Abstract:

It is essential for computers to understand the emotions conveyed in views, comments, and
textual conversations. Artificial Intelligence (AI) and Natural Language Processing (NLP)
techniques have revolutionized human-computer interaction by enabling computers to
comprehend human languages. Emotion detection is a challenging task, but it is becoming
increasingly important in a variety of applications, such as customer service, healthcare, and
social media monitoring. Additionally, there is a need to address the vulnerabilities of deep
learning models to adversarial attacks in the field of emotion detection in text analysis. This
paper proposes a novel approach Adversarial Emotion Network (AdvEmoNet) for the
classification of emotion. The proposed approach addresses the limitations of existing
approaches by using adversarial training to make the model more robust to adversarial
attacks. The proposed approach was trained on multiple datasets and was shown to
outperform state-of-the-art methods. The results suggest that the proposed approach has
shown improved classification accuracy. The proposed method using the AdvEmoNet
method exhibits an improvement of 17.35% accuracy compared to existing approaches.

Keywords: Emotion detection, Natural Language Processing, Adversarial transfer learning,

Text analysis, Convolutional neural networks, Deep learning.

1. INTRODUCTION
In today's world, the internet is constantly adding new data, which leads to an increase in the data
size ranging from terabytes to petabytes. The definition of "emotions" is a intricate interplay of
various subjective and objective factors, which are influenced by both neurological and hormonal
systems. Emotions encompass an individual's personal experience and interpretation of the feeling, as
well as external factors that may trigger or influence the emotion. In the brain, emotions are processed
through a network of interconnected regions, such as the amygdala, prefrontal cortex, and insula.
Additionally, hormonal systems, such as the release of cortisol and adrenaline, can impact emotional
responses, particularly during stressful situations. Overall, emotions are a complex phenomenon that
involves a combination of subjective and objective factors, along with neurological and hormonal
processes [1].
Emotions play a crucial role in cognition and behavior, triggering cognitive processes and
physiological changes in response to external stimuli. They can be classified into positive, negative,
and neutral categories. Emotion recognition from web text documents poses challenges due to the
absence of explicit emotional terms. Sentiment and emotions are often used interchangeably. Text
classification, involving supervised machine learning methods like SVM, decision trees, and deep
learning techniques such as LSTM, is used to categorize emotions in web text documents. The use of
adversarial methods aids in generalizing classifiers across different domains. Paul Ekman identified
emotions like surprise, anger, joy, disgust, fear, and sadness[2]. Emotions in interpersonal
relationships are crucial for communication [3]. Emotions in written text, audio, and video can be
recognized in computer science[4]. Wen and Wan used a labeled dataset to classify positive and
negative emotions based on a threshold value [5]. The dimensional technique can handle complex
emotional situations, whereas the categorical approach has limitations in capturing nuanced emotional
experiences[6]. Computational techniques for emotion extraction include rule-based, ML, keyword-
based, corpus-based, and DL methods [7] [8].
The ML algorithms, particularly supervised learning, are commonly used for emotion
classification[9]. Emotion recognition from web documents using machine learning-based methods is
a significant area of research [19]. Emotion classes are assigned to text using both supervised and
unsupervised learning algorithms [20]. Deep learning (DL) techniques allow algorithms to learn
hierarchically, improving their ability to grasp new concepts [28]. This has various practical
applications, including understanding customer emotions for business strategies on social media
platforms like Youtube, Twitter, and Facebook, analyzing patient mental health in the healthcare
sector, and detecting issues like fake news and hate speech through emotion detection in text.
Common text classification methods used in research include support vector machines, decision trees,
random forests, convolutional neural networks, long short-term memory, and bi-directional LSTM
[28].
Adversarial training helps the classifier become more resilient, decreases sensitivity to domain-
specific noise or biases, and increases accuracy on target domains where labelled data may be hard to
come by or expensive to acquire. Additionally, adversarial approaches provide higher generalisation
and transfer learning capabilities by solving the domain shift problem, which eventually results in
superior text classification performance in real-world contexts. In the paper, we use a framework
AdvEmoNet and CNN for the classification of web text documents.
The structure of the work presented in this paper is mentioned as containing emotional and non-
emotional documents as follows. Section 2 is about the previous works related to emotion and
emotion categorization. We describe the concerned methodology and system architecture of our
proposed models in Section 3. And Section 4 presents the experimental outcomes of the proposed
models. Finally, section 5 discusses the conclusion.

2. BACKGROUND

The study of text-based emotion recognition has become important research work. In
psychological research, there are prevalent models related to emotions, such as the categorical
emotion model or discrete model, the dimensional model, and the appraisal-based model [10].
In the brain, multiple neural subsystems correspond to various emotions. Six fundamental
categories are used by Ekman's model to classify emotions [11-12]. Happy, anger, sadness, disgust,
surprise, and fear are some examples of these fundamental feelings. The eight emotions in Plutchik's
model are listed as joy vs. sad, contempt vs trust, fear vs. anger, and anticipation vs surprise [13]. The
dimensional model perceives emotional states as interconnected rather than separate from one another.
The model is represented in dimensional space, i.e., unidimensional and multidimensional, expressing
the relation between emotions and an event depending on the intensity (low to high) of the emotions.
Many dimensional emotion models employ two or three dimensions: "valence" (which shows the
emotions of positive or negative), arousal which represents the intensity of enthusiasm of emotion,
and dominance, a measure of how well one can manage emotions [14].

