Intelligent customer life cycle management AiCRM method and system

Abstract

The invention relates to the technical field of artificial intelligence customer service, provides an intelligent customer life cycle management AiCRM method and system, and is used for solving the problems of low customer service accuracy and poor response timeliness in the prior art. The method comprises the steps that in the interaction process of a target client and an enterprise, client voice, expression and text data are acquired, multi-modal sentiment analysis is carried out, and a client sentiment feature set is generated; dynamically updating a customer emotion evolution graph based on the customer emotion feature set, and generating an accurate customer emotion score by adopting an emotion calculation model in combination with historical data and real-time interaction information; and when the score is lower than a preset baseline, the system automatically generates a personalized service strategy adjustment scheme matched with the current emotional state and the historical change trend of the customer. According to the application, the customer service accuracy and the response timeliness are improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence customer service technology, and in particular to an intelligent customer lifecycle management AiCRM method and system. Background Art

[0002] In the customer service sector, companies need to accurately identify customer emotions in real time and dynamically adjust service strategies based on historical interaction data to improve customer satisfaction and loyalty. Traditional manual analysis is inefficient and difficult to quantify emotional trends. Therefore, an intelligent approach is urgently needed to achieve multimodal emotion fusion analysis and automated decision-making.

[0003] Currently, there is a customer relationship management system based on text sentiment analysis, which uses natural language processing technology to extract emotional tendencies in customer service conversations and combines it with a simple rule engine to generate service recommendations.

[0004] This solution relies only on single text data and ignores multi-dimensional emotional information such as voice intonation and facial expressions, resulting in one-sided emotional judgment; at the same time, the rule engine lacks modeling of the customer's historical emotional evolution trends, and the strategy adjustment is rigid and difficult to adapt to personalized needs. Summary of the Invention

[0005] This application provides an intelligent customer lifecycle management AiCRM method and system to solve the problems of low customer service accuracy and poor response timeliness in the existing technology.

[0006] In a first aspect, the present application provides an intelligent customer lifecycle management AiCRM method, comprising:

[0007] During the interaction between target customers and enterprises, obtain target customers' voice call records, real-time facial expression images and text feedback data;

[0008] Performing semantic analysis on the voice call records and the text feedback data, performing dynamic feature recognition on the facial expression images, and cross-modally correlating the semantic analysis results with the dynamic feature recognition results to generate a customer emotion feature set;

[0009] Based on the customer emotion feature set, dynamically updating the customer emotion evolution graph in the preset multimodal customer emotion database;

[0010] Based on the updated customer emotion evolution graph and the customer emotion feature set, a sentiment computing model is used to generate a target customer emotion score for the target customer;

[0011] When the customer's emotional score is lower than the preset service baseline, service strategy adjustment information is generated to match the customer's current emotional state and historical emotional change trends to achieve intelligent customer lifecycle management AiCRM.

[0012] Optionally, the generating of a target customer emotion score of the target customer using an emotion computing model based on the updated customer emotion evolution graph and the customer emotion feature set includes:

[0013] Dividing the updated customer emotion evolution graph into multiple historical emotion segments according to a preset time window, and extracting the emotion polarity label distribution density corresponding to the voice call record, the real-time facial expression image, and the text feedback data in each historical emotion segment;

[0014] Discretizing and marking the current emotion polarity label in the customer emotion feature set;

[0015] Calculating a target association strength value between the current emotion polarity label and each historical emotion segment based on the emotion polarity label distribution density;

[0016] Assigning a weight value to each historical emotion segment according to the target association strength value, and linearly superimposing the weight value with the discretization labeling result to generate an emotion state fusion vector;

[0017] The emotional state fusion vector is input into the emotional computing model to output the target customer emotional score.

[0018] Optionally, inputting the emotional state fusion vector into an emotional computing model and outputting a target customer emotional score includes:

[0019] By using an emotion calculation model, the emotion state fusion vector is divided into a historical emotion weight component and a current emotion tag component according to a preset dimension;

[0020] Performing hierarchical processing on the historical sentiment weight components, and adjusting the superposition coefficient of each layer according to the target association strength value in each processing layer;

[0021] Interlacing and splicing the adjusted superposition coefficient with the current emotion tag component to generate a hierarchical fusion signal;

[0022] Using the dynamic aggregation nodes of the sentiment computing model, the difference between the historical sentiment weight component and the current sentiment label component in the hierarchical fusion signal of each level is compared layer by layer to generate the hierarchical contribution parameter;

[0023] selectively activating transmission paths in the dynamic aggregation node according to the hierarchical contribution parameter;

[0024] Perform scalar compression on the output signal of the activated transmission path to generate an initial customer sentiment score;

[0025] Based on the statistical deviation of the distribution density of the sentiment polarity labels, boundary constraints are performed on the initial customer sentiment scores, and the target customer sentiment scores are output.

[0026] Optionally, the dynamic aggregation node of the emotion calculation model is used to compare the difference in proportion between the historical emotion weight component and the current emotion label component in the hierarchical fusion signal of each level layer by layer to generate the hierarchical contribution parameter, including:

[0027] Split the hierarchical fusion signal of each level into historical signal segments and current signal segments according to a preset ratio;

[0028] Performing amplitude accumulation calculation on the historical signal segment and the current signal segment respectively to obtain a historical cumulative amplitude and a current cumulative amplitude;

[0029] generating a level difference amplitude according to an absolute value of a difference between the historical cumulative amplitude and the current cumulative amplitude;

[0030] Comparing the level difference amplitude with a preset difference threshold, and activating a difference accumulator corresponding to the level if the level difference amplitude exceeds the preset difference threshold;

[0031] Directionally mark the signal output by the activated difference accumulator;

[0032] Based on the directional marking results, polarity weighting is performed on the signals output by the difference accumulators in the same level to generate a level difference index;

[0033] The hierarchical difference index of each level is accumulated in a sliding window according to the preset aggregation rules to generate the hierarchical contribution parameter.

[0034] Optionally, the calculating, based on the distribution density of the emotion polarity label, a target association strength value between the current emotion polarity label and each historical emotion segment includes:

[0035] The emotional polarity labels in each historical emotional segment are classified and counted according to the preset emotional categories, and the emotional category distribution histogram of each historical emotional segment is generated;

[0036] Performing category matching on the current emotion polarity label to determine the emotion category to which the current emotion polarity label belongs;

[0037] Calculate the frequency of occurrence of the emotion category in each emotion category distribution histogram;

[0038] Generating an initial correlation strength value according to the degree of deviation between the occurrence frequency ratio and a preset benchmark frequency;

[0039] Based on the time decay coefficient of each historical emotion segment, the initial association strength value is corrected to generate a target association strength value.

[0040] Optionally, when the customer emotion score is lower than a preset service baseline, generating service strategy adjustment information that matches the customer's current emotion state and historical emotion change trends includes:

[0041] According to the customer emotion evolution graph, establish an emotion polarity label change curve of N interaction cycles, where N is greater than or equal to 2;

[0042] Matching the emotion polarity label corresponding to the customer's current emotion state with the key turning point in the emotion polarity label change curve to determine the stage position of the customer's current emotion state in the historical emotion change trend;

[0043] Selecting a corresponding strategy combination in a preset strategy template library according to the stage position;

[0044] comparing the customer sentiment score to a preset service baseline;

[0045] According to the comparison result, parameter enhancement processing is performed on the policy combination to generate service policy adjustment information.

