International Journal of Trend in Scientific Research and Development (IJTSRD)

Volume 8 Issue 3, May-June 2024 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470

Opinion Text Analysis Using Artificial Intelligence

Rahul Sagar1, Sumit Dalal2, Sumiran3
1
Student, ECE, Sat Kabir Institute of Technology and Management, Bahadurgarh, Haryana, India
2,3
Assistant Professor, ECE, Sat Kabir Institute of Technology and Management, Bahadurgarh, Haryana, India

ABSTRACT How to cite this paper: Rahul Sagar |

This paper presents a robust methodology for sentiment analysis of Sumit Dalal | Sumiran "Opinion Text
comments leveraging advanced techniques such as Bag of Words Analysis Using Artificial Intelligence"
(BoW), K-Nearest Neighbors (K-NN), Support Vector Machine Published in
International Journal
(SVM), and Discriminant Analysis. Sentiment analysis plays a
of Trend in
crucial role in understanding user opinions, attitudes, and emotions Scientific Research
expressed in textual data. By employing BoVW, we extract and Development
discriminative features from comments, capturing both semantic and (ijtsrd), ISSN: 2456-
visual cues. These features are then utilized in conjunction with 6470, Volume-8 | IJTSRD64847
machine learning algorithms including K-NN, SVM, and Issue-3, June 2024,
Discriminant Analysis to classify sentiments accurately. The pp.168-174, URL:
proposed approach offers a comprehensive framework for sentiment www.ijtsrd.com/papers/ijtsrd64847.pdf
classification, achieving high accuracy and reliability across diverse
datasets. Experimental results demonstrate the effectiveness and Copyright © 2024 by author (s) and
scalability of the proposed methodology, showcasing its potential for International Journal of Trend in
Scientific Research and Development
real-world applications in sentiment analysis of comments across Journal. This is an
various domains. Open Access article
KEYWORDS: Opinion Analysis, K-NN, BOW, SVM, Discriminant distributed under the
terms of the Creative Commons
Analysis
Attribution License (CC BY 4.0)
(http://creativecommons.org/licenses/by/4.0)

INTRODUCTION
A social media platform like Facebook, provides an specific issue utilizing various sentiments including
opportunity for users to share their views and positive, negative, and neutral [4]. Sentiment analysis
opinions, as well as connect, communicate, and can be applied to social media data to investigate
contribute to specific topics through short-character variations in people's behavior, feelings, and
messages referred to as comments. This can be opinions, like by categorizing the spread tendency of
accomplished using text, images, and videos, among political campaigns. In this work, we use social media
other things, and users can interact by clicking the research to explore young sentiments. This article
like, comment, and repost icons. As more individuals examines the feelings of tweets using multiple
utilize social media, the study of data available online methodologies, including lexical and machine
can be utilized to shed light on evolving people's learning techniques. The time required is a major
views, conduct, and cognition [1]. As a result, issue for existing machine learning approaches,
employing Twitter or Facebook data for sentiment posing a hurdle for all firms seeking to transition their
analysis has grown more common. The growing operations to be processed by automated workflows.
interest in social media analysis has increased the Deep learning techniques have been applied to a
focus on text analysis technologies such as Natural variety of real-world applications, including
Language Processing (NLP) and Artificial sentiment analysis. Techniques for deep learning use
Intelligence (AI)[2]. Text analysis allows you to various algorithms to extract information from raw
assess the sentiments and attitudes of specific target data, including texts or tweets, and express it in
groups. Most of the existing literature concentrates on specific types of models. These models are used to
English texts, however there is an increasing interest derive information from novel datasets that have not
in multilingual analysis[3]. Text analysis can be before been represented.
performed by retrieving subjective comments about a

@ IJTSRD | Unique Paper ID – IJTSRD64847 | Volume – 8 | Issue – 3 | May-June 2024 Page 168
 International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470

