2.1 Online reviews and group characteristics

As formerly observed by ^{Oliveira et al. (2020}), previous studies have made use of online reviews to analyse different groups of travellers in the same destination. ^{Antonio, De Almeida, Nunes, Batista and Ribeiro (2018}) and ^{Phillips, Antonio, De Almeida and Nunes (2020}) analysed reviews from a set of Portuguese hotels in three different dialects and observed that cultural backgrounds as well as geographic distance influence hotel online reviews as travellers of different origins may have distinct expectancies. ^{Ahani et al. (2019}a; 2019c) collected reviews from 5-star hotels in Wellington, New Zealand, and four and five-star hotels in the Canary Islands, Spain, respectively. The findings of both researches detected various degrees of satisfaction with dissimilar preferences among customers and recommended hoteliers to segment travellers’ preference and satisfaction through data mining as a way to improve the quality of hotel products and services.

Besides collecting reviews from four different world capi^{tals, Francesco and Roberta (2019}), investigated if travellers from different countries put different emphasis on hotel attributes and if these attributes were perceived differently. After analysing reviews on Tripadvisor written by Italian, American and Chinese travellers, the results also suggested considerable dissimilarities in the sentiments customers were having regarding the hotels. According to the findings provided by ^{Li, Liu, Tan and Hu (2020}), guests’ expectations towards hotel performance might differ with respect to their geography. Although some similarities were found, different responses from local and international travellers regarding hotel attributes were registered.

^{Mariani and Predvoditeleva (2019}) examined the role and influence of guests' cultural traits and perceived experience through online review ratings of Russian hotels and their findings show that different online customer groups can be clustered into segments, as they display different online behaviours and give different online evaluations. Similar online review behaviour, was also observed by ^{Xu (2018}). Table 1 presents the main studies observed.

All aforementioned studies registered dissimilarities among guest’s satisfaction and preferences when customers presented different cultural traits and backgrounds, nationalities or group composition. Accordingly, this study intends to verify if the four chosen European markets also display disparities concerning satisfaction and dissatisfaction factors about hotel products and services offered in the same region.

2.2 Hotel attributes and satisfaction

Many previous studies aimed to identify customer’s sentiment towards specific hotel attributes. ^{Ban et al. (2019}) performs opinion mining techniques over 6,500 reviews to identify key hotel attributes that influence customer satisfaction. According to the findings of that particular study, staff service stands out as a fundamental feature to enhance customer experience and satisfaction, overcoming facilities, showing that intangible service has a great impact on customer experience. Also, ^{Almeida and Pelissari (2019}), ^{Annisa and Surjandari (2019}), ^{Chaves, Gomes and Pedron (2012}), ^{Li et. Al (2020}), ^{Xu (2019}, 2020), ^{Ying, Chan and Qi (2020}), identify a range of intangible services that are considered essential by guests. Although many of the attributes found by these studies seem to be aligned, it becomes noticeable that the type of hotel, geography and type of traveller influences customer perceptions and therefore, hotel ratings.

When it comes to tangible attributes, studies demonstrate that hotel facilities, as well as food and beverage, exert a very important influence over customer experience and satisfaction, with those attributes alternating ranking positions according to the geography and samples of each study (^{Almeida and Pelissari, 2019}; ^{Annisa & Surjandari, 2019}; ^{Chaves et al. (2012}); ^{Ban et al., 2019}, ^{Egresi, Puiu, Zotic & Alexandru, 2020}; ^{Gunasekar, & Sudhakar, 2019} and ^{Li et al., 2020}). Moreover, results from these studies demonstrate that room quality can directly sway profitability and guests’ willingness to recommend the accommodation, besides leading to customer grievances and unfavourable reviews.

Many studies also refer to location as a key attribute of satisfaction. This feature presents a significant impact on online rates and textual reviews as presented by ^{Anagnostopoulou, Buhalis, Kountouri, Manousakis and Tsekrekos (2019}), ^{Chaves et al. (2012}), ^{Gunasekar and Sudhakar (2019}) and ^{Li et al. (2020}). Outcomes of these studies demonstrated that hotel location is very significant when considering profitability and volume of bookings, which means that hoteliers should be aware that unchangeable attributes have also a considerable influence on business performance. Finally, online textual reviews can prove the quality of tangible and intangible services, which, in turn, must be the hotelier’s priorities, particularly when the main purpose of the business is to promote the well-being of their guests (Li et al. 2020; ^{Xu, 2019}).

