A Large-Model Intelligent Customer Service Method and System Based on Retrieval Enhancement

Abstract

This application discloses a large-scale intelligent customer service method and system based on retrieval enhancement. The method includes: receiving a query message; performing intent recognition on the query message, and if the intent of the query message is determined to be to obtain recommended financial products, sending a follow-up message to the target user; receiving the preference weights input by the target user, and determining the target user profile corresponding to the target user in a user profile database based on the target user's identification information; analyzing the target user profile, query message, and preference weights using a vertical domain intent recognition model; calculating the total investment amount and preference weights in the query message using a multi-objective optimization algorithm to obtain a second recommended financial product combination; and simultaneously sending the first and second recommended financial product combinations to the target user. This application solves the technical problem in intelligent customer service in the financial field of being unable to accurately recommend suitable financial products to users based on their needs.

Description

Technical Field

[0001] This application relates to the field of data analysis technology, and more specifically, to a large-scale intelligent customer service method and system based on retrieval enhancement. Background Technology

[0002] In practical applications of intelligent customer service in the financial sector, several unsatisfactory situations often arise. A prominent issue is the inability to accurately recommend suitable financial products based on the user's true needs. Specifically, when users consult intelligent customer service with their specific financial situation, investment goals, risk tolerance, and other factors, the system may fail to accurately grasp the user's core needs due to limitations in its algorithms, incomplete data collection, or misunderstandings of the user's expression. For example, a user might be looking for a low-risk investment product suitable for short-term cash flow, but the intelligent customer service might recommend a longer-term, higher-risk investment product. Or, a user might expect recommended products to match their existing financial asset allocation for risk diversification and portfolio optimization, but the intelligent customer service fails to consider this, arbitrarily recommending seemingly popular products that don't align with the user's asset portfolio. This leads to a significant discrepancy between the recommendations and the user's actual needs, impacting user satisfaction with the intelligent customer service and the accuracy of product selection.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This application provides a large-scale intelligent customer service method and system based on retrieval enhancement, which aims to at least solve the technical problem in intelligent customer service in the financial field of being unable to accurately recommend suitable financial products to users based on their needs.

[0005] According to one aspect of this application, a large-scale intelligent customer service method based on retrieval enhancement is provided, comprising: receiving a query message input by a target object; performing intent recognition on the query message, and if it is determined that the intent of the query message is to obtain recommended financial products, sending a follow-up message to the target object, wherein the follow-up message is used to prompt the target object to input preference weights for returns, risks, and terms; receiving the preference weights input by the target object, and determining the target user profile corresponding to the target object in a user profile database based on the target object's identification information; analyzing the target user profile, query message, and preference weights using a vertical domain intent recognition model to obtain a first recommended financial product combination output by the vertical domain intent recognition model; calculating the total investment amount and preference weights in the query message using a multi-objective optimization algorithm to obtain a second recommended financial product combination; simultaneously sending the first and second recommended financial product combinations to the target object, receiving a selection response message from the target object regarding the first and second recommended financial product combinations, and updating the vertical domain intent recognition model and the target user profile based on the selection response message.

[0006] Optionally, a vertical domain intent recognition model is used to analyze the target user profile, query message, and preference weights to obtain the first recommended financial product combination output by the vertical domain intent recognition model. This includes: extracting the target values ​​corresponding to preset keywords from the query message, encoding the preset keywords and target values ​​together as numerical features to obtain the first numerical feature; encoding the target user profile as numerical features to obtain the second numerical feature, and encoding the preference weights as numerical features to obtain the third numerical feature; weighting and concatenating the first, second, and third numerical features to obtain the fused feature; determining the matching score between the fused feature and the candidate products through a neural network ranking model, and determining the top n financial products with the highest matching scores as the first recommended financial product combination. The neural network ranking model is a dual-tower structure, including a first feature encoding tower and a second feature encoding tower. The first feature encoding tower is used to receive the fused feature and output the user representation vector, and the second feature encoding tower is used to receive the attribute features of the financial products and output the financial product representation vector; n is a positive integer greater than 1.

[0007] Optionally, a multi-objective optimization algorithm is used to calculate the total investment amount and preference weights in the query message to obtain a second recommended financial product portfolio, including: Step S1, determining the target number of financial products to be recommended; Step S2, generating an initial population, wherein each individual in the initial population represents a different combination of financial products to be recommended, wherein each combination includes n financial products to be recommended; Step S3, calculating the objective function for each individual in the current population, wherein the objective function includes: maximizing the first function value representing the return, minimizing the second function value representing the risk, and minimizing the third function value representing the term; Step S4, performing a non-dominated ranking of the individuals, and dividing the current population into different non-dominated groups based on the ranking results. Step S3: For each non-dominated level, calculate the crowding distance and select individuals from the current population to enter the breeding pool based on the crowding distance calculation results. Perform crossover and mutation operations on the individuals in the breeding pool to generate offspring populations. Step S5: Merge the initial population and offspring populations into the target population. Perform non-dominated sorting and crowding distance calculation on the target population, and retain individuals whose non-dominated level and crowding distance meet the preset requirements to form a new population. Repeat steps S3 to S4 until the maximum number of iterations is reached to obtain the final population. Step S6: In the final population, determine the target individuals with the highest level in the non-dominated sorting, and determine the investment ratio of each financial product to be recommended corresponding to the target individuals as the second recommended financial product portfolio.

[0008] Optionally, determining the target number of financial products to be recommended includes: selecting the top m financial products with the highest matching scores in the neural network ranking model, and determining the top m financial products with the highest matching scores as financial products to be recommended, where m is the target number and m is a positive integer greater than n.

[0009] Optionally, the method further includes: determining the first function value using the following formula. : in, Weight for expected returns Let i be the i-th combination of financial products to be recommended, and k be the number of combinations of financial products to be recommended, k = , Let i be the expected rate of return for product i, where i is a positive integer not greater than k; the value of the second function is determined by the following formula. : in, As the risk tolerance weight, Let i be the risk coefficient of product i; the value of the third function is determined by the following formula. : in, Let i be the risk coefficient of product i. Let D be the timeframe for product i, and D be the user's expected timeframe. The objective function also includes constraint penalty terms, which include: *C is less than the preset minimum investment amount for product i, where C is the total investment amount.

[0010] Optionally, intent recognition is performed on the query message, including: using an intent recognition model to recognize the intent of the query message, obtaining the recognition result output by the intent recognition model, wherein the recognition result includes: the intent of the query message is to obtain recommended financial products, and the intent of the query message is not to obtain recommended financial products; the intent recognition model is trained by the following method: acquiring a user query text dataset, labeling the user query text data in the user query text dataset, generating binary labels containing intent to recommend financial products and intent not to recommend financial products, obtaining labeled user query text data; generating feature vectors corresponding to the labeled user query text data, wherein the feature vectors... The parameters include at least one of the following: a term frequency-inverse document frequency vector generated by a term frequency-inverse document frequency algorithm, and a contextual semantic embedding vector generated based on a pre-trained language model; the feature vector is input into a neural network classifier including fully connected layers, and the activation function is used to output the prediction result, which is the probability value of the prediction result being the intention to recommend financial products; the relationship between the probability value and a preset threshold is judged, and when the probability value is greater than the preset threshold, it is determined to be the intention to recommend financial products, and when the probability value is less than the preset threshold, it is determined to be the intention not to recommend financial products; based on the error between the prediction result output by the neural network classifier and the labeled label, the model parameters of the neural network classifier are updated through a backpropagation algorithm to obtain the intention recognition model.

[0011] Optionally, the user profile database is constructed using the following method: extracting user behavior feature sets from transaction records, log traces, and historical session records, wherein the user behavior feature sets include at least: numerical features, time series features, and categorical features; standardizing and normalizing the numerical features using principal component analysis, extracting the latent vector representation of the time series features using a long short-term memory autoencoder, and mapping the categorical features to a continuous vector space using one-hot encoding to obtain the target feature set; clustering the features in the target feature set based on density peaks to obtain clustering results, and constructing the user profile database based on the clustering results, wherein the clustering results carry user group labels.

[0012] According to another aspect of this application, a large-scale intelligent customer service system based on retrieval enhancement is also provided, comprising: a receiving module for receiving query messages input by a target object; an identification module for performing intent recognition on the query messages, and sending follow-up messages to the target object when the intent of the query messages is determined to be to obtain recommended financial products, wherein the follow-up messages are used to prompt the target object to input preference weights for returns, risks, and terms; a determination module for receiving preference weights input by the target object, and determining the target user profile corresponding to the target object in a user profile database based on the target object's identification information; an analysis module for analyzing the target user profile, query messages, and preference weights using a vertical domain intent recognition model to obtain a first recommended financial product combination output by the vertical domain intent recognition model; a calculation module for calculating the total investment amount and preference weights in the query messages using a multi-objective optimization algorithm to obtain a second recommended financial product combination; and an optimization module for simultaneously sending the first and second recommended financial product combinations to the target object, receiving the target object's selection response messages for the first and second recommended financial product combinations, and updating the vertical domain intent recognition model and the target user profile based on the selection response messages.

[0013] According to another aspect of this application, a non-volatile storage medium is also provided, the storage medium including a stored program, wherein the program, when running, controls the device where the storage medium is located to execute the above-mentioned large-model intelligent customer service method based on retrieval enhancement.

[0014] According to another aspect of this application, an electronic device is also provided, comprising: a memory and a processor, the processor being configured to run a program stored in the memory, wherein the program, when running, executes the above-described retrieval-enhanced large-model intelligent customer service method.

[0015] According to another aspect of this application, a computer program is also provided, wherein when the computer program is executed by a processor, it implements the above-described search-enhanced big model intelligent customer service method.

[0016] According to another aspect of this application, a computer program product is also provided, the computer program product including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-described search-enhanced big model intelligent customer service method.

[0017] In this application, the following steps are employed: receiving a query message input from a target object; performing intent recognition on the query message; and, if the intent of the query message is determined to be to obtain recommended financial products, sending a follow-up message to the target object, wherein the follow-up message is used to prompt the target object to input preference weights for return, risk, and term; receiving the preference weights input by the target object; and determining the target user profile corresponding to the target object in a user profile database based on the target object's identification information; analyzing the target user profile, query message, and preference weights using a vertical domain intent recognition model to obtain the first recommended financial product combination output by the vertical domain intent recognition model; and using a multi-objective optimization algorithm to process the query message. The system calculates the total investment amount and preference weights in the information to obtain a second recommended financial product portfolio. The first and second recommended financial product portfolios are simultaneously sent to the target user. The system receives the target user's response message regarding their selection of the first and second recommended financial product portfolios. Based on the response message, the system updates the vertical domain intent recognition model and the target user profile. This achieves the goal of accurately recommending suitable financial products to users based on their needs, thereby enhancing user stickiness and solving the technical problem in intelligent customer service in the financial field of not being able to accurately recommend suitable financial products based on user needs. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This is a flowchart of a large-model intelligent customer service method based on retrieval enhancement according to an embodiment of this application;

[0020] Figure 2 This is a structural diagram of a large-model intelligent customer service system based on retrieval enhancement, according to an embodiment of this application.

[0021] Figure 3 This is a hardware structure block diagram of a computer terminal for a large-model intelligent customer service method based on retrieval enhancement, according to an embodiment of this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0023] It should be noted that the terms "first," "second," etc., in the specification, claims, 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 apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] According to an embodiment of this application, a method embodiment of a large-model intelligent customer service method based on retrieval enhancement is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0025] Figure 1 This is a flowchart of a large-model intelligent customer service method based on retrieval enhancement according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0026] Step S101: Receive the query message input by the target object.

[0027] The query message may include information such as the total investment amount and investment type.

[0028] Step S102: Perform intent recognition on the query message. If it is determined that the intent of the query message is to obtain recommended financial products, send a follow-up message to the target object. The follow-up message is used to prompt the target object to input the preference weights of return, risk and term.

[0029] Follow-up messages can use a structured template, such as:

[0030] To provide you with accurate recommendations, please specify the weighting of the following preferences (total weighting is 100%):

[0031] 1. Profit requirement: ____%

[0032] 2. Risk tolerance: ____%

[0033] 3. Investment period: ____%

[0034] Please enter numbers (e.g., 50, 30, 20).

[0035] This can be achieved through the following simplified code:

[0036] if intent == "recommend financial products":

[0037] send_followup_message(template_id=3) # Trigger the preset follow-up question template

[0038] log_user_state("awaiting_preferences") # Marks the user's state as waiting for input.

[0039] Step S103: Receive the preference weights input by the target object, and determine the target user profile corresponding to the target object in the user profile database based on the target object's identification information.

[0040] The user profile database includes, but is not limited to, the following: Personal identity information: including the user's name, gender, age, ID number, and contact information. Occupational information: the user's occupation type (e.g., white-collar, blue-collar, freelancer, entrepreneur, etc.) and the nature of their employer. Income level: including monthly income, annual salary, and sources of income (salary, bonuses, investment income, rental income, etc.). Assets: covering the valuation and distribution of the user's real estate, vehicles, and financial assets (e.g., stocks, bonds, funds, bank deposits, etc.). Liabilities: mainly including the amount and repayment progress of the user's loans (mortgage, car loan, consumer loans, etc.). Investment experience: including the types of financial products the user has invested in in the past (e.g., stock investment duration, fund investment period, etc.), and successful and unsuccessful investment cases. Risk preference: for example, categorized as conservative, moderate, or aggressive. Investment goals: clarifying whether the user's investment is short-term (e.g., for purchasing electronics, travel, etc.) or long-term. Online behavioral data: Browsing history on financial institution websites or apps (e.g., which financial product pages were viewed, types of investment information followed, etc.), and search keywords (e.g., searching for "high-yield wealth management products"). Transaction behavior data: Includes user transaction records of financial products conducted at financial institutions (transaction amount, frequency, and channels). Product preferences: Records the degree of user preference for different financial products (e.g., funds, insurance, futures, etc.), including the types of financial products users actively collect, inquire about, or purchase. Product feedback: Includes user evaluations of purchased financial products (e.g., satisfaction level, suggestions for improvement, etc.).

[0041] Step S104: Analyze the target user profile, query messages, and preference weights using the vertical domain intent recognition model to obtain the first recommended financial product combination output by the vertical domain intent recognition model.

[0042] In step S104, the user profile includes, for example: static attributes: age, occupation, asset size, risk tolerance rating (e.g., R1-R5). Dynamic behaviors: historical investment records (product type, holding period, return performance), frequency of financial consultations. Query message text, for example: "Recommended stable funds" or "High-yield short-term financial products". Preference weights, for example: return: 50%, risk: 30%, term: 20%.

[0043] The query message is processed using a BERT model fine-tuned for the financial domain, outputting a text embedding vector. Numerical features of the user profile (such as age and assets) and categorical features (such as occupational codes) are concatenated and mapped to a profile vector through a fully connected layer. The return, risk, and term weight vectors are normalized and concatenated with the text embedding vector and the profile vector, then input into a cross-attention mechanism to generate a joint representation. The fully connected layer maps the joint representation to the matching score vector of the financial product library. A probability distribution is generated through Softmax, and the Top-K products are selected to form the first recommendation combination.

[0044] Step S105: Calculate the total investment amount and preference weights in the query message using a multi-objective optimization algorithm to obtain the second recommended financial product portfolio.

[0045] Step S106: Simultaneously send the first recommended financial product portfolio and the second recommended financial product portfolio to the target object, receive the target object's selection response message for the first recommended financial product portfolio and the second recommended financial product portfolio, and update the vertical domain intent recognition model and the target user profile based on the selection response message.

[0046] The selection response message is used to indicate whether the target object has selected the first recommended financial product portfolio and the second recommended financial product portfolio.

