key: cord-0707697-bhei7baz
authors: Kostromitina, Maria; Keller, Daniel; Cavusoglu, Muhittin; Beloin, Kyle
title: “His lack of a mask ruined everything.” Restaurant customer satisfaction during the COVID-19 outbreak: An analysis of Yelp review texts and star-ratings
date: 2021-08-10
journal: Int J Hosp Manag
DOI: 10.1016/j.ijhm.2021.103048
sha: d52cbad8062c5fd00c6fa5c9c68cc47dd30c7d20
doc_id: 707697
cord_uid: bhei7baz

The aim of the study was to provide practical advice to restaurant managers for improving star ratings as well as information for researchers on how the pandemic has impacted established determinants of satisfaction. The study examined criteria used by restaurant customers in assigning star-ratings on Yelp during the COVID-19 pandemic using keyword analysis and Multiple Correspondence Analysis. In evaluating restaurants, the reviewers focused on service, overall experience, and food quality. Service was discussed in relation to the pandemic and included safety of the dine-in experience, contrasted with take-out options and compliance with COVID-19 guidelines. These criteria applied differently with lower-star reviews focusing on safety, social distancing, and mask policies. Higher-star reviews focused on take-out/delivery services, high-quality food, and an overall positive experience. The study provides valuable contributions to our understanding of how the COVID-19 pandemic will impact the restaurant sector in a post-pandemic world.

Due to the COVID-19 pandemic, the hospitality industry has suffered daily losses estimated at USD 534 million. The restaurant sector is one of the most severely affected areas (Burrow, 2020; Yang et al., 2020) . A confluence of factors including (a) stay-at-home orders and travel/mobility restrictions, (b) social distancing requirements for public spaces, and (c) the fact that "almost all restaurants [have been] asked to limit their operations to only take-outs" (Gursoy and Chi, 2020, p. 527) , has led to restaurant demand falling since the start of the COVID-19 pandemic (Bartik et al., 2020; Ozili and Arun, 2020) . Accordingly, the National Restaurant Association (2020) estimates that more than 110,000 restaurants (17% of the restaurants in the United States) closed permanently or for the long term and more than 500,000 are in economic distress. Because of the severe impact of COVID-19 on the restaurant industry, businesses have had to adapt to stay in business.

From the perspective of restaurant patrons, uncertainty over the safety of dining out during the COVID-19 pandemic can be mitigated by seeking out and sharing information online about restaurants' sanitation and hygiene protocols. Online review platforms are particularly useful for this purpose as they provide unfiltered, up-to-date information during crises (Kim et al., 2021; Schroeder et al., 2013) . Siering (2020) reported that online review platforms (e.g., Yelp) provide timely and relevant information about restaurants' newly introduced hygiene practices as well as other safety precautions. Online review texts thus offer analysts information on both the types of adjustments that restaurant managers have made in response to the pandemic (the so-called new normal; Brandau, 2020; Breier et al., 2021) , as well as how customers view those adjustments.

Despite this, much is still unknown regarding how specific COVID-19 precautions are perceived by reviewers and how these precautions (or lack thereof) have impacted reviewers' choice of star ratings. suggested that "while using previous conceptual and theoretical frameworks may benefit future research, it is critical to generate new knowledge that can provide insight to the industry about how to transform their operations according to newly emerging customers' needs and wants due to . Given the heightened importance of online reviews during the current outbreak, it is important to understand COVID-19 related criteria that impact review star-rating as well as how these criteria differ across the different star-ratings.

Toward this end, we apply the method developed in Keller and Kostromitina, (2020) to Yelp restaurant reviews that discuss the COVID-19 outbreak (a) to identify criteria that are important to consumers during the pandemic, and (b) to understand how these criteria vary in importance in high, low, and mid-star reviews. The aim of the study was to provide both practical advice to restaurant managers and owners for improving star-ratings as well as information for researchers on how the pandemic has impacted established determinants of satisfaction (e.g., food quality, service, overall experience; see e.g., Keller and Kostromitina, 2020) . Thus, in the current study, we investigated the following questions: RQ 1. : Which COVID-19 safety precautions (or lack thereof) are discussed in the texts of reviews and how do customers respond to them? RQ 2. : Which COVID-19 precautions (or lack thereof) are more likely to be discussed in reviews at different star-ratings?

To answer these questions, we applied the text analysis method described in Keller and Kostromitina, (2020) to sets of keywords from pandemic-focused Yelp reviews, words that are statistically more likely to occur in reviews that refer to the pandemic. This allowed us to identify dimensions of co-occurring keywords and then explore how these dimensions interacted with star-rating. In Section 2, we review recent literature on pandemic-focused safety procedures, as well as Uncertainty Reduction Theory (URT; Berger and Calabrese, 1975) , a theoretical explanation for why consumers produce and consult reviews more often during the pandemic and why it is reasonable to expect reviews to contain relevant information about pandemic-related safety procedures. Finally, we discuss recent text-analytic research on Yelp restaurant reviews. In Sections 3 and 4, we describe our approach and results. In Section 5, we discuss how our findings relate to the current understanding of the new normal. We conclude with a discussion of practical managerial implications as well as implications for theory.

Nation-wide pandemic lockdowns have changed the way business in hospitality is done and will be done in the future with social distancing remaining "a key strategy to manage COVID-19 in many countries for several months" (Go¨ssling et al., 2020, p. 12) . According to Hu et al. (2021) , hospitality remains highly susceptible to health-related crises due to the high volume of patrons, large staff work teams, exposure to travelers, and the potential for contagion through cross-contamination. Restaurants have had to learn a new way of conducting business while remaining safe by implementing new rules and regulations concerning hygiene and social distancing to accommodate customers who are nervous about dining out; these rules and regulations have become the new norm for the industry (Brandau, 2020; Breier et al., 2021) . Specifically, restaurateurs have sought opportunities to adapt to the new normal, which includes offering contactless order and delivery/pick-up options, social distancing measures (e.g., placing tables 6 feet away from each other or installing barriers between tables; limiting the number of dine-in customers), implementing hygiene practices according to Centers for Disease Control and Prevention (CDC) guidelines (e.g., providing hand sanitizers at entryways; sanitizing tables and menus; building ventilation system; avoiding shared items which are reusable), and limiting menu items (CDC, 2020; Lane, 2020; McCarthy, 2020) . According to the new normal perspective on post-pandemic recovery, some of these practices may become standard parts of restaurants' operations (e.g., Kim et al., 2021; Lane, 2020) .

