Customer reputation index prediction method and device, electronic equipment and storage medium

Abstract

The application provides a customer reputation index prediction method and device, electronic equipment and a computer readable storage medium, and relates to the technical field of business intelligent analysis. The prediction method comprises the following steps: obtaining a dialogue text of a customer and a current service of a customer service; performing service classification on the dialogue text to obtain a customer label of the current service; performing emotion classification on the dialogue text to obtain a customer emotion label of the current service; and predicting a customer reputation index of the current service based on the customer label and the customer emotion label of the current service. At least, the problems of large amount of text analysis data, poor emotion analysis accuracy and poor customer reputation prediction accuracy in the related art are solved. The application is suitable for business satisfaction prediction and evaluation scenarios.

Description

Technical Field

[0001] This invention relates to the field of business intelligence analysis technology, and in particular to a method, apparatus, electronic device, and computer-readable storage medium for predicting customer reputation index. Background Technology

[0002] Currently, the basic communication services and data traffic services provided by major telecommunications operators are largely similar, while the operators' reputation and customer satisfaction significantly influence customers' choices of communication services. Each customer has a comprehensive rating for communication service products (such as mobile, fixed-line, repair, and after-sales service), which is known as the customer reputation index.

[0003] However, existing technologies typically use customer conversations directly for analysis, resulting in large amounts of text analysis data and poor accuracy in sentiment analysis, ultimately affecting the accuracy of customer reputation prediction results. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the above-mentioned shortcomings of the prior art by providing a method, apparatus, electronic device, and computer-readable storage medium for predicting customer reputation index. This method can improve the efficiency and accuracy of customer sentiment classification and service classification, enhance the adaptability of customer sentiment tags in predicting customer reputation index for different services, reduce the interference and errors caused by customer sentiment tags in the process of predicting customer reputation index, thereby improving the efficiency and accuracy of customer reputation index prediction and achieving rapid and accurate prediction of customer reputation index.

[0005] In a first aspect, the present invention provides a method for predicting customer reputation index, comprising: acquiring the dialogue text between a customer and customer service regarding the current service; classifying the dialogue text into services to obtain customer tags for the current service; classifying the dialogue text into emotions to obtain customer emotion tags for the current service; and predicting the customer reputation index for the current service based on the customer tags and customer emotion tags for the current service.

[0006] Preferably, the step of classifying the dialogue text for emotion to obtain the current service's customer emotion tag specifically includes: extracting a summary of the dialogue text based on an abstract text summarization model to obtain the summary text, wherein the abstract text summarization model includes a pre-trained PEGASUS model for generating summaries; and classifying the summary text for emotion based on an emotion classification model to obtain the current service's customer emotion tag.

[0007] Preferably, the emotion classification model includes the self-attention mechanism of the BERT model and a fully connected layer of the classifier. The step of classifying the summary text based on the emotion classification model to obtain the current service's customer emotion label specifically includes: converting the summary text into a context-dependent word vector matrix based on the self-attention mechanism of the BERT model; and classifying the word vector matrix based on the fully connected layer of the classifier to obtain the current service's customer emotion label.

[0008] Preferably, the self-attention mechanism of the BERT model includes a masked language modeling (MLM) task and a next-sentence prediction (NSP) task. The BERT-based self-attention mechanism transforms the summary text into a context-sensitive word vector matrix, specifically including: predicting the relationships between words in the summary text based on the masked language modeling (MLM) task; predicting the relationships between sentences in the summary text based on the next-sentence prediction (NSP) task; and transforming the summary text into a context-sensitive word vector matrix based on the relationships between words and the relationships between sentences.

[0009] Preferably, the step of predicting the customer reputation index of the current service based on the customer tags and customer sentiment tags of the current service specifically includes: determining the score corresponding to the customer sentiment tag of the current service based on the preset relationship between the customer sentiment tag and the score; assigning weights to the customer tags of the current service based on the preset relationship between the customer tags and the weights, thereby obtaining the weights of the customer tags of the current service; and predicting the customer reputation index of the current service based on the weights of the customer tags of the current service and the scores corresponding to the customer sentiment tags.

[0010] Preferably, the customer tags include service type tags and various service indicator tags under the service type. Before assigning weights to the customer tags of the current service based on the preset relationship between customer tags and weights, and obtaining the weights of the customer tags of the current service, the customer reputation index prediction method further includes: obtaining a sample evaluation data matrix A = (X... ij ) m×n Where m represents the total number of customers who have rated the service, n represents the total number of service indicator tags under each service type, and X ij Let Y represent the score of the j-th service indicator label for the i-th customer; standardize the sample evaluation data matrix to obtain the sample evaluation standard data matrix B = (Y ij ) m×n , where Y ij This represents the score of the j-th service indicator label for the i-th customer after standardization; based on the entropy weight method and the sample evaluation standard data matrix B = (Y ij ) m×nCalculate the weight of each service indicator tag; associate each service indicator tag with its corresponding weight to form a preset relationship between customer tags and weights.

[0011] Preferably, the data matrix B = (Y) based on the entropy weight method and sample evaluation criteria... ij ) m×n The weights of each service indicator label are calculated, specifically including: according to formulas (1) and (2), the weights of each service indicator label are calculated:

[0012]

[0013] in, If P ij =0, then define P ij ·ln(P ij ) = 0, P ij E represents the percentage of the score for the j-th service metric tag for the i-th customer. j W represents the information entropy of the j-th service metric label. j This represents the weight of the j-th service metric label.

[0014] Preferably, the step of predicting the customer reputation index of the current service based on the weight of the customer tags and the scores corresponding to the customer sentiment tags specifically includes: predicting the customer reputation index of the current service according to formula (3):

[0015] Among them, S i This represents the customer reputation index of the current service, where n represents the total number of service indicator tags under the current service type, and W represents the customer reputation index of the current service. j X represents the weight of the j-th service metric label under the current service type. ij This indicates the score of the customer's sentiment label for the current service.

[0016] Secondly, the present invention also provides a customer reputation index prediction device, comprising an acquisition module, a first classification module, a second classification module, and a prediction module. The acquisition module is used to acquire the dialogue text between the customer and the customer service representative for the current service. The first classification module is connected to the acquisition module and is used to classify the dialogue text for the current service to obtain a customer tag for the current service. The second classification module is connected to the acquisition module and is used to classify the dialogue text for the current service for the current service for the current emotion tag. The prediction module is connected to the first classification module and the second classification module respectively and is used to predict the customer reputation index for the current service based on the customer tag and the customer emotion tag for the current service.

[0017] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to implement the customer reputation index prediction method provided in the first aspect above.

[0018] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the customer reputation index prediction method provided in the first aspect above.

[0019] This invention provides a method, apparatus, electronic device, and computer-readable storage medium for predicting customer reputation index. By classifying customer and customer service dialogue text into emotion and service categories, it can more quickly and accurately identify customer needs and emotional states, improving the efficiency and accuracy of customer emotion and service classification. This also enhances the data accuracy of customer emotion tags and customer tags in the customer reputation index prediction process, thereby improving the efficiency and accuracy of customer reputation index prediction. Furthermore, by comprehensively considering the influence of customer emotion tags and customer tags during prediction, the invention allows customer emotion tags to better adapt to different service scenarios, reducing interference and errors caused by customer emotion tags in the prediction process, thus improving the efficiency and accuracy of customer reputation index prediction. Therefore, this invention can improve the efficiency and accuracy of customer emotion and service classification, enhance the adaptability of customer emotion tags to customer reputation index prediction across different services, reduce interference and errors caused by customer emotion tags in the prediction process, and ultimately improve the efficiency and accuracy of customer reputation index prediction, achieving rapid and accurate prediction of the customer reputation index. Attached Figure Description

[0020] Figure 1 This is a flowchart of a customer reputation index prediction method according to Embodiment 1 of the present invention;

[0021] Figure 2 This is a flowchart illustrating the extraction of a summary of dialogue text based on an abstract text summarization model in Embodiment 1 of the present invention.

