helpfn

Received August 23, 2020, accepted September 10, 2020, date of publication September 22, 2020, date of current version October 1, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.3025823

Aspect Based Sentiment Analysis of Ridesharing Platform Reviews for Kansei Engineering SAQIB ALI 1,2, GUOJUN WANG 1, AND SHAZIA RIAZ 2 1School of Computer Science, Guangzhou University, Guangzhou 510006, China 2Department of Computer Science, University of Agriculture, Faisalabad 38000, Pakistan

Corresponding author: Guojun Wang ([email protected])

This work was supported in part by the National Natural Science Foundation of China under Grant 61632009, in part by the Guangdong Provincial Natural Science Foundation under Grant 2017A030308006, and in part by the High-Level Talents Program of Higher Education in Guangdong Province under Grant 2016ZJ01.

ABSTRACT At present online reviews are becoming an important source for Kansei engineering of the services provided by ridesharing platforms. Kansei engineering deals with incorporating customer feedback and demands into product and service design. Thus, it is used as a tool for organizations to uplift their businesses by considering customer reviews and feedback. Customer reviews available on social media are in unstructured form; therefore, sentiment analysis is employed to extract customer’s opinions in a systematic manner. In India-Pakistan, these reviews are mostly in Roman Urdu/Hindi and English, which are of great value for ridesharing platforms as a part of their Kansei engineering strategy. However, sentiment analysis cannot be performed directly on these reviews as they are mostly in Roman Urdu/Hindi. Therefore, the objective of this paper is to conduct aspect based sentiment analysis on these reviews after translating them into English for Kansei engineering of the service. Consequently, sentiment analysis is carried out to extract the most frequent features along with nouns and adjectives used by the customers to express their sentiments. We extracted prominent aspects of the service (i.e., ‘Driver’, ‘Company’, ‘Service’, and ‘Ride’) based on their highest frequencies using aspect based sentiment analysis. The customer sentiments are then clustered into these main aspects using unsupervised machine learning technique. Each aspect is further analyzed based on their polarity, which serves as an input for Kansei engineering of the service. As a result, it can facilitate ridesharing companies to enhance their businesses by improving services in accordance with customer demands.

INDEX TERMS Roman Urdu sentiment, Kansei engineering, aspect based sentiment analysis, polarity classification, ridesharing platform reviews.

I. INTRODUCTION The technical advancements in businesses have changed the traditional business strategies from product-oriented design to a customer-oriented design by linking it with customer’s satisfaction and feedback. It makes the customer’s require- ments as much important as other functional requirements of a successful business. Similarly, in ridesharing platforms, the customer’s satisfaction is a key to their success in a highly competing market. Now, it has become a challenge for these companies to achieve maximum customer satisfaction covering all the aspects of their services like performance, pricing, ease of use, etc. The technique to integrate the cus- tomer’s feedback into product and service design is known as

The associate editor coordinating the review of this manuscript and

approving it for publication was Yiming Tang .

‘Kansei engineering’ [1]. The Kansei engineering aims to develop a relationship between customer’s sentiments and product design parameters [2]. This will help the product to achieve maximum customer satisfaction, which in turn uplifts the company’s business in the market [3].

In the past, customer opinions were obtained through questionnaires, literature reviews, interviews, and large scale web-based surveys [4]–[7]. In questionnaires and surveys, the surveyor typically tries to map the product design param- eters with customer satisfaction [8], [9]. In these methods, a scale called a semantic differential scale (SD) is developed based on Kansei attributes. The scale consists of some arbi- trary points to get the respondent’s degree of satisfaction to evaluate the product design [10]. These methods are used to get high-quality data on a small scale. For example, in past customer sentiments about ten products were

173186 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ VOLUME 8, 2020

S. Ali et al.: Aspect Based Sentiment Analysis of Ridesharing Platform Reviews for Kansei Engineering

evaluated by only a group of seven users [11]. Therefore, the results of these interviews and questionnaires cannot be trusted. Whereas, in large scale web-based surveys, there are hundreds of online respondents. However, these responses were highly biased by the attitude and commitment of the participants.

Another important source of getting customer feedback is customer reviews available on online social networks. Customers post their reviews on social networks without any biases [12]. Thus, these reviews depict customer’s sentiment in the true sense at large scale. This interactive content is of great importance for corporations, product designers, and government agencies, etc. [13]. As the reviews are enormous and unstructured, therefore, sentiment analysis is carried out to extract the sentiments conveyed by the customers about the quality of the product [14], [15]. These sentiments are then used as an input for Kansei engineering of product design and improvement, which enable the businesses to remain in market competition [16].

The ridesharing platforms are growing rapidly all over the world especially in South Asian countries like India and Pakistan during the last few years. The majority of the people do not have their own transport in this region, which contains a one-fifth population of the world. Therefore, it has raised the competition among the ridesharing companies like Uber, Careem, Ola, Taxi, Blacklane, Talixo, and Lyft. It makes them much more conscious about the quality of their services. In the India-Pakistan region, the most dominant language is Urdu/Hindi. Therefore, most of the ridesharing customers prefer to post their comments/reviews in Roman Urdu instead of English. The existing work available in literature applied the sentiment analysis only on the reviews posted in the English language [17]. Thus, it does not depict the actual sentiments of customers about the quality of service in this region. Further, the existing work on sentiment analysis in ridesharing platforms considers the service as a whole with- out splitting it into its key aspects, i.e., ‘Driver’, ‘Company’, ‘Service’, and ‘Ride’. As a result, it failed to reveal which aspect of the service is adversely affecting the performance. Moreover, for Kansei engineering, we need customer’s feed- back for each and every aspect of the service in order to improve them according to customer demands.