[0047] Based on the above steps, the system receives query messages from the target user; performs intent recognition on the query messages; and, if the intent of the query message is determined to be to obtain recommended financial products, sends follow-up messages to the target user, prompting them to input their preference weights for returns, risks, and terms. It receives the preference weights input by the target user and, based on the target user's identifier information, determines the target user profile corresponding to the target user in the user profile database. A vertical domain intent recognition model is used to analyze the target user profile, query messages, and preference weights to obtain a first recommended financial product combination output by the vertical domain intent recognition model. A multi-objective optimization algorithm is used to calculate the total investment amount and preference weights in the query messages to obtain a second recommended financial product combination. The first and second recommended financial product combinations are simultaneously sent to the target user, and the target user's selection response messages for the first and second recommended financial product combinations are received. Based on the selection response messages, the vertical domain intent recognition model and the target user profile are updated. This approach achieves the goal of accurately recommending suitable financial products to users based on their needs, thereby enhancing user stickiness.

[0048] The following are Figure 1 The steps shown are illustrated and explained by way of example.

[0049] According to some optional embodiments of this application, the first recommended financial product combination output by the vertical domain intent recognition model is obtained by analyzing the target user profile, query message, and preference weights. This can be achieved by the following method: extracting the target value corresponding to the preset keyword in the query message, encoding the preset keyword and the target value together as a numerical feature to obtain the first numerical feature; encoding the target user profile as a numerical feature to obtain the second numerical feature, and encoding the preference weight as a numerical feature to obtain the third numerical feature; weighting and concatenating the first, second, and third numerical features to obtain the fused feature; determining the matching score between the fused feature and the candidate products through a neural network ranking model, and determining the top n financial products with the highest matching scores as the first recommended financial product combination. The neural network ranking model is a dual-tower structure, including a first feature encoding tower and a second feature encoding tower. The first feature encoding tower is used to receive the fused feature and output the user representation vector, and the second feature encoding tower is used to receive the attribute features of the financial products and output the financial product representation vector; n is a positive integer greater than 1.

[0050] According to some optional embodiments of this application, a second recommended financial product portfolio is obtained by calculating the total investment amount and preference weights in the query message using a multi-objective optimization algorithm. This can be achieved through the following steps: Step S1, determine the target number of financial products to be recommended; Step S2, generate an initial population, where each individual in the initial population represents a different combination of financial products to be recommended, and each combination includes n financial products to be recommended; Step S3, calculate the objective function for each individual in the current population, where the objective function includes: maximizing a first function value representing return, minimizing a second function value representing risk, and minimizing a third function value representing term; Step S4, perform non-dominated sorting on the individuals, and based on the sorting results, select the current... The population is divided into different non-dominated levels; crowding distance is calculated for individuals in each non-dominated level, and individuals are selected from the current population to enter the breeding pool based on the crowding distance calculation results; crossover and mutation operations are performed on individuals in the breeding pool to generate offspring populations; in step S5, the initial population and offspring populations are merged into the target population; non-dominated sorting and crowding distance calculation are performed on the target population, and individuals whose non-dominated level and crowding distance meet the preset requirements are retained to form a new population; steps S3 to S4 are repeated until the maximum number of iterations is reached to obtain the final population; in step S6, the target individuals with the highest level in the non-dominated sorting are determined in the final population, and the investment ratio of each financial product to be recommended corresponding to the target individuals is determined as the second recommended financial product portfolio.

[0051] Step S4 can be implemented as follows: Using Pareto dominance relations, hierarchical levels are defined. Non-dominated solutions (those not completely outperformed by other solutions) are placed in the first level, suboptimal solutions in the second level, and so on. Within the same non-dominated level, the distribution density of each individual in the target space is calculated, prioritizing the retention of sparsely distributed individuals to maintain diversity. The population is filled from high to low non-dominated levels until the population capacity is reached. A cutoff point is randomly selected, and some products from the parent individuals are exchanged. A product is replaced with another product from the candidate pool with a preset probability.

[0052] Step S5 can be implemented as follows: merge the parent generation (e.g., 200 individuals) and the offspring generation (e.g., 200 individuals) into 400 individuals, and recalculate the non-dominated sorting and crowding. Select the top 200 individuals from the mixed population in descending order of non-dominated hierarchy and crowding as the new generation. Stop when the maximum number of iterations (e.g., 100 generations) is reached or the Pareto front change rate is less than a threshold (e.g., 0.1%).

[0053] In step S6, individuals with the highest crowding density can be selected from the non-dominated first layer of the final population (balancing diversity and excellence).

[0054] The above steps, through the adaptively improved NSGA-II algorithm, achieve the effect of accurately determining the second recommended financial product portfolio.

[0055] Specifically, the target number of financial products to be recommended can be determined as follows: Select the top m financial products with the highest matching scores in the neural network ranking model, and determine the top m financial products with the highest matching scores as the financial products to be recommended, where m is the target number and m is a positive integer greater than n.

[0056] It is worth noting that in the above embodiments, it is assumed that the neural network ranking model determines the top 5 financial products with the highest matching scores as the first recommended financial product portfolio. In the multi-objective optimization algorithm, the top 10 financial products with the highest matching scores are selected. Among these 10 financial products, a multi-objective optimization algorithm based on risk, return, and maturity is used to select the optimal 5 financial products. This is used as the second recommended financial product portfolio. The purpose is to verify the accuracy of the vertical domain intent recognition model in recommending financial products, so as to select financial products that better match the customer's investment preferences. In addition, the top m financial products with the highest matching scores are selected in the neural network ranking model. Determining the top m financial products with the highest matching scores as the starting parameters in the multi-objective optimization algorithm can simplify the calculation process of the multi-objective optimization algorithm and improve the computational efficiency.

