Hypergraph knowledge fused service quality dynamic analysis method

Abstract

The invention belongs to the technical field of natural language processing, and particularly relates to a hypergraph knowledge-fused service quality dynamic analysis method, which comprises the following steps of: mapping a comment theme to five dimensions of an SERVQUAL model by using a service quality dimension point mutual information algorithm; constructing the customer comments into a hypergraph, and inputting the hypergraph into a hypergraph attention network for processing so as to obtain customer service quality perception values contained in the customer comments; using a Wilson interval method to solve an expected score of the merchant, and then correcting the expected score of the merchant through a Bayesian average method; and adding the expected score of the merchant and the historical score mean value of the customer, normalizing the sum and taking the sum as a customer service quality expected value, and obtaining the satisfaction degrees of the customer in five different dimensions of the SERVQUAL model. The invention provides a customer service quality evaluation scheme independent of questionnaire survey, and provides a feedback mechanism for merchants, thereby enhancing the competitiveness and influence of the merchants.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing, and in particular relates to a service quality dynamic analysis method integrating hypergraph knowledge. Background Art

[0002] Service quality is different from physical products that can be directly touched. Its production and consumption are integrated. Therefore, it depends more on the interaction process between customers and service providers, and has the characteristics of intangibility. Since different people have different needs and expectations for services, the evaluation criteria of service quality are also subjective and flexible. These characteristics make service quality evaluation difficult, because multiple factors such as service providers, consumers, and service processes must be considered at the same time. In order to better evaluate service quality, the SERVQUAL model based on gap analysis was proposed and gradually became the mainstream model in the field of service quality analysis. The model regards service quality as a multidimensional concept, and specifically defines five dimensions: tangibility, reliability, responsiveness, assurance, and empathy. And service quality is defined as the gap between customers' expectations of services and the actual service level they experience, that is, it is believed that the customer's service quality perception value is equal to the difference between the customer's service quality perception value and the customer's service quality expectation value. Although the SERVQUAL model is widely used, it still has some shortcomings. First, the implementation of the SERVQUAL model requires the design of a scale based on specific areas, and the expansion of the scale into a questionnaire to be distributed to customers for surveys. However, in reality, only a small number of customers are willing to fill out the questionnaire. Subsequently, a series of complex operations such as data collection and result analysis are required, which requires a large investment of manpower and material resources. Secondly, the SERVQUAL model can only reflect the service quality level at the current moment, and cannot track and predict the trend of service quality changes, which limits its application in the continuous improvement of service quality. Summary of the invention

[0003] The purpose of this invention is to propose a service quality dynamic analysis method that integrates hypergraph knowledge. This method first obtains comment topics and their keywords from large-scale online comments through a comment topic extraction method based on a pre-trained model and maps them to the dimensions of the SERVQUAL model. Secondly, each comment is constructed into a hypergraph using three hyperedge construction methods: text structure, intrinsic semantics, and external knowledge, and is processed using a hypergraph attention network to extract the customer's actual perceived service quality level. Subsequently, the customer's service quality expectation is extracted, and the difference between it and the actual perception is calculated as the customer's service quality perception value reflected by each comment. Combined with the comment topic corresponding to the evaluation, the customer's satisfaction on five different dimensions of the SERVQUAL model is obtained.

[0004] The present invention is achieved through the following technical solutions.

[0005] A method for dynamic analysis of service quality integrating hypergraph knowledge, the steps are as follows: Step 1: Collect online network platform data and construct a data set; the online network platform data at least includes customer reviews; Step 2: Extract the review topics contained in the customer reviews and determine the keywords and their weights for each review topic; Step 3: Adjust the extracted review topics and keywords to optimize the review topics; Step 4: Use the service quality dimension point mutual information algorithm to map the review topics to the five dimensions of the SERVQUAL model; Step 5: Perform confidence learning on the dataset to find samples with incorrect labels and manually correct them; Step 6: Construct customer reviews into a hypergraph containing three types of hyperedges: text structure, intrinsic semantics, and external knowledge. Then input the hypergraph into the hypergraph attention network for processing, thereby obtaining the customer service quality perception value contained in the customer reviews. Step 7: Use the Wilson interval method to calculate the expected score of the merchant, and then use the Bayesian average method to correct the expected score of the merchant; add the expected score of the merchant and the average of the customer's historical score and normalize them as the customer's service quality expectation; Step 8: Combining the three factors of review topic, customer service quality expectations, and customer service quality perception, we can derive customer satisfaction on the five different dimensions of the SERVQUAL model.

