Intelligent customer relationship management system based on multi-modal data fusion

Abstract

The invention relates to the technical field of customer management, in particular to an intelligent customer relationship management system based on multi-modal data fusion, and the system comprises a multi-modal data acquisition module which is used for extracting a semantic contradiction index, a service request frequency change rate and a historical performance quality score of a text work order of a voice conversation from a customer interaction log; the state vector generation module is used for generating customer relationship state vectors based on health degree scores; the customer relationship network topology construction module is used for constructing a causal association edge of'customer-service defect-relationship attenuation 'based on the customer relationship state vector, outputting a customer relationship network topology structure with an interpretable path, and forming a relationship maintenance knowledge graph; and a self-adaptive relation repair strategy generation module. According to the invention, through identification of the high-weight relation attenuation path, key nodes and paths of customer relation degradation can be positioned.

Description

Technical Field

[0001] The present invention relates to the technical field of customer management, and in particular to an intelligent customer relationship management system based on multi-modal data fusion. Background Art

[0002] With the rapid development of information technology and the continuous change of customer needs, enterprises are facing unprecedented challenges in improving customer satisfaction and loyalty. Traditional customer relationship management (CRM) methods mainly rely on customer interaction data from a single channel, usually analyzing based on historical transaction records or customer feedback. However, this method has the following deficiencies:

[0003] Only relying on a single interaction data source (such as phone calls, emails, online chats, etc.), ignoring the integration and in-depth mining of cross-channel and multi-modal data. This single data source method cannot comprehensively understand customers' needs, behaviors, and emotional fluctuations, and it is also difficult to accurately evaluate the overall picture of customer relationships.

[0004] Most existing customer relationship management systems rely on surface data such as customer satisfaction surveys and historical purchase records to evaluate the health of customer relationships, lacking an effective early warning mechanism. When problems occur in customer relationships, they are often discovered only after a long period of accumulation, resulting in enterprises missing the best opportunity to take timely repair measures. Summary of the Invention

[0005] The present invention provides an intelligent customer relationship management system based on multi-modal data fusion.

[0006] An intelligent customer relationship management system based on multi-modal data fusion includes:

[0007] A multi-modal data acquisition module: extracting the semantic contradiction index, service request frequency change rate, and historical fulfillment quality score of the text work order of the voice conversation from the customer interaction log, where:

[0008] Extraction of semantic contradiction index: performing two-channel semantic encoding on the voice conversation text and the corresponding work order text, respectively extracting the emotional polarity and responsibility attribution labels, and by comparing the emotional polarity and responsibility attribution labels, determining whether it is a contradictory event. When the determination result is a contradictory event, it is marked as a contradictory event, and the semantic contradiction index is generated according to the frequency of contradictory events;

[0009] Extraction of service request frequency change rate: counting the total amount of customer cross-channel service requests according to a preset period (7 days / 30 days), and calculating the service request frequency change rate ΔF of the request volume in adjacent periods using the time series difference algorithm;

[0010] Extraction of historical fulfillment quality score: extracting the fulfillment index set from the service work order library and generating the comprehensive fulfillment quality score Q by weighted summation;

[0011] Status vector generation module: Based on the semantic contradiction index, the service request frequency change rate, and the historical performance quality score, generate a customer relationship status vector V based on the health score 客户 ;

[0012] Customer relationship network topology construction module: Based on the customer relationship status vector V 客户 , create a triple structure of customer nodes, service resource nodes, and relationship degradation incentive nodes. Based on the triple structure, construct causal association edges of "customer - service defect - relationship attenuation", output a customer relationship network topology structure with an interpretable path, and form a relationship maintenance knowledge graph;

[0013] Adaptive relationship repair strategy generation module: Execute the following according to the relationship maintenance knowledge graph:

[0014] Identify high - weight relationship attenuation paths;

[0015] Match countermeasures in the preset strategy library.

[0016] Optionally, in the extraction of the semantic contradiction index, based on natural language processing technology, perform dual - channel semantic encoding on the voice dialogue text and the corresponding work order text, and extract the sentiment polarity and responsibility attribution labels respectively. The sentiment polarity vector includes positive sentiment and negative sentiment. The responsibility attribution label is used to identify the designated responsible party in the work order, including the customer, the customer service, or a third party;

[0017] For each service event, directly compare the sentiment polarity in the voice text with the responsibility attribution label in the work order text. When the sentiment polarity is positive sentiment (such as the customer service's voice commitment to the solution) and the work order responsibility attribution is the customer or a third party, mark it as a contradiction event. According to the occurrence frequency of the contradiction events, statistically generate a semantic contradiction index, which is used to measure the potential relationship contradictions existing in the customer service process and further affect the customer relationship health assessment.

[0018] Optionally, the service request frequency change rate ΔF is calculated as:

[0019] where ΔF is the service request frequency change rate, representing the relative change percentage of the service request volume in two adjacent periods. F t is the total service request volume in the current period t, that is, the total number of customer cross - channel service requests within the time window. F t-1 is the total service request volume in the previous period t - 1, that is, the total number of customer cross - channel service requests in the previous time window.

