Language Differences in Online Complaint Responses between Generative Artificial Intelligence and Hotel Managers

: Since November 2022, the use of generative artificial intelligence (GAI) technology has increased in many customer service industries. However, little is known about AI’s language choices and meaning-making resources compared to human responses from a systematic linguistic point of view. The present study is a discourse analysis that explores negative online guest complaints made to four luxury heritage hotels in Hong Kong that are classified as cultural heritage sites with rich interpersonal and historical values. We collected authentic guest complaints and responses from hotel managers from April 2012 to October 2022 in online travel forums, and then had GAI draft response letters on behalf of the hotel managers. Our total dataset was 65,539 words and consisted of three subcorpora: guest complaints (Text a of 115 complaints totaling 26,224 words), hotel manager responses (Text b of 115 response letters totaling 14,975 words), and AI-generated responses (Text c of 115 response letters totaling 24,340 words). This study used systemic functional linguistics to explore interpersonal meanings in texts; for example, appraisal resources, verb processes, and personal pronouns were compared between texts. First, we identified the most frequent words of the common themes across the three subcorpora and found significant differences in lexicogrammatical features between hotel managers and AI-generated responses using the log-likelihood ratio. The results suggest that AI-generated texts are able to provide a tailored and empathetic response to guests, but hotel managers may need to introduce some modifications, such as time indicators, sensory verbs used, and complimentary offers. This study explores the differences in word choices and communication strategies, which have implications and insights for the hospitality industry, especially luxury heritage hotels where caring and personalized customer service are considered important.


Introduction
The trend of reusing heritage buildings has received much attention in recent years.This involves the conversion of historic buildings into tourist accommodations, known as heritage hotels [1].For instance, Pongsermpol and Upala [2] have addressed the impacts of adaptive reuse in small hotels in Bangkok, and Yusran et al. [3] discuss the preservation of heritage buildings in resort-hotel areas in Indonesia, focusing on the importance of preserving heritage buildings.Heritage hotels are unique in that they often have historical or cultural significance, which means that service providers are expected to have more personal and warm interactions with guests [4][5][6].Heritage hotels not only provide adequate hospitality services, but also preserve heritage sites and promote cultural narratives.Unlike commercial hotels, which prioritize efficiency and standardization, heritage hotels require greater interpersonal communication skills and establish a stronger sense of humanity and cultural engagement.When responding to online complaints, hotel management may make cultural and historical references, such as "we are proud of our heritage facilities and aim to provide the highest standards of service" to maintain the hotel's special identity and value.
Due to the rapid development of social media, numerous hotel guests leave reviews on public online travel forums after a stay at a hotel [7][8][9].Some of these comments are positive, but others express dissatisfaction, which can lead to a negative reputation for the hotel [10,11].After locating negative comments on the online platform, hotel managers use their contextual and hospitality knowledge to respond to guest complaints, convey empathy, recover from service failures, and rebuild goodwill [12,13].Previous studies have shown that language that can demonstrate understanding, and construct care and empathy is the core solution for handling online complaints [14][15][16].The rise of generative artificial intelligence (GAI) may have a significant impact on hospitality management, particularly in addressing guest complaints.GAI is a computer language program that mimics human intelligence by integrating complex software and hardware components with large data sets [17].GAI programs use pre-programmed templates or existing algorithms to generate text [18][19][20], enabling them to produce response texts within seconds.However, little is known about the effectiveness of GAI systems in responding to authentic and complex complaints in the hospitality context, and how these responses differ from human responses.For example, while GAI provides responses, it also raises concerns about the lack of personal touch and empathy in the constructed text.Therefore, further research is needed to fill this knowledge gap by comparing GAI responses with human responses.This study is motivated by the need to explore the potential role of GAI in the hospitality industry, balancing efficiency with the rich interpersonal and cultural connections in luxury heritage hotels.It also examines the differences in interpersonal meanings embedded in GAI and hotel managers' responses, and explores how these differences affect the emotional tone and empathy delivered to dissatisfied guests.This study contributes to the application of the GAI technology in the hospitality sector.Understanding the strengths and limitations of GAI in customer service can help hotel managers optimize its use, and GAI developers can better understand the linguistic aspects needed to improve the responsiveness of GAI systems.
This study aims to analyze and compare the language strategies employed by human hotel managers and generative artificial intelligence (GAI) in responding to negative online guest complaints.By exploring these differences, the present study hopes to provide insights for hotel management to effectively integrate the strengths of both human and AI responses to enhance customer service experiences in luxury heritage hotels.The present study has two research questions (RQ): RQ1.What are the common themes mentioned in the negative reviews, hotel managers, and AI-generated responses for Hong Kong's luxury heritage hotels?RQ2.How do human managers and generative artificial intelligence differ in their use of language features such as verbs, personal pronouns, and meaning-making resources in apologies to guest complaints?

Literature Review 2.1. Demand for GAI Application in the Hospitality Industry
Following the release of GAI technology in 2022, potential language features associated with the integration of GAI technologies have become a topic of interest among academic researchers.For example, Cabanac et al. [21] discovered the presence of tortured phrases and writing styles produced by text generators in scientific journal articles, which could threaten the integrity of academic accuracy.Jakesch et al. [22] investigated how participants distinguished between human and AI-generated text based on the presence of first-person pronouns, contractions, and choice of family topic.They discovered some interesting findings, such as that human judgments were more predictable and manipulable, while AIgenerated text was interpreted as more human than human.Research has also explored the impact of GAI technologies on the hospitality sector.For example, Limna and Kraiwanit [23] interviewed 15 hotel professionals in a Thai hotel and found that ChatGPT significantly improved language skills, provided recommendations, and improved productivity and workflow.Furthermore, Rasheed et al. [24] investigated the factors influencing the adoption of GAI services in the Pakistani hospitality industry.For Pakistani customers, cultural values positively influence adoption, while technological complexity, security concerns, and ethical issues are significant barriers.Table 1 summarizes the updated language studies that discuss the implications of GAI text for the hospitality industry.Zhu et al. [17] examined the acceptance and use of GAI gadgets in the tourism and hospitality industry.Service providers believe that GAI gadgets can increase efficiency and safety, reduce costs, improve service quality, facilitate sustainable workforce, and increase employee satisfaction.Subsequent studies have shown similar positive results for the effectiveness of the GAI system ChatGPT.For example, Jo and Park [28] and Wang [30] both point out that ChatGPT consistently remembers user input, enhancing personalized interactions and accurate information retrieval efficiency to improve service quality and guest satisfaction.Christensen et al. [29] found that consumers prefer a GAI itinerary, even if it contains errors, over other sources such as the government tourism website or social media influencers.Possible reasons for choosing the ChatGPT itinerary trusted it because it was more impartial and personalized, with GAI being easier to use.Gajić et al. [31] discovered that GAI's interactive features of providing tailored responses significantly enhanced customer engagement.The study emphasizes the significance of robust AI system development in building customer perceptions and interactions with brands.
The human-like features of ChatGPT contribute to user satisfaction, as interactions with human-like AI create a sense of familiarity, comfort, and emotional recognition [28,30].Emotions such as safety, gratitude, intimacy, and curiosity can influence user click-through and booking rates on social media [27].In particular, Cheong and Law [26] and Remountakis et al. [27] discovered that empathy realized in ChatGPT texts is a key dimension that can create an emotional connection between users and the GAI system, leading to increased satisfaction and bookings.However, there are contrasting viewpoints on AI's limitations in providing emotional and personalized communication.There have been some studies on the limitations of ChatGPT's capabilities in emotion recognition [32,33].Koco ń et al. [32] examined ChatGPT's performance on 25 diverse Natural language processing tasks for sentiment analysis and emotion recognition.They found that the ChatGPT may not excel in complex and difficult tasks.Wake et al. [33] used a variety of datasets and labels to evaluate the emotion recognition performance of ChatGPT.They discovered that ChatGPT generally has a reasonable level of reproducibility; however, the use of different labels and datasets can also lead to instability and bias.Thus, they highlighted the importance of dataset and label selection for the better integration of emotion analysis in ChatGPT applications.Gender seems to have no significant effect on word-of-mouth communication about ChatGPT [28].However, several studies (see Table 1) have examined the relationship between user age and satisfaction with ChatGPT applications with slightly different results.For example, Jo and Park [28] and Wang et al. [25] found that older travelers have different preferences for GAI interactions.Jo and Park [28] found that older people place a higher value on interpersonal relationships and share more experiences than their younger counterparts.Older travelers may be reluctant to use ChatGPT due to a lack of familiarity with the technology or concerns about the authenticity of AI-generated interactions.However, Wang et al. [25] found that older travelers were more satisfied with robotic AI applications and apologies, possibly due to their neutral and unbiased response and a desire for clarity of communication.These studies suggest that older travelers' preferences may vary depending on specific contexts, such as the different nature of complaints or experiences.Thus, the present study highlights that, while GAI can improve operational efficiency, luxury heritage hotels need to balance AI with personalized human interactions to maintain their unique interpersonal and cultural values in different contexts and motivations.
Finally, concerns have been raised regarding the use of GAI technology, which remains controversial.According to Zhu et al. [17] and Remountakis et al. [27], some customers continue to show strong resistance to GAI devices due to reduced social connectedness, perceived inhumanity, privacy concerns, security, and ethics issues.Wang et al. [25] discovered that human apologies increase revisit intentions, whereas robot apologies do not.When faced with unexpected technical crises, GAI may exhibit biases and limitations [30].Furthermore, the widespread use of GAI may clash with local hospitality cultures that value personalized services and warm interactions [30].According to these studies, hospitality and tourism managers should train employees on the importance of human adaptability and emotional intelligence and prepare them to serve travelers effectively [25,30].Some of these training areas include the language and interpersonal aspects suggested in the present study.

Theoretical Framework
Systemic functional linguistics (SFL) is a linguistic theory that focuses on the relationship between language use and its social context [34].Language is viewed as a primary social semiotic system used to express meanings, reflect experiences, exchange shared values, negotiate relationships, and construct social sense between speakers and writers [34][35][36][37].The concepts of the context of situation and culture were first introduced by Polish-British anthropologist Bronisław Malinowski in the 1920s and later further developed by distinguished linguist J. R. Firth and his student M. A. K. Halliday, who integrated situation as a kind of context into a general theory of language to explain meaning [38][39][40][41][42]. Martin [43] discusses different text-forming resources in English and practical procedures for analyzing English texts in their contexts.Martin [44] reflects on the evolution of SFL theory, especially in relation to genre, discourse semantics, and appraisal systems.These studies on SFL emphasize the importance of understanding language in relation to its social context, including discourse semantics, genre, and interpersonal communication.
A growing number of researchers have used a discourse approach to analyze the language used in tourist texts [45,46], online reviews [47], and responses in the hospitality industry [12,13].Understanding customer satisfaction and dissatisfaction through language reviews is important for hoteliers to improve service quality and meet customer needs [48][49][50][51].The relationship between online reviews and hotel service demand has been explored through SFL theory.For example, Wan [12] examines the communication strategies used to deal with negative electronic word-of-mouth (eWOM) in the luxury hotel industry in Hong Kong.The study examines external conjunctions and resources in the appraisal system [37,52] and focuses on service standards in complaints and how hotel management responds.Wan and Forey [13] explored electronic word-of-mouth via travelers' online reviews and the responses from hotel managers.They used engagement resources to look at heteroglossic voice and typical generic stages of management responses to online complaints posted online, as follows: "Acknowledgement ˆApology ˆ(Investigation) ˆ(Resolution) ˆFollow-up ˆGratitude" (p.367).Over the past two years, GAI has attracted significant attention in marketing and consumer research, with applications in the hospitality industry for text analysis and data comparison [53].In particular, ChatGPT has revolutionized content creation and language understanding in the hotel industry [30].This study extends these studies by investigating negative eWOM in luxury heritage hotels and responses created by a GAI system, establishing a new area of research in customer service development for the hospitality and tourism industry.Heritage hotels are not merely profit-driven, but also preserve the historical and cultural values of the city; thus, their operation requires management to focus on interpersonal factors and offer sufficient human warmth to guests [5,6].This study investigates negative guest complaints and the corresponding responses from four luxury heritage hotels in Hong Kong.The current research design and methods are discussed in the next section.

