Abstract
Online review manipulation poses significant threat to the digital trust in tourism industry. Inspired by iceberg metaphor, this study conceptualizes emotional clues of deception into obvious, surface-level emotional signals and subtle, deeply embedded affective information. To capture the full affective architecture, we propose an Affective Iceberg Model (AIM) for review manipulation detection (AIM-RMD) incorporating GPT-based Multidimensional Affective Extraction (MAE) pipelines. Evaluated on Chinese hotel reviews, AIM-RMD achieved an accuracy of 0.8875 and significantly outperformed existing benchmarks. Detailed module-level analysis revealed that emotional diversity and specific fine-grained emotions serve as critical differentiators between genuine and fake reviews. Additionally, a second experiment demonstrated the model’s strong adaptability through GPT prompt engineering across languages and tourism sub-sectors. Our work provides novel insights into how manipulators’ emotional expressions differ from those of real users, and offers a scalable, explainable, and cross-culturally applicable solution for identifying fake reviews on hotel platforms and broader tourism entities.
Keywords
Introduction
Online review manipulation poses a pervasive threat to e-commerce, particularly in the hospitality industry, where fraudulent reviews undermine consumer trust and distort decision-making processes (Barnes, 2018; Choi et al., 2017). Existing detection methods predominantly rely on semantic, linguistic, or simple sentiment features (e.g., positive/negative polarity; A. Kumar et al., 2022; Wang et al., 2022), which can be easily replicated or imitated by manipulators. For instance, through basic online searches of hotel descriptions and image references, manipulators may fabricate seemingly authentic reviews by appropriating details about room layouts and surrounding environments without firsthand experience. Moreover, by inserting common positive or negative lexical cues (e.g., “comfortable,” “clean,” “terrible,” “poor”), manipulators can artificially simulate the evaluative tone of genuine reviews. Consequently, detection approaches based solely on such superficial features exhibit significant limitations, failing to identify meticulously fabricated reviews or address deeper theoretical underpinnings. There is a pressing need to develop detection methods capable of uncovering latent deceptive features to achieve a more comprehensive understanding and identification of fraudulent reviews.
Deception leakage theory posits that, beyond the semantic content of lies, the psychological stress inherent to deceptive behavior often manifests through paralinguistic or nonverbal cues, such as affective signals (Ekman & Friesen, 1969; Porter et al., 2011). This theoretical lens suggests that latent affective features may serve as a critical, pivotal entry point for detecting review manipulation. Authentic experiences typically entail complex, multidimensional, and dynamic emotional trajectories, reflecting not only direct reactions to events or objects but also deeper cognitive and psychological responses (Domínguez-Quintero et al., 2021). Hotel reviews based on actual stays often incorporate rich details and nuanced emotional variations, which are inherently challenging for manipulators to replicate. For instance, a guest may simultaneously experience satisfaction with room conditions, disappointment regarding service quality, pleasant surprise at dining offerings, and frustration over transportation logistics. Such multifaceted and concurrent affective states tend to manifest in their review as intricately interwoven sentiment patterns—a level of emotional granularity that proves particularly challenging to artificially replicate. Few studies have systematically quantified these latent complex affective states as discriminative cues for review manipulation detection. Thus, we propose the necessity of an analytical framework to unveil the hidden affective deceptive cues in reviews—specifically, through fine-grained multidimensional emotional leakage patterns.
Accordingly, our research objectives are three fold: to design and evaluate a GPT-based artifact that (1) extracts these affective features from hotel reviews, (2) detects manipulation through emotional cues, and (3) ensures robustness across languages, cultures, and tourism-specific contexts. To address this, we propose the Affective Iceberg Model of Review Manipulation Detection (AIM-RMD), which conceptualizes emotional features in manipulated reviews as “explicit emotional cues” and “implicit emotional cues.” This metaphorical framework positions binary sentiment (e.g., positive/negative polarity) as the visible “tip” of the iceberg—readily be replicated by fraudsters, while burying richer affective complexity (e.g., fine-grained and mixed emotions) beneath the surface. The model crystallizes our meta-requirement: to transcend superficial features by excavating deeper emotional strata for enhanced detection accuracy. Operationalizing this framework, we developed a GPT-based Multidimensional Affective Extraction (MAE) pipeline, integrating text normalization, dynamically engineered prompts design, structured API calls, and output processing. The final module synergizes AIM-derived emotional cues with conventional semantic features through ensemble machine learning for manipulation detection. We evaluated AIM-RMD through two experiments: Experiment 1 validated the model on our proprietary dataset of 2,000+ Chinese hotel reviews (spanning 20 Shanghai hotels of varying scales over 3 years). Results demonstrated AIM-RMD’s efficacy in improving detection accuracy, with interpretability analyses revealing emotional diversity as a critical latent discriminator. Experiment 2 assessed cross-domain adaptability, adapted AIM-RMD into the context of different languages and different sub-domains of the tourism sector. The model maintained robust performance, confirming its generalizability and practical utility for global platforms.
This study makes several important contributions. Practically, we propose a new solution to address review manipulation in the hospitality industry, offering scalable tools for platforms to assess review credibility. Consider a simplified example: A fraudulent positive review might state, “The stay was great, very happy!” (exhibiting singular positive polarity—the superficial “tip” of the affective iceberg), while conspicuously lacking the multidimensional emotional cadence characteristic of authentic evaluations (e.g., “Satisfied with the view but angry about the noisy room and sad about the poor service”). Under our proposed AIM-RMD framework, such manipulation is detectable due to the absence of emotional diversity and depth. Our artifact employs GPT to extract these latent layers, enabling platforms to flag suspicious reviews more effectively. The above example would be classified as manipulated, thereby safeguarding consumers and assisting hotel managers in authentic feedback analysis. The GPT-powered solution can be adapted to other tourism sectors and multilingual/cross-cultural contexts through customized prompting. Theoretically, through the Affective Iceberg metaphor, we instantiate the emotional leakage hypothesis, conceptualizing explicit and implicit affective cues in review manipulation. This model extends the theoretical understanding of emotions in fake review detection by introducing latent cues (e.g., Ekman’s six basic emotions, emotional diversity) that expose deeper, less forgeable affective signatures. Beyond this, the paper presents an integrative analytical framework that weaves together deception theory, emotional theory, and text-analytic techniques. This unified scaffold not only furnishes a novel methodology for fake-review detection but also opens a fresh avenue for cross-theoretical inquiry at the intersection of these domains.
Literature Review
Literature of Review Manipulation and Detection
Review manipulation refers to the act of artificially influencing the visibility, ranking, or content of reviews to mislead potential customers (N. Kumar et al., 2018). This is a phenomenon that deserves attention in the tourism and hospitality industries, as many businesses engage in review manipulation by offering discounts, free products, or other incentives in exchange for favorable customer reviews (Mayzlin et al., 2014), or by hiring third parties to post a plethora of irrelevant or negative comments (Li et al., 2023). These manipulations are used to enhance their own electronic word-of-mouth or to tarnish the reputation of competitors (He et al., 2022; Sparks & Browning, 2011). Studies estimate that 10% to 15% of online reviews are manipulated, with significant economic impacts on platforms like TripAdvisor and Booking.com (Luca & Zervas, 2016; Mayzlin et al., 2014). These manipulations undermine consumer trust and distort market dynamics, underscoring the critical need for advanced detection methodologies.
