International Journal of

INTELLIGENT SYSTEMS AND APPLICATIONS IN

ENGINEERING
ISSN:2147-67992147-6799www.ijisae.org Original Research Paper

Enhancing Customer Engagement in Fashion: Strategies for

Optimizing Chatbot Performance
Dr. DikshendraDaulat Sarpate1*, Dr. Manju D Pawar2, Dr. Lekshmy P L3, Dr.Shanthaamani V4, Dr. Swati
V Sakhare5,Dr. Megha Kadam6,Dr. Sarita Avinash Patil7
Submitted:14/03/2024 Revised: 29/04/2024 Accepted: 06/05/2024
Abstract: In recent years, chatbots have become integral in the fashion industry, enhancing customer-brand interaction, facilitating
communication, and contributing to e-commerce growth through personalized shopping experiences. This research aims to reveal
strategic approaches in luxury fashion, translating insights for small and medium-sized enterprises. It investigates the impact of chatbot
integration on customer sentiment, aiming to provide practical recommendations for improving customer interactions. Emphasizing the
importance of intelligent chatbot design, the research recognizes their crucial role in enhancing customer experience through valuable
insights. The primary objective is to optimize chatbot performance, ensuring precise and effective responses to user queries and,
consequently, adding significant value to the business. The proposed approach incorporates optimally configured Long Short-Term
Memory (LSTM) networks, Seq2Seq Architecture, Attention Mechanism, Bag of Words (BOW) model, and Beam Search decoding. The
effectiveness of the optimally configured LSTM, in combination with attention mechanisms, extends to both longer and shorter
sentences. Through rigorous testing, the proposed model achieves an exceptional accuracy rate of 99%, surpassing other state-of-the-art
techniques. This outcome underscores the efficacy of the implemented strategies in elevating chatbot performance.

