Insurance customer service optimization system based on artificial intelligence

Abstract

The invention discloses an insurance customer service optimization system based on artificial intelligence, and relates to the technical field of intelligent customer service, and the system comprises a data collection and analysis module which collects the multi-mode interaction data of a customer in real time, extracts the fundamental frequency change rate and MFCC coefficient of voice voiceprint, and calculates an emotion polarity value; meanwhile, analyzing the semantic content of the text to obtain an intention classification label and an emergency degree score, and detecting the coordinates of a text missing region in the image and the micro-expression action frequency in the video stream; and the closed-loop optimization module is used for pushing an audit conclusion and the revised visual decision graph to the client, synchronously returning feedback data, and dynamically optimizing dynamic weight distribution model parameters and causal knowledge graph clause node connection weights. According to the method, accurate quantification of customer emotion and service emergency degree is realized through multi-modal data fusion and PPO dynamic weight distribution.

Description

Technical Field

[0001] The present invention relates to the field of intelligent customer service technology, and in particular to an insurance customer service optimization system based on artificial intelligence. Background Art

[0002] Insurance customer service systems are gradually incorporating artificial intelligence (AI) technologies to improve service efficiency, with significant progress particularly in multimodal data processing. Existing technologies often use static rule engines to classify service requests, combined with natural language processing algorithms to analyze text intent and image recognition to detect document integrity. Some advanced systems have attempted to integrate speech sentiment analysis technology, extracting speech features through Mel-Frequency Cepstral Coefficients (MFCCs) and combining them with convolutional neural networks for emotion recognition. In terms of service routing, dynamic scheduling methods based on reinforcement learning are being used to optimize resource allocation, while knowledge graph technology is being incorporated into the clause matching process to enhance logical reasoning capabilities.

[0003] However, existing technologies still have significant defects in multimodal data fusion and dynamic decision-making: first, sentiment analysis, text parsing and image detection modules mostly use independent weight distribution mechanisms, which cannot dynamically adjust the contribution of each modal data according to the interaction scenario, resulting in deviations in service priority determination; second, the clause matching process relies on a static knowledge graph traversal algorithm, which makes it difficult to capture the logical conflicts between user supplementary materials and the original chain of evidence in real time, which easily leads to inconsistencies in the review conclusions. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an insurance customer service optimization system based on artificial intelligence to solve the problems of low efficiency of multimodal data fusion and insufficient accuracy of dynamic decision-making in existing insurance customer service systems.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] The present invention provides an insurance customer service optimization system based on artificial intelligence, which includes: a data acquisition and analysis module, which collects customers' multimodal interaction data in real time, extracts the fundamental frequency change rate and MFCC coefficient of the voiceprint and calculates the emotional polarity value, and simultaneously analyzes the text semantic content to obtain the intention classification label and the urgency score, and detects the coordinates of the text missing area in the image and the frequency of micro-expression movement in the video stream; a weight dynamic calculation module, which inputs the emotional polarity value, the urgency score, the coordinates of the text missing area and the frequency of micro-expression movement into a dynamic weight allocation model based on the PPO algorithm, and outputs the emotional weight value of the service request; an evidence chain generation module, which allocates the service request to the automated review queue based on the emotional weight value, and matches the intention classification label with the pre-built causal knowledge graph through the graph neural network. The node path of the clause in the spectrum is used to generate an evidence chain containing the coordinates of the conflicting clause nodes; the interactive visualization module converts the evidence chain into a visual decision map, superimposes the coordinates of the text missing area to form a heat map, and triggers the interactive interface to receive the supplementary materials uploaded by the user for the heat map nodes; the consistency verification module compares the pixel difference between the supplementary materials and the original uploaded materials, verifies the consistency of the supplementary text content with the logic of the causal knowledge graph nodes, and updates the audit conclusion when the pixel difference and the node logic conflict are both lower than the corresponding preset thresholds. Otherwise, it is marked as abnormal and triggers manual review; the closed-loop optimization module pushes the audit conclusion and the revised visual decision map to the client, synchronously returns the feedback data, and dynamically optimizes the dynamic weight allocation model parameters and the causal knowledge graph clause node connection weights.