[0046] Optionally, after generating service strategy adjustment information that matches the customer's current emotional state and historical emotional change trends, the method further includes:

[0047] According to the service strategy adjustment information, the service content display order in the interaction process between the target customer and the enterprise is adjusted in real time, and the emotional tendency label in the customer emotional evolution map is updated;

[0048] After the customer completes the current interaction, a closed-loop feedback signal is generated based on the adjusted service content display order and the changing trend of the customer's emotional score. The closed-loop feedback signal is used to iteratively optimize the mapping relationship between the emotional computing model and the service strategy adjustment information.

[0049] In a second aspect, the present application provides an intelligent customer lifecycle management AiCRM system, comprising:

[0050] The acquisition module is used to obtain the target customer's voice call records, real-time facial expression images and text feedback data during the interaction between the target customer and the enterprise;

[0051] An analysis module is configured to perform semantic analysis on the voice call records and the text feedback data, perform dynamic feature recognition on the facial expression images, and perform cross-modal correlation between the semantic analysis results and the dynamic feature recognition results to generate a customer emotion feature set;

[0052] An updating module, configured to dynamically update a customer emotion evolution graph in a preset multimodal customer emotion database based on the customer emotion feature set;

[0053] A first generating module is configured to generate a target customer sentiment score of the target customer using a sentiment computing model based on the updated customer sentiment evolution graph and the customer sentiment feature set;

[0054] The second generation module is used to generate service strategy adjustment information that matches the customer's current emotional state and historical emotional change trends when the customer's emotional score is lower than the preset service baseline, so as to realize intelligent customer lifecycle management AiCRM.

[0055] In a third aspect, the present application provides a computing device comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute an intelligent customer lifecycle management AiCRM method as described in any one of the first aspects.

[0056] In a fourth aspect, the present application provides a computer storage medium having computer program instructions stored thereon, wherein the computer program instructions, when executed by a processor, implement an intelligent customer lifecycle management AiCRM method as described in any one of the first aspects.

[0057] In the present application, an intelligent customer lifecycle management AiCRM method is provided, which includes: obtaining voice call records, real-time facial expression images and text feedback data of target customers during the interaction between target customers and enterprises; performing semantic analysis on the voice call records and the text feedback data, performing dynamic feature recognition on the facial expression images, cross-modally correlating the semantic analysis results with the dynamic feature recognition results to generate a customer emotion feature set; based on the customer emotion feature set, dynamically updating a customer emotion evolution map in a preset multimodal customer emotion database; based on the updated customer emotion evolution map and the customer emotion feature set, using an emotion computing model to generate a target customer emotion score for the target customer; when the customer emotion score is lower than a preset service baseline, generating service strategy adjustment information that matches the customer's current emotion state and historical emotion change trends to realize intelligent customer lifecycle management AiCRM.

[0058] The technical solution provided by this application has the following beneficial effects:

[0059] This application solves the problem of one-sidedness of sentiment analysis based on a single data source by simultaneously collecting voice, expression, and text data, and provides a multi-dimensional data foundation for a comprehensive assessment of the customer's emotional state. The correlation analysis of speech semantics, text semantics, and expression features breaks through the limitations of traditional single-modal analysis and improves the accuracy of sentiment feature extraction. It realizes the temporal modeling of the customer's emotional state, enabling the system to continuously track the trajectory of customer emotional changes and provide data support for trend analysis. By integrating real-time emotional features and historical evolution laws, a quantitative emotional score is generated, which solves the problem of strong subjectivity of manual judgment. Personalized service strategies are automatically triggered based on the score, realizing closed-loop management from emotion recognition to service optimization, and improving the accuracy and timeliness of customer service.

[0060] Furthermore, this application also divides the customer emotion evolution map into segments according to time windows, extracts the multimodal emotion distribution characteristics of each segment, calculates the correlation strength between the current emotion label and the historical segment, assigns weights accordingly and generates a fusion vector, and finally outputs the emotion score.

[0061] Moreover, this process realizes the dynamic weighted fusion of historical sentiment data and real-time interaction information, taking into account the temporal evolution of customer emotions and combining the detailed characteristics of the current interaction, so that the generated sentiment score has both historical continuity and can reflect the latest changes, thereby improving the accuracy and reliability of sentiment status assessment.

[0062] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0064] Figure 1 A flowchart of an intelligent customer lifecycle management AiCRM method provided in an embodiment of the present application;

[0065] Figure 2 A schematic diagram of the structure of an intelligent customer lifecycle management AiCRM system provided in an embodiment of the present application;

[0066] Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0067] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0068] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0069] Researchers have found that existing customer relationship management systems have problems with emotion recognition, such as a single data source and limited analysis dimensions. It is difficult to fully and accurately capture the true emotional state of customers, and lacks the ability to dynamically track changes in customer emotions, resulting in delayed adjustments to service strategies. Based on this, an embodiment of the present application provides an intelligent customer lifecycle management AiCRM method, which achieves a comprehensive quantitative assessment of customer emotions through multimodal data collection and fusion analysis, and dynamically tracks emotional evolution trends in combination with time series modeling technology, automatically triggering personalized service strategies when abnormal customer emotions are detected. The technical solution of this application can be applied to customer relationship management scenarios that require accurate grasp of customer emotional changes and the provision of intelligent services, such as high-value customer service areas such as financial customer service and e-commerce after-sales.

[0070] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0071] Figure 1 A flowchart of an intelligent customer lifecycle management AiCRM method provided in an embodiment of the present application is shown as follows: Figure 1 As shown, the method includes:

[0072] Step 101: During the interaction between the target customer and the enterprise, obtain the target customer's voice call records, real-time facial expression images, and text feedback data.

[0073] In step 101, the voice call record represents the audio data of a conversation between a customer and a company's customer service representative, including voice features such as intonation and speaking speed. The real-time facial expression image represents the customer's facial expression captured by a video capture device. The text feedback data represents the textual information entered by the customer during the interaction, such as chat logs and comments.

[0074] In an embodiment of the present application, the audio data of the customer's voice call is collected in real time through the customer service system, the camera is called to obtain the customer's facial expression video stream, and the text information entered by the customer in the interactive interface is recorded. These three types of original data are used as the input source for subsequent sentiment analysis.

[0075] For example, when a bank customer consults about financial products on a mobile banking app, the system records the call voice, captures the customer's video footage, and records the text questions entered by the customer in the chat window. The three are collected simultaneously and timestamped.

[0076] Step 102: Perform semantic analysis on the voice call record and the text feedback data, perform dynamic feature recognition on the facial expression image, perform cross-modal association between the semantic analysis result and the dynamic feature recognition result, and generate a customer emotion feature set.

[0077] In step 102, semantic analysis involves identifying intent and emotional tendencies in text and speech content. Dynamic feature recognition involves extracting expression features from video sequences. Cross-modal association involves fusing and correlating emotional features from different data sources. The customer emotional feature set represents the structured emotional feature data generated by integrating multi-source data.

[0078] In an embodiment of the present application, after the speech data is processed for speech-to-text conversion, emotional keyword extraction and emotional tendency analysis are performed together with the original text data; at the same time, face detection and expression feature extraction are performed on the video data; the text emotion analysis results and the expression recognition results are feature aligned and weightedly fused to generate a structured data set containing multi-dimensional emotional features.

[0079] For example, the system converts a customer's words "the profit is too low" into text and analyzes the negative emotion; at the same time, it detects the customer's frowning expression; and combines these two features according to preset weights to generate a set of emotional features containing the "negative-strong" label.

[0080] Step 103: Based on the customer emotion feature set, dynamically update the customer emotion evolution graph in the preset multimodal customer emotion database.

[0081] In step 103, the multimodal customer emotion database represents a structured database storing historical emotion features. The customer emotion evolution graph represents a graph of emotion state changes with time as the axis, where nodes represent emotion states and edges represent state transitions.

