Identification of tourists’ dynamic risk perception—the situation in Tibet

The study addresses how tourists’ risk perceptions of a destination change dynamically from pre-travel (formed via indirect information and prior knowledge) to post-travel (revised through direct, first-hand experience). Grounded in risk communication theory, which posits that risk perceptions evolve as information accumulates, the paper highlights that static assessments (common in prior work) may misrepresent true perceptions that shift across the travel process. The context is Tibet, a destination with strong appeal and distinctive risk characteristics (high altitude, complex terrain, and cultural specificity). The purpose is to (1) identify types and salience of perceived risks before and after travel, (2) portray their dynamics using large-scale online textual data (Q&As and travel notes), and (3) provide implications for guiding tourists toward more accurate risk assessments and for destination managers to adjust image promotion and risk management. The study is important because pre-travel overestimation can deter trips, while underestimation can endanger safety and reduce satisfaction; understanding the dynamics supports better decision-making and management.

The paper reviews four strands of research on tourists’ risk perception identification: (1) Destination-focused studies identifying types and degrees of perceived risks at specific destinations (e.g., physical, financial, performance, socio-psychological, time), often before, during, or after travel, and sometimes classifying risks into controllable vs. uncontrollable. (2) Event-focused studies examining perceptions of specific risk events (terrorism, natural disasters, environmental risks like air pollution) and their impacts on destination image and revisit intentions. (3) Special event and e-commerce contexts, including sporting mega-events and mobile booking risks; recent COVID-19 studies show heightened safety, health, and cleanliness risk perceptions. (4) Social and individual differences perspectives analyzing effects of gender, experience, motivation, and culture; however, most samples are cross-sectional and static. Few dynamic studies compare pre- and post-travel perceptions (e.g., health risk reassessments), but these typically rely on questionnaires/interviews with small samples and predefined categories, risking high costs, sample bias, and incomplete risk coverage. The review underscores the need for scalable, dynamic identification using broader textual data to capture the full spectrum of perceived risks and their evolution.

Data and approach: The study uses two large-scale textual datasets from Ctrip: 3,117 Q&As (cleaned to 2,627 Tibet-specific questions posted pre-travel) and 17,523 travel notes (post-travel). The design leverages methodological complementarity: short Q&As with fewer risks per record are manually multilabel-annotated to identify a comprehensive set of pre-travel risk categories; lengthy, multi-topic travel notes are processed via a thematic dictionary and sentiment analysis to identify post-travel risk mentions and their emotional valence. Pre-travel (naive) risk perception identification via manual labeling: Two trained annotators conducted multilabel classification of each Q&A, iteratively defining and updating risk categories. Steps: (1) Briefing on objectives, (2) previewing samples, agreeing on multilabel protocol and initial categories (e.g., accommodation, transportation, routes, safety, religious culture), (3) annotating each Q&A with multiple risk labels, updating the category list as new risks emerged, (4) computing inter-annotator agreement (kappa index = 0.8903) and resolving discrepancies, and (5) computing category frequencies to reflect perceived importance. Post-travel (revised) risk perception identification via dictionary and sentiment analysis: Five steps: (1) Construct a thematic risk dictionary seeded by the finalized Q&A risk list and preview of travel notes (e.g., health risk keywords such as "altitude stress"). (2) Segment travel notes into sentences (using punctuation) and identify risk mentions by matching dictionary keywords at sentence level. (3) Validate and iteratively refine keywords using Precision and Recall; low precision prompted keyword specificity improvements; low recall prompted adding missing terms. After five rounds, both metrics stabilized at high levels; the dictionary was finalized by round seven. (4) Measure risk perception frequency at the document level (a travel note counts as 1 for a risk if any sentence mentions it, regardless of multiple mentions). (5) Conduct sentiment analysis on risk-related sentences using five pretrained models from Baidu Senta (LSTM, BiLSTM, CNN, GRU, BOW) to estimate negative sentiment probabilities for each risk; compute average negative sentiment per risk across mentions to approximate post-travel perceived importance (frequency × mean negative sentiment). Comparative analysis: Compare pre-travel importance (frequency in Q&As) with post-travel importance (frequency × negative sentiment) to classify risks into increased, decreased, or unchanged importance after travel, revealing under- or overestimation pre-travel.