The objective of this research is to carry out aspect based sentiment analysis to extract customer’s sentiments from online reviews posted in Roman Urdu/Hindi and English as required for the Kansei engineering of the service. We believe that this is the first work that considers Roman Urdu/Hindi reviews along with English for the aspect based sentiment analysis of the service in the region. In this work, the target ride sharing service is Uber as it has more than 50%1 mar- ket share in India and Pakistan. The customer reviews are extracted from a third-party platform (i.e., Facebook) instead of the company’s own webpage/mobile app for maintaining

1https://investor.uber.com/news-events/news/press-release-details/2020/ Uber-Announces-Results-for-First-Quarter-2020/default.aspx

their transparency. In the first step, the reviews available in Roman Urdu/Hindi are translated into English using cus- tomized APIs. Next, the sentiment analysis is carried out on combined reviews to extract the most frequent features as well as nouns and adjectives to mine the sentiments of the customers. This leads to mine the prominent aspects (i.e., ‘Driver’, ‘Company’, ‘Service’, and ‘Ride’) of the service based on their highest frequencies using aspect based sen- timent analysis. This unique approach provides a compre- hensive insight into the service in terms of the customer’s perspective. Finally, the sentiments of the customers are clus- tered with positive and negative categories into these main aspects of the service using unsupervised machine learning. This classification helps to find the most discussed aspect of the company in terms of quality of service. It serves as an input for Kansei engineering of the service in order to design and improve it according to customer’s demands. The main contributions of the paper are as follows.

1) This work is first of its kind as it analyzed customer reviews posted in two languages (i.e., Roman Urdu/Hindi and English) collectively for Kansei engineering of ridesharing platform.

2) We mined customers’ sentiments to extract the most frequent features, nouns, and adjectives to express their sentiments in online reviews.

3) We carry out aspect based sentiment analysis to extract the prominent aspects (i.e., ‘Driver’, ‘Company’, ‘Service’, and ‘Ride’) of the service based on their highest frequencies. Later, customer sentiments are clustered into these aspects using an unsupervised machine learning approach.

4) Finally, the positive and negative sentiments of the customers are separated out for each aspect to act as an input for Kansei engineering of the service.

The rest of this paper is organized as follows. Section II reviews the related work. Section III outlines the method- ology of this work. Results and discussions are detailed in Section IV. Finally, the conclusion and future work are drawn in Section V.

II. RELATED WORK In this section, we discussed the existing work on the basis of Kansei engineering, sentiment analysis, and their applica- tions in ridesharing platforms.

A. KANSEI ENGINEERING The history witnessed countless research studies in the field of Kansei engineering to discover the association between product design parameters and customer intentions [18]. Kansei is a word taken from the Japanese language used to depict customers’ sentiments and expressions. In recent work, Kansei engineering was carried out using different product parameters. In these studies, the researchers take the end user’s opinion about mobile phones on a five-point scale [19]–[21]. The same scale with four bipolar parameters

VOLUME 8, 2020 173187

S. Ali et al.: Aspect Based Sentiment Analysis of Ridesharing Platform Reviews for Kansei Engineering

is used to get customer perception about home service robots [22]. Chou [11] formulates the computing with the word approach for Kansei engineering and uses a seven-point scale to evaluate customer perception. Further, the author finds out the fuzzy relationship among Kansei parameters and classify them accordingly using a questionnaire. The ques- tionnaire surveys are extensively used for Kansei evaluation to know the customer intentions. For example, about home delivery service or product warranty. Moreover, these surveys are used to find an association between the sentiments of the user (represented by simple words in the questionnaire) and the design elements such as parameters, properties, and characteristics of the service design [23]. In [24], the authors evaluate the product by considering all important steps of product manufacturing including marketing strategies using a Kansei engineering approach. However, all these examples have a small number of respondents who are solely selected by the researchers. Thus, it does not truly represent the population and hence highly biased.

B. SENTIMENT ANALYSIS Online reviews available on e-commerce web site and social media can solve the above-said problem [15]. Customer reviews available on these sites are increasing at a very fast pace due to rapid Internet growth [2]. Therefore, sentiment analysis is extensively used to extract and evaluate customer opinions and sentiments from these online reviews in var- ious fields [16], [25], [26]. For example, online customer reviews are used to check people’s attitudes towards ad block- ing [27]. The researcher used the topic modeling approach and showed that people expressed negative attitude towards the websites that contain ads and positive attitude that do not. Similarly, online customer reviews are collected from popular e-commerce website i.e., Amazon and used to analyze the customer opinion about the ‘‘recliner’’ to carry out its Kansei engineering [15].

Further, the researchers used deep learning techniques to conduct sentiment analysis of product reviews available on Amazon. They developed a framework to judge a review sentence’s orientation based on polarity classification [28]. Similarly, Majumder et al. [29] emphasized that sentiment analysis can be benefited with sarcasm detection showing the intensity of user sentiments. They used a deep neural network to build a multitask learning-based framework. Similarly, Wang et al. [30] identified improvement in twitter sentiment analysis by combining diffusion patterns with text messages.

C. ASPECT BASED SENTIMENT ANALYSIS An important category of sentiment analysis is aspect based mining, which provides comprehensive opinions or senti- ments about different aspects of the product or service [31]. Aspect based mining is executed in two stages, in the first stage, opinion attributes are identified and in the second stage, their polarities are figured out. A similar approach was used for review mining of restaurants to analyze customer feedback [14]. Similarly, Yang et al. [32] developed a deep

neural network with contextual, lexical, and syntactic clues for aspect based sentiment analysis. However, Do et al. [33] reviewed and suggested that deep leaning techniques used for aspect based sentiment analysis achieved better accuracy but still faces challenges in data, languages, and domains.