[0020] Optionally, the performance index set includes the problem - solving rate and the customer satisfaction score. The comprehensive performance quality score is expressed as:

[0021] Q = w1·S 响应 + w2·R 解决 + w3·C 满意 , where:

[0022] Q is the comprehensive performance quality score, representing the comprehensive quality score of customer service performance. S 响应 is the response time score, R 解决 is the problem-solving rate score, C 满意 is the customer satisfaction score, and w1, w2, w3 are the weight coefficients of each index.

[0023] Optionally, the state vector generation module specifically includes:

[0024] Semantic contradiction index fusion unit: Based on the semantic contradiction index, quantify the inconsistency between emotional conflict and responsibility attribution in customer interaction to obtain the emotional health score H of the customer relationship 情感 , and the higher the semantic contradiction index, the lower the emotional health score of the customer.

[0025] Service request frequency change rate fusion unit: According to the value of the service request frequency change rate ΔF, evaluate the volatility of customer service needs. The larger the change rate ΔF, the greater the fluctuation of customer needs, which may be an indication of the instability of the customer relationship. After normalizing the service request frequency change rate to the [0,1] interval, obtain the customer demand fluctuation score H 需求 .

[0026] Performance quality score fusion unit: Based on the historical performance quality score Q, evaluate the performance health of the customer relationship in combination with the customer's past service performance. The performance quality score Q directly reflects the customer's satisfaction with the service and the effectiveness of problem-solving. The performance health score H 履约 directly uses the normalized historical performance quality score Q, and the normalized performance health score is;

[0027] The final customer relationship state vector V 客户 = [H 情感 , H 需求 , H 履约 , where each dimension represents the health status of the customer relationship.

[0028] Optionally, the creation of the customer node includes creating a customer node based on each health score extracted from the customer relationship state vector V 客户 , including the following information:

[0029] The basic identity identifier of the customer;

[0030] The emotional health score H of the customer 情感;

[0031] Customer demand volatility score H 需求 ;

[0032] Customer fulfillment health score H 履约 ;

[0033] The creation of the service resource node includes creating a service resource node according to various service resources in the service system, including customer service personnel, service platforms, and problem-solving solutions. The service resource node represents the interaction between the customer and the service resource, including the service history of the service resource;

[0034] The creation of the relationship degradation incentive node includes creating a relationship degradation incentive node by analyzing customer historical interaction data, identifying potential factors leading to the decline of the customer relationship, including blame-shifting (corresponding to emotional health), service defects (corresponding to the change rate of service request frequency), and fulfillment problems (corresponding to fulfillment quality);

[0035] Triple structure creation: Based on the creation of customer nodes, service resource nodes, and relationship degradation incentive nodes, a triple structure is constructed to represent the relationship between the customer, service resources, and service defects. The triple form is: [customer node, service resource node, relationship degradation incentive node].

[0036] Optionally, the causal association edge is constructed according to the customer state vector V 客户 to form a causal link of customer-service-relationship attenuation. The causal association edge also includes an edge weight w 边 , and the edge weight w 边 is used to quantify the severity of customer relationship degradation and the impact of different degradation factors on the customer relationship;

[0037] The customer node, service resource node, relationship degradation incentive node, and causal association edge form a complete relationship maintenance knowledge graph, outputting a customer relationship network topology structure with an interpretable path.

[0038] Optionally, the edge weight is calculated as:

[0039] w 边 = α·|ΔF| + β·(1 - H 情感 ) + γ·(1 - Q 履约 ), where ΔF is the change rate of service request frequency. The larger it is, the more unstable the customer's demand is, which in turn triggers the degradation of the customer relationship. Take its absolute value |ΔF|. H 情感 is the customer emotional health score. The lower the emotional health, the stronger the customer's dissatisfaction, resulting in the easier degradation of the customer relationship. To reflect the severity of degradation, 1 - H 情感 is used. Q 履约is the historical performance quality score. The lower it is, the worse the past service performance quality is, the worse the customer experience is, resulting in customer churn. Use 1 - Q 履约 The historical performance quality score, where α, β, and γ are weight coefficients that respectively control the influence of the service request frequency change rate, emotional health score, and performance quality score on the edge weights.

[0040] Optionally, the adaptive relationship repair strategy generation module specifically includes:

[0041] Identify high - weight relationship decay paths: In the relationship maintenance knowledge graph, high - weight relationship decay paths represent paths with a high probability of customer relationship degradation, which are composed of multiple causally related edges. Starting from the customer node, use the Dijkstra algorithm to traverse all causal relationship paths in the graph (i.e., the causal path of customer - service defect - relationship decay), and calculate the total weight of each path. The total weight W 路径 is the weighted sum of the weights of each edge on the path: where w 边,i is the weight of the i - th edge in the path, n is the number of edges in the path, and select the paths with a total weight greater than the weight threshold as high - weight relationship decay paths;

[0042] Match countermeasures in the preset policy library: According to the identified high - weight relationship decay paths, use the countermeasures in the preset policy library to repair the customer relationship. The countermeasures include predefined strategies for dealing with customer relationship decline.

