Tourism review sentiment prediction method and system based on knowledge graph and multi-modal feature fusion

Abstract

This invention provides a method and system for predicting tourism review sentiment based on knowledge graph and multimodal feature fusion, relating to the field of natural language processing. The method includes: collecting multiple tourism reviews from a target region; extracting a comprehensive feature matrix from the multiple tourism reviews based on LDA and TF-IDF; extracting context-enhanced features from the multiple tourism reviews based on a knowledge graph; filtering the comprehensive feature matrix and context-enhanced features of the multiple tourism reviews to select multiple optimal features; and performing sentiment prediction of the reviews in the target region based on multiple optimal features using multiple models. This method has the advantages of improving the accuracy of tourism review sentiment prediction and further providing scientific support for smart tourism services.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and in particular to a method and system for predicting sentiment in tourism reviews based on the fusion of knowledge graphs and multimodal features. Background Technology

[0002] With the rapid development of internet technology, user-generated content has experienced explosive growth. In the tourism sector, travel reviews, as a crucial component of user-generated content, contain a wealth of genuine tourist experiences and feedback, becoming a key data source for reflecting tourist experiences and optimizing service quality. How to extract valuable information from massive amounts of travel review data and use this information to build an intelligent tourism service system has become a research hotspot of common interest in both the tourism industry and academia.

[0003] Negative sentiment in travel reviews is often complex and diverse, potentially involving varying degrees of dissatisfaction with tourist attractions, services, and facilities. Existing technologies mostly employ relatively simple sentiment analysis methods, which struggle to accurately interpret complex negative sentiments. For example, some analysis methods based on word frequency statistics and simple sentiment dictionaries can only determine whether a review is positive or negative, failing to delve into the specific direction and intensity of negative sentiment. This makes it difficult to implement effective improvement measures targeting different types of negative sentiment when optimizing tourism services, reducing the efficiency of service optimization. Tourists' emotional expressions are often closely related to the semantic information of attractions; for instance, tourists' evaluations of a site's historical and cultural value, natural scenery, etc., directly influence their emotional inclination. However, existing technologies in travel review sentiment prediction struggle to accurately uncover the deep connections between tourist emotions and attraction semantics. Some studies simply associate attraction information with sentiment tags without deeply analyzing the inherent logical relationship between attraction characteristics and tourist emotions. This prevents targeted improvements to specific aspects of attractions based on tourist emotional feedback when optimizing tourism services, hindering the achievement of a complete closed loop from semantic understanding to service optimization, and failing to fully leverage the value of travel review data in improving tourism service quality.

[0004] Therefore, there is a need to provide a method and system for predicting the sentiment of tourism reviews based on the fusion of knowledge graphs and multimodal features, in order to improve the accuracy of tourism review sentiment prediction and further provide scientific support for smart tourism services. Summary of the Invention

[0005] This invention provides a tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion, comprising: collecting multiple tourism reviews of a target region; extracting a comprehensive feature matrix of multiple tourism reviews based on LDA and TF-IDF; extracting context-enhanced features of multiple tourism reviews based on knowledge graph; filtering the comprehensive feature matrix and context-enhanced features of multiple tourism reviews to select multiple optimal features; and performing review sentiment prediction of the target region based on multiple optimal features using multiple models.

[0006] Furthermore, based on LDA and TF-IDF, a comprehensive feature matrix of multiple travel reviews is extracted, including: preprocessing multiple travel reviews to generate preprocessed travel reviews; determining the optimal number of topics based on the preprocessed travel reviews, and generating a topic distribution matrix based on the LDA model corresponding to the optimal number of topics; extracting the TF-IDF feature matrix of multiple travel reviews based on TF-IDF; and generating a comprehensive feature matrix of multiple travel reviews based on the topic distribution matrix and the TF-IDF feature matrix.

