Intelligent lead-based bidirectional long short term memory for COVID-19 sentiment analysis

Abstract

Social media is an online platform with millions of users and is utilized to spread news, information, world events, discuss ideas, etc. During the COVID-19 pandemic, information and ideas are shared by users both officially and by citizens. Here, the detection of useful content from social media is a challenging task. Hence, natural language processing (NLP) and deep learning are widely utilized for the analysis of the emotions of people during the COVID-19 pandemic. Hence, this research introduces a deep learning mechanism for identifying the sentiment of the people by considering the online Twitter data regarding COVID-19. The intelligent lead-based BiLSTM is utilized to analyze people's sentiments. Here, the loss of the classifier while learning the data is eliminated through the incorporation of the intelligent lead optimization. Hence, the loss is reduced, and a more accurate analysis is obtained. The intelligent lead optimization is devised by considering the role of the informer in identifying the enemy base to safeguard the territory from attack along with the Monarch's knowledge. The performance of the intelligent lead-based BiLSTM for the sentiment analysis is assessed using the metrics like accuracy, sensitivity, and specificity and obtained the values of 96.11, 99.22, and 95.35%, respectively, which are 14.24, 10.45, and 26.57% enhanced performance compared to the baseline KNN technique.

Introduction

The World Health Organization (WHO) proclaimed COVID-19 as a pandemic disease by considering the severity and the outrageous spreading that lost people's lives along with several diseases like respiratory failure, pneumonia, and various dangerous diseases. The coronavirus with severe respiratory syndrome is characterized as coronavirus 2 (SARS-CoV-2). As no appropriate treatment is available for COVID-19, several preventive measures like self-isolation, social distancing, hand hygiene, sanitizing, and several other public health practices are followed. The pandemic outbreak affects almost all nations that severely affect the economy globally. Therefore, the WHO devises several restrictions as corona affects severe damage to lives; hence, the Public Emergency of International Concern (PEIC) is declared. The restrictions lead to the closure of several businesses and people traveling. Thus, people started to share their views on social media regarding the happenings in the world because the pandemic makes social media more noticeable than before. Hence, the impact of the pandemic resulted in social media platforms as a big global data center, and people started to spend more time on these platforms and their applications. The people's reactions and conditions during the corona pandemic are analyzed using Twitter's social media platform. The people's emotions analysis by considering the products, disaster management, crime, election, stock market, and several other criteria are employed using Twitter data. Thus, the people's changes and reactions to COVID-19 are analyzed using Twitter data by evaluating the sentiment (Choudrie et al. 2021).

The detection of people's sentiments by considering the Twitter data under three categories, neutral, positive, and negative, is termed sentiment analysis. Sentiment analysis is widely utilized in several application areas like health informatics, recommendation systems, opinion mining, etc. For example, while considering the business, profit enhancement is employed by providing the best services through strategic planning, which is offered by detecting consumer opinions through various sentiments. Likewise, the health status of the person, psychology, and behaviors are analyzed through the Tweets. In addition, the information acquisition in the text format is employed in several other ways, like the posts on Facebook social media for analyzing the sentiment. Here, the analysis of the sentiment from the textual data is employed through the natural language processing technique, in which the features are extracted through subjective knowledge. The information obtained from electronic devices is either consumer-generated or user-generated content for sentiment analysis. The view of each person varies for the information obtained from Twitter and other social media, hence the consolidation through evaluation and analysis for expanding the perceptions (Saifullah et al. 2021). Thus, the claims concerning the consumer's feelings regarding the products, negative or positive, are obtained by analyzing the specific subject based on evidence. However, analyzing the sentiment from the online data by considering the keyword is still challenging. Thus, the machine learning technique is utilized for extracting the sentiment from the online data for evaluating the different text analyses (Elgeldawi et al. 2021).

NLP and deep learning are widely utilized for the analysis of the emotions of people during the COVID-19 pandemic (Kaur et al. 2021; Kumari et al. 2021). Text-based sentiment analysis is employed through emotion detection and recognition in the artificial intelligence (AI) arena (Chole and Gadicha 2022). The goal behind the sentiment analysis is to detect the person's feelings from the textual data through the detection and recognition stages for evaluating expressions like surprise, sadness, happiness, fear, disgust, and anger. The consumer's perception or happiness of purchasing the product is evaluated through the detection of emotion for business decision-making. The techniques like XG-boost, Random Forest, Support Vector Classifier, Decision Tree Classifier, Bernoulli, and K-Nearest Neighbor (K-NN) are some of the machine learning techniques (Chole and Gadicha 2020) utilized for emotion recognition. The significant attributes are fed as input to the classifier for processing the text data obtained online, in which the deep learning technique learns the features automatically by adjusting its parameters. The non-self-adjustable parameters are termed hyper-parameter and are adjusted through optimization-based tuning to obtain the desired output. Thus, the hyper-parameter decides the choice of the sentiment analysis output based on deep learning.

The research aims to devise a novel sentiment analysis by considering the online Twitter data regarding COVID-19 for analysis of people's sentiments. Here, the optimized deep learning is designed for the sentiment analysis of the COVID-19 data is proposed. Furthermore, the novel intelligent lead optimization is designed for tuning the bidirectional long short-term memory (BiLSTM), in which the tunable hyper-parameters are tuned using the intelligent lead algorithm. Initially, from the online COVID-19 data, by using the essential keyword, the COVID-19 data are acquired from the input data for the reduction in computational complexity. Then, by using the distance-based similarity measure, the consumers with similar features are grouped using the graph-based strategy. Finally, using the BiLSTM, the sentiments like positive, negative, and neutral are employed. The contributions are:

• Proposed Intelligent lead optimization: The intelligent lead optimization is designed by considering the role of the informer in identifying the enemy base to safeguard the territory from attack along with the Monarch's knowledge. Thus, the proposed intelligent lead optimization based on the hybrid behavior improves the convergence rate.

