An Intelligent System to Identify the Emotions in the Text Using a
Hybrid Deep Learning Model
Chevella Anil Kumar
1
a
, Vinay Kumar Chikoti
2
b
,Sree Ram Gandla
3
c
, Akshitha Ganji
1
,
M. Dharma Teja
1
and B. Nikhilesh
1
1
Department of Electronics and Communication Engineering, VNR Vignana Jyothi Institute of Engineering and
Technology, Hyderabad, Telangana, India
2
Software Engineer, Softech International Resources Inc, North Carolina, U.S.A.
3
E-Giants Technologies LLC, Texas, U.S.A.
Keywords: Natural Language Processing (NLP), Bidirectional Encoder Representations from Transformers (BERT),
Hybrid Models, LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), Contextual Embeddings,
Attention Mechanism, Feature Extraction.
Abstract: This research introduces a novel deep learning approach to effectively identifying emotions in text. This
model is a hybrid structure that integrates the advantages of BERT, LSTM, GRU, and Transformer encoder
layers to detect nuanced emotional signals and intricate linguistic patterns. By incorporating attention
mechanisms, we enhance the model's ability to focus on significant details and comprehend context. We
trained our model using the dataset 'emotion_dataset.csv' to accurately classify emotions across different text
formats. This approach has a wide range of applications, including sentiment analysis, complex storytelling,
human-computer interaction, personalized content generation, and monitoring mental health.
1 INTRODUCTION
As digital communication grows rapidly, there is a
wealth of unstructured textual data from sources such
as social media, online chats, emails, and consumer
reviews. This data contains significant emotional
insights that are critical for applications such as
human-computer interaction, mental health support,
sentiment analysis (Salam, Gupta, et al. 2018), and
tailored content production. By identifying emotions
in text, AI systems across a range of fields can
become far more perceptive and sympathetic.
However, conventional emotion detection systems,
which are usually based on rule-driven techniques or
simple machine learning algorithms, struggle with the
complex and context-sensitive nature of human
emotions.
By capturing the complex linguistic patterns and
relationships within a sentence, deep learning (Souza,
Souza, et al. 2019)] models—such as, Bidirectional
a
https://orcid.org/0000-0003-2277-8748
b
https://orcid.org/0009-0006-3041-6519
c
https://orcid.org/0009-0009-5495-2246
Encoder Representations from Transformers
(BERT)—have demonstrated great promise in natural
language processing (NLP) applications in recent
years. Due to its bidirectional architecture, BERT can
understand the underlying context from both forward
and backward, which makes it useful for tasks
requiring a thorough comprehension of context.
Although BERT is widely used in many NLP tasks,
its ability to decipher text's complex and multi-
layered emotions is somewhat constrained by its
inability to handle long-term dependencies and subtle
emotional cues
In order to overcome these challenges, our study
introduces a hybrid model that combines Transformer
encoder layers with BERT and Long Short-Term
Memory (LSTM) and Gated Recurrent Unit (GRU)
architectures. The suggested approach seeks to
increase the model's sensitivity to emotional nuances
by utilizing the contextual benefits of BERT, the
sequential memory capacities of LSTM and GRU,
Anil Kumar, C., Chikoti, V. K., Gandla, S. R., Ganji, A., Dharma Teja, M. and Nikhilesh, B.
An Intelligent System to Identify the Emotions in the Text Using a Hybrid Deep Learning Model.
DOI: 10.5220/0013590700004664
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 277-283
ISBN: 978-989-758-763-4
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
277