[0022] Figure 3 This is a flowchart of the emotion classification of summary text based on an emotion classification model in Embodiment 1 of the present invention;

[0023] Figure 4 This is a flowchart of a customer reputation index prediction method according to Embodiment 2 of the present invention;

[0024] Figure 5 This is a schematic diagram of a customer reputation index prediction device according to Embodiment 3 of the present invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solution of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0026] It is understood that the specific embodiments and accompanying drawings described herein are merely for explaining the invention and are not intended to limit the invention.

[0027] It is understood that, without conflict, the various embodiments and features in the embodiments of the present invention can be combined with each other.

[0028] It is understood that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, while the parts unrelated to the present invention are not shown in the drawings.

[0029] It is understood that each unit or module involved in the embodiments of the present invention may correspond to only one entity structure, or may be composed of multiple entity structures, or multiple units or modules may be integrated into one entity structure.

[0030] It is understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of this invention may occur in a different order than that marked in the accompanying drawings.

[0031] It is understood that the flowcharts and block diagrams of this invention illustrate the possible architecture, functions, and operations of systems, apparatuses, devices, and methods according to various embodiments of this invention. Each block in the flowchart or block diagram may represent a unit, module, program segment, or code, containing executable instructions for implementing the specified function. Furthermore, each block or combination of blocks in the block diagram and flowchart can be implemented using a hardware-based system to achieve the specified function, or using a combination of hardware and computer instructions.

[0032] It is understood that the units and modules involved in the embodiments of the present invention can be implemented by software or by hardware. For example, the units and modules can be located in a processor.

[0033] Example 1:

[0034] like Figure 1 As shown, this embodiment provides a method for predicting customer reputation index.

[0035] In this embodiment, the prediction of customer reputation index is mainly applied to mobile network service scenarios, but the same steps but with different details can still be taken to predict the customer reputation index for fixed network service scenarios. Fault repair and after-sales service belong to the service ticket category and are typically used as customer satisfaction perception tags for participating in the prediction of customer reputation index for mobile or fixed network services.

[0036] Methods for predicting customer reputation index include:

[0037] S101, retrieve the current conversation text between the customer and customer service.

[0038] In this embodiment, the dialogue text data between the customer and the customer service department is obtained, and the dialogue text data is preprocessed to obtain the dialogue text between the customer and the customer service department. The data preprocessing includes, but is not limited to, desensitization processing, removal of irrelevant symbols and emoticons, cleaning of text content, and counting the text length of the dialogue text.

[0039] S102, classify the dialogue text into services to obtain the customer tags for the current service.

[0040] Specifically, customer tags include service type tags and various service indicator tags under each service type.

[0041] In this embodiment, a large amount of training dialogue text data between customers and customer service personnel is obtained from the customer service department. This training dialogue text data covers different service scenarios and customer groups. The training dialogue text data is preprocessed to obtain training dialogue texts, and the text length of these training dialogue texts is calculated. Based on in-depth analysis of customer service expert experience, a portion of the training dialogue texts is manually labeled, and the labeled dataset is divided into training and test sets. A text classification model is selected, such as logistic regression, support vector machines (SVM), LSTM (Long Short-Term Memory) model, or BERT (Bidirectional Encoder Representations from Transformers) model. Based on the training and test sets, the text classification model is trained and optimized to obtain a service classification model. Based on the service classification model, the dialogue texts are classified according to service, resulting in customer tags for the current service. In this embodiment, the customer tags for the current service are presented in a customer-perceived view.

[0042] In this embodiment, all service customer tags include at least one of the following: complaint service perception, perception experience evaluation, business product perception, network service perception, channel service perception, customer loyalty, customer activity, customer sentiment perception, and service indicator tags under each service type. As shown in Table 1, the customer tags include 8 major service type tags and 29 service indicator tags under the 8 major service types.

[0043] Table 1 Customer Tags for All Services

[0044]

[0045]

[0046] S103, perform emotion classification on the dialogue text to obtain the current customer emotion label.

[0047] Specifically, S103: Perform emotion classification on the dialogue text to obtain the current customer emotion label, including steps S1031-S1032:

[0048] S1031, based on the abstract text summarization model, extracts a summary of the dialogue text to obtain the summary text. The abstract text summarization model includes a pre-trained PEGASUS model for generating summaries.

[0049] In this embodiment, the abstract text summarization model includes, but is not limited to: PEGASUS (Pre-training with Extracted Gap-sentences for Abstractive Summarization), S2S (Sequence-to-Sequence Models, a traditional sequence-to-sequence model suitable for summarization tasks), and Led (Longformer Encoder-Decoder, a model specifically designed for processing long texts, suitable for summarizing long documents). Among these, the PEGASUS model performs exceptionally well in summarization tasks and has a wide range of applications. Due to different summarization task requirements (such as Chinese and English formats), the PEGASUS model is typically optimized and fine-tuned. The Randeng-Pegasus-238M-Summary-Chinese model is a variant of the PEGASUS model, specifically optimized and fine-tuned for Chinese text summarization tasks. Therefore, this embodiment takes the Randeng-Pegasus-238M-Summary-Chinese model, which uses the PEGASUS model as the abstract text summarization model, as an example. After obtaining a large amount of training dialogue text data between customers and customer service personnel from the customer service department, preprocessing the training dialogue text data to obtain training dialogue text, and counting the text length of the training dialogue text, the embodiment further includes: training and optimizing the Randeng-Pegasus-238M-Summary-Chinese model based on the training dialogue text to obtain the abstract text summarization model. Based on the abstract text summarization model, the dialogue text is summarized to obtain the summary text. The abstract text summarization model selects important sentences for prediction by occluding complete sentences and using uneven sampling, combining autoregressive and autoencoder structures to generate a coherent and well-organized summary, such as... Figure 2As shown in the figure. This embodiment utilizes the PEGASUS model to capture key information from dialogue text. The generated summary text maintains the integrity and coherence of the information, making the summary text more readable and grammatically complete. This makes the foundation for sentiment analysis more robust. Furthermore, the PEGASUS model can generate high-quality summaries in a relatively short time, making it suitable for fast-response business scenarios. It not only improves the quality and accuracy of the summaries, providing a clearer and more direct information basis for subsequent sentiment classification, but also improves the efficiency of subsequent sentiment classification and its applicability to business scenarios.

[0050] The PEGASUS model is a generative text summarization model based on a Transformer decoder architecture. Therefore, the core architecture of this abstract text summarization model consists of multiple identical layers, each containing three sub-layers: a self-attention mechanism, multi-head attention, and a feedforward neural network. The calculation formula for the self-attention mechanism is as follows: Where Q, K, and V represent the query, key, and value matrices, respectively, and d k This represents the dimension of the key vector. The multi-head attention mechanism uses the formula MultiHead(Q,K,V) = Concat(head1,…,head). h W I Self-attention is broken down into multiple "heads". Furthermore, the calculation formula for the feedforward neural network is FFN(x) = max(0, xW1+b1)W2+b2.

[0051] Abstract text summarization models include an optimized Masked Language Model (MLM) task. The optimized MLM task extracts and masks a series of sentence fragments (i.e., real fragments) from dialogue text, then trains a masked language model to predict the real fragments, thus obtaining the predicted fragments. The loss function of the optimized MLM task is... During the pre-training and fine-tuning phases, the masked language model uses the cross-entropy loss function. This is used to measure the difference between predicted and actual segments. Through the optimized MLM task, the abstract text summarization model can learn how to generate concise and accurate summaries from long texts, and optimize for specific tasks during the fine-tuning phase.

