CN-121981837-A - Insurance personalized intelligent recommendation and search method, device, equipment and medium

CN121981837ACN 121981837 ACN121981837 ACN 121981837ACN-121981837-A

Abstract

The invention relates to the field of artificial intelligence and is applied to financial and medical scenes, and discloses an insurance personalized intelligent recommending and searching method, device, equipment and medium; the method comprises the steps of extracting potential motivations of a user based on a large language model, constructing motivation perception characterization, enhancing motivation semantics of different dimensions through an attention expert mixed model, carrying out semantic alignment and interaction fusion on the enhanced motivation characterization and insurance product characteristics and user history preference respectively to generate user comprehensive characterization, and finally calculating matching scores of the comprehensive characterization and candidate products and outputting a recommendation list in a side-by-side mode. The invention can deeply understand the real insurance motivation of the user, realizes multi-level accurate matching of motivation, products and preference, and remarkably improves the accuracy, individuation and user satisfaction of intelligent recommendation in financial and medical insurance scenes.

Inventors

WANG JIANZONG
ZHANG NAN
QU XIAOYANG

Assignees

平安科技（深圳）有限公司

Dates

Publication Date: 20260505
Application Date: 20260202

Claims (10)

1. An insurance personalized intelligent recommendation and search method is characterized by comprising the following steps: Acquiring multiple rounds of consultation dialogue data generated by a target user on an insurance platform, wherein the multiple rounds of consultation dialogue data comprise a user question text and a system reply text; extracting at least one potential motivation of the user based on the multi-round consultation dialogue data, and constructing motivation perception characterization; Inputting the motivation perception representation into an attention expert mixed model, respectively extracting features of different dimensionality semantics of the motivation perception representation through a plurality of expert sub-networks, and dynamically fusing the output of each expert sub-network through a gating network to obtain enhanced user motivation semantic representation; Carrying out semantic alignment on the enhanced user motivation semantic representation and the product feature representation to obtain a first alignment result; performing interactive fusion on the enhanced user motivation semantic representation and the user history preference representation to obtain a second alignment result; fusing the first alignment result and the second alignment result to generate a comprehensive characterization of the user; calculating the recommendation score of each candidate insurance product based on the comprehensive characterization of the user and the product characteristic characterization of the candidate insurance product; and sequencing the candidate insurance products according to the recommendation scores, generating a personalized recommendation list and outputting the personalized recommendation list.
2. The personalized intelligent recommendation and search method for insurance according to claim 1, wherein said step of extracting at least one potential motivation of said user based on said multi-round advisory dialog data and constructing motivational perception characterizations comprises: semantic analysis is carried out on the multi-round consultation dialogue data by utilizing a large language model, and at least one motivation theme representing potential demands of users is identified; converting each identified motivational topic into a corresponding motivational vector using a semantic encoder, respectively; calculating semantic relevance weights between each motivation vector and each round of consultation dialogue through a contextual attention mechanism; carrying out weighted aggregation on semantic vectors of each round of consultation dialogue according to the semantic relevance weight to generate a context enhancement representation corresponding to each motivation vector; And aggregating all the context enhancement characterizations to generate a motivational perception characterization of the user.
3. The personalized intelligent insurance recommendation and search method according to claim 1, wherein the step of inputting the motivational perception representation into an attention expert hybrid model, respectively extracting features of different dimensionality semantics of the motivational perception representation through a plurality of expert sub-networks, dynamically fusing the output of each expert sub-network through a gating network, and obtaining the enhanced motivational semantic representation of the user comprises the following steps: Pre-constructing an attention expert mixed model comprising N expert sub-networks, wherein each expert sub-network is pre-trained to pay attention to semantics of different dimensions, and the dimensions comprise at least one of economic motivation dimension, emotion motivation dimension and guarantee motivation dimension; Inputting the motivation perception representation to the N expert sub-networks simultaneously, and outputting feature vectors of corresponding dimensions by each expert sub-network; Inputting the motivation perception representation to a gating network, and dynamically calculating the weight for fusing the output of each expert sub-network by the gating network; and carrying out weighted fusion on the feature vectors output by the N expert sub-networks according to the weights to obtain the enhanced semantic representation of the user motivation.