D. MULTILINGUAL SENTIMENT ANALYSIS With the increased use of social media throughout the world, it is not appropriate to consider users’ sentiments only in English. There is a dire need to work on multilingual reviews to extract users’ true sentiments. As a result, multilingual sentiment analysis is gaining attention in Chinese, Spanish, Arabic, German, Italian, French, Japanese, Malay, Urdu, and Romanian languages along with English [34]–[36]. For example, Babelsenticnet [37] is a concept level knowledge- base for multilingual sentiment analysis that provides support for more than forty languages including Hindi/Urdu. Simi- larly, in [38], authors build a multilingual sentiment classifier on the Malaysian social media disaster dataset to find the sentiment conveyed in Malay and English languages.

In the South Asia region, most people are unable to speak or even understand English. Consequently, the majority of people prefer to type reviews in Urdu. However, they do not have enough skills to type in Urdu directly. Thus, they switch to Roman Urdu for expressing their sentiments on social media or e-commerce sites. As discussed earlier, Babelsen- ticnet provides support for Urdu/Hindi languages. However, it does not support Roman Urdu as it is a poor resource language. Therefore, in this paper, we have analyzed social media dataset containing multilingual reviews of users in Roman Urdu and English to extract their true sentiments using unsupervised machine learning techniques.

E. RIDE SHARING PLATFORMS The ridesharing service is a transportation network that pro- vides a platform to connect drivers having private automo- biles and customers looking for a taxi. There are many issues in these platforms (e.g., Uber, Careem, Ola, Taxi, Blacklane, Talixo, and Lyft) regarding the QoS as mentioned in [39]. For example, Shokoohyar [40] investigates the ride sharing platforms from driver’s perspective by using online reviews of drivers to compare Lyft and Uber. The results conclude that Uber is slightly better than Lyft in terms of driver’s job perspective. Saragih and Girsang [41] compare prominent ride sharing platforms in Indonesia and discuss the feedback of users about these platforms. Similarly, Baj-Rogowska [17] classifies customer opinions about Uber services as positive, negative, and neutral by using customer reviews posted in the English language. The author performs sentence-level sentiment analysis and results are generalized by considering the Uber service as a whole.

Comparing with the above work, this paper deals with the reviews available in both Roman Urdu/Hindi and English languages in India-Pakistan region considering all important aspects of the service. The reviews are extracted from third party online social network site (i.e., Facebook) to maintain

173188 VOLUME 8, 2020

S. Ali et al.: Aspect Based Sentiment Analysis of Ridesharing Platform Reviews for Kansei Engineering

their transparency. In Roman Urdu/Hindi, a customer uses Roman alphabets to express their sentiments in a free and open manner. Therefore, it is important to analyze these Roman Urdu/Hindi along with English comments to depict the actual sentiments of the customers about the service in this region.

III. METHODOLOGY This work is centered on aspect based sentiment analysis of customer reviews for Kansei engineering of Uber service. The snapshot of the proposed methodology is illustrated in Figure 1. In this work, we have selected Uber as a tar- get service. Next, we have collected the online customer reviews expressing their sentiments about the service in Roman Urdu/Hindi and English languages. In the next step, Roman Urdu/Hindi comments are translated into the English

FIGURE 1. Proposed methodology - Aspect based sentiment analysis of ridesharing platform reviews for Kansei engineering.

language. After pre-processing, aspect based sentiment anal- ysis is carried out on most frequent features along with nouns and adjectives to extract prominent aspects of the service. Finally, customers’ sentiments are clustered into these aspects based on their polarities using unsupervised machine learning technique to perform Kansei engineering of the service.

A. TARGET SERVICE - UBER In this work, we have selected Uber as a target service. It is operational in 85 countries comprising of almost 785 metropolitan areas with over 110 million active users. Besides ridesharing, Uber is also offering other services to its customers such as food delivery, electric bicycle-sharing system, etc. Uber started his business in India and Pakistan since 2013 and 2016, receptively. At present, the service is operational in more than 50 metropolitan areas of the region. The service is a highly popular mode of transport in the region as most of the people in these countries do not have their own vehicle. A huge amount of customer reviews are available online about the performance of the service, which can serve as an input for Kansei engineering after sentiment analysis.

B. DATA ACQUISITION Facebook is an excellent platform to collect customer reviews about any service as it has more than 2.5 billion monthly active users. It provides a neutral platform for the customers to post their sentiments freely without any biasedness. Further, companies have no right to change them for their business purposes. Therefore, the Uber Face- book page is selected to collect the customer reviews with the help of a web crawler (i.e., Facepager available at https://github.com/strohne/Facepager/releases) instead of Uber’s own app/web page. The reviews are saved directly into the CSV file. Each review has a unique userID, postID, date, time, and review text. After filtering, the resulting file contains 3853 reviews in Roman Urdu and English languages collectively.

C. SORTING OUT COMMENTS This step separates Roman Urdu and English reviews in two different files. Roman Urdu reviews are translated into the English language to get all reviews in one language and make them ready for the next pre-processing step. Customized APIs are used to translate these reviews into English. The example of converting Roman Urdu review into the English language is shown in Table 1.

D. DATA PRE-PROCESSING The sorted reviews are in an unstructured form therefore, data pre-processing is carried out in five different steps. These steps include Tokenization, Parts of Speech (POS) tagging, Removal of stop words, Stemming & Lemmatization, and Vectorization. Each step is elaborated with the target review text in Table 2.

VOLUME 8, 2020 173189

S. Ali et al.: Aspect Based Sentiment Analysis of Ridesharing Platform Reviews for Kansei Engineering

TABLE 1. Example - conversion of roman urdu review into english.