Research Design and Methodology
To answer the research questions, the data were analyzed textually using a mixed approach.We used the log-likelihood ratio to investigate the quantitative differences between different language choices and subcorpora.We also focused on appraisal resources and lexicogrammatical features as part of a systemic functional linguistic approach to interpret apology strategies and meaning making in texts between GAI and human managers.The sample consisted of four heritage hotels in Hong Kong.Visitor reviews for these four hotels were manually collected from travel forums.Our dataset consisted of three subcorpora: (a) guest complaints, (b) hotel manager responses, and (c) AI-generated responses from ChatGPT version 4.0.We first selected four luxury heritage hotels with 4.5-and 5-star rankings, each with a distinct historical background and cultural significance in Hong Kong, and then conducted a collection of authentic English language complaints and replies from international visitors and hoteliers on different online travel forums.
This study adopted a systematic approach for selecting data for discourse analysis.Criteria for selecting complaints and responses included the nature and seriousness of the complaints, the level of detail provided, and the overall attitude and language used in communication.Our aim was to capture a wide range of issues such as accommodation service quality, food and beverages, cleanliness, and staff performance to reflect the diverse experiences of guests in luxury heritage hotels.Before selection, each complaint was assessed for content, length, and attitude expressed by the dissatisfied guests.We also ensured that the selected texts included both complaint letters and management responses, allowing for further comparison with GAI response letters.The study also examined the heritage context of Hong Kong hotels, collecting complaints about outdated facilities and decoration patterns.As a result, the dataset included a broad and representative sample of online guest complaints and responses in heritage hotel contexts.
For data collection, we set this cut-off date to November 2022 to ensure that the complaints and responses included in our study were written by humans before the widespread adoption of GAI technology in customer service contexts.We focused on responses generated by hotel managers and those authentically created in response to guest complaints rather than responses that may have been generated by AI systems in more recent reviews.This decision was important for our analysis, as it allowed us to focus on interactions that reflected authentic linguistic differences between human and AI responses.This decision may reduce the timeframe for data collection and result in less data being collected; however, it improves the accuracy and relevance of our findings and is consistent with the study's objectives and research questions.
In addition, it should be noted that most of the reviews left for Hong Kong luxury heritage hotels are positive, indicating a comfortable environment and above-average service quality.Usually, only negative reviews attract the attention of hotel managers and induce them to write response letters; therefore, they were selected for the present study.The data collection process involved systematic collection and analysis of complaints and responses to ensure a representative sample, as shown in Figure 1.In this study, pre-processing steps were carried out to ensure data quality and consistency prior to discourse analysis.Text cleaning was performed to remove irrelevant elements, such as hyperlinks, tags, emoticons, special characters, and usernames from online travel forum posts.Word cloud feature extraction was used to identify key linguistic elements such as appraisal resources, verb processes, and personal pronouns.The most frequent words and common themes across the three subcorpora-guest complaints, hotel manager responses, and AI-generated responses-were identified and categorized.These steps were important for preparing the data for analysis and ensuring that the findings were based on accurate and reliable information.
The dataset of Text a included 115 complaints from these four hotels, totaling 26,224 words (an average of 228 words per complaint text).Together, we collected the hotel managers' responses as Text b, which contains 115 replies totaling 14,975 words (an average of 130 words per manager reply).We used the average word limit of the human manager, which is 130 words, to create an instruction for generating response letters by ChatGPT (Text c): "Please respond in 130 words to this guest complaint posted on the online travel forum.Write this response letter on behalf of the luxury heritage hotel manager in Hong Kong." We tried to keep the instructions brief and simple, allowing ChatGPT to decide how to frame and construct a response letter.We copied and pasted all the original complaints (Text a) one by one into the ChatGPT software.We later obtained 115 AI-generated responses, totaling 24,340 words, with an average of 211 words per AI-generated response letter, as shown in Table 2.Each AI-generated response letter was different (See File S1).Initially, we hoped to create an AI-generated response letter similar in length to that written by a human manager to enable direct comparison, so the word limit was set at 130 words.However, we discovered that ChatGPT met the word limit in the instruction only 2 out of 115 times (1.7%); the responses ranged from 107 to 362 words.A possible explanation is that some of these complaints are very complex, and the GAI uses more words to paraphrase issues and respond to them specifically.The paraphrasing part of the complaints in the response letters will be discussed further in the following findings and discussion section.After collecting the data, we corrected the spelling mistakes and removed sensitive details in Texts a and b.All sensitive information about guests and hotels was anonymized to protect privacy.The present study only aims to discover language features, and we avoided using specific guest names or identifiable information from hotels.In this study, pre-processing steps were carried out to ensure data quality and consistency prior to discourse analysis.Text cleaning was performed to remove irrelevant elements, such as hyperlinks, tags, emoticons, special characters, and usernames from online travel forum posts.Word cloud feature extraction was used to identify key linguistic elements such as appraisal resources, verb processes, and personal pronouns.The most frequent words and common themes across the three subcorpora-guest complaints, hotel manager responses, and AI-generated responses-were identified and categorized.These steps were important for preparing the data for analysis and ensuring that the findings were based on accurate and reliable information.
The dataset of Text a included 115 complaints from these four hotels, totaling 26,224 words (an average of 228 words per complaint text).Together, we collected the hotel managers' responses as Text b, which contains 115 replies totaling 14,975 words (an average of 130 words per manager reply).We used the average word limit of the human manager, which is 130 words, to create an instruction for generating response letters by ChatGPT (Text c): "Please respond in 130 words to this guest complaint posted on the online travel forum.Write this response letter on behalf of the luxury heritage hotel manager in Hong Kong." We tried to keep the instructions brief and simple, allowing ChatGPT to decide how to frame and construct a response letter.We copied and pasted all the original complaints (Text a) one by one into the ChatGPT software.We later obtained 115 AI-generated responses, totaling 24,340 words, with an average of 211 words per AI-generated response letter, as shown in Table 2.Each AI-generated response letter was different (See File S1).Initially, we hoped to create an AI-generated response letter similar in length to that written by a human manager to enable direct comparison, so the word limit was set at 130 words.However, we discovered that ChatGPT met the word limit in the instruction only 2 out of 115 times (1.7%); the responses ranged from 107 to 362 words.A possible explanation is that some of these complaints are very complex, and the GAI uses more words to paraphrase issues and respond to them specifically.The paraphrasing part of the complaints in the response letters will be discussed further in the following findings and discussion section.After collecting the data, we corrected the spelling mistakes and removed sensitive details in Texts a and b.All sensitive information about guests and hotels was anonymized to protect privacy.The present study only aims to discover language features, and we avoided using specific guest names or identifiable information from hotels.After collecting three subcorpora (Texts a, b, and c), we compared the linguistic features and styles used by human hotel managers and GAI systems in their responses by searching for the 100 most frequent words by word cloud and later identifying specific lexical features such as theme-related content words, appraisal resources, sensing and mental verb processes, personal pronouns, modal adjuncts, and temporal indicators and testing them by log-likelihood ratio and p-value.By comparing and contrasting these features, we discovered significant differences in linguistic features and meaning-making resources between human managers and GAI responses.To ensure the reliability of our findings, two of our research team members independently reviewed and doublechecked the log-likelihood test results as well as the coding and categorization of evaluation resources and relevant lexicogrammatical features.

