Method, system, device and medium for distributing insurance policy clients based on a relationship graph

Abstract

The application relates to a policy customer allocation method, system, device and medium based on a relationship graph, and relates to the technical field of insurance management. The method comprises the following steps: collecting customer data containing unstructured interactive text, calculating a sentiment index through pre-labeled sentiment word weight extraction of semantic information, constructing a dynamic knowledge graph in combination with structured business data, evaluating customer comprehensive value and risk level based on the graph and the business data, and finally realizing customer re-allocation when a businessperson leaves and risk level triggered collection when a policy is overdue. The application can realize accurate quantification of customer value and dynamic perception of risk, thereby driving intelligent allocation and compliance collection.

Description

Technical Field

[0001] This application relates to the field of insurance management technology, and in particular to a method, system, device and medium for policyholder allocation based on relationship graphs. Background Technology

[0002] In the insurance industry, technological advancements have driven continuous innovation in insurance business models. The intersection of insurtech, financial risk control, and artificial intelligence has brought new ideas and methods to the industry. By integrating various advanced technologies, insurance businesses can achieve better results in customer management and risk control, improving overall operational efficiency and service quality, and injecting new vitality into the industry's development. Simultaneously, with intensifying industry competition, the effective management and utilization of customer resources has become increasingly important. How to accurately allocate customers and conduct compliant debt collection have become critical issues that the insurance industry urgently needs to address.

[0003] In traditional insurance operations, when a frontline agent leaves, their clients need to be reassigned, typically using methods such as geographical rotation or simple average allocation. For clients with overdue premiums, a one-size-fits-all approach is often used when initiating collection processes. Furthermore, a large amount of unstructured interaction data, such as call recordings and online chat logs, containing information about customer emotions, intentions, and relationship networks, is not effectively utilized. The entire allocation and collection process lacks transparency and auditing capabilities, and decision-making relies heavily on human experience.

[0004] However, these traditional methods have significant drawbacks. First, customer redistribution relies on static rules, neglecting the multidimensional value of customers and easily leading to the loss of high-value customers. Second, collection strategies lack dynamic assessment of customers' willingness and ability to repay, resulting in low efficiency and a high likelihood of complaints. Furthermore, unstructured interactive data is not effectively utilized, leading to the waste of a large amount of valuable information. At the same time, the lack of transparency and auditing capabilities in the process makes it difficult to meet increasingly stringent financial compliance regulatory requirements. Summary of the Invention

[0005] The purpose of this application is to provide a policyholder allocation method based on relationship graphs, which can achieve accurate quantification of customer value and dynamic risk perception, thereby driving intelligent allocation and compliant collection.

[0006] Firstly, this application provides a policyholder allocation method based on a relationship graph, which employs the following technical solution: A policyholder allocation method based on relationship graphs, comprising: Collect customer data, which includes unstructured text interactions between customers and salespersons; Semantic information is extracted from the interactive text, and a weighted sentiment calculation is performed on a single interactive text based on the pre-labeled sentiment word weights to generate customer-side sentiment indicators and salesperson-side sentiment indicators, respectively. Based on the structured business data in the customer data, the customer-side sentiment indicators and the salesperson-side sentiment indicators, a knowledge graph containing customer nodes and salesperson nodes is constructed and dynamically updated. Based on the knowledge graph and business data, the comprehensive value score and risk level of the customer are calculated and obtained through customer value assessment and collection risk assessment. In response to the event that a salesperson's status changes to resignation, customer reallocation is performed based on the comprehensive value score of the customers under the resigned salesperson's name and the correlation between each customer and the candidate current salesperson in the corresponding knowledge graph. In response to an event that changes the policy status to overdue, a collection process is triggered and executed based on the risk level stated by the corresponding customer.

[0007] By employing the aforementioned technical solutions, unstructured interaction texts between customers and sales representatives can be collected, effectively utilizing previously overlooked data containing information such as customer emotions, intentions, and relationship networks. Weighted sentiment analysis of the interaction text generates sentiment indicators for both the customer and sales representative sides, quantifying customer emotional attitudes during interactions. A knowledge graph is constructed and dynamically updated based on structured business data and sentiment indicators, intuitively presenting the relationship between customers and sales representatives and reflecting changes in that relationship in real time based on new interaction data. Calculating a comprehensive value score and risk level through customer value assessment and collection risk assessment helps to accurately... Understanding customer value and risk profiles; redistributing assets based on customer comprehensive value scores and relevance when sales agents leave, avoiding inefficient asset redistribution, reducing high-value customer churn, and improving the productivity of new sales agents; triggering and executing collection processes based on risk levels when policies are overdue, enabling personalized and differentiated collection efforts, improving collection success rates, and reducing ineffective collection time; simultaneously, the constructed knowledge graph effectively reduces compliance risks and can be accumulated as knowledge assets for application in other scenarios, achieving precise quantification of customer value and dynamic risk perception, thereby driving intelligent allocation and compliant collection.

[0008] In a preferred embodiment, this application can be further configured as follows: the step of extracting semantic information from the interactive text and performing weighted sentiment calculation on a single interactive text based on pre-labeled sentiment word weights to generate customer-side sentiment indicators and salesperson-side sentiment indicators respectively includes: Identify customer-side and salesperson-side statements from a single interaction text; Sentiment analysis is performed on both customer-side and salesperson-side statements. The sentiment analysis is based on a preset sentiment word library and corresponding sentiment word weights to calculate and obtain the customer-side sentiment score and salesperson-side sentiment score for a single interaction. Based on the customer-side sentiment scores and salesperson-side sentiment scores from multiple interactions within a preset time window, a weighted calculation is performed to generate periodic customer-side sentiment indicators and salesperson-side sentiment indicators.

[0009] By adopting the above technical solutions, we can reliably quantify the emotional indicators of customers and salespersons from unstructured interactive text, providing data support for the construction of knowledge graphs. This helps to accurately identify customer emotions and intentions, thereby improving the accuracy of customer asset reallocation and collection strategies, reducing the loss of high-value customers, increasing collection success rates, reducing compliance risks, and accumulating knowledge assets.

[0010] In a preferred embodiment, this application can be further configured as follows: the step of constructing and dynamically updating a knowledge graph containing customer nodes and salesperson nodes based on structured business data in the customer data, the customer-side sentiment indicators, and the salesperson-side sentiment indicators includes: Customers and salespersons are used as nodes in the graph; Based on the policy ownership relationship between customers and agents, construct the first type of entity edge; Based on the periodic customer-side sentiment indicators and salesperson-side sentiment indicators, the correlation degree between customer nodes and salesperson nodes is calculated and obtained, and the second type of implicit relationship edge is constructed or updated using the correlation degree as the weight. In response to the processing of new interactive data, the first type of entity edges and the second type of implicit relationship edges are reconstructed, triggering an update of the knowledge graph.

