A multi-dimensional matching method, system, and equipment for insurance products based on behavioral profiling.

Abstract

This invention discloses a method, system, and device for multi-dimensional matching of insurance products based on behavioral profiling, belonging to the field of data processing technology. The method includes: extracting multi-turn dialogue statements from users, extracting semantic fuzziness feature vectors, performing fuzzy scoring based on the semantic fuzziness feature vectors to obtain fuzzy score values; constructing a field decomposition granularity template library, inputting the fuzzy score values ​​into the field decomposition granularity template library to obtain the matching field decomposition granularity, and parsing the multi-dimensional intent vectors of the multi-turn dialogue statements; constructing a user behavior profile according to the multi-dimensional intent vectors, performing matching retrieval in an insurance product database, and outputting a set of matched insurance products. This invention solves the technical problem that existing insurance recommendation methods cannot accurately handle the fuzziness of user dialogues, leading to low accuracy in demand analysis and product matching. It achieves the technical effect of improving the accuracy of demand analysis and product matching through multi-granularity field intent guidance analysis using fuzzy scoring.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method, system, and device for multi-dimensional matching of insurance products based on behavioral profiling. Background Technology

[0002] In modern insurance services, as consumer demands become increasingly complex and diverse, traditional insurance product recommendation systems are gradually revealing numerous shortcomings. Most existing systems rely on fixed questionnaires filled out by users or simple keyword matching, making it difficult to capture the complex needs expressed by users in natural language interactions. This is especially true when users use ambiguous expressions in their conversations, such as words like "probably" or "maybe," or frequently change topics; existing systems often fail to accurately understand the user's true intentions. Furthermore, the high complexity and diversity of insurance products themselves, with numerous types and varying terms, further increases the difficulty of accurate recommendations. Summary of the Invention

[0003] This application provides a method, system, and device for multi-dimensional matching of insurance products based on behavioral profiles, which is used to solve the technical problem that existing insurance recommendation methods cannot accurately handle the ambiguity of user dialogue, resulting in low accuracy in demand analysis and product matching.

[0004] The first aspect of this application provides a multi-dimensional matching method for insurance products based on behavioral profiles. The method includes: extracting multi-turn dialogue statements from users; extracting semantic fuzziness feature vectors from the multi-turn dialogue statements; performing fuzzy scoring based on the semantic fuzziness feature vectors to obtain fuzzy score values; wherein the semantic fuzziness feature vectors at least include the proportion of high-frequency fuzzy words, semantic drift degree, number of missing fields, and context jump frequency; constructing a field decomposition granularity template library; inputting the fuzzy score values ​​into the field decomposition granularity template library to obtain matching field decomposition granularity; parsing the multi-dimensional intent vectors of the multi-turn dialogue statements using the matching field decomposition granularity; constructing a user behavior profile according to the multi-dimensional intent vectors; calling the user behavior profile to perform matching retrieval in an insurance product database; and outputting a set of matched insurance products.

[0005] A second aspect of this application provides a multi-dimensional matching system for insurance products based on behavioral profiles. The system includes: a fuzzy scoring module, which extracts multi-turn dialogue statements from users, extracts semantic fuzziness feature vectors from the multi-turn dialogue statements, performs fuzzy scoring based on the semantic fuzziness feature vectors, and obtains a fuzzy score value. The semantic fuzziness feature vectors at least include the proportion of high-frequency fuzzy words, semantic drift, number of missing fields, and context jump frequency; a multi-dimensional intent parsing module, which constructs a field decomposition granularity template library, inputs the fuzzy score value into the field decomposition granularity template library to obtain the matching field decomposition granularity, and parses the multi-dimensional intent vector of the multi-turn dialogue statements using the matching field decomposition granularity; and a product matching retrieval module, which constructs a user behavior profile according to the multi-dimensional intent vector, calls the user behavior profile to perform matching retrieval in an insurance product database, and outputs a set of matched insurance products.

[0006] A third aspect of this application provides an electronic device comprising: a processor coupled to a memory for storing a program that, when executed by the processor, causes the system to perform the method described in any of the first aspects.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] This application provides a method, system, and device for multi-dimensional matching of insurance products based on behavioral profiles, which relates to the field of data processing technology. It extracts semantically ambiguous feature vectors from multi-turn user dialogues and scores them, dynamically selects the granularity of field decomposition to parse multi-dimensional intent vectors, constructs user behavior profiles, and uses a vector similarity function to accurately match insurance products in an insurance product database. This achieves multi-dimensional and accurate insurance product recommendations, solving the technical problem that existing insurance recommendation methods cannot accurately handle the ambiguity of user dialogues, leading to low accuracy in demand analysis and product matching. It achieves the technical effect of improving the accuracy of demand analysis and product matching by using fuzzy scoring for multi-granularity field intent guidance analysis to construct user behavior profiles for matching insurance products. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0010] Figure 1A schematic diagram of the multi-dimensional matching method for insurance products based on behavioral profiling provided in this application embodiment;

[0011] Figure 2 A schematic diagram of the structure of a multi-dimensional matching system for insurance products based on behavioral profiling provided in this application embodiment;

[0012] Figure 3 This application provides a schematic diagram of the structure of an electronic device.