Figure 1: Opinion Classification

Literature review
The key procedure in sentiment analysis is to categorize retrieved data into sentiment polarity such as positive,
neutral, and negative. A wide spectrum of emotions can also be thought of, which is fundamental to the growing
fields of affective computing and sentiment analysis [5]. Sentiments can be separated into satisfied and angry
categories based on the research topic. Sentiment analysis with ambivalence handling can be used to account for
more fine-grained outcomes and categorize emotions into such specific areas as anxiety, sadness, anger,
excitement, and happiness[6]. Sentiment evaluation is normally performed on text data, but it may also be
utilized for analyzing data from devices that use audio- or audio-visual formats such as webcams to study
expression, body movement, or noises, referred to as multimodal sentiment analysis[7]. Multimodal sentiment
analysis enhances text-based assessment into something more complicated, allowing for the application of NLP
for a variety of applications. Progress of NLP is also quickly expanding, driven by different studies, such as
neural networks[8]. One instance is the use of Neurosymbolic AI, which mixes deep learning and symbolic
reasoning and is seen as a promising tool in NLP for understanding reasonings (Sarker et al. 2021). This
demonstrates the broad possibilities for the path of NLP research[9]. There are three major ways to detecting and
classifying emotions represented in text: lexicon-based, machine-learning-based, and hybrid techniques.
The lexicon-based methodology relies on word polarity, whereas the machine learning approach views texts as
an issue of classification and further divides as unsupervised, semi-supervised, and supervised learning. In real-
world situations, machine learning and lexicon-based techniques may be utilized in conjunction. When dealing
with big text datasets, such as those from Twitter, it is critical to perform data pre-processing before beginning
analysis. This involves substituting upper-case letters, eliminating unnecessary words or links, extending
contractions, eliminating non-alphabetical letters or symbols, eliminating stop words, and eliminating redundant
datasets. Tokenization, stemming, lemmatization, and Part of Speech (POS) labeling should all be used in
addition to the fundamental data cleaning process.
Tokenization breaks down texts into smaller parts and converts them into a set of tokens. This makes it easier to
compute the frequency of each word in the text and determine its sentiment polarity. Stemming and
lemmatization substitute words with their roots. Utilizing stemming, the words "feeling" and "felt" can be
mapped to their stem word, "feel".
In contrast, lemmatization makes use of the words' contexts. This reduces the dimensionality and intricacy of a
bag of words while simultaneously improving the effectiveness of searching for the term in the lexicon when
using the lexicon-based approach. POS Tagging mechanically tags the POS of words in the text, including
nouns, verbs, and adjectives, which is beneficial to choosing features and retrieval.
LEXICON-BASED APPROACH
The lexicon-based methodology works by first splitting sentences into a bag of words, then comparing them to
terms in the sentiment polarity lexicon and their relevant semantic relations, and then calculating the polarity
score of the entire text. These techniques can accurately assess if the text's sentiment is favorable, negative, or
neutral[10]. The lexicon-based technique tags words with semantic orientation employing either dictionary-
based or corpus-based methods. The former is easier and we may calculate the polarity score of words or phrases
in the text employing a sentiment dictionary that includes opinion terms.

@ IJTSRD | Unique Paper ID – IJTSRD64847 | Volume – 8 | Issue – 3 | May-June 2024 Page 169
 International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470
MACHINE LEARNING APPROACH
Machine learning algorithms can design classifiers that finish sentiment categorization by getting feature
vectors, which mostly comprises phases such as data collection and cleaning, feature extraction, training data
with the classifier, and evaluating outcomes[11]. Employing machine learning techniques, the dataset should be
separated into two parts: training and testing. The training sets are designed to help the classifier learn text
features, while the test dataset evaluates its efficiency. Classifiers (for example, Support Vector Machines)
classify text into predefined categories. Machine learning is a common technique for text classification among
scholars. Furthermore, the accuracy of the same classifier for multiple types of text can vary significantly, so the
feature vectors for all kinds of text ought to be trained independently. In the following step, the tweeted data
must be vectorized and divided into a training set and a test set, after which the sentiment labels can be predicted
by employing various categorization models.