One of the most common adjustments among restaurant businesses have been delivery, take-out or drive-through services to sustain businesses until normal operations resumed (Go¨ssling et al., 2020; Luo and Xu, 2021) . Moreover, those restaurants that were already offering delivery experienced substantial increases in customers' use of this option. With this increase in delivery services, new options for contactless delivery negan to appear. Kim et al. (2021) investigated the role of drone-based food delivery services during the pandemic and argued that drones in food delivery services should be considered a new food-delivery mode. Usage of tamper-proof packaging for delivery or take out and social window shields for drive-throughs have also become standard for the restaurant industry (Hand and Reinstain, 2020) . The spike in the popularity of contactless delivery, take-out, or drive-through is best explained by diners' concerns about food safety during the pandemic. Byrd et al. (2021) found that consuming food at a restaurant during the pandemic has been associated with the highest level of concern among consumers, whereas curbside pickup or drive through food carried less concern. Kim et al. (2021) also suggested that offering delivery services is part of a strategy of "[minimizing] customers' uncertainty toward the restaurant industry" (p. 3), presumably born out of fear of contracting the disease through contact with staff without proper protective equipment.

Another aspect of new normal practices in restaurant operations is associated with hygiene and employee compliance with sanitation norms. Brandau (2020) found that "regularly/visibly wiping down tables, kiosks, other things people touch" and "employees visibly wearing food safety apparel" are the two most effective behaviors for making restaurant consumers comfortable while dining in restaurants (p. 8). Similarly, Cobanoglu and his collaborators (2020) conducted a study on consumer perceptions about mask-wearing at hotels and restaurants. They found that consumers "perceive higher service quality, show greater trust, feel more gratitude, and express reciprocal behaviors in establishments where servers wear masks" (para. 8). More importantly, unmasked staff cause consumers to feel anxious (Cobanoglu et al., 2020) . Additionally, in an examination of facilitatory measures for employee compliance with health and safety regulations and procedures at restaurants, Hu et al. (2021) found that managers played an important role in demonstrating genuine commitment to workplace safety and integrating safety measures into the work routines.

To some extent, pandemic-related safety precautions are driven by consumer demand. Recent research suggests that restaurant customers have been an important driver of the aforementioned innovations related to COVID-19. In particular, McCarthy (2020) emphasized that customers often observe restaurants' hygiene practices or efforts to provide a safe dining experience, and that these practices impact customers' perceptions of cleanliness. Supporting this claim, Hand and Reinstain (2020) mentioned that since sanitization has become a priority among restaurant consumers, restaurant patrons are likely to develop a set of criteria with a greater emphasis on "the ease of ordering, pickup and delivery options, food safety, and sanitization" when choosing a restaurant (Hand and Reinstain, 2020, para 11) . Accordingly, recent survey research has focused on determining the factors or topics that restaurant patrons consider in evaluating dining experiences. found that the current top three customer expectations from restaurants are providing hand sanitizers at entryways, employees wearing masks and gloves, and placing tables six feet away from each other.

Pandemic-driven shifts in customers' attitudes toward the aspects of a dining experience that determine satisfaction may best be understood in light of uncertainty reduction. Given the possibility of transmission of infection at restaurants, there is inherent uncertainty around the safety of dining out during the pandemic. Consequently, consumers assess dining out to be a high-risk activity (Byrd et al., 2021) and seek information to reduce their sense of risk (Sontag, 2020; Tidwell and Walther, 2002) . Uncertainty Reduction Theory (URT) provides fundamental background for understanding consumer reaction to uncertainties (Berger and Calabrese, 1975) . According to URT, people seek information to reduce uncertainty and resulting risk (Berger and Calabrese, 1975) . Consumers refrain from dining out and seek other options due to their concerns over safety and fear of infection. Also, the level of uncertainty impacts consumers' decisions (Lacey et al., 2009) . Supporting this perspective, Foroudi et al. (2021) asserted "the high level of social uncertainty caused by the COVID-19 outbreak leads customers to a higher risk judgment and to develop a high level of negative emotion" (p. 8). This assertion is supported by Jia (2021) who found that since the beginning of the pandemic, consumers have visited fewer restaurants, assigned lower star ratings in online reviews, and spent less money.

As URT predicts, online reviews were an important source of information for customers during the pandemic. Research suggests that due to uncertainty over the safety of dining out and the possibility of transmission of infection at restaurants, consumers assessed the risk of dining out by seeking information from online review platforms (Byrd et al., 2021; Sontag, 2020) . Previous research has found that online restaurant reviews, specifically on Yelp, affect the popularity and success of restaurants (Anderson and Magruder, 2012; Nakayama and Wan, 2018) . However, the pandemic highlighted the importance of reviews for restaurant customers. The restaurant industry ranks first in terms of consumer use of reviews and according to the Local Consumer Review Survey, 31% of consumers read more reviews during the pandemic than they did before (Brightlocal, 2020) . Almost a quarter of consumers avoided businesses due to a perceived lack of safety measures and 17% wrote negative reviews for the same reason. Similarly, 20% of customers expected responses from management within one day of writing the review and 96% of consumers indicated that they read managerial responses to negative reviews (Brightlocal, 2020) . This information supports Mehta et al. (2020) claim that "in crisis times, new trends in consumer behaviour emerge" (p. 293). During the pandemic, consumer behavior was shaped by attitudes toward and perception of the risk of dining out (Mehta et al., 2020) .

While reviews provide information to consumers to reduce uncertainty, they also provide restaurateurs with information on how managerial decisions meant to accommodate patrons' concerns over cleanliness are perceived by customers and thus serve as a reflection of customer satisfaction with an establishment. As Hao et al. (2017) note, the combination of star ratings and the text of customer reviews in online platforms forms a complete description of customer satisfaction. Therefore, online review platforms such as Yelp and Google Reviews offer a text-analytic alternative to survey data for analysts interested in understanding the post-pandemic new-normal. Approaching the investigation of customer criteria from this perspective, Luo and Xu (2021) conducted a keyword analysis and sentiment analysis of Yelp reviews published during the pandemic. Consistent with the studies of restaurant reviews prior COVID-19 (e.g., Keller and Kostromitina, 2020; Gan et al., 2017) , the four main topics that influenced customers' evaluation were service, food, place, and experience. However, these topics were marked by keywords related to the restaurants adjusting to the pandemic (e.g., "delivery", "online ordering", "UberEats", "outdoor seating", "hygiene practices"). The subsequent analysis of sentiment scores of customer attitudes conducted in the study suggested that when dining in, customers cared about whether the restaurant staff had good personal hygiene practices. Moreover, unavailability of outdoor seating as well as dissatisfaction with delivery services (e.g., food arriving cold) negatively impacted patrons' emotional sentiments about a restaurant (Luo and Xu, 2021) . More recently, Cao et al. (2021) analyzed a corpus of 3.1 million Yelp reviews using several machine learning techniques (e.g., sentiment analysis, topic modeling) and found that more than 10% of reviews contained COVID-related keywords and that the prevalence of COVID-related topics (e.g., hygiene, social distancing) increased over the course of the pandemic, while the prevalence of other topics decreased. To the best of our knowledge, these are the only studies that have taken text-analytic approaches to understanding the effect of the pandemic on review texts.