[0082] In an embodiment of the present application, the currently generated emotional feature set is associated with the historical records of the customer in the database, and the emotional state nodes and transfer edge weights are updated to form a time series graph reflecting the long-term changing trend of the customer's emotions.

[0083] For example, the system compares the "negative-strong" feature with the customer's past three interaction records and finds that the customer's negative emotions continue to worsen in the financial product consultation scenario. It adds a new state node to the evolution graph and adjusts the transfer edge weight.

[0084] Step 104: Based on the updated customer emotion evolution graph and the customer emotion feature set, a target customer emotion score of the target customer is generated using an emotion calculation model.

[0085] In step 104, the sentiment calculation model represents a sentiment evaluation model that comprehensively considers real-time features and historical trends. The target customer sentiment score represents a quantified customer sentiment state value.

[0086] In an embodiment of the present application, the model receives current emotional features and evolutionary graph data, first calculates the similarity between the current features and the historical state, then weights them in combination with the time decay factor, and finally outputs a standardized score through normalization processing.

[0087] For example, the model calculates the similarity between current negative emotions and historical records to be 0.8, and after weighting by the time factor, it is 0.72, which is normalized and converted into a percentage score of 58 points (calculation formula: 100*(1-0.72)=58).

[0088] Step 105: When the customer emotion score is lower than the preset service baseline, service strategy adjustment information matching the customer's current emotion state and historical emotion change trend is generated to implement intelligent customer lifecycle management AiCRM.

[0089] In step 105, the preset service baseline represents a preset emotion score threshold. The customer's current emotional state is the instant emotion analysis result generated by extracting emotion features and processing the emotion calculation model from the voice call records, facial expression images and text feedback data obtained in real time from the latest customer interaction. The current emotional state of the customer is characterized by the emotion polarity label generated by the current interaction data. Each emotion polarity label (such as positive / negative / neutral) corresponds to a specific emotion state classification to form a quantitative expression. Service strategy adjustment information refers to personalized service optimization instructions dynamically generated according to the customer's emotional state, specifically including communication language adjustment plans (such as tone and urgency, word choice), service priority ranking (such as response speed, processing level), service content adjustment (such as preferential strength, solution) and service channel selection (such as transfer to manual, video communication) in four dimensions.

[0090] In an embodiment of the present application, when the score is lower than the baseline, the system selects a matching response plan from the strategy library based on the emotional feature type and evolution trend, adjusts the plan strength parameter based on the score difference, and generates a final execution strategy.

[0091] For example, if the baseline score is 70 and the current score is 58, triggering a strategy adjustment, the system selects the "Wealth Management Manager Follow-up + Profit Compensation" plan. Based on the 12-point difference, the follow-up time limit is adjusted from 48 hours to within 24 hours, and the compensation ratio is increased by 5%.

[0092] This method achieves accurate identification and quantitative assessment of customers' emotional states through multi-dimensional data collection and fusion analysis. It can dynamically track the changing trends of customers' emotions and automatically generate and optimize service strategies when emotional anomalies are detected, thereby improving the timeliness and accuracy of customer service and providing intelligent decision-making support for the company's customer relationship management.

[0093] To address the accuracy and timeliness issues of customer sentiment scores, in some embodiments, step 104: generating a target customer sentiment score for the target customer using a sentiment computing model based on the updated customer sentiment evolution graph and the customer sentiment feature set, includes:

[0094] Step 201: Divide the updated customer emotion evolution map into multiple historical emotion segments according to a preset time window, and extract the emotion polarity label distribution density corresponding to the voice call record, the real-time facial expression image and the text feedback data in each historical emotion segment.

[0095] In step 201, a preset time window is set based on the business scenario, such as weekly or monthly divisions. Historical sentiment segments are units of historical customer interaction data divided into fixed time periods. The distribution density of sentiment polarity labels reflects the frequency of occurrence of different sentiment types (e.g., positive, neutral, negative) within each time period.

[0096] In an embodiment of the present application, the system reads the customer's interaction data for the past three months, divides it into three historical segments by month, and counts the proportion of emotional tags in voice, expression and text data in each segment to form three emotional distribution histograms.

[0097] Step 202: Discretely mark the current emotion polarity label in the customer emotion feature set.

[0098] In step 202, the current sentiment polarity label refers to the immediate sentiment characteristics obtained from the most recent customer interaction. It is distinguished from historical data by a timestamp, and its currency is determined by the real-time nature of data collection and processing. Discretization is the process of converting continuous sentiment values into several pre-defined categorical labels.

[0099] In the embodiment of the present application, the sentiment value of 0.7 (range 0-1) obtained from the current call analysis is discretized into a "positive" label to facilitate comparison with historical clips.

[0100] Step 203: Calculate the target association strength value between the current emotion polarity label and each historical emotion segment based on the emotion polarity label distribution density.

[0101] In step 203 , the target association strength value quantifies the similarity between the current sentiment and the historical segments, and is calculated by comparing the position of the current label in the distribution of each segment.

[0102] In the embodiment of the present application, the percentile position of the current "positive" label in the distribution of three historical segments is calculated to obtain three association strength values.

[0103] Step 204: assigning a weight value to each historical emotion segment according to the target association strength value, and linearly superimposing the weight value and the discretization labeling result to generate an emotion state fusion vector.

[0104] In step 204, weights reflect the degree of influence of each historical segment on the current score. Each historical emotional segment is assigned a corresponding weight, and each segment's weight is independently calculated based on its target association strength with the current emotional state. The emotional state fusion vector is a feature representation that combines the weights with the current label.

[0105] In the embodiment of the present application, the three historical segments are given weights of 0.3, 0.5, and 0.2, and are weighted with the current "positive" label to generate a five-dimensional fusion vector.

[0106] Step 205: Input the emotional state fusion vector into the emotional calculation model, and output the target customer emotional score.

[0107] In the embodiment of the present application, after the fusion vector is input into the model, it is processed through three layers to output a scoring result of 0-100 points.

[0108] Here's a specific example:

[0109] When a bank customer consulted about the same wealth management product for the fourth time via mobile banking, the system divided the customer's past three consultations into three historical sentiment segments by month. The system calculated the occurrence rate of the "negative" sentiment label in each segment to be 60%, 70%, and 80%, respectively, to form the distribution density of the sentiment polarity label. During this consultation, the system detected the customer's voice saying "still too low" (converted to text and labeled "negative") and the facial expression "frowning" (labeled "strong"). The system combined these two factors to generate the current sentiment polarity label "negative-strong." The percentile method was used to calculate the correlation strength of this label with the three historical segments, yielding values of 0.6, 0.7, and 0.8, respectively (calculated as: current label strength / historical maximum strength). This was then corrected by applying time decay coefficients of 0.9, 0.8, and 0.7 (calculated as: 1 - 0.1 × number of months between intervals), resulting in final correlation strengths of 0.54, 0.56, and 0.56, respectively. The weights are proportionally allocated as 0.32, 0.34, and 0.34 (calculation formula: association strength value of each segment / total), and linearly superimposed with the current label value 0.8 (negative 0.5 + strong 0.3) to generate a fusion vector [0.16, 0.17, 0.17, 0.5, 0.3]. After input into the model, the output sentiment score is 55 points (calculation formula: 100×(1-0.8×0.7)=55, where 0.7 is the model weighting parameter).

[0110] In the embodiment of the present application, the method realizes dynamic quantitative assessment of the customer's emotional state through the fusion of time series data analysis and real-time features, taking into account both historical change patterns and current characteristics, providing a reliable basis for precise service.

