Language model is an important part of natural language processing (NLP), and the word vector depends on the training of language model. Word vector method involves two important models: CBOW model and skip-grams model. CBOW model predicts wt under the premise of known w_t-2, w_t-1, w_t + 1, w_t + 2, i.e., the frequency of occurrence of the current word is predicted by context. Skip-grams model is just the opposite of CBOW model, which is to predict w_t-2, w_t-1, w_t + 1, w_t + 2 on the premise of known w_t (Ning and Liu, 2016).

This study uses Word vector method to find tourism risk perception dimensions. The word vector is constructed by using Python Gensim toolkit and the skip-grams model based on Hierarchical Softmax, as shown in Figure 1. The input layer contains only one word w, and the word vector v (w) ∈Rm; it projects v(w) onto v(w) by means of the projection layer. The projection is identical. In skip-grams model, the projection layer is actually redundant. The output layer is a Huffman tree. The word vector of the root node of Huffman tree is the word vector that is mapped by the projection layer. All the corresponding leaf nodes correspond to each word in the training corpus and the number of leaf nodes represents the number of words in the training corpus.

After separating the words, just keep nouns; use blanks to separate each noun, and output word segmentation results to form a noun corpus.

The word frequency of the noun corpus is counted, and the top 10 high-frequency words are shown in Table 4. By comparing the word frequency of questions and answers in Table 4, it is found that seven words are the same, such as exit, problem, entry, passport, Thailand, airport, and air ticket.

TF-IDF method is used to extract keywords from the noun corpus, and the top 10 high-frequency keywords are shown in Table 5. As can be seen from the results of word frequency statistics and keyword extraction, the high-frequency words are almost the same as the keywords, with only a slight difference in order.

The vector dimension is 200, the size of the training window is 5, the skip-gram algorithm based on Hierarchical Softmax is adopted, the frequency of minimum word is 5, the number of iterations is 5, and the thread is 25. The word vector training is carried out to obtain the vector representation of all the words in the corpus and a word vector model. For example, the word vector representation of the word “exit” in the “question” section is −0.0046794955–0.14208734 0.12448783.

In addition, the cosine similarity is used to calculate the words that are similar to each keyword. The first five similar words of some keywords are shown in Table 6. It can be seen from Table 6 that the first five words similar to the keyword “exit” in the question are Shanghai, Shenzhen, San Francisco, Chicago and take a plane.

According to the steps of cluster analysis of K-means algorithm, the elbow rule is adopted to determine the cluster number (K value). According to the elbow rule, the accuracy of each sample classification increases with the increase of K value. The clustering effect is closely related to the sum of squares of errors (SEE) value, which means, when the SSE is smaller, the aggregate effect of each cluster is better. As a result, the relationship between the value of k and the sum of the squares of the error shows a shape similar to that of a bent elbow. The best value is K at the elbow. The analysis results are shown in Figure 2. “Q & A” in the Word vector method, k = 7. When k = 7, K-means clustering is performed on the keywords in the training model using Python GenSim library, and 7 clusters were obtained. TF-IDF method is used to obtain the keywords of each cluster. See Table 7 for some keywords in question clustering and Table 7 for some keywords in answer clustering.

From Table 7, the first category mainly revolved around airport related problems, corresponding to functional problems such as airport traffic and duty-free shopping. The second category is mainly service, mostly ticketing and visa issues. Both the first and second categories have traffic problems, and words related to flight precautions such as airport, consignment, and transfer account for a large proportion. It can be seen that because the plane is the main means of transportation out of the country, users account for a large proportion of questions about flights. The third category is financial, which is mostly related to expenses and includes guarantee fund, service fee, bank, and other keywords. The fourth category is similar to the third category. The key word is preferential treatment. Generally, outbound travel costs are high, and tourists are more sensitive to preferential information. The fifth category is the package tour, which is related to service. Since the risk of overseas tourism is much higher than that of domestic tourism and there are too many unknown factors, most tourists choose to travel in groups. There are many words related to group travel, such as leader, group and partner. The sixth category is tourism destination planning. Tourists have different preferences for the choice of tourist destinations. Tourists tend to choose Europe, Southeast Asia and other places as tourist destinations, and this category also includes destination visa issues. The seventh category has poor clustering effect, mainly including keywords related to trip and route planning, which belongs to planning risk.

From Table 8, the first category is related to planning and mostly answers to the choice of tourist destinations. The second category is related to the function, which is the answer to the questions about the tourist attractions and the project, including the introduction of various tourist projects and tourist attractions. The third category is similar to the fourth category, which is about traffic questions. The answers mainly focus on the matters needing attention in air travel, and the whole process of air travel is reflected in the answers. There are also a small number of functional responses. Tourist satisfaction can be improved by developing a detailed marketing strategy for tourist commodities. This category not only includes duty-free goods, security checks, and other airport duty-free shop commodity trading keywords, but also includes souvenirs, stores, and other overseas local souvenirs to buy. The fifth category is mainly about communication, finance, etc., which not only gives a detailed introduction to overseas communication, but also gives an introduction to overseas currency exchange and local currency circulation. At the same time, there are corresponding answers to questions about foreign currency and exchange rate conversion. There are also answers to service-oriented questions such as visas, and materials. The sixth category is mainly service problems. Tourists answered more questions about travel restrictions such as visas, passports, and information. The seventh category is communication problems, in which the customs and safety of tourist destinations are the issues that tourists pay more attention to. Tourists’ responses to such questions were comprehensive, and they also shared many travel experiences, such as route planning and hotel accommodations.