Existing detection approaches primarily rely on linguistic and semantic features to identify manipulated reviews. Early studies, such as Ott et al. (2013), utilized n-gram models to analyze word choice and syntactic patterns, noting that fake hotel reviews often exhibit distinct linguistic characteristics, such as fewer first-person pronouns (Catelli et al., 2023) and a lack of spatial details (Martinez-Torres & Toral, 2019). Additionally, manipulated reviews tend to overuse adjectives and focus on peripheral aspects, such as location, staff appearance, or venue reputation, rather than core service experiences (Tuomi, 2021). However, these methods have notable limitations. Linguistic and semantic features are easily replicable by manipulators who can access hotel descriptions or images online to craft convincing reviews (Hu et al., 2012).
Recent studies have begun to explore emotional and tonal cues as potential indicators of review manipulation, recognizing that these features may reveal deception in ways that linguistic patterns cannot. Catelli et al. (2022) and Zaeem et al. (2020) found a statistically significant relationship between positive emotions and genuine texts, as well as between negative emotions and fake texts. Existing literature posits that review fabricators often exhibit emotional leakage when crafting inauthentic reviews (Wang et al., 2022), which is also referred to review exaggeration (Catelli et al., 2023). This phenomenon suggests that individuals who do not accurately state their experiences tend to express emotions more intensely, as they may overcompensate to appear convincing, according to a finding by Ott et al. (2013). By Wang et al. (2022)’s theory, genuine reviews tend to incorporate a greater amount of cognitive descriptions, detailing the aspects of the experience with more factual depth. In contrast, fabricated reviews, lacking the support of objective experience, often resort to an overreliance on emotional language to substantiate their claims. However, the theory that associates exaggeration with deception remains an open question, as some scholars have found that deceptive reviews, whether expressing positive or negative emotions, tend to be less intense than genuine reviews, contrary to the exaggeration suggested by other studies (Chen et al., 2015).
Despite these advancements, the focus on emotional cues in existing RMD studies is limited to binary emotional polarities (positive/negative), which oversimplifies the complex, multi-dimensional emotional profiles of authentic reviews. This reliance on coarse-grained emotional analysis fails to capture the nuanced interplay of emotions, such as a mix of satisfaction and frustration that characterizes genuine experiences.
Relevant Theories and Their Implications for RMD
The Leakage Hypothesis of Deception
The leakage hypothesis, proposed by Ekman and Friesen (1969), posits that deceptive behaviors inadvertently reveal cues due to the psychological stress associated with lying. These cues often manifest through nonverbal channels, such as facial expressions, body language, or vocal tone, which are difficult to control consciously during face-to-face deception (Porter et al., 2012). Ekman and Friesen argued that liars focus primarily on crafting the content of their deception, diverting cognitive resources away from managing subtle paraverbal or nonverbal signals (Ekman et al., 1991). As a result, these signals “leak” emotional inconsistencies that betray the deceit, such as fleeting micro-expressions of guilt or discomfort. This theory has been widely applied in face-to-face deception detection, with studies demonstrating its efficacy in identifying subtle emotional cues in interpersonal interactions (Porter & Ten Brinke, 2008).
While the leakage hypothesis has primarily been applied to verbal communication, its core principle that deception leads to unintended emotional leakage, offers valuable insights for detecting manipulated online reviews. In the context of written communication, such as hotel reviews, manipulators may exhibit subtle emotional expressions that differ from those of authentic reviewers. Similar to face-to-face liars who prioritize the content of their deception over their emotional tone or body language, review manipulators focus on generating content that aligns with their persuasive goals, such as promoting a hotel or disparaging a competitor. This focus often results in emotional expressions that lack the depth, complexity, or authenticity found in reviews based on genuine experiences (Ott et al., 2011). For example, a manipulated review might emphasize generic positive descriptors (e.g., “amazing stay”) or negative complaints (e.g., “terrible service”) without the nuanced emotional interplay, such as a mix of satisfaction and frustration, that characterizes authentic reviews.
Despite its potential, the application of the leakage hypothesis to textual deception detection remains underexplored. This gap highlights the opportunity to adapt the leakage hypothesis to online reviews by focusing on emotional leakage—subtle, multi-dimensional emotional expressions that manipulators struggle to replicate due to their lack of firsthand experience.
Discrete Emotion Theory
Discrete emotion theory posits that human emotions extend beyond simple positive or negative polarities, encompassing a range of distinct categories that reflect nuanced psychological states. Ekman (1992) identified six basic emotions: happiness, sadness, anger, fear, surprise, and disgust, which are universal across cultures and characterized by specific physiological and expressive markers. Tomkins (1962) proposed nine basic emotions: interest, enjoyment, surprise, distress, anger, disgust, shame, and fear. These basic emotions serve as building blocks that combine to form more complex emotional states, capturing the rich spectrum of human affective experiences (Plutchik, 1982). This theory has become a cornerstone of emotion computation, widely applied in fields such as natural language processing and speech analysis to quantify emotional information in human-generated data (Shiota et al., 2017).
In the context of review manipulation detection (RMD), discrete emotion theory offers a powerful framework for moving beyond the limitations of binary sentiment analysis (see the illustration in Figure 1). Online hotel reviews are rich with emotional information, often expressing a dynamic interplay of emotions rather than mere praise or criticism (Aeron & Rahman, 2023). Hotel reviews are no exception: authentic reviews may encompass various aspects of the hotel experience, and we should aim to capture the emotional depth of these reviews while examining whether fake reviews differ in their depth of emotional expression, such as exhibiting more flattened emotional profiles. It inspires our study to integrate discrete emotion theory into RMD, leveraging advanced algorithms, such as those based on GPT, to extract and analyze multiple discrete emotions in review texts as potential discriminators for manipulation.
Ekman’s six basic emotions and their manifestations in hotel reviews.
The Affective Iceberg Model for Review Manipulation Detection
Design Principle
The iceberg metaphor has been widely used to describe the hierarchical relationship between explicit and implicit aspects, offering a vivid framework for understanding observable and hidden aspects of complex phenomena. In reality, only a small fraction of an iceberg is visible above water, with the vast majority hidden below, symbolizing that what is apparent often represents only a fraction of the whole. The visible tip of the iceberg, floating above the water’s surface, represents outward, easily observable characteristics, while the submerged mass beneath the surface conceals deeper, less accessible elements that require careful exploration to uncover. This metaphor was first applied to describe human psychology by Breuer and Freud (2009), who likened conscious thought to the visible tip and the preconscious and unconscious to the submerged layers. Building on this foundation, subsequent models, such as Satir’s Iceberg Model (1964) and McClelland’s Competency Iceberg Model (1973), also use the metaphor to describe the surface-level expressions (e.g., self-presentation or workplace behaviors) and deeper, intrinsic attributes (e.g., emotions or competencies; Jackson & Weakland, 1961; McClelland, 1973). These models have a profound influence on fields such as psychology, behavioral science, and management, promoting the understanding of the underlying human inner states within these domains.