[0007] Furthermore, based on the preprocessed multiple travel reviews, the optimal number of topics is determined, including: determining multiple candidate topic numbers; for each candidate topic number, based on the preprocessed multiple travel reviews, determining the perplexity and topic consistency index value of the LDA model corresponding to the candidate topic number; and determining the optimal number of topics based on the perplexity and topic consistency index value of the LDA model corresponding to each candidate topic number.

[0008] Furthermore, based on TF-IDF, the TF-IDF feature matrix of multiple travel reviews is extracted, including: for each word, calculating the word frequency and inverse document frequency of the word in multiple travel reviews; calculating the TF-IDF value of the word based on the word frequency and inverse document frequency of the word in multiple travel reviews; and generating the TF-IDF feature matrix based on the TF-IDF value of each word.

[0009] Furthermore, based on the knowledge graph, context-enhanced features of multiple travel reviews are extracted, including: determining the sentiment tendency of multiple travel reviews based on social signals; and determining five levels of intermediate voting features for each attraction in the target area based on multiple travel reviews. The context-enhanced features include the sentiment tendency of multiple travel reviews and multi-level intermediate voting features for each attraction in the target area. The multi-level intermediate voting features for attractions include attraction-level weight, sentiment distribution probability, global visit popularity percentage, attraction-level sentiment contribution, category-level sentiment contribution, sentiment tendency index, and sentiment enhancement value.

[0010] Furthermore, based on the social signals of multiple travel reviews, the sentiment tendency of the multiple travel reviews is determined, including: calculating weights based on the mean and standard deviation of the number of likes of the multiple travel reviews; calculating the piecewise augmentation function value based on the weights and the piecewise augmentation function; and determining the sentiment tendency of the multiple travel reviews based on the piecewise augmentation function value.

[0011] Furthermore, the method also includes: obtaining a tourism knowledge graph of the target area, wherein the tourism knowledge graph includes the dimensions of attractions, users, time and path; and generating personalized recommendation results based on context-enhanced features.

[0012] Furthermore, the comprehensive feature matrix and context-enhanced features of multiple travel reviews are screened to select several optimal features, including: for each feature in the comprehensive feature matrix and context-enhanced features of multiple travel reviews, the linear correlation between the feature and the attraction rating is calculated; based on the linear correlation between the feature and the attraction rating, several optimal features are selected.

[0013] Furthermore, comment sentiment prediction is performed using multiple models based on multiple optimal features, including: generating multiple initial comment sentiment prediction results using multiple models based on multiple optimal features; and performing a weighted average of the multiple initial comment sentiment prediction results to generate a final comment sentiment prediction result.

[0014] This invention provides a tourism review sentiment prediction system based on knowledge graph and multimodal feature fusion. Applying the aforementioned tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion, the system includes: a review collection module for collecting multiple tourism reviews from a target region; a feature extraction module for extracting a comprehensive feature matrix from multiple tourism reviews based on LDA and TF-IDF, extracting context-enhanced features from multiple tourism reviews based on a knowledge graph, filtering the comprehensive feature matrix and context-enhanced features from multiple tourism reviews, and selecting multiple optimal features; and a sentiment prediction module for predicting the sentiment of reviews in the target region using multiple models based on multiple optimal features.

[0015] Compared with existing technologies, the tourism review sentiment prediction method and system based on knowledge graph and multimodal feature fusion provided by this invention has at least the following beneficial effects: 1. A comprehensive feature matrix of tourism reviews is extracted based on LDA and TF-IDF, while context-enhanced features are obtained using knowledge graphs. By integrating multi-dimensional features, the emotional information in the reviews can be captured more comprehensively and accurately. Compared with single feature extraction methods, this can effectively improve the accuracy of sentiment prediction, providing a reliable basis for the tourism industry to gain a deeper understanding of tourists' emotional tendencies and helping tourism enterprises to improve services and optimize products in a targeted manner.