3.1 Sampling and data collection

In order to accomplish the tasks proposed by this research, online reviews of hotels located in the Algarve and written by guests who reside in the UK, France and Portugal need to be collected from Tripadvisor. This study opted to work with this platform because it stands out as the leading global advice website dedicated to tourist products and services with more than 300 million members and 500 million reviews of hotels and other tourism-related featured items (^{Liang, Liu & Wang, 2019}; ^{Zhao et al., 2019}). Also, the website allows hotel practitioners to identify hotel characteristics closely connected to customer overall satisfaction at lower costs compared to traditional approaches (^{Litvin, 2019}).

Before starting the review collection process, it was necessary to check if ^{all 170 hotels listed by Registo Nacional de Turismo (2019}), were available on Tripadvisor’s website. All hotels were searched on the review platform, according to their names, address, postal code and the official website. A total of 161 hotels were considered eligible to have their reviews collected. Among the excluded sample, four hotels had duplicated entries, two establishments had not been reviewed, one accommodation had only one review, one hotel had been converted into a hostel and one unit was not found.

The second stage consisted of extracting the reviews of the selected hotels. First, each hotel unit has been given a singular ID number. Second, each hotel website link on Tripadvisor was included in an Excel spreadsheet and a bot built in C# has been used as a web scraper to collect online reviews posted from January 2019 to December 2019. A total of 15,585 ratings and text reviews written in three different languages were obtained from the platform. 10,744 reviews were written in English, 2,716 in Portuguese and finally 2,125 reviews were registered in French. Also, username and country of residence related to each rating and text review were collected.

3.2. Data refining

Data refining is an essential process when analysing a large volume of text. This extensive procedure consists mainly of converting unstructured text data into structured forms (^{Ban, Choi, Choi, Lee & Kim, 2019}). Consequently, a five-stage process shown in Figure 1 is conducted with the purpose of promoting as accurate data analysis as possible. First, the text reviews are segmented by language in an Excel spreadsheet to facilitate further market segmentation. Second, reviews missing the city or country of residence are excluded from the sample, since it precludes market segmentation. Next, all reviews in which users are not from the target markets are also excluded leading to a sample of 8,596 valid reviews. Then, with the help of an online translator, all reviews written in the foreign languages covered by this study were translated to their original language and segmented accordingly (for example, a review posted by a French resident in English is translated into French and segmented as French market). Finally, with the assistance of an online spell checker, all reviews are revised in order to be correctly interpreted by further analysis software.

Figure 1 Stages of data refining 

As aforementioned, the United Kingdom is Algarve’s most representative foreign market, representing almost 40% of the total overnight guests in the area. Its representation is reflected in the total number of reviews (5,677), which surpasses the number of Portuguese (1,664) and French (1,255) reviews combined. Figure 2 presents the final amount of valid reviews for each of the selected markets.

Figure 2 Number of valid reviews per market 

Although the number of valid reviews for each language differs, previous studies, such as ^{García-Pablos, Duca, Cuadros, Linaza and Marchetti (2016}), ^{Antonio et. Al (2018}), ^{Phillips et al. (2020}), also work with different amounts of reviews per language, with goals that include the verification of sentiment strength from different languages towards accommodations or destinations. Furthermore, all valid reviews correspond to the same time period and region, therefore avoiding bias when it comes to the sentiment analysis of the reviewers. For this reason, this study takes into consideration all valid reviews from each language in order to obtain results for the chosen region and timeline.

3.3 Sentiment analysis

Sentiment analysis - also known as opinion mining - is the field of study responsible for analysing individuals’ reviews, opinions, sentiments and emotions towards products and services. This type of investigation mainly focuses on positive or negative sentiments expressed on people’s opinions (^{Liu, 2012}). It is also described as a computational study that investigates texts, which include someone’s opinion about a specific business, such as hotel reviews (Çalı & Balaman, 2019).