In the current study, we seek to understand how the actions of nonchain restaurant owners and managers taken in response to the COVID-19 pandemic have impacted review star ratings in order to provide practical recommendations for restaurant managers. Previous research has established a systematic relationship between review texts and starratings, namely that reviewers use the text of their review to justify their choice of star rating (see e.g., Fan and Khademi, 2014; Huang et al., 2014; Keller and Kostromitina, 2020; Kong et al., 2016; Ladhari et al., 2008; Linshi, 2014; Yu et al., 2017) . Consequently, statistical regularities in review texts can provide insight into the factors that influence reviewers' choice of star rating. In the current study, however, we follow Keller and Kostromitina, (2020) in their use of Multiple Correspondence Analysis (MCA). MCA was chosen for use in that study because it provides a flexible, open-ended method of dimension reduction that is uniquely suited for linguistic analysis of texts (cf. Section 3.3). We depart from the method employed in that study slightly, however, in that instead of performing the MCA on all words in the corpus, we first identify the keywords for reviews that mention the COVID-19 pandemic using the text dispersion keyness approach described in Egbert and Biber (2019) and then locate the barycenters for reviews of each star rating in the low-dimensional space identified by the MCA. The precise steps in this analysis and the purpose of each is summarized in Table 1 .

As the method of keyword extraction that we adopted requires two corpora (c.f., Section 3.2) -a reference corpus for establishing baseline statistics regarding word usage, and a focal corpus for identifying keywords, two corpora were used in this study: a COVID-19 corpus (the focal corpus) spanning the period between 4/1/2020 and 9/9/2020 and a corpus of pre-COVID-19 reviews (the reference) spanning the same period in the previous year (4/1/2019 to 9/9/2019). The period covered by the COVID-19 corpus was chosen to coincide with the initial surge of COVID-19 cases in many urban areas of North America and the ensuing "lockdown" orders from various state and municipal governments that limited restaurant-goers' ability to patronize restaurants, as well as the eventual loosening of restrictions that began in the summer.

All review texts for the pre-COVID-19 corpus were taken from the Yelp Academic dataset (YAD: Yelp, 2020), a publicly available dataset containing more than eight million reviews of more than 200,000 establishments in ten metropolitan areas. All reviews in the 'restaurant' category were selected from the dataset, but then all reviews of chain restaurants and food-serving businesses other than restaurants (e.g.,

Steps taken in the analysis of review star rating and review texts in COVID-19 focused reviews.

Step Draw conclusions about the relationship between restaurants' handling of COVID-19 precautions and review star ratings hotels, bowling alleys) were filtered out following the procedure described in Keller and Kostromitina, (2020) . To maximize the generalizability of our findings, we stratified the corpus according to two variables (price point and star-rating), discovered the number of texts in the smallest stratum, and then sampled a number of reviews randomly from the remaining strata equal to the size of the smallest stratum. This created a corpus balanced across both star-ratings and price points. As described at length in Keller and Kostromitina, (2020) , previous research on review texts and star-ratings has been performed on non-random samples of unstratified corpora. As a result, the findings from these studies reflect more variation within the most common star-ratings and price points (i.e., 5-star reviews of $$ restaurants) than variation across star ratings and price points. By stratifying the corpus and sampling randomly and equally from each stratum, however, we ensure that the findings will be accurate for all star-ratings and price points.

As the most recent version of the YAD does not contain any reviews published in 2020, it was necessary to scrape review texts for the COVID-19 corpus directly from Yelp.com. To do this, we extracted the business identification numbers for the restaurants in the YAC dataset and used them to scrape reviews of the same restaurants in the COVID-19 period (a custom scraper was written in the Python programming language). Thus, reviews in the two corpora are roughly paired in that they are all from the same set of restaurants, though precise 1:1 pairing was not possible due to the random sampling procedure described above. We also chose to use the full set of scraped reviews in the COVID-19 corpus rather than apply the random sampling procedure. This was because balancing the COVID-19 corpus across price points or star-ratings may obfuscate the impact of the pandemic on restaurant-goers' dining choices. They may, for example, choose to eat more or less frequently at high priced restaurants, or be more or less inclined to give a high starrating. Balancing the corpus across strata may thus hide or disguise these effects. Finally, in order to focus our analysis on only those reviews which demonstrate some awareness of the COVID-19 pandemic, we selected reviews containing one of the following terms: covid, corona, pandemic, or epidemic. The remaining reviews were removed from the corpus. Descriptive statistics for these corpora appear in Table 2 .

Linguists interested in describing the semantic or topical content of a corpus often turn to keyword analysis. Informally, keywords for a corpus are those words with particular importance for the domain of language the texts in the corpus are drawn from (Stubbs, 2010) . Though there is disagreement over how the construct of keyness is best understood and what different keyness metrics actually measure (see e.g., Gabrielatos, 2018; Pojanapunya and Todd, 2018 for reviews), here, we adopt Egbert and Biber's definition of keywords as words that are both distinct for the corpus (i.e., they appear consistently in texts in the corpus, but only infrequently in texts outside the corpus: Egbert and Biber, 2019, p. 78) and generalizable to the corpus as a whole (i.e., they reflect the content of texts throughout the corpus rather than the content of a few influential texts: Egbert and Biber, 2019, p. 79) . In Egbert and Biber's method, keywords are identified by comparing the number of texts a word appears in (its range) in a corpus of interest (the focal corpus) with that word's range in a second corpus (the reference corpus). Words with significantly greater ranges in the focal corpus are keywords. In the current study, we used the pre-COVID-19 corpus as our reference corpus and the COVID-19 corpus as the focal. This allowed us to identify words that are distinctly associated with reviews discussing the pandemic and which were used consistently across review texts. To extract these keywords, we wrote custom Python scripts to calculate the log likelihood ratio of each word's range in the focal corpus to its range in the reference corpus (G 2 ; Dunning, 1993; Egbert and Biber, 2019, p. 94; Rayson et al., 2004; Scott, 2001) . Following Rayson et al. (2004) , keywords with log likelihood ratios greater than G 2 = 3.84 and with greater ranges in the focal corpus were taken as being statistically significantly more frequent in the focal corpus than the reference at the α = 0.05 level.

Keywords may be thought of as contributing dimensions of variation to the dataset. Each text varies along these dimensions by either containing the keywords or not. Because keywords co-occur, however, some dimensions provide information about multiple keywords. Consequently, we used Multiple Correspondence Analysis to identify the dimensions which provided information about the most regularly cooccurring sets of keywords as well as clusters of keywords within each dimension. Multiple Correspondence Analysis (MCA: see Greenacre and Blasius (2006) for a detailed description, or Abdi and Valentin (2007) for a short overview) is a statistical procedure for dimension reduction. It is thus similar to other methods of dimension reduction such as Factor Analysis and Principal Component Analysis, but is designed to work with categorical data (here 'present' or 'absent'). Each dimension identified by the MCA has both positive and negative poles where keywords that occur at the positive pole are in complementary distribution with those at the negative end. These complementary distributions allow MCA to account for the absence as well presence of words, revealing sets of keywords that co-occur with the absence of complementary sets.