In the field of emotion encoding, Russell introduced the circumplex of effect, a significant
two-dimensional model where valence is represented on the vertical axis and arousal on the horizontal
axis, forming an emotional wheel [15]. This model is essential in understanding dimensional aspects
of emotions. The dimensional model, incorporating valence and arousal, can be combined with the
appraisal-based model to enrich emotion analysis [17]. This integration allows for a broader
methodology in studying emotions. Componential emotion models are integrated based on the
evaluation concept [16]. The appraisal hypothesis explains how the same experience can trigger
multiple emotions in different individuals and at different times [18]. Understanding the various
factors influencing emotions, such as knowledge, communication, physiology, motivation, movement,
and responses, is crucial [5].
 Hai Huan et. al [21] proposed an improved text classification technique using the CBM
(Convolutional and Bi-LSTM Models) approach. The CBM approach utilizes both global and local
semantic features extracted from text by employing the Glove model for text vectorization in the
embedding layer. Basiri, M. E. et al. [22] introduced the ABCDM model, an Attention-based
Bidirectional CNN-RNN Deep Model for sentiment analysis. The model utilizes pre-trained Glove
word embedding vectors as initial weights for the embedding layer. It employs bidirectional LSTM
and GRU. Thi-Thanh-Thuy Huynh and Anh-Cuong Le [23] proposed a CNN-based framework for
emotion categorization. They addressed the challenge of representing emotions in deep learning
methods by incorporating pre-defined grammatical patterns containing emotional data to extract
external features. Experimental results on Vietnamese emotion and ISEAR1 datasets demonstrated the
success of their approach.

Pradhan et al. [24] presented CLAVER, an attention mechanism-based system for recommending
scholarly venues. The model utilized a combination of convolutional layer, LSTM, and bi-LSTM to
make recommendations solely based on the title and abstract. Experimental results on the DBLP
dataset showed superior performance compared to state-of-the-art techniques in terms of
precision@k, nDCG@k, accuracy, MRR, and diversity and stability of recommended venues. Liang et
al. [25] proposed an improved Double Channel (DC) approach to enhance the performance of CNN-
LSTM models. The DC approach incorporated word-level and char-level embeddings simultaneously
on separate channels. The authors introduced a hybrid attention mechanism and trade-off learning to
improve the model's generalizability. Experimental results demonstrated that the proposed DC CNN-
LSTM model outperformed the basic CNN-LSTM algorithm in terms of accuracy and F1-measure.

Deepanway Ghosal et al. [26] introduced the Dialogue Graph Convolutional Network
(DialogueGCN) to improve context awareness for word-uttered emotion recognition in dialogues.
DialogueGCN achieved better performance than strong baselines and current state-of-the-art
techniques on three ERC benchmark datasets, achieving a weighted average accuracy of 65.25% for
identifying emotions in conversational dialogues. Trueman and Cambria [27] proposed a
convolutional stacked bidirectional LSTM model with a multiplicative attention technique for aspect
categorization and sentiment identification tasks. The model incorporated a convolutional layer to
extract high-level aspects and emotion features, along with two bidirectional LSTM layers to regulate
information flow. The fixed-width multiplicative attention technique was used to find context vectors
for the input sequence, leading to improved performance compared to other methods on the SemEval-
2015 and SemEval-2016 datasets.
Alsmadi et al. [28] provided an overview of adversarial attacks and defensive systems in
social media applications. The review focused on major challenges, research directions, and discussed
applications such as sentiment analysis, hate speech detection, clickbait & spam identification, rumor
detection, satire detection, and misinformation detection. The authors provided insights into the
vulnerabilities of machine learning models in these domains and highlighted the need for robust
defensive mechanisms to mitigate the impact of adversarial attacks. Hajek et al. [29] presented 2-deep
neural network models for detecting fraudulent customer reviews. These models incorporated n-gram,
skip-gram, and emotion models and demonstrated high effectiveness on larger datasets with mixed
polarities. The researchers utilized various emotion representations and pre-trained word embeddings
to enhance the models' accuracy, suggesting their applicability in real-world scenarios.

Alsmadi et al. [30] and Han and Zhang [31] focused on adversarial attacks in text processing
and sentiment analysis. Alsmadi et al. developed a pre-trained model to improve the resilience of text
generation models against adversarial attacks. Han and Zhang [31] explored the use of adversarial
training to enhance the robustness of emotion detection and sentiment analysis models. Both studies
highlighted the significance of recognizing and addressing vulnerabilities of machine learning models
to adversarial attacks in these domains.
A few shortcomings were identified and listed as follows:

1. Lack of accuracy in conventional construction models.

2. Limited consideration of adversarial attacks: The literature lacks effective strategies to address
adversarial attacks on text documents, which can manipulate the decisions of machine or deep
learning models and potentially compromise their performance and reliability.

This work proposes an approach to overcome the above shortcomings by adopting the following
steps:

 Feature extraction using pre-trained weights from the GloVe model: To enhance feature
representation, the model utilizes pre-trained weights from the GloVe model. By freezing and
transferring these weights to the CNN, the model benefits from the knowledge learned from a
large dataset. This improves its ability to capture important features relevant to emotion
classification.

 Integration of CNN architecture: The proposed model adopts a Convolutional Neural Network
(CNN) architecture for text classification and emotion analysis.

 Learning from a large dataset: The proposed model learns from a large dataset, which helps it
gain a deeper understanding of emotions expressed in text. By training on a diverse range of
samples, the model can better generalize its knowledge and improve its overall accuracy in
emotion categorization. Adversarial transfer learning is adopted to increase the robustness and
resilient of the model.

By integrating CNN architecture, utilizing pre-trained weights from the GloVe model, addressing
adversarial attacks, and leveraging a large dataset for training, the proposed model aims to overcome
limitations in the existing approaches. These strategies enhance the accuracy, robustness, and
generalizability of the model for emotion analysis in text.

