Property service satisfaction improvement method based on big language model reasoning

Abstract

The invention discloses a property service satisfaction improvement method based on big language model reasoning, and belongs to the field of artificial intelligence and property service management. In order to solve the problems that customer satisfaction improvement measures depend on experience, native comments are insufficient in utilization and the like, the customer satisfaction improvement method based on fusion of a large language model and multi-source data is adopted in the scheme. Multi-source data of enterprise WeChat, interviews, business systems and the like are integrated, a training analysis model is designed, and a semantic recognition system is constructed to realize theme cognition and appeal analysis. And establishing a passenger full knowledge base, determining a hot scene according to historical data, and generating an evaluation report through a large language model in combination with index analysis. And fusing the hotspot and the native content by using the recommendation model, and outputting a TOP5 lifting measure by using the large language model and pushing the TOP5 lifting measure to a tracking system. According to the method, the value of multi-source data is mined, customer concerns are accurately positioned through analysis of the large model, the service level and the satisfaction degree are improved, and the model can be autonomously optimized along with data accumulation to adapt to different project requirements.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and specifically relates to a method for improving satisfaction with new service projects based on a large language model and multi-source data fusion. Background Art

[0002] Property services primarily address the diverse needs of property owners in their residential or work environments. Encompassing asset maintenance, lifestyle services, and security, they provide owners with comfortable, convenient, and secure living and working conditions, thus playing a vital role in modern communities and commercial spaces. With the improvement of people's quality of life and the increasing demands of commercial operations, the demand for property services is growing, and New Hope's new service projects are continuously expanding their scope and scale of services.

[0003] Customer satisfaction has become a core performance metric in New Hope's new service projects. For a long time, measures to improve customer satisfaction relied primarily on the personal experience of project managers, lacking a scientific and systematic approach. Furthermore, satisfaction evaluations were limited to traditional scoring systems, with a significant underutilization of native reviews (such as those from WeChat chats, interviews, surveys, and existing data from various business systems). These native reviews contain a wealth of customer needs, opinions, and suggestions, but they were not fully explored and analyzed, making it difficult to accurately formulate improvement measures.

[0004] To improve customer satisfaction, the project team has tried various approaches. These include strengthening employee training to enhance service personnel's skills and service awareness, and optimizing service processes to reduce delays and errors. However, these measures have been less than ideal and have failed to fundamentally address the problem. While some data analysis methods are currently being used to improve customer satisfaction, most are unable to fully integrate and deeply analyze multi-source data, struggle to adapt to the complex needs of projects of varying types and sizes, and lack the ability to autonomously learn and continuously optimize as data accumulates.

[0005] Amid the rapid development of artificial intelligence, large language model technology has demonstrated powerful capabilities in natural language processing and data analysis. It enables in-depth understanding and analysis of large amounts of text data, uncovering the underlying insights. Incorporating large language model technology into new service projects to improve customer satisfaction promises to overcome existing challenges. However, in practical applications, how to effectively integrate large language models with multi-source data, accurately analyze native reviews, precisely recommend satisfaction-enhancing measures, and enable autonomous learning and continuous optimization of the model remain pressing challenges. Summary of the Invention

[0006] In order to solve the problems existing in the current process of improving customer satisfaction of new service projects, such as over-reliance on experience, insufficient utilization of native comments, lack of targeted improvement measures and dynamic optimization, this solution provides a new service project satisfaction improvement method based on a large language model and multi-source data fusion. This solution uses a customer satisfaction improvement method based on a large language model and multi-source data fusion to collect various types of data for in-depth mining and analysis around the factors affecting the satisfaction score. A native comment analysis model is designed using a large language model, and a semantic recognition system is constructed to accurately analyze customer demands and hot issues. At the same time, a full house practice knowledge base is established, and improvement measures are recommended in combination with hot spot analysis. The closed-loop management of measures and experience archiving are achieved through a full house measure tracking system. With the continuous accumulation of data, the large language model realizes autonomous learning and continuous optimization, adjusts and improves improvement measures to meet the needs of different projects, thereby effectively improving service levels and customer satisfaction. The technical problems proposed by the present invention are solved as follows:

[0007] Specific content:

[0008] A method for improving satisfaction with new service projects based on a large language model and multi-source data fusion, comprising the following steps:

[0009] Step 1: Obtain multi-source data related to the new service project, including original comment data such as corporate WeChat chat records, interview content, survey data, and existing data from various business systems. This data is then preliminarily cleaned and integrated, and classified according to data source and business type.