[0016] 2. By employing multiple models based on multiple optimal features for sentiment prediction, the advantages of different models are fully utilized, avoiding the limitations that may exist with a single model. Multi-model collaboration enables sentiment analysis and judgment from multiple perspectives, enhancing the stability and robustness of prediction results. It maintains good predictive performance under different scenarios and data distributions, providing a more reliable reference for tourism decision-making.

[0017] 3. Obtaining a tourism knowledge graph encompassing attractions, users, time, and route dimensions enriches the semantic information of tourism reviews. The knowledge graph clearly presents the relationships between elements, providing a solid foundation for extracting context-enhanced features. This helps to deeply understand the complex semantics behind the reviews, uncover potential emotional cues, and make sentiment predictions more accurate.

[0018] 4. The comprehensive feature matrix and context-enhanced features are filtered, and multiple optimal features are selected for prediction. This step removes redundant and irrelevant features, reduces data dimensionality, and lowers computational complexity. While maintaining prediction accuracy, it improves computational efficiency, enabling the method to process large amounts of tourism review data more quickly and making it suitable for real-time analysis of large-scale tourism data. Attached Figure Description

[0019] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is a flowchart illustrating a tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion, as shown in some embodiments of this specification. Figure 2 This is a schematic diagram showing the perplexity and topic consistency index values ​​of the LDA model corresponding to the number of candidate topics as illustrated in some embodiments of this specification. Figure 3 This is a schematic diagram illustrating the performance metrics of prediction models under four different feature sets, as shown in some embodiments of this specification. Figure 4 This is a schematic diagram of a tourism review sentiment prediction system based on the fusion of knowledge graph and multimodal features, as shown in some embodiments of this specification. Detailed Implementation

[0020] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0021] Figure 1 This is a flowchart illustrating a tourism review sentiment prediction method based on the fusion of knowledge graphs and multimodal features, as shown in some embodiments of this specification. Figure 1 As shown, the tourism review sentiment prediction method based on the fusion of knowledge graph and multimodal features may include the following steps.

[0022] Step 110: Collect multiple travel reviews for the target area.

[0023] For example, multiple travel reviews of a target area can be collected from various travel application platforms.

[0024] Step 120: Extract the comprehensive feature matrix of multiple travel reviews based on LDA (Latent Dirichlet Allocation) and TF (Term Frequency)-IDF (Inverse Document Frequency).

[0025] Specifically, it includes: Multiple travel reviews are preprocessed to generate preprocessed travel reviews. Specifically, in order to explore the potential themes of the travel review texts, multiple travel reviews are first preprocessed, and the gensim dictionary and corpus are constructed using the preprocessed texts, while extreme words are filtered out. Based on the preprocessed multiple travel reviews, the optimal number of topics is determined, and a topic distribution matrix is ​​generated based on the LDA model corresponding to the optimal number of topics. Based on TF-IDF, extract the TF-IDF feature matrix of multiple travel reviews; Based on the topic distribution matrix and the TF-IDF feature matrix, a comprehensive feature matrix of multiple travel reviews is generated.

[0026] In some embodiments, determining the optimal number of topics based on multiple preprocessed travel reviews includes: Determine the number of candidate topics, for example, 3, 5, 7, 8, etc.; For each candidate topic number, based on the preprocessed multiple travel reviews, determine the perplexity and topic consistency index values ​​of the LDA (Latent Dirichlet Allocation) model corresponding to the candidate topic number; The optimal number of topics is determined based on the perplexity and topic consistency index of the LDA model corresponding to each candidate topic number.

[0027] Specifically, the perplexity of the LDA model corresponding to the number of candidate topics can be calculated using the following formula: , in, It is a sequence of comments containing N words. ， It is the joint probability of the comment sequence. ， It is the cross-entropy of the sequence: , in, The LDA model is for the first The probability of predicting each word. The lower the perplexity, the stronger the predictive ability of the LDA model for the text; the higher the perplexity, the worse the performance of the LDA model.