[0013] Figure labeling: Fuzzy scoring module 11, multi-dimensional intent parsing module 12, product matching retrieval module 13, line protection module 16, electronic device 300, memory 301, processor 302, communication interface 303, bus architecture 304. Detailed Implementation

[0014] This application provides a method, system, and device for multi-dimensional matching of insurance products based on behavioral profiles, which is used to solve the technical problem that existing insurance recommendation methods cannot accurately handle the ambiguity of user dialogue, resulting in low accuracy in demand analysis and product matching.

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0016] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0017] Example 1, as Figure 1 As shown, this application provides a multi-dimensional matching method for insurance products based on behavioral profiles, which includes:

[0018] P10: Extract user multi-turn dialogue statements, extract semantic fuzziness feature vectors from the multi-turn dialogue statements, perform fuzzy scoring based on the semantic fuzziness feature vectors, and obtain fuzziness score values. The semantic fuzziness feature vectors include at least the proportion of high-frequency fuzzy words, semantic drift degree, number of missing fields, and context jump frequency.

[0019] Furthermore, based on the semantic fuzziness feature vector, fuzzy scoring is performed to obtain a fuzziness score value. Step P10 in this embodiment of the application further includes:

[0020] P11: Standardize the semantic ambiguity feature vector to output a standardized semantic ambiguity feature vector; P12: Construct an ambiguity scoring function, which is obtained by training the training corpus data using a BERT word segmenter and a regression head network. The training corpus data includes multi-turn dialogue sentence samples, semantic ambiguity feature vectors extracted from the multi-turn dialogue sentence samples, and ambiguity score annotation values; P13: Input the standardized semantic ambiguity feature vector into the ambiguity scoring function to perform ambiguity scoring and obtain the ambiguity score value.

[0021] It should be understood that semantic fuzziness feature vectors are extracted from user multi-turn dialogue statements, and fuzzy scoring is performed based on these feature vectors to quantify the degree of fuzziness in the user dialogue. Specifically, it is first necessary to extract user multi-turn dialogue statements, which are natural language texts generated during the interaction between the user and the system, containing information such as the user's needs, preferences, and historical behavioral data. In order to perform quantitative analysis on these dialogue statements, semantic fuzziness feature vectors need to be extracted. A semantic fuzziness feature vector is a multi-dimensional vector used to characterize the degree of fuzziness in a user dialogue, and its feature dimensions can reflect the fuzziness of the dialogue from different perspectives. In this application, the semantic fuzziness feature vector includes at least four feature dimensions: the proportion of high-frequency fuzzy words, semantic drift, the number of missing fields, and the frequency of context jumps.

[0022] For example, the proportion of high-frequency vague words is calculated first. Vague words refer to words whose semantics are not clear and have multiple interpretations or meanings, such as "probably," "possibly," and "more or less." The proportion of high-frequency vague words refers to the ratio of the frequency of vague words appearing in a user's dialogue to the total vocabulary. The higher this ratio, the stronger the ambiguity of the user's dialogue. For example, if a user frequently uses words such as "probably" and "possibly" in a dialogue, their proportion of high-frequency vague words is high, and the dialogue is also more ambiguous.

[0023] Secondly, semantic drift is calculated. Semantic drift refers to the degree to which the semantic direction expressed by a user shifts or changes during a conversation. For example, a user might mention "health insurance" at the beginning of the conversation but gradually shift to "critical illness insurance" in later parts. Semantic drift is measured by calculating the difference in similarity between different semantic units in the conversation. For example, a pre-trained language model (such as BERT) can be used to semantically encode each round of the conversation, and then the semantic similarity between adjacent rounds can be calculated. The greater the difference in semantic similarity, the higher the semantic drift, and the greater the ambiguity of the conversation.

[0024] Next, the number of missing fields is calculated. The number of missing fields refers to the number of key information fields missing from the user's dialogue. In user dialogue, some key information may be missing. For example, when describing insurance needs, the user may not explicitly mention important information such as the insurance amount or insurance period. Predefined field templates can be used to detect whether certain key fields are missing from the user's dialogue. The more missing fields, the higher the ambiguity of the dialogue.

[0025] Finally, the frequency of context jumps is calculated. Context jumps refer to the frequency with which a user jumps from one topic to another during a conversation. For example, a user might be discussing insurance products and suddenly bring up the insurance claims process. This can be measured by calculating the ratio of the number of topic jumps in the conversation to the total number of conversation rounds. Specifically, topic detection algorithms can be used to identify different topics in the conversation and count the number of topic jumps. A higher frequency of context jumps indicates poorer coherence and greater ambiguity in the conversation.

[0026] After extracting the semantic ambiguity feature vectors, feature vector standardization is performed to ensure comparability of values ​​across different feature dimensions. Feature vector standardization involves adjusting the values ​​of each dimension of the feature vector to the same dimensional range through mathematical transformations. For example, the mean of each feature dimension can be subtracted, and then divided by its standard deviation. After standardization, the output standardized semantic ambiguity feature vectors more accurately reflect the ambiguity of user dialogue.

