Ratio Mathematica                                                                            Volume 42, 2022 

 

 
 

 

 

 

A Deep Learning Model for Classifying the 

Hate and Offensive Language in Social Media 

Text 

 
 

Nidhi Bhandari1 

Rachna Navlakhe2 

G.L. Prajapati3 
 
 

Abstract 

Social media is a big data source for analyzing content. It requires a significant 

amount of computational cost. Recently introduced model for identifying and 

removing toxic content from Twitter, using an Information Retrieval (IR) model is 

working accurately but with small amount of data. Therefore, in this paper, a deep 

learning technique is used to process large-scale social media text data. First, it 

uses Natural Language Processing (NLP) based feature extraction to create four 

different sets of training samples i.e. TF-IDF-based features, POS Tagged Features, 

a reduced feature vector of POS and the combined vector of TF-IDF and POS 

tagged features. The deep Convolutional Neural Networks (CNN) is used to train 

the model and to classify hate and offensive language. The dataset has been 

obtained from Kaggle. The performance in terms of training accuracy, validation 

accuracy, training loss and validation loss has been measured with the time 

complexity. In addition, the class-wise Precision, Recall, F1-score and Mean 

accuracy have also been investigated. From experimental results, we found TF-IDF 

and POS-based combined features provide 82% and 83% of accuracy, respectively. 

Keywords: Text mining, social media, semantic knowledge, sentiment 

analysis, deep learning, hate and offensive language.* 

 
1Department of Applied Mathematics, IET, DAVV, Indore. nidhi.bhandari1@gmail.com. 
2Department of Applied Mathematics & Compu. Sci., SGSITS, Indore, sgsits.rachna@gmail.com. 
3Department. of Comp. Engi., IET, DAVV, Indore, glprajapati1@gmail.com. 
*Received on January 21st, 2022. Accepted on June 20th, 2022. Published on June 30th, 2022. doi: 

10.23755/rm.v39i0.705. ISSN: 1592-7415. eISSN: 2282-8214. ©The Authors. This paper is published 

under the CC-BY licence agreement. 

73



Nidhi Bhandari, Rachna Navlakhe and G.L.Prajapati 

 

1. Introduction 

Information Retrieval (IR) is a technique used to locate precise information and 

supports different data formats, i.e. text, image, video, etc. Among these data formats, 

the text has a significant contribution. The text IR model uses text mining techniques. In 

IR techniques, text mining or data mining algorithms are employed to recover user 

query relevance information [1].The IR model contains three key components: (1) User 

query (2) Query processing (3) Generation of outcomes[2]. However, the deficiency of 

these components can impact the performance of IR model such as lack of specific user 

query keywords, inappropriate keyword selection, lack of similar data, ranking of 

results etc. [3].  

Our study is initiated with an IR model named SOIR (Semantic query Optimization-

based Information Retrieval). This IR model incorporates query optimization and an 

FCM clustering technique. The results indicate the performance of the SOIR is better 

than previous models [4]. However, there are various applications of the IR systems. 

Beyond these applications, the IR model can also be applied for pattern recognition. In 

this context, the proposed work is extended in order to be used with the social media 

toxic content filtering. This paper is an extension of the SOIR model, which will be 

used for classifying the toxic contents from the social media posts. This model is a 

promising technique for handling various negative tweets from social media by using 

lexical as well as semantic pattern analysis. 

In this paper, we are going to evaluate the SOIR and DGM based toxic content 

classification system on a large dataset. Additionally, we do an investigation of a deep 

learning technique for classifying the different NLP based text features. We organize 

the contents in the following manner: 

1. Background: This section discusses the previously introduced SOIR and 
Directed Graph Model (DGM) for text IR and toxic content classification. 

2. Proposed work: This Section provides the understanding of the proposed Deep 
learning model for classifying hate speech, offensive language and normal text. 

3. Results analysis: This section contains the experimental analysis of the proposed 
model and the different parameters have been provided. 