[0028] Calculation in progress Point-based mutual information: , in, For words and Point-based mutual information, Indicator and The probability of them appearing together in the same window. and Indicator and The probability of each occurring. Normalized to the interval [-1, 1], denoted as the normalized pointwise mutual information. The higher the normalized pointwise mutual information, the stronger the correlation between words. Then, for each topic's keyword set... Constructing a matrix , in the form of For the matrix For each row in the array, calculate its cosine similarity to all other rows: , And take the average value, where, and It is a matrix The row and number row vectors For matrix The row and number Cosine similarity of row vectors.

[0029] The topic consistency index value corresponding to the number of candidate topics is calculated using the following formula: , in, This represents the topic consistency index value corresponding to the number of candidate topics. For matrix number of rows.

[0030] In some embodiments, based on TF-IDF, the TF-IDF feature matrix of multiple travel reviews is extracted, including: For each word, calculate the word frequency and inverse document frequency in multiple travel reviews, and calculate the TF-IDF value of the word based on the word frequency and inverse document frequency in multiple travel reviews; A TF-IDF feature matrix is ​​generated based on the TF-IDF value of each word.

[0031] Specifically, term frequency (TF) refers to the frequency with which a particular word appears in a travel review document, i.e. ,in These are the words for which TF (Test Values) are calculated. It is a document containing words. It represents the number of times the i-th word appears in the j-th document containing the i-th word. It represents the total number of words in the j-th document containing the i-th word. A higher TF value indicates a more important word within a single document. Inverse Document Frequency (IDF) reflects the importance of a word within the entire collection of travel review documents. , It is a complete collection of comment documents. It is the total number of documents in the document collection. It contains words The number of documents. A higher IDF value indicates that the word is less common in the entire collection of review documents, and therefore more important in travel reviews. Finally, the calculated TF-IDF value is the product of the two, which takes into account both the importance of the word in a single review and its global importance in the entire collection of reviews.

[0032] The TF-IDF value of a word can be calculated using the following formula: , in, Let be the TF-IDF value of the i-th word.

[0033] The LDA model captures the deep semantic structure of travel review texts through topic probability distribution, but weakens the strength of specific words; TF-IDF preserves the distinguishability between sentiment words and entity words through word frequency weights, but lacks topic relevance. To achieve effective integration of multi-dimensional text features, matrix column concatenation is used to fuse text features from different sources, constructing a comprehensive feature matrix, specifically represented as follows: Where N represents the number of documents and D represents the total dimension of the features. Represents the metadata matrix, Represents the LDA topic distribution matrix. This represents the TF-IDF feature matrix.

[0034] in, Used to capture non-textual attribute information of a document. This information is unrelated to the document content but can provide additional context or structured features, such as time information, source information, user attributes, user behavior (e.g., number of views, likes, comments, shares, etc.).

[0035] Step 130: Extract contextual enhancement features from multiple travel reviews based on the knowledge graph.

[0036] Since the sentiment of travel reviews is closely related to information such as attractions, users, time, and routes, integrating knowledge graphs into travel review sentiment analysis provides rich contextual information for prediction, which is of significant value in improving recognition accuracy. This paper constructs a knowledge graph for travel reviews and related content, with the entity type set as follows: ,in For entity types such as category, attraction, user, time, path, etc. Number of entities. Set of relations. in, This refers to relation types, such as contain, visit, comment, distance, etc. The knowledge graph triple set is... ,in, This is a knowledge triple, such as a user commenting on a tourist attraction, etc. This represents the total number of knowledge triples. Entity attributes are cleaned and added to the semi-structured data collected, such as appending comment text information from a user's data collection process as the user's comment attribute. Therefore, a context-enhanced approach based on knowledge graphs is used to extract features, further enriching the multimodal and diversified feature fusion.

[0037] Specifically, this includes: dynamic evaluation feature extraction based on graph network social signal weighting, and attraction sentiment value feature extraction based on hierarchical Bayesian weighting.

