Exploring Audience Reactions on YouTube: An Approach to Sentiment

and Toxicity Analysis

Andr

e F. Rollwagen

1, 2 a

, Gabriel Zurawski

2 b

, Stefano Carraro

, Roberto Tietzmann

2 c

Marcelo C. Fontoura

2 d

and Isabel H. Manssour

2 e

Federal Institute of Education, Science and Technology Sul-Rio-Grandense, IFSUL, Brazil

Pontiﬁcal Catholic University of Rio Grande do Sul - PUCRS, Porto Alegre, Brazil

Keywords:

YouTube, Social Media Analysis, Sentiment Analysis, Toxic Speech Detection, Portuguese Language.

Abstract:

Social media platforms like YouTube have become central spaces for public expression, allowing users to share

their emotional reactions and, at times, toxic content. Understanding these reactions requires scalable and

reproducible approaches that can handle informal, user-generated content. This paper presents an integrated

approach for analyzing sentiment polarity and detecting toxic speech in YouTube video comments written in

Portuguese. For this, we developed a set of Python scripts that automate data collection and apply Natural

Language Processing (NLP) techniques to perform both tasks. These scripts are publicly available and can be

adapted for use in various video and social analysis contexts. Interactive visualizations were also generated

to support the interpretation of results. The applicability of the approach is demonstrated through two case

studies involving highly controversial videos, which allow us to explore the relationship between sentiment,

toxicity, and audience engagement patterns. The results provide valuable insights into the dynamics of public

discourse and offer tools for future research on audience speech analysis on YouTube.

1 INTRODUCTION

Social media platforms like YouTube have become

central spaces for public debate. Comments posted

by users on videocasts or communication channels,

e.g., often express opinions, emotions, and sometimes

manifestations of toxic language. Although rich in

content, such environments are also prone to offen-

sive, discriminatory, or aggressive messages (Fortuna

et al., 2021), reinforcing the need for automated meth-

ods to detect and interpret these expressions.

In this context, it is essential to clarify key con-

cepts. Hate speech, as deﬁned by (Fortuna et al.,

2021), targets individuals or groups based on traits

like race, gender, or religion, often carrying legal im-

plications. Toxicity is a broader concept that includes

hate speech along with offensive language, threats,

and verbal abuse. Here, we use the term toxic speech

to encompass this wider range of harmful expressions.

https://orcid.org/0000-0002-1106-7222

https://orcid.org/0009-0007-4594-1938

https://orcid.org/0000-0002-8270-0865

https://orcid.org/0000-0002-3229-0167

https://orcid.org/0000-0001-9446-6757

Most existing works focus exclusively either on

sentiment analysis (Kurtz et al., 2025; Dharini et al.,

2025) or on the detection of toxic speech (Bonetti

et al., 2023; Maity et al., 2024). However, few works

explore both dimensions jointly, especially in high-

visibility contexts such as videocasts, where the com-

bination of emotional tone and the presence of toxic

speech can offer a more comprehensive understand-

ing of audience reactions. Additionally, while much

research centers on platforms like Twitter (Campan

and Holtke, 2024; Siegel, 2020) (now called X),

YouTube, despite its 2.7 billion users in 2023 (Muneer

and Khan, 2025), remains underexplored. These ﬁg-

ures underscore its strategic relevance for studies on

large-scale interactions and discursive behavior.

Language bias is another limitation, with a

predominance of English-centered analyses (Siegel,

2020). Portuguese remains underexplored, especially

in works that provide open repositories and repro-

ducible methodologies (Leite et al., 2020). In this

context, we focus on the Portuguese language, moti-

vated by both linguistic proximity and the nature of

the conﬂicts analyzed, which emerged in Brazilian

media and are expressed in that language.

Thus, this work aims to develop an approach

Rollwagen, A. F., Zurawski, G., Carraro, S., Tietzmann, R., Fontoura, M. C. and Manssour, I. H.

Exploring Audience Reactions on YouTube: An Approach to Sentiment and Toxicity Analysis.

DOI: 10.5220/0013718200003985

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

In Proceedings of the 21st International Conference on Web Information Systems and Technologies (WEBIST 2025), pages 481-488

ISBN: 978-989-758-772-6; ISSN: 2184-3252

481

for performing sentiment analysis and toxicity de-

tection in Portuguese-language YouTube comments.

The proposed solution involves the development of

Python scripts that enable comment collection, sen-

timent classiﬁcation (positive, negative, or neutral),

and the identiﬁcation of different levels of toxicity.

