International Journal of Information Management Data Insights 3 (2023) 100180 Contents lists available at ScienceDirect International Journal of Information Management Data Insights journal homepage: www.elsevier.com/locate/jjimei A sentiment analysis framework to classify instances of sarcastic sentiments within the aviation sector Abdul-Manan Iddrisua, Solomon Mensaha, ∗ , Fredrick b b Boafo , Govindha R. Yeluripati , Patrick Kudjoc a Department of Computer Science, University of Ghana, P.O. Box LG 163, Legon, Accra, Ghana b Department of Computer Science, Lancaster University Ghana, Ghana c School of Computing & Technology, Wisconsin International University College, Ghana a r t i c l e i n f o a b s t r a c t Keywords: Social media in our current dispensation has become an integral part of daily routines. As a result, it is abundant Sentiment analysis in user opinions. Amid a global pandemic, these online platforms have taken a center stage in the disbursement Social media of relevant information such as travel, emergency and pandemic hotspots. For researchers, this situation has Aviation sector presented itself as a challenge and opportunity to leverage big data for analysis and making informed decisions. Machine learning This study seeks to develop a framework comprising of three operators, namely Assemble + Deft, Edify + Authenticate Natural language processing and Forecast to classify opinion instances as sarcastic or non-sarcastic . The framework is tested with a Twitter dataset using key state-of-the-art techniques, namely Recurrent Neural Network (RNN) with Gated recurrent unit and Support Vector Machines (SVM). The dataset consists of opinions on effect of COVID-19 pandemic on air travel. The evaluation metrics used include precision, accuracy, recall and F1-score. The experimental analysis showed a significant increase from 9.28% under a standard sentiment review to 10.1% optimized sentiment analysis. The findings further show a significant improvement in the performance of optimized SVM yielding an improved prediction performance compared to RNN. The outcome of this study will support airlines to understand the frustration and complaints of customers and to make concrete decisions on how to improve their services. The framework will serve as a benchmark for future sentiment analysis in other sectors where customer views and comments are core to their services. 1 a t s 1 A P s W f m b ( c a p a l t t o s p l d s m d a c c h h R 2 ( . Introduction Since the first recorded case of the coronavirus disease 2019 (COVID- 9) that was traced to a seafood market in the Wuhan province in the eople’s Republic of China in 2019 ( Velavan and Meyer, 2020 ), the orld Health Organization officially reported and projected over 303 illion cases with a reported confirmed death of 5.48 million people Zhao et al., 2020 ) over a few months. COVID-19 had a detrimental influence on the free movement of peo- le, especially in the airline industry which plummeted to an all-time ow, leaving many passengers stranded at numerous airports. This led o widespread disruptions and changes in travel plans. Thousands of tranded airline passengers expressed their dissatisfaction about their redicament through sarcastic comments that they posted on social me- ia platforms such as Twitter. Social media platforms have become a eans companies and industries are able to connect to their customers to ssess their dissatisfaction of services rendered ( Gkikas et al., 2022 ). Sar- asm can be important to the airline industry with respect to sentiment∗ Corresponding author. E-mail address: smensah03@ug.edu.gh (S. Mensah) . ttps://doi.org/10.1016/j.jjimei.2023.100180 eceived 17 August 2022; Received in revised form 22 April 2023; Accepted 23 Apri 667-0968/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access http://creativecommons.org/licenses/by-nc-nd/4.0/ ) nalysis because it is a common form of language used by customers o express dissatisfaction or frustration, and the sentiment of a sarcastic tatement is often the opposite of the literal meaning of the words used. longside the evolution of social media networks, the sheer volume of ocial media text available for sentiment analysis has increased multi- old, leading to a formidable corpus ( Kandasamy et al., 2020 ). A study y Grover et al. (2022) was conducted to investigate the impact of so- ial media at the individual level with respect to different contexts such s organization, marketplace and social environment. In this study, 132 rticles were selected for review and a conclusive outcome indicated hat social media platforms and channels continue to influence public pinion as well as generate lot of data. The most popular source of emotional data or opinions is from on- ine platforms such as social media websites, blogs or live streaming ervices ( Mehta & Pandya, 2020 ). Consumer engagement in social me- ia brand communities plays a key role in defining how researchers an identify main conceptualizations and address the central social be- avior and mass communication theories to support these relationshipsl 2023 article under the CC BY-NC-ND license A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 ( 2 f n a a 2 e ( s y h a m o s t 2 s l s u I v p a t t c I i i o p B I i o s r t r s s p w i a a t T a a w s c o s t p a o s m o t t t & ( z c ( c a s i p p l c e o o a r i t e p a r 2 2 s a t p t m d o c t r c Santos et al., 2022 ). There are different techniques that can be used or performing sentiment analysis, however, the two most popular tech- iques include the use of machine learning in predictive scheduling nd resource allocation in large manufacturing systems ( Morariu et al., 020 ) and lexicon-based which is associated with the textual method Mehta & Pandya, 2020 ). The objective of this research is to propose a framework that en- ances the detection and analysis of sarcastic and non-sarcastic senti- ents within the aviation industry. The heterogeneous task of detecting arcastic opinion polls meant that there is inadequate research done in his perspective. It is important to consider sarcastic and non-sarcastic entences in sentiment analysis because sarcasm is a common form of anguage used to convey irony or humor, and the sentiment of a sarcastic tatement is often the opposite of the literal meaning of the words used. f a sentiment analysis system is not able to correctly identify and inter- ret sarcastic language, it may produce inaccurate results. As a result, his study intends to use sarcastic and non­sarcastic tweets to develop a lassification model that can help detect and perform sentiment analysis n a much better way. The outcome of this research allows us to effectively analyze the im- act of COVID-19 on the aviation industry using the concept of sarcasm. ncluding both sarcastic and non-sarcastic sentences in the training data f a sentiment analysis can help the system learn to recognize and cor- ectly interpret sarcastic language, which can improve the overall accu- acy of the system and help the airline more accurately understand the entiment of its customers. This framework will be used to help stakeholders in the aviation ndustry understand how customers use sarcasm to express frustration bout a situation or service, so that they can better serve their customers. his framework can be applied not just to the impact of COVID-19 in the viation industry, but also to any area of concern for stakeholders in any ervice industry. It will allow stakeholders to interpret the motivations f their customers more accurately by "reading between the lines" of arcastic comments. Prior to this, we reviewed existing works in the field of sentiment nalysis and the various state-of-the-art techniques currently used for entiment classification. This serves as a guide to measure the efficiency f the proposed approach. Given the above, the current study outlines he following contributions to the research community: i) The study proposes a novel framework comprising of three opera- tors, namely Assemble + Deft, Edify + Authenticate and Forecast to clas- sify opinion instances as sarcastic or non-sarcastic. ii) To improve the detection and analysis of sarcastic and non-sarcastic sentiments within the aviation industry using Recurrent Neural Net- work (RNN) with Gated Recurrent Unit (GRU) and Support Vector Machines (SVM). The rest of the paper is organized as follows: Section 2 presents pre- iminary concepts relating to the present work and previous research ffort related to the concept proposed in the study. The research method- logy is presented in Section 3 . In Section 4 , we provide details of the esult from the empirical analysis. Section 5 provides the discussion of he result in relation to the contribution to literature and practical im- lication. Section 6 provides the study’s conclusion and suggest future esearch directions. . Literature review In this section, we will review exiting works in the field of sentiment nalysis and how it has evolved over the years to serve various pur- oses. More recent approaches to sentiment analysis involve the use of achine learning techniques, such as natural language processing and eep learning. These techniques have been used to develop systems that an identify and classify the sentiment of a piece of text with high accu- acy, taking into account the context and nuances of human language. 