[0057] Specifically, the first function value can be determined using the following formula. : in, Weight for expected returns Let i be the i-th combination of financial products to be recommended, and k be the number of combinations of financial products to be recommended, k = , Let i be the expected rate of return for product i, where i is a positive integer not greater than k; the value of the second function is determined by the following formula. : in, As the risk tolerance weight, Let i be the risk coefficient of product i; the value of the third function is determined by the following formula. : in, Let i be the risk coefficient of product i. Let D be the timeframe for product i, and D be the user's expected timeframe. The objective function also includes constraint penalty terms, which include: *C is less than the preset minimum investment amount for product i, where C is the total investment amount.

[0058] As some optional embodiments of this application, intent recognition of query messages can be achieved through the following methods: using an intent recognition model to identify the intent of the query message, obtaining the recognition result output by the intent recognition model, wherein the recognition result includes: the intent of the query message is to obtain recommended financial products, and the intent of the query message is not to obtain recommended financial products; the intent recognition model is trained through the following methods: acquiring a user query text dataset, labeling the user query text data in the user query text dataset, generating binary labels containing intents for recommended financial products and intents for non-recommended financial products, obtaining labeled user query text data; generating features corresponding to the labeled user query text data. The vector, wherein the feature vector includes at least one of the following: a term frequency-inverse document frequency vector generated by the term frequency-inverse document frequency algorithm, and a context semantic embedding vector generated based on a pre-trained language model; the feature vector is input into a neural network classifier including fully connected layers, and the activation function is used to output the prediction result, which is the probability value of the prediction result as the intention to recommend financial products; the relationship between the probability value and a preset threshold is judged, and when the probability value is greater than the preset threshold, it is judged as the intention to recommend financial products, and when the probability value is less than the preset threshold, it is judged as the intention not to recommend financial products; based on the error between the prediction result output by the neural network classifier and the labeled label, the model parameters of the neural network classifier are updated through the backpropagation algorithm to obtain the intention recognition model.

[0059] In some optional embodiments of this application, the user profile library is constructed by the following method: extracting user behavior feature sets from transaction flow, log traces, and historical session records, wherein the user behavior feature sets include at least: numerical features, time series features, and categorical features; standardizing and normalizing the numerical features using principal component analysis algorithm, extracting the latent vector representation of the time series features using long short-term memory autoencoder, and mapping the categorical features to a continuous vector space using one-hot encoding to obtain the target feature set; clustering the features in the target feature set based on density peaks to obtain clustering results, and constructing the user profile library based on the clustering results, wherein the clustering results carry user group labels.

[0060] Specifically, when building a user profile database, business tags can be defined based on common characteristics within clusters, such as: High-net-worth active users: high transaction frequency, large-amount consumption, VIP level; Risk-sensitive users: frequent balance inquiries, conversations containing keywords such as "security" and "risk control"; Long-term planning users: regular fixed-amount investments, focus on retirement financial products. Furthermore, user IDs, feature vectors, and cluster tags are written into the database, supporting real-time queries.

[0061] Figure 2 This is a structural diagram of a large-scale intelligent customer service system based on retrieval enhancement, according to an embodiment of this application. Figure 2As shown, the system includes:

[0062] The receiving module 21 is used to receive query messages input by the target object.

[0063] The identification module 22 is used to identify the intent of the query message. If it is determined that the intent of the query message is to obtain recommended financial products, a follow-up message is sent to the target object. The follow-up message is used to prompt the target object to input the preference weights of return, risk and term.

[0064] The determination module 23 is used to receive the preference weights input by the target object and determine the target user profile corresponding to the target object in the user profile database based on the identification information of the target object.

[0065] Analysis module 24 is used to analyze the target user profile, query messages and preference weights using the vertical domain intent recognition model to obtain the first recommended financial product combination output by the vertical domain intent recognition model.

[0066] The calculation module 25 is used to calculate the total investment amount and preference weight in the query message using a multi-objective optimization algorithm to obtain the second recommended financial product portfolio.

[0067] The optimization module 26 is used to simultaneously send the first recommended financial product portfolio and the second recommended financial product portfolio to the target object, receive the target object's selection response message for the first recommended financial product portfolio and the second recommended financial product portfolio, and update the vertical domain intent recognition model and the target user profile based on the selection response message.

[0068] Optionally, a vertical domain intent recognition model is used to analyze the target user profile, query message, and preference weights to obtain the first recommended financial product combination output by the vertical domain intent recognition model. This includes the following steps: extracting target values ​​corresponding to preset keywords from the query message; encoding the preset keywords and target values ​​together as numerical features to obtain the first numerical feature; encoding the target user profile as numerical features to obtain the second numerical feature; and encoding the preference weights as numerical features to obtain the third numerical feature; weighted concatenation of the first, second, and third numerical features to obtain a fused feature; and determining the matching score between the fused feature and candidate products using a neural network ranking model, and identifying the top n financial products with the highest matching scores as the first recommended financial product combination. The neural network ranking model is a dual-tower structure, including a first feature encoding tower and a second feature encoding tower. The first feature encoding tower receives the fused feature and outputs a user representation vector, while the second feature encoding tower receives the attribute features of the financial products and outputs a financial product representation vector; n is a positive integer greater than 1.