[0111] In order to solve the problem of fusing historical and current data in sentiment score calculation, in some embodiments, step 205: inputting the sentiment state fusion vector into the sentiment calculation model and outputting the target customer sentiment score includes:

[0112] Step 301: Using an emotion calculation model, the emotion state fusion vector is divided into a historical emotion weight component and a current emotion label component according to a preset dimension.

[0113] In step 301, the preset dimensions are set based on the model structure. Typically, the first N dimensions of the vector are classified as historical components, while the last M dimensions are current components. The historical sentiment weight component refers to the portion of the fused vector that reflects the influence of historical sentiment. The current sentiment tag component refers to the sentiment features extracted from the most recent customer interaction data. This component is distinguished from historical data by the data collection timestamp. Its current nature is reflected in the fact that processing always refers to the data generated by the most recent interaction.

[0114] In an embodiment of the present application, the system classifies the first three dimensions of the five-dimensional fusion vector as historical weight components and the last two dimensions as current label components.

[0115] Step 302: hierarchically process the historical sentiment weight components, and adjust the superposition coefficient of each level according to the target association strength value in each processing level.

[0116] In step 302, hierarchical processing refers to multi-level feature extraction and conversion of historical components. The superposition coefficient controls the contribution of each level to the final result and is dynamically adjusted according to the correlation strength. For example, in a customer service scenario, the customer's historical interactions within 3 months are divided into 12 segments by week. According to the correlation strength between each segment and the current complaint event (such as the matching degree of complaint-related conversations), the influence weight of each week's data in the calculation model is adjusted separately. If the correlation strength in the most recent week is high, its superposition coefficient is increased. The superposition coefficient of each level is obtained by dynamically calculating the target correlation strength value between the historical emotional segment corresponding to each level and the current interaction when hierarchically processing the historical emotional weight components.

[0117] In the embodiment of the present application, the model performs three-layer processing on the historical components, and each layer adjusts the conversion coefficient according to the association strength value, and the stronger the association, the higher the coefficient.

[0118] Step 303: interlacing and splicing the adjusted superposition coefficient with the current emotion tag component to generate a hierarchical fusion signal.

[0119] In step 303, interleaving is to alternately combine the adjusted historical features with the current features according to specific rules to form a new feature representation containing spatiotemporal correlations. For example, the adjusted weekly weight coefficients (such as [0.2, 0.3, 0.5]) are alternately combined with the emotional tags of the current complaint event (such as "anger") in the format of [weight-tag-weight...] to generate a signal sequence containing temporal weights and real-time emotions. The hierarchical fusion signal is a composite feature vector formed by interleaving the adjusted historical emotional weight components with the current emotional tag components in a specific order. It not only retains the hierarchical characteristics of historical emotional changes, but also integrates the real-time characteristics of current emotions, providing a unified feature representation for subsequent analysis.

[0120] In an embodiment of the present application, the three-layer historical processing results and the current mark are spliced in the order of "history-current-history" to generate a seven-dimensional hierarchical fusion signal.

[0121] Step 304: Using the dynamic aggregation node of the emotion calculation model, the difference in proportion between the historical emotion weight component and the current emotion label component in the hierarchical fusion signal of each level is compared layer by layer to generate a hierarchical contribution parameter.

[0122] In step 304, the dynamic aggregation node, a special processing unit in the model, evaluates the importance of each layer's contribution by comparing the relative proportions of historical and current features. The "proportion difference" refers to the difference between the proportions of the historical sentiment weight component and the current sentiment tag component in the fusion signal at each layer. This reflects the relative strength of historical sentiment influence and current sentiment expression and is used to assess the importance of each layer. The layer contribution parameter is a quantitative value of each layer's contribution to the final sentiment score, calculated based on the proportion difference. A larger parameter value indicates a greater impact of the layer's features on the score result, and is used to guide subsequent path selection.

[0123] In an embodiment of the present application, the node calculates the difference between the proportion of historical and current features in each layer of signal. The greater the difference, the higher the contribution parameter of the layer.

[0124] Step 305: Selectively activate the transmission paths in the dynamic aggregation node according to the hierarchical contribution parameters.

[0125] In step 305, the transmission path refers to the connection channel between different processing units within the dynamic aggregation node, not involving the connection between nodes. Each path corresponds to a specific signal processing method. Selective activation is to open the most relevant processing path based on the contribution parameter.

[0126] In an embodiment of the present application, the node activates two paths whose contribution parameters exceed a threshold value and blocks other low-contribution paths.

[0127] Step 306: Perform scalar compression on the output signal of the activated transmission path to generate an initial customer sentiment score.

[0128] In step 306, the output signal is the result of processing the hierarchical fusion signal after filtering through the activation path. This signal retains the key features of the original signal while undergoing dimensionality reduction and normalization. Scalar compression is the process of converting a multidimensional signal into a single numerical value through methods such as weighted summation. The initial customer sentiment score is a preliminary score obtained by compressing the output signal of the activation path. This score does not yet take into account the constraints of historical sentiment fluctuations and is an intermediate calculation result of the target customer sentiment score.

[0129] In an embodiment of the present application, the multidimensional signals output by the activated paths are weighted and summed according to the contribution parameters of each path, and compressed into an initial score in the range of 0-1. Specifically: first, the amplitude of the multidimensional signals output by each activated path is summed, and then the summation result is mapped to the interval [0,1] by a preset normalization factor, and finally the median of the output results of all paths is taken as the initial score. For example, in a customer complaint handling scenario, the three activated paths output signal vectors [0.6, 0.2, 0.8], [0.4, 0.5, 0.3] and [0.7, 0.1, 0.9] respectively. Each path is first summed to obtain 1.6, 1.2 and 1.7, and then compressed using a normalization factor of 2.0 to obtain 0.8, 0.6 and 0.85, and the median 0.8 is taken as the initial sentiment score.

[0130] Step 307: Based on the statistical deviation of the distribution density of the sentiment polarity labels, perform boundary constraints on the initial customer sentiment scores and output the target customer sentiment scores.

[0131] In step 307, the statistical deviation is derived from the standard deviation of the frequency of various emotion labels (e.g., positive / negative) in the customer's historical interactions, reflecting the degree of customer emotional fluctuation. Boundary constraints are used to rationalize the initial score based on this fluctuation range. For example, if a customer's historical emotional distribution is detected to be highly skewed (emotionally volatile), the initial score is constrained from 0.3 to the range [0.4, 0.6] to avoid misjudgments due to single extreme emotions. A final score of 0.45 is output for service strategy formulation.

[0132] In an embodiment of the present application, when a large historical sentiment fluctuation is detected, the initial score of 0.4 is constrained to the range of 0.3-0.5, and a final score of 0.45 is output.

[0133] Here's a specific example:

[0134] When a bank customer inquires about the same wealth management product for the fifth time, the system splits the emotional state fusion vector [0.16, 0.17, 0.17, 0.5, 0.3] into a historical component [0.16, 0.17, 0.17] and a current component [0.5, 0.3]. The historical component undergoes three-level processing, adjusting the stacking coefficients to 0.3, 0.4, and 0.3 based on the correlation strengths of 0.54, 0.56, and 0.56, respectively (calculated as: correlation strength value × layer weight coefficient 0.5). These adjusted coefficients are interleaved with the current component to generate a seven-dimensional hierarchical fusion signal [0.3, 0.5, 0.4, 0.3, 0.3, 0.5, 0.3]. The dynamic aggregation node calculated the difference between the historical and current contributions of each layer, finding the largest difference in the second layer (historical contribution 0.4 / 0.9 = 44%, current contribution 0.3 / 0.9 = 33%, a difference of 11%). The layer contribution parameters [0.2, 0.4, 0.2] were generated. Two primary paths were activated based on these parameters, and the output signals were compressed by weighted summation (weights 0.4 and 0.4) to yield an initial score of 0.38 (calculated as (0.3 × 0.4 + 0.4 × 0.4) / 0.8). Taking into account the historical sentiment fluctuation range (maximum deviation 0.2), the score was constrained to the range of 0.3-0.5, ultimately outputting a target score of 0.4.