To perform the MCA, we wrote custom Python scripts to convert the text of each review into a binary matrix where each column represented one keyword and each row represented one text. The value in each cell was either 1 or 0 where 1 indicated that the keyword (the column value) occurred in the text (the row value) at least one time and 0 indicated that the keyword did not occur in the text at all. Finally, we filtered out all keywords that appeared in fewer than 5% of texts. This reduced the number of keywords from 935 to 194. This was done because performing MCA on the full set of keywords would have been impossible given the limits of computational power available to us. It also allowed for a cleaner interpretation of the MCA dimensions as the least dispersed (and therefore least generalizable) keywords were removed from the analysis. Finally, the MCA was performed using the MCA function in the FactoMineR package for the R programming language.

In the current study, we consulted a scree plot to determine how many dimensions to interpret. The scree plot indicated that we should interpret the first five dimensions. As shown in Table 3 , these five dimensions together explained 11.1% of variance in the dataset. We stopped interpreting dimensions when the additional variance they explained dropped below 1% and the scree plot (Fig. 1 ) indicated a flattening of eigenvalues. Fig. 1 displays a common situation with MCA with language datathe first dimension explains substantially more variance than the other dimensions (see e.g., Clarke and Grieve, 2017; Keller and Kostromitina, 2020) . In previous research, the first dimension has been interpreted as encoding an opposition between long and short texts as the positive pole is associated with a wide variety of words with little semantic overlap, while the negative pole is associated with the absence of all words, but especially common function words like the, to, and of. Since it is impossible to produce a coherent English text of any length without including function words such as these and since the length of a text is often orthogonal to the content of the text, the first dimension is often not interpreted. As expected, in the current study, dimension 1 scores were highly correlated with text length (r = 0.91), so following previous research, we discarded dimension one and began interpretation with dimension two. Fig. 2 displays the scree plot for dimensions 2-10.

To validate the relationship between the MCA dimensions and star ratings, we fitted a regression model with review star rating as the response variable and the first five dimensions of the MCA as predictors. This allowed us to test whether each of the five dimensions predicted star rating in some capacity (e.g., that reviews with higher or lower dimension scores systematically receive higher or lower star-ratings). In previous work, two of this study's authors used ordinal regression for this purpose. However, a key finding of Keller and Kostromitina, (2020) , was that star-rating was not necessarily ordinal. Reviews in the 1-to 3-star range focused on qualitatively different criteria from reviews in the 3-to 5-star range. Consequently, we used multinomial regression in this study as it does not assume an ordered relationship between the levels of the outcome variable. This was done using the multinom function in the nnet package for the R programming language.

As noted above, when applied to language data, the dimensions extracted by the MCA are often interpretable as explaining variation in topic or linguistic functions across the texts in the corpus (see e.g., Grieve, 2017, 2019; Keller and Kostromitina, 2020) . Each dimension's keywords define a semantic space for that dimension, or more commonly, opposition between two semantic spaces. To interpret the dimensions in the current study, we examined the keywords that contributed to each pole of dimensions 2-5 noting overlap in meaning of keywords, as well as clusters of co-occurring keywords. Based on these observations, we formed hypotheses about the semantic, functional, and topical distinctions encoded by each dimension. Then we tested and refined these hypotheses by examining review texts with extreme scores (high and low) on one dimension, but neutral scores on the others. This allowed us to discover systematic differences in review texts.

Finally, we located the barycenters for each star rating in the lowdimensional space defined by dimensions 2-5 (the point with the smallest average distance along the four dimensions to each individual review of that star rating). We then interpreted differences in locations of the star rating barycenters in reference to the underlying dimensions across which they varied. This allowed us to determine ways in which reviews varied systematically within single dimensions, as well as locations in low-dimensional space in which star-rating barycenters clustered together.

As explained above, the main assumption of MCA conducted in this study was that the text of a review serves as a justification for the star rating given by the review author. Thus, we assumed that the retained dimensions were able to predict star rating. The results of the multinomial regression supported this prediction demonstrating that Dimensions 2-5 significantly predicted star ratings, as shown in Table 4 .

To interpret the extracted dimensions, we created keyword coordinate plots with keywords for each dimension (see Fig. 1 for an example and the rest of the plots in Appendix A-C). Then, to refine our understanding of the keyword plots, we examined reviews with high and low scores on one dimension and neutral scores on the other dimensions. We determined that Dimension 4 primarily encapsulated stylistic variation. It did not include specific criteria that the reviewers referred to in their evaluation of restaurants; moreover, we observed substantial overlap of keywords of Dimension 4 with Dimensions 2, 3, and 5 meaning that the semantic content encoded in this dimension was recoverable from the other dimensions. In other words, the dimension was not contributing additional criteria used by the reviewers in evaluating the restaurants. Therefore, we decided to discard this dimension from subsequent analysis and focus instead on the other three dimensions. In the following subsections, we interpret Dimensions 2, 3, and 5 in detail.

Dimension 2 was characterized by the keywords in Table 5 on the positive pole.

The keywords in Table 6 contributed to the formation of negative pole of Dimension 2:

The keywords that loaded on the positive end of the dimension clearly indicate the reviewers' focus on the precautionary safety measures enforced in the restaurants. In particular, the first two clusters of co-occurring keywords suggest that the reviewers discussed the practice of social distancing and wearing masks at the establishments. The subsequent clusters also contribute to the formation of the dimension, although less strongly. The clusters include words such as precautions, patio, safe, bar, seating, inside, outside that imply that safety evaluations were particularly strongly linked with reviewers' dine-in experience. Looking at how the keywords co-occurred, if a restaurant had a bar, for example, the reviewers focused on the safety of being seated there. Contrastingly, the negative end of the dimension indicates a review's emphasis on take-out and delivery experience, as demonstrated by the first two clusters on this pole. The reviews on this end of the dimension appeared to evaluate the process of ordering food, the ease of delivery or pick-up, and the quality of the order itself, as suggested by words like home co-occurring with love, chicken, family. Additionally, reviews on this end of the dimension did not include the co-occurrence of words related to COVID safety measures for in-restaurant dining (masks, wearing, distancing, social, tables).

Review texts 1 and 2 below further support this interpretation of Dimension 2. Texts 1 and 2 have the highest coordinates on Dimension 2 thus representing the positive pole while Text 2 has the lowest coordinates representing the negative pole. Text 2. : I'm from Hawaii and we was craving Thai food. Thanks to Yelp reviews, we tried this place. Currently it's this COVID 19 but their delivery service was simple in ordering through yelp and in time. Infact, this is the best service ever. As for the meal, it was even better than the service!! We are here for a visit but will definitely eat here again before we leave Vegas!!! We ordered the Pad Thai, yellow curry, basil fried rice and fried egg rolls. You will enjoy every single bite!! Love this place!!!. (Dim.2 ¡0.24 ; Dim.3 − 0.01; Dim.5 − 0.03).

The two texts demonstrate the dine in vs. delivery and take-out opposition. Specifically, when reviewers described their dine-in experience (Texts 1 and 2), they paid particular attention to how well a restaurant was following the COVID-19 guidelines for restaurants and bars. Consequently, if violations were detected (e.g., someone was not wearing a mask), the review acquired a negative tone, even though other aspects, like food or service, were positively evaluated. On the other hand, when reviewing delivery and take-out services, the reviewers did not comment so much on the safety measures but rather the overall experience of ordering, receiving the meal as well as the quality of delivered food.