[0043] Optionally, the preset policy library specifically includes:

[0044] Automatically trigger compensation agreements: If there is a problem of shifting responsibility on the path, automatically trigger compensation measures (such as giving points, coupons, etc.);

[0045] Assign an exclusive customer manager: If there are performance problems on the path and there is a long - term negative interaction history, assign an exclusive customer manager to the customer to ensure more accurate and personalized communication with the customer;

[0046] Prioritize service issues: If there are service defect problems on the path, immediately prioritize handling their service requests and repair the problems.

[0047] Advantages of the present invention:

[0048] In the present invention, through multimodal data fusion, combining voice dialogue text, service request frequency change rate, and fulfillment quality score, a dynamic evaluation system for customer relationship health is established. Through comprehensive analysis of these data, potential factors for customer relationship degradation can be accurately identified. By calculating the semantic contradiction index, conflict events of emotion and responsibility attribution are discovered and marked, which provides key clues and early warning information for subsequent relationship repair. This technology makes customer relationship management more intelligent, can reflect the microscopic changes of customer relationships in real time, and timely identify key factors affecting customer satisfaction.

[0049] In the present invention, a relationship maintenance knowledge graph is constructed. By identifying high-weight relationship attenuation paths, the key nodes and paths of customer relationship degradation can be located. Using the dynamic calculation of the weights of nodes and edges in the causal link, the actual impact of different relationship degradation incentives on customer relationships can be reflected. Through in-depth analysis of these causal association paths, the system can accurately match corresponding countermeasures, such as automatically triggering a service compensation agreement or assigning an exclusive customer manager, so as to efficiently formulate relationship maintenance strategies and prioritize the implementation of the most urgent repair measures, greatly improving the pertinence and response speed of customer relationship repair. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only those of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0051] Figure 1 It is a schematic diagram of the function modules of the management system according to an embodiment of the present invention;

[0052] Figure 2 It is a schematic diagram of the state vector generation module according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] The present invention will be described in detail below in conjunction with the drawings and specific embodiments. At the same time, it should be noted here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments. For some well-known technologies, those skilled in the art can also adopt other alternative methods for implementation; moreover, the drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0054] It should be noted that in the specification, when referring to "an embodiment", "embodiment", "exemplary embodiment", "some embodiments", etc., it indicates that the described embodiments may include specific features, structures or characteristics, but not necessarily every embodiment includes such specific features, structures or characteristics. Additionally, when combining an embodiment to describe a specific feature, structure or characteristic, implementing such feature, structure or characteristic in combination with other embodiments (whether explicitly described or not) should be within the knowledge scope of those skilled in the relevant art.

[0055] Generally, terms can be understood at least in part from their use in the context. For example, at least in part depending on the context, the term "one or more" used herein can be used to describe any feature, structure or characteristic in a singular sense, or can be used to describe a combination of features, structures or characteristics in a plural sense. Additionally, the term "based on" can be understood as not necessarily aiming to convey a set of exclusive factors, but instead, at least in part depending on the context, allowing for the existence of other factors that may not be explicitly described.

[0056] As Figure 1 - Figure 2 shown, an intelligent customer relationship management system based on multimodal data fusion includes:

[0057] Multimodal data acquisition module: Extract the semantic contradiction index, service request frequency change rate and historical fulfillment quality score of the text work order of the voice conversation from the customer interaction log, where:

[0058] Semantic contradiction index extraction: Perform dual-channel semantic encoding on the voice conversation text and the corresponding work order text, respectively extract the emotional polarity and responsibility attribution labels, and by comparing the emotional polarity and responsibility attribution labels, determine whether it is a contradictory event. When the determination result is a contradictory event, mark it as a contradictory event, and generate a semantic contradiction index according to the frequency of contradictory events;

[0059] Service request frequency change rate extraction: Statistically calculate the total cross-channel service request volume of customers according to a preset period (7 days / 30 days), use the time series difference algorithm to calculate the service request frequency change rate ΔF of the request volume in adjacent periods, and attribute the source of abnormal fluctuations (the absolute value of the change rate ≥ 50%) to the channel, and mark the main fluctuation channel;

[0060] Historical fulfillment quality score extraction: Extract the fulfillment index set from the service work order library, and generate a comprehensive fulfillment quality score Q by weighted summation;

[0061] State vector generation module: Generate a customer relationship status vector V based on the health score based on the semantic contradiction index, service request frequency change rate and historical fulfillment quality score 客户 ;

[0062] Customer relationship network topology construction module: Based on the customer relationship status vector V 客户, create a triple structure of customer nodes, service resource nodes, and relationship degradation inducement nodes, and based on the triple structure, construct the causal correlation edge of "customer-service defect-relationship decay", output the customer relationship network topology with an explainable path, and form a relationship maintenance knowledge graph;

[0063] Adaptive relationship repair strategy generation module: performs the following based on the relationship maintenance knowledge graph:

[0064] Identify high-weight relationship decay paths (e.g., a customer service agent’s response delay causes a surge in relationship entropy for a specific customer group);

[0065] Match the countermeasures in the preset policy library (such as automatically triggering a service compensation agreement or assigning a dedicated account manager).