[0008] As a preferred solution of the artificial intelligence-based insurance customer service optimization system of the present invention, the extraction of the fundamental frequency change rate and MFCC coefficient of the voiceprint includes the following steps:

[0009] Calculate the absolute value of the difference between adjacent frames of the fundamental frequency in the speech signal and calculate the average rate of change of the fundamental frequency within the preset time window;

[0010] Perform frequency domain energy distribution analysis on the speech signal through the Mel filter bank and extract MFCC coefficients;

[0011] The mean of the fundamental frequency change rate and the MFCC coefficient are Z-score normalized to generate the speech emotion feature vector.

[0012] As a preferred solution of the artificial intelligence-based insurance customer service optimization system of the present invention, the step of parsing the text semantic content to obtain the intent classification label includes the following steps:

[0013] Use the pre-trained BERT model to encode the text into word vectors and generate contextual semantic representations;

[0014] Input the contextual semantic representation into the bidirectional LSTM network to capture long-distance dependencies and output temporal features;

[0015] The time series features are weightedly fused through the attention mechanism to generate attention-weighted context features, and the Softmax classifier is used to output the intent classification label and confidence score.

[0016] As a preferred solution of the artificial intelligence-based insurance customer service optimization system of the present invention, the dynamic weight allocation model includes the following steps:

[0017] Initialize the policy network of the PPO algorithm and define the state space as the joint features of the sentiment polarity value, urgency score, coordinates of the text missing area, and micro-expression action frequency;

[0018] Construct a reward function as a weighted combination of service request processing efficiency and customer satisfaction score;

[0019] The network parameters of the strategy are optimized through importance sampling, and the distribution ratio of the sentiment weight values ​​is dynamically adjusted.

[0020] As a preferred solution of the artificial intelligence-based insurance customer service optimization system of the present invention, the generating of the evidence chain containing the node coordinates of the conflict clauses includes the following steps:

[0021] In the graph neural network, the knowledge graph clause nodes are defined as clause semantic vectors, and the edges are defined as the logical association strength of clauses.

[0022] Calculate the multi-hop path similarity between the intent classification label and the clause node through the graph attention mechanism;

[0023] Path nodes with similarity higher than the preset similarity threshold are selected to generate an evidence chain containing node coordinates and associated weights.

[0024] As a preferred solution of the artificial intelligence-based insurance customer service optimization system of the present invention, the converting of the chain of evidence into a visual decision map and superimposing the coordinates of the text-missing area to form a heat map includes the following steps:

[0025] Map the nodes of the evidence chain into interactive elements in a three-dimensional topological graph;

[0026] Perform kernel density estimation on the coordinates of the text missing area to generate a semi-transparent heat map covering the original material;

[0027] The spatial overlay display of heat map and topology map is realized through the WebGL rendering engine.

[0028] As a preferred solution of the artificial intelligence-based insurance customer service optimization system described in the present invention, the verification of the supplementary materials includes calculating the difference measure between the supplementary content and the original materials through image similarity evaluation, and analyzing the logical consistency between the supplementary text and the knowledge graph clause nodes in combination with the graph structure traversal algorithm. When the difference measure is lower than the critical value of the image difference measure and the logical consistency meets the conditions, the verification is determined to be passed.

[0029] As a preferred solution of the artificial intelligence-based insurance customer service optimization system described in the present invention, the interactive interface adopts a multimodal input protocol driven by a temporal memory network and supports uploading supplementary materials in structured data format.

[0030] As a preferred solution of the artificial intelligence-based insurance customer service optimization system described in the present invention, the connection relationship between the clause nodes of the causal knowledge graph is defined by the logical relationship through the semantic ontology framework, and the connection weights between the clause nodes are dynamically optimized based on historical service data.

[0031] As a preferred solution of the artificial intelligence-based insurance customer service optimization system described in the present invention, the policy network of the PPO algorithm adopts an adaptive learning rate mechanism to update parameters, and at the same time explores the intensity of the strategy through KL divergence regularization constraint.

[0032] The beneficial effects of the present invention are: through multimodal data fusion and PPO dynamic weight allocation, accurate quantification of customer emotions and service urgency can be achieved; combined with graph neural networks and causal knowledge graphs, clause conflicts can be dynamically identified and a visual evidence chain can be generated to ensure the compliance of service decisions; based on heat map interaction and logical consistency verification, traceability of material supplementation and abnormal blocking can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 Block diagram of a system for optimizing AI-based insurance customer service.