and the focused attention qualities of Transformer
layers. By taking into account both the immediate
context and the more complex, deeper linkages seen
in emotionally charged language, our suggested
approach seeks to improve emotion recognition. The
suggested deep learning (Rashid, Iqbal, et al. 2020)
approach has broad and important potential
ramifications, with improvements anticipated in
domains including automated customer support,
personalized content distribution, and mental health
support systems.
2 LITERATURE SURVEY
We strongly encourage authors to use this document
Emotion recognition in text has emerged as a critical
area of research, driven by its potential applications
in fields such as human-computer interaction,
sentiment analysis, and mental health assessment.
Recent studies have introduced innovative models
that leverage deep learning and hybrid approaches to
improve emotion detection accuracy. In "Hybrid
Feature Extraction for Multi-Label Emotion
Classification in English Text Messages
"
(Ahanin,
Ismail, et al. 2023) by Zahra Ahanin, Maizatul
Akmar Ismail, Narinderjit Singh Sawaran Singh, and
Ammar AL-Ashmori (2022), the authors propose a
hybrid feature extraction model for multi-label
emotion classification. This approach combines
human-engineered features, such as sentiment
polarity derived from lexical resources, with deep
learning-based features generated by Bi-LSTM and
BERT. The model addresses the challenges of small
training datasets through data augmentation, enabling
it to effectively capture linguistic and contextual
information. The study achieved Jaccard accuracies
of 68.40% on the SemEval-2018 dataset and 53.45%
on GoEmotions, demonstrating that combining
handcrafted and automated features can significantly
enhance performance in emotion detection tasks.
Another notable contribution is AHRNN: Attention-
Based Hybrid Robust Neural Network for Emotion
Recognition (Xu, Liu, et al. 2022) by Ke Xu, Bin Liu,
Jianhua Tao, Zhao Lv, Cunhang Fan, and Leichao
Song (2022). This study introduces the Attention-
Based Hybrid Robust Neural Network (AHRNN),
designed to improve semantic emotion recognition
and cross-language sentiment analysis.
.
The model
integrates CNNs for extracting local semantic
features, Bi-LSTM for capturing contextual
dependencies, and attention mechanisms for
emphasizing emotionally salient words. Pre-trained
embeddings are used to infuse prior semantic
knowledge, and the architecture exhibits robustness
against noisy data. AHRNN achieved 86% accuracy
in single-language tasks, improved fine-grained
classification by 9.6%, and enhanced cross-language
recognition by 1.5%. These results highlight the value
of attention mechanisms and hybrid architectures in
addressing complex emotion recognition challenges.
Building on these advancements, we propose a novel
hybrid architecture that integrates BERT, LSTM,
GRU, and a Transformer Encoder layer to further
improve emotion detection. Unlike previous studies
that combine BERT with either LSTM or attention-
based models, our model leverages all four
components to simultaneously capture deep
contextual, sequential, and global information. BERT
provides robust contextual embeddings but lacks the
sequential memory necessary for tasks involving
gradual emotional shifts. To address this limitation,
our model incorporates LSTM and GRU layers to
capture long-term dependencies and track the
progression of emotional cues. The Transformer
Encoder layer introduces an attention mechanism,
refining feature representations and enhancing the
model's focus on critical emotional signals while
balancing local and global context. The proposed
architecture capitalizes on the unique strengths of
each component: BERT for context-rich embeddings,
LSTM and GRU for sequential memory, and
Transformer Encoder for attention-based refinement.
Initial experimental results demonstrated a promising
accuracy of 73%, indicating the potential of this
hybrid design to outperform simpler model
combinations. By addressing key gaps in the
literature, this architecture presents a comprehensive
and well-rounded approach for emotion recognition,
paving the way for future advancements in NLP
applications requiring nuanced emotional
understanding.
3 DATASET USED
In this research, we used a dataset
emotion_dataset_2.csv of 34,785 sentences, each
labeled with one of eight emotion categories: anger,
disgust, fear, joy, neutral, sadness, shame, and
surprise. This dataset was particularly selected to
enable the model to recognize a wide range of human
emotions, which is valuable in tasks such as sentiment
analysis (Singh, Sharma, et al. 2023), social media
monitoring, and mental health assessment.
The dataset has an imbalanced distribution, with
certain emotions like joy (11,044 sentences) and
sadness (6,721 sentences) having a higher frequency
INCOFT 2025 - International Conference on Futuristic Technology
278