[0011] By employing the aforementioned technical solutions, unstructured interactive text and structured business data between customers and sales agents are collected. Weighted sentiment analysis is then performed on the interactive text to generate sentiment indicators for both the customer and sales agent sides. Based on the business data and sentiment indicators, a knowledge graph is constructed and dynamically updated. This quantifies the implicit relationship between customers and sales agents, enabling the dynamic evolution of the knowledge graph and accurately reflecting their relationship. Through customer value assessment and collection risk assessment, a customer's comprehensive value score and risk level can be obtained, thereby achieving precise reallocation of high-value customers and differentiated, compliant premium collection. Customer reallocation and collection processes are triggered based on sales agent departures and policy overdue events, improving system flexibility and scalability to meet the needs of different scenarios. Simultaneously, it avoids value loss caused by indiscriminate reallocation of customer assets, increases collection success rates, and reduces compliance risks.

[0012] In a preferred embodiment, this application can be further configured as follows: in response to an event where a salesperson's status changes to resignation, a customer reallocation step is performed based on the comprehensive value score of customers under the resigned salesperson's name and the correlation between each customer and candidate current salespersons in the corresponding knowledge graph, including: Obtain the weights of the second type of implicit relationship edges between the customer nodes and each candidate salesperson node in the knowledge graph, and use them as the first degree of association; Based on all the connection paths between the customer node and the candidate salesperson node in the knowledge graph, the weights of each path are aggregated and calculated to obtain the multi-hop path connection strength as the second correlation degree. The first correlation and the second correlation are input into a preset correlation fusion function to calculate the total correlation between the customer and the candidate salesperson.

[0013] By adopting the above technical solution, unstructured interactive text between customers and salespersons is collected, semantic information is extracted, and weighted sentiment calculation is performed to generate sentiment indicators on the customer side and the salesperson side. Combined with structured business data, a knowledge graph is constructed and dynamically updated. Through customer value assessment and collection risk assessment, the comprehensive value score and risk level of customers are calculated. When a salesperson leaves, the weight of the second type of implicit relationship edge between the customer node and the candidate salesperson node is obtained as the first degree of association. The weights of all connected paths are aggregated to obtain the multi-hop path connection strength as the second degree of association. The two are then input into the association degree fusion function to obtain the total degree of association for customer reallocation. This can achieve accurate reallocation of high-value customers, reduce the churn rate of high-value density customers, and improve the productivity utilization rate of new salespersons.

[0014] In a preferred embodiment, this application can be further configured such that: in the step of calculating and obtaining the customer's comprehensive value score and risk level based on the knowledge graph and business data through customer value assessment and collection risk assessment, the step of obtaining the customer's comprehensive value score includes: The first indicator is the financial value dimension of the customer; the second indicator is the potential value dimension predicted based on policy data; and the third indicator is the relational value dimension measured by the customer's network centrality in the knowledge graph. The first, second, and third indicators are input into a trained customer value assessment model, which uses a machine learning algorithm to non-linearly fuse the first, second, and third indicators to output the comprehensive value score.

[0015] By adopting the above technical solutions, customer indicators from different dimensions, including financial value, potential value, and relationship value, can be obtained, enabling a more comprehensive measurement of customer value. By using a trained customer value assessment model to non-linearly fuse multi-dimensional indicators, the overall value score of customers can be calculated more accurately, providing a precise basis for subsequent customer reallocation. This helps reduce the churn rate of high-value, high-density customers and improve the productivity utilization rate of new sales staff.

[0016] In a preferred embodiment, this application can be further configured such that: in the step of calculating and obtaining the customer's comprehensive value score and risk level based on the knowledge graph and business data through customer value assessment and collection risk assessment, the step of obtaining the customer's risk level includes: Obtain static risk factors based on the policy's overdue status; Obtain dynamic risk signals based on the changing trends of the customer-side sentiment indicators and the salesperson-side sentiment indicators; The static risk factors and the dynamic risk signals are input into a trained risk assessment model. The risk assessment model uses a machine learning algorithm to perform correlation analysis on the static risk factors and the dynamic risk signals, calculates and obtains a risk score, and classifies customers into at least three risk levels: low, medium, and high, based on preset thresholds.

[0017] By adopting the above technical solutions, we can comprehensively consider the static risk factors of the policy's overdue status and the dynamic risk signals of the changing trends of emotional indicators on the customer side and the agent side. We can use a trained risk assessment model to calculate risk scores and classify risk levels, providing a basis for triggering different collection processes based on risk levels. This will improve collection efficiency, reduce ineffective collection time, and better achieve differentiated and precise collection under compliance control.

[0018] In a preferred embodiment, this application can be further configured as follows: the step of triggering and executing a collection process in response to an event that the policy status changes to overdue, based on the corresponding customer's risk level, includes: If the risk level is low, an automated alert message will be sent. If the risk level is medium, then push the standard collection script guidelines containing customer interest tags to the assigned salesperson. If the risk level is high, an auxiliary decision package containing historical customer sentiment indicators and compliance points will be generated, and a manual review and processing process will be triggered.

[0019] By adopting the above technical solutions, different collection processes are triggered based on the different risk levels of customers. For low-risk customers, automated reminder messages are sent, which can improve collection efficiency and reduce ineffective collection time. For medium-risk customers, standard collection scripts containing customer interest tags are pushed to assigned salespersons, which can achieve differentiated collection and improve collection success rate. For high-risk customers, auxiliary decision packages containing historical customer sentiment indicators and compliance points are generated and manual review and processing are triggered, which not only helps to improve collection results, but also ensures the compliance of the collection process, thereby improving collection efficiency, reducing compliance risks, and realizing the preservation of customer assets.

[0020] Secondly, this application provides a policyholder allocation system based on a relationship graph, employing the following technical solution: A policyholder allocation system based on a relationship graph includes: Data acquisition module: used to collect customer data, which includes unstructured interactive text between customers and salespersons; The sentiment quantification module is used to extract semantic information from the interactive text and perform weighted sentiment calculation on a single interactive text based on the pre-labeled sentiment word weights, generating customer-side sentiment indicators and salesperson-side sentiment indicators respectively. Knowledge graph construction module: used to construct and dynamically update a knowledge graph containing customer nodes and salesperson nodes based on the structured business data in the customer data, the customer-side sentiment indicators and the salesperson-side sentiment indicators; Customer scoring module: Based on the knowledge graph and business data, it calculates and obtains the customer's comprehensive value score and risk level through customer value assessment and collection risk assessment; The reallocation module is used to respond to the event that a salesperson's status changes to resignation. Based on the comprehensive value score of the customers under the resigned salesperson's name and the correlation between each customer and the candidate current salesperson in the corresponding knowledge graph, the module performs customer reallocation. Collection module: In response to an event where the policy status changes to overdue, the collection process is triggered and executed based on the risk level stated by the corresponding customer.