Keywords: Chatbots, Fashion Industry, Long Short-Term Memory, Zebra Optimization Algorithm
1. Introduction luxury fashion retail industries.
In recent years, chatbots have emerged as pivotal tools in The proposed approach integrates various techniques,
the fashion industry, reshaping customer-brand including LSTM networks [9], Seq2Seq Architecture,
interactions [1] and the e-commerce landscape [2]. Attention Mechanism, Bag of Words (BOW) model,
Particularly in luxury fashion, these AI-driven chatbots Beam Search decoding [10], and hyperparameter tuning.
offer round-the-clock personalized shopping experiences, LSTM networks are tailored for sequential data
providing valuable insights and fostering brand loyalty in processing [11], while Seq2Seq models facilitate
a fiercely competitive market [3]. Embracing lessons mapping input sequences to output sequences, vital for
from luxury brands, small and medium enterprises tasks like machine translation [12]. Attention
(SMEs) [4] can leverage similar strategies to thrive in mechanisms enhance the handling of long inputs by
this dynamic environment, prioritizing authenticity, focusing on relevant parts. The BOW model represents
innovation, and customer-centricity. text as word vectors, disregarding grammar and order
[13], while Beam Search decoding enables exploration of
However, chatbots face challenges in understanding
multiple potential responses [14], thereby enriching
conversational context, handling complex queries, and
chatbot output quality and diversity. Additionally, the
maintaining coherence in responses [5]. These limitations
Zebra Optimization Algorithm (ZOA) fine-tunes
can lead to inappropriate replies, potentially harming
hyperparameters [15], drawing inspiration from zebras'
customer satisfaction. Hence, this study aims to optimize
collective behavior to efficiently optimize model
chatbot performance to ensure precise and effective
performance. Applied in luxury fashion retail, this
responses, thereby enhancing the overall customer
enhances chatbot efficacy with optimized LSTM
experience [6]. Leveraging advanced natural language
networks capable of processing inquiries of varying
processing techniques [7] and machine learning
lengths, ensuring seamless conversational experiences
algorithms [8], our chatbot aims to blend personalized
[16].
service with digital convenience, setting new standards in
1 This study's significance lies in demonstrating the
Assistant Professor,AI & DS Department,Zeal College of Engineering
and Research, Narhe, Pune, India. effectiveness of implemented strategies in elevating
Email:[email protected] chatbot performance within luxury fashion contexts. By
2
Associate Professor,AI & DS Department,Zeal College of Engineering
showcasing the positive outcomes of optimizing chatbot
and Research, Narhe, Pune, India.
Email:[email protected] capabilities, this research contributes valuable insights to
3
Assistant Professor, Department of Computer Science Engineering, academia and industry practitioners alike. Moreover, it
LBS Institute of Technology for Women, Kerala, India.
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(4), 4759 - 4768 |4759
Email:[email protected]
4
Professor, Department of Management Studies,
Hindusthan College of Arts & Science, India.
Email:[email protected]
5
Assistant Professor,Faculty of Engineering and Technology,
Parul Institute of Engineering and Technology,Parul
 underscores the importance of engaging digital 91.57%. In contrast, generative-based chatbots, boasting
interactions in luxury fashion retail, highlighting encoder-decoder designs, achieved an accuracy of
enhanced chatbot performance as a strategic tool for 94.45%.
driving customer engagement and retention amidst
In 2021 Dhyani et al [19] discussed the modeling and
evolving consumer expectations.
performance in deep learning computation for a Chatbot.
2. Literature Review The Bidirectional Recurrent Neural Networks (BRNN)
containing attention layers is used, so that input sentence
A comprehensive literature review on chatbot utilization
with large number of tokens (or sentences with more than
in the fashion industry and its impact on customer
20–40 words) can be replied with more appropriate
experience would provide valuable insights into current
conversation. The dataset used in the paper for training of
trends, challenges, and advancements in this domain.
model is used from Reddit. The main purpose of this
In 2020 Anki et al [17] introduced chatbot programs for work is to increase the perplexity and learning rate of the
storing data gathered from question and answer systems. model and find Bleu Score for translation in same
Their program utilizes the BiLSTM model, aiming for language. The experiments are conducted using
high accuracy in output alignment with user expectations. Tensorflow using python 3.6. The perplexity, leaning
Through program evaluation and tests with various rate, Bleu score and Average time per 1000 steps are
parameter pairs, they determined that Parameter Pair 1, 56.10, 0.0001, 30.16 and 4.5 respectively. One epoch is
utilizing the BiLSTM model, yielded the highest completed at 23,000 steps.
accuracy for the BiLSTM Chatbot.
In 2022 Trojovská et al [20] introduced the Zebra
In 2023 Aslam and Usman conducted a study on retail Optimization Algorithm (ZOA), a bio-inspired
fashion chatbots. They found that current chatbots often metaheuristic algorithm. ZOA mimic the foraging
give generic responses, lacking depth for complex behavior and defense strategy of zebras in nature. The
inquiries. To enhance usability, the study explored how algorithm's steps are outlined and mathematically
social experiences shape customer interactions. Data modeled. ZOA's effectiveness is evaluated across sixty-
from online reviews, interviews, and focus groups eight benchmark functions, including various types of
emphasized the need for a wider product range and more functions. Comparative analysis with nine established
sophisticated features to enrich shopping experiences. algorithms demonstrates ZOA's superior performance in
While chatbots can recommend items aligned with achieving a balance between exploration and
preferences, developers must address usability challenges exploitation. Additionally, ZOA's applicability to real-
and offer multilingual support. Optimization is crucial to world engineering design problems, such as
reduce battery drain, prevent freezing, and improve tension/compression spring, welded beam, speed reducer,
response times. The study proposes a research framework and pressure vessel, showcases its effectiveness in
to enhance retail chatbot experiences through optimizing design variables when compared to the
interactivity, compatibility, and credibility. competitor algorithms.
In 2023 Pandey et al [18] explored the health impacts of 2.1. Research methodology
increased screen time, emphasizing the detrimental
The research explores how customers interact with
effects on mental health. They utilized Deep Learning
chatbots in luxury fashion, examining satisfaction,
(DL) and Machine Learning (ML) techniques to
experience, purchase likelihood, engagement, and
investigate the association between technological
privacy concerns. It suggests that integrating chatbots can
obsessions and health. The deployment of chatbots across
benefit small and medium fashion enterprises by
various industries emerged as a transformative
streamlining operations, improving customer service, and
development. Their study focused on conversational
potentially increasing sales. These insights can guide
Artificial Intelligence (AI) systems, enabling human-like
SMEs in effectively adopting chatbot technology,
interactions with machines. Two types of chatbots were
enhancing customer satisfaction, and maximizing overall
designed and developed: retrieval-based and generative-
engagement in the fashion market, as showcased in
based. Among retrieval-based chatbots, various designs
Figure 1.
were tested, with accuracies ranging from 65.57% to