[0135] In the embodiment of the present application, the method achieves an optimized combination of historical sentiment patterns and real-time features through multi-level feature fusion and dynamic path selection, so that the final score reflects both long-term trends and current status, providing a reliable basis for precise services.

[0136] To address the accuracy issue of historical and current feature fusion in sentiment scoring, in some embodiments, step 304: utilizing the dynamic aggregation node of the sentiment computing model to compare the difference in proportion between the historical sentiment weight component and the current sentiment tag component in the hierarchical fusion signal of each level, layer by layer, to generate a hierarchical contribution parameter, includes:

[0137] Step 401: Split the hierarchical fusion signal of each level into historical signal segments and current signal segments according to a preset ratio.

[0138] In step 401, the preset ratio is set based on the signal dimensions, typically splitting the signal based on the original ratio of historical to current features. The historical signal segment refers to the portion of data in the hierarchically fused signal that reflects historical emotional features, including the adjusted superposition coefficients for the historical emotional weight components. The current signal segment refers to the portion of data that reflects real-time emotional features, including the discretized labeling results for the current emotional label component.

[0139] In an embodiment of the present application, the system divides the seven-dimensional hierarchical fusion signal into historical 5-dimensional and current 2-dimensional segments to form historical signal segments and current signal segments.

[0140] Step 402: Perform amplitude accumulation calculation on the historical signal segment and the current signal segment respectively to obtain a historical cumulative amplitude and a current cumulative amplitude.

[0141] In step 402, the amplitude accumulation calculation refers to the weighted summation of the values of each dimension in the signal segment. The historical cumulative amplitude reflects the overall strength of the historical feature. The current cumulative amplitude reflects the overall strength of the real-time feature.

[0142] In an embodiment of the present application, each dimension of the historical signal segment is multiplied by the corresponding weight and then summed to obtain the historical cumulative amplitude, and the current signal segment is directly summed to obtain the current cumulative amplitude.

[0143] Step 403: Generate a level difference amplitude according to the absolute value of the difference between the historical cumulative amplitude and the current cumulative amplitude.

[0144] In step 403 , the level difference amplitude is a quantitative indicator obtained by calculating the absolute difference between the historical and current cumulative amplitudes, reflecting the degree of deviation between the historical characteristics and the real-time characteristics of the level.

[0145] In the embodiment of the present application, the historical cumulative amplitude is subtracted from the current cumulative amplitude and the absolute value is taken to obtain the level difference amplitude.

[0146] Step 404: Compare the level difference amplitude with a preset difference threshold. If the level difference amplitude exceeds the preset difference threshold, activate the difference accumulator corresponding to the level.

[0147] In step 404, a preset difference threshold is set based on historical data analysis. Exceeding the threshold indicates that the difference requires special processing. The difference accumulator is a computing unit specifically used to process the difference signal.

[0148] In the embodiment of the present application, when the level difference amplitude exceeds a threshold, the level difference accumulator is activated to start recording and amplifying the difference signal; otherwise, the input signal of the difference accumulator is suppressed.

[0149] Step 405: Directionally mark the signal output by the activated difference accumulator.

[0150] In step 405, the signal refers to the processed quantized emotional feature data, specifically the feature vector generated by the difference accumulator after comparing historical and current emotional data. The activated difference accumulator can output a signal because when the hierarchical difference amplitude exceeds a threshold, a specific processing unit within the accumulator is triggered, automatically performing preset feature extraction and conversion operations on the input historical signal segments and the current signal segment, generating a structured output signal containing difference information. Directional marking refers to determining whether the strength of the historical feature is greater than or less than the current feature, marking it as positive or negative, respectively, to reflect the direction of emotional change.

[0151] In the embodiment of the present application, the historical and current cumulative amplitudes are compared, and a positive sign is given when the historical value is larger, and a negative sign is given otherwise.

[0152] Step 406: Perform polarity weighting on the signals output by the difference accumulators in the same level based on the directionality marking result to generate a level difference index.

[0153] In step 406 , polarity weighting is to apply a positive or negative weight coefficient to the difference signal according to the directional flag to generate a directional difference index.

[0154] In this embodiment of the present application, positive markers are weighted 1.2 times, and negative markers are weighted 0.8 times, generating a signed hierarchical difference index. For example, in a customer complaint handling scenario, when the cumulative amplitude of the current angry emotion marker (current signal segment) is detected to be higher than the historical calm emotion data (historical signal segment), the directionality is marked as "negative," and a negative weight coefficient of 0.8 is applied to the difference signal, generating a hierarchical difference index of -0.64, indicating worsening emotions. Conversely, if the current satisfaction level is higher than the historical level, it is marked as "positive" and assigned a positive weight of 0.6, generating a difference index of +0.52.

[0155] Step 407: The level difference index of each level is accumulated in a sliding window according to a preset aggregation rule to generate a level contribution parameter.

[0156] In step 407, the preset aggregation rule refers to the calculation method used when performing a sliding window accumulation of the hierarchical difference indices. Specifically, it includes three key parameters: window size setting, weight distribution method, and normalization process. These parameters are used to control the fusion method and influence of the difference indices of different hierarchies. Sliding window accumulation refers to the local summation of the difference indices of adjacent hierarchies in a time series, smoothing random fluctuations and highlighting major trends.

[0157] In the embodiment of the present application, a sliding window with a window size of 3 is used to sum and average the difference indices of three adjacent levels to generate the final contribution parameter.

[0158] Here's a specific example:

[0159] When a bank customer inquired about a wealth management product for the sixth time, the system segmented the seven-dimensional hierarchical fusion signal [0.3, 0.5, 0.4, 0.3, 0.3, 0.5, 0.3] according to the ratio of the historical five dimensions to the current two dimensions, resulting in three historical signal segments: [0.3, 0.5], [0.4, 0.3], and [0.3, 0.5], and the current signal segment [0.3]. The cumulative amplitudes of each historical signal segment were calculated to be 0.8, 0.7, and 0.8, respectively (calculated by adding the dimension values), and the cumulative amplitude of the current signal segment was 0.3. The hierarchical difference amplitudes were generated to be 0.5, 0.4, and 0.5 (calculated by subtracting the absolute value of the current amplitude from the historical amplitude). A threshold of 0.45 was set, the second layer was deactivated, and the difference accumulators for the first and third layers were activated (calculated by difference amplitude > threshold). The first layer's historical amplitude of 0.8 is greater than the current amplitude of 0.3, marking the positive direction. The same applies to the third layer, weighted by 1.2, to obtain difference indices of 0.6 and 0.6 (calculated by: difference amplitude × weighting coefficient). A sliding window with a window size of 2 is used for accumulation (calculated by: sum of adjacent indices divided by 2) to generate the contribution parameters [0.6, 0.4, 0.6].

[0160] In the embodiment of the present application, the method achieves a precise fusion of historical sentiment patterns and real-time features by quantitatively analyzing the differences in features at each level and highlighting changes, so that the scoring results can better reflect the real changing trends of customer emotions.