[0066] In the extraction of semantic contradiction index, dual-channel semantic encoding is performed on the voice dialogue text and the corresponding work order text based on natural language processing technology to extract sentiment polarity and responsibility attribution labels respectively. The sentiment polarity vector includes positive sentiment and negative sentiment. The responsibility attribution label is used to identify the responsible party specified in the work order, including the customer, customer service or third party.

[0067] For each service event, we directly compare the sentiment polarity in the voice text with the responsibility attribution label in the work order text. When the sentiment polarity is positive (such as the customer service voice promises a solution) and the work order responsibility is attributed to the customer or a third party, it is marked as a contradictory event. Based on the frequency of occurrence of contradictory events, we statistically generate a semantic contradiction index to measure the potential relationship contradictions in the customer service process, thereby affecting the assessment of customer relationship health.

[0068] Natural language processing technology includes sentiment analysis models (BERT) to extract sentiment polarity from voice conversation text. The sentiment polarity vector represents the direction and intensity of the sentiment in the text, which is positive (such as solution commitment) or negative (such as service shirking): S = Sentiment (T voice ), where S is the emotion polarity vector, indicating the emotion polarity of the speech dialogue, and T voice is the speech dialogue text, and Sentiment(·) represents the calculation process of the sentiment analysis model;

[0069] The purpose of the sentiment analysis model is to extract the sentiment polarity and intensity of the speech dialogue text. The sentiment analysis model processes the text input and outputs the sentiment polarity vector. The calculation process is as follows:

[0070] Text preprocessing:

[0071] Preprocess the speech dialogue text, including:

[0072] Word segmentation, stop word removal, stemming or lemmatization. For example, for the text "The customer service has promised to solve the problem", after processing, we get: {customer service, promise, solve, problem};

[0073] Sentiment classification: Use a pre-trained sentiment analysis model to classify the sentiment of the text. The sentiment analysis model maps the input text to a sentiment polarity and outputs a sentiment score or sentiment category.

[0074] The sentiment analysis model Sentiment(·) outputs a sentiment polarity score s and the corresponding label:

[0075] If s > 0, the label is "positive";

[0076] If s ≤ 0, the label is "negative".

[0077] Responsibility attribution label: Extract the responsibility attribution label from the work order text to identify the responsible party (such as customer service, third party or customer) in the service event. The responsibility attribution label is a discrete value:

[0078] R = Responsibility(T ticket ), where R is the responsibility attribution label, representing the responsibility attribution in the work order text, T ticket is the work order text, and Responsibility(·) represents the calculation process of the responsibility attribution extraction model. The purpose of the responsibility attribution extraction model is to extract the responsible party label (such as customer service, third party or customer) from the work order text, relying on named entity recognition technology. The calculation process is as follows:

[0079] Text preprocessing: Similarly, first preprocess the work order text. The steps are similar to the preprocessing in sentiment analysis. For example, for the text "The third-party vendor did not deliver on time", after processing, we get: {third party, vendor, did not deliver}

[0080] Named entity recognition: Use a named entity recognition model to identify the relevant responsible entities in the text. In this example, the named entity recognition model identifies "third-party vendor" as a responsible party.

[0081] Responsibility attribution classification: Classify the extracted entities to determine whether they are "customer service", "third party" or "customer". It is a simple classification task that can be based on keyword matching.

[0082] The semantic contradiction index is normalized according to the frequency of contradictory events to obtain an index value reflecting the degree of relationship contradiction.

[0083] The change rate of service request frequency ΔF is calculated as:

[0084] Among them, ΔF is the service request frequency change rate, which represents the relative change percentage of the service request volume in two adjacent cycles. F t is the total service request volume in the current cycle t, that is, the total number of cross-channel service requests of customers within the time window (for example: the total number of requests in the past 7 days or 30 days), F t-1 is the total service request volume in the previous cycle t - 1, that is, the total number of cross-channel service requests of customers in the previous time window.