[0038] Dynamic evaluation feature extraction based on graph network social signal weighting. First, features are constructed using the social dimension contextual information of the knowledge graph to enhance the recognition of predicted targets. The entity types are mainly based on the category of the attraction (e.g., mountain climbing, exhibitions, water activities, historical sites, etc.), the attraction itself, users, time, and related entity attributes. To ensure the strength of the social signal in the knowledge graph network, for any path depth... User entity node Get the comment attributes of this node. The set of comment attributes of subgraph nodes that meet the path depth requirement is denoted as set. Retrieve the rating attributes of the corresponding node. Base score (Z is a set of integers), number of likes for social signal strength attribute: Text features: , A triple mapping function and social signal weights are constructed using hyperbolic tangent transform. , , This represents the mean and standard deviation of the number of likes attribute in the subgraph of entity nodes of the same type, avoiding bias in sentiment weights due to differences in the popularity of likes across different types. This function has the following properties: when... , (Benchmark weight) (Upper bound constraint) (Lower bound constraint), construct piecewise enhancement function ,For example, , Its characteristics are manifested in its positive evaluation Weight Positive reinforcement occurs during periods of positive evaluation; negative evaluations are also positive. Weight Negative enhancement occurs at times; neutral evaluations remain stable, and the final score... The calculation uses the Sigmoid function. Impose constraints to ensure the results remain within a valid range. The sentiment classification decision surface approach is used to define a non-linear classification boundary, for example... , in, This is an adjustable sensitivity parameter.

[0039] Based on hierarchical Bayesian weighted feature extraction of attraction sentiment values, similarly, tourist review attributes form a tourism review dataset. ,Include Each attraction is categorized into several types (such as mountain climbing, museums and exhibitions, water activities and summer retreats, historical sites, etc.), and each category is... , For category Total number of comments for Class A tourist attraction. For category emotions Number of comments. (Positive / Neutral / Negative, dynamic evaluation features derived from graph network social signals weighted). For tourist attractions The number of comments was used to construct five levels of intermediate voting features, with scenic spot-level weights. ,in, For tourist attractions In category Weights in the distribution of sentiment probability , As a tourist attraction, The corresponding emotion is Comment percentage, overall visit popularity percentage , For category The overall visitor popularity percentage For the total number of reviews in the dataset, attractions On emotions Contribution level, scenic spot-level emotional contribution Category-level emotional contribution , among which, category On emotions The total contribution is calculated as follows: Emotional Tendency Index = (Positive Percentage - Negative Percentage) × 100, and Emotional Enhancement Value = Emotional Tendency Index × Category Weight.

[0040] Step 140: Filter the comprehensive feature matrix and context-enhanced features of multiple travel reviews, and select several optimal features.

[0041] Specifically, it includes: For each feature in the comprehensive feature matrix and context-enhanced features of multiple travel reviews, calculate the linear correlation between the feature and the attraction rating; Based on the linear correlation between features and attraction ratings, several optimal features are selected.

[0042] Specifically, the ANOVA F-value test is used to filter features in the fused feature set to identify the most effective features for predicting attraction ratings. First, data normalization is performed using MinMaxScaler to map feature values ​​to the [0,1] interval, ensuring non-negativity and avoiding the influence of unit weight. Second, feature importance is assessed by using SelectKBest combined with the f_classif statistic to calculate the F-value of each feature with the target variable (e.g., user ratings of attractions on a travel website), measuring the linear correlation between features and categories. Finally, feature selection is performed, ranking features according to their F-values ​​and selecting the top k optimal features. This method filters features through hypothesis testing, making it suitable for multi-class classification problems. Compared to traditional methods, it effectively eliminates redundant and irrelevant features. In processing the fused feature set (Fused), this method selects 100 key features from all features, reducing feature dimensionality while retaining essential information, laying the foundation for subsequent model training.

[0043] Step 150: Using multiple models, perform sentiment prediction for comments in the target region based on multiple optimal features.

[0044] Specifically, it includes: Multiple initial sentiment prediction results for comments are generated using multiple models and based on multiple optimal features. The initial sentiment prediction results of multiple comments are weighted and averaged to generate the final sentiment prediction result.