Additionally, visualizations are generated to facili-

tate the interpretation of results and support the ex-

ploratory analysis of audience reaction patterns. To

demonstrate the applicability of our approach, we an-

alyzed two high-impact YouTube videos, examining

how content type shapes the tone of comments and

engagement patterns, supported by visualizations and

metrics that reveal discursive trends.

Therefore, the main contributions of this work are:

• The integrated application of sentiment analy-

sis and toxicity detection on the same textual

dataset, enabling the investigation of relationships

between emotional polarity and offensive speech.

• The set of Python scripts that automate the collec-

tion, sentiment analysis, and toxicity detection of

comments on YouTube videos.

• The possibility of visually exploring correlations

between sentiment and toxicity, allowing for the

visual analysis of comment dynamics.

• The demonstration of the proposed approach

through two case studies involving widely dis-

cussed videos, highlighting its ability to reveal

discursive patterns in contexts marked by high au-

dience engagement and social controversy.

The remainder of this paper is organized as fol-

lows. Section 2 presents related work. Section 3 de-

scribes the methodology. The results are discussed

in Section 4, together with the description of the two

case studies. Section 5 outlines our ﬁndings.

2 RELATED WORK

Sentiment analysis and toxic speech detection on so-

cial media have been widely investigated as strategies

to understand public opinion and engagement patterns

in political, social, and cultural contexts. Recent ad-

vances apply Natural Language Processing (NLP) and

deep learning techniques to automate the collection,

preprocessing, and analysis of textual data (Yadollahi

et al., 2022).

The study by (Parraga-Alava et al., 2021) employs

an API-based pipeline to analyze tweet sentiment dur-

ing Ecuador’s presidential election, revealing correla-

tions with the results. (Vargas et al., 2022) proposed

an Unsupervised method for Aspect Term Extraction

(UnATE) that combines topic models, word embed-

dings, and a ﬁne-tuned BERT model, proving effec-

tive for sentiment analysis in product reviews without

labeled data.

In YouTube, (Adesina and Howe, 2024) apply

a multilabel model using Graph Convolutional Net-

works (GCNs) and LLMs to classify sentiments in

political comments. Also, (Bindhumol et al., 2024)

combine Convolutional LSTM and Convolutional

Gated Recurrent Unit (CGRU) networks to support

recommendation systems based on sentiments ex-

tracted from user comments. (Guzman et al., 2025)

conduct a manual sentiment analysis to reveal pat-

terns of distrust toward vaccines and public policies

during the pandemic, while (Schmidt et al., 2023)

analyze georeferenced sentiment on X after Elon

Musk’s acquisition, exposing regional polarization.

In toxic speech detection, (Shahi and Majchrzak,

2025) demonstrates that leveraging data from multi-

ple platforms improves classiﬁer performance when

corpora are similar. (Kamma et al., 2025) propose a

Bi-LSTM with hierarchical attention, achieving high

precision and recall in comment classiﬁcation.

Table 1 presents a comparative summary of re-

lated work, highlighting the data sources, models and

libraries, and key contributions.

In contrast to related works, we propose an ap-

proach for analyzing YouTube comments written in

Portuguese that combines sentiment analysis (posi-

tive, negative, and neutral) and toxic speech detec-

tion in a uniﬁed manner. While most works focus

on English data analysis and typically address only

one of these dimensions, we provide both classiﬁ-

cations simultaneously. Interactive visualizations are

also available to support the interpretation and explo-

ration of the results. We made the code available on

GitHub

to facilitate reproducibility and contribute to

research on public speech in digital media.

3 RESEARCH METHODOLOGY

This section outlines the methodology used for col-

lecting comments from YouTube videos, performing

sentiment analysis, detecting toxic speech, and pre-

senting the results through visualizations. To support

this process, a set of Python scripts has been devel-

oped and is available, enabling the systematic execu-

tion of data gathering, processing, and analysis, the

latter supported by LLMs. Figure 1 illustrates the

methodological workﬂow adopted.

https://github.com/DAVINTLAB/Toxicytube

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

482

Table 1: Comparison of related work.

