An artificial intelligence-based insurance customer service optimization system

Abstract

The application discloses an insurance customer service optimization system based on artificial intelligence, and relates to the technical field of intelligent customer service.The system comprises a data acquisition and analysis module, which acquires multi-modal interaction data of customers in real time, extracts the fundamental frequency variation rate and MFCC coefficient of voice voiceprints, calculates emotional polarity values, simultaneously analyzes text semantic content to obtain intention classification labels and emergency degree scores, and detects text missing area coordinates in images and micro-expression action frequencies in video streams; a closed-loop optimization module, which pushes an audit conclusion and a revised visual decision graph to a client, synchronously returns feedback data, dynamically optimizes dynamic weight distribution model parameters and causal knowledge graph clause node connection weights.The application realizes accurate quantification of customer emotions and service emergency degrees through multi-modal data fusion and PPO dynamic weight distribution.

Description

Technical Field

[0001] This 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 Technology

[0002] Current insurance customer service systems are gradually incorporating artificial intelligence (AI) technology to improve service efficiency, with significant progress made, particularly in multimodal data processing. Existing technologies largely employ static rule engines for service request classification, combine natural language processing algorithms to analyze text intent, and utilize image recognition technology to detect document integrity. Some advanced systems have attempted to integrate speech sentiment analysis technology, extracting speech features through Mel-frequency cepstral coefficients (MFCC) and combining it with convolutional neural networks to achieve emotion recognition. In terms of service routing, reinforcement learning-based dynamic scheduling methods are used to optimize resource allocation, while knowledge graph technology is being introduced into the clause matching process to enhance logical reasoning capabilities.

[0003] However, existing technologies still have significant shortcomings in multimodal data fusion and dynamic decision-making: First, the sentiment analysis, text parsing and image detection modules mostly adopt independent weight allocation mechanisms, which cannot dynamically adjust the contribution of each modality of data according to the interaction scenario, resulting in bias in service priority determination; Second, the clause matching process relies on static knowledge graph traversal algorithms, which makes it difficult to capture logical conflicts between user-supplemented materials and the original evidence chain in real time, which can easily lead to inconsistencies in review conclusions. Summary of the Invention

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

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

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an AI-based insurance customer service optimization system, comprising: a data acquisition and analysis module, which collects multimodal interaction data from customers in real time, extracts the fundamental frequency change rate and MFCC coefficient of voiceprints and calculates sentiment polarity values, and simultaneously analyzes text semantic content to obtain intent classification labels and urgency scores, and detects the coordinates of text missing regions in images and the frequency of micro-expressions in video streams; a dynamic weight calculation module, which inputs the sentiment polarity value, urgency score, coordinates of text missing regions, and frequency of micro-expressions into a dynamic weight allocation model based on the PPO algorithm, and outputs the sentiment weight value of the service request; and an evidence chain generation module, which allocates service requests to an automated review queue based on the sentiment weight value, and matches the intent classification labels with a pre-constructed causal knowledge graph through a graph neural network. The system generates an evidence chain containing the coordinates of conflicting clause nodes by analyzing the clause node paths in the spectrum. An interactive visualization module converts this evidence chain into a visual decision graph, overlaying the coordinates of missing text areas to form a heatmap, and triggering an interactive interface to receive supplementary materials uploaded by the user for the heatmap nodes. A consistency verification module compares the pixel differences between the supplementary materials and the original uploaded materials to verify the consistency between the supplementary text content and the logic of the causal knowledge graph nodes. When both the pixel difference and the degree of node logic conflict are below their respective preset thresholds, the review conclusion is updated; otherwise, an anomaly is marked and manual review is triggered. A closed-loop optimization module pushes the review conclusion and the revised visual decision graph to the client, synchronously transmits feedback data, and dynamically optimizes the dynamic weight allocation model parameters and the connection weights of the causal knowledge graph clause nodes.

[0008] As a preferred embodiment of the AI-based insurance customer service optimization system of the present invention, the extraction of the fundamental frequency change rate and MFCC coefficients 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 statistically analyze the average rate of change of the fundamental frequency within a preset time window;

[0010] Frequency domain energy distribution analysis of the speech signal is performed using the Mel filter bank to extract the MFCC coefficients.

[0011] The mean fundamental frequency change rate and MFCC coefficients are Z-score normalized to generate speech emotion feature vectors.

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

[0013] The text is encoded with word vectors using a pre-trained BERT model to generate a contextual semantic representation.

