1.

INTRODUCTION

1.1 BACKGROUND INFORMATION ON THE PROJECT:

The world we see nowadays is becoming more digitalized. In this digitalized

world e- commerce is taking the ascendancy by making products available within the
reach of customers where the customer doesn't have to go out of their house. As now a
day's people are relying on online products so the importance of a review is going higher.
For selecting a product, a customer needs to go through thousands of reviews to
understand a product. But in this prospering day of machine learning, going through
thousands of reviews would be much easier if a model is used to polarize those reviews
and learn from it. We used self-supervised learning method on a largescale amazon
dataset to polarize it and get satisfactory accuracy. World we see nowadays is becoming
more digitalized. In this digitalized world e-commerce is taking the ascendancy by
making products available within the reach of customers where the customer doesn't
have to go out of their house. As now a day's people are relying on online products so
the importance of a review is going higher.

1.2 GOALS AND OBJECTIVES (POBJs):

1. Develop an efficient sentiment analysis system: The project aims to build a robust and
accurate sentiment analysis system that can effectively classify Amazon product reviews
into positive, negative, and neutral categories. This involves leveraging the strengths of
both VADER, a rule- based approach, and RoBERTa, a deep neural network-based
approach, to capture different aspects of sentiment in the text data.

2. Compare the performance of VADER and RoBERTa models: The project seeks to
compare the performance of VADER and RoBERTa models in the context of sentiment
analysis. This involves evaluating and analyzing the strengths and weaknesses of each
model.

3. Develop a user-friendly web interface: The project aims to build a user-friendly web
interface using the Flask framework, allowing users to input product reviews and obtain
real-time sentiment predictions. The web interface should be intuitive, interactive, and
easy to use, making the system accessible and usable for businesses and practitioners
who may not have expertise in machine learning or programming.

1
4. Ensure scalability and deployment: The project aims to develop a scalable sentiment
analysis system that can handle a large volume of Amazon product reviews efficiently.
The use of the Flask web framework allows for easy deployment of the system on web
servers or cloud platforms, ensuring scalability and availability for businesses of different
sizes and requirements.

5. Provide real-time sentiment analysis: The project aims to provide real-time sentiment
analysis of Amazon product reviews, allowing businesses to receive immediate feedback
on customer sentiments. This involves developing a system that can process and analyze
reviews in real-time, enabling businesses to promptly respond to customer feedback and
take necessary actions to improve their products or services.

6. Add value to businesses: The project aims to create a system that can add value to
businesses by providing valuable insights from customer reviews on Amazon products.
The sentiment analysis system can help businesses understand customer sentiments,
identify areas of improvement, and make data-driven decisions to enhance their products,
services, and customer satisfaction, ultimately leading to improved business
performance.

1.3 PROBLEM STATEMENT :

The proposed project aims to address the problem of efficiently and accurately analyzing
Amazon product reviews for sentiment classification. With the increasing volume of
online product reviews, businesses need a reliable and automated way to understand the
sentiment of customer feedback. Manually analyzing a large number of reviews can be
time-consuming, error-prone, and impractical. Hence, the project focuses on developing
a sentiment analysis system using the Vader and Roberta models, integrated with the
Flask framework, to provide an automated solution. It is important as businesses rely on
customer feedback to make informed decisions, such as improving product quality,
marketing strategies, and customer service. By automating the sentiment analysis
process, the project aims to provide a scalable and efficient solution for analyzing a large
volume of Amazon product reviews, saving time and resources.

1.4 SCOPE OF THE PROJECT :

1. Data Collection and Preprocessing: The project requires collecting a large dataset
of Amazon product reviews from different categories and preprocessing the data to
ensure it is clean and ready for analysis. This includes tasks such as text cleaning,
tokenization, and feature extraction.

2
2. Sentiment Analysis Models: The project involves implementing and integrating two
different sentiment analysis models - VADER and RoBERTa. VADER is a rule-based
approach that relies on lexical and grammatical rules, while RoBERTa is a deep learning
model that captures complex contextual information from text. The scope of the project
includes training and fine-tuning these models to optimize their performance for
sentiment classification.