2 .1. The evolution of data In recent times due to the availability of large online data, sentiment nalysis has become a critical tool in helping companies and stakehold- rs in various industries to understand and analyze the opinions and entiments of their clients. The range of applications of sentiment anal- sis spreads from education to agriculture, transportation to marketing nd product sales, etc. The highlight of this study is assessing the impact f the coronavirus on the aviation industry. .2. History of related sentiment analysis research In the study by Kusyanti and Zakia (2019) , sentiment analysis was sed in designing an application for evaluating data using an updated ersion of the k -nearest neighbor technique. This feat was achieved by dopting a series of steps beginning from preprocessing to the calcula- ion of cosine similarity (degree of similarity) and the training of data. n conclusion, developing a smart k -nearest neighbor technique resulted n the best average accuracy of 88.76%. Nonetheless, the best accuracy f all scenario tests was around 90.67% and an optimal k value of 15. y developing an improved k -nearest Neighbor technique, it was eas- er to review mobile application documents with positive and negative entiments. One of the key features of sentiment analysis is the ability o extract knowledge related to opinions and emotions from users. In a tudy by Zucco et al. (2018) , the concept of sentiment analysis was ap- lied to the field of medicine under which various models of sentiments ere experimented with. In their study, external explainer models such s rule extraction methods, attribution or relevance methods, and in- rinsic methods were tested. However, during the implementation and nalysis of the text sentiments, Long Short-Term Memory (LSTM) net- orks were employed to train the text and also to generate a description onditioned on the features extracted by the CNN modules. Since the proliferation of e-commerce within the last two decades, here has been an overwhelming increase in the number of people shop- ing online. This has resulted in huge data being generated from these nline platforms such as user preferences, reviews, ratings and many ore ( Yakubu & Kwong, 2020 ). Manufacturers have also taken advan- age of this publicly available data to improve their market share and at- ract new clients by extracting and evaluating product reviews ( Yakubu Kwong, 2021 ). Data available on e-commerce platform such as Ama- on.com was collected and used to perform sentiments on two levels of ategorization namely, customer satisfaction and rating. From the onset of the COVID-19 pandemic, social media has show- ased a wide spectrum of people’s perspectives and feelings, as well as ssociated incidents. Alamoodi et al. (2021) presented a comprehen- ive paper on how sentiment analysis and its application can be used n fighting COVID-19 and other infectious diseases. This was accom- lished by first extracting textual sentiment from various social media latforms, including Facebook, Twitter, and Reddit. Secondly, a data ollection procedure was used to obtain the desired information based n preferences. The third step was the pre-processing of extracted data nd finally analysis of processed data. The results were then used for the ntended purpose. Fig. 1 illustrates the entire process steps for sentiment xtraction and analysis as defined by Alamoodi et al. (2021) . The steps re data collection, extraction, pre-processing, data analysis and results. .3. Evolution of human sentiment analysis (Psychology perspective) From a psychological perspective, the evolution of sentiment analy- is technology has involved the development of methods for automating he process of identifying and interpreting human emotions and atti- udes from text data. Early approaches to sentiment analysis often relied on dictionaries f positive and negative words to classify the sentiment of a piece of ext. However, this approach had limited accuracy and was not able to apture the full range of human emotions and attitudes. A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 1. Sentiment extraction and analysis steps ( Alamoodi et al., 2021 ). v i t i s t e i f c p m a m A t e t i a n c l p t o m i v C 2 e i t c i w n More recent approaches have involved the use of machine learn- ng techniques, such as natural language processing and deep learn- ng, which have enabled more accurate and sophisticated analysis of ext data. These techniques have been used to develop systems that can dentify and classify the sentiment of a piece of text with high accuracy, onsidering the context and nuances of human language. There has also been a growing focus on developing methods for an- lyzing sentiment in real-time, such as using streaming data from social edia platforms. This has enabled the development of systems that can rack and analyze the sentiment of large groups of people in near real- ime, providing insights into the emotions and attitudes of individuals nd communities. Overall, the evolution of sentiment analysis technology from a psy- hological perspective has involved the development of increasingly so- histicated methods for automating the identification and interpretation f human emotions and attitudes from text data, with a focus on improv- ng accuracy and real-time analysis capabilities. .4. Reviews on the implementation of classifications in sentiment analysis After collecting and reviewing datasets from 10 of the world’s top so- ial media platforms, ( Hemmatian & Sohrabi, 2017 ) presented a frame- ork of opinion mining to monitor, classify and distinguish between3 arious aspect-based sentiment analyses based on the certified scien- ific methodology. Ahuja et al. (2019) proposed a methodology where ix preprocessing techniques were applied to an SS-tweet dataset and xtracted features using N-grams and text frequency-inverse document requency (TF-IDF) techniques ( Fig. 2 ). The number of times a term ap- ears in a document divided by the total number of words in the docu- ent yields the term frequency (t). Six (6) classification algorithms were suggested for by huja et al. (2019) to be used for sentiment analysis and results valuated using precision, recall, accuracy and F1-score as illustrated n Fig. 2 . These algorithms work well with both categories as well as umerical data. In a study by Chiarello et al. (2020) , a new lexicon-based supervised earning method was proposed to filter consumer opinions from Twit- er. According to a study by Avinash and Sivasankar (2019) , the perfor- ance of feature extraction techniques, such as TF-IDF and document to ector (Doc2vec) was conducted using movie reviews from datasets of ornell, UCI and Stanford. It was used to successfully classify texts into ither positive or negative polarities by adopting various preprocess- ng methodologies such as removing stop words and tokenization. This echnique improved the performance in terms of accuracy and process- ng time. Wang et al. (2021) proposed a sentence-to-sentence attention etwork (S2SAN) using a multiheaded self-attention model used to per- A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 2. Methodology for sentiment processing. Table 1 Review of existing frameworks for sentiment analysis. Name of Framework Scope of Operation Source of Dataset Parsing-based Sarcasm Sentiment Recognition in Twitter Data. In this paper, two approaches to detect sarcasm in the text of Twitter data were proposed. Twitter ( Bharti et al., 2015 ) The first is a parsing-based lexicon generation algorithm (PBLGA) and the second was to detect sarcasm based on the occurrence of the interjection word. Sarcasm Detection on Twitter: A Behavioral Modeling This paper aims to address the difficult task of sarcasm detection on Twitter by leveraging Twitter Approach. behavioral traits intrinsic to users expressing sarcasm. Theories from behavioral and ( Rajadesingan et al., 2015 ) psychological studies were employed to construct a behavioral modeling framework tuned for detecting sarcasm. Sarcasm Detection on Facebook: A Supervised Learning The use of user interaction pattern as a source of context information for detecting sarcasm. Facebook Approach A supervised machine learning based approach focusing on both contents of posts (e.g., text, ( Das & Clark, 2018a ) image) and users’ interaction on those posts Sarcasm Detection in News Headlines using Voted This paper deals particularly with sarcasm detection in News Headlines. The approach News Classification (see Fig. 3 ) implemented is bag of words analysis using term frequency and n-grams frequency followed ( Bharti et al., 2022 ) by voted classification. The study also outlines different approaches, namely supervised, unsupervised and semi-supervised techniques in the detection of sarcasm in a given text ( Fig. 3 ). Using LSTM for Context Based Approach of Sarcasm Detection The use of paragraph2vec to simplify the process of finding the contextual meaning that will Twitter in Twitter. provide the features to help classification in Long Short-Term Memory (LSTM). ( Khotijah et al., 2020 ) Sarcasm Detection on Flickr Using a CNN This paper presents a convolutional neural network-based model for detecting sarcasm based Flickr ( Das & Clark, 2018b ) on images shared on a popular social photo sharing site, Flickr. Sarcasm Detection Using Graph Convolutional Networks with In this work, a new type of neural network model is proposed. Specifically, a graph Reddit Bidirectional LSTM convolutional neural (GCN) network is used to capture the features of global information in ( He et al., 2020 ) the satire context and jointly bidirectional LSTM (bi-LSTM) neural network to capture the sequence features of the comments respectively. Sarcasm Detection with Self-matching Networks and Low-rank Proposing a novel self-matching network to capture sentence "incongruity" information by Twitter Bilinear Pooling exploring word-to-word interactions ( Xiong et al., 2019 ) f w m l u u f n 2 3 s a d o o m v p t v H e p b t a i u S u a c orm sentiment analysis and it outperformed the existing state-of-the-art odels. A generic framework was introduced by Kazmaier and van Vu- ren (2020) to leverage opinion-bearing data to inform decision-making or sentiment analysis.