4. The personalized intelligent insurance recommendation and search method according to claim 1, wherein the step of semantically aligning the enhanced user motivation semantic representation with product feature representation to obtain a first alignment result comprises: acquiring a product characteristic representation of an insurance product, wherein the product characteristic representation is obtained by carrying out coding fusion on text description and structural characteristics of the insurance product; Taking an insurance product with the user actually generating the positive interaction as a positive sample, and taking an insurance product without the positive interaction as a negative sample, so as to construct a training sample pair; Through comparison with a learning target, training a model to maximize the similarity between the enhanced user motivation semantic representation and the positive sample product feature representation in a semantic space, and minimizing the similarity between the enhanced user motivation semantic representation and the negative sample product feature representation; and carrying out semantic similarity calculation on the enhanced user motivation semantic representation after training convergence and the product feature representation of the candidate insurance product to be recommended, wherein the obtained similarity is used as the first alignment result.
5. The personalized intelligent recommendation and search method for insurance according to claim 1, wherein the step of interactively fusing the enhanced semantic representation of user motivation with the historical preference representation of user to obtain a second alignment result comprises: Constructing a user history preference characterization reflecting long-term preferences of the user based on the history behavior data of the user; calculating a first attention response of the enhanced user motivation semantic representation to the user history preference representation, and generating a first fusion vector; Calculating a second attention response of the user history preference characterization to the enhanced user motivation semantic characterization, and generating a second fusion vector; and fusing the first fusion vector and the second fusion vector to generate the second alignment result.
6. The personalized intelligent recommendation and search method for insurance according to claim 1, wherein said step of fusing said first alignment result with said second alignment result to generate a user comprehensive characterization comprises: The semantic matching information of the motivation and the product, which are represented by the first alignment result, and the coupling information of the motivation and the preference, which are represented by the second alignment result, are spliced or weighted and fused to obtain comprehensive information; and encoding the comprehensive information to generate the comprehensive characterization of the user.
7. The personalized intelligent insurance recommendation and search method according to claim 1, wherein the step of calculating a recommendation score for each candidate insurance product based on the comprehensive characterization of the user and the characterization of product features of the candidate insurance product comprises: Calculating a first similarity between the user comprehensive characterization and product feature characterization of each candidate insurance product; Calculating a second similarity between the enhanced user motivational semantic representation and the product feature representation of each candidate insurance product; and carrying out weighted summation on the first similarity and the second similarity to obtain a final recommendation score of each candidate insurance product.
8. An insurance personalized intelligent recommendation and search device, characterized by comprising: The data acquisition unit is used for acquiring multiple rounds of consultation dialogue data generated by the target user on the safety platform, wherein the multiple rounds of consultation dialogue data comprise user question texts and system reply texts; a representation construction unit for extracting at least one potential motivation of the user based on the multi-round consultation dialogue data and constructing motivation perception representations; The characterization enhancing unit is used for inputting the motivation perception characterization into an attention expert mixed model, extracting features of different dimension semantics of the motivation perception characterization through a plurality of expert sub-networks respectively, and dynamically fusing the output of each expert sub-network through a gating network to obtain enhanced user motivation semantic characterization; the characterization alignment unit is used for carrying out semantic alignment on the enhanced semantic characterization of the user motivation and the product characteristic characterization to obtain a first alignment result; the characterization fusion unit is used for fusing the first alignment result and the second alignment result to generate a user comprehensive characterization; The score calculating unit is used for calculating the recommendation score of each candidate insurance product based on the comprehensive characterization of the user and the product characteristic characterization of the candidate insurance product; and the product recommending unit is used for sequencing the candidate insurance products according to the recommending scores, generating a personalized recommending list and outputting the personalized recommending list.
9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the insurance personalized intelligent recommendation and search method according to any one of claims 1 to 7 when the computer program is executed by the processor.
10. A computer readable storage medium, characterized in that it stores a computer program which, when executed by a processor, causes the processor to perform the insurance personalized intelligent recommendation and search method according to any of claims 1 to 7.