[0006] Further preferably, the online network platform data also includes merchant information, customer information and customer ratings.

[0007] Further preferably, a review topic modeling method based on the pre-trained model RoBERTa is used to extract review topics contained in customer reviews.

[0008] Further preferably, the c-TF-IDF algorithm is used to determine the keywords and their weights of each review topic; Further preferably, the extracted comment topics and keywords are adjusted to optimize the comment topics by setting stop words, changing the minimum word frequency, and adjusting N-gram parameters.

[0009] Further optimization, in step 4, seed words on the five dimensions of the SERVQUAL model are constructed, and the keywords in each review topic and the seed words on the five dimensions of the SERVQUAL model are constructed into a word set; the service quality dimension point mutual information algorithm (SQ-PMI) is used to calculate the correlation value between the sets, and the review topics are dynamically mapped to the five dimensions of the SERVQUAL model.

[0010] Further preferably, in step five, all customer review data in the data set are divided into five equal parts, and then the FastText model is used to perform a five-fold cross-validation on the data. During the cross-validation process, the joint distribution matrix of the estimated noise labels and the true labels is obtained, and finally, the samples below the confidence threshold are screened out by comparing the joint distribution matrix, which are the incorrectly labeled samples, and the rest are correctly labeled samples, and then the incorrectly labeled samples are manually corrected.

[0011] Further preferably, in step six, each customer review is constructed as a hypergraph. First, the customer reviews are segmented and the words are used as nodes of the graph. Then, a sliding window of a fixed size is used to obtain global word co-occurrences as sequential contexts and used as text structure hyperedges of the hypergraph. Words belonging to the same review topic in the nodes are connected by the same hyperedge as the intrinsic semantic hyperedge. External knowledge is obtained from the ConceptNet knowledge graph and external knowledge hyperedges are constructed. In the hypergraph processing link, the constructed hypergraph is sent to the hypergraph attention network, and the hypergraph attention network outputs the customer service quality perception values ​​contained in the customer reviews.

[0012] Based on the five dimensions of the SERVQUAL model and the calculation method of customer service quality perception value, the present invention uses the text sequence structure, internal semantic association and knowledge graph triple knowledge of customer review text to construct a hypergraph, and then embeds the hypergraph into the graph neural network for processing, which has the following advantages: 1. Combine the traditional SERVQUAL model with deep learning technology to conduct sentiment analysis and comment topic modeling on online comments, so as to dynamically and continuously analyze customer satisfaction, avoiding the tedious process of designing, distributing and collecting questionnaires in traditional analysis methods.

[0013] 2. A dynamic mapping method between the review topic model and the five dimensions of the SERVQUAL model is designed, so that each review can be mapped to a dimension of the SERVQUAL model. A method for extracting customer service quality expectations that are difficult to quantify is proposed. The fit between the deep learning model and the SERVQUAL model is improved, so that they can better form an organic whole to analyze the reviews together.

[0014] 3. Hyperedges are constructed for the evaluation hypergraph using the sequential structure of the evaluation text, the internal potential semantic information, and the external knowledge provided by the knowledge graph. The three different types of hyperedges are complementary and can provide multi-source heterogeneous features for the evaluation hypergraph and enrich the semantic information contained in the hypergraph.