[0014] The contextual semantic representation is input into a bidirectional LSTM network to capture long-distance dependencies and output temporal features;

[0015] The temporal features are weighted and fused using an attention mechanism to generate attention-weighted contextual features. The Softmax classifier is then used to output the intent classification label and confidence score.

[0016] As a preferred embodiment of the AI-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 a joint feature of sentiment polarity value, urgency score, coordinates of missing text regions, and frequency of micro-expression actions;

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

[0019] The importance sampling optimization strategy optimizes network parameters and dynamically adjusts the allocation ratio of sentiment weights.

[0020] As a preferred embodiment of the AI-based insurance customer service optimization system of the present invention, the step of generating an evidence chain containing the coordinates of conflicting clause nodes includes the following steps.

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

[0022] The multi-hop path similarity between intent classification labels and clause nodes is calculated using a graph attention mechanism.

[0023] Select path nodes with similarity higher than a preset similarity threshold and generate an evidence chain containing node coordinates and association weights.

[0024] As a preferred embodiment of the AI-based insurance customer service optimization system of the present invention, the step of converting the evidence chain into a visual decision graph and overlaying the coordinates of missing text areas to form a heatmap includes the following steps.

[0025] Map the nodes of the evidence chain to interactive elements in a 3D topology graph;

[0026] Kernel density estimation is performed on the coordinates of the missing text regions to generate a semi-transparent heatmap covering the original material.

[0027] A spatial overlay display of heatmaps and topology maps is achieved using the WebGL rendering engine.

[0028] As a preferred embodiment of the AI-based insurance customer service optimization system of the present invention, the verification of supplementary materials includes: calculating the difference measure between the supplementary content and the original material through image similarity assessment; analyzing the logical consistency between the supplementary text and the knowledge graph clause nodes by combining graph structure traversal algorithm; and determining that the verification is passed when the difference measure is lower than the critical value of the image difference measure and the logical consistency meets the condition.

[0029] As a preferred embodiment of the AI-based insurance customer service optimization system of the present invention, the interactive interface adopts a multimodal input protocol driven by a time-series memory network, which supports the uploading of supplementary materials in structured data format.

[0030] As a preferred embodiment of the AI-based insurance customer service optimization system described in this invention, the connection relationships of the clause nodes in the causal knowledge graph are defined logically through a semantic ontology framework, and the connection weights between clause nodes are dynamically optimized based on historical service data.

[0031] As a preferred embodiment of the AI-based insurance customer service optimization system of the present invention, the policy network of the PPO algorithm updates parameters using an adaptive learning rate mechanism, while exploring strength through a KL divergence regularization constraint strategy.

[0032] The beneficial effects of this invention are as follows: by multimodal data fusion and PPO dynamic weight allocation, the accuracy of customer sentiment and service urgency can be quantified; by combining graph neural networks and causal knowledge graphs, conflicts in terms can be dynamically identified and a visual evidence chain can be generated to ensure the compliance of service decisions; and based on heatmap interaction and logical consistency verification, the traceability of material supplementation and the prevention of anomalies can be achieved. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 A module diagram for optimizing an AI-based insurance customer service system.

[0035] Figure 2 This is a flowchart of the multimodal data acquisition and analysis process.

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

[0037] Figure 4 This is a flowchart for consistency verification. Detailed Implementation

[0038] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0039] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0040] Secondly, the term "one embodiment" or "embodiment" as used 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 different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0041] Reference Figures 1-4 This is one 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 multimodal interaction data from customers 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 text missing areas 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 a preset time window is statistically analyzed. The speech signal is subjected to frequency domain energy distribution analysis through Mel filter bank to extract MFCC coefficients. The mean fundamental frequency change rate and MFCC coefficients are Z-score standardized to generate speech emotion feature vectors.

[0045] Ideally, the continuous speech signal is segmented into short time windows of fixed length, with a fixed step size between adjacent windows to preserve continuity. A window function that suppresses spectral leakage is used to weight each signal segment, reducing frequency domain distortion caused by truncation. The fundamental frequency is calculated using a signal periodicity detection algorithm, and the statistical characteristics of fundamental frequency changes between adjacent windows are analyzed. The fundamental frequency search range is set to cover the typical frequency range of human speech, eliminating outlier interference. A filter bank covering the audible frequency range is constructed. After converting the speech signal to the frequency domain, coefficients characterizing timbre are extracted through the energy distribution of the filter bank. Dynamic range compression and standardization are applied to the coefficients to generate robust feature vectors.