The positive end of Dimension 3 included the keywords in Table 7 : The negative end of Dimension 3 included the keywords given in Table 8 :

The positive pole of Dimension 3 is characterized by clusters of cooccurring positive evaluative adjectives and adverbs as well as positively charged verbs. This is particularly noticeable in clusters 1,2, and 5 that contain strongly loading keywords that seem to describe primarily food quality (tasty, sweet, perfect, enjoyed, super) . Furthermore, verbs like recommend and try in clusters 3 and 6 suggest that the reviews included recommendations about visiting a restaurant for the readers. Although the aforementioned clusters reflected the reviewer's main focus on food, they also contained keywords that represented factors other than food (e.g., seating, patio) implying that the reviewers also evaluated their *p < 0.05. **p < 0.01. ***p < 0.001.

Dimension 2 -positive pole keywords by cluster. Note: Bold font indicates keywords that loaded on the dimension with considerable contribution. Underlining indicates keywords that loaded on the dimension with small contribution. Cursive indicates words that loaded on other dimensions but contribute to this dimension considerably. _0 indicates the absence of a word. The order of the clusters reflects their coordinates on the dimension starting with more extreme coordinates at the beginning of the list. Note: See table note for Table 5 .

overall experience at an establishment. The orientation of reviews on an overall positive evaluation of experience is reinforced by cluster 7 that includes positively colored adjectives (friendly, excellent, amazing), indefinite pronouns (everything), and keywords indicating the intention of future patronage (visit, next). These keywords also contributed to the formation of this dimension, although less strongly. In contrast, the negative pole of Dimension 3 indicates reviewers' orientation on the failure of a restaurant to comply with COVID guidelines. Clusters 1 and 3 include keywords that strongly define the dimension and reveal the reviewers' focus on the business and its customers following the practices of wearing masks and distancing. Moreover, the semi-modal verb need co-occurring with people, distancing and masks in Cluster 3 suggests that the reviewers felt the necessity for the businesses to follow the safety regulations. Some of the keywords that loaded on the dimension less strongly, but still contributed to its formation, include second person pronouns (your, you) and present participial (progressive) action verb forms (taking, going). These words indicate that the reviewers adapted an oral interactional style discussing the restaurant's performance in following the safety measures. That is, the reviewers seemed to be addressing a restaurant's managers and staff in expressing their concerns and criticizing their behavior. Finally, the co-occurrence of words explicitly referring to the evaluation of reviewers' experience (review, am, going) implies that the reviewers used the arguments about (not) following the COVID guidelines to justify the star rating they gave to a restaurant.

The example reviews with the extreme positive (Text 3) and negative (Text 4) scores on this dimension are presented below.

Text 3. : We made reservations for brunch and so happy we did. Between here and the Peppermill in Reno I dont know which is better, but this is definitely top 2! The servings were the cutest! Food was delicious and the serving of the fried chicken and waffle was adorable. I also got the bottomless mimosa/bloody mary, they were just ok but reasonable for the price. They did an amazing job of COVID compliance as far as assisting patrons with choosing of plates so that the public did not touch others food. This was a great experience and perfect way to start our day. (Dim.2 − 0.007; Dim.3 0.30; Dim.5 0.05). The two texts clearly present the contrast between a positive overall experience at a restaurant and a negative experience caused by a restaurant's inability to follow the safety guidelines. In particular, following the COVID guidelines in combination with other factors such as good food quality and presentation results in positive evaluation. On the other hand, failure to comply with COVID guidelines, specifically with regards to workers not wearing masks or wearing them incorrectly, results in a negative evaluation of the establishment. Importantly, Dimension 3 had the strongest relationship with star-rating in the multinomial regression model (see Table 4 ).

Dimension 5 was characterized by the keywords in Table 9 that loaded on the positive pole:

The keywords in Table 10 loaded on the negative end of Dimension 5:

The positive pole of Dimension 5 demonstrates a pronounced focus on supporting local restaurants. It is defined most strongly by Cluster 1 and COVID-related keywords (covid-19, pandemic, during) suggesting the reviewers' intention to support local businesses that are having a hard time during the pandemic. Keywords like distancing and social in Cluster 2 as well as wearing, masks, and mask imply that the intent of supporting local businesses is linked to their fulfillment of the anti-COVID precautionary measures. Finally, several positive evaluative adjectives and verbs in Cluster 4 (enjoyed, delicious) and Cluster 5 (fresh, excellent) emphasize that reviewers' overall experience at the restaurant was still good despite the pandemic.

The negative pole of Dimension 5 is related to the worsening of the customers' experience at a restaurant due to COVID. The keywords in Cluster 2 that defined the dimension most strongly include because and covid indicating that the reviewers connect this worsening in with the pandemic. Moreover, adverbs of place and time (before, here) and first person pronouns combined with auxiliary verbs (I'm, I've) imply that reviewers might have had better experience with a business before the pandemic. The dimension is also defined by words go, there, and back in Cluster 5 suggesting that the worsened experience due to COVID may affect the people's intention of repeat visits.

This interpretation is also supported by the review texts that received the highest (Text 6) and lowest (Text 7) scores on this dimension. amazing, everything, excellent, spot, visit, next, enough, menu, friendly 8 lunch, which, got, outside, our, decided, dining 9 due, new, were, well, we, us, great, also, first, some, really, had 10 is_0 4 , covid, with, so, are_0, pandemic_0, you_0

Note: See table note on Table 5 . were_0, up, order, you, when, are, covid_0, their, had_0, with_0, our_0, great_0 9 delicious_0, is, definitely_0, chicken_0, try_0

Note: See table note on Table 5 . In Text 6, the reviewer offers a positive evaluation of their take-out experience focusing on the cleanliness and safety of the restaurant, the pleasant service, and the quality of food. The end part of the review reinforces the call to support local businesses. Text 7, however, presents a negative evaluation with the author admitting that the decline in the overall experience might have been related to the pandemic.

After interpreting Dimensions 2, 3, and 5 that emerged as a result of MCA, we located the barycenters for each star rating in a series of twodimensional spaces each including two out of three dimensions. In each plot, we interpreted the distance between the star rating barycenters with regard to our interpretations of dimensions. To better understand the location of the barycenters, we divided the plots into quadrants that were defined by the poles of the dimensions on the plane. In what follows, we discuss each plot in detail. Fig. 3 represents a factorial plane map of Dimensions 2 and 3. Quadrant 1 (top right) represents reviews that focused on the experience, especially safety of dining-in, and good food. None of the star ratings are located in this quadrant. Moving counterclockwise, Quadrant 2 includes reviews that discussed good food and the availability and quality of take-out. Quadrant 3 contains reviews that focused on restaurants not following safety measures and the availability and quality of take-out. Quadrant 4 includes reviews that addressed the safety of dine-in experience and service, especially restaurants' non-compliance with COVID safety guidelines.