The iceberg concept could also have a relation to review manipulation detection (RMD). Deception leakage theory suggests that while manipulators can fabricate surface emotions, they consistently fail to replicate the unconscious affective leakage characteristic of authentic reviews. It inspires us to conceptualize the emotional cues embedded in online reviews and map the emotional leakage at different levels of depth to the iceberg’s structure. When writing reviews, consumers express binary sentiments—positive or negative—based on their experiences relative to expectations. These explicit emotional polarities can serve as visible clues for detecting manipulation, as manipulators often produce exaggerated praise or criticism to achieve persuasive goals, resulting in anomalous sentiment patterns as summarized in Section 2.1. Beyond these surface-level binary emotions lies a wealth of finer-grained, implicit emotional cues that hold greater potential for identifying deceptive reviews. Discrete emotion theory informs us to quantify these deeper emotions as combinations of basic emotions (e.g., happiness, sadness, anger, fear, surprise, disgust) and their complex derivatives. We conceptualize this distinction as the Affective Iceberg Model (AIM) for Review Manipulation Detection, where the iceberg represents visible binary sentiment (the tip) versus hidden fine-grained emotions (the submerged mass). Manipulators falter in replicating the submerged affective complexity, making these deeper emotional cues a critical marker for distinguishing authentic from deceptive reviews.
To quantify the underlying emotional cues in the iceberg, we developed a GPT-based Multidimensional Affective Extraction (MAE) pipeline, which extracts fine-grained emotional profiles from review texts. This pipeline, combined with subsequent machine learning algorithms for identifying manipulation patterns, forms an instantiation of the AIM-RMD artifact. To summarize, the whole design artifact is guided by the iceberg design metaphor and grounded in emotional leakage theory and discrete emotion theory, as illustrated in Figure 2.
The Affective Iceberg Model.
Affective Iceberg Conceptual Module
The Affective Iceberg Model (AIM) conceptualizes emotional cues in review manipulation detection (RMD) as an iceberg, where distinct layers of emotional features correspond to the visible and hidden portions of the structure. This module delineates two critical components: surface sentiment features, represented by the iceberg’s visible tip above water, and latent sentiment features, embodied by the submerged mass below water. By distinguishing these layers, AIM provides a conceptual model to depict emotional patterns that differentiate authentic from manipulated reviews, emphasizing the need to explore deeper emotional cues beyond traditional detection methods.
Above Water—Easily Replicated Surface Sentiment Features
The tip of the iceberg, visible above the water’s surface, represents overt, low-dimensional emotional features, primarily binary sentiment polarities (positive or negative). These surface sentiment features are simple, direct, and easily observable, forming the superficial emotional facade of online reviews. For instance, a review might express unreserved praise (“amazing stay, highly recommended”) or stark criticism (“terrible service, avoid at all costs”), reflecting a clear positive or negative tone. Traditional RMD methods have focused on these features, employing tools like sentiment analysis (e.g., VADER, BERT) to quantify emotional polarity and detect anomalies, such as exaggerated sentiment that may indicate manipulation (Catelli et al., 2023; Wang et al., 2022).
However, the simplicity of the surface sentiment features makes them easily replicable by manipulators, who can mimic positive or negative tones. In addition, binary sentiment captures only a coarse approximation of emotional expression, failing to reflect the full spectrum of a reviewer’s experience (Bouazizi & Ohtsuki, 2016). For example, a genuine review might blend satisfaction with minor frustrations, but such nuance is often absent in manipulated reviews that prioritize persuasive impact over authenticity. It could be insufficient to robustly distinguish authentic from manipulated reviews relying solely on surface sentiment, as these features do not capture the deeper emotional complexity that characterizes genuine experiences (Yucel et al., 2022).
Below Water—Hardly Replicated Latent Sentiment Features
The submerged mass of the iceberg represents latent, high-dimensional emotional features that are less observable but more revealing of review authenticity. These include but are not limited to fine-grained emotions (e.g., happiness, sadness, anger, surprise), emotional diversity, and emotional interaction cues, which collectively capture the nuanced and multifaceted nature of human experiences (Larsen et al., 2017). Unlike surface sentiment, these latent features can reflect rich emotional dimensions from multiple perspectives, better capturing the nuanced and sincere emotions in authentic reviews or the hollow and flat emotions in manipulated reviews, thereby distinguishing the emotional expression differences between genuine and fake reviews. Such emotional richness is difficult for manipulators to replicate due to their lack of genuine experiential grounding or constraints on time and effort, thus revealing their deception. The leakage hypothesis suggests that manipulators, focusing on crafting persuasive surface content, inadvertently reveal their deception through the absence or flattening of these latent emotional features. For instance, manipulated reviews may exhibit limited emotional diversity, relying on a single dominant emotion (e.g., exaggerated happiness) to achieve their goal, resulting in an emotionally hollow or overly uniform tone.
The core insight of AIM lies in recognizing that emotional cues in review manipulation extend beyond the surface-level positive or negative polarities. The model underscores the importance of delving into the submerged, fine-grained emotional features to uncover patterns that distinguish authentic expressions from manipulative ones.
GPT-Based Multidimensional Affective Extraction Module
Building on the AIM, which conceptualizes RMD through explicit binary sentiments (visible tip) and latent fine-grained emotions (submerged mass), this module addresses the challenge of quantifying the complex, multi-dimensional emotions beneath the iceberg’s surface. Traditional sentiment analysis methods struggle to capture the nuanced emotional profiles critical for distinguishing authentic from manipulated reviews. Therefore, we leverage the emotional intelligence of large language models (LLMs) by designing a GPT-based Multidimensional Affective Extraction (MAE) module. This module automatically extracts fine-grained sentiment scores from online hotel reviews and provides reasoning for each score, enabling robust detection of emotional discrepancies indicative of manipulation.
Input Preprocessing
Raw review texts are cleaned by removing HTML tags, punctuation, and stopwords (e.g., “the,” “and”) to reduce noise and focus on meaningful content. Then the texts are normalized to a consistent encoding (e.g., UTF-8) and, for multilingual datasets, converted to simplified Chinese where applicable to ensure compatibility with the GPT model and maintain linguistic consistency across reviews. To handle lengthy reviews within the model’s token limits, texts are then segmented into manageable chunks while preserving semantic coherence, allowing comprehensive analysis without truncating critical emotional content.
Prompt Design
Carefully crafted prompts are critical for GPT to focus on relevant emotional dimensions, provide intensity scores, and justify its reasoning. By configuring GPT as a hotel sentiment analysis specialist, we ground its interpretations in a domain-specific context, allowing it to discern subtle affective patterns unique to hospitality reviews. This expertise is channeled through precise instructions that direct the model to: (1) identify six basic emotions of Ekman (1992), (2) quantify their intensity (0–1 scale), and (3) provide evidential reasoning, for instance, linking “spotless bathroom” to disgust avoidance (0.1 intensity) versus “stained towels” to disgust activation (0.7). To ensure analytical rigor and computational utility, we enforce JSON-structured outputs that capture both quantitative scores and qualitative rationales. This machine-readable format seamlessly integrates with downstream detection algorithms while maintaining human-interpretable logic.