[0027] Next, a fuzzy scoring function is constructed. A fuzzy scoring function is a mathematical model used to map standardized semantic fuzziness feature vectors to fuzzy score values. In this application, the fuzzy scoring function is obtained by training the training corpus data using a BERT word segmenter and a regression head network. The BERT word segmenter is a pre-trained language model based on the Transformer architecture, capable of efficiently segmenting natural language text and extracting semantic information. The regression head network is a neural network structure used for regression tasks, capable of predicting the corresponding numerical output based on the input feature vector. The training corpus data includes multi-turn dialogue sentence samples, semantic fuzziness feature vectors extracted from these samples, and manually annotated fuzzy score values. By using this training corpus data to train the fuzzy scoring function, the function can learn the mapping relationship between different semantic fuzziness feature vectors and fuzzy scores, thereby accurately scoring the fuzziness of user dialogues in subsequent applications.

[0028] Finally, the standardized semantic fuzziness feature vector is input into the fuzziness scoring function for fuzziness scoring, yielding a fuzziness score value. The fuzziness score value can be a numerical value between 0 and 1, used to quantify the degree of fuzziness in the user's dialogue. The closer the score value is to 1, the higher the fuzziness of the user's dialogue; the closer it is to 0, the lower the fuzziness of the dialogue. In this way, the fuzziness of multi-turn user dialogue statements can be quantitatively evaluated, providing an important basis for subsequent field decomposition granularity matching and insurance product matching.

[0029] Furthermore, in constructing the fuzzy scoring function, step P12 of this embodiment also includes:

[0030] P12-1: The BERT segmenter is used to semantically encode the input text of the multi-turn dialogue sentence samples and output the encoded semantic ambiguity feature vector. The regression head network is used to map the semantic ambiguity feature vector to ambiguity score values ​​and output ambiguity score mapping values. P12-2: A loss function is introduced to perform multiple rounds of iterative training on the ambiguity score annotation data and the ambiguity score annotation values ​​until the error between the ambiguity score mapping data and the ambiguity score annotation data in the current iteration round is less than a preset threshold, thus obtaining a converged ambiguity score function.

[0031] Specifically, the fuzziness scoring function is obtained by training the training corpus data using a BERT word segmenter and a regression head network. In constructing the fuzziness scoring function, the BERT word segmenter first performs semantic encoding on the input text of multi-turn dialogue sentence samples, outputting the encoded semantic fuzziness feature vector. The BERT word segmenter is a pre-trained language model based on the Transformer architecture, capable of transforming natural language text into high-dimensional semantic vectors, capturing the contextual information and semantic features of the text. Through the encoding of the BERT word segmenter, multi-turn dialogue sentence samples are transformed into semantic fuzziness feature vectors, providing the basic data for subsequent fuzziness scoring.

[0032] Next, a regression head network is used to map the semantic ambiguity feature vectors to ambiguity rating values, outputting the ambiguity rating mapping values. A regression head network is a neural network structure specifically designed for regression tasks, aiming to map input feature vectors to a specific numerical output. By learning the relationship between semantic ambiguity feature vectors and ambiguity ratings, the regression head network can accurately predict the ambiguity rating of a user's dialogue. The synergy between the BERT word segmenter and the regression head network enables the system to extract semantic features from natural language text and transform them into specific ambiguity ratings, providing technical support for subsequent training and applications.

[0033] To ensure that the fuzzy scoring function accurately maps the relationship between semantically fuzzy feature vectors and fuzzy scores, this application introduces a loss function and performs multiple rounds of iterative training on fuzzy score annotation data and fuzzy score annotation values. The loss function measures the difference between the model output and the true annotations; in this application, it is used to measure the error between the mapped fuzzy score values ​​and the fuzzy score annotation values. By minimizing the value of the loss function, the output of the fuzzy scoring function can be made closer to the true annotations, thereby improving the accuracy and reliability of the model.

[0034] In actual training, fuzzy rating labeled data is used as input. Forward propagation is performed using a BERT word segmenter and regression head network to obtain fuzzy rating mapping values. Then, a loss function is used to calculate the error between the fuzzy rating mapping values ​​and the fuzzy rating labeled values, and the model parameters are updated through backpropagation. This process is repeated for multiple iterations until the error between the fuzzy rating mapping values ​​and the fuzzy rating labeled values ​​in the current iteration is less than a preset threshold. The preset threshold is an error range set according to the specific application scenario. When the error is less than this threshold, the model can be considered to have converged, achieving the expected training effect. Through this multi-round iterative training method, the fuzzy rating function can continuously optimize its parameters, improve its ability to fit the relationship between semantically fuzzy feature vectors and fuzzy ratings, and ultimately obtain a converged fuzzy rating function, providing an accurate fuzzy rating tool for insurance product recommendation systems.

[0035] P20: Construct a field decomposition granularity template library, input the fuzziness score into the field decomposition granularity template library to obtain the matching field decomposition granularity, and use the matching field decomposition granularity to parse the multi-dimensional intent vector of the multi-turn dialogue statement.

[0036] The field decomposition granularity template library stores multiple field decomposition granularity templates, which correspond to multiple pre-defined granularity levels. The pre-defined granularity levels include multiple fuzzy score value ranges. Each field decomposition granularity template selects a set of subordinate decomposition fields at the corresponding granularity from the decomposition field set. The decomposition field set includes fields such as insurance type, insured object, usage scenario, coverage period, budget range, and additional liability.