• Proposed Intelligent lead-based BiLSTM: The intelligent lead-based BiLSTM is utilized for analyzing people's sentiments during COVID-19 by considering the online Twitter data. Here, the loss of the classifier while learning the data is eliminated through the incorporation of the intelligent lead optimization; hence, the loss is reduced, and a more accurate analysis is obtained.

The organization of the intelligent lead-based sentiment analysis is organized as Sect. 2 details the conventional COVID-19-based sentiment analysis methods, and the detailed methodology is elucidated in Sect. 3. Finally, the analysis is presented in Sect. 4, and Sect. 5 concludes the work.

Motivation

The inaccurate identification of the sentiment and the complexities in finding the best solution are the most challenging issues in traditional sentimental analysis. Thus, this research introduces a novel technique to fulfill the challenges by reviewing the related works by the existing researchers, which are detailed in this section.

Related works

The sentimental analysis (SA) concerning the COVID-19 situation is detailed in this section and highlighted in Table 1. The SVM-based SA with the Twitter data was presented by Basiri et al. (2021) using the R programming technique, in which the keyword was utilized for identifying the sentiment based on the feature selection. Then, the score was assigned for the classified sentiment to analyze the polarity. The better performance was evaluated based on accuracy and obtained enhanced performance. The analysis depicts the positive sentiment sometimes and the negative sentiment daily basis. The SVM-based SA devised by Sontayasara et al. (2021) using the shuffle split cross-validation with the data processing acquired an accuracy of 71% using the linear kernel, which is higher than the sigmoid kernel. However, they failed to analyze the performance using large data. The NLP-based SA using Twitter data was suggested by Garcia and Berton (2021) and Hitesh et al. (2019) through sentence embedding and word embedding. Here, the summarization of the text was employed from the extracted features based on the topic modeling and obtained the negative sentiment during the pandemic's peak. This method failed to consider the keyword to retrieve the content which leads to the loss of some information, which is considered a drawback of the method.

Table 1.

Related works

S. no	Author	Method	Advantage	Disadvantage
1	Basiri et al. (2021)	A hybrid deep learning technique	Complexity was reduced	Failed in selecting the keyword
2	Sontayasara et al. (2021)	The SVM-based SA	High accuracy in the data processing	Failed to analyze the performance using large data
3	Garcia and Berton et al. (2021)	NLP-based SA using the Twitter data	Enhanced the performance	Loss of some information
4	Shamrat et al. (2021)	KNN-based SA	Time efficient	The prediction stage might be slow
5	Chowdhury et al. (2021)	Data tokenization, normalization, and lemmatization are based on KNN	Versatile to different calculations of proximity	Inefficient and it's difficult to pick the correct value
6	Sitaula et al. (2021)	Deep learning-based SA using the CNN	Very High accuracy	Failed to consider the essential attributes
7	Mishra et al. (2021)	SA based on deep learning	Helped to uncover the required topic from the corpus	Failed to consider the significant aspects like economy, transportation
8	Abdulaziz et al. (2021)	The topic modeling using the Latent Dirichlet Allocation (LDA)	Minimized loss function	Failed to evaluate the performance

The KNN-based SA concerning the vaccination of COVID-19 was devised by Shamrat et al. (2021) for identifying three different sentiments using Twitter data (Chowdhury et al. 2021). In this, the data tokenization, normalization, and lemmatization were performed to remove noise, and the polarity was identified for the sentiment analysis using the KNN. The method analyzed the sentiment for three vaccines based on negative, neutral, and positive sentiments. The deep learning-based SA using the CNN was presented by Sitaula et al. (2021), in which the extraction of the informative attributes is employed using domain-specific, cross-domain, and fastText data, and the sentiment was analyzed separately using three various CNN. Finally, the decision taken by the three methods was combined to find the sentiment. The method obtained poor accuracy due to the failure to consider the essential attributes. SA based on deep learning was presented by Mishra et al. (2021) using long short-term memory (LSTM) by considering the rule-based sentiment estimation, in which the dictionary was utilized for estimating sentiment. In addition, the topic modeling helped to uncover the required topic from the corpus. It was evaluated based on accuracy, which helps to identify citizens' sentiments concerning tourism during COVID-19. However, while modeling the topic, the method failed to consider significant aspects like economy, transportation, and other factors to obtain a more accurate analysis.

The SA with the topic modeling using the Latent Dirichlet Allocation (LDA) was suggested by Abdulaziz et al. (2021) and Monish et al. (2018) for obtaining the informative features by considering the large-scale data. Here, the SA was employed using the lexicon-based strategy, in which the polarity was evaluated by clustering the words into positive and negative groups. Then, the score was generated to identify the sentiment. Unfortunately, the method failed to evaluate the performance. A hybrid deep learning technique devised by combining five different deep learning techniques was presented by Basiri et al. (2021) using the Tweets obtained from social media. In this, the opinions of the people were identified using the metal learner, in which the decision regarding the sentiments for the tweets was fused for obtaining the stacked generalization capability, and the complexity was reduced using the iterative gradient boosting technique. As a result, the method obtained positive sentiment while the recovery rate is higher and evaluated higher negative sentiment during the high death charts that depict the method's accuracy. The method failure's drawback is selecting the keyword for analyzing the sentiment by considering individual countries, which may enhance the analysis accuracy further.         

Challenges

The challenges faced by the existing methods of sentimental analysis by considering the COVID-19 data are:

• The method devised by Garcia and Berton (2021) failed to consider the keyword to retrieve the content, which led to the loss of some information, which is considered a drawback of the method (Basiri et al. 2021).

• The poor performance was obtained by the method (Sitaula et al. 2021) due to the failure to consider the essential attributes that assist in elevating the accuracy of the method.

• The method (Sontayasara et al. 2021) failed to analyze the performance using large data. The insufficiency of labeled data and the inefficiency in evaluating complex sentences degrades the system's performance (Abdulaziz et al. 2021).

• Identifying the sentiment from the sentences with both positive and negative sentiment and detect as neutral is challenging (Abdulaziz et al. 2021).