[0045] Specifically, an RF-XGB-LGBM fusion algorithm was generated by integrating three base models: RandomForest, XGBoost, and LightGBM, using a geometrically weighted average ensemble strategy. First, the base models are trained, outputting the probability distribution of samples belonging to individual learners. Then, a geometric ensemble (GeoEnsemble) is performed on the probability outputs of the base models. By enhancing the weights of high-confidence classes, the multi-model divergence problem is addressed, effectively reducing the error rate in sentiment analysis tasks. Furthermore, the dynamic weight settings based on model training performance improve the overall recognition accuracy. Finally, normalization is performed; after normalizing the fused probabilities, the class corresponding to the maximum probability is taken as the prediction result. This strategy leverages the complementarity of different models, enhancing the discriminative power of high-probability classes through geometric ensemble, thereby improving overall prediction accuracy. , in, To comment on the sentiment prediction results, Let be the probability output of the i-th model. C represents the weights of the individual learners, and C represents the number of emotion categories, such as positive, negative, neutral, etc.

[0046] The following section, through experiments, illustrates the beneficial effects of the tourism review sentiment prediction method based on the fusion of knowledge graphs and multimodal features.

[0047] The experimental data came from 17,691 Tibet tourism reviews crawled from multiple tourism websites. Of these, 47.4% were related to mountain climbing, 39.1% to scenic spots and historical sites, 12.2% to water activities and summer retreats, and 1.3% to exhibitions and museums. A three-layer progressive cleaning strategy was adopted to improve the text data quality. The first layer used regular expressions to remove numbers, symbols, and other redundant characters. The second layer loaded a combined stop word library, namely the Harbin Institute of Technology's general stop word library and a Tibet-specific stop word library, to deeply filter semantic noise. The third layer set a double-character length threshold to remove single-character words. After cleaning, word clouds showed that tourists preferred Tibet's natural landscapes and cultural heritage. In the mountain climbing category, Tibet, scenic spots, the Grand Canyon, and Namcha Barwa Peak were particularly popular. Since the original rating data consisted of discrete values ​​from 1 to 5, with a high score of 5 accounting for 77.64% and a low score of 2 accounting for less than 2%, the SMOTE algorithm was used in this experiment to balance the imbalance of training sample categories.

[0048] By normalizing perplexity and consistency, a comprehensive score is calculated to select the optimal number of topics. Since lower perplexity is better, while higher topic consistency is better, subtracting perplexity from consistency yields a higher comprehensive score, indicating better model performance. Figure 2 As shown, the highest overall score is 1.0 when the number of topics is 5, corresponding to a model perplexity of 158.8 and a consistency of 0.6.

[0049] By combining the optimal number of topics and using the LDA model, tourism reviews are divided into five categories, including Tibetan religious architecture, tourism service processes, etc. This includes not only keywords related to physical attractions such as Jokhang Temple and the Grand Canyon, but also service elements such as tour guides and tickets, as well as emotional expressions such as worthiness and sacredness. This provides a relatively comprehensive understanding of tourists' focus on Tibetan tourism from the semantic dimension of the topics.

[0050]

[0051] To balance the deep semantics at the topic level captured by the LDA model and the discriminative strength at the word frequency level of the TF-IDF model, the experiment fused the features extracted by both to form a two-layer representation of "topic semantics + word-level details," merging them to construct a multi-dimensional feature system. The TF-IDF semantic word frequency feature matrix extraction process incorporated dynamic screening using a chi-square test. By constructing a joint index system of TF-IDF feature matrix and chi-square test screening, low-frequency words with an occurrence rate below 0.5% and common words with an occurrence rate above 90% were filtered out, initially screening the original 1200+ dimensional features. Based on the chi-square test, with a p-significance level of 0.05, the dimensionality was reduced to 346 dimensions. Table 2 shows the top 10 features with the strongest correlation.

