A method and system for matching insurance products based on customer multi-dimension reference

Abstract

The application provides a kind of matching method and system of insurance product based on customer multidimensional reference.The method is: collecting multi-modal data and extracting dynamic characteristics;Multiple agents respectively convert dynamic characteristics into feature vectors and probability vectors;Intention fusion agent calculates conflict coefficient, and different processing strategies are selected according to the threshold interval where the conflict coefficient is located;Product matching agent splices out customer query vector, then updates product node embedding in product knowledge graph through scene-aware heterogeneous graph neural network, calculates matching degree score through bilinear matching, and filters candidate products through business rules, outputs the product with the highest matching degree and score.The application improves the accuracy, compliance and marketability of insurance product recommendation through consistency detection and closed-loop verification of multi-modal information such as voice, expression and behavior, combined with scene-aware knowledge graph matching and business rule filtering.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for matching insurance products based on multi-dimensional customer references. Background Technology

[0002] In the insurance sales field, especially in instant messaging channels such as WeChat for businesses, intelligent recommendation systems typically rely on static or quasi-static customer profiles for product matching. These customer profiles are primarily built upon structured information such as historical insurance records, age, gender, region, and occupation, as well as tags obtained through offline data mining (e.g., "price-sensitive," "health-conscious"). These profiles are usually updated daily or even weekly, failing to reflect the dynamically changing needs and states of customers in real-time conversations.

[0003] In existing technologies, some solutions attempt to assist recommendations by matching keywords in conversations, such as triggering preset products after recognizing customers mentioning words like "critical illness insurance" or "premium." However, these methods are still limited to capturing explicit text signals, neglecting the analysis of other dimensions of information and their inclusion in product recommendation references. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method and system for matching insurance products based on multi-dimensional references.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: S1. Real-time acquisition of multimodal data, extraction of customer profile vectors, preprocessing of the acquired multimodal data, and generation of dynamic features; S2. Dynamic features are transformed into feature vectors and probability vectors; S3. The intent fusion agent performs conflict detection on the probability vector, obtains the result of the conflict detection, and selects to perform weighted fusion on the feature vector to output a comprehensive intent vector or trigger follow-up questions based on the result of the conflict detection. S4. The intent fusion agent fuses feature vectors into a comprehensive intent vector; S5, the product matching agent performs product matching based on the comprehensive intent vector and outputs the product matching result.

[0006] Furthermore, step S2 includes the following steps: S2. Dynamic features are transformed into feature vectors and probability vectors; S21. The semantic understanding agent transforms text features into semantic feature vectors. and probability vector semantic intent distribution and semantic intent tags ; The semantic understanding agent is based on a lightweight BERT (TinyBERT) real-time semantic parsing model, deployed using ONNXRuntime, and supports dynamic word vector updates. Dynamic word vectors utilize the Transformer's self-attention mechanism to dynamically adjust word embeddings based on the current dialogue context, allowing the same word to have different semantic representations in different contexts. The final output is a semantic feature vector. and probability vector semantic intent distribution and semantic intent tags .

[0007] S22. The emotion-oriented agent transforms emotion features into emotion feature vectors. And probability vector sentiment distribution and emotional tags ; The emotional tendency intelligence integrates a multimodal input emotional classification network and adjusts the text sentiment using a pre-trained BERT-base-Chinese algorithm to achieve three main categories: positive, neutral, and negative. It also integrates emoji mapping features and speech tone features (speech rate, pauses). Specifically: For text sentiment analysis, bert-base-Chinese is used for fine-tuning, and the probability distribution of sentiment categories is output. For facial expression features, an expression-emotion mapping table is constructed, which maps emojis to emotion vectors and extracts meta-features such as expression location and frequency; To target speech intonation features, acoustic features such as fundamental frequency, energy, speech rate, and MFCC are extracted from speech messages, and speech intonation feature vectors are extracted using CNN. Features from the three modalities are weighted and fused through an attention mechanism to ultimately output a sentiment feature vector. And probability vector sentiment distribution and emotional tags .

[0008] S23. The behavioral pattern agent transforms behavioral features and interaction frequency features into behavioral feature vectors. and probability vector behavior pattern distribution and behavioral tags ; The behavioral pattern agent is based on an LSTM-based temporal pattern mining network. It takes historical dialogue sequences and current behavioral features as input, including: whether a product card was clicked, the duration of viewing terms, whether a message was withdrawn, message length, response time, number of price inquiries, and the number of products compared. The LSTM network then outputs behavioral pattern encodings to identify customer types (price-sensitive, comparison-oriented, hesitant, decisive, and skeptical), and outputs a behavioral feature vector. and probability vector behavior pattern distribution and behavioral tags .

[0009] Furthermore, step S3 includes the following steps: S3. The intent fusion agent performs conflict detection on the probability vector, obtains the result of the conflict detection, and selects to perform weighted fusion on the feature vector to output a comprehensive intent vector or trigger follow-up questions based on the result of the conflict detection. The result after conflict detection is the conflict coefficient C; S31. The intention is to fuse the intelligent agent to perform conflict detection on the probability vector and calculate the conflict coefficient C; Specifically, the semantic intent distribution is calculated using weighted relative entropy (KL divergence). Emotional distribution Distribution of behavioral patterns The difference between pairwise probability distributions, combined with the maximum probability class conflict indicator function. Quantify the conflict coefficient using the following formula : ; In the formula: Indicates KL divergence; This is the maximum probability category conflict indicator function, used when the semantic main category and the behavioral main category are inconsistent. =1, otherwise =0; S32. Different processing strategies are triggered based on the threshold range in which the conflict coefficient C falls; Low conflict: The probability vectors output by the semantic understanding agent, the sentiment agent, and the behavior pattern agent are highly consistent, and the final intention is directly weighted and fused. Medium-level conflict: There is some divergence between the probability vectors output by the semantic understanding agent, the sentiment agent, and the behavior pattern agent, which triggers mild follow-up questions to reduce the conflict level to the low-conflict range. The weights are then adjusted based on the responses before being fused and output. High conflict: There is a serious discrepancy between the probability vectors output by the semantic understanding agent, the sentiment agent, and the behavior pattern agent, which triggers forced follow-up questions. If the conflict drops to a low conflict range after the follow-up questions, the weights are adjusted according to the responses and then the outputs are merged. If the conflict cannot be reduced to a low level after multiple follow-up questions or after reaching the follow-up question limit, the final comprehensive intent vector will directly adopt the output of the semantic understanding agent.

[0010] The follow-up question is generated dynamically by the intelligent agent based on the source and content of the conflict and using a "conflict-problem" mapping strategy when the conflict reaches a preset trigger threshold.

[0011] When a conflict reaches a preset trigger threshold, the intent fusion agent dynamically generates follow-up questions based on the source and content of the conflict, using a "conflict-problem" mapping strategy, and sends the generated questions to the client.

