StreamVis: An Analysis Platform for YouTube Live Chat Audience

Interaction, Trends and Controversial Topics

Gabriela B. Kurtz

1 a

, St

efano de P. Carraro

, Carlos R. G. Teixeira

2 b

, Leonardo D. Bandeira

3 c

Bernardo L. M

uller

, Roberto Tietzmann

2 d

, Milene S. Silveira

2 e

and Isabel H. Manssour

2 f

University Canada West, Vancouver, Canada

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

Federal University of Rio Grande do Sul (UFRGS), Porto Alegre, Brazil

{stefano.c, b.lorenzi}@edu.pucrs.br, {rtietz@pucrs.br, milene.silveira, isabel.manssour}@pucrs.br

Keywords:

YouTube, Social Media Analysis, Visualization, Streaming, Chat Interaction.

Abstract:

This paper presents StreamVis, an easy-to-use platform that provides stats and visual representations to ana-

lyze live chat data from YouTube. StreamVis uses Python and Google’s YouTube Data API for data gathering,

combined with libraries such as NLTK for natural language processing, Pandas for data analysis, and Mat-

plotlib for visualization. Its interactive dashboard facilitates real-time data visualization through frequency

charts, word clouds, and sentiment analysis, providing deep insights into audience engagement patterns. A

case study analyzing the NFL’s ﬁrst game in Brazil broadcast on Caz

e TV demonstrates how StreamVis reveals

trends in audience interactions during critical moments, like game highlights and performances. StreamVis

is different from previous tools because it has a user-friendly interface, enabling non-technical users (such as

journalists and other media professionals) to perform complex data analysis with a large volume of content,

helping them to understand how live chat dynamics inﬂuence media consumption.

1 INTRODUCTION

Television producers have always used various meth-

ods to measure audience response and guide produc-

tion efforts (Barnouw, 1990), (Sterling and Kittross,

2001), (Bourdon and M

eadel, 2014). The most tradi-

tional, inherited from the radio, included letters and

phone calls to the broadcaster regarding aired pro-

grams. This was followed by audience measurement

techniques, which became more automated over the

years. In parallel, qualitative studies have also been

regularly used to assess audience appreciation beyond

mere quantitative measurements (Marc, 1996). More

recently, online data collection has played a crucial

role in shaping strategies and narratives, providing

immediate and detailed insights into audience behav-

ior. With the restrictions on data collection on X (for-

https://orcid.org/0000-0002-8730-3383

https://orcid.org/0000-0001-6829-1682

https://orcid.org/0009-0006-0043-912X

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

https://orcid.org/0000-0003-2159-551X

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

merly Twitter) after the acquisition by Elon Musk, the

attention ﬂowed to other platforms.

Building on the growing popularity of live broad-

casts on YouTube (YT) and similar platforms, the

analysis of audience comments becomes increasingly

signiﬁcant. These comments not only reﬂect the

viewers’ immediate reactions but also reveal broader

social dynamics (Burgess and Green, 2018), such as

public engagement, sentiment, and community build-

ing around the content being shared.

The massive engagement found on such platforms

has shaped sports institutions and promoters such as

the International Olympic Committee (IOC) and the

National Football League (NFL) to adopt an inno-

vative broadcasting strategy. During the Paris 2024

Olympic Games, Brazilian audiences had multiple

viewing options: they could watch the broadcasts on

TV Globo, the ofﬁcial broadcasting partner of the

event, on both free-to-air and pay TV, through the

IOC’s own channels on its applications, or via the

streaming partner and YT channel of Caz

e TV

, also

available for free. On September 6, 2024, the ﬁrst

https://www.youtube.com/cazetv

630

Kurtz, G. B., Carraro, S. P., Teixeira, C. R. G., Bandeira, L. D., Müller, B. L., Tietzmann, R., Silveira, M. S. and Manssour, I. H.

StreamVis: An Analysis Platform for YouTube Live Chat Audience Interaction, Trends and Controversial Topics.

DOI: 10.5220/0013446200003929

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 2, pages 630-640

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

NFL game played in Brazil showcased the Green Bay

Packers facing the Philadelphia Eagles on a Friday

night, featuring a similar media arrangement. This

multi-channel approach underscored the growing sig-

niﬁcance of digital platforms in sports broadcasting,

reﬂecting a shift towards more ﬂexible and audience-

driven viewing experiences.

The main goal of this work is to provide a deeper

understanding of how interactive platforms inﬂuence

audience engagement and media consumption by ana-

lyzing user-generated comments. We explore how the

comment streams accompanying live broadcasts, par-

ticularly in informal and interactive YouTube chan-

nels like Caz

e TV, contribute to the viewer experi-

ence. By analyzing the content, frequency, and senti-

ment of these comments, we seek to uncover patterns

that highlight the role of live-streaming platforms in

shaping contemporary media consumption. An in-

terdisciplinary team featuring Computer Science and

Communication researchers deﬁned the scope, tool

feature set, and prototype case study (NFL in Brazil).

To achieve this goal, the authors developed a plat-

form called StreamVis that collects comments accom-

panying live broadcasts on YouTube and processes the

collected data, generating a dashboard with frequency

charts, word clouds, sentiment analysis, and other vi-

sual analyses described further in this study. These

ﬁndings are presented as a report, allowing users to

capture and interpret the dynamics of audience inter-

actions in a short time, offering valuable insights into

how viewers engage with live content.