Findings and Discussion
We began by generating top keywords from word clouds in the three subcorpora to obtain an overall view.After removing the conjunctions, articles, and numbers, the top 100 content words from the three corpora were examined.Figure 2 shows the distribution of all words, with the top 30 words listed in Table 3 among the three subcorpora.After collecting three subcorpora (Texts a, b, and c), we compared the linguistic features and styles used by human hotel managers and GAI systems in their responses by searching for the 100 most frequent words by word cloud and later identifying specific lexical features such as theme-related content words, appraisal resources, sensing and mental verb processes, personal pronouns, modal adjuncts, and temporal indicators and testing them by log-likelihood ratio and p-value.By comparing and contrasting these features, we discovered significant differences in linguistic features and meaning-making resources between human managers and GAI responses.To ensure the reliability of our findings, two of our research team members independently reviewed and double-checked the log-likelihood test results as well as the coding and categorization of evaluation resources and relevant lexicogrammatical features.

Findings and Discussion
We began by generating top keywords from word clouds in the three subcorpora to obtain an overall view.After removing the conjunctions, articles, and numbers, the top 100 content words from the three corpora were examined.Figure 2 shows the distribution of all words, with the top 30 words listed in Table 3 among the three subcorpora.
Frequent words in Text a-guest complaints  As shown in Figure 2 and Table 3, keywords and noun phrases such as room service, experience, food, breakfast, reception, lobby, and stay appeared more frequently in Text A complaint reviews.This is because guests used these words to express their problems and describe their experiences, with disappointment and disappointing being the most explicit reflection of attitudes.In the hotel managers' responses, they responded to issues by using phrases such as your feedback, thank you, and they emphasized their hope that customers would revisit, with words like further, next time, again, back to the hotel, hope, and looking forward, whereas the GAI was more direct and its replies focused more on apologize and regret and claimed to appreciate the complaints very much in Text C.   As shown in Figure 2 and Table 3, keywords and noun phrases such as room service, experience, food, breakfast, reception, lobby, and stay appeared more frequently in Text A complaint reviews.This is because guests used these words to express their problems and describe their experiences, with disappointment and disappointing being the most explicit reflection of attitudes.In the hotel managers' responses, they responded to issues by using phrases such as your feedback, thank you, and they emphasized their hope that customers would revisit, with words like further, next time, again, back to the hotel, hope, and looking forward, whereas the GAI was more direct and its replies focused more on apologize and regret and claimed to appreciate the complaints very much in Text C.
The data also revealed that hotel managers prefer to use positive language, emphasizing gratitude and the constructive nature of guest complaints.In addition, they often encourage guests to contact them through personal channels, such as phone calls or emails, for further follow-up, rather than engaging in open discussions about solutions on social media platforms.In contrast, AI-generated responses take a more direct approach, expressing appreciation for the complaints and clearly stating that certain actions detailed in the complaints are unacceptable, which is rarely seen in human managers' responses.After reviewing common words and testing with the log-likelihood ratio, we identified and categorized five specific areas of difference in the data.Examples are extracted from the texts for illustration.

A. Defining problems and responding to guests' negative comments B. Paraphrasing guest problems C. Sensing and mental verb processes toward investigation, apology, and solution D. Setting timeframe and intensity through graduation resources E. Aligning the guest with the use of personal pronouns
After running the log-likelihood test on hundreds of frequent words from three subcorpora, log-likelihood ratios were generated, and only those with a ratio of 3.84 or higher were chosen for further analysis, as shown in Tables 4-8.The log-likelihood ratio was calculated using Rayson and Garsides' [54] online log-likelihood wizard [55].Items with a log-likelihood greater than 3.84 are marked with an asterisk *, indicating a significant difference in the corpus at p < 0.05.A ratio greater than 6.63 ** indicates p < 0.01, greater than 10.83 *** indicates p < 0.001, and greater than 15.13 **** indicates p < 0.0001, according to the significance level scale used on the website (http://ucrel.lancs.ac.uk/llwizard.html,accessed on 1 February 2024).We focused on items with significant differences for further analysis.Figures 3-7 show a comparison of the relative frequency of elements between hotel managers and GAI texts, where red dots indicate hotel managers in Text b and blue dots indicate GAI in Text c.The red line indicates that the relative frequency for hotel managers is higher than that for the GAI text, while the blue line indicates that the relative frequency in the GAI text is higher than that in hotel managers.