[0052] S1032, Based on the emotion classification model, perform emotion classification on the summary text to obtain the current service's customer emotion label.

[0053] In this embodiment, customer sentiment tags are used to characterize the emotional state or attitude expressed by customers during service interactions. Customer sentiment tags can help service providers better understand customer needs and feelings, thereby improving services and products. The abstract text generated by the abstract text summarization model is used as input data for the sentiment classification model. The sentiment classification model performs sentiment classification on the abstract text and outputs the current service's customer sentiment tag, such as... Figure 3 As shown, the abstract text can undergo further data preprocessing before being input into the sentiment classification model to ensure clarity and highlight key information.

[0054] It should be noted that before classifying the sentiment of the abstract text based on the sentiment classification model, the process also includes: fine-tuning the sentiment classification model through backpropagation algorithm to adapt it to the text classification task in a specific scenario; evaluating the performance of the sentiment classification model using a validation set or test set; and adjusting the hyperparameters or model architecture of the sentiment classification model according to actual needs to achieve the best results.

[0055] Fine-tuning the emotion classification model involved dividing the summary text into training and validation sets. Two strategies were tested: freezing the parameters of the bottom 1-6 layers of the emotion classification model and not freezing the parameters of the bottom 1-6 layers. Results showed that freezing the parameters significantly improved the performance of emotion classification, reduced the risk of overfitting, and enhanced the generalization ability of the emotion classification model. Therefore, this embodiment adopts the strategy of freezing the parameters of the bottom 1-6 layers of the pre-trained model.

[0056] Specifically, the emotion classification model includes the self-attention mechanism of the BERT model and the fully connected layer of the classifier.

[0057] Specifically, S1032: Based on the sentiment classification model, the summary text is classified into sentiment to obtain the current service's customer sentiment label, including: a self-attention mechanism based on the BERT model to convert the summary text into a context-related word vector matrix; and a fully connected layer based on the classifier to classify the word vector matrix into sentiment to obtain the current service's customer sentiment label.

[0058] Specifically, the self-attention mechanism of the BERT model includes the Masked Language Model (MLM) task and the Next Sentence Prediction (NSP) task.

[0059] Specifically, based on the self-attention mechanism of the BERT model, the summarized text is transformed into a context-dependent word vector matrix, including: predicting the relationship between words in the summarized text based on the Masked Language Model (MLM) task; predicting the relationship between sentences in the summarized text based on the Next Sentence Prediction (NSP) task; and transforming the summarized text into a context-dependent word vector matrix based on the relationship between words and the relationship between sentences.

[0060] In this embodiment, similarly to choosing the Randeng-Pegasus-238M-Summary-Chinese model from the PEGASUS model as the abstract text summarization model, the BERT model is usually optimized and fine-tuned due to different text vectorization task requirements (such as Chinese and English formats). BERT-base-chinese is a specific variant of the BERT model, specifically pre-trained for Chinese text. Therefore, this embodiment uses the BERT-base-chinese model. The BERT model is based on the Transformer architecture and utilizes a self-attention mechanism to understand the contextual information of the input text. The BERT model itself does not have a built-in classification module; therefore, as the foundational model for sentiment classification, its main function is to generate context-related word vector matrices. Combining the BERT model's self-attention mechanism and classification module (i.e., the fully connected layer of the classifier), a sentiment classification model is generated, such as... Figure 3 As shown in the figure. This embodiment adds a fully connected layer on top of the BERT model and uses the word vector matrix processed by the self-attention mechanism for emotion classification. Compared with the traditional word vector method, it can significantly improve the accuracy of emotion classification, better identify complex emotional expressions and make more accurate decisions, and analyze customer emotions in real-time dialogues to provide rapid feedback and help provide more personalized services.

[0061] The BERT model's self-attention mechanism performs two tasks: MLM (Multi-Level Modeling), which randomly masks a subset of words in a sentence and predicts the masked words; and NSP (Next Sentence Prediction), which predicts whether two sentences have logical continuity. Therefore, this embodiment uses MLM and NSP tasks to learn the internal representations of words and the relationships between sentences in the summary text, transforming the summary text into a context-dependent word vector matrix. Through MLM, the BERT model can understand the meaning of each word in different contexts, allowing the same word to be represented as different vectors in different contexts. This captures the subtle differences between polysemous and synonymous words, and the generated word vectors are not merely fixed embeddings but dynamically generated based on surrounding words, making the BERT model more flexible in handling context. Through the NSP task, the BERT model determines whether two sentences are adjacent in the text, helping to understand the logical relationships and semantic connections between sentences. This allows for a better understanding of how sentences in the summary connect, improving the context-dependent nature of the word vectors. This makes the BERT model more resistant to noise or bias in the text, improving its performance in real-world applications.

[0062] Through MLM and NSP tasks, the internal representations of words and the relationships between sentences in the summarized text are learned. Specifically, the summarized text is transformed into three embedding vectors: Token Embeddings, Segment Embeddings, and Position Embeddings. These three embedding vectors are then added together to form the final input vector E = T + S + P, where T is the token embeddings, S is the segment embeddings, and P is the position embeddings. The input vector is then fed into the self-attention mechanism of the BERT model. In the encoding layer (i.e., the self-attention mechanism of the BERT model), the input vector at each position is processed by a multi-head self-attention mechanism. This multi-head self-attention mechanism allows the BERT model to consider all positions in the sequence simultaneously, and its calculation method is equivalent to that of the self-attention mechanism in the abstract text summarization model. After the input vector at each position is processed by the multi-head self-attention mechanism, it is further processed by feedforward neural networks (FFNNs) and Layer Normalization (LN).

[0063] Before the BERT model's self-attention mechanism performs the two tasks, it also includes: preprocessing the summary text. This preprocessing includes, but is not limited to, word segmentation and adding special symbols. The special symbols include, but are not limited to, the `[CLS]` flag and the `[SEP]` flag. The `[CLS]` flag is placed at the beginning of the first sentence of the summary text. After the BERT model's self-attention mechanism processes the summary text, the `[CLS]` flag is transferred to the word vector corresponding to the first sentence for use in the subsequent classification task of the fully connected layer. The `[SEP]` flag is used to separate the two sentences of the summary text. For example, if the summary text includes sentences A and B, the `[SEP]` flag needs to be added between sentences A and B. Accordingly, after the BERT model's self-attention mechanism processes the summary text, the `[SEP]` flag is transferred between the word vectors corresponding to sentences A and B.

[0064] After the self-attention mechanism of the BERT model performs two tasks, the fully connected layer receives the hidden state h from the `[CLS]` token (i.e., the word vector carrying the `[CLS]` flag). [cls] and hide the state h [cls] Converting to a class probability distribution specifically includes: according to the formula logits = W cls h [cls] +b, calculate the hidden state h [cls]The unstandardized scores, and according to the formula The unstandardized scores are transformed into a class probability distribution using the softmax function, where W cls Let and represent the weight matrix and bias term of the classification layer, respectively. P(y = c|x) represents the probability of belonging to the c-th category given input x. If the classification task has N categories, the output dimension of the fully connected layer should be N, and the output of the fully connected layer is the probability of N categories. In this embodiment, N is 4, and the categories include satisfied, calm, dissatisfied, and untouched. The output of the fully connected layer includes the probabilities of satisfied, calm, dissatisfied, and untouched. The category with the highest probability is taken as the customer sentiment label for the current service.

[0065] S104, based on the customer tags and customer sentiment tags of the current service, predicts the customer reputation index of the current service.

[0066] Specifically, S104: Based on the customer tags and customer sentiment tags of the current service, predict the customer reputation index of the current service, including steps S1041-S1043:

[0067] S1041, Based on the preset relationship between customer sentiment tags and scores, determine the score corresponding to the customer sentiment tag for the current service.

