The Differences and Dynamics of eWOM on Travel Review Platforms Investigating Reviews from TripAdvisor and Expedia
Lisa de Klerk – 11384182 Master’s Thesis
Graduate School of Communication
Research Master’s programme Communication Science Damian Trilling
Abstract
Review platforms in the tourism industry provide opportunities for consumers to share their experiences with other travelers in the form of reviews. This type of electronic word-of-mouth (eWOM) forms a rich data source that reflects consumers’ evaluation of travel products, and has been becoming more popular in social media analytics. Previous eWOM research often relied on only one data source, overlooking the differences and dynamics of reviews on different travel review platforms. The aim of this study was twofold: 1) identify the textual differences between reviews on different types of review platforms, and 2) investigate the dynamics between reviews on different types of review platforms. Using text analytics techniques to analyze reviews from TripAdvisor (i.e. a consumer-generated platform) and Expedia (i.e. a marketer-generated platform), this study comparatively analyzed the review length, topics and sentiment of both platforms. Secondly, time series analysis was performed to test the influence of review topics between the platforms over time. The findings show that TripAdvisor has lengthier reviews and higher mentions of major review topics, while similar positively-skewed distributions were found for the sentiment on TripAdvisor and Expedia. Additionally, a small short-term effect was found from the presence of the topics on Expedia on the presence of the topics on TripAdvisor. This study provides evidence for the inherent differences between consumer-generated and marketer-generated travel platforms, and presents preliminary evidence for the dynamic nature of eWOM on review platforms. Implications are presented for using travel review platforms as a data source.
Keywords: eWOM, hospitality, review platforms, social media analytics, text analytics, topic modelling, time series
The Dynamics and Differences of eWOM on Travel Review Platforms Investigating Travel Reviews from TripAdvisor and Expedia
In many consumer decision situations it is necessary to make choices about products or services without having any personal experience with the product or alternative products. In order to reduce uncertainty and perceived risks in such situations (Bronner & de Hoog, 2011), consumers often rely on product information from advertising or the experiences and opinions of others by word-of-mouth (WOM) (Rabanser & Ricci, 2005). Traditionally WOM has referred to offline interpersonal information sources (Bronner & de Hoog, 2011), such as family members, friends, or neighbors, but a new type of WOM has emerged online during the past two decades. This type of WOM has been coined as electronic WOM (eWOM) and is considered to be the most powerful and influential source of information about products and services for consumers in their decision-making process (e.g Bore, Rutherford, Glasgow, Taheri, & Antony, 2017; Sun, Youn, Wu, & Kuntataporn, 2006). This electronic aspect has increased the influence of WOM “due to its speed, convenience, one-to-many reach, and its absence of face-to-face human pressure” (Sun et al., 2006, p. 1105). An important underlying aspect of eWOM’s considerable influence is the fact that it is user-generated content (UGC), which often tends to be perceived as more credible, reliable, and accurate, and thus more useful in making a decision than information created by marketers (Erkan & Evans, 2016; Zhu & Zhang, 2010). Furthermore, the increased reliance of consumers on online information sources over interpersonal sources, and the enduring growth of social media (King, Racherla, & Bush, 2014), emphasizes the power and importance of eWOM.
This study focuses specifically on the hospitality and tourism domain, where eWOM plays an unprecedented role. Due to the risky nature of the travel domain, where “no product or service is like the other” (Rabanser & Ricci, 2005, p. 161), the need for adequate assistance and information when ‘hotel shopping’ is especially high for travelers. Travelers often resort
to online reviews shared by fellow travelers on travel review platforms, such as TripAdvisor, Expedia, or Hostelworld. About 90 percent of travelers consult the Internet in their decision-making process, of which over 60 percent rely on reviews (Bore et al., 2017).
Abundant research has demonstrated the influential nature of eWOM1. Conventional
approaches on studying the effects of eWOM on consumer behavior primarily relied on experiments and panel studies (King et al., 2014). A drawback of these approaches is that it raises the question of ecological validity (Schmuckler, 2001), where the observed consumer behavior resulting from exposure to simulated reviews might not be generalizable to real-life situations. A new approach has been emerging in research on online reviews in the past years, where studies are taking advantage of the readily available and rapidly growing data resources on eWOM platforms (Bore et al., 2017; Xiang, Du, Ma, & Fan, 2017). More and more
studies2 are extracting online reviews from travel platforms to detect, describe, and predict
meaningful patterns, and contributing to the knowledge on eWOM effects (e.g. Crotts, Mason, & Davis, 2009; Liu, Law, Rong, Gang, & Hall, 2013). However, most research using real-world data in this area fails to acknowledge that review platforms differ tremendously, and selecting only one data source can lead to generalizability problems. Xiang et al. (2017) provided an overview of recent studies using online travel reviews, showing that only three out of 22 studies between 2009 and 2017 used more than one data source. To my knowledge, only one study comparatively analyzed the content on multiple travel review platforms. The study by Xiang et al. (2017) used a big data analytics approach to comparatively examine popular review platforms (i.e. TripAdvisor, Expedia, and Yelp) and found huge discrepancies in terms of linguistics and semantic features, sentiment, and ratings between the platforms. This underlines the importance of platform selection.
1 For reviews of eWOM research see: Cheng & Zhou, 2010; Cheung & Thadani, 2012; King, Racherla, & Bush, 2014. For reviews of eWOM research in the travel domain see: Bore et al., 2017; Cantallops & Salvi, 2014. 2 Xiang et al. (2017, p.53) provide an overview of recent studies using data analytics to study online reviews from travel platforms.
Furthermore, current eWOM studies using these data resources also fail to
acknowledge the dynamic nature of eWOM. Common definitions of eWOM disregard the dynamic processes that happen over time on review platforms, and only touch on eWOM as a singular statement of opinion by a customer. Liu (2006) conducted one of the few studies that acknowledges the dynamic patterns of eWOM, which they clarify as “how it evolves” (p. 75). This evolving and dynamic character is an important aspect of eWOM and it is argued here that it should be studied as such. My study sees eWOM communication as an ongoing process over time, whereby consumers influence each other, contributing to the complex and dynamic eWOM sphere. Some early studies (e.g. Allsop, Bassett, & Hoskins, 2007; Trusov, Bucklin, & Pauwels, 2009) mention the dynamic patterns of eWOM, but no effort is made to study these patterns with data sources readily available online.
This study makes use of actual hotel reviews gathered from travel review platforms, and will build on the work of Xiang et al. (2017) by verifying the results in a different context (i.e. Amsterdam). Furthermore, it contributes to eWOM knowledge by examining the
dynamic pattern of eWOM with an exploratory study regarding the dynamics of topics mentioned by eWOM contributors. Therefore, two research questions are formulated:
RQ1: Does the content of eWOM reviews on Amsterdam-based properties differ in terms of linguistic features, semantic features, and sentiment for different types of travel review platforms?
RQ2: To what extent does the presence of topics in the past influence the presence of topics in the future between platforms?
Theoretical Framework
In order to investigate the differences and dynamics of eWOM content on different types of review platforms in the travel industry, it is fundamental to understand how travel review platforms operate. In this section, reviews as a type of eWOM communication and the
application of travel review platforms as a channel for eWOM are discussed. Next, a typology of travel review platforms is provided which constitutes a basis for the selection of platforms in research. Lastly, related studies with regards to the methods, platforms, and theory are reviewed before formulating hypotheses and additional research questions.