Our main contributions are outlined as follows:

• An innovative platform for collecting and analyz-

ing live broadcast comments on YouTube, which

enables interactive visualizations of audience en-

gagement through frequency charts, word clouds,

and sentiment analysis.

• A step-by-step approach that combines data col-

lection and visual analysis, providing new insights

into audience behavior during live events on inter-

active platforms.

• A case study demonstrating the application of the

proposed platform in analyzing comments from

Caz

e TV’s NFL in Brazil broadcast, illustrating

how informal and interactive broadcasting for-

mats impact viewer experiences.

The remainder of the paper is organized as fol-

lows. Sections 2 and 3 present the background and

related work. The proposed approach is described in

Section 4. Section 5 details the obtained results by

presenting one use case. A discussion on contribu-

tions, limitations, and future directions is presented in

Section 6. The last section outlines our conclusions.

2 BACKGROUND

This section presents an introduction related to

YouTube Lives and the Caz

e TV channel in the con-

text of sports broadcasts.

2.1 YouTube Lives

YouTube, owned by Alphabet, is one of the largest so-

cial media and streaming platforms in the world. Ac-

cording to the Digital 2024 Global Overview Report,

YouTube is the second-ranked social media channel

when it comes to time spent on platforms, with an av-

erage of slightly over 28 hours per month per user-

it only loses to TikTok, with 34 hours

. As of July

2024, India had the largest YouTube audience in the

world, with roughly 476 million users actively engag-

ing with the video-sharing platform, a little over one-

third of its population. Both the United States, ranked

second, and Brazil, ranked third, had over two-thirds

of their populations watching YouTube. The US had

approximately 238 million viewers, while Brazil had

about 147 million people using YouTube

Live streaming, once a niche technology used by

early internet pioneers, has evolved into one of the

most popular and inﬂuential forms of broadcasting.

Initially introduced in the 1990s, live streaming has

since become a cornerstone of digital communica-

tion, utilized by tech giants such as Google, Mi-

crosoft, and Apple. Its growth was further acceler-

ated by the global shift to remote work and study

during the COVID-19 pandemic. The platform that

truly revolutionized the accessibility and popularity

of live streaming, however, was YouTube. Beginning

with its ﬁrst live event in 2008, YouTube’s foray into

live streaming signaled a pivotal moment in the tech-

nology’s mainstream adoption, paving the way for a

broader cultural and technological shift that redeﬁned

media consumption

YouTube Live offers a variety of features that

make live streaming accessible and ﬂexible for cre-

ators. It allows users to stream directly from a web-

cam, mobile device, or through more advanced setups

with encoders. Creators can use real-time interactions

like polling and visual overlays to engage their audi-

ence, while options like monetization and scheduling

enhance the overall live experience. The live chat fea-

ture is a key component of YouTube Live, enabling

real-time communication between creators and view-

ers. Creators can moderate the chat through tools like

slow mode, where messages are spaced out to control

https://tinyurl.com/ydvmec8t

https://tinyurl.com/39bshxj7

https://tinyurl.com/2kv9w2we

StreamVis: An Analysis Platform for YouTube Live Chat Audience Interaction, Trends and Controversial Topics

631

the ﬂow, and can even ﬁlter or block inappropriate

comments. This interaction fosters stronger commu-

nity engagement, encouraging real-time feedback and

participation. YouTube Live also provides creators

with detailed insights about their streams. Metrics

such as concurrent viewers, chat activity, and over-

all audience demographics help creators evaluate their

content’s performance. These analytics allow stream-

ers to better understand their audience, reﬁne their

content, and optimize future live broadcasts, making

YouTube Live not only a platform for real-time inter-

action but also a tool for long-term content strategy

development

2.2 Caz

e TV and Sports Streaming

The media strategy used by the league deserves atten-

tion: the event was broadcasted across three different

platforms (RedeTV on free-to-air TV, ESPN Brazil

on cable TV, and Caz

eTV via online streaming), and

scheduled for the NFL’s ﬁrst Friday night primetime

slot. This approach contributed to the success of the

event, not just in the host state but throughout Brazil

Casimiro Miguel, popularly known as Caz

e, is a

reference in sports broadcasting on YouTube, with

over 16 million subscribers to his channel, more

than 7,000 videos published, and over 1.8 billion

views. The streamer has gained worldwide recog-

nition for his coverage of major events such as the

World Cup and the Olympics, ranking 300th glob-

ally in terms of subscribers

. During the 2022 World

Cup, Casimiro made history by becoming the ﬁrst

Brazilian streamer to broadcast tournament matches

on YouTube. Thanks to an agreement with FIFA and

LiveMode, he streamed 22 matches, including those

of the Brazilian national team, the semiﬁnals, and the

ﬁnal. His channel, Caz

eTV, peaked at 6 million si-

multaneous viewers, setting a live viewership record

on YouTube, ranking among the top 4 of the 5 most-

watched live streams

This success marked a turning point for sports

streaming, demonstrating that digital platforms could

compete with traditional television broadcasters

Another important milestone was his role in

broadcasting the ﬁrst NFL game in Brazil. The event

offered Casimiro an opportunity to expand his reach

beyond soccer. Although American football does

not have the same penetration as traditional soccer in

https://tinyurl.com/2nwj4eb8

https://tinyurl.com/mwznhf66

https://tinyurl.com/p6zm5wcy

https://tinyurl.com/yau4h6vs

https://tinyurl.com/djh43p7t

Brazil, the broadcast lasted over 7 hours and garnered

nearly 3 million views, as shown in Figure 1

Figure 1: Total views of the NFL broadcast. Source:

https://playboard.co/en/video/W049t2oII4Q.