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(4), 4759 - 4768 |4760
 Fig. 1.Benefits of Chatbots for Small and Medium Enterprises (SMEs)
The research analyzed consumer sentiment towards personalized assistance, 24/7 support, and engaging
chatbots and online human agents on X-social media interactions. The analysis shown in Figure 2 found that
platform (formerly Twitter). It aimed to identify consumer among negative sentiments, 226 customers preferred the
sentiment towards these agents. The results show that chatbot, while 1116 leaned towards the human agent. In
chatbots generally evoke a more positive sentiment the neutral category, 197 customers favored the chatbot,
compared to online human agents. Chatbots offer the and 371 opted for the human agent. For positive
advantage of handling multiple customers simultaneously, sentiments, 127 customers expressed satisfaction with the
unlike human agents who can only manage one customer chatbot, while 243 preferred the human agent. The total
at a time. Integrating chatbots into luxury fashion retail interactions were 550 with the chatbot and 1730 with the
can enhance customer experience by providing human agent across all sentiment categories [21].

Fig. 2. Comparison of Consumer Sentiment with Online Human Agents with the chatbot
The investigation highlights a significant negative impact 3. Proposed Methodology
of 41.09% on customer sentiments due to chatbots,
Over the years, extensive research has been conducted to
signaling a notable concern. To enhance customer service,
implement chatbots effectively, with a shift towards
engagement, data monitoring, cost savings, and privacy
leveraging advanced algorithms like artificial intelligence.
protection, improving chatbot performance is crucial. The
In addition to traditional rule-based techniques,
proposed approach integrates advanced techniques
contemporary approaches rely on sophisticated algorithms,
including optimally configured LSTM networks, Seq2Seq
necessitating the use of datasets for model training. The
Architecture, Attention Mechanism, BOW model, and
proposed methodology adopts a multifaceted approach to
Beam Search decoding. This framework aims to enhance
develop and optimize a chatbot system, the architecture
the efficiency and accuracy of natural language
exhibits in Figure 3. Initially, raw data undergoes thorough
processing. The optimally configured LSTM, combined
pre-processing to enhance its quality and suitability for
with attention mechanisms, effectively processes both
machine learning models. Subsequently, the Sequence-to-
longer and shorter sentences, addressing nuanced
Sequence (Seq2Seq) architecture is implemented,
linguistic structures. This comprehensive model captures
leveraging LSTM networks to handle sequential data
intricate patterns and dependencies across varying
effectively. Attention mechanisms are integrated to
sentence lengths, demonstrating versatility in language
improve the model's focus on relevant input information,
processing.
particularly crucial for longer sequences. Additionally, the
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(4), 4759 - 4768 |4761
Bag-of-Words (BOW) model is employed for text LSTM model, aiming to boost its performance. Through
representation, while Beam Search Decoding enhances rigorous evaluation and validation, the research endeavors
response generation diversity. Notably, the optimization to deliver an advanced chatbot system capable of
techniques, specifically the Zebra Optimization Algorithm delivering accurate and contextually relevant responses
(ZOA), are introduced to fine-tune hyperparameters of the across various domains.

