Tobacco enterprise customer classification and grading method and system based on sentiment analysis

Abstract

The invention relates to a tobacco enterprise customer classification and grading method and system based on sentiment analysis, and the method comprises the steps: obtaining the session data inputted by a customer in a tobacco customer service scene, and judging the intention of the customer; inputting the session data into a pre-trained multi-mode sentiment analysis model, and outputting a sentiment analysis result of the current customer dialogue data; acquiring target customer information initiating the current session, and inputting the target customer information into a pre-trained customer classification and grading model for processing; and generating a recommended customer strategy based on a Q-Learning algorithm, optimizing a service strategy by combining a preset generative AI model with the customer intention of the current customer, the sentiment analysis result and the recommended customer strategy, and generating a response mode and response content based on the optimized service strategy. According to the scheme, a tobacco industry dynamic knowledge base is constructed, and personalized service strategies are formulated in combination with emotional appeals and behavioral habits of customers, so that the answer accuracy and satisfaction of intelligent customer service staff are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent customer service in the tobacco industry, and particularly to a method and system for classifying and grading tobacco enterprise customers based on sentiment analysis. Background Art

[0002] In the current business environment, enterprises' demands for customer service efficiency and quality continue to increase. Although the question-and-answer system based on natural language processing has become a core tool, it faces systematic challenges in key aspects.

[0003] 1. Insufficient depth of emotional interaction: It is difficult to accurately identify emotions in complex contexts (such as sarcasm, puns), resulting in a fragmented human-machine collaboration; 2. Limited service accuracy: Dynamic information such as industry policies and product rules requires manual annotation and updating, and the update of the knowledge base depends on manual work, resulting in a lag in response; for tobacco enterprise customers, their consultations often involve multi-dimensional information (such as order records, complaint history, brand preferences, etc.), but the intelligent customer service of existing technologies is difficult to quickly associate multi-dimensional information, resulting in insufficient accuracy of cross-scenario information retrieval; 3. Difficulty in balancing cost and experience: Customer service data is disconnected from the business decision-making link, and it is difficult for enterprises to dynamically allocate resources according to customer value; moreover, the functions of existing systems (such as intelligent robots, FAQ libraries) are concentrated on answering standardized questions such as simple order inquiries and policy interpretations, lacking in-depth insights into customers' emotional demands (such as anxiety, expectation, dissatisfaction, etc.) and behavioral habits (such as sensitivity to ordering cycles, brand dependence); 4. The tobacco industry is strictly regulated, and customer service responses need to strictly comply with policies and regulations. However, general generation models are prone to output non-compliant content, posing legal risks; existing systems lack structured knowledge support and rely mostly on manual intervention, driving up operating costs. At the same time, static knowledge bases cannot adapt to sudden events and have weak emergency response capabilities. Summary of the Invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide a method and system for classifying and grading tobacco enterprise customers based on sentiment analysis, construct a dynamic knowledge base for the tobacco industry, and formulate personalized service strategies in combination with customers' emotional demands and behavioral habits to improve the accuracy and satisfaction of intelligent customer service answers.

[0005] A method for classifying and grading tobacco enterprise customers based on sentiment analysis, the method comprising: Obtaining the conversation data input by a customer in a tobacco customer service scenario, parsing the input conversation data, and retrieving a pre-constructed dynamic knowledge base of the tobacco industry based on the parsed conversation data to determine the customer's intention; Inputting the conversation data into a pre-trained multi-modal sentiment analysis model to output the sentiment analysis result of the current customer's conversation data; Obtain the target customer information that initiated the current session, and input the target customer information into a pre-trained customer classification and grading model for processing. The input data of the customer classification and grading model includes the historical behavior data of the current customer, and the historical data includes historical consultation records and order information. The output data of the customer classification and grading model includes the customer grading of the current customer, and the customer grading includes multiple preset types; Generate a recommended customer strategy based on the Q-Learning algorithm: Define the policy action space and dynamically adjust the policy action space according to the customer grading level; The state space includes at least the customer grading level and the sentiment analysis result; The preset generative AI model combines the customer intention, sentiment analysis result of the current customer, and the recommended customer strategy to optimize the service strategy, and generates a response method and response content based on the optimized service strategy.