[0012] Based on the output labels of the three agents—the semantic understanding agent, the sentiment-oriented agent, and the behavioral pattern agent—the semantic intent label is... Emotional tags Behavioral tags And the distribution of the conflict coefficient C, to determine the main sources of conflict. Define the conflict type encoding vector T=(t1,t2,t3,t4): in: t1 is and A predefined pair of contradictions represents a semantic-emotional conflict; if a conflict exists, t1=1; otherwise, t1=0. t2 is and A predefined pair of contradictions represents a semantic-behavioral conflict; if a conflict exists, then t2 = 1; otherwise, t2 = 0. t3 is and A predefined pair of contradictions represents an emotional-behavioral conflict; if a conflict exists, t3 = 1; otherwise, t3 = 0. t4 is , , A predefined pair of contradictions represents a complete conflict between semantics, emotion, and behavior. If a conflict exists, then t4 = 1; otherwise, t4 = 0. The predefined contradictions are constructed using domain knowledge; for example, the semantic "price inquiry" contradicts the emotional "negative", the semantic "purchase intention" contradicts the behavioral "comparative", and the emotional "positive" contradicts the behavioral "hesitant".

[0013] A conflict-problem mapping table is constructed, where each record contains the conflict type, the specific contradictory pair, and a problem template (including fillable slots). The mapping table is designed offline manually or learned using historical dialogue data; the problem template selection function is as follows: ; In the formula: A set of mapping tables; Cosine similarity between conflict type encoding and template encoding; This is the contradiction pair matching exponential function, which takes the value 1 if the current tag pair matches the contradiction pair defined in the template, and 0 otherwise.

[0014] The slots in the template (such as [Product Type], [Premium Range]) need to be dynamically extracted from the current dialogue context. The slot filling function is: ; The problem template can be formalized as follows: The template string is , which includes One slot { The problem after filling in the blanks is: .

[0015] Furthermore, step S4 includes the following steps: S4. The intent fusion agent fuses feature vectors into a comprehensive intent vector; S41. The intent fusion agent will integrate the semantic feature vectors output by the three agents: the semantic understanding agent, the sentiment agent, and the behavior pattern agent. Emotional feature vector behavioral feature vector Transformed to the same dimension d through a linear mapping, we obtain , ; ; ; In the formula: The scene feature mapping matrix; For the bias vector, and To complete the linear transformation; The scene feature mapping matrix; For the bias vector, and To complete the linear transformation; The scene feature mapping matrix; For the bias vector, and To complete the linear transformation; semantic feature vector The value obtained through linear mapping; For sentiment feature vectors The value obtained through linear mapping; behavioral feature vector The value obtained through linear mapping; S42. Use an attention mechanism to calculate the attention weights corresponding to the semantic understanding agent, the sentiment-oriented agent, and the behavioral pattern agent; ; ; ; In the formula: Weights corresponding to semantic understanding agents; Weights corresponding to agents with emotional tendencies; The weights corresponding to the behavioral pattern agents; These are learnable weight vectors; This is a bias term (scalar); S33. Weighted fusion yields the final comprehensive intent vector; The weighted fusion calculation formula is as follows: ; In the formula: This is the overall intent vector.

[0016] Furthermore, step S5 includes the following steps: S5, the product matching agent performs product matching based on the comprehensive intent vector and outputs the product matching result; S51, Generate customer query vector ; The product matching agent will integrate intent vectors. Emotional feature vector behavioral feature vector and customer profile vector Concatenate into a unified customer query vector ; through customer query vectors Indicates the customer's query intent; ; In the formula, This represents a vector concatenation operation; The dimensions after concatenation are: ; In the formula, D is the dimension of the concatenated customer query vector; The dimension of the comprehensive intent vector; The dimension of the sentiment feature vector; The dimension of the behavioral feature vector; Dimensions of the customer profile vector; S52, Generate product node vector ; The product node vector represents a product in the insurance product knowledge graph. The correspondence between the product node vector and the product in the insurance product knowledge graph is that one product node vector corresponds to one product in the insurance product knowledge graph. Product Node Vector The generation steps are as follows: In the insurance product knowledge graph, each product node The embedding vectors are updated through multi-layer heterogeneous graph convolution. Layer representation It is given by the following formula: ; In the formula: For the first The product node vector of the layer; A set of edge types; For type With products The set of connected neighboring nodes; The weight matrix is ​​a learnable weight matrix; The weight matrix is ​​a learnable weight matrix; Scene-aware attention weights; Furthermore, in the above formula The calculation formula is as follows: ; In the formula: This involves concatenating vectors. For scene feature vectors; For product nodes Feature vectors, subscripts Indicates product nodes; Neighboring nodes The feature vector is used to represent other types of nodes (such as keyword nodes and clause nodes) connected to the product node. The scene mapping matrix; edge type The corresponding weight matrix uses different linear transformations for different relationships (edge ​​types); This is the attention parameter vector; The superscript T indicates the transpose of a matrix or vector, and is used to calculate the inner product.

[0017] Through the above methods L After layer propagation, the final product node vector is obtained. .

[0018] S53, Bilinear matching calculates the matching score; The customer query vector is calculated using a bilinear calculation method. With product node vector Match score ; ; In the formula: It is a learnable bilinear weight matrix; The Sigmoid activation function is used to compress the matching degree to the (0,1) interval; The superscript T indicates the transpose of a matrix or vector, and is used to calculate the inner product.

[0019] The matching score is used to indicate the degree of match between the customer's intent and the product.

[0020] S54, based on the matching score, recalls products and outputs the product with the highest matching score and matching score; The product matching agent calculates all product and customer query vectors. Match score Sort the products by score from highest to lowest, recall the Top-K candidate products, filter them according to business rules, and finally output the m products with the highest matching degree and their matching scores.