[0085] The performance indicator set includes the problem resolution rate and the customer satisfaction score. The comprehensive performance quality score is expressed as:

[0086] Q = w1·S 响应 + w2·R 解决 + w3·C 满意 , where:

[0087] Q is the comprehensive performance quality score, which represents the comprehensive quality score of customer service performance and is used to evaluate the overall service quality obtained by customers during the service process. S 响应 is the response time score, which measures the response speed of the customer service after the customer's request. The faster the response speed, the higher the score. Services with shorter response times can better meet customer expectations and also have higher scores. R 解决 is the problem resolution rate score, which measures whether the customer's problem is effectively resolved. The higher the problem resolution rate, the better the service quality and the higher the score. C 满意 is the customer satisfaction score, which measures the overall satisfaction of customers with the service. The customer's feedback (such as satisfaction surveys, scores, etc.) determines this score. w1, w2, and w3 are the weight coefficients of each indicator, which are used to adjust the relative importance of different indicators in the comprehensive score. The sum of the weight coefficients is 1. According to business requirements, different weights can be assigned to different indicators. In the present invention, w1, w2, and w3 can take values of 0.4, 0.3, and 0.3.

[0088] The state vector generation module specifically includes:

[0089] Semantic contradiction index fusion unit: Based on the semantic contradiction index, quantify the inconsistency of emotional conflicts and responsibility attribution in customer interactions to obtain the emotional health degree score H of the customer relationship 情感 , the higher the semantic contradiction index, the lower the emotional health degree score of the customer. The emotional health degree score H 情感 can be calculated by the following formula:

[0090] Among them, E contradiction is the occurrence frequency of contradiction events, E total is the total number of all customer interaction events, and the emotional health degree score H 情感 The closer it is to 1, the healthier the emotional relationship is, and the closer it is to 0, the more unstable the emotional relationship is.

[0091] Service request frequency change rate fusion unit: According to the value of the service request frequency change rate ΔF, evaluate the volatility of customer service demand. The larger the change rate ΔF, the greater the fluctuation of customer demand, which may be an indication of the instability of the customer relationship. After normalizing the service request frequency change rate to the interval [0, 1], the customer demand volatility score H is obtained. 需求 , and the specific calculation method is as follows: Among them, ΔF max is the preset maximum change rate threshold, which can be set to 50%, indicating extreme service request fluctuations.

[0092] Performance quality score fusion unit: Based on the historical performance quality score Q, combined with the customer's past service performance, evaluate the performance health of the customer relationship. The performance quality score Q directly reflects the customer's satisfaction with the service and the effectiveness of problem-solving. The performance health score H 履约 directly uses the standardized historical performance quality score Q. The standardized performance health score is: Among them, Q min and Q max are respectively the minimum and maximum values of the historical performance scores;

[0093] The final customer relationship status vector V 客户 = [H 情感 , H 需求 , H 履约 , where each dimension represents the health status of the customer relationship.

[0094] The creation of customer nodes includes creating customer nodes based on each health score extracted from the customer relationship status vector V 客户 , including the following information:

[0095] The basic identity identifier of the customer;

[0096] The emotional health score H of the customer 情感 ;

[0097] The demand volatility score H of the customer 需求 ;

[0098] The performance health score H of the customer 履约 ;

[0099] These scores reflect the health status of each dimension (emotion, demand, performance) of the customer relationship. Each customer node will carry the information of these dimensions for subsequent relationship analysis;

[0100] The creation of service resource nodes involves creating service resource nodes based on various service resources in the service system, including customer service staff, service platforms, and problem-solving solutions. Service resource nodes represent the interaction between customers and service resources, including the service history of service resources.

[0101] The creation of relationship degradation incentive nodes involves creating relationship degradation incentive nodes by analyzing customer historical interaction data to identify potential factors leading to the decline of customer relationships, including blame shifting (corresponding to emotional health), service defects (corresponding to the change rate of service request frequency), and performance issues (corresponding to performance quality). These incentive nodes are factors directly related to the decline of customer relationships.

[0102] Creation of the triple structure: Based on the creation of customer nodes, service resource nodes, and relationship degradation incentive nodes, a triple structure is constructed to represent the relationship between customers, service resources, and service defects. The triple form is: [customer node, service resource node, relationship degradation incentive node]. For example:

[0103] (Customer A, Customer service staff B, Response delay), this triple indicates that the service relationship of Customer A is affected by the response delay of Customer service staff B.

[0104] The creation of relationship degradation incentive nodes is as follows:

[0105] 1. Collect historical interaction data:

[0106] From the customer's historical interaction data, extract the following key data points:

[0107] Customer service request records: including customer complaints, consultations, and problem-solving records.

[0108] 2. Identify potential relationship degradation incentives:

[0109] Based on the collected historical interaction data, identify potential relationship degradation incentive nodes:

[0110] a. Service defects: By analyzing the customer's feedback records (complaints, refund requests, service interruptions, etc.), identify frequently occurring types of service defects, such as:

[0111] Function failure: System crash, service interruption.

[0112] Operation error: Service inadequacy caused by customer service staff or system errors.

[0113] b. Performance issues: Extract the time difference between customer requests and responses to identify customer service requests with long response times, which may be an important incentive for customer dissatisfaction. Based on sentiment analysis of chat records or voice conversations, identify inappropriate communication issues.

[0114] c. Shifting the blame: There is unclear responsibility in communication.