The system’s precision is further refined through strategic few-shot examples, which is also called in-context learning technique. We include 2 to 3 review samples with annotated emotion scores and reasoning to guide GPT toward consistent and nuanced emotional extraction. These exemplars prime the model for complex emotional profiling, particularly crucial when analyzing mixed-affect scenarios where authentic reviews naturally exhibit emotional oscillations that manipulated content fails to replicate.
Finally, inspired by Wei et al. (2022), we innovatively introduced three structured reminders into the prompt design to enhance the large language model’s organizational clarity and align its reasoning process more closely with human-like logic. For instance, the model is guided to reason step by step: first extracting the explicit binary sentiment (representing the “tip of the iceberg”), followed by identifying the implicit fine-grained emotions (reflecting the “submerged layer”). Additionally, it is instructed to systematically execute each subtask through a standardized procedure of analysis, scoring, and normalization.
Our prompt design extends beyond basic task descriptions by incorporating state-of-the-art in-context learning and Chain-of-Thought (CoT) techniques, forming an innovative technical template for interaction with large language models. This template is abstracted and illustrated in Figure 3 to facilitate replication and adaptation across various domains.
Illustration of the prompt template of GPT-based Multidimensional Affective Extraction (MAE) pipelines.
OpenAI API Connection
Integration with the OpenAI ChatGPT API enables scalable and efficient sentiment analysis, the process is illustrated in Figure 4. The setup involves establishing a secure authentication layer through a rigorously configured API key firstly. Then, parameters such as temperature and max tokens are optimized to balance response quality and computational efficiency. Also, the version of the LLM needs to be specified, and reviews are processed in batches or via a pipeline to handle large datasets efficiently, with rate-limiting mechanisms to manage API quotas finally.
Illustration on calling ChatGPT API to analyze texts.
Output Standardization
To ensure reliable and usable results, the module standardizes GPT outputs through the following steps. First, JSON outputs are parsed to extract emotion scores and reasoning, ensuring structured data for downstream analysis. Then, responses are checked for errors, such as missing scores or contradictory reasoning, using predefined validation rules (e.g., scores within 0–1 range). Next, invalid outputs are either repaired (e.g., by re-running the prompt with adjusted parameters) or discarded to maintain data integrity. After that, validated emotion scores and reasoning are stored in a database, enabling integration with machine learning algorithms for manipulation detection.
This GPT-based MAE module operationalizes AIM by quantifying the submerged emotional complexity of reviews, providing a foundation for identifying manipulation through fine-grained emotional cues that traditional methods overlook.
Detection Module Using Multiple Machine Learning Classifiers
As previously discussed, our approach is inspired by the iceberg model to quantize both the overt binary sentiment (representing the visible tip) and the latent fine-grained affective complexity (reflecting the submerged structure) as discriminative cues for review manipulation detection. These emotions serve as indicators for identifying comment manipulation. In this regard, we further integrated these indicators into our machine learning model to generate a final judgment on whether a given comment is a manipulated review. To achieve this, we initially employed a Bag-of-Words (BoW) model to extract high-frequency terms from the reviews that are independent of sentiment, thereby forming our foundational features. We first compile a corpus-specific vocabulary from a stratified sample of hotel reviews:
After lower-casing, de-accenting and removing standard stop-words, we retain the top-K most frequent unigrams plus any domain-relevant bigrams (e.g., “front_desk,” “late_checkout”) that exceed a
where
To weigh the basic information of the review by probability, the term frequency and inverse document frequency are also considered.
So the fundamental semantic features in each review are represented as:
Subsequently, we incorporated the results from the previously established GPT-based Multidimensional Affective Extraction (MAE) module as advanced features. Role of GPT, task description, output format, GPT-parameters, n-shot examples and CoT reminders are specified in Section 3.3.
These comprehensive features were then fed into five widely used classifiers. Random Forest is an ensemble learning method that operates by constructing a multitude of decision trees during training and outputting the class that is the mode of the classes of the individual trees. The final decision of the review is determined by majority voting, where
Logistic Regression models the probability that a given input belongs to a particular class using the logistic sigmoid function, where
Support vector machine (SVM) is another typical machine learning model, which aims to find the optimal hyperplane that maximizes the margin between different classes in a high-dimensional space. It can handle both linear and non-linear classification via the kernel trick, which is suitable to capture the interaction between the fundamental feature and complex affective cues in reviews. The decision function and the objective function for SVM are:
Naïve Bayes is also adopted due to its advantages of being simple, fast, and particularly effective for high-dimensional data like text. It decides based on applying Bayes’ theorem with strong (naive) independence assumptions between the features:
We also consider neural network models like Multilayer Perceptron (MLP) to produce decisions on whether the reviews are manipulated or not. As a class of feedforward artificial neural network (ANN) that consists of at least three layers of nodes: an input layer, one or more hidden layers, and an output layer, MLP has extraordinary flexibility and is also adept at handling nonlinear relationships. It processes the calculation layer by layer, and the output of a neuron in layer
where
Evaluation Experiment
To evaluate the practical performance of the proposed AIM-RMD artifact, we designed and conducted two experiments based on real-world reviews and integrated the technical infrastructure within a tourism management environment, with corresponding results summarized in Table 1. The first experiment was designed to assess the proposed artifact across multiple dimensions—including metric computation, interpretability, and robustness, following recommendations from Hevner et al. (2004)—by measuring both error and accuracy in detecting manipulated Chinese hotel reviews using a range of evaluation metrics. We further conducted an effectiveness analysis of each module in AIM-RMD, with results offering new insights for knowledge bases such as the operational boundaries of LLMs in review manipulation contexts. Inspired by Abbasi et al. (2024), the second experiment emphasized domain-specific adaptability, examining the capability of AIM-RMD to identify manipulated reviews across different languages, cultural contexts, and tourism sectors. This investigation aimed to explore the generalizability of the proposed model under varied and realistic conditions.
Summary of Experiments.
Experiment 1: Model Performance on Chinese Hotel Review Manipulation Detection
Experiment Design
Dataset Construction
As one of the world’s largest tourism markets, China accounts for a significant portion of international travel movements. The hotel industry, as a key component of the tourism value chain, is at the forefront of tourist-business interactions, making it a critical barometer to understand tourist behaviors (Francesco & Roberta, 2019). Therefore, in Experiment 1, we collect hotel reviews from China’s largest travel platform—Ctrip, 1 to implement and test our designed AIM for RMD. During the actual harvesting of reviews from the platform, we adhered to a tripartite sampling doctrine of diversity, representativeness, and bounded randomness, and collected reviews for 20 hotels in Shanghai, spanning various geographical locations and star ratings. Considering the timeliness of the data, we have only collected reviews from 2022 up to the first half of 2024.