[0021] A matching system for insurance products based on multi-dimensional customer references includes a data acquisition module, a semantic understanding agent, a sentiment agent, a behavioral pattern agent, an intent fusion agent, and a product matching agent. These components communicate and interface with each other through standardized information interfaces. The data acquisition module collects multimodal data in real time, extracts customer profile vectors, preprocesses the collected multimodal data, and generates dynamic features. The semantic understanding agent converts textual features into semantic feature vectors. and probability vector semantic intent distribution and semantic intent tags The emotional tendency agent is used to transform emotional features into emotional feature vectors. And probability vector sentiment distribution and emotional tags The behavioral pattern agent is used to transform behavioral features and interaction frequency features into behavioral feature vectors. and probability vector behavior pattern distribution and behavioral tags The intent fusion intelligent body includes a conflict detection module, a question generation module, and a vector fusion module. The conflict detection module detects conflicts in the probability vectors, calculates the conflict coefficient C, and triggers different processing strategies based on the threshold range of the conflict coefficient C. The question generation module generates follow-up questions when there are discrepancies in the meaning of the probability vectors output by the semantic understanding agent, the sentiment agent, and the behavior pattern agent, and sends these follow-up questions to the client. The vector fusion module fuses the feature vectors into a comprehensive intent vector. The product matching intelligent body includes a product matching module, an insurance product knowledge graph, and a rule filtering module. The product matching module generates customer query vectors. and product node vector And calculate all product and customer query vectors. Match score The product node vector is then output, along with the products filtered according to business rules and their matching scores. It corresponds to the products in the insurance product knowledge graph; the rule filtering module is used to filter candidate products through constraint rules after product recall.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: I. This invention constructs an insurance product knowledge graph and designs a scene-aware heterogeneous graph neural network (HGNN) to update product node embeddings. During the message passing process of the HGNN, scene feature vectors extracted from the dialogue context are introduced as attention biases to dynamically adjust the aggregation weights of different neighboring nodes, enabling product embeddings to perceive the specific scenario of the current dialogue. Combined with a bilinear matching function to calculate the matching score between the customer query vector and the product embedding, the system can recall candidate products that best match the customer's real-time needs from a massive number of products, solving the problems of lagging traditional static profile recommendations and their inability to adapt to changes in dialogue context.

[0023] Second, this invention uses three intelligent agents—semantic understanding, sentiment tendency, and behavioral pattern—to analyze multimodal interaction data of customers in parallel. It employs weighted relative entropy (KL divergence) and a maximum probability category conflict indicator function to quantitatively calculate the conflict coefficient, achieving real-time quantitative assessment of the consistency among customer language, sentiment, and behavior. When the conflict coefficient is in the medium to high threshold range, the system proactively triggers a follow-up questioning mechanism. Through a "conflict-question" mapping strategy, it dynamically generates targeted questions to guide the customer to clarify their true intentions. Based on the responses, the system adjusts the fusion weights of each modality until the conflict level is reduced to a low level. This closed-loop verification mechanism effectively solves the decision-making dilemma of traditional single-modality or simple weighted fusion when information contradictions arise, significantly improving the reliability and confidence of the comprehensive intent vector, laying a solid foundation for subsequent accurate product matching.

[0024] Third, this invention also includes a follow-up questioning mechanism. The core of follow-up questioning is to proactively eliminate ambiguity, enabling the system to obtain clearer signals from subsequent customer responses, thereby correcting inconsistencies in the output meanings among the three agents: semantic understanding agent, sentiment agent, and behavioral pattern agent. When the judgments of semantics, sentiment, and behavior diverge, the system cannot determine the customer's true intention, and forced weighted fusion may lead to incorrect recommendations. Through follow-up questioning, the system can guide customers to clarify their needs and supplement information, thereby reducing the conflict coefficient. The effect of reducing the conflict coefficient is to improve the reliability and confidence of the comprehensive intent vector. When the conflict coefficient drops to a low level, the meanings of the outputs of the three dimensions of semantics, sentiment, and behavior are consistent. At this time, the intent vector obtained by weighted fusion can accurately reflect the customer's true state, and subsequent product matching is also based on this high-confidence input, thereby significantly improving the accuracy of recommendations, reducing false pushes, and avoiding business risks caused by contradictory information (such as forcibly selling to hesitant customers or ignoring customers' negative emotions). In addition, low conflict also means that the system's understanding of customers is more interpretable, facilitating subsequent auditing and optimization.

[0025] Fourth, after the intelligent agent matches and recalls Top-K candidate products, this invention further introduces a business rule filtering layer. The rule engine sequentially executes non-model-based constraints such as product sales status, customer qualification verification, product mutual exclusion, regulatory compliance, sales personnel permissions, and historical recommendation deduplication, retaining only the m products that simultaneously satisfy all business rules and have the highest matching degree. This design effectively avoids invalid recommendations, regulatory compliance risks, internal product conflicts, and duplicate recommendations of rejected products, significantly improving the actual conversion rate and customer satisfaction of the recommendation results, while reducing the company's compliance risks and sales disputes.

[0026] Fifth, this invention incorporates non-textual modal information such as speech tone, emoticons, and behavioral sequences into the feature acquisition and fusion scope, and dynamically weights each modal feature through an attention mechanism, breaking through the limitations of traditional methods that rely solely on text keyword matching. Simultaneously, the various intelligent agents (semantic understanding, sentiment tendency, behavioral patterns, intent fusion, and product matching) collaborate through standardized vector interfaces, forming a complete processing chain from "acquisition to parsing to conflict detection to follow-up correction to fusion to matching to filtering." Each module supports the others rather than operating in isolation, significantly improving overall service efficiency. Attached Figure Description

[0027] Figure 1 This is a flowchart of an insurance service risk control method based on multi-agent collaboration, according to an embodiment of the present invention. Detailed Implementation

[0028] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0029] This invention proposes an insurance product matching method and system based on multi-dimensional reference, comprising the following steps: S1. Real-time acquisition of multimodal data, extraction of customer profile vectors, preprocessing of the acquired multimodal data, and generation of dynamic features; The multimodal data includes: text messages, voice messages, images, files, emoticons, retraction information, message interval time, time length changes, click behavior, and session status; The data collection method involves collecting multimodal data through a data collection intelligent agent; specifically, in this embodiment, multimodal data is collected in real time through the WeChat API. The method for extracting customer profile vectors is as follows: obtain customer profile vectors from the database and historical behavior records. ; The preprocessing method involves using a real-time stream processing engine to clean, time-align, and extract features from the collected multi-source data, and generating a dynamic feature vector at fixed intervals (30 seconds) and storing it in a real-time feature library. The dynamic features include: text features, emotional features, behavioral features, and interaction frequency features.

[0030] S2. Dynamic features are transformed into feature vectors and probability vectors; The feature vector includes semantic feature vectors. Emotional feature vector behavioral feature vector The probability vector includes a semantic intent distribution. Emotional distribution Distribution of behavioral patterns Specifically: S21. The semantic understanding agent transforms text features into semantic feature vectors. and probability vector semantic intent distribution and semantic intent tags ; The semantic understanding agent is based on a lightweight BERT (TinyBERT) real-time semantic parsing model, deployed using ONNXRuntime, and supports dynamic word vector updates. Dynamic word vectors utilize the Transformer's self-attention mechanism to dynamically adjust word embeddings based on the current dialogue context, allowing the same word to have different semantic representations in different contexts. The semantic understanding agent of this invention is domain-adjusted on an insurance dialogue dataset and employs INT8 quantization to reduce inference latency, ultimately outputting semantic feature vectors. Probability vector semantic intent distribution and semantic intent tags .

