Sentiment Analysis of YouTube Comments Using Bidirectional

Encoder Representations from Transformers Neural Network Model

Pramila R. Gadyanavar

, Mahantesh Laddi

, Prashant S. Jadhav

, Vaishali S. Katti

Jambukeshwar Pujari

and Azeem Javed Jamadar

Department of CSBS, Kolhapur Institute of Technology's College of Engineering (Autonomous), Kolhapur, India

Department of CSE, Bharatesh Institute of Technology, Belagavi, India

Department of ME, Rajarambapu Institute of Technology, Rajaramnagar, Shivaji University, Kolhapur, India

Keywords: BERT (Bidirectional Encoder Representations), AMB (Adapted Multimodal BERT), TCN (Temporal

Convolutional Network), CNN (Convolutional Neural Network), NLP (Natural Language Processing).

Abstract: Everything in the today’s world based on Sentiment. Sentiments are the Feelings that is based on Socially,

Mentally, Economically, Psychologically Factors of the audience. Suppose you are multinational brand, and

you want to know more about your Consumers Sentiments by figuring out by Looking at the people comment

in your video at YouTube. It’s very hard to analyse comments line by line, word by word. Practically it’s not

possible at that stage, because dealing with n numbers of comments are not possible. To Overcome these

technical Situation, we are Introducing Our Sentiment Model that can Filtered the Audience Comments or

sentiments. Sentiment Analysis is a natural Languages Processing Technique that is use know about the

Sentiments in the text Mainly Positive Negative and Our model will Classify the YouTube comments

Outcomes into five different Labels 1: "Very Negative", 2: "Negative" ,3: "Neutral" ,4: "Positive", 5: "Very

Positive”. It will also give some Insights from the data using some Visualization Technique like Bar Graph,

Pie chart and Word Cloud.

1 INTRODUCTION

Everything in the today’s world based on Sentiment.

Sentiments are the Feelings that is based on Socially,

Mentally, Economically, Psychologically Factors of

the audience. Suppose you are multinational brand,

and you want to know more about your Consumers

Sentiments by figuring out by Looking at the people

comment in your video at YouTube. It’s very hard to

analyse comments line by line, word by word.

Practically it’s not possible at that stage, because

dealing with n numbers of comments are not possible.

Not only have you had to deal you also wanted to

know more about consumer's Sentiments.

https://orcid.org/0009-0009-0586-918X

https://orcid.org/0009-0002-9119-0657

https://orcid.org /0000-0002-3102-6460

https://orcid.org/0009-0004-4198-8171

https://orcid.org/0000-0003-3422-4515

https://orcid.org/0009-0001-1834-4667

This Problem is not only limited to Brands, when

it comes to YouTube Content creator they have the

huge impact of audience Sentiments. Many YouTube

Creators Makes videos on the basis based on audience

polarity or Sentiments, if they themselves did not

have the knowledge of Audience How they will boost

their channels and audience. Similarly, it applicable

on constraints.

Comments are included in local language with

multiple emoji or gifs, so classifying such comments

using normal technique of sentiment analysis is

efficient one. Because is sentiment analysis we have

work with “context-based model”. And all traditional

methods are aspect-based models so here in this paper

Gadyanavar, P. R., Laddi, M., Jadhav, P., Katti, V. S., Pujari, J. and Jamadar, A. J.

Sentiment Analysis of YouTube Comments Using Bidirectional Encoder Representations from Transformers Neural Network Model.

DOI: 10.5220/0013640000004664

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

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 3, pages 683-689

ISBN: 978-989-758-763-4

683

we used sentence BERT technique for efficient

classification of YouTube comments.