3. Flask Web Application: The project includes developing a web application using the
Flask framework for user interaction. The scope of the project includes designing and
implementing an interactive user interface, along with visualizations such as word clouds
and sentiment distribution charts.

4. Deployment and Application: The project includes deploying the developed

sentiment analysis system to a production environment, making it accessible for real-
world use. The scope of the project includes testing the system with actual Amazon
product reviews, and showcasing its potential applications in areas such as e-commerce,
market research, and customer feedback analysis.

The project has potential applications in the field of e-commerce, customer

feedback analysis, and market research, providing valuable insights and actionable
information to businesses for decision making and product improvement.

1.5 LIMITATIONS OF THE PROJECT:

1. Limited accuracy

2. Real-time data processing

3. Language and cultural biases

4. External factors and context

5. Subjectivity of Sentiment

3
 2. PROBLEM DEFINITION

“To classify the amazon product reviews in to positive, negative and neutral reviews and
to find the final sentiment and sentiment score of the reviews with the visual
representation.”

2.1 FLOW CHART OF SENTIMENTAL ANALYSIS PROCESS:

Fig 2.1 flow chart for sentimental analysis process

2.2 ALGORITHMS USED TO SOLVE THE PROBLEM:

The proposed project involves using two different algorithms for sentiment analysis:
(i) VADER (Valence Aware Dictionary and Sentiment Reasoner) algorithm
(ii) RoBERTa(Robustly Optimized BERT Approach) model.

1. VADER Model: The VADER model is a rule-based sentiment analysis approach

that relies on a pre-built lexicon of words and phrases with assigned sentiment scores. It
uses a combination of lexical and grammatical heuristics to determine sentiment polarity
(positive, negative, or neutral) and sentiment intensity (strength of the sentiment) of a
given text. The algorithm also accounts for punctuation, capitalization, and context-
specific rules to improve accuracy. VADER is known for its fast and efficient
processing and is commonly used for social media sentiment analysis.

4
 INTERNAL WORKING OF VADER:
The VADER (Valence Aware Dictionary and sEntiment Reasoner) model is a widely
used rule- based sentiment analysis tool that is specifically designed for social media text
analysis. It is capable of determining the sentiment polarity (positive, negative, or neutral)
of a piece of text, as well as providing a sentiment intensity score that represents the
strength of the sentiment.
The internal working of the VADER model in the proposed project can be summarized
as follows:

1. Lexicon-based approach: The VADER model uses a pre-built sentiment lexicon, which
is a dictionary containing words or phrases with associated sentiment scores. The
lexicon contains words that are commonly used in social media text and their sentiment
scores are based on human judgment. The sentiment scores in the lexicon are typically
calibrated to capture the sentiment expressed in social media text more accurately.

2. Sentiment intensity scoring: The VADER model calculates a sentiment intensity score
for a given piece of text by summing the sentiment scores of individual words in the
lexicon. The sentiment scores can be positive, negative, or neutral, and they are weighted
based on their relative importance in the text. The intensity score represents the overall
sentiment polarity and strength of the sentiment in the text.

3. Handling negations and intensifiers: The VADER model also takes into account the
impact of negations and intensifiers in the text. It uses a set of rules to detect negations
and intensifiers and appropriately adjusts the sentiment scores of the words that are
affected by them. For example, words following negations like "not" or "but" are given
opposite sentiment scores, and intensifiers like "very" or "extremely" can increase the
sentiment score of a word.

4. Handling emoticons and punctuation: The VADER model also considers emoticons
and punctuation marks in the text as additional cues for sentiment analysis. Emoticons,
such as :) or :(, are assigned sentiment scores based on their typical sentiment
connotations. Punctuation marks, such as exclamation marks or question marks, can
also influence the sentiment intensity score by indicating the intensity of the sentiment
expressed.

5. Integration with Flask framework: The VADER model is integrated into the Flask
framework in the proposed project to enable sentiment analysis of Amazon product
reviews. The Flask framework provides a web interface for users to input the product
reviews, and the VADER model is called to perform sentiment analysis on the input
text. The sentiment analysis results, including sentiment polarity (positive, negative,
or neutral) and sentiment intensity score, are then displayed to the users through the Flask
interface.