[0031] S22. The emotion-oriented agent transforms emotion features into emotion feature vectors. And probability vector sentiment distribution and emotional tags ; The emotional tendency intelligence integrates a multimodal input emotional classification network and adjusts the text sentiment using a pre-trained BERT-base-Chinese algorithm to achieve three main categories: positive, neutral, and negative. It also integrates emoji mapping features and speech tone features (speech rate, pauses). Specifically: For text sentiment analysis, bert-base-Chinese is used for fine-tuning, and the probability distribution of sentiment categories is output. For facial expression features, an expression-emotion mapping table is constructed, which maps emojis to emotion vectors and extracts meta-features such as expression location and frequency; To target speech intonation features, acoustic features such as fundamental frequency, energy, speech rate, and MFCC are extracted from speech messages, and speech intonation feature vectors are extracted using CNN. Features from the three modalities are weighted and fused using an attention mechanism to ultimately output a sentiment feature vector. And probability vector sentiment distribution and emotional tags .

[0032] S23. The behavioral pattern agent transforms behavioral features and interaction frequency features into behavioral feature vectors. and probability vector behavior pattern distribution and behavioral tags ; The behavioral pattern agent is based on an LSTM-based temporal pattern mining network. It takes historical dialogue sequences and current behavioral features as input, including: whether a product card was clicked, the duration of time spent viewing terms, whether a message was withdrawn, message length, response time, number of price inquiries, and the number of products compared. The LSTM network then outputs behavioral pattern encodings to identify customer types (price-sensitive, comparison-oriented, hesitant, decisive, and skeptical), and outputs a behavioral feature vector. and probability vector behavior pattern distribution and behavioral tags .

[0033] S3. The intent fusion agent performs conflict detection on the probability vector, obtains the result of the conflict detection, and selects to perform weighted fusion on the feature vector to output a comprehensive intent vector or trigger follow-up questions based on the result of the conflict detection. Furthermore, the result after conflict detection is the conflict coefficient C; S31. The intention is to fuse the intelligent agent to perform conflict detection on the probability vector and calculate the conflict coefficient C; Specifically, the semantic intent distribution is calculated using weighted relative entropy (KL divergence). Emotional distribution Distribution of behavioral patterns The difference between pairwise probability distributions, combined with the maximum probability class conflict indicator function. Quantify the conflict coefficient using the following formula : ; In the formula: Indicates KL divergence; This is the maximum probability category conflict indicator function, used when the semantic main category and the behavioral main category are inconsistent. =1, otherwise =0; S32. Different processing strategies are triggered based on the threshold range in which the conflict coefficient C falls; Specifically: Low conflict ( The probability vectors output by the semantic understanding agent, the sentiment agent, and the behavior pattern agent have highly consistent meanings, and are directly weighted and fused to output the final intent. China-US conflict ( The semantic understanding agent, the sentiment agent, and the behavior pattern agent have conflicting interpretations of their output probability vectors. This triggers a mild follow-up question to reduce the conflict level to a low-conflict range. The weights are then adjusted based on the responses before the vectors are fused and output. High conflict ( The semantic understanding agent, the sentiment agent, and the behavior pattern agent have serious discrepancies in their output probability vectors, which triggers forced follow-up questions. If the conflict drops to a low-conflict range after the follow-up questions, the weights are adjusted based on the responses and then the outputs are fused. If the conflict cannot be reduced to a low level after multiple follow-up questions or after reaching the follow-up question limit, the final comprehensive intent vector will directly adopt the output of the semantic understanding agent (instead of the weighted fusion result).

[0035] Specifically, the implementation of the follow-up questioning is as follows: when the conflict reaches a preset trigger threshold, the intent fusion agent dynamically generates follow-up questions based on the source and content of the conflict and a "conflict-question" mapping strategy, and sends the generated questions to the client.

[0036] Based on the output labels of the three agents—the semantic understanding agent, the sentiment-oriented agent, and the behavioral pattern agent—the semantic intent label is... Emotional tags Behavioral tags And the distribution of the conflict coefficient C, to determine the main sources of conflict. Define the conflict type encoding vector T=(t1,t2,t3,t4): in: t1 is and A predefined pair of contradictions is formed, representing a semantic-emotional conflict. If a conflict exists, then t1=1; otherwise, t1=0. t2 is and A predefined pair of contradictions is formed, representing a semantic-behavioral conflict. If a conflict exists, then t2=1; otherwise, t2=0. t3 is and A predefined pair of contradictions represents an emotional-behavioral conflict. If a conflict exists, then t3=1; otherwise, t3=0. t4 is , , A predefined pair of contradictions represents a complete conflict between semantics, emotion, and behavior. If a conflict exists, then t4 = 1; otherwise, t4 = 0. The predefined contradictions are constructed using domain knowledge; for example, the semantic "price inquiry" contradicts the emotional "negative", the semantic "purchase intention" contradicts the behavioral "comparative", and the emotional "positive" contradicts the behavioral "hesitant".

[0037] A conflict-problem mapping table is constructed, where each record contains the conflict type, the specific contradictory pair, and a problem template (including fillable slots). The mapping table is designed offline manually or learned using historical dialogue data; the problem template selection function is as follows: ; In the formula: A set of mapping tables; Cosine similarity between conflict type encoding and template encoding; This is the contradiction pair matching exponential function, which takes the value 1 when the current tag pair matches the contradiction pair defined in the template, and 0 otherwise.

[0038] The slots in the template (such as [Product Type], [Premium Range]) need to be dynamically extracted from the current dialogue context. The slot filling function is: ; The problem template can be formalized as follows: The template string is It contains K slots { The problem after filling in the blanks is: .

[0039] Finally, the generated question text is sent to the client, either immediately or after confirmation from the sales staff.

[0040] S4. The intent fusion agent fuses feature vectors into a comprehensive intent vector; S41. The intent fusion agent will integrate the semantic feature vectors output by the three agents: the semantic understanding agent, the sentiment agent, and the behavior pattern agent. Emotional feature vector behavioral feature vector Transformed to the same dimension d through a linear mapping, we obtain , ; ; ; In the formula: The scene feature mapping matrix; For the bias vector, and To complete the linear transformation; The scene feature mapping matrix; For the bias vector, and To complete the linear transformation; The scene feature mapping matrix; For the bias vector, and To complete the linear transformation; semantic feature vector The value obtained through linear mapping; For sentiment feature vectors The value obtained through linear mapping; behavioral feature vector The value obtained through linear mapping.

[0041] S42. Use an attention mechanism to calculate the attention weights corresponding to the semantic understanding agent, the sentiment-oriented agent, and the behavioral pattern agent; ; ; ; In the formula: Weights corresponding to semantic understanding agents; Weights corresponding to agents with emotional tendencies; The weights corresponding to the behavioral pattern agents; These are learnable weight vectors; This is a bias term (scalar); S33. Weighted fusion yields the final comprehensive intent vector; The weighted fusion calculation formula is as follows: ; In the formula: This is the overall intent vector.