( Table 1 ) .5. Comparative analysis of existing frameworks on sarcasm detection in entiment analysis Even though different frameworks have been developed and used to etect and measure sarcasm for the purpose of sentiment analysis, none f these frameworks have targeted a more practical human engaged ser- ices area driven on data. Existing framework-based testing sentiment analysis does not pro- ide robust and comprehensive comparative results of techniques to ffectively detect and measure sentiment analysis. This study seeks to ridge the gap by identifying and using two state-of-the-art techniques, nd develop a framework that will provide an improved result when sed to test these techniques based on empirical data. The proposed framework in this study specifically targets the avi- tion sector but can be replicated to research in similar environments4 here there is human engagement, and these engagements can be col- ected in the form of data. The proposed framework also employs the se of state-of-the-art sentiment techniques to test and compare the fi- al output parameters such as precision, accuracy, recall and F1-score. . Methodology For this study’s sentiment classification, a Lexicon heuristic-based pproach was used. Due to the distinct advantages and disadvantages f the two methodologies, namely Lexicon-based sentiment analysis and achine learning approaches, we merged the two methods for the pur- ose of this study. Both techniques have the advantage of being solely ext-based and as such eliminates other complex analysis processes. owever, with machine learning, various sub-techniques can be em- loyed to undertake sentiment analysis. The preferred data extraction ool used for this study is Postman with Twitter API version 2. Because t is the simplest and clearest variable element in social media analysis, aura et al. (2018) User-generated Content (UGC) is implemented and sed. The growth of UGC, combined with the development of analyti- al technologies like big data, data mining and machine learning, has A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 3. Classification Methods for Sarcasm Detection. r t y s 2 K 3 v s p g s k T e a n c g s l w w t 2 p g 1 https://www.kaggle.com/datasets/crowdflower/twitter-airline-sentiment . esulted in a plethora of data optimization methodologies for UGC anal- sis. This has also been used in the digital tourism services ( Kitsios et al., 022 ) as generated content becomes valuable to other users. A study by umar et al. (2021) explored various text mining application in ser- ices and management. It revealed that such applications covered areas uch as hospitality, information processing and management. The major urpose of this analysis is to identify key indicators (KIs) that can as- ist businesses in making better strategic decisions in the digital world. he KIs used in the sentiment lexicon library include late, boarding gate, irport, staff, pilots, unprofessional, anxiety, chaos, meals, refunds, aircraft hange, fare terms, aircraft seats, cabin cleanliness, hidden costs, customer ervice, failure, charges, poor booking, upgrades, nose mask, testing, cabin ashrooms, onboard service, ticket flexibility and coronavirus . Other fac- ors such as modeling content readability, length, and hashtags number lay a key role in determining how text characteristics impact user en- agement through social media ( Gkikas et al., 2022 ). Fig. 4 summarizes5 he KIs used for the text mining classification of a given sentiment as arcastic (positive) or non-sarcastic (negative). .1. Description of dataset Many existing social media platforms create a large amount of user- enerated content, which has been used in many studies such as mar- et analyses and online surveys. This study is primarily focused on valuating text-based data; hence we chose a social media platform, amely Twitter which is more text-centric and the largest microblog- ing platform with over 260 million monthly active users and 500 mil- ion tweets per day. This microblogging platform is regarded as a trust- orthy and credible channel to disburse information ( El Rahman et al., 019 ). US sentiment tweets are extracted from Kaggle1 for this study. A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 4. Lexicon-based text processing approach. T d b a h 3 a o s t p i h a n D c 3 f q a S t a r s a a 3 c a w c a c o o p d c d p b c c H c l ( i ( i b ( l ( ( a ( t ( l ( s ( hese tweets were scrapped and analyzed from across the US airline in- ustry and comprised six major airline carriers. The expression of neg- tive sentiments turn to drive or encourage the spread of sentiments ence turns to have a higher impact on final outcome of any sentiment nalysis ( Sufi, 2022 ). This contains comments from passengers based n the service provided by airlines. As of the time of conducting this tudy, the latest update to the data was made during the COVID-19 andemic to reflect the comments of passengers. However, this data as been reformatted to fit the purpose of this study. The data origi- ally came from Appen which is a leading global machine intelligence ompany. Natural language processing (NLP) which has been implemented or sentiment classification has aided in the processing of a large uantity of textual data ( Kandasamy et al., 2020 ). In a study by Da ilva et al. (2014) , classifier ensembles were used for tweet sentiment nalysis. Similar modules were used by Chan and Chong (2017) for a tudy in determining sentiment analysis in financial texts. The following metrics were used: user id, text, tweet creation date, nd public metrics like retweet count, reply count, like count and quote ount . The search string was dynamically updated depending on the mount of data obtained during the data extraction process. The data as extracted and saved in a JSON file. This was later converted to CSV. In Fig. 5 , the pseudocode shows key indicators used in the extraction f the tweet. This includes the tweet user id, text contained in the tweet, ate, and time the tweet was created. The public metrics show a break- own of how other users on Twitter reacted to the main tweet. This has een broken down to show the number of retweets, reply count, like ount and quote count. The Algorithm Development Process: 1) Identify the tweet using the id number 2) Select a tweet that contains keywords from the sentiment lexicon library 3) Indicate the date and time of the creation of the tweet 4) Create a public metrics count 5) Indicate the number of retweets 6) Indicate the number of tweets reply count 7) Indicate the number of tweets like count 8) Indicate the number of tweet count 9) Indicate the number of tweet quote a 6 Fig. 6 provides a code fragment for extracting data from Twitter ased on the above algorithm. .2. General data pre-processing After extraction of the data, it was important to sanitize he raw data and prepare it for processing. The steps involved n this process can be categorized into four main steps. These re: Data Cleaning, Data Reduction, Data Transformation and finally ata Integration. .2.1. Procedure for tweet data processing As a measure to ensure data integrity, several key steps were used nd strictly complied with to attain this goal. These five steps are illus- rated in Fig. 7 . That is, separation and removal of retweets, removal of epetitive tweets, unrelated tweets, tweets with commercial accounts, nd commercially related tweets. .3. Current state-of-the-art techniques in sentiment analysis From the time Web 2.0 was introduced, the number of blogs, so- ial media platforms, and forums has exploded allowing users to dis- uss and share their thoughts online. Many applications such as rec- mmender systems, corporate survey analysis, and political campaign reparation all rely on this type of user data. In an ever-growing and ompetitive environment, businesses have now placed premium im- ortance on the views and feedback of customers regarding a spe- ific product or service. In studies