• The higher computational complexity is due to the failure to extract the essential attributes before the sentimental analysis.

Proposed methodology for COVID-19 sentiment analysis

The prediction and the analysis of people's thinking are obtained using the social media platform as the issues concerning personal concerns, societal issues, and thoughts are shared by citizens. Thus, the prediction of the sentiments concerning COVID-19 among the people is analyzed using the Twitter social media data in the sentiment analysis method. Let the input data concerning the COVID-19 information be obtained from the Kaggle database to evaluate the sentimental analysis. The sentences on the COVID-19 information are obtained by searching the sentences with the significant keywords corresponding to COVID-19. Then, the graph-based sentence tracking is performed based on the distance measures with Euclidian and cosine similarity measures. Finally, using the tracked information, the sentiment analysis is performed using the BiLSTM, which is tuned using the intelligent lead algorithm. Here, the informer and the Monarch's role in safeguarding the territory by identifying the enemy base is considered for designing the Intelligent Lead algorithm. Figure 1 depicts the block diagram of the intelligent Lead based BiLSTM for the sentiment analysis of COVID-19 data.

Fig. 1.

Block diagram of intelligent lead-based BiLSTM for sentiment analysis using COVID-19 data

Data acquisition

The input data for the sentiment analysis are acquired from the Kaggle COVID-19 database for the sentiment analysis (COVID-19,2022; SemEval 2022)). Let, the database be denoted as $C$ with data samples corresponding to the sentiments of COVID-19 and expressed as,

$$C=\{C_a\};(1\le a\le n)$$ 1

where the total sample in the database is notated as $n$, and $a\text{th}$ sample in the database is notated as $C_a$.

Keyword searching from the tweets

The keywords are searched from the input text regarding the COVID-19 sentiment for selecting the sentences associated with COVID. Thus, the COVID-19-based sentiment analysis using the Intelligent Lead Optimization-based BiLSTM searches for the keywords in the data corpus as provided in Table 2.

Table 2.

Keywords utilized for searching the COVID-19-based Tweets from Kaggle

Domain	Keywords
Treatments	Hydroxychloroquine, plasma, treatment, trial, drug, hospital, patient, and vaccine
Entertainment	Show, love, video, like, and watch
Politics	Minister, state, Fauci, president, response, government, and Trump
Proliferation care	Reopen, spread, social, test, school, people, wear, stay at home, and mask
Reports of case	Confirm, update, record, patient, total, people, number, die, report, death, and case
Economic aspects	Pay, crisis, business, impact, and work
Online events	Host, tomorrow, virtual, impact, discuss, talk, live, webinar, and join
Anti-racism	War, right, die, death, American, black, kill, police, and protest
Sports	Sport, play, game, league, team, football, season, and player
Charity	Million, food, community, provide, donate, relief, fund, help, and support

Thus, from the input data, the Tweets with essential attributes presented in Table 2 are selected for identifying the inter-word relationship for tracking COVID-19.

COVID tracking using graph-based sentiment tracking

COVID-19 is tracked by evaluating the sentiment tracking between the words based on the graph-based strategy. Instead of taking into consideration all the words, using graphs for keyword extraction offers a superior situation for finding the main group of terms. The Vector Space Model (VSM) and the Graph-based model are two well-liked methods for extracting keywords from the text. The given dataset is turned into a network for keyword extraction by treating each word as a node. Two words are connected by an edge to convey a connection if they co-occur in the dataset. Numerous centrality metrics may be used on graphs to identify the key phrases that should be utilized as the document's keywords. The following is a discussion of two of the common centrality measurements used in graphs cosine similarity and euclidean distance. Here, the distance between the words is estimated for the tracking of the sentiment based on the graph. The distances like the cosine similarity and the Euclidean distance are utilized for the evaluation of the distances and to form the graph for tracking COVID-19. Figure 2 depicts the graph representation for sentiment tracking based on distance measures. Based on the distance measures, the sentences with three different sentiments positive, negative, and neutral, are tracked (Wang et al. 2020).

Fig. 2.

Graph representation for tracking the sentiment

Euclidean distance (Dis₁)

The distance between the two points is evaluated using the Euclidean measure, in which the line segment between the two points is obtained for tracking the sentiments. The Euclidean distance measure between the sentences is evaluated for tracking the sentences with the sentiments of COVID-19 and is expressed as,

$${\text{Dis}}_1(u,v)=\sum _{x=1}^n{(u_x-v_x)}^{1/2}$$ 2

where the two different sentences are referred to as $u$ and $v$, and the Euclidean distance is notated as ${\text{Dis}}_1$. Here, the total sentence is notated as $n$.

Cosine similarity (Dis₂)

The cosine similarity between the sentences is measured between the two different sentences for tracking the sentiment regarding COVID-19 by considering the angle between the two vectors and is expressed as,

$${\text{Dis}}_2(u,v)=\frac{u+v}{∥u∥+∥v∥}$$ 3

where the cosine similarity is notated as ${\text{Dis}}_2$.

Thus, based on the two different distance measures, the similarity between the COVID-19 sentiments sentences are identified based on the graphical representation.

Sentiment classification using intelligent lead optimization-based BiLSTM

The sentiment classification by considering COVID-19 using the intelligent Lead based BiLSTM (Li et al. 2021) here the classifier is trained using the intelligent lead algorithm to reduce the loss that occurs while training the classifier for making generalizations while testing the unknown data to enhance the accuracy of the analysis. Computation efficiency and reliability can be impacted by a model’s design and choice of configuration parameters. With default parameter settings, the majority of models operate precisely and effectively, but some models perform better when the solver parameters are changed. When giving the keywords, knowledge of a model’s behavior can help you execute simulations more effectively. The optimization algorithm improves the proposed model's comprehension and updates the model's parameters to increase efficiency and precision. In order to retrieve only the data pertinent to the performance, it is necessary to filter out particular keywords from the data. By measuring the distance cosine similarity and euclidean distance, the text similarity-based method is used to visualize opinions in order to achieve that. The sentiments polarities like the neutral, positive and negative are evaluated using the intelligent Lead-based BiLSTM by considering the keywords. A detailed explanation is provided in the following subsections.