Subsequently, we proceed to obtain ground truth labels indicating whether each review is manipulated, which will guide the training of our proposed detection model and facilitate the evaluation of its performance on the test set. In real-life review data, it is extremely challenging for researchers to obtain labels indicating whether hotel reviews are genuine or not. The filtering algorithms of platforms have access to a vast amount of user behavior data, including the reviewer’s IP address, device information, account history, purchase records, review frequency, and more, using a unified standard to authenticate the authenticity of reviews. For instance, Li et al. (2023) utilized the filtering algorithm results from the platform to label online reviews as either genuine or fake, marking those filtered out by the platform as fake reviews and the remaining ones as genuine. Similarly, we followed the labeling methods of previous literature and used the platform’s filtering mechanism to determine the authenticity of hotel reviews. Reviews identified by the platform as not having stayed at the hotel, malicious reviews (fabricating false information to tarnish a hotel for improper gains), promotional reviews (written by friends, family, or employees associated with the merchant, or left by consumers through exchanges of benefits or emotional coercion), and other reviews deemed unauthentic by the platform are labeled as “manipulated reviews.” The rest are considered genuine reviews without manipulation.
Preliminary collection results indicate that we obtained 35,452 genuine reviews and 11,595 fake reviews, which also implies that the platform has folded approximately 24.6% of the reviews. We then cleaned all the reviews, which included removing duplicates and filtering out reviews that were too short or too long. Considering the ratio of genuine to fake reviews in the cleaned dataset is 3:1 and the ratio of positive to negative reviews is 30:1, which can significantly undermine the training effectiveness of a classifier due to the imbalanced training data (Thabtah et al., 2020), we conducted under-sampling. After sampling, we reduced the dataset to a more balanced subset of 2,210 reviews, ensuring a 1:1 ratio between genuine and fake reviews, while maintaining a 2:1 ratio between positive and negative reviews.
Baselines
The innovation of the proposed AIM-RMD, distinguishing it from existing methods for detecting review manipulation, lies primarily in two key aspects. First, it leverages the iceberg metaphor to argue that, beyond superficial binary sentiment cues, latent fine-grained emotional information should be recognized and excavated as critical signals for identifying manipulation. Second, it constructs a novel Multidimensional Affective Extraction (MAE) pipeline powered by advanced LLMs to systematically quantify complex emotions in online reviews. Our baseline selection aligns with these two directions. To examine the contribution of incorporating fine-grained emotional information, we established baselines that either exclude emotional features entirely or incorporate only binary sentiment, thereby enabling a comparative evaluation against AIM-RMD’s holistic integration of the full affective iceberg. Furthermore, for the task of emotion analysis and scoring from reviews, we adopted widely adopted lexicon-based methods as baselines to benchmark against the proposed GPT-based MAE module.
Metrics
We employ a series of metrics to evaluate the performance of our classification models, all of which are derived from the confusion matrix associated with classification problems. The confusion matrix is a table that allows for the visualization of the performance of an algorithm, particularly in the context of a binary classification. It delineates the true positive (TP), false positive (FP), true negative (TN), and false negative (FN) outcomes, providing a comprehensive overview of the model’s predictive accuracy. Based on this matrix, we calculate precision and recall. Precision, also known as the positive predictive value, is the ratio of correctly predicted positive observations to the total predicted positives. Recall, which measures the sensitivity or true positive rate, is the ratio of correctly predicted positive observations to all actual positives. Precision and recall are often inversely related metrics in the context of binary classification models. Generally, high precision indicates that the model is conservative in its predictions, ensuring that the positive instances it predicts are likely to be correct. On the other hand, high recall indicates that the model is aggressive in capturing positive instances, aiming to detect as many actual positive cases as possible. The F-score is a harmonic mean of precision and recall, providing a single score that balances both metrics. Accuracy, on the other hand, is the ratio of correctly predicted observations to the total observations, which provides an overall measure of the model’s effectiveness. In addition to these metrics, we also compute the Area Under the Curve (AUC), which is based on the Receiver Operating Characteristic (ROC) curve. The ROC curve is a graphical representation that illustrates the diagnostic ability of a binary classifier system as its discrimination threshold is varied. The AUC, which ranges from 0 to 1, provides a measure of the model’s ability to distinguish between the two classes, with an AUC of 1 indicating perfect classification performance.
Results
Detection Performance
The assessment outcomes are displayed in Table 2. As stated before, we evaluated the breakthrough of AIM-RMD from two distinct dimensions: the first includes methods that utilize only surface-level binary sentiment (corresponding to the “Above water” sections in the table), and the second consists of lexicon-based approaches that capture emotional cues using sentiment vocabulary ontologies from Dalian University of Technology. In contrast, our proposed method, which leverages GPT to comprehensively integrate the full spectrum of the affective iceberg for deception detection, is presented in the final section of the table.
Comparison of Fake Review Detection Metrics.
First, our findings under the proposed AIM abstraction reveal that as emotional features transition from the superficial layer (above water) to the deeper, submerged layer of the iceberg, nearly all accuracy metrics across the compared methods demonstrate significant improvement, leading to more precise identification of review manipulation. The highest detection accuracy is achieved when the full spectrum of the affective iceberg—encompassing both explicit and implicit emotional cues—is incorporated into the model. Specifically, the accuracy reaches 0.8875, substantially outperforming the result obtained using only surface-level sentiments, which yielded an accuracy of 0.7919. This demonstrates that the proposed AIM conceptualization for review manipulation possesses substantial practical value, offering a theoretically grounded and effective approach to enhance the detection accuracy of manipulated reviews on hospitality and tourism platforms. By guiding the identification process through in-depth emotional mining—from superficial polarity to nuanced affective layers—our work provides a novel sentiment-based methodology to tackle review manipulation.
Second, upon examining the second set versus the third set, or the fourth set versus the fifth set of features in the table, it is evident that the proposed GPT-based affective extraction framework exhibits higher performance in detecting fake reviews than the popular lexicon-based approaches. This indicates that employing advanced models like GAI to enhance the accuracy of sentiment analysis is beneficial for differentiating between genuine and fake reviews. Also, we notice that the proposed AIM-RMD framework generates not only better, but also stabler results across diverse model architectures, potentially because of the effectiveness of the fine-grained affections captured by the GPT-based extraction modules, which we will evaluate in the next sub-section.
Additionally, we have observed that our AIM-RMD framework offers a better balance in detecting both truthful and deceptive reviews, maintaining well-rounded performance for both. In contrast, other methods may sacrifice one for the sake of the other. For instance, the Naive Bayes classifier, when utilizing coarse-grained GPT sentiment scores, exhibits superior performance in identifying truthful reviews with a recall of 0.9050, yet it struggles with deceptive reviews, achieving only a 0.6018 recall. However, AIM-RMD provides a more equitable recall between the two classes, indicating that it can recognize a broader spectrum of patterns characteristic of both deceptive and truthful reviews.
Effectiveness of Fine-Grained Sentiment
The previous section empirically confirms that our proposed AIM-RMD framework enhances the accuracy of review manipulation detection. A natural follow-up question arises: which component (or components) within AIM-RMD primarily contributes to this detectable improvement? Addressing this question not only allows for a more granular evaluation of the artifact but also enhances interpretability regarding how manipulation is identified through the inputs and outputs of its constituent modules. We begin by examining the first module of the framework—the role played by the hidden, submerged emotions revealed by AIM in detecting review manipulation. To this end, we fitted the distribution of fine-grained emotional scores (across six distinct emotions) for both genuine and fraudulent reviews and present the results in Figure 5. It can be observed that there are distinct specificities in the fine-grained emotions between genuine and fake reviews.