It is also through the conduction of opinion mining that a more precise interpretation of text reviews is assured since this method aims to classify texts according to their sentiment orientation. This categorization can be binary (positive, negative), ternary (positive, neutral, negative) or to the context of reviews, as for instance, thumbs up or thumbs down, approval or disapproval (Çalı & Balaman, 2019; ^{Duan, Yu, Cao & Levy, 2016}; ^{Han, Mankad, Gavirneni & Verma, 2016}).

Travellers’ online reviews contain objective and subjective statements about hotels and travel experiences. Consequently, the sentiment analysis’ main aim is to deal with the subjective expressions utilized to describe emotions, opinions and feelings (Çalı & Balaman, 2019; ^{Liu, 2012}). Moreover, sentiment analysis is proven to be an important tool to quantify guest’s opinions in the textual component of reviews (^{Antonio et al. 2018}). Thereby, as a means to reach the goals proposed by this research, a lexicon-based approach sentiment analysis is adopted to identify the polarity of each selected online review. By doing so, it will be possible to uncover and classify any relevant semantic and emotional information regarding the targeted markets selected by this study.

3.4 Data analysis

Throughout the years, Tripadvisor’s database has become noticeable in the hospitality industry due to the fact that its extraction and handling eases the detection of business strengths and weaknesses, benchmarking, measurement of customer overall satisfaction, hotels image improvement and market visibility (^{Lima & Viana, 2017}). The platform is also widely used by scholars in pursuit of useful and strategic information regarding the hotel industry. ^{Ahani et al. (2019}a, 2019c), Ahani, Nilashi, Ibrahim, Sanzogni and Weaven (2019b), ^{Han et al. (2016}), ^{Hu and Chen (2016}), Li, Law, Vu, Rong and Zhao (2015), ^{Litvin (2019}), ^{Yi, Li and Jai (2018}), ^{Yu, Li and Jai (2017}), Zhao, ^{Xu and Wang (2019}) are recent examples of studies that made use of Tripadvisor as a primary source of information to identify customer overall satisfaction, preferences and segment markets to provide new insights as well as managerial and theoretical implications. Hence, the online textual reviews analysed in this study are open and unstructured data acquired from Tripadvisor. Even though the data has been previously refined, some of the challenges faced by this research are the impossibility to employ direct analysis methods, such as standard statistical or econometric procedures, and also to deal with a substantial number of reviews, which can be very time consuming and lead to biased results if done manually (Han et al., 2016; ^{Limberger, Meira, Añaña & Sohn, 2016}; ^{Loo & Leung, 2018}).

For this reason and as pointed out by recent studies such as ^{Ahani et al. (2019}a); ^{Ban et al. (2019}); Çalı and Balaman, (2019); ^{Francesco and Roberta, (2019}) and ^{Oliveira et al. (2020}), an appropriate software devoted to analysing large quantity of data is required. Therefore NVivo 12, a tool developed to assist researchers in qualitative data analysis, was chosen to assist the following content analysis of this study. The chosen tool is then used for the following procedures: first, all refined data has been imported into NVivo 12 separately; after, three different projects were created. One for each market and local language (English, Portuguese and French); next, a general sentence-level sentiment analysis is carried out to detect the sentiment polarity in each sentence of the text reviews. By doing so, this study aims to identify the overall sentiment of each market towards their hotel experience in the Algarve. In this first inquiry, the total amount of reviews and all their content were taken into consideration. Then, a frequency analysis is conducted in order to rank the 30 most frequent nouns related to hotel attributes. As a result, it was possible to uncover which hotel attributes are mostly discussed by each market in the online reviews posted on Tripadvisor. Also, frequency analysis uncovered common codes, also known as themes, which helped to define which hotel attributes could be compared afterwards. Finally, an aspect level sentiment analysis was conducted to identify the sentiment polarity of opinions expressed on seven different hotel attributes, namely: room, pool, staff, breakfast, restaurant, beach and view. Finally, a comparison between all three markets is performed.