[0069] Optionally, a multi-objective optimization algorithm is used to calculate the total investment amount and preference weights in the query message to obtain a second recommended financial product portfolio, including the following steps: Step S1, determine the target number of financial products to be recommended; Step S2, generate an initial population, where each individual in the initial population represents a different combination of financial products to be recommended, and each combination includes n financial products to be recommended; Step S3, calculate the objective function for each individual in the current population, where the objective function includes: maximizing the first function value representing the return, minimizing the second function value representing the risk, and minimizing the third function value representing the term; Step S4, perform non-dominated ranking of the individuals, and divide the current population into different groups based on the ranking results. Step S3: For each non-dominated level, calculate the crowding distance for individuals and select individuals from the current population to enter the breeding pool based on the crowding distance calculation results; perform crossover and mutation operations on individuals in the breeding pool to generate offspring populations; Step S5: merge the initial population and offspring populations into the target population; perform non-dominated sorting and crowding distance calculation on the target population, and retain individuals whose non-dominated level and crowding distance meet the preset requirements to form a new population; repeat steps S3 to S4 until the maximum number of iterations is reached to obtain the final population; Step S6: determine the target individuals with the highest level in the non-dominated sorting in the final population, and determine the investment ratio of each financial product to be recommended corresponding to the target individuals as the second recommended financial product portfolio.

[0070] Optionally, determining the target number of financial products to be recommended includes the following steps: selecting the top m financial products with the highest matching scores in the neural network ranking model, and determining the top m financial products with the highest matching scores as financial products to be recommended, where m is the target number and m is a positive integer greater than n.

[0071] Optionally, the large-model intelligent customer service system based on retrieval enhancement is also used to perform the following steps: determining the value of the first function using the following formula. : in, Weight for expected returns Let i be the i-th combination of financial products to be recommended, and k be the number of combinations of financial products to be recommended, k = , Let i be the expected rate of return for product i, where i is a positive integer not greater than k; the value of the second function is determined by the following formula. : in, As the risk tolerance weight, Let i be the risk coefficient of product i; the value of the third function is determined by the following formula. : in, Let i be the risk coefficient of product i. Let D be the timeframe for product i, and D be the user's expected timeframe. The objective function also includes constraint penalty terms, which include: *C is less than the preset minimum investment amount for product i, where C is the total investment amount.

[0072] Optionally, intent recognition is performed on the query message, including: using an intent recognition model to recognize the intent of the query message, obtaining the recognition result output by the intent recognition model, wherein the recognition result includes: the intent of the query message is to obtain recommended financial products, and the intent of the query message is not to obtain recommended financial products; the intent recognition model is trained by the following method: acquiring a user query text dataset, labeling the user query text data in the user query text dataset, generating binary labels containing intent to recommend financial products and intent not to recommend financial products, obtaining labeled user query text data; generating feature vectors corresponding to the labeled user query text data, wherein the feature vectors... The parameters include at least one of the following: a term frequency-inverse document frequency vector generated by a term frequency-inverse document frequency algorithm, and a contextual semantic embedding vector generated based on a pre-trained language model; the feature vector is input into a neural network classifier including fully connected layers, and the activation function is used to output the prediction result, which is the probability value of the prediction result being the intention to recommend financial products; the relationship between the probability value and a preset threshold is judged, and when the probability value is greater than the preset threshold, it is determined to be the intention to recommend financial products, and when the probability value is less than the preset threshold, it is determined to be the intention not to recommend financial products; based on the error between the prediction result output by the neural network classifier and the labeled label, the model parameters of the neural network classifier are updated through a backpropagation algorithm to obtain the intention recognition model.

[0073] Optionally, the user profile database is constructed using the following method: extracting user behavior feature sets from transaction records, log traces, and historical session records, wherein the user behavior feature sets include at least: numerical features, time series features, and categorical features; standardizing and normalizing the numerical features using principal component analysis, extracting the latent vector representation of the time series features using a long short-term memory autoencoder, and mapping the categorical features to a continuous vector space using one-hot encoding to obtain the target feature set; clustering the features in the target feature set based on density peaks to obtain clustering results, and constructing the user profile database based on the clustering results, wherein the clustering results carry user group labels.

[0074] It should be noted that the above Figure 2 The modules in can be program modules (e.g., a set of program instructions that implements a specific function) or hardware modules. For the latter, they can be represented in the following forms, but are not limited to these: each of the above modules is represented by a processor, or the functions of each of the above modules are implemented by a processor.

[0075] It should be noted that, Figure 2 Preferred embodiments of the shown examples can be found in [reference needed]. Figure 1 The relevant descriptions of the embodiments shown will not be repeated here.

[0076] Figure 3 A hardware block diagram of a computer terminal for implementing a large-model intelligent customer service method based on retrieval enhancement is shown. Figure 3 As shown, the computer terminal 30 may include one or more processors 302 (shown as 302a, 302b, ..., 302n in the figure) 302 (processor 302 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 304 for storing data, and a transmission module 306 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 3 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 30 may also include... Figure 3 The more or fewer components shown, or having the same Figure 3 The different configurations shown.

[0077] It should be noted that the aforementioned one or more processors 302 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be wholly or partially embodied in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be wholly or partially integrated into any other element in the computer terminal 30. As involved in the embodiments of this application, the data processing circuits serve as processor control.