[0021] Thirdly, this application provides an electronic device that adopts the following technical solution: An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described policyholder allocation method based on a relationship graph.

[0022] Fourthly, this application provides a computer storage medium, as follows: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described policyholder allocation method based on a relationship graph.

[0023] In summary, this application has the following beneficial technical effects: 1. This application achieves customer redistribution by collecting unstructured interactive text, generating sentiment indicators, and constructing knowledge graphs, thereby reducing the churn rate of high-value density customers, improving the productivity utilization rate of new sales staff, avoiding the problem of high-value customer churn, and realizing the preservation of customer assets. 2. This application triggers and executes different collection processes based on the customer's risk level, which improves the collection success rate, reduces ineffective collection time, solves the problems of lack of personalized and dynamic risk perception and low success rate in collection, and improves collection efficiency. 3. This application process is fully traceable and undergoes policy compliance verification, which reduces the risk of regulatory penalties and customer complaints, and solves the problems of lack of process transparency, audit capabilities and high compliance risks, making compliance risks controllable. Attached Figure Description

[0024] Figure 1 This is a flowchart of a policyholder allocation method based on a relationship graph, as described in one embodiment of this application.

[0025] Figure 2 This is a flowchart of a sub-step of step S2 in one embodiment of this application.

[0026] Figure 3 This is a flowchart of a sub-step of step S3 in one embodiment of this application.

[0027] Figure 4 This is a flowchart of a sub-step of step S5 in one embodiment of this application.

[0028] Figure 5 This is a flowchart of the steps for obtaining the customer's comprehensive value score in step S4 of one embodiment of this application.

[0029] Figure 6 This is a flowchart of the steps for obtaining the customer's risk level in step S4 of one embodiment of this application.

[0030] Figure 7 This is a flowchart of a sub-step of step S6 in one embodiment of this application.

[0031] Figure 8 This is a schematic diagram of the structure of a policy customer allocation system based on a relationship graph, which is one embodiment of this application.

[0032] Figure 9This is a schematic block diagram of an electronic device in one embodiment of this application.

[0033] Attached labels: 1. Data acquisition module; 2. Sentiment quantification module; 3. Graph construction module; 4. Customer scoring module; 5. Redistribution module; 6. Collection module. Detailed Implementation

[0034] The following is in conjunction with the appendix Figure 1-9 This application will be described in further detail.

[0035] It should be noted that, in the embodiments of this invention, the data related to object information, when applied to specific products or technologies, requires the permission or consent of the object. Furthermore, the collection, use, and processing of this data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. In other words, if the embodiments of this invention involve data related to an object, it must be obtained with the object's authorization and consent, the authorization and consent of relevant departments, and in accordance with the relevant laws, regulations, and standards of the country and region. If the embodiments involve personal information, the acquisition of all personal information requires the individual's consent. If sensitive information is involved, the separate consent of the information subject is required. The embodiments also need to be implemented with the object's authorization and consent.

[0036] refer to Figure 1 A policyholder allocation method based on relationship graphs, specifically including: S1. Collect customer data, which includes unstructured text interactions between customers and salespersons.

[0037] Specifically, by connecting to the insurance company's customer relationship management system, call center system, and online customer service platform, the system collects unstructured interactive text such as transcripts of recorded calls between customers and sales representatives, online chat logs, and email bodies. During the collection process, the system first confirms that it has obtained explicit authorization from the customer for the relevant data processing, and performs real-time anonymization processing on sensitive personal information in the interactive text. For example, it uses named entity recognition technology to replace names, ID numbers, bank account numbers, etc., with irreversible identifiers.

[0038] In summary, by transforming previously isolated and overlooked unstructured interactive text into analyzable data assets, a data foundation has been laid for uncovering customers' implicit intentions and relationship networks. At the same time, through strict authorization and de-identification mechanisms, compliance of data processing has been ensured from the source, reducing the risk of information leakage.

[0039] S2. Extract semantic information from the interactive text and perform weighted sentiment calculation on the single interactive text based on the pre-labeled sentiment word weights to generate customer-side sentiment indicators and salesperson-side sentiment indicators respectively.

[0040] Specifically, firstly, natural language processing technology is applied to segment the text of a single interaction and determine its role, accurately distinguishing between customer-side and salesperson-side statements. Next, the system calls upon a pre-built sentiment lexicon for the financial insurance field, containing positive and negative sentiment words with weights annotated by domain experts. The system then segments the segmented text, matches sentiment words, and performs a weighted sum based on their weights to calculate the independent customer-side and salesperson-side sentiment scores for a single interaction.

[0041] The above transforms subjective and vague emotional states in dialogue into objective and quantifiable values, achieving a preliminary quantification of interactive emotions and providing basic signals for subsequent analysis.

[0042] S3. Based on structured business data, customer sentiment indicators, and salesperson sentiment indicators in customer data, construct and dynamically update a knowledge graph containing customer nodes and salesperson nodes.

[0043] Specifically, the system first extracts structured business data such as customer attributes, policy details, payment records, and agent files from the core insurance business system. Then, using customers and agents as entity nodes, the system constructs the first type of entity edges based on explicit policy affiliation relationships, forming the basic framework of the knowledge graph. Next, the system uses calculated sentiment scores to generate a correlation degree describing the strength of the relationship between customers and agents according to a predefined rule, and uses this correlation degree as weight to construct or update the second type of relationship edges in the knowledge graph.

[0044] In summary, by integrating structured business relationships with quantitative sentiment data into a unified knowledge graph, a hybrid network containing explicit business links and implicit relationship links is constructed, providing a structured data relationship foundation for subsequent value and risk assessment.

[0045] S4. Based on knowledge graphs and business data, calculate and obtain the customer's comprehensive value score and risk level through customer value assessment and collection risk assessment.

[0046] Specifically, in customer value assessment, the system integrates financial information from business data with network structure information from knowledge graphs, calculates the customer's overall value score using a customer value assessment model, and outputs the customer's comprehensive value score. In collection risk assessment, the system integrates overdue status information from business data with sentiment analysis information from interactive text, calculates the customer's risk level using a collection risk assessment model, and outputs the customer's risk level based on preset thresholds.

[0047] The above achieves the quantitative scoring of customer value and the classification of customer risk levels, providing a clear and actionable basis for subsequent business decisions.

[0048] S5. In response to the event that a salesperson's status changes to resignation, perform customer reassignment based on the comprehensive value score of the customers under the resigned salesperson's name and the correlation between each customer and the candidate current salesperson in the corresponding knowledge graph.