Based on the clustering results of questions and answers, Chinese tourists’ risk perception of outbound tourism can be mainly divided into traffic risk, planning risk, service risk, communication risk, financial risk, and functional risk. See Figure 3 for attribute classification.

It can be seen from the clustering results that the clustering of questions and answers does not reach the optimal effect, and there are problems such as category repetition and crossover.

Evaluation of the model accuracy. The accuracy of the “Q & A” model before feature extraction are 59.18% (question) and 17.61% (answer) respectively. After feature extraction, the accuracy of the “Q & A” models are 70.69 and 76.69%, respectively. It shows that after the features are extracted, the accuracy of the model has been improved.

The word vector method is used in this paper, with “outbound travel” as the key word, to crawl “Q & A” data from online travel operator platforms such as Ctrip.com and Tuniu.com, and then python is used to preprocess the data to form an initial corpus. Finally, the k-means method is used to cluster the data, and six outbound tourism risk perception dimensions are identified, which are traffic risk, planning risk, service risk, communication risk, financial risk and functional risk. It can be seen that there are some differences between our results and findings of previous researches. (1) Chinese tourists have the highest perception of traffic risk among the six outbound tourism risk perception dimensions discovered in the study, such as transportation issues such as boarding, transfers, and road safety. It is similar to the findings that flight delays and cancellations caused by airline operations are medium-high frequency risk events for Chinese tourists to travel abroad (Huang et al., 2022). (2) Chinese tourists pay more attention to service risks when travelling broad, such as scenic spots, hotels, and group tours in tourist destinations, and this result also applies to the study of Malaysian students’ risk perceptions when travelling to India (Khan et al., 2019), because low expectation about service quality develops a high perception of performance risk that negatively influences the purchase intention (Wood and Scheer, 1996; Garretson and Clow, 1999). (3) Financial risk refers to the possibility of not getting a corresponding return on the money spent in the process of travel, or the possibility of losing property (Roehl and Fesenmaier, 1992). Tourists are concerned about issues such as the protection of overseas property and consumption methods, as well as factors that influence tourist destinations’ itinerary. (4) Chinese tourists pay attention to their travel plan, that is, they make plans in advance when travelling abroad. Roselius (1971) argued that when a product does not perform according to expectations, people waste their precious time and convenience. (5) Chinese tourists are concerned about functional risk when traveling abroad, the objective factors affecting tourism safety perception include the social and natural environment in tourist destinations and the security situation of “food, housing, transportation, travel, shopping, entertainment” in the process of travel (Reisinger and Mavondo, 2005; Yüksel and Yüksel, 2007), and functional risk includes these aspects. (6) Chinese tourists have communication risk but the level of perception is not high, especially in group tours. On the one hand, due to the high quality of tour guides, the risk of communication is seldom paid attention to by tourists. On the other hand, experiencing exotic scenery is one of the motivations for Chinese tourists to travel abroad. Studies have proved that the culture of the destination country, the individual culture of the tourists, the culture of the destination country and the difference between the culture of the source country and the culture of the destination country are the cultural factors that affect the tourists’ behavior (Ng et al., 2007),

The classic outbound tourism risk perception dimensions are based on the integration of the research of Roehl and Fesenmaier (1992), Lepp and Gibson (2003), Lepp et al. (2011), including 9 dimensions of hardware facilities, money and finance, physical health, psychological reasoning, expectation satisfaction, social relations, time cost, political events, and cultural differences and obstacles. Comparing our research results with the 9 dimensions, it is found that Chinese tourists do not consider the risks such as physical health, time cost, and political events.

First, Chinese outbound tourists have insufficient perception of physical health risks. The possible reason for this phenomenon is that the main destinations of Chinese tourists’ outbound travel are Thailand, Japan, the United States, the Maldives, Canada, the Netherlands, Australia and other countries. In the perception of Chinese tourists, these countries are generally safe. However, even one major incident can malign the image of a destination (Yang et al., 2017). Chinese visitors have no perception of physical risk does not mean that these nations and areas should not emphasis on providing a better environment in terms of safety and security. Therefore, providing physical safety and security is a must to sustain the growth of tourism of a destination. Second, the time cost is not within the risk perception of Chinese outbound tourists. This situation is likely to be that many Chinese outbound tourists are people who are retired or have not entered the workplace. This group of people has a lot of free time and generally does not consider the cost of time. According to the 2019 China Cross-Border Travel Consumption Report jointly released by Ctrip.com and Mastercard, the age distribution of Chinese outbound tourists is mainly located in the 50s and 60s (retirees) and 90s and 00s (most of them did not enter the workplace) (Luo, 2019). Third, Chinese outbound tourists do not attain a good understanding of the risks of political events. The reason for this is that tourists actively avoid countries and regions with high risk of political incidents, such as Mali, Libya, Syria, Afghanistan, etc., when choosing a destination.

Chinese tourists lack awareness of risks such as physical health, time costs, and political events. In view of risk perception of Chinese outbound tourists, relevant Chinese government agencies and enterprises ought to take the following measures: