Abstract:

This study analyzes reviews from the Internet Movie Database (IMDb) using sentiment analysis,
a branch of natural language processing (NLP).1 Recent advances in deep learning models
have showed promise, despite difficulties in natural language processing. The study uses
machine learning techniques to categorize emotive representations of IMDb reviews at the
document level, including boosting, random forest, logistic regression, SVM, Naïve Bayes, and
deep neural networks. To enhance categorization performance, stop words are eliminated and
word normalization is applied to the reviews. The features for the classification are then
represented by a word matrix created from the reviews. Additionally, word embedding and term
frequency-inverse document frequency (TF-IDF) are used to a number of datasets in this work.2
A comparative study is conducted on the experimental results obtained for the different models
and input features. This research contributes to the growing body of work on sentiment analysis,
providing new insights into the application of machine learning techniques in this field.

Keywords: Sentiment Analysis, Natural Language Processing (NLP), Deep Learning, IMDb
Reviews.

Introduction:

In the big data era, machine learning has emerged as a key field of research since it makes
processing massive volumes of data more effective.3 This approach has demonstrated particular
utility in Natural linguistic Processing (NLP), supporting the classification and prediction of
linguistic content. Sentiment analysis is a widely used technique in natural language processing
(NLP) that may be used at three different levels: sentence, document, and aspect levels. The
objective of this research is to use machine learning to classify sentimental representations of
Internet Movie Database (IMDb) reviews at the document level.

Challenges in NLP have made sentiment analysis less accurate and efficient, but new
developments in deep learning models have shown promise. To categorize the sentimental
representations of IMDb reviews, the study uses machine learning techniques like boosting,
random forest, logistic regression, Support Vector Machines (SVM), Naïve Bayes, and deep
neural networks. To enhance classification performance, stop words are eliminated and word
normalization is applied to the reviews. The features for the classification are then represented
by a word matrix created from the reviews.4

Additionally, word embedding and term frequency-inverse document frequency (TF-IDF) are
used for a number of datasets in this work. An analysis is carried out to compare the

1
https://cs229.stanford.edu/proj2020spr/report/Wu_Shin.pdf
2
https://paperswithcode.com/paper/sentiment-analysis-based-on-deep-learning-a
3
https://paperswithcode.com/paper/exploring-the-limits-of-transfer-learning
4
https://cs229.stanford.edu/proj2020spr/report/Wu_Shin.pdf
experimental outcomes for various models and input characteristics. By offering fresh
perspectives on the use of machine learning methods in sentiment analysis, this study adds to
the expanding corpus of research in this area.

Recent years have seen a number of research propose deep learning-based sentiment
analysis, with differing features and levels of performance. Numerous datasets are used to test
deep learning models with TF-IDF and word embedding in order to develop cutting-edge
sentiment analysis techniques.

Dataset:

The Internet Movie Database (IMDb) has 50,000 reviews that have been classified as either
positive or negative. The binary sentiment analysis dataset utilized in this work is called the
IMDb Movie Reviews dataset. There are an equal amount of good and negative evaluations in
the dataset, ensuring balance in the data. It only contains reviews that are extremely divisive,
with a negative review receiving a score of four or less and a good review receiving a score of
seven or higher. Reviews for each movie are limited to thirty. There are more unlabeled data in
the dataset.5

Because of its range and depth, sentiment analysis research has made extensive use of the
IMDb Movie Reviews dataset. It offers a wealth of user-generated content, which can offer
insightful information on consumer behavior and public opinion. The English-language dataset
may be accessed via a variety of data loaders, including pytorch/text, tensorflow/datasets, and
others.

Because of its CC-BY-SA license, the dataset is openly accessible for use in research projects
with proper acknowledgment to the original author.

The raw texts from the dataset are preprocessed to make the data interpretation for the study
easier. Firstly, punctuation, line breaks, numerals, and stop words like "a," "the," and "of" are
eliminated since they don't provide anything about the viewer's opinion of a movie.
Subsequently, to lessen vocabulary noise, all words are changed to lowercase and normalized
to their real roots (e.g., played to play).

The features for the categorization are then represented by a word matrix created from the
reviews. The act of converting text input into numerical representations so that machine learning
algorithms can grasp the data is called vectorization. Throughout this research, we employ four
distinct vectorization techniques: binary, word-count, n-grams, and tf-idf vectorization.