[0068] In this embodiment, based on the experience of customer service experts, different scores are assigned to different customer emotion tags to determine the preset relationship between customer emotion tags and scores. For example, considering that most customers have a positive opinion of the company's services, which is in line with the principle of conservative estimation, the scores are set between the scores of satisfaction and calmness, with satisfaction being 2 points, calmness being 1 point, dissatisfaction being 0 points, and null value (i.e. no contact) being set at 1.5 points.

[0069] It should be noted that before determining the score corresponding to the customer sentiment tag of the current service based on the preset relationship between customer sentiment tags and scores, and assigning weights to the customer tags of the current service based on the preset relationship between customer tags and weights, binning and null value imputation methods are used to preprocess the customer sentiment tags and customer tags. This enhances the resistance of customer reputation index prediction to data noise and outliers, reduces overfitting, and improves the stability and robustness of customer reputation index prediction.

[0070] S1042, based on the preset relationship between customer tags and weights, assign weights to the customer tags of the current service to obtain the weights of the customer tags of the current service.

[0071] Optionally, before step S1042: assigning weights to the customer tags of the current service based on the preset relationship between customer tags and weights, and obtaining the weights of the customer tags of the current service, the customer reputation index prediction method further includes steps S1044-S1047:

[0072] S1044, Obtain the sample evaluation data matrix A = (X ij ) m×n Where m represents the total number of customers who have rated the service, n represents the total number of service indicator tags under each service type, and X ij This represents the score of the j-th service indicator label for the i-th customer.

[0073] S1045, standardize the sample evaluation data matrix to obtain the sample evaluation standard data matrix B = (Y ij ) m×n , where Y ij This represents the score of the j-th service indicator label for the i-th customer after standardization.

[0074] S1046, Based on the entropy weight method and sample evaluation standard data matrix B = (Y ij ) m×n Calculate the weight of each service indicator label.

[0075] Specifically, S1046: Based on the entropy weight method and sample evaluation criteria, the data matrix B = (Y ij ) m×n Calculate the weight of each service indicator label, including: Calculate the weight of each service indicator label according to formulas (1) and (2):

[0076]

[0077] in, If P ij =0, then define P ij ·ln(P ij ) = 0, P ij E represents the percentage of the score for the j-th service metric tag for the i-th customer. j W represents the information entropy of the j-th service metric label. j This represents the weight of the j-th service metric label.

[0078] S1047, associate each service indicator label with its corresponding weight to form a preset relationship between customer labels and weights.

[0079] In this embodiment, the weights are used to characterize the influence of customer sentiment tags under different service indicator labels on the prediction of the service's customer reputation index. Therefore, the weights can reflect the importance of the customer sentiment tag score to the service. The entropy weight method is an objective weighting method based on information entropy theory. Its principle is to determine the weight based on the amount of information provided by each indicator; the greater the information, the higher the weight. Specifically, it calculates the entropy value of each indicator to reflect the differences between them and thus determine the weight of each indicator. The smaller the entropy value, the greater the differences, the more information provided, and the higher the weight. This embodiment uses the entropy weight method and a sample evaluation standard data matrix to determine the weights of each service indicator label, thereby establishing a preset relationship between customer tags and weights. This method is highly objective, reducing the influence of subjective judgment. Furthermore, it dynamically adjusts the weights of customer sentiment tags in the prediction of the customer reputation index for different services using service tags, improving the accuracy and fairness of customer sentiment tags in the customer reputation index prediction process.

[0080] It should be noted that after calculating the weights of each service indicator tag, the service indicator tags are sorted based on their weights to form a pre-defined relationship between customer tags and weights. This allows for a visual representation of the impact of customer sentiment tags under each service indicator tag on the prediction of the service's customer reputation index. For example, customer sentiment tags under the tags of customer emotional value, terminal contract customers, special monthly package product orders, uninformed customization, and appealing customers have the greatest impact on the prediction of the customer reputation index, while simulated assessments, official account message customers, real-time assessments, and self-payment have the least impact on the prediction of the customer reputation index.

[0081] S1043, based on the weight of the customer tags of the current service and the scores corresponding to the customer sentiment tags, predict the customer reputation index of the current service.

[0082] Specifically, S1043: Based on the weights of the customer tags and the scores corresponding to the customer sentiment tags of the current service, predict the customer reputation index of the current service, including: predicting the customer reputation index of the current service according to formula (3):

[0083] Among them, S i This represents the customer reputation index of the current service, where n represents the total number of service indicator tags under the current service type, and W represents the customer reputation index of the current service. j X represents the weight of the j-th service metric label under the current service type. ij This indicates the score of the customer's sentiment label for the current service.

[0084] In this embodiment, the customer reputation index S for the current service is calculated according to formula (3). i Then, calculate the customer satisfaction index (S) for each customer regarding the current service. iWe obtain the customer reputation index of all customers for the current service, and then convert the customer reputation index of all customers for the current service into a percentage system.

[0085] It should be noted that the customer satisfaction index S for each customer regarding the current service is calculated. i Next, based on customer satisfaction with the service, customers were categorized into four groups: 1-2 for venting customers, 3-6 for expectant customers, 7-8 for neutral customers, and 9-10 for appreciative customers. Then, a customer reputation index (S) was assigned to each customer regarding the current service. i The process is then extended to the corresponding customers. The customer reputation index of all customers under each customer type for the current service is observed, and the distribution threshold of the customer reputation index for each customer type is determined. There are three distribution thresholds, namely a, b, and c, which are used to directly determine the prediction accuracy of each customer type after calculating the customer reputation index of each customer for the service.

[0086] The prediction accuracy of each type of customer is determined based on the distribution threshold. Specifically, this includes: counting the total number of each type of customer (i.e., the number of customers of each type) and the number of customers whose customer reputation index meets the distribution threshold under each type of customer (i.e., the number of accurate predictions); and calculating the prediction accuracy of each type of customer based on the total number of each type of customer and the number of customers whose customer reputation index meets the distribution threshold under each type of customer, as shown in Table 2.

[0087] Table 2 Prediction Accuracy of Different Customer Types

[0088] Type of customer Number of customers by type Prediction accuracy Prediction accuracy Approval type 923 835 90％ Neutral 442 248 56％ Expectant 223 122 55％ Confessional type 57 28 49％ total 1645 1233 75％

[0089] After counting the total number of customers of each type (i.e., the number of customers by type) and the number of customers whose customer reputation index meets the distribution threshold for each customer type (i.e., the number of accurate predictions), this embodiment further calculates the customer reputation index hit rate based on the total number of customers and the number of customers whose customer reputation index meets the distribution threshold for all customer types. Specifically, this includes:

[0090] According to the formula P = (N_praise + N_neutral + N_expect + N_confide) / N_total, calculate the hit rate of the customer word-of-mouth index. Among them, N_total represents the total number of customers, N_praise represents the number of accurate predictions of praise-type customers, and N_praise is used to represent the number of customers who are actually praise-type and are also predicted as praise-type (i.e., m > c). N_neutral represents the number of accurate predictions of neutral-type customers, and N_neutral is used to represent the number of customers who are actually neutral-type and are also predicted as neutral-type (b < m ≤ c). N_expect represents the number of accurate predictions of expectant-type customers, and N_expect is used to represent the number of customers who are actually expectant-type and are also predicted as expectant-type (a < m ≤ b). N_confide represents the number of accurate predictions of confiding-type customers, and N_confide is used to represent the number of customers who are actually confiding-type and are also predicted as confiding-type (m ≤ a). m represents the customer word-of-mouth index calculated for each customer's service.

[0091] In addition, by comparing the hit rates of the customer word-of-mouth index under different distribution threshold combinations, select the three distribution thresholds corresponding to the optimal hit rate as the target distribution thresholds. The overall prediction accuracy rate reaches 75%, meeting the actual operation needs and having the application conditions.