[0037] Optionally, a field decomposition granularity template library can be constructed, and an appropriate field decomposition granularity can be selected based on the fuzziness score to parse the multidimensional intent vector of multi-turn dialogue statements.

[0038] First, a field decomposition granularity template library is constructed. This library stores multiple field decomposition granularity templates, each corresponding to a granularity level. These granularity levels are predefined, and each granularity level corresponds to a fuzziness score range. For example, the granularity levels can be divided into "coarse-grained," "medium-grained," and "fine-grained," each corresponding to different fuzziness score ranges. Specifically, "coarse-grained" may correspond to a range with higher fuzziness scores, such as [0.7, 1.0]; "medium-grained" corresponds to a range with medium fuzziness scores, such as [0.4, 0.7]; and "fine-grained" corresponds to a range with lower fuzziness scores, such as [0.0, 0.4].

[0039] Each field decomposition granularity template in the field decomposition granularity template library selects a set of subordinate decomposition fields at the corresponding granularity from the decomposition field set. The decomposition field set includes, but is not limited to, the following fields: insurance type, insured object, usage scenario, coverage period, budget range, and additional liability fields. These fields cover all aspects that users may involve in expressing their insurance needs. For example, the insurance type field may include health insurance, life insurance, property insurance, etc.; the insured object field may include individuals, family members, enterprises, etc.; the usage scenario field may include daily protection, travel protection, specific disease protection, etc.; the coverage period field may include short-term (e.g., one year), long-term (e.g., lifetime), etc.; the budget range field may include low budget, medium budget, high budget, etc.; and the additional liability field may include additional liabilities such as accidental injury, critical illness, and dividends.

[0040] Once the system obtains the fuzziness score, it inputs it into the field decomposition granularity template library. Based on the range to which the fuzziness score belongs, the system selects the corresponding field decomposition granularity template. For example, if the fuzziness score is 0.8, the "coarse-grained" template is selected; if the fuzziness score is 0.5, the system selects the "medium-grained" template; and if the fuzziness score is 0.2, the system selects the "fine-grained" template.

[0041] After selecting a field decomposition granularity template, this template can be used to parse the multi-dimensional intent vector of multi-turn dialogue statements. A multi-dimensional intent vector is a high-dimensional vector used to represent a user's intent and needs across multiple dimensions. For example, if the "medium granularity" template is selected, the system can extract the following intent vectors from multi-turn dialogue statements: insurance type (health insurance), insured (individual), usage scenario (daily protection), coverage period (one year), budget range (medium budget), and additional liabilities (none). These intent vectors can more accurately reflect the user's needs, providing a basis for subsequent insurance product matching.

[0042] In this way, the field decomposition granularity template library can dynamically adjust the granularity of field decomposition according to the degree of ambiguity in the user's dialogue, thereby effectively parsing the user's true intent under different ambiguity conditions and enhancing the accuracy of insurance product recommendations.

[0043] Furthermore, step P20 in this embodiment of the application also includes:

[0044] P21: Extract the set of membership decomposition fields at the granularity of the matching field decomposition, parse the multi-turn dialogue statement based on the set of membership decomposition fields, and obtain the semantic fragments of each field; P22: Map the semantic fragments extracted from each field in the intent vector space to obtain the intent vector of each field, and output the multi-dimensional intent vector.

[0045] Specifically, the process of parsing the multi-dimensional intent vector of multi-turn dialogue statements from the granularity of the matching field decomposition can be further refined. After selecting the matching field decomposition granularity template, the set of subordinate decomposition fields of this template is first extracted. For example, if the matching field decomposition granularity template is "medium granularity", its subordinate decomposition field set may include fields such as insurance type, insured object, usage scenario, and coverage period. Subsequently, multi-turn dialogue statements are parsed based on this set of subordinate decomposition fields, and semantic fragments related to these fields are extracted from each round of dialogue. Semantic fragments refer to text fragments related to a specific field that can reflect the user's specific needs or preferences for that field. For example, the semantic fragment extracted from the dialogue for the "insurance type" field might be "health insurance", the semantic fragment for the "insured object" field might be "individual", and the semantic fragment for the "usage scenario" field might be "daily protection", etc.

[0046] Next, the semantic fragments extracted from each field are mapped into the intent vector space. The intent vector space is a high-dimensional space where each dimension corresponds to a possible intent or need. For example, for the "Insurance Type" field, the intent vector space might include multiple dimensions such as health insurance, life insurance, and property insurance. Semantic fragments can be mapped into the intent vector space using predefined mapping rules or natural language processing techniques (such as word embedding models) to obtain the intent vector for each field. For example, the semantic fragment "health insurance" might be mapped to a high value in the "health insurance" dimension of the intent vector space, and the semantic fragment "individual" might be mapped to a high value in the "individual" sub-dimension of the "insured object" dimension, and so on. Finally, the intent vectors of all fields are combined to output a complete multi-dimensional intent vector. This multi-dimensional intent vector comprehensively reflects the user's needs and preferences across multiple dimensions, providing accurate input for subsequent insurance product matching.

[0047] P30: Construct a user behavior profile based on the multidimensional intent vector, call the user behavior profile to perform matching and retrieval in the insurance product database, and output a set of matching insurance products.