[0006] In the above solution, by constructing a dynamic knowledge base for the tobacco industry and integrating multi-scenario features, the customer intention determined based on the dynamic knowledge base of the tobacco industry is more accurate; Based on the integration of multi-modal sentiment analysis and dynamic customer characteristics, the refinement of customer grading is realized; Combine the Q-Learning algorithm to dynamically adjust the policy action space, and generate personalized service strategies for different graded customers (such as angry type, satisfied type); And further optimize the service strategy based on the generative AI model (such as deepseek R1) to ensure that the response content complies with industry policies, can ensure the accuracy and relevance of the response content, improve satisfaction, and reduce the intervention of human customer service.

[0007] Preferably, it also includes updating the tobacco industry knowledge base. Specifically: within a preset time, when the occurrence frequency of the same tobacco industry policy document keyword involved in the conversation data of different customers in the tobacco customer scenario reaches a preset growth rate and is not included in the tobacco industry knowledge graph, trigger an update of the tobacco industry knowledge base; Among them, the conversation data includes the conversation data with the intervention of human customer service.

[0008] After the tobacco industry policy is updated, based on the above settings, it can automatically trigger an update of the tobacco industry knowledge base, which can reduce the manual maintenance of the knowledge base while ensuring the timeliness of policy response. And after the update, it can reduce the intervention of human customer service for similar policy questions.

[0009] Preferably, the tobacco industry knowledge base is constructed based on the native graph database Neo4j graph database, covering the knowledge graph of multiple types of entities and businesses in the tobacco industry chain; Based on the GraphRAG framework, divide the subgraph index to provide structured context for the generative AI model; Parse the tobacco industry policy documents through the generative AI model, extract key entities, and dynamically update the nodes of the knowledge graph.

[0010] The Neo4j knowledge base supports complex path queries of more than 30 hops, which is beneficial to improving the response speed; GraphRAG provides accurate business scenario context for the generative AI, making the knowledge base highly adaptable to the generative AI and capable of accurately judging the customer's intention.

[0011] Preferably, within a preset time period, when the consultation volume for the same keyword reaches a preset threshold, an emergency subgraph is automatically associated with historical complaint cases, and the generative AI model performs similarity retrieval based on the emergency subgraph to generate a processing template for similar problems.

[0012] Generate an emergency subgraph for similar problems. The subgraph supports similarity retrieval and can directly call the processing template for similar problems when encountering similar problems, improving the problem processing efficiency in emergency situations.

[0013] Preferably, the multi-modal sentiment analysis model performs the following steps: Based on the pre-trained BERT-base model, combined with the pre-constructed sentiment dictionary in the tobacco industry field, perform sentiment recognition on the text expressions of the obtained conversation data and calculate the sentiment amplitude index; Extract the customer voice acoustic features through the Wav2Vec2.0 model, combined with the pre-constructed sentiment dictionary in the tobacco industry field, and use the Bi-LSTM model to identify the voice sentiment and calculate the sentiment amplitude index; Align the time stamps of the text and voice sentiment analysis results, and perform weighted fusion through the gated fusion mechanism and the attention mechanism to generate a multi-modal joint feature vector; Optimize the multi-modal joint feature vector based on the RoBERTa model and output the sentiment analysis result.

[0014] For the same content, perform sentiment analysis on the voice and text respectively and then fuse them to make the output sentiment analysis result more accurate.

[0015] Preferably, the customer classification and grading model performs the following steps: Construct a multi-dimensional customer feature vector based on the historical behavior data of the target customer, including behavior features, sentiment features, and value features; After standardizing the customer feature vector, use an improved DBSCAN clustering algorithm to dynamically calculate the neighborhood radius and hierarchically set the minimum number of samples to generate an initial clustering cluster and a noise point set; Perform secondary clustering or manual correction on the noise points to finally complete multi-level customer grading.