Samples and reliability: 2,627 Tibet-specific Q&As (Aug 2008–Apr 2021) and 17,523 travel notes were analyzed. Manual Q&A labeling achieved high agreement (kappa = 0.8903). The dictionary approach for travel notes achieved high, stable Precision and Recall after iterative refinement. Pre-travel perceived risks (20 categories; top frequencies): Route selection (23.27%, 907), Traffic/transportation (17.01%, 663), Expense (9.75%, 380), Equipment (7.80%, 304), Season (7.70%, 300), Entry procedures (5.23%, 204), Time (4.82%, 188), Climate (4.54%, 177), Health (4.00%, 156), Accommodation (3.69%, 144). Less frequent: Security (2.49%, 97), Tickets (2.36%, 92), Infrastructure (2.05%, 80), Travel agency selection (1.77%, 69), Openness (1.39%, 54), Dining & shopping (0.85%, 33), Traditional customs (0.64%, 25), Epidemic policy (0.59%, 17), Communication (0.44%, 8), Other. Post-travel negative sentiment (five-model consensus): Security showed the highest negative sentiment probability (≈0.48–0.54), followed by Health and Infrastructure (≈0.39–0.45), and Time (≈0.34–0.39). Lower negative sentiment included Season (≈0.20–0.23) and Traditional customs (≈0.18–0.22). Four-quadrant analysis (frequency vs. negative sentiment) and comparative dynamics:

• Increased importance after travel (pre-travel underestimation): Health, Accommodation, Time, Security, Infrastructure, Dining & shopping—these were more frequently perceived post-travel and carried higher negative sentiment (e.g., narratives of severe altitude sickness, poor/unsafe road conditions, accidents, and basic service shortfalls).
• Decreased importance after travel (pre-travel overestimation): Traffic/transportation, Route selection, Equipment, Season, Entry procedures—concerns eased after direct experience (e.g., convenience of car rental, appreciation of seasonal diversity).
• Relatively unchanged: Expense and Climate remained strongly perceived before and after travel; Traditional customs, Communication, Openness of attractions, Travel agency selection, and Epidemic policy remained weakly perceived (note low epidemic salience reflects data time span and travel restrictions). Synthesis: Before travel, tourists tend to underestimate safety, health, and time-related risks while overestimating transportation, route selection, and season-related risks. Post-travel, risks tied to safety/health/infrastructure and service availability emerge as critical to satisfaction. The findings support risk amplification/attenuation processes and optimism bias effects.

The findings directly address the research goal of portraying dynamic risk perceptions by demonstrating significant shifts from naive (pre-travel) to revised (post-travel) assessments. Pre-travel overestimation of route/season/transport risks can deter visitation, while underestimation of safety, health, time, and infrastructure risks can endanger tourists and depress satisfaction. Results align with risk communication theory and the social amplification of risk framework: information sources (media, reviews) shape initial perceptions, while direct experience amplifies or attenuates specific risks. Optimism bias likely contributes to underestimation of personal susceptibility to altitude sickness and road hazards. Managerial implications:

• For risks that intensify post-travel (health, safety, infrastructure, dining/accommodation, time), destinations should strengthen preventive measures (e.g., altitude acclimatization guidance, road safety management), improve basic services and infrastructure, and communicate clear, actionable risk information to set realistic expectations.
• For risks that attenuate post-travel (season, route, transport, equipment, entry), targeted pre-travel communications (using visuals, infographics, and narratives rather than numeric probabilities alone) can reduce undue anxiety and barriers to travel.
• Two-way risk communication, leveraging both official channels and user-generated content, can calibrate expectations and enhance satisfaction. For consistently salient risks (expense, climate), optimize pricing transparency and educate visitors on climate preparedness; for consistently weak risks, improvements (e.g., cultural promotion, communications infrastructure) can provide incremental satisfaction gains.

The paper introduces a scalable framework combining pre-travel Q&As and post-travel notes with text mining (manual multilabeling, thematic dictionaries, and deep-learning sentiment analysis) to capture the dynamics of tourists’ risk perceptions. Using Tibet, it shows that perceived importance of health and safety risks increases after travel, while season/route/transport concerns decline; expense and climate remain persistently salient, and several other risks remain peripheral. These insights help managers tailor risk communication and management strategies to correct pre-travel misperceptions and enhance satisfaction. Future research can test generalizability across destinations, incorporate finer-grained factors (e.g., gender, experience), and enrich temporal coverage (e.g., during public health crises) to refine understanding of dynamic risk trajectories.

Generalizability may be limited by Tibet’s unique geographical, cultural, and infrastructural context. COVID-19-related perceptions are underrepresented due to data gaps from travel restrictions and the sampling window. The approach relies on keyword dictionaries and pretrained sentiment models, which, despite validation, may miss nuanced context. Future work should include subgroup analyses (e.g., gender differences), broader temporal coverage, and cross-destination comparisons.