[0115] 3. Create nodes for relationship degradation incentives

[0116] Convert the above potential factors into nodes for relationship degradation incentives. The creation of each incentive node is based on the following information:

[0117] Node name: Identify the specific content of the incentive (service defect, response delay, poor communication).

[0118] Node attributes: The specific attributes of the node, such as defect type, delay time, sentiment analysis result.

[0119] 4. Identify the relationships between incentive nodes: After creating the incentive nodes, we need to consider the associations between these nodes and how they affect the customer relationship. Consider the following types of associations:

[0120] Causal relationship: Some incentives directly lead to the decline of the customer relationship.

[0121] Joint influence: Multiple incentives may act together to increase the risk of customer relationship degradation.

[0122] Temporal relationship: The order of occurrence of some incentives may affect the degree of customer relationship degradation. For example, a long service response delay may be the direct cause of customer complaints.

[0123] 5. Finally, through the above steps, nodes for relationship degradation incentives can be generated. These nodes can form the node part in the knowledge graph and, when constructing the relationship network topology, are connected by edges to other customer nodes and service resource nodes to form a complete relationship maintenance graph.

[0124] The causal association edges are constructed according to the customer state vector V 客户 to form a causal link of customer - service - relationship attenuation. The health status V of the customer relationship 客户 will affect the relationship strength between the customer and service defects. The causal association edges also include edge weights w 边 The edge weight w 边 is used to quantify the severity of customer relationship degradation and the impact of different degradation factors on the customer relationship;

[0125] Customer nodes, service resource nodes, nodes for relationship degradation incentives, and causal association edges form a complete relationship maintenance knowledge graph, outputting a customer relationship network topology with interpretable paths. The paths can help identify the key factors leading to customer relationship decline and provide guidance for subsequent relationship repair. For example, the path may show that the relationship decline of customer A is due to a chain reaction of factors such as response delay and service defects.

[0126] The triple structure (customer node, service resource node, relationship degradation incentive node) defines the basic relationship between customers, service resources, and service defects. Causal association edges, based on these triples, further represent the causal chain of customer relationship decline. For example, a triple represents the service defect caused by the delayed response between customer A and customer service staff B, while the causal association edge represents the specific impact and intensity of this service defect on the decline of the customer relationship. Therefore, the triple structure is the basis for causal association edges, and the latter quantifies the impact of service defects on customer relationships.

[0127] The edge weight is calculated as:

[0128] w 边 = α·|ΔF| + β·(1 - H 情感 ) + γ·(1 - Q 履约 ), where ΔF is the change rate of service request frequency. The larger it is, the more unstable the customer's demand is, which in turn triggers the degradation of the customer relationship. Take its absolute value |ΔF|. H 情感 is the customer emotional health score. The lower the emotional health score, the stronger the customer's dissatisfaction, and the easier it is for the customer relationship to degrade. To reflect the severity of the degradation, 1 - H 情感 is used. Q 履约 is the historical performance quality score. The lower it is, the worse the past service performance quality and the customer experience, resulting in customer churn. Use 1 - Q 履约 for the historical performance quality score. α, β, and γ are weight coefficients that respectively control the influence of the change rate of service request frequency, emotional health score, and performance quality score on the edge weight and can be adjusted according to the actual business situation.

[0129] The adaptive relationship repair strategy generation module specifically includes:

[0130] Identify high-weight relationship decay paths: In the relationship maintenance knowledge graph, high-weight relationship decay paths represent paths with a high probability of customer relationship degradation, which are composed of multiple causal association edges. Starting from the customer node, use the Dijkstra algorithm to traverse all causal relationship paths in the graph (i.e., the causal path of customer - service defect - relationship decay), calculate the total weight of each path. The total weight W 路径 is the weighted sum of the weights of each edge on the path: where w 边,i is the weight of the i-th edge in the path, and n is the number of edges in the path. Select the paths with a total weight greater than the weight threshold as high-weight relationship decay paths. The weight threshold is based on the percentile of historical data. Select the paths with a total weight above the 90% percentile as high-weight relationship decay paths;

[0131] Match countermeasures in the preset policy library: Based on the identified high-weight relationship decay path, countermeasures in the preset policy library will be adopted to repair the customer relationship. The countermeasures include predefined strategies for dealing with the decline of customer relationships.

[0132] The preset policy library specifically includes:

[0133] Automatically trigger compensation protocol: If there is a problem of shifting responsibility on the path, automatically trigger compensation measures (such as giving points, coupons, etc.);

[0134] Allocate an exclusive customer manager: If there are performance problems on the path and there is a long-term negative interaction history, allocate an exclusive customer manager to the customer to ensure more accurate and personalized communication with the customer;

[0135] Prioritize service issues: If there are service defect problems on the path, immediately prioritize and process their service requests to fix the problems.

[0136] Total weight W 路径 Refers to the sum of the weights of all edges of a complete path formed by starting from the customer node and passing through one or more relationship degradation incentive nodes and service resource nodes. The total weight of the path reflects the comprehensive impact of all factors on the degradation of the customer relationship on this path.