[0010] Step 2: Input the classified data into a pre-designed native review analysis model. Combined with customer information, the large language model is used to deeply analyze the data and extract key information and latent semantics.

[0011] Step 3: Based on the information analyzed in Step 2, a semantic recognition system is constructed. Using a large language model, a precise understanding of discussion topics and in-depth analysis of customer demands are achieved, resulting in structured topic and demand data.

[0012] Step 4: Build a knowledge base for full occupancy practices, identify hot service scenarios based on historical original reviews, combine weekly original review analysis information with recent indicators, use a large language model to conduct a detailed analysis of hot service scenarios, and generate an evaluation report.

[0013] Step 5: Utilize the recommendation model, combined with hot topics and native content, and leverage the large language model to filter out suitable improvement measures from the knowledge base. The recommended top five improvement measures are then sent to the full-occupancy measure tracking system.

[0014] Step 6: Implement and track the recommended improvement measures in the full occupancy measures tracking system, and record various data and feedback information during the implementation process;

[0015] Step 7: Regularly evaluate the implementation effect. Based on the evaluation results, use the large language model to optimize and update the native review analysis model, recommendation model, and knowledge base to continuously improve customer satisfaction with new service projects.

[0016] Step 1 of this protocol includes:

[0017] Step 1-1: Export chat history data from the WeChat Work backend and filter out conversation records containing key information such as customer reviews, feedback, and complaints.

[0018] Step 1-2: Extract existing data related to customer service from the database of each business system, such as work order processing records, service evaluation data, etc.

[0019] Steps 1-3: Organize the original documents collected from interviews and surveys, digitize the text information, and remove data with incorrect formatting and incomplete content.

[0020] Steps 1-4: De-duplicate the integrated multi-source data to avoid interference with subsequent analysis.

[0021] Step 2 of this plan includes:

[0022] Step 2-1: Select the Large Language Model (LLM) based on the Transformer architecture as the native review analysis model, leveraging its ability to understand and process natural language.

[0023] Step 2-2: Use word vector embedding technology to convert text data into vector form, so that the model can better capture the semantic relationship between words.

[0024] Step 3 of this plan includes:

[0025] Step 3-1: Input the preliminarily processed native review data into the native review analysis model. Use the model to extract key information from the text, such as customer demands, service scenarios, etc., and convert it into feature vectors. Use the customer demand feature vector to calculate the correlation with historical similar demand vectors. Use the service scenario feature vector to obtain common problems and solutions in the corresponding scenario. Use the cosine similarity formula to calculate the correlation between different feature vectors. The formula is: The feature vectors are weighted according to the correlation results to obtain the mutual attention weighted feature vectors, and then the clustering algorithm is used to align the weighted feature vectors to achieve fine-grained semantic understanding and classification.

[0026] Step 3-2: Taking the calculation of the mutual attention weighted features of customer demands in a certain service scenario as an example, first use the linear transformation layer to perform a linear transformation on the current customer demand feature vector V. The formula is as follows: Q = WQ V, thereby obtaining the query vector Q of the customer demand feature vector V. At the same time, another linear transformation layer is used to transform the historical similar demand feature vector V history For linear transformation, the formula is as follows: K = W K V history , thereby obtaining the keyword vector K of the current customer demand feature vector. During the model optimization process, the query vector Q and the keyword vector K are transformed through the linear transformation layer W Q and W K Then, the following formula is used to calculate the current customer demand feature vector V and the historical similar demand feature vector V history Mutual attention weight

[0027] Step 3-3: After calculating the mutual attention weight α, multiply the mutual attention weight α corresponding to each current customer demand feature vector V to obtain the mutual attention weighted customer demand feature vector V weighted .

[0028] Step 3-4: In order to improve the accuracy of semantic understanding and classification, the residual connection mechanism is introduced to reduce the information loss caused by model training errors. The formula for calculating the feature vector of mutual attention weight is as follows: V final =V weighted +V.

[0029] Step 3-5: Process different types of native comment data according to steps 3-1 to 3-4, and finally obtain the mutual attention weighted native comment feature vectors of each category.

[0030] Step 4 of this plan includes:

[0031] Step 4-1: Design domain discriminators for different categories of native review feature vectors. Each domain discriminator is composed of a multi-layer perceptron (MLP), and a total of n domain discriminators are designed (n is the number of native review categories).