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

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

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

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

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

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

[0052] Based on a predefined time-sensitive vocabulary list according to the classification probability distribution, the number and position of matching words in the text are counted. Time-sensitive keywords (such as 'urgent', 'expedited', 'within 24 hours', etc.) in the text are detected by combining keyword matching algorithm with semantic analysis. Combined with the pre-set basic urgency value of intent classification label, an urgency score is generated.

[0053] Image features are extracted using a residual convolutional neural network, and candidate box coordinates are generated through a region proposal network. The coordinates of the candidate boxes with the highest confidence are selected as the missing region localization results using a non-maximum suppression algorithm. At the same time, based on a facial muscle motion coding system, the activation state of specific action units is tracked, and the trigger frequency per unit time is statistically analyzed to generate a micro-expression frequency index.

[0054] The dynamic weight calculation module takes the sentiment polarity value, urgency score, coordinates of missing text regions, and micro-expression frequency as inputs to the dynamic weight allocation model based on the PPO algorithm, and outputs the sentiment weight value of the service request.

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

[0056] Range standardization was performed on the emotional polarity value and urgency score, the coordinates of missing text regions were converted into the proportion of missing area, and the frequency of micro-expression movements was logarithmically transformed.

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

[0058] The reward function is constructed to include two indicators: service timeliness improvement and customer satisfaction. A weight allocation scheme is output through a policy network. For example, a weight vector of [0.4, 0.3, 0.2, 0.1] indicates that the emotional weight accounts for 40%.

[0059] The policy network synchronizes its parameters every 1000 iterations, setting a 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 to 0.01.

[0060] A superior approach employs a dual-evaluation network architecture to enhance policy stability. This architecture refines the decision-making logic through a collaborative mechanism where the main network evaluates state value and the auxiliary network evaluates action value. The training process incorporates priority-based empirical sampling and adaptive entropy adjustment strategies, combined with gradient constraint techniques to balance exploration and efficiency. Simultaneously, interpretable analysis components are deployed to provide interpretive outputs such as feature importance quantification and decision path tracing, meeting transparency requirements.

[0061] The evidence chain generation module assigns service requests to an automated review queue based on sentiment weight values. It uses a graph neural network to match intent classification labels with the clause node paths of a pre-built causal knowledge graph, generating an evidence chain containing the coordinates of conflicting clause nodes. It constructs a graph neural network model, maps intent classification labels to the root node of the causal knowledge graph, and uses a breadth-first search algorithm to traverse the clause node paths.

[0062] It should be noted that Graph Neural Network (GNN) refers to the algorithm type of Graph Neural Network (GNN) in general, without specifying a particular instance, and is usually used for functional description; building a Graph Neural Network model specifically emphasizes the specific model building process (such as architecture design, parameter training, domain adaptation, etc.) for the knowledge graph of insurance terms.

[0063] Specifically, insurance clauses are abstracted into nodes, and relationships between nodes are established through semantic similarity calculation and historical data analysis, forming a weighted directed graph structure. A graph traversal algorithm is used starting from the root node, combined with an attention mechanism to evaluate the path weights of clause nodes, and select 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 a "car accident medical claim" application, starting from the root node, the system evaluates the weight of each clause path using an attention mechanism, prioritizing the high-weight path "car insurance main clause → medical expense coverage" (0.9). At the same time, it automatically detects that the "medical expense coverage" clause in this path requires a hospital with a qualification of level two or above, while the clinic invoice provided by the customer does not meet this requirement. The system immediately marks the conflict node and highlights it on the visualization interface, prompting the customer to supplement compliant hospital proof, thus achieving intelligent clause verification and conflict warning.

[0065] The Jaccard coefficient is used to calculate the association strength of clause nodes, match the optimal path, and detect conflicting clause nodes. Example: A claim application intends to trigger the path "Accidental Injury Insurance → Medical Expense Clause → Deductible Provisions," and a geographical conflict is detected between "Deductible" and "Contracted Hospitals." For the detected conflicting clause nodes, an evidence chain containing the coordinates of the conflicting node and the content of the associated clauses is generated.

[0066] In a superior manner, the evidence chain generation module enables the dynamic evolution of the knowledge graph. Newly added clauses are adaptively embedded through a contrastive learning framework, and relationships are mined based on inductive graph learning. Conflict rules are automatically generated using pattern mining algorithms. For query performance optimization, a hierarchical index structure, an approximate nearest neighbor search scheme, and a parallel traversal engine are constructed to support efficient access to large-scale knowledge graphs.

