10 VI June 2022

https://doi.org/10.22214/ijraset.2022.44500
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VI June 2022- Available at www.ijraset.com

Toxic Comments Classification

Neha Reddy1, Neha Ram2, Pallavi3, Dr. K. Kranthi Kumar4, Dr. K Sree Rama Murthy5
1, 2, 3 4, 5
Students, , Associate Professor, Department of Information Technology, Sreenidhi institute of Science and Technology, India,

Abstract: Conversational toxicity is a problem that might drive people to cease truly expressing themselves and seeking out other
people's opinions out of fear of being attacked or harassed. The purpose of this research is to employ natural language
processing (NLP) techniques to detect toxicity in writing, which might be used to alert people before transmitting potentially
toxic informational messages. Natural language processing (NLP) is a part of machine learning that enables computers to
comprehend natural language. Understanding, analysing, manipulating, and maybe producing human data language with the
help of a machine are all possibilities. Natural Language Processing (NLP) is a type of artificial intelligence that allows
machines to understand and interpret human language instead of simply reading it. Machines can understand written or spoken
text and execute tasks such as speech recognition, sentiment analysis, text classification, and automatic text summarization
using natural language processing (NLP).

I. INTRODUCTION
In the early days of the Internet, people communicated only by e-mail that encountered spam. At the time, classifying emails as
good or bad, whether or not spam was a challenge. Internet communications and data flows have changed significantly over time,
especially since the development of social networks. With the advent of social media, classifying content as “good” or “bad” has
become more important than ever to prevent social harm and prevent people from engaging in antisocial behavior.

A. Purpose
Every day, vast amounts of data are released by social media networks. This massive volume of data is having a big impact on the
quality of human existence However, because there is so much toxicity on the Internet, it can be harmful. Because toxic comments
limit people's ability to express themselves and have different opinions, as a result of the negative, there are no positive
conversations on social networks. As a result, detecting and restricting antisocial behaviour in online discussion forums is a pressing
requirement. These toxic comments might be offensive, menacing, or disgusting. Our goal is to identify these harmful remarks.

B. Project Overview
To classify noxious comments, this study will use different methods of machine learning. To handle the problem of text
categorization, Some of the methods we use are logistic entertainment, random forest, SVM classifier, multi-navigation database,
and XGBoost classifier. As a result, we maintain a data set using six machine learning algorithms, evaluate their accuracy, and
compare them.. We choose the model with the highest accuracy and forecast the toxicity based on unseen data based on its accuracy
level.

II. LITERATURE SURVEY

A. Existing Systems
1) Convolutional Neural Networks (CNN): CNN structures for categorization are Neural Network structures that can adapt to
many stages. Each of these levels has a number of layers. The Embedding Layer is a CNN component that is used to solve
problems with text categorization. The translation of the embedded text into CNN's description format is a function of the input
layer. This approach converts each word in a text document into a standard compressed vector.

Fig: A Convolutional Neural Network

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2839
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VI June 2022- Available at www.ijraset.com

2) LSTM (Long Short Term Memory): Long-term memory is abbreviated as LSTM. According to memory, LSDM is a kind of
continuous neural network that works better than conventional continuous neural networks. LSTM is certainly better when it
comes to memorizing specific patterns. Like other NNs, LSTM can have multiple hidden layers, storing the relevant data in
each cell as it passes through each level, and rejecting inappropriate data. LSTM has a memory function that allows you to store
data sequences. It also has another function: it works to delete unwanted data, and since we all know that text data has a lot of
unnecessary data, LSTM can delete it to save time and money on calculations. As a result, LSTM's ability to memorize data
sequences while deleting additional data makes it an effective text classification tool.

Fig:A Long Short-Term Memory Network.

B. Proposed System
We will use five machine learning algorithms to classify the data the online harmful comments: logistic regression, random forest,
SVM classifier, Naive Bayes, and XGBoost classifier. We'll use logistic regression to determine whether or not a comment is
harmful because it either belongs to the poisonous group or does not. SVM classifiers can be used to separate data values, and can
also be used in XGBoost and Random Forest systems. Because we can classify ideas into a wide range of destructive and non-toxic,
we will use the concept of results in both directions. Because our data are independent of each other and have nothing in common
with two different concepts, we use the Nave Bayes classifier to classify them. Because the data is marked, we can immediately use
a controlled machine learning algorithm.
1) Logistic Regression: One of the few approaches to the classification of utility data is logistic regression. Suppose you have a
medical history of a patient with a tumor. It is necessary to determine if the tumor is malignant (dangerous). The concept also
suggests whether it is toxic or not. 1 means positive class and 0 means negative class. Positive(1) mean that it is toxic and
negative(0) means that it is non-toxic. This is calculated by the sigmoid function. Suppose we have only one piece of
information. It has different features. Suppose we have the right weight mattress. Depending on the data point, we now need to
provide a class tag (classify it as 1 or 0). The weight vector is applied, and the input vector is multiplied by it, which gives a
scalable output. To get a value between 0 and 1, this output is placed into a Sigmoid function. We'll name this likelihood of the
projected class 1 for now. If the probability is greater than 0.5, the predicted class is 1. If the probability is less than 0.5, the
predicted class is 0.