[0016] Adopt dynamic neighborhood radius design, and perform secondary clustering and manual correction for noise points, making the classification more accurate.

[0017] Preferably, the improved DBSCAN clustering algorithm includes: The dynamic neighborhood radius is calculated as , where is the Euclidean distance from the k-th nearest neighbor to the point xi, is the p-th quantile of the local density values in the dataset, and the p value is adaptively adjusted according to the density distribution. Specifically, the data is divided into density layers according to the average k-nearest neighbor distance of the data points. For high-density regions, the p value is reduced and a smaller eps is used; for high-density regions, the p value is increased and a larger eps is used.

[0018] The adaptive p value adjustment optimizes the cluster boundary, making the regional division more reasonable and taking into account both the clustering speed and accuracy.

[0019] Preferably, the service response control generates a recommended customer strategy based on a preset service strategy, including: The emotional features include the mean value of the historical consultation emotional intensity and the variance of the emotional fluctuation; when receiving a request for manual customer service intervention, the customer service resources are automatically allocated based on the variance of the emotional fluctuation, and the value of the variance of the emotional fluctuation is positively correlated with the level of the customer service resources.

[0020] It can automatically allocate customer service resources of different levels according to the emotional characteristics of customers. For example, for those with a large variance of emotional fluctuations, high-level customer service is allocated, which can improve the problem-solving rate and satisfaction and reduce the complaint rate.

[0021] Preferably, the customer satisfaction is dynamically updated according to the emotional polarity, complaint frequency, and consultation response duration of the customer; an accuracy evaluation from the customer is obtained, and the reward function of the Q-Learning algorithm is determined based on the accuracy, customer satisfaction, and consultation response duration. Optimizing the customer strategy recommendation algorithm based on customer satisfaction and the like can improve the accuracy of the recommendation results.

[0022] The purpose of this application is also to provide a tobacco enterprise customer classification and grading system based on sentiment analysis, which applies the method described in any one of the above. Due to applying the above method, this system also has the advantages of the above method.

[0023] In summary, due to adopting the above solution, this application uses generative AI to retrieve the tobacco industry dynamic knowledge base to identify customer intentions, and the parsing results are more accurate. Moreover, the finally generated response method and response content take into account the customer's emotions and historical behaviors, enabling the intelligent customer service to formulate personalized service strategies for customers, which is beneficial to solving complex conversation content and improving customer satisfaction. Detailed Implementation Modes

[0024] The embodiments of the present invention will be described in detail below.

[0025] This embodiment relates to a method for classifying and grading tobacco enterprise customers based on sentiment analysis, which is characterized in that the method includes: S1, obtaining the conversation data input by the customer in the tobacco customer service scenario, parsing the input conversation data, and retrieving the pre-constructed tobacco industry dynamic knowledge base based on the parsed conversation data to judge the customer's intention.

[0026] The conversation data input by the customer in the tobacco customer service scenario refers to the conversation data input by the customer in the customer service scenario, such as in the customer service dialogue window. The conversation data can be text data or voice data. In this embodiment, a generative AI model is accessed to parse the input conversation data, and the tobacco industry dynamic knowledge base is retrieved based on the parsed conversation data to judge the customer's intention.

[0027] The method includes constructing a tobacco industry dynamic knowledge base, which includes relevant policy documents and specifications of the tobacco industry. Thus, by retrieving the tobacco industry knowledge base, the relevant policy documents and specifications involved in the customer's consultation can be determined.

[0028] In this embodiment, the tobacco industry dynamic knowledge base is dynamically updated. Specifically, within a preset time, when the occurrence frequency of the same tobacco industry policy document keyword in the conversation data of different customers in the tobacco customer scenario reaches a preset growth rate and is not included in the tobacco industry knowledge graph, the update of the tobacco industry knowledge base is triggered; wherein, the conversation data includes the conversation data intervened by the artificial customer service. According to the above settings, the automatic update of the industry knowledge base can be triggered, and there is no need for manual update of the latest industry policy documents to the knowledge base by humans. For example, when the daily ring growth rate of the policy document keyword occurrence frequency reaches 50%, the update is triggered. Since the supply customers in the tobacco industry are relatively stable, when the occurrence frequency of the same policy document (benefiting the supply volume) increases significantly, it can be presumed that the relevant policy has been updated, and the automatic update is triggered based on this setting.