Reference Source

Sentiment

Analysis

Toxic Detection

Visual

Analysis

Language

Code

Available

Main Contribution

(Parraga-Alava et al.,

2021)

TextBlob

✓ Spanish Correlation between sentiment polarity

and electoral outcomes

(Vargas et al., 2022)

Product

Reviews

UNATE English Proposal of UNATE method for aspect-

level sentiment analysis

(Adesina and Howe,

2024)

YouTube

GCNs,

ChatGPT-4,

VADER

✓ English Multilabel political sentiment classiﬁca-

tion using NLP

(Guzman et al., 2025) X Not speciﬁed ✓

Filipino,

English

Patterns of distrust and negative senti-

ment during pandemic

(Bindhumol et al.,

2024)

YouTube CLSTM, CGRU ✓ English Sentiment-based product recommenda-

tion from YouTube comments

(Schmidt et al., 2023) X RoBERTa ✓ English Detection of negative sentiment and re-

gional polarization

(Shahi and Majchrzak,

2025)

YouTube,

X, Gab,

Wikipedia

SVM, LSTM,

BERT

English,

German

✓ Improved hate speech detection using

cross-platform dataset similarity

(Kamma et al., 2025) X

Bi-LSTM,

Word2Vec

English Hierarchical Bi-LSTM for toxic com-

ment detection

Our Approach YouTube

XLM-T,

GPT-3.5-Turbo

XLM-T,

GPT-3.5-Turbo,

Detoxify

✓ Portuguese ✓ Combined analysis of sentiment, toxic

speech detection, and visual exploration

using dashboards

TextBlob: Python library for processing textual data.

VADER: Valence Aware Dictionary and Sentiment Reasoner.

Word2Vec: Neural

embedding technique that represents words as continuous vectors in a latent space, capturing their semantic similarity based on context.

Figure 1: Methodological workﬂow adopted for the senti-

ment analysis and toxic speech detection.

3.1 Data Gathering

The data used in the proposed approach is obtained

by collecting comments from videos published on

YouTube, excluding live streams. To accomplish

this, we developed a Python script (main.py) that

uses the YouTube Data API

to perform comment

collection from videos already available on the plat-

form. To run the script, the user must provide a

https://developers.google.com/youtube/v3

Google API key, the video ID, and the desired num-

ber (N) of comments. The system retrieves the top

N most liked comments and saves them in a ﬁle

named comments <VIDEO ID>.json, containing the

ﬁelds id, author, message, publishedAt, and like-

Count, which will be used in analysis steps.

3.2 Data Processing

This section presents the automated processing of

YouTube comments, combining pre-trained NLP

models and GPT-3.5-Turbo for toxic speech detection

and sentiment classiﬁcation.

3.2.1 Toxic Speech Detection

Toxic speech detection is performed by the Python

script ToxicClassification.py, using the JSON

ﬁle generated during the data collection. The clas-

siﬁcation process uses the Detoxify (Hanu and Uni-

tary team, 2020) library, which provides pre-trained

transformer-based models to identify seven types

of toxicity, including toxicity (general rude or dis-

respectful content), severe toxicity (highly aggres-

sive or hateful speech), obscene (use of offensive

or profane language), identity attack (targeting peo-

ple based on attributes like race, gender, or reli-

gion), insult (personal attacks or name-calling), threat

(statements expressing intent to harm), and sex-

ual explicit (sexual explicit language). The output is

a toxic comments <VIDEO ID>.json ﬁle that con-

tains the original comment data along with the pre-

dicted scores for each toxicity type and serves as input

for subsequent analyses.

A comment is assigned to a category if its pre-

dicted probability score exceeds the 0.7 threshold.

Exploring Audience Reactions on YouTube: An Approach to Sentiment and Toxicity Analysis

483

The Detoxify library was selected for its multilin-

gual coverage and practical effectiveness in identi-

fying toxic content on social media, including in

Portuguese-language data.

3.2.2 Sentiment Analysis with XLM-T

Sentiment analysis is performed using two ap-

proaches. The ﬁrst, via the Python script

SentimentsClassificationWithXLM.py, ap-

plies the pre-trained model cardiffnlp/twitter-

xlm-roberta-base-sentiment, ﬁne-tuned for mul-

tilingual X data and based on the XLM-RoBERTa

architecture (Barbieri et al., 2022).

It was chosen for its performance on short user-

generated texts in the TweetEval benchmark (Barbieri

et al., 2020). The model generalizes well to other lan-

guages, including Portuguese, despite being mainly

trained on English data.