[0032] Step 4-2: Domain Discriminator D i Responsible for predicting the probability of the category (such as complaint, suggestion, etc.) to which the feature vector of the i-th native review belongs. The predicted probability is compared with the true category label to calculate the cross-entropy loss. Then, through an adversarial training mechanism, it competes with the native review analysis model, prompting the model to extract features more accurately, thereby achieving fine-grained semantic classification.

[0033] Step 4-3: Calculate the classification loss of the k-th category native comment feature vector through the semantic classification objective function. The objective function is as follows: where y i is the true category label, p iis the predicted probability, and m is the number of samples.

[0034] Step 5 of this plan includes:

[0035] Step 5-1: Similar to step 4-1, design separate domain discriminators for native review feature vectors at different scales (such as by service process stage, customer group, etc.) to perform fine-grained semantic classification simultaneously at multiple scales.

[0036] Step 5-2: Through multi-scale semantic classification objective function Get the classification loss of native review feature vectors at different scales, where is the classification loss of the kth scale, and s is the number of scales.

[0037] Step 6 of this protocol includes:

[0038] Step 6-1: Construct the objective function of the improvement measure recommendation task, which mainly consists of three parts: recommendation accuracy loss, recommendation relevance loss, and recommendation diversity loss. By designing the objective function of the recommendation task: L recommend =w1L accuracy +w2L relevance +w3L diversity Calculate the loss of the recommendation task, where w1, w2, and w3 are weight coefficients, and L accuracy is the recommendation accuracy loss, L relevance is the recommendation relevance loss, L diversity is the recommendation diversity loss.

[0039] Step 6-2: Use the semantic classification objective function obtained in steps 4 and 5 to constrain the domain discriminator and the native review analysis model, prompting the domain discriminator to more accurately judge the native review category. At the same time, through adversarial training, the native review analysis model can extract more discriminative features, achieving the purpose of fine-grained semantic classification and feature extraction, thereby improving the performance of the recommendation model.

[0040] Step 6-3: Through the final objective function: L final =L recommend+ L multi-scale The final model loss is calculated, and the performance of the recommendation model and semantic classification model is optimized to improve the quality of recommended improvement measures. Finally, the optimized model is used to recommend corresponding improvement measures for different customer demands and service scenarios.

[0041] This solution proposes a method for improving satisfaction with new service projects based on a large language model and multi-source data fusion. Compared with existing technologies, it has the following advantages:

[0042] This solution combines a large language model with multi-source data fusion methods, which can effectively improve the pertinence and effectiveness of customer satisfaction improvement measures for new service projects. Compared with existing methods, the method proposed in this solution can effectively mine the potential information in multi-source native review data when conducting customer satisfaction analysis, thereby solving problems such as traditional methods relying on experience and lacking scientific analysis, and improving the efficiency and effectiveness of improving customer satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is the overall flow chart of the satisfaction analysis method;

[0044] Figure 2 It is a multi-source data classification framework diagram based on the classification model;

[0045] Figure 3 It is a native data classification logic diagram;

[0046] Figure 4 It is the logic diagram for calculating the weight of each line and extracting hot issues;

[0047] Figure 5 It is the native review analysis model LLM interaction diagram;

[0048] Figure 6 This is a diagram of the semantic recognition architecture;

[0049] Figure 7 This is a diagram of the working principle of the recommended model;

[0050] Figure 8 This is the system front-end, back-end and AI module deployment architecture diagram; DETAILED DESCRIPTION

[0051] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more precise definition of the protection scope of the present invention.

[0052] The implementation of this plan includes:

[0053] The overall process of the method proposed in this scheme is as follows Figure 1 As shown in the figure, the multi-source data classification framework based on the classification model is as follows Figure 2 As shown, the specific steps for constructing the satisfaction analysis in this embodiment are as follows:

[0054] Step 1: Obtain the required multi-source data, including enterprise WeChat chat records, call center reception records, etc., and perform preliminary data sorting to remove invalid data;

[0055] Step 2: Input the sorted data into the classification model to obtain the enterprise WeChat chat record classification result table and the call center reception record classification result table, such as Figure 2 As shown;

[0056] Step 3, based on Figure 3 The native data classification logic shown above performs structured classification on the classified data, dividing it into multiple categories such as environment, emergency, and real estate;

[0057] Step 4, follow Figure 4 The weight calculation of each line and the logic of extracting hot issues are shown. The weight of each line is calculated and the hot issues are extracted.