[0029] In this embodiment, the tobacco industry knowledge base is constructed based on the native graph database Neo4j graph database, covering the knowledge graph of multiple types of entities and businesses in the tobacco industry chain. The Neo4j graph database can achieve efficient traversal and real-time reasoning of complex relationships, construct a knowledge graph covering 12 types of entities and 28 types of business relationships in the tobacco industry chain, support real-time query of complex paths of more than 30 hops, and pre-define compliance rules and customer grading rules to achieve real-time risk warning and strategy recommendation.

[0030] The dynamic knowledge graph of the tobacco industry is based on the GraphRAG framework, realizes the partitioning of the knowledge graph subgraph, constructs a retrieval index for business scenarios, and provides structured context for generative AI such as the DeepSeek R1 large model, so that the generative AI model can retrieve the dynamic knowledge graph of the tobacco industry.

[0031] In this embodiment, when triggering the automatic update of the industry knowledge base, the generative AI model parses the tobacco industry policy document, extracts key entities and dynamically updates the knowledge graph nodes. Specifically, a tobacco knowledge graph is constructed based on the Neo4j graph database, business events are accessed in real time through Kafka, and the APOC plugin of Neo4j is called for batch writing to ensure the real-time update of the graph.

[0032] When this policy is involved later, the generative AI can directly generate corresponding responses based on this policy and customer information, without the need to connect to a human customer service, which is beneficial to accelerating the response speed and improving customer satisfaction.

[0033] In a preferred embodiment, within a preset time period, when the consultation volume for the same keyword reaches a preset threshold, an emergency subgraph is automatically associated with historical complaint cases, and the generative AI model performs similarity retrieval based on the emergency subgraph to generate a processing template for similar problems. For example, when the single-day consultation volume for the same policy document keyword or the same question exceeds the threshold, an emergency subgraph is automatically associated with historical complaint cases. Store the data of the past 3 months in the Neo4j memory, configure the LRU cache policy, and the hit rate ≥ 95%. The historical data (more than 3 months) is exported as a subgraph snapshot and stored in the graph database, partitioned by timestamp.

[0034] The knowledge base partitions subgraphs according to business scenarios, and each subgraph contains relevant entities and relationships. The GraphSAGE algorithm is used to generate subgraph embedding vectors and store them in the Faiss index to support similarity retrieval.

[0035] In this embodiment, the generative AI model such as DeepSeek R1 converts the conversation data expressed in natural language into Cypher, so as to be able to retrieve the Neo4j graph database, retrieve relevant subgraphs, and then extract key entities and relationships to generate a prompt template. And a dynamic weight mechanism is adopted. In the attention layer of DeepSeek R1, a weight gain is applied to the knowledge prompt (such as weight × 1.2). Check whether the model output result contains policy keywords, and if missing, trigger regeneration. It can improve the accuracy of the output result.

[0036] S2, input the conversation data into a pre-trained multi-modal sentiment analysis model, and output the sentiment analysis result of the current customer conversation data.

[0037] The multi-modal sentiment analysis model performs the following steps: Based on the pre-trained BERT-base model, combined with the pre-constructed sentiment dictionary in the tobacco industry domain, perform sentiment recognition on the text expressions of the obtained conversation data and calculate the sentiment amplitude index; Extract the customer voice acoustic features through the Wav2Vec2.0 model, combined with the pre-constructed sentiment dictionary in the tobacco industry domain, and use the Bi-LSTM model to identify the voice sentiment and calculate the sentiment amplitude index. Among them, first for the tobacco industry corpus, combined with the Word2Vec model and the K-means clustering algorithm, construct a special sentiment dictionary for the tobacco domain, including more than 3,000 professional terms and their sentiment weights; in a specific embodiment, the construction of the special sentiment dictionary for the tobacco domain specifically includes: Select 1 million tobacco industry corpora, including complaint records, policy documents, customer service conversations, etc. Train with Word2Vec, set vector_size = 300, window = 5, min_count = 10, and generate word vectors. Set the number of clusters K = 15 for K-means++ clustering, with the silhouette coefficient ≥ 0.65, and extract the top 5% high-frequency words for manual annotation.

