A method and system for classifying and grading tobacco enterprise customers based on sentiment analysis

Abstract

The present invention relates to a method and system for classifying and grading tobacco enterprise customers based on sentiment analysis. The method comprises: obtaining conversation data input by customers in tobacco customer service scenarios to determine customer intent; inputting the conversation data into a pre-trained multimodal sentiment analysis model to output the sentiment analysis results of the current customer conversation data; obtaining target customer information initiating the current conversation, inputting the target customer information into a pre-trained customer classification and grading model for processing; generating a recommended customer strategy based on a Q-Learning algorithm; optimizing the service strategy using a preset generative AI model based on the current customer's intent, sentiment analysis results, and the recommended customer strategy; and generating a response method and response content based on the optimized service strategy. This solution constructs a dynamic knowledge base for the tobacco industry, combines customer emotional demands and behavioral habits, and formulates personalized service strategies to improve the accuracy and satisfaction of intelligent customer service responses.

Description

Technical Field

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

[0002] In the current business environment, businesses continue to demand more efficient and high-quality customer service. While question-answering systems based on natural language processing have become core tools, they face systemic challenges in key areas.

[0003] 1. Insufficient emotional interaction depth: It is difficult to accurately identify emotions in complex contexts (such as sarcasm and puns), resulting in a disconnect between human-machine collaboration;

[0004] 2. Limited service accuracy: Dynamic information such as industry policies and product regulations requires manual annotation and updating, and knowledge base updates rely on manual labor, resulting in delayed responses. For tobacco company clients, their inquiries often involve multi-dimensional information (such as order records, complaint history, and brand preferences). However, existing intelligent customer service technologies struggle to quickly correlate these multi-dimensional information, resulting in insufficient accuracy in cross-scenario information retrieval.

[0005] 3. Difficulty Balancing Cost and Experience: Customer service data is disconnected from business decision-making processes, making it difficult for companies to dynamically allocate resources based on customer value. Existing systems (such as intelligent robots and FAQ libraries) focus on answering simple, standardized questions like order inquiries and policy interpretations, lacking in-depth insights into customer emotions (e.g., anxiety, expectations, dissatisfaction) and behavioral habits (e.g., sensitivity to order cycles and brand reliance).

[0006] 4. The tobacco industry is strictly regulated, and customer service responses must strictly comply with policies and regulations. However, general generation models are prone to outputting non-compliant content, posing legal risks. Existing systems lack structured knowledge support and rely heavily on manual intervention, driving up operating costs. Furthermore, static knowledge bases cannot adapt to sudden events, and emergency response capabilities are weak. Summary of the Invention

[0007] 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, build a dynamic knowledge base for the tobacco industry, combine customer emotional demands and behavioral habits, and formulate personalized service strategies to improve the accuracy and satisfaction of intelligent customer service responses.

[0008] A method for classifying and grading tobacco enterprise customers based on sentiment analysis, the method comprising:

[0009] Acquire conversation data entered by the customer in the tobacco customer service scenario, parse the input conversation data, and search a pre-built tobacco industry dynamic knowledge base based on the parsed conversation data to determine the customer's intention;

[0010] Input the conversation data into a pre-trained multimodal sentiment analysis model and output a sentiment analysis result of the current customer conversation data;

[0011] Obtaining target customer information for initiating the current session, and inputting 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 data including 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 including multiple pre-set types;

[0012] Generate recommended customer strategies based on the Q-Learning algorithm: define a strategy action space and dynamically adjust it based on customer grading. The state space includes at least customer grading and sentiment analysis results.

[0013] The preset generative AI model combines the current customer's customer intention, 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.

[0014] The above solution, by building a dynamic knowledge base for the tobacco industry and integrating features from multiple scenarios, makes the customer intent determined based on the dynamic knowledge base more accurate. It also achieves refined customer segmentation through multimodal sentiment analysis and the fusion of dynamic customer features. It dynamically adjusts the strategy action space using the Q-Learning algorithm to generate personalized service strategies for customers in different segments (such as angry and satisfied). It further optimizes service strategies based on generative AI models (such as deepseek R1) to ensure that the response content complies with industry policies, ensuring the accuracy and relevance of the response content, improving customer satisfaction, and reducing manual customer service intervention.

[0015] Preferably, it also includes updating the tobacco industry knowledge base, specifically: within a preset time, when the frequency of occurrence of keywords in the same tobacco industry policy document that are not included in the tobacco industry knowledge graph in the conversation data of different customers in the tobacco customer scenario reaches a preset growth rate, the update of the tobacco industry knowledge base is triggered; wherein, the conversation data includes conversation data with the intervention of manual customer service.

[0016] After the tobacco industry policy is updated, the above settings can automatically trigger the update of the tobacco industry knowledge base, which can reduce manual maintenance of the knowledge base while ensuring the timeliness of policy responses. After the update, it can also reduce manual customer service intervention for similar policy issues.

[0017] Preferably, the tobacco industry knowledge base is built based on the native graph database Neo4j, covering the knowledge graph of multiple entities and businesses in the tobacco industry chain;

[0018] Dividing subgraph indexes based on the GraphRAG framework to provide structured context for the generative AI model;

[0019] The tobacco industry policy document is parsed through the generative AI model to extract key entities and dynamically update the knowledge graph nodes.

[0020] The Neo4j knowledge base supports complex path queries with more than 30 hops, which helps improve response speed; GraphRAG provides precise business scenario context for generative AI, making the knowledge base highly compatible with generative AI and able to accurately judge customer intentions.

[0021] Preferably, when the number of consultations for the same keyword reaches a preset threshold within a preset time period, historical complaint cases are automatically associated to generate an emergency subgraph, and the generative AI model performs similarity retrieval based on the emergency subgraph to generate a template for processing similar problems.

[0022] Generate emergency subgraphs for similar problems. The subgraphs support similarity retrieval. When encountering similar problems, similar problem processing templates can be directly called to improve the efficiency of problem handling in emergency situations.

[0023] Preferably, the multimodal sentiment analysis model performs the following steps:

[0024] 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 magnitude index.

[0025] The Wav2Vec2.0 model is used to extract the acoustic features of the customer's speech. Combined with the pre-built tobacco industry sentiment dictionary, the Bi-LSTM model is used to identify the speech emotion and calculate the emotion amplitude index.

[0026] The timestamps of text and speech sentiment analysis results are aligned, and weighted fusion is performed through a gated fusion mechanism and an attention mechanism to generate a multimodal joint feature vector.

[0027] The multimodal joint feature vector is optimized based on the RoBERTa model, and the sentiment analysis result is output.

[0028] For the same content, sentiment analysis of speech and text is performed separately and then fused to make the output sentiment analysis results more accurate.

[0029] Preferably, the customer classification and grading model performs the following steps:

[0030] Construct a multi-dimensional customer feature vector based on the input target customer's historical behavior data, including behavioral characteristics, emotional characteristics, and value characteristics;

[0031] After normalizing the customer feature vector, an improved DBSCAN clustering algorithm is used to dynamically calculate the neighborhood radius and set the minimum number of samples in layers to generate initial clusters and noise point sets;

[0032] Perform secondary clustering or manual correction on noise points to finally complete multi-level customer classification.

[0033] The dynamic neighborhood radius design is adopted, and secondary clustering and manual correction are performed on noise points to make the classification more accurate.

[0034] Preferably, the improved DBSCAN clustering algorithm includes:

[0035] The dynamic neighborhood radius is calculated as ,

[0036] in, is the Euclidean distance from the kth nearest neighbor to point xi, The p-th quantile of the local density value in the data set is used, 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 lowered and a smaller eps is used; for high-density areas, the p-value is increased and a larger eps is used.

[0037] Adaptive p-value adjustment optimizes cluster boundaries, making region division more reasonable while taking into account both clustering speed and accuracy.

[0038] Preferably, the service response control generates a recommended customer policy based on a preset service policy, including:

[0039] 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.

[0040] It can automatically allocate different levels of customer service resources based on the customer's emotional characteristics. For example, for requests with large fluctuation variance, they can be allocated to high-level customer service, which can improve the problem-solving rate and satisfaction and reduce the complaint rate.

[0041] Preferably, customer satisfaction is dynamically updated based on customer sentiment, complaint frequency, and consultation response time. Accuracy ratings are obtained from customers, and the reward function of the Q-Learning algorithm is determined based on these accuracy, customer satisfaction, and consultation response time. Optimizing the customer strategy recommendation algorithm based on customer satisfaction and other factors can improve the accuracy of recommendation results.

[0042] The present application also aims to provide a tobacco enterprise customer classification and grading system based on sentiment analysis, applying any of the above methods. Due to the application of the above method, the system also has the advantages of the above method.

[0043] To sum up, this application adopts the above-mentioned solution, which uses generative AI to retrieve the tobacco industry dynamic knowledge base to identify customer intentions, and the analysis results are more accurate. The final generated response method and response content take into account customer emotions and customer historical behavior, which enables intelligent customer service to formulate personalized service strategies for customers, which is conducive to solving complex conversation content and improving customer satisfaction. DETAILED DESCRIPTION

[0044] The embodiments of the present invention are described in detail below.

[0045] This embodiment relates to a method for classifying and grading tobacco enterprise customers based on sentiment analysis, characterized in that the method includes:

[0046] S1, obtaining conversation data input by the customer in the tobacco customer service scenario, parsing the input conversation data, searching a pre-built tobacco industry dynamic knowledge base based on the parsed conversation data, and judging the customer's intention.

[0047] The conversation data entered by customers in tobacco customer service scenarios refers to conversation data entered by customers in customer service scenarios, such as in the customer service dialogue window. This conversation data can be text data or voice data. This embodiment uses a generative AI model to parse the input conversation data and, based on the parsed conversation data, searches the tobacco industry dynamic knowledge base to determine customer intent.

[0048] The method includes constructing a tobacco industry dynamic knowledge base, which includes relevant policy documents and specifications of the tobacco industry, so that the relevant policy documents and specifications involved in customer consultation can be determined by searching the tobacco industry knowledge base.

[0049] In this embodiment, the dynamic knowledge base of the tobacco industry is dynamically updated. Specifically, within a preset time, when the frequency of occurrence of keywords in the same tobacco industry policy document that are not included in the tobacco industry knowledge graph in the conversation data of different customers in the tobacco customer scenario reaches a preset growth rate, the update of the tobacco industry knowledge base is triggered; wherein, the conversation data includes conversation data intervened by manual customer service. According to the above settings, the automatic update of the industry knowledge base can be triggered, and there is no need to manually update the latest industry policy documents to the knowledge base. For example, when the frequency of occurrence of policy document keywords increases by 50% day-on-day, an update is triggered. Since the supply customers of the tobacco industry are relatively stable, when the frequency of occurrence of the same policy document (which is beneficial to the supply volume) increases significantly, it can be inferred that the relevant policy has been updated, and automatic update is triggered based on this setting.

[0050] In this example, the tobacco industry knowledge base is built on the native graph database Neo4j, covering a knowledge graph of multiple entities and businesses across the tobacco industry chain. The Neo4j graph database enables efficient traversal and real-time reasoning of complex relationships, building a knowledge base covering 12 entity types and 28 business relationships across the tobacco industry chain. It supports real-time queries of complex paths with more than 30 hops, pre-defined compliance rules and customer tiering rules, and enables real-time risk warnings and strategy recommendations.

[0051] The tobacco industry dynamic knowledge graph, based on the GraphRAG framework, implements knowledge graph subgraph partitioning and constructs a retrieval index for business scenarios, providing a structured context for generative AI such as the DeepSeek R1 large model, thereby enabling retrieval of the tobacco industry dynamic knowledge graph through the generative AI model.

[0052] In this embodiment, when the automatic update of the industry knowledge base is triggered, the generative AI model parses the tobacco industry policy documents, extracts key entities, and dynamically updates the knowledge graph nodes. Specifically, the tobacco knowledge graph is constructed based on the Neo4j graph database, and business events are accessed in real time through Kafka. The Neo4j APOC plug-in is called to batch write to ensure real-time updates of the graph.

[0053] When this policy is involved again in the future, generative AI can directly respond based on this policy and customer information without the need to contact manual customer service, which is conducive to speeding up response time and improving customer satisfaction.

[0054] In a preferred embodiment, when the number of consultations for the same keyword reaches a preset threshold within a preset time period, historical complaint cases are automatically associated to generate an emergency subgraph, and the generative AI model performs similarity retrieval based on the emergency subgraph to generate a template for processing similar problems. For example, when the number of consultations per day regarding the same policy document keyword or the same issue exceeds a threshold, historical complaint cases are automatically associated to generate an emergency subgraph. The data for the past three months is stored in Neo4j memory, and an LRU cache strategy is configured with a hit rate of ≥95%. Historical data (over three months) is exported as a subgraph snapshot, stored in a graph database, and partitioned by timestamp.

[0055] The knowledge base is divided into subgraphs based on business scenarios, each containing related entities and relationships. The subgraph embedding vectors are generated using the GraphSAGE algorithm and stored in the Faiss index, supporting similarity retrieval.

[0056] In this example, a generative AI model, using DeepSeek R1 as an example, converts conversation data expressed in natural language into Cypher, enabling it to search the Neo4j graph database for relevant subgraphs, extract key entities and relationships, and generate prompt word templates. Furthermore, a dynamic weighting mechanism is employed to apply a weight gain (e.g., a weight × 1.2) to knowledge prompt words in DeepSeek R1's attention layer. The model output is then checked for policy keywords, triggering a regeneration process if they are missing. This improves the accuracy of the output.

[0057] S2: Input the conversation data into a pre-trained multimodal sentiment analysis model and output a sentiment analysis result of the current customer conversation data.

[0058] The multimodal sentiment analysis model performs the following steps:

[0059] 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 magnitude index.

[0060] The acoustic features of customer speech are extracted through the Wav2Vec2.0 model. Combined with the pre-built tobacco industry sentiment dictionary, the Bi-LSTM model is used to identify speech emotions and calculate the sentiment amplitude index. First, for the tobacco industry, a special sentiment dictionary for the tobacco field is constructed by combining the Word2Vec model and the K-means clustering algorithm. The dictionary contains more than 3000 professional terms and their sentiment weights. In a specific embodiment, the construction of the special sentiment dictionary for the tobacco field specifically includes:

[0061] We selected a corpus of 1 million tobacco industry records, including complaint records, policy documents, and customer service conversations. We used Word2Vec training with settings like vector_size=300, window=5, and min_count=10 to generate word vectors. We also used K-means++ clustering with the number of clusters K=15 and a silhouette coefficient ≥0.65. We extracted and manually annotated the top 5% of high-frequency words.

[0062] We use clustering algorithms to discover sentiment clusters and combine them with manual annotation to build a sentiment lexicon. We use the sentiment weights in the lexicon as features to fine-tune the pre-trained language model and enhance the influence of domain keywords in the model's attention layer.

[0063] The training method of the Wav2Vec2.0 model is as follows:

[0064] The pre-trained wav2vec2-base-960h model was used to extract features and output frame-level features. Each frame was 50ms long and average-pooled to generate a 128-dimensional acoustic vector. This model was based on a bidirectional LSTM (with 256 hidden layers) followed by an 8-head self-attention layer, and finally output a sentiment label using a softmax filter. Random noise (SNR = 10dB), speed variation (±10%), and pitch perturbation (±50 cents) were added. 10,000 annotated speech samples were used, divided into 80% training, 10% validation, and 10% test sets.