[0058] Step 5: Based on the hot issues extracted in step 4 and the improvement measures knowledge base, formulate a satisfaction improvement plan;

[0059] Step 6: Use the report generation model and the satisfaction analysis report template to generate a satisfaction analysis report, such as Figure 1 As shown;

[0060] Step 7: Set the system to Figure 5 The front-end, back-end, and AI module deployment architecture shown in the figure is deployed to ensure stable system operation.

[0061] Step 1 of this protocol includes:

[0062] Step 1-1: Collect multi-source data such as WeChat chat records and call center reception records from channels such as WeChat and call centers;

[0063] Step 1-2: Pre-process the collected data to remove duplicate, erroneous and other invalid data, such as Figure 3 Data related to the “invalid” category shown in ;

[0064] Step 1-3: Follow Figure 3 The native data classification logic divides the remaining valid data into multiple categories such as environment, emergency, and real estate to prepare for subsequent analysis;

[0065] Step 2 of this plan includes:

[0066] Step 2-1, select a suitable classification model, such as Figure 2 As shown, it is used to classify enterprise WeChat chat records and call center reception records;

[0067] Step 2-2: Use text cleaning, keyword extraction and other technologies to enhance the feature extraction capability of text data in order to achieve more accurate data classification;

[0068] Step 3 of this plan includes:

[0069] Step 3-1: Input the enterprise WeChat chat records and call center reception record data classified in Step 2 into the classification model. Use the model to extract key features from the text, such as the customer's problem type and degree of need, and convert them into feature vectors. Use the problem type feature vector to calculate the correlation with historical similar problem type vectors, and use the degree of need feature vector to obtain common solutions under the corresponding degree of need. The correlation between different feature vectors is calculated using the cosine similarity formula, which is: The feature vectors are weighted according to the correlation results to obtain the mutual attention weighted feature vectors. The weighted feature vectors are then aligned using a clustering algorithm to achieve fine-grained data classification and problem analysis.

[0070] Step 3-2: Taking the calculation of the mutual attention weighted features of the customer demand level under a certain question type as an example, first use the linear transformation layer to perform a linear transformation on the current customer demand level feature vector V. The formula is as follows: Q = W Q V, thereby obtaining the query vector Q of the customer demand degree feature vector V. At the same time, another linear transformation layer is used to transform the historical similar demand degree feature vector V history For linear transformation, the formula is as follows: K = W K V history , thereby obtaining the keyword vector K of the current customer demand level feature vector. During the model optimization process, the query vector Q and the keyword vector K are transformed through the linear transformation layer to obtain relevant information. Then the current customer demand level feature vector V and the historical similar demand level feature vector V are calculated by the following formula history The mutual attention weight α:

[0071] Step 3-3: After calculating the mutual attention weight α, multiply the mutual attention weight α corresponding to each current customer demand level feature vector V to obtain the mutual attention weighted customer demand level feature vector V weighted ;

[0072] Step 3-4: In order to improve the accuracy of data classification and problem analysis, the residual connection mechanism is introduced to reduce the information loss caused by model training errors. The formula for calculating the feature vector of mutual attention weight is as follows: V final =V weighted +V;

[0073] Step 3-5: Process different types of enterprise WeChat chat records and call center reception record data according to steps 3-1 to 3-4, and finally obtain the mutual attention weighted feature vectors of each category of data;

[0074] Step 4 of the present invention comprises:

[0075] Step 4 of this plan includes:

[0076] Step 4-1: Design domain discriminators for different categories of enterprise WeChat chat records and call center reception record data feature vectors. Each domain discriminator is composed of a multi-layer fully connected neural network, and a total of n domain discriminators are designed (n is the number of native comment categories).

[0077] Step 4-2: Domain Discriminator D i Responsible for predicting the probability of the category (such as complaint, suggestion, etc.) to which the feature vector of the i-th native review belongs. The predicted probability is compared with the true category label to calculate the cross-entropy loss. Then, through an adversarial training mechanism, it competes with the native review analysis model, prompting the model to extract features more accurately, thereby achieving fine-grained semantic classification.

[0078] Step 4-3: Calculate the classification loss of the k-th category native comment feature vector through the semantic classification objective function. The objective function is as follows: where y i is the true category label, p i is the predicted probability, and m is the number of samples.

[0079] Step 5 of the present invention comprises:

[0080] Step 5 of this plan includes:

[0081] Step 5-1: Similar to step 4-1, design separate domain discriminators for native review feature vectors at different scales (such as by service process stage, customer group, etc.) to perform fine-grained semantic classification simultaneously at multiple scales.

