The research data in this paper are mainly taken from the raw data of online travel platforms and microblogs. To convert these raw data into a language that can be recognized by computers, a systematic pre-processing work has to be performed. This specifically includes denoising, normalizing the representation, and word separation.

Assuming that the vocabulary of the current evaluation dimension consists of i characters, match the first i characters of the current string. It also matches the i characters after the evaluation dimension (i values are 2–10). If this i character is the evaluation vocabulary, the field is sliced out and corresponds to the previous evaluation dimension, thus forming the initial correspondence data.

When the corresponding data are formed, the vectorization process begins. The model can also be represented by the mathematical formula (Minaee et al., 2021):

where D is the set of feature terms, t_n is each feature term, and W_n is the weight corresponding to each feature term.

The specific process of data vectorization is divided into two parts: data vectorization and error tolerance (Hartmann et al., 2019).

The core of the SVM algorithm is to find the support vector and the optimal hyperplane (Varatharajan et al., 2018; Zeng et al., 2018; Hu et al., 2020), so the key problem is to find the interval threshold between the vector corresponding to the data points and the support vector. Many linearly indistinguishable scenarios often occur in practice. Therefore, the selection of the kernel function is a key for the linearly indistinguishable problem. Since the data volume of travel online reviews is too large and the dimensionality is relatively low, the selection of polynomial type kernel function is particularly suitable.

For the case of linear indivisibility, the innovation of this paper is to improve the operation method by expanding the point vector in the SVM classifier into individual sub-vectors and then substituting them into the polynomial in turn. The original polynomial kernel function is as:

The SVM classifier decision function when linearly indistinguishable is as:

where x is the vector to be classified, x_i is the support vector, and x is an n-dimensional vector in the calculation of this formula.

It is obvious that the total time complexity is as:

It can be seen that n inner product calculations are performed during the operation and the number of times is d.

Therefore, the computational effort of the improved SVM algorithm is not related to the number of support vectors but to the spatial dimensionality of the study context and the order of the kernel function in the polynomial. The use of the improved SVM algorithm is particularly effective for the case of this paper where the number of vectors is large and the dimensionality and order are small and low. When computing the vectors, there is no effect on the selection of the classification results and the calculation of the threshold value because the spreading vectors of the point vectors are discarded. Only the normal vectors need to be substituted into the computation to perform the classification, so that the efficiency of the algorithm is greatly improved. The penalty function in the linearly indistinguishable case is as:

where C is a free parameter, and based on practical experience, C is set to 2 in this study. R is the number of vectors.

Another important issue, the choice of the order d-value has a great impact on the accuracy and time complexity of the kernel function. Regarding the effect of different d-value choices on the recognition rate of the algorithm, it is illustrated here by the experimental results shown in Table 3.

In this paper, the data dimension is 3 and the number of training samples is 3,776 data points. The polynomial type kernel function is used, and the parameter C is 2. From Table 3, we can see that the data recognition rate is highest when d = 2, so the d of the polynomial kernel function is 2 in this study.

When deploying an HDFS storage structure on a Hadoop platform, a Master is used as the host, which is responsible for the NameNode and JobTracker. Each Slave usually has a DateNode for storing data information and its backups, and executes Map tasks and Reduce tasks in conjunction with a local TaskTraker according to the application requirements. The design idea of the Map-Reduce model is shown in Figure 2.

The Map-Reduce framework treats the input of an application as a set of <key,{list of value} > keys and the <key,{list of value} > keys input to the Map function are raw vector form data, where key represents the evaluation dimension weights and {list of value} corresponds to the “evaluation vocabulary” and “positive/negative meaning,” respectively. When <key,value> is derived from the Map, it comes to the Reduce class, and the Reduce task starts to process the healthy values scattered in many different nodes. The Reduce function saves the classification results of each data node into different classifiers and counts the number and percentage of nodes in each classifier.