[0137] Relationship degradation incentive nodes are key nodes in the path, representing potential factors leading to the degradation of the customer relationship. Each node is connected to one or more causal association edges, indicating that a certain factor affects the health of the customer relationship. The identification methods of these nodes include:

[0138] Identify nodes from the causal link: When constructing the knowledge graph for relationship maintenance, the nodes correspond to specific events or service problems.

[0139] Determine the influence of nodes according to the causal link of the path: By analyzing the path from the customer node to the relationship decay node, determine which nodes have a greater impact on the customer relationship. For example, if a certain node (service defect node) in the path has a strong causal relationship with the subsequent node (relationship decay node), then this node can be regarded as the main relationship degradation incentive node.

[0140] Generally speaking, it is to first find the high-weight relationship decay path based on the total weight W 路径 Then, based on the high-weight relationship decay path, find the edge that causes the high-weight relationship decay path. Based on the edge weight of this edge, find the corresponding customer relationship state vector H 情感 ,H 需求 ,H 履约 One of them, and then find the relationship degradation incentive nodes based on the corresponding customer relationship state vector, so as to perform the matching strategy of the preset policy library.

[0141] The present invention encompasses any alternatives, modifications, equivalent methods, and solutions made to the essence and scope of the present invention. For the public to have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, and those skilled in the art can fully understand the present invention without the description of these details. In addition, well-known methods, processes, procedures, components, and circuits are not described in detail to avoid unnecessary confusion to the essence of the present invention.

[0142] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An intelligent customer relationship management system based on multimodal data fusion, characterized in that: include: Multimodal data acquisition module: extracts the semantic contradiction index, service request frequency change rate and historical performance quality score of the text work order of the voice dialogue from the customer interaction log, including: Semantic contradiction index extraction: Perform dual-channel semantic encoding on the voice dialogue text and the corresponding work order text, extract the sentiment polarity and responsibility attribution labels respectively, and judge whether it is a contradictory event by comparing the sentiment polarity and responsibility attribution labels. If the judgment result is a contradictory event, it is marked as a contradictory event, and a semantic contradiction index is generated according to the frequency of contradictory events; Service request frequency change rate extraction: The total number of cross-channel service requests from customers is counted according to the preset period, and the service request frequency change rate ΔF of the request volume in adjacent periods is calculated using the time series difference algorithm; Extraction of historical performance quality scores: Extract the performance indicator set from the service work order database, and generate a comprehensive performance quality score Q by weighted summation; State vector generation module: Generates a customer relationship state vector V based on the health score based on the semantic contradiction index, service request frequency change rate and historical performance quality score 客户 ; Customer relationship network topology building module: based on the customer relationship state vector V 客户 , create a triple structure of customer nodes, service resource nodes, and relationship degradation inducement nodes, and based on the triple structure, construct the causal correlation edge of "customer-service defect-relationship decay", output the customer relationship network topology with an explainable path, and form a relationship maintenance knowledge graph; Adaptive relationship repair strategy generation module: performs the following according to the relationship maintenance knowledge graph: Identify high-weight relationship decay paths; Match the countermeasures in the preset strategy library.

2. The intelligent customer relationship management system based on multimodal data fusion according to claim 1, characterized in that: In the extraction of the semantic contradiction index, dual-channel semantic encoding is performed on the voice dialogue text and the corresponding work order text based on natural language processing technology, and the sentiment polarity and responsibility attribution labels are extracted respectively. The sentiment polarity vector includes positive sentiment and negative sentiment, and the responsibility attribution label is used to identify the responsible party specified in the work order, including the customer, customer service or a third party; For each service event, we directly compare the sentiment polarity in the voice text with the responsibility attribution label in the work order text. When the sentiment polarity is positive and the work order responsibility is attributed to the customer or a third party, it is marked as a contradictory event. Based on the frequency of occurrence of contradictory events, we statistically generate a semantic contradiction index to measure the potential relationship contradictions in the customer service process.

3. The intelligent customer relationship management system based on multimodal data fusion according to claim 1, characterized in that: The service request frequency change rate ΔF is calculated as: Among them, ΔF is the change rate of service request frequency, which means the relative change percentage of service request volume in two adjacent cycles, F t is the total number of service requests in the current cycle t, that is, the total number of cross-channel service requests from customers within the time window, F t-1 It is the total number of service requests in the previous cycle t-1, that is, the total number of cross-channel service requests from customers in the previous time window.

4. The intelligent customer relationship management system based on multimodal data fusion according to claim 1, characterized in that: The performance indicator set includes problem resolution rate and customer satisfaction score, and the comprehensive performance quality score is expressed as: Q=w1·S 响应 +w2·R 解决 +w3·C 满意 ,in: Q is the comprehensive score of performance quality, which represents the comprehensive quality score of customer service performance. 响应 is the response time score, R 解决 is the problem solving rate score, C 满意 is the customer satisfaction score, w1, w2, w3 are the weight coefficients of each indicator.