Consumer Reviews on Travel Review Platforms
Based on the eWOM literature, it can be concluded that the Internet has had a large impact on the ways in which consumers share product information. The definition of eWOM states that it is “any positive or negative statement made by potential, actual, or former customers about a product or company, which is made available to a multitude of people and institutions via the Internet” (Hennig-Thurau, Gwinner, Walsh, & Gremler, 2004, p. 39). As mentioned, for the hotel industry, the Internet and the subsequent existence of travel review platforms has played an especially important role. On these platforms, the statements of customers about the visited hotels are shaped in the form of online reviews. Consumer reviews are “widely considered a rich data source that reflects consumer experiences and evaluation of products” (Xiang et al., 2017, p. 51). Since these reviews are UGC, they are seen as an independent source of information to other travelers (Gretzel & Yoo, 2008). In order to understand the different types of platforms, it is important to first understand how travel review platforms operate.
Consumers can write reviews on the review platforms provided by sites such as TripAdvisor, Hostelworld, Expedia, or Booking.com. These travel websites employ
recommendation systems (RSs), which “align destination information with booking facilities to convert destination search into online bookings” (Rabanser & Ricci, 2005, p. 160). Without getting into too much detail on the technological functioning of these platforms, RSs function by providing personalized or non-personalized ‘advice’ to consumers based on their assumed or predicted interests (Ricci, Rokach, & Shapira, 2011). Besides offering property information
to help consumers in their decision-making process, these travel platforms also devote a section of the property’s page to reviewing. Although most travel review platforms have dedicated very different portions of their website to reviewers, it remains an important aspect of their function.
Travel Review Platform Typology
The type of review platform and the characteristics of the review platforms determine, for a large part, what reviewers will contribute. This is demonstrated by Bronner and de Hoog’s (2011) study, which found that the choice of travel review platform by consumers is determined by the type of motivation they have for sharing their experiences. The choice for type of platform in turn influences the content that is contributed and made available to other consumers. Therefore, the business orientation of a travel platform attracts certain consumers and has an influence on the content, both technologically and ideologically. The study also provided a framework for determining the type of travel review platform. This framework is used to select review platforms for analyzing eWOM content.
Bronner and de Hoog (2011) identified three types of travel review platforms: 1) marketer-generated platforms, 2) consumer-generated platforms, and 3) mixed forms. Marketer-generated platforms, such as online travel agents (OTA), are characterized by a relatively small review platform. This is in contrast with consumer-generated sites, where consumers exchange product information on platforms created by consumers, such as travel blogs. Furthermore, there are also platforms which mixed characteristics, such as commercial platforms with a larger section devoted reviews.
In addition, review platforms that fall under the same type can still differ greatly in terms of the eWOM content they technologically support. Bronner and de Hoog (2011) have identified three characteristics that classify what readers are exposed to: 1) expressiveness, 2) perceived similarity, and 3) accessibility. Expressiveness refers to the opportunities that
platforms provide for reviewers to express their opinions, such as textual reviews, ratings, sub-ratings, and images. Platforms can also differ in terms of the amount of user information they show, which allows for the assessment of similarity between the reader and the poster. The level of accessibility is determined by how available the contributions are to other users. Selection of Travel Review Platforms
Based on the mentioned typology of travel review platforms, TripAdvisor and Expedia are chosen to embody a consumer-generated and commercial-marketer-generated platform respectively. Since part of this study is a replication, the selection of review platforms was highly influenced by Xiang et al.’s (2017) study. However, the original study included Yelp, which was not included in the current study since this platform is less popular in the
Netherlands. The reasons for selecting TripAdvisor and Expedia were three-fold. Firstly, these platforms are among the most popular RSs (Rabanser & Ricci, 2005), making them comparable in terms of review quantity. Secondly, based on the typology of review platforms by Bronner and de Hoog (2011), TripAdvisor and Expedia are contrasting platforms, which is described in more detail below. Finally, Xiang et al.’s (2017) comparative study supports this selection of review platforms by providing evidence that suggests inherent differences in eWOM content between platforms. Both platforms are discussed in detail below.
An example of a consumer-generated platform: TripAdvisor. TripAdvisor is a consumer-facing online travel community (O’Connor, 2008). The RS is applied as an external site (Rabanser & Ricci, 2005), meaning that they are not attached to any commercial partners. Therefore, TripAdvisor can be categorized as a consumer-generated platform. With regards to the technological characteristics, TripAdvisor scores highly on all three factors. Firstly, users can contribute both quantitative and qualitative content, in many different forms (O’Connor, 2008). User can add a textual review, overall rating and sub-ratings (e.g. sleep, cleanliness, location), when they traveled, who they were travelling with (e.g. alone, as a couple, for
business), and even photos taken during their trip, thus there is high expressiveness. Secondly, users of TripAdvisor can create their own profile, which is visible to others, and determine which information they display on their accounts. Thus, users are in charge of the similarity that other users perceive, but the technological options for perceived similarity are extensive. TripAdvisor also scores high on accessibility, since all reviews are visible to others users.
An example of a commercial-marketer-generated platform: Expedia. In contrast to TripAdvisor, Expedia is an OTA that has devoted only a small section of their website to reviewing (Lee, Denizci Guillet, & Law, 2013). Thus, it is categorized as a
marketer-generated platforms. Users of Expedia can also contribute to both qualitative and quantitative content, however, the options for expressing opinions are limited in comparison to
TripAdvisor. Users can write textual reviews and give an overall rating. Sub-ratings by reviewers are only visible as aggregated sub-ratings. Furthermore, Expedia users can create a personal account, but in contrast to TripAdvisor this account is not visible to other users, and the user information that is displayed with the review is limited to the username and place of residence. Thus, there are limited options for perceived similarity. Lastly, Expedia scores highly on accessibility, similar to TripAdvisor, since all reviews are visible.
Differences in eWOM Content
This study replicates parts of Xiang et al.’s (2017) study by investigating reviews in a different context (i.e. Amsterdam). Therefore, the same analytical framework for assessing the features of reviews is used. The framework by Xiang et al. (2017) focuses on six review-related measurements, including four basic review components (i.e. linguistic features, semantic features, sentiment, and source), as well as variables connected to (i.e. rating), or influenced by (i.e. helpfulness) the reviews. Only three of the characteristics will be used to assess the reviews in this study: 1) linguistic features, 2) semantic features, and 3) sentiment. These features are chosen since they reflect the objective content of the textual reviews.
Linguistic features refer to the characteristics of the textual content, measured here by review length. Semantic features refers to words and topic choices, measured by the review topics. Lastly, sentiment is seen as the valence of a review.