[0161] In order to solve the problem of quantifying the association between historical emotion data and the current emotion state, in some embodiments, step 203: calculating the target association strength value between the current emotion polarity label and each historical emotion segment based on the emotion polarity label distribution density includes:

[0162] Step 501: classify and count the emotion polarity labels in each historical emotion segment according to preset emotion categories, and generate an emotion category distribution histogram for each historical emotion segment.

[0163] In step 501, the emotional polarity labels in each historical emotional segment and the current emotional polarity labels are temporally correlated. The historical labels provide a reference for the customer's long-term emotional pattern, and the current labels are real-time data that need to be compared and analyzed. The two are emotional feature records of the same customer at different time points. The preset emotional category refers to a predefined emotional classification standard, which usually includes basic categories such as positive, neutral, and negative, and is used to standardize and classify emotional polarity labels to ensure the consistency of emotional analysis. The emotional category distribution histogram refers to a distribution chart formed by counting the various emotional polarity labels in the historical emotional segment according to the preset emotional categories. The horizontal axis represents the emotional category and the vertical axis represents the frequency of occurrence.

[0164] In an embodiment of the present application, the system first reads the original emotion label data stored in each historical emotion segment, classifies and organizes them according to the three preset emotion categories of "positive", "neutral" and "negative", counts the number of occurrences of each category in the segment, and then visualizes the statistical results in the form of a histogram, with the emotion categories arranged on the horizontal axis and the frequency of occurrence of the corresponding category on the vertical axis, forming an intuitive emotion distribution chart to provide a data basis for subsequent correlation analysis.

[0165] Step 502: performing category matching on the current emotion polarity label to determine the emotion category to which the current emotion polarity label belongs.

[0166] In step 502, category matching involves comparing the current emotion polarity label with a pre-defined library of emotion categories to determine the standard emotion category to which it belongs. Emotion categories refer to the specific categories to which the emotion polarity label belongs, such as "anger" or "satisfaction." These categories are used to quantitatively describe a customer's emotional state and are the basic unit of sentiment analysis.

[0167] In an embodiment of the present application, the system receives the emotional polarity label generated during the current interaction process, compares the similarity of the label with the standard definition in the preset emotion category library, and determines the emotion category to which it belongs through semantic analysis and feature matching. For example, the "anger" label is classified into the "negative" category, and the "satisfaction" is classified into the "positive" category, thereby completing the standardized classification of the current emotional state.

[0168] Step 503: Calculate the frequency ratio of the emotion category in each emotion category distribution histogram.

[0169] In step 503 , the frequency of occurrence refers to the frequency of the current emotion category appearing in the historical segment, which is calculated as the number of occurrences of the category divided by the total number of tags in the segment.

[0170] In an embodiment of the present application, based on the determined current emotion category, the system scans the emotion category distribution histogram of each historical segment, extracts the frequency data of the category in each histogram, and then divides it by the total number of emotion tags of the corresponding segment to calculate the relative occurrence ratio of the emotion category in each historical segment and quantify the historical emotion distribution characteristics.

[0171] Step 504: Generate an initial association strength value according to the degree of deviation between the occurrence frequency ratio and a preset reference frequency.

[0172] In step 504, the preset baseline frequency refers to the average frequency of occurrence of a certain emotion category in historical data, serving as a reference for calculating the association strength. This frequency is typically derived from historical data analysis. The degree of deviation refers to the difference between the frequency of the current emotion category and the preset baseline frequency. It measures the deviation of the current emotional state from the historical norm; a greater difference indicates a stronger association. The initial association strength value is generated by comparing the current emotion's proportion relative to the baseline frequency; a greater deviation indicates a stronger association.

[0173] In an embodiment of the present application, the system pre-sets the baseline occurrence frequency of each type of emotion as a reference value, compares the calculated actual proportion with the baseline value, and calculates the relative deviation between the two. The greater the deviation, the higher the initial association strength value generated, indicating that the current emotional state is more strongly associated with the historical segment.

[0174] Step 505: Based on the time decay coefficient of each historical emotion segment, the initial association strength value is corrected to generate a target association strength value.

[0175] In step 505, a time decay coefficient is calculated based on the interval between the recording time of the historical emotion segment and the current time. An exponential decay function is typically used, with the decay coefficient decreasing as the interval increases. Correction processing involves calculating a time decay coefficient for each historical emotion segment individually, meaning each segment has its own independent time decay coefficient used to correct its corresponding initial association strength value.

[0176] In an embodiment of the present application, the system calculates the time attenuation coefficient based on the collection time of each historical segment. The closer the segment is to the time, the smaller the attenuation. Then, the obtained initial association strength value is multiplied by the corresponding attenuation coefficient to appropriately weaken the long-term association strength. Finally, the target association strength value considering timeliness is obtained, and the timing correction is completed.

[0177] Here's a specific example:

[0178] When a bank customer inquires about a wealth management product for the fifth time, the system divides the past four inquiries into four historical emotional segments by month. The percentage of "negative" sentiment in each segment is calculated, with the percentage being 50%, 60%, 70%, and 80%, respectively. This time, the customer's expression "getting worse" (labeled "negative") and a narrowed-eye expression (labeled "suspicious") are detected and fused to generate the current label "negative-suspicious." The frequency percentage of the "negative" category in the four historical segments is calculated (step 503). Comparing this with the baseline value of 40%, the initial correlation strengths are: (50-40) / 40 = 0.25 for the first segment, (60-40) / 40 = 0.5 for the second segment, (70-40) / 40 = 0.75 for the third segment, and (80-40) / 40 = 1.0 for the fourth segment (step 504). The decay coefficients are calculated based on the time intervals (4, 3, 2, and 1 month): 1-0.1×4=0.6, 1-0.1×3=0.7, 1-0.1×2=0.8, and 1-0.1×1=0.9 (step 505). The corrected target association strengths are: 0.25×0.6=0.15, 0.5×0.7=0.35, 0.75×0.8=0.6, and 1.0×0.9=0.9.

[0179] In the embodiment of the present application, the method achieves accurate calculation of the intensity of emotional association by quantifying the degree of matching between historical emotional distribution and current state and taking into account the time decay effect, thereby providing a reliable time series reference basis for emotional scoring.

[0180] To accurately match service strategies with customer emotional states, in some embodiments, step 105 , when the customer emotional score is lower than a preset service baseline, generates service strategy adjustment information that matches the customer's current emotional state and historical emotional change trends, including:

[0181] Step 601: Based on the customer emotion evolution graph, establish an emotion polarity label change curve of N interaction cycles, where N is greater than or equal to 2.

[0182] In step 601, the emotion polarity label change curve refers to a continuous change curve with time as the horizontal axis and the emotion polarity label value as the vertical axis, which reflects the temporal evolution trend of the customer's emotional state. N interaction cycles are set according to business needs, usually 3-6 cycles. Establishing an emotion polarity label change curve for N interaction cycles means establishing a curve that is connected by multiple emotion polarity label data points within N interaction cycles. That is, the number of curves is one, but the label data points that constitute the curve are multiple (corresponding to N cycles).

[0183] In an embodiment of the present application, the system reads the emotional polarity label data of the customer's most recent five interactions, draws a continuous change curve, and marks obvious peaks and troughs.

[0184] Step 602: Match the emotion polarity label corresponding to the customer's current emotion state with the key turning point in the emotion polarity label change curve to determine the stage position of the customer's current emotion state in the historical emotion change trend.

[0185] In step 602, key turning points are defined as the points in the slope of the sentiment polarity label change curve where the slope changes. These points are automatically identified by calculating where the second-order derivative of the curve exceeds a preset threshold (e.g., a rate of change > 15%), marking the turning point in the sentiment trend. A stage position refers to the relative position of the current sentiment state within the historical trend, such as an upward, downward, or stable period.