4. Conclusion: In this section, finally the work is concluded and future extension of 
work has been suggested. 

 

2. Background  

This section provides insight into SOIR model and previously developed DGM 

classification model. 

 

A. SOIR Model 
This model is aimed to improve the IR model for running time and relevancy. Due to 

the large amount of data, a significant amount of time is required to locate the precise 

information. In the documents database, a number of documents are present which do 

not belong to the similar category and content. This nature of database increases search 

74



A Deep Learning Model for Classifying the Hate and Offensive Language in Social Media Text 

 

 

space. Therefore, some improvements have been made to enhance the user query 

representation and subjective data categorization. The query optimization involves a 

semantic model to recognize similar words to optimize the query.  

First, we pre-process the data to improve the quality of data and exclusion of noise by 

removing stop words and special characters. Next, feature selection techniques are used 

to reduce data dimensions and speed up the search process. Thus, the TF-IDF is used 

[5], to compute the weight W for identifying important tokens. The fixed size of the 

feature i.e. 30 tokens is considered. Further, FCM clustering is being used to categorize 

feature vectors [6]. The clustering results are organized as a list of features: 

 

𝐹 =< 𝐹𝑛 , 𝑘1,2,…𝑛, 𝐶 > 
 

where, F is the feature set, 𝐹𝑛 is the file name or index, 𝑘1,2,…𝑛 is the list of keywords, 
and C is the class name or subject. The training feature vector F is stored in a database.  

The categorized features are helpful for efficient data retrieval. On the other hand, the 

user query is transformed into a vector Q as a set of keywords: 

 

𝑄 = {𝑞1, 𝑞2, … , 𝑞𝑘 } 
 

In order to optimize the query, we initialize a set of queries by using synonyms. 

Additionally, a map is prepared that contains the keywords and the synonyms. In this 

algorithm, a single keyword is twisted multiple times to generate new queries using 

similar words. The different search query increases the chances of finding accurate data. 

The search process is developed on the basis of the k-NN (k-nearest neighbour) 

algorithm [7]. The distance between the query and data less than 0.25 is counted as the 

result. The developed SOIR model has been compared with the Cosine similarity-based 

and k-NN-based IR models. The precision, Recall and F-Score are calculated and 

visualized in figure 1. 

 

 
Figure 1. Mean Performance of SOIR Model 

75



Nidhi Bhandari, Rachna Navlakhe and G.L.Prajapati 

 

In this diagram, X-axis shows the measured matrix, and the Y-axis shows the precision, 

Recall and F-Score. According to the results, we found that the technique improves the 

precision, recall and f-score as we utilizing a large size of learning data. 

 
B. SOIR for toxic content classification 

In order to keep clean social media, we need an accurate model to identify malicious 

and toxic posts. We used an SOIR model for this task [4]. We summarized the negative 

emotions identified in table 1 below, which consists of Emotion classes and the relevant 

flow of emotions [8]. There are five main classes, and each class consists of its own 

subclasses. Thus, we need a multiclass classification system. So, we solve this problem 

by using the SOIR-based model. 

 

Table 1. Emotion Classes 

Emotions  Content 

Distressed Sad, Disappointed, Guilty, Missed 

Surprised Surprised 

Fearful Panic, Frightened, Shy 

Angry Angry 

Disgusted Dissatisfied, Annoyed, Doubtful, Hateful 

 

In order to train the model, we have collected more than 3329 tweets and categorized 

them. Then we considered 2002 tweets for the experiment. The pre-processing has been 

applied for removing tags, special characters, and stop words. Further, part of speech 

tagging (POS) is used to understand the lexical structure [9]. The POS tags and 

sentiment classes are used to prepare a new dataset. On the other hand, tweets are also 

processed using the TF-IDF weights. The top 20 weighted tokens are picked and used. 

In the next process, we prepared thresholds using the tweets' POS tags. In the first step, 

we compute the mean feature of each feature group using: 

𝑀𝐹 =
1

𝑁
∑ 𝐹𝑖

𝑁

𝑖=1

 

Where 𝑀𝐹  is the mean of the feature, N is the total number of samples in the groups. 
After measuring the mean value, we compute the distance from each point as the limit. 