The script reads comments from a JSON ﬁle and

classiﬁes each as positive, negative, or neutral us-

ing the model. The predicted labels are added as a

new column called sentiment, and the results are ex-

ported to sentiments <VIDEO ID> XLM-T.json, in-

dicating the model used for the classiﬁcation.

3.2.3 Sentiment Analysis with GPT-3.5-Turbo

The second sentiment analysis approach

was implemented using OpenAI’s GPT-3.5-

Turbo model via API, through the script

SentimentClassificationWithGPT.py. GPT-

3.5-Turbo was selected due to its high linguistic

accuracy and contextual understanding, making it

suitable for more nuanced sentiment analysis, partic-

ularly in user-generated content that often includes

informal or ambiguous language (Zhang et al., 2024).

However, its use requires access to the OpenAI API,

which is subject to usage limitations and costs.

The script receives a JSON ﬁle containing user

comments and applies a speciﬁc prompt designed to

instruct the model to classify the sentiment of each

comment, such as:

You are a sentiment analyzer.

Classify the sentiment of the following comment

as ‘‘positive’’, ‘‘negative’’, or ‘‘neutral’’.

Respond with the corresponding word only.

The result is exported to a JSON ﬁle named

sentiments <VIDEO ID> GPT.json, where the suf-

ﬁx GPT indicates that the ﬁle was generated using the

GPT-3.5-Turbo model.

3.2.4 Models Agreement

This step aimed to evaluate the consistency be-

tween the sentiment classiﬁcations produced by

the XLM-T and GPT-3.5-Turbo models. To that

end, we developed a script in Python named

ComparisonRoBERTa GPT.py (“RoBERTa” referring

to the architecture underlying XLM-T). The input

consists of two JSON ﬁles: one containing the sen-

timent labels assigned by XLM-T and the other by

GPT-3.5-Turbo.

The script processes these ﬁles to calculate agree-

ment rates between the models’ classiﬁcations and

generates summary statistics to assess alignment and

potential divergences in interpretation. The output is

a sentiments comparison <VIDEO ID>.json ﬁle

containing only the comments for which both mod-

els assigned the same sentiment rating.

3.3 Results Visualization

To support the interpretation of the classi-

ﬁed data, we developed two Python scripts:

SentimentsVisualization.py, for visu-

alizing sentiment classiﬁcation results, and

ToxicCommentsVisualization.py, for visual-

izing toxic speech detection results. Both use the

Plotly library

to generate interactive visualizations

based on their respective JSON input ﬁles.

The script generates several visualizations, includ-

ing time series plots that display temporal variations

in sentiment polarity or toxicity categories, bar charts

showing the total number of comments per sentiment

or toxicity label, pie charts representing these propor-

tions as percentages, and word clouds to highlight the

most frequent terms. The resulting visualizations are

organized into dashboards and exported as interactive

HTML ﬁles, as shown in Figure 2.

4 RESULTS AND DISCUSSION

This section presents and discusses the results of ap-

plying our methodology to two YouTube case stud-

ies: the Monark case

(Section 4.1) and the Nego do

Borel case

(Section 4.2). Section 4.3 compares both

cases. The analysis addresses sentiment, toxicity, and

user engagement patterns.

The selected videos were chosen for their high vis-

ibility and controversial nature, making them suitable

for analyzing patterns of sentiment and toxic speech.

The Monark case centers on a politically charged

statement about Nazism, which sparked public out-

rage and moral condemnation. In contrast, the Nego

https://plotly.com/python/

https://www.youtube.com/watch?v=Qo2kYS2 XnI

https://www.youtube.com/watch?v=FY3m6hMyh3g

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

484

Figure 2: Interactive visualizations in HTML format.

do Borel case revolves around a music video that gen-

erated diverse reactions, inﬂuenced by cultural refer-

ences and the artist’s public image. Both were se-

lected for their thematic relevance, high engagement,

and suitability for contrasting public responses.

We deﬁned three hypotheses based on the themes

and public impact of the videos. First, we expected a

peak in comment activity for the Monark case shortly

after the video’s release, due to the controversy’s in-

tensity. Second, we assumed that political polariza-

tion could intensify user engagement, leading users

to actively seek out related content. Lastly, we hy-

pothesized that top-liked comments, while critical,

would avoid overtly toxic or hateful language, favor-

ing provocative but moderate discourse. These hy-

potheses are examined in Sections 4.1, 4.2, and 4.3.