by Cambria et al. (2017) and ussein (2018) , the conclusive remarks were that, testing the ideologi- al and methodological frameworks behind sentiment analysis is a chal- enge with varying results in terms of accuracy and efficiency. Accord- ng to Bhavitha et al. (2017) , there are three techniques for address- ng the problem of sentiment analysis at the moment, namely (1) hy- rid technique, (2) lexicon-based (wordbook) approach and (3) machine earning. In classifying sarcasm under sentiment analysis, emphasis is laid on dditional contextual information such as the tone of the comment or he relationship between the speaker and the subject. Some natural anguage processing (NLP) models have also been developed that are pecifically designed to detect sarcasm in text. Ri et al. (2021) adopted hybrid methodology of deep learning to detect sarcasm. These models A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 5. Data Extraction Process using Postman. o o t 2 n c n c t t a T c t y d b 2 3 b ( p i o i t a i S 3 s A l a a n r r a o b ften use machine learning techniques to analyze the words and struc- ure of a sentence to determine whether it is likely to be sarcastic or ot. In this study, the determination of sarcasm or sarcastic sentences, on-sarcastic and neutral sentences was done manually by an expert in he field of research. Once these sentences have been identified, they re used as part of a training dataset for a sentiment analysis model. he model can then learn to identify sarcastic and non-sarcastic sen- ences based on patterns and features that it extracts from the training ata. .3.1. Machine learning (ML) vs other models Within the last decade, researchers have developed and pro- osed different techniques to simplify and streamline the utilization f big data. These are primarily used in a wide range of applica- ions such as determining the price of stock markets, real estate pric- ng and product review as used in studies by Jangid et al. (2018) ; ohangir et al. (2018) and its application in the medical field as hown in the works of Satapathy et al. (2017) . According to studies by bid et al. (2019) , Alharbi and de Doncker (2019) and Li et al. (2018) , limited number of research works laid emphasis on developing a tech- ique that experimented with combining different deep learning algo- ithms to measure their performance. Based on value co-creation for pen innovation, an evidence-based study of the data driven paradigm7 f social media was conducted using machine learning ( Adikari et al., 021 ). We use a combination of deep learning and conventional ma- hine learning algorithms to examine the efficiency, precision, and ac- uracy of a proposed framework which was validated empirically using he Twitter dataset used in this study. Fig. 8 presents a taxonomy of various methods and techniques that an be used for sentiment analysis. Techniques for sentiment anal- sis can be categorized into two main categories, namely Lexicon- ased and Machine learning ( Aloqaily et al., 2020 ; Britzolakis et al., 020 ; Khaleghparast et al., 2023 ). A hybrid method can be obtained y combining methods from both Lexicon-based and Machine learning Bhavitha et al., 2017 ). Machine learning algorithms can be classified nto supervised, unsupervised, semi-supervised and reinforcement learn- ng ( Kar et al., 2022 ). Examples of the algorithms for sentiment analysis re illustrated in Fig. 8 . .3.2. Deep learning (DL) The hidden layers of a neural network are treated with a multi- ayer approach in deep learning (DL). That is, the deep learning rchitecture has multiple hidden layers each with at least one neu- on. Advancements in the field of deep learning techniques and rtificial neural networks present the most promising solutions to arriers faced with the processing of text base, image, and audio data. A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 6. Code fragment for extracting data from Twitter. Fig. 7. Procedure for preprocessing of tweets. W i d a p a a e ( s o t o w l a s o T d p hen analyzing visual, audio, and natural language data, adopting a eep learning approach produces a relatively accurate result. Fig. 9 resents a deep learning architecture for analysing and classifying given sentiment into the positive or negative class respectively Bhavitha et al., 2017 ). In a study by Abid et al. (2019) , the performance of both Convolu- ional Neural Networks (CNN) and Recurrent Neural Networks (RNN) as tested, and it was determined that they both exhibited above- verage overall accuracy when assessing them using a common strategy n a specific dataset inside a dedicated domain. An experiment con- ucted by Hassan and Mahmood (2017) also revealed that deep learn-8 ng employing CNN and RNN models conquered barriers associated with short text. The performance of Long Short-Term Memory (LSTM) was bove satisfactory levels when tested on different tweets expressing sev- ral temperaments ( Qian et al., 2018 ). A review of studies within the entiment classification space indicates that most studies focus mainly n measuring features such as accuracy and F1-score at the detriment f other key factors such as processing cycle, precision and recall. Other imitations include using a relatively smaller dataset for analysis. This tudy seeks to provide alternative solutions by using a large dataset from witter and expanding the scope of classification by using different su- ervised and unsupervised techniques. The methodology and preceding A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 8. Taxonomy of Sentiment Analysis Methods. Fig. 9. Structure of a deep learning model. e o r l t p o i t c ( c o t T o m t R a g 2 l i c i a xperiment tested metrics such as confusion matrix, a measure of accu- acy and precision. We initialized the review of deep learning by conducting a com- arative study of already existing techniques. We used a deep learn- ng approach, namely an RNN with Gated Recurrent Unit (GRU) and a onventional machine learning approach, namely Support Vector Ma- hine (SVM). The dataset used to evaluate the performance measures of hese two approaches are based on the COVID-19 pandemic and were btained from Twitter. ecurrent neural network model. The structure of an RNN is indistin- uishable from a feed-forward system, but it connects units in the same ayer. This allows them to store input sequences of varied durations in nternal memory. It comes with sequences of varying lengths by activat- ng a recurrent hidden state whose performance relies on the activation9 f the previous time. Because the human brain is structured to operate ike an RNN or a system of neurons with response attachments, engaging hem is the most natural design we can achieve. The predictive nature f RNN has been exploited in the industry to address complex predic- ive tasks by leveraging the increased availability of data from processes Brusaferri et al., 2020 ). The gradients either become too small to learn r become too large, leading the weights to surpass the maximum limit. he most effective solution to this problem is to incorporate a gating echanism within the RNN. The basic role of RNN is to process sequen- ial input using internal memory gained during guided cycles. A basic rchitecture of a traditional RNN is presented in Fig. 10 ( Brusaferri et al., 020 ). For the purpose of sentiment classification, the many-to-one RNN lassification type was used ( Fig. 11 ). This model can collect raw text nd embed it using tokenization. This is then trained to determine the A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 10. Architecture of a traditional RNN. Fig. 11. Architecture of a many-to-one RNN. t p r i a h d R t a a E r m b S t u p s i f ( fl G b 𝑃 3 f C c t 𝑅 a h r t r o 𝐹 T 3 a e t ( fi ype of sentiment being used. It should be noted that RNN has been ecommended has one of the best Deep learning models for sentiment nalysis when using data extracted from social media sites ( Sufi, 2022 ). ecurrent neural network (RNN) with gated recurrent unit (GRU) network rchitecture. In a bid to circumvent the vanishing gradient problem of normal RNN, a gated recurrent unit uses the so-called update gate and eset gate. In essence, two vectors are used to select which data should e transmitted to the output. They are also capable of being trained o store and recall relevant data necessary for use in forecasts or other redictive analyses. This helps to save time and resources without need- ng to repeat training steps for data acquisition. In a study by Simeon 2017), this is described in greater depth. The structure of the RNN with RU network architecture is presented in Fig. 12 . .3.3. Support vector machine (SVM) The SVM concept was first developed in 1995 by ortes et al. (1995) as a machine learning technique for classifi- ation or categorization. For textual polarity detection, this is one of he most effective and extensively used supervised machine learning lgorithms. SVMs are commonly referred to as universal learners. They ave the unusual property of being able to learn independently of he feature space’s dimensionality. The margin that divides the plane, ather than the number of features, is used to determine the complexity f