[0092] This embodiment provides a method for predicting customer reputation index. By classifying customer and customer service dialogue texts into emotion and service categories, it can more quickly and accurately identify customer needs and emotional states, improving the efficiency and accuracy of customer emotion and service classification. This enhances the data accuracy of customer emotion tags and customer tags in the customer reputation index prediction process, thereby improving the efficiency and accuracy of customer reputation index prediction. Furthermore, by comprehensively considering the impact of customer emotion tags and customer tags during prediction, the method allows customer emotion tags to better adapt to different service scenarios, reducing interference and errors caused by customer emotion tags in the prediction process. This further improves the efficiency and accuracy of customer reputation index prediction, enhances the efficiency and accuracy of customer emotion and service classification, and improves the adaptability of customer emotion tags to customer reputation index prediction across different services. Ultimately, this method achieves rapid and accurate prediction of the customer reputation index. Furthermore, the PEGASUS model can capture key information from dialogue text, generating summary text that maintains information integrity and coherence, making the summary text more readable and grammatically complete. This strengthens the foundation for sentiment analysis. The PEGASUS model can also generate high-quality summaries in a relatively short time, suitable for fast-response business scenarios. It not only improves the quality and accuracy of summaries, providing a clearer and more direct information basis for subsequent sentiment classification, but also improves the efficiency and applicability of subsequent sentiment classification to business scenarios. By adding a fully connected layer on top of the BERT model and using a word vector matrix processed by a self-attention mechanism for sentiment classification, compared with traditional word vector methods, it can significantly improve the accuracy of sentiment classification, better identify complex emotional expressions and make more accurate decisions. It can also analyze customer emotions in real-time dialogues, provide rapid feedback, and help provide more personalized services. Through MLM, the BERT model can understand the meaning of each word in different contexts, allowing the same word to be represented as different vectors in different contexts. This captures the subtle differences between polysemous and synonymous words. The generated word vectors are not just fixed embeddings, but are dynamically generated based on surrounding words, making the BERT model more flexible in handling context. Through the NSP task, the BERT model determines whether two sentences are adjacent in the text, helping to understand the logical relationships and semantic connections between sentences. This allows for a better understanding of how sentences in a summary connect with each other, improving the contextual relevance of word vectors. This makes the BERT model more resistant to noise or bias in the text, thus improving its performance in real-world applications.By using the entropy weight method and a sample evaluation standard data matrix, the weights of each service indicator label are determined to establish a pre-defined relationship between customer labels and weights. This method is highly objective and can reduce the influence of subjective judgment. Subsequently, the weights of customer sentiment labels in the prediction of customer reputation index for different services are dynamically adjusted using service labels, thereby improving the accuracy and fairness of customer sentiment labels in the prediction of customer reputation index.

[0093] Example 2:

[0094] like Figure 4 As shown, this embodiment provides a method for predicting customer reputation index. The method for predicting customer reputation index includes:

[0095] S201. Obtain the dialogue text between the customer and customer service for the current service, classify the dialogue text for the service, obtain the customer tag for the current service, and extract the summary of the dialogue text based on the abstract text summarization model to obtain the summary text. The abstract text summarization model includes the pre-trained PEGASUS model for generating summaries.

[0096] In this embodiment, the abstract text summarization model is... Figure 4 In the abstract model, the abstract text is... Figure 4 The user dialogue summary is obtained from the customer service department. The dialogue text data between the customer and the customer service is obtained, and the dialogue text data is preprocessed to obtain the dialogue text between the customer and the customer service. The data preprocessing includes, but is not limited to, de-identification, removal of irrelevant symbols and emoticons, text cleaning, and counting the text length of the dialogue text.

[0097] A large amount of training dialogue text data between customers and customer service personnel was obtained from the customer service department. This training dialogue text data covers different service scenarios and customer groups. Data preprocessing was performed on the training dialogue text data to obtain the training dialogue texts, and the text length of the training dialogue texts was calculated. Based on in-depth analysis of customer service expert experience, a portion of the training dialogue texts was manually labeled, and the labeled dataset was divided into training and test sets. A text classification model was selected, such as logistic regression, support vector machine, LSTM model, or BERT model, and the text classification model was trained and optimized based on the training and test sets to obtain a service classification model. Based on the service classification model, the dialogue texts were classified according to service to obtain the customer tags for the current service. In this embodiment, the customer tags for the current service are presented in a customer-perceived view.

[0098] In this embodiment, all service customer tags include at least one of the following: complaint service perception, perception experience evaluation, business product perception, network service perception, channel service perception, customer loyalty, customer activity, customer sentiment perception, and service indicator tags under each service type. As shown in Table 1, the customer tags include 8 major service type tags and 29 service indicator tags under the 8 major service types.

[0099] Abstract text summarization models include, but are not limited to, the PEGASUS model, the S2S model, and the Led model. Among them, the PEGASUS model performs exceptionally well in summarization tasks and has a wide range of applications. Due to different summarization task requirements (such as Chinese and English formats), the PEGASUS model is usually optimized and fine-tuned. The Randeng-Pegasus-238M-Summary-Chinese model is a variant of the PEGASUS model, specifically optimized and fine-tuned for Chinese text summarization tasks. Therefore, this embodiment takes the Randeng-Pegasus-238M-Summary-Chinese model, which uses the PEGASUS model as its abstract text summarization model, as an example. After obtaining a large amount of training dialogue text data between customers and customer service personnel from the customer service department, preprocessing the training dialogue text data to obtain the training dialogue text, and calculating the text length of the training dialogue text, the embodiment further includes: training and optimizing the Randeng-Pegasus-238M-Summary-Chinese model based on the training dialogue text to obtain the abstract text summarization model. Abstract text summarization (AMS) is used to extract summaries from dialogue text. The AMS model occludes complete sentences and uses uneven sampling to select important sentences for prediction. Combining autoregressive and autoencoder structures, it generates coherent and well-organized summaries, such as... Figure 2 As shown in the figure. This embodiment utilizes the PEGASUS model to capture key information from dialogue text. The generated summary text maintains the integrity and coherence of the information, making the summary text more readable and grammatically complete. This makes the foundation for sentiment analysis more robust. Furthermore, the PEGASUS model can generate high-quality summaries in a relatively short time, making it suitable for fast-response business scenarios. It not only improves the quality and accuracy of the summaries, providing a clearer and more direct information basis for subsequent sentiment classification, but also improves the efficiency of subsequent sentiment classification and its applicability to business scenarios.

[0100] The PEGASUS model is a generative text summarization model based on a Transformer decoder architecture. Therefore, the core architecture of this abstract text summarization model consists of multiple identical layers, each containing three sub-layers: a self-attention mechanism, multi-head attention, and a feedforward neural network. The calculation formula for the self-attention mechanism is as follows: Where Q, K, and V represent the query, key, and value matrices, respectively, and d k This represents the dimension of the key vector. The multi-head attention mechanism uses the formula MultiHead(Q,K,V) = Concat(head1,…,head). h W O Self-attention is broken down into multiple "heads". Furthermore, the calculation formula for the feedforward neural network is FFN(x) = max(0, xW1+b1)W2+b2.

[0101] Abstract text summarization models include an optimized masked language modeling (MLM) task. The optimized MLM task extracts and masks a series of sentence fragments (i.e., real fragments) from dialogue text, then trains a masked language model to predict the real fragments, thus obtaining the predicted fragments. The loss function of the optimized MLM task is... During the pre-training and fine-tuning phases, the masked language model uses the cross-entropy loss function. This is used to measure the difference between predicted and actual segments. Through the optimized MLM task, the abstract text summarization model can learn how to generate concise and accurate summaries from long texts, and optimize for specific tasks during the fine-tuning phase.