[0038] Discover sentiment word clusters through the clustering algorithm, and construct a sentiment dictionary in combination with manual annotation. Use the sentiment weights in the dictionary as features to fine-tune the pre-trained language model and strengthen the influence of domain keywords in the model attention layer.

[0039] The training method of the Wav2Vec2.0 model is as follows: Configure the use of the wav2vec2-base-960h pre-trained model to extract features, output frame-level features, 50ms per frame, perform mean pooling on the output features to generate 128-dimensional acoustic vectors. Based on a bidirectional LSTM (256-dimensional hidden layer), connect an 8-head self-attention layer, and finally output the sentiment label through Softmax. Add random noise (SNR = 10dB), variable speed (±10%), and pitch perturbation (±50 cents), and use 10,000 labeled voices to divide into an 80% training set, a 10% validation set, and a 10% test set.

[0040] Emotion fluctuation detection uses sliding window analysis, with a 30-second window and a step size of 10 seconds, calculate the entropy value of the sentiment labels within the window, and mark the entropy value mutation points as emotion fluctuations. Analyze the emotion fluctuation period of the voice segment, generate an emotion change curve, and associate it with the type of consultation question.

[0041] Align the timestamps of the text and voice sentiment analysis results, and generate a multi-modal joint feature vector through weighted fusion by the gating fusion mechanism and the attention mechanism; Among them, the fusion weight g = \sigma(w\times [text feature; speech feature]+b) , where g represents the weight of the text feature, \(\sigma\) is the Sigmoid function, the symbol ";" represents the vertical concatenation operation of matrices, and the weight biases towards the high-confidence modality; w and b are training parameters, w is the weight value, and b is the offset; Optimize the multi-modal joint feature vector based on the RoBERTa model and output the sentiment analysis result. In this embodiment, the output sentiment analysis result is a pre-classified sentiment label (7 classifications: anger, satisfaction, neutral, etc.), and the accuracy rate is ≥89%.

[0042] S3. Obtain the target customer information initiating the current session, and input the target customer information into a pre-trained customer classification and grading model for processing. The input data of the customer classification and grading model includes the historical behavior data of the current customer. The historical data includes historical consultation records and order information. The output data of the customer classification and grading model includes the customer grading of the current customer, and the customer grading includes multiple pre-set types; The historical data is stored in a unified data platform. This method includes constructing a data platform, including: Collect customer historical data, including multi-source data such as customer consultation records, ordering behaviors, brand preferences, etc. The multi-source data includes, but is not limited to, historical consultation records, ordering frequencies, payment methods, brand preferences, complaint records, etc. of the customer; Integrate the collected multi-source data into a unified data platform to ensure the integrity and consistency of the data. Use ETL (Extract, Transform, Load) tools for data cleaning, transformation, and loading to ensure data quality; Adopt a combination of a distributed storage system OSS and a relational database MySQL to store structured and unstructured data to ensure efficient access and storage of the data.

[0043] Store the historical data of the customer in a unified data platform in the above manner for easy calling and access when needed.

[0044] The customer classification and grading model performs the following steps: S31. Adopt multi-dimensional customer feature engineering, collect behavioral data such as the historical consultation records, ordering frequencies, brand preferences, etc. of the customer, and construct a customer feature vector containing 18 dimensions. The historical behavior data among them includes: Behavioral data: Extract historical order records of customers from the management platform, including order frequency (times / month), single order quantity (pieces), and brand concentration (the proportion of the order quantity of the top 3 brands = the order quantity of the top 3 brands / total order quantity).

[0045] Payment method classification: Proportion of cash payment, proportion of online payment, proportion of payment on credit.