A. Defining problems and responding to guests' negative comments
In the appraisal system, appreciation is used by writers or speakers to assess the value of things [37].Appreciation refers to a mental reaction to something, the composition of its components, and the evaluation of appraised items [37] (p.69).Table 4 shows the frequent appreciation words used by hotel managers in GAI.Most appreciation items are positive.The results of the appreciation analysis revealed several key differences between human hotel managers' responses and GAI responses in the context of negative guest complaints, as shown in Figure 3.The results of the appreciation analysis revealed several key differences between human hotel managers' responses and GAI responses in the context of negative guest complaints, as shown in Figure 3.No hotel managers in the data used the term "complaints" to describe dissatisfied guests' experiences.They used positive words to define the nature of the issues that arose, emphasizing the positive construction and benefits of knowing about them.However, we did not find similar language strategies, such as constructive comments, generated by GAI.Instead, it directly referred to these complaints as unacceptable, displaying an intensity that was missing from human managers' responses: o Their rudeness and lack of professionalism are unacceptable (Text 34c); o It is unacceptable that you were made to wait (Text 41c); o We understand that it is unacceptable for a five-star hotel to not provide a comfortable temperature for our guests (Text 85c).No hotel managers in the data used the term "complaints" to describe dissatisfied guests' experiences.They used positive words to define the nature of the issues that arose, emphasizing the positive construction and benefits of knowing about them.However, we did not find similar language strategies, such as constructive comments, generated by GAI.
Instead, it directly referred to these complaints as unacceptable, displaying an intensity that was missing from human managers' responses: Their rudeness and lack of professionalism are unacceptable (Text 34c); It is unacceptable that you were made to wait (Text 41c); We understand that it is unacceptable for a five-star hotel to not provide a comfortable temperature for our guests (Text 85c).

GAI responses characterized the feedback as invaluable:
Your feedback is invaluable in helping us maintain our standards (Text 44c); Your detailed feedback is invaluable to us (Text 110c).
However, the GAI emphasized pride in being a luxury hotel, which was absent from the heritage hotel managers' responses.
As a luxury heritage hotel, we take great pride in delivering exceptional service and culinary experiences to our guests.Your comments regarding the poor quality of the food and the subpar service are unacceptable, and we will address these issues immediately.(Text 3c) One possible reason is that the GAI was instructed to perform tasks through a single channel of written words, with no additional channels such as visual or gestural hits [56].All meanings can only be expressed in writing.For example, GAI uses the emotional response pride to fulfill our instruction "on behalf of the manager of the luxury heritage hotel", while the human response uses pride differently in the complaint response letter.
We failed to deliver the level of service that we pride ourselves in delivering and please accept my sincere apologies for the inconvenience you have encountered.(Text 115b) In text 115b, the pride (positive appraisal resources) in this example is not genuine pride that reinforces positive meanings about oneself.Instead, the manager used it to facilitate the subsequent apology.GAI, in this example, showed some inaccuracies in understanding complex emotions constructed in the apology scenario.

B. Paraphrasing guest problems
This section, comparing paraphrasing practices between human hotel managers and AI, was inspired by a study conducted by Min et al. [16], who identified empathy statements, paraphrasing statements, and response time as the key elements in responding to hotel reviews.Table 5 contains the common noun phrases used in Texts b and c corpora, which helps us investigate paraphrasing statements in the data.To visualize the relative difference between Text b and Text c, Figure 4 highlights the different paraphrasing approaches adopted by AI and human managers when addressing guest concerns.Our interpretation is that human hotel managers paraphrased less and focused more on generalization, while AI paraphrased more accurately, point by point, on the issues raised in the complaint texts (Text a).To visualize the relative difference between Text b and Text c, Figure 4 highlights the different paraphrasing approaches adopted by AI and human managers when addressing guest concerns.Our interpretation is that human hotel managers paraphrased less and focused more on generalization, while AI paraphrased more accurately, point by point, on the issues raised in the complaint texts (Text a).GAI has a key advantage in terms of paraphrasing: it is competent at identifying many core issues in complaint texts, such as the size of the gym, dining experience, food safety, Hotel managers tended to generalize the complaints as the details, your comments, your stay, or concern and not to repeat specific details.They encouraged guests to contact them directly via email for further private communication and assistance: We greatly appreciate you taking the time to bring these matters to our attention (Text 3b); Reading through the details of your stay (Text 114b); Thank you very much for your comment dated 01 August 2020 (Text 50b); The impression that you left with is of great concern to us (Text 61b); Welcoming you back again in the near future (Text 82b).GAI has a key advantage in terms of paraphrasing: it is competent at identifying many core issues in complaint texts, such as the size of the gym, dining experience, food safety, or problems with front desk staff, quickly and efficiently.It can provide concise paraphrases for specific issues.Examples from the data include the following: The issues you have raised regarding the size of the gym (Text 8c); We will personally oversee your dining experience to ensure . . .(Text 3c); We take matters of food safety and guest well-being very seriously (Text 18c); Will ensure better communication between our reservation team and front desk staff to prevent . . .(Text 22c); We deeply regret the unpleasant experience you encountered during check-in and check-out (Text 43c); We will use your feedback to refine our breakfast offerings in the restaurant (Text 28c).
The GAI reply to a guest's negative review is 1.62 times longer than the human reply, with an average of 211 words.This is due to the GAI's intention of paraphrasing the guest's issues in greater detail.However, our thoughts are consistent with those of Min et al. [16], who propose that a more detailed paraphrase can be beneficial because customers perceive it as more personal and tailored.Furthermore, in the data from Text 101b to 105b, we discovered that a hotel manager at Hotel D copied and pasted the same response letters five times to reply to five different guests with different problems.This may save time and effort, but responding to negative online word-of-mouth in this way may not reflect well on the hotel's reputation.