𝐿𝐹 =
1

𝑁
∑ |𝐹𝑖 − 𝑀𝐹 |

𝑁

𝑖=1

 

Thus the threshold of the feature can be defined as: 

𝑇𝐹
𝐺𝑟𝑜𝑢𝑝 = 𝑀𝐹 ± 𝐿𝐹 

76



A Deep Learning Model for Classifying the Hate and Offensive Language in Social Media Text 

 

 

Using this equation, we calculate the feature map FM for all the POS features. After 

that, we append features of the TF-IDF. Here the FCM clustering is used for creating 

the dictionary learning. We are just using the membership function of FCM for 

preparing the dictionary. The membership between data instance 𝑖 and centroid j is 
measured using: 

𝜇𝑖,𝑗 =
1

∑ (
𝑑𝑖,𝑗

𝑑𝑖,𝑘
)

2

𝑚−1
𝑐
𝑘=1

 

According to this process, the emotion-labeled data is being used from each group. A 

tweet is tokenized and then inserted into the dictionary D with the TF-IDF weight. If a 

token exists in the dictionary, then we update the weight of the token and if the token is 

not available, then we simply insert the token. In order to compute the updated weight, 

we used: 

𝑁𝑒𝑤𝑊𝑒𝑖𝑔ℎ𝑡 = 0.5 ∗ 𝑂𝑙𝑑𝑤𝑒𝑖𝑔ℎ𝑡 + 0.5 ∗ 𝜇𝑖,𝑗  

where, 𝜇𝑖,𝑗 is the membership between previous and new weight for the token. 

The dictionaries are preserved in a database. Now we classify new tweets. For the 

classification, 50% of training samples and 50% of new tweets from Tweeter have been 

used. Then, we perform the POS tagging of test data. The tagged feature vector 𝑇𝑣 is 
used to compare with the thresholds. For both pattern matching and decision making 

here the mean value 𝑀 and limit 𝐿 are used for computing upper threshold 𝐿𝑚𝑎𝑥 and 
lower threshold 𝐿𝑚𝑖𝑛.The patterns between these two limits are used to compute the 
distance among queried tweets. POS tag feature and all the features threshold are used 

from feature map FM. Finally the higher matched value based class label is predicted as 

decision. After lexical pattern based decision making, we use the semantics for 

categorizing a tweet as final class label. Thus, the tweets are tokenized to regenerate it, 

as in SOIR. Let us have a tweet such that: 

𝑇𝑊 = {𝑘1, 𝑘2, … , 𝑘𝑛} 

After regeneration, we get: 

𝑇𝑊𝑚,𝑛 = {

𝑘1,1, 𝑘1,𝑠, … … … … . , 𝑘1,𝑛
𝑘2,1, 𝑘2,2, … … … … , 𝑘2,𝑛

𝑘𝑚,1, 𝑘𝑚,2, … … … … , 𝑘𝑚,𝑛

 

After preparing the set of similar keywords, we train the model using keywords and 

relevant weights for all the classes. The trained model can be defined as: 

𝐷𝑜,𝑝 = {

𝐷1 = {[𝑇1,1, 𝑊1,1], [𝑇1,2, 𝑊1,2] … . . , [𝑇1,𝑝, 𝑊1,𝑝]}

𝐷2 = {[𝑇2,1, 𝑊2,1], [𝑇2,2, 𝑊2,2] … . . , [𝑇2,𝑝, 𝑊2,𝑝]}

𝐷𝑜 = {[𝑇𝑜,1, 𝑊𝑜,1], [𝑇𝑜,2, 𝑊𝑜,2] … . . , [𝑇𝑜,𝑝, 𝑊𝑜,𝑝]}

 

77



Nidhi Bhandari, Rachna Navlakhe and G.L.Prajapati 

 

This is a function that generates a virtual directed graph using a weight matrix for 

decision-making. The matrix in the form of a graph model describes the association of a 

tweet with the given sentiment dictionary. An algorithm is developed to get the class 

label of the tweet, as given in table 2.  