System model of BiLSTM for sentiment classification

The sentiment analysis using the recurrent neural network (RNN_-based technique provided enhanced results while considering the continuous data, but the vanishing gradient issue still exists that fails to update the weights, and hence the learning process can't be able to complete, which makes the performance degradation in the sentiment analysis process. Thus, for solving the issue regarding the vanishing gradient, long-term dependencies among the data are employed with the inclusion of a memory unit and are named as long short-term memory (LSTM). Besides, the time-series data classification for classifying the sentiment is obtained efficiently using the LSTM. In addition, the inclusion of the bidirectional layers named forward layer and backward layer in the LSTM constitutes the bidirectional LSTM (Bi-LSTM) for considering both the future and past context because the LSTM propagates the data from back to front. Still, it fails to process the data from front to back, which limits the learning capability. Thus, the BiLSTM is utilized for classifying the sentiment in the sentiment analysis technique. The deep learning methods comprise several layers like input, hidden, output, softmax, and activation function, in which the loss associated with the training limits the performance. Thus, the optimization is incorporated for tuning the weights and bias of the classifier optimally to reduce the loss while training the classifier. Here, the need for training is to learn the classifier with appropriate data to generalize to predict the sentiment more efficiently while providing the unknown test data. In this condition, when the loss occurs while training the classifier, the system's performance gets degraded. So to minimize the loss in training and to speed up the learning process optimization algorithm is utilized. In the sentiment classification, Intelligent Lead optimization is utilized for training the BiLSTM. The system model of the BiLSTM is depicted in Fig. 3.

Fig. 3.

System model of BiLSTM

Let $R_i,P_i\text{and}{K}_i$ be the output, input, and forget gates of the BiLSTM utilized for the sentiment analysis of COVID-19, which has modifiable parameters like $N$ and $W$ that depict the weight and bias. The multiplication process that is performed element-wise is notated as $\ast$ and the point-based multiplication is represented as $\bullet$, respectively.

The input gate processes the input and provides the output by considering the previous state information ($e_{i-1}$) and the current input ($R_i$), which is represented as,

$$R_i=\text{sigmoid}\left(N_R·\left[e_{i-1},Y_i\right]+W_R\right)$$ 4

where the weight is notated as $Y$, and the bias is notated as $W$.

The role of the forget gate is to discard or retain the information by considering the ($e_{i-1}$) and present input and its output is represented as

$$P_i=\text{sigmoid}\left(N_P·\left[e_{i-1},Y_i\right]+W_P\right)$$ 5

The output of the memory cell utilizes the input gate and the forget gate's output and processes the output as $M_i$,

$$\overline{M_i}=tanh\left(N_M·\left[e_{i-1},Y_i\right]+W_M\right)$$ 6

$$M_i=P_i\ast m_{i-1}+R_i\ast \overline{m_i}$$ 7

The controlling function of the memory cell is employed by the output layer and hence the control function is represented as,

$$K_i=\text{sigmoid}\left(N_K·\left[e_{i-1},Y_i\right]+W_K\right)$$ 8

Finally, the sentiment analysis by considering the current input is presented as,

$$e_i=K_i\ast tanh\left(M_i\right)$$ 9

The sentiment analysis by considering the two additional layers, namely the forward and the backward layer constitute the BiLSTM, in which ($\overleftarrow{e_i}$) and ($\overrightarrow{e_i}$) refers to the backward and forward layer of the BiLSTM and are represented as,

$$\overrightarrow{e_i}=\text{LSTM}\left(Y_i,\overrightarrow{e_{i-1}}\right)$$ 10

$$\overleftarrow{e_i}=\text{LSTM}\left(Y_i,\overleftarrow{e_{i-1}}\right)$$ 11

$$G_i=a_i\overrightarrow{e_i}+b_i\overleftarrow{e_i}+W_i$$ 12

The sentiment analysis by considering the COVID-19 data in the forward and the backward layer encoding provides the enhanced solution compared to the standard LSTM. The tunable weights and bias $Y\text{and}W$ are tuned using the Intelligent lead optimization algorithm.

Proposed intelligent lead optimization algorithm

In ancient days, the role of an informer is crucial in locating the enemy base and is updated to the king for safeguarding the country. Thus, the role of an informer is crucial for safeguarding the country and the quality of the information acquired by the informer should be higher to safeguard the country. Thus, the role of the informer (Pambudi and Kawamura 2022) along with the Monarch (Soradi-Zeid et al. 2020) is considered for designing the Intelligent Lead algorithm.

Motivation

The Monarch is considered sovereign over the territory and is the protector of the citizens of the country. The role of the Monarch is to protect the territory from the enemies, punish the misbehavers, and maintain the dharma. While considering the decision-making, Monarch discusses the particular topic with the ministers and provincial heads and gets their ideas, and implements the best idea for the betterment of citizens in the territory. In the intelligent lead algorithm, the role of safeguarding the territory is considered as it is considered as a crucial factor for living in a safer environment. Thus, for safeguarding the territory, Monarch appoints a group of informers (spies) under the head of a minister to locate the enemy base, who intrudes on the territory to capture the territory and plunder it. Here, the role of informers plays an essential role in safeguarding the country. The informer's moves to identify the enemy base, which is unknown initially based on three different movements. They are:

• Random search This is a non-cooperative movement, in which the informer explores more areas in the search space to obtain the global best solution, but this informer is a low-rank informer.

• SwingMove search This is also a non-cooperative movement, in which the informer exploits deeply inside the search space to locate the enemy base. Here, the informer moves with his location within a small perimeter and is a high-ranked informer with high-quality information concerning the enemy base.

• MoveToward search This is a cooperative movement, in which the movement of the informer is based on another informer and is a mid-ranked informer. In this movement, the informer initially explores more areas and after some iteration, he enters into the exploitation phase, when he is closer to the solution (enemy base).