[0048] Furthermore, in constructing a user behavior profile based on the multi-dimensional intent vector, step P30 of this embodiment further includes:

[0049] P31: Determine whether the multidimensional intent vector is a composite intent. If the multidimensional intent vector is a composite intent, obtain composite intent clues. P32: Perform semantic segmentation according to the composite intent clues and output multiple semantic candidate units. P33: Use field-aligned clustering to perform semantic clustering on the multiple semantic candidate units, identify multiple independent intent vectors, and construct a user behavior profile based on the multiple independent intent vectors.

[0050] It should be understood that, in order to build user behavior profiles more accurately, especially to handle complex user needs, user behavior profiles can be built based on multi-dimensional intent vectors, and these profiles can be used to perform matching and retrieval in insurance product databases to output a set of matching insurance products.

[0051] First, a user behavior profile is constructed based on the multidimensional intent vector. This profile is then used to perform matching and retrieval in the insurance product database, outputting a set of matching insurance products. During this process, the system first determines whether the multidimensional intent vector represents a composite intent. A composite intent refers to a user expressing multiple different needs or preferences in the conversation. These needs or preferences may involve different insurance types, coverage objects, usage scenarios, etc. For example, a user might mention both "health insurance" and "travel insurance," or "personal protection" and "family protection." If the system determines that the multidimensional intent vector represents a composite intent, it proceeds to the next step to obtain clues about the composite intent.

[0052] Composite intent cues refer to features or markers that can distinguish different intents. By analyzing the fields in a multidimensional intent vector, it is possible to identify which fields point to different intents. For example, if the "insurance type" field contains both "health insurance" and "travel insurance," then these two field values ​​can serve as composite intent cues. Semantic segmentation is then performed based on these cues, decomposing the multidimensional intent vector into multiple semantic candidate units. A semantic candidate unit is a fragment that may correspond to an independent intent; for example, one semantic candidate unit might contain "insurance type: health insurance, coverage: individual," and another semantic candidate unit might contain "insurance type: travel insurance, coverage: individual."

[0053] Next, field-aligned clustering is used to semantically cluster multiple semantic candidate units. Field-aligned clustering is a clustering method based on the similarity of field combinations. By comparing the combinations and values ​​of fields in semantic candidate units, similar semantic candidate units are grouped into one class. For example, if two semantic candidate units both contain "Insurance Type: Health Insurance, Coverage Object: Individual", they can be grouped into one class and identified as an independent intent vector. In this way, the system can identify multiple independent intent vectors from composite intents.

[0054] Finally, a user behavior profile is constructed based on the identified multiple independent intent vectors. This user behavior profile is a data structure that comprehensively reflects user needs and preferences, integrating multiple independent intents to form a comprehensive user needs model. For example, a user behavior profile might include: "Health insurance needs: personal protection, budget range: mid-range; travel insurance needs: personal protection, usage scenario: overseas travel." Based on this user behavior profile, a matching search is performed in the insurance product database, outputting a set of matching insurance products that meet the user's multiple independent intents.

[0055] The above steps effectively process complex intents, accurately construct user behavior profiles, and perform precise matching and retrieval in the insurance product database, outputting a set of insurance products that meet the user's needs. This processing mechanism not only improves the system's flexibility and adaptability but also enhances the relevance and satisfaction of insurance product recommendations.

[0056] Furthermore, in this embodiment of the application, step P33 further includes constructing a user behavior profile based on the multiple independent intent vectors:

[0057] P33-1: Obtain the set of multiple candidate matching insurance products corresponding to the multiple independent intent vectors; P33-2: Perform product correlation analysis on the set of multiple candidate matching insurance products, sort the set of multiple candidate matching insurance products according to product correlation, and output the updated set of multiple candidate matching insurance products.

[0058] Optionally, the specific process of constructing user behavior profiles based on multiple independent intent vectors and optimizing the candidate matching insurance product set can be further refined.

[0059] First, the system performs a matching search in the insurance product database based on each independent intent vector to obtain a set of candidate matching insurance products corresponding to each independent intent vector. For example, for the intent vector "Health insurance need: personal protection, budget range: mid-budget," all health insurance products that meet this need are retrieved; for the intent vector "Travel insurance need: personal protection, usage scenario: overseas travel," the system will retrieve all travel insurance products that meet this need. This process continues, generating a set of candidate matching insurance products for each independent intent vector, resulting in multiple sets of candidate matching insurance products.

[0060] Next, product correlation analysis is performed on these candidate matching insurance product sets. The purpose of product correlation analysis is to assess the relevance or complementarity between different insurance products. For example, health insurance and travel insurance may be highly correlated in certain situations because users may need both types of insurance to meet different protection needs. The system can use predefined association rules or data analysis techniques (such as collaborative filtering, association rule mining, etc.) to assess the correlation between products. Then, based on the results of the product correlation analysis, the multiple candidate matching insurance product sets are ranked according to product correlation. The purpose of ranking is to place insurance product sets with higher correlation at the top, so that users can more intuitively see insurance product combinations that meet their comprehensive needs. For example, if health insurance and travel insurance are highly correlated, then sets containing both types of insurance may be ranked higher. Finally, the updated multiple candidate matching insurance product sets are output through ranking. These sets not only meet the user's individual intent but also consider the correlation between products, thereby improving the relevance and usability of the recommendations.