[0035] Figure 2 Flowchart for multimodal data collection and analysis.

[0036] Figure 3 Generate a flow chart for the chain of evidence.

[0037] Figure 4 This is a flow chart for consistency verification. DETAILED DESCRIPTION

[0038] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0039] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0040] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0041] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides an insurance customer service optimization system based on artificial intelligence, including the following steps:

[0042] The data acquisition and analysis module collects customers' multimodal interaction data in real time, extracts the fundamental frequency change rate and MFCC coefficient of the voiceprint and calculates the emotional polarity value. At the same time, it analyzes the semantic content of the text to obtain the intent classification label and urgency score, and detects the coordinates of the text missing area in the image and the frequency of micro-expression movements in the video stream.

[0043] Deploy multi-channel audio acquisition equipment to capture customer voice signals in real time, and use the Mel-frequency cepstral coefficient algorithm to extract the 39-dimensional feature vector of the voice signal.

[0044] Specifically, the absolute value of the difference between adjacent frames of the fundamental frequency in the speech signal is calculated, and the mean of the fundamental frequency change rate within the preset time window is counted; the frequency domain energy distribution of the speech signal is analyzed through the Mel filter group to extract the MFCC coefficients; the mean of the fundamental frequency change rate and the MFCC coefficients are Z-score normalized to generate the speech emotion feature vector.

[0045] Optimally, the continuous speech signal is divided into short time segments (time windows) of fixed length, with a fixed step size set between adjacent windows to preserve continuity. A window function that suppresses spectral leakage is used to weight each signal segment to reduce frequency domain distortion caused by truncation effects. The fundamental frequency value is calculated using a signal periodicity detection algorithm, and the statistical characteristics of the fundamental frequency changes in adjacent windows are analyzed. The fundamental frequency search range is set to cover the typical frequency range of human speech to eliminate interference from outliers. A filter bank covering the audible frequency range of the human ear is constructed. After converting the speech signal to the frequency domain, the coefficients representing the timbre characteristics are extracted through the energy distribution of the filter bank. The coefficients are dynamically compressed and normalized to generate a robust feature vector.

[0046] Example: For a 5-second speech segment, with a frame length of 10 ms, calculate the standard deviation of the fundamental frequency as the rate of change indicator.

[0047] Input the speech emotion feature vector into the pre-trained SVM classifier and output the emotion score (in the range of [-1, 1]);

[0048] The sentiment polarity value is calculated by combining the text sentiment dictionary (detecting the proportion of negative words) with weighted calculation.

[0049] Example: The speech anger sentiment score is 0.8, the text contains 30% negative words, and the sentiment polarity value is 0.62.

[0050] The text semantic content is processed through a bidirectional long short-term memory network model, and the intent classification label (including three categories: claim application, terms consultation, and policy cancellation request) is output through a classifier pre-trained by domain adaptability.

[0051] Specifically, a bidirectional recurrent neural network is used to process text sequences containing semantic content. The input is a sequence of word vectors, which are used to generate hidden states by capturing contextual dependencies. An attention mechanism is applied to the hidden states to weightedly aggregate key vocabulary information and output a classification probability distribution. A word vector sequence is a numerical representation formed by mapping each word in the text into a dense vector of fixed dimension using a pre-trained word embedding model (such as Word2Vec, GloVe, or BERT), and then arranging them in their original word order.

[0052] A time-sensitive vocabulary is predefined based on the classification probability distribution. The number and position of matching words in the text are counted. Time-sensitive keywords in the text (such as "urgent," "expedited," "within 24 hours," and other time-sensitive expressions) are detected by combining a keyword matching algorithm with semantic analysis. A basic urgency value is preset in combination with the intent classification label to generate an urgency score.

[0053] A residual convolutional neural network is used to extract image features, and the candidate box coordinates are generated through the region proposal network. The non-maximum suppression algorithm is used to screen out the candidate box coordinates with the highest confidence as the missing area positioning result; at the same time, based on the facial muscle motion coding system, the activation state of specific action units is tracked, and the triggering frequency per unit time is counted to generate a micro-expression frequency index.

[0054] The weight dynamic calculation module inputs the emotional polarity value, urgency score, text missing area coordinates and micro-expression action frequency into the dynamic weight distribution model based on the PPO algorithm, and outputs the emotional weight value of the service request.