3 RELATED WORK

Research about YouTube has been quite abundant,

since it is a well-established platform. In our work,

we aim to expand on those ﬁndings, acknowledg-

ing the previous works. When looking for those

works, we ﬁltered investigations that were related to

YouTube comments, both in regular uploaded videos

and Lives. We noticed a lower volume of comment

analysis in YouTube livestreams, suggesting room for

further analysis.

When it comes to sentiment analysis on YouTube,

recent studies have made progress in using machine

learning and deep learning to understand user opin-

ions from the platform’s large amount of comments.

For instance, Chalkias et al.(Chalkias et al., 2023)

used tools like VADER and TextBlob to study com-

ments on educational YouTube videos, ﬁnding that

most comments were neutral, but many were posi-

tive when videos used engaging features like anima-

tions. Mehta and Deshmukh(Mehta and Deshmukh,

2022) focused on YouTube ads, using machine learn-

ing models like Decision Trees, which showed strong

results in predicting how viewers would react. These

studies highlight how automated systems are improv-

ing in handling large datasets and identifying patterns

in user sentiment across different types of YouTube

content.

Earlier research laid the groundwork for these ad-

vancements. Singh and Tiwari (Singh and Tiwari,

2021) used several machine learning models, ﬁnd-

ing that Support Vector Machine (SVM) worked best

for analyzing comment sentiment. Sentiment analysis

is also a topic for multi-language adaptations other

than English. Yasin et al. (Yasin et al., 2020) tack-

led the challenge of bilingual sentiment analysis in

English and Roman Urdu, successfully using logistic

regression to rank videos based on comments. This

is just one example of this analysis being done in

multiple languages. Uryupina et al.(Uryupina et al.,

https://tinyurl.com/3hcejrm9

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

632

2014) introduced the SenTube corpus, a dataset de-

signed for sentiment analysis in multiple languages.

Together, these studies show how sentiment analysis

on YouTube has grown from basic dataset creation to

more advanced models that can handle a wider range

of content and languages.

Sentiment analysis has been widely used by re-

searchers on YouTube, but there were also other ap-

proaches regarding user participation in the com-

ments. It is important to note that the following re-

searches were done in pre-recorded posted videos on

YouTube, not live streams. Recent studies have ex-

plored various aspects of YouTube comments, from

informal learning to public discourse and content

moderation. Dubovi and Tabak (Dubovi and Tabak,

2020) focused on how YouTube comments foster

knowledge co-construction, particularly in science

videos. Their research demonstrated that user com-

ments often go beyond simple information sharing,

engaging in discussions and debates that lead to

deeper learning.

Other researchers have examined the broader dy-

namics of YouTube comments, particularly their im-

pact on engagement and content quality. Siersdorfer

et al. (Siersdorfer et al., 2010) analyzed over six mil-

lion comments to understand the factors that inﬂu-

ence comment ratings and usefulness. They found

a clear correlation between positive sentiment and

higher community ratings, while offensive language

led to negative ratings. Their work also introduced

machine learning classiﬁers to predict which com-

ments would be rated positively or negatively by the

community, providing insights into how sentiment

and language shape user interactions on the platform.

Comment analysis on YouTube has also been ap-

plied to speciﬁc issues such as hate speech and spam.

Latorre and Amores (Latorre and Amores, 2021)

used topic modeling to identify racist and xenopho-

bic comments targeting migrants and refugees, show-

ing that 19.35 percent of the analyzed comments con-

tained hate speech. This highlights the darker side of

YouTube comments, where far-right rhetoric and na-

tionalist views are prevalent in certain channels. Ab-

dullah et al.(Abdullah et al., 2018) took a different ap-

proach, focusing on spam detection in YouTube com-

ments. They compared various spam ﬁltering tech-

niques, ﬁnding that low-complexity algorithms can

achieve high accuracy in identifying spam content,

suggesting that YouTube’s built-in tools could be en-

hanced to combat the spread of malicious content.

These studies show that while YouTube comments of-

fer valuable insights into user sentiment and interac-

tion, they also present challenges related to content

moderation and the spread of harmful speech.

More speciﬁcally about YouTube comments anal-

ysis in live streams, we have some relevant ﬁndings.

Recent research focuses on understanding user behav-

ior, emotional intensity, and content moderation dur-

ing live events. Sentiment analysis was one of the

tools used to analyze live stream comments. Tirpude

et al. (Tirpude et al., 2024) developed a system to

analyze sentiments in live chat through Natural Lan-

guage Processing (NLP) techniques. By using Fast-

Text for sentiment scoring and emoji analysis, they

provided real-time insights into audience reactions,

enabling content creators to adjust their approach dy-

namically during live streams. Similarly, Liebeskind

et al. (Liebeskind et al., 2021) investigated engage-

ment patterns in YouTube live chats, speciﬁcally dur-

ing political events such as Donald Trump’s 2020

campaign. Their study revealed that live comments

were highly emotional, with a signiﬁcant portion be-

ing abusive, but frequent commenters tended to use

less offensive language, emphasizing how emotional

involvement plays a role in live chat dynamics.