 

Table 2. DGM based Classification 

Input: set of semantically similar twits 𝑇𝑊𝑚,𝑛, trained dictionary model 𝐷𝑜,𝑝 

Output: Class label C 

Process: 

1. 𝑓𝑜𝑟(𝑖 = 1; 𝑖 < 𝑚; 𝑖 + +) 
a. 𝑓𝑜𝑟(𝑗 = 1; 𝑗 < 𝑛; 𝑗 + +) 

i. 𝑖𝑓 (𝐷𝑖 . 𝑐𝑜𝑛𝑡𝑎𝑖𝑛𝑠(𝑇𝑊𝑖,𝑗 )) 

1. 𝑊𝑖 = 𝑊𝑖−1 + 𝑊𝑖,𝑗 

ii. 𝑒𝑛𝑑 𝑖𝑓 
b. 𝑒𝑛𝑑 𝑓𝑜𝑟 

c. 𝐷𝑊𝑖 =
𝑊𝑖

𝑚
 

2. End for 
3. 𝐶 = 𝑔𝑒𝑡𝑀𝑎𝑥𝑉𝑎𝑙(𝐷𝑊𝑖 ). 𝑐𝑙𝑎𝑠𝑠𝑙𝑎𝑏𝑒𝑙 
4. Return C 

 

The above algorithm searches each word in the dictionary and the relevant weights are 

aggregated. A greater value of weights is used as the final sentiment label. In order to 

make final decision, we use a function as: 

 

𝑓(𝐵, 𝐶) = {
𝐵 = 𝐶                                     𝑡ℎ𝑒𝑛  𝐶
𝐵! = 𝐶       𝑖𝑠 𝐵𝑖−1 == 𝐶      𝑡ℎ𝑒𝑛 𝐶

 

 

Finally, if we find the Class label C belongs to shy, panic, sad and guilty then we label 

the tweet as nontoxic else C returns the toxic labels. This model is named as Directional 

Graph Model. Figure 2 below shows the precision, recall and f-score of different 

techniques. The results indicate the DGM shows better accuracy as compared to other 

models. However, the DGM performance is higher but the SOIR is consistent as 

compared to DGM. 

78



A Deep Learning Model for Classifying the Hate and Offensive Language in Social Media Text 

 

 

 

Figure 2. Performance summaries of classification models 

 

Secondly, we measured the required time for training and reported it in figure 3. 

According to the results, the cosine-based and k-NN based techniques are winners. The 

main advantage of this model is that we can preserve the previously trained model for 

future use. Additionally, new data will also learn continuously. The model's 

performance replicates the efficient and accurate identification of toxic tweets. 

 

 

Figure 3. Time consumption 

0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

Precision Recall F-Score

P
e

rf
o

rm
a

n
ce

 (
%

)

Parameters

cosin KNN SOIR DGM

79



Nidhi Bhandari, Rachna Navlakhe and G.L.Prajapati 

 

3. The proposed work 

Emotion classification has become a common and classical domain of research and 

development but still there are many opportunities available for discovering something 

new. Initially, we started our work with the study of text mining and information 

retrieval. But due to new kinds of problems associated with text mining attracted us to 

classify social media text for finding toxic content. In order to design an effective 

method, we have implemented a DGM approach which is accurate and efficient for 

modelling negative emotions. But during the investigation, we found the following 

issues we tried to resolve in this study. 
1. Social media content these days is full of offensive language even during formal 

communication, comments and others. Therefore, there is a need of distinguishing 

between the content as normal and toxic. 

2. The previous experiments have been done with a small size of samples, but social 
media involve big data problems. Therefore, a previously defined model is 

reframed using the Deep learning technique to deal with a large amount of data. 

In this context, we have obtained a dataset of Twitter from Kaggle. The dataset is hate 

speech and offensive language dataset. The dataset is distributed in three class labels 

hate speech, offensive and neither. The dataset involves 1430 instances as hate speech 

post of twitter, 19190 tweets as the offensive language and 4163 instances as the normal 

text. The dataset consists of 24783 instances of data. Among them, we make a split of 

18587 instances for training and 6196 for testing. However, this dataset contains more 

data than the dataset on which we trained our previously designed models. 