3. METHODOLOGY
An overview of the proposed architecture of AdvEmoNet is as shown in Figure 1. The process
begins with the collection of a dataset from a social media platform. To prepare the text for
classification, several preprocessing steps are applied, including punctuation removal, tokenization,
lowercasing, stop word removal, and stemming. Once the text has been preprocessed, feature
extraction techniques are used to transform the text into numerical representations.
Two methods, namely one-hot encoding and GloVe embedding, are employed for feature
extraction. One-hot encoding represents each word as a binary vector. Each dimension corresponds to
a unique word in the dataset. On the other hand, GloVe embedding represents pre-trained word
vectors to capture semantic information once the features are extracted. The weights of the GloVe
embedding features are frozen. These frozen features, with one-hot encoded features and word
embeddings, serves as inputs to a CNN algorithm. It utilizes the extracted features, including the
frozen GloVe embedding features, to learn patterns and make accurate predictions based on the
provided dataset.
 Figure1: Proposed AdvEmoNet for classification of text.
3.1. Dataset
The embedding model used in this work is using training multiple datasets to ensure a comprehensive
understanding of emotions. These datasets include the go-emotions dataset [32], the Twitter dataset
[33], the Emotion Detection from Text dataset [34], and the Amazon review data [21]. By
incorporating diverse dataset are used to train the model to make the model robust and resilience.
The go-emotions dataset consists of 58,009 sentences extracted from Reddit, representing
both positive and negative emotions in a well-balanced manner. It covers 27 different emotions. The
Twitter dataset, comprises 1.2 million tweets, offers a broad collection of emotions, provides valuable
insights for sentiment analysis. The Emotion Detection from Text dataset includes 40,001 sentences
with labelled emotions. The Amazon review dataset contains 3.4 million reviews and offers a wide
range of emotions expressed by customers. These datasets were carefully divided into training and
testing splits to evaluate the model's performance are 50% - 50% , 60% - 40% and 80% - 20% as
shown in Table 4.

3.2 Text Pre-processing

The text preprocessing steps include tokenization, removal of prepositions, determiners,
punctuation, and stop words. Tokenization facilitated independent analysis of individual words, which
is vital for subsequent emotion detection tasks. To reduce noise and irrelevant elements to identify
meaningful sentences the prepositions, determiners, and punctuation marks were removed. These
elements typically contribute little semantic meaning in the context of emotion detection and were
deemed unnecessary for accurate analysis. Additionally, stop words, lacks specific semantic value,
were removed to streamline the data and enhance analysis efficiency. Porter’s stemming is applied to
reduce words to their base standardizing the text's terms and emotion category mapping and analysis
is done by standardizing the dataset. The number of stemmed words that are associated with each
category is calculated using Eq.(1). This pre-processing technique, transforms raw text data into a
cleaner and more manageable format to extract meaningful insights from the text.
Sw = Sw(t) * P(w) (1)
Where,
Sw - No. of stemmed words
Sw(t) - No. of stemmed words in text
P(w) - probability of word
3.3 Feature Extraction and Classification
This section discusses in detail the feature extraction methods phases to classify the text as
emotional and non-emotional using AdvEmoNet.
3.3.1 Feature Extraction using AdvEmoNet:
The first step in this process involves representing words using one-hot encoding. Each word is
represented by a vector with a dimension for every possible word in the vocabulary. In a one-hot
vector, the dimension corresponding to the word is set to 1, while all other dimensions are set to 0.
The GloVe model is used to map words to dense vectors in a high-dimensional space. These vectors
capture semantic information by encoding relationships between words. The GloVe model is trained
on large text corpora, allowing it to capture and incorporate rich semantic context. By utilizing the
GloVe model, raw text phrases are transformed into numerical vectors that encompass the semantic
context of the words. This conversion enables more meaningful analysis and interpretation of the text,
facilitating tasks such as sentiment analysis and emotion detection.
The computational steps involved in developing the AdvEmoNet approach using Glove and CNN
model:
1. Each word in the input text is represented as a sparse binary vector, with a dimension equal to
the size of the vocabulary.
2. In this vector, all elements are zero except for the index corresponding to the word, which is
set to one. This representation allows us to identify each word uniquely such as: one-
hot_vector(word) = [0, 0, ..., 1, ..., 0].
3. One-hot encoded word vectors are mapped into the d-dimensional space of the GloVe
embeddings. By multiplying the one-hot vectors with the pre-trained word embedding matrix
as in Eq.(2). The resulting vectors are dense, captures the semantic information of each word
in the input text.
word_embedding(word) = GloVe_matrix * one-hot_vector(word) (2)
Step 1: Computing Glove embedding matrix
Given a corpus of emotion-labeled text, a co-occurrence matrix is constructed. Each element
X(i, j) of the matrix represents the co-occurrence count of word i with word j in the corpus.
This matrix captures the relationships between words based on their co- occurrence patterns.
Step 2: Initialization of Word Vectors
The word vectors (w i ¿ and context word vector (wi) , are initialized with pre-trained Glove
word embeddings. These vectors are high-dimensional, dense representations of the words.
Step 3: Objective Function
The objective function is applied to an objective function as shown in Eq.(3):
J=∑ f ( X ( i , j ) ) ¿ – log ( X ( i , j ) )2 E(i,j) (3)
ⅈj

where:
J - cost function
X(i, j) - co-occurrence count between word i and word j
T
w i - target word vector for word i
T
w j - context word vector for word j
b I & bj - biases
 E(i, j) - binary indicator to determine words associated with emotion.
f(X(i, j)) - weighting function