TABLE 2. Data pre-processing of review text.

1) Tokenization: This process breaks down a text sequence into different smaller pieces such as word, phrase, symbol, etc. based on the grammar of the lan- guage. For this purpose, first of all, sentence segmen- tation is used to divide every customer’s review into individual sentences. The punctuation marks such as period and comma are used as a marker of the sentence. Afterward, in word segmentation, the sentences are divided into words and phrases where space is used as an indicator of the next word in a sentence. These small pieces are called tokens. Python NLP package and Natural Language Toolkit (NLTK) [42] are used to carry out tokenization in this work. The example of the sentence and word segmentation are shown in row 1 and 2 of Table 2.

2) Parts of Speech (POS) Tagging: In this process, part of speech tag is assigned to each word in the text. Python NLTK package is used for this purpose. It uses different combinations of letters for each part of speech. For example, ‘NN’ is a tag used to represent noun, ‘RB’ for an adverb, ‘VB’ for a verb, and ‘JJ’ for an adjective. The example of POS tagging on review text is shown in row 3 of Table 2. For example, ‘ride’ is a noun, therefore, ‘NN’ is assigned to it. Likewise, ‘VB’ is attached to the word ‘use’ to represent it as a verb. In this way, all the words obtained after tokenization are assigned with respective POS tags. This tagging is an important step of the data pre-processing phase, as a result, the required features are identified and extracted easily.

3) Removal of Stop words: The English language sen- tences contain an extensive portion of stop words that have little or no value in text mining and natural language processing. Removal of stop words is the process, which is used to eliminate the inappropriate words in the reviews i.e., is, am, are, including different punctuation marks. The objective of this work is to extract the most frequent features from the reviews, however, stop words always have a high frequency in any type of textual data. Therefore, it is extremely important to remove these stop words before perform- ing any analysis on the text. This process reduces the size of the review text to a great extent as shown in row 4 of Table 2. Python Scikit Learn (SKLEARN) and NLTK libraries provide built-in lists of stop words for 16 different languages. The English stop words built-in list provided by SKLEARN is used to eliminate the frequently used stop words from the review text. Moreover, some custom words are also included in the built-in list to remove them while extracting the most frequent features from the text.

4) Stemming & Lemmatization: It is the process of gen- erating root words from different modified forms. The process of stemming truncates the suffix or prefix of the word without considering its meaning. For example, it extracts ‘driv’ from ‘driving’ which is not a real word. However, in lemmatization, lemma will be an actual language word as it considers the morphology of the word. For example, it generates the word ‘drive’ from ‘driving’, ‘drives’, or ‘drove’. The purpose of this step

173190 VOLUME 8, 2020

S. Ali et al.: Aspect Based Sentiment Analysis of Ridesharing Platform Reviews for Kansei Engineering

FIGURE 2. Vectorization of reviews.

is to reduce the volume of the vocabulary set generated after pre-processing of the reviews.

5) Vectorization: The pre-processed reviews as obtained through previous steps, are in an unstructured form. Therefore, it is required to convert them into structured data, which is known as a bag of words model (BoW) or vector space model. The BoW is simply a bag of words containing the review text without considering its grammar and sequence. For this purpose, a dic- tionary is established that consists of all the words present in Review 1 and Review 2 as shown in Figure 2. Afterward, each review becomes a fixed-length vector with dimension equal to the size of the dictionary. The value of each word in the vector is filled with the frequency of that word in the review text as shown by Vector 1 and Vector 2 in Figure 2. The CountVector- izer package of scikit-learn library is used for the said vectorization process [43].

E. EXTRACTION OF MOST FREQUENT FEATURES, NOUNS, AND ADJECTIVES After pre-processing, the most frequent features are extracted from the reviews using SKLEARN CountVectorizer, Collec- tion, and RE packages of the Python. We extracted more than a hundred features from the reviews. However, after sorting and filtering, the total features are reduced to fifty as shown in Figure 3. The descriptive statistics reveal that ‘Driver’, ‘Company’, ‘Service’, and ‘Ride’ have the highest frequency in the data. In order to get comprehensive sen- timents about the service, most frequent nouns, and adjec- tives are also mined separately form the reviews and are summarized in Figure 4 and Figure 5. We found that the

highest frequent features like ‘Driver’, ‘Company’, ‘Service’, and ‘Ride’ are almost the same as nouns with the highest frequency. Afterward, the extraction of most frequent adjec- tives demonstrates that ‘New’, ‘Late’, ‘Wrong’, ‘Disappoint’, and ‘Fraudulent’ adjectives have the highest frequency as compared to others.

F. ASPECT EXTRACTION Aspects are those words that are frequently used by the customers to express their sentiments about the service, thus, representing their significance. These aspects are actually concealed in the reviews since customers do not specify them clearly in the reviews. Therefore, the highest frequency fea- tures from the most frequent feature list are taken as aspects of the service i.e., ‘Driver’, ‘Company’, ‘Service’, and ‘Ride’. Similar aspects have been revealed when we parsed the most frequent nouns, which denote their prominence for Kansei engineering of the service. Thereafter, customer sentiments are clustered based on these aspects using a Probabilistic Latent Semantic Analysis (PLSA) [44] as explained in the next section.

G. CLUSTERING AND POLARITY CLASSIFICATION The clustering is done by applying the PLSA using Matrix Factorization Model. As a result, we get four clusters contain- ing the most frequent words used by the customers. Each clus- ter has aspect word as a top word with the highest frequency i.e., ‘Driver’, ‘Company’, ‘Service’, or ‘Ride’. It confirms that these highest frequency words are the main aspects of the Uber service. The snippet of clustering is shown in Table 3. Next, we use a lexicon-based approach using SentiWord-

Net [45] for the polarity classification of the sentiments used

VOLUME 8, 2020 173191

S. Ali et al.: Aspect Based Sentiment Analysis of Ridesharing Platform Reviews for Kansei Engineering

FIGURE 3. Most frequent features from reviews.