The comments were collected from each video’s

publication date until May 7, 2025. For the Monark

case, the collection period began on February 9, 2022,

yielding 3,798 comments. In the Nego do Borel case,

it started on July 9, 2018, with a total of 114,695 com-

ments. It is important to note that the script used does

not collect comment replies.

To ensure a fair comparison despite the differ-

ence in comment volume, we selected a sample of

the 1,000 most liked comments per case study. This

sampling enabled a balanced analysis in terms of vol-

ume, focusing on the most relevant and highly en-

gaged content.

We applied two models for sentiment classiﬁca-

tion: GPT-3.5-Turbo (via prompts) and XLM-T (mul-

tilingual, ﬁne-tuned). Both models were applied to

the datasets in parallel, allowing us to compare their

outputs and assess their consistency. This model

agreement step was applied to both case studies, with

77.2% agreement for Monark and 56.9% for Nego do

Borel. Due to the low alignment between the models,

we based our analyses on GPT classiﬁcations, given

its strong zero-shot and few-shot performance, as well

as its previously discussed advantages (Zhang et al.,

2024), requiring no additional training.

4.1 Monark Case

This case study refers to the audience reaction to con-

troversial statements by Bruno Aiub (Monark) on the

Flow Podcast

. As the original video was removed

from the platform, making direct analysis unfeasible,

we used an alternative source, a video published by

the Jornalismo TV Cultura channel, which reported

his dismissal from the program under the headline

“Monark

e demitido, ap

os defesa de partido nazista”

(translated as “Monark is ﬁred after defending the

existence of a Nazi party”). The removed video,

which lasted 1 minute and 46 seconds, had garnered

1,414,319 views and 6,081 comments.

We analyzed sentiment and toxicity to assess pub-

lic reaction. As shown in Figure 3a, the results in-

dicate a strong prevalence of criticism, with 77.9%

of the comments classiﬁed as negative. Although the

video had a journalistic focus, users used the com-

ment section to voice direct criticism and moral judg-

ment.

Figure 3: Classiﬁcation of audience reactions by sentiment

and toxicity Monark case.

Just 15.4% of comments conveyed positive senti-

ment, reﬂecting limited support, irony, or partial de-

fense. In the sample, 33% of the comments showed

some degree of toxicity: 21.2% were toxic and 11.3%

insults (212 and 113 comments, respectively), as

shown in Figure 3b). These results indicate that, al-

though criticism prevailed, a signiﬁcant portion of

the engagement occurred through offensive language.

More severe categories like threats and identity at-

tacks were rare, supporting the hypothesis that top-

https://www.youtube.com/@FlowPodcast.

Exploring Audience Reactions on YouTube: An Approach to Sentiment and Toxicity Analysis

485

liked comments, though critical, avoid extreme or ex-

plicitly hateful speech.

Figure 4 shows the most frequent terms in

the comments, providing an overview of dominant

themes and expressions used by the audience.

Figure 4: Audience reactions to the Monark case.

Daily comment volume was analyzed to under-

stand engagement over time. Activity peaked on the

ﬁrst day (391 comments), then declined rapidly, in-

dicating a brief but intense reaction typical of viral

controversies, with limited sustained discussion.

To support contextualization, Table 2 presents the

top 10 most liked comments that reveal a discursive

plurality, combining direct criticism and political re-

ﬂections. While some contain toxic elements, they are

not dominant. The most liked comment — “Monark

e a melhor propaganda anti drogas q existe kkkkjkkk”

(translated as “Monark is the best anti-drug advertise-

ment that exists LOL”) — classiﬁed as positive, ex-

empliﬁes the use of humor as an indirect criticism.

Table 2: Partial list with top 10 most liked comments.

Likes Comments

5590 Monark

e a melhor propaganda anti drogas q existe kkkkjkkk

5504 O maior erro dele foi misturar lazer com coisa s

eria. Se drogar num debate

pol

ıtico

e o c

umulo da irresponsabilidade.

2219 Esse cara

e o esp

ecime perfeito para oq as pessoas chamam de idiota

1733 O professor que explicou sobre o holocausto foi bem assertivo: QUANDO

VOC

E TOLERANTE COM O INTOLER

AVEL...

1665 Monark acha que liberdade de express

e nao ter responsabilidade. Cara de-

fender que alguem seja antissemita racista . . .

1504 Perdeu dinheiro e se arrependeu, se n

ao fosse isso...