[0078] The memory 304 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the search-enhanced large-model intelligent customer service method in this embodiment. The processor 302 executes various functional applications and data processing by running the software programs and modules stored in the memory 304, thereby realizing the aforementioned search-enhanced large-model intelligent customer service method. The memory 304 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 304 may further include memory remotely located relative to the processor 302, and these remote memories can be connected to the computer terminal 30 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0079] The transmission module 306 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 30. In one example, the transmission module 306 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission module 306 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0080] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 30.

[0081] It should be noted here that, in some optional embodiments, the above... Figure 3 The computer terminal shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 3 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computer terminal.

[0082] It should be noted that, Figure 3 The computer terminal shown is used to execute Figure 1 The large-model intelligent customer service method based on retrieval enhancement shown above also applies to this electronic device, and will not be repeated here.

[0083] This application embodiment also provides a non-volatile storage medium, which includes a stored program, wherein the program, when running, controls the device where the storage medium is located to execute the above-mentioned large-model intelligent customer service method based on retrieval enhancement.

[0084] A non-volatile storage medium performs the following functions: receiving a query message input by a target object; performing intent recognition on the query message, and if the intent of the query message is determined to be to obtain recommended financial products, sending a follow-up message to the target object, wherein the follow-up message is used to prompt the target object to input preference weights for returns, risks, and terms; receiving the preference weights input by the target object, and determining the target user profile corresponding to the target object in the user profile database based on the target object's identification information; analyzing the target user profile, query message, and preference weights using a vertical domain intent recognition model to obtain a first recommended financial product portfolio output by the vertical domain intent recognition model; calculating the total investment amount and preference weights in the query message using a multi-objective optimization algorithm to obtain a second recommended financial product portfolio; simultaneously sending the first and second recommended financial product portfolios to the target object, receiving the target object's selection response messages for the first and second recommended financial product portfolios, and updating the vertical domain intent recognition model and the target user profile based on the selection response messages.

[0085] This application also provides an electronic device, including: a memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the program executes the above-described large model intelligent customer service method based on retrieval enhancement.

[0086] The processor runs a program that performs the following functions: receiving a query message input by a target object; performing intent recognition on the query message, and if the intent of the query message is determined to be to obtain recommended financial products, sending a follow-up message to the target object, wherein the follow-up message is used to prompt the target object to input preference weights for returns, risks, and terms; receiving the preference weights input by the target object, and determining the target user profile corresponding to the target object in the user profile database based on the target object's identification information; analyzing the target user profile, query message, and preference weights using a vertical domain intent recognition model to obtain a first recommended financial product portfolio output by the vertical domain intent recognition model; calculating the total investment amount and preference weights in the query message using a multi-objective optimization algorithm to obtain a second recommended financial product portfolio; simultaneously sending the first and second recommended financial product portfolios to the target object, receiving the target object's selection response messages for the first and second recommended financial product portfolios, and updating the vertical domain intent recognition model and the target user profile based on the selection response messages.

[0087] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0088] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0089] In the above embodiments of this application, the information collected is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with relevant laws, regulations and standards, take necessary protective measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0090] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0091] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0092] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0093] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to related technologies, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0094] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A large-scale intelligent customer service method based on retrieval enhancement, characterized in that, include: Receive query messages input by the target object; The query message is subjected to intent recognition. If it is determined that the intent of the query message is to obtain recommended financial products, a follow-up message is sent to the target object. The follow-up message is used to prompt the target object to input the preference weights of return, risk and term. The system receives the preference weights input by the target object and determines the target user profile corresponding to the target object in the user profile database based on the identification information of the target object. By analyzing the target user profile, the query message, and the preference weight using a vertical domain intent recognition model, the first recommended financial product combination output by the vertical domain intent recognition model is obtained. A second recommended financial product portfolio is obtained by using a multi-objective optimization algorithm to calculate the total investment amount and the preference weight in the query message; The first recommended financial product portfolio and the second recommended financial product portfolio are sent to the target object simultaneously. The target object's selection response message for the first recommended financial product portfolio and the second recommended financial product portfolio is received. Based on the selection response message, the vertical domain intent recognition model and the target user profile are updated. The vertical domain intent recognition model is used to analyze the target user profile, the query message, and the preference weights to obtain a first recommended financial product combination output by the vertical domain intent recognition model. This includes: extracting target values ​​corresponding to preset keywords from the query message; encoding the preset keywords and the target values ​​together as numerical features to obtain a first numerical feature; encoding the target user profile as numerical features to obtain a second numerical feature; and encoding the preference weights as numerical features to obtain a third numerical feature; weighted concatenation of the first, second, and third numerical features to obtain a fusion feature; and determining the matching score between the fusion feature and candidate products using a neural network ranking model, and identifying the top n financial products with the highest matching scores as the first recommended financial product combination. The neural network ranking model is a dual-tower structure, including a first feature encoding tower and a second feature encoding tower. The first feature encoding tower receives the fusion feature and outputs a user representation vector, while the second feature encoding tower receives the attribute features of the financial products and outputs a financial product representation vector; n is a positive integer greater than 1.