Therefore, initially, we have employed our DGM model for extracting features. During 

this, we found a significant drawback of this method, which we tried to demonstrate in 

figure 4. Figure 4 shows the time consumed during the experiment with only 50% of 

entire dataset for constructing the required features for DGM technique. However, the 

DGM based technique has accurate results but having huge time complexity for training 

with a large amount of data. 

 

Figure 4. Time consumed with DGM 

0

500

1000

1500

2000

2000 3000 5000 7000 8000 9000 12000

T
im

e
 i

n
 S

e
c

Dataset SIze

Time Consumed

80



A Deep Learning Model for Classifying the Hate and Offensive Language in Social Media Text 

 

 

 

Therefore we drop the idea of utilizing the traditional classification technique for 

processing such a huge amount of data. The main reason behind the time consumption 

is computation of TF-IDF features and then computing the feature representation, which 

increases the time consumption significantly. Additionally, the measured features that 

need to train with the machine learning algorithms consume a significant amount of 

time for training. Therefore, the idea of working with the large amount of data with the 

classical and traditional feature construction method is becoming expensive. Finally, we 

proposed to use the Deep Learning technique for classifying the hate speech and 

offensive text to deal with the large amount of data. However, the deep learning models 

are mainly developed for image classification, but these techniques can also be used for 

other relevant classification and prediction tasks. The proposed model consists of the 

following steps to conduct the experiments. 

 

Data preparation 

In any machine learning model, data preparation is an essential step of data analysis. We 

have placed the obtained data from Kaggle into our Google drive and use the Google 

Colab to process data using the different techniques of text mining and machine 

learning.  

 

Figure 5. Dataset samples 

In the first step, we load the dataset, the initial dataset samples are demonstrated in 

figure 5 with their relevant attributes. However, the dataset contains various attributes, 

but we are just utilizing the class labels and tweets here for data analysis. 

 

Data pre-processing 

In basic terminology, the data pre-processing is a technique by which we prepare the 

data to apply to the learning algorithm. Therefore in this context, we studied various 

NLP based text pre-processing techniques. From them, some essential techniques are 

applied for pre-processing of data. Here we applied the following pre-processing 

techniques: 

a. Punctuation Removal: In this phase, we removed all the Punctuations from the input 
dataset instances by analyzing each character of the input strings. 

81



Nidhi Bhandari, Rachna Navlakhe and G.L.Prajapati 

 

b. Tokenization: In this step, the aim is to divide the tweet strings into chunks of 
words.  

c. Stop-word Removal: In this step, the aim is to remove English stop words. The stop 
words are those words which frequently occur in the text but not essential for 

domain identification such as, ''this', ''that', ''is', '''am' so on. In this context, we apply 

the Natural Language Toolkit (NLTK) library for stop word removal.  

d. Stemming and Lemmatization: Both are generated form of the same words. The 
difference is that stem might not be an actual word and lemma is an actual language 

word. Stemming usage an algorithm to which makes word processing faster. On the 

other hand, in lemmatization, we used WordNet to produce lemma, which is slower 

than stemming. 

 

 

Figure 6. Pre-processed data 

 

Features analysis 

In this phase, we transform the tweets into a structured manner. So, we have used the 

following different ways to prepare four different set of feature vector.  

a. POS Tagging: In this phase, we utilized the NLTK based POS Tagger to process 
each string of the dataset. After that we prepared the vector for making it training 

and testing set. In this process we get 37 tags or features for training data. The 

transformed data is demonstrated in figure 7. 

 

Figure 7. Sample tagged data 

b. POS with reduced features: In this phase, the 37 features of NLP-POS tags are used 
for experimentation. In order to reduce the amount of features, we have used 

Principle Component Analysis (PCA) and reduced the features, finally we have 

getting only 20 features for applying the algorithms. 