[0049] Specifically, this step is automatically triggered when the system detects a salesperson's status change to "resigned." The system first obtains the overall value score of all customers under the resigned salesperson's name. Then, the system queries and retrieves the correlation between each customer and each candidate current salesperson from the customer-specific knowledge graph. Based on each customer's value score and its correlation with the candidate, the system calculates and executes an allocation plan using a matching logic.

[0050] In the customer reallocation decision-making process, two key dimensions—the customer's own value and the strength of their relationship with the target service provider—are introduced simultaneously. This ensures that the allocation results can better guarantee service continuity and customer satisfaction, and initially realizes the intelligentization of allocation.

[0051] S6. In response to an event where the policy status changes to overdue, trigger and execute the collection process based on the corresponding customer's risk level.

[0052] Specifically, this step is automatically triggered when the system detects that the policy status has changed to overdue. The system first reads the risk level assessed for the customer corresponding to the policy, and then selects and triggers the corresponding collection process from the predefined strategy library based on the different risk levels.

[0053] The above-mentioned differentiated processing procedures based on the risk level of customers have changed the past model of using a single collection method for all overdue customers, and provided an execution framework for improving the targeting and compliance of collection.

[0054] refer to Figure 2 Furthermore, in one embodiment, step S2 is refined into the following sub-steps: S20. Identify customer-side statements and salesperson-side statements from a single interaction text.

[0055] Specifically, a sentence boundary recognition and role classification model from natural language processing technology is applied to process the text of a single interaction. This model, trained on a corpus of dialogues from the insurance sector, can accurately identify each independent sentence in the conversation. Then, based on the initiation phrase, specific address patterns, and semantic roles within the sentences, the system automatically categorizes the identified sentences as either customer-side or salesperson-side statements, thus achieving precise separation of the interaction participants.

[0056] In summary, the analysis of each interactive text laid a structured data foundation for subsequent independent analysis of customer and salesperson sentiment, ensuring the accuracy and relevance of sentiment analysis and avoiding analytical errors caused by role confusion.

[0057] S21. Perform sentiment analysis on the customer-side statements and the salesperson-side statements respectively. The sentiment analysis is based on a preset sentiment word library and the corresponding sentiment word weights, and calculates the customer-side sentiment score and the salesperson-side sentiment score for a single interaction.

[0058] Specifically, sentiment analysis is performed separately for the separated customer-side and salesperson-side statements. The system calls upon a sentiment lexicon built specifically for financial and insurance service scenarios. This lexicon not only includes general sentiment terms but also focuses on a large number of domain-specific sentiment terms related to insurance consultation, claims, complaints, and sales. Each term is pre-labeled with a weight value reflecting its sentiment polarity and intensity. Then, the system performs word segmentation and part-of-speech tagging on the input statement, matches it with terms from the sentiment lexicon, and calculates the independent customer-side and salesperson-side sentiment scores for that interaction using a weighted summation algorithm based on the weights of the matched terms.

[0059] In summary, by using a domain-customized sentiment dictionary and quantitative calculations, the subjective emotional attitudes contained in unstructured text are transformed into objective and comparable data indicators, thus achieving accurate measurement of the emotions in a single interaction.

[0060] S22. Based on the customer-side sentiment scores and salesperson-side sentiment scores from multiple interactions within a preset time window, calculate and generate periodic customer-side sentiment indicators and salesperson-side sentiment indicators using a weighted average.

[0061] Specifically, a configurable observation time window is set, such as the last 7 days, and the system collects the single sentiment scores corresponding to all interactions that occur within this window. Then, the system uses a time-decay weighted model to calculate the sentiment scores, assigning different weights to sentiment scores at different points in time within the window. Generally, recent interactions are given higher weights, while the weights of older interactions decrease according to a predefined pattern. Next, the system calculates a weighted average of all scores from both the customer and salesperson sides within the time window, ultimately outputting periodic sentiment indicators for both sides, enabling the system to capture the evolution trend of sentiment over time.

[0062] In summary, by integrating time-series information and decay weighting, a dynamic indicator reflecting recent emotional state and trends is generated, overcoming the randomness that may exist in a single analysis and providing a stable, reliable, and time-meaning quantitative input for constructing a dynamically updated knowledge graph.

[0063] In addition, refer to Figure 3Furthermore, in one embodiment, step S3 is refined into the following sub-steps: S30. Treat customers and salespersons as nodes in the graph.

[0064] Specifically, each customer and each salesperson is instantiated as an independent node in the knowledge graph. Each customer node is associated with structured attributes extracted from the business system, such as customer ID and insurance period, while each salesperson node is associated with their team and job level, ultimately defining the core entities of the knowledge graph.

[0065] The above abstracts real-world insurance business participants into computable digital objects, establishing the main framework for constructing a network model describing the complex relationship between them, and forming the basis for subsequent relationship linking and graph analysis.

[0066] S31. Based on the policy ownership relationship between the customer and the agent, construct the first type of entity edge.

[0067] Specifically, by traversing the policy data, whenever the insured of a policy is a customer and the salesperson is a salesperson, a first-type entity edge is created between the corresponding customer node and salesperson node in the graph. The first-type entity edge has a clear type label and directionality, such as being labeled as "policy ownership" and pointing from the salesperson to the customer. Attributes such as policy number and product type can be attached to the edge.

[0068] The above transforms discrete business records into an interconnected network structure, intuitively solidifying the most fundamental service and ownership relationship between customers and salespersons, forming the static backbone of the graph, and providing stable anchor points for superimposed dynamic interactive relationships.

[0069] S32. Based on periodic customer-side sentiment indicators and salesperson-side sentiment indicators, calculate the correlation between customer nodes and salesperson nodes, and use the correlation as weight to construct or update the second type of implicit relationship edge.

[0070] Specifically, the system acquires sentiment indicators from both the customer and salesperson sides over a specific time period. This data, combined with interaction frequency data within the same period, is then processed using a predefined correlation calculation function. For example, the function uses interaction frequency as the base and a weighted sum of sentiment indicators from both sides as an adjustment factor to calculate a comprehensive value representing the closeness and harmony of the relationship—the correlation degree. Using this correlation degree as weight, the system creates or updates a second-type implicit relationship edge between the corresponding customer and salesperson nodes, thus quantifying unstructured interactive sentiment into a calculable relationship strength.

[0071] The above mapping of subjective and dynamic interaction quality into the quantitative weights of edges in the knowledge graph makes the previously implicit and difficult-to-capture service affinity and customer relationship temperature explicit and continuously measurable, greatly enriching the semantic connotation and computational value of the graph.

[0072] S33. In response to the processing of new interactive data, the first type of entity edges and the second type of implicit relationship edges are reconstructed, triggering an update of the knowledge graph.