1336 Bem que diziam que um dia, o Monark ia falar uma porcaria MUITO grande

1267 Bons tempos em que monark era apenas uma marca de bicicletas. Saudades.

870 O cara foi demitido n

ao por causa do que ele falou e sim pela perdas dos patroci-

nadores. O mundo gira em torno do dinh. . .

603 Se esse cara vai trabalhar b

ebado n

ao tinham que trabalhar com ele.

E uma

irresponsabilidade

Negative sentiments are shown in red, positive in green, and toxic com-

ments are highlighted with a light gray background.

The ﬁndings indicate a predominance of nega-

tive sentiment and offensive discourse. However, the

low presence of severe toxicity among top-liked com-

ments suggests a social moderation, with less ap-

proval for more aggressive content.

4.2 Nego Do Borel Case

This case study examines the audience reception of

the music video “Me Solta” by Nego do Borel. At the

time of data collection, the 3 minutes and 46 seconds

video, published on YouTube, had over 231 million

views and 114.695 comments, including the replies.

Figure 5a depicts the proportions of positive, neu-

tral, and negative comments, while Figure 5b shows

the frequency of toxic content. The sentiment analy-

sis reveals that over half of the comments were nega-

tive. Despite being entertainment content, many users

used the comment section to express criticism or con-

troversy toward the artist or the music video. Toxicity

analysis revealed a low incidence of harmful speech:

only 8.2% of comments were classiﬁed as toxic, and

less than 1% into severe categories.

Figure 5: Classiﬁcation of audience reactions by sentiment

and toxicity in the Nego do Borel case.

A word cloud showing audience reactions to the

video, displayed in Figure 2, highlights the most fre-

quently used terms and offers qualitative insights into

the main themes and expressions present in public dis-

course.

To analyze engagement in the Nego do Borel case,

we examined comment distribution over time. The

peak occurred on July 10, 2018 (270 comments), one

day after publication, followed by a gradual decline -

173 and 71 comments on the two subsequent days.

This pattern reﬂects typical viral behavior: intense

initial reaction followed by rapid drop-off.

The table with the most liked comments pre-

sented in Figure 2 helps illustrate the tone and fram-

ing adopted by the audience, often marked by infor-

mal language and cultural references. For instance,

“ME LEMBREI DA VAI LACRAIA KKKK” (trans-

lated as “REMINDED ME OF VAI LACRAIA LOL”)

and “Borel t

a andando muito com a Anitta KKKK”

(“Borel is hanging out with Anitta too much LOL”)

reference pop culture ﬁgures, use humor and sarcasm.

Overall, the Nego do Borel case suggests that au-

dience engagement was driven more by cultural ref-

erences and entertainment than by moral judgment or

ideological polarization.

4.3 Comparative Analysis

The comparison between the Monark and Nego do

Borel cases reveals distinct patterns in public recep-

tion, sentiment, and toxicity. While both videos gen-

erated high engagement, the tone and nature of com-

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

486

ments reﬂect their differing sociopolitical and cultural

contexts. Figure 6 summarizes these contrasts, show-

ing sentiment polarity (6a) and toxicity levels (6b) for

each case.

Figure 6: Sentiment and toxicity distribution in case studies.

In the Monark case, 78% of comments were nega-

tive, reﬂecting moral and political criticism, and turn-

ing the comment section into a space for public ac-

countability and ideological positioning. In contrast,

the Nego do Borel case showed a more balanced sen-

timent, with lighter, pop culture–driven reactions de-

spite a high rate of negative comments ( 50%).

Regarding toxicity, the Monark case exhibited a

higher incidence of toxic speech, though severe cat-

egories were rare, indicating audience tolerance for

criticism within certain limits. In contrast, toxicity in

the Nego do Borel case was lower and milder, often

expressed through humor, irony, or nostalgia.

The nature and context of each case shaped au-

dience responses: the Monark case, rooted in politi-

cal controversy, triggered moral and hostile reactions,

while the Nego do Borel case, framed as entertain-

ment, prompted lighter, culturally driven commen-

tary. The comparison shows that audience engage-

ment depends not only on controversy severity but

also on how the audience frames the content. While

the Monark case sparked explicit outrage, the Nego

do Borel case elicited more ambiguous and emotion-

ally driven responses, with less toxicity. These ﬁnd-

ings suggest that sentiment polarity and toxicity are

shaped by content themes and social framing.