Here, for the optimization algorithm, the solution is referred to as an informer, the optimal best solution for solving the optimization issue is the detection of an unknown enemy base and the location of the informer is the solution vector that is evaluated based on the fitness function. The information obtained only by the high-rank and mid-rank informer makes the solution converge at a local solution. Thus, to enhance the search space, the role of low ranked informer based on the random search is utilized. In addition, safeguarding the territory is employed by the Monarch based on the information obtained from the informer. Thus, in the decision-making to safeguard the territory, the combined role of an informer and the Monarch are considered and hence the role of both are hybridized in the intelligent lead optimization to obtain the global best solution for tuning the parameters of the classifier.

Mathematical modeling

The intelligent lead algorithm is detailed in this section with its mathematical modeling.

Initialization: The maximal number of iterations in the intelligent lead optimization is referred as $i_{max}$, and the dimension of the search space is notated as $S$. Let the variable $B$ is utilized for defining the objective function and is expressed as $B=\left(b_1,b_2,...b_S\right)$, and the total number of informers is the same as the population size and is referred to as $D$. Here, $B^{\alpha }\text{and}f\left(B^{\alpha }\right)$ refers to the position of the informers and the corresponding value, respectively, in which $\alpha =1,2,3,...,D$.

Fitness evaluation: The fitness function is utilized for guiding the optimization to find the global best solution by estimating the goodness of the position updated by the search agents in the search space. Here, in the intelligent lead optimization, the mean square error is evaluated to estimate the fitness and is expressed as,

$$\text{IL}{}_{\text{fit}}=\frac{1}{a}\sum _{i=1}^a{\left(\text{Predicted}\text{value}-\text{Actual}\text{value}\right)}^2$$ 13

where the fitness function of the intelligent lead optimization is notated as $\text{IL}{}_{\text{fit}}$, and the total samples are notated as $a$.

Here, the fitness is evaluated for all the informers in the search space and the informer with the least fitness is considered as an informer with the highest quality of information regarding the enemy-base identification. Thus, based on fitness all the informers are ranked, in which the position of the informer is not changed until the informer's position is referenced by another informer.

Exploration phase: In the exploration phase, the informer with the lowest rank among all the informer searches in the search space randomly explores more areas to analyze the features in the feature space. Here, the random search movement is followed by the informer.

Phase switching: In this phase, the move toward criteria is followed, in which the cooperative movement is employed, which means one informer follows the other informer to identify the enemy base. Initially, the informer explores more area in the search space and then identifies an informer $B^m$ and move toward him to find the enemy base and the position updation is represented as,

$${\left[B^{\alpha }(i+1)\right]}_{\text{informer}}=B^{\alpha }(i)+c[-1,1]\left[B^m(i)-B^{\alpha }(i)\right]$$ 14

where the position of the $\alpha \text{th}$ informer is updated as $B^{\alpha }(i+1)$ for the iteration $(i+1)$, $B^m(i)$ refers to the randomly selected best informer for entering into the exploiting phase. Here, the position of the informer is enhanced by hybridizing the Monarch knowledge with the informer's best informer selection in a random manner and is expressed as,

$${\left[B(i+1)\right]}_{\text{Monarch}}=B_{\text{best}}+E\ast \left(B_{\text{rand1}}-B_{\text{rand2}}\right)$$ 15

$$\left[B(i+1)-B(i)\right]=B_{\text{best}}+E\ast \left(B_{\text{rand1}}-B_{\text{rand2}}\right)-B(i)$$ 16

where the best knowledge of the Monarch is notated as $B_{\text{best}}$, the position of the Monarch is notated as $\left[B(i+1)\right]$, and the two randomly selected ideas of the Monarch to choose the best informer randomly is notated as $B_{\text{rand1}}\text{and}{B}_{\text{rand2}}$, respectively. The term $E$ refers to the coefficient vector utilized for updating the Monarch knowledge. Then, the fractional derivative concept is applied for the knowledge updation of the Monarch and is expressed as,

$$B(i+1)-\delta B(i)-\frac{1}{2}B(i-1)=B_{\text{best}}+E\ast \left(B_{\text{rand1}}-B_{\text{rand2}}\right)-B(i)$$ 17

Then, the position updation of the informer with the Monarch's knowledge in selecting the random informer for the intelligent lead optimization is expressed based on the rule (Binu and Kariyappa 2020) as,

$${\left[B(i+1)\right]}_{\text{IL}}=0.5{\left[B(i+1)\right]}_{\text{Monarch}}+0.5{\left[B(i+1)\right]}_{\text{Informer}}$$ 18

where the position updation based on the intelligent lead optimization is notated as ${\left[B(i+1)\right]}_{\text{IL}}$, the position of the informer is notated as ${\left[B(i+1)\right]}_{\text{Informer}}$, the knowledge of the Monarch is notated as ${\left[B(i+1)\right]}_{\text{Monarch}}$.

$$\begin{array}{ccc}\hfill {\left[B(i+1)\right]}_{\text{IL}}& =0.5\left[B^{\alpha }(i)+c[-1,1]\left[B^m(i)-B^{\alpha }(i)\right]\right]\hfill & \hfill +\\ \hfill 0.5\left[B(i)\left(\delta -1\right)+\frac{1}{2}\delta B(i-1)+B_{\text{best}}+E\ast \left(B_{\text{rand1}}-B_{\text{rand2}}\right)\right]\\ \hfill \end{array}$$ 19

$${\left[B(i+1)\right]}_{\text{IL}}=\frac{1}{2}\left\{B(i)\left[\delta -c(-1,1)\right]+B(i)c(-1,1)+\frac{1}{2}\delta B(i-1)+B_{\text{best}}+E\ast \left(B_{\text{rand1}}-B_{\text{rand2}}\right)\right\}$$ 20

Thus, using the intelligent lead optimization the feature space is explored randomly to obtain the global best solution among the search space in terms of detecting the enemy base. Here, the incorporation of the Monarch knowledge in identifying the enemy base with the informer's knowledge assists to identify the best solution.