Emotional intensity in live streams has been an-

other focus, as evidenced by the works of Luo et

al. (Luo et al., 2020) and Guo and Fussell (Guo and

Fussell, 2020). Luo et al. explored how emotions are

ampliﬁed in live chat compared to comments posted

after events, ﬁnding that shared real-time experiences

intensify both positive and negative sentiments. This

emotional ampliﬁcation has implications for content

moderation, as heightened emotions can lead to an in-

crease in abusive or toxic comments. Guo and Fussell

took this further by examining emotional contagion in

live-streaming environments, showing that the senti-

ments in chat messages have a stronger inﬂuence on

subsequent messages than the content of the video it-

self. Their ﬁndings suggest that audience interactions

can signiﬁcantly shape the overall sentiment of live

chat, often more than the live content itself.

In addition to understanding emotional dynam-

ics, other researchers have focused on combat-

ing challenges such as spam and toxicity in live

chats. Yousukkee and Wisitpongphan (Yousukkee

and Wisitpongphan, 2021) analyzed spammers’ be-

havior in live streams, using machine learning models

to differentiate between spam and legitimate user en-

gagement. Their decision tree classiﬁer achieved high

accuracy in detecting repetitive, irrelevant content.

Complementing this, Tarafder (Tarafder et al., 2023)

developed an automated tool to identify and ﬂag

toxic comments in real-time, addressing the increas-

ing need for content moderation during live streams,

especially as platforms like YouTube have seen a

surge in usage during the pandemic.

An interesting recent tool that is worth mentioning

StreamVis: An Analysis Platform for YouTube Live Chat Audience Interaction, Trends and Controversial Topics

633

is the “CatchLive”, developed by Yang et al. (Yang

et al., 2022). The system provides a real-time sum-

mary by segmenting the stream into high-level sec-

tions and highlighting key moments based on both

stream content and user interaction data, such as chat

messages and likes. The paper discusses the devel-

opment of two core algorithms—one for stream seg-

mentation and another for identifying highlight mo-

ments. Even though this tool was not developed for

analyzing comments on YouTube Livestreams, it was

an interesting inspiration for our team when develop-

ing our solution, since the main goal is to provide a

comprehensive summarization of user comments and

engagement in Live Streamings.

4 StreamVis PLATFORM

To facilitate the analysis of YT live chat, we devel-

oped an online platform for gathering and analyzing

data through an interactive dashboard. The dashboard

enables temporal analysis of chat interactions, allow-

ing users to identify inﬂuential viewers, trending top-

ics, and controversial discussions. Its easy-to-use in-

terface allows users to gain insights from live chat

data without the need for technical expertise or pro-

gramming skills.

Figure 2 presents the main components of

StreamVis. The following sections describe these

components and the platform’s main functionalities.

Figure 2: StreamVis main components.

4.1 Implementation Environment

We used the Python programming language and

Google’s ofﬁcial YouTube Data API for data gather-

ing. This API returns a payload with relevant data,

such as comment message, author, and timestamp.

The decision to use Python as the primary program-

ming language for this project was based on its library

support and performance in data analysis and natural

language processing.

In addition to the Python language, we also used

Streamlit

to develop the dashboard interface.

In Table 1, we outline the technologies utilized

and their essential roles in the project. Integrating

these libraries with Python provided a robust frame-

work for conducting large-scale analysis of YouTube

comments, enabling seamless interaction with APIs,

efﬁcient data processing, natural language analysis,

and the creation of insightful visualizations.

Table 1: Libraries used and their respective functions.

Library Description

requests Used to retrieve YouTube com-

ments in real-time by interacting

with the YouTube Data API, en-

abling dynamic data collection di-

rectly from the platform.

python-dateutil Used for parsing and manipulat-

ing dates, ensuring consistent pro-

cessing and visualization by nor-

malizing date and time data.

pytz Used for managing time zones

and converting timestamps to a

uniform time reference, facilitat-

ing accurate time-based analysis

across different geographic loca-

tions.

python-dotenv Used for securely managing envi-

ronment variables and safeguard-

ing sensitive information in a

.env ﬁle.

nltk Used for preprocessing com-

ments, particularly for remov-

ing common stopwords in Por-

tuguese, ensuring that only mean-

ingful words remained in text-

based analyses.

pandas Used for organizing comments

into DataFrames and conducting

detailed analysis, such as identi-

fying peaks in commenting activ-

ity.

WordCloud Used for generating visual repre-

sentations of the most frequently

occurring words in the comments,

providing intuitive insights into

prominent themes and topics.

matplotlib Used for generating charts and

plots, aiding in the interpretation

of resampled data and identiﬁca-

tion of trends in user behavior.

4.2 Data Gathering

As a tool for data gathering, we developed a small

and reusable script in Python that uses the YouTube

Data API to make requests. This script inter-

acts with YouTube’s API to retrieve a payload con-

https://streamlit.io/

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

634

taining relevant data, such as the comment mes-

sage, author details, and timestamp. Key func-

tions within the script, get live details() and

get chat messages(), respectively, are responsible

for retrieving live stream details and the live chat ID,

and gathering chat messages, calculating elapsed time

from the stream’s start, and processing each comment,

author, and time data into a structured format.