[0042] S5, the product matching agent performs product matching based on the comprehensive intent vector and outputs the product matching result; S51, Generate customer query vector ; The product matching agent will integrate intent vectors. Emotional feature vector behavioral feature vector and customer profile vector Concatenate into a unified customer query vector ; through customer query vectors Indicates the customer's query intent; ; In the formula, This represents a vector concatenation operation; The dimensions after concatenation are: ; In the formula, D is the dimension of the concatenated customer query vector; The dimension of the comprehensive intent vector; The dimension of the sentiment feature vector; The dimension of the behavioral feature vector; Dimensions of the customer profile vector; S52, Generate product node vector ; The product node vector represents a product in the insurance product knowledge graph. The correspondence between the product node vector and the product in the insurance product knowledge graph is that one product node vector corresponds to one product in the insurance product knowledge graph. Specifically, product node vector The generation steps are as follows: In the insurance product knowledge graph, each product node The embedding vectors are updated through multi-layer heterogeneous graph convolution. Layer representation It is given by the following formula: ; In the formula: For the first The product node vector of the layer; A set of edge types; For type With products The set of connected neighboring nodes; The weight matrix is ​​a learnable weight matrix; The weight matrix is ​​a learnable weight matrix; Scene-aware attention weights; Furthermore, in the above formula The calculation formula is as follows: ; In the formula: This involves concatenating vectors. For scene feature vectors; For product nodes Feature vectors, subscripts Indicates product nodes; For neighboring nodes The feature vector is used to represent other types of nodes (such as keyword nodes and clause nodes) connected to the product node. The scene mapping matrix; edge type The corresponding weight matrix uses different linear transformations for different relationships (edge ​​types); This is the attention parameter vector; The superscript T indicates the transpose of a matrix or vector, used to calculate the inner product; Through the above methods L After layer propagation, the final product node vector is obtained. .

[0043] S53, Bilinear matching calculates the matching score; The customer query vector is calculated using a bilinear calculation method. With product node vector Match score ; ; In the formula: It is a learnable bilinear weight matrix; The Sigmoid activation function is used to compress the matching degree to the (0,1) interval; The superscript T indicates the transpose of a matrix or vector, used to calculate the inner product; The matching score is used to indicate the degree of match between the customer's intent and the product.

[0044] S54, based on the matching score, recalls products and outputs the product with the highest matching score and matching score; The product matching agent calculates all product and customer query vectors. Match score Sort the products by score from highest to lowest, recall the Top-K candidate products, filter them according to business rules, and finally output the m products with the highest matching degree and their matching scores. The business rules refer to non-model-based constraints used to filter candidate products after a product recall, ensuring that the recommended results meet actual business compliance and marketability. Specific business rules include: Product availability status: Only currently available products are retained (not discontinued, not removed from shelves); Customer eligibility verification: For example, age restrictions (some critical illness insurance policies are only available to people aged 0-55), geographical restrictions (some products cannot be purchased in certain regions), occupational restrictions, etc. Product exclusivity rule: The same customer cannot recommend two products with overlapping coverage (such as two million-dollar medical insurance policies) at the same time. Regulatory compliance requirements: For example, high-risk investment-type insurance cannot be recommended to the elderly; Sales personnel permissions: Some products can only be promoted by sales personnel of specific ranks; Deduplication of historical recommendations: Products that have been recommended to customers and explicitly rejected in the short term will not be recommended again; The business rule filtering is implemented by using a lightweight rule engine (such as Drools or EasyRules) or a configurable list of conditional judgments to chain rules together. The rule engine receives a list of candidate products, applies business rules, and outputs "pass" or "reject" for each rule. Products that pass all rules are output; products rejected by any rule are no longer included in the subsequent sorting.

[0045] The specific implementation plan is as follows: Check if the "is_on_sale" field in the product database is true; products that are discontinued, removed from the shelves, or whose pre-orders have ended are directly eliminated; then compare the customer profile with the product terms (e.g., critical illness insurance requires an age of ≤55 years old; some products are only available locally in certain regions), and eliminate products that do not meet the eligibility criteria; then determine if there is any overlap in coverage between candidate products (e.g., two different million-dollar medical insurance policies), and if both exist, retain the one with the higher matching degree and eliminate the lower-scoring product among mutually exclusive products. Mutually exclusive relationships are defined using "mutually exclusive edges" in the knowledge graph; finally, according to regulatory provisions (e.g., the State Administration of Financial Supervision and Administration of China...), ... The assessment process involves several steps: first, considering the restrictions imposed by the State Administration of Financial Supervision and Administration on the sale of high-risk products to the elderly; second, based on regulatory provisions (such as the restrictions imposed by the State Administration of Financial Supervision and Administration on the sale of high-risk products to the elderly), and again considering the customer's age and the product's risk level, non-compliant products are eliminated; third, checking whether the current salesperson's rank is on the authorized sales list for the product (e.g., some participating insurance policies are limited to senior agents); eliminating products for which the salesperson lacks the authority to sell; and fourth, checking whether the same product has been recommended to the customer recently (e.g., in the past 7 days) and the customer's feedback was "rejection" or "ignoring"—if so, eliminating products that have been repeatedly rejected.

[0046] Business rule filtering does not change the original matching score of a product; it only prunes the candidate set. The final output of m products represents the products with the highest matching score, while satisfying all business constraints.

[0047] Furthermore, this invention also proposes a system for matching insurance products based on multi-dimensional customer references, including a data acquisition module, a semantic understanding agent, an emotional tendency agent, a behavioral pattern agent, an intent fusion agent, and a product matching agent; the components communicate and connect with each other through standardized information interfaces.

[0048] The data acquisition module is used to collect multimodal data in real time, extract customer profile vectors, preprocess the collected multimodal data, and generate dynamic features. The acquisition method is to collect multimodal data in real time via the WeChat API. The customer profile vector is extracted from the database and historical behavior records. The preprocessing method involves using a real-time stream processing engine to clean, time-align, and extract features from the collected multi-source data, generating a dynamic feature vector at fixed intervals and storing it in a real-time feature library.

[0049] The semantic understanding agent is used to transform text features into semantic feature vectors. and probability vector semantic intent distribution and semantic intent tags The semantic understanding agent is based on a lightweight BERT (TinyBERT) real-time semantic parsing model, deployed using the ONNX Runtime, and supports dynamic word vector updates. Dynamic word vectors utilize the Transformer's self-attention mechanism to dynamically adjust word embeddings based on the current dialogue context, allowing the same word to have different semantic representations in different contexts. The semantic understanding agent of this invention is domain-adjusted on an insurance dialogue dataset and employs INT8 quantization to reduce inference latency, ultimately outputting semantic feature vectors. Probability vector semantic intent distribution and semantic intent tags .