C. Sensing and mental verb processes toward investigation, apology, and solutions
According to Wan and Forey [13], two generic stages of response letters are critical: the investigation stage and the apology stage.The investigation stage involves steps like reviewing records, speaking with employees, and gathering additional information to solve problems, whereas the apology is the stage in which management sincerely apologizes to the guest for the inconvenience.By analyzing the frequent verbs used by the hotel manager and GAI, we discovered significant differences in expressing apologies and investigating problems, as shown in Table 6 and Figure 5.In terms of the verb processes (see [34]) used in the investigation phase, human managers used more sensing verbs such as note, look, hear, and read.These examples of sensing verbs used by hotel managers are extracted from the data: On the other hand, GAI does not use many sensing verbs in the process of investigating matters.It prioritizes the apology and problem-solving strategies rather than gathering details for the investigation process.GAI responses also showed much stronger expressions of apology, including words such as apologize, regret, lack, prevent, ensure, and assure.Examples of apology and problem-solving verbs are extracted as follows: We apologize for any confusion or dissatisfaction (Text 7c); o We deeply regret that the lobby atmosphere was not to your liking (Text 35c); o The food variety was lacking (Text 77c); o We have addressed this internally to prevent similar oversights in the future (Text 64c); o We would ensure that guests are allowed entry at the earliest possible time (Text 82c); o I assure you that we take them seriously (Text 22c).
Roschk and Kaiser [57] initiated a golden rule for hotel professionals to apologize to In terms of the verb processes (see [34]) used in the investigation phase, human managers used more sensing verbs such as note, look, hear, and read.These examples of sensing verbs used by hotel managers are extracted from the data: We have noted your comment (Text 27b); We could look into things which did not go right (Text 20b); We would like to hear more about your experience (Text 23b); We are saddened to read that your experience . . .(Text 44b).
On the other hand, GAI does not use many sensing verbs in the process of investigating matters.It prioritizes the apology and problem-solving strategies rather than gathering details for the investigation process.GAI responses also showed much stronger expressions of apology, including words such as apologize, regret, lack, prevent, ensure, and assure.Examples of apology and problem-solving verbs are extracted as follows: We apologize for any confusion or dissatisfaction (Text 7c); We deeply regret that the lobby atmosphere was not to your liking (Text 35c); The food variety was lacking (Text 77c); We have addressed this internally to prevent similar oversights in the future (Text 64c); We would ensure that guests are allowed entry at the earliest possible time (Text 82c); I assure you that we take them seriously (Text 22c).
Roschk and Kaiser [57] initiated a golden rule for hotel professionals to apologize to travelers immediately after a service failure, even if the service provider is not the sole cause of the problem.Later, Wang et al. [25] supported this rule but emphasized that the golden rule assumes that the apology comes from a human employee.We are aligned with Roschk and Kaiser [57] and Wang et al. [25] in believing that travelers' satisfaction is based on emotional understanding and empathy from hotel service providers.GAI is a relatively new language technology, and different customer preferences are still being studied.For example, Wang [30] states that younger generations may prefer digital interactions, whereas older guests may value traditional services.In addition, following the solution after the apology, we discovered that only GAI offered rich complimentary offerings in its responses, such as We would like to offer you a complimentary upgrade to one of our Presidential Suites during your next visit.(Text 43c); We would like to offer you a complimentary weekend stay, along with a dining credit, so that we may have the opportunity to restore your faith in our hotel (Text 72c).
ChatGPT, a language robot, frequently included many complimentary offers in its response letters.However, due to the complexity of discussing complimentary items on the online travel forum, the hotel manager's response texts did not reveal any of these complimentary specifics.The manager, with their industry knowledge, may speak with customers further after gathering additional information for the investigation.Furthermore, according to the above language features associated with the apology, we observed that when hotel managers apologize, they do so to encourage customers to revisit and make a good impression of other potential customers.However, ChatGPT, at least in its current version, is more dedicated to obtaining forgiveness by using verbs such as regret, prevent, and assure to address complaints.

D. Setting timeframe and intensity through graduation resources
Graduation is a subcategory of the appraisal system that grades attitudinal meanings up and down, and can be divided into force and focus [37,52].Force includes intensifications such as "very," while focus includes quantifications, e.g., numbers, mass, and extent [52].The focus of authors' semantic categorizations can be sharpened or blurred with words such as "exactly," or by using hedges such as "kind of" [52].Hood and Forey [58] later extended the graduation resources to time and space.Frequent words related to graduation resources are shown in Table 7 and visualized in Figure 6.In contrast, AI-generated responses focused on sharpening issues using resources such as regarding and continuously.Regarding can be categorized as a focus resource tha helps to make the discussed matter more specific, and continuously in Text c is generally associated with a future promise, as shown in the following examples: o I am truly sorry for the confusion regarding your anniversary booking (Text 1c); o We continuously strive to maintain the highest standards in all aspects of our hotel.(Text 44c).

E. Aligning the guest with the use of personal pronouns
This section discusses the differences in the use of personal pronouns between hote Regarding establishing the timeframe and intensity in the response letters, hotel managers were more likely to employ a range of graduation resources to establish intensifiers and focus, such as again, very, near, utmost, so, and only, and specific time indicators, such as further, back, and forward.These graduation examples are extracted from the data: We sincerely hope this isolated incident will not deter you from staying with us again.In contrast, AI-generated responses focused on sharpening issues using resources, such as regarding and continuously.Regarding can be categorized as a focus resource that helps to make the discussed matter more specific, and continuously in Text c is generally associated with a future promise, as shown in the following examples: I am truly sorry for the confusion regarding your anniversary booking (Text 1c); We continuously strive to maintain the highest standards in all aspects of our hotel.(Text 44c).