Figure 2: Machine Learning Based Classification

FEATURE REPRESENTATION MODELS
Bag of words (BoW): It translates textual input to numerical data using a fixed-length vector by calculating the
frequency of every phrase in tweets/comments. Compute the frequency of terms, which will result in a sparse
matrix with clean tokens. A 'Bag-of-Words (BoW)' model is being developed for this paradigm. A 'BoW' (which
is additionally referred to as a term-frequency counter) keeps track of how many times a given word comes in a
collection's documents. Yet, the MATLAB function 'bagOfWords()' does not separate the text into words. Every
single tweet is tokenized using the 'tokenizedDocument()' function.
The provided data is kept in a file containing a table of comments and their category sentiment. Column 1
displays category sentiment, whereas column 2 displays content tweets as text. Comments are in sorted in an
alphabetical order. The four exclusive opinions are enthusiasm, happiness, relief, and surprise. After tokenizing
tweets, a filter may eliminate English punctuation, common stopwords, and rare words (less than 100 times) for
counting purposes. This filter distinguishes meaningful words from irrelevant bits of speech.
Term frequency–inverse document frequency (TF–IDF) Matrix: It assesses the relevance of a word to the
overall text and the importance of the word in the tweet dataset. A TF-IDF matrix can be generated by
multiplying the word frequency metric and the inverse document frequency meter for each word in clean tweets.
The final step in preparing the data is to pick the label vectors and feature matrix. There are two groups of these:
one for training and another for testing and evaluation. The training set generates a feature matrix and label
vectors by picking the top m rows of the TF-IDF matrix and all columns. The test set for feature matrix and label
vectors includes rows following m rows.
METHODOLOGY
Once data preparation has been finished, the label vectors and features matrix can be utilized for categorizing
sentiment using various Classification methods. The prepared data is sent into the categorization methods, which
generate a model. The model used the training data to make prediction.
Models for classification: The classification of sentiment is the technique of estimating whether a user's tweet is
positive, negative, or neutral based on its feature representation. Classifiers in supervised machine learning
approaches, such as random forests, may categorize and predict unlabeled text after training on plenty of
sentiment-labeled tweets. The categorization models utilized in this paper are outlined below:

@ IJTSRD | Unique Paper ID – IJTSRD64847 | Volume – 8 | Issue – 3 | May-June 2024 Page 170
 International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470
Support Vector Machine Classification
The goal of this system is to find linear separators in vector space to help separate distinct types of input vector
data. After the hyperplane has been retrieved, the retrieved text features can be fed into the classifier to predict
the outcomes. Furthermore, the primary goal is to identify a line that is closest to the support vectors. The steps
for setting up SVC involve computing the distance between the nearest support vectors, also known as the
margin, maximizing the margin to find an optimal hyperplane between support vectors from given data, and
utilizing this hyperplane as a decision boundary to separate the support vectors.
K-Nearest Neighbor (K-NN)
The K-Nearest Neighbors (KNN) method is a prominent machine learning methodology for classification and
regression problems. It is based on the assumption that similar data points would have similar labels or
values.Throughout the training phase, the KNN algorithm keeps the complete training dataset as a reference.
When making predictions, it estimates the distance between the input data point and all of the training instances
using a distance metric such as Euclidean distance. The method then identifies the K nearest neighbors of the
input data point based on their distances. In the case of classification, the algorithm uses the most prevalent class
label among the K neighbors to predict the label for the input data point. Regression uses the average or
weighted average of the target values of the K neighbors to forecast the value of the input data point[12].
Linear discriminant analysis (LDA), as the name implies, is a linear framework for classification and
reduction of dimensionality. It is a statistical approach that divides data into categories. It detects patterns in
features that differentiate between classes. LDA seeks to identify a straight line or plane that best divides these
groups while reducing overlap between each class. It allows for accurate classification of fresh data points by
increasing the spacing between classes. Simply said, LDA helps make sense of data by determining the most
effective way to split various groups, which aids tasks such as pattern detection and classification.
Data Base: we have used Sentiment140 - Automatically labelled database of tweets. We have also used
Facebook comments Sentiment analysis database[13]. We combined data from these two databases and feed into
our model.
In this research, strategies for text cleaning, polarity calculation, and sentiment classification models are devised
and optimized utilizing two distinct sentiment analysis approaches: lexical and machine-learning-based
methodologies. We then compared the results of the various approaches, including output and prediction
accuracy. Machine-learning-based techniques require tweet labels, but manually annotating a significant amount
of data typically takes too long. As a result, 8000 tweets/comments are picked at random in this study, with an
average of roughly 1000 tweets each sentiment category.

Figure 3: Confusion Matrix for K-NN

@ IJTSRD | Unique Paper ID – IJTSRD64847 | Volume – 8 | Issue – 3 | May-June 2024 Page 171
 International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470

Figure 4: Confusion Matrix for Discriminant Analysis

Figure 5: Confusion Matrix for SVM

Accuracy, Precision, and Recall are the assessment indicators used in this research to assess the performance of
each categorization model. Before computing them, the values of the confusion matrix must be known, which
are TP (True Positive), TN (True Negative), FP (False Positive), and FN (False Negative). Using the formula
below, accuracy is expressed as the proportion of correct observations to total occurrences.
(1)

Precision is the percentage of positive observations that accurately forecast the total number of positive forecasts
using the calculation method.
(2)

Recall is the proportion of genuine positive observations that are accurately identified, computed using:
(3)

@ IJTSRD | Unique Paper ID – IJTSRD64847 | Volume – 8 | Issue – 3 | May-June 2024 Page 172
 International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470
The F1 Score is an in-depth assessment and balancing of precision and recall values. It can be computed as
follows:
(4)