GPT-extracted fine-grained emotional distributions between authentic and deceptive reviews.
For example, in the happiness dimension, deceptive reviews show a higher peak and concentrated range than genuine ones, indicating that these reviews are more likely to express inflated happiness, while truthful reviews display a wider distribution. Also, compared to fake reviews, genuine reviews exhibit a multi-peak characteristic, especially in the emotion of sadness. Genuine reviews have high density around scores of 0, 0.5, and 0.7. This indicates that authentic reviews contain a variety of emotional expressions in terms of intensity and type. Some are more restrained or objective, while others involve emotional venting. In contrast, the sadness emotion in fake reviews is characterized by a prominent single peak, with a large number of reviews having scores concentrated at 0. There is also a slightly higher value around 0.5, but it is barely noticeable. This suggests that fake reviews either do not contain the emotion of sadness or express it very weakly. If present, such expressions are extremely rare. In the score distributions of anger and disgust, genuine and fake reviews also show similar differences. The intensity of these emotional expressions in genuine reviews is significantly higher than that in fake reviews. This is possibly because genuine reviews reflect users’ direct experiences and reactions after using a product or service. When users encounter unsatisfactory situations, they tend to naturally express emotions such as anger or disgust. In contrast, fake reviewers lack authentic emotional experiences while creating their content, making it difficult for them to naturally convey strong emotions like anger or disgust. This artificiality results in a more uniform and weakened expression of emotions in fake reviews. In the case of fear and surprise, the differences between genuine and fake reviews are not as pronounced as with other emotions. However, it can still be observed that the range of scores covered by genuine reviews is slightly broader than that of fake reviews.
With the incorporation of fine-grained sentiment information extracted by GPT, fake reviews can be identified more effectively. Another question is: which specific emotional dimensions are more useful in the task of detecting fake reviews? This is important for understanding the theory of review manipulation, as it can help pinpoint the core emotional features of fake reviews and provide emotional cues for distinguishing between genuine and fake reviews. To this end, this paper examines the importance of six fine-grained sentiment features and ranks them accordingly. To ensure the robustness of the results, three classifiers capable of naturally assessing feature contributions are employed to calculate the importance of each fine-grained sentiment individually, with the results displayed in Figure 6.
Fine-grained sentiment importance ranking in fake reviews detection task: (a) RF, (b) logistic, and (c) SVM.
Although the feature importance rankings generated by the three classifiers are not identical, it can be observed that happiness, sadness, and anger consistently rank as the top three most important sentiment features. Disgust ranks fourth, while fear and surprise are the least important features. This indicates that happiness, sadness, and anger play a crucial role in the detection of fake reviews, and these emotions are more likely to be manipulated in deceptive reviews. In light of the previously discussed score distributions of these emotions in genuine and fake reviews (Figure 5), it appears that fake reviews may exaggerate or fabricate the advantages of a product or service by overexpressing happiness. Conversely, due to the lack of authentic experiences and emotional grounding, fake reviews tend to underrepresent strong emotions such as sadness or anger.
Implication for RM: Sentiment Diversity
Building upon our previous discussion of latent fine-grained emotions, we observed that the emotional expressions in genuine reviews typically exhibit a multi-peak distribution and frequently involve a mixture of multiple emotions. This inspires us to examine whether there are significant differences in the emotional richness of genuine and fake reviews. Therefore, we further analyzed the emotional diversity in genuine and fake hotel reviews, measured by the number of non-zero sentiment categories. The resulting distribution is shown in Figure 7. The figure presents the distribution of different sentiment categories in genuine and fake hotel reviews in the form of histograms and kernel density curves. The x-axis represents the number of sentiment categories (e.g., a review containing happiness, anger, and surprise would correspond to a value of 3), while the y-axis indicates the number of reviews with the corresponding sentiments.
Sentiment diversity distributions in real and fake hotel reviews: (a) real reiviews and (b) fake reviews.
It can be observed that genuine hotel reviews typically contain two to four different sentiments per review, with a significant number even reaching five or six sentiments. In contrast, fake reviews mostly contain only two sentiments per review, and a considerable portion have just one sentiment. When we calculate the average number of different emotions in genuine and fake reviews, we find that each genuine hotel review contains an average of 3.38 distinct emotions, while fake reviews contain only an average of 2.07 distinct emotions. This is an interesting phenomenon: Genuine reviews often reflect a complex interplay of emotions based on their actual experiences. In contrast, manipulated reviews typically exhibit a singular emotional tone, either overly positive or overly negative, to achieve their intended impact. This distinction provides a novel theoretical perspective on the emotional landscape of review manipulation.
Case Study: Quality of Emotional Quantifications
Building upon the above findings, which solidify the importance of submerged fine-grained emotions (as conceptualized in the Affective Iceberg Model) for detecting review manipulation, we now turn to evaluating the performance of the second core component of the AIM-RMD artifact: the GPT-based Multidimensional Affective Extraction (MAE) module. This module is centered on leveraging carefully designed prompts to activate GPT’s affective intelligence for analyzing and extracting complex emotions from online reviews. A natural way to assess its utility is to examine the quality of its emotional interpretations. Given the large volume of reviews, we adopted a case study approach, evaluating and discussing the outputs for a selected set of representative examples. We summarize the aspects where the module’s outputs align closely with human judgment in Table 3, while its current limitations are detailed in Table 4. Since a key innovation of the GPT-based MAE module lies in its use of generative AI instead of traditional lexicon-based methods, we also include results from a lexicon-based approach in both tables for comparative analysis.
Advantages of GPT-Based MAE Pipeline in Analyzing Complex Sentiment of Hotel Reviews.
Limitations of the GPT-Based MAE Pipeline in Analyzing Complex Sentiment of Hotel Reviews.
By examining examples with complex contexts and semantics that convey emotional information, we found that the proposed GPT-based MAE module has an overwhelming advantage over the Dalian University of Technology Sentiment Dictionary in the following aspects. It excels at understanding personalized expressions of emotions such as irony, metaphor, and slang. Additionally, it demonstrates capabilities that conventional dictionary-based sentiment analysis lacks, such as judging positive or negative sentiment and emotional orientation based on context, as well as identifying emotional shifts in subtle or long-sentence transitions. Guests with different personalities may express their emotions in various ways, which can pose significant challenges for lexicon-based extraction methods. For example, when a tourist uses a sarcastic tone to express irony, such as saying “really impressive!,” methods based on the Dalian University of Technology Sentiment Dictionary can only recognize the literal meaning, interpreting the sarcasm as praise. Similarly, when a guest humorously describes a flat pillow as being like their wallet to express dissatisfaction, the lexicon-based approach fails to capture the underlying emotion effectively. In such cases, our GPT-based MAE module demonstrates a higher level of “emotional intelligence,” capable of understanding both positive and negative expressions and interpreting emotions conveyed through humorous or poetic language styles. There are also situations that require flexible judgment of emotions based on context. The proposed module can discern positive intent within negative phrases and understand emotional shifts. In contrast, dictionary-based methods are more rigid and may lead to misjudgments.