[0082] Step 5-2: Through multi-scale semantic classification objective function Get the classification loss of native review feature vectors at different scales, where is the classification loss of the kth scale, and s is the number of scales.

[0083] Step 6 of this protocol includes:

[0084] Step 6-1: Construct the objective function of the improvement measure recommendation task, which mainly consists of three parts: recommendation accuracy loss, recommendation relevance loss, and recommendation diversity loss. By designing the objective function of the recommendation task: L recommend =w1L accuracy +w2L relevance +w3L diversity Calculate the loss of the recommendation task, where w1, w2, and w3 are weight coefficients, and L accuracy is the recommendation accuracy loss, L relevance is the recommendation relevance loss, L diversity is the recommendation diversity loss.

[0085] Step 6-2: Use the semantic classification objective function obtained in steps 4 and 5 to constrain the domain discriminator and the native review analysis model, prompting the domain discriminator to more accurately judge the native review category. At the same time, through adversarial training, the native review analysis model can extract more discriminative features, achieving the purpose of fine-grained semantic classification and feature extraction, thereby improving the performance of the recommendation model.

[0086] Step 6-3: Through the final objective function: L final =L recommend +L multi-scale The final model loss is calculated, and the performance of the recommendation model and semantic classification model is optimized to improve the quality of recommended improvement measures. Finally, the optimized model is used to recommend corresponding improvement measures for different customer demands and service scenarios.

[0087] This solution proposes a method for improving satisfaction with new service projects based on a large language model and multi-source data fusion. Compared with existing technologies, it has the following advantages:

[0088] This solution combines a large language model with multi-source data fusion methods, which can effectively improve the pertinence and effectiveness of customer satisfaction improvement measures for new service projects. Compared with existing methods, the method proposed in this solution can effectively mine the potential information in multi-source native review data when conducting customer satisfaction analysis, thereby solving problems such as traditional methods relying on experience and lacking scientific analysis, and improving the efficiency and effectiveness of improving customer satisfaction.

[0089] This solution provides a method for improving satisfaction with new service projects based on a large language model and multi-source data fusion. The above is only an embodiment of this solution and does not limit the scope of protection of this solution. Any equivalent structure or equivalent process transformation made using the contents of the description and drawings of this solution, or directly or indirectly used in other related technical fields, are also included in the scope of protection of this solution.

Claims

1. A method for improving satisfaction with new service items based on a large language model and multi-source data fusion, characterized in that: The following steps are involved: Step 1: Obtain multi-source data related to the new service project, including original comment data such as corporate WeChat chat records, interview content, survey data, and existing data from various business systems. This data is then preliminarily cleaned and integrated, and classified according to data source and business type. Step 2: Input the classified data into a pre-designed native review analysis model. Combined with customer information, the large language model is used to deeply analyze the data and extract key information and latent semantics. Step 3: Based on the information analyzed in Step 2, a semantic recognition system is constructed. Using a large language model, a precise understanding of discussion topics and in-depth analysis of customer demands are achieved, resulting in structured topic and demand data. Step 4: Build a knowledge base for full occupancy practices, identify hot service scenarios based on historical original reviews, combine weekly original review analysis information with recent indicators, use a large language model to conduct a detailed analysis of hot service scenarios, and generate an evaluation report. Step 5: Utilize the recommendation model, combined with hot topics and native content, and leverage the large language model to filter out suitable improvement measures from the knowledge base. The recommended top five improvement measures are then sent to the full-occupancy measure tracking system. Step 6: Implement and track the recommended improvement measures in the full occupancy measures tracking system, and record various data and feedback information during the implementation process; Step 7: Regularly evaluate the implementation effect. Based on the evaluation results, use the large language model to optimize and update the native review analysis model, recommendation model, and knowledge base to continuously improve customer satisfaction with new service projects.

2. The method according to claim 1, characterized in that Step 1 includes: Step 1-1: Export chat history data from the WeChat Work backend and filter out conversation records containing key information such as customer reviews, feedback, and complaints. Step 1-2: Extract existing data related to customer service from the database of each business system, such as work order processing records, service evaluation data, etc. Steps 1-3: Organize the original documents collected from interviews and surveys, digitize the text information, and remove data with incorrect formatting and incomplete content. Steps 1-4: De-duplicate the integrated multi-source data to avoid interference with subsequent analysis.

3. The method according to claim 1, characterized in that Step 2 includes: Step 2-1: Select the Large Language Model (LLM) based on the Transformer architecture as the native review analysis model, leveraging its ability to understand and process natural language. Step 2-2: Use word vector embedding technology to convert text data into vector form, so that the model can better capture the semantic relationship between words.