FIGURE 4. Most frequent nouns from reviews.

by the customers. In this process, each aspect with its opin- ion words are passed to SentiWordNet as a key value pair e.g., (service, good), (ride, complaint), (driver, disappoint). Afterward, SentiWordNet retrieves the opinion from this key value pair and calculates its polarity. Based on the calcu- lated polarity, each opinion word is regarded as positive or negative.

IV. RESULTS AND DISCUSSION A. FREQUENCY ANALYSIS OF MOST FREQUENT FEATURES, NOUNS, AND ADJECTIVES The extraction of most frequent features revealed that the word ‘Driver’ was the most discussed feature (with the highest frequency of 2428) in the reviews. Similarly, the other highly discussed features are ‘Service’, ‘Ride’, and

173192 VOLUME 8, 2020

S. Ali et al.: Aspect Based Sentiment Analysis of Ridesharing Platform Reviews for Kansei Engineering

FIGURE 5. Most frequent adjectives from reviews.

TABLE 3. Snippet of clustering generated using PLSA.

‘Company’ as shown in Figure 3. In the case of most frequent nouns, the same features (‘Driver’, ‘Company’, ‘Service’, and ‘Ride’) have the highest frequency as shown in Figure 4. As a result, these four features and nouns collectively become the prominent aspects of the service. In terms of adjectives, most repeated ones are ‘New’, ‘Late’, ‘Wrong’, ‘Disappoint’, ‘Horrible’, and ‘Good’ as shown in Figure 5. These most fre- quent aspects and adjectives define the relationship between service providers i.e., Uber and customer demands. This rela- tionship is an aid for Kansei engineering of Uber to fine-tune their services according to the intentions of the customers.

B. FREQUENCY ANALYSIS OF CLUSTERS In order to mine hidden sentiments about the service, cus- tomer sentiments are clustered based on each prominent aspect of the service (i.e., ‘Driver’, ‘Company’, ‘Service’, and ‘Ride’). Additionally, the polarity classification of sentiments in each cluster is carried out. It helps to reveal the customer feedback about the service in a true sense. This feedback serves as an input for Kansei engineering of the service.

In this work, the Driver’s aspect constitutes a biggest clus- ter containing maximum number of features with highest frequency as shown in Table 4. In this cluster, the customers

TABLE 4. Frequency chart of Driver’s cluster.

discussed Driver’s aspect using nouns and adjectives like ‘Issue’, ‘Late’, ‘Disappoint’, ‘Terrible’, ‘Horrible’, ‘Worst’, ‘Ridiculous’, ‘Bad’ and ‘Poor’ to express their negative sen- timents about the service. Similarly, they used adjective like ‘Better’ and ‘Good’ to express positive feedback about the service. However, the frequency of negative sentiment is much higher as compared to the positive one. The higher frequency of negative sentiments conclude that customers are not satisfied with the behavior of the ‘Driver’. Therefore, it is an important result for Kansei engineering of the Uber service to tame their drivers according to expectations of the customers.

The next cluster is about the aspect ‘Company’ (i.e., Uber). It forms the second highest frequency cluster and is shown in Table 5. In this cluster, customers used words like ‘Issue’, ‘Bad’, ‘Worst’, ‘Terrible’, ‘Horrible’, ‘Fraudulent’ and ‘Poor’ to show their dissatisfaction towards the service. On the other

VOLUME 8, 2020 173193

S. Ali et al.: Aspect Based Sentiment Analysis of Ridesharing Platform Reviews for Kansei Engineering

TABLE 5. Frequency chart of Company’s cluster.

hand, they used ‘Good’, ‘Better’, ‘like’ and ‘New’ to express their positive sentiments about the Company. The frequency of negative words used in this cluster is less as compared to Driver’s cluster, which shows that customers have an opti- mistic behavior towards the Company.

The third cluster represents the ‘Service’ aspect of the Uber as shown in Table 6. Again, the customers mostly used negative words like ‘Issue’, ‘Helpless’, ‘Worst’, ‘Bad’, ‘Terrible’, ‘Horrible’, etc. while used positive words ‘Like’, ‘New’, ‘Good’ and ‘Better’ to express their sentiments about the Uber service. Finally, in the fourth cluster, aspect ‘Ride’ is evaluated and results are summarized in Table 7. The cluster is also dominant with negative sentiments showing customers bad experience about the ‘Ride’ aspect. The frequency of negative words used in this cluster concludes that customers are highly concerned about the ‘Ride’ aspect of the service after driver’s behavior.

TABLE 6. Frequency chart of service cluster.

In summary, the negative words like ‘Issue’, ‘Terrible’, ‘Horrible’, ‘Bad’, and ‘Worst’ are present in all clusters with moderate to high frequency. However, the positive words like ‘Good’ and ‘Better’ have low to moderate frequency. These findings indicate that the customers are mostly complain- ing about different aspects of the Uber service. In order to improve the service, Uber can incorporate this aspect based sentiment analysis in their Kansei engineering strategy to uplift their business in the region.

TABLE 7. Frequency chart of ride cluster.