The script continuously runs in a loop, collecting

data at speciﬁed intervals and automatically saving

the gathered data into a JSON ﬁle. This data stor-

age ensures persistence and can be later used as in-

put to the developed dashboard, enabling further vi-

sualizations like frequency charts, word clouds, and

sentiment analysis. Additional functions manage the

reading and writing of data and ﬁlter out duplicate en-

tries, ensuring only new, relevant comments are added

to the dataset.

This script’s main beneﬁt is its ease of use

users only need to provide their Google API Key

and the YouTube video ID, which can be easily ob-

tained from the video URL. This simplicity allows

even non-technical users to deploy the script effort-

lessly. Furthermore, the script’s automated real-

time data collection with subsequent visual analysis

through the dashboard described in the next section

enhances the understanding of audience behavior dur-

ing live YouTube events. The proposed platform is

particularly advantageous for informal and interactive

broadcasting formats, as it captures the dynamic ﬂow

of viewer responses, which traditional data-gathering

methods might overlook.

4.3 Dashboard Description

StreamVis is designed to provide insights into user en-

gagement through YouTube comments and live chat

data. Its dashboard allows users to examine critical

metrics, such as chat volume, trending topics, and in-

ﬂuential users.

The dashboard offers a simple interface, incorpo-

rating line and bar charts and word clouds. It en-

ables users to ﬁlter data, zoom, and access detailed

insights, allowing for an in-depth exploration of au-

dience behavior without requiring programming ex-

pertise. Additionally, customizable views enhance

the tool’s adaptability to diverse use cases, support-

ing various analytical needs.

A sidebar menu allows users to navigate between

various features of the dashboard. The available op-

tions included are shown in Figure 3 and explained in

detail below.

The script is available at

https://github.com/DAVINTLAB/StreamVis

Figure 3: StreamVis main page presenting menu options.

• Upload JSON - Allows you to load a JSON ﬁle

with the data collected by the script described in

Section 4.2.

• Comment Peaks - One of the primary features of

StreamVis is the ability to identify and analyze

peaks in comment activity. The Comment Peaks

page exempliﬁed in Figure 4 generates an inter-

active time-series chart that visualizes moments of

high comment volume. These peaks often corre-

spond to key events in the video, such as signif-

icant announcements or moments of heightened

viewer interest. This chart allows users to zoom

in on speciﬁc time intervals and explore trends in-

side the collected data. In addition to visualizing

the peaks, StreamVis generates a corresponding

word cloud for each peak based on the comments

made during that time frame.

• Top Commenters - Highlights the most active par-

ticipants and offers a detailed breakdown of all

their comments, including timestamps and con-

tent, allowing analysts to explore the volume of

engagement and the nature of the contributions

made by key users. It is also possible to conﬁg-

ure the number to the top commenters you want

to analyze, as shown in Figure 5.

• Partitions - For further analysis, allow the user to

divide the dataset into custom time intervals for

further analysis. Sliders allow users to adjust the

granularity of the data displayed, such as resam-

pling comment data by minutes or hours for more

reﬁned insights.

• Stats - Provides general statistics about the

dataset, such as the total number of comments,

unique commenters, and average comments per

user.

• New Members - As illustrated in Figure 6, dis-

plays new users who have recently joined a paid

membership plan, helping to sponsor the stream-

ing channel and its projects. This function shows

the timing and volume of memberships through-

out the live stream.

StreamVis: An Analysis Platform for YouTube Live Chat Audience Interaction, Trends and Controversial Topics

635

Figure 4: StreamVis comments peak page.

Figure 5: StreamVis top commenters page.

StreamVis provides an easy-to-use platform for

analyzing YouTube comments and live chat data.

Through its interactive components, including time-

series charts, word clouds, and user activity tracking,

the tool facilitates in-depth exploration of audience

engagement.

The ﬂexibility afforded by customizable time in-

tervals, ﬁltering capabilities, and visual representa-

tions allows the platform to adapt to diverse analytical

contexts. Whether analyzing overall chat dynamics to

understand general audience engagement, zooming in

on key moments of signiﬁcance, or identifying stand-

out trends and commenters, StreamVis provides the

tools necessary to explore the data from various per-

spectives.

Figure 6: StreamVis new members page.

5 CASE STUDY

In this section, we present a case study to exemplify

the use and potential of the developed approach.

5.1 NFL and Its First Game in Brazil

The NFL has been heavily investing in its internation-

alization strategy since 2007, when it held its ﬁrst

game outside the United States, between the New

York Giants and the Miami Dolphins at Wembley Sta-

dium in London. This event marked the beginning

of the NFL International Games, which have since

expanded, with London becoming the main hub for

the league’s international games. Between 2007 and

2023, more than 30 regular season games were played

in the English capital, at venues such as Wembley and

Tottenham Hotspur Stadium

The peak of this expansion reached Brazil in 2024,

with the ﬁrst NFL game on Brazilian soil, held on

September 6th at Neo Qu

ımica Arena in S

ao Paulo,

the stadium of the Corinthians football team. The

event attracted more than 47,000 people to the sta-

dium and was a milestone for the NFL’s presence in

South America, breaking audience records both in the

American and Brazilian broadcasts

5.2 Game Analysis Through the

Proposed Platform

The Eagles and Packers game started at 9:15 PM

Brasilia time (Brazil); however, the ofﬁcial live broad-

cast on Caz

e TV’s YouTube channel began two hours

https://tinyurl.com/hj6622mk

https://tinyurl.com/ny6y2rr4

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

636

earlier. The analysis presented below refers exclu-

sively to the game period, from the kickoff (9:15

PM) until its conclusion (past midnight). In total, the

match lasted approximately 3 hours and 50 minutes,

during which data collection occurred, resulting in

28,524 comments during the live stream, representing

an average of 2 comments per second. This ﬁgure is

noteworthy, as a higher number of posts was initially

expected. The authors feared a bottleneck in data col-

lection due to limitations of the YouTube API, which

did not occur, making it possible to capture nearly all

the comments made.