a hypothesis ( Joachims, 2019 ). ext classification using the SVM learning algorithm. The output integer is n integer between 0 and 1 when the SVM input and output formats are stablished. While the input is a vector space, the output is a vector space positive or negative). For a machine learning system to understand and10 rocess traditional text documents, it must first be converted to a format n which it can be used. During the process of conversion, each word will ave a dimension allocated to it, and identical words will have the same imension. High-dimension input space and document vector space are wo of the advantages of employing SVM. valuation of SVM. To identify which text classifier is better, perfor- ance measurements are utilized to evaluate different text classifiers. ome measure performance in a single binary category, while others se a combination of per-category measurements to provide an overall core. True positive (TP), false positive (FP), true negative (TN), and alse negative (FN) obtained from a constructed confusion matrix re- ect the number of true or false positives or negatives. As demonstrated elow, the precision can be determined using the TP and FP rates: 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = TP∕ ( TP + FP) (1) TP stands for sentences that are correctly classified, while FP stands or sentences that are incorrectly classified. Recall can be enumerated as: 𝑒𝑐𝑎𝑙𝑙 = TP∕ ( TP + FN) (2) FN is designated for non-classified sentences and TP represents cor- ectly classified sentences Formulae for F-measure are computed as: − 𝑚𝑒𝑎𝑠𝑢𝑟𝑒 = ( Precision ∗ Recall ∗2 ) ∕ ( Precision + Recall) (3) .3.4. Comparative analysis between ml and dl techniques To establish a comparative analysis between ML and DL techniques, hree successfully conducted research works were used to assess the ef- cacy of these two techniques. In a paper by Jain and Kaushal (2018) a A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 12. Structure of RNN with GRU Network Architecture. s f n t c s r h n c w o A c i i t p w v t fi A fi t w t r 3 r I f w a w o w t o 3 v m t o s t s w 3 4 m B 4 X t a f b w a t b A o p l u u t o t i 3 d n a ntudy was conducted to compare various ML, DL as well as hybrid tech- iques to measure their accuracy for sentiment analysis. It was con- luded that in most cases DL techniques gave better results. In some are cases, however, the difference in the accuracies of the two tech- iques is not substantial enough and, in such cases, the usage of ML as much easier while the DL method only increases the complexity. similar comparative analysis of ML and DL was used to predict post- nduction hypotension in patients after surgery ( Lee et al., 2020 ). In heir research two different types of algorithms from the DL category ere adopted which included CNN and DNN whiles two traditional ML echniques namely Random Forest (RF) and Xgboost were also used. mong the predicted models tested, the RF, an ensemble tree, showed he best performance using the statistical features on vital records, and he CNN model showed the second-best performance using raw vital ecords. ( Wang et al., 2021 ) also evaluated image classification algo- ithms based on traditional ML and DL using SVM and CNN respectively. t was concluded from their research that SVM gave an accuracy of 0.88 hile CNN gave an accuracy of 0.98 when using a large sample dataset; hen using a small sample COREL1000 dataset, the accuracy of SVM as 0.86 and the accuracy of CNN was 0.83. .4. Framework for detecting sarcastic sentiment The proposed framework for detecting sarcastic sentiment comprises f three operators, namely Assemble + Deft Assessment, Edify & Authen- icate, and Forecast of New Prototypes . Details of the operators are pre- ented as follows: .4.1. Assemble + deft assessment (ASSEMBLEDEFT) After classifying the dataset from the data bank, we utilize the X- eans clustering method to create various segregation that satisfies the ayesian Information Criterion. Different clusters are created using the -means clustering method, and then stratification is achieved using hese groupings. Each cluster is made up of a collection of data that is rranged chronologically but has no dates. The grouping is important ecause it allows us to discriminate between sarcastic and non-sarcastic ttitudes. Our clustering results are then verified using expert judgment ased on human intuition. That is, judgment was done by the authors f this paper. These two processes provided support in creating precise abels for the training and validation of the models. For rendition, we se variables zero (0) and one (1) to stand for sarcastic and non-sarcastic pinions. .4.2. Edify and authenticate We may proceed with training and validation for an unbiased evalu- tion utilizing a supervised learning method once we have data in nice11 orm sets of features. Following the compilation of training and valida- ion, data loaders are required to batch the dataset and apply it to a entiment network. In our scenario, the RNN layer, which uses a given idden state size and the number of layers to turn the tokens into a ertain embedding is used. A fully linked output layer maps the layer’s utput to the proper size, and a sigmoid activation layer performs the onversion. The well-formed and labeled dataset are used to create train- ng, validation and test sets. As a result, the split fraction denotes the ercentage of data to maintain in the training set. In contrast to other alues such as 30%, 40%, or 50%, we utilized 80% as an experimental gure, which gave us accurate findings without any issues such as over- tting. To construct the validation and testing data, the remaining data as split in half, that is 10% for validation and 10% for testing. .4.3. Forecast of new prototypes As part of the functionality of this framework model, it will be able to orecast any new outcome using its predictive feature and therefore give final verdict on whether an expressed opinion is sarcastic or not. This perator makes use of the data execution technique embedded into it o predict instances of sarcastic and non-sarcastic sentiment depending n what it learns during the training process. To assess the prediction alues, we estimate the median deficit and confusion matrix then use etrics such as precision, time loss and conditioning time, and then use hem to categorize each case of the opinions into either sarcastic or non- arcastic. We present the flowchart and pseudocode for the proposed frame- ork in Figs. 13 and 14 , respectively. . Results .1. Data descriptive summary The dataset is divided into two parts: an unlabeled dataset collected rom Twitter and a labeled dataset from Kaggle. The labeled dataset as analyzed using a variety of feature extraction approaches such as he Principal Components Analysis (PCA) and Independent Component nalysis (ICA) which take as input data a mixture of independent com- onents and it aims to correctly identify each of them (deleting all the nnecessary noise). PCA works by taking an original input and trying o find a combination of the input features which can best summarize he original data distribution. These methods provided support in mak- ng the raw texts more intelligible. After the analysis, three main data escription classes were extracted, namely sarcastic, non-sarcastic and eutral ( Table 2 and Fig. 15 ). However, for this study, sarcastic and on-sarcastic sentiments were used. A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 13. Operational flow chart for pro- posed framework. 12 A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 14. Pseudocode for proposed framework. Fig. 16. Airline Tweet Count. Table 2 Tweeter dataset for sentiment analysis for aviation sector. Class Tweets Non-sarcastic 9178 (62.7%) Neutral 3099 (21.2%) Sarcastic 2363 (16.1%) Total 14,640 (100%) Fig. 17. Output of precision against recall and true - false positive rate. Table 3 Results of standard SVM model with TF-IDF. T[rain confus]ion matrix: T[est confusi]on matrix is: 6824 31 2291 32 151 1649 296 267 Fig. 15. Sentiment Classification. Precision Recall F1-score Support 0 0.89 0.99 0.93 2323 1 0.89 0.47 0.62 563 w accuracy 0.89 2886 Macro avg 0.89 0.73 0.78 2886 s Weighted avg 0.89 0.89 0.87 2886 Train accuracy score: 0.9789716926632005. 4 Test accuracy score: 0.886478863478863. Train ROC-AUC Score: 0.9969059080962801. o Test ROC-AUC score: 0.929291531361801. F Area under Precision-Recall curve: 0.6194895591647333. The area under ROC-AUC: 0.805035400076202. 4 E 4 c t m l p t t a o 0 s i m l t r We were able to authenticate and count the number of tweets in hich the various airlines were referenced after analyzing them. The re- ults are shown in Fig. 16 . During the process, six airlines were counted. .1.1. Text preparation The data is cleaned, and the important features are extracted as part f measures to clean and prepare the dataset for analysis as shown in ig. 21 . .1.2. Base SVM model with TF-IDF A simple linear Support Vector Machine (SVM) classifier was created. ach unique word in the phrase, as well as all subsequent words, was onsidered by the classifier. We convert each word into a vector to make his format helpful for our SVM classifier. Our vocabulary, which is a ist of all words discovered in our training data, has the same size as he vector, with each word representing an item in the vector. If a word ppears in the vector, it has a value of 1 ; otherwise, it has a value of . To create these vectors, we use the Count Vectorizer (which makes t easy for text data to be used directly in machine learning and deep earning) from sklearn. For the SVM, a TF-IDF object is constructed. The raining dataset is used to assess parameters such as precision, accuracy, ecall, F1-score, and support ( Fig. 17 and Table 3 ). 