[0050] The emotion-oriented agent is used to transform emotion features into emotion feature vectors. And probability vector sentiment distribution and emotional tags ; The sentiment-oriented intelligent system integrates a multimodal input sentiment classification network and adjusts the text sentiment using a pre-trained BERT-base-Chinese algorithm to achieve three major classifications: positive, neutral, and negative. It also integrates emoji mapping features and speech tone features (speech rate, pauses).

[0051] For text sentiment analysis, bert-base-Chinese is used for fine-tuning, and the probability distribution of sentiment categories is output. For facial expression features, an expression-emotion mapping table is constructed, which maps emojis to emotion vectors and extracts meta-features such as expression location and frequency; To target speech intonation features, acoustic features such as fundamental frequency, energy, speech rate, and MFCC are extracted from speech messages, and speech intonation feature vectors are extracted using CNN. Features from the three modalities are weighted and fused through an attention mechanism to ultimately output a sentiment feature vector. And probability vector sentiment distribution and emotional tags .

[0052] The behavioral pattern agent is used to transform behavioral features and interaction frequency features into behavioral feature vectors. and probability vector behavior pattern distribution and behavioral tags ; The behavioral pattern agent is based on an LSTM-based temporal pattern mining network. It takes historical dialogue sequences and current behavioral features as input, including: whether a product card was clicked, the duration of time spent viewing terms, whether a message was withdrawn, message length, response time, number of price inquiries, and the number of products compared. The LSTM network then outputs behavioral pattern encodings to identify customer types (price-sensitive, comparison-oriented, hesitant, decisive, and skeptical), and outputs a behavioral feature vector. and probability vector behavior pattern distribution and behavioral tags .

[0053] The intent fusion intelligent body includes a conflict detection module, a question generation module, and a vector fusion module; The conflict detection module is used to detect conflicts in the probability vector, calculate the conflict coefficient C, and trigger different processing strategies based on the threshold range in which the conflict coefficient C falls. The conflict detection module uses weighted relative entropy (KL divergence) to calculate the semantic intent distribution. Emotional distribution Distribution of behavioral patterns The difference between pairwise probability distributions, combined with the maximum probability class conflict indicator function. Quantify the conflict coefficient using the following formula : ; In the formula: Indicates KL divergence; This is the maximum probability category conflict indicator function, used when the semantic main category and the behavioral main category are inconsistent. =1, otherwise =0.

[0055] The question generation module is used to generate follow-up questions when there are discrepancies in the meaning of the probability vectors output by the semantic understanding agent, the sentiment agent, and the behavior pattern agent, and then sends the follow-up questions to the client. The follow-up questioning method is as follows: when the conflict reaches a preset trigger threshold, the question generation module dynamically generates follow-up questions based on the source and content of the conflict and the "conflict-question" mapping strategy, and sends the generated questions to the client.

[0056] Based on the output labels of three agents—the semantic understanding agent, the sentiment-oriented agent, and the behavioral pattern agent—the semantic intent label is... Emotional tags Behavioral tags And the distribution of the conflict coefficient C, to determine the main sources of conflict. Define the conflict type encoding vector T=(t1,t2,t3,t4): in: t1 is and A predefined pair of contradictions represents a semantic-emotional conflict; if a conflict exists, t1=1; otherwise, t1=0. t2 is and A predefined pair of contradictions represents a semantic-behavioral conflict; if a conflict exists, then t2 = 1; otherwise, t2 = 0. t3 is and A predefined pair of contradictions represents an emotional-behavioral conflict; if a conflict exists, t3 = 1; otherwise, t3 = 0. t4 is , , A predefined pair of contradictions represents a complete conflict between semantics, emotion, and behavior. If a conflict exists, then t4 = 1; otherwise, t4 = 0. The predefined contradictions are constructed using domain knowledge.

[0057] A conflict-problem mapping table is constructed, where each record contains the conflict type, the specific contradictory pair, and a problem template (including fillable slots). The mapping table is designed offline manually or learned using historical dialogue data; the problem template selection function is as follows: ; In the formula: A set of mapping tables; Cosine similarity between conflict type encoding and template encoding; This is the contradiction pair matching exponential function, which takes the value 1 if the current tag pair matches the contradiction pair defined in the template, and 0 otherwise.

[0058] The slots in the template are dynamically extracted from the current dialogue context, and the slot filling function is: ; The problem template can be formalized as follows: The template string is named "template" and contains K slots. The problem after filling in the blanks is: .

[0059] The vector fusion module is used to fuse feature vectors into a comprehensive intent vector; the vector fusion module combines the semantic feature vectors output by the semantic understanding agent, the sentiment agent, and the behavior pattern agent respectively. Sentiment feature vector behavioral feature vector Transformed to the same dimension d through a linear mapping, we obtain , ; ; ; In the formula: The scene feature mapping matrix; For the bias vector, and To complete the linear transformation; The scene feature mapping matrix; For the bias vector, and To complete the linear transformation; The scene feature mapping matrix; For the bias vector, and To complete the linear transformation; semantic feature vector The value obtained through linear mapping; For sentiment feature vectors The value obtained through linear mapping; behavioral feature vector The value obtained through linear mapping; The attention mechanism is used to calculate the attention weights corresponding to the semantic understanding agent, the sentiment-oriented agent, and the behavioral pattern agent; ; ; ; In the formula: Weights corresponding to semantic understanding agents; Weights corresponding to agents with emotional tendencies; The weights corresponding to the behavioral pattern agents; These are learnable weight vectors; This is a bias term (scalar); The weighted fusion yields the final comprehensive intent vector; The weighted fusion calculation formula is as follows: ; In the formula: This is the overall intent vector.

[0060] The product matching intelligent body includes a product matching module, an insurance product knowledge graph, and a rule filtering module; The product matching module is used to generate customer query vectors. and product node vector And calculate all product and customer query vectors. Match score It outputs the products filtered by business rules and their matching scores; one of the product node vectors. This corresponds to a product in the insurance product knowledge graph. The product matching module will integrate intent vectors. Sentiment feature vector behavioral feature vector and customer profile vectors Concatenate into a unified customer query vector ; through customer query vectors Indicates the customer's query intent; ; In the formula, This represents a vector concatenation operation; The dimensions after concatenation are: ; In the formula, D is the dimension of the concatenated customer query vector; This refers to the dimension of the comprehensive intent vector; The dimension of the sentiment feature vector; The dimension of the behavioral feature vector; Dimensions of the customer profile vector; Meanwhile, in the insurance product knowledge graph, each product node The embedding vectors are updated through multi-layer heterogeneous graph convolution. Layer representation It is given by the following formula: ; In the formula: For the first The product node vector of the layer; A set of edge types; For type With products The set of connected neighboring nodes; The weight matrix is ​​a learnable weight matrix; The weight matrix is ​​a learnable weight matrix; Scene-aware attention weights; In the above formula The calculation formula is as follows: ; In the formula: This involves concatenating vectors. For scene feature vectors; For product nodes Feature vectors, subscripts Indicates product nodes; Neighboring nodes The feature vector is used to represent other types of nodes (such as keyword nodes and clause nodes) connected to the product node. This is the scene mapping matrix; edge type The corresponding weight matrix uses different linear transformations for different relationships (edge ​​types); This is the attention parameter vector; The superscript T indicates the transpose of a matrix or vector, and is used to calculate the inner product.