Fig. 3.Flowchart illustrating the Proposed Methodology

3.1
.Data Pre-Processing employ LSTM networks instead of traditional Recurrent
Neural Networks. Additionally, integrating the Attention
Data pre-processing is a crucial step that enhances dataset
Mechanism enhances the model's predictive accuracy by
quality by extracting meaningful insights. It involves
handling lengthy sentences [23].
preparing the data to make it suitable for machine learning
and artificial intelligence models, essentially transforming 3.3.Attention mechanism
raw data into a format that is meaningful and
The purpose of Attention Mechanisms is to enhance the
understandable. This process, often referred to as data
performance of the chatbot system by improving its ability
mining, involves cleaning the data by removing
to focus on relevant information within input sequences.
unnecessary characters such as punctuations and
Specifically, Attention Mechanisms are crucial for
abbreviations [22]. Additionally, standardization
handling longer input sequences effectively [24], where
procedures such as converting all alphabets to lowercase
traditional Seq2Seq models might struggle due to memory
and removing symbols are implemented to ensure
limitations. By dynamically allocating attention to
consistency across the dataset.
different parts of the input sequence during the decoding
3.2.Sequence-to-Sequence process, Attention Mechanisms ensure that the model can
selectively attend to relevant information, thereby
The Sequence-to-Sequence (Seq2Seq) architecture,
improving the accuracy and contextuality of the generated
consisting of an encoder and decoder, is a versatile
responses. This is particularly important in natural
network widely used in natural language processing tasks,
language processing tasks, where context plays a crucial
originally designed for language translation like English to
role in understanding and generating meaningful
Spanish. Its utility has expanded to tasks such as text
responses. Therefore, integrating Attention Mechanisms
summarization, chatbot development, and more. Seq2Seq
into the chatbot system helps to address the challenge of
excels in handling input and output sequences of varying
processing longer sequences and enhances the overall
lengths, making it suitable for diverse NLP datasets. For
quality of the conversation experience for users.
our retrieval-based chatbot model, we choose Seq2Seq due
to its adaptability and effectiveness. While it performs The attention weights (aij) are calculated using softmax
well with shorter inputs, memory limitations hinder its function by Equation (1) and (2).
performance with longer inputs. To address this, we
exp eij 
aij  (1)
k 1 exp eik 
Tx

eij  l g i 1 , h j  (2)

A context vector (Ci) is created by Equation (3).

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(4), 4759 - 4768 |4762
 Tx interactions, ultimately enhancing user satisfaction and
Ci   aij h j (3) usability.
j 1
3.6.Long Short-Term Memory
Where the output score eij is described by function l which
In our chatbot system, LSTM networks are employed to
is used to capture the alignment among input at j and
generate responses when customers ask questions. LSTM
output at i. Tx is the annotation such as h0, h1, h2, h3, h4.
networks are a type of recurrent neural network designed
The attention weights calculated are a10, a11, a12, a13, a14
to handle sequential data, such as natural language
are 0.05, 0.2, 0.05, 0.6, and 0.1 respectively. This context
sentences. These networks are particularly effective in
vector Ci is now fed to g1 along with the output of
capturing long-term dependencies in the input sequence,
previous decoder layers.
allowing the chatbot to understand the context of the
3.4.Bag-of-Words question and provide appropriate responses. By leveraging
LSTM networks, our chatbot aims to deliver accurate and
The purpose of Bag-of-Words (BOW) model is to provide
coherent answers, enhancing the overall customer
a simplified yet effective representation of text data. The
experience.Hyperparameters tuning is essential in LSTM
BOW model serves as a baseline technique for encoding
networks because it directly impacts the performance and
input text, where each word in the vocabulary is
effectiveness of the model. LSTM networks are powerful
represented as a unique feature[25]. By disregarding the
tools for processing sequential data, such as text or time
order and grammar of the words in the text, the BOW
series data, but their performance heavily depends on the
model focuses solely on the presence or absence of words
values of various hyperparameters.The hyperparameter
within the document. This approach enables efficient
tuning is essential in LSTM networks to maximize
processing of text data and facilitates subsequent analysis
performance, prevent overfitting, stabilize training, adapt
and modeling tasks. In the context of the chatbot system,
to dataset characteristics, and enhance generalization,
the BOW model helps in encoding user queries and
ultimately leading to more effective and reliable models
responses, enabling the model to understand the basic
for various sequential data processing tasks. The research
content and context of the input text. While simplistic, the
involves utilizing the ZOA to identify the optimal
BOW model serves as a valuable starting point for text
hyperparameters for the LSTM.
representation, laying the foundation for more
sophisticated techniques used in the chatbot system. 3.7.Zebra Optimization Algorithm
Zebra’s social interactions mainly focus on finding food
3.5.Beam Search Decoding
and protecting themselves from predators. In the foraging
The purpose of Beam Search Decoding is to enhance the process, a leading zebra within the herd guides others to
quality and diversity of the responses generated by the grazing areas across open plains. In times of danger from
chatbot system. Beam search decoding is a search predators, zebras often resort to a zigzagging escape
algorithm used during the generation of output sequences, pattern, although they may also gather together to
particularly in natural language processing tasks like disorient or intimidate the threat. The ZOA is a
language translation or dialogue generation. The primary metaheuristic optimization algorithm inspired by the
aim of Beam Search Decoding is to explore multiple collective behavior of zebras. It mimics the social
potential response candidates and select the most likely behaviors observed in zebras, such as herding, foraging,
ones based on a scoring criterion. By considering a and defense strategies against predators.To iteratively
predetermined number of candidate sequences (referred to search for optimal solutions to optimization problems,
as the "beam width"), Beam Search Decoding allows the ZOA employs mechanisms like communication,
model to generate a range of plausible responses[26]. This cooperation, and competition among solution candidates,
approach helps to mitigate the issue of the model aiming to efficiently explore the search space and
producing overly deterministic or repetitive responses, converge towards promising solutions. This algorithm has
leading to more diverse and contextually appropriate been applied to various optimization tasks in fields such as
outputs. In the context of the chatbot system, Beam Search engineering, computer science, and operations research,
Decoding enables the generation of varied and nuanced showcasing its effectiveness in finding high-quality
responses to user queries, improving the overall solutions across diverse problem domains [21].The flow
conversational experience. By considering multiple chart of ZOA shown in Figure 4 is used to configure the
potential responses simultaneously, the chatbot system can LSTM for performance enhancement.
offer more engaging and contextually relevant