[0102] S202, based on the emotion classification model, performs emotion classification on the summary text to obtain the current service's customer emotion label. The emotion classification model includes the self-attention mechanism of the BERT model and the fully connected layer of the classifier.

[0103] In this embodiment, the customer emotion tag is... Figure 4 User sentiment in the context of service interaction. Customer sentiment tags are used to characterize the emotional state or attitude expressed by customers during service interactions. Customer sentiment tags help service providers better understand customer needs and feelings, thereby improving services and products. Using summary text generated by an abstract text summarization model as input data for the sentiment classification model, the model categorizes the sentiment of the summary text and outputs the current service's customer sentiment tag, such as... Figure 3 As shown, the abstract text can undergo further data preprocessing before being input into the sentiment classification model to ensure clarity and highlight key information.

[0104] It should be noted that before classifying the sentiment of the abstract text based on the sentiment classification model, the process also includes: fine-tuning the sentiment classification model through backpropagation algorithm to adapt it to the text classification task in a specific scenario; evaluating the performance of the sentiment classification model using a validation set or test set; and adjusting the hyperparameters or model architecture of the sentiment classification model according to actual needs to achieve the best results.

[0105] Fine-tuning the emotion classification model involved dividing the summary text into training and validation sets. Two strategies were tested: freezing the parameters of the bottom 1-6 layers of the emotion classification model and not freezing the parameters of the bottom 1-6 layers. Results showed that freezing the parameters significantly improved the performance of emotion classification, reduced the risk of overfitting, and enhanced the generalization ability of the emotion classification model. Therefore, this embodiment adopts the strategy of freezing the parameters of the bottom 1-6 layers of the pre-trained model.

[0106] Similarly, just as the Randeng-Pegasus-238M-Summary-Chinese variant of the PEGASUS model is chosen as the abstract text summarization model, the BERT model is usually optimized and fine-tuned to meet the needs of different text vectorization tasks (such as Chinese and English formats). BERT-base-chinese is a specific variant of the BERT model, specifically pre-trained for Chinese text. Therefore, this embodiment uses the BERT-base-chinese variant. The BERT model is based on the Transformer architecture and utilizes a self-attention mechanism to understand the contextual information of the input text. The BERT model itself does not have a built-in classification module; therefore, as the foundational model for sentiment classification, its main function is to generate context-dependent word vector matrices. Combining the BERT model's self-attention mechanism and classification module (i.e., the fully connected layer of the classifier), a sentiment classification model is generated, such as... Figure 3 As shown in the figure. This embodiment adds a fully connected layer on top of the BERT model and uses the word vector matrix processed by the self-attention mechanism for emotion classification. Compared with the traditional word vector method, it can significantly improve the accuracy of emotion classification, better identify complex emotional expressions and make more accurate decisions, and analyze customer emotions in real-time dialogues to provide rapid feedback and help provide more personalized services.

[0107] The BERT model's self-attention mechanism performs two tasks: MLM (Multi-Level Modeling), which randomly masks a subset of words in a sentence and predicts the masked words; and NSP (Non-Contextual Contextualization), which predicts whether two sentences have logical continuity. Therefore, this embodiment uses MLM and NSP tasks to learn the internal representations of words and the relationships between sentences in the summary text, transforming the summary text into a context-dependent word vector matrix. Through MLM, the BERT model can understand the meaning of each word in different contexts, allowing the same word to be represented as different vectors in different contexts, thus capturing the subtle differences between polysemous and synonymous words. The generated word vectors are not just fixed embeddings but are dynamically generated based on surrounding words, making the BERT model more flexible in handling context. Through the NSP task, the BERT model determines whether two sentences are adjacent in the text, helping to understand the logical relationships and semantic connections between sentences. This allows for a better understanding of how sentences in the summary connect with each other, improving the context-dependent nature of word vectors. This makes the BERT model more resistant to noise or bias in the text, improving its performance in real-world applications.

[0108] Through MLM and NSP tasks, the internal representations of words and the relationships between sentences in the summarized text are learned. Specifically, the summarized text is transformed into three embedding vectors: Token Embeddings, Segment Embeddings, and Position Embeddings. These three embedding vectors are then added together to form the final input vector E = T + S + P, where T is the token embeddings, S is the segment embeddings, and P is the position embeddings. The input vector is then fed into the self-attention mechanism of the BERT model. In the encoding layer (i.e., the self-attention mechanism of the BERT model), the input vector at each position is processed through a multi-head self-attention mechanism. This multi-head self-attention mechanism allows the BERT model to consider all positions in the sequence simultaneously, and its calculation method is equivalent to that of the self-attention mechanism in the abstract text summarization model. After the input vector at each position is processed through the multi-head self-attention mechanism, it is further processed by feedforward neural networks (FFNNs) and Layer Normalization.

[0109] Before the BERT model's self-attention mechanism performs the two tasks, it also includes: preprocessing the summary text. This preprocessing includes, but is not limited to, word segmentation and adding special symbols. The special symbols include, but are not limited to, the `[CLS]` flag and the `[SEP]` flag. The `[CLS]` flag is placed at the beginning of the first sentence of the summary text. After the BERT model's self-attention mechanism processes the summary text, the `[CLS]` flag is transferred to the word vector corresponding to the first sentence for use in the subsequent classification task of the fully connected layer. The `[SEP]` flag is used to separate the two sentences of the summary text. For example, if the summary text includes sentences A and B, the `[SEP]` flag needs to be added between sentences A and B. Accordingly, after the BERT model's self-attention mechanism processes the summary text, the `[SEP]` flag is transferred between the word vectors corresponding to sentences A and B.

[0110] After the self-attention mechanism of the BERT model performs two tasks, the fully connected layer receives the hidden state h from the `[CLS]` token (i.e., the word vector carrying the `[CLS]` flag). [cls] and hide the state h [cls] Converting to a class probability distribution specifically includes: according to the formula logits = W cls h [cls] +b, calculate the hidden state h [cls] The unstandardized scores, and according to the formula The unstandardized scores are transformed into a class probability distribution using the softmax function, where W cls Let and represent the weight matrix and bias term of the classification layer, respectively. P(y = c|x) represents the probability of belonging to the c-th category given input x. If the classification task has N categories, the output dimension of the fully connected layer should be N, and the output of the fully connected layer is the probability of N categories. In this embodiment, N is 4, and the categories include satisfied, calm, dissatisfied, and untouched. The output of the fully connected layer includes the probabilities of satisfied, calm, dissatisfied, and untouched. The category with the highest probability is taken as the customer sentiment label for the current service.

[0111] S203, based on the entropy weight method and the sample evaluation data matrix, determines the preset relationship between customer tags and weights.

[0112] In this embodiment, the preset relationship between customer tags and weights is... Figure 4 The customer perception view in the system.

[0113] S204. Based on the preset relationship between customer sentiment tags and scores, determine the score corresponding to the customer sentiment tag of the current service, and based on the preset relationship between customer tags and weights, assign weights to the customer tags of the current service to obtain the weights of the customer tags of the current service.

[0114] In this embodiment, the score corresponding to the customer's sentiment tag for the current service is determined. Figure 4 The system assigns emotional scores. Based on the experience of customer service experts, different scores are assigned to different customer emotional tags to determine the preset relationship between customer emotional tags and scores. For example, considering that most customers have a positive opinion of the company's services, which is in line with the principle of conservative estimation, the scores are set between satisfaction and calmness, with satisfaction being 2 points, calmness being 1 point, dissatisfaction being 0 points, and blank value (i.e., no contact) being set at 1.5 points.