compared to underrepresented emotions such as
shame (145 sentences) and disgust (855 sentences) as
shown in table 1. Though the code does not explicitly
apply techniques (Malagi, Y. R et al. 2023) such as
data augmentation, oversampling, or class weighting
during training, this limitation is mitigated in part by
using a robust architecture that combines BERT
embeddings with LSTM (Ren, and She, 2021) and
GRU layers (Ren, and She, 2021) followed by a
Transformer encoder. This model structure,
combined with careful data splitting and validation
procedures, helps capture both contextual and
sequential dependencies in text data, enhancing the
model’s ability to generalize across diverse emotions.
To prepare this, we applied label encoding to
convert categorical emotion labels into numerical
values, ensuring compatibility with our model’s loss
function. For the text data, we used tokenization with
the BertTokenizer from Hugging Face, converting
each sentence into BERT-compatible tokens, with
sequences either truncated or padded to a fixed length
of 128. This standardization facilitated efficient
processing within our BERT-LSTM-GRU-
Transformer model, enabling it to effectively learn
patterns within the emotion categories.
Table 1: Dataset Distribution per emotion label.
Emotion Number of Sentences
Joy 11,044
Sadness 6,721
Fea
r
5,409
An
g
e
r
4,297
Surprise 4,061
Neutral 2,253
Dis
g
ust 855
Shame 145
Total 34,785
The dataset was split into training and validation
sets using an 80-20 ratio. This split allowed us to train
the model on a majority of the data while reserving a
substantial portion for performance evaluation. The
pre-processed data was then converted into PyTorch
datasets and subsequently loaded into Data Loaders
for batch processing during training.
The model training involves tracking metrics such
as accuracy, precision, recall, and F1 score. These
metrics provide insight into the model's performance
on each emotion category, allowing us to identify
potential biases and limitations related to class
imbalance.
4 METHODOLOGY
The suggested model uses a number of deep learning
components to process and extract local and global
textual data for the purpose of identifying emotions.
The BERT model, a transformer-based model that has
already been trained and offers contextual
embeddings for every token in the input text, is the
first component of the architecture. After that, these
embeddings are sent to an LSTM (TGDK, Selvarai,
et al. 2023), (Su, Wu, et al. 2018),
(Su, Wu, et al. 2018) layer, which records long-term
dependencies between tokens and sequential
information. A bidirectional GRU layer refines these
features after the LSTM, with an emphasis on
maintaining important sequential information while
lowering computational cost. The Transformer
Encoder layer, which uses self-attention methods to
improve the model's capacity to capture global
context, receives the output from the GRU.
To enable the model to concentrate on the most
pertinent passages for emotion categorization, the
Transformer encoder applies attention to different
areas of the text based on the information that has
been successively processed. This output conforms to
the model’s structure through dimension permutation,
facilitating efficient feature extraction and preserving
interoperability with subsequent layers.
Figure 1: Workflow Diagram of the Proposed Model.
Figure 2: Hybrid Deep Learning Model for the Fine-
Grained Emotion Recognition
An Intelligent System to Identify the Emotions in the Text Using a Hybrid Deep Learning Model
279