82



A Deep Learning Model for Classifying the Hate and Offensive Language in Social Media Text 

 

 

c. Term Frequency – Inverted Document Frequency (TF-IDF): In this feature vector, 
we utilized TF-IDF Vectorizer to extract weighed keywords from the data. The 

dimensions of data are kept maximum 5000 keywords. The extracted features are 

demonstrated in figure 8. 

 

Figure 8. extracted TF-IDF features 

d. Combined features: In this phase, we just combined TF-IDF features, and POS 
Tagged Features in a vertical manner to construct total of 5037 columns and 24783 

rows of data. 

 

Applying CNN 

Here, we trained the high-dimensional data using Convolutional neural network (CNN). 

Therefore, we have developed a simple three-layer sequential model. First layer is 

configured with the total number of attributes in training features, and activation 

function '''ReLu' is used. The second hidden layer consists of 10 neurons and the 

activation function '''Sigmoid' has been used. The final layer contains three output 

neurons and 'Soft-'max' activation function. The model takes very less amount of time to 

train the CNN model. The training time of the input training vectors has given in figure 

9. The model will be trained in a few times even the combined features, which consist 

of 5037 columns and 24783 data. Therefore, we can use the large data for learning with 

deep learning techniques. 

 

Figure 9. Training time 

0

50

100

150

200

250

TF-IDF POS POS F20 Comined

T
ra

in
in

g
 T

im
e

 i
n

 S
e

c

Training Time

83



Nidhi Bhandari, Rachna Navlakhe and G.L.Prajapati 

 

 

Performance analysis   

In this phase, we evaluated our model for obtaining the performance matrix for learning 

as well as the validation of the model. Therefore, the training accuracy, validation 

accuracy, training loss and validation loss of the model have been measured. These 

performance parameters are discussed in next section. In addition, the class-wise 

performance report has also been prepared. 

 

4. Results 

This section details the conducted experiments and the obtained performance results. In 

deep learning, the accuracy metric is used to measure the 'algorithm's performance in an 

interpretable manner. The accuracy is determined in the form of a percentage. It 

measures how accurate our model's prediction is with the true data.  

 

 

Figure 10. (A) Training Accuracy (%) 

 

In this experiment, we have used four different set of feature vectors namely TF-IDF 

based weighted feature vector, second the POS tagged features, third we have reduce 

the size of POS tagged data for classification and finally the combined features of POS 

and TF-IDF has been employed with the CNN model. The performance of the model 

with the different sets of selected features in terms of accuracy (%) has been reported 

using figure 10(A). Additionally, the accuracy for validation or test dataset has been 

demonstrated in 10(B). According to the obtained consequences in terms of these two 

parameters, we found that the TF-IDF based and the combination of TF-IDF and POS 

84



A Deep Learning Model for Classifying the Hate and Offensive Language in Social Media Text 

 

 

produces a higher performance than the other two algorithms. During training, we found 

the POS tagged data shows approximately 74.5 to 75.2% performance. 

 

 

Figure 10. (B) Validation Accuracy (%) 

On the other hand, when we used the reduced 20 features with the POS tagged data then 

we found a small improvement and we get an accuracy of 75.5%, which is higher than 

the POS tagged complete dataset, which has 36 features.  

 

 
Figure 10. (C) Training Loss  

 

On the other hand, when we train the model with the same data but TF-IDF features, 

model returns the total of 98.8% accurate results in consistent manner. And when we 

apply the combination of POS and TF-IDF features, we get improved results up to 

85



Nidhi Bhandari, Rachna Navlakhe and G.L.Prajapati 

 

99.8%. However, when working with the validation set, the performance of algorithm 

has dropped continuously. In figure 10(B), we can see the performance of all the feature 

based technique has been decreased significantly. 

 

 

Figure 10. (D) Validation Loss 

 

In addition, we have also computed the loss for training and validation. The loss 

function is used to optimize a machine learning algorithm. The loss is calculated on the 

training set and validation set and its interpretation is based on how well the model is 

doing in these two sets. It is the sum of errors made for each sample in training or test 

sets. Loss value implies how poorly a model behaves after each round of optimization. 