[0055] It should be noted that the policy network constructed includes a multilayer perceptron. The input layer receives a four-dimensional feature vector, the hidden layer uses an activation function to prevent gradient vanishing, and the output layer generates weight assignment probabilities through normalization. The proximal policy optimization algorithm (PPO) is used to update the network parameters.

[0056] The emotional polarity values ​​and urgency scores were normalized, the coordinates of the missing text areas were converted into missing area percentages, and the frequencies of micro-expression movements were logarithmically transformed.

[0057] A dynamic weight allocation model is constructed based on the proximal strategy optimization algorithm, and the state space is set as a four-dimensional standardized feature vector (emotion polarity, urgency, missing area, and micro-expression frequency), and the action space is set as a weight allocation vector.

[0058] The reward function is constructed to include two indicators: service efficiency improvement and customer satisfaction. The policy network outputs a weight distribution scheme. For example, the weight vector = [0.4, 0.3, 0.2, 0.1] indicates that the sentiment weight accounts for 40%.

[0059] The policy network synchronizes parameters every 1000 iterations, and sets the discount factor and entropy coefficient to optimize convergence efficiency. For example, during training, the discount factor is set to 0.99 and the entropy coefficient is set to 0.01.

[0060] The optimal approach employs a dual-evaluation network architecture to enhance policy stability. This collaborative mechanism, where a primary network evaluates state value and a secondary network evaluates action value, improves decision-making logic. Prioritized empirical sampling and adaptive entropy adjustment strategies are introduced during training, complemented by gradient-constrained techniques to balance exploration and efficiency. Simultaneously, an interpretable analysis component is deployed, providing interpretable outputs such as feature importance quantification and decision path tracing to ensure transparency.

[0061] The evidence chain generation module assigns service requests to the automated review queue based on the sentiment weight value, matches the intent classification label with the clause node path of the pre-built causal knowledge graph through the graph neural network, generates an evidence chain containing the coordinates of the conflicting clause nodes, builds a graph neural network model, maps the intent classification label to the root node of the causal knowledge graph, and uses the breadth-first search algorithm to traverse the clause node path.

[0062] It should be noted that graph neural network refers generally to the algorithm type of graph neural network (GNN), and does not specifically refer to specific instances usually used for functional description; building a graph neural network model specifically emphasizes the specific model construction process for the insurance terms knowledge graph (such as architecture design, parameter training, domain adaptation, etc.).

[0063] Specifically, the system abstracts insurance clauses into nodes, establishes node associations through semantic similarity calculation and historical data analysis, and forms a weighted directed graph structure. Starting from the root node, a graph traversal algorithm is used, combined with an attention mechanism to evaluate the weights of clause node paths, selecting the optimal clause node path that matches the intent label. Logically conflicting nodes in the clause node path are detected in real time and marked.

[0064] For example, when a customer submits an application for "car accident medical claim", starting from the root node, the attention mechanism is combined to evaluate the weights of each clause path, and the high-weight path (0.9) of "car insurance main clause → medical expense coverage" is prioritized. At the same time, it is automatically detected that the "medical expense coverage" clause in this path requires a second-level or above hospital qualification, and the clinic invoice provided by the customer does not meet this requirement. The conflicting node is immediately marked and highlighted in the visual interface, prompting the customer to supplement the compliant hospital certificate, realizing intelligent clause verification and conflict warning.

[0065] The Jaccard coefficient is used to calculate the strength of the association between clause nodes, match the optimal path, and detect conflicting clause nodes. For example, a claim application intends to trigger the path from "accidental injury insurance → medical expense clause → deductible provision," and a geographical conflict is detected between "deductible" and "contracted hospital." For each conflicting clause node detected, an evidence chain is generated, including the coordinates of the conflicting node and the associated clause content.

[0066] The evidence chain generation module optimally implements the dynamic evolution of the knowledge graph. New terms are adaptively embedded using a comparative learning framework, associations are mined based on inductive graph learning, and conflict rules are automatically generated using a pattern mining algorithm. To optimize query performance, a hierarchical index structure, an approximate nearest neighbor search scheme, and a parallel traversal engine are built to support efficient access to large-scale graphs.