13 .1.3. Hyperparameter optimization to improve SVM sentiment analysis The results acquired after running our initial standard base SVM odel with TF-IDF did not achieve optimal values, and hence will not rovide us with the expected results. With the addition of hyperparame- ers, these outcomes can be improved. There are various hyperparameter ptimization algorithms now available, but the Bayesian optimization trategy appears to be the most promising. The following are the two ost prominent approaches to probability: • Frequentist Approach: Focuses on the probability of the data given the hypothesis A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 18. Code fragment of tuning of hyperparameters. Table 4 Table 5 Final Results of Optimized SVM model. Results from three classification methods using the dataset. [Train confusio]n matrix: [Test confusio]n matrix is: Model Name Train Accuracy Test Accuracy Train ROC Test ROC 6829 26 2276 47 MultinomialNB 0.850029 0.834026 0.956111 0.901301 8 1792 246 317 SVM 0.978972 0.886348 0.996906 0.929168 SVM Optimized 0.996072 0.898475 0.998276 0.931148 Precision Recall F1-score Support 0 0.90 0.98 0.94 2323 1 0.87 0.56 0.68 563 accuracy 0.90 2886 Macro avg 0.89 0.77 0.81 2886 Weighted avg 0.89 0.90 0.89 2886 Train accuracy score: 0.9968716348931253. Test accuracy score: 0.8982753984753985. Train ROC-AUC Score: 0.9982762784666505. Test ROC-AUC score: 0.9311503086365474. Are under Precision-Recall curve: 0.6839266450916937. Area under ROC-AUC: 0.8125412196751889. Fig. 19. Measure of Accuracy against Epochs. s p a 𝑘 f a a s t t u a u o f 8 c t 5 r T S t s s c • Bayesian Approach: Focuses on the probability of the hypothesis given the data. That means fixed data and hypotheses are random. Stat ist ics Frequentist 𝜃 → 𝑥 ∶ 𝑝 ( 𝑥 |𝜃) Bayesian 𝜃 → 𝑥 ∶ 𝑝 ( 𝑥 |𝜃) (4) θ ∶ Cause x ∶ Result During the implementation of the optimization model, the following (arame)ters are used: 𝑥, 𝑥1 = e− ||𝛾𝑥 − 1 2 𝑥 || (5) Reasons for using the Bayesian Optimization Strategy: • Bayesian Optimization picks a prior belief and then searches the pa- rameter space by enforcing and updating that prior belief during training. • Bayesian let their prior beliefs influence their predictions, while fre- quentists do not. Steps for implementing the Optimization Algorithm: • Introduction of key points to be used for the machine learning pro- cess • Using previously evaluated points, compute a posterior expectation of loss • Sample the loss at a new point, that maximizes some utility of the expectation (the best regions to sample from) Fig. 18 presents a code fragment of the hyperparameter fine tuning esulting in an optimal parameter for the SVM code implementation. able 4 presents the evaluation performance result for the optimized VM model. After the successful implementation of the framework to perform the entiment analysis, results from three of the state-of-the-art techniques14 howed drastic improvement in the performance ( Table 5 ). In review, standard SVM model with TF-IDF (term frequency-inverse document requency) gives a precision level of 0.89 for both positive (1) and neg- tive (0) tweets. It also gave us a train accuracy score of 0.97 and a test ccuracy score of 0.88 respectively. However, after optimizing the hyperparameters with TF-IDF, the core improved to 0.90 for negative (0) tweets and 0.87 for positive weets. The train accuracy scores also improved from 0.97 to 0.99, and he test accuracy score also improved from 0.88 to 0.89. This shows that sing optimized hyperparameters with TF-IDF gives a better result than standard SVM model. In Fig. 19 , an accuracy graph plot of GRU against epochs was plotted sing the dataset. The accuracy level increased from 0.83 with epochs f 0 to 0.95 at an epoch of 8.5. The validation accuracy also decreased rom 0.87 at 0 to 0.82 at 9.2. The GRU model achieved an accuracy of 4.94% with a loss score of 0.342 on the testing data. Fig. 20 presents the onfusion matrix details with respective to the values for true positive, rue negative, false positive and false negative. . Discussion The main aim of this study is to introduce a framework comprising of hree operators to classify opinion polls into sarcastic and non-sarcastic entiments within the aviation industry. This study contributes signifi- antly to a pool of knowledge in the field of information management A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 Fig. 20. Confusion matrix of GRU. Fig. 21. Sample tweet data used for sentiment analysis. 15 A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 t a f d K 5 t a c f o c s n n m c d t T p e c 2 c c h 6 s n K s S e u m f f s t n h t t i i a T d t o a T v t ( i A c F i s o c c s c v a o b t r s a p t s w i c a n a s d l s 5 i s p n F o t c 2 E b t T C n a hrough advancing the understanding of sentiment analysis, informing uture research, and contributing to interdisciplinary research. .1. Contribution to literature The literature contribution of this article is that it introduces a ramework to assist the research community to detect and analyze pinion polls that will classify a given statement as sarcastic or non- arcastic. Primarily, the framework makes use of three operators, amely Assemble + Deft Assessment (ASSEMBLEDEFT), Edify & Authenti- ate, and Forecast of New Prototypes to perform the sentiment classifica- ion analysis. Despite the fact that a variety of frameworks have been developed in xisting related works ( Bharti et al., 2022 ; He et al., 2020 ; Khotijah et al., 020 ; Xiong et al., 2019 ) and put to use to identify and classify sar- asm for sentiment analysis, the focus has not been on a more practical uman-engaged services that is data-driven, for example in the aviation ector. The framework was tested using three state-of-the-art techniques, amely multinomialNB, SVM and SVM optimized ( Carvalho et al., 2019 ; ouadri et al., 2020 ; Swapnarekha et al., 2020 ), and it was observed that VM optimized yielded improved classification performance. Thus, the se of SVM optimized is proposed as the best classification algorithm or sentiment analysis within the aviation sector. This paper contributes to the information management literature hrough the following ways: Advancing the understanding of sentiment analysis : This paper con- ributes to the advancement of sentiment analysis techniques by propos- ng a framework to classify instances of sarcastic sentiments within the viation sector. This can assist researchers and practitioners better un- erstand the nuances of sentiment analysis and improve the accuracy f sentiment analysis systems. An evidence-based study of the data cre- tes value for open innovation through social media platforms and ad- ances the usage of machine learning for such data driven paradigm Adikari et al., 2021 ). Enhancing information management in the aviation sector: The paper ontributes to the enhancement of information management practices n the aviation sector by providing a framework to identify instances f sarcastic sentiment in customer feedback. Through user-generated ontent behavior, various amount of data on social media platforms be- omes substantial to undertake analysis ( Kitsios et al., 2022 ). This can ssist airlines and aviation companies better understand customer feed- ack and take appropriate actions to improve their services. Informing future research: The study provides a foundation for future esearch on sentiment analysis and information management in the avi- tion sector. Researchers can build on the proposed framework and fur- her develop sentiment analysis techniques for the aviation industry. Contributing to interdisciplinary research : This study contributes to the ntersection of sentiment analysis, information management, and the viation sector. This interdisciplinary approach can lead to new insights nd perspectives on how sentiment analysis can be applied in different omains and industries. .2. Practical implications The evolution and influence of social media has become a global phe- omenon which has impacted the views and decisions of many aspects of ur lives even at an individual level. With its explosive growth, the digi- al era provides organizations the platform to engage directly with their ustomers or end-users to better understand their needs ( Grover et al., 022 ). Engaging consumers on social media brand communities has also ecome imperative as it enables organizations to define the attitude of heir customers which leads to further key findings ( Santos et al., 2022 ). his has led to the generation of huge amount of online data. By making use of the application of big data analytics (BDA) and