[0061] Furthermore, the user behavior profile is invoked to perform matching and retrieval in the insurance product database, and a set of matching insurance products is output. Step P30 in this embodiment of the application also includes:

[0062] P34: Obtain the structured product vector for each product in the insurance product database; P35: Calculate the similarity score between the user behavior profile and the structured product vector using the vector similarity function, and output the similarity score set; P36: Based on the similarity score set, filter the top N insurance products with a similarity score greater than a preset value as the matching insurance product set for output.

[0063] Specifically, the process can be further refined to call upon user behavior profiles to perform matching and retrieval in the insurance product database in order to obtain a set of matching insurance products.

[0064] First, the system obtains the structured product vector for each product in the insurance product database. A structured product vector is a high-dimensional vector used to represent the characteristics and attributes of an insurance product. These characteristics and attributes may include insurance type, insured party, usage scenario, coverage period, budget range, and additional liabilities. For example, the structured product vector for a health insurance product might contain: "Insurance type: Health insurance, Insured party: Individual, Usage scenario: Daily protection, Coverage period: One year, Budget range: Medium budget, Additional liabilities: None." By transforming the various characteristics and attributes of insurance products into structured product vectors, the system can represent and process insurance product information in a standardized way.

[0065] Next, a similarity score is calculated between the user behavior profile and the structured product vector using a vector similarity function. A vector similarity function is a mathematical function used to measure the degree of similarity between two vectors; common similarity functions include cosine similarity and Euclidean distance. In this application, a suitable vector similarity function can be selected to calculate the similarity score between the user behavior profile and each structured product vector. For example, if the vector representation of the user behavior profile is U, and the structured product vector of an insurance product is P, then their similarity score S can be calculated using the cosine similarity function to obtain the similarity score between the user behavior profile and each product in the insurance product database, and these scores are then combined into a similarity score set.

[0066] Finally, based on the similarity score set, the top N insurance products with similarity scores greater than a preset threshold are selected as the output set of matched insurance products. The preset similarity score is a threshold set according to the specific application scenario, used to filter out insurance products with low similarity. The system selects the top N insurance products with scores higher than this threshold from the similarity score set and outputs them as the set of matched insurance products. For example, if the preset similarity score is 0.7, the system selects the top 5 insurance products with scores greater than 0.7 from the similarity score set as the set of matched insurance products. This ensures that the output set of insurance products not only highly matches the user's behavioral profile but also has a suitable number, facilitating further selection and comparison by the user. This mechanism not only improves the accuracy of insurance product recommendations but also enhances the user experience, enabling users to quickly find insurance products that meet their needs.

[0067] In summary, the embodiments of this application have at least the following technical effects:

[0068] This application quantifies the ambiguity of user dialogue to accurately measure the uncertainty of user expression; it dynamically adjusts the granularity of field decomposition based on ambiguity scores to precisely adapt to the parsing of needs under different levels of ambiguity, improving the flexibility and accuracy of parsing; it uses field decomposition granularity templates to parse multi-turn dialogue statements, generating multi-dimensional intent vectors to comprehensively capture user needs across multiple dimensions such as insurance type, insured objects, and usage scenarios; it uses vector similarity functions to accurately match insurance products, filtering the most suitable options from a massive number of products, significantly improving recommendation accuracy; and it combines dynamic interaction strategies and reward mechanisms to continuously optimize recommendation results based on user feedback, enhancing user satisfaction and user experience.

[0069] The technology achieves the goal of using fuzzy scoring to conduct multi-granularity field intent-guided analysis, constructing user behavior profiles to match insurance products, and improving the accuracy of demand analysis and product matching.

[0070] Example 2, based on the same inventive concept as the multi-dimensional matching method for insurance products based on behavioral profiling in the foregoing examples, such as... Figure 2 As shown, this application provides a multi-dimensional matching system for insurance products based on behavioral profiling. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0071] The fuzzy scoring module 11 is used to extract user multi-turn dialogue statements, extract semantic fuzziness feature vectors from the multi-turn dialogue statements, perform fuzzy scoring based on the semantic fuzziness feature vectors, and obtain fuzzy score values. The semantic fuzziness feature vectors include at least the proportion of high-frequency fuzzy words, semantic drift degree, number of missing fields, and context jump frequency.

[0072] The multidimensional intent parsing module 12 is used to construct a field decomposition granularity template library, input the fuzziness score value into the field decomposition granularity template library to obtain the matching field decomposition granularity, and use the matching field decomposition granularity to parse the multidimensional intent vector of the multi-turn dialogue statement.

[0073] Product matching and retrieval module 13 is used to construct a user behavior profile according to the multi-dimensional intent vector, call the user behavior profile to perform matching and retrieval in the insurance product database, and output a set of matching insurance products.