[0046] Sentiment data: Extract the sentiment polarity (range [-1, 1]) of historical consultation records, the variance of sentiment fluctuations ( ) and the statistics of complaint frequency from the sentiment analysis engine.

[0047] Value data: ARPU value (average monthly consumption amount), extract the total consumption amount of customers in the recent 6 months from the financial system and calculate the monthly average value. Customer life cycle stage, divided based on the RFM model, such as new customer, growth stage, maturity stage, decline stage. Potential churn risk score: Predicted using a logistic regression model, and the input features include the growth rate of order interval and the downward trend of sentiment polarity.

[0048] The extracted customer feature vectors of 18 dimensions include: Behavioral characteristics: Order frequency, brand concentration, payment method; Sentiment characteristics: Mean value of historical consultation sentiment intensity, variance of mood fluctuations; Value characteristics: ARPU value, customer life cycle stage, potential churn risk score.

[0049] Based on the above extracted features and sentiment analysis results, a customer portrait can be generated. Based on the customer portrait, an interpretable customer grading result is output to support business decision-making and analysis. And in this solution, different customer maintenance strategies are executed based on the customer portrait. For example, limited edition new products are pushed to level 1 customers (high value), and customer relationship maintenance is strengthened for level 6 customers (potential churn).

[0050] S32. Process the feature vectors using Min-Max normalization and Z-Score normalization methods.

[0051] S33. Adopt an improved DBSCAN clustering algorithm, dynamically calculate and set the dynamic neighborhood radius based on the feature space distribution, and hierarchically set the minimum number of samples according to the customer density distribution.

[0052] Calculation of dynamic neighborhood radius (eps): , where is the Euclidean distance from the k-th nearest neighbor to the point xi, k is selected according to the data dimension, is the p-th percentile of all local density values in the dataset, p is adaptively selected according to the local density, and the dynamic parameter allows simultaneous processing of high-density compact clusters and low-density sparse clusters.

[0053] Divide the data into density layers according to the average k-nearest neighbor distance of the data points. Decrease p in the high-density area and use a smaller eps, while increase p in the high-density area and use a larger eps.

[0054] S34. Perform secondary clustering on the noise points, i.e., customers who cannot be classified, and use KNN to fill them into the nearest neighbor category. Based on the improved DBSCAN, generate the initial clusters and the set of noise points. Calculate the distance from each noise point to the cluster centers and assign it to the nearest cluster. If the KNN result does not conform to the business logic, manually set the classification boundary to correct the pure data-driven deviation. Finally, perform six-level classification according to the clustering result. The six-level classification in this embodiment is shown in Table 1.

[0055] Table 1 Correspondence table of customer characteristics and customer classification

[0056] S4. Generate recommended customer strategies based on the Q-Learning algorithm: Define the strategy action space and dynamically adjust the strategy action space according to the customer classification level. The state space includes at least the customer classification level and the sentiment analysis result.

[0057] In this embodiment, the state space of the Q-Learning algorithm is defined as the customer classification level (6 dimensions), the real-time sentiment label (7 categories), and the historical behavior characteristics (such as the order frequency interval). Thus, dynamic decision-making driven by classification is realized. The service response action control includes 8 basic actions and 12 combined strategies. The basic actions include emotion appeasement, quick solution, service upgrade channel, expert intervention, emotional account, etc.; the combined strategies are composed based on factors such as basic actions, scenario conditions, and customer types. Thus, generate the customer strategies adopted by the recommended service response action control based on the Q-Learning algorithm. For example, the combined action of emotion appeasement + quick solution + service upgrade channel can respond to angry customers and customers with emotional fluctuations, achieving a 30-second golden response; the emotional account action can respond to satisfied and silent customers, providing personalized care and enhancing the customer's emotional value.

[0058] Reward function: , where the accuracy rate is the problem-solving rate after the strategy is executed, and the satisfaction degree is dynamically calculated based on the sentiment analysis result.

[0059] S5. Combine the preset generative AI model with the current customer's customer intention, sentiment analysis result, and the recommended customer strategy to optimize the service strategy, and generate the response method and response content based on the optimized service strategy.