[0073] Specifically, whenever a new customer interaction occurs and is processed, generating a new periodic sentiment indicator, a graph update event is automatically triggered. This update not only maintains the accuracy of the first type of entity edges based on the latest policy data, but more importantly, it calls the correlation calculation logic to recalculate the correlation between the affected customer and agent node pairs using the latest sentiment indicator, and immediately updates the weights of their second type of implicit relationship edges.

[0074] In summary, by updating the knowledge graph with new interactive data, the knowledge graph can be transformed from a static snapshot of business relationships into a dynamic system that can evolve in real time with each customer interaction. This ensures the timeliness of the weights of implicit relationship edges in the graph, so that any analysis, evaluation, and decision based on the knowledge graph can reflect the latest interpersonal interaction status, thus guaranteeing the accuracy and timeliness of subsequent intelligent decisions.

[0075] In addition, refer to Figure 4 Furthermore, in one embodiment, step S5 is refined into the following sub-steps: S50. Obtain the weights of the second type of implicit relationship edges between customer nodes and each candidate salesperson node in the knowledge graph, and use them as the first degree of association.

[0076] Specifically, when it is necessary to assess the correlation between a specific customer and a candidate salesperson, the system first locates the corresponding customer node and salesperson node in the graph, and then searches whether there is a second type of implicit relationship edge between these two nodes, calculated by sentiment indicators and interaction frequency. If the second type of implicit relationship edge exists, its currently stored weight value is directly read. This weight represents the closeness of the two based on historical direct interaction, and therefore this weight value is used as the first correlation degree R1.

[0077] S51. Based on all the connection paths between customer nodes and candidate salesperson nodes in the knowledge graph, aggregate the weights of each path and obtain the multi-hop path connection strength as the second correlation degree.

[0078] Specifically, starting with customer nodes in the knowledge graph and ending with candidate salesperson nodes, a graph traversal algorithm is executed. A maximum number of hops is set to balance the computational depth and breadth, enumerating all connected paths that meet the conditions. Each path consists of a series of connected entity edges and second-type implicit relationship edges. For any path, its path weight is defined as the product of the weights of all edges on that path, reflecting the attenuation effect of relationship strength during multi-hop transmission. After enumerating all possible paths and calculating the weights of each path, the system performs aggregation calculations to merge the contributions of all indirect relationships into a comprehensive index. In this embodiment, the aggregation strategy is to calculate the sum of all path weights. The calculated sum is the multi-hop path connection strength, which is used as the second correlation degree R2.

[0079] S52. Input the first correlation degree and the second correlation degree into the preset correlation degree fusion function to calculate and obtain the total correlation degree between the customer and the candidate salesperson.

[0080] Specifically, the preset correlation fusion function is a weighted summation function that merges the first correlation R1 and the second correlation R2 to calculate the total correlation R between the customer and the candidate salesperson. The preset weight coefficients can adjust the relative importance of direct and indirect network relationships in the total correlation calculation. When prioritizing the maintenance of historical strong relationships, a higher weight coefficient is assigned to the first correlation R1; when exploring potential network relationships, a higher weight coefficient is assigned to the second correlation R2.

[0081] Subsequently, the system executes the core matching decision, assigning the customer's overall value score V... C Substituting the total relevance R into the matching degree calculation formula M=α·V C The system calculates +β·R, where α and β are preset weighting coefficients used to ultimately balance the contribution of customer value and total relevance in the matching decision. Based on the calculated matching score, a customer redistribution plan is generated, enabling precise control of the matching strategy. This allows the system to adapt to different business objectives and transform multi-dimensional relationship analysis and customer value quantification results into the optimal matching plan.

[0082] In addition, refer to Figure 5 Furthermore, in one embodiment, step S4, obtaining the customer's overall value score, is further refined into the following sub-steps: S40, the first indicator for acquiring the financial value of customers, the second indicator for the potential value predicted based on policy data, and the third indicator for the relational value measured by the network centrality of customers in the knowledge graph.

[0083] Specifically, the first indicator is the financial value dimension, calculated by aggregating structured business data such as the customer's historical total premiums, average annual premium contribution, and on-time payment rate, reflecting the robustness of their historical contributions. The second indicator is the potential value dimension, which analyzes the customer's policy type, coverage amount, policy term, product portfolio, and coverage increase history, using predictive models such as gradient boosting decision trees to predict their lifetime value over a specific future period, thus measuring growth potential. The third indicator is the relationship value dimension, which uses graph computing algorithms to measure the network centrality of customer nodes in the aforementioned knowledge graph, such as calculating their degree centrality or eigenvector centrality, to quantify the customer's influence and pivotal position within their relationship network. By integrating traditional financial indicators, data-driven future value prediction, and network influence quantification based on dynamic relationship graphs into three orthogonal dimensions for assessing customer value, comprehensive structured input features are provided for subsequent integrated evaluation.

[0084] S41. Input the first indicator, the second indicator, and the third indicator into the trained customer value assessment model. The customer value assessment model uses a machine learning algorithm to non-linearly fuse the first indicator, the second indicator, and the third indicator, and outputs a comprehensive value score.

[0085] Specifically, a customer value assessment model is trained and applied to achieve nonlinear fusion of the aforementioned multi-dimensional indicators. This model is trained using a supervised machine learning algorithm, with historical customer data as training samples and labels defined as composite value scores based on subsequent actual business performance. During model training, the algorithm iteratively optimizes and automatically learns the complex interactions and nonlinear mappings between several primary, secondary, and tertiary indicators, ultimately forming a predictive function that captures the synergistic effect of the three. In the application phase, the system inputs the three indicators of the customer to be evaluated into the trained customer value assessment model, which directly outputs a comprehensive value score based on the nonlinear functional relationships learned internally.

[0086] By using a trained customer value assessment model to non-linearly fuse multi-dimensional indicators, the overall value score of a customer can be calculated more accurately, providing an accurate data foundation for subsequent customer reallocation. This helps reduce the churn rate of high-value, high-density customers and improve the productivity utilization rate of new sales staff.

[0087] In addition, refer to Figure 6 Furthermore, in one embodiment, step S4, obtaining the customer's risk level, is further refined into the following sub-steps: S42. Obtain static risk factors based on the policy's overdue status.

[0088] Specifically, static risk quantification factors that directly reflect a customer's ability to fulfill their obligations are extracted from the structured data of the core insurance business system. These static risk quantification factors include, but are not limited to, the current number of overdue days for the policy, the longest historical overdue days, the ratio of total unpaid premiums to the annual premiums due, and the rate of change in recent payment intervals. Furthermore, the system generates a set of static risk indicators that are updated over time for each customer through periodic data queries and aggregation calculations.