5
It is important to note that the VADER model has its limitations, such as potential biases,
subjectivity, and reliance on pre-built sentiment lexicons. Care should be taken to
understand and mitigate these limitations in the project, and further evaluation and
validation may be needed to ensure the accuracy and reliability of the sentiment analysis
results.

Fig 2.2 : flow chart of VADER model

2. Roberta Model: The Roberta model is a deep learning-based approach for sentiment
analysis that utilizes a variant of the BERT (Bidirectional Encoder Representations from
Transformers) architecture. BERT is a pre-trained language model that learns
contextualized word representations from a large corpus of text data. Roberta is a robustly

6
 optimized version of BERT that further improves its performance by incorporating
additional training data and optimization techniques. The Roberta model is capable of
capturing complex contextual information from text and achieving state-of-the-art
performance in various natural language processing tasks, including sentiment analysis.

INTERNAL WORKING OF RoBERTa MODEL:

The Roberta model is a state-of-the-art pre-trained language model that uses a

transformer-based architecture for natural language processing tasks, including sentiment
analysis. The internal working of the Roberta model in the proposed project can be
summarized as follows:

1. Pre-training on large text corpus: The Roberta model is pre-trained on a large corpus
of text data, typically containing billions of words, to learn language patterns, syntax,
and semantic representations. During pre-training, the model learns to predict masked
words in sentences, which helps it capture contextual information and contextualized
word representations.

2. Fine-tuning for sentiment analysis: After pre-training, the Roberta model is fine-
tuned on a labeled dataset of Amazon product reviews for sentiment analysis. Fine-
tuning involves training the model on a smaller, task-specific dataset to adapt it to the
specific sentiment analysis task. The labeled dataset contains product reviews with
sentiment labels (positive, negative, or neutral) for training the model to predict
sentiment polarity.

3. Tokenization: Text input for the Roberta model is tokenized, which involves breaking
it into smaller units called tokens. Tokens can be words, sub-words, or characters,
depending on the language and model configuration. Tokenization is an important step
as it helps the model to process text input efficiently.

4. Embedding generation: The Roberta model generates contextualized word

embeddings, which are vector representations of words that capture their contextual
meaning in the given text. These embeddings are generated using the transformer
architecture, which allows the model to capture long-range dependencies and contextual
information.

5. Sentiment prediction: The contextualized word embeddings are fed into a

sentiment classification layer in the Roberta model, which predicts the sentiment
polarity of the input text. The sentiment classification layer is trained during the fine-
tuning process to classify the text into one of the sentiment categories: positive,
negative, or neutral.

7
6. Integration with Flask framework: The Roberta model is integrated into the Flask
framework in the proposed project to enable sentiment analysis of Amazon product
reviews. The Flask framework provides a web interface for users to input the product
reviews, and the Roberta model is called to perform sentiment analysis on the input
text. The sentiment analysis results, including sentiment polarity (positive, negative, or
neutral), are then displayed to the users through the Flask interface.

It is important to note that the Roberta model may require significant computational
resources for training and inference due to its large size and complexity. Care should be
taken to optimize the performance and efficiency of the model in the project, and further
evaluation and validation may be needed to ensure the accuracy and reliability of the
sentiment analysis results.

Fig 2.3 flowchart of RoBERTa model

8
 3. DATA COLLECTION AND PROCUREMENT

3.1 LIST OF INPUTS AND EXPECTED OUTPUTS:

INPUTS:
Amazon product reviews dataset: This is the primary input to the project, which contains the
text reviews written by users for various Amazon products. The dataset should be in a
structured format, such as CSV, with relevant fields like review text, rating, product ID, etc.

Amazon review dataset:

https://www.kaggle.com/datasets/snap/amazon-fine-food-reviews

OUTPUTS EXPECTED:

1. Sentiment classification labels: The main expected output of the project is the sentiment
classification of the Amazon product reviews into positive, negative, or neutral categories.
Each review in the dataset should be classified into one of these categories based on its
sentiment polarity.