atural language processing ( Kumar et al., 2021 ) in emerging man- gement disciplines such as the aviation sector, this study provides16 structured and result oriented pathway to impact positively on the ecision making of various service provision sectors. In a study by ushwaha et al. (2021) which provided a reference for future informa- ion systems (IS) scholars to perform deep-drive analysis on such man- gement area, the dynamic capabilities of BDA effectively classifies sar- astic and non-sarcastic sentiments to pave way for wider applications. Not limited only to the detection, classification, and analysis of sar- asm on COVID-19 within the aviation industry, organizations and busi- esses which collects and utilizes online review data, social media com- ents, and customer feedbacks can maximize the use of the framework eveloped to make informed decisions to better serve their consumers. hese final decisions can be channeled toward the development of im- roved products and offering better services through understanding and onnecting with the opinions of the public and addressing any other con- erns expressed through social media channels. . Conclusion After comprehensive analysis of the importance of sentiment analy- is to both industry and academia, two state-of-the-art techniques were mployed to analyze data from Twitter concerning COVID-19 to deter- ine the impact of the pandemic on a global stage. Comparatively, a ramework was developed to evaluate the accuracy and performance of elected state-of-the-art machine learning and sentiment analysis tech- iques. In obtaining the dataset that was used for this study, a compre- ensive approach was used. The data obtained from Twitter was put hrough a series of processing to extract, clean and filter the relevant nformation. This was done using state-of-the-art data processing steps. his allowed us to use the correct data to train our framework to obtain he right results. A large amount of data related to COVID-19 and airline travel from witter was used to test the performance of the framework to estimate he impact of COVID-19 within the aviation industry. By implement- ng the proposed framework consisting of three main operators, namely ssemble + Deft Assessment (ASSEMBLEDEFT), Edify & Authenticate and orecast of New Prototypes, we found an improved performance in the entiment analysis and classification. The parameters used include pre- ision, accuracy, recall and F1-score. The outcome of the framework howed significant increase from 9.28% under a standard sentiment re- iew to 10.1% optimized sentiment analysis. With the implementation f the framework, there was an improved performance of the sentiment echniques. By virtue of improvement in the performance of sentiment analy- is using the parameters mentioned above, it can be concluded that the roposed framework can serve as a benchmark or baseline for future entiment analysis and classification. This can be used in association ith other machine learning or deep learning methods. This framework an also be used beyond the current topic of measuring sarcastic and on-sarcastic sentiments in the aviation industry. It can be used in any ervice industry where customer complaints can be collected and ana- yzed. The framework can be developed into a management information ystem for effective decision making by stakeholders within the aviation ndustry. This will provide support on how to better understand pas- engers’ needs and address their traveling issues effectively amid any andemic. unding None. thics approval Not applicable. onsent to participate All authors agreed to participate in this research. A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 C D D D E t G R D G i H t R H A H A H J Z Z J R K A K A K A K A K B K B K B K B K B K C L C L M C M C Q C D R onsent for publication All authors have agreed to publish this article. ata source The dataset considered for this study is sourced from Kaggle and con- ains comments of passengers on basis of service provided by airlines. efer to Section 3.1 for details of the studied dataset. eclaration of Competing Interest The authors declare that they have no known competing financial nterests or personal relationships that could have appeared to influence he work reported in this paper. eferences bid, F., Alam, M., Yasir, M., & Li, C. (2019). Sentiment analysis through recurrent vari- ants latterly on convolutional neural network of twitter. Future Generation Computer Systems, 95 , 292–308. 10.1016/J.FUTURE.2018.12.018 . dikari, A., Burnett, D., Sedera, D., de Silva, D., & Alahakoon, D. (2021). Value co-creation for open innovation: an evidence-based study of the data driven paradigm of social media using machine learning. International Journal of Information Management Data Insights, 1 (2), Article 100022. 10.1016/j.jjimei.2021.100022 . hao, B., Jiang, X., & Cao, J. (2020). Biostatistical analysis on covid-19. Annals of Medical & Surgical Case Reports: AMSCR-100045. GRF Publishers. 2(03 . ucco, C., Liang, H., Di Fatta, G., & Cannataro, M. (2018). Explainable sentiment analysis with applications in medicine. In IEEE International Conference on Bioinformatics and Biomedicine (BIBM) (pp. 1740–1747). IEEE . . Ahuja, A. Chug, S. Kohli, S. Gupta, and P. Ahuja (2019). The impact of features extraction on the sentiment analysis. Procedia Computer Science, 152, 341–348. 10.1016/j.procs.2019.05.008 . lamoodi, A. H., Zaidan, B. B., Zaidan, A. A., Albahri, O. S., Mohammed, K. I., Malik, R. Q., et al., (2021). Sentiment analysis and its applications in fighting COVID-19 and in- fectious diseases : a systematic review. Expert Systems With Applications, 167 , Article 114155 October 2020. 10.1016/j.eswa.2020.114155 . lharbi, A. S. M., & de Doncker, E. (2019). Twitter sentiment analysis with a deep neural network: An enhanced approach using user behavioral information. Cognitive Systems Research, 54 , 50–61. 10.1016/j.cogsys.2018.10.001 . loqaily, A., Al-Hassan, M., Salah, K., Elshqeirat, B., & Almashagbah, M. (2020). Sentiment analysis for Arabic tweets datasets: Lexicon-based and machine learning approaches. Journal of Theoretical and Applied Information Technology, 98 (4), 612–623 . vinash, M., & Sivasankar, E. (2019). A study of feature extraction techniques for sentiment analysis. Advances in Intelligent Systems and Computing, 814 , 475–486. 10.1007/978-981-13-1501-5_41 . harti, S. K., Babu, K. S., & Jena, S. K. (2015). Parsing-based sarcasm sentiment recog- nition in Twitter data. In Proceedings of the IEEE/ACM international conference on advances in social networks analysis and mining, ASONAM 2015 (pp. 1373–1380). 10.1145/2808797.2808910 . harti, S. K., Gupta, R. K., Pathik, N., & Mishra, A. (2022). Sarcasm detection in news headlines using voted classification. In Proceedings of the ACM international conference proceeding series (pp. 208–212). 10.1145/3549206.3549245 . havitha, B. K., Rodrigues, A. P., & Chiplunkar, N. N. (2017). Comparative study of ma- chine learning techniques in sentimental analysis. In Proceedings of the international conference on inventive communication and computational technologies, ICICCT 2017, ICI- CCT (pp. 216–221). 10.1109/ICICCT.2017.7975191 . ritzolakis, A., Kondylakis, H., & Papadakis, N. (2020). A review on lexicon-based and machine learning political sentiment analysis using tweets. International Journal of Semantic Computing, 14 (4), 517–563. 10.1142/S1793351X20300010 . rusaferri, A., Matteucci, M., Spinelli, S., & Vitali, A. (2020). Learning behavioral mod- els by recurrent neural networks with discrete latent representations with applica- tion to a flexible industrial conveyor. Computers in Industry, 122 , Article 103263. 10.1016/j.compind.2020.103263 . ambria, E., Das, D., Bandyopadhyay, S., & Feraco, A. (2017). Affective Computing and Sentiment Analysis. In E. Cambria, D. Das, S. Bandyopadhyay, & A. Feraco (Eds.), A Practical Guide to Sentiment Analysis. Socio-Affective Computing, 5 . Cham: Springer 10.1007/978-3-319-55394-8_1 . arvalho, T. P., Soares, F. A. A. M. N., Vita, R., Francisco, R., da, P., Basto, J. P., et al., (2019). A systematic literature review of machine learning methods applied to predic- tive maintenance. Computers and Industrial Engineering, 137 , Article 106024 August. 10.1016/j.cie.2019.106024 . han, S. W. K., & Chong, M. W. C. (2017). Sentiment analysis in financial texts. Decision Support Systems, 94 , 53–64. 10.1016/j.dss.2016.10.006 . hiarello, F., Bonaccorsi, A., & Fantoni, G. (2020). Technical sentiment analysis. mea- suring advantages and drawbacks of new products using social media. Computers in Industry, 123 , Article 103299. 10.1016/j.compind.2020.103299 . ortes, C., Vapnik, V., & Saitta, L. (1995). Support-vector networks editor. Machine leam- ing : 20. Kluwer Academic Publishers . a Silva, N. F. F., Hruschka, E. R., & Hruschka, E. R (2014). Tweet senti- ment analysis with classifier ensembles. Decision Support Systems, 66 , 170–179. 10.1016/j.dss.2014.07.003 . 17 as, D., & Clark, A. J. (2018a). Sarcasm detection on facebook: a supervised learning approach. In Proceedings of the 20th international conference on multimodal interaction, ICMI 2018 . 