[0015] 4. The introduction of focal loss improves the robustness of the model on sample imbalanced data sets. It can dynamically adjust the weights on sample data sets with extreme positive and negative ratios, increase the loss weights of minority class samples, and thus improve network performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of a service quality dynamic analysis method integrating hypergraph knowledge provided by the present invention; Figure 2 It is a schematic diagram of the framework structure of a service quality dynamic analysis method integrating hypergraph knowledge; Figure 3 It is a scatter plot of the customer service quality perception value distribution of the reviews; Figure 4 It is a histogram of the number of reviews in each customer service quality perception value interval; Figure 5 It is a graph showing the changes in merchant service quality levels over time series. DETAILED DESCRIPTION

[0017] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0018] Reference Figure 1 ,A service quality dynamic analysis method integrating hypergraph knowledge, the steps are as follows: Step 1: Collect online platform data such as merchant information, customer information, customer comments and customer ratings from online platforms, and build a data set after cleaning and preprocessing the online platform data; Step 2: Use the review topic modeling method based on the pre-trained model RoBERTa to extract the review topics contained in customer reviews, and use the c-TF-IDF algorithm to determine the keywords and their weights of each review topic; Step 3: Adjust the extracted comment topics and keywords to optimize the comment topics by setting stop words, changing the minimum word frequency, and adjusting N-gram parameters; Step 4: Use the service quality dimension point mutual information algorithm to map the review topics to the five dimensions of the SERVQUAL model; Step 5: Perform confidence learning on the dataset to find samples with incorrect labels and manually correct them; Step 6: Construct customer reviews into a hypergraph containing three types of hyperedges: text structure, intrinsic semantics, and external knowledge. Then input the hypergraph into the hypergraph attention network for processing, thereby obtaining the customer service quality perception value contained in the customer reviews. Step 7: Use the Wilson interval method to calculate the expected score of the merchant, and then use the Bayesian average method to correct the expected score of the merchant; add the expected score of the merchant and the average of the customer's historical score and normalize them as the customer's service quality expectation; Step 8: Combining the three factors of review topic, customer service quality expectations, and customer service quality perception, we can derive customer satisfaction on the five different dimensions of the SERVQUAL model.

[0019] Figure 2 The basic framework of the present invention is demonstrated, and the framework has three branches: SERVQUAL dimension mapping branch, customer service quality perception value prediction branch, and customer service quality expectation value prediction branch.

[0020] For the dataset composed of online network platform data, the pre-trained RoBERTa model is first used in the SERVQUAL dimension mapping branch to embed the words in the text. The embedded word vectors will be used to extract comment topics and construct nodes of the hypergraph. Subsequently, the comment topics contained in the dataset are obtained by reducing the dimension and clustering the word vectors, and then the keywords and weights of each comment topic are determined by the c-TF-IDF algorithm. The algorithm can be expressed by formula (1): (1); in, represents the result of c-TF-IDF algorithm operation of keyword t under cluster c, t represents the keyword in the comment topic cluster, c represents the cluster, represents the frequency of keyword t in cluster c, represents the frequency of keyword t in all corpora, is the average number of words in cluster c. To ensure that the values ​​are only positive, add 1 to the division in the logarithm.

[0021] After obtaining the preliminary review topics and their keywords, the extracted review topics and keywords are adjusted to optimize the review topic representation by setting stop words, changing the minimum word frequency, and adjusting N-gram parameters. After obtaining the optimized review topics and keywords, there may be a problem of inconsistent emphasis between the review topics and the dimensions of the SERVQUAL evaluation model. Based on this, the present invention constructs seed words on the five dimensions of the SERVQUAL model, and constructs the keywords in each review topic and the seed words of the five dimensions of the SERVQUAL model into a word set. The present invention adopts the service quality dimensional point mutual information algorithm (SQ-PMI) to calculate the correlation value between the sets, which can dynamically map the review topics to the five dimensions of the SERVQUAL model. The algorithm can be expressed by formula (2): (2); Where T represents the set of comment topic keywords, D represents the set of dimension seed words, represents the service quality dimension point mutual information value of the review topic T on dimension D, d represents a dimension seed word, represents the probability of the comment topic keyword t appearing, represents the probability of the dimension seed word d appearing, Indicates the probability of the review topic keyword t and dimension seed word d appearing at the same time. A positive value of the service quality dimension point mutual information indicates that the two sets are positively correlated. The larger the value, the stronger the correlation. If it is 0, it means that the two sets are almost statistically independent and have no correlation. A negative number indicates that the two sets are negatively correlated, that is, mutually exclusive. For each review topic, use SQ-PMI to calculate the correlation value between the review topic and each of the five dimensions of the SERVQUAL model and map the review topic to the dimension with the largest correlation value.