Exploitation phase: In the exploitation phase, the SwingMove is adapted by the informer for identifying the enemy base. Here, the position of the informer is updated as,

$$B^{\alpha }(i+1)=B^{\alpha }(i)+c(-1,1)\left[\frac{Q}{i}\right]$$ 21

The position of the informer is further strengthened by incorporating the knowledge of the King in decision-making, which is provided in Eq. (21). Thus, the enemy-base identification is safeguarding the territory using the intelligent lead optimization as per the rule (Binu and Kariyappa 2020) is expressed as,

$$\begin{array}{ccc}\hfill {\left[B(i+1)\right]}_{\text{IL}}& =0.5\left[B^{\alpha }(i)+c(-1,1)\left[\frac{Q}{i}\right]\right]\hfill & \hfill +\\ \hfill 0.5\left[B(i)\left(\delta -1\right)+\frac{1}{2}\delta B(i-1)+B_{\text{best}}+E\ast \left(B_{\text{rand1}}-B_{\text{rand2}}\right)\right]\\ \hfill \end{array}$$ 22

The position updation of the informer using the intelligent lead optimization is notated as,

$${\left[B(i+1)\right]}_{\text{IL}}=\frac{1}{2}\left\{\delta B(i)+c(-1,1)\left[\frac{Q}{i}\right]+\frac{1}{2}\delta B(i-1)+B_{\text{best}}+E\ast \left(B_{\text{rand1}}-B_{\text{rand2}}\right)\right\}$$ 23

Thus, the enemy-based identification in safeguarding the territory by the combined knowledge of Monarch and the informer ensures more accurate detection with a fast convergence rate. Besides, the in-depth analysis of the exploitation phase helps to identify the local best solution for solving the optimization issues. Here, in the intelligent lead optimization, the balanced exploration and exploitation phase in identifying the global best solution and the local best solution ensures the more accurate solution for solving the optimization issues in detecting the sentiment of the COVID-19 data.

Fitness re-evaluation: After identifying the enemy base by the informer, the fitness is re-evaluated to find the correctness of the solution obtained by the current iteration. If the present solution is best, then the position is updated using the present global best solution, else remains in the old position.

Stopping criterion: The above-mentioned steps are repeated until $i_{max}$. The pseudo-code for the intelligent lead optimization is presented in Algorithm 1.

Algorithm 1 Pseudo-code for the intelligent lead algorithm.

Pseudo-code for the intelligent lead algorithm
1	Begin:
2	Initialize the parameters: swing factor $Q$, iteration $i_{max}$, and population size D
3	Exploration:
4	Generate the initial solution $I$
5	Evaluate the fitness
6	Sort according to the ranks //*The highest-ranked informer is the one with minimal fitness value
7	for
8	{
9	$i=1\text{to}{i}_{max}$
10	Low-ranked informer explores to identify the enemy base based on a random search $I$
11	Mid-ranked informer follows another informer by considering the Monarch knowledge to identify the enemy base using MoveToward movement
12	High-ranked informer tries to locate the enemy base based on the Monarch Knowledge using the SwingMove movement
13	Re-evaluate the fitness
14	Update the position of the low-ranked informer based on the exploration phase using Eq. (20)
15	Update the position of the mid-ranked informer based on the phase switching using Eq. (14)
16	Update the position of the high-ranked informer based on the exploitation phase using Eq. (23)
17	Rank based on the fitness
18	$i=i++$
19	}
20	Return the best solution
21	end

Results and discussion

The MATLAB tool is utilized for the implementation of the sentiment analysis method using intelligent Lead-based BiLSTM using Windows 10 OS with 8 GB RAM.

Description of dataset

The dataset utilized for evaluating the performance is taken from Kaggle open source for sentimental analysis (COVID-19,2022). It comprises of 9.17 MB data with user descriptions like location, name, and ID. The data corresponding to the COVID-19 vaccine are obtained from the Tweets of the massive population across 180 countries that express the feelings of the people.

Laptop and restaurant data: The data taken from the restaurant and laptop trail data are obtained from SemEval-2014 Task-4 dataset (2022). In this, the reviews concerning the laptops and restaurants with two different polarities are included and are utilized for the assessment of the proposed method.

Assessment metrics

The specificity, sensitivity, and accuracy are evaluated for assessing the performance of the intelligent Lead based BiLSTM technique.

Specificity: For the sentiment analysis by the intelligent Lead-based BiLSTM, the capability of predicting the negative sentiments from the available data is termed specificity and is expressed as,

$${\text{SA}}_{\text{spe}}=\frac{{\text{SA}}_{\text{tn}}}{{\text{SA}}_{\text{tn}}+{\text{SA}}_{\text{fp}}}$$ 24

Sensitivity: For the sentiment analysis by the intelligent Lead-based BiLSTM, the capability of predicting the positive sentiments from the available data is termed sensitivity and is expressed as,

$${\text{SA}}_{\text{sen}}=\frac{{\text{SA}}_{\text{tp}}}{{\text{SA}}_{\text{tp}}+{\text{SA}}_{\text{fn}}}$$ 25

Accuracy: The closeness of the sentiments estimated by the intelligent Lead based BiLSTM with the target is termed accuracy and is expressed as,

$${\text{SA}}_{\text{acc}}=\frac{{\text{SA}}_{\text{tn}}+{\text{SA}}_{\text{tp}}}{{\text{SA}}_{\text{tn}}+{\text{SA}}_{\text{tp}}+{\text{SA}}_{\text{fp}}+{\text{SA}}_{\text{fn}}}$$ 26

where the true positive is notated as ${\text{SA}}_{\text{tp}}$, the true negative is notated as ${\text{SA}}_{\text{tn}}$, the false positive is notated as ${\text{SA}}_{\text{fp}}$, and the false negative is notated as ${\text{SA}}_{\text{fn}}$.