2. The method according to claim 1, characterized in that, A second recommended financial product portfolio is obtained by using a multi-objective optimization algorithm to calculate the total investment amount and preference weights in the query message, including: Step S1: Determine the target number of financial products to be recommended; Step S2: Generate an initial population, wherein each individual in the initial population represents a different combination of financial products to be recommended, and each combination includes n financial products to be recommended; Step S3: Calculate the objective function for each individual in the current population, wherein the objective function includes: maximizing a first function value representing the return, minimizing a second function value representing the risk, and minimizing a third function value representing the duration; Step S4: Perform non-dominated sorting on individuals, and divide the current population into different non-dominated levels based on the sorting results; calculate the crowding distance for individuals in each non-dominated level, and select individuals from the current population to enter the breeding pool based on the crowding distance calculation results; perform crossover and mutation operations on the individuals in the breeding pool to generate offspring population. Step S5: Merge the initial population and the offspring population into a target population; perform non-dominated sorting and crowding distance calculation on the target population, and retain individuals whose non-dominated level and crowding distance meet the preset requirements to form a new population; repeat steps S3 to S4 until the maximum number of iterations is reached to obtain the final population. Step S6: In the final population, determine the target individual with the highest rank in the non-dominated ranking, and determine the investment ratio of each of the financial products to be recommended corresponding to the target individual as the second recommended financial product portfolio.

3. The method according to claim 2, characterized in that, Determining the target number of financial products to be recommended includes: selecting the top m financial products with the highest matching scores in the neural network ranking model, and determining the top m financial products with the highest matching scores as the financial products to be recommended, where m is the target number and m is a positive integer greater than n.

4. The method according to claim 1, characterized in that, Intent identification of the query message includes: The intent of the query message is identified using an intent recognition model, and the recognition result output by the intent recognition model is obtained. The recognition result includes: the intent of the query message is to obtain recommended financial products, and the intent of the query message is not to obtain recommended financial products. The intent recognition model is trained by the following method: acquiring a user query text dataset, labeling the user query text data in the user query text dataset, generating binary labels containing intent to recommend financial products and intent not to recommend financial products, and obtaining labeled user query text data. Generate feature vectors corresponding to the labeled user query text data, wherein the feature vectors include at least one of the following: a term frequency-inverse document frequency vector generated by the term frequency-inverse document frequency algorithm, and a context semantic embedding vector generated based on a pre-trained language model; The feature vector is input into a neural network classifier including fully connected layers, and an activation function is used to output a prediction result, which is a probability value of the intention to recommend financial products. The relationship between the probability value and a preset threshold is determined. When the probability value is greater than the preset threshold, it is determined to be an intention to recommend financial products. When the probability value is less than the preset threshold, it is determined to be an intention not to recommend financial products. Based on the error between the prediction result output by the neural network classifier and the labeled data, the model parameters of the neural network classifier are updated through the backpropagation algorithm to obtain the intent recognition model.

5. The method according to claim 1, characterized in that, The user profile database is constructed using the following method: User behavior feature sets are extracted from transaction records, log traces, and historical session records, wherein the user behavior feature sets include at least: numerical features, time series features, and categorical features; The numerical features are standardized and normalized using principal component analysis, the latent vector representation of the time series features is extracted using a long short-term memory autoencoder, and the categorical features are mapped to a continuous vector space using one-hot encoding to obtain the target feature set. Clustering is performed on the features in the target feature set based on density peaks to obtain clustering results. Based on the clustering results, the user profile database is constructed, wherein the clustering results carry user group labels.

6. A large-scale intelligent customer service system based on retrieval enhancement, characterized in that, include: The receiving module is used to receive query messages input by the target object; The identification module is used to identify the intent of the query message. If it is determined that the intent of the query message is to obtain recommended financial products, a follow-up message is sent to the target object. The follow-up message is used to prompt the target object to input the preference weights of return, risk and term. The determination module is used to receive the preference weights input by the target object, and determine the target user profile corresponding to the target object in the user profile database based on the identification information of the target object; The analysis module is used to analyze the target user profile, the query message, and the preference weight using a vertical domain intent recognition model to obtain the first recommended financial product combination output by the vertical domain intent recognition model. The calculation module is used to calculate the total investment amount and the preference weight in the query message using a multi-objective optimization algorithm to obtain a second recommended financial product portfolio; The optimization module is used to simultaneously send the first recommended financial product portfolio and the second recommended financial product portfolio to the target object, receive the target object's selection response message for the first recommended financial product portfolio and the second recommended financial product portfolio, and update the vertical domain intent recognition model and the target user profile based on the selection response message. The analysis module is further configured to perform the following steps: extracting target values ​​corresponding to preset keywords from the query message; encoding the preset keywords and the target values ​​together as numerical features to obtain a first numerical feature; encoding the target user profile as a numerical feature to obtain a second numerical feature; and encoding the preference weight as a numerical feature to obtain a third numerical feature; weighting and concatenating the first numerical feature, the second numerical feature, and the third numerical feature to obtain a fusion feature; determining the matching score between the fusion feature and the candidate products through a neural network ranking model, and determining the top n financial products with the highest matching scores as the first recommended financial product combination, wherein the neural network ranking model is a dual-tower structure, including: a first feature encoding tower and a second feature encoding tower, wherein the first feature encoding tower is used to receive the fusion feature and output a user representation vector, and the second feature encoding tower is used to receive the attribute features of the financial products and output a financial product representation vector; n is a positive integer greater than 1.

7. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein, when the program is executed, it controls the device where the non-volatile storage medium is located to execute the large-model intelligent customer service method based on retrieval enhancement as described in any one of claims 1 to 5.

8. An electronic device, characterized in that, include: A memory and a processor, the processor being configured to run a program stored in the memory, wherein the program, when running, executes the large-model intelligent customer service method based on retrieval enhancement as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the large-model intelligent customer service method based on retrieval enhancement as described in any one of claims 1 to 5.

Images