2. Sentiment scores or probabilities: Along with the sentiment classification labels, the project
may also output sentiment scores or probabilities for each review, indicating the level of
positive, negative, or neutral sentiment. For example, a review may be classified as
positive with a sentiment score of 0.9, indicating high positive sentiment, while another review
may be classified as negative with a sentiment score of 0.3, indicating low negative sentiment.

3. Count of reviews: The number of positive, negative, neutral reviews in the total inputted
reviews.

4. Visualizations or graphical representations: The project may also generate visualizations

or graphical representations of the sentiment analysis results, such as bar charts, pie charts
to provide a visual understanding of the sentiment distribution among the Amazon product
reviews.

5. Flask web application : If the project includes the development of a Flask web application,
the expected output would be a functional web application that allows users to input Amazon
product reviews and displays the sentiment analysis results in an interactive manner, such as
displaying the sentiment labels, scores, and visualizations on a web page.

3.2 SOURCES OF DATA USED:

1. Amazon website
2. Kaggle

9
3.3 HARDWARE/SOFTWARE REQUIRED FOR THE PROJECT:

HARDWARE REQUIREMENTS:

1. Hard Disk : 256 GB.

2. Input Devices : Keyboard, Mouse
3. Ram : 8 GB.

SOFTWARE REQUIREMENTS:

1. Operating system : Windows 10

2. Coding Language : python , HTML, CSS, javascript
3. Tool : jupyter Notebook, VS code

10
 4. DESIGN AND IMPLEMENTATION

4.1 MODULES USED:

1. Pandas : Pandas is a popular data manipulation library in Python that is commonly used
in machine learning for tasks such as data preparation, data exploration, feature
engineering, data visualization, and data preprocessing for machine learning models.

2. Numpy : NumPy is a fundamental numerical computing library in Python used in

machine learning for tasks such as numerical operations, mathematical calculations, and
array manipulation. It provides efficient array operations, linear algebra, and
mathematical functions, making it essential for many machine learning algorithms and
data processing tasks.

3. Matplotlib : Matplotlib is a widely used data visualization library in Python that plays a
crucial role in machine learning for tasks such as creating visualizations to analyze data,
understand patterns, and communicate results. It provides a wide range of plot types and
customization options, making it an essential tool for visualizing data in machine
learning projects.

4. Seaborn : Seaborn is a Python data visualization library based on Matplotlib that is

commonly used in machine learning for creating attractive and informative statistical
visualizations. It provides a high-level interface for creating visually appealing plots,
such as scatter plots, bar plots, and heatmaps, which can aid in data analysis, model
evaluation, and result communication in machine learning projects.

5. Nltk : NLTK (Natural Language Toolkit) is a popular open-source Python library for
natural language processing (NLP). It provides a wide range of tools, resources, and
algorithms for processing and analyzing text data, making it a valuable tool for NLP-
related tasks in machine learning. NLTK is widely used in machine learning projects for
text processing, feature extraction, and data preparation for NLP tasks. It offers an
extensive collection of functions and algorithms that can be combined with other
machine learning libraries to build end-to-end NLP applications.

6. Scipy : SciPy is a popular scientific computing library in Python that is commonly used
in machine learning for tasks such as numerical optimization, linear algebra, integration,
interpolation, and signal processing. It provides a wide range of mathematical and
scientific functions that are essential for implementing machine learning algorithms and
performing advanced data analysis and computation.

11
 7. Huggingface : Hugging Face is a popular open-source organization and platform for
natural language processing (NLP) that provides various tools and resources for NLP
tasks. Hugging Face is known for its contributions to the NLP community, including pre-
trained models such as BERT, GPT-2, and T5, as well as libraries like Transformers,
Tokenizers, and Datasets that make it easier to work with NLP tasks in Python.

4.2 CODING:
1. import required libraries

Fig 4.1 importing required libraries

12
 2. Read the data from dataset

Fig 4.2 Reading the data in to a data frame

3. Performing Quick Exploratory data analysis

The count of 5 star rating is higher for the fine food reviews dataset which was taken.