[0074] Furthermore, the fuzzy scoring module 11 is also used to perform the following steps:

[0075] The semantic ambiguity feature vector is standardized to output a standardized semantic ambiguity feature vector. A ambiguity scoring function is constructed, which is obtained by training the training corpus data using a BERT word segmenter and a regression head network. The training corpus data includes multi-turn dialogue sentence samples, semantic ambiguity feature vectors extracted from the multi-turn dialogue sentence samples, and ambiguity score annotation values. The standardized semantic ambiguity feature vector is input into the ambiguity scoring function to perform ambiguity scoring and obtain the ambiguity score value.

[0076] Furthermore, the fuzzy scoring module 11 is also used to perform the following steps:

[0077] The BERT segmenter is used to semantically encode the input text of the multi-turn dialogue sentence samples and output the encoded semantic ambiguity feature vector. The regression head network is used to map the semantic ambiguity feature vector to ambiguity score values ​​and output ambiguity score mapping values. A loss function is introduced to perform multiple rounds of iterative training on the ambiguity score annotation data and the ambiguity score annotation values ​​until the error between the ambiguity score mapping data and the ambiguity score annotation data in the current iteration round is less than a preset threshold, thus obtaining a converged ambiguity score function.

[0078] Furthermore, the multidimensional intent parsing module 12 is also used to perform the following steps:

[0079] The field decomposition granularity template library stores multiple field decomposition granularity templates. These multiple field decomposition granularity templates correspond to multiple pre-defined granularity levels, and the pre-defined granularity levels include multiple corresponding fuzzy scoring value ranges. Each field decomposition granularity template selects a set of subordinate decomposition fields under the corresponding granularity from the decomposition field set. The decomposition field set includes fields such as insurance type, insured object, usage scenario, coverage period, budget range, and additional liability.

[0080] Furthermore, the multidimensional intent parsing module 12 is also used to perform the following steps:

[0081] Extract the set of membership decomposition fields at the granularity of the matching field decomposition, parse the multi-turn dialogue statement based on the set of membership decomposition fields, and obtain semantic fragments of each field; map the semantic fragments extracted from each field in the intent vector space to obtain the intent vector of each field, and output a multi-dimensional intent vector.

[0082] Furthermore, the product matching and retrieval module 13 is also used to perform the following steps:

[0083] Obtain the structured product vector for each product in the insurance product database; calculate the similarity score between the user behavior profile and the structured product vector using a vector similarity function, and output a set of similarity scores; based on the set of similarity scores, filter out the top N insurance products with similarity scores greater than a preset value as a set of matching insurance products.

[0084] Furthermore, the product matching and retrieval module 13 is also used to perform the following steps:

[0085] Determine whether the multidimensional intent vector is a composite intent. If the multidimensional intent vector is a composite intent, obtain composite intent clues. Perform semantic segmentation according to the composite intent clues and output multiple semantic candidate units. Use field-aligned clustering to perform semantic clustering on the multiple semantic candidate units to identify multiple independent intent vectors. Construct a user behavior profile based on the multiple independent intent vectors.

[0086] Furthermore, the product matching and retrieval module 13 is also used to perform the following steps:

[0087] Obtain multiple candidate matching insurance product sets corresponding to the multiple independent intent vectors; perform product correlation analysis on the multiple candidate matching insurance product sets, sort the multiple candidate matching insurance product sets according to product correlation, and output the updated multiple candidate matching insurance product sets.

[0088] Example 3, Exemplary Electronic Device

[0089] The following is for reference. Figure 3 The electronic device described in the embodiments of this application is used to illustrate this application.

[0090] Based on the same inventive concept as the multi-dimensional matching method for insurance products based on behavioral profiling in the foregoing embodiments, this application also provides a multi-dimensional matching system for insurance products based on behavioral profiling, including: a processor coupled to a memory, the memory being used to store a program, and when the program is executed by the processor, the system performs the steps of the method described in Embodiment 1.

[0091] The electronic device 300 includes a processor 302, a communication interface 303, and a memory 301. Optionally, the electronic device 300 may also include a bus architecture 304. The communication interface 303, processor 302, and memory 301 can be interconnected via the bus architecture 304; the bus architecture 304 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus architecture 304 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0092] Processor 302 may be a CPU, microprocessor, ASIC, or one or more integrated circuits used to control the execution of programs according to the present application.

[0093] Communication interface 303 uses any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area network (WLAN), wired access network, etc.

[0094] Memory 301 can be ROM or other types of static storage devices capable of storing static information and instructions, RAM or other types of dynamic storage devices capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory can exist independently and be connected to the processor via bus architecture 304. Memory can also be integrated with the processor.

[0095] The memory 301 stores computer execution instructions for implementing the solution of this application, and the processor 302 controls the execution. The processor 302 executes the computer execution instructions stored in the memory 301, thereby realizing the multi-dimensional matching method for insurance products based on behavioral profiles provided in the above embodiments of this application.