E. Aligning the guest with the use of personal pronouns
This section discusses the differences in the use of personal pronouns between hotel managers and GAI, as shown in Table 8. Figure 7 shows the relative frequency of personal pronouns used in the subcorpora of Text b and Text c.In contrast, AI-generated responses relied more on plural personal pronouns like w and our, which can form a collective voice encompassing the hotel manager, staff, an guests.

o
We will address these issues with our team to ensure that such instances do not recur.Luxury heritage hotels in Hong Kong have a unique historical and cultural signi cance.In the data, we found some complaints about the heritage infrastructure, as show in Table 9.In contrast, AI-generated responses relied more on plural personal pronouns like we and our, which can form a collective voice encompassing the hotel manager, staff, and guests.
We will address these issues with our team to ensure that such instances do not recur.(Text 110c); We have reinforced our privacy protocols to prevent similar incidents.(Text 111c); We apologize for any offense caused by our staff's remarks (Text 115c).
Luxury heritage hotels in Hong Kong have a unique historical and cultural significance.In the data, we found some complaints about the heritage infrastructure, as shown in Table 9.We found that both hotel managers and the ChatGPT drew on this heritage context and tailored their responses accordingly with positive messages.They were able to take a personalized and empathetic approach to responding to the details, while also relating it to the characteristics of heritage hotels as further marketing and to rebuild goodwill.GAI responses are also contextual, able to show understanding and provide effective responses, e.g., our design aims to blend heritage elements with modern comforts and show understanding to customers.This finding is consistent with Jo and Park's [28] conclusion that ChatGPT displays "human-like characteristics".

Conclusions
The hospitality industry must be aware of customer perceptions of GAI adoption and use [17].This study responds to the call for further research to explore the benefits and complexities of GAI technology in the hospitality industry [17].In this study, human hotel managers' responses to complaints about online travel forums were compared with GAI responses in the context of the negative reviews of luxury heritage hotels in Hong Kong.By exploring the linguistic features, communication strategies, and meaning-making resources employed by both human managers and GAI, this research contributes to a better understanding of using GAI in customer service operations.The log-likelihood test and discourse analysis using the SFL approach allowed us to compare the frequent words and different lexicogrammatical resources used in human and AI responses in the hospitality context.Our research has shown that both GAI and human managers can provide effective personal and empathetic responses to guest complaints and are able to refer to the heritage features of their hotels.Table 10 summarizes the key findings of this study.• GAI responses use plural personal pronouns such as we and our to form a collective voice.
To summarize some general differences, GAI is highly skilled at paraphrasing customer complaints, which is an advantage over human hotel managers.The hotel managers used sufficient sensing verbs, such as read, look, and hear, and temporal indicators, such as further, again, forward, and near, which show interpersonal skills and express personal warmth.GAI, on the other hand, emphasizes apology and future prevention, but may sometimes offer unrealistic complimentary offerings, skipping the processes of investigation and responsibility.Finally, all AI texts analyzed in this study were generated within a few seconds, which is in line with what previous researchers [27,28,30] suggest-that GAI can improve work efficiency and enhance staff productivity.Based on the data, we also discovered that GAI responses can reflect both effective language use and the specific communication strategies used in managerial roles.However, in order to optimize its use, hotel professionals will need to modify the text it produces.
The theoretical implications of this study contribute to the analysis of language differences between GAI and humans using a systemic functional linguistic approach, with a focus on interpersonal meanings.Traditional text analysis focuses on human writing, and the present study applies the linguistic theory to artificial intelligence technology.GAI developers can improve AI speech generation capabilities to match human-like characteristics and interpersonal touches expected of staff in luxury heritage hotels.The findings of this study, which examined language differences in online complaint responses between generative artificial intelligence (GAI) and hotel managers, have several practical implications for the hospitality industry.First, the findings highlight the ability of GAI to provide tailored and empathetic communication to address various guest dissatisfactions.Hotel managers can use this language technology in their interactions with guests.In addition, using GAI technology, hotel managers can save time and focus on complex issues that require human interaction.This can help the hotel allocate human resources better.In addition, GAI can assist in the analysis of successful responses and subsequent development of good language templates for staff training, thereby increasing staff productivity and service quality.Finally, the findings in relation to Text b of the human hotel managers' responses can provide insights into how empathy and problem-solving skills can be incorporated into hotel language training in different departments, for example, at the front desk.The study's findings were analyzed from four luxury heritage hotels in Hong Kong, potentially limiting their applicability to other types of hotels or geographical areas.Cultural and historical factors, as well as hotels' respective operational strategies, may influence the language used to respond to guest complaints.In addition, the reputation and mission of each hotel and guest demographic were not taken into account.Future research could expand the sample size and explore the use of GAI in different customer service industries and cultural contexts to improve the customer experience and satisfaction.

Informatics 2024 , 23 Figure 1 .
Figure 1.Data analysis process in the present study.

Figure 1 .
Figure 1.Data analysis process in the present study.

FrequentFigure 2 .
Figure 2. Frequent words appear in three subcorpora in the present study.(Source: These images were generated by www.simplewordcloud.com(accessed on 2 September 2024)).

Figure 2 .
Figure 2. Frequent words appear in three subcorpora in the present study.(Source: These images were generated by www.simplewordcloud.com(accessed on 2 September 2024)).

Figure 3 .
Figure 3. Relative frequency of appreciation resources.Human managers tended to acknowledge the complaints with the words sincere, grateful, or constructive comments.Extracts from the data include the following: o Constructive comments (Text 8b)/constructive feedback (Text 26b); o Please accept our sincere apologies (Text 4b)/sincere appreciation (Text 16b); o We are most grateful for your feedback (Text 99b).