[0061] Through the above methods L After layer propagation, the final product node vector is obtained. .

[0062] The customer query vector is calculated using a bilinear calculation method. With product node vector Match score ; ; In the formula: It is a learnable bilinear weight matrix; The Sigmoid activation function is used to compress the matching degree to the (0,1) interval; The superscript T indicates the transpose of a matrix or vector, used to calculate the inner product; The matching score is used to indicate the degree of match between the customer's intent and the product.

[0063] The rule filtering module is used to filter candidate products by constraint rules after a product recall. The rules module checks if the "is_on_sale" field in the product database is true. Products that are discontinued, removed from shelves, or whose pre-orders have ended are directly removed. Next, it compares the customer profile with the product terms and conditions, removing ineligible products. It then determines if there is overlap in coverage between candidate products; if both exist, the one with the higher matching degree is retained, and the lower-scoring product among mutually exclusive products is removed. Mutually exclusive relationships are defined using "mutually exclusive" edges in the knowledge graph. Next, it judges compliance based on regulatory requirements, customer age, and product risk level, removing non-compliant products. It then checks if the current salesperson's rank is on the authorized sales list for the product; products without sales authority are removed. Finally, it checks if the same product has been recommended to the customer recently, and if the customer's feedback is "rejected" or "ignored"; if so, the duplicate rejected product is removed. Products that pass all rules are output; products removed by any rule are no longer included in subsequent sorting.

[0064] Compared with existing technologies, the beneficial effects of this invention are as follows: I. This invention constructs an insurance product knowledge graph and designs a scene-aware heterogeneous graph neural network (HGNN) to update product node embeddings. During the message passing process of the HGNN, scene feature vectors extracted from the dialogue context are introduced as attention biases to dynamically adjust the aggregation weights of different neighboring nodes, enabling product embeddings to perceive the specific scenario of the current dialogue. Combined with a bilinear matching function to calculate the matching score between the customer query vector and the product embedding, the system can recall candidate products that best match the customer's real-time needs from a massive number of products, solving the problems of lagging traditional static profile recommendations and their inability to adapt to changes in dialogue context.

[0065] Second, this invention uses three intelligent agents—semantic understanding, sentiment tendency, and behavioral pattern—to analyze multimodal interaction data of customers in parallel. It employs weighted relative entropy (KL divergence) and a maximum probability category conflict indicator function to quantitatively calculate the conflict coefficient, achieving real-time quantitative assessment of the consistency among customer language, sentiment, and behavior. When the conflict coefficient is in the medium to high threshold range, the system proactively triggers a follow-up questioning mechanism. Through a "conflict-question" mapping strategy, it dynamically generates targeted questions to guide the customer to clarify their true intentions. Based on the responses, the system adjusts the fusion weights of each modality until the conflict level is reduced to a low level. This closed-loop verification mechanism effectively solves the decision-making dilemma of traditional single-modality or simple weighted fusion when information contradictions arise, significantly improving the reliability and confidence of the comprehensive intent vector, laying a solid foundation for subsequent accurate product matching.

[0066] Third, this invention also includes a follow-up questioning mechanism. The core of follow-up questioning is to proactively eliminate ambiguity, enabling the system to obtain clearer signals from subsequent customer responses, thereby correcting inconsistencies in the output meanings among the three agents: semantic understanding agent, sentiment agent, and behavioral pattern agent. When the judgments of semantics, sentiment, and behavior diverge, the system cannot determine the customer's true intention, and forced weighted fusion may lead to incorrect recommendations. Through follow-up questioning, the system can guide customers to clarify their needs and supplement information, thereby reducing the conflict coefficient. The effect of reducing the conflict coefficient is to improve the reliability and confidence of the comprehensive intent vector. When the conflict coefficient drops to a low level, the meanings of the outputs of the three dimensions of semantics, sentiment, and behavior are consistent. At this time, the intent vector obtained by weighted fusion can accurately reflect the customer's true state, and subsequent product matching is also based on this high-confidence input, thereby significantly improving the accuracy of recommendations, reducing false pushes, and avoiding business risks caused by contradictory information (such as forcibly selling to hesitant customers or ignoring customers' negative emotions). In addition, low conflict also means that the system's understanding of customers is more interpretable, facilitating subsequent auditing and optimization.

[0067] Fourth, after the intelligent agent matches and recalls Top-K candidate products, this invention further introduces a business rule filtering layer. The rule engine sequentially executes non-model-based constraints such as product sales status, customer qualification verification, product mutual exclusion, regulatory compliance, sales personnel permissions, and historical recommendation deduplication, retaining only the m products that simultaneously satisfy all business rules and have the highest matching degree. This design effectively avoids invalid recommendations, regulatory compliance risks, internal product conflicts, and duplicate recommendations of rejected products, significantly improving the actual conversion rate and customer satisfaction of the recommendation results, while reducing the company's compliance risks and sales disputes.

[0068] Fifth, this invention incorporates non-textual modal information such as speech tone, emoticons, and behavioral sequences into the feature acquisition and fusion scope, and dynamically weights each modal feature through an attention mechanism, breaking through the limitations of traditional methods that rely solely on text keyword matching. Simultaneously, the various intelligent agents (semantic understanding, sentiment tendency, behavioral patterns, intent fusion, and product matching) collaborate through standardized vector interfaces, forming a complete processing chain from "acquisition to parsing to conflict detection to follow-up correction to fusion to matching to filtering." Each module supports the others rather than operating in isolation, significantly improving overall service efficiency.

[0069] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A method for matching insurance products based on multi-dimensional customer references, characterized in that: Includes the following steps: S1. Real-time acquisition of multimodal data, extraction of customer profile vectors, preprocessing of the acquired multimodal data, and generation of dynamic features; S2. Dynamic features are transformed into feature vectors and probability vectors; S3. The intent fusion agent performs conflict detection on the probability vector, obtains the result of the conflict detection, and selects to perform weighted fusion on the feature vector to output a comprehensive intent vector or trigger follow-up questions based on the result of the conflict detection. S4. The intent fusion agent fuses feature vectors into a comprehensive intent vector; S5, the product matching agent performs product matching based on the comprehensive intent vector and outputs the product matching result.