4.1 Comparison of overall market sentiment analysis

As shown in Figure 3, the results obtained from the general sentence-level sentiment analysis present a binary categorization of the sentiment found in the total number of reviews of each market. The automatic process carried by NVivo 12 applies a scoring system where each word containing sentiment has a pre-defined score and it is classified on a Likert-type sentiment scale under the following labels: very negative, moderately negative, moderately positive and very positive. It is possible to visualize that English and Portuguese reviews followed a similar pattern, while French reviews remained centred in the two moderate groups. Still, all three markets wrote more positive sentences than negative ones in their reviews, which validates ^{Bayer and Emir, (2017}) as well as ^{Tontini, Dos Santos Bento, Milbratz, Volles and Ferrari (2017})’s findings that travellers mainly write reviews to share their positive experiences instead of the negative ones. The English, Portuguese and French markets presented respectively 71%, 69%, 75% of absolute positive sentiment towards their hotel experience in the Algarve, which represents an average of 71.6%. In terms of overall satisfaction, results show that, if on the one hand, the local market shows a higher percentage of very positive sentiment, on the other hand, it is behind the two other markets in terms of general positive sentiment. Furthermore, although presenting a very low percentage of very positive sentiment, the French residents’ reviews display three-quarters of positive sentiment.

Figure 3 Overall sentiment of the reviews 

When conducting the sentiment analysis of the unstructured data, NVivo searches for expressions of sentiment in the source material, however, it is important to highlight that the tool only can analyse one language at a time and does not recognize sarcasm, double negatives, slangs, dialect variations, idioms or ambiguity. Still, the software provides a handy overview when it comes to the analysis of a large volume of data in a short period of time.

Regarding the total number of sentiment coding references in relation to the total number of reviews detected in each language, it was verified that slightly over 10% of French reviews were classified as neutral. In other words, these coding references were considered to have no sentiment at all. In turn, English and Portuguese text reviews had less than 2% of neutral coding references detected. Also, French reviews presented more negative coding references (5.8%), which represents more negative mentions than the two other markets summed. Concerning mixed references, when you have both positive and negative sentiments identified in the same text passage, results show that more than half of the English reviews (52.6%) mention pros and cons about their hotel experience on a single Tripadvisor review. This pattern of writing reviews was followed by the Portuguese (40%) and only by a quarter of the French posts. With regard to positive coding references, these were mainly detected on French reviews (57.1%), followed by Portuguese (55.1%) and English samples (43.5%). Figure 4 shows the comparison among all three markets.

Figure 4 Overall sentiment of coding references 

4.2 Frequency analysis

As previously performed by ^{Antonio et al. (2018}), ^{Ban et al. (2019}), ^{Gerdt, Wagner and Schewe (2019}), as well as ^{Khoo-Lattimore and Ekiz (2014}) a frequency analysis was conducted in order to uncover the 30 most mentioned words associated with hotel experience in each language. This type of analysis is essential, especially at the beginning of the textual data analysis, since it can assist in identifying which hotel attributes are mostly discussed and assessed in the reviews, besides revealing significant results about the interaction between what is written in all three different languages.

Taking this into account, unrelated words, verbs, adjectives, articles, conjunctions, prepositions, pronouns and others were excluded from the query. Only nouns related to hotel attributes were maintained. Also, the word hotel, which was the most frequent noun in the reviews was removed since it is not a hotel attribute, but the hotel unit itself. As shown in Table 2, after conducting the word search query in each one of the languages, it was found that room, staff and pool are respectively the three most mentioned words, regardless of the language used. Similar results are found in ^{Khoo-Lattimore and Ekiz (2014}), which found room, staff and food as the most recurrent themes. Also, six terms amongst the top ten most mentioned words are the same in all languages. Apart from the three words aforementioned, breakfast, restaurant and beach figure amongst the ten most-cited hotel attributes.

The outcomes of this query support the selection of hotel attributes that are going to have their sentiment detected and contrasted among all selected markets in order to answer whether satisfaction and dissatisfaction attributes are similar among the British, French and Portuguese markets.

4.3. Automatic and manual theme coding

After conducting the word search query, an automatic coding process is executed on NVivo to identify the most relevant themes in each one of the three languages. The unstructured data are analysed by NVivo 12 with the assistance of a language pack. Themes are then automatically detected by analysing the content and the sentence structure within it. In doing so, the significance is assigned to some themes over others based on how frequently they occur in the reviews.