2 LITERATURE SURVEY

Text classification, named entity recognition (NER),

and sentiment analysis benefit from the ability to

assess contextual meanings in language using

bidirectional context. In contrast, the autoregressive

structure of the GPT model enables it to generate

more coherent content and achieve effective results in

tasks such as chatbots and creative writing. However,

applying these models to agglutinative and

morphologically complex languages, like Turkish,

presents significant challenges. (M. Salıcı, 2024)

An Aspect-based sentiment Analysis method

based on transformer-based deep learning models. It

uses the Hugging Face Transformers library, which

gives access to cutting-edge pre-trained models. For

sentiment categorization, the approach used the

BERT model, a potent transformer architecture. This

study adds to the expanding field of sentiment

analysis by offering a scalable and reliable

transformer-based learning method for ABSA. This

approach is a useful tool for applications in product

reviews, and customer feedback analysis. (G. P. M S,

2024)

RoBERT sentiment analysis generates negative,

positive and neutral sentiment in the analysed text.

These scores are described independently, it is very

difficult to complete understanding about the

opinions in the comments. To propose a more reliable

analysis of mixed emotions found in unstructured

data, a composite sentiment summarizer combines

positive, negative, and neutral scores into one. This

composite score gives a more accuracy and reliable

representation of the sentiment elaborated in

comments. (Z. -Y. Lai, 2023)

For multimodal tasks, the Adapted Multimodal

BERT (AMB) is a BERT-based model that integrates

adapter modules with intermediate fusion layers.

These fusion layers perform task-specific, layer-by-

layer integration of audio-visual data and textual

BERT representations. Meanwhile, the adapter

adjusts the pre trained language model to suit the

specific task at hand. This approach enables fast and

parameter-efficient training by keeping the

parameters of the pre trained language model frozen

during the adaptation process. Research shows that

this method yields effective models that are resilient

to input noise and can outperform their fine-tuned

counterparts. (O. S. Chlapanis, 2023)

The architecture comprising a temporal

convolutional network (TCN), a convolutional layer,

a bidirectional long short-term memory (BiLSTM),

and robustly optimized bidirectional encoder

representations from transformers pre-training

approach (RoBERTa) is proposed to address issues.

Dual branch feature coding network based on

RoBERTa (DBN-Ro) is the name of the suggested

architecture. The stitched vectors undergo

dimensionality compression via the convolutional

layer. (F. Wang, 2021)

To assess the performance of lexicon-based and

sentence-BERT sentiment analysis models used as

code-mixed, low-resource texts as input, it

summarizes the results of experiments. Some code-

mixed texts in Javanese and Bahasa Indonesia are

utilized as a sample of low-resource code-mixed

languages in this study. Google Machine Translation

is used first to translate the raw dataset into English.

The input text is translated into English and then

classified using a pre-trained Sentence-BERT model.

The dataset used in this study is divided into positive

and negative categories. The experimentation found

that the combined Google machine translator and

Sentence-BERT model achieved 83 % average

accuracy, 90 % average precision, 76 % average

recall, and 83 % average F1 Score. (C. Tho, 2021)

Classifying sentiment is a crucial step in figuring

out how individuals feel about a good, service, or

subject. Sentiment classification and numerous

models for natural language processing have been put

forth. But most of them have focused on categorizing

sentiment into two or three groups. The model tackles

the fine-grained sentiment categorization task using a

promising deep-learning model named BERT.

Without a complex design, experiments demonstrate

that the model performs better than other well-known

models for this task. (M. Munikar, 2019)

Sentiment classification used as an Indonesian

dataset, was explored with the problem utilizing a

two-step procedure: sentiment classification and

aspect detection. The bag-of-words vector, which is

handled by a fully connected layer, and the word

embedding vector, which is handled by a gated

recurrent unit (GRU), are two deep neural network

models with different input vectors and topologies for

aspect detection that are compared. They also contrast

two deep neural network methods for sentiment

classification. Word embedding, sentiment lexicon,

and POS tags are input vectors in the first method,

which has a bi-GRU-based architecture. In the

second, the word embedding vector is rescaled using

an aspect matrix as the input vector, and the topology

INCOFT 2025 - International Conference on Futuristic Technology

684

is based on a convolutional neural network (CNN).