In analyzing the review length across platforms, Xiang et al. (2017) found that the average length of reviews on Expedia was shorter than the average length of reviews on TripAdvisor. Where 87.5% of the reviews on Expedia were shorter than 50 words, this was only true for 50% of reviews on TripAdvisor and further, 61% of Expedia reviews were shorter than 25 words, whereas just 21.9% of reviews on TripAdvisor were this length. Xiang et al. (2017) furthermore identified five major review topics mentioned by reviewers of New York-based hotels: 1) Basic Service, 2) Value, 3) Landmarks and Attractions, 4) Dining and Experience, and 5) Core Product. However, they mention that topics are destination-specific. For TripAdvisor and Expedia, the authors found differences in the presence of the topics. For example, Expedia members mention the topics Value and Landmarks and Attractions more often than TripAdvisor members, while TripAdvisor members more often mention the topics Dining and Experience. For the topics Basic Service and Core Product the same presence was found. However, since the topics might differ for another destination, this study will
independently examine review topics for hotels in Amsterdam and will assume differences in the presence of the topics. Lastly, Xiang et al. (2017) reported a comparable average in sentiment scores for TripAdvisor (M = 0.70) and Expedia (M = 0.66)3, although they did not
statistically test the difference. The averages showed that the reviews are, in general, very positive. Additionally, they reported that the distributions of review sentiments showed identical patterns on the platforms, with strongly left-skewed distributions. Based on these results, the following hypotheses and sub-research question are formulated:
H1: Reviews on Expedia are shorter than reviews on TripAdvisor for Amsterdam-based properties.
RQ1: What are the 5 most characteristic review topics for hotels in Amsterdam mentioned by reviewers on TripAdvisor and Expedia?
H2: Expedia and TripAdvisor differ in presence of the major review topics in the reviews of Amsterdam-based hotels.
Since the statement that the averages of review sentiment are equal among platforms was not statistically tested, and the statistical testing of such a statement is questionable, the following additional sub-research question is formulated:
RQ2: How equal are the distributions of review sentiment on TripAdvisor and Expedia?
Additional support was found in eWOM research regarding review sentiment. Research has shown that positive reviews are more common than negative reviews (e.g. Melián-González, Bulchand-Gidumal, & González López-Valcárcel, 2013; Pantelidis, 2010). A popular
explanation for the abundance of positive reviews is that positive experiences are more likely to make someone favorable to communicate their experience (Melián-González et al., 2013). Therefore, left-skewed distributions are expected for both platforms. In sum, the formulated hypotheses and research questions reflect the assumption that there are structural differences between consumer- and marketer-generated platforms, with the exception of review
sentiment.
Dynamics of eWOM Content
This study also examines the dynamic nature of eWOM content by investigating the influence of major review topics between review platforms over time. Several studies have investigated the topics that are mentioned in online reviews (e.g. Dong, Schaal, O'Mahony, & Smyth, 2013; Xiang et al., 2017; Xianghua, Guo, Yanyan, & Zhiqiang, 2013), however, no
studies thus far have investigated whether the topics have an effect on a larger scale. As Tufekci (2014) mentions, research on social media data should not be confined to a single platform, since there is a flow of information between channels on the Internet. Platforms cannot be studied as closed systems, since they simply are not. However, can the salience of a review topic on one platform be transferred over time to reviews on another platform? For example, when reviewers on one platform mention a certain theme often, such as the delicious Dutch pancakes at breakfast, another previous or future guest might ‘copy’ the salience of this topic in their review on another platform. Thus, there might be a certain degree of interplay between reviews. A theory that deals with the transfer of salience between outlets is agenda setting. As first-level agenda setting theory states, the salience of issues is transferred from one agenda to another (McCombs & Shaw, 1972). Although this theory originates in the field of journalism, and focuses on the transfer of salience between the news media and the public agenda, it may be applicable to the transfer of salience between review outlets. Especially relevant is the intermedia-agenda setting theory, which deals with the transfer of salience and imitation processes between different (news) outlets (McCombs & Shaw, 1993). This transfer of salience is described in a field where the only clue for issue priority comes from other outlets, since the news media are not in contact with their audience members (Vliegenthart & Walgrave, 2008). Although this theory describes the transfer of salience, the processes of salience selection differs with regards to reviewing. When a consumer reviews a hotel, the importance of a topic is, for a large part, determined by internal motivation (Bronner & de Hoog, 2011). However, the communication mentioned in the intermedia-agenda setting theory and eWOM communications show similarities. Both types of communication can be characterized as “one-to-many” communication (Bronner & de Hoog, 2011). This
communication scope, where content is dispersed to a large audience, might ask for more thought, for example, which topics do others write about? Therefore, some of the salience
selection might be determined by social processes. This is especially likely since both types of platform (i.e. consumer- and marketer-generated) are seen by reviewers and readers as
different but complementary channels (Bronner & de Hoog, 2010). Therefore, both types of platforms play an important role in the decision-making process and, although no data is available on this, reviewers likely visit both customer- and marketer-generated platforms.
In terms of the direction of the transfer of salience, it is likely to think that consumer-generated platforms like TripAdvisor influence the topics on marketer-consumer-generated platforms like Expedia. Since TripAdvisor is a platform without direct commercial influence from travel service providers and is generated by consumers, it is more powerful and influential than marketer-generated platforms. This is supported by the comScore Media Metrix (2018) that states that TripAdvisor is the world’s largest travel site, and TripAdvisor’s claims to receive an average of 455 million unique visitors per month (TripAdvisor, 2018). Expedia, in
comparison, is listed below TripAdvisor on the comScore Media Metrix (2018) and claims to receive an average of 131 million unique visitors per month (Expedia, 2018). Based on these number, it is expected that TripAdvisor influences Expedia.
To conclude, no studies have investigated whether the topics that reviewers mention are influenced by the topics previously mentioned by other reviewers. Thus, the following sub-research question is formulated:
RQ3: Do the topics in eWOM communication on TripAdvisor and Expedia influence each other over time?
Methods
This study used the data resources from TripAdvisor and Expedia for a comparative analysis of eWOM content on the consumer- and marketer-generated platforms. Firstly, the reviews from Amsterdam-based hotels were collected from both platforms. The social media analytics procedure was followed (e.g. Fan & Gordon, 2014; Xiang et al., 2017), is based on
three separate steps. Firstly, data was gathered, pre-processed, and quantifiable metrics were extracted from the unstructured data. Then, advanced analytics were performed, and lastly, the findings are evaluated in the next chapter.
Data Collection and Cleaning
The data collection took place in early and mid-2018. The same procedures were carried out for both TripAdvisor and Expedia. Information was extracted from all Amsterdam-based properties using custom-built web scrapers written in the Python
programming language, creating a unique database (see Appendix A for a description of the full datasets). Only relevant variables are discussed here. From each hotel, English-written reviews from the first 100 review-pages were collected, focusing on reviews posted between January 2012 and February 2018, since older reviews will not be relevant for the current media landscape. The process for each platform is described below.
TripAdvisor. For TripAdvisor, the web scraper extracted information from requested URLs. Each scraped review from the English review-pages existed of the textual review and multiple identification marks (e.g. date, username, property name). In total, 146,585 reviews were collected from 402 Amsterdam-based properties. Removal of cases is described in order of exclusion. Firstly, reviews outside the specified time period (N = 13,411) were excluded. Secondly, due to the dynamic structure of the TripAdvisor website, resulting from its RS, a large number of duplicate reviews (N = 44,878) were identified and excluded. Next, the reviews that were not successfully scraped (N = 3,662) within Python due to changes in the website during the scraping process were removed. Lastly, the language detection package (i.e. langdetect) in Python (see https://pypi.org/project/langdetect/) was used to check if all reviews were written in English. A manual check of the non-English identified reviews (N = 24) assured that all reviews could be removed. Exclusion of all mentioned cases lead to a reduction rate of 42%, resulting in a final dataset of 84,610 reviews.