[0089] The above provides objective and traceable factual basis for risk assessment and establishes a static dimension in the customer risk profile that describes the objective severity of their default.

[0090] S43. Obtain dynamic risk signals based on the changing trends of customer-side sentiment indicators and salesperson-side sentiment indicators.

[0091] Specifically, the system analyzes the changing patterns of periodic customer-side and salesperson-side sentiment indicators with time-series characteristics. In this embodiment, the system calculates the slope or mean change rate of the customer-side sentiment indicator within a recent time window to quantify the deteriorating trend of their sentiment tendencies. Simultaneously, it may also analyze the accumulation of communication difficulties or negative feedback revealed in the salesperson-side sentiment indicators. These derived indicators based on sentiment trends are defined as dynamic risk signals.

[0092] In summary, by transforming emotional time-series data into quantifiable risk warning signals, it is possible to capture the downward changes in customer willingness to cooperate and communication status before the actual deterioration of financial data, thereby achieving earlier and more forward-looking perception of risks.

[0093] S44. Based on the input of static risk factors and dynamic risk signals into the trained risk assessment model, the risk assessment model uses machine learning algorithms to perform correlation analysis on the static risk factors and dynamic risk signals, calculates and obtains risk scores, and classifies customers into at least three risk levels: low, medium, and high, according to preset thresholds.

[0094] Specifically, a supervised learning-trained risk assessment model correlates and comprehensively analyzes static risk factors and dynamic risk signals. This model is trained using historical customer samples, whose features include both static factors and dynamic signals, labeled with the customers' subsequent repayment behavior. Furthermore, during training, the machine learning algorithm automatically learns the complex correlation patterns between the two types of risk information, as well as the combined impact weights of static factors and dynamic signals on the final risk.

[0095] In application, the system inputs the collected static risk factors and dynamic risk signals of the customers to be assessed into the risk assessment model. The risk assessment model calculates a comprehensive risk score using an internally learned nonlinear discriminant function. Finally, the system compares this risk score with a preset threshold range to clearly classify customers into three risk levels: low, medium, and high.

[0096] The above uses a trained risk assessment model to calculate risk scores and classify risk levels, providing a basis for triggering different collection processes based on risk levels, thereby improving collection efficiency, reducing ineffective collection time, and better achieving differentiated and precise collection under compliant control.

[0097] In addition, refer to Figure 7 Furthermore, in one embodiment, step S6 is refined into the following sub-steps: S60. If the risk level is low, an automated reminder message will be sent.

[0098] Specifically, when a customer is assessed as low-risk, a pre-set lightweight collection process is automatically triggered. In this embodiment, the system calls a standardized message template library that has undergone compliance review, and automatically generates and sends a reminder message based on the customer's commonly used contact channels, such as SMS or mobile application in-app messages. The message content typically includes a friendly payment reminder, a convenient online payment portal link, and contact information for consultation services, and the sending time can be set to non-rest hours to improve the customer experience.

[0099] The above methods enable efficient processing of a large number of low-risk overdue cases through full automation, completing basic performance reminders without human intervention. This reduces the daily administrative workload of the collection team, allowing human resources to be concentrated on more complex cases, thereby improving overall collection operational efficiency and reducing unit costs.

[0100] S61. If the risk level is medium, then push the standard collection script guidelines containing customer interest tags to the assigned on-duty salespersons.

[0101] Specifically, when a customer is assessed as medium-risk, a standardized human-machine collaborative collection mode is initiated. In this embodiment, firstly, based on the allocation results generated by the correlation between the customer and candidate current sales representatives in the knowledge graph or existing service relationships, the current sales representative responsible for the customer is determined. Subsequently, the system extracts interest tags from the customer profile, derived from analysis of historical interactions, sentiment indicators, etc., and matches them to a pre-stored, compliance-approved standard collection script library to generate a personalized communication guide. This communication guide will be proactively pushed to the corresponding sales representative's work terminal, containing suggestions for communication entry points based on customer interests and standard response statements.

[0102] In summary, for medium-risk clients, the collection script guidelines provided by the platform offer sales staff a data-driven, precise communication tool. This tool transforms unstructured client interest information into actionable script suggestions, enabling a shift from a uniform to a differentiated collection strategy. This enhances the relevance of communication and client acceptance, effectively increasing the success rate of collections for medium-risk cases.

[0103] S62. If the risk level is high, an auxiliary decision package containing historical customer sentiment indicators and compliance points will be generated, and a manual review and processing process will be triggered.

[0104] Specifically, when a customer is assessed as high-risk, the most stringent escalation processing and compliance assurance mechanisms need to be activated. In this embodiment, the system automatically integrates the customer's historical customer sentiment indicators to generate a summary of the current customer's sentiment trend. Simultaneously, it retrieves core compliance points related to the customer type and overdue scenario from the rule base. Subsequently, this information, key interaction records, and case background are packaged to generate a structured decision support package. This package, along with the collection task, is assigned to a certified senior specialist or a dedicated compliance collection team for processing and is highlighted in the corresponding salesperson's interface to aid decision-making. Simultaneously, the system records this assignment and marks the task as requiring manual review and closure.

[0105] The above-mentioned auxiliary decision-making package, which includes historical customer sentiment indicators and compliance points, is generated for high-risk customers and triggers manual review and processing. This not only helps to improve collection effectiveness but also ensures the compliance of the collection process, thereby improving overall collection efficiency, reducing compliance risks, and realizing the preservation of customers' assets.

[0106] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0107] This application also provides a policy customer allocation system based on a relationship graph, which corresponds one-to-one with the policy customer allocation method based on a relationship graph in the embodiments.

[0108] refer to Figure 8 A policyholder allocation system based on relationship graphs includes: a data acquisition module 1, an emotion quantification module 2, a graph construction module 3, a customer scoring module 4, a reallocation module 5, and a collection module 6. Detailed descriptions of each functional module are as follows: Data Acquisition Module 1: Used to collect customer data, which includes unstructured interactive text between customers and salespersons; Sentiment Quantification Module 2: This module is used to extract semantic information from interactive text and perform weighted sentiment calculation on single interactive text based on pre-labeled sentiment word weights, generating customer-side sentiment indicators and salesperson-side sentiment indicators respectively. Knowledge graph construction module 3: Used to build and dynamically update a knowledge graph containing customer nodes and salesperson nodes based on structured business data, customer sentiment indicators and salesperson sentiment indicators in customer data; Customer scoring module 4: Based on knowledge graphs and business data, it calculates and obtains the customer's comprehensive value score and risk level through customer value assessment and collection risk assessment; Reassignment Module 5: In response to the event that a salesperson's status changes to resignation, it performs customer reassignment based on the comprehensive value score of the customers under the resigned salesperson's name and the correlation between each customer and the candidate current salesperson in the corresponding knowledge graph. Collection Module 6: In response to events where the policy status changes to overdue, it triggers and executes the collection process based on the corresponding customer's risk level.