A total of 9,500 users made comments during the

live stream, with an average of 3 comments per user.

This number can also be considered low, as the num-

ber of views approached 3 million, with comments

representing less than 0.5 of the views. The most

active user posted 91 comments, averaging approx-

imately 17 comments during the 30-minute periods

in which they were most active. The 10 most active

users averaged around 80 comments, showing varied

frequency patterns throughout the game, as illustrated

in Figure 7.

Figure 7: Comments throughout the live stream by the top

10 users.

Figure 7 shows the distribution of all comments

over the course of the game, allowing for a deeper

analysis of user behavior. What one can observe in

the ten charts shown in Figure 7 is that there are dis-

tinct behaviors among users. Primarily, users’ posting

behavior does not follow a consistent curve, as one

might expect from a professional commentator or a

game narrator. Driven by fan support for the team or

attention to speciﬁc plays, the curves suggest different

motivations for engagement, which becomes evident

when examining individual messages more closely.

Considering the overall numbers of the comments,

there were 135,498 words with 23,172 unique words,

and an average of 4.75 words per comment. When all

comments made during the game broadcast are ag-

gregated into a line chart based on the comment fre-

quency one can observe an increasing interest toward

the middle of the match and a decreasing interest in

the second half, represented in Figure 8.

Figure 8: Comments throughout the live stream.

The three highest peaks of interest in terms of

comments are clustered in the ﬁrst half. Typically,

fan enthusiasm and key plays prompt increased com-

menting. Our platform can analyze each peak in de-

tail using segmented word clouds. The ﬁrst peak,

comprising 1,655 comments (5.8% of the total), oc-

curred during one of the initial commercial breaks

(Figure 9), highlighting a common second-screen be-

havior where viewers shift their attention to chat and

comments during pauses in the main broadcast. This

break followed a ﬁeld goal by the Green Bay Pack-

ers, a moment that sparked a trend seen throughout

the broadcast: references and comparisons between

American football (NFL) and soccer (the dominant

sport in Brazil and the location of the event, held at

the Corinthians football club stadium). In this con-

text, the recurrence of the word ‘Goi

as’ stands out, re-

ferring to Goi

as Futebol Clube, which shares a similar

visual identity with the Green Bay Packers. Viewers

played on this similarity with comments like ‘VAMO

GOI

AS, 3X0 J

A’ (LET’S GO GOI

AS, 3-0 NOW),

‘Goi

as 3 X 0 Corinthians?’ and ‘CARACA KKK

GOI

AS T

A JOOOGANDO!!!!’ (WOW LOL GOI

IS PLAYIIING!!!!). Moreover, much of the discus-

sion centered on ‘Anitta,’ the Brazilian singer per-

forming at the halftime show, with comments such

as ‘we want Anitta,’ ‘When is Anitta’s show???,’ and

‘Come on, Anitta.’

The second peak also occurred during the break

following another score, with 2,349 comments

(8.2%), already indicating a higher frequency of

viewers as the main break and the show approached.

At this point, mentions of the singer ‘Anitta’ were al-

ready considerably higher than any other topic. Dur-

ing the third peak, Anitta’s show ﬁnally took place,

directly mentioned in the comments, with 2,877 com-

ments (10% from the total). The comments reveals

a high volume of criticism directed at the show,

with words like ‘playback’ and ‘shame’ standing

out. Throughout Anitta’s performance, the comments

StreamVis: An Analysis Platform for YouTube Live Chat Audience Interaction, Trends and Controversial Topics

637

Figure 9: Wordcloud of peak one comments.

were divided between criticism of her use of playback

and her decision to sing only her international songs

rather than her native Portuguese repertoire of songs.

It is also worth noting the word ‘better,’ which, when

analyzed in isolation, might seem positive, but upon

reviewing the comments, its tone aligned with the crit-

ical sentiment of the audience: ‘wasn’t there a bet-

ter singer???,’ ‘the game deserved something better,’

‘so many better and quality artists to choose from,’

among others. Few comments used ‘better’ in a com-

plimentary manner. The presence of emojis in the

comments was also noteworthy, with a signiﬁcant vol-

ume of expressions representing shame and criticism.

‘shame’ / Fuchsia Poop / poop shape (poop emoji?) /

face fuchsia (shame?)”