Step 4: Adam optimization function

The word vectors and biases are adjusted and minimized the objective function.
Step 5: AdvEmoNet using the Glove model
The model is trained using the co-occurrence matrix and the modified cost function. Through
iterative optimization, the word vectors and biases are updated to capture the semantic
relationships between words and emotions.
Step 6: Embedding Matrix
In this, the optimized word vectors are extracted and used to construct an embedding
matrix. Each row of the embedding matrix corresponds to a word and its associated vector.
These vectors capture the learned semantic representations of words in the context of
emotions. Then the weights learned are frozen.
4. The feature extraction and word embedding for the sample input text :
1): "I feel happy and excited." 2): “I feel sad and disappointed.”
For the above feature vector are computed using Eq.(4). The vector values are as shown in
Table 1 and Table 2 represents for positive and negative emotions.
Fv = we1 we2 we3... we(n-1) we(n) (4)
Here, we1, we2, ..., wen represent the word embeddings of individual words in the input text.
Table1: Matrix Representation of Vectors for Positive Phrase.
One-hot GloVe Word
[1, 0, 0, ..., 0] [0.249, 0.785, -0.539, ..., 0.923] "I"
[0, 1, 0, ..., 0] [0.155, 0.614, 0.384, ..., 0.868] "feel"
Positive
[0, 0, 1, ..., 0] [0.096, 0.486, 0.623, ..., 0.795] "happy"
[0, 0, 0, 1, ..., 0] [0.433, 0.485, 0.693, ..., 0.823] "and"
[0, 0, 0, 0, 1, ..., 0] [-0.418, 0.298, -0.921, ..., 0.729] "excited"

Tabel 2: Matrix Representation of Vectors for Negative Phrase

One-hot GloVe Word
Negative [1, 0, 0, ..., 0] [0.249, 0.785, -0.539, ..., 0.923] "I"
[0, 1, 0, ..., 0] [0.155, 0.614, 0.384, ..., 0.868] "feel"
[0, 0, 1, ..., 0] [-0.182, -0.485, 0.615, ..., -0.792] "sad"
[0, 0, 0, 1, ..., 0] [0.433, 0.485, 0.693, ..., 0.823] "and"
[0, 0, 0, 0, 1, ..., 0] [0.454, -0.298, 0.919, ..., -0.728] "disappointed"

The feature vector is formed by concatenating the GloVe embeddings are calculated using Eq.(5).
feature_vector = concat(GloVe_embeddings(sentence)) Eq. (5)

3.3.2. CNN Architecture for classification of emotional text:

 The CNN model consists of two main parts: the convolutional layer and the fully connected
layer. The convolutional layer is responsible for detecting local patterns in the input feature vector,
while the fully connected layer learns the relationships between the extracted features and the classes.
The convolutional layer in the CNN model applies filters to the feature vector, generating feature
maps highlighting specific local patterns. These filters are learned during the training process to build
the model to identify the input patterns. Pooling layers further contribute to the model's effectiveness
by reducing the spatial dimensions by retaining essential information. This helps in condensing the
extracted features of the dataset samples.
Convolutional Neural Network (CNN) model with pre-trained GloVe word embeddings with
softmax classifier is used to classify the emotional text. The use of GloVe word embeddings allows
the model to capture the semantic relationships between words. The words in dense vectors represents
a high-dimensional space, GloVe embeddings enable the model to understand the context and
meaning of the input text.
Adam (Adaptive Moment Estimation) is used to optimize the algorithm in deep learning to
updates the weights of a neural network during training [35]. Momentum and adaptive learning rates
are used to accelerate the convergence of the model. Eq. 6 updates the weights of the network based
on gradients and adaptive learning rates. It adjusts the learning rate for each parameter individually,
enhancing the optimization process. Also Relu function [35] is adopted to map the negative values to
‘0’ is given from Eq.(11).
The Adam updates are calculated using Eq.(6):
^ t /( √ v^ t + ε )
θt =θt −1−α∗m (6)
Where,
mt & vt = first and second moments
mt = β1 * m{t-1} + (1 - β1) * gt (7)
vt = β2 * v{t-1} + (1 - β2) * gt2 (8)
gt = gradient of the weights at time t.
m̂ t = mt / (1 - β1t) (9)
v̂ t= vt/ (1 - β2t) (10)
β1 and β2 = decay rates for the first and second moments, set to 0.9 and 0.999, respectively.
α = learning rate, which determines the step size for weight updates.
θt = updated weights at time t.
ε = small constant (e.g., 10-8) added for numerical stability.
f(x) = max(0, x) (11)

Input layer: By utilizing the Adam optimizer and the ReLU activation function in the training
process, the emotion classification model can effectively optimize the network parameters and
introduces non-linearity, for features to improve the accuracy and performance.
Output layer: The softmax classifier [35] is applied final layer of the CNN to obtain the
probability distribution of emotional classes and non-emotional classes. By applying the softmax
function to the output scores, the model generates probabilities that sum up to 1, providing a reliable
measure of confidence for each class prediction using Eq. (12).
To classify softmax is represented as follows:
 exp( F−1)
P(emotioni | text) = (12)
∑ ¿ j exp(F −1) ¿
Where,

∑ ¿ j exp(F −1 ) ¿ - sum of the exponentials of all the feature vectors.