As shown in Fig. 3 , reviews with 1, 2, and 3 stars are located in the fourth quadrant that represents bad service, specifically the restaurants' non-compliance with COVID safety guidelines paired with the overall evaluation of dine-in experience, particularly its safety. This suggests that the authors of these reviews are mostly concerned with restaurants' poor service shown in their failure to provide a safe environment for dining. It is not surprising, therefore, that 1-star reviews are situated closer to the negative end of Dimension 3. As the star rating increases, the barycenters of 2-and 3-star reviews move closer to the middle of the dimension, although still located in Quadrant 4. Contrastingly, 4-and 5star reviews find themselves in the second quadrant that represents food quality and the availability of take-out and delivery. Since service, and specifically compliance with COVID safety guidelines, doesn't play as big of a role in take-out and delivery, the reviewers are able to focus more on the quality of food. Taken together, these results suggest that while service and food are both important criteria, they have different weight when it comes to the differences in star ratings. Fig. 4 represents the factorial plane map of Dimensions 3 and 5. Quadrant 1 consists of reviews that focused on the good overall experience and included a call to support local restaurants. Quadrant 2 represents reviews that focused on supporting local businesses and bad service reflected in the non-compliance with COVID guidelines. It is no surprise that no star rating barycenters were found in this quadrant. Note: See table note on Table 5 . Quadrant 3 includes reviews that discussed COVID-related bad service and worsening of the overall experience at an establishment. Quadrant 4 represents reviews that would have focused on good overall experience and worsening of experience due to COVID, an unlikely combination. No star rating barycenters were located in this quadrant. Fig. 4 shows that the lower star rating barycenters are situated in the third quadrant that represents bad service and worsened experience caused by COVID. The barycenters align against Dimension 3 with 1-star reviews located closer to the negative end of the dimension. Interestingly, looking at the distribution of the 1-, 2-, and 3-star ratings along Dimension 5, the 2-star barycenter is closest to the negative pole of this dimension. When examining the texts of reviews with the extreme negative scores on this dimension, we noticed that reviewers often indicated previous patronage at the restaurant being reviewed. It is therefore possible that because of previous positive experience at a restaurant, reviewers were more inclined to give the restaurant 2 stars instead of the lowest rating, even though their experience at the establishment was worse. In comparison, the barycenters of reviews with higher star ratings (4 and 5 stars) are located in the first quadrant that represents good experience at a restaurant and reviewers' predisposition to support local businesses. The coordinates of the 5-star barycenter are especially high along Dimension 5 suggesting that reviewers are inclined to encourage others to support local restaurants only if they follow the safety guidelines and provide a consistently good overall experience.

Fig. 5 shows the factorial plane map of Dimensions 2 and 5. Quadrant 1 represents reviews that focused on the overall dine-in experience and its safety as well as on supporting local restaurants. Quadrant 2 includes reviews that discussed the availability and quality of take-out and delivery options and encouraged readers to support local restaurants. Quadrant 3 includes reviews that discussed take-out and delivery quality and COVID-related worsening of experience. Quadrant 4 includes reviews that addressed worsening of the dine-in experience and the safety of dining-in.

As seen in Fig. 5, 1 -, 2-, and 3-star barycenters are situated in the fourth quadrant suggesting that reviews with lower star ratings primarily discussed the decline in the overall experience at a restaurant potentially connecting it to COVID and restaurants' performance in ensuring the safety of their customers. In parallel with the plot in Fig. 4 , the barycenter of 2-star reviews is also found to be closest to the negative end of Dimension 5, likely for the same reasons explained above. The coordinates of the three barycenters do not vary along Dimension 2. Reviews with higher star ratings are located in the second quadrant that represents supporting local businesses through ordering food for takeout or delivery. Since reviewers evaluate take-out or delivery services at a restaurant, they do not focus as much on dine-in safety, or safety overall, possibly due to take-out presenting less of a threat to customers' safety as opposed to dining-in. Consequently, the location of the 5-star barycenter suggests that if reviewers are satisfied with the quality of restaurants' take-out or delivery services, they are more inclined to recommend supporting that restaurant.

The present study investigated patterns of co-occurring keywords that characterize COVID-19-aware Yelp reviews and examined the way these patterns relate to the star-rating given to restaurants by review authors. Specifically, two questions were addressed in the study: the first question focused on the customers' response to COVID-19 precautions (or lack thereof) in review texts overall, while the second question identified precautions that were more likely to be discussed in reviews at different star ratings. Regarding the first research question, our results revealed that in evaluating restaurants during the pandemic, the reviewers focused on specific topics that motivated their star-rating assignment: service, overall experience, and food quality. Similar topics were identified by Keller and Kostromitina, (2020) in Yelp reviews before COVID-19. Previous research of customer evaluations of restaurants also highlighted the significance of food quality and service (e.g., Bae et al., 2018; Gan et al., 2017; Josiam and Henry, 2014; Luo and Xu, 2021) . These overlapping findings suggest that the topics identified in this analysis represent universal aspects of dining in a restaurant that are likely to be evaluated by the consumers regardless of the external conditions (like the pandemic). In contrast to previous research, however, service was discussed mostly in relation to the pandemic and included safety of dine-in experiences, contrasted with the availability and efficiency of take-out, and compliance with COVID-19 guidelines. Additionally, we identified topics related to the effect of the pandemic on repeat patronage: desire to support local businesses versus excusing poor experiences due to the pandemic. The identified topics reflect customers' concerns about dining at a restaurant during a pandemic that were found in recent research (e.g., Brandau, 2020; Sontag, 2020) .

With regards to the second research question, the present study demonstrated that, in line with the previous findings in Keller and Kostromitina, (2020) , the identified topics were distributed differently across star-ratings. The distribution of reviews on the factor map for Dimensions 2 and 3 suggests that the criterion of effective or ineffective COVID-19 procedures serves as an important consideration in the decision-making process for assigning the star ratings. Specifically, bad service was associated with a restaurant's failure to provide a safe environment for a dine-in experience demonstrated through the non-compliance with the COVID-19 safety guidelines. These results are supported by studies conducted before and during the pandemic that identified food safety and restaurant hygiene as diners' major concerns (Sharma and Radhakrishna, 2015; Byrd et al., 2021) . Moreover, recent research has suggested that the lack of a safe dining environment at restaurants causes negative emotions towards restaurants, which could be reflected in the lower star-rating assigned to an establishment (e.g., Foroudi et al., 2021; Luo and Xu, 2021) . While the centers for the lower star ratings do not differ in their position along the dine-in safety positive pole of Dimension 2, lack of social distancing and strict and consistent mask rules differentiate 1, 2, and 3-star reviews along the negative end of Dimension 3 with the lowest star rating being awarded to the restaurants that offer unsafe dine-in services. The positive pole of Dimension 2 and the negative pole of Dimension 3 are characterized in particular by the shared variables related to wearing, social, masks) . These dimensions complement the findings of customer perception studies (e.g., Cobanoglu et al., 2020; Brandau, 2020; Wang et al., 2021) as well as text-analysis studies (Luo and Xu, 2021) that showed that compliance with COVID safety guidelines (e.g., staff visibly wearing masks) as well as good personal hygiene practices of the staff increases customers' trust towards an establishment, comfort about dining in, and perceived service quality.