A loss is a number indicating how bad the 'model's prediction was on a sample. If the 

'model's prediction is perfect, the loss is zero; otherwise, the loss is greater. The goal of 

a model is to find the set of weights and biases that have low loss. Higher loss is worse 

for any model. A loss is not a percentage. It is a sum of the errors made for each sample 

in training or validation. 

 

 

Figure 11. shows the loss measuring example 

86



A Deep Learning Model for Classifying the Hate and Offensive Language in Social Media Text 

 

 

Figure 11 shows two graphs demonstrating the losses of two different models, the left 

graph has a high loss and the right graph has a low loss.    

• The arrows represent a loss. 

• The blue lines represent predictions. 

The loss of the proposed deep learning model has demonstrated in figure 10(C) and 

10(D). The figure 10(C) shows the training loss and 10(D) shows the validation loss of 

the model. In the diagram 10(C), we found that the loss of training reduced and become 

consistent. On the other hand, 10(D) demonstrates the loss during the validation of the 

model. The performance indicates when number of epoch has increases the loss of 

validation has also been increases in the case of TF-IDF based features learning, as well 

as also for the combined features, but the for combined features the loss function is less 

incline. 

On the other hand for POS tagged data we can see the model produces similar or 

consistent performance in all the cases. Finally, for more simple understanding we have 

also measured the classification report. 

 

 

The classification report for the implemented techniques has been given in table 3.The 

class-wise performance analysis is demonstrated in the table 3. That demonstrates the 

precision, recall and F1-score. Additionally, shows the accuracy, Marco avg, and 

weighted avg. according to these values we can say the combined and TF-IDF based 

features are providing higher accurate results. The accuracy has also visualized in figure 

12. According to the given results of the implemented techniques, we found that the 

accuracy of POS based both the techniques i.e. entire features and top 20 features based 

approach demonstrate the similar accuracy of 77%. On the other hand the TF-IDF and 

combined features based technique provide the 82% and 83% of accuracy. Thus we can 

say the combined features based classification approach provide higher accurate results. 

 

 

5. Conclusions 

Social media is one of the user-centric and user-generated data source. The data may 

have a significant amount of knowledge, challenges and opportunity for diverse areas of 

Table 3. Performance summary of the models  
 POS with 20 feature POS Tagged TF-IDF Combined (TF-IDF + 

POS) 

 Precisi

on  

Recal

l  

F1-

sscor
e  

Precisio

n  

Recal

l  

F1-

sscor
e  

Precisio

n  

Recal

l  

F1-

sscor
e  

Precisio

n  

Recal

l  

F1-

sscor
e  

Class 0 0. 0 0.0 0.0 0.0 0.0 0.0 0.25 0.31 0.28 0.23  0.29 0.26 

Class 1 0.78 0.99 0.87 0.77 0.99 0.87 0.90 0.89 0.89 0.90 0.89 0.90 

Class 2 0.51 0.06 0.51 0.42 0.05 0.09 0.70 0.70 0.70 0.74 0.72 0.73 

Accuracy  0.77 0.77 0.82 0.83 

Marco avg  0.43 0.35 0.43  0.40 0.35 0.32 0.62 0.63 0.62 0.62 0.63 0.63 

Weighted 

avg 

0.69 0.77 0.69 0.67 0.77 0.68 0.83 0.82 0.83 0.84 0. 83 0.83 

87



Nidhi Bhandari, Rachna Navlakhe and G.L.Prajapati 

 

applications. However, there are a number of opportunities hidden in the social media 

data, but toxic intension based data contains can create the biggest issues in healthy 

social media surroundings. In this context, a DGM model is analysed, which is limited 

to working with a small amount of data. Therefore, in this paper, a deep learning Model 

is proposed for classifying hate speech, offensive language and the normal social media 

posts. The finding says the deep learning techniques are better for classifying large 

datasets. 