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(4), 4759 - 4768 |4763
 Fig. 4. Flow Chart of Zebra Optimization Algorithm
4. Results and Discussion Number of TP  Number of TN
Accuracy 
Number of TP  Number of TN  Number of FP  Number of FN
The research demonstrates that our optimally configured (4)
LSTM with ZOA-based chatbot system excels across various
evaluation metrics, showcasing high Accuracy, Precision, In evaluating performance metrics across classifiers,
F1-Score, Bilingual Evaluation Understudy (BLEU) Score notable findings emerge. Optimally configured LSTM
and Recall-Oriented Understudy for Gisting Evaluation models with ZOA and Gravitational Optimization
(ROUGE). The response quality of our proposed approach, Algorithm (GOA) demonstrated outstanding accuracy,
integrated with the fashion industry chatbot, is marked by its with ZOA leading at 99% and GOA closely following
ability to offer precise, relevant, and insightful responce to at 98% shown in Figure 5. This underscores the
user queries pertaining to fashion products, trends, and efficacy of optimization algorithms in fine-tuning
services. Our chatbot uses advanced natural language LSTM architectures for superior performance. The
processing to enhance user experience, boost satisfaction, LSTM model with Seq-Seq architecture also achieved
increase purchase likelihood, foster engagement, and address high accuracy at 98%, showcasing its effectiveness in
privacy concerns, bolstering the brand's competitiveness in capturing underlying patterns. Ensemble learning
the fashion industry. techniques like Random Forest and AdaBoost
performed well with accuracies of 97.82% and 97.42%
4.1.Performance Evaluation Metrics respectively, highlighting the strength of combining
In the assessment of a fashion industry chatbot response, multiple models. Decision Tree and KNN classifiers
several performance evaluation metrics such as Accuracy, achieved respectable accuracies of 97.26% and 96.13%
Precision, F1-Score, BLEU Score and ROUGE can be respectively, though slightly trailing LSTM-based
utilized to measure its effectiveness and quality. models and ensemble methods. Conversely, the
Bidirectional LSTM (Bi-LSTM) model exhibited lower
Accuracy is the proportion of correctly classified responses accuracy at 87%, suggesting potential limitations in
out of the total responses evaluated provides an overall capturing long-term dependencies.
measure of correctness, as shown in Equation (4).

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(4), 4759 - 4768 |4764
 Fig. 5. Accuracy to evaluate the performance of chatbot's response
Precision is the proportion of correctly identified optimization techniques in fine-tuning LSTM
relevant responses out of all responses identified as architectures. The LSTM model with Seq-Seq
relevant by the chatbot. It measures the accuracy of architecture also performed exceptionally well, with a
relevant response identification, as shown in Equation precision score of 98%, highlighting its ability to
(5). capture patterns accurately. Ensemble methods like
Random Forest and AdaBoost achieved precision
TP
Pr ecision  (5) scores close to LSTM models, while simpler models
TP  FP like Decision Trees and K-Nearest Neighbors (KNN)
The precision results offer valuable insights into each had respectable scores. However, models like Gated
classifier's effectiveness in identifying relevant Recurrent Unit (GRU) and Bidirectional LSTM (Bi-
responses for fashion industry chatbots. The optimally LSTM) had slightly lower precision, suggesting
configured LSTM with ZOA demonstrated in Figure 6 challenges in capturing relevant patterns. Overall, these
shows the highest precision at 98.5%, indicating the results emphasize the superiority of LSTM-based
majority of positive classifications were relevant. models, especially when optimized with advanced
Similarly, the LSTM with GOA) achieved a high algorithms, in classifying relevant responses for fashion
precision of 97%, showcasing the effectiveness of industry chatbots.