2. The matching method for insurance products based on multi-dimensional customer reference as described in claim 1, characterized in that: Step S2 includes the following steps: S21. The semantic understanding agent transforms text features into semantic feature vectors. and probability vector semantic intent distribution and semantic intent tags ; S22. The emotion-oriented agent transforms emotion features into emotion feature vectors. And probability vector sentiment distribution and emotional tags ; S23. The behavioral pattern agent transforms behavioral features and interaction frequency features into behavioral feature vectors. and probability vector behavior pattern distribution and behavioral tags .

3. The matching method for insurance products based on multi-dimensional customer reference as described in claim 2, characterized in that: Step S3 includes the following steps: The result after conflict detection is the conflict coefficient C; Calculate semantic intent distribution using weighted relative entropy. Emotional distribution Distribution of behavioral patterns The difference between pairwise probability distributions, combined with the maximum probability class conflict indicator function. Quantify the conflict coefficient using the following formula : ； In the formula: Indicates KL divergence; This is the maximum probability category conflict indicator function, used when the semantic main category and the behavioral main category are inconsistent. =1, otherwise =0; S32. Different processing strategies are triggered based on the threshold range in which the conflict coefficient C falls; (1) When the conflict coefficient C is in the low conflict threshold range, the weighted fusion output is the final intention; (2) When the conflict coefficient C is in the medium conflict threshold range, a mild follow-up question is triggered, the conflict level is reduced to the low conflict range, and the weight is adjusted according to the response before fusion and output. (3) When the conflict coefficient C is in the medium-high threshold range, a forced follow-up question is triggered. If the conflict drops to the low conflict range after the follow-up questioning ends, the weight is adjusted according to the response and then the output is fused. If the conflict cannot be reduced to a low level after multiple follow-up questions or after reaching the upper limit of follow-up questions, the final integrated intent vector is directly adopted from the output of the semantic understanding agent.

4. The matching method for insurance products based on multi-dimensional customer reference as described in claim 3, characterized in that: Step S4 includes the following steps: S41. The intent fusion agent will integrate the semantic feature vectors output by the three agents: the semantic understanding agent, the sentiment agent, and the behavior pattern agent. Emotional feature vector behavioral feature vector Transformed to the same dimension d through a linear mapping, we obtain ; S42. Use the attention mechanism to calculate the attention weights corresponding to the semantic understanding agent, the sentiment-oriented agent, and the behavioral pattern agent; S43. Weighted fusion yields the final integrated intent vector.

5. The insurance product matching method based on multi-dimensional reference as described in claim 4, characterized in that: Step S5 includes the following steps: S51, Generate customer query vector ; The product matching agent will integrate intent vectors. Emotional feature vector behavioral feature vector and customer profile vector Concatenate into a unified customer query vector ; through customer query vectors Indicates the customer's query intent; ； In the formula, This represents a vector concatenation operation; The dimensions after concatenation are: ； In the formula, D is the dimension of the concatenated customer query vector; This refers to the dimension of the comprehensive intent vector; The dimension of the sentiment feature vector; The dimension of the behavioral feature vector; Dimensions of the customer profile vector; S52, Generate product node vector ; The product node vector represents a product in the insurance product knowledge graph. The correspondence between the product node vector and the product in the insurance product knowledge graph is that one product node vector corresponds to one product in the insurance product knowledge graph. Product Node Vector The generation steps are as follows: In the insurance product knowledge graph, each product node The embedding vectors are updated through multi-layer heterogeneous graph convolution. Layer representation It is given by the following formula: ； In the formula: For the first The product node vector of the layer; A set of edge types; For type With products The set of connected neighboring nodes; The weight matrix is ​​a learnable weight matrix; The weight matrix is ​​a learnable weight matrix; Scene-aware attention weights; Furthermore, in the above formula The calculation formula is as follows: ； In the formula: This involves concatenating vectors. For scene feature vectors; For product nodes Feature vectors, subscripts Indicates product nodes; Neighboring nodes The feature vector is used to represent other types of nodes connected to the product node; This is the scene mapping matrix; edge type The corresponding weight matrix uses different linear transformations for different relationships; This is the attention parameter vector; The superscript T indicates the transpose of a matrix or vector, used to calculate the inner product; Through the above methods L After layer propagation, the final product node vector is obtained. ; S53, Bilinear matching calculates the matching score; The customer query vector is calculated using a bilinear calculation method. With product node vector Match score ; ； In the formula: It is a learnable bilinear weight matrix; The Sigmoid activation function is used to compress the matching degree to the (0,1) interval; The superscript T denotes the transpose of a matrix or vector, used for calculating the inner product. The matching score is used to represent the degree of matching between customer intent and product; S54, based on the matching score, recalls products and outputs the product with the highest matching score and matching score; The product matching agent calculates all product and customer query vectors. Match score Sort the products by score from highest to lowest, recall the Top-K candidate products, filter them according to business rules, and finally output the m products with the highest matching degree and their matching scores.

6. The matching method for insurance products based on multi-dimensional customer reference as described in claim 3, characterized in that: The follow-up question is generated dynamically by the intelligent agent based on the source and content of the conflict and using a "conflict-problem" mapping strategy when the conflict reaches a preset trigger threshold.

7. A matching system for insurance products based on multi-dimensional customer references, characterized in that: It includes a data acquisition module, a semantic understanding agent, an emotion tendency agent, a behavior pattern agent, an intent fusion agent, and a product matching agent; the components communicate and connect with each other through standardized information interfaces; the data acquisition module is used to collect multimodal data in real time, extract customer profile vectors, preprocess the collected multimodal data, and generate dynamic features; The semantic understanding agent is used to transform text features into semantic feature vectors. and probability vector semantic intent distribution and semantic intent tags ; The emotion-oriented agent is used to transform emotion features into emotion feature vectors. And probability vector sentiment distribution and emotional tags The behavioral pattern agent is used to transform behavioral features and interaction frequency features into behavioral feature vectors. and probability vector behavior pattern distribution and behavioral tags The intent fusion intelligent body includes a conflict detection module, a question generation module, and a vector fusion module. The conflict detection module detects conflicts in the probability vectors, calculates the conflict coefficient C, and triggers different processing strategies based on the threshold range of the conflict coefficient C. The question generation module generates follow-up questions when there are discrepancies in the meaning of the probability vectors output by the semantic understanding agent, the sentiment agent, and the behavior pattern agent, and sends these follow-up questions to the client. The vector fusion module fuses the feature vectors into a comprehensive intent vector. The product matching intelligent body includes a product matching module, an insurance product knowledge graph, and a rule filtering module. The product matching module generates customer query vectors. and product node vector And calculate all product and customer query vectors. Match score The product node vector is then output, along with the products filtered according to business rules and their matching scores. It corresponds to the products in the insurance product knowledge graph; the rule filtering module is used to filter candidate products through constraint rules after product recall.

Images