(AIlmania, 2018)

Four sentiment categories of data will be of the

form positive, negative, neutral, and mixed

in text messages using sentiment analysis and a soci

al adaptive fuzzy similarity-

based categorization technique. It can also determine

which emotion categories are most prevalent in the

messages, such as happiness, excitement, rage,

sadness, anxiety, and satisfaction. Additionally, it is

integrated into a comprehensive social media analysis

system that can gather, filter, categorize, and analyse

text data from social media platforms and present a

dashboard of descriptive and predictive metrics for a

particular idea. The suggested approach has been

created and is prepared for user licensing. (Z. Wang,

2018)

3 METHDOLOGY

Using Visualization and sentiments reports any

Technical or non-technical person can have the

Audience's Point of View what they are conveying in

the comments and what they believe in your Videos.

If these Sentiment Reports are used properly and

analysed efficiently then it can be helping Brand and

Content creator to boost their YouTube channel.

Cause every Comment in the video have the specific

value to the data.

Types of data used, and purpose means why you

choose. For The Sentiment Report the Model is just

getting ' Video ID ' as the input then Model will

Automatically generate the Original Comments CSV

file and then it will Generate another CSV file that

contains Sentiments reports and Shows Visualization

Using Graph as a output. For the sentiments we

require the NLP model so that our Comments can be

analysed on the Dataset, so we used “Bert-base-

multilingual-uncased-sentiment " which is the variant

of BERT fine-tuned specifically for multilingual

sentiment analysis. It was developed by NLP Town,

this model can be used for six different languages

English, Dutch, German, French, Spanish, and

Italian. And it predicts the sentiment of the review as

a number of stars (between 1 and 5) which I have

described earlier.

3.1 Sentiment Analysis

Main Methods for sentiments start with the

Initializing Sentiment Pipelines, Basically Pipeline

are nothing but the single function or module contains

different processes of analysis, it is generally the

subpart of scikit-learn which is a free and open-source

machine learning library In our use case we require

only two passing Parameters first is method name that

is "Sentiment-Analysis" in our case and second we

require to pass model which we used " BERT-base-

multilingual-uncased-sentiment " that developed by

NLP town Community.

After initializing and passing the parameters in

pipelines function. The processor further

Will be working on the Sentiment Mapping. In

our case we have described 1 star to 5 star

Based on Sentiments Score.

"1 ": "1 (Very Negative)",

"2 ": "2 (Negative)",

"3 ": "3 (Neutral)",

"4 ": "4 (Positive)",

"5 ": "5 (Very Positive)".

Now the mapping process has been also finished.

Now we need to handle 'Truncate of tokens' Suppose

if the length of the comment is to large. Then to avoid

tokenisation limit we have to shorten the input texts

to the first 512 characters to prevent issues with the

model's limit on text length. This is the number of

product reviews used for fine-tuning this model:

Language Number of reviews

English 150k

Dutch 80k

German 137k

French 140k

Italian 72k

Spanish 50k

Using this multilingual sentiment, we can easily

conclude sentiments of any comments. This model

claims Accuracy of 67% the exact match for the

number of stars. And When Accuracy Off by 1 it

predicates 95% of a accuracy numbers were given by

the human reviews.

3.1.1 Sentence-BERT (Bidirectional

Encoder Representation from

transformers) technique:

A transformer-based machine learning model called

BERT (Bidirectional Encoder Representations from

Transformers) was created for problems involving

natural language processing (NLP). Google

researchers created it, and its capacity to pre-train on

large volumes of text and then fine-tune for tasks has

transformed natural language processing (NLP) and

resulted in notable advancements in tasks such as

sentiment analysis, linguistic inference, and question

answering. Following are the variants of BERT

Sentiment Analysis of YouTube Comments Using Bidirectional Encoder Representations from Transformers Neural Network Model

685

technique with varying numbers of layers (Encoders),

attention heads, and hidden units.