 

 

Figure 12. Comparison of Mean Accuracy 

 

The DGM model utilizes the classical techniques for accurate classification of text but a 

limitation of the model motivates us to work with deep learning models. The proposed 

model includes the text feature analysis techniques and CNN algorithm for providing an 

accurate text classification technique. This model aims to differentiate between hate 

speech, offensive text and normal text by only considering text in social media post. 

During the experiments we conclude the following observations: 

1. DGM is accurate but is limited to work with the less dimensional data on the 
other hand the deep learning is effective for learning with large scale data  

2. Deep learning technique provides relevant or higher accuracy in less amount of 
time which is proved in the obtained performance.  

3. In classification of offensive text and hate speech text, the key words are more 
essential then the sentence structures or sentence formation.  

4. POS tagger is less effective for representing the text features to model difference 
between offensive and hate speech text. 

5. The combination of keyword based features and POS tagged data can improve 
the performance significantly. 

0,74

0,75

0,76

0,77

0,78

0,79

0,8

0,81

0,82

0,83

0,84

POS POS F20 TF-IDF TF-IDF+ POS

A
cc

u
ra

cy
(%

)

Methods

Accuracy

88



A Deep Learning Model for Classifying the Hate and Offensive Language in Social Media Text 

 

 

This study is helpful as a tutorial to demonstrate how the classical machine learning 

techniques are accurate but can be expensive to solve real-world problems. On the other 

hand, this study also tried to demonstrate how we can make use the deep learning 

techniques for processing the text data and get higher classification accuracy. In the near 

future, the following directions can be helpful for the extension of this study: 

1. The deep learning techniques involve more complex data models, which will 
provide more accurate data analysis. Among them the LSTM has become much 

more popular. For extension, we can use them in future implementations. 

2. The combination of two different kinds of features increases the dimension of 
the data to be trained. At the same time, in the selected TF-IDF features, various 

non-meaningful keywords are available. Therefore, the need is to design an 

enhanced semantic keyword selection technique for feature representation. 

 

References 

[1] H. Wang, Q. Zhang, & J. Yuan, "Semantically Enhanced Medical Information 

Retrieval System: A Tensor Factorization Based Approach”, IEEE, 2169-3536, 

2017. 

[2] S. Bergamaschi, E. Domnor, F. Guerra, M. Orsini, R. T. Lado, Y. Velegrakis, 

“Keymantic: Semantic Keyword-based Searching in Data Integration Systems”, 

Proceedings of the VLDB Endowment, ACM, Vol. 3, No. 2, 2010. 

[3] M. Chahal, “Information Retrieval using Jaccard Similarity Coefficient”, 

International Journal of Computer Trends and Technology, Volume 36, Number 3, 

June 2016. 

[4] N. Bhandari, R. Navlakhe, G. L. Prajapati, “Semantic Query Optimization based 

Information Retrieval Technique”, The Journal of Oriental Research Madras, 

ISSN:0022-3301,Vol. XCII-LXXVII, 2021 

[5] P. Bafna, D. Pramod, A. Vaidya, “Document Clustering: TF-IDF approach”, 

International Conference on Electrical, Electronics, and Optimization Techniques, 

IEEE, 2016. 

[6] R. K. Roul, J. K. Sahoo, K. Arora, “Modified TF-IDF Term Weighting Strategies 

for Text Categorization”, IEEE, 2017. 

[7] J. Chen, X. Tang, “Ensemble Of Multiple K-NN Classifiers For Societal Risk 

classification”, J Syst Sci Syst Eng, 26(4): 433-447, Aug 2017. 

[8] H. Xu, W. Yang, J. Wang, “Hierarchical emotion classification and emotion 

component analysis on chinese micro-blog posts”, Expert Systems With 

Applications, 42, 8745–8752, 2015. 

[9] C. Pasquier, C. da Costa Pereira, A. G. B. Tettamanzi, “Extending a Fuzzy Polarity 

Propagation Method for Multi-Domain Sentiment Analysis with Word Embedding 

and POS Tagging”, ECAI, The IOS Press, 2020. 

89