[0067] The interactive visualization module converts the chain of evidence into a tree-like visual decision map, in which the node size dynamically reflects the importance of the clauses, and the edge weights intuitively display the strength of logical associations. At the same time, the coordinates of the text-missing areas are superimposed to generate a heat map, and the module receives supplementary materials uploaded by users for the heat map nodes through an interactive interface.

[0068] It should also be noted that the control logic for node size and edge attributes (term importance and association strength) is as follows: Node size is dynamically adjusted based on term importance, and association strength determines edge thickness and color depth. Dynamic interactive operations within three-dimensional spatial layouts are supported. Spatial density analysis is performed on the coordinates of missing regions, generating a semi-transparent overlay. High-density areas are highlighted with a low transparency.

[0069] Among them, the spatial density analysis of the missing area coordinates specifically refers to using the kernel density estimation algorithm to calculate the distribution density of all detected missing area coordinates on the original image plane, constructing a Gaussian distribution function with each coordinate point as the center, and superimposing the density contributions of all points to generate a continuous probability density field, thereby intuitively reflecting the degree of aggregation of text missing in the image.

[0070] Specifically, the node size is dynamically adjusted according to the importance of the clause, and the importance score is calculated based on the citation frequency and conflict probability in historical service data; the edge weight is calculated through a graph neural network to reflect the causal relationship strength between clauses, and the weight value range is limited to [0,1].

[0071] Among them, historical service data includes: intent classification labels, sentiment polarity values, review conclusion feedback, clause path access frequency and conflict marking records in customer service requests.

[0072] The WebGL rendering engine is used to implement three-dimensional spatial layout, allowing users to explore topological relationships through dragging and zooming operations.

[0073] Perform kernel density estimation on the coordinates of the detected text missing area to generate Gaussian distribution thermal values:

[0074] Specifically, the local density of each coordinate point is calculated to generate a semi-transparent gradient overlay; a thermal value threshold interval is set, and different intervals correspond to different transparency levels (such as 30% transparency in high-density areas and 50% transparency in low-density areas).

[0075] Align the heat map with the original material image at the pixel level and achieve overlay display through alpha channel blending.

[0076] Deploy an interactive SVG-based canvas and bind click event listeners to the heat map nodes:

[0077] Specifically, when the user clicks on the text-missing area, an HTTP POST request is triggered to call the file upload interface; the interface supports the structured data format (JSON Schema) to define the supplementary material metadata.

[0078] File transfer protocol configuration: limit uploaded file types to PDF / JPG, verify file header information to prevent format tampering; use a block upload mechanism, and set the upper limit of a single file size to 10MB;

[0079] After the upload is complete, the file is temporarily stored in a distributed file system (such as HDFS) and a unique access token is generated.

[0080] It should be noted that by building a high-density visual rendering system and adopting a next-generation graphical interface, smooth interaction among millions of nodes is achieved, and rendering efficiency is improved by combining optimized view loading strategies. A multi-layered security system is deployed, including permission control, operation auditing, and attack protection. File transfers implement format verification and block checksum mechanisms to ensure data integrity.

[0081] The consistency verification module compares the pixel differences between the supplementary materials and the original uploaded materials, and verifies the logical matching between the supplementary text content and the knowledge graph terms. After the double verification is passed, the review conclusion is automatically updated. Otherwise, an abnormality is marked and manual review is triggered.

[0082] Among them, the original uploaded materials refer to the unmodified insurance application documents submitted by the customer for the first time (such as scanned copies of medical invoices, contract texts, etc.), which serve as the comparison benchmark for subsequent supplementary materials.

[0083] The structural similarity index was used to calculate the pixel difference between the original uploaded material and the supplementary material, and the key elements in the supplementary text (such as the name of the medical institution and the date of diagnosis and treatment) were extracted through named entity recognition.

[0084] Specifically, the BiLSTM-CRF model is used to identify entity boundaries, and the recall rate is improved by combining the domain dictionary; the entity attributes are semantically matched with the causal knowledge graph nodes, and the cosine similarity is calculated.

[0085] Based on cosine similarity, a logical conflict rule library is constructed to define mutually exclusive clause combinations (such as "deductible clause" and "full compensation clause"); when the supplementary content triggers two or more conflicting rules at the same time, an exception report is generated and the violating node ID is recorded.