[0186] In the embodiment of the present application, the current emotion label is compared with the turning point on the curve to determine that it is in the third stage of "continuous decline period".

[0187] Step 603: Select a corresponding strategy combination in a preset strategy template library according to the stage position.

[0188] In step 603, the preset strategy template library contains standardized response plans for different emotional stages. The strategy combination is an organic combination of multiple strategies selected according to the characteristics of the stage. The strategy combination specifically includes: the degree of urgency of the tone (such as adopting a more moderate tone when angry), the proportion of professional terms used (such as reducing the use of technical terms for dissatisfied customers), and the frequency of inserting emotional resonance sentences (such as inserting one empathetic sentence for every three sentences). The service priority adjustment plan specifically includes: the response time level (such as increasing from 24 hours to 2 hours), the processing personnel level (such as upgrading from general customer service to supervisor), and the compensation plan authority (such as the ability to provide higher compensation amounts).

[0189] In the embodiment of the present application, a combination strategy of "quick response + compensation plan + upgrade service" is selected for the "continuous decline period".

[0190] Step 604: Compare the customer sentiment score with a preset service baseline.

[0191] In step 604, the preset service baseline is a sentiment score threshold set based on historical service data, and the comparison result reflects the degree of abnormality in the current sentiment state. The comparison result indicates that when the comparison result exceeds a preset first threshold, a primary service strategy adjustment is triggered, and when the comparison result exceeds a preset second threshold, a high-level service strategy adjustment is triggered.

[0192] In the embodiment of the present application, the current score is compared with the baseline value, and a difference result of 15 points below the baseline is obtained.

[0193] Step 605: Based on the comparison result, parameter enhancement processing is performed on the policy combination to generate service policy adjustment information.

[0194] In step 605, parameter enhancement is performed to adjust the policy execution intensity based on the score difference, such as shortening response time or increasing compensation. For example, if the customer sentiment score (0.3) falls below the service baseline (0.5) by a 40% difference, the original "apology + coupon" policy combination is enhanced: the coupon value is increased from 50 yuan to 50 × (1 + 40%) = 70 yuan, and the priority of senior customer service intervention is increased. The enhancement magnitude is calculated as (baseline value - score value) / baseline value = 40%.

[0195] In the embodiment of the present application, parameter enhancement processing is to adjust the strategy execution intensity according to the score difference, such as shortening the response time, increasing the compensation strength, etc.

[0196] Here's a specific example:

[0197] When a bank customer inquired about a wealth management product again, the system created a sentiment curve for the customer's last four interactions, showing a gradual increase in the proportion of "negative" sentiment from 50% to 70%. The system detected the customer's statement of "totally unacceptable" (labeled "negative") and fist-clenching gesture (labeled "angry"), merging them to generate a "negative-angry" label. Matching the curve revealed a "sharp deterioration" phase. The system selected a strategy combination of "senior manager intervention + emergency compensation + product adjustment." The current sentiment score was 53 (calculated by: 100 × (1 - 0.8 × 0.6), a difference of 17 points from the baseline score of 70. Based on this difference, the strategy was strengthened to: "video communication with the president within two hours + an 8% increase in earnings compensation (original 5% + difference × 0.2%) + product plan redesign." The compensation increase was calculated as: base compensation 5% + (70 - 53) × 0.2% = 8.4%, rounded up to 8%.

[0198] In the embodiment of the present application, the method achieves accurate matching and dynamic optimization of service strategies by analyzing sentiment change trends and quantifying score differences, effectively improving the pertinence and timeliness of customer service.

[0199] To solve the problem of dynamic optimization of service strategies, in some embodiments, after generating service strategy adjustment information that matches the customer's current emotional state and historical emotional change trends in step 105 , the method further includes:

[0200] Step 701: According to the service strategy adjustment information, the service content display order in the interaction process between the target customer and the enterprise is adjusted in real time, and the emotional tendency label in the customer emotional evolution map is updated.

[0201] In step 701, the service content display order refers to the priority of the interactive interface elements rearranged according to the strategy adjustment result. The emotional tendency label is the latest emotional state mark recorded in the customer's emotional evolution map.

[0202] In an embodiment of the present application, the system adjusts the information according to the strategy, displays the compensation plan and the advanced customer service entrance at the top, and updates the emotional label in the graph to the "negative-angry-processing" status.

[0203] Step 702: After the customer completes the current interaction, a closed-loop feedback signal is generated based on the adjusted service content display order and the changing trend of the customer's emotional score. The closed-loop feedback signal is used to iteratively optimize the mapping relationship between the emotional calculation model and the service strategy adjustment information.

[0204] In step 702, the closed-loop feedback signal is a comprehensive evaluation indicator that records the effectiveness of the strategy execution, including data such as the change in sentiment score and the click-through rate of service content, and is used to optimize the model strategy mapping relationship.

[0205] In an embodiment of the present application, the system tracks the entire customer interaction process, records the changes in sentiment scores and service option clicks after the strategy is executed, and generates a feedback signal including an effect evaluation.

[0206] Here's a specific example:

[0207] When a bank customer was upset over a financial product yield issue, the system implemented a modified "wealth manager follow-up within 24 hours + 5% yield compensation increase" strategy. The system immediately moved the "Dedicated Wealth Manager" portal and the "Yield Compensation Application" button to the top of the mobile banking screen (originally located on the third screen). The label on the customer's sentiment evolution graph was also updated to "Negative - Strong - Resolved." Within 24 hours, after the wealth manager completed the follow-up visit and provided a compensation plan, the system detected an increase in the customer's sentiment score from 58 to 68 (calculated using the formula: 100 × (1 - 0.32), where sentiment similarity was reduced to 0.6 and the time weighting factor was 0.8). The system also recorded three clicks on the "Yield Compensation" button and two clicks on the "Wealth Manager" portal. This closed-loop feedback signal, consisting of a 10-point score increase (68-58) and an 83% click-through rate (5 clicks / 6 impressions), was used to optimize the sentiment computing model's strategy matching weights for "negative-strong" sentiment.

[0208] In an embodiment of the present application, the method achieves closed-loop optimization of service strategies by adjusting service presentation in real time and continuously collecting feedback, enabling the system to continuously improve the accuracy of emotion recognition and service recommendations.

[0209] Figure 2 This is a schematic diagram of the structure of an intelligent customer lifecycle management AiCRM system provided in an embodiment of the present application, such as Figure 2 As shown, the system includes:

[0210] The acquisition module 21 is used to obtain the target customer's voice call records, real-time facial expression images and text feedback data during the interaction between the target customer and the enterprise.

[0211] The analysis module 22 is used to perform semantic analysis on the voice call records and the text feedback data, perform dynamic feature recognition on the facial expression images, and perform cross-modal association between the semantic analysis results and the dynamic feature recognition results to generate a customer emotion feature set.

[0212] The updating module 23 is configured to dynamically update the customer emotion evolution graph in the preset multimodal customer emotion database based on the customer emotion feature set.

[0213] The first generating module 24 is configured to generate a target customer emotion score of the target customer using an emotion calculation model based on the updated customer emotion evolution graph and the customer emotion feature set.

[0214] The second generation module 25 is used to generate service strategy adjustment information that matches the customer's current emotional state and historical emotional change trends when the customer's emotional score is lower than the preset service baseline, so as to realize intelligent customer lifecycle management AiCRM.