Conventional methods

The conventional methods utilized for the comparative assessment of the intelligent lead-based BiLSTM are KNN (Shamrat et al. 2021), stochastic gradient descent-based neural network (SGD-NN) (Akçay 2020), Random Forest (Neogi et al. 2021), TD-LSTM (Wang et al. 2016), ATAE-LSTM (Tang et al. 2015), BiLSTM (Li et al. 2021), Spy based BiLSTM (Pambudi and Kawamura 2022), and King based BiLSTM (Soradi-Zeid et al. 2020).

Comparative assessment

The comparative assessment of the intelligent Lead based BiLSTM for the sentiment analysis is employed by varying the training data.

Assessment using twitter data

Figure 4 portrays the assessment of the intelligent Lead-based BiLSTM using the Twitter data by varying the training data, in which the accuracy is presented in Fig. 4a, the sensitivity is presented in Fig. 4b and the specificity is presented in Fig. 4c. The accuracy acquired by the proposed method with 70% of training data is 92.40%, which is 14.73% enhanced performance compared to the 40% training data. Besides, the proposed method with the same training data is 16.74, 15.39, 15.59, 17.57, 12.71, 14.71, 13.64, and 12.08% superior to the KNN, SGD-NN, Random Forest, TD-LSTM, ATAE-LSTM, BiLSTM, Spy algorithm based BiLSTM, and King algorithm based BiLSTM conventional methods. Table 3 depicts the detailed analysis of Fig. 4.

Fig. 4.

Assessment using Twitter data based on a accuracy, b sensitivity, and c specificity

Table 3.

Assessment using Twitter data

Methods/Training percentage	KNN	SGD-NN	Random forest	TD-LSTM	ATAE-LSTM	BiLSTM	Spy algorithm based BiLSTM	King algorithm-based BiLSTM	Intelligent lead based BiLSTM
Accuracy
40	74.51	76.33	75.42	72.98	78.15	77.23	77.27	78.60	78.79
50	75.53	77.01	76.73	74.05	78.78	77.33	77.96	79.48	80.01
60	76.70	77.97	77.72	74.75	79.38	78.23	79.00	80.75	91.77
70	76.93	78.18	78.00	76.17	80.66	78.80	79.80	81.24	92.40
80	77.63	78.44	78.20	76.47	82.20	79.17	79.88	85.51	95.51
Sensitivity
40	82.02	83.86	82.79	81.14	85.92	85.12	85.16	86.57	86.87
50	83.27	85.19	84.79	81.46	87.05	85.21	86.24	88.00	88.74
60	84.80	86.59	86.30	82.60	88.04	86.88	87.89	89.60	91.07
70	84.86	86.92	86.64	83.75	89.46	87.94	89.16	90.20	91.51
80	86.02	87.29	86.87	84.08	89.69	88.55	89.25	96.27	96.63
Specificity
40	67.74	69.56	68.81	65.54	71.16	70.11	70.15	71.43	71.50
50	68.55	69.60	69.44	67.39	71.31	70.23	70.46	71.76	72.08
60	69.37	70.12	69.91	67.64	71.52	70.36	70.90	72.71	93.39
70	69.77	70.23	70.13	69.35	72.66	70.45	71.24	73.09	94.22
80	70.02	70.38	70.31	69.62	75.53	70.59	71.32	75.61	95.35

Assessment using laptop data

Figure 5 portrays the assessment of the intelligent Lead based BiLSTM using the laptop data by varying the training data, in which the accuracy is presented in Fig. 5a, the sensitivity is presented in Fig. 5b and the specificity is presented in Fig. 5c. The sensitivity acquired by the proposed method with 60% of training data is 87.06%, which is 1.87% enhanced performance compared to the 40% training data. Besides, the proposed method with the same training data is 5.25, 5.14, 7.20, 5.82, 2.32, 4.82, 1.73, and 0.74% superior than the KNN, SGD-NN, Random Forest, TD-LSTM, ATAE-LSTM, BiLSTM, Spy algorithm based BiLSTM, and King algorithm based BiLSTM conventional methods. Table 4 depicts the detailed analysis of Fig. 5.

Fig. 5.

Assessment using laptop data based on a accuracy, b sensitivity and c specificity

Table 4.

Assessment using laptop data

Methods/Training percentage	KNN	SGD-NN	Random forest	TD-LSTM	ATAE-LSTM	BiLSTM	Spy algorithm based BiLSTM	King algorithm based BiLSTM	Intelligent lead based BiLSTM
Accuracy
40	77.34	77.57	73.75	77.06	79.24	78.09	80.31	80.48	90.13
50	78.57	86.40	76.97	77.78	80.74	79.02	81.62	82.33	90.99
60	78.77	79.01	77.44	78.32	81.25	79.43	81.90	82.47	91.56
70	84.20	80.23	77.88	78.83	82.21	81.26	82.41	83.82	92.53
80	79.65	80.69	78.49	78.95	83.40	82.83	83.77	92.17	95.32
Sensitivity
40	80.29	80.46	75.98	79.99	82.56	81.18	84.53	84.79	85.43
50	82.42	82.42	80.62	80.99	84.94	82.54	85.27	86.16	86.60
60	82.49	82.59	80.79	81.99	85.04	82.86	85.55	86.41	87.06
70	83.05	84.65	81.52	82.66	86.30	86.00	86.38	87.92	88.25
80	83.95	85.46	82.16	82.70	88.51	87.99	88.58	91.39	91.84
Specificity
40	70.94	71.22	68.26	70.69	72.37	71.52	72.44	72.52	91.22
50	71.18	71.81	69.85	71.08	72.89	71.94	74.30	74.79	91.73
60	71.50	71.88	70.61	71.11	73.80	72.42	74.56	74.80	92.39
70	71.64	72.15	70.72	71.44	74.40	72.81	74.72	75.93	93.09
80	71.74	72.23	71.28	71.64	74.46	73.88	75.14	89.06	94.92

Assessment using restaurant data

Figure 6 portrays the assessment of the intelligent Lead based BiLSTM using the restaurant data by varying the training data, in which the accuracy is presented in Fig. 6a, the sensitivity is presented in Fig. 6b and the specificity is presented in Fig. 6c. The specificity acquired by the proposed method with 80% of training data is 95.32%, which is 5.06% enhanced performance compared to the 40% training data. Besides, the proposed method with the same training data is 19.41, 19.15, 18.74, 19.26, 16.62, 19.91, 17.52, and 15.99% superior than the KNN, SGD-NN, Random Forest, TD-LSTM, ATAE-LSTM, BiLSTM, Spy algorithm based BiLSTM, and King algorithm based BiLSTM conventional methods. Table 5 depicts the detailed analysis of Fig. 6.