Expedia. The web scraper for Expedia was built with the Selenium package in Python (see https://pypi.org/project/selenium/). This package automates a Firefox WebDriver and simulates a user’s browsing behavior, which makes it possible to scrape an interactive JavaScript platform. Each English review4 consisted of the textual review and multiple
identification marks (e.g. date, username, property name). In total, 2,969,768 reviews were collected from 709 Amsterdam-based properties. Exclusion of cases is, again, described in order of removal. Firstly, reviews without dates (N = 6,451) and reviews outside the specified time period (N = 381,682) were excluded. Secondly, due to the technical workings of
Expedia, a large number of duplicate reviews (N = 2,539,817) were collected and removed from the dataset. Thirdly, reviews that returned as ‘None’ type objects (i.e. missing; N = 800) were also removed. Lastly, the langdetect package identified non-English reviews (N = 769), which were manually checked. Of the non-English identified reviews, 460 reviews were falsely identified as non-English, leaving 309 reviews for removal. After removal of all mentioned cases, 40,709 reviews were included in the final dataset, which is a reduction rate of 98%.
Final Sample
As shown in Table 1, the total dataset consisted of 125,319 English reviews of
Amsterdam-based hotels posted between January 2012 and February 2018, of which 67.5% of the reviews were collected from TripAdvisor and 32.5% from Expedia5. In the final dataset,
TripAdvisor shows a smaller number of non-English reviews (99.99%) than Expedia
(73.07%). On Expedia, reviews from 584 unique hotels were found, while reviews from only 343 unique hotels on TripAdvisor were collected. Due to the dynamic nature of TripAdvisor’s RS, many hotels were displayed and thus scraped multiple times, while some hotels were not
4 Since Expedia does not provide all-English review-pages, the English reviews were selected based on the absence of a ‘translate’ button.
5 This ratio of 2:1 is due to the scraping process, since the same number of review pages were scraped on both platforms, however, the non-English reviews on Expedia were not collected.
displayed at all. Furthermore, Expedia showed a large number of properties that had no reviews (N = 125), while this number was much smaller for TripAdvisor hotels (N = 58). Additionally, TripAdvisor showed an average number of reviews per property that was three times higher than that of Expedia, again reflecting the issue of non-English reviews on Expedia.
Data Pre-Processing
The reviews were pre-processed in three steps, influenced by Xiang et al.’s (2017) study: 1) tokenizing, 2) stop word removal, 3) lowercasing, and 4) stemming. Denny and Spirling (2018) state that pre-processing decisions are of great importance, since “such decisions have profound effects on the results of real models for real data” (p. 168),
supporting the decision to modify pre-processing steps to the current context. Tokenization is the process of splitting text into meaningful units called tokens (Webster & Kit, 1992), which was done using the RegexTokenizer function from the ‘nltk.tokenize’ package (see
http://www.nltk.org/api/nltk.tokenize.html) in Python’s natural language toolkit ‘nltk’. This function uses sets of regular expressions in retrieving meaningful tokens, which can be adjusted according to the research purpose. The pattern was adjusted to only selects words, leaving out any punctuation, special characters, or numbers. Next, stop words were removed from the tokens. Stop words are words without a specific meaning to a text and should be removed for meaningful text processing (Trilling, 2017). An existing stop word list (see
http://www.lextek.com/manuals/onix/stopwords1.html) consisting of 429 English stop words Table 1
Summary of the two datasets
Review Platform N of properties N of reviews Average N of reviews per property
TripAdvisor 343 84,610 247
Expedia 584 40,709 70
was modified to be more relevant for the hospitality domain, removing relevant words from the list (e.g. ‘room’, ‘booking’) and adding contractions (e.g. ‘you’re’, ‘we’d’) and written numbers (e.g. ‘one’, ‘ten’). The final list consisted of 560 words. Next, all tokens were lowercased, and lastly, all words were stemmed using the SnowballStemmer from the ‘nltk’ library (see http://www.nltk.org/howto/stem.html). Stemming is used to extract the stem from a word, so that words such as ‘booking’, ‘book’, and ‘bookings’ return the same. The pre-processing lead to a reduction rate of the text corpus of 39.68% for TripAdvisor, and 40.56% for Expedia.
Development of Key Measurements and Analyses
Review length. The basic linguistic feature review length was calculated by measuring the number of tokens per text (Xiang et al., 2017), and a score was assigned to each review. The Python module ‘statistics’ (see https://docs.python.org/3/library
/statistics.html) was used to calculate the descriptives of review length per platform. To test whether there are differences in review length, an independent t-test was conducted and the effect size (Cohen’s d) was calculated. Although large samples are likely to return significant results, the t-test is conducted for the sake of completeness.
Review topics. Topic modeling, a form of unsupervised machine learning, was used to identify prevalent topics mentioned by consumers in their evaluation of Amsterdam-based hotels on TripAdvisor and Expedia. Following Xiang et al.’s (2017) study, the Latent Dirichlet Allocation (LDA) model was applied to the review corpus, which extracts potentially meaningful latent aspects (topics) from the review corpus by examining the co-occurrence of words (Titov & McDonald, 2008). Following Xiang et al.’s study (2017), the model was specified with five topics6. Four different LDA models were tested to find the
model that generates the most meaningful topics. The first model was estimated with the
default count vectorizer, which uses simple term frequencies (Digital Research Infrastructure for the Arts and Humanities [DARIAH]). The second model was estimated with a tf-idf vectorizer, which weighs the frequency of a term (tf) by the inverse document frequency (idf), adjusting for words that occur very frequently and are less informative (Trilling, 2017). After identifying the most appropriate vectorizer, a third model was estimated without extremes. Extremes are words that occur very frequently or almost never, and thus contribute very little in distinguishing the topics (Puschmann, Scheffler, & Trilling, 2017). Words that occur in less than 5 reviews, or words that occur in more than 50% of all reviews were filtered out. After determining if filtering extremes returns more meaningful topics, a last model was specified with bigrams as added features to the unigrams. Bigrams are formed by joining adjacent words, and thus takes into account word order (Puschmann et al., 2017). After meaningful topics were found, the presence of the topics on TripAdvisor and Expedia was compared using independent t-tests. Again, the t-tests are conducted for the sake of completeness, and Cohen’s d was calculated to indicate the effect size.
Review sentiment. The sentiment, or valence, of a review was measured using the sentiment analysis module Vader (Hutto, & Gilbert, 2014; see
https://github.com/cjhutto/vaderSentiment). The module deals with punctuation, negations, intensifiers (e.g. ‘really’, ‘extremely’), dampeners (e.g. ‘a bit’, ‘kind of’), emoticons, and emojis, making it an appropriate module for social media data. Vader was applied to the plain review texts, and returned a positive, negative, and neutral score. Furthermore, the module also calculated a compound (i.e. aggregated, overall) score, ranging from -1 (very negative) to 1 (very positive). Although a different procedure was used in the study by Xiang et al. (2017), the scales of the overall sentiment scores are comparable. Each review was assigned an overall sentiment score. Again, an independent t-test was performed to compare the group means and Cohen’s d was used to indicate the effect size.