V. CONCLUSION AND FUTURE WORK The purpose of this work is to perform aspect based sentiment analysis to extract the customer sentiments for Kansei engi- neering of the Uber service in the India-Pakistan region. For this purpose, online reviews are collected from the Facebook page of Uber service where users post their sentiments openly without any bias. As the major portion of these reviews is in Roman Urdu/Hindi, therefore, they are translated into English using customized APIs. Afterward, aspect based sentiment analysis is carried out on most frequent features, nouns, and adjectives to get prominent aspects of the service i.e., ‘Driver’, ‘Company’, ‘Service’, and ‘Ride’. Further, pos- itive and negative sentiments of the customers are clustered for every aspect of the service. This revealed that the majority of the customers are complaining about the service and they are highly concerned about the ‘Driver’ and the ‘Ride’ aspect of the service. Thus, Uber can integrate this aspect based sen- timent analysis in its Kansei engineering process to enhance its business in the region.

As Roman Urdu is a poor resource language and does not contain a standardized dictionary. The spellings of a word vary greatly from person to person or even the same person uses different spellings on different occasions. Therefore, in this work lexicon-based approaches with base machine learning models are giving quite satisfactory results. However, for a huge amount of Roman Urdu reviews that con- tain different spellings for the same word, deep learning mod- els can be used in the future to provide fine-tuned analysis.

REFERENCES [1] M. Nagamachi and A. M. Lokman, Innovations of Kansei Engineering,

vol. 2. Boca Raton, FL, USA: CRC Press, 2016, doi: 10.1201/b16799. [2] Z. Li, Z. G. Tian, J. W. Wang, and W. M. Wang, ‘‘Extraction of affective

responses from customer reviews: An opinion mining and machine learn- ing approach,’’ Int. J. Comput. Integr. Manuf., vol. 33, no. 7, pp. 670–685, 2020, doi: 10.1080/0951192X.2019.1571240.

[3] G. Zhong, K. Zhang, H. Wei, Y. Zheng, and J. Dong, ‘‘Marginal deep architecture: Stacking feature learning modules to build deep learning models,’’ IEEE Access, vol. 7, pp. 30220–30233, 2019, doi: 10.1109/ ACCESS.2019.2902631.

[4] S. Smith and S.-H. Fu, ‘‘The relationships between automobile head-up display presentation images and drivers’ Kansei,’’ Displays, vol. 32, no. 2, pp. 58–68, Apr. 2011, doi: 10.1016/j.displa.2010.12.001.

173194 VOLUME 8, 2020

S. Ali et al.: Aspect Based Sentiment Analysis of Ridesharing Platform Reviews for Kansei Engineering

[5] J. Park and S. H. Han, ‘‘A fuzzy rule-based approach to modeling affective user satisfaction towards office chair design,’’ Int.J. Ind.Ergonom., vol. 34, no. 1, pp. 31–47, Jul. 2004, doi: 10.1016/j.ergon.2004.01.006.

[6] H.-H. Lai, Y.-C. Lin, C.-H. Yeh, and C.-H. Wei, ‘‘User-oriented design for the optimal combination on product design,’’ Int. J. Prod. Econ., vol. 100, no. 2, pp. 253–267, Apr. 2006, doi: 10.1016/j.ijpe.2004.11.005.

[7] H.-B. Yan, V.-N. Huynh, T. Murai, and Y. Nakamori, ‘‘Kansei evalu- ation based on prioritized multi-attribute fuzzy target-oriented decision analysis,’’ Inf. Sci., vol. 178, no. 21, pp. 4080–4093, Nov. 2008, doi: 10. 1016/j.ins.2008.06.023.

[8] K. Y. Chan, C. K. Kwong, T. S. Dillon, and K. Y. Fung, ‘‘An intelligent fuzzy regression approach for affective product design that captures nonlin- earity and fuzziness,’’ J. Eng. Des., vol. 22, no. 8, pp. 523–542, Aug. 2011, doi: 10.1080/09544820903550924.

[9] H. Jiang, C. K. Kwong, K. W. M. Siu, and Y. Liu, ‘‘Rough set and PSO- based ANFIS approaches to modeling customer satisfaction for affec- tive product design,’’ Adv. Eng. Informat., vol. 29, no. 3, pp. 727–738, Aug. 2015, doi: 10.1016/j.aei.2015.07.005.

[10] H.-B. Yan and Y. Nakamori, ‘‘A probabilistic approach to Kansei profile generation in Kansei engineering,’’ in Proc. IEEE Int. Conf. Syst., Man Cybern., Oct. 2010, pp. 776–782, doi: 10.1109/ICSMC.2010. 5641848.

[11] J.-R. Chou, ‘‘A Kansei evaluation approach based on the technique of computing with words,’’ Adv. Eng. Informat., vol. 30, no. 1, pp. 1–15, Jan. 2016, doi: 10.1016/j.aei.2015.11.001.

[12] R. Wang, D. Zhou, M. Jiang, J. Si, and Y. Yang, ‘‘A survey on opin- ion mining: From stance to product aspect,’’ IEEE Access, vol. 7, pp. 41101–41124, 2019, doi: 10.1109/ACCESS.2019.2906754.

[13] G. Vinodhini and R. Chandrasekaran, ‘‘A sampling based sentiment mining approach for e-commerce applications,’’ Inf. Process. Manage., vol. 53, no. 1, pp. 223–236, Jan. 2017, doi: 10.1016/j.ipm. 2016.08.003.

[14] M. Al-Smadi, M. Al-Ayyoub, Y. Jararweh, and O. Qawasmeh, ‘‘Enhancing aspect-based sentiment analysis of Arabic hotels’ reviews using morpho- logical, syntactic and semantic features,’’ Inf. Process. Manage., vol. 56, no. 2, pp. 308–319, Mar. 2019, doi: 10.1016/j.ipm.2018.01.006.

[15] W. Kim, T. Ko, I. Rhiu, and M. H. Yun, ‘‘Mining affective experience for a Kansei design study on a recliner,’’ Appl. Ergonom., vol. 74, pp. 145–153, Jan. 2019, doi: 10.1016/j.apergo.2018.08.014.