However, we also point out the limitations of the GPT-based MAE pipelines in sentiment analysis. As shown in Table 4, our module sometimes confuses similar emotional categories, leading to misjudgments, such as between anger and disgust. Additionally, its lack of in-depth reasoning about event logic results in errors in specific situations. For instance, it does not assign higher weight to the most recent feelings, and sometimes fails to grasp the guest’s emphasized expressions. These deficiencies in weight assessment, causal reasoning, and key-point identification may lead to inaccurate sentiment analysis. In addition, it could struggle to understand implicit emotions behind polite euphemisms, failing to detect the underlying sentiments in more subtle and indirect statements.
Robustness
Our AIM-RMD framework exhibits a degree of stochasticity in its GPT-based MAE module, as the responses generated by the module are influenced by parameters configured in the prompt design—notably the temperature parameter, which directly modulates the randomness, stability, and diversity of the output. The closer the value is to 0, the more deterministic the responses generated by GPT will be. Conversely, the closer the value is to 1, the more random the responses will become. To examine whether the results would be influenced by this parameter, we set the temperature parameter to 0.3 (low), 0.6 (medium), and 1 (high) respectively, and observed the sentiment features extracted by GPT and its performance in detecting review manipulation under these different temperature settings.
There are no significant differences in the distribution of happiness, sadness, anger, fear, and disgust features across different temperature settings. However, the extraction of surprise shows notable differences (see Figure 8). When the temperature is high, the distribution range of this emotion becomes wider, with a longer tail and a more random overall shape. This indicates that different emotional categories have varying sensitivities to temperature changes. When identifying mainstream emotions such as happiness, sadness, and anger, the GPT-based module is relatively stable because the vocabulary associated with these emotions is relatively fixed.
Distribution of surprise scores across different temperatures.
In terms of detecting review manipulation across different temperature settings, we did not observe significant differences in performance, which indicates the stability of the proposed GPT-based MAE module. Although the results of sentiment extraction and review manipulation detection were similar across these temperature settings, we still recommend a temperature setting of 0.3. When the temperature parameter was increased, we noticed a significant rise in the occurrence of garbled text and responses that did not follow the required format in the feedback. This greatly increased the cost of data processing.
Experiment 2: Cross-Cultural and Domain Adaptation Evaluation
As suggested by Abbasi et al. (2024), an artifact must demonstrate strong domain adaptation capabilities to be of practical value. Accordingly, Experiment 2 was designed to evaluate the proposed AIM-RMD artifact using review datasets spanning multiple countries/regions, languages, and specialized sectors within the tourism industry.
Experiment Design
Data
English is one of the most widely used languages on the internet, validating our method on English hotel reviews can help promote its application for fake hotel review detection in more countries. Therefore, we selected an English gold standard hotel review dataset known as the Deceptive Opinion Spam Corpus v1.4 (DOSC). 2 It comprises a total of 1,600 reviews for 20 distinct hotels located in Chicago. The dataset collects 800 truthful reviews from reputable online platforms such as TripAdvisor and Expedia and 800 deceptive reviews through Amazon’s Mechanical Turk. The truthful or deceptive reviews are further divided, featuring 400 positive and 400 negative reviews. This balance ensures a comprehensive assessment of classification models across different types of reviews.
Apart from this, we also extended our experiments to the widely used YelpZip restaurant review dataset, which contains over 600,000 reviews written by more than 260,000 users. Given the significant class imbalance in the original dataset (with genuine reviews far outnumbering fake ones, and positive reviews dominating negative ones), we followed the construction method of the DOSC dataset and sampled 1,600 reviews, ensuring an equal split between genuine and fake reviews, as well as positive and negative sentiments.
Adaptation Strategy
When encountering novel datasets, particularly those involving different languages, cultures, or domains, traditional methods often struggle to adapt autonomously. For instance, lexicon-based approaches require the reconstruction of language- and domain-specific dictionaries, while conventional machine learning or deep learning models demand large-scale retraining on new corpora. In contrast, the GPT-based MAE module within our proposed AIM-RMD framework is designed to effectively handle such challenges. By leveraging appropriate prompt engineering techniques, such as in-context learning and Chain-of-Thought (CoT), it achieves adaptive performance across languages and knowledge contexts with minimal migration cost, eliminating the need for cumbersome fine-tuning.
Specifically, within the second module of the artifact, we designed a CoT prompt component (as illustrated in Figure 3). This component guides the model to proactively reason about the linguistic and cultural context of the review before performing affective analysis—much like how a human would implicitly acclimatize to a different cultural and linguistic setting. This “pre-thinking” step enhances the model’s contextual awareness and cross-cultural adaptability. We followed the advice of Cao et al. (2023) to craft prompts in the target language and embedded the cultural context of the language that GPT will encounter into the prompts. This approach enables GPT to adapt to the cultural nuances of the language and handle emotional expressions appropriately across different linguistic and cultural settings.
When processing reviews from diverse tourism sectors, we explicitly defined GPT-3.5’s “expert identity” by specifying its role as, for example, a “world-class expert in luxury hospitality” or a “specialist in budget travel,” as research has demonstrated that such persona assignment can markedly improve the model’s accuracy and response quality in sentiment analysis tasks (Amin et al., 2023). Furthermore, we employed the in-context learning technique by providing two carefully curated demonstration cases, enabling the model to automatically learn through analogy and generalize from the given examples.
Baselines
For hotel reviews from other countries, platforms, and online reviews from the restaurant domain, we have chosen to compare the results of our approach with the baselines. These studies focus on either feature extraction or the use of more complex deep learning models, and they include the latest state-of-the-art (SOTA) results from the past year. Comparing our method with these studies provides a credible and comprehensive evaluation, as it benchmarks our approach against the most current and advanced techniques in the field.
Results
Cross-Cultural Performance
The results and comparisons on the DOSC English hotel reviews data are presented in Table 5, our method outperforms these benchmarks in terms of accuracy and precision, and is on par with them in metric recall. This suggests our approach is better at capturing the manipulated reviews without missing many cases. The comprehensive performance indicator, metric F1-score, also shows significant improvement, indicating that your model more effectively identifies true positive cases (correctly classified manipulated reviews).
Validation of Review Manipulation Detection Performance on the Gold-Standard Chicago Hotels Dataset DOSC.
Note. Due to varying metrics used across different studies, only common metrics are retained to ensure comparability. The results listed correspond to the same training and testing set splits as used in this study. The “Avg. Time per 1,000 Samples (s)” represents the average inference time for processing 1,000 reviews.
Domain Adaptation Performance
As shown in Table 6, the findings demonstrate that our method consistently outperforms all previous benchmark approaches across multiple key metrics, further validating its robustness and cross-domain applicability. Specifically, when evaluated on the same Yelp restaurant review dataset for deceptive review detection, our approach achieved an accuracy of 0.8438 and an AUC score of 0.8413, surpassing all compared methods including that of Hyder et al. (2024) (accuracy: 0.8108), Mohawesh et al. (2023) (precision: 0.8318, recall: 0.7786), and A. Kumar et al. (2022) (F1-score: 0.8380), indicating a more reliable and balanced performance in identifying manipulated content with reduced false positives and improved harmonic mean between precision and recall and validating the adaptability and applicability of our approach across different domains.