Temporal dynamics of review topics. To investigate the temporal dynamics of review topics between platforms, time series analysis was performed with Vector
Autoregression (VAR) models. VAR models are appropriate when there is no theory-based motivation for the causal relationship between variables, since it considers the
interdependence of variables within the model (Strycharz, Strauss, & Trilling, 2018). More specifically, a VAR model estimates both directions in separate equations, thus testing the effects of TripAdvisor topics on Expedia topics, and vice versa (Vliegenthart, 2014). First, the datasets containing the presence of the five topics for each platform were merged using the pandas module (McKinney, 2010; see https://pandas.pydata.org/). Since data was gathered over an extensive time period (2012-2018), the data was aggregated in weeks. After the final dataset was create, five separate VAR models were specified, one for each topic.
Following previous time series analysis research (Strycharz et al., 2018; Vliegenthart, 2014), a step-wise approach to VAR-modeling was used. First, the stationarity of the time series was tested using the Dickey-Fuller (DF) test. The DF test indicates whether the statistical properties of variables are constant over time (i.e. stationary), with insignificance indicating that all series need to be differenced in order to achieve stationarity. Next, the appropriate number of lags (i.e. weeks) was selected using fit statistics (Akaike’s information criterion, Bayesian information criterion). The lags were theoretically limited to 6 weeks, since holiday seasons are often shorter than a month-and-a-half. Third, after estimation of the models, Granger-causality tests were performed to test whether the topics on one platform influenced the topics on the other platform, regardless of its own influence. Lastly, the impulse response function (IRF) was used to estimate the duration of effects, cumulative IRF (CIRF) was used to summarize the effect over time, and the forecast error variance (FEV) was used to estimate the variation in a topic on a platform explained by the presence of the topic on the other platform.
Results Review Length
The first hypothesis considers the length of reviews written on TripAdvisor and Expedia of Amsterdam-based hotels, by stating that TripAdvisor would have lengthier reviews than Expedia. Table 2 displays the descriptive statistics of review length. The results showed that the average review length, measured in number of tokens, is about twice as large on TripAdvisor (M = 46.15, SD = 36.97) as the average review length on Expedia (M = 22.59, SD = 20.06)7. Thus, reviewers on TripAdvisor use twice the number of words when sharing
their evaluation of Amsterdam-based hotels than those on Expedia. Moreover, it is interesting to note that TripAdvisor’s largest review length (717) is more than three times the largest review length on Expedia (225). Figure 1 shows the distributions of review length among both platforms, with the X-axis indicating the review length scores, and the percentage of reviews on each platform on the Y-axis. Both distributions are heavily right-skewed, but Expedia shows a more narrow distribution around the median than TripAdvisor. On Expedia, 57.1% of the reviews has less than 20 words, while this is true for only 16.8% on
TripAdvisor. Furthermore, 90.56% of Expedia reviews are shorter than 50 words, while this is again much lower on TripAdvisor (68.7%). Although the distributions of review length are not normally distributed, an independent t-test was conducted to compare the review length on TripAdvisor and Expedia. The results show that the review length on TripAdvisor (M =
7 Although Xiang et al. (2017) did not remove outliers, the analysis was repeated without outliers (N TripAdvisor = 80.618, N Expedia = 39.103), using percentiles. The results still indicate that reviews on TripAdvisor (M = 43.06, SD = 26.85) are lengthier than reviews on Expedia (M = 21.06, SD = 16.10) for Amsterdam-based properties. The distributions of review length without outliers are displayed in Appendix B.
Table 2
Descriptive Statistics of the review length (in tokens) on both platforms
Min Max Mean Median SD CI
TripAdvisor 4 717 46.15 35 36.97 [45.90,46.40]
46.15, SD = 36.97) significantly differed from the review length on Expedia (M = 22.59, SD = 20.06); t(125317) = 145.99, p < .001. The effect size indicated that this effect was quite large (Cohen, 1988); d = 0.728. To conclude, reviews on Expedia are overall shorter then reviews on TripAdvisor for properties in Amsterdam, therefore, H1 is accepted.
Review Topics
To answer sub-RQ1, which asks about the five most characteristic topics mentioned by reviewers of Amsterdam-based properties, four LDA models were conducted and the results were compared. Based on logical comparison of the generated topics, the LDA-count vectorizer model with unigrams and bigrams as features and filtered out extremes returned the most meaningful topics8. Table 3 displays the five most common topics identified by the LDA
model in the review corpus, reflected by the ten most relevant tokens and the probability that the token belongs to the topic. The topics were labeled as Value, Basic Service, Location and Landmarks, Core Product, and Dining and Experience. These topics illustrate common
8 For the results of the other LDA models, see Appendix C.
Table 3
Review topics extracted with the LDA-count vectorizer model, with unigrams and bigrams as features and filtered out extremes Topic 1. Value Topic 2. Basic Service Topic 3. Location and
Landmarks
Topic 4. Core Product Topic 5. Dining and Experience Token Probability Token Probability Token Probability Token Probability Token Probability
location* 0.026 stay 0.010 walk 0.025 bed* 0.017 stay 0.022
clean 0.021 night 0.009 minute 0.016 bathroom* 0.013 lovely* 0.018
nice* 0.021 check* 0.009 museum 0.012 shower* 0.012 helpful 0.010
stay 0.020 booking 0.008 station* 0.011 floor 0.011 amsterdam 0.010
friendly* 0.018 time 0.007 tram 0.010 breakfast 0.010 view* 0.008
helpful 0.016 day 0.006 amsterdam 0.009 stair 0.008 amazing 0.008
amsterdam 0.012 arrive 0.006 min 0.009 clean 0.008 location 0.008
close 0.012 door* 0.005 airport 0.009 location 0.006 service* 0.007
city 0.012 pay 0.004 minut_walk 0.008 night 0.006 breakfast 0.007
comfort 0.011 look 0.004 central* 0.007 coffee* 0.006 bar* 0.007
themes that were discussed by guests of Amsterdam-based properties on travel review
platforms. The topic Value describes a theme relating to the evaluation of the value of hotels, with words like ‘comfort’, ‘clean’, and ‘nice’. Basic Service is characterized by words reflecting the standard services or characteristics of the services provided by hotels, such as ‘night’, ‘arrive’, and ‘pay’. Location and Landmarks reflects a theme covering the location of hotels and the landmarks of the destination, indicated by words like ‘station’, ‘museum’, and ‘Amsterdam’. The topic Core Product is characterized by words that describe the product (i.e. the hotel/hotel room), such as ‘room’, ‘shower’, and ‘bed’. The last topic, Dining and
Experience, is reflected by words describing the available dining possibilities and the overall experience, such as ‘bar’, ‘amazing’, and ‘view’. Some of these topics are generic topics for hotel guests, and others are destination specific. For example, the topics Value, Basic Service, and Core Product are relevant themes for all hotel guests, since these are important aspects when evaluating a hotel. In contrast, Location and Landmarks and Dining and Experience are subject to the travel destination. In this particular case, these topics are identified by words describing the landmarks and experiences in Amsterdam. To conclude, the topics Value, Basic Service, Location and Landmarks, Core Product, and Dining and Experience are characteristic topics of reviews of Amsterdam-based hotels.