[0215] Figure 2 The intelligent customer lifecycle management AiCRM system can execute Figure 1 The implementation principles and technical effects of the intelligent customer lifecycle management AiCRM method described in the illustrated embodiment are not further elaborated. The specific manner in which each module and unit performs operations in the intelligent customer lifecycle management AiCRM system in the above embodiment has been described in detail in the relevant embodiment of the method and will not be elaborated on here.

[0216] In one possible design, Figure 2 An intelligent customer lifecycle management AiCRM system of the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0217] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0218] The processing component 32 is as follows Figure 1 The embodiment provides an intelligent customer lifecycle management AiCRM method.

[0219] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0220] The storage component 31 is configured to store various types of data to support operations in the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random-access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0221] Of course, a computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0222] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0223] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0224] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0225] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The embodiment shown is an intelligent customer lifecycle management AiCRM method.

[0226] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0227] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0228] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0229] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent customer lifecycle management AiCRM method, characterized by: include: During the interaction between target customers and enterprises, obtain target customers' voice call records, real-time facial expression images and text feedback data; Performing semantic analysis on the voice call records and the text feedback data, performing dynamic feature recognition on the facial expression images, and cross-modally correlating the semantic analysis results with the dynamic feature recognition results to generate a customer emotion feature set; Based on the customer emotion feature set, dynamically updating the customer emotion evolution graph in the preset multimodal customer emotion database; Based on the updated customer emotion evolution graph and the customer emotion feature set, a sentiment computing model is used to generate a target customer emotion score for the target customer; When the customer's emotional score is lower than the preset service baseline, service strategy adjustment information is generated to match the customer's current emotional state and historical emotional change trends to achieve intelligent customer lifecycle management AiCRM.

2. The method according to claim 1, characterized in that The method of generating a target customer emotion score for the target customer using an emotion calculation model based on the updated customer emotion evolution graph and the customer emotion feature set includes: Dividing the updated customer emotion evolution graph into multiple historical emotion segments according to a preset time window, and extracting the emotion polarity label distribution density corresponding to the voice call record, the real-time facial expression image, and the text feedback data in each historical emotion segment; Discretizing and marking the current emotion polarity label in the customer emotion feature set; Calculating a target association strength value between the current emotion polarity label and each historical emotion segment based on the emotion polarity label distribution density; Assigning a weight value to each historical emotion segment according to the target association strength value, and linearly superimposing the weight value with the discretization labeling result to generate an emotion state fusion vector; The emotional state fusion vector is input into the emotional computing model to output the target customer emotional score.

3. The method according to claim 2, characterized in that The step of inputting the emotional state fusion vector into an emotional computing model and outputting a target customer emotional score includes: By using an emotion calculation model, the emotion state fusion vector is divided into a historical emotion weight component and a current emotion tag component according to a preset dimension; Performing hierarchical processing on the historical sentiment weight components, and adjusting the superposition coefficient of each layer according to the target association strength value in each processing layer; Interlacing and splicing the adjusted superposition coefficient with the current emotion tag component to generate a hierarchical fusion signal; Using the dynamic aggregation nodes of the sentiment computing model, the difference between the historical sentiment weight component and the current sentiment label component in the hierarchical fusion signal of each level is compared layer by layer to generate the hierarchical contribution parameter; selectively activating transmission paths in the dynamic aggregation node according to the hierarchical contribution parameter; Perform scalar compression on the output signal of the activated transmission path to generate an initial customer sentiment score; Based on the statistical deviation of the distribution density of the sentiment polarity labels, boundary constraints are performed on the initial customer sentiment scores, and the target customer sentiment scores are output.

4. The method according to claim 3, characterized in that The dynamic aggregation node of the emotion calculation model is used to compare the difference between the historical emotion weight component and the current emotion label component in the hierarchical fusion signal of each level layer by layer to generate the hierarchical contribution parameter, including: Split the hierarchical fusion signal of each level into historical signal segments and current signal segments according to a preset ratio; Performing amplitude accumulation calculation on the historical signal segment and the current signal segment respectively to obtain a historical cumulative amplitude and a current cumulative amplitude; generating a level difference amplitude according to an absolute value of a difference between the historical cumulative amplitude and the current cumulative amplitude; Comparing the level difference amplitude with a preset difference threshold, and activating a difference accumulator corresponding to the level if the level difference amplitude exceeds the preset difference threshold; Directionally mark the signal output by the activated difference accumulator; Based on the directional marking results, polarity weighting is performed on the signals output by the difference accumulators in the same level to generate a level difference index; The hierarchical difference index of each level is accumulated in a sliding window according to the preset aggregation rules to generate the hierarchical contribution parameter.

5. The method according to claim 2, characterized in that The calculating, based on the distribution density of the emotion polarity label, the target association strength value between the current emotion polarity label and each historical emotion segment includes: The emotional polarity labels in each historical emotional segment are classified and counted according to the preset emotional categories, and the emotional category distribution histogram of each historical emotional segment is generated; Performing category matching on the current emotion polarity label to determine the emotion category to which the current emotion polarity label belongs; Calculate the frequency of occurrence of the emotion category in each emotion category distribution histogram; Generating an initial correlation strength value according to the degree of deviation between the occurrence frequency ratio and a preset benchmark frequency; Based on the time decay coefficient of each historical emotion segment, the initial association strength value is corrected to generate a target association strength value.

6. The method according to claim 1, wherein When the customer's emotion score is lower than the preset service baseline, generating service strategy adjustment information that matches the customer's current emotion state and historical emotion change trend, including: According to the customer emotion evolution graph, establish an emotion polarity label change curve of N interaction cycles, where N is greater than or equal to 2; Matching the emotion polarity label corresponding to the customer's current emotion state with the key turning point in the emotion polarity label change curve to determine the stage position of the customer's current emotion state in the historical emotion change trend; Selecting a corresponding strategy combination in a preset strategy template library according to the stage position; comparing the customer sentiment score to a preset service baseline; According to the comparison result, parameter enhancement processing is performed on the policy combination to generate service policy adjustment information.

7. The method according to claim 1, characterized in that After generating service strategy adjustment information that matches the customer's current emotional state and historical emotional change trends, the method further includes: According to the service strategy adjustment information, the service content display order in the interaction process between the target customer and the enterprise is adjusted in real time, and the emotional tendency label in the customer emotional evolution map is updated; After the customer completes the current interaction, a closed-loop feedback signal is generated based on the adjusted service content display order and the changing trend of the customer's emotional score. The closed-loop feedback signal is used to iteratively optimize the mapping relationship between the emotional computing model and the service strategy adjustment information.

8. An intelligent customer lifecycle management AiCRM system, characterized by: include: The acquisition module is used to obtain the target customer's voice call records, real-time facial expression images and text feedback data during the interaction between the target customer and the enterprise; An analysis module is configured to perform semantic analysis on the voice call records and the text feedback data, perform dynamic feature recognition on the facial expression images, and perform cross-modal correlation between the semantic analysis results and the dynamic feature recognition results to generate a customer emotion feature set; An updating module, configured to dynamically update a customer emotion evolution graph in a preset multimodal customer emotion database based on the customer emotion feature set; A first generating module is configured to generate a target customer sentiment score of the target customer using a sentiment computing model based on the updated customer sentiment evolution graph and the customer sentiment feature set; The second generation module is used to generate service strategy adjustment information that matches the customer's current emotional state and historical emotional change trends when the customer's emotional score is lower than the preset service baseline, so as to realize intelligent customer lifecycle management AiCRM.

9. A computing device, characterized in that It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an intelligent customer lifecycle management AiCRM method as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, an intelligent customer lifecycle management AiCRM method according to any one of claims 1 to 7 is implemented.

Images