5. The intelligent customer relationship management system based on multimodal data fusion according to claim 1, characterized in that: The state vector generation module specifically includes: Semantic contradiction index fusion unit: Based on the semantic contradiction index, quantify the inconsistency between emotional conflict and responsibility attribution in customer interaction, and obtain the emotional health score H of the customer relationship. 情感 ,The higher the semantic contradiction index is, the lower the customer’s emotional health score is; Service request frequency change rate fusion unit: According to the value of the service request frequency change rate ΔF, the volatility of customer service demand is evaluated. The larger the change rate ΔF, the greater the fluctuation of customer demand. After normalizing the service request frequency change rate to the [0,1] interval, the customer demand volatility score H is obtained. 需求 ; Performance quality score fusion unit: Based on the historical performance quality score Q, combined with the customer's past service performance, the performance health of the customer relationship is evaluated. The performance quality score Q directly reflects the customer's satisfaction with the service and the effectiveness of problem solving. The performance health score H 履约 Directly use the standardized historical performance quality score Q, and the standardized performance health score is; The final customer relationship state vector V 客户 =[H 情感 ,H 需求 ,H 履约 ], where each dimension represents the health status of the customer relationship.

6. The intelligent customer relationship management system based on multimodal data fusion according to claim 1, characterized in that: The creation of the customer node includes based on the customer relationship state vector V 客户 The health scores extracted from the client node are used to create a client node, including the following information: The basic identity of the customer; The client's emotional health score H 情感 ; Customer demand volatility score H 需求 ; The customer's fulfillment health score H 履约 ; The creation of the service resource node includes creating a service resource node according to various service resources in the service system, including customer service personnel, service platform, and problem solution, and the service resource node represents the interaction between the customer and the service resource, including the service history of the service resource; The creation of the relationship degradation inducement node includes creating the relationship degradation inducement node by analyzing the customer's historical interaction data, and identifying potential factors that lead to the decline of customer relationships, including shirking of responsibility, service defects, and contract performance issues; Creation of triple structure: Based on the creation of customer nodes, service resource nodes and relationship degradation inducement nodes, a triple structure is constructed to represent the relationship between customers and service resources and service defects. The triple form is: [customer node, service resource node, relationship degradation inducement node].

7. The intelligent customer relationship management system based on multimodal data fusion according to claim 6, characterized in that: The causal relationship edge is based on the customer state vector V 客户 Construct and form a causal link of customer-service-relationship decay, and the causal association edge also includes edge weight w 边 , edge weight w 边 Used to quantify the severity of customer relationship degradation and the impact of different degradation factors on customer relationships; The customer nodes, service resource nodes, relationship degradation inducement nodes and causal association edges form a complete relationship maintenance knowledge graph, and output a customer relationship network topology structure with an explainable path.

8. The intelligent customer relationship management system based on multimodal data fusion according to claim 7, characterized in that: The edge weight is calculated as: w 边 =α·|ΔF|+β·(1-H 情感 )+γ·(1-Q 履约 ), where ΔF is the change rate of service request frequency. The larger the value, the more unstable the customer demand. 情感 Customer emotional health score. The lower the emotional health, the more dissatisfied the customer is, which makes the customer relationship more likely to deteriorate. In order to reflect the severity of the degradation, 1-H is used. 情感 , Q 履约 is the historical performance quality score. The lower the score, the poorer the service performance in the past, the worse the customer experience, and the more likely it is to lose customers. Use 1-Q 履约 Historical fulfillment quality score, α, β, γ are weight coefficients, which respectively control the impact of service request frequency change rate, emotional health score and fulfillment quality score on edge weight.

9. The intelligent customer relationship management system based on multimodal data fusion according to claim 7, characterized in that: The adaptive relationship repair strategy generation module specifically includes: Identify high-weight relationship decay paths: High-weight relationship decay paths represent paths with a high probability of customer relationship degradation. They are composed of multiple causal association edges. Use the Dijkstra algorithm to start from the customer node, traverse all causal relationship paths in the graph, and calculate the total weight of each path. The total weight W 路径 is the weighted sum of the edge weights on the path: Among them, w 边,i is the weight of the ith edge in the path, n is the number of edges in the path, and the path with a total weight greater than the weight threshold is selected as the high-weight relationship attenuation path; Matching countermeasures in the preset strategy library: Based on the identified high-weight relationship decay paths, countermeasures in the preset strategy library will be used to repair customer relationships. The countermeasures include predefined strategies for dealing with customer relationship decay.

10. The intelligent customer relationship management system based on multimodal data fusion according to claim 9, characterized in that: The preset strategy library specifically includes: Automatically trigger compensation agreements: If there is a problem of shirking responsibility on the path, compensation measures will be automatically triggered; Assign a dedicated account manager: If there are performance issues on the path and there is a long history of negative interactions, a dedicated account manager will be assigned to the customer; Prioritize service issues: If a service defect occurs on the path, immediately prioritize its service request and fix the problem.

Images