[0022] The customer service quality perception value prediction branch constructs hyperedges for the hypergraph based on the text structure, intrinsic semantics, and external knowledge. Then, through confidence learning, low-confidence samples are found and corrected to prevent the propagation of incorrect labels. The hypergraph is then sent to the hypergraph attention network, which outputs the customer service quality perception value. The reason for using confidence learning is that when customers rate online, the scores given do not match the emotions contained in the comments, such as consumers giving negative comments but giving full marks or giving positive comments but only giving low marks. This will reduce the credibility of the labels of customer comment data and affect subsequent model training and customer service expectation extraction. In order to alleviate this situation, the present invention divides all customer comment data in the data set into five equal parts, and then uses the FastText model to perform five-fold cross-validation on the data. During the cross-validation process, the joint distribution matrix of the estimated noise labels and the true labels is obtained. This process can be expressed by formula (3) and formula (4): (3); (4); in represents the sample number, m represents the total number of label categories, Indicates one of the categories. Representation sample The predicted value of Indicates category The number of labels on . It is a count matrix with m rows and m columns. The values ​​are obtained from the samples and the columns represent the categories. The number of labels above, the row indicates the model prediction value is the category The number of . The pseudo label is The total number of samples. is the calibration count matrix. Formula (4) shows the process of regularizing it. After regularization, the joint distribution matrix of the estimated noise label and the true label is obtained. Finally, the samples below the confidence threshold are screened out by comparing the joint distribution matrix, which are the samples with incorrect labels, and the rest are the samples with correct labels. The samples with incorrect labels are then manually corrected.

[0023] The customer service quality expectation prediction branch first combines the Wilson interval method with the Bayesian average to obtain the expected score of the merchant. Specifically, the present invention uses the Wilson interval method to calculate the expected score of the merchant at a 95% confidence level. Since some merchants have fewer reviews, using only the Wilson interval algorithm will cause unpopular merchants to be restricted by the lower limit of the interval, resulting in lower scores. Therefore, the present invention uses the Bayesian average method to provide a compensation value for unpopular merchants so that they do not generate unreliable scores due to a small number of reviews. This process can be expressed by formulas (5) and (6): (5); (6); in, represents the Wilson score, It is the upper limit of the score and plays a scaling role. represents the favorable comment rate, which is obtained by dividing the average by the total score. n is the total number of comments, and K is a statistic of the normal distribution at a fixed confidence level. h represents the number of merchant comments, and C represents the correction value for the merchant, which is equal to the average number of votes for all items. The smaller the C value, the more biased the Bayesian correction result will be towards the original data, and the impact on unpopular merchants will be weakened. The overall score will be closer to the Wilson score. The larger the C value, the greater the correction value for unpopular merchants, which will reduce the difference in the scores of all merchants. H represents the total average score of all merchants. Indicates the final expected merchant rating.

[0024] After obtaining the expected score of the merchant, it is summed and normalized with the customer's historical average score to extract the customer service quality expectation. Finally, the difference between the customer service quality perception value and the customer service quality expectation value in the five dimensions of the SERVQUAL model is calculated to obtain the customer satisfaction in the five dimensions.

[0025] As a generalization of graphs, hypergraphs can more flexibly model complex network structures and show high-order relationships that graphs cannot fully describe. The present invention constructs each customer review as a hypergraph. First, the customer reviews are segmented and the words are used as nodes of the graph. Then, a fixed-size sliding window is used to obtain global word co-occurrences as sequential contexts, which are used as text structure hyperedges of the hypergraph. In order to capture the high-order correlation between words and potential review topics, the words belonging to the same review topic in the nodes are connected by the same hyperedge as the intrinsic semantic hyperedge. Since most customer reviews are short texts, in order to enrich the semantics therein, the present invention obtains external knowledge from the ConceptNet knowledge graph and constructs external knowledge hyperedges. The hypergraph obtained using the above three hyperedge construction methods can be represented by G(V, E), where V={v1,…,v a} represents a set of a nodes in the hypergraph, v1 represents the first node, v a represents the ath node, E={e1,…,e b} represents a set of b hyperedges, e1 represents the first hyperedge, e b represents the bth hyperedge. For any hyperedge e, it can connect two or more nodes. The topological structure of the hypergraph G can be represented by an association matrix with a rows and b columns. express, (7); Where p is the node number, q is the hyperedge number, represents the pth node, represents the pth hyperedge.