[0109] The system comprises several modules: Data Acquisition Module 1 collects customer data, including unstructured interactive text between customers and salespersons, providing a rich data foundation for subsequent analysis; Sentiment Quantification Module 2 extracts semantic information from the interactive text and performs weighted sentiment calculations to generate sentiment indicators for both the customer and salesperson sides, thus quantifying the sentiment information in unstructured interactive text; Knowledge Graph Construction Module 3 constructs and dynamically updates a knowledge graph based on structured business data and sentiment indicators, which can intuitively display the relationship between customers and salespersons; Customer Scoring Module 4 calculates the comprehensive value score and risk level of customers through customer value assessment and collection risk assessment, providing a quantitative basis for customer allocation and collection; Reallocation Module 5 performs customer reallocation based on the comprehensive value score and relevance of customers, which can avoid the loss of high-value customers; and Collection Module 6 triggers and executes the collection process based on the customer's risk level, enabling personalized collection. By combining the various modules, customer assets were preserved, the churn rate of high-value customers was reduced, and the productivity utilization rate of new sales staff was improved. At the same time, collection efficiency was improved, the collection success rate was increased, and ineffective collection hours were reduced. Furthermore, compliance risks were kept under control, and the risks of regulatory penalties and customer complaints were reduced. Knowledge assets were also accumulated, and the knowledge graph constructed can be reused in other scenarios.

[0110] Specific limitations regarding the relationship graph-based policy customer allocation system can be found in the context of the limitations on the relationship graph-based policy customer allocation method, and will not be repeated here. Each module in the aforementioned relationship graph-based policy customer allocation system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in an electronic device, or stored in the memory of the electronic device as software, so that the processor can call and execute the corresponding operations of each module. In one embodiment, an electronic device is provided, which is a user terminal. (Reference) Figure 9 The electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores detection data tables. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a relationship graph-based policyholder allocation method.

[0111] In one embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: S1. Collect customer data, which includes unstructured text interactions between customers and salespersons.

[0112] S2. Extract semantic information from the interactive text and perform weighted sentiment calculation on the single interactive text based on the pre-labeled sentiment word weights to generate customer-side sentiment indicators and salesperson-side sentiment indicators respectively.

[0113] S3. Based on structured business data, customer sentiment indicators, and salesperson sentiment indicators in customer data, construct and dynamically update a knowledge graph containing customer nodes and salesperson nodes.

[0114] S4. Based on knowledge graphs and business data, calculate and obtain the customer's comprehensive value score and risk level through customer value assessment and collection risk assessment.

[0115] S5. In response to the event that a salesperson's status changes to resignation, perform customer reassignment based on the comprehensive value score of the customers under the resigned salesperson's name and the correlation between each customer and the candidate current salesperson in the corresponding knowledge graph.

[0116] S6. In response to an event where the policy status changes to overdue, trigger and execute the collection process based on the corresponding customer's risk level.

[0117] In one embodiment, the sub-steps of step S2 refinement include: S20. Identify customer-side statements and salesperson-side statements from a single interaction text.

[0118] S21. Perform sentiment analysis on the customer-side statements and the salesperson-side statements respectively. The sentiment analysis is based on a preset sentiment word library and the corresponding sentiment word weights, and calculates the customer-side sentiment score and the salesperson-side sentiment score for a single interaction.

[0119] S22. Based on the customer-side sentiment scores and salesperson-side sentiment scores from multiple interactions within a preset time window, calculate and generate periodic customer-side sentiment indicators and salesperson-side sentiment indicators using a weighted average.

[0120] In one embodiment, the sub-steps of step S3 refinement include: S30. Treat customers and salespersons as nodes in the graph.

[0121] S31. Based on the policy ownership relationship between the customer and the agent, construct the first type of entity edge.

[0122] S32. Based on periodic customer-side sentiment indicators and salesperson-side sentiment indicators, calculate the correlation between customer nodes and salesperson nodes, and use the correlation as weight to construct or update the second type of implicit relationship edge.

[0123] S33. In response to the processing of new interactive data, the first type of entity edges and the second type of implicit relationship edges are reconstructed, triggering an update of the knowledge graph.

[0124] In one embodiment, the refined sub-steps of step S5 include: S50. Obtain the weights of the second type of implicit relationship edges between customer nodes and each candidate salesperson node in the knowledge graph, and use them as the first degree of association.

[0125] S51. Based on all the connection paths between customer nodes and candidate salesperson nodes in the knowledge graph, aggregate the weights of each path and obtain the multi-hop path connection strength as the second correlation degree.

[0126] S52. Input the first correlation degree and the second correlation degree into the preset correlation degree fusion function to calculate and obtain the total correlation degree between the customer and the candidate salesperson.

[0127] In one embodiment, step S4, the detailed sub-steps for obtaining the customer's overall value score, include: S40, the first indicator for acquiring the financial value of customers, the second indicator for the potential value predicted based on policy data, and the third indicator for the relational value measured by the network centrality of customers in the knowledge graph.

[0128] S41. Input the first indicator, the second indicator, and the third indicator into the trained customer value assessment model. The customer value assessment model uses a machine learning algorithm to non-linearly fuse the first indicator, the second indicator, and the third indicator, and outputs a comprehensive value score.

[0129] In one embodiment, step S4, the detailed sub-steps for obtaining the customer's risk level include: S42. Obtain static risk factors based on the policy's overdue status.

[0130] S43. Obtain dynamic risk signals based on the changing trends of customer-side sentiment indicators and salesperson-side sentiment indicators.

[0131] S44. Based on the input of static risk factors and dynamic risk signals into the trained risk assessment model, the risk assessment model uses machine learning algorithms to perform correlation analysis on the static risk factors and dynamic risk signals, calculates and obtains risk scores, and classifies customers into at least three risk levels: low, medium, and high, according to preset thresholds.

[0132] In one embodiment, the sub-steps of step S6 are further refined as follows: S60. If the risk level is low, an automated reminder message will be sent.

[0133] S61. If the risk level is medium, then push the standard collection script guidelines containing customer interest tags to the assigned on-duty salespersons.

[0134] S62. If the risk level is high, an auxiliary decision package containing historical customer sentiment indicators and compliance points will be generated, and a manual review and processing process will be triggered.