Fig. 6.Precision to evaluate the performance of chatbot's response

F1-Score is the harmonic mean of precision and recall, configured LSTM models with ZOA and GOA
offering a balanced measure of the chatbot's achieved the highest scores, demonstrating excellent
performance in identifying relevant responses, as shown balance between precision and recall. Other models,
in Equation (6). such as Seq-Seq LSTM, decision trees, Random Forest,
and AdaBoost, also performed well. However, GRU
2 * Pr ecision
F1 - Score  and Bi-LSTM exhibited lower F1-scores, indicating
Pr ecision  Re call challenges in achieving balanced performance. Overall,
(6) LSTM-based models optimized with advanced
algorithms proved most effective for fashion industry
The F1-score analysis in Figure 7 reveals that optimally
chatbots.
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(4), 4759 - 4768 |4765
 Fig. 7.F1-score to evaluate the performance of chatbot's response
4.2.Bilingual Evaluation Understudy Score The BLEU score calculation involves factors such as the
brevity penalty (BP), maximum n-gram size (N), and
BLEU is the score to evaluate how well the translation
weights assigned to the precision of n-grams. However,
from one language to another was done. Perfect BLEU is
while BLEU score is commonly used to evaluate chatbots
considered to be 1, where the mismatch score is 0. BLEU
and other natural language processing systems, it's crucial
score is used to evaluate the predicted responses, in which
to consider additional metrics and qualitative assessments
we compared the predicted response with the reference
to capture aspects like fluency, coherence, and semantic
from the dataset. Calculating the BLEU score for a chatbot
quality. In assessing the effectiveness of the optimally
involves comparing the generated responses of the chatbot
configured LSTM with ZOA model, BLEU scores are
to a set of reference responses, which are considered to be
compared. The Figure 8 shows that the optimally
correct or ideal. BLEU score is commonly used to evaluate
configured LSTM with ZOA outperforms existing models,
the quality of machine-generated text by comparing it to
achieving a BLEU score of 0.82. Conversely, K-Nearest
human-generated text. Equation (7) describes the BLEU
Neighbors (KNN) exhibits the weakest performance
score.
among classifiers, with a BLEU score of 0.20.
 N 
BLEU  BP  exp   wn  log p n  (7)
 n1 

Fig. 8. Bilingual Evaluation Understudy Score to evaluate the performance of chatbot's response
4.3.Recall-Oriented Understudy for Gisting between the generated text and reference text based on the
Evaluation overlap of n-grams (contiguous sequences of n items,
typically words) between the two. Our experiment reveals
ROUGE is a statistic used to automatically evaluate the
that the optimally configured LSTM with (ZOA) stands
quality of a computer-generated word summary to that of
out, demonstrating the better performance of ROGUE
a person, which is a more evolved way of recall. ROGUE
Score of 0.79, as illustrated in Figure 9. Moreover, the
is a metric commonly used to evaluate the quality of
KNN displays the weakest performance among the
machine-generated text, such as chatbot responses, in
classifiers, achieving a bleu rate of 0. 07 in this scenario.
comparison to reference texts. It calculates the similarity