Table 1: Variants of BERT technique with varying

numbers of layers

Variants BER

Base

BER

Lar

BER

Tin

BER

Mini

BERT

Small

Encoders

(L)

12 24 2 4 4

Attention

Heads (A)

Hidden

(

)

Units

768

1024

128

256

512

Working of BERT will encode the sentences

with all possible keys in same sentence, so it is

having unique encoding code for all words. And

it also generates different code for same words

in two different sentences. In following

sentences, the word “python” means different.

First sentence python is snake, and second

sentence python is programming language.

Sentence 1: He got bit by python.

Sentence 2: Python is a programming

language.

By Sentence- BERT technique can generate two

different codes for two pythons, so it is context based

model. And for sentiment analysis it is efficient

model. In BERT it encodes each word to a vector it

can be called as word2vec. Each of the word from the

sentence carries unique encoded vector.

Figure. 1: Architecture of Sentence-BERT

Truncation can also refer to the process of

converting data into a new record with smaller field

lengths than the original. Truncates tokens that

exceed a specified character limit. This limit defaults

to 10 but can be customized using the length

parameter. In fact, Many NLP couldn't Support Large

Tokens in our Use Case BERT have a maximum

token (Word or characters) limit that they can access

in a single input. If the input text exceeds this limit, it

can cause errors or lead to incomplete processing. By

truncating, we ensure that The Model Can Process the

Text Smoothly Staying within the model's toset next

pageken limit avoids errors. Efficiency shortening

long texts improves processing speed without

significantly affecting sentiment accuracy.

Figure. 2: Architecture Diagram

3.2 Data Cleaning

Most Important Part of any Sentiment Analysis is

Data Cleaning Without cleaning the data, Model is

imperfect. There are multiple data cleaning

technique, but we have chosen the most appropriated

and simple Technique that fits in our model that is

"Regex". Regular Expressions, regex or regexp in

short is extremely and amazingly powerful in

searching and manipulating text strings, particularly

in pre-processing text. Using this python library, we

can easily check if a string contains the specified

search pattern. Python has a built-in package called

re, which can be used to work with Regular

Expressions. In our Use Case, YouTube comments

can contain URL’s, numbers, special characters,

punctuations and some white spaces which do not

play any role in our Analysis, so we need to remove

all unessential part of comments, So We used "re" to

manipulate pre-process Comments. Cleaning is

essential because it increase model's Accuracy. And

re is simplest version to get fitted.

3.3 Data Preparation

Data Preparation technique start with our first step in

Research where we are trying to fetch All Possible

YouTube Comments Using YouTube API. Using

API, we are fetching Top Level Comments that can

INCOFT 2025 - International Conference on Futuristic Technology

686

play an important our analysis. We are not taking

comments reply,

Comment deleted, Spam Comment because it

doesn't play an important role in sentiment analysis.

In top levels comments that are organized by the

YouTube we are fetching 'author' that is username of

the account, 'published at' date when it was originally

published, 'updated at' date when the comment was

edited, 'like count' no of likes counts, 'text' is the Main

comment. There is a limit of fetching a comment in

one request so in our case it was 100 so to overcome

this problem we used "Pagination" technique that can

fetch all comments in one request. First Just we have

to set “ maxResults=100 " next it will check and get

the result in one output. YouTube API, It stand for

Application Programming Interference that allows

developer to embed videos and offer other YouTube

functionalities on your code and YouTube

functionalities on your application. It functions as a

client-server model, where the API acts as the

intermediary between the client (your application)

and the YouTube servers. This contract defines how

the two communicate with each other using requests

and Google Itself provides an environment to develop

Such Applications. We just need to generate

"Developer key" and do specific tasks.