[0060] In this step, based on the current customer emotions, customer intentions, and customer strategies recommended by the Q-Learning algorithm obtained by the multimodal sentiment analysis model, the deepseek model is used to optimize the service strategy and output it to the customer. For example, customer A is a satisfied customer, his current emotions are neutral, his intention is information consultation, and the recommended strategy is the emotional account. Deepseek will optimize the response content and form (picture and text replies, blessing closing remarks, etc.) based on customer intentions and emotional account information to ensure that neutral emotions are transformed into positive emotions.

[0061] In this embodiment, customer satisfaction is dynamically updated according to the customer's emotional polarity, complaint frequency, and consultation response time, and a dynamic update mechanism for customer satisfaction is established: customer satisfaction = α×emotional polarity+β×complaint frequency+γ×response time, where weights α, β, and γ are optimized through grid search to ensure that the contribution to customer grading complies with industry rules. The accuracy evaluation from the customer is obtained, and the above-mentioned reward function of the Q-Learning algorithm is determined based on the accuracy, customer satisfaction, and consultation response time, so as to provide feedback on the service strategy formulated by the given S5, and optimize Q-Learning based on the feedback data drive.

[0062] Furthermore, in this solution, full-link feedback data is collected in real time to build a closed-loop optimization foundation. Closed-loop optimization strategy effects improve customer satisfaction. Through direct feedback such as customer satisfaction ratings (1-5 stars), complaint resolution rates and indirect feedback such as changes in sentiment analysis results (such as a decrease in the intensity of negative emotions), follow-up behavior data, etc., dynamic and personalized customer service is achieved, significantly improving the operational efficiency and customer experience of tobacco companies. System performance and business effects are quantified through satisfaction index, strategy hit rate, average response time, and manual customer service intervention rate to drive optimization direction.

[0063] Continuously improve system performance based on feedback data, including Q-Learning strategy optimization, sentiment analysis model iteration, and knowledge graph dynamic correction.

[0064] The solution of this embodiment closely matches the supervision and business needs of the tobacco industry, and has the following benefits in terms of emotional dimension integration, closed-loop service design, and industry knowledge graph customization: 1. Enrich the customer classification dimensions. Based on historical data such as sales data and customer value, sentiment analysis is introduced as a customer classification dimension to further segment customer groups; 2. Improve customer satisfaction. Through the linkage between emotional tags and knowledge graphs, a closed loop of "emotion recognition → classification and grading → targeted service → data feedback" is realized, which can accurately respond to customer needs and effectively improve customer satisfaction and loyalty. 3. Reduce operating costs: The customized knowledge graph for the tobacco industry effectively shares most of the routine customer problems, reduces the workload and operating costs of human customer service, and improves the overall customer service efficiency.

[0065] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for classifying and grading tobacco enterprise customers based on sentiment analysis, characterized in that: The method comprises: Acquire the session data input by the customer in the tobacco customer service scenario, parse the input session data, and retrieve a pre-built tobacco industry dynamic knowledge base based on the parsed session data to determine the customer's intention; Inputting the conversation data into a pre-trained multimodal sentiment analysis model, and outputting sentiment analysis results of the current customer conversation data; Acquire target customer information for initiating the current session, and input the target customer information into a pre-trained customer classification and grading model for processing, wherein the input data of the customer classification and grading model includes historical behavior data of the current customer, the historical behavior data includes historical consultation records and order information, and the output data of the customer classification and grading model includes the customer grading of the current customer, the customer grading includes a plurality of pre-set types; Generate recommended customer strategies based on the Q-Learning algorithm: define the strategy action space and dynamically adjust the strategy action space according to the customer classification level; the state space at least includes the customer classification level and sentiment analysis results; The preset generative AI model combines the current customer's customer intent, sentiment analysis results and the recommended customer strategy to optimize the service strategy, and generates a response method and response content based on the optimized service strategy.