[0086] It should be noted that this system defines mutually exclusive clause sets and describes conflict triggering conditions through logical expressions. It also supports rule syntax extension and versioning. When a conflict is detected, an exception report is generated, recording the node information of the offending clause and updating the node status identifiers in the visual map.

[0087] After the review conclusion is updated, modify the visual attributes of the nodes in the visual decision map.

[0088] Specifically, the color of the conflicting node changes from red (RGB 255, 0, 0) to green (RGB 0, 255, 0); the edge weight values ​​are updated and the topology layout is recalculated.

[0089] It should be noted that by establishing a cross-modal verification framework, multi-dimensional consistency verification is achieved through semantic alignment, content comparison, and biometric analysis. Dynamic resource allocation is achieved based on task complexity, and support for hardware acceleration and breakpoint resumption ensures an efficient and reliable verification process.

[0090] The closed-loop optimization module pushes the review conclusions and the revised visual decision map to the client, synchronously transmits feedback data, and dynamically optimizes the dynamic weight allocation model parameters and the causal knowledge graph clause node connection weights.

[0091] The review conclusion is pushed to the client mobile app via the message queue, along with a thumbnail of the revised decision map. For example, the push content includes a JSON summary of the conclusion and a PNG image of the map.

[0092] It should be noted that the decision map thumbnail refers to a tree structure diagram dynamically generated based on the clause conflict detection results, in which the node size reflects the importance weight of the clause and the edge thickness indicates the strength of the logical association.

[0093] Collect client operation logs and satisfaction scores, build a feedback dataset, and store it in a time series database. For example, record the time it takes for users to view conclusions and the number of times they zoom in and out of the graph.

[0094] Among them, the satisfaction score refers to the quantitative service evaluation value obtained by weighted calculation of the user's 5-level Likert scale subjective evaluation (1-5 points) collected through the client interface and objective behavioral indicators (such as graph interaction duration and operation frequency).

[0095] An incremental learning algorithm is used to update the policy network parameters of the dynamic weight allocation model. An adaptive learning rate mechanism dynamically adjusts the update step size based on the policy gradient variance (e.g., fine-tuning parameters every 24 hours). KL divergence regularization is used to constrain the intensity of policy exploration. A backpropagation algorithm is used to optimize the connection weights of clause nodes in the causal knowledge graph, increasing the strength of associations for frequently used paths.

[0096] Specifically, the latest feedback data is regularly loaded, network parameters are adjusted using a gradient descent algorithm, and historical versions are retained to support rapid rollbacks. Statistical analysis is performed on the access frequency of clause paths, dynamically adjusting node association weights. The conflict rule base is incrementally updated to maintain business logic synchronization.

[0097] It should be noted that by building a streaming data processing pipeline and an automated testing framework, combined with an intelligent rollback mechanism, system stability is ensured. We deepen feedback analysis, conduct value-added analysis such as user behavior modeling and service bottleneck diagnosis, and achieve continuous evolution through incremental learning and graph optimization.

[0098] In summary, the present invention achieves accurate quantification of customer emotions and service urgency through multimodal data fusion and PPO dynamic weight allocation; combines graph neural networks with causal knowledge graphs to dynamically identify clause conflicts and generate a visual evidence chain to ensure the compliance of service decisions; based on heat map interaction and logical consistency verification, it achieves traceability of material supplementation and anomaly blocking.

[0099] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An insurance customer service optimization system based on artificial intelligence, characterized by: include, The data acquisition and analysis module collects customers' multimodal interaction data in real time, extracts the fundamental frequency change rate and MFCC coefficients of the voiceprint, and calculates the emotional polarity value. It also analyzes the semantic content of the text to obtain the intent classification label and urgency score, and detects the coordinates of text-missing areas in images and the frequency of micro-expressions in video streams. The dynamic weight calculation module inputs the sentiment polarity value, urgency score, text missing area coordinates, and micro-expression action frequency into the dynamic weight allocation model based on the PPO algorithm, and outputs the sentiment weight value of the service request; The evidence chain generation module assigns service requests to an automated review queue based on sentiment weights. It uses a graph neural network to match intent classification labels with the clause node paths of a pre-built causal knowledge graph to generate an evidence chain containing the coordinates of conflicting clause nodes. The interactive visualization module converts the evidence chain into a visual decision map, overlays the coordinates of the text-missing areas to form a heat map, and triggers the interactive interface to receive supplementary materials uploaded by users for the heat map nodes; The consistency verification module compares the pixel differences between the supplementary materials and the original uploaded materials, verifies the consistency between the supplementary text content and the node logic of the causal knowledge graph, and updates the review conclusion when both the pixel difference and the node logic conflict are below their respective preset thresholds. Otherwise, it marks an anomaly and triggers manual review. The closed-loop optimization module pushes the review conclusions and the revised visual decision map to the client, synchronously transmits feedback data, and dynamically optimizes the dynamic weight allocation model parameters and the causal knowledge graph clause node connection weights.