Figure 3 .
Figure 3. Relative frequency of appreciation resources.Human managers tended to acknowledge the complaints with the words sincere, grateful, or constructive comments.Extracts from the data include the following: Constructive comments (Text 8b)/constructive feedback (Text 26b); Please accept our sincere apologies (Text 4b)/sincere appreciation (Text 16b); We are most grateful for your feedback (Text 99b).

Figure 4 .
Figure 4. Relative frequency of top noun phrases.Hotel managers tended to generalize the complaints as the details, your comments, your stay, or concern and not to repeat specific details.They encouraged guests to contact them directly via email for further private communication and assistance: o We greatly appreciate you taking the time to bring these matters to our attention (Text 3b); o Reading through the details of your stay (Text 114b); o Thank you very much for your comment dated 01 August 2020 (Text 50b); o The impression that you left with is of great concern to us (Text 61b); o Welcoming you back again in the near future (Text 82b).

Figure 4 .
Figure 4. Relative frequency of top noun phrases.

Figure 5 .
Figure 5. Relative frequency of verb processes used in investigation and coming up with solutions.

o
We have noted your comment (Text 27b); o We could look into things which did not go right (Text 20b); o We would like to hear more about your experience (Text 23b); o We are saddened to read that your experience … (Text 44b).

Figure 5 .
Figure 5. Relative frequency of verb processes used in investigation and coming up with solutions.

Figure 6 .
Figure 6.Relative frequency of graduation resources in the appraisal system.Regarding establishing the timeframe and intensity in the response letters, hote managers were more likely to employ a range of graduation resources to establish inten sifiers and focus, such as again, very, near, utmost, so, and only, and specific time indicators such as further, back, and forward.These graduation examples are extracted from the data o We sincerely hope this isolated incident will not deter you from staying with us again.(Text 9b); o I am very keen to know which areas within the hotel could improve in this regard (Text 112b); o We look forward to having the pleasure of welcoming you back in the near future (Text 15b); o The team will do our utmost to ensure your next stay is a more enjoyable one!(Text 5b); o Thank you ever so much for your comments (Text 107b); o As it is only through such valuable feedback that we are able to perfect our service (Text 53b); o I would like to drill down further with you on your experience … to ensure we get it right the first time and to that point (Text 114b); o To welcome you back in the not-too-distant future (Text 60b); o I appreciate your making me aware of the situation and very much look forward to hearing from you (Text 21b).

Figure 6 .
Figure 6.Relative frequency of graduation resources in the appraisal system.
(Text 9b); I am very keen to know which areas within the hotel could improve in this regard (Text 112b); We look forward to having the pleasure of welcoming you back in the near future (Text 15b); The team will do our utmost to ensure your next stay is a more enjoyable one!(Text 5b); Thank you ever so much for your comments (Text 107b); As it is only through such valuable feedback that we are able to perfect our service (Text 53b); I would like to drill down further with you on your experience . . . to ensure we get it right the first time and to that point (Text 114b); To welcome you back in the not-too-distant future (Text 60b); I appreciate your making me aware of the situation and very much look forward to hearing from you (Text 21b).

Figure 7 .
Figure 7. Relative frequency of personal pronouns.Both hotel managers and GAI texts used personal pronouns to create alignment an persuade guests; however, they had slightly different lexicogrammatical patterns.Huma managers mainly used pronouns such as I, my, and yours, aiming to create a unique pe sonal connection: o I could exactly feel for you how upset you have been about this visit.(Text 50b); o I fully understand your frustrations and please allow me to reiterate my sincerest apologi for the mentioned service failure.(Text 45b); o as it is only through constructive feedback such as yours that we are able to constantly mo itor and improve our hotel operation more closely (Text 86b).

(
Text 110c); o We have reinforced our privacy protocols to prevent similar incidents.(Text 111c); o We apologize for any offense caused by our staff's remarks (Text 115c).

Figure 7 .
Figure 7. Relative frequency of personal pronouns.Both hotel managers and GAI texts used personal pronouns to create alignment and persuade guests; however, they had slightly different lexicogrammatical patterns.Human managers mainly used pronouns such as I, my, and yours, aiming to create a unique personal connection: I could exactly feel for you how upset you have been about this visit.(Text 50b); I fully understand your frustrations and please allow me to reiterate my sincerest apologies for the mentioned service failure.(Text 45b); as it is only through constructive feedback such as yours that we are able to constantly monitor and improve our hotel operation more closely (Text 86b).

Table 1 .
Summary of language studies on GAI in hospitality.

Table 2 .
Corpus summary of guest complaints, management responses, and GAI responses.

Table 2 .
Corpus summary of guest complaints, management responses, and GAI responses.

Table 3 .
Thirty most frequent words, relative frequency in guest complaints, hotel manager responses, and GAI texts.

Table 3 .
Thirty most frequent words, relative frequency in guest complaints, hotel manager responses, and GAI texts.

Table 4 .
Appreciation words used by hotel managers and AI generative texts.

Table 4 .
Appreciation words used by hotel managers and AI generative texts.

Table 5 .
Frequent noun phrases used by hotel managers and AI generative text.

Table 6 .
Verb processes used by hotel managers and GAI generative texts.

Table 7 .
Graduation resources used by hotel managers and GAI-generated responses.

Table 8 .
Personal pronouns used by hotel managers and GAI generative texts.

Table 8 .
Personal pronouns used by hotel managers and GAI generative texts.

Table 9 .
Complaints on heritage infrastructure.

Table 10 .
Summary of major findings of the present study.

Table 10 .
Cont.Hotel managers use more graduation resources to establish intensifiers such as again, very, near, utmost, so, and only, as well as specific time indicators such as further, back, and forward to create action timelines in response letters.•GAIresponsesrarelyseta timeframe but sharpen issues using graduation (focus) resources like regarding and continuously.E.Aligning the guest with the use of personal pronouns•Human managers use pronouns such as I, my, and yours to create personal connections.