H2 deals with the presence of the major review topics in reviews of Amsterdam-based hotels on TripAdvisor and Expedia. To test whether the presence of the topics differs for TripAdvisor and Expedia, the second output of the LDA model was used, which returns scores for the presence of each topic in the review texts. Figure 2 shows the presence of the five topics. Noteworthy is the clear difference between the platforms, with a higher presence of all topics on TripAdvisor. An evident reason for the large difference in presence of the topics among TripAdvisor and Expedia is the difference in sample size, however, this
explanation was eliminated when controlling for the number of reviews still showed the same results9. Although the topics for both TripAdvisor and Expedia are not normally distributed,
five separate independent t-tests were conducted. For topic 1 (Value), a significant difference was found between TripAdvisor (M = 16.63, SD = 12.77) and Expedia (M = 10.48, SD = 9.17); t(125317) = 97.37, p < .001. For the second topic (Basic Service), a significant
difference between TripAdvisor (M = 15.98, SD = 26.19) and Expedia (M = 7.34, SD = 13.27) was also found; t(125317) = 77.48, p < .001. Similarly, a significant difference between
9 Since TripAdvisor has lengthier reviews, an additional figure was created controlling for the length of the
reviews (see Appendix D). The figure indicates that when controlling for length, Expedia shows a higher presence of the topics than TripAdvisor. However, it is important to note that the length of the reviews is an important attribute of the eWOM content and should not be controlled. Moreover, this might suppress how informative the TripAdvisor reviews are. Therefore, a second additional figure was created with dichotomized topic scores (0 = no presence, 1 = presence). The threshold was determined based on the distribution, cutting off scores representing very low presence of the topics. The figure shows a smaller difference between the presence of topics on TripAdvisor and Expedia, indicating the a larger number of reviews on TripAdvisor show low presence of the topics. This makes sense since TripAdvisor also has a larger number of reviews.
TripAdvisor (M = 12.31, SD = 17.50) and Expedia (M = 6.48, SD = 10.51) was found for topic 3 (Location and Landmarks); t(125317) = 73.29, p < .001. For topic 4 (Core Product) there was also a significant difference between TripAdvisor (M = 14.30, SD = 19.32) and Expedia (M = 7.23, SD = 11.35); t(125317) = 81.18, p < .001. Finally, for the last topic (Dining and Experience) there was also a significant difference between TripAdvisor (M = 15.87, SD = 18.90) and Expedia (M = 7.29, SD = 10.12); t(125317) = 104.60, p < .001. These results support the observation that TripAdvisor shows a higher presence of the topics than Expedia. However, the effect sizes are relatively small (Cohen, 1988), with statistics for topic 1 (d = 0.525) and topic 5 (d = 0.518) indicating medium effect sizes, and for topic 2 (d = 0.379), topic 3 (d = 0.374), and topic 4 (d = 0.412), small effect sizes. Based on these results, H2 is accepted.
Review Sentiment
To investigate sub-RQ2, which asks how equal the distributions of review sentiment on TripAdvisor and Expedia are, a sentiment analysis was performed. The descriptives of review sentiment are displayed in Table 4, with scores ranging from -1 (very negative) to 1 (very positive), and 0 indicating neutral valence. The distribution of sentiment on both platforms is displayed in Figure 3, with on the X-axis the sentiment scores and on the Y-axis the percentage of reviews for that specific sentiment score. Noteworthy is that reviews from TripAdvisor and Expedia are overall very positive, reflected by the high means. Figure 3
Table 4
Descriptives of review sentiment on both platforms
Min Max Mean Median SD CI
TripAdvisor -0.998 1.000 0.743 0.946 0.48 [0.739,0.746]
Expedia -0.996 0.999 0.640 0.852 0.47 [0.635,0.644]
shows that the distributions of review sentiment are very heavily left-skewed on both
platforms. For TripAdvisor, most sentiment scores fall narrowly around the median, showing a very positive kurtosis. Expedia shows a higher percentage of neutral reviews and more moderately positive reviews, and thus has a lower kurtosis.
Even though review sentiment was not normally distributed, an independent t-test was performed to test whether review sentiment differed on TripAdvisor and Expedia. The results show that TripAdvisor (M = 0.74, SD = 0.48) significantly differed from Expedia (M = 0.64, SD = 0.47); t(125317) = 35.93, p < .001. However, the effect size indicated a small effect (Cohen, 1988); d = 0.215.
To conclude, the review sentiments on TripAdvisor and Expedia show very similar distributions; both platforms have overall very positive reviews. Expedia, however, shows more variance in sentiment scores, therefore accounting for the small difference found in the means of review sentiment for TripAdvisor and Expedia.
Transfer of Review Topics
Sub-RQ3 deals with the influence of topics in reviews of Amsterdam-based hotels on TripAdvisor and Expedia over time, and was answered using VAR-modeling. The DF stationarity tests of the time series were insignificant, and the series were differenced to achieve stationarity. The AIC and BIC model fit statistics indicated that 3 weeks were the most appropriate number of lags. The alternative was 6 weeks, however, 3 weeks theoretically makes more sense based on two reasons: 1) review topics become less relevant and up-to-date to other reviewers over longer periods of time 2) holiday seasons generally fall between a week and a month. For each topic, a VAR model was estimated with the presence of the topic on TripAdvisor and Expedia interchangeably used as dependent and independent variables. The results of the Granger-causality tests, CIRFs, and FEVs for each model are displayed in Table 5, with separate results for each topic, with TripAdvisor as the dependent variable and with Expedia as the dependent variable in the VAR model. The table indicates that for all topics (i.e. Value, Basic Service, Location and Landmarks, Core Product, Dining and
Experience), the presence of the topic on TripAdvisor is significantly Granger-caused by the presence of the topic on Expedia. After determining the direction of the effects, the CIRFs for TripAdvisor summarizes the cumulative effects after three weeks. The plotted impulse
response functions, both impulse and cumulative, are displayed in Appendix E. The IRF plots show after how many weeks the effect of Expedia on TripAdvisor fades, within the selected time period. The CIRF plots show the cumulative effects of Expedia on TripAdvisor, and the CIRF value in Table 5 indicates the increase of the topic presence on one platform after one additional increase in the topic presence on the other platform. The topic Basic Service shows the lowest increase in the presence of the topic on TripAdvisor after one additional increase of the topic presence on Expedia, while this was highest for the topic Core Product. For all topics, most of the effects occurred within the first week, indicating that there are only
Table 5
The findings for all five VAR-models
Dependent variable
VAR model Topic TripAdvisor Topic Expedia
Topic 1 Granger-causality test F(3, 612) = 5.97** F(3, 612) = 0.74
CIRF .287 .004
FEV 1.76% 1.54%
Topic 2 Granger-causality test F(3, 612) = 3.25* F(3, 612) = 0.22
CIRF .018 .006
FEV 1.51% 6.47%
Topic 3 Granger-causality test F(3, 612) = 4.94** F(3, 612) = 1.22
CIRF .254 .025
FEV 1.74% 0.90%
Topic 4 Granger-causality test F(3, 612) = 3.97** F(3, 612) = 0.67
CIRF .343 .015
FEV 1.54% 5.13%
Topic 5 Granger-causality test F(3, 612) = 3.79* F(3, 612) = 0.48
CIRF .186 .012
FEV 1.28% 1.95%
Note. For each dependent variable, the other variable was estimated as the independent variable; Significances for Granger-causality tests: *p < .05. **p < .01. ***p < .001; Cumulative impulse response function (CIRF) and forecast error variance (FEV) after selected number of lags (3 weeks); For the 95% CI of the CIRF, see Appendix E.
short-term effects. Furthermore, the FEVs show the explained variance of the topic presence on one platform by the topic presence on the other platform, within three weeks. The scores indicate that the effects are quite weak, with 1.76% for the Value topic being the highest explained variance, and 1.28% for the topic Dining and Experience being the lowest
explained variance. Thus, only a small variation in the topics on TripAdvisor are explained by the presence of topics on Expedia.