In contrast, the good service that characterizes 4 and 5-star reviews primarily encompasses efficient and straightforward take-out and delivery services located on the negative pole of Dimension 2. Reviewers' attention to these services is not surprising as previous research has detected the increase in delivery and take-out during the pandemic as an alternative to dine-in (e.g., Go¨ssling et al., 2020; Yang et al., 2020) . The variables contributing to this dimension include words directly related to this type of service (delivery, takeout), words related to the process or ordering and receiving food (ordering, order, home, pick) , and absence of words related to social, distancing, masks) . Thus, since safety presents less of a concern in the case of take-out and delivery (Byrd et al., 2021) , the criterion that further differentiates restaurants with higher star ratings is the overall experience formed on the positive pole of Dimension 3. In discussing their overall experience, the reviewers paid attention to the food and its quality (sweet, tasty, fried, delicious, fresh) combined with positive evaluative adjectives (amazing, everything, excellent, perfect) , words indicating intention of future visit (recommend, visit, next) , and overall experience-related words (friendly, patio, seating).

Intent of future patronage is represented by Dimension 5. The variables that meaningfully contribute to this dimension include adverbs of place and time (before, there, here, back) and words explaining the decline in quality (because, covid). That is, failure to provide a safe dining environment resulted in the worsening of the overall experience and, consequently, a lower star rating. The diminishment of customer satisfaction with a restaurant's failure to provide a safe environment has been suggested by previous research of the effects of COVID-19 on the hospitality industry (e.g., Luo and Xu, 2021) . Interestingly, however, our findings indicate that despite a negative experience, the reviewers, due to previous relationships with an establishment, avoid giving it the lowest rating and instead tend to assign 2 stars. On the other hand, the factor maps of Dimensions 2 and 5 as well as 3 and 5 show that 4 and 5-star reviews were situated in the quadrants that represented the criteria of take-out/delivery, good overall experience, and encouragement of supporting local businesses. With the keywords support, local, restaurants, during, covid-19, pandemic representing the positive end of Dimension 5, as well as words like enjoyed, delicious, tasty, excellent that loaded on Dimension 3 and delivery, pick, ordering, and takeout that were shared with Dimension 2, our findings indicate that if the criteria of good and safe service, particularly with regards to take-out and delivery, and good overall experience were met, the reviewers were inclined to recommend a restaurant for future patronage and encourage supporting it during the pandemic.

In aggregate, our findings suggest that reviewers who chose to dine in a restaurant were likely to leave reviews in the 1-3-star range. The primary concern of these customers was the safety of their experience related to the compliance with COVID safety guidelines. If these guidelines were not followed, in particular, if the staff and other customers were not wearing masks and if there was no proper social distancing, reviewers were likely to leave a 1-star review. However, this star rating was mediated by a previously established positive experience with a business, which resulted in customers being more lenient and leaving 2-star reviews. Contrastingly, reviewers were prone to write a 5star review for a restaurant that offered efficient and safe COVIDcompliant take-out or delivery services together with high quality food and pleasant overall experience. If these criteria were met, the reviewers were likely to recommend others support the local businesses.

The present study offers significant theoretical and methodological contributions for the domain of hospitality management. First, the study presents further empirical support of Keller and Kostromitina, (2020) findings, as well as findings from previous studies of customer perceptions (e.g., Bae et al., 2018; Gan et al., 2017) , that service quality, food quality, and overall experience are meaningful predictors of restaurant customer satisfaction and that their importance differs across star-ratings. That is, our findings suggest that the evaluation of a restaurant and assignment of a star-rating may represent two-step process with good or bad service being the first step of differentiating between 1 and 3-star reviews vs. 4-5-star reviews followed by the application of specific criteria (COVID precautions for 1-3-star reviews and overall experience for 4-5-star reviews). Additionally, the study expands the previous findings from Keller and Kostromitina, (2020) by identifying additional criteria of customer satisfaction pertaining to the context of the COVID-19 pandemic, that included the availability of take-out or delivery and safety of service, which, in turn, can motivate future patronage. Importantly, however, the criteria identified in this study and the relationships between them may not solely apply to restaurants in times of a pandemic; rather, it is possible that they present universal criteria that, with further research, may inform hospitality management practices (see Section 5.2).

Second, the current study presents a methodological refinement of Keller and Kostromitina, (2020) in that it uses keyword analysis to get a more precise profile of the topics in the COVID-19 corpus reviews and, consequently, more defined dimensions in MCA. Furthermore, taking into consideration the relationship between the levels of the star ratings noted in Authors-a&b, we used multinomial regression in validating the relationship between the dimensions and star ratings since this regression type does not assume an ordinal relationship in the outcome variable.

The study offers implications for the managerial staff and business owners that are interested in increasing the star rating of their restaurants on Yelp. In agreement with Keller and Kostromitina, (2020), we also found service to be a "threshold criterion" differentiating between lower and higher star ratings. However, since the present study focuses primarily on restaurant reviews during the pandemic, the service criterion appears to address the way managers enforce the COVID safety precautions for their staff and customers. If the guidelines are not followed (e.g., staff and customers not wearing masks, no proper social distancing), especially if a restaurant offers dine-in services, the reviewers are likely to assign 1-, 2-, or 3-star ratings irrespective of the overall experience and previously established patronage. However, several important factors are at play here as key differentiators among these lower star ratings. First, the decline in customers' experience due to COVID moderates the star rating with reviewers assigning 1 star to restaurants that, on top of disregarding the safety guidelines, also underperformed in other areas. At the same time, however, if a reviewer had a previous positive experience at an establishment, they are more likely to give 2 stars in their reviews instead of the lowest rating. When service and the implemented precautions meet the expectations, the reviewers focus on evaluating the quality of food and overall experience at a restaurant. Another factor that seems to yield positive reviews with high star-rating is the availability of take-out and/or delivery services with a simple and efficient ordering process.

Based on the findings in this study, we suggest that depending on the current star rating of a restaurant, managers should choose to focus on improving certain aspects of customer experience. First, we suggest that managers may reduce the amount of 1-, 2-, and 3-star reviews by improving restaurants' service by establishing and maintaining consistent COVID-19 precautionary measures. At a minimum, these measures need to include a mask-wearing policy and proper social distancing for both customers and the staff. If the service threshold is achieved, maintaining the quality of food and overall experience during the pandemic is what differentiates restaurants with higher star-ratings. Thus, if precautionary measures are already in place, managers need to shift their focus on the quality of food to improve the star rating of the restaurant. Second, we argue that offering quick and simple take-out and delivery is a more effective way to get a 5-star rating than offering a safe dine-in option. The results of the current and previous studies have indicated that during the pandemic, customers are more likely to leave a positive review about a delivery or take-out service rather than a dine-in experience. Accordingly, we suggest that restaurant managers implement such services to improve the star-rating of their business, provided that take-out or delivery is paired with high-quality food and is efficient.