The fourth peak occurred during a rare game mo-

ment: a ﬁeld goal hitting the post, which allows for

a direct comparison to soccer. This peak had 964

comments (3.3%), with most of them once again fo-

cusing on the game, consistently making references

to soccer. The pitch and its quality were frequently

mentioned. Athlete LeBron James criticized the pitch

on his social media, which reignited discussions in

the live chat about its quality and concerns that the

sport might be ‘ruining’ the stadium’s pitch. Some

of the comments included: ‘LeBron criticized the

pitch, oh boy,’ ‘Is it just me, or is the pitch deterio-

rating?’, ‘EVEN LEBRON COMPLAINED ABOUT

THE PITCH,’ ‘COME ON CORINTHIANS, WHAT

ARE THEY DOING TO OUR PITCH, IT’S A DIS-

GRACE.’”

Finally, peak 5 had 1,247 comments (4.3%),

around the time of an error in the ofﬁcial broadcast

that incorrectly announced a touchdown and mistak-

enly changed the score, a fact that was highlighted by

the commentators and caught viewers’ attention. In

spite of that, this incident itself did not stand out in

the comments, once more focused on the connection

to soccer. The presence of the words ‘Corinthians,’

‘Tim

ao’ (a popular nickname for the team in Brazil),

and ‘Vai Corinthians’ (a characteristic chant among

the fans) versus ‘Goi

as’ brought soccer back into the

spotlight. Another analysis enabled by the tool re-

lates to the channel’s members. The channel has a

membership area that offers beneﬁts in exchange for

a monthly subscription. These beneﬁts range from a

badge in the chat to the opportunity to make more fre-

quent comments. Regarding memberships, there were

57 new members during the game. It is possible to

observe a trend that the beginning of the games has

a higher potential for attracting new members, with

this number decreasing signiﬁcantly as the match pro-

gresses.

6 DISCUSSION

While many studies, such as those by Mehta and

Deshmukh (Mehta and Deshmukh, 2022), focused

on pre-recorded videos and YouTube ads, this

work speciﬁcally targets live streaming environments,

which present unique challenges and opportunities in

real-time data gathering and its subsequent analysis.

The idea of collecting and analyzing live com-

ments and chats during YT broadcasts has proven

rich in possibilities and relatively simple to imple-

ment. Firstly, StreamVis does not automatically pro-

vide a way to receive live video or audio feeds, as

typical in TV broadcasts or platforms like Stream-

Yard. Thus, the primary feedback channel chosen by

YouTube is text, which includes using slang, hash-

tags, and emojis, adding different inﬂections to the in-

teraction. Previous research on comment analysis has

mainly addressed content moderation and spam de-

tection in pre-recorded content, as seen in the works

of Latorre and Amores (Latorre and Amores, 2021)

or Abdullah et al. (Abdullah et al., 2018). In con-

trast, our approach focuses on providing insights into

audience engagement and interaction dynamics dur-

ing live streams, making it more relevant to dynamic

broadcasting scenarios.

Besides, we are not directly analyzing video or au-

dio content, which would require signiﬁcantly more

computational power. This allows our platform to be

used on various machines without needing high-end

hardware. This accessibility is advantageous, as it

opens opportunities for analysts, consultants, and pro-

fessionals interested in understanding and evaluating

engagement with YT live content. Our platform im-

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

638

plementation emphasizes visual analytics using, e.g.,

frequency charts and word clouds, allowing users to

interpret trends in live chat data interactively. This

visual approach sets this tool apart from traditional

text-based sentiment analysis methods used in earlier

studies.

The ability to perform a visual analysis of the

data allows the StreamVis users to develop a macro

understanding of the dynamics during the broadcast

and audience engagement, and a granular view that

enables them to examine conversations in detail and

identify situations or moments that acted as turning

points. This dual perspective helps users capture

broader trends while also pinpointing speciﬁc inter-

actions or events that signiﬁcantly inﬂuenced the ﬂow

of the live broadcast. When aligned with an under-

standing of what was happening in the broadcast dur-

ing those speciﬁc minutes, the visual nature of peaks

allows us to build an insight into how people collec-

tively react to what is being shown. We recognize that

different events can elicit varying behaviors, and even

the behavior of the most active users is not necessarily

convergent.

The choice of an NFL game as the proof-of-

concept of our platform was intentional for several

reasons. First, the novelty of the sporting event guar-

anteed attention and was likely to drive engagement.

Additionally, the growing interest in American foot-

ball in Brazil is a proxy to the richer strata of the popu-

lation, one with widely available internet services and

devices, consequently abundant participation on YT

and other platforms.

7 FINAL REMARKS

The work done in this study contributes to the existing

area research by introducing a platform speciﬁcally

designed for YouTube live chat analysis, which allows

real-time data gathering, visualization of user engage-

ment, trending topics, and emotional intensity during

live events. We provide all the codes of our data gath-

ering and visual analytics approach at the GitHub

and the StreamVis app is available online

for any-

one who wants to use it. Thus, it is possible to gather

and analyze data from different contexts.

As future possibilities for this research, we be-

lieve that integrating other natural language process-

ing models and AI infrastructures will enable a more

detailed sentiment analysis, potentially allowing for

seamless operation in multiple languages. This would

https://github.com/DAVINTLAB/StreamVis

https://davint-live-comments.streamlit.app

make it possible to observe the reactions of inter-

national audiences to the same event that attracts

viewers from various countries, such as the Olympic

Games, the UEFA Champions League, or other simi-

lar sporting events.

ACKNOWLEDGEMENTS

Carraro and M

uller would like to thank the Tutorial

Education Program (PET). Manssour would like to

thank the ﬁnancial support of the CNPq Scholarship -

Brazil (303208/2023-6).