Figure 6: Comparative analysis of Classifier Accuracy

Table 1: Data retrieved After Simulations
Number of Number of Discriminant SVM
K-NN Accuracy
Training tweets tested Tweets analysis Accuracy Accuracy
6532 1468 0.4133 0.5140 0.5109
6500 1500 0.4108 0.5101 0.5081
6000 2000 0.4121 0.5179 0.5135
5500 7450 0.4168 0.5211 0.5187
5000 3000 0.4153 0.5186 0.5166
It has been observed that accuracy of Discriminant analysis classifier is better than other two classifiers.
CONCLUSION 2, pp. 5200–5205, 2024, [Online]. Available:
Twitter sentiment analysis falls under the topic of text https://ijarasem.com/admin/img/62_Classificati
and opinion mining. This approach analyzes tweet on.pdf
sentiment and trains a machine learning model for [3] S. L. Lo, E. Cambria, R. Chiong, and D.
future use based on its accuracy. The method includes
Cornforth, “Multilingual sentiment analysis:
data collection, text pre-processing, sentiment
from formal to informal and scarce resource
detection, classification, training, and testing the languages,” Artif. Intell. Rev., vol. 48, pp. 499–
model. primarily we clean the data and utilize
527, 2017.
unsupervised lexicon-based methods to determine the
sentiment orientations of the tweets at every stage. [4] K. Arun and A. Srinagesh, “Multi-lingual
Then, we employ supervised machine learning Twitter sentiment analysis using machine
algorithms utilizing a sample of annotated data to learning,” Int. J. Electr. Comput. Eng., vol. 10,
train the K-NN, SVM, and discriminant analysis. no. 6, pp. 5992–6000, 2020.
REFERENCES [5] E. Cambria, D. Das, S. Bandyopadhyay, and A.
[1] A. H. Alamoodi et al., “Sentiment analysis and Feraco, “Affective computing and sentiment
its applications in fighting COVID-19 and analysis,” A Pract. Guid. to Sentim. Anal., pp.
infectious diseases: A systematic review,” 1–10, 2017.
Expert Syst. Appl., vol. 167, p. 114155, 2021. [6] Z. Wang, V. Joo, C. Tong, and D. Chan,
[2] R. Sagar, S. Dalal, R. Sharma, and Sumiran, “Issues of social data analytics with a new
“Classification of Sentiment Analysis method for sentiment analysis of social media
Techniques: A Comprehensive Approach,” Int. data,” in 2014 IEEE 6th International
J. Adv. Res. Arts, Sci. Eng. Manag., vol. 11, no. Conference on Cloud Computing Technology

@ IJTSRD | Unique Paper ID – IJTSRD64847 | Volume – 8 | Issue – 3 | May-June 2024 Page 173
 International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470
and Science, 2014, pp. 899–904. Technologies and Optimization (Trends and
[7] M. Soleymani, D. Garcia, B. Jou, B. Schuller, Future Directions)(ICRITO), 2020, pp. 537–
542.
S.-F. Chang, and M. Pantic, “A survey of
multimodal sentiment analysis,” Image Vis. [11] O. Adwan, M. Al-Tawil, A. Huneiti, R. Shahin,
Comput., vol. 65, pp. 3–14, 2017. A. A. Zayed, and R. Al-Dibsi, “Twitter
sentiment analysis approaches: A survey,” Int.
[8] Y. Kim, “Convolutional neural networks for
sentence classification. arXiv [J],” preprint, J. Emerg. Technol. Learn., vol. 15, no. 15, pp.
2014. 79–93, 2020.

[9] M. K. Sarker, L. Zhou, A. Eberhart, and P. [12] “K-NN”, [Online]. Available:

Hitzler, “Neuro-symbolic artificial https://www.analyticsvidhya.com/blog/2018/03
intelligence,” AI Commun., vol. 34, no. 3, pp. /introduction-k-neighbours-algorithm-
197–209, 2021. clustering/

[10] S. Zahoor and R. Rohilla, “Twitter sentiment [13] “Facebook comments Sentiment analysis”,
analysis using lexical or rule based approach: a [Online]. Available:
case study,” in 2020 8th International https://www.kaggle.com/code/mortena/faceboo
Conference on Reliability, Infocom k-comments-sentiment-analysis

@ IJTSRD | Unique Paper ID – IJTSRD64847 | Volume – 8 | Issue – 3 | May-June 2024 Page 174