Fig. 6.

Assessment using restaurant data based on a accuracy, b sensitivity and c specificity

Table 5.

Assessment using restaurant data

Methods/Training percentage	KNN	SGD-NN	Random forest	TD-LSTM	ATAE-LSTM	BiLSTM	Spy algorithm based BiLSTM	King algorithm based BiLSTM	Intelligent lead based BiLSTM
Accuracy
40	76.98	83.08	83.60	80.07	84.74	70.00	83.98	86.07	91.91
50	79.60	83.20	83.85	81.75	85.05	75.25	84.45	86.44	91.95
60	79.81	83.85	84.21	82.70	85.99	75.71	85.13	88.49	92.23
70	80.82	84.48	85.03	83.37	86.52	76.43	85.87	89.83	94.23
80	82.43	84.89	85.32	83.93	86.36	81.67	86.11	91.50	96.11
Sensitivity
40	78.99	90.31	91.16	84.49	92.10	66.15	91.47	93.13	94.25
50	83.55	90.44	91.61	87.79	92.38	75.82	92.18	93.72	94.27
60	83.92	91.63	92.03	89.64	93.89	76.70	93.08	94.08	95.94
70	85.92	92.89	93.47	90.98	94.86	77.91	94.19	96.46	97.75
80	88.86	93.57	94.04	91.73	95.33	87.81	94.46	96.65	99.22
Specificity
40	75.74	76.69	76.87	76.45	78.22	74.44	77.32	78.22	90.49
50	76.44	76.80	76.93	76.54	78.58	75.44	77.56	79.22	91.22
60	76.50	76.91	77.24	76.58	78.95	75.48	78.04	79.56	91.78
70	76.53	76.92	77.44	76.60	79.06	75.71	78.40	79.87	93.82
80	76.82	77.06	77.46	76.96	79.48	76.34	78.62	80.08	95.32

Comparative discussion

The assessment by considering the best measure acquired by the intelligent Lead based BiLSTM for the sentiment analysis is portrayed in Table 6. The maximal accuracy and sensitivity are acquired while analyzing using the restaurant data with the value of 96.11 and 99.22%. Likewise, the maximal specificity is acquired using the Twitter data with a value of 95.35%. The performance superiority of the proposed method in terms of accuracy is 14.24, 11.67, 11.22, 12.67, 10.14, 15.03, 10.41, and 4.80%, in terms of sensitivity is 10.45, 5.69, 5.22, 7.55, 3.92, 11.50, 4.80, and 2.59% and in terms of specificity is 26.57, 26.19, 26.26, 26.99, 20.79, 25.97, 25.20, and 20.70% for the traditional methods like KNN, SGD-NN, Random Forest, TD-LSTM, ATAE-LSTM, BiLSTM, Spy based BiLSTM, and King based BiLSTM, respectively.

Table 6.

Comparative assessment with best measures

Methods/Training percentage	KNN	SGD-NN	Random forest	TD-LSTM	ATAE-LSTM	BiLSTM	Spy algorithm based BiLSTM	King algorithm based BiLSTM	Intelligent lead based BiLSTM
Accuracy	82.43	84.89	85.32	83.93	86.36	81.67	86.11	91.50	96.11
Sensitivity	88.86	93.57	94.04	91.73	95.33	87.81	94.46	96.65	99.22
Specificity	70.02	70.38	70.31	69.62	75.53	70.59	71.32	75.61	95.35

The intelligent lead-BiLSTM acquired superior performance compared to all the conventional sentimental analysis methods due to the incorporation of the intelligent lead optimization in training the BiLSTM for the minimization of the loss in data learning. Besides, the proposed method acquired fast convergence by integrating the informer's knowledge with the Monarch in decision-making regarding the detection of the enemy base for safeguarding the country. In addition, the balanced exploration and exploitation phase based on the informer's ranks helps to identify the global best solution for tuning the modifiable parameters of the sentiment analysis method. Thus, the enhanced performance is acquired by the intelligent Lead based BiLSTM.

Conclusion

This research introduces a sentiment analysis technique using the intelligent Lead based BiLSTM. Here, in the intelligent lead optimization, the role of the informer along with the Monarch knowledge in decision-making for locating the enemy-based is considered. The optimal best solution for solving the optimization issue is the detection of an unknown enemy-base and the location of the informer is the solution vector that is evaluated based on the fitness function. In addition, the loss of the classifier while learning the data is eliminated through the incorporation of the intelligent lead optimization hence the loss is reduced and more accurate analysis is obtained in terms of assessment metrics. Besides, the balanced exploration and exploitation phase helps to obtain the global best solution for solving the optimization issues in sentiment analysis by considering the COVID-19 data. The assessment based on the accuracy, sensitivity, and specificity acquired by the intelligent Lead based BiLSTM is 96.11, 99.22, and 95.35%, respectively. The accuracy of the detection still needs to be improved for further online applications. Hence, in the future, a novel deep learning technique will be devised with a graph-based technique for identifying the polarity of the sentiment to enhance the performance further.

Author contributions

All authors have made substantial contributions to conception and design, revising the manuscript, and the final approval of the version to be published. Also, all authors agreed to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Declarations

Competing interests

The authors declare no competing interests.