In sum, this section shows the results of a preliminary analysis testing whether the presence of topics on travel review platforms influences each other. To answer sub-RQ2, the results suggest that the presence of major review topics mentioned on Expedia has an impact on the presence of these topics on TripAdvisor, however, these effects seem rather small and are very short-term.
Discussion & Conclusion Summary of Findings
This study examined the content of eWOM communications in the form of reviews on two contrasting types of travel review platforms, namely consumer- and
commercial-marketer-generated platforms, embodied by TripAdvisor and Expedia respectively. Using reviews collected from the platforms, this study examined actual consumer behavior. The purpose of this study was two-fold: firstly, it partially replicated Xiang et al.’s (2017) study by comparatively examining reviews on travel review platforms in a different context (i.e. Amsterdam), and secondly, this study preliminarily investigated the influence of the presence of major review topics between travel review platforms over time. By comparing two types of review platforms, this study shows the importance of the selection of review platforms.
Following Xiang et al.’s (2017) study, the first section of this study investigated the objective characteristics of the reviews on both platforms. In examining the length of the reviews, interestingly but expectedly I found that consumers on TripAdvisor use about twice
as many words as consumers on Expedia. This result is in line with Xiang et al.’s (2017) study and, based on the theoretical framework of platforms types, it was also expected that consumers on consumer-generated platforms are more likely to write longer reviews due to the more ‘other-oriented’ motivations that bring consumers to TripAdvisor.
Furthermore, this study investigated the topics that were characteristics for the evaluation of Amsterdam-based hotels. I found that from the five topics that were identified (i.e. Value, Basic Service, Location and Landmarks, Core Product, Dining and Experience), some topics were generic review topics that are important to reviewers in general, and others were more destination-dependent, which was in line with the characteristics of topics found in reviews of New York-based hotels by Xiang et al. (2017). More importantly, and surprisingly, the major review topics that were found were almost identical to the review topics of New York-based hotels. Some topics were even described using identical words, which was even seen for the destination-specific topics. Therefore, I conclude that there are indeed
destination-specific topics, however, the degree to which these topics are specific to the destination is lower than was originally expected. Even for these destination-specific topics, there are certain aspects that are still considered by all reviewers. For example, although the Location and Landmarks topic is dependent on the destination, it is also described with generic words such as ‘station’ or ‘museum.’ The destination-specific Dining and Experience topic also showed that the ways in which customers of hotels describe their experience and the food options of a hotel is still described with the same aspects, such as ‘view’ and ‘bar’. Given the high comparability of New York and Amsterdam in general, future research should look into whether these generic and somewhat seemingly universal concepts are also
applicable in very different destinations, such as the Global South or non-Western countries. The presence of the major topics on TripAdvisor and Expedia was additionally examined and revealed that users of TripAdvisor mention all topics more than the users of
Expedia. Although a difference in presence of the topics was expected, this result was not completely in line with Xiang et al.’s study (2017), since they found some topics to be more prevalent on Expedia. The authors concluded that the presence of a topic was determined by the platform’s business orientation. The same explanation can be used in clarifying why TripAdvisor reviews were found to be more informative. TripAdvisor attracts more reviewers that are motivated to help others and are looking for social benefits (Bronner & de Hoog, 2011) due to their consumer-oriented business approach. People with such motivations are more likely to exchange more information in their review. For example, Bronner and de Hoog (2011) suggest that consumers directed at helping other travelers like to express their opinion in as many ways as possible, including both qualitative and quantitative ways, while self-directed consumers contribute less content. This seems to explain why TripAdvisor reviews are so widely consulted by travelers and perceived as helpful in their decision-making process (O’Connor, 2008). Therefore, it is concluded here that reviews posted on consumer-generated platforms are more informative than reviews posted on marketer-generated platforms. The consequence of this finding for the theoretical framework of Bronner and de Hoog (2011) is that the types of review platforms should include the type of customer that it attracts, since this is determining for the eWOM content that is contributed. Future research should therefore investigate the process of customers’ platform selection.
Lastly, the sentiment of reviews placed on both platforms was compared. Expectedly, the study found that the distribution of review sentiments were largely equal on both
platforms, which was in line with Xiang et al.’s (2017) study. Only minor differences were found, with slightly more diverse sentiments in reviews on Expedia. Additionally, the reviews on Amsterdam-based properties were overall extremely positive, which was in line with the theoretical framework stating that positive reviews are more common in the travel industry.
To conclude, this study shows that there are fundamental differences in the review length, topics and sentiment of eWOM content on consumer-generated and commercial-marketer generated platforms. By replicating Xiang et al.’s (2017) study for Amsterdam-based properties, I showed that, regardless of the destination, the same patterns exist in
eWOM content on travel review platforms. This study provides better insights in the nature of eWOM content on consumer-generated and commercial-marketer-generated platforms, and shows that differences in business orientation are determinants for the content that is
contributed by consumers. Furthermore, this study shows the importance of review platform selection in research on eWOM communication by underlining the discrepancies that were found on different types of platforms. Researchers should be aware of the differences when selecting platforms, and further research should focus on an even better system of mapping these differences in eWOM content. For example, mixed forms of travel review platforms could be incorporated in the analyses to investigate to what extent platforms should be consumer- or marketer-generated to display the inherent characteristics of the types.
Motivated by the lack of research on the dynamic patterns of eWOM communication mentioned in the literature, this study additionally tested whether major review topics on travel review platforms influence each other over time. The results suggest that there is indeed a dynamic pattern over time, however, the direction was unexpected. The review topics mentioned by consumers on Expedia were found to influence the review topics mentioned by consumers on TripAdvisor, which contradicts the assumption that was made based on the influential and consumer-oriented nature of TripAdvisor. A possible explanation for the direction of the effect is that Expedia is an online travel agent and all customers that have booked through the platform will be asked to write a review for the visited property. Since TripAdvisor is an independent travel community, reviewers have to actively decide to write a review. Therefore, it seems likely that marketer-generated platforms influence
consumer-generated platforms regarding the topics that are mentioned in the reviews. However, no clear support can be attributed to the transfer of salience between reviewers as stated by the
intermedia agenda setting theory, since it was not known whether reviewers from TripAdvisor and Expedia were exposed to reviews on the other platform. Future research should focus on investigating whether the transfer of topics is caused by the exposure to other reviewers on other platforms. For example, tracking data could be used to check what reviewers are exposed to regarding online travel reviews. Additionally, the effects are short-term, which indicates that only very recent reviews influence reviewers in what they are contributing. To conclude, more research is needed to better understand the transfer of topic salience between platforms. For example, the same study could be done in a different context or on more types of platforms. Based on these findings, this study offers implications for tourism research and the tourism industry.