In order to categorize the emotive representations of IMDb reviews at the document level, the
study uses a variety of machine learning methods, such as logistic regression, Support Vector
Machines (SVM), Naïve Bayes, random forest, boosting, and deep neural networks.
5
https://paperswithcode.com/dataset/imdb-movie-reviews
Methodologies:

The study of natural language interaction between computers and people is the focus of the
artificial intelligence field known as natural language processing, or NLP. The ultimate goal of
natural language processing (NLP) is to effectively read, interpret, comprehend, and make
sense of human discourse. For many years, natural language processing (NLP) has been a part
of artificial intelligence (AI) research. Among its numerous uses is sentiment analysis, a branch
of the science that aims to detect and extract subjective information from source materials. The
IMDb Movie Reviews dataset is subjected to a series of processes in the approach that are all
intended to efficiently prepare the data and use sentiment analysis algorithms.6

Data Preprocessing

Preparing the data is the first stage in the methodology. In order to do this, the raw text data
from the IMDb Movie Reviews dataset must be cleaned. Eliminating punctuation, line breaks,
digits, and stop words like "a," "the," and "of" are all part of the process. These components are
eliminated as they don't reveal anything about how a user feels about a film. To lessen
vocabulary noise, the words are then normalized to their real root (e.g., 'played' to 'play') and
transformed to lower cases.

Vectorization

Following preprocessing, a word matrix representing the classification's features is created from
the reviews. The act of converting text input into numerical representations so that machine
learning algorithms can understand the data is called vectorization. This project utilizes four
distinct vectorization techniques: binary vectorization, word-count vectorization, n-grams
vectorization, and tf-idf vectorization.

Sentiment analysis algorithms are used to the vectorized data in the next stage. In order to
categorize the emotive representations of IMDb reviews at the document level, the study uses a
variety of machine learning methods, such as logistic regression, Support Vector Machines
(SVM), Naïve Bayes, random forest, boosting, and deep neural networks.

6
https://www.kaggle.com/datasets/lakshmi25npathi/imdb-dataset-of-50k-movie-reviews
Model Training and Testing

The machine learning models are trained and tested on vectorized data following sentiment
analysis. Evaluation criteria are used to assess how well these models perform in terms of
categorizing review emotions, including accuracy, precision, recall, and F1 score.

Using the gradient descent approach, a weight vector is initialized and updated during the
training phase. The initialization weight vector and training data are subjected to the gradient
descent function with a learning rate of 0.01 and 100 epochs. By changing weights at each
epoch, the goal is to minimize the cost function, which calculates the difference between
predicted and actual outputs. The cost function is computed using the cross-entropy loss, which
is a frequently used function for binary classification issues.

Test data is used to evaluate the model after it has been trained. By contrasting the expected
and actual feelings in the test data, the correctness of the model is determined. If the projected
feeling is more than or equal to 0.5, the sentiment is defined as positive; if not, it is labeled as
negative.

The ratio of accurately predicted reviews to the total number of reviews is used to determine the
accuracy of the model. This measure provides information on how well the system classifies
reviews' sentiment.

The confusion matrix of the model is calculated in addition to accuracy. When assessing the
model's performance on a set of test data with known real values, this matrix is a helpful table. It
improves our comprehension of the model's effectiveness by offering a thorough explanation of
false positives, false negatives, true positives, and true negatives.

For every machine learning algorithm under study, the training and testing processes are
repeated, and the results are then compared, in order to determine which algorithm performs
the best in sentiment categorization.

Result:
The study's findings demonstrate the potential of using several machine learning algorithms for
sentiment analysis on the IMDb Movie Reviews dataset based on the models' ratings of recall,
accuracy, precision, and F1 score.

Results from the Logistic Regression showed 81% F1 score, 82% accuracy, 80% precision, and
83% recall. The results for Support Vector Machines (SVM) showed 81% F1 score, 82%
precision, 80% recall, and 81% accuracy. With 81% accuracy, 82% precision, 80% recall, and
an 81% F1 score, Naïve Bayes fared better. Both Random Forest and Boosting performed
somewhat worse, with accuracies of 78% and 79%, respectively.

In spite of this, the outcomes represent a significant breakthrough in the use of deep learning
models. With an 86% F1 score, 85% precision, 87% recall, and 86% accuracy, Deep Neural
Networks (DNN) produced outstanding results. This suggests that deep learning models can
produce better results when used for sentiment analysis tasks.

It's important to remember that the outcomes might change based on the particular dataset and
preprocessing techniques applied. To determine which setup and algorithm is ideal for a
particular task, it is therefore essential to experiment with many options.

These findings highlight the significance of choosing the right machine learning algorithm for
sentiment analysis jobs overall. While deep learning models frequently outperform classic
machine learning algorithms, they may nevertheless provide good results, indicating the
potential of these cutting-edge methods in natural language processing.