[0026] In the hypergraph processing link, the constructed hypergraph is sent to the hypergraph attention network, and the hypergraph attention network will output the customer service quality perception value contained in this customer review. The hypergraph attention network will aggregate the vertices connected to the hyperedge to obtain the representation vector of the hyperedge, and then aggregate the hyperedge vector to update the node vector. Since the importance of each node to the hyperedge is not the same during the node aggregation process; the influence of each hyperedge on the node is not consistent during the hyperedge aggregation process, the present invention introduces an attention mechanism in the hypergraph attention network to dynamically adjust the weights in the aggregation process. Formula (8) represents the process of aggregating all nodes connected to the hyperedge to obtain the hyperedge representation vector. Formula (9) represents the process of aggregating all hyperedge vectors connected to the node to update the node representation. In this process, the number of hyperedges aggregated by the node is always equal to the degree of this node.

[0027] (8); (9); in, Represents a hyperedge In the Hyperedge representation vectors in layer-by-layer hypergraph attention networks. Represents a hyperedge The node connected to, Indicates that the node is The representation vector corresponding to the layer, As the attention coefficient represents the node Super edge The importance of. W1 and W2 both represent weight matrices. Indicates Aggregation operations of layers, Indicates the aggregation Layer each hyperedge is connected to all nodes, thus obtaining the hyperedge representation vector process. Indicates the aggregation The process of updating the node representation vector by calculating the hyperedges to which each node in the layer is connected. Indicates that node p is in The representation vector of the layer, Representation Node In the The representation vector of the layer. Representation and Node The set of connected edges, Represents a hyperedge when aggregated For Node The importance of.

[0028] Customer service quality perception values ​​often have an imbalanced distribution. Generally speaking, positive perceptions far outweigh negative perceptions. This imbalance will cause minority class samples to be overwhelmed by the majority class, making the output of the hypergraph attention network all majority class, resulting in prediction failure. In order to cope with the sample imbalance problem faced in the hypergraph processing link, the present invention uses the focal loss (FL) with dynamic class weights to train the hypergraph attention network. The focal loss is an improvement on the cross entropy loss. By setting a dynamic scaling factor, the loss value weight of minority class samples is increased to avoid being overwhelmed by the majority class.

[0029] The data used in the embodiment of the present invention comes from the Yelp website. The comments on this website are different from the comments related to physical products. Customers of the former are more inclined to evaluate the service quality level of service providers and other service entities, while the latter focuses on evaluating an actual product itself. Due to the intangibility of services and the diversity of evaluation criteria, it is more difficult to measure the level. The present invention stores the merchant information, customer information, customer comments, and customer ratings provided by the Yelp website in different tables of the database. The merchant information table contains three fields: merchant number, merchant category, and merchant corresponding comment number. The present invention collects all relevant data of merchants engaged in the tree care industry on the Yelp website and integrates them into a data set, and then preprocesses the text in the data set. First, remove stop words, mainly including some prepositions and conjunctions that have no practical meaning, then convert all letters to lowercase, and finally perform part of speech restoration.

[0030] The effect of the present invention is measured using four indicators, namely accuracy and weighted precision, recall and F1 value. They use the number of samples in each category as weights. Compared with the macro method of directly averaging the indicators of each category, it can ensure that more important categories have greater influence when evaluating the model effect.

[0031] The present invention automatically extracts 9 review topics and the keywords and weights under each review topic in the data set through the pre-trained RoBERTa model, where the keywords and weights of the first three review topics are shown in Table 1.

[0032] Table 1 Comment topic keywords and their weights of the first three comment topics extracted

[0033] Then, the service quality dimension point mutual information algorithm is used to map each review topic to the five dimensions of the SERVQUAL model. The service quality dimension point mutual information value (SQ-PMI value) of each review topic on the five dimensions of the SERVQUAL model is shown in Table 2. The dimension with the largest value is the SERVQUAL dimension to which this review topic belongs.