2. A method for classifying and grading tobacco enterprise customers based on sentiment analysis according to claim 1, characterized in that: It also includes updating the tobacco industry dynamic knowledge base, specifically: When, within a preset time, the frequency of occurrence of keywords in the same tobacco industry policy document that are not included in the tobacco industry knowledge graph in conversation data among different customers in the tobacco customer scenario reaches a preset growth rate, the update of the tobacco industry dynamic knowledge base is triggered; wherein, the conversation data includes conversation data with the intervention of manual customer service.

3. A method for classifying and grading tobacco enterprise customers based on sentiment analysis according to claim 2, characterized in that: The tobacco industry dynamic knowledge base is built based on the native graph database Neo4j graph database, covering the knowledge graph of multiple entities and businesses in the tobacco industry chain; Dividing subgraph indexes based on the GraphRAG framework to provide structured context for the generative AI model; The tobacco industry policy document is parsed through the generative AI model, key entities are extracted and the knowledge graph nodes are dynamically updated.

4. A method for classifying and grading tobacco enterprise customers based on sentiment analysis according to claim 3, characterized in that: Within a preset time period, when the number of consultations for the same keyword reaches a preset threshold, historical complaint cases are automatically associated to generate an emergency subgraph. The generative AI model performs similarity retrieval based on the emergency subgraph to generate a template for processing similar problems.

5. The method for classifying and grading tobacco enterprise customers based on sentiment analysis according to claim 1, characterized in that: The multimodal sentiment analysis model performs the following steps: Based on the pre-trained BERT-base model and the pre-built tobacco industry sentiment dictionary, we perform sentiment recognition on the textual expressions of the acquired conversation data and calculate the sentiment amplitude index. The acoustic features of customer voice are extracted by the Wav2Vec2.0 model, and combined with the pre-built tobacco industry sentiment dictionary, the Bi-LSTM model is used to identify voice sentiment and calculate the sentiment amplitude index; Align the timestamps of text and speech sentiment analysis results, and generate a multimodal joint feature vector through weighted fusion through gated fusion mechanism and attention mechanism; The multimodal joint feature vector is optimized based on the RoBERTa model, and the sentiment analysis result is output.

6. The method for classifying and grading tobacco enterprise customers based on sentiment analysis according to claim 1, characterized in that: The customer classification and grading model performs the following steps: Construct a multi-dimensional customer feature vector based on the input target customer's historical behavior data, including behavioral features, emotional features, and value features; After the customer feature vector is standardized, an improved DBSCAN clustering algorithm is used to dynamically calculate the neighborhood radius and set the minimum number of samples in layers to generate an initial clustering cluster and a noise point set; Perform secondary clustering or manual correction on noise points to ultimately complete multi-level customer classification.

7. A method for classifying and grading tobacco enterprise customers based on sentiment analysis according to claim 6, characterized in that: The improved DBSCAN clustering algorithm comprises: The dynamic neighborhood radius is calculated as , in, is the Euclidean distance from the kth nearest neighbor to point xi, Qp is the pth quantile of the local density value in the data set, and the p value is adaptively adjusted according to the density distribution. Specifically, the data is divided into density layers according to the average k nearest neighbor distance of the data points. For high-density areas, the p value is reduced and a smaller eps is used; for high-density areas, the p value is increased and a larger eps is used.

8. The method for classifying and grading tobacco enterprise customers based on sentiment analysis according to claim 6 is characterized in that: The customer strategy for generating recommendations based on the Q-Learning algorithm includes: The emotional characteristics include the mean emotional intensity of historical consultations and the variance of emotional fluctuations. When a request for manual customer service intervention is received, customer service resources are automatically allocated based on the variance of emotional fluctuations, and the variance value of emotional fluctuations is positively correlated with the level of customer service resources.

9. The method for classifying and grading tobacco enterprise customers based on sentiment analysis according to claim 1, characterized in that: Dynamically update customer satisfaction based on the customer's emotional polarity, complaint frequency, and consultation response time; obtain accuracy evaluation from the customer, and determine the reward function of the Q-Learning algorithm based on the accuracy, customer satisfaction, and consultation response time.

10. A tobacco enterprise customer classification and grading system based on sentiment analysis, characterized in that: Use the method as claimed in any one of claims 1 to 9.

Images