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(4), 4759 - 4768 |4766
 Fig. 9. Recall-Oriented Understudy for Gisting Evaluation to evaluate the performance of chatbot's response
5. Conclusion [2] Jiang, L., Sun, W., and Ren, L., "Bridging the Gap:
Anxiety’s Role in Shaping Consumption Patterns of
The incorporation of chatbots has become increasingly
Chinese University Students in the O2O E-
essential within the fashion sector, transforming
Commerce Landscape”. Journal of the Knowledge
interactions between customers and brands and fostering
Economy, pp.1-22, (2023).
e-commerce growth through tailored shopping
https://doi.org/10.1007/s13132-023-01642-w
experiences. This study aims to narrow the divide between
[3] Kumar, A., Bapat, G., Kumar, A., Hota, S. L.,
high-end fashion labels and smaller businesses by
Abishek, G. D., and Vaz, S., "Unlocking Brand
investigating strategic approaches and translating insights
Excellence: Harnessing AI Tools for Enhanced
into actionable suggestions. By examining the impact of
Customer Engagement and Innovation". Engineering
chatbot integration on customer sentiment, this research
Proceedings, 59(1), pp.204, (2024).
emphasizes the importance of intelligent chatbot design in
https://doi.org/10.3390/engproc2023059204
improving overall customer satisfaction. The primary goal
[4] Li, W., Liu, K., Belitski, M., Ghobadian, A., and
is to enhance chatbot performance, ensuring accurate and
O'Regan N., "e-Leadership through strategic
efficient responses to user inquiries, thereby delivering
alignment: An empirical study of small-and medium-
significant value to businesses. Notably, the optimally
sized enterprises in the digital age”. Journal of
configured LSTM, combined with attention mechanisms,
Information Technology, 31, pp.185-206, (2016).
demonstrates outstanding effectiveness in handling both
https://doi.org/10.1057/jit.2016.10
lengthy and concise sentences, achieving an impressive
[5] Almansor, E. H., and F. K., Hussain. "Survey on
accuracy rate of 99% during thorough testing. These
intelligent chatbots: State-of-the-art and future
results underscore the efficacy of the applied techniques in
research directions". In Complex, Intelligent, and
substantially improving chatbot performance, leading to
Software Intensive Systems: Proceedings of the 13th
enhanced customer interactions and satisfaction within the
International Conference on Complex, Intelligent,
fashion industry.
and Software Intensive Systems (CISIS-2019), pp.
Declaration of Conflicting Interests 534-543. Springer International Publishing, (2020).
The authors declared no potential conflicts of interest with https://doi.org/10.1007/978-3-030-22354-0_47
respect to the research, authorship, and/or publication of [6] Kaushal, V., and Yadav, R., "Learning successful
this article. implementation of Chatbots in businesses from B2B
customer experience perspective". Concurrency and
Funding Computation: Practice and Experience, 35(1), pp.
The authors received no financial support for the research, e7450, (2023). https://doi.org/10.1002/cpe.7450
authorship, and/or publication of this article. [7] Tahayori, B., Chini‐ Foroush, N., and Akhlaghi H.,
"Advanced natural language processing technique to
Reference predict patient disposition based on emergency triage
[1] Aslam, U., "Understanding the usability of retail notes". Emergency Medicine Australasia, 33(3),
fashion brand chatbots: Evidence from customer pp.480-484, (2021). https://doi.org/10.1111/1742-
expectations and experiences". Journal of Retailing 6723.13656
and Consumer Services, 74, pp.103377, (2023). [8] Sarker, I. H., "Machine learning: Algorithms, real-
https://doi.org/10.1016/j.jretconser.2023.103377 world applications and research directions". SN