[0034] Table 2 SQ-PMI values ​​of each review topic on the five dimensions of the SERVQUAL model

[0035]

[0036] Then, the hypergraph attention network is used to obtain the service quality perception value of each comment. The hypergraph attention network is trained on a part of the data set. After training, it is used to analyze the newly added samples to ensure that the solution in this paper can incrementally extract customer perception. The accuracy of the hypergraph attention network on the test set reached 85.2%, and the weighted precision, recall, and F1 value were 83.3%, 81.5%, and 82.6%, respectively. Then, the customer's expectation of customer service quality is calculated, and their difference is used as the customer service quality perception value corresponding to the evaluation. The distribution of customer service quality perception values ​​of all comments in the data set is as follows: Figure 3 The number of customer reviews in each customer service quality perception value interval is shown in Figure 4 shown.

[0037] In order to verify the effectiveness of the proposed method in the generation of standardized documents, an ablation experiment was designed and six ablation schemes were formulated. Scheme 1 regards a comment as a node, constructs the entire dataset into a graph, and then performs the node classification task of the graph, using focal loss to avoid the classification result from shifting to multiple categories. Scheme 2 introduces an attention mechanism in the process of graph node aggregation based on Scheme 1. Scheme 3 constructs each review as a graph, constructs the hyperedge of the graph according to the text structure, and then uses a graph convolutional network to process the hypergraph. Scheme 4 integrates the internal semantics and external knowledge of the comment text on the basis of Scheme 3, and constructs the internal potential comment topic and the external knowledge triple as hyperedges respectively. Scheme 5 introduces an additional attention mechanism in the process of graph convolution of the hypergraph that integrates text structure, internal semantics, and external knowledge. Scheme 6 removes the focal loss on the basis of Scheme 5 and uses the cross entropy loss function to observe whether the model prediction result is unbalanced. Scheme 7 is the scheme of the present invention. Before model training, confidence learning is performed on the dataset to correct some samples that may have incorrect labels, so as to prevent the wrong labels from propagating in the model and interfering with the prediction effect of the model. The ablation experiment results are shown in Table 3. It can be seen that the various indicators of the method of the present invention are better than those of the other six ablation schemes, which reflects the effectiveness of the scheme of the present invention and the rationality of the module combination.

[0038] Table 3 Ablation experiment

[0039] Since the present invention can quickly obtain the customer satisfaction in the five dimensions of the SERVQUAL evaluation model, the review sequence can be dynamically analyzed in the time dimension. The changes in satisfaction in the Tree Services dataset in the time series are as follows: Figure 5As shown in the figure, it can be seen that the service level of the merchants in the empathy evaluation dimension in this data set is always lower than the expected level, which reflects that merchants should strengthen personalized care and service for customers, improve their understanding and attention to customer needs, and give priority to the interests of customers to establish a deeper emotional connection and thus improve customer satisfaction in terms of empathy. There have also been cases where customers were dissatisfied with the responsiveness dimension, which also requires merchants to focus on and continuously improve.

[0040] This embodiment provides a computer-readable storage medium on which computer instructions are stored. When the instructions are executed by a processor, the service quality dynamic analysis method integrating hypergraph knowledge is implemented.

[0041] The invention described above only expresses the implementation methods of the embodiments of the present invention, and cannot be understood as limiting the scope of the invention patent, nor does it impose any form of limitation on the structure of the embodiments of the present invention. It should be pointed out that for ordinary technicians in this field, several changes and improvements can be made without departing from the concept of the embodiments of the present invention, which all belong to the protection scope of the embodiments of the present invention.

Claims

1. A method for dynamic analysis of service quality integrating hypergraph knowledge, characterized in that: Here are the steps: Step 1: Collect online network platform data and construct a data set; the online network platform data at least includes customer reviews; Step 2: Extract the review topics contained in the customer reviews and determine the keywords and their weights for each review topic; Step 3: Adjust the extracted review topics and keywords to optimize the review topics; Step 4: Use the service quality dimension point mutual information algorithm to map the review topics to the five dimensions of the SERVQUAL model; Step 5: Perform confidence learning on the dataset to find samples with incorrect labels and manually correct them; Step 6: Construct customer reviews into a hypergraph containing three types of hyperedges: text structure, intrinsic semantics, and external knowledge. Then input the hypergraph into the hypergraph attention network for processing, thereby obtaining the customer service quality perception value contained in the customer reviews. Step 7: Use the Wilson interval method to calculate the expected score of the merchant, and then use the Bayesian average method to correct the expected score of the merchant; add the expected score of the merchant and the average of the customer's historical score and normalize them as the customer's service quality expectation; Step 8: Combining the three factors of review topic, customer service quality expectations, and customer service quality perception, we can derive customer satisfaction on the five different dimensions of the SERVQUAL model.