Fig 4.3 Bar plot which shows the count of reviews by rating

13
4. Perform basic NLTK

Fig 4.4 performing basic nltk steps

14
5. VADER model

Fig 4.5 construction of vader model

15
6. RoBERTa MODEL

Fig 4.6 Construction of RoBERTa model

16
 5. RESULTS AND DISCUSSION

5.1 EXPECTED OUTCOMES :

• Sentiment classification labels
• Sentiment scores or probabilities
• Count of reviews (positive, negative and neutral)
• Visualizations or graphical representations

5.2 RESULTS OBTAINED:

Fig 5.1: Home page of the sentiment analysis web application

17
Fig 5.2: Index page of the web app where we were provided with 2 fields. In the first field ,
we can give any single review as an input to analyse the sentiment. In the next field, we can
upload a csv file of reviews to be analysed.

Fig 5.3: Single review is given as input in the first input field provided

18
Fig 5.4: The sentiment analysis result obtained for the review entered. It depicts
that the given review has negative sentiment score with negative sentiment.

Fig 5.5: The product reviews are exported from amazon website and the csv file of that
reviews is given as input in the second input field provided for analysis.

19
Fig 5.6: The Sentiment analysis results obtained for the reviews dataset uploaded for
analysis.
It depicts the total number of reviews in the file, count of positive reviews, negative
reviews and neutral reviews, final sentiment and final sentiment score. And also the
visual representation of reviews percentage using a pie chart.

20
 6. CONCLUSION

6.1 SUMMARY OF THE PROJECT:

The project involves the sentiment analysis of Amazon product reviews using
VADER and RoBERTa models, implemented within a Flask framework. The aim of the
project is to classify the sentiment of the reviews as positive, negative, or neutral, and
provide valuable insights for decision making, user experience enhancement, and
business outcomes.

Based on the analysis of Amazon product reviews using the VADER and
RoBERTa models for sentiment analysis, it can be concluded that both models are
effective in accurately identifying the sentiment of the reviews. However, the RoBERTa
model outperformed the VADER model indicating that it is a more robust and powerful
model for sentiment analysis.

This project involves collecting a large dataset of labeled Amazon reviews, pre-
processing the data, developing a neural network model, training and testing the model,
and deploying it for real-time analysis of sentiments. The deployment of the trained
model can provide valuable insights to businesses that can be used to improve their
products and services.

Overall, sentimental analysis using deep learning methods is a powerful tool that
can be used to gain valuable insights into customer sentiments and preferences. With the
increasing importance of online reviews in shaping customer decisions, this project can
help businesses stay competitive and stay ahead of the competition.

6.2 LIMITATIONS:
1. Limited accuracy: Despite the high performance of VADER and RoBERTa models,
sentiment analysis may still have limitations in accurately capturing the nuances of
sentiment in text, such as sarcasm, irony, or context-dependent sentiment. This could
result in misclassification or incomplete analysis of certain reviews, leading to potential
inaccuracies in the overall sentiment analysis results.

2. Real-time data processing: Real-time processing of large volumes of Amazon

product reviews may pose challenges in terms of computational resources, processing
speed, and system scalability. If the system is not optimized for handling real-time data,
it could result in delays or inefficiencies in obtaining real-time sentiment insights from
customer reviews.

21
3. Language and cultural biases: The sentiment analysis models may be biased towards
the language and cultural context in which they are trained. If the dataset used for training
and fine- tuning the models is predominantly from a specific language or cultural
background, it may result in biased sentiment analysis results when applied to reviews
in other languages or cultures.

4. External factors and context: Sentiment in reviews can be influenced by external

factors, such as product price, brand reputation, marketing campaigns, or current events.
The sentiment analysis models may not take into account these external factors and
context, which could impact the accuracy and interpretation of sentiment analysis results.

5. Subjectivity of sentiment: Sentiment is inherently subjective, and different people

may interpret and express sentiment differently. The sentiment analysis models may not
capture the subjective nature of sentiment accurately, leading to potential discrepancies
between the predicted sentiment and the actual sentiment perceived by users, which
could affect the reliability of the results

22