[0067] The interactive visualization module transforms the chain of evidence into a tree-like visual decision graph, where the node size dynamically reflects the importance of the clauses, the edge weights intuitively display the strength of logical connections, and a heat map is generated by overlaying the coordinates of missing text areas. The module also 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 (clause importance, association strength) is as follows: Node size is dynamically adjusted based on clause importance, and association strength determines the thickness and color depth of the edges; dynamic interactive operations in 3D spatial layout are supported. Spatial density analysis is performed on the coordinates of missing areas to generate a semi-transparent overlay layer, with high-density areas highlighted using low transparency.

[0069] Specifically, spatial density analysis of the coordinates of missing regions refers to using a kernel density estimation algorithm to calculate the distribution density of all detected missing region coordinates on the original image plane, constructing a Gaussian distribution function centered on each coordinate point, and generating a continuous probability density field by superimposing the density contributions of all points, 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 clauses, and the importance score is calculated based on the frequency of citation and the probability of conflict in historical service data; the edge weights are calculated through a graph neural network to reflect the strength of the causal relationship between clauses, and the weight value range is limited to [0,1].

[0071] Historical service data includes: intent classification tags, sentiment polarity values, review conclusion feedback, frequency of access to terms and conditions, and conflict marker records in customer service requests.

[0072] The WebGL rendering engine is used to realize the 3D spatial layout, allowing users to explore the topological relationships through drag and zoom operations.

[0073] Kernel density estimation is performed on the coordinates of the detected missing text regions to generate Gaussian distributed heatmap values:

[0074] Specifically, the local density of each coordinate point is calculated to generate a semi-transparent gradient overlay layer; a thermal value threshold range is set, with different ranges corresponding to different transparency levels (e.g., 30% transparency in high-density areas and 50% transparency in low-density areas).

[0075] The heat map and the original material image are aligned pixel-level and then overlaid using alpha channel mixing.

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

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

[0078] File transfer protocol configuration: Restrict uploaded file types to PDF / JPG, verify file header information to prevent format tampering; adopt a chunked upload mechanism, with a single file size limit set 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 a high-density visualization rendering system is constructed, employing a new generation of graphics interfaces to achieve smooth interaction with millions of nodes, and combined with view optimization loading strategies to improve rendering efficiency. A multi-layered security system is deployed, including access control, operation auditing, and attack protection, and file transfer implements format verification and block verification 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. If the double verification passes, the review conclusion is automatically updated; otherwise, it is marked as abnormal and a manual review is triggered.

[0082] The original uploaded materials refer to the insurance application documents submitted by the customer for the first time without modification (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 is used to calculate the pixel difference between the original uploaded material and the supplementary material, and key elements (such as the name of the medical institution and the date of treatment) in the supplementary text are 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 it with a domain dictionary; entity attributes are semantically matched with causal knowledge graph nodes, and cosine similarity is calculated.

[0085] Based on cosine similarity, a logical conflict rule base is constructed, defining 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 anomaly report is generated and the violation node ID is recorded.

[0086] It should be noted that the system defines a set of mutually exclusive clauses and describes conflict triggering conditions using logical expressions; it supports rule syntax extension and version control. When a conflict is detected, an exception report is generated, recording the information of the violating clause node; and the node status identifiers in the visualization graph are updated.

[0087] After the review conclusion is updated, the visual attributes of the nodes in the visualized decision graph will be modified.

[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 a cross-modal verification framework is established to achieve multi-dimensional consistency verification through semantic alignment, content comparison, and biometric analysis. Resources are dynamically allocated based on task complexity, and hardware acceleration and breakpoint resumption are supported to ensure the efficiency and reliability of the verification process.

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

[0091] The review results are pushed to the client mobile application via message queue, along with a revised decision graph thumbnail. For example, the push content includes a summary of the results in JSON format and the graph in PNG format.

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

[0093] Collect client operation logs and satisfaction ratings to build a feedback dataset and store it in a time-series database. For example, record the time users spend viewing conclusions and the number of times the map is zoomed.

[0094] The satisfaction score is a quantitative service evaluation value calculated by weighting user subjective evaluations (1-5 points) on a 5-level Likert scale collected through the client interface with objective behavioral indicators (such as interaction time and operation frequency).

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

[0096] Specifically, the system periodically loads the latest feedback data, adjusts network parameters using gradient descent, and retains historical versions for rapid rollback. It also statistically analyzes the access frequency of terms and conditions, dynamically adjusts node association weights, and incrementally updates the conflict rule base to keep business logic synchronized.