Generalizability Test on Yelp Restaurant Reviews With Benchmark Models.
Note. Due to varying metrics used across different studies, only common metrics are retained to ensure comparability.
Conclusions
The accuracy of review manipulation detection (RMD) in the tourism industry is critical for maintaining trust and ensuring the reliability of online reviews. While existing methods have made strides, they fall short in recognizing the rich affective clues in review manipulation. This paper proposes an Affective Iceberg Model (AIM) for conceptualizing these affective clues, which categorizes the emotional features of deception into a visible “above-water” component (surface-level sentiments) and a submerged, latent component (fine-grained emotions). Based on this model, we designed a GPT-based Multidimensional Affective Extraction (MAE) framework to capture deeper and more nuanced affective signals indicative of manipulation. The extracted features are then integrated with Bag-of-Words (BoW) representations and processed through machine learning classifiers to identify suspicious reviews. Experimental results demonstrate that the proposed AIM-RMD artifact achieves higher accuracy in detecting manipulated content across both Chinese and English hotel and restaurant reviews compared to current state-of-the-art benchmarks. Further internal module analysis reveals a distinctive pattern in the nuanced emotional expression of genuine reviews compared to manipulated ones: specifically, the intensity distribution of fine-grained sentiments in authentic reviews exhibits a multi-peak pattern, as opposed to the strong uni-modal distribution observed in fake reviews. Moreover, genuine reviews cover a significantly wider variety of fine-grained sentiments (avg. 3.38) than their deceptive counterparts (avg. 2.07). This work not only provides a novel affective perspective for identifying fake reviews, but also offers a scalable and interpretable framework for future research in cross-lingual and cross-domain deception detection. It also deepens the theoretical understanding of emotional features in review manipulation and offers new insights into deception mechanisms.
Contributions
This study makes three direct contributions. First, recognizing that reviews in the hospitality and tourism industries often convey complex and multifaceted emotions that are frequently overlooked in existing manipulation detection approaches, we developed the Affective Iceberg Model (AIM) as a metaphorical representation of the underlying emotional richness in user-generated reviews. Second, we designed a GPT-based module to analyze and quantify the fine-grained emotions corresponding to the submerged part of the iceberg. The effectiveness of this model was empirically estimated and evaluated using machine learning classifiers, resulting in improved accuracy in detecting review manipulation. This contribution aligns with the design science research paradigm as advocated by Gregor and Hevner (2013), fulfilling the criterion of “improvement” through the introduction of an innovative and actionable artifact.
Second, to evaluate the effectiveness of the AIM-RMD artifact, we conducted two sets of experiments. The first experiment focused on depth and interpretability using a Chinese hotel review dataset. It not only demonstrated the superiority of AIM-RMD over baseline methods in the final task of review manipulation detection but also revealed, through detailed output analysis at each stage, how fine-grained emotional features and emotional diversity contribute to distinguishing genuine from fake reviews. These findings offer deeper insights into the complex affective patterns underlying review manipulation and enhance theoretical understanding of how emotions are expressed and manipulated in online reviews. The second experiment emphasized breadth and generalizability, testing the artifact’s performance on English hotel and restaurant reviews across different languages and tourism-related domains. The results confirmed that AIM-RMD outperforms recently published state-of-the-art models in detecting fake reviews. Together, these experiments reflect our contribution, as suggested by Hevner et al. (2004), to designing domain-aware and developing highly adaptable and universal artifacts for practical deception detection.
Third, in quantifying the complex emotions residing beneath the surface of the iceberg, this study constructed a structured GPT prompt template. This template offers a feasible and comprehensive methodological framework for a wider range of complex text-processing tasks using GPT and other generative AI models. The incorporation of Chain-of-Thought (CoT) reasoning and iterative feedback mechanisms not only enhances the interpretability and trustworthiness of such otherwise opaque “black-box” systems (Mitchell et al., 2023), but also enables human intelligence to more effectively guide and improve AI-driven processes. This approach reflects our contribution, which is in line with Abbasi et al.’s (2024) vision toward developing more interactive, human-guided, and user-centric AI systems.
Practical and Managerial Implications
This study holds significance in offering actionable tourism management practices and policy formulations. Firstly, for hoteliers and travel operators, integrating GPT-based sentiment analysis into review monitoring systems can significantly improve the accuracy of identifying authentic and deceptive reviews. Also, for travel agencies, the study’s focus on the detection of emotional manipulation in reviews can assist in combating the spread of misinformation and fraudulent practices in tourism behaviors. By filtering out deceptive reviews, tourism platforms can present a more authentic picture of traveler experiences, crucial for building consumer trust and confidence in online review platforms. For managers, the ability to distinguish nuanced emotions in reviews provides a richer understanding of customer experiences. It enables targeted service improvements, personalized marketing campaigns, and the enhancement of customer loyalty through responsive customer service. The fine-grained analysis can also help in identifying emerging trends and preferences among travelers, which is invaluable for product development and service innovation. From a policy perspective, the findings underscore the importance of adopting advanced GAI technologies in maintaining the integrity of online review systems. Policymakers can leverage these insights to develop regulations that promote transparency and trust in online reviews, ultimately protecting consumers and fostering fair competition among tourism businesses.
Limitations and Future Research Directions
While this study provides valuable insights into the detection of fake hotel reviews through building an Affective Iceberg Model for Review Manipulation Detection (AIM-RMD), several limitations should be acknowledged. First, as depicted in our AIM, the emotional cues related to manipulation can be largely latent and deeply buried, the use of Ekman’s six basic emotions in this study captures only one dimension of this complex affective landscape, whereas the full spectrum of emotions experienced during a hotel stay includes more nuanced sentiments such as satisfaction, gratitude, disappointment, confusion, etc. Future work could explore additional emotional cues in fake reviews to improve detection. Second, the second GPT-based module within the proposed AIM-RMD artifact still presents several limitations, leaving room for further improvement. Currently, this component can struggle with distinguishing between similar emotions, lack logical reasoning capabilities, and overlook implicit sentiments sometimes. Future research could address these issues by refining prompt engineering, utilizing more advanced model versions, and integrating human review processes to enhance accuracy. Third, beyond emotional expression, fake reviews may exhibit other features, such as factual inconsistencies or patterns in reviewer behavior and social network characteristics. Future research should investigate these additional markers of review manipulation to develop more comprehensive detection strategies. Lastly, this paper relies on the platform’s filtering algorithms to generate labels for review manipulation. Although similar methods have been widely adopted in existing literature, future research could enhance the accuracy of these labels by combining multiple detection methods, such as manual verification and user behavior analysis.
Footnotes
Author Contributions
Shuihan Liu: Conceptualization; Data curation; Formal analysis; Writing—original draft; Writing—review & editing. Yunjie Wei: Conceptualization; Formal analysis; Validation; Writing—original draft; Writing—review & editing. Shouyang Wang: Formal analysis; Supervision; Validation; Writing—review & editing.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was partly supported by the National Natural Science Foundation of China under Grants Nos. 72171223 and 71988101 and the Youth Innovation Promotion Association Chinese Academy of Sciences.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.