3.4 Statistical Method

When the process of sentiment analysis has finished,

we required some statistical tools to demonstrate our

progress and visualize the reports and conclude some

results using this visualization. For the visualization

of Sentiments Currently we have used "Matplotlib"

library of python which is used to create static,

animated and interactive visualization in any

applications. Using matplotlib we have drawn Bar

Graph and Pie chart. Bar Graph typically Shows

Sentiments Distribution of comments that shows how

many comments fall in each category positive,

negative or neutral.

Similarly, we have also draw pie chart that use to

Show the proportion of each sentiment as a

percentage. It exactly defines percentage of each

Sentiments. Here each sentiment is represented as a

slice of the pie. After Visualizing with Bar Graph and

Pie chart we have also used Word cloud which

represents the Most Frequent Word clouds.

4 RESULT AND DISCUSSION

YouTube API is fetched the comments, these

comments will be encoded and classified using

BERT.

Figure. 3: Original comment Dataset and Sentiment

Reports

Following visualizations will describe very

positive comments are above 55, positive comments

are 15 in number, neutral comments near about 30,

very negative comments are above 45, and negative

comments are near about 5 in dataset.

Figure. 4: Bar Graph

Each word's magnitude in a word cloud, which is a

visual representation of text data, reflects how

frequently or how important it is in the dataset. Larger

font sizes are used for words that occur more

frequently in the text, whereas smaller font sizes are

used for words that occur less frequently. To make it

easier to see which terms are most common, word

clouds are frequently used to summarize, analyse, or

illustrate textual data.

Sentiment Analysis of YouTube Comments Using Bidirectional Encoder Representations from Transformers Neural Network Model

687

Figure 5: Word Cloud

Heatmap representing the confusion matrix,

offering a detailed analysis of the model's

classification performance and accuracy. For very

negative statements model has predicted 4 comments

accurately, for negative 48 comments are predicted

accurately, foe neutral 48 comments are predicted

accurately, for positive 12 comments are predicted

accurately, very positive 67 comments are predicted

accurately.

Figure 6: Confusion Matrix

5 CONCLUSIONS

The YouTube Comment Sentiment Analysis system

successfully demonstrated the application of machine

learning and natural language processing (NLP)

techniques to analyse audience feedback. By

automating the retrieval and sentiment classification

of comments, the system provided actionable insights

into how content is perceived. BERT algorithm is

provided the accuracy for classification of average

comments. We used fine tuning to increase the

accuracy of model. If we go with the optimization,

then again accuracy of the model will get increased.

REFERENCES

M. Salıcı and Ü. E. Ölçer, (2024) "Impact of Transformer-

Based Models in NLP: An In-Depth Study on BERT

and GPT," 2024 8th International Artificial

Intelligence and Data Processing Symposium (IDAP),

Malatya, Turkiye, 2024, pp. 1-6, doi:

10.1109/IDAP64064.2024.10710796.

G. P. M S, P. G. J, R. Kumar, P. K. H R, V. S. K and A.

Dahiya, (2024) "Sentiment Analysis Using Transfer

Learning for E-Commerce Websites," 2024

International Conference on Advances in Modern Age

Technologies for Health and Engineering Science

(AMATHE), Shivamogga, India, 2024, pp. 1-5, doi:

10.1109/AMATHE61652.2024.10582143.

Z. -Y. Lai, L. -Y. Ong and M. -C. Leow, (2023) "A

Composite Sentiment Summarizer Score for Patient

Reviews: Extending RoBERTa," 2023 11th

International Conference on Information and

Communication Technology (ICoICT), Melaka,

Malaysia, 2023, pp. 405-410, doi:

10.1109/ICoICT58202.2023.10262598

O. S. Chlapanis, G. Paraskevopoulos and A. Potamianos,

(2023) "Adapted Multimodal Bert with Layer-Wise

Fusion for Sentiment Analysis," ICASSP 2023 - 2023

IEEE International Conference on Acoustics, Speech

and Signal Processing (ICASSP), Rhodes Island,

Greece, 2023, pp. 1-5, doi:

10.1109/ICASSP49357.2023.10094923.