[16] L. Yue, W. Chen, X. Li, W. Zuo, and M. Yin, ‘‘A survey of sentiment analysis in social media,’’ Knowl. Inf. Syst., vol. 60, no. 2, pp. 617–663, Aug. 2019, doi: 10.1007/s10115-018-1236-4.

[17] A. Baj-Rogowska, ‘‘Sentiment analysis of Facebook posts: The Uber case,’’ in Proc. 8th Int. Conf. Intell. Comput. Inf. Syst. (ICICIS), Dec. 2017, pp. 391–395, doi: 10.1109/INTELCIS.2017.8260068.

[18] W. M. Wang, Z. Li, Z. G. Tian, J. W. Wang, and M. N. Cheng, ‘‘Extracting and summarizing affective features and responses from online product descriptions and reviews: A Kansei text mining approach,’’ Eng. Appl. Artif. Intell., vol. 73, pp. 149–162, Aug. 2018, doi: 10.1016/j.engappai. 2018.05.005.

[19] K. Y. Fung, C. K. Kwong, K. W. M. Siu, and K. M. Yu, ‘‘A multi-objective genetic algorithm approach to rule mining for affective product design,’’ Expert Syst. Appl., vol. 39, no. 8, pp. 7411–7419, Jun. 2012, doi: 10. 1016/j.eswa.2012.01.065.

[20] C. K. Y. Fung, C. K. Kwong, K. Y. Chan, and H. Jiang, ‘‘A guided search genetic algorithm using mined rules for optimal affective product design,’’ Eng. Optim., vol. 46, no. 8, pp. 1094–1108, Aug. 2014, doi: 10. 1080/0305215X.2013.823196.

[21] H. Jiang, C. K. Kwong, Y. Liu, and W. H. Ip, ‘‘A methodology of integrat- ing affective design with defining engineering specifications for product design,’’ Int. J. Prod. Res., vol. 53, no. 8, pp. 2472–2488, Apr. 2015, doi: 10.1080/00207543.2014.975372.

[22] Y. Zhu and G. Chen, ‘‘Research on the head form design of service robots based on Kansei engineering and BP neural network,’’ Proc. SPIE, vol. 10322, Jan. 2017, Art. no. 103222D, doi: 10.1117/12.2267419.

[23] M.-C. Chen, C.-L. Hsu, K.-C. Chang, and M.-C. Chou, ‘‘Applying Kansei engineering to design logistics services—A case of home delivery ser- vice,’’ Int. J. Ind. Ergonom., vol. 48, pp. 46–59, Jul. 2015, doi: 10.1016/j. ergon.2015.03.009.

[24] C. K. Kwong, H. Jiang, and X. G. Luo, ‘‘AI-based methodology of integrating affective design, engineering, and marketing for defining design specifications of new products,’’ Eng. Appl. Artif. Intell., vol. 47, pp. 49–60, Jan. 2016, doi: 10.1016/j.engappai.2015.04.001.

[25] E. Cambria, ‘‘Affective computing and sentiment analysis,’’ IEEE Intell. Syst., vol. 31, no. 2, pp. 102–107, Mar. 2016, doi: 10.1109/MIS.2016.31.

[26] S. Poria, N. Majumder, R. Mihalcea, and E. Hovy, ‘‘Emotion recogni- tion in conversation: Research challenges, datasets, and recent advances,’’ IEEE Access, vol. 7, pp. 100943–100953, 2019, doi: 10.1109/ACCESS. 2019.2929050.

[27] A. A. Tudoran, ‘‘Why do Internet consumers block ads? New evidence from consumer opinion mining and sentiment analysis,’’ Internet Res., vol. 29, no. 1, pp. 144–166, 2019, doi: 10.1108/IntR-06-2017-0221.

[28] W. Zhao, Z. Guan, L. Chen, X. He, D. Cai, B. Wang, and Q. Wang, ‘‘Weakly-supervised deep embedding for product review sentiment analy- sis,’’ IEEETrans.Knowl.DataEng., vol. 30, no. 1, pp. 185–197, Jan. 2018, doi: 10.1109/TKDE.2017.2756658.

[29] N. Majumder, S. Poria, H. Peng, N. Chhaya, E. Cambria, A. Gelbukh, and E. Cambria, ‘‘Sentiment and sarcasm classification with multitask learning,’’ IEEE Intell. Syst., vol. 34, no. 3, pp. 38–43, May 2019, doi: 10. 1109/MIS.2019.2904691.

[30] L. Wang, J. Niu, and S. Yu, ‘‘SentiDiff: Combining textual information and sentiment diffusion patterns for Twitter sentiment analysis,’’ IEEE Trans. Knowl. Data Eng., vol. 32, no. 10, pp. 2026–2039, Oct. 2020, doi: 10.1109/TKDE.2019.2913641.

[31] F. Tang, L. Fu, B. Yao, and W. Xu, ‘‘Aspect based fine-grained sentiment analysis for online reviews,’’ Inf. Sci., vol. 488, pp. 190–204, Jul. 2019, doi: 10.1016/j.ins.2019.02.064.

[32] T. Yang, Q. Yin, L. Yang, and O. Wu, ‘‘Aspect-based sentiment anal- ysis with new target representation and dependency attention,’’ IEEE Trans. Affect. Comput., early access, Oct. 1, 2019, doi: 10.1109/TAFFC. 2019.2945028.

[33] H. H. Do, P. Prasad, A. Maag, and A. Alsadoon, ‘‘Deep learning for aspect- based sentiment analysis: A comparative review,’’ Expert Syst. Appl., vol. 118, pp. 272–299, Mar. 2019, doi: 10.1016/j.eswa.2018.10.003.