Collectively, these steps enable the CNN model to process the text data, extract meaningful
features using GloVe embeddings, identify local patterns and dependencies, to classify the text into
emotional and non-emotional classes. The combination CNN and Glove embedding techniques
ensures accurate emotion classification by leveraging the semantic information embedded in the text.
The following procedure is adopted to construct the CNN model at various layers [Eq.(13) – Eq.(10)]:
i) Convolutional layer:
The filters are learned during training and are applied to the feature vector to generate feature maps.
Each filter represents a specific pattern or feature that the model identifies in the input.
feature_maps = convolve(feature_vector, filters) (13)
ii) Max-pooling layer:
This layer reduces the spatial dimensions of the feature maps, by applying optimization function.
pooled_feature_maps = max_pooling(feature_maps) (14)
iii) Fully connected layer:
This layer learns the relationships between the extracted features from the pooled feature maps and
the classes. It computes scores or activations that indicate the model's confidence in each class.
class_scores = fully_connected(pooled_feature_maps) (15)
iv) Classification: activation function Softmax:
The softmax function converts the class scores into probabilities, providing a normalized distribution
representing the model's predictions for the input text.
probabilities = softmax(class_scores) (16)

3.4 Algorithm
The brief algorithm of the proposed approach encloses all the steps discussed above:

Step 1: Text Preprocessing

- Input: "I am feeling happy today!"
- Tokenize the text data: ["I", "am", "feeling", "happy", "today"]
- Perform data cleaning: ["feeling", "happy", "today"]
- Remove stop words: ["feeling", "happy", "today"]

Step 2: Dataset Split : The dataset is divided into 20% for testing and 80% for training.
- Training Dataset: ["I am feeling happy today!", "I don't have any emotions."]
- Testing Dataset: ["This is a neutral sentence.", "I feel indifferent."]

Step 3: Build Vocabulary (Bag-of-Words)

- Vocabulary: ["feeling", "happy", "today", "emotions"]

Step 4: Word Tokenization and Frequency Calculation

 - Tokens: [["I", "am", "feeling", "happy", "today"], ["I", "don't", "have", "any",
"emotions"]]
- Word Frequency: {"I": 2, "am": 1, "feeling": 1, "happy": 1, "today": 1, "don't": 1,
"have": 1, "any": 1, "emotions": 1}

Step 5: GloVe Word Embeddings

- Sample GloVe embedding: ["happy": [0.1, 0.2, 0.3, ...], "sad": [0.4, 0.5, 0.6, ...], ...]

Step 6: Co-occurrence Matrix and GloVe Probability Ratio

- Co-occurrence matrix example:
feeling happy today emotions
feeling 0 1 1 0
happy 1 0 1 0
today 1 1 0 0
emotions 0 0 0 0

Step 7: Embedding Matrix Construction

- Embedding Matrix:

Embedding Vector
feeling [0.2, 0.3, 0.4, ...]
happy [0.1, 0.2, 0.3, ...]
today [0.5, 0.6, 0.7, ...]
emotions [0.0, 0.0, 0.0, ...]

Step 8: Adversarial Training (AdvEmoNet) the weights are freezed learned from the Glove
embedding model and passed to the classification model CNN.

Step 9: Convolutional Neural Network (CNN) Training

- Design the CNN model architecture with convolutional layers, pooling layers, fully
connected layers, and softmax activation for classification.
- Initialize the weights from the AdvEmoNet to CNN model.and obtain prediction.
- Compute the loss between the predicted labels and the ground truth labels.
- Perform backpropagation to calculate the gradients with respect to the weights.
- Update the weights of the model using an optimization algorithm (e.g., Adam).
- Repeat the training process for multiple epochs to improve the model's performance.

Step 10: Evaluation and Testing

Input: "I am feeling ecstatic after winning the competition!"
Predicted Output: Emotion (based on the trained model)
Input: "The weather today is quite pleasant."
Predicted Output: No emotion (based on the trained model)

4. EXPERIMENTAL SETUP AND EVALUATION:

Python IDE 3.6 [21] is used for simulating the experiment. The 200-dimensional GloVe word
vectors [54] are utilized for word embedding. Deep learning frameworks, specifically Keras and
TensorFlow, are employed for model development and training. Keras and TensorFlow played a
pivotal role in the research pipeline, offering a comprehensive set of tools and functionalities for
effective model development and optimization. The results are evaluated by assessing various
performance metrics and comparing the obtained results with those of several existing works.
4.1 Parameters used
 The performances of the model are accessed using various evaluation metrics such as accuracy,
precision, recall, and F1-score.

Precision(P) : the ratio of true positives (TP) to the sum of true positives and false positives (FP)
[41]. It is often used to evaluate the performance of a binary classification model. The formula for
precision can be expressed as follows:

TP
P= (17)
TP+ FP

Recall(R) sensitivity: performance metric commonly used to evaluate a binary classification

model [41]. It measures the proportion of actual positives that were correctly identified by the model.
This is done by calculating the ratio of true positives (TP) to the sum of true positives and false
negatives (FN). The formula for recall is as follows:

TP
R= (18)
TP+ FN

F1 Score: performance metric in binary classification tasks, providing a balanced evaluation of a

classifier's precision and recall. It is the harmonic mean of precision and recall and is used to assess a
model's ability to correctly identify positive instances while minimizing false positives and false
negatives. The F1 score ranges between 0 and 1, with higher values indicating better classifier
performance. It is particularly valuable in scenarios with imbalanced class distributions or when both
precision and recall are equally important evaluation criteria. The formula to calculate the F1 score is:
2∗Precision∗Recall
F 1 Score= (19)
Precision+ Recall
Where,
TP -True positive, FN – False negative and FP – False positive.
4.2 Results and performance analysis:
We evaluated the accuracy, precision, recall, and F1-score to comprehensively acesses the
effectiveness of our approach. The results, presented in Table 3, depict the performance of various
approach applied to the challenge of the same Amazon review dataset.
The first result in Table 3 represents the work done by the researchers [42]. They have used the
BAC model (Bi-LSTM Self Attention-based Convolutional Neural Network), for text classification.
By leveraging the power of bidirectional LSTM and self-attention mechanisms, they achieved
accuracy of 78% and an F1-score of 0.83. This model effectively captured the contextual information
and long-range dependencies in the text data, contributing to its impressive performance. In the
second result in Table 3, the researcher represents the adopted combination of Glove word
embeddings and Convolutional Neural Network (CNN) for text classification [43]. This approach
harnessed the semantic information captured by the pre-trained Glove embeddings and the local
pattern detection capability of CNN. The authors reported an accuracy of 79.5% for this model,
showcasing its effectiveness in capturing relevant features for classification tasks. However, the F1-
score is not reported, limiting the comprehensive evaluation of its performance.
The third result in Table 3 represents CNN architecture for text categorization[44]. The authors
achieved an impressive F1-score of 0.806 and an accuracy of 81%. The CNN model captured
effectively the relevant information and learned discriminative features for accurate categorization.
The fourth result in Table 3 represents work done by the researcher using Word2vec embeddings
combined with a CNN for text classification[45]. Word2vec embeddings capture semantic
relationships between words, and when integrated with CNN, enable effective classification. This
approach achieved an accuracy of 68%, highlighting the potential of combining distributional word
representations with convolutional operations. However, similar to the second study, the F1-score for
this approach was not reported, limiting the comprehensive assessment of its performance.
Table3: Comparison of Proposed work with Similar Works in Literature