Fig : Sigmoid Function

Fig : Logistic Regression curve

2) Random Forest: Random Forest is a controlled machine learning approach to classification and regression problems. Uses an
explicit majority for classification and a moderate delay to create a decision tree of different data. One of the main features of
the random forest algorithm is the ability to manage a set of data with classified and continuous variables, such as regression
and classification. As for the rating, it performs better than its competitors.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2840
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VI June 2022- Available at www.ijraset.com

STEPS
Step 1: Random Forest takes random entries from the Q-record database..
Step 2: A separate end tree is created for each model.
Step 3: Each end tree will have an output.
Step 4: The final decision on classification and regression is based on an appropriate or intermediate majority.

Fig : Random Forest

3) SVM Classifier: The reference vector machine (SVM) is a machine learning technique with classification and regression
monitoring. SVM defines a cloud page that classifies the boundaries between two data sets. SVM is often used to classify data,
although it can also be used for regression. This is a fast and reliable method that works well with low data. Another advantage
of SVM is that it can explore various input functions without increasing system problems, using different types of core
functions.

Fig : Flow chart of SVM

4) Naive Bayes: The Bayes theorem is a mathematical procedure for estimating conditional probabilities. A probability, as you
may know, is the possibility of an event occurring. We call an event's probability if it has a chance of occurring.

Fig : Formula for Naïve Bayes

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2841
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VI June 2022- Available at www.ijraset.com

Fig : Naïve Bayes Classification

The Bayes theorem is the foundation of the Naive Bayes algorithm. It's primarily utilized for categorization tasks. In classification,
we teach the model what each class belongs to using a labeled dataset, and then the model learns and classifies or labels a new
dataset that has never been seen before. All of the features or variables in the Naive Bayes model are treated as independent of one
another.

5) XGBoost Classifier: Gradual gain, such as Random Forest (another end-of-tree algorithm), is a controlled machine learning
approach for topics such as classification (men, women) and regression (expected value). The most common names to
implement this method are Gradient Boosting Machines (GBM) and XGBoost. Gradient Boosting is a training group similar to
Random Forest. This means that it links the final model to a set of individual models. These individual models have a poor
prognosis and are very consistent, but combining several of these weak models into one group will give the best results.
Random forest-related end trees are the most popular weak sample used in gradient reinforcement machines.

Figure :XG boost

Fig : XGBoost process

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2842
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VI June 2022- Available at www.ijraset.com

Fig : Flow chart for detecting toxic comments

III. PROJECT REQUIREMENTS

A. Python
Python is an advanced interactive and semantic programming language. It is written dynamically and has a ridiculously simple
syntax. It is written in plain English and has no syntax like Java or C. Seriously, it follows a stroke. Using language keywords, such
as when, how, in, etc., makes the code much clearer and simpler. It is a more semantic language because it is a structured language
with all the complete structures necessary for a specialized synthetic language. Python is an interactive programming language.
Python does not have a compiler, so the compiler executes the code directly at runtime.This is similar to how JS and PHP work.
1) Python is a universal programming language that can be used for a variety of applications. It was established in 1991 by Guido
van Rossum as a computer language with a design philosophy that emphasises code readability and the use of large spaces.
2) Dynamic kind systems and memory management square measure 2 of Python’s options. Object-oriented and imperative
programming, still as useful and procedural approaches, square measure all supported by the big and in-depthlibrary.
3) Python is a language for writing program descriptions. Use Python Prompt or Idle to build programs on Python Translator and
test variables and environments. Python is an object-oriented programming language. Python supports object- oriented
programming with full use of technology and its programming capabilities.

B. Numpy
NumPy is a Python module that uses arrays to perform a wide range of functions. To ensure faster execution, it can be
implemented using the C and C ++ libraries. It works faster with the NumPy module than similar performance without it.
This ability to fill signs with pure math is still a correction in February. It will not be added automatically when Python is
inserted. Must be installed at the same time as the NumPy pip installation command. The Python number is indicated by
NumPy.

C. Pandas
Pandas are a free, open source architecture that facilitates and directly manages linked and encrypted data. It includes several
data structures and methods for working with numerical data and time series data. This library is based on the NumPy Python
library. Pandas are a fast-paced program, and its users benefit from its speed and performance.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2843
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VI June 2022- Available at www.ijraset.com

D. Matplotlib
For 2D series diagrams, Matplotlib is a great Python imaging library. Matplotlib is a cross-platform NumPy-based data imaging
software compatible with the entire SciPy stack. First launched in 2002 by John Hunter.
One of the most important advantages of imaging is that it allows you to see large amounts of data in easy-to-understand images.
Line, strip, skater, histogram and many other maps are available on Matplotlib.

