOX
Número 29 / AGOSTO, 2026 (13-35)
34
137–166. https://doi.org/10.1017/S0269888910000032
McDonough, M. (2023). Cross-validation. Encyclopedia Britannica.
https://www.britannica.com/technology/cross-validation-
computer-science
Molina, J., Fàbregues, S., & Escalante, E. (2024). Métodos mixtos de
investigación. Ediciones Pirámide.
Morreale, D., & Rosa, A. (2024). Media Technologies and
Epistemologies: The Platforming of Everything| Roblox and
the Pervasiveness of Play: What Game-Making Communities
Can Teach Us About Participatory Practices in Anity Spaces.
International Journal of Communication, 18, 4281–4299.
https://ijoc.org/index.php/ijoc/article/view/21902
Ortiz, N. (2024). El grooming como delito trasnacional: Avances y
desafíos de la legislación y la cooperación entre Colombia,
Argentina y México [Master’s Thesis, Universidad
Externado de Colombia]. https://bdigital.uexternado.edu.co/
handle/001/26312
Partington, A., Duguid, A., & Taylor, C. (2013). Patterns and meanings
in discourse: theory and practice in corpus-assisted discourse
studies (CADS). John Benjamins Publishing Company.
Pérez, J. M., Furman, D. A., Alonso Alemany, L., & Luque, F. M. (2022).
RoBERTuito: A pre-trained language model for social media text
in Spanish. In N. Calzolari, F. Béchet, P. Blache, K. Choukri, C.
Cieri, T. Declerck, S. Goggi, H. Isahara, B. Maegaard, J. Mariani,
H. Mazo, J. Odijk, & S. Piperidis (Eds.), Proceedings of the
Thirteenth Language Resources and Evaluation Conference (1st
ed., pp. 7235–7243. https://aclanthology.org/2022.lrec-1.785
Pienczykowski, B., & Madella, P. (2026). ‘I was so Afraid All You
Wanted from Me Was Sex’: A Corpus-Assisted Study in the
Pragmatics of Manipulation in Online Child Sexual Groomers’
Discourse. Corpus Pragmatics, 10, article number 11. https://
doi.org/10.1007/s41701-025-00218-0
Pyslar, M. (2025). The Evolution of Community Management in
Gaming: From Forums to AI-driven Engagement. Актуальные
Исследования, 9(244), 69–73. https://doi.org/10.5281/
zenodo.14968205
Red Grooming Latam. (2024). Los riesgos de niñas, niños y adolescentes
en internet [[PDF le]. https://www.groominglatam.org/
wp-content/uploads/2025/02/INFORME-2024-FLYER-
INFORMATIVO.pdf
Rozgonjuk, D., Schivinski, B., Pontes, H. M., & Montag, C.
Ana Paola Castañón Marroquín, Brenda Ailed Rodríguez Colis, Andrea Bazán Durán, Luis Enrique Colmenares-Guillen
CHAKIÑAN. Revista de Ciencias Sociales y Humanidades / ISSN 2550 - 6722
35
(2023). Problematic Online Behaviors Among Gamers: the
Links Between Problematic Gaming, Gambling, Shopping,
Pornography Use, and Social Networking. International Journal
of Mental Health and Addiction, 21(1), 240–257. https://doi.
org/10.1007/s11469-021-00590-3
Scheglo, E. (2007). Sequence organization in interaction: A primer in
conversation analysis I. Cambridge University Press.
Schweinberger, M. (2024). Analyzing Collocations and N-grams in R.
Language Technology and Data Analysis Laboratory (LADAL).
https://slcladal.github.io/coll.html
Sorlin, S. (2017). The pragmatics of manipulation: Exploiting im/
politeness theories. Journal of Pragmatics, 121, 132–146.
https://doi.org/10.1016/j.pragma.2017.10.002
Steward, K. (2023). Sensitivity vs specicity: Test accuracy explained.
Technology Networks. https://www.technologynetworks.com/
analysis/articles/sensitivity-vs-specicity-318222
Straka, M., & Straková, J. (2020). UDPipe at EvaLatin 2020:
Contextualized Embeddings and Treebank Embeddings. En R.
Sprugnoli & M. Passarotti (Eds.), Proceedings of LT4HALA
2020 - 1st Workshop on Language Technologies for Historical
and Ancient Languages (1st ed., pp. 124–129). European
Language Resources Association (ELRA). https://aclanthology.
org/2020.lt4hala-1.20/
Ting, K. M. (2010). Confusion Matrix. In G. I. Sammut Claude & Webb
(Eds.), Encyclopedia of Machine Learning (1st ed., p. 209).
Springer. https://doi.org/10.1007/978-0-387-30164-8_157
Vaamode, G., & González, F. (2008). Clasicación verbal, etiquetación
semántica e información lexicográca en el proyecto ADESSE.
En M. Verdejo & A. Serrano (Eds.), Acceso y visibilidad de la
información multilingüe en la red: El rol de la semántica (1st
ed., pp. 225–236). UNED.
Van Dijk, T. (1997). Discourse as Structure and Process. Sage
Publications.
Walsh, M. J. (2022). About ‘face’’: Reconsidering Goman’s theory
of face-work for digital culture.’ In M. Jacobsen & G. Smith
(Eds.), The Routledge International Handbook of Goman
Studies (1st ed., pp. 207–218). Routledge. https://doi.
org/10.4324/9781003160861-20