Sl. Approaches Accuracy F1 score

No.
1 Self attention based[42] 0.78 0.83
2 Glove+CNN[43] 0.79 -*
3 CNN[44] 0.8130 0.806
4 Word2vec+CNN[45] 0.68 -*
5 Proposed solution : AdvEmoNet 0.97 0.98
*- data not provided in the literature

Our proposed solution exhibits better performance than the existing works in terms of both accuracy
and F1-score. We are able to achieve a better accuracy of 98% and an F1-score of 0.99, demonstrating
the robustness in learning activity by our model. Since our proposed approach combines the strengths
of advanced architectures, efficient feature extraction techniques, and data preprocessing has
influenced to achieve better classification accuracy for different training and testing ratio of the data.
Table 4 shows the training and testing dataset splits for the Amazon dataset with their performance
accuracies.
Table 4: Result of Training and Test Data set

Sl.No Dataset Split Classification Precision Recall F1 Accuracy

Of Text Score
1 Training: 50% Non-Emotional 0.86 1.00 0.92 97
Testing : 50% Emotional 1.00 0.97 0.98
2 60% training Non-Emotional 0.85 1.00 0.92 97
40% testing Emotional 1.00 0.97 0.98
3 80% training Non-Emotional 0.87 1.00 0.93 98
20% testing Emotional 1.00 0.97 0.99

Overall, our findings highlight the efficacy of our proposed model and its potential for practical
applications in sentiment analysis and text classification tasks. The better results obtained in terms of
accuracy and F1-score emphasize that our approach is better than the existing approach and validate
the effectiveness of our methodology in capturing and understanding the details of textual emotions.
Figure 2 provides a comprehensive comparison between the standard Glove + CNN model and our
proposed approach AdvEmoNet for different datasets. While both models utilize Glove word
embeddings and CNN architecture, our AdvEmoNet approach incorporates adversarial training
method. The results clearly demonstrate the superiority of our model, with significant improvement in
accuracy, precision, recall, and F1-score metrics. By exposing the model to perturbed or noisy
examples during training, it learns to distinguish genuine emotional text from adversarial text, leading
to more accurate and reliable emotion detection. This enhanced resilience against adversarial
examples empowers our proposed model to outperform the existing method and paves the way for
more effective emotion detection systems.
 Figure 2: Pperformance of Proposed Preposed Aapproach on Dafferent Training datasets.

Figure 3 presents the comparison of our proposed model's AdvEmoNet performance on various
datasets with that of existing works. The effectiveness of our model is achieved through the
integration of adversarial training, in ensuring robustness and generalization. By employing the
adversarial method and fine-tuning the pre-trained model, we successfully mitigate the impact of
adversarial examples, resulting in significantly improved accuracy, precision, recall, and F1-score.
The experimental results demonstrate our model's proficiency in handling diverse and real-world data.

Figure 3: Comparison of various datasets performance and the Proposed approach.

5. CONCLUSION

The analysis of social media text presents an interesting possibility for advanced technology
research. Leveraging vast amounts of digital data from social media platforms, researchers employ a
diverse range of deep-learning algorithms to draw meaningful conclusions from textual content. In
our study, we focused on adversarial transfer learning for text classification in web text documents,
with the aim of accurately discerning between emotional and non-emotional text. Our proposed
approach AdvEmoNet achieved better results, boosting a remarkable 97% accuracy, precision of 1.00,
a recall rate of 0.97, and an F1-score of 0.98. Compared to previously published literature, our work
showcased a substantial improvement of 17.35% and also the advantage of using adversarial transfer
learning helps to capture the semantic features effectively. This research holds promising potential for
extending its capabilities to classify different levels of emotions and adapting to other deep-learning
models for classification tasks. Such advancements can significantly enrich the understanding of
textual content in the realm of sentiment analysis and emotion detection on social media platforms.
The side-by-side analysis illustrates proposed approach consistently outperforms the hybrid
conventional Glove and CNN method using adversarial transfer learning across different datasets.