2. The artificial intelligence-based insurance customer service optimization system according to claim 1, characterized in that: The extraction of the fundamental frequency change rate and MFCC coefficient of the speech voiceprint includes the following steps: Calculate the absolute value of the difference between adjacent frames of the fundamental frequency in the speech signal and calculate the average rate of change of the fundamental frequency within the preset time window; Perform frequency domain energy distribution analysis on the speech signal through the Mel filter bank and extract MFCC coefficients; The mean of the fundamental frequency change rate and the MFCC coefficient are Z-score normalized to generate the speech emotion feature vector.

3. The artificial intelligence-based insurance customer service optimization system according to claim 2, characterized in that: The method of parsing the text semantic content to obtain the intent classification label includes the following steps: Use the pre-trained BERT model to encode the text into word vectors and generate contextual semantic representations; Input the contextual semantic representation into the bidirectional LSTM network to capture long-distance dependencies and output temporal features; The time series features are weightedly fused through the attention mechanism to generate attention-weighted context features, and the Softmax classifier is used to output the intent classification label and confidence score.

4. The artificial intelligence-based insurance customer service optimization system according to claim 1, characterized in that: The dynamic weight allocation model includes the following steps: Initialize the policy network of the PPO algorithm and define the state space as the joint features of the sentiment polarity value, urgency score, coordinates of the text missing area, and micro-expression action frequency; Construct a reward function as a weighted combination of service request processing efficiency and customer satisfaction score; The network parameters of the strategy are optimized through importance sampling, and the distribution ratio of the sentiment weight values ​​is dynamically adjusted.

5. The artificial intelligence-based insurance customer service optimization system according to claim 1, characterized in that: The steps of generating an evidence chain containing the coordinates of the conflict clause nodes include: In the graph neural network, the knowledge graph clause nodes are defined as clause semantic vectors, and the edges are defined as the logical association strength of clauses. Calculate the multi-hop path similarity between the intent classification label and the clause node through the graph attention mechanism; Path nodes with similarity higher than the preset similarity threshold are selected to generate an evidence chain containing node coordinates and associated weights.

6. The artificial intelligence-based insurance customer service optimization system according to claim 1, characterized in that: The process of converting the chain of evidence into a visual decision map and superimposing the coordinates of the text missing area to form a heat map includes the following steps: Map the nodes of the evidence chain into interactive elements in a three-dimensional topological graph; Perform kernel density estimation on the coordinates of the text missing area to generate a semi-transparent heat map covering the original material; The spatial overlay display of heat map and topology map is realized through the WebGL rendering engine.

7. The artificial intelligence-based insurance customer service optimization system according to claim 1, characterized in that: Verification of supplementary materials includes calculating the difference measure between the supplementary content and the original material through image similarity evaluation, and analyzing the logical consistency between the supplementary text and the knowledge graph clause nodes in combination with the graph structure traversal algorithm. When the difference measure is lower than the critical value of the image difference measure and the logical consistency meets the conditions, the verification is judged to be successful.

8. The artificial intelligence-based insurance customer service optimization system according to claim 1, characterized in that: The interactive interface adopts a multimodal input protocol driven by a temporal memory network and supports uploading supplementary materials in structured data format.

9. The artificial intelligence-based insurance customer service optimization system according to claim 5, characterized in that: The connection relationship between the clause nodes of the causal knowledge graph is defined by the semantic ontology framework, and the connection weights between the clause nodes are dynamically optimized based on historical service data.

10. The artificial intelligence-based insurance customer service optimization system according to claim 4, characterized in that: The policy network of the PPO algorithm adopts an adaptive learning rate mechanism to update parameters, and constrains the policy exploration intensity through KL divergence regularization.

Images