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

[0098] In summary, this invention achieves precise quantification of customer sentiment and service urgency through multimodal data fusion and dynamic PPO weight allocation; it dynamically identifies clause conflicts and generates a visual evidence chain by combining graph neural networks and causal knowledge graphs to ensure the compliance of service decisions; and it achieves traceability of material replenishment and anomaly prevention based on heatmap interaction and logical consistency verification.

[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 it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An insurance customer service optimization system based on artificial intelligence, characterized in that: include, The data acquisition and analysis module collects multimodal interaction data from customers 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 text missing areas in the image and the frequency of micro-expression movements in the video stream. The dynamic weight calculation module inputs the sentiment polarity value, urgency score, coordinates of missing text regions, and micro-expression 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 weight values, and generates an evidence chain containing the coordinates of conflicting clause nodes by matching intent classification tags with the clause node paths of a pre-built causal knowledge graph through a graph neural network. The interactive visualization module converts the chain of evidence into a visual decision map, overlays the coordinates of missing text areas to form a heat map, and triggers an 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 to verify the consistency between the supplementary text content and the logic of the causal knowledge graph nodes. When both the pixel difference and the node logic conflict are lower than their respective preset thresholds, the review conclusion is updated; otherwise, it is marked as abnormal and manual review is triggered. The closed-loop optimization module pushes the review conclusions and revised visual decision graphs to the client, synchronously transmits feedback data, and dynamically optimizes the dynamic weight allocation model parameters and the connection weights of the causal knowledge graph clause nodes.

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

3. The insurance customer service optimization system based on artificial intelligence as described in claim 2, characterized in that: The process of parsing the semantic content of the text to obtain the intent classification label includes the following steps. The text is encoded with word vectors using a pre-trained BERT model to generate a contextual semantic representation. The contextual semantic representation is input into a bidirectional LSTM network to capture long-distance dependencies and output temporal features; The temporal features are weighted and fused using an attention mechanism to generate attention-weighted contextual features. The Softmax classifier is then used to output the intent classification label and confidence score.

4. The insurance customer service optimization system based on artificial intelligence as described in 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 a joint feature of sentiment polarity value, urgency score, coordinates of missing text regions, and frequency of micro-expression actions; The reward function is constructed as a weighted combination of service request processing efficiency and customer satisfaction score; The importance sampling optimization strategy optimizes network parameters and dynamically adjusts the allocation ratio of sentiment weights.

5. The insurance customer service optimization system based on artificial intelligence as described in claim 1, characterized in that: The process of generating a chain of evidence containing the coordinates of nodes with conflicting clauses includes the following steps. In graph neural networks, knowledge graph clause nodes are defined as clause semantic vectors, and edges represent the logical association strength of clauses. The multi-hop path similarity between intent classification labels and clause nodes is calculated using a graph attention mechanism. Select path nodes with similarity higher than a preset similarity threshold and generate an evidence chain containing node coordinates and association weights.

6. The insurance customer service optimization system based on artificial intelligence as described in claim 1, characterized in that: The process of converting the chain of evidence into a visual decision graph and overlaying it with the coordinates of missing text areas to form a heatmap includes the following steps. Map the nodes of the evidence chain to interactive elements in a 3D topology graph; Kernel density estimation is performed on the coordinates of the missing text regions to generate a semi-transparent heatmap covering the original material. A spatial overlay display of heatmaps and topology maps is achieved using the WebGL rendering engine.

7. The insurance customer service optimization system based on artificial intelligence as described in claim 1, characterized in that: Verification of supplementary materials involves calculating the difference between the supplementary content and the original material through image similarity assessment, and analyzing the logical consistency between the supplementary text and the knowledge graph clause nodes using a graph structure traversal algorithm. When the difference measure is lower than the critical value of the image difference measure and the logical consistency condition is met, the verification is deemed successful.

8. The insurance customer service optimization system based on artificial intelligence as described in claim 1, characterized in that: The interactive interface adopts a multimodal input protocol driven by a time-series memory network, which supports the uploading of supplementary materials in structured data format.

9. The insurance customer service optimization system based on artificial intelligence as described in claim 5, characterized in that: The connection relationships between the clause nodes in the causal knowledge graph are defined logically through a semantic ontology framework, and the connection weights between clause nodes are dynamically optimized based on historical service data.

10. The insurance customer service optimization system based on artificial intelligence as described in claim 4, characterized in that: The policy network of the PPO algorithm uses an adaptive learning rate mechanism to update parameters, while exploring the strength through a KL divergence regularization constraint strategy.

Images