[0065] Emotional fluctuation detection uses a sliding window analysis with a 30-second window and a 10-second step size. The entropy of the emotion tag within the window is calculated, and the entropy mutation points are marked as emotional fluctuations. The emotional fluctuation cycle of the voice clip is analyzed to generate an emotional change curve and associate it with the consultation question type.

[0066] The timestamps of text and speech sentiment analysis results are aligned, and weighted fusion is performed through a gated fusion mechanism and an attention mechanism to generate a multimodal joint feature vector.

[0067] Among them, the fusion weight g=\sigma \left ( {w\times \left [ {text features; speech features} \right ]+b} \right ) , g represents the weight of text features, σ is the Sigmoid function, the symbol “;” represents the vertical concatenation operation of the matrix, and the weight is biased towards the high confidence mode; w and b are training parameters, w is the weight value, and b is the offset;

[0068] The multimodal joint feature vector is optimized based on the RoBERTa model to output a sentiment analysis result. In this embodiment, the sentiment analysis result output is a pre-classified sentiment label (7 categories: anger, satisfaction, neutral, etc.), with an accuracy rate of ≥ 89%.

[0069] S3, obtaining target customer information for initiating the current session, and inputting 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 data including 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 including multiple pre-set types;

[0070] Historical data is stored in a unified data platform. This method involves building a data platform, including:

[0071] Collect customer historical data, including customer consultation records, ordering behavior, brand preferences, and other multi-source data. The multi-source data includes but is not limited to customer historical consultation records, ordering frequency, payment methods, brand preferences, complaint records, etc.;

[0072] Integrate the collected multi-source data into a unified data platform to ensure data integrity and consistency. Use ETL (Extract, Transform, Load) tools to clean, transform, and load data to ensure data quality.

[0073] It combines the distributed storage system OSS with the relational database MySQL to store structured and unstructured data, ensuring efficient data access and storage.

[0074] In the above manner, customers' historical data are uniformly stored in a unified data platform for easy retrieval and access when needed.

[0075] The customer segmentation model performs the following steps:

[0076] S31. Use multi-dimensional customer feature engineering to collect customer behavior data such as historical consultation records, order frequency, brand preferences, etc., and construct a customer feature vector containing 18 dimensions. The historical behavior data includes:

[0077] Behavioral data: Extract customer historical order records from the management platform, including order frequency (times / month), single order quantity (items), and brand concentration (the proportion of order volume of the top three brands = the order volume of the top three brands / the total order volume).

[0078] Payment method classification: proportion of cash payment, proportion of online payment, and proportion of payment on account.

[0079] Sentiment data: Extract sentiment polarity (range [-1,1]), sentiment fluctuation variance ( ), complaint frequency statistics.

[0080] Valuable data: ARPU (average monthly spending), calculated by extracting the total customer spending over the past six months from the financial system. Customer lifecycle stages, categorized based on the RFM model, such as new customer, growth, maturity, and decline. Potential churn risk score: predicted using a logistic regression model, with input features including the order interval growth rate and the downward trend in sentiment polarity.

[0081] The extracted 18-dimensional customer feature vectors include:

[0082] Behavioral characteristics: order frequency, brand concentration, payment method;

[0083] Emotional characteristics: mean emotional intensity of historical consultations and variance of emotional fluctuations;

[0084] Value characteristics: ARPU value, customer lifecycle stage, potential churn risk score.

[0085] Based on the extracted features and sentiment analysis results, a customer profile can be generated. Based on this profile, interpretable customer segmentation results are output to support business decision-making and analysis. Furthermore, this solution implements different customer retention strategies based on the profile, such as promoting limited-edition new products to Level 1 customers (high-value customers) and strengthening customer relationship maintenance for Level 6 customers (potential churn).

[0086] S32. Use Min-Max standardization and Z-Score standardization to process the feature vector.

[0087] S33. Use the improved DBSCAN clustering algorithm, set the dynamic neighborhood radius based on the dynamic calculation of feature space distribution, and set the minimum number of samples according to the customer density distribution layer.