[0115] It should be noted that before determining the score corresponding to the customer sentiment tag of the current service based on the preset relationship between customer sentiment tags and scores, and assigning weights to the customer tags of the current service based on the preset relationship between customer tags and weights, binning and null value imputation methods are used to preprocess the customer sentiment tags and customer tags. This enhances the resistance of customer reputation index prediction to data noise and outliers, reduces overfitting, and improves the stability and robustness of customer reputation index prediction.

[0116] Weights are used to characterize the influence of customer sentiment labels under different service indicator tags on the prediction of the service's customer reputation index. Therefore, weights can reflect the importance of the customer sentiment label score to the service. The entropy weight method is an objective weighting method based on information entropy theory. Its principle is to determine the weight based on the amount of information provided by each indicator; the greater the information, the higher the weight. Specifically, it calculates the entropy value of each indicator to reflect the differences between them, thereby determining the weight of each indicator. The smaller the entropy value, the greater the differences, the more information provided, and the higher the weight. This embodiment uses the entropy weight method and a sample evaluation standard data matrix to determine the weight of each service indicator tag, thus establishing a preset relationship between customer tags and weights. This method is highly objective, reducing the influence of subjective judgment. Subsequently, the weight of customer sentiment labels in the prediction of the customer reputation index for different services is dynamically adjusted using service tags, improving the accuracy and fairness of customer sentiment labels in the customer reputation index prediction process.

[0117] It should be noted that after calculating the weights of each service indicator tag, the service indicator tags are sorted based on their weights to form a pre-defined relationship between customer tags and weights. This allows for a visual representation of the impact of customer sentiment tags under each service indicator tag on the prediction of the service's customer reputation index. For example, customer sentiment tags under the tags of customer emotional value, terminal contract customers, special monthly package product orders, uninformed customization, and appealing customers have the greatest impact on the prediction of the customer reputation index, while simulated assessments, official account message customers, real-time assessments, and self-payment have the least impact on the prediction of the customer reputation index.

[0118] S205, based on the weight of the customer tags for the current service and the scores corresponding to the customer sentiment tags, predicts the customer reputation index for the current service.

[0119] In this embodiment, the customer reputation index S for the current service is calculated according to formula (3). i Then, calculate the customer satisfaction index (S) for each customer regarding the current service. i We obtain the customer reputation index of all customers for the current service, and then convert the customer reputation index of all customers for the current service into a percentage system.

[0120] It should be noted that the customer satisfaction index S for each customer regarding the current service is calculated. i Next, based on customer satisfaction with the service, customers were categorized into four groups: 1-2 for venting customers, 3-6 for expectant customers, 7-8 for neutral customers, and 9-10 for appreciative customers. Then, a customer reputation index (S) was assigned to each customer regarding the current service. i The process is then extended to the corresponding customers. The customer reputation index of all customers under each customer type for the current service is observed, and the distribution threshold of the customer reputation index for each customer type is determined. There are three distribution thresholds, namely a, b, and c, which are used to directly determine the prediction accuracy of each customer type after calculating the customer reputation index of each customer for the service.

[0121] The prediction accuracy of each type of customer is determined based on the distribution threshold. Specifically, this includes: counting the total number of each type of customer (i.e., the number of customers of each type) and the number of customers whose customer reputation index meets the distribution threshold under each type of customer (i.e., the number of accurate predictions); and calculating the prediction accuracy of each type of customer based on the total number of each type of customer and the number of customers whose customer reputation index meets the distribution threshold under each type of customer, as shown in Table 2.

[0122] After counting the total number of customers of each type (i.e., the number of customers by type) and the number of customers whose customer reputation index meets the distribution threshold for each customer type (i.e., the number of accurate predictions), this embodiment further calculates the customer reputation index hit rate based on the total number of customers and the number of customers whose customer reputation index meets the distribution threshold for all customer types. Specifically, this includes:

[0123] According to the formula P = (N_praise + N_neutral + N_expect + N_confide) / N_total, calculate the hit rate of the customer word-of-mouth index, where N_total represents the total number of customers, N_praise represents the number of accurate predictions of customers with a praise attitude, N_praise is used to represent the number of customers who are actually of a praise attitude and are also predicted to be of a praise attitude (i.e., m > c), N_neutral represents the number of accurate predictions of customers with a neutral attitude, N_neutral is used to represent the number of customers who are actually of a neutral attitude and are also predicted to be of a neutral attitude (b < m ≤ c), N_expect represents the number of accurate predictions of customers with an expectant attitude, N_expect is used to represent the number of customers who are actually of an expectant attitude and are also predicted to be of an expectant attitude (a < m ≤ b), N_confide represents the number of accurate predictions of customers with a confiding attitude, N_confide is used to represent the number of customers who are actually of a confiding attitude and are also predicted to be of a confiding attitude (m ≤ a), and m represents the customer word-of-mouth index calculated for each customer's service.

[0124] In addition, by comparing the hit rates of the customer word-of-mouth index under different combinations of distribution thresholds, select the three distribution thresholds corresponding to the optimal hit rate as the target distribution thresholds. The overall prediction accuracy rate reaches 75%, meeting the actual operation requirements and having the application conditions.

[0125] A prediction method for the customer word-of-mouth index provided in this embodiment can more quickly and accurately identify the needs and emotional states of customers by classifying the emotions and services in the conversation texts between customers and customer service, improving the efficiency and accuracy of customer emotion classification and service classification, enhancing the data accuracy of customer emotion labels and customer labels in the process of predicting the customer word-of-mouth index, and further improving the efficiency and accuracy of predicting the customer word-of-mouth index; comprehensively considering the influence brought by customer emotion labels and customer labels when predicting the customer word-of-mouth index, enabling customer emotion labels to better adapt to the prediction of the customer word-of-mouth index in different service scenarios, reducing the interference and errors brought by customer emotion labels in the process of predicting the customer word-of-mouth index, and further improving the efficiency and accuracy of predicting the customer word-of-mouth index, improving the efficiency and accuracy of customer emotion classification and service classification, enhancing the adaptability of customer emotion labels in the prediction of the customer word-of-mouth index for different services, reducing the interference and errors brought by customer emotion labels in the process of predicting the customer word-of-mouth index, and further improving the efficiency and accuracy of predicting the customer word-of-mouth index, achieving the rapid and accurate prediction of the customer word-of-mouth index.

[0126] Example 3:

[0127] As Figure 5As shown, this embodiment also provides a customer reputation index prediction device, including an acquisition module 31, a first classification module 32, a second classification module 33, and a prediction module 34. The acquisition module 31 is used to acquire the dialogue text between the customer and the customer service representative for the current service. The first classification module 32 is connected to the acquisition module 31 and is used to classify the dialogue text for the service to obtain the customer tag for the current service. The second classification module 33 is connected to the acquisition module 31 and is used to classify the dialogue text for the emotion to obtain the customer emotion tag for the current service. The prediction module 34 is connected to the first classification module 32 and the second classification module 33 respectively and is used to predict the customer reputation index for the current service based on the customer tag and the customer emotion tag for the current service.

[0128] Specifically, the second classification module 33 includes an extraction unit 331 and a classification unit 332. The extraction unit 331 is used to extract a summary of the dialogue text based on an abstract text summarization model to obtain the summary text. The abstract text summarization model includes a pre-trained PEGASUS model for generating summaries. The classification unit 332 is used to classify the summary text for emotion based on an emotion classification model to obtain the current service's customer emotion label.

[0129] Specifically, the classification unit 332 includes a transformation subunit and a classification subunit. The transformation subunit is used to transform the summary text into a context-dependent word vector matrix based on the self-attention mechanism of the BERT model. The classification subunit is used to perform sentiment classification on the word vector matrix based on the fully connected layer of the classifier to obtain the current service's customer sentiment label.