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(4), 4759 - 4768 |4767
 computer science, 2(3), pp.160, (2021). Chatbot Using Deep Recurrent Neural Networks
https://doi.org/10.1007/s42979-021-00592-x Based on BiLSTM Model". In 2020 3rd
[9] Muzaffar, S., and Afshari, A., "Short-term load International Conference on Information and
forecasts using LSTM networks". Energy Procedia, Communications Technology (ICOIACT), pp. 382-
158, pp.2922-2927, (2019). 387, IEEE, (2020). DOI:
https://doi.org/10.1016/j.egypro.2019.01.952 10.1109/ICOIACT50329.2020.9332074
[10] Choudhary, P., and Chauhan, S., "An intelligent [18] Pandey, S., and Sharma, S., "A comparative study of
chatbot design and implementation model using long retrieval-based and generative-based chatbots using
short-term memory with recurrent neural networks Deep Learning and Machine Learning". Healthcare
and attention mechanism". Decision Analytics Analytics, 3, pp.100198, (2023).
Journal, 9, pp.100359, (2023). https://doi.org/10.1016/j.health.2023.100198
https://doi.org/10.1016/j.dajour.2023.100359 [19] Dhyani, M., and Kumar, R., “An intelligent Chatbot
[11] Tax, N., Verenich, I., La Rosa, M., and Dumas, M., using deep learning with Bidirectional RNN and
"Predictive business process monitoring with LSTM attention model”. Materials today: proceedings, 34,
neural networks". In Advanced Information Systems pp.817-824, (2021).
Engineering: 29th International Conference, CAiSE https://doi.org/10.1016/j.matpr.2020.05.450
2017, Essen, Germany, June 12-16, 2017, [20] Trojovská, E., Dehghani, M., and Trojovský, P.,
Proceedings 29, pp. 477-492. Springer International "Zebra optimization algorithm: A new bio-inspired
Publishing, (2017). https://doi.org/10.1007/978-3- optimization algorithm for solving optimization
319-59536-8_30 algorithm". IEEE Access, 10, pp.49445-49473,
[12] Aalipour, G., Kumar, P., Aditham, S., Nguyen, T., (2022). DOI: 10.1109/ACCESS.2022.3172789
and Sood, A., "Applications of sequence to sequence [21] Tran, A. D., Pallant, J. I., and Johnson, L.W.,
models for technical support automation". In 2018 "Exploring the impact of chatbots on consumer
IEEE International Conference on Big Data (Big sentiment and expectations in retail". Journal of
Data), pp. 4861-4869. IEEE, (2018). DOI: Retailing and Consumer service, 63, pp.102718,
10.1109/BigData.2018.8622395 (2021).
[13] Qader, W. A., Ameen, M. M., and Ahmed, B.I., "An https://doi.org/10.1016/j.jretconser.2021.102718
overview of bag of words; importance, [22] Assayed, S. K., Alkhatib, M., and Shaalan, K.,
implementation, applications, and challenges". In "Artificial intelligence based chatbot for promoting
2019 international engineering conference (IEC), pp. equality in high school advising". In 2023 4th
200-204, IEEE, (2019). DOI: International Conference on Intelligent Engineering
10.1109/IEC47844.2019.8950616 and Management (ICIEM), pp. 1-4, IEEE, (2023).
[14] Vijayakumar, A. K., Cogswell, M., Selvaraju, R. R., DOI: 10.1109/ICIEM59379.2023.10167112
Sun, Q., Lee, S., Crandall, D., and Batra, D., [23] Zhong, Q., Ding, L., Liu, J., Du, B., and Tao, D.,
"Diverse beam search: Decoding diverse solutions "E2s2: Encoding-enhanced sequence-to-sequence
from neural sequence models". arXiv preprint pretraining for language understanding and
arXiv:1610.02424 (2016). generation". IEEE Transactions on Knowledge and
https://doi.org/10.48550/arXiv.1610.02424 Data Engineering, (2023). DOI:
[15] Rana, A., Khurana, V., Shrivastava, A., Gangodkar, 10.1109/TKDE.2023.3341917
D., Arora, D., and Dixit, A. K., "A ZEBRA [24] Luong, M.T., Pham, H., and Manning,
Optimization Algorithm Search for Improving C.D.,"Effective approaches to attention-based neural
Localization in Wireless Sensor Network". In 2022 machine translation". arXiv preprint
2nd International Conference on Technological arXiv:1508.04025 (2015).
Advancements in Computational Sciences https://doi.org/10.48550/arXiv.1508.04025
(ICTACS), pp. 817-824, IEEE, (2022). DOI: [25] Zhao, R., and Mao, K., "Fuzzy bag-of-words model
10.1109/ICTACS56270.2022.9988278 for document representation". IEEE transactions on
[16] Ortiz-Garces, I., Govea, J., Andrade, R. O., and fuzzy systems, 26(2), pp.794-804, (2017). DOI:
Villegas-Ch, W., "Optimizing Chatbot Effectiveness 10.1109/TFUZZ.2017.2690222
through Advanced Syntactic Analysis: A [26] Freitag, M., and Al-Onaizan, Y., "Beam search
Comprehensive Study in Natural Language strategies for neural machine translation". arXiv
Processing". Applied Sciences, 14(5), pp.1737, preprint arXiv:1702.01806
(2024). https://doi.org/10.3390/app14051737 (2017).https://doi.org/10.48550/arXiv.1702.01806
[17] Anki, P., Bustamam, A., Al-Ash, H. S., and
Sarwinda, D., "High Accuracy Conversational AI

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(4), 4759 - 4768 |4768