[0096] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0097] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0098] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A multi-dimensional matching method for insurance products based on behavioral profiling, characterized in that, The method includes: Extract user multi-turn dialogue statements, extract semantic fuzziness feature vectors from the multi-turn dialogue statements, perform fuzzy scoring based on the semantic fuzziness feature vectors, and obtain fuzzy score values. The semantic fuzziness feature vectors include at least the proportion of high-frequency fuzzy words, semantic drift degree, number of missing fields, and context jump frequency. Construct a field decomposition granularity template library, input the fuzziness score into the field decomposition granularity template library to obtain the matching field decomposition granularity, and use the matching field decomposition granularity to parse the multi-dimensional intent vector of the multi-turn dialogue statement; A user behavior profile is constructed based on the multidimensional intent vector. The user behavior profile is then used to perform matching and retrieval in the insurance product database, and a set of matching insurance products is output. The method for constructing a user behavior profile based on the multidimensional intent vector further includes: Determine whether the multidimensional intent vector is a composite intent; if the multidimensional intent vector is not a composite intent, obtain composite intent clues. Semantic segmentation is performed based on the composite intent clues, and multiple semantic candidate units are output; The multiple semantic candidate units are semantically clustered using a field-aligned clustering method to identify multiple independent intent vectors, and a user behavior profile is constructed based on the multiple independent intent vectors.

2. The method as described in claim 1, characterized in that, Fuzzy scoring is performed based on semantic fuzziness feature vectors to obtain fuzzy score values. Methods include: The semantically ambiguous feature vector is subjected to feature vector standardization processing to output a standardized semantically ambiguous feature vector; A fuzziness scoring function is constructed. The fuzziness scoring function is obtained by training the training corpus data through a BERT word segmenter and a regression head network. The training corpus data includes multi-turn dialogue sentence samples, semantic fuzziness feature vectors extracted from the multi-turn dialogue sentence samples, and fuzziness scoring label values. The standardized semantic fuzzy feature vector is input into the fuzzy scoring function to perform fuzzy scoring, and a fuzzy score value is obtained.

3. The method as described in claim 2, characterized in that, The training steps for the fuzzy scoring function include: The BERT word segmenter is used to perform semantic encoding on the input text of the multi-turn dialogue sentence samples and output the encoded semantic ambiguity feature vector. The regression head network is used to map the semantic ambiguity feature vector to ambiguity score values ​​and output ambiguity score mapping values. A loss function is introduced to perform multiple rounds of iterative training on the fuzzy rating labeled data and the fuzzy rating labeled values ​​until the error between the fuzzy rating mapping data and the fuzzy rating labeled data in the current iteration is less than a preset threshold, thus obtaining a converged fuzzy rating function.

4. The method as described in claim 1, characterized in that, Construct a field decomposition granularity template library, which stores multiple field decomposition granularity templates. These multiple field decomposition granularity templates correspond to multiple pre-defined granularity levels, and the pre-defined granularity levels include multiple corresponding fuzzy score value ranges. Each field decomposition granularity template selects a set of subordinate decomposition fields at the corresponding granularity from the set of decomposition fields. The set of decomposition fields includes fields such as insurance type, insured object, usage scenario, coverage period, budget range, and additional liability.

5. The method as described in claim 4, characterized in that, The method for parsing the multi-dimensional intent vector of the multi-turn dialogue statements at the granularity of the matching fields includes: Extract the set of subordinate decomposition fields at the granularity of the matching field decomposition, parse the multi-turn dialogue statements based on the set of subordinate decomposition fields, and obtain semantic fragments of each field; The semantic fragments extracted from each field are mapped in the intent vector space to obtain the intent vector of each field, and a multidimensional intent vector is output.

6. The method as described in claim 1, characterized in that, The method involves calling the user behavior profile to perform matching and retrieval in the insurance product database, and outputting a set of matched insurance products. Obtain the structured product vector for each product in the insurance product database; The similarity score between the user behavior profile and the structured product vector is calculated using a vector similarity function, and a similarity score set is output. Based on the set of similarity scores, the top N insurance products with similarity scores greater than the preset value are selected as the output set of matching insurance products.

7. The method as described in claim 1, characterized in that, The method for constructing user behavior profiles based on the multiple independent intent vectors further includes: Obtain a set of multiple candidate matching insurance products corresponding to the multiple independent intent vectors; Perform product correlation analysis on the multiple candidate matching insurance product sets, sort the multiple candidate matching insurance product sets according to product correlation, and output the updated multiple candidate matching insurance product sets.

8. A multi-dimensional matching system for insurance products based on behavioral profiling, characterized in that, The system is used to implement the multi-dimensional matching method for insurance products based on behavioral profiling as described in any one of claims 1-7, the system comprising: The fuzzy scoring module is used to extract user multi-turn dialogue statements, extract semantic fuzziness feature vectors from the multi-turn dialogue statements, perform fuzzy scoring based on the semantic fuzziness feature vectors, and obtain fuzzy score values. The semantic fuzziness feature vectors include at least the proportion of high-frequency fuzzy words, semantic drift degree, number of missing fields, and context jump frequency. A multidimensional intent parsing module is used to construct a field decomposition granularity template library, input the fuzziness score value into the field decomposition granularity template library to obtain the matching field decomposition granularity, and use the matching field decomposition granularity to parse the multidimensional intent vector of the multi-turn dialogue statement; The product matching and retrieval module is used to construct a user behavior profile according to the multi-dimensional intent vector, call the user behavior profile to perform matching and retrieval in the insurance product database, and output a set of matching insurance products.

9. An electronic device, characterized in that, include: A processor coupled to a memory for storing a program that, when executed by the processor, causes the system to perform the steps of the method as claimed in any one of claims 1 to 7.

Images