[0130] Specifically, the transformation subunit includes: a first prediction minimum unit, a second prediction minimum unit, and a transformation minimum unit. The first prediction minimum unit is used to predict the relationship between words in the summary text based on the Masked Language Model (MLM) task. The second prediction minimum unit is used to predict the relationship between sentences in the summary text based on the Next Sentence Prediction (NSP) task. The transformation minimum unit is used to transform the summary text into a context-dependent word vector matrix based on the relationship between words and the relationship between sentences.

[0131] Specifically, the prediction module 34 includes: a determination unit 341, a weighting unit 342, and a prediction unit 343. The determination unit 341 is used to determine the score corresponding to the customer sentiment tag of the current service based on the preset relationship between customer sentiment tags and scores. The weighting unit 342 is used to assign weights to the customer tags of the current service based on the preset relationship between customer tags and weights, thereby obtaining the weights of the customer tags of the current service. The prediction unit 343 is used to predict the customer reputation index of the current service based on the weights of the customer tags of the current service and the scores corresponding to the customer sentiment tags.

[0132] Optionally, the prediction module 34 further includes: an acquisition unit 344, a standardization processing unit 345, a calculation unit 346, and an association unit 347. The acquisition unit 344 is used to acquire the sample evaluation data matrix A = (X... ij ) m×n Where m represents the total number of customers who have rated the service, n represents the total number of service indicator tags under each service type, and X ij This represents the score of the j-th service indicator label for the i-th customer. Standardization unit 345 is used to standardize the sample evaluation data matrix, resulting in the sample evaluation standard data matrix B = (Y...). ij ) m×n , where Y ij This represents the score of the j-th service indicator label for the i-th customer after standardization. Calculation unit 346 is used based on the entropy weight method and the sample evaluation standard data matrix B = (Y ij ) m×n The weight of each service indicator tag is calculated. The association unit 347 is used to associate each service indicator tag with its corresponding weight to form a preset relationship between customer tags and weights.

[0133] Specifically, the calculation unit 346 includes: a calculation subunit for calculating the weights of each service indicator label according to formulas (1) and (2):

[0134]

[0135] in, If P ij =0, then define P ij ·ln(P ij ) = 0, P ij E represents the percentage of the score for the j-th service metric tag for the i-th customer. j W represents the information entropy of the j-th service metric label. j This represents the weight of the j-th service metric label.

[0136] Understandably, the customer reputation index prediction device provided above executes the customer reputation index prediction method corresponding to Embodiment 1 provided above. Therefore, the beneficial effects it can achieve can be referred to the beneficial effects of the scheme corresponding to the customer reputation index prediction method of Embodiment 1 above, which will not be repeated here.

[0137] Example 4:

[0138] This embodiment also provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to implement the customer reputation index prediction method in Embodiment 1 or Embodiment 2 described above.

[0139] Example 5:

[0140] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the customer reputation index prediction method in Embodiment 1 or Embodiment 2 above.

[0141] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting customer reputation index, characterized in that, include: Obtain the text of the current conversation between the customer and customer service; The dialogue text is categorized by service to obtain the customer tags for the current service; The dialogue text is classified by emotion to obtain the current customer emotion tag; Based on the customer tags and customer sentiment tags of the current service, predict the customer reputation index of the current service.

2. The method for predicting customer reputation index according to claim 1, characterized in that, The step of classifying the dialogue text for emotion to obtain the current customer emotion tag includes: Based on the abstract text summarization model, a summary of the dialogue text is extracted to obtain the summary text. The abstract text summarization model includes a pre-trained PEGASUS model for generating summaries. Based on the emotion classification model, the summary text is classified into emotions to obtain the current customer emotion tags.

3. The method for predicting customer reputation index according to claim 2, characterized in that, The emotion classification model includes the self-attention mechanism of the BERT model and the fully connected layer of the classifier. The process of classifying the summary text based on the emotion classification model to obtain the current customer emotion tag includes: Based on the self-attention mechanism of the BERT model, the summarized text is transformed into a context-dependent word vector matrix; Based on the fully connected layer of the classifier, the word vector matrix is ​​classified for sentiment to obtain the current customer sentiment label.

4. The method for predicting customer reputation index according to claim 3, characterized in that, The self-attention mechanism of the BERT model includes the Masked Language Modeling (MLM) task and the Next Sentence Prediction (NSP) task. The self-attention mechanism based on the BERT model transforms the summarized text into a context-sensitive word vector matrix, specifically including: Based on the Masked Language Modeling (MLM) task, predict the relationships between words in the summary text; Based on the next sentence prediction NSP task, predict the relationship between sentences in the summary text; Based on the relationships between the words and sentences, the summary text is transformed into a context-dependent word vector matrix.

5. The method for predicting customer reputation index according to claim 1, characterized in that, The prediction of the customer reputation index for the current service based on customer tags and customer sentiment tags specifically includes: Based on the preset relationship between customer sentiment tags and scores, the score corresponding to the customer sentiment tag for the current service is determined; Based on the preset relationship between customer tags and weights, the customer tags of the current service are assigned weights to obtain the weights of the customer tags of the current service. Based on the weight of customer tags and the scores corresponding to customer sentiment tags for the current service, predict the customer reputation index for the current service.

6. The method for predicting customer reputation index according to claim 5, characterized in that, The customer tags include service type tags and various service indicator tags under each service type. Before assigning weights to customer tags for the current service based on the preset relationship between customer tags and weights, and obtaining the weights of customer tags for the current service, the customer reputation index prediction method further includes: Obtain the sample evaluation data matrix A = (X ij ) m×n Where m represents the total number of customers who have rated the service, n represents the total number of service indicator tags under each service type, and X ij This represents the score of the j-th service metric label for the i-th customer; The sample evaluation data matrix is ​​standardized to obtain the sample evaluation standard data matrix B = (Y ij ) m×n , where Y ij This represents the score of the j-th service indicator label for the i-th customer after standardization. Based on the entropy weight method and sample evaluation criteria, the data matrix B = (Y) ij ) m×n Calculate the weight of each service indicator label; Associate each service metric tag with its corresponding weight to form a pre-defined relationship between customer tags and weights.

7. The method for predicting customer reputation index according to claim 6, characterized in that, The data matrix B = (Y) based on entropy weighting method and sample evaluation criteria ij ) m×n Calculate the weight of each service metric label, specifically including: Calculate the weights of each service indicator label according to formulas (1) and (2): in, If P ij =0, then define P ij ·ln(P ij ) = 0, P ij E represents the percentage of the score for the j-th service metric tag for the i-th customer. j W represents the information entropy of the j-th service metric label. j This represents the weight of the j-th service metric label.

8. The method for predicting customer reputation index according to claim 5, characterized in that, The prediction of the customer reputation index for the current service based on the weights of customer tags and the scores corresponding to customer sentiment tags specifically includes: Based on formula (3), predict the customer reputation index of the current service: Among them, S i This represents the customer reputation index of the current service, where n represents the total number of service indicator tags under the current service type, and W represents the customer reputation index of the current service. j X represents the weight of the j-th service metric label under the current service type. ij This indicates the score of the customer's sentiment label for the current service.

9. A device for predicting customer reputation index, characterized in that, It includes an acquisition module, a first classification module, a second classification module, and a prediction module. The acquisition module is used to acquire the text of the current conversation between the customer and customer service. The first classification module, connected to the acquisition module, is used to classify the dialogue text into services and obtain the customer tags for the current service. The second classification module, connected to the acquisition module, is used to classify the dialogue text by emotion to obtain the current customer's emotion tag. The prediction module, connected to the first classification module and the second classification module respectively, is used to predict the customer reputation index of the current service based on the customer tags and customer sentiment tags of the current service.

10. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to implement a customer reputation index prediction method as described in any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a method for predicting a customer reputation index as described in any one of claims 1 to 8.

Images