While preparing and revising this manuscript, we

used ChatGPT, Grammarly, and Google Translate to

enhance clarity and grammatical precision, as English

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

sibility for the content and its technical accuracy.

REFERENCES

Abdullah, A. O., Ali, M. A., Karabatak, M., and Sengur, A.

(2018). A comparative analysis of common youtube

comment spam ﬁltering techniques. In 2018 6th Inter-

national Symposium on Digital Forensic and Security

(ISDFS), pages 1–5.

Barnouw, E. (1990). Tube of Plenty: The Evolution of

American Television. Oxford University Press.

Bourdon, J. and M

eadel, C. (2014). Television Audiences

Across the World: Deconstructing the Ratings Ma-

chine. Palgrave Macmillan UK.

Burgess, J. and Green, J. (2018). YouTube: Online Video

and Participatory Culture. Digital Media and Society.

Polity Press.

Chalkias, I., Tzaﬁlkou, K., Karapiperis, D., and Tjortjis, C.

(2023). Learning analytics on youtube educational

videos: Exploring sentiment analysis methods and

topic clustering. Electronics, 12(18).

Dubovi, I. and Tabak, I. (2020). An empirical analysis

of knowledge co-construction in youtube comments.

Computers & Education, 156:103939.

Guo, J. and Fussell, S. R. (2020). A preliminary study of

emotional contagion in live streaming. In Compan-

ion Publication of the 2020 Conference on Computer

Supported Cooperative Work and Social Computing,

CSCW ’20 Companion, page 263–268, New York,

NY, USA. Association for Computing Machinery.

Latorre, J. P. and Amores, J. J. (2021). Topic mod-

elling of racist and xenophobic youtube comments.

analyzing hate speech against migrants and refugees

spread through youtube in spanish. In Ninth Inter-

national Conference on Technological Ecosystems for

Enhancing Multiculturality (TEEM’21), TEEM’21,

page 456–460, New York, NY, USA. Association for

Computing Machinery.

StreamVis: An Analysis Platform for YouTube Live Chat Audience Interaction, Trends and Controversial Topics

639

Liebeskind, C., Liebeskind, S., and Yechezkely, S. (2021).

An analysis of interaction and engagement in youtube

live streaming chat. In 2021 IEEE SmartWorld,

Ubiquitous Intelligence & Computing, Advanced &

Trusted Computing, Scalable Computing & Commu-

nications, Internet of People and Smart City Inno-

vation (SmartWorld/SCALCOM/UIC/ATC/IOP/SCI),

pages 272–279.

Luo, M., Hsu, T. W., Park, J. S., and Hancock, J. T. (2020).

Emotional ampliﬁcation during live-streaming: Evi-

dence from comments during and after news events.

Proc. ACM Hum.-Comput. Interact., 4(CSCW1).

Marc, D. (1996). Demographic Vistas: Television in Amer-

ican Culture. University of Pennsylvania Press.

Mehta, T. and Deshmukh, G. (2022). Youtube ad view senti-

ment analysis using deep learning and machine learn-

ing. International Journal of Computer Applications,

184(11):10–14.

Siersdorfer, S., Chelaru, S., Nejdl, W., and San Pedro, J.

(2010). How useful are your comments? analyzing

and predicting youtube comments and comment rat-

ings. In Proceedings of the 19th International Confer-

ence on World Wide Web, WWW ’10, page 891–900,

New York, NY, USA. Association for Computing Ma-

chinery.

Singh, R. and Tiwari, A. (2021). Youtube comments sen-

timent analysis. International Journal of Scientiﬁc

Research in Engineering and Management (IJSREM),

5(5):1–11.

Sterling, C. and Kittross, J. M. (2001). Stay Tuned: A His-

tory of American Broadcasting. Routledge Communi-

cation Series. Taylor & Francis.

Tarafder, T., Vashisth, H. K., and Arora, M. (2023). Au-

tomated tool for toxic comments identiﬁcation on live

streaming youtube. In International Conference on

MAchine inTelligence for Research & Innovations,

pages 47–56. Springer.

Tirpude, S., Thakre, Y., Sudan, S., Agrawal, S., and

Ganorkar, A. (2024). Mining comments and senti-

ments in youtube live chat data. In 2024 4th In-

ternational Conference on Intelligent Technologies

(CONIT), pages 1–6.

Uryupina, O., Plank, B., Severyn, A., Rotondi, A., Mos-

chitti, A., et al. (2014). Sentube: A corpus for sen-

timent analysis on youtube social media. In LREC,

pages 4244–4249.

Yang, S., Yim, J., Kim, J., and Shin, H. V. (2022). Catchlive:

Real-time summarization of live streams with stream

content and interaction data. In Proceedings of the

2022 CHI Conference on Human Factors in Comput-

ing Systems, CHI ’22, New York, NY, USA. Associa-

tion for Computing Machinery.

Yasin, S., Ullah, K., Nawaz, S., Rizwan, M., and Aslam,

Z. (2020). Dual language sentiment analysis model

for youtube videos ranking based on machine learn-

ing techniques. Pakistan Journal of Engineering and

Technology, 3(2):213–218.

Yousukkee, S. and Wisitpongphan, N. (2021). Analy-

sis of spammers’ behavior on a live streaming chat.

IAES International Journal of Artiﬁcial Intelligence,

10(1):139.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

640