[0052]

[0053] From the perspectives of social signal weighting and attraction sentiment stratification, this study utilizes dynamic ratings weighted by social signals and attraction sentiment to construct a content feature extraction model based on knowledge graph contextual information, resulting in four categories of comment text sentiments, as shown in Table 3. The sentiment indices for mountain climbing and exhibition categories are +35.24 and +29.13, respectively, significantly higher than other types of attractions, indicating that these two types of attractions can more accurately capture tourists' emotional characteristics in target recognition effectiveness detection. Combined with the chi-square test results in Table 3, at the 95% confidence level, the P-values ​​are all less than the significance level of 0.05. This indicates a significant statistical difference in the sentiment weight distribution among different tour types, further validating the effectiveness of the knowledge graph and sentiment fusion model in distinguishing the sentiment characteristics of attraction types.

[0054]

[0055] For the fused feature set, LDA topic features, TF-IDF word frequency features, knowledge graph context enhancement features, and structured features were selected using the ANOVA F-value method. By evaluating different k values ​​within the range of 100 to 394 (step size 10) using 5-fold cross-validation and the f1_macro scoring index, it was found that when k=100, most of the selected features had p-values ​​much less than 0.05 in the ANOVA F-value test, indicating that at a significance level of p<0.05, a 95% confidence interval was met, and 100 was determined to be the optimal result.

[0056] For the selected features, the performance metrics of the prediction models under four different feature sets were compared, such as... Figure 3As can be seen, after multimodal feature fusion, the accuracy, F1 score, and Area Under the Curve of the comment sentiment prediction results reached 0.9915, 0.9794, and 0.9933, respectively, which is about 0.25 higher than that of a single feature set such as the TF-IDF feature set in terms of AUC. This indicates that multimodal feature fusion can effectively capture more comprehensive rating influencing factors, and the fusion strategy has a significant advantage in predicting the target.

[0057] Further comparison of the performance of the proposed multi-model fusion method on a multimodal feature fusion dataset reveals that the geometric mean ensemble strategy model has a significant advantage over individual learners, as shown in Table 4. The confusion matrix shows that the ensemble model significantly reduces the misclassification rate for the 2-3 point critical score, and by fusing the advantages of different models in the high and low score ranges, it achieves an overall reduction in prediction error across all score ranges. This demonstrates the model ensemble strategy's ability to capture score details and its good robustness in multi-classification scenarios.

[0058]

[0059] Figure 4 This is a schematic diagram of a tourism review sentiment prediction system based on knowledge graph and multimodal feature fusion, as shown in some embodiments of this specification. Figure 4 As shown, a tourism review sentiment prediction system based on knowledge graph and multimodal feature fusion can include a review collection module, a feature extraction module, and a sentiment prediction module.

[0060] The review collection module is used to collect multiple travel reviews for the target area. The feature extraction module is used to extract a comprehensive feature matrix of multiple travel reviews based on LDA and TF-IDF, extract context-enhanced features of multiple travel reviews based on knowledge graph, and filter the comprehensive feature matrix and context-enhanced features of multiple travel reviews to select multiple optimal features; The sentiment prediction module is used to predict the sentiment of comments in a target region using multiple models and based on multiple optimal features.

[0061] The tourism review sentiment prediction system based on knowledge graph and multimodal feature fusion can apply the above-mentioned tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion, which will not be elaborated here.

[0062] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. A tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion, characterized in that, include: Collect multiple travel reviews for the target area; Based on LDA and TF-IDF, a comprehensive feature matrix of multiple travel reviews is extracted; Based on knowledge graphs, context-enhanced features are extracted from multiple travel reviews; The comprehensive feature matrix and context-enhanced features of multiple travel reviews are filtered to select several optimal features; Sentiment prediction for comments in the target region is performed using multiple models and based on multiple optimal features.