Implications for Tourism Research
This study’s most important implication for tourism research is demonstrated by the supporting evidence that was found for the inherent differences of travel review platforms, especially with regards to consumer- and marketer-generated platforms. By showing that the results of Xiang et al.’s (2017) study holds ups for eWOM in a different context, this study places even more importance on the selection of travel review platforms in research. Reviews vary considerably on platforms and in using the available online data sources, researchers should be aware of the characteristics of the platforms with regards to review length, topics and sentiment. This study furthermore implies that research with multiple data sources is necessary. While single-data studies are also valuable, this study shows that researchers should not overlook the wider interaction and dynamics between platforms.
Implications for the Tourism Industry
The results of this study also offer implications for hoteliers. First of all, this study shows that hoteliers should be aware of the inherent characteristics of the travel review platforms with which they engage. Since the different types of platforms provide different content, hoteliers should pay attention to the characteristics of the content in dealing with reviews. For example, evaluations on consumer-generated platforms are likely to be written in more words and show higher mentions of major review topics than reviews on marketer-generated
platforms, and responses should be tailored to these inherent characteristics. Since eWOM research already indicated that such reviews, which are likely to be perceived as more informative, are more helpful in consumers’ decision-making process, hoteliers should especially handle negative reviews on consumer-generated platforms with care. Negative reviews that are considered to be helpful to others are likely to do more damage than negative reviews that not informative. Additionally, as mentioned by Xiang et al. (2017), different platforms “may represent different consumer segments” (p. 63). Hoteliers should be aware of these segments when addressing consumers in advertising on review platforms, and tailoring to the type of travel review platforms might be a good solution. Finally, by showing the dynamic structure of major review topics between platforms in general, and consumer- and marketer-generated platforms in specific, this study implies that it is crucial for hoteliers to be more aware of the interplay between platforms. Since the agenda of marketer-generated platforms is most influential, hoteliers should be especially careful of negative reviews on these platforms and prevent the transfer of these negative topics to other platforms. Strengths and Limitations
First of all, reviews by real customers written on travel review platforms were used in investigating the dynamics and differences of eWOM content, which is a valuable strength of the study. In using real-world data, the advantage lies that actual customer behavior is studied
and ecological validity is not jeopardized as opposed to the traditional methods of studying eWOM. Using real data can allow for crucial insights in consumer behavior and generalizable results. Another strength of this study is the comparison of platforms. Since studies have persistently used one platform to generalize results on a population, this study has advanced from this practice by investigating multiple platforms and showed the structural biases of the platforms. Additionally, this study also investigated the interaction between platforms, which has not been done until now. Therefore, this study is the first to recognize the wider social ecology, which is crucial since review platforms are not closed, singular systems.
A limitation of this study is that, of the linguistic features, only one aspect was
considered (i.e. review length), thus, no claims can be made on the full linguistic nature of the reviews. Future research should study additional linguistic features in order to better
understand the linguistic differences between consumer-and marketer-generated platforms. This is especially important since I found that the platforms attract different consumer-segments with their business orientation, and these consumer-segments are likely to use different linguistic styles. A second limitation regards the data collection procedure. Since a maximum page was selected for scraping the reviews (i.e. 100 pages), not all existing reviews between the selected time period were collected, especially for hotels with a high review quantity. This might create a bias in the dataset, however, such a bias is less relevant in a very large dataset.
To conclude, this study highlights the differences and dynamics of eWOM communication on review platforms in the travel industry, and encourages researchers to challenge their perspective of travel review platforms, by investigating the complex interplay and nature of eWOM on different types of platforms.
References
Allsop, D. T., Bassett, B. R., & Hoskins, J. A. (2007). Word-of-mouth research: principles and applications. Journal of advertising research, 47(4), 398-411.
doi:10.2501/S0021849907070419
Bore, I., Rutherford, C., Glasgow, S., Taheri, B., & Antony, J. (2017). A systematic literature review on eWOM in the hotel industry: Current trends and suggestions for future research. Hospitality & Society, 7(1), 63-85. doi:10.1386/hosp.7.1.63_1
Bronner, F., & De Hoog, R. (2010). Consumer-generated versus marketer-generated websites in consumer decision making. International Journal of Market Research, 52(2), 231-248. doi:10.2501/S1470785309201193
Bronner, F., & De Hoog, R. (2011). Vacationers and eWOM: Who posts, and why, where, and what? Journal of Travel Research, 50(1), 15-26. doi:10.1177/0047287509355324 Cantallops, A. S., & Salvi, F. (2014). New consumer behavior: A review of research on
eWOM and hotels. International Journal of Hospitality Management, 36, 41-51. doi:10.1016/j.ijhm.2013.08.007
Cheng, X., & Zhou, M. (2010, August). Study on effect of ewom: A literature review and suggestions for future research. In Management and Service Science (MASS), 2010 International Conference on (pp. 1-4). IEEE. doi10.1109/ICMSS.2010.5576663 Cheung, C. M., & Thadani, D. R. (2012). The impact of electronic word-of-mouth
communication: A literature analysis and integrative model. Decision support systems, 54(1), 461-470. doi:10.1016/j.dss.2012.06.008
Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Hillsdale, NJ: Erlbaum.
comScore Media Metrix (2018). Top 50 Multi-Platform Properties (Desktop and Mobile). Retrieved from https://www.comscore.com/Insights/Rankings
Crotts, J. C., Mason, P. R., & Davis, B. (2009). Measuring guest satisfaction and competitive position in the hospitality and tourism industry. Journal of Travel Research, 48(2), 139–51. doi:10.1177/0047287508328795
Denny, M. J., & Spirling, A. (2018). Text preprocessing for unsupervised learning: why it matters, when it misleads, and what to do about it. Political Analysis, 26(2), 168-189. doi:10.1017/pan.2017.44
Digital Research Infrastructure for the Arts and Humanities. Text Analysis with Topic Models for the Humanities and Social Sciences [Jupyter Notebook]. Retrieved 8 June 2018, from https://de.dariah.eu/tatom/working_with_text.html.
Dong, R., Schaal, M., O'Mahony, M. P., & Smyth, B. (2013). Topic Extraction from
Online Reviews for Classification and Recommendation. Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence, 13, 1310-1316. Retrieved from http://aaai.org
Erkan, I., & Evans, C. (2016). The influence of eWOM in social media on consumers’ purchase intentions: An extended approach to information adoption. Computers in Human Behavior, 61, 47–55. doi:10.1016/j.chb.2016.03.003
Expedia (2018). Audience. Retrieved from https://advertising.expedia.com/audience
Fan, W., & Gordon, M. D. (2014). The power of social media analytics. Communications of the ACM, 57(6), 74-81. doi:10.1145/2602574
Gretzel, U., & Yoo, K.H., (2008). Use and Impact of Online Travel Reviews. In: O’Connor, P., Höpken, W., & Gretzel, U. (eds). Information and Communication Technologies in Tourism 2008: Proceedings of the International Conference in Innsbruck, Austria, 35-46. Vienna, Austria: Springer. doi:10.1007/978-3-211-77280-5_4
Hennig-Thurau, T., Gwinner, K. P., Walsh, G., & Gremler, D. D. (2004). Electronic word-of- mouth via consumer-opinion platforms: What motivates consumers to articulate