Figure 3: BERT Transformer Encoder Architecture
The architecture starts with the BERT model,
which accepts tokenized text as its input. Bert model
produces contextual embeddings that capture the
meaning of each token in context. The resulting
output shape from BERT is (number of tokens, 768),
indicating the contextual representation of each token
as shown in Figure 3.
Figure 4: LSTM cell with Input, Forget, Output, and Cell
Gates.
The output generated by BERT is then fed into an
LSTM (Long Short-Term Memory) layer as shown in
Figure 4. This layer captures the temporal
dependencies among the tokens, yielding an output
shape of (number of tokens, 256). It aids in
identifying patterns and relationships within the
sequence.
Figure 5: GRU: Capturing Long Term dependencies in
Sequential Data with Reduced Complexity.
A GRU (Gated Recurrent Unit) layer processes
the output after the LSTM layer. By generating an
output shape of (number of tokens, 128), the GRU
improves comprehension of the links between tokens.
This layer enhances the model's ability to represent
complex dependencies.
Subsequent to the GRU layer, the output is
directed into a Transformer Encoder layer, which
employs multi-head attention to enrich the contextual
embeddings. The output shape remains (number of
tokens, 128). The following layers include Global
Figure 6: Transformer Encoder Layer with Self-Attention
and Feed-Forward Neural Network
Average Pooling, Dropout, and a Fully Connected
layer, which ultimately yields class scores with an
output shape of (1, number of classes), representing
the predicted sentiment classification.
Deep learning (Baruah Chutia, et al. 2024)
concepts utilized in our model include Cross-Entropy
Loss, which measures the difference between
predicted probabilities and true labels, Adam W
Optimizer, an adaptive learning rate optimization
algorithm, Accuracy, Precision, Recall, and F1-Score
for evaluation, Matrix Multiplication for neural
network operations, ReLU Activation Function for
introducing non-linearity, Tensor Operations
(element-wise addition, subtraction, multiplication,
and division) for data manipulation, Permutation for
rearranging tensor dimensions, and Mean Pooling for
reducing dimensionality. The Adam W Optimizer
uses moving averages, hyperparameters, and
gradients for efficient training. Precision, Recall, and
F1-Score evaluate classification performance. Matrix
Multiplication enables neural network layer
interactions. ReLU Activation Function introduces
non-linearity, allowing complex pattern learning.
Tensor Operations facilitate data processing, while
Permutation and Mean Pooling enable dimension
rearrangement and reduction.
Let us consider an example sentence “I am so
happy today!” using a BERT tokenizer. The tokenizer
converts words into token IDs.
The tokenized output is:
INCOFT 2025 - International Conference on Futuristic Technology
280

[101, 1045, 2572, 2061, 3407, 2651, 102]
Where [101] and [102] are [cls] and [sep] tokens.
These tokens are embedded into vectors. Assume
each token has an embedding vector of dimension
d=768, so the sentence embedding matrix X has the
shape (7,768), where each row represents a word.
The self-attention layer computes attention
weights using:
𝐴𝑡𝑡𝑒𝑛𝑡𝑖𝑜𝑛(𝑄,𝐾,𝑉) = 𝑠𝑜𝑓𝑡𝑚𝑎𝑥(
𝑄𝐾


𝑑

) 𝑉
For example, if our embedding dimension is 768,
we may set dk=64 per attention head. Suppose after
calculating softmax on the attention weights, we get
a weighted sum that represents the context for each
token based on other tokens.
The LSTM takes in the last hidden state from
BERT (size 7×768) and processes it through its gates
at each timestep t.
For the token “happy” at the input “I am so happy
today!”, the hidden state is processed as:
Forget gate: 𝑓

=𝜎(𝑊

·[ℎ
,
𝑥

]+𝑏

) ,
controlling how much of the previous state is
retained.
Input gate: 𝑖

=𝜎(𝑊

·[ℎ
,
𝑥

]+𝑏

) , which
decides the amount of new information to add.
Cell update:
C

=f

⊙C

+i

⊙tanhW

·[ℎ
,
𝑥

]+b


These gates collectively update the hidden state
for each word, capturing sequential dependencies in
the text. After processing, LSTM produces output
with hidden size 128 (bidirectional output size 256).
Later, the LSTM output serves as input to the GRU
layer, which further refines sequential
representations.
For the “happy” token,
Update gate: 𝑧

=𝜎(𝑊𝑧·[ℎ
,
𝑥

]+𝑏

)
controls how much of the previous state is retained.
Reset gate: 𝑟

=𝜎(𝑊

·[ℎ
,
𝑥

]+𝑏

) controls
what previous information to ignore.
The output for each token is calculated using:
h