2. The tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion according to claim 1, characterized in that, Based on LDA and TF-IDF, a comprehensive feature matrix of multiple travel reviews is extracted, including: Multiple travel reviews are preprocessed to generate preprocessed travel reviews; Based on the preprocessed multiple travel reviews, the optimal number of topics is determined, and a topic distribution matrix is ​​generated based on the LDA model corresponding to the optimal number of topics. Based on TF-IDF, extract the TF-IDF feature matrix of multiple travel reviews; Based on the topic distribution matrix and the TF-IDF feature matrix, a comprehensive feature matrix of multiple travel reviews is generated.

3. The tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion according to claim 2, characterized in that, Based on the preprocessed multiple travel reviews, the optimal number of topics is determined, including: Determine the number of candidate topics; For each candidate topic number, based on the preprocessed multiple travel reviews, the perplexity and topic consistency index values ​​of the LDA model corresponding to the candidate topic number are determined; The optimal number of topics is determined based on the perplexity and topic consistency index of the LDA model corresponding to each candidate topic number.

4. The tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion according to claim 3, characterized in that, Based on TF-IDF, the TF-IDF feature matrix of multiple travel reviews is extracted, including: For each word, calculate the word frequency and inverse document frequency in multiple travel reviews, and calculate the TF-IDF value of the word based on the word frequency and inverse document frequency in multiple travel reviews; A TF-IDF feature matrix is ​​generated based on the TF-IDF value of each word.

5. The tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion according to any one of claims 1-4, characterized in that, Based on knowledge graphs, context-enhanced features are extracted from multiple travel reviews, including: Based on the social signals of multiple travel reviews, determine the sentiment of multiple travel reviews; Based on multiple travel reviews, five levels of intermediate voting features are determined for each attraction in the target area. The context-enhanced features include the sentiment tendency of multiple travel reviews and the multi-level intermediate voting features of each attraction in the target area. The multi-level intermediate voting features of attractions include attraction-level weight, sentiment distribution probability, global visit popularity ratio, attraction-level sentiment contribution, category-level sentiment contribution, sentiment tendency index, and sentiment enhancement value.

6. The tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion according to claim 5, characterized in that, Based on the social signals from multiple travel reviews, the sentiment trends of these reviews were determined, including: The weights are calculated based on the mean and standard deviation of the number of likes for multiple travel reviews; Calculate the piecewise augmentation function value based on the weights and the piecewise augmentation function; Based on the piecewise enhancement function values, the sentiment tendency of multiple travel reviews is determined.

7. The tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion according to claim 5, characterized in that, Also includes: Obtain the tourism knowledge graph of the target area, which includes the dimensions of attractions, users, time, and path. Personalized recommendation results are generated based on context-enhanced features.

8. The tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion according to any one of claims 1-4, characterized in that, The comprehensive feature matrix and context-enhanced features of multiple travel reviews were filtered to select several optimal features, including: For each feature in the comprehensive feature matrix and context-enhanced features of multiple travel reviews, calculate the linear correlation between the feature and the attraction rating; Based on the linear correlation between features and attraction ratings, several optimal features are selected.

9. The tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion according to any one of claims 1-4, characterized in that, Sentiment prediction for comments is performed using multiple models based on multiple optimal features, including: Multiple initial sentiment prediction results for comments are generated using multiple models and based on multiple optimal features. The initial sentiment prediction results of multiple comments are weighted and averaged to generate the final sentiment prediction result.

10. A tourism review sentiment prediction system based on knowledge graph and multimodal feature fusion, characterized in that, The tourism review sentiment prediction method based on knowledge graph and multimodal feature fusion as described in claim 1 includes: The review collection module is used to collect multiple travel reviews for the target area. The feature extraction module is used to extract a comprehensive feature matrix of multiple travel reviews based on LDA and TF-IDF, extract context-enhanced features of multiple travel reviews based on knowledge graph, and filter the comprehensive feature matrix and context-enhanced features of multiple travel reviews to select multiple optimal features; The sentiment prediction module is used to predict the sentiment of comments in a target region using multiple models and based on multiple optimal features.

Images