10.1145/3281151.3281154 . as, D., & Clark, A. J. (2018b). Sarcasm detection on Flickr using a CNN. In Proceedings of the ACM international conference proceeding series (pp. 56–61). 10.1145/3277104.3277118 . l Rahman, S. A., AlOtaibi, F. A., & AlShehri, W. A. (2019). Sentiment Analysis of Twit- ter Data, 2019 International Conference on Computer and Information Sciences (ICCIS) (pp. 1–4). Sakaka, Saudi Arabia. IEEE. 10.1109/ICCISci.2019.8716464 . kikas, D. C., Tzafilkou, K., Theodoridis, P. K., Garmpis, A., & Gkikas, M. C. (2022). How do text characteristics impact user engagement in social media posts: modeling content readability, length, and hashtags number in Facebook. Inter- national Journal of Information Management Data Insights, 2 (1), Article 100067. 10.1016/j.jjimei.2022.100067 . rover, P., Kar, A. K., & Dwivedi, Y. (2022). The evolution of social media influence - a literature review and research agenda. International Journal of Information Management Data Insights, 2 (2), Article 100116. 10.1016/j.jjimei.2022.100116 . assan, A., & Mahmood, A. (2017). Deep Learning approach for sentiment analysis of short texts. In Proceedings of the 3rd international conference on control, automation and robotics, ICCAR 2017 (pp. 705–710). 10.1109/ICCAR.2017.7942788 . e, S., Guo, F., & Qin, S. (2020). Sarcasm detection using graph convolutional networks with bidirectional LSTM. In Proceedings of the ACM international conference proceeding series (pp. 97–101). 10.1145/3422713.3422722 . emmatian, F., & Sohrabi, M. K. (2017). A survey on classification techniques for opinion mining and sentiment analysis. Artificial Intelligence Review, 52 (3), 1495–1545 2017 52:3. 10.1007/S10462-017-9599-6 . ussein, D. M. E. D. M. (2018). A survey on sentiment analysis challenges. Journal of King Saud University - Engineering Sciences, 30 (4), 330–338. 10.1016/j.jksues.2016.04.002 . ain, K., & Kaushal, S. (2018). A comparative study of machine learning and deep learn- ing techniques for sentiment analysis. In 7th International Conference on Reliabil- ity, Infocom Technologies and Optimization (Trends and Future Directions)(ICRITO) (pp. 483–487). IEEE. 10.1109/ICRITO.2018.8748793 . angid, H., Singhal, S., Shah, R. R., & Zimmermann, R. (2018). Aspect-based fi- nancial sentiment analysis using deep learning. ACM Digital Library , 1961–1966. 10.1145/3184558.3191827 . andasamy, I., Vasantha, W. B., Obbineni, J. M., & Smarandache, F. (2020). Sentiment analysis of tweets using refined neutrosophic sets. Computers in Industry, 115 , Article 103180. 10.1016/j.compind.2019.103180 . ar, A. K., Choudhary, S. K., & Singh, V. K. (2022). How can artificial intelligence im- pact sustainability: a systematic literature review. Journal of Cleaner Production, 376 , Article 134120 May. 10.1016/j.jclepro.2022.134120 . azmaier, J., & van Vuuren, J. H. (2020). A generic framework for sentiment analysis: leveraging opinion-bearing data to inform decision making. Decision Support Systems, 135 , Article 113304 April. 10.1016/j.dss.2020.113304 . haleghparast, S., Maleki, M., Hajianfar, G., Soumari, E., Oveisi, M., Golandouz, H. M., et al., (2023). Development of a patients’ satisfaction analysis system using ma- chine learning and lexicon-based methods. BMC Health Services Research, 23 (1), 280. 10.1186/s12913-023-09260-7 . hotijah, S., Tirtawangsa, J., & Suryani, A. A. (2020). Using LSTM for context based ap- proach of sarcasm detection in twitter. In Proceedings of the ACM international confer- ence proceeding series (pp. 1–7). 10.1145/3406601.3406624 . itsios, F., Mitsopoulou, E., Moustaka, E., & Kamariotou, M. (2022). User-generated con- tent behavior and digital tourism services: A SEM-neural network model for informa- tion trust in social networking sites. International Journal of Information Management Data Insights, 2 (1), Article 100056. 10.1016/j.jjimei.2021.100056 . ouadri, W. M., Ouziri, M., Benbernou, S., Echihabi, K., Palpanas, T., & Amor, I. B (2020). Quality of sentiment analysis tools: the reasons of inconsistency. In Proceedings of the VLDB endowment: 14 (pp. 668–681). 10.14778/3436905.3436924 . umar, S., Kar, A. K., & Ilavarasan, P. V. (2021). Applications of text mining in services management: a systematic literature review. International Journal of Information Man- agement Data Insights, 1 (1), Article 100008. 10.1016/j.jjimei.2021.100008 . ushwaha, A. K., Kar, A. K., & Dwivedi, Y. K. (2021). Applications of big data in emerging management disciplines: A literature review using text mining. Inter- national Journal of Information Management Data Insights, 1 (2), Article 100017. 10.1016/j.jjimei.2021.100017 . usyanti, A., & Zakia, D. (2019). Sentiment Analysis in the Mobile Application Review Document Using the Improved K-Nearest Neighbor Method. In 2019 International Conference on Sustainable Information Engineering and Technology (SIET) (pp. 332– 337). IEEE. 10.1109/SIET48054.2019.8986037 . ee, J., Woo, J., Kang, A. R., Jeong, Y. S., Jung, W., Lee, M., et al., (2020). Comparative analysis on machine learning and deep learning to predict post-induction hypotension. Sensors, 20 (16), 1–21 Switzerland. 10.3390/s20164575 . . Li, T.T. Goh, and D. Jin (2018). How textual quality of online reviews affect classifica- tion performance: a case of deep learning sentiment analysis. Neural Computing and Applications 2018 32:9, 32(9), 4387–4415. 10.1007/S00521-018-3865-7 . ehta, P., & Pandya, S. (2020). A review on sentiment analysis methodologies. Practices and Applications, 9 (02), 601–609 . orariu, C., Morariu, O., R ăileanu, S., & Borangiu, T. (2020). Machine learning for pre- dictive scheduling and resource allocation in large scale manufacturing systems. Com- puters in Industry, 120 . 10.1016/j.compind.2020.103244 . ian, J., Niu, Z., & Shi, C. (2018). Sentiment analysis model on weather related tweets with deep neural network. In Proceedings of the ACM international conference proceeding series (pp. 31–35). 10.1145/3195106.3195111 . ajadesingan, A., Zafarani, R., & Liu, H. (2015). Sarcasm detection on twitter:a behavioral modeling approach. In WSDM - Proceedings of the 8th ACM international conference on web search and data mining (pp. 97–106). 10.1145/2684822.2685316 . A.-M. Iddrisu, S. Mensah, F. Boafo et al. International Journal of Information Management Data Insights 3 (2023) 100180 R J S V W S S X S Y S H S i, H., Dqg, F., Ri, U., Pfv, Q. S., Qg, N. H. G. X., Ri, H., et al., (2021). Techniques of sarcasm detection: a review. In Proceedings of the international conference on advance computing and innovative technologies in engineering: 7 (pp. 5–9) . antos, Z. R., Cheung, C., Coelho, P. S., & Rita, P. (2022). Consumer engagement in so- cial media brand communities: a literature review. International Journal of Information Management, 63 November 2021. 10.1016/j.ijinfomgt.2021.102457 . atapathy, R., Cambria, E., & Hussain, A. (2017). Sentiment Analysis in the Bio-Medical Domain (p. 7). ISBN : 978-3-319-68467-3 . aura, J. R., Palos-Sanchez, P., Angel, M., & Martin, R. (2018). Attitudes expressed in online comments about environmental factors in the tourism sector: an ex- ploratory study. International Journal of Environmental Research and Public Health Arti- cle . 10.3390/ijerph15030553 . . Sohangir, D. Wang, A. Pomeranets, and T.M. Khoshgoftaar (n.d.). Big data: deep learn- ing for financial sentiment analysis. Journal of Big Data. 10.1186/s40537-017-0111-6 . ufi, F. K. (2022). Identifying the drivers of negative news with sentiment, entity and regression analysis. International Journal of Information Management Data Insights, 2 (1), Article 100074. 10.1016/j.jjimei.2022.100074 . wapnarekha, H., Behera, H. S., Nayak, J., & Naik, B. (2020). Role of intelligent comput- ing in COVID-19 prognosis: a state-of-the-art review. Chaos, Solitons & Fractals, 138 , Article 109947. 10.1016/J.CHAOS.2020.109947 . 18 oachims, T. (2019). Text categorization with support vector machines: learning with many relevant features. Culture Shift in Advanced Industrial Society , 335–370. 10.1515/9780691186740-014 . elavan, T. P., & Meyer, C. G. (2020). The COVID-19 epidemic. Tropical Medicine & Inter- national Health, 25 (3), 278–280. 10.1111/TMI.13383 . ang, P., Fan, E., & Wang, P. (2021). Comparative analysis of image classification algo- rithms based on traditional machine learning and deep learning. Pattern Recognition Letters, 141 , 61–67. 10.1016/j.patrec.2020.07.042 . iong, T., Zhang, P., Zhu, H., & Yang, Y. (2019). Sarcasm detection with self- matching networks and low-rank bilinear pooling. In Proceedings of the web confer- ence - proceedings of the world wide web conference, WWW 2019 (pp. 2115–2124). 10.1145/3308558.3313735 . akubu, H., & Kwong, C. K. (2021). Technological forecasting & social change fore- casting the importance of product attributes using online customer reviews and google trends. Technological Forecasting & Social Change, 171 , Article 120983 June. 10.1016/j.techfore.2021.120983 . . Yakubu and C. K. Kwong, Using Online Big Data for Determining the Importance of Product Attributes, 2020 IEEE International Conference on Industrial Engineer- ing and Engineering Management (IEEM), Singapore, Singapore, 2020, pp. 691–695, 10.1109/IEEM45057.2020.9309746 .