=(1−z

)⊙h

+z

⊙h′

Where h′

is the candidate hidden state at t.
For "happy," let’s assume this layer outputs a
vector of shape (7, 256), where 256 is the
bidirectional hidden size.
The GRU [14][15] output shape of (7,256) is
transposed to (7, batch size,256) for the transformer.
With 4 attention heads, the Transformer applies
multi-head self-attention:
MultiHead(Q,K,V)
=Concat(head
,
....,head

)W

where each head computes an attention score.
For "happy," the final output vector (after all
heads and FFN layers) is a contextualized
representation that considers other tokens in the
sentence. This helps capture complex relationships in
the text.
After passing through the Transformer, the output
is pooled by averaging across the sequence:
pooled_output =
1
7
𝑡𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚𝑒𝑟_𝑜𝑢𝑡𝑝𝑢𝑡


t
Suppose this pooled vector has a dimension of
256.
Fully Connected Layer: The pooled output is then
passed through a fully connected layer to produce
class logits. For a 4-class problem (Happy, Angry,
Joy, Sad):
output = pooled_output· W + b
where W has dimensions 256×4, resulting in
logits for each emotion.
With the logits, we apply cross-entropy loss:
for the example "I am so happy today!", the
predicted logits are [0.2, -0.5, 1.0, -1.5]
𝐬oftmax
(
[
0.2,−0.5,1.0,−1.5
]
)
=
[0.24,0.12,0.54,0.10]
Suppose the correct label is "Happy" (label 0).
The loss is:
loss = −
∑
𝑦


log(softmax(output)

)
loss = −log
(
0.24
)
≈1.43
This setup, along with these metrics, provides a
comprehensive analysis of the model's performance
across the training and validation datasets.
5 RESULTS AND DISCUSSION
In this study, we evaluated the performance of four
BERT-based model architectures for emotion
classification: BERT+CNN, BERT+ biLSTM,
BERT+GRU, and a proposed model that combines
BERT with LSTM, GRU, and Transformer Encoder
layers. The goal was to identify the most effective
model for accurately classifying emotions based on
text input, measured by validation accuracy, F1 score,
precision, and recall. The BERT+CNN model
showed moderate performance, achieving a
validation accuracy of 0.71, with an F1 score,
precision, and recall all at 0.71. This suggests that the
CNN layer effectively captures local text features but
may lack the ability to process deeper sequential
context, which is crucial in understanding the nuances
of emotional expressions. The BERT+biLSTM
model, by comparison, performed slightly lower, with
An Intelligent System to Identify the Emotions in the Text Using a Hybrid Deep Learning Model
281