4. The method according to claim 1, characterized in that Step 3 includes: Step 3-1: Input the preliminarily processed native review data into the native review analysis model. Use the model to extract key information from the text, such as customer demands, service scenarios, etc., and convert it into feature vectors. Use the customer demand feature vector to calculate the correlation with historical similar demand vectors. Use the service scenario feature vector to obtain common problems and solutions in the corresponding scenario. Use the cosine similarity formula to calculate the correlation between different feature vectors. The formula is: The feature vectors are weighted according to the correlation results to obtain the mutual attention weighted feature vectors, and then the clustering algorithm is used to align the weighted feature vectors to achieve fine-grained semantic understanding and classification. Step 3-2: Taking the calculation of the mutual attention weighted features of customer demands in a certain service scenario as an example, first use the linear transformation layer to perform a linear transformation on the current customer demand feature vector V. The formula is as follows: Q = W Q V, thereby obtaining the query vector Q of the customer demand feature vector V. At the same time, another linear transformation layer is used to transform the historical similar demand feature vector V history For linear transformation, the formula is as follows: K = W K V history , thereby obtaining the keyword vector K of the current customer demand feature vector. During the model optimization process, the query vector Q and the keyword vector K are transformed through the linear transformation layer W Q and W K Then, the following formula is used to calculate the current customer demand feature vector V and the historical similar demand feature vector V history Mutual attention weight Step 3-3: After calculating the mutual attention weight α, multiply the mutual attention weight α corresponding to each current customer demand feature vector v to obtain the mutual attention weighted customer demand feature vector V weighted . Step 3-4: In order to improve the accuracy of semantic understanding and classification, the residual connection mechanism is introduced to reduce the information loss caused by model training errors. The formula for calculating the feature vector of mutual attention weight is as follows: V final =V weighted +V. Step 3-5: Process different types of native comment data according to steps 3-1 to 3-4, and finally obtain the mutual attention weighted native comment feature vectors of each category.

5. The method according to claim 1, characterized in that Step 4 includes: Step 4-1: Design domain discriminators for different categories of native review feature vectors. Each domain discriminator is composed of a multi-layer perceptron (MLP), and a total of n domain discriminators are designed (n is the number of native review categories). Step 4-2: Domain Discriminator D i Responsible for predicting the probability of the category (such as complaint, suggestion, etc.) to which the feature vector of the i-th native review belongs. The predicted probability is compared with the true category label to calculate the cross-entropy loss. Then, through an adversarial training mechanism, it competes with the native review analysis model, prompting the model to extract features more accurately, thereby achieving fine-grained semantic classification. Step 4-3: Calculate the classification loss of the k-th category native comment feature vector through the semantic classification objective function. The objective function is as follows: where y i is the true category label, p i is the predicted probability, and m is the number of samples.

6. The method according to claim 1, characterized in that Step 5 includes: Step 5-1: Similar to step 4-1, design separate domain discriminators for native review feature vectors at different scales (such as by service process stage, customer group, etc.) to perform fine-grained semantic classification simultaneously at multiple scales. Step 5-2: Through the multi-scale semantic classification objective function: Get the classification loss of native review feature vectors at different scales, where is the classification loss of the kth scale, and s is the number of scales.

7. The method according to claim 1, characterized in that Step 6 includes: Step 6-1: Construct the objective function of the improvement measure recommendation task, which mainly consists of three parts: recommendation accuracy loss, recommendation relevance loss, and recommendation diversity loss. By designing the objective function of the recommendation task: L recommend =w1L accuracy +w2L relevance +w3L diversity Calculate the loss of the recommendation task, where w1, w2, and w3 are weight coefficients, and L accuracy is the recommendation accuracy loss, L relevance is the recommendation relevance loss, L diversity is the recommendation diversity loss. Step 6-2: Use the semantic classification objective function obtained in steps 4 and 5 to constrain the domain discriminator and the native review analysis model, prompting the domain discriminator to more accurately judge the native review category. At the same time, through adversarial training, the native review analysis model can extract more discriminative features, achieving the purpose of fine-grained semantic classification and feature extraction, thereby improving the performance of the recommendation model. Step 6-3: Through the final objective function: L final =L recommend +L multi-scale The final model loss is calculated, and the performance of the recommendation model and semantic classification model is optimized to improve the quality of recommended improvement measures. Finally, the optimized model is used to recommend corresponding improvement measures for different customer demands and service scenarios.

Images