[0135] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0136] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

Claims

1. A method for allocating policyholders based on relationship graphs, characterized in that, include: Collect customer data, which includes unstructured text interactions between customers and salespersons; Semantic information is extracted from the interactive text, and a weighted sentiment calculation is performed on a single interactive text based on the pre-labeled sentiment word weights to generate customer-side sentiment indicators and salesperson-side sentiment indicators, respectively. Based on the structured business data in the customer data, the customer-side sentiment indicators and the salesperson-side sentiment indicators, a knowledge graph containing customer nodes and salesperson nodes is constructed and dynamically updated. Based on the knowledge graph and business data, the comprehensive value score and risk level of the customer are calculated and obtained through customer value assessment and collection risk assessment. In response to the event that a salesperson's status changes to resignation, customer reallocation is performed based on the comprehensive value score of customers under the name of the salesperson whose status has changed to resignation, and the correlation between each customer and the candidate current salesperson in the corresponding knowledge graph. The steps of performing the customer reallocation include: obtaining the weight of the second type of implicit relationship edge between the customer node and each candidate salesperson node in the knowledge graph, as the first correlation. Based on all the connection paths between the customer node and the candidate salesperson node in the knowledge graph, the weights of each path are aggregated and calculated to obtain the multi-hop path connection strength as the second correlation degree. The first correlation degree and the second correlation degree are input into a preset correlation degree fusion function to calculate and obtain the total correlation degree between the customer and the candidate salesperson; The customer's overall value score and the total relevance are substituted into the matching degree calculation formula M=α·VC+β·R for calculation, where α and β are preset weight coefficients, VC is the overall value score, and R is the total relevance. A customer reassignment plan is generated based on the calculated matching degree score. In response to an event that changes the policy status to overdue, a collection process is triggered and executed based on the risk level stated by the corresponding customer.

2. The method according to claim 1, characterized in that, The steps of extracting semantic information from the interactive text and performing weighted sentiment calculation on a single interactive text based on pre-labeled sentiment word weights to generate customer-side sentiment indicators and salesperson-side sentiment indicators respectively include: Identify customer-side and salesperson-side statements from a single interaction text; Sentiment analysis is performed on both customer-side and salesperson-side statements. The sentiment analysis is based on a preset sentiment word library and corresponding sentiment word weights to calculate and obtain the customer-side sentiment score and salesperson-side sentiment score for a single interaction. Based on the customer-side sentiment scores and salesperson-side sentiment scores from multiple interactions within a preset time window, a weighted calculation is performed to generate periodic customer-side sentiment indicators and salesperson-side sentiment indicators.

3. The method according to claim 2, characterized in that, The step of constructing and dynamically updating a knowledge graph containing customer nodes and salesperson nodes based on the structured business data in the customer data, the customer-side sentiment indicators, and the salesperson-side sentiment indicators includes: Customers and salespersons are used as nodes in the graph; Based on the policy ownership relationship between customers and agents, construct the first type of entity edge; Based on the periodic customer-side sentiment indicators and salesperson-side sentiment indicators, the correlation degree between customer nodes and salesperson nodes is calculated and obtained, and the second type of implicit relationship edge is constructed or updated using the correlation degree as the weight. In response to the processing of new interactive data, the first type of entity edges and the second type of implicit relationship edges are reconstructed, triggering an update of the knowledge graph.

4. The method according to claim 1, characterized in that, The step of calculating and obtaining the customer's comprehensive value score and risk level based on the knowledge graph and business data through customer value assessment and collection risk assessment includes the following steps for obtaining the customer's comprehensive value score: The first indicator is the financial value dimension of the customer; the second indicator is the potential value dimension predicted based on policy data; and the third indicator is the relational value dimension measured by the customer's network centrality in the knowledge graph. The first, second, and third indicators are input into a trained customer value assessment model, which uses a machine learning algorithm to non-linearly fuse the multi-dimensional indicators and outputs the comprehensive value score.

5. The method according to claim 1, characterized in that, The step of calculating and obtaining a customer's comprehensive value score and risk level based on the knowledge graph and business data through customer value assessment and collection risk assessment includes the following steps for obtaining the customer's risk level: Obtain static risk factors based on the policy's overdue status; Obtain dynamic risk signals based on the changing trends of the customer-side sentiment indicators and the salesperson-side sentiment indicators; The static risk factors and the dynamic risk signals are input into a trained risk assessment model. The risk assessment model uses a machine learning algorithm to perform correlation analysis on the static risk factors and the dynamic risk signals, calculates and obtains a risk score, and classifies customers into at least three risk levels: low, medium, and high, based on preset thresholds.

6. The method according to claim 5, characterized in that, The steps for triggering and executing the collection process in response to an event that changes the policy status to overdue, based on the corresponding customer's risk level, include: If the risk level is low, an automated alert message will be sent. If the risk level is medium, then push the standard collection script guidelines containing customer interest tags to the assigned salesperson. If the risk level is high, an auxiliary decision package containing historical customer sentiment indicators and compliance points will be generated, and a manual review and processing process will be triggered.

7. A policyholder allocation system based on a relationship graph, characterized in that, include: Data acquisition module (1): used to collect customer data, which includes unstructured interactive text between customers and salespersons; Emotion quantification module (2): used to extract semantic information from the interactive text and perform weighted emotion calculation on the single interactive text based on the pre-labeled emotion word weights, and generate customer-side emotion indicators and salesperson-side emotion indicators respectively. Graph construction module (3): Used to construct and dynamically update a knowledge graph containing customer nodes and salesperson nodes based on the structured business data in the customer data, the customer-side sentiment indicators and the salesperson-side sentiment indicators; Customer scoring module (4): Based on the knowledge graph and business data, it calculates and obtains the customer's comprehensive value score and risk level through customer value assessment and collection risk assessment; The reallocation module (5) is used to respond to the event of a salesperson's status changing to resignation. Based on the comprehensive value score of the customers under the salesperson whose status has changed to resignation, and the correlation between each customer and the candidate in-service salesperson in the corresponding knowledge graph, the module performs customer reallocation. The customer reallocation includes: obtaining the weight of the second type of implicit relationship edge between the customer node and each candidate salesperson node in the knowledge graph as the first correlation; based on all the connected paths between the customer node and the candidate salesperson node in the knowledge graph, the weight of each path is aggregated and calculated to obtain the multi-hop path connection strength as the second correlation; the first correlation and the second correlation are input into a preset correlation fusion function to calculate the total correlation between the customer and the candidate salesperson; the comprehensive value score of the customer and the total correlation are substituted into the matching degree calculation formula M=α·VC+β·R for calculation, where α and β are preset weight coefficients, VC is the comprehensive value score, and R is the total correlation; and a customer reallocation scheme is generated based on the calculated matching degree score. Collection module (6): In response to an event where the policy status changes to overdue, the collection process is triggered and executed according to the risk level of the corresponding customer.

8. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as any one of the relationship graph-based policy customer allocation methods as claimed in claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The system stores a computer program capable of being loaded by a processor and executing any of the relationship graph-based policyholder allocation methods as described in claims 1 to 6.

Images