[34] K. Dashtipour, S. Poria, A. Hussain, E. Cambria, A. Y. A. Hawalah, A. Gelbukh, and Q. Zhou, ‘‘Multilingual sentiment analysis: State of the art and independent comparison of techniques,’’ Cognit. Comput., vol. 8, no. 4, pp. 757–771, Aug. 2016, doi: 10.1007/s12559-016-9415-7.

[35] S. L. Lo, E. Cambria, R. Chiong, and D. Cornforth, ‘‘Multilingual senti- ment analysis: From formal to informal and scarce resource languages,’’ Artif. Intell. Rev., vol. 48, no. 4, pp. 499–527, Dec. 2017, doi: 10.1007/ s10462-016-9508-4.

[36] L. R. C. Pessutto, D. S. Vargas, and V. P. Moreira, ‘‘Multilingual aspect clustering for sentiment analysis,’’ Knowl.-Based Syst., vol. 192, Mar. 2020, Art. no. 105339, doi: 10.1016/j.knosys.2019.105339.

[37] D. Vilares, H. Peng, R. Satapathy, and E. Cambria, ‘‘BabelSenticNet: A commonsense reasoning framework for multilingual sentiment anal- ysis,’’ in Proc. IEEE Symp. Ser. Comput. Intell. (SSCI), Nov. 2018, pp. 1292–1298, doi: 10.1109/SSCI.2018.8628718.

[38] M. J. Fuadvy and R. Ibrahim, ‘‘Multilingual sentiment analysis on social media disaster data,’’ in Proc. Int. Conf. Electr., Electron. Inf. Eng. (ICEEIE), vol. 6, Oct. 2019, pp. 269–272, doi: 10.1109/ICEEIE47180. 2019.8981479.

[39] S. Banerjee and R. Johari, ‘‘Ride sharing,’’ in Sharing Economy: Making Supply Meet Demand, vol. 6, M. Hu, Ed. Cham, Switzerland: Springer, 2019, pp. 73–97, doi: 10.1007/978-3-030-01863-4_5.

[40] S. Shokoohyar, ‘‘Ride-sharing platforms from drivers’ perspective: Evi- dence from uber and LYFT drivers,’’ Int. J. Data Netw. Sci., pp. 89–98, 2018, doi: 10.5267/j.ijdns.2018.10.001.

[41] M. H. Saragih and A. S. Girsang, ‘‘Sentiment analysis of customer engagement on social media in transport online,’’ in Proc. Int. Conf. Sus- tain. Inf. Eng. Technol. (SIET), Nov. 2017, pp. 24–29, doi: 10.1109/SIET. 2017.8304103.

[42] E. Loper and S. Bird, ‘‘NLTK: The natural language toolkit,’’ 2002, arXiv:cs/0205028. [Online]. Available: https://arxiv.org/abs/cs/ 0205028

[43] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, and J. Vanderplas, ‘‘Scikit-learn: Machine learning in Python,’’ J. Mach. Learn. Res., vol. 12, pp. 2825–2830, Oct. 2011.

[44] T. Hofmann, ‘‘Unsupervised learning by probabilistic latent semantic analysis,’’ Mach. Learn., vol. 42, nos. 1–2, pp. 177–196, 2001, doi: 10. 1023/A:1007617005950.

[45] S. Baccianella, A. Esuli, and F. Sebastiani, ‘‘Sentiwordnet 3.0: An enhanced lexical resource for sentiment analysis and opinion mining,’’ in Proc. Int. Conf. Lang. Resour. Eval. (LREC), vol. 10, 2010, pp. 2200–2204.

VOLUME 8, 2020 173195

S. Ali et al.: Aspect Based Sentiment Analysis of Ridesharing Platform Reviews for Kansei Engineering

SAQIB ALI received the Ph.D. degree in com- puter science from the Universiti Teknologi Malaysia. He is currently working as a Postdoc- toral Research Fellow with the School of Com- puter Science, Guangzhou University, Guangzhou, China. He is a part of the PingER project led by the SLAC National Accelerator Laboratory, Stanford, USA, in collaboration with PERN, MYREN, CERNET, and Guangzhou University, to monitor and analyze the performance of the

Internet links using advance machine learning techniques. His research interests include big data analytics, mobile computing, and artificial intelligence.

GUOJUN WANG received the B.Sc. degree in geophysics, the M.Sc. degree in computer science, and the Ph.D. degree in computer science from Central South University, China, in 1992, 1996, and 2002, respectively. He had been a Profes- sor with Central South University, an Adjunct Professor with Temple University, USA, a Vis- iting Scholar with Florida Atlantic University, USA, a Visiting Researcher with the University of Aizu, Japan, and a Research Fellow with The

Hong Kong Polytechnic University, Hong Kong. He is currently a Pearl River Scholarship Distinguished Professor of higher education, Guangdong. He is also a Ph.D. Supervisor with the School of Computer Science and Cyber Engineering, Guangzhou University, China. His research interests include artificial intelligence, big data, cloud computing, mobile computing, trustworthy/dependable computing, cyberspace security, recommendation systems, and mobile healthcare systems.

SHAZIA RIAZ received the B.S. and M.S. degrees in computer science, in 2005 and 2011, respec- tively. She is currently pursuing the Ph.D. degree in computer science with the Department of Computer Science, University of Agriculture, Faisalabad, Pakistan. She is also working as an Assistant Professor with the Department of Com- puter Science, Government College Women Uni- versity Faisalabad, Pakistan. Her current research interests include data mining, machine learning,

and application of deep learning in big data analytics.

173196 VOLUME 8, 2020