Additionally, review management is an important aspect of restaurant management in the pandemic era. As Brightlocal (2020) noted, a sizable majority of review consumers read management responses to negative reviews. This suggests that restaurant managers need to (1) be proactive in implementing the suggested changes in their business models and (2) monitor and promptly respond to the negative reviews on platforms like Yelp. Finally, we stress that while diners are more supportive of their local favorites in the pandemic, they are not unwilling to overlook COVID safety issues, which may result in lower star-rating. Therefore, restaurant managers should take the implementation of safety precautions seriously to maintain the achieved star-rating.

There are several limitations in this study that must be recognized. First, although we took a principled approach to corpus construction for the analyses and employed web scraping techniques following the suggestions in Keller and Kostromitina, (2020) , this method did not come without constraints. Specifically, we were only able to scrape reviews that appeared on the first page for each restaurant. The reviews were organized by recency; thus, the most recent reviews at the time the corpus collection took place were included. While this method could limit the generalizability of the findings, it was the most effective approach at the time since Yelp had not published the new dataset. Future studies may refine the web scraping techniques used to access the review texts to include all the reviews from the beginning of the pandemic.

Furthermore, our study did not take into account take-out and delivery options while sampling -that is, we did not stratify our sample according to this variable. Notwithstanding this limitation, our results suggest that offering such services may lead to a higher star rating and that this effect may extend to the post-COVID context. Future research may look into the differences in reviews for restaurants that only offered take-out or delivery and the criteria applied by the reviewers in evaluating these services. Such research could be beneficial especially since "ghost kitchens," or restaurants that only work with take-out or delivery, present a business format that is gaining popularity in the restaurant industry. Future studies may further investigate the benefits of delivery and contactless payments for restaurant businesses, if these options are maintained after the end of the pandemic. Finally, it should be noted that the data analysis described in this manuscript was conducted in December 2020, when the pandemic was in full swing. While businesses start to open back up across the country and COVID-19-related mandates are being lifted, the findings in the present study provide valuable insights into the dynamic nature of reviews on online platforms like Yelp as well as their impact on the star ratings of restaurants as they adjust to the new normal.

The aim of the present study was to examine the impact of restaurant customers' views on safety during the pandemic, expressed in their Yelp reviews, on their choice of star-rating. In particular, through analyzing the keywords in the reviews, we investigated the customers' response to COVID-19 precautions (or lack thereof) along with the importance of certain precautions at different star ratings. Based on the conducted analyses, the study provided restaurant managers and owners with recommendations about improving the star-rating of their restaurant in the times of COVID-19. These recommendations were based on an analysis of the texts of pandemic-related restaurant reviews with different star-ratings using corpus linguistics methods. Our results showed that restaurant customers pay particular attention to restaurants' implementation of COVID safety guidelines when dining in. Noncompliance with these guidelines may lead to a restaurant's low star rating. Additionally, our results suggested that offering safe and efficient take-out and delivery services may increase a restaurant's star rating. Finally, the findings indicate that despite previous positive experiences at an establishment, customers are reluctant to patronize a restaurant if their overall experience is worsened due to COVID-19.

Multiple correspondence analysis. Encycl

Learning from the crowd: Regression discontinuity estimates of the effects of an online review database

The effects of restaurant attributes on satisfaction and return patronage intentions: Evidence from solo diners' experiences in the United States

How are small businesses adjusting to COVID-19? Early evidence from a survey

Some explorations in initial interaction and beyond: toward a developmental theory of interpersonal communication

Coronavirus & The impact on eating

The role of business model innovation in the hospitality industry during the COVID-19 crisis

Local consumer review survey 2020

The economic impact of COVID-19 on hospitality

Restaurants and COVID-19: what are consumers' risk perceptions about restaurant food and its packaging during the pandemic?

Quantifying the effects of COVID-19 on restaurant reviews

Considerations for restaurant and bar operators

Dimensions of abusive language on Twitter

Stylistic variation on the Donald Trump Twitter account: a linguistic analysis of tweets posted between

Do masks make a difference in customer perception of service quality in hotels and restaurants? HospitalityNet

Accurate methods for the statistics of surprise and coincidence

Incorporating text dispersion into keyword analyses

Predicting a business star in yelp from its reviews text alone

The gloom of the COVID-19 shock in the hospitality industry: a study of consumer risk perception and adaptive belief in the dark cloud of a pandemic

Keyness analysis

A text mining and multidimensional sentiment analysis of online restaurant reviews

Pandemics, tourism and global change: a rapid assessment of COVID-19

Multiple Correspondence Analysis and Related Methods. CRC

Tracking the effects of COVID 19 pandemic on restaurant and hotel customers' sentiments: Customers' sentiments in November towards dining out, traveling to a destination and staying at hotels. Hospitality and Tourism Management Academy

Effects of COVID-19 pandemic on hospitality industry: Review of the current situations and a research agenda

COVID-19 for restaurants: What's working and what will become part of the new normal. FSR Magazine

A tale of two countries: International comparison of online doctor reviews between China and the United States

Creating a safe haven during the crisis: How organizations can achieve deep compliance with COVID-19 safety measures in the hospitality industry

Improving restaurants by extracting subtopics from yelp reviews

Analyzing restaurant customers' evolution of dining patterns and satisfaction during COVID-19 for sustainable business insights

Eatertainment: utilitarian and hedonic motivations for patronizing fun experience restaurants

Characterizing non-chain restaurants' Yelp starratings: Generalizable findings from a representative sample of Yelp reviews

Uncertainty risks and strategic reaction of restaurant firms amid COVID-19: evidence from China

Predicting international restaurant success with yelp

The role of perceived risk in wine purchase decisions in restaurants

Determinants of dining satisfaction and postdining behavioral intentions

How restaurant brands can continue to grow in the 'new normal

Personalizing Yelp star ratings: A semantic topic modeling approach

Comparative study of deep learning models for analyzing online restaurant reviews in the era of the COVID-19 pandemic

The new normal for daily restaurant operations

The new consumer behaviour paradigm amid COVID-19: permanent or transient?

Is culture of origin associated with more expressions? An analysis of Yelp reviews on Japanese restaurants

Spillover of COVID-19: Impact on the global economy

Log-likelihood and odds ratio: keyness statistics for different purposes of keyword analysis

Extending the Cochran rule for the comparison of word frequencies between corpora

Using social media in times of crisis

Mapping key words to problem and solution

Consumer knowledge and perceptions towards food safety practices: implications for consumer education programs

Leveraging online review platforms to support public policy: predicting restaurant health violations based on online reviews

Imperfect COVID-era takeout? One star

Three concepts of keywords

Computer-mediated communication effects on disclosure, impressions, and interpersonal evaluations: Getting to know one another a bit at a time

The effects of crowdedness and safety measures on restaurant patronage choices and perceptions in the COVID-19 pandemic

COVID-19 and restaurant demand: early effects of the pandemic and stay-at-home orders

Yelp dataset documentation

Identifying restaurant features via sentiment analysis on yelp reviews