The initial hypothesis assumed intense comment

activity in the Monark case immediately after the

video’s release, which the data conﬁrmed. For the

Nego do Borel video, we anticipated longer-term en-

gagement, but it also peaked early, with a signiﬁcant

majority of comments concentrated within the initial

days. These ﬁndings suggest that the viral dynamics

of YouTube engagement tend to follow an intense but

short-lived cycle, regardless of content type.

In the Monark case, many comments appeared on

a journalistic video. This supports our second hypoth-

esis: that political polarization drives users to engage

with related content even after its removal. It is plau-

sible that many users had watched the original video

or excerpts and used the comment section to react to

the controversy.

5 CONCLUSIONS

This work proposed an integrated and reproducible

approach to analyze sentiment and detect toxic speech

in Portuguese-language YouTube comments. By de-

veloping a set of Python scripts, we automated the

processes of comment collection, sentiment classiﬁ-

cation using both GPT-3.5-Turbo and XLM-T, and

toxicity detection using the Detoxify library. The

generated outputs, complemented by interactive vi-

sualizations, provided a comprehensive view of au-

dience reactions to controversial content. We have

made the source code of our approach openly avail-

able on GitHub for anyone who wants to use it, along

with the case studies, including data and results.

The two case studies revealed distinct patterns of

audience response. The Monark case demonstrated

a predominance of negative sentiment and toxic lan-

guage, underscoring the impact of political polariza-

tion and moral judgment. The Nego do Borel case

exhibited a more heterogeneous emotional distribu-

tion, with pop culture references contributing to a dis-

course that was less aggressive overall. These ﬁnd-

ings suggest that public sentiment and toxicity are

shaped not only by the content itself but also by

how the audience socially frames the message and the

broader context in which the communication occurs.

As future work, we plan to expand the scope to

multilingual datasets and other social platforms. We

also aim to enhance the interpretability of classiﬁca-

tion outcomes by utilizing explainable AI techniques,

thereby contributing to greater transparency and ac-

countability in the analysis of digital public discourse.

ACKNOWLEDGEMENTS

Carraro would like to thank the Tutorial Educa-

tion Program (PET). Manssour would like to thank

the ﬁnancial support of the CNPq Scholarship -

Brazil (303208/2023-6). This paper was supported

by the Ministry of Science, Technology, and In-

novations, with resources from Law No. 8.248,

dated October 23, 1991, within the scope of PPI-

SOFTEX, coordinated by Softex, and published in

the Resid

encia em TIC 02 - Aditivo, Ofﬁcial Gazette

01245.012095/2020-56.

While preparing and revising this manuscript, we

used ChatGPT, Grammarly, and Google Translate to

enhance clarity and grammatical precision, as English

Exploring Audience Reactions on YouTube: An Approach to Sentiment and Toxicity Analysis

487

is not our ﬁrst language. The authors take full respon-

sibility for the content and its technical accuracy.

REFERENCES

Adesina, M. T. and Howe, L. (2024). Social media

(youtube) political sentiment multi-label analysis. In-

ternational Journal of Science and Research Archive,

12(02):2063–2071.

Barbieri, F., Camacho-Collados, J., Espinosa-Anke, L., and

Ronzano, F. (2020). Tweeteval: Uniﬁed benchmark

and comparative evaluation for tweet classiﬁcation.

In Proceedings of the 2020 Conference on Empirical

Methods in Natural Language Processing: Findings,

pages 1644–1650.

Barbieri, F., Espinosa Anke, L., and Camacho-Collados,

J. (2022). XLM-T: Multilingual language models in

Twitter for sentiment analysis and beyond. In Pro-

ceedings of the Thirteenth Language Resources and

Evaluation Conference, pages 258–266.

Bindhumol, M., Singh, T., and Patra, P. (2024). Sentiment

analysis using youtube comments. In 2024 15th Inter-

national Conference on Computing Communication

and Networking Technologies (ICCCNT), pages 1–7.

Bonetti, A., Mart

ınez-Sober, M., Torres, J. C., Vega, J. M.,

Pellerin, S., and Vila-Franc

es, J. (2023). Compari-

son between machine learning and deep learning ap-

proaches for the detection of toxic comments on social

networks. Applied Sciences, 13(10).

Campan, A. and Holtke, N. (2024). Beyond twitter: Explor-

ing alternative api sources for social media analytics.

In Proceedings of the 16th International Joint Confer-

ence on Knowledge Discovery, Knowledge Engineer-

ing and Knowledge Management (IC3K 2024), Vol.1,

pages 441–447.