[0088] Dynamic neighborhood radius (eps) calculation: ,in, is the Euclidean distance from the kth nearest neighbor to point xi, where k is selected according to the data dimension. is the p-th quantile of all local density values ​​in the dataset, where p is adaptively selected based on the local density. The dynamic parameter allows processing both high-density compact clusters and low-density sparse clusters simultaneously.

[0089] The data is divided into density layers according to the average k-nearest neighbor distance of the data points. The high-density area reduces p and uses a smaller eps, while the high-density area increases p and uses a larger eps.

[0090] S34: Perform secondary clustering on noise points, i.e., customers who cannot be categorized, using KNN to fill in the nearest neighbor tier. Based on the improved DBSCAN, generate initial clusters and noise point sets. For each noise point, calculate its distance to the center of each cluster and assign it to the nearest cluster. If the KNN result does not conform to the business logic, manually set the tier boundaries to correct for purely data-driven biases. Finally, perform six-tier categorization based on the clustering results. The six-tier categorization in this embodiment is shown in Table 1.

[0091] Table 1 Correspondence between customer characteristics and customer classification

[0092]

[0093] 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 at least includes the customer classification level and sentiment analysis results.

[0094] In this embodiment, the state space of the Q-Learning algorithm is defined as customer tier levels (6 dimensions), real-time emotional labels (7 categories), and historical behavioral characteristics (such as order frequency ranges). This enables dynamic decision-making driven by tiering. The service response action control includes eight basic actions and 12 combination strategies. Basic actions include emotional soothing, quick resolution, escalation channel, expert intervention, and emotional accounting. Combination strategies are based on basic actions, scenario conditions, customer type, and other factors. The Q-Learning algorithm then generates recommended customer strategies for the service response action control. For example, the combination of emotional soothing, quick resolution, and escalation channel strategies can respond to angry and emotionally unstable customers, achieving a golden 30-second response. The emotional accounting action can respond to satisfied and silent customers, providing personalized care and enhancing customer emotional value.

[0095] Reward function:

[0096] ,Among them, accuracy rate: problem solving rate after strategy execution, satisfaction rate: dynamically calculated based on sentiment analysis results.

[0097] S5, the preset generative AI model combines the customer intention of the current customer, the 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.

[0098] In this step, DeepSeek optimizes the service strategy based on the current customer sentiment and intent derived from the multimodal sentiment analysis model, as well as the customer strategy recommended by the Q-Learning algorithm. This strategy is then delivered to the customer using the DeepSeek model. For example, if customer A is a satisfied customer, their current sentiment is neutral, and their intent is information seeking, the recommended strategy is based on the emotional account. DeepSeek will combine this information with the customer's intent and emotional account to optimize the response content and format (e.g., text-based replies, congratulatory closing remarks), ensuring that neutral sentiment shifts to positive sentiment.

[0099] In this embodiment, customer satisfaction is dynamically updated based on the customer's emotional polarity, complaint frequency, and consultation response time. A dynamic customer satisfaction update mechanism is established: Customer Satisfaction = α × Emotional Polarity + β × Complaint Frequency + γ × Response Time, where the weights α, β, and γ are optimized through grid search to ensure that their contribution to customer grading complies with industry regulations. Accuracy ratings are obtained from customers, and the aforementioned reward function of the Q-Learning algorithm is determined based on these accuracy, customer satisfaction, and consultation response time. This provides feedback on the service strategy formulated by S5, and Q-Learning is optimized based on this feedback data.

[0100] Furthermore, this solution collects feedback data from across the entire supply chain in real time, building a foundation for closed-loop optimization. This closed-loop optimization strategy improves 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 (e.g., a decrease in negative sentiment intensity), and subsequent behavioral data, dynamic and personalized customer service is achieved, significantly improving tobacco companies' operational efficiency and customer experience. System performance and business effectiveness are quantified through satisfaction index, strategy hit rate, average response time, and customer service intervention rate, driving optimization strategies.

[0101] Continuously improve system performance based on feedback data, including Q-Learning strategy optimization, sentiment analysis model iteration, and dynamic correction of knowledge graphs.