References

[1] Tyng, C. M., Amin, H. U., Saad, M. N., & Malik, A. S. (2017). The influences of
emotion on learning and memory. Frontiers in psychology, 8, 1454.
[2] Paul Ekman, Wallace Friesen (1980) Facial Signs of Emotional Expressions Journal of
personality and social psychology, Vol 39, No. 6, 1125-1134.
[3] Kopelman, Shirli, Ashleigh Shelby Rosette, and Leigh Thompson. "The three faces of
Eve: Strategic displays of positive, negative, and neutral emotions in negotiations."
Organizational Behavior and Human Decision Processes 99.1 (2006): 81-101
[4] Gross, James J., et al. "Emotion and aging: Experience, expression, and control."
Psychology and aging 12.4 (1997): 590.
[5] Gunes H, Pantic M (2010) Automatic, dimensional and continuous emotion recognition.
Int J Synth Emot (IJSE) 1(1):68–99.
[6] Nandwani, P., & Verma, R. (2021). A review on sentiment analysis and emotion
detection from text. Social Network Analysis and Mining, 11(1), 1-19.
[7] Alswaidan, N., & Menai, M. E. B. (2020). A survey of state-of-the-art approaches for
emotion recognition in text. Knowledge and Information Systems, 62(8), 2937-2987.
[8] Jamal, N., Xianqiao, C., Hussain Abro, J., & Tukhtakhunov, D. (2020, December).
Sentimental analysis based on hybrid approach of latent dirichlet allocation and machine
learning for large-scale of imbalanced twitter data. In 2020 3rd International Conference
on Algorithms, Computing and Artificial Intelligence (pp. 1-7).
[9] Wen, S., & Wan, X. (2014, June). Emotion classification in microblog texts using class
sequential rules. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol.
28, No. 1).
[10] Sander, D., Grandjean, D., & Scherer, K. R. (2005). A systems approach to appraisal
mechanisms in emotion. Neural networks, 18(4), 317-352.
[11] Lench, H. C., Flores, S. A., & Bench, S. W. (2011). Discrete emotions predict changes in
cognition, judgment, experience, behavior, and physiology: a meta-analysis of
experimental emotion elicitations. Psychological bulletin, 137(5), 834.
[12] Paul Ekman, Wallace Friesen (1980) Facial Signs of Emotional Expressions Journal of
personality and social psychology, Vol 39, No. 6, 1125-1134.
[13] Plutchik R. A general psychoevolutionary theory of emotion. Amsterdam, Netherlands:
Elsevier; 1980 (pp. 3–33).
[14] Warriner, A. B., Kuperman, V., & Brysbaert, M. (2013). Norms of valence, arousal, and
dominance for 13,915 English lemmas. Behavior research methods, 45, 1191-1207.
[15] Russell JA. A circumplex model of affect. J PersSoc Psychol. 1980;39(6):1161
[16] Acheampong, F. A., Wenyu, C., & Nunoo‐Mensah, H. (2020). Text‐based emotion
detection: Advances, challenges, and opportunities. Engineering Reports, 2(7), e12189.
[17] Plutchik R. A general psychoevolutionary theory of emotion. Amsterdam, Netherlands:
Elsevier; 1980 (pp. 3–33).
[18] Algoe, S. B., & Haidt, J. (2009). Witnessing excellence in action: The ‘other-
praising’emotions of elevation, gratitude, and admiration. The journal of positive
psychology, 4(2), 105-127.
[19] Kusal, S., Patil, S., Choudrie, J., Kotecha, K., Vora, D., & Pappas, I. (2022). A Review
on Text-Based Emotion Detection--Techniques, Applications, Datasets, and Future
Directions. arXiv preprint arXiv:2205.03235.
[20] Lan, M., Tan, C. L., Su, J., & Lu, Y. (2008). Supervised and traditional term weighting
methods for automatic text categorization. IEEE transactions on pattern analysis and
machine intelligence, 31(4), 721-735.
[21] Huan, H., Guo, Z., Cai, T., & He, Z. (2022). A text classification method based on a
convolutional and bidirectional long short-term memory model. Connection Science,
34(1), 2108-2124.
[22] Basiri, M. E., Nemati, S., Abdar, M., Cambria, E., & Acharya, U. R. (2021). ABCDM:
An attention-based bidirectional CNN-RNN deep model for sentiment analysis. Future
Generation Computer Systems, 115, 279-294.
[23] Le, A. C. (2018, November). Integrating grammatical features into cnn model for
emotion classification. In 2018 5th NAFOSTED Conference on Information and
Computer Science (NICS) (pp. 243-249). IEEE.
[24] Pradhan, T., Kumar, P., & Pal, S. (2021). CLAVER: an integrated framework of
convolutional layer, bidirectional LSTM with attention mechanism based scholarly
venue recommendation. Information Sciences, 559, 212-235.
[25] Liang, S., Zhu, B., Zhang, Y., Cheng, S., & Jin, J. (2020, December). A Double Channel
CNN-LSTM Model for Text Classification. In 2020 IEEE 22nd International Conference
on High Performance Computing and Communications; IEEE 18th International
Conference on Smart City; IEEE 6th International Conference on Data Science and
Systems (HPCC/SmartCity/DSS) (pp. 1316-1321). IEEE.
[26] Ghosal, D., Majumder, N., Poria, S., Chhaya, N., & Gelbukh, A. (2019). Dialoguegcn: A
graph convolutional neural network for emotion recognition in conversation. arXiv
preprint arXiv:1908.11540.
[27] Trueman, T. E., & Cambria, E. (2021). A convolutional stacked bidirectional LSTM with
a multiplicative attention mechanism for aspect category and sentiment detection.
Cognitive Computation, 13, 1423-1432.
[28] Alsmadi, I., Ahmad, K., Nazzal, M., Alam, F., Al-Fuqaha, A., Khreishah, A., &
Algosaibi, A. (2021). Adversarial attacks and defenses for social network text processing
applications: Techniques, challenges and future research directions. arXiv preprint
arXiv:2110.13980.
[29] Hajek, P., Barushka, A., & Munk, M. (2020). Fake consumer review detection using
deep neural networks integrating word embeddings and emotion mining. Neural
Computing and Applications, 32, 17259-17274.
[30] Alsmadi, I., Aljaafari, N., Nazzal, M., Alhamed, S., Sawalmeh, A. H., Vizcarra, C. P., ...
& Al-Humam, A. (2022). Adversarial machine learning in text processing: A literature
survey. IEEE Access, 10, 17043-17077.
[31] Han, J., Zhang, Z., Cummins, N., & Schuller, B. (2019). Adversarial training in affective
computing and sentiment analysis: Recent advances and perspectives. IEEE
Computational Intelligence Magazine, 14(2), 68-81.
[32] Olah, J., Baruah, S., Bose, D., & Narayanan, S. (2021). Cross domain emotion
recognition using few-shot knowledge transfer. arXiv preprint arXiv:2110.05021.
[33] Sailunaz, K., & Alhajj, R. (2019). Emotion and sentiment analysis from Twitter text.
Journal of Computational Science, 36, 101003.
[34] Acheampong, F. A., Nunoo-Mensah, H., & Chen, W. (2021). Transformer models for
text-based emotion detection: a review of BERT-based approaches. Artificial
Intelligence Review, 1-41.
[35] Jamal. N., Xianqiao, C. & Aldabbas, H.(2019). Deep learning-based sentimental analysis
for large-scale imbalance twitter data. Future Internet, 11(9),190.
[36] Danisman, T., & Alpkocak, A. (2008, April). Feeler: Emotion classification of text using
vector space model. In AISB 2008 convention communication, interaction, and social
intelligence (Vol. 1, No. 4, pp. 53-59).
[37] Rassadin, A., Gruzdev, A., & Savchenko, A. (2017, November). Group-level emotion
recognition using transfer learning from face identification. In Proceedings of the 19th
ACM international conference on multimodal interaction (pp. 544-548).
[38] Al-Omari, H., Abdullah, M. A., & Shaikh, S. (2020, April). Emodet2: Emotion detection
in English textual dialogue using bert and bilstm models. In 2020 11th International
Conference on Information and Communication Systems (ICICS) (pp. 226-232). IEEE.
[39] Kumari, R., Ashok, N., Ghosal, T., & Ekbal, A. (2022). What is the fake? Probing
misinformation detection standing on the shoulder of novelty and emotion. Information
Processing & Management, 59(1), 102740.
[40] She, X., & Zhang, D. (2018, December). Text classification based on hybrid CNN-
LSTM hybrid model. In 2018 11th International symposium on computational
intelligence and design (ISCID) (Vol. 2, pp. 185-189). IEEE.
[41] Alm, C. O., Roth, D., & Sproat, R. (2005, October). Emotions from the text: machine
learning for text-based emotion prediction. In Proceedings of human language
technology conference and conference on empirical methods in natural language
processing (pp. 579-586).
[42] Bhuvaneshwari, P., Rao, A. N., Robinson, Y. H., & Thippeswamy, M. N. (2022).
Sentiment analysis for user reviews using Bi-LSTM self-attention-based CNN model.
Multimedia Tools and Applications, 81(9), 12405-12419.
[43] Huan, H., Guo, Z., Cai, T., & He, Z. (2022). A text classification method based on a
convolutional and bidirectional long short-term memory model. Connection Science,
34(1), 2108-2124
[44] Hajek, P., Barushka, A., & Munk, M. (2020). Fake consumer review detection using
deep neural networks integrating word embeddings and emotion mining. Neural
Computing and Applications, 32(23), 17259-17274
[45] Tammina, S., & Annareddy, S. (2020, January). Sentiment analysis on customer reviews
using convolutional neural network. In 2020 International Conference on Computer
Communication and Informatics (ICCCI) (pp. 1-6). IEEE.
[46] Kumari, R., Ashok, N., Ghosal, T., & Ekbal, A. (2021, July). A multitask learning
approach for fake news detection: novelty, emotion, and sentiment lend a helping hand.
In 2021 International Joint Conference on Neural Networks (IJCNN) (pp. 1-8). IEEE
[47] Stojanovski D, Strezoski G, Madjarov G, Dimitrovski I. Twitter sentiment analysis
using deep convolutional neural network. International Conference on Hybrid Artificial
Intelligence Systems 2015 Jun 22 (pp. 726-737). Springer, Cham
[48] https://www.kaggle.com/code/sayakd6/emotion-detection-with-glove-and-cnn-bilstm/
notebook
[49] Shen, Q., Wang, Z., & Sun, Y. (2017). Sentiment analysis of movie reviews based on
cnn-blstm. In Intelligence Science I: Second IFIP TC 12 International Conference, ICIS
2017, Shanghai, China, October 25-28, 2017, Proceedings 2 (pp. 164-171). Springer
International Publishing.
[50] Garg, A., Vats, S., Jaiswal, G., & Sharma, A. (2022, January). Analytical Approach for
Sentiment Analysis of Movie Reviews Using CNN and LSTM. In Artificial Intelligence
 and Speech Technology: Third International Conference, AIST 2021, Delhi, India,
November 12–13, 2021, Revised Selected Papers (pp. 99-115). Cham: Springer
International Publishing.
[51] Poria, S., Hazarika, D., Majumder, N., Naik, G., Cambria, E., & Mihalcea, R. (2018).
Meld: A multimodal multi-party dataset for emotion recognition in conversations. arXiv
preprint arXiv:1810.02508.
[52] Le, A. C. (2018, November). Integrating grammatical features into cnn model for
emotion classification. In 2018 5th NAFOSTED Conference on Information and
Computer Science (NICS) (pp. 243-249). IEEE.
[53] Pradhan, A., Ranjan Senapati, M., & Sahu, P. K. (2022). A multichannel embedding and
arithmetic-optimized stacked Bi-GRU model with semantic attention to detect emotion
over text data. Applied Intelligence, 1-18.
[54] Yao, L., Mao, C., & Luo, Y. (2019, July). Graph convolutional networks for text
classification. In Proceedings of the AAAI conference on artificial intelligence (Vol. 33,
No. 01, pp. 7370-7377).