Description

Insurance personalized intelligent recommendation and search method, device, equipment and medium Technical Field The invention relates to the field of artificial intelligence, and is applied to financial and medical scenes, in particular to an insurance personalized intelligent recommending and searching method, device, equipment and medium. Background In the fields of finance and medical insurance, an intelligent recommendation and search system has become an important technical means for improving user experience and enhancing service conversion rate. The existing system mainly depends on traditional methods such as keyword matching, user portrait retrieval, semantic similarity calculation, collaborative filtering and the like. For example, when a user inputs query words such as serious illness risks or care, the system matches corresponding products through semantic analysis and collaborative filtering algorithm to recommend. However, these methods have the following significant technical drawbacks in the highly specialized and decision-making complex scenario of finance and medical insurance: First, existing systems generally ignore deep implicit motivation information that users have in multiple rounds of consultation interactions prior to formal searching. In actual insurance business, users often carry out multi-round inquiry through customer service dialogue or intelligent consultation robots, the content relates to multi-dimensional implicit requirements such as guarantee intention, family structure, economic condition, risk bearing capacity and the like, and the traditional model only carries out recommendation based on final inquiry words, so that a large amount of context information is lost, and the real purchasing motivation of the users cannot be accurately reflected. Secondly, insurance user inquiry usually has the characteristics of short text, ambiguity and ambiguity, such as how to buy insurance, how to buy serious insurance cost-effective, and the like, and the back of the insurance user inquiry can correspond to various different motivations such as guarantee, financial management, family responsibility, and the like, so that the accurate inquiry and intention mapping of the existing semantic model are difficult to realize, and recommendation deviation is easy to generate. Furthermore, insurance products have the characteristics of complex structure and multidimensional characteristics, including terms and descriptions of text descriptions, and classification labels (such as life insurance, accident insurance and the like) and structural attributes (such as insurance amount, term, rate and the like). The existing method has a semantic gap when the natural language consultation, the product category and the structural feature are subjected to unified modeling, so that the matching degree of the recommendation result and the user requirement is not high. In addition, the historical consultation sequence of the user often contains topic skipping, emotional expression or non-decision related content, so as to form sequence noise, interfere with capturing of a real motivation of the user by a model, and reduce accuracy and stability of a recommendation system. Therefore, there is a need for an intelligent recommendation and search method that can deeply integrate user consultation context, accurately sense its potential motivation, and realize multi-level alignment of motivation, product and preference, so as to improve the accuracy and intelligent level of personalized services in the financial and medical insurance fields. Disclosure of Invention The invention aims to provide an insurance personalized intelligent recommendation and search method, device, equipment and storage medium, and aims to solve the problem that the existing financial and medical insurance intelligent recommendation method is insufficient in personalized recommendation precision because hidden motivations in a user consultation context are ignored, and the user consultation intention and product semantics are difficult to match accurately. In a first aspect, an embodiment of the present invention provides a personalized intelligent recommendation and search method for insurance, including: Acquiring multiple rounds of consultation dialogue data generated by a target user on an insurance platform, wherein the multiple rounds of consultation dialogue data comprise a user question text and a system reply text; extracting at least one potential motivation of the user based on the multi-round consultation dialogue data, and constructing motivation perception characterization; Inputting the motivation perception representation into an attention expert mixed model, respectively extracting features of different dimensionality semantics of the motivation perception representation through a plurality of expert sub-networks, and dynamically fusing the output of each expert sub-network through a gating network to obtain enhance