Dharini, N., Madhuvanthi, M., Aswini, C. S., Lakshya, R.,

Triumbika, M., and Saranya, N. (2025). Ensemble-

driven multilingual sentiment analysis framework for

youtube comments with dashboard. In 2025 Interna-

tional Conference on Visual Analytics and Data Visu-

alization (ICVADV), pages 1524–1529.

Fortuna, P., Soler-Company, J., and Wanner, L. (2021).

How well do hate speech, toxicity, abusive and offen-

sive language classiﬁcation models generalize across

datasets? Information Processing & Management,

58(3):102524.

Guzman, J. P. S., Cheng, C. K., Bernadas, J. M. A. C., and

Lao, A. R. (2025). Aspect-level sentiment analysis of

ﬁlipino tweets during the covid-19 pandemic. In Pro-

ceedings of the 18th International Joint Conference

on Biomedical Engineering Systems and Technologies

(BIOSTEC 2025) - Vol.2, pages 343–350.

Hanu, L. and Unitary team (2020). Detoxify. Github.

https://github.com/unitaryai/detoxify.

Kamma, V., Wbaid, S., K, S., A, N., and Saravanan, T.

(2025). Toxic comment detection based on improved

bi directional long short term memory. In 2025 Inter-

national Conference on Intelligent Systems and Com-

putational Networks (ICISCN), pages 1–6.

Kurtz, G. B., de P. Carraro, S., Teixeira, C. R. G., Bandeira,

L. D., M

uller, B. L., Tietzmann, R., Silveira, M. S.,

and Manssour, I. H. (2025). Streamvis: An analy-

sis platform for youtube live chat audience interaction,

trends and controversial topics. In Proceedings of the

27th International Conference on Enterprise Informa-

tion Systems (ICEIS 2025) - Vol.2, pages 630–640.

Leite, J. A., Silva, D., Bontcheva, K., and Scarton, C.

(2020). Toxic language detection in social media for

Brazilian Portuguese: New dataset and multilingual

analysis. In Proceedings of the 1st Conference of the

Asia-Paciﬁc Chapter of the Association for Compu-

tational Linguistics and the 10th International Joint

Conference on Natural Language Processing, pages

914–924.

Maity, A., More, R., Patil, P. A., Oza, J., and Kambli, G.

(2024). Toxic comment detection using bidirectional

sequence classiﬁers. In 2024 2nd International Con-

ference on Intelligent Data Communication Technolo-

gies and Internet of Things (IDCIoT), pages 709–716.

Muneer, Q. and Khan, M. A. (2025). Role of youtube in

creating awareness of sustainable transportation: A

latent dirichlet allocation approach. Sustainable Fu-

tures, 9:100607.

Parraga-Alava, J., Rodas-Silva, J., Quimi, I., and Alcivar-

Cevallos, R. (2021). Opinion and sentiment analy-

sis of twitter users during the 2021 ecuador presiden-

tial election. In Proceedings of the 17th International

Conference on Web Information Systems and Tech-

nologies (WEBIST 2021), pages 257–266.

Schmidt, S., Zorenb

ohmer, C., Ariﬁ, D., and Resch, B.

(2023). Polarity-based sentiment analysis of georef-

erenced tweets related to the 2022 twitter acquisition.

Information, 14(2):71.

Shahi, G. K. and Majchrzak, T. A. (2025). Hate speech de-

tection using cross-platform social media data in en-

glish and german language. In Proceedings of the 20th

International Conference on Web Information Systems

and Technologies (WEBIST 2024), pages 131–140.

Siegel, A. A. (2020). Online Hate Speech, pages 56–88.

SSRC Anxieties of Democracy. Cambridge University

Press, Cambridge.

Vargas, D. S., Pessutto, L. R. C., and Moreira, V. P. (2022).

Unsupervised aspect term extraction for sentiment

analysis through automatic labeling. In Proceedings

of the 18th International Conference on Web Informa-

tion Systems and Technologies (WEBIST 2022), pages

344–354.

Yadollahi, A., Shahraki, A. G., and Zaiane, O. R. (2022).

Deep learning in sentiment analysis: Recent architec-

tures and trends. ACM Computing Surveys, 54(2):1–

36.

Zhang, W., Deng, Y., Liu, B., Pan, S. J., and Bing, L. (2024).

Sentiment analysis in the era of large language mod-

els: A reality check. In Findings of the Association

for Computational Linguistics: NAACL 2024, pages

3883–3906.

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

488