E. Seaborn
Seaborn is a great Python imager for programming statistical graphics. It has attractive default styles and a color palette that
make statistical maps very attractive. It uses matplotlib software and is closely related to the panda data structure.Seaporn's
mission is to develop data visualization as an important part of data recognition and understanding. It provides a database-based
API that allows you to switch between multiple visual representations of a variable for a better understanding of the data.

F. Scikit-Learn
Scikit-Learn is a free Python machine learning library. Supports both controlled and uncontrolled machine learning and includes a
number of algorithms for classification, regression, grouping, and size reduction. This library was created using many libraries you
know, such as NumPy and SciPy. It can also be used with other libraries, such as Panda and Seaborn.

IV. SYSTEM DESIGN

A. Architecture

Fig : Architecture

We upload our dataset first, then perform preprocessing such as stemming and removing stop words, before applying our machine
learning models and calculating accuracy for each model. Finally, the most accurate model is chosen.
The best model is then used to estimate toxicity based on unknown data.

V. OUTPUTS

Fig : Distribution of Tweets in Dataset

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2844
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VI June 2022- Available at www.ijraset.com

Fig : Accuracies of Algorithms

Now we predict on unseen data by taking our best algorithms Logistic Regression.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2845
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VI June 2022- Available at www.ijraset.com

Fig : Unseen data prediction

VI. CONCLUSION
We tested the accuracy of five machine learning methods: logistic rest, ship bay, random forest, SVM classifier, and XGboost
classifier. After careful consideration, we can conclude that the logistic delay has the best performance in terms of accuracy, as it is
the most accurate model among other models. As a result, we choose our final model based on its precise nature. If there is a logistic
regression model, we get the maximum. We will use the logistic delay model as our newest machine learning strategy, as it works
best with our data.
VII. FUTURE SCOPE
Other machine learning models can be used to calculate accuracy, interference loss, and loss recording to improve performance in
future research. We consider RNN, BERT, multilayer perceptron and GRU, all of which are deep learning methods. As a result, we
can explore a variety of alternative strategies to help improve the end result.

REFRENCES
[1] Rahul, H. Kajla, J. Hooda and G. Saini, "Classification of Online Toxic Comments Using Machine Learning Algorithms," 2020 4th International Conference
on Intelligent Computing and Control Systems (ICICCS), 2020, pp. 1119-1123, doi: 10.1109/ICICCS48265.2020.9120939
[2] M. Duggan, “Online harassment 2017,” Pew Res., pp. 1–85, 2017, doi: 202.419.4372.
[3] M. A. Walker, P. Anand, J. E. F. Tree, R. Abbott, and J. King, “A corpus for research on deliberation and debate,” Proc. 8th Int . Conf. Lang. Resour. Eval. Lr.
2012, pp. 812–817, 2012.
[4] J. Cheng, C. Danescu-Niculescu-Mizil, and J. Leskovec, “Antisocial behavior in online discussion communities,” Proc. 9th Int. Conf. Web Soc. Media,
ICWSM 2015, pp. 61–70, 2015.
[5] B. Mathew et al., “Thou shalt not hate: Countering online hate speech,” Proc. 13th Int. Conf. Web Soc. Media, ICWSM 2019, no. August, pp. 369–380, 2019.
[6] C. Nobata, J. Tetreault, A. Thomas, Y. Mehdad, and Y. Chang, “Abusive language detection in online user content,” 25th Int. World Wide Web Conf. WWW
2016, pp. 145–153, 2016, doi: 10.1145/2872427.2883062.
[7] E. K. Ikonomakis, S. Kotsiantis, and V. Tampakas, “Text Classification Using Machine Learning Techniques,” no. August, 2005.
[8] M. R. Murty, J. V. . Murthy, and P. Reddy P.V.G.D, “ Text Document Classification basedon Least Square Support Vector Machines with Singular Value
Decomposition,” Int. J. Comput. Appl., vol. 27, no. 7, pp. 21–26, 2011, doi: 10.5120/3312-4540.
[9] E. Wulczyn, N. Thain, and L. Dixon, “Ex machina: Personal attacks seen at scale,” 26th Int. World Wide Web Conf. WWW 2017, pp. 1391– 1399, 2017, doi:
10.1145/3038912.3052591.
[10] H. Hosseini, S. Kannan, B. Zhang, and R. Poovendran, “Deceiving Google’s Perspective API Built for Detecting Toxic Comments,” 2017, [Online]. Available:
http://arxiv.org/abs/1702.08138.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 2846