a validation accuracy of 0.69, an F1 score of 0.68,
precision of 0.67, and recall of 0.69. The bi-
directional LSTM layer within this configuration
captures dependencies in both forward and backward
directions, enhancing contextual understanding, but it
may increase computational complexity and
susceptibility to overfitting, which could account for
its slightly lower performance.
Similarly, the BERT+GRU model achieved a
validation accuracy of 0.69, with an F1 score of 0.69,
precision of 0.71, and recall of 0.69. GRU, with a
simpler architecture compared to LSTM,
demonstrated efficiency but did not significantly
improve the model’s generalization capacity. In
contrast, the proposed hybrid model combining
BERT with LSTM, GRU, and Transformer Encoder
layers achieved the highest performance metrics, with
validation accuracy, F1 score, precision, and recall
each reaching 0.73. This hybrid approach leverages
the strengths of each component, with BERT
providing robust contextual embeddings, LSTM and
GRU capturing sequential information, and the
Transformer Encoder enhancing global context
understanding. This combined architecture appears to
effectively balance model complexity and
generalization, resulting in superior classification
accuracy and robustness, as evidenced by the higher
scores across all metrics.
Figure 7: Confusion matrix for proposed BERT- based
hybrid model for emotion classification.
The confusion matrix analysis provides further
insight into the model's classification performance.
The diagonal entries in the matrix indicate correctly
classified instances, while off-diagonal entries
represent misclassifications. Notably, the model
shows strong classification accuracy for the most
frequently occurring emotions in the dataset,
especially for classes with substantial data
representation. However, there were
misclassifications between emotions with similar
linguistic patterns, such as "joy" and "neutral,"
suggesting that overlapping language expressions in
certain emotions contribute to these classification
challenges. An illustrative example input, “I am
feeling very excited today!” was accurately classified
by the model as “joy,” demonstrating its capability in
real-world usage scenarios such as sentiment
analysis, where accurate emotion recognition is
essential.
Overall, The proposed model which is a
combination of BERT+LSTM+GRU+Transformer
encoder demonstrated not only high performance
metrics but also balanced precision and recall,
underscoring its suitability for real-world
applications that require robust generalization.
Table 2: Comparative Performance Metrics of BERT-based
Models for Emotion Classification as per our dataset
emotion_dataset_2.csv.
Model Validation
Accuracy
F1
score
Precision Recall
BERT+CN
N
0.71 0.70 0.71 0.71
BERT+bils
tm
0.69 0.68 0.67 0.69
BERT+GR
U
0.69 0.69 0.71 0.69
Proposed
model-
BERT+LS
TM+GRU
+Transfor
mer
encode
r
0.73 0.73 0.73 0.73
Compared to traditional machine learning
classifiers, which often struggle with overfitting and
generalization to unseen data, this deep learning
approach provides enhanced contextual and
sequential processing, making it well-suited for
complex emotion recognition tasks in diverse
applications such as mental health assessments,
customer feedback analysis, and conversational
sentiment analysis. The superior performance of this
model highlights the potential of advanced hybrid
architectures in accurately detecting nuanced
emotions in text data. Future work could focus on
expanding the dataset with a broader variety of
emotional expressions and increasing the number of
training epochs to further improve accuracy.
Additionally, integrating other advanced
components, such as attention mechanisms within the
recurrent layers, may further enhance classification
INCOFT 2025 - International Conference on Futuristic Technology
282