2. The method for dynamic analysis of service quality integrating hypergraph knowledge according to claim 1 is characterized in that: The online network platform data also includes merchant information, customer information and customer ratings.

3. The method for dynamic analysis of service quality integrating hypergraph knowledge according to claim 1 is characterized in that: A review topic modeling method based on the pre-trained model RoBERTa is used to extract review topics contained in customer reviews.

4. The method for dynamic analysis of service quality integrating hypergraph knowledge according to claim 1 is characterized in that: The c-TF-IDF algorithm is used to determine the keywords and their weights for each review topic.

5. The method for dynamic analysis of service quality integrating hypergraph knowledge according to claim 1 is characterized in that: By setting stop words, changing the minimum word frequency, and adjusting N-gram parameters, the extracted comment topics and keywords are adjusted to optimize the comment topics.

6. The method for dynamic analysis of service quality integrating hypergraph knowledge according to claim 1 is characterized in that: In step 4, the seed words of the five dimensions of the SERVQUAL model are constructed, and the keywords in each review topic and the seed words of the five dimensions of the SERVQUAL model are constructed into a word set; The service quality dimension point mutual information algorithm is used to calculate the correlation value between sets, and the review topics are dynamically mapped to the five dimensions of the SERVQUAL model. The service quality dimension point mutual information algorithm is expressed as follows: ； Where T represents the set of comment topic keywords, D represents the set of dimension seed words, represents the service quality dimension point mutual information value of the review topic T on dimension D, d represents a dimension seed word, represents the probability of the comment topic keyword t appearing, represents the probability of the dimension seed word d appearing, Represents the probability of the review topic keyword t and dimension seed word d appearing at the same time.

7. The method for dynamic analysis of service quality integrating hypergraph knowledge according to claim 1 is characterized in that: In step five, all customer review data in the dataset are divided into five equal parts, and then the FastText model is used to perform a five-fold cross-validation on the data. During the cross-validation process, the joint distribution matrix of the estimated noise labels and the true labels is obtained. Finally, the samples below the confidence threshold are screened out against the joint distribution matrix, which are the incorrectly labeled samples, and the rest are correctly labeled samples. The incorrectly labeled samples are then manually corrected.

8. The method for dynamic analysis of service quality integrating hypergraph knowledge according to claim 1 is characterized in that: The Wilson interval method is used to find the expected score of the merchant, which is expressed as follows: ； in, represents the Wilson score, is the upper limit of the rating. represents the rate of favorable comments, n is the total number of comments, and K is the statistic of the normal distribution at a fixed confidence level.

9. The method for dynamic analysis of service quality integrating hypergraph knowledge according to claim 8 is characterized in that: The expected rating of the merchant is corrected by the Bayesian average method, which is expressed as follows: ； Among them, h represents the number of merchant reviews, C represents the correction value of the merchant, and H represents the total average score of all merchants. Indicates the final expected merchant rating.

10. The method for dynamic analysis of service quality integrating hypergraph knowledge according to claim 1 is characterized in that: In step six, each customer review is constructed as a hypergraph. First, the customer reviews are segmented and the words are used as nodes of the graph. Then, a fixed-size sliding window is used to obtain global word co-occurrence as sequential context, which is used as the text structure hyperedge of the hypergraph. The words belonging to the same review topic in the node are connected with the same hyperedge as the intrinsic semantic hyperedge. External knowledge is obtained from the ConceptNet knowledge graph and external knowledge hyperedges are constructed. In the hypergraph processing stage, the constructed hypergraph is sent to the hypergraph attention network, which will output the customer service quality perception value contained in the customer reviews.

Images