The PC environment used for the test experiment is Intel I7 processor, 16G RAM, Windows 10 operating system, Eclipse, and Matlab. The cooperation with an online tourism platform in terms of big data was conducted and the online review data were captured using the big data cloud computing service platform owned by it. In 2019, the total tourism revenue of Hangzhou city was 219.741 billion yuan, an increase of 12.72% year-on-year. This study used this platform to capture online review data of Hangzhou West Lake Scenic Area in 2019. There are two main types of these data: (1) microblog data, such as Sina Weibo and Tencent Weibo; (2) OTA online review data, such as Ctrip.com, GoWhere.com, and Tongcheng.com. The collected raw data are pre-processed, among which the most important work is to manually classify these data, and the obtained division results yield the dataset shown in Table 4.

Travel sentiment classification can be summarized as a matter of binary classification. Those that contain negative evaluations are not recommended reviews and those that do not contain negative evaluations are recommended reviews. In general, there are two major dimensions for the evaluation criteria of classification models, one is the accuracy of classification, and the other is the efficiency of classification. In this paper, we will compare the travel sentiment classification models from these two perspectives. The evaluation metrics of classification accuracy use recall, correctness, and precision. Among them, correctness is an evaluation of the degree of classification of correct results in the classification results; precision is the percentage of each correct classification target in its category; and recall is a metric to evaluate the proportion of the category of recall targets. These evaluation metrics are calculated as follows:

To verify the superiority of the improved SVM algorithm with respect to the original algorithm, two aspects of accuracy and time consumption are verified separately.

Figure 3 represents the comparison of the result accuracy of the original SVM and the improved SVM algorithm. It can be seen that the original SVM algorithm is higher at small data size. However, when the data volume grows to more than 600 G, the accuracy gap between the two algorithms gradually decreases. When the data volume grows to more than 1 T, the accuracy rate of the original SVM algorithm starts to decline, while the accuracy rate of the improved SVM algorithm remains around 80%. It can be inferred that the accuracy of the improved SVM algorithm has been well stabilized in the context of big data.

Figure 4 shows the result accuracy of the improved SVM algorithm in the Hadoop system. It can be seen that after the introduction of Hadoop platform, the result accuracy of the algorithm has improved, especially when dealing with large data. From Figure 4, we can see that the accuracy rate is maintained at about 80%, which is basically in line with the actual application requirements. A comparison of the time consumption of the improved SVM algorithm and the original SVM algorithm is shown in Figure 5.

From Figure 5, it is clear that when the amount of data is small, there is no significant advantage of the improved SVM algorithm after decomposing the vectors due to the limited number of both point vectors and support vectors. Therefore, there will not be much difference between the two in terms of time cost. However, when the data size gradually increases, the improved SVM algorithm has a great advantage in the computation of kernel function due to the linear indistinguishability gradually increases, especially when the data size is more than 20 GB, this advantage is especially obvious. Overall, in the context of massive data, the improved SVM algorithm greatly improves the execution efficiency of the algorithm while ensuring a certain accuracy of the results.

Sentiment analysis models based on SVM, plain Bayes and dictionary rules were continuously trained and optimized according to five major measures for assessing model accuracy. The classification results of different models at the same test set are shown in Table 5.

The results of the accuracy analysis of these three models are shown in Figure 6. From the results, it can be seen that the commonality of the three is that the accuracy rate is high, above 83%, and all of them can achieve the accurate management of tourism emotion. However, there are still some differences among the three. Compared with the model based on sentiment dictionary rules, the classification accuracy of the machine learning-based model is more stable, and the improved SVM classifier is generally better than the plain Bayesian classifier. Among them, the classification accuracy of the dictionary model is higher than that of the machine learning classification when the number of test sets is small, but it gradually decreases as the number of test sets increases.

In summary, the travel sentiment classification model based on sentiment dictionary has better classification results when dealing with a small amount of text. However, as the number of test sets increases, the improved support vector machine-based travel sentiment classification model is able to achieve more desirable classification results. Therefore, the improved support vector machine-based sentiment analysis method has great superiority in dealing with large amount of travel review data.