performance. These findings highlight the promise of
BERT-based hybrid architectures in providing
reliable and accurate emotion detection, crucial for
applications that demand high precision in sentiment
recognition from textual data
6 CONCLUSION
In terms of potential developments, this initiative
could be broadened by investigating different
enhancements and fine-tuning methods designed for
the emotional classification model's unique needs.
Integrating alternative pre-trained language models,
such as RoBERTa or DeBERTa, may enhance the
model's capacity for contextual comprehension, while
applying ensemble techniques could bolster its
robustness by merging the outputs from various
architectures. Furthermore, implementing advanced
data augmentation approaches could offer a wider
range of linguistic contexts for training, thereby
increasing the model's resilience to varied inputs.
Looking into hyperparameter optimization
techniques, including grid search or Bayesian
optimization, might refine the training parameters,
resulting in improved accuracy and efficiency.
To summarize, The Proposed model(BERT-based
model, augmented with LSTM, GRU, and
Transformer layers)shows remarkable potential for
emotion classification within text. Its multi-faceted
approach merges BERT's contextual capabilities with
the sequential analysis of LSTM and GRU, alongside
the potent sequence-to-sequence functions of
Transformers. This integration effectively utilizes the
strengths of each layer, thereby enhancing the
precision of emotion detection. Future endeavours
focused on adapting the model to different datasets
and contexts will further boost its flexibility, making
it a significant asset in sentiment analysis, mental
health assessments, and various NLP applications.
REFERENCES
S. A. Salam and R. Gupta, ‘‘Emotion detection and
recognition from text using machine learning,’’ Int. J.
Comput. Sci. Eng., vol. 6, no. 6, pp. 341–345, Jun.
2018.
A. D. Souza and R. D. Souza, ‘‘Affective interaction based
hybrid approach for emotion detection using machine
learning,’’ in Proc. Int. Conf. Smart Syst. Inventive
Technol. (ICSSIT), Nov. 2019, pp. 337–342.
Rashid, U., Iqbal, M.W., Skiandar, M.A., Raiz, M.Q.,
Naqvi, M.R. and Shahzad, S.K. (2020) Emotion
Detection of Contextual Text Using Deep Learning.
2020 4th International Symposium on
Multidisciplinary Studies and Innovative Technologies
(ISMSIT), Istanbul, Turkey, 22-24 October 2020, 1-5.
https://doi.org/10.1109/ISMSIT50672.2020.9255279
Z. Ahanin, M. A. Ismail, N. S. S. Singh, and A. AL-
Ashmori, "Hybrid Feature Extraction for Multi-Label
Emotion Classification in English Text Messages,"
Sustainability, vol. 15, no. 16, p. 12539, 2023, doi:
10.3390/su151612539.
K. Xu, B. Liu, J. Tao, Z. Lv, C. Fan, and L. Song,
"AHRNN: Attention-based hybrid robust neural
network for emotion recognition," Cogn. Comput.
Syst., vol. 4, no. 1, pp. 85–95, Mar. 2022, doi:
10.1049/ccs2.12038.
V. Singh, M. Sharma, A. Shirode, and S. Mirchandani,
"Text emotion detection using machine learning
algorithms," in Proc. 8th Int. Conf. Commun. Electron.
Syst. (ICCES), Coimbatore, India, 2023, pp.1264-1268.
R. R. Malagi, Y. R., S. P. T. K., A. Kodipalli, T. Rao, and
R. B. R., "Emotion detection from textual data using
supervised machine learning models," in Proc. 5th Int.
Conf. Emerg. Technol. (INCET), Belgaum, India,
2023, pp. 1–5.
F. Ren and T. She, "Utilizing external knowledge to
enhance semantics in emotion detection in
conversation," IEEE Access, vol. 9, 2021.
S. TGDK, V. Selvarai, U. Raj, V. P. Raiu, and J. Prakash,
"Emotion detection using bi-directional LSTM with an
effective text pre-processing method," in Proc. 12th Int.
Conf. Comput. Commun. Netw. Technol. (ICCCNT),
Kharagpur, India, 2021.
M.-H. Su, C. H. Wu, K.-Y. Huang, and Q.-B. Hong,
"LSTM-based text emotion recognition using semantic
and emotional word vectors," in Proc. Conf. Affect.
Comput. Intell. Interact. (ACII Asia), Beijing, China,
2018.
M.-H. Su, C. H. Wu, K.-Y. Huang, and Q.-B. Hong,
"LSTM-based text emotion recognition using semantic
and emotional word vectors," in Proc. Conf. Affect.
Comput. Intell. Interact. (ACII Asia), Beijing, China,
2018.
"Contextual emotion detection in text using deep learning
and big data," in Proc. 2nd Int. Conf. Comput. Sci., Eng.
Appl. (ICCSEA), Gunupur, India, 2022, pp. 1–5.
N. Baruah and T. Chutia, "An overview of deep learning
methods for detecting emotions," Artif. Intell., 2024.
K. Sakthiyavathi and K. Saruladha, "Design of hybrid deep
learning model for text-based emotion recognition," in
Proc. 2nd Int. Conf. Adv. Comput. Intell. Commun.
(ICACIC), Puducherry, India, 2023.
N. Priyadarshini and J. Aravinth, "Emotion recognition
based on fusion of multimodal physiological signals
using LSTM and GRU," in Proc. 3rd Int. Conf. Secure
Cyber Comput. Commun. (ICSCCC), Jalandhar, India,
2023, pp. 1–6.
An Intelligent System to Identify the Emotions in the Text Using a Hybrid Deep Learning Model
283