[0102] The solution of this embodiment closely aligns with the regulatory and business needs of the tobacco industry, and offers the following benefits in terms of emotional dimension integration, closed-loop service design, and customized industry knowledge graphs:

[0103] 1. Enrich customer segmentation dimensions. Based on historical data such as sales data and customer value, sentiment analysis is introduced as a customer segmentation and classification dimension to further segment customer groups.

[0104] 2. Improve customer satisfaction. By linking emotional tags with knowledge graphs, we can achieve a closed loop of "emotion recognition → classification and grading → targeted service → data feedback," responding precisely to customer needs and effectively improving customer satisfaction and loyalty.

[0105] 3. Reduce operating costs: The customized tobacco industry knowledge graph effectively shares most of the common customer issues, reduces the workload and operating costs of manual customer service, and improves overall customer service efficiency.

[0106] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the 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 conversation data entered by the customer in the tobacco customer service scenario, parse the input conversation data, and search a pre-built tobacco industry dynamic knowledge base based on the parsed conversation data to determine the customer's intention; Input the conversation data into a pre-trained multimodal sentiment analysis model and output a sentiment analysis result of the current customer conversation data; Obtaining target customer information for initiating the current session, and inputting 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 including 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 including multiple pre-set types; Generate recommended customer strategies based on the Q-Learning algorithm: define a strategy action space and dynamically adjust it based on customer grading. The state space includes at least customer grading 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; The multimodal sentiment analysis model performs sentiment analysis on the input conversation data in combination with a pre-built tobacco industry sentiment dictionary. The tobacco industry sentiment dictionary is a tobacco industry-specific sentiment dictionary constructed by combining the Word2Vec model and the K-means clustering algorithm with tobacco industry corpus. It also includes updating the tobacco industry dynamic knowledge base, specifically: When, within a preset time period, the frequency of occurrence of keywords in tobacco industry policy documents that are not included in the tobacco industry knowledge graph in conversation data between different customers in tobacco customer scenarios reaches a preset growth rate, an update of the tobacco industry dynamic knowledge base is triggered; wherein the conversation data includes conversation data involving manual customer service intervention; The tobacco industry dynamic knowledge base is built based on the native graph database Neo4j, 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 to extract key entities and dynamically update the knowledge graph nodes.

2. The method for classifying and grading tobacco enterprise customers based on sentiment analysis according to claim 1, characterized in that: When the number of consultations for the same keyword reaches a preset threshold within a preset time period, 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.

3. 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 magnitude index. The Wav2Vec2.0 model is used to extract the acoustic features of the customer's speech. Combined with the pre-built tobacco industry sentiment dictionary, the Bi-LSTM model is used to identify the speech emotion and calculate the emotion amplitude index. The timestamps of text and speech sentiment analysis results are aligned, and weighted fusion is performed through a gated fusion mechanism and an attention mechanism to generate a multimodal joint feature vector. The multimodal joint feature vector is optimized based on the RoBERTa model, and the sentiment analysis result is output.

4. 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 characteristics, emotional characteristics, and value characteristics; After normalizing the customer feature vector, an improved DBSCAN clustering algorithm is used to dynamically calculate the neighborhood radius and set the minimum number of samples in layers to generate initial clusters and noise point sets; Perform secondary clustering or manual correction on noise points to finally complete multi-level customer classification.

5. The method for classifying and grading tobacco enterprise customers based on sentiment analysis according to claim 4, characterized in that: The improved DBSCAN clustering algorithm includes: The dynamic neighborhood radius is calculated as espi = Qp{dk(x1),…,dk(xi)}, Where dk(xi) 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 lowered and a smaller eps is used; for high-density areas, the p-value is increased and a larger eps is used.

6. The method for classifying and grading tobacco enterprise customers based on sentiment analysis according to claim 4, 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.

7. 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.

8. A tobacco enterprise customer classification and grading system based on sentiment analysis, characterized in that: Use the method according to any one of claims 1 to 7.

Images