The themes found by the tool are combined into groups and the results are presented as a code for each broad idea, with child codes attached to each main theme. All the relevant content is coded to the theme codes that are created. The results are summarized in a code matrix which shows the codes for each broad idea, and the number of coding references from each source. Still, human supervision is necessary since text analytics is a complex process and manual coding is always going to be more accurate owing to the fact that human sentiments, such as sarcasm, slangs, idioms or ambiguity cannot be identified by the tool.

During the human checking process, it was verified that the French and Brazilian Portuguese expressions for breakfast, respectively ‘petit-déjeuner’ and ‘café da manhã’ were broken into different themes. When separate, the word ‘petit’ means small and ‘déjeuner’ means lunch. Hence, human supervision was necessary to identify if customers were evaluating breakfast, lunch or saying that something was small. The same applies to the expression ‘café da manhã’. It was necessary to check if comments were related to breakfast (café da manhã), coffee (café) or morning (manhã). Also, similar themes were merged. For example, ‘hôtel’ and ‘l’hôtel’ (the hotel) were unified among French themes. ‘Pool’ and ‘pool area’ were merged among English themes, since their comments were related to the same hotel attribute.

Finally, taking into account the theme coding results and the frequency word outcomes, six hotel attributes are chosen to have their sentiment compared: staff; breakfast; restaurant; pool; room and beach. Besides figuring among the most ten frequent words in all languages, these expressions are also among the most mentioned detected themes. Furthermore, sentiment towards the theme ‘hotel’ is also checked in the interest of verifying how hotels in the Algarve are evaluated by the selected markets.

4.4. Sentiment analysis towards selected hotel attributes

After concluding the previous process, each hotel attribute is individually submitted to a sentiment analysis performed by NVivo 12. Next, the outcomes are compared between all markets to verify if satisfaction and dissatisfaction attributes are similar among them. However, a prior sentiment analysis towards the theme hotel is conducted with the purpose of verifying the selected markets’ general opinion about the topic.

As shown in Figure 5, Portuguese reviews show the highest overall level of dissatisfaction over hotels (32.8%), followed by English (28.2%) and French reviews (26.2%). With respect to positive sentiment, French reviews present almost three quarters (73.9%) of approving text content, whilst only 4.9% were detected as very positive, which is almost seven times weaker in positive sentiment when compared to English reviews (34.8%) and Portuguese reviews (36.0%). According to previous results shown in Figure 4, French reviews sentiments are significantly centred in moderate sentiment levels, whilst English and Portuguese reviews are similarly distributed. Still, as shown in Figure 5, considerable sentiment differences are observed amongst all languages.

Figure 5 Sentiment towards hotel 

With regard to the verified hotel attributes, other dissimilarities are observed as presented in Figure 6. When both moderately positive and very positive sentiment are summed up, reviews written in French stand out among all six selected attributes. That leads to the conclusion that French reviews are in general more positive than the two other analysed languages. Nevertheless, a significant small amount of very positive sentiment is found in French reviews. Staff is the hotel attribute in which the highest percentage of very positive sentiment was found (11%). Breakfast is the feature with the highest rate of moderately positive sentiment (93.8%) and absolute positive sentiment in French reviews (93.8%), while room is the most criticized. Not only approximately 30 percent of negative sentiment towards hotel room is detected, but also the highest amount of very negative sentiment (4.3%).

With respect to English reviews, the highest absolute value of positive sentiment is found towards beach (81.3%), followed by staff (73.7%) and restaurant (73.5%). When it comes to very positive sentiment, staff is the most praised hotel feature among English residents (40.2%). At the other end, as observed in the French reviews, room presents the highest negative sentiment rate, both on an absolute (33%) and very negative sentiment rate (20.9%). Finally, Portuguese reviews present the highest level of very positive sentiment towards one single attribute: beach (50.6%), followed by breakfast (40.7%) and pool (37.4%). Beach also corresponds to the highest absolute positive rate of Portuguese reviews (77.1%). In regard to wholly negative sentiment, Portuguese reviews also demonstrated most of their dissatisfaction towards room (34.1%), followed by staff (33.1%), which was also the attribute where most of the very negative reviews were focused on (16.7%). Figure 6 shows the sentiment comparison between all three languages towards each hotel feature.

Figure 6 Sentiment towards hotel attributes