F. Wang, G. Liu, Z. Wang and X. Wu, (2021) "Sentiment

Analysis of Movie Reviews based on Pre training and

Dual Branch Coding," 2021 11th IEEE International

Conference on Intelligent Data Acquisition and

Advanced Computing Systems: Technology and

Applications (IDAACS), Cracow, Poland, 2021, pp.

1051-1055,doi: .1109/IDAACS53288.2021.9661049.

C. Tho, Y. Heryadi, I. H. Kartowisastro and W.

Budiharto,(2021) "A Comparison of Lexicon-based and

Transformer-based Sentiment Analysis on Code-mixed

of Low-Resource Languages," 2021 1st International

Conference on Computer Science and Artificial

Intelligence (ICCSAI), Jakarta, Indonesia, 2021,pp. 81-

85, doi: 10.1109/ICCSAI53272.2021.9609781

M. Munikar, S. Shakya and A. Shrestha,(2019) "Fine-

grained Sentiment Classification using BERT," 2019

Artificial Intelligence for Transforming Business and

Society (AITB), Kathmandu, Nepal, 2019, pp. 1-5, doi:

10.1109/AITB48515.2019.8947435.

AIlmania, Abdurrahman, S. Cahyawijaya and A.

Purwarianti, (2018) "Aspect Detection and Sentiment

Classification Using Deep Neural Network for

Indonesian Aspect-Based Sentiment Analysis," 2018

International Conference on Asian Language

Processing (IALP), Bandung, Indonesia, 2018, pp. 62-

67, doi: 10.1109/IALP.2018.8629181.

INCOFT 2025 - International Conference on Futuristic Technology

688

Z. Wang, C. S. Chong, L. Lan, Y. Yang, S. Beng Ho and J.

C. Tong,(2016) "Fine-grained sentiment analysis of

social media with emotion sensing," 2016 Future

Technologies Conference (FTC), San Francisco, CA,

USA, 2016, pp. 1361-1364, doi:

10.1109/FTC.2016.7821783.

Talaat, A.S, (2023) Sentiment analysis classification

system using hybrid BERT models. J Big Data 10, 110.

https://doi.org/10.1186/s40537-023-00781-w

Manuel-Ilie, Dorca & Gabriel, Pitic & George, Crețulescu.

(2023). Sentiment Analysis Using Bert Model.

International Journal of Advanced Statistics and IT&C

for Economics and Life Sciences. 13. 59-66.

10.2478/ijasitels-2023-0007.

Alaparthi, S., Mishra, M. (2021) BERT: a sentiment

analysis odyssey. J Market Anal 9, 118–126 (2021).

https://doi.org/10.1057/s41270-021-00109-8

R. Man and K. Lin, (2021) "Sentiment Analysis Algorithm

Based on BERT and Convolutional Neural

Network," 2021 IEEE Asia-Pacific Conference on

Image Processing, Electronics and Computers (IPEC),

Dalian, China, 2021, pp. 769-772, doi:

10.1109/IPEC51340.2021.9421110.

P. R. Gadyanavar and D. A. Kulkarni,(2018)"Temperature

measurement using soft computing techniques based on

computer vision," 2017 International Conference on

Energy, Communication, Data Analytics and Soft

Computing (ICECDS), Chennai, India, 2017, pp. 896-

900, doi: 10.1109/ICECDS.2017.8389566.

Sudharsan Ravichandiran, (2021) Getting Started with

Google BERT: Build and train state-of-the-art natural

language processing models using BERT , Packt

Publishing, 2021.

J. Yadav, D. Kumar and D. Chauhan,(2020)"Cyberbullying

Detection using Pre-Trained BERT Model," 2020

International Conference on Electronics and

Sustainable Communication Systems (ICESC),

Coimbatore, India, 2020, pp. 1096-1100, doi:

10.1109/ICESC48915.2020.9155700.

Sentiment